diff --git a/.ci/windows_amd_base_files/README_VERY_IMPORTANT.txt b/.ci/windows_amd_base_files/README_VERY_IMPORTANT.txt index 2cbb00d99..2c72c8a13 100755 --- a/.ci/windows_amd_base_files/README_VERY_IMPORTANT.txt +++ b/.ci/windows_amd_base_files/README_VERY_IMPORTANT.txt @@ -1,5 +1,4 @@ -As of the time of writing this you need this driver for best results: -https://www.amd.com/en/resources/support-articles/release-notes/RN-AMDGPU-WINDOWS-PYTORCH-7-1-1.html +As of the time of writing this you need a recent driver. Updating to the latest driver is recommended. HOW TO RUN: @@ -7,9 +6,9 @@ If you have a AMD gpu: run_amd_gpu.bat -If you have memory issues you can try disabling the smart memory management by running comfyui with: +If you have memory issues you can try enabling the new dynamic memory management by running comfyui with: -run_amd_gpu_disable_smart_memory.bat +run_amd_gpu_enable_dynamic_vram.bat IF YOU GET A RED ERROR IN THE UI MAKE SURE YOU HAVE A MODEL/CHECKPOINT IN: ComfyUI\models\checkpoints diff --git a/.github/workflows/backport_release.yaml b/.github/workflows/backport_release.yaml new file mode 100644 index 000000000..ede6bde33 --- /dev/null +++ b/.github/workflows/backport_release.yaml @@ -0,0 +1,519 @@ +name: Backport Release + +on: + workflow_dispatch: + inputs: + commit: + description: 'Full 40-char SHA of the tip commit of the backport source branch (the PR head commit that passed tests). The branch is resolved from this SHA and must be unique.' + required: true + type: string + +permissions: + contents: read + pull-requests: read + checks: read + +jobs: + backport-release: + name: Create backport release + runs-on: ubuntu-latest + environment: backport release + + steps: + - name: Generate GitHub App token + id: app-token + uses: actions/create-github-app-token@bcd2ba49218906704ab6c1aa796996da409d3eb1 + with: + app-id: ${{ secrets.FEN_RELEASE_APP_ID }} + private-key: ${{ secrets.FEN_RELEASE_PRIVATE_KEY }} + + - name: Checkout repository + uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd + with: + token: ${{ steps.app-token.outputs.token }} + fetch-depth: 0 + fetch-tags: true + + - name: Configure git + run: | + git config user.name "fen-release[bot]" + git config user.email "fen-release[bot]@users.noreply.github.com" + + - name: Resolve source branch from commit SHA + id: resolve + env: + SOURCE_COMMIT: ${{ inputs.commit }} + DEFAULT_BRANCH: ${{ github.event.repository.default_branch }} + run: | + set -euo pipefail + + # Require a full 40-char lowercase-hex SHA. Short SHAs are ambiguous + # and we will be comparing this value against API responses (PR head + # SHA, ref tips) that always return the full form. + if [[ ! "${SOURCE_COMMIT}" =~ ^[0-9a-f]{40}$ ]]; then + echo "::error::Input commit '${SOURCE_COMMIT}' is not a full 40-char lowercase hex SHA." + exit 1 + fi + + # Fetch all remote branches so we can search for which one(s) point + # at this SHA. `actions/checkout` with fetch-depth: 0 fetches full + # history of the checked-out ref but does not necessarily populate + # every refs/remotes/origin/*, so do it explicitly. + git fetch --prune origin '+refs/heads/*:refs/remotes/origin/*' + + # Verify the commit actually exists in this repo's object DB. + if ! git cat-file -e "${SOURCE_COMMIT}^{commit}" 2>/dev/null; then + echo "::error::Commit ${SOURCE_COMMIT} was not found in the repository." + exit 1 + fi + + # Find every remote branch whose tip == SOURCE_COMMIT. Exactly one + # branch must point at it. If zero, the commit isn't anyone's tip + # (likely stale, force-pushed past, or never the PR head). If more + # than one, the (branch -> SHA) mapping is ambiguous and we refuse + # to guess — the operator must give us a unique branch to release. + mapfile -t matching_branches < <( + git for-each-ref \ + --format='%(refname:strip=3)' \ + --points-at="${SOURCE_COMMIT}" \ + refs/remotes/origin/ \ + | grep -vx 'HEAD' || true + ) + + if [[ "${#matching_branches[@]}" -eq 0 ]]; then + echo "::error::No branch on origin has ${SOURCE_COMMIT} as its tip." + echo "::error::Either the branch was updated after you copied this SHA, or this commit was never the head of a branch." + exit 1 + fi + + if [[ "${#matching_branches[@]}" -gt 1 ]]; then + echo "::error::More than one branch on origin has ${SOURCE_COMMIT} as its tip; cannot pick one:" + for b in "${matching_branches[@]}"; do + echo "::error:: - ${b}" + done + echo "::error::Refusing to proceed with an ambiguous source branch." + exit 1 + fi + + source_branch="${matching_branches[0]}" + + if [[ "${source_branch}" == "${DEFAULT_BRANCH}" ]]; then + echo "::error::Source branch must not be the default branch ('${DEFAULT_BRANCH}')." + exit 1 + fi + + echo "Resolved commit ${SOURCE_COMMIT} to branch '${source_branch}'." + echo "source_branch=${source_branch}" >> "$GITHUB_OUTPUT" + + - name: Determine latest stable release + id: latest + env: + GH_TOKEN: ${{ steps.app-token.outputs.token }} + run: | + set -euo pipefail + + # List all tags matching vMAJOR.MINOR.PATCH and pick the highest by numeric + # comparison of each component. We DO NOT use `sort -V` because it treats + # v0.19.99 as higher than v0.20.1. + latest_tag="$( + git tag --list 'v[0-9]*.[0-9]*.[0-9]*' \ + | grep -E '^v[0-9]+\.[0-9]+\.[0-9]+$' \ + | awk -F'[v.]' '{ printf "%010d %010d %010d %s\n", $2, $3, $4, $0 }' \ + | sort -k1,1n -k2,2n -k3,3n \ + | tail -n1 \ + | awk '{print $4}' + )" + + if [[ -z "${latest_tag}" ]]; then + echo "::error::No stable release tags (vMAJOR.MINOR.PATCH) were found." + exit 1 + fi + + # Parse components + ver="${latest_tag#v}" + major="${ver%%.*}" + rest="${ver#*.}" + minor="${rest%%.*}" + patch="${rest#*.}" + + new_patch=$((patch + 1)) + new_version="v${major}.${minor}.${new_patch}" + release_branch="release/v${major}.${minor}" + + latest_sha="$(git rev-list -n 1 "refs/tags/${latest_tag}")" + + echo "latest_tag=${latest_tag}" >> "$GITHUB_OUTPUT" + echo "latest_sha=${latest_sha}" >> "$GITHUB_OUTPUT" + echo "major=${major}" >> "$GITHUB_OUTPUT" + echo "minor=${minor}" >> "$GITHUB_OUTPUT" + echo "patch=${patch}" >> "$GITHUB_OUTPUT" + echo "new_version=${new_version}" >> "$GITHUB_OUTPUT" + echo "new_version_no_v=${major}.${minor}.${new_patch}" >> "$GITHUB_OUTPUT" + echo "release_branch=${release_branch}" >> "$GITHUB_OUTPUT" + + echo "Latest stable release: ${latest_tag} (${latest_sha})" + echo "New version will be: ${new_version}" + echo "Release branch: ${release_branch}" + + - name: Validate source branch is cut directly from the latest stable release + env: + SOURCE_BRANCH: ${{ steps.resolve.outputs.source_branch }} + SOURCE_COMMIT: ${{ inputs.commit }} + LATEST_TAG_SHA: ${{ steps.latest.outputs.latest_sha }} + LATEST_TAG: ${{ steps.latest.outputs.latest_tag }} + run: | + set -euo pipefail + + # Use the user-provided SHA directly rather than re-resolving the branch + # tip — the resolve step already proved the branch tip equals SOURCE_COMMIT, + # and pinning to the SHA here makes the rest of the job TOCTOU-safe against + # someone pushing to the branch mid-run. + source_sha="${SOURCE_COMMIT}" + + # Walking first-parent from the source tip must reach LATEST_TAG_SHA. + # We capture rev-list into a variable and grep against a here-string + # rather than piping `rev-list | grep -q`: under `set -o pipefail`, + # `grep -q` would exit on first match and SIGPIPE the still-streaming + # `rev-list`, propagating exit 141 as a spurious "not found". + first_parent_chain="$(git rev-list --first-parent "${source_sha}")" + if ! grep -Fxq "${LATEST_TAG_SHA}" <<< "${first_parent_chain}"; then + echo "::error::Source branch '${SOURCE_BRANCH}' is not cut from '${LATEST_TAG}'." + echo "::error::Its first-parent history does not include ${LATEST_TAG_SHA}." + exit 1 + fi + + # Additionally, every commit added on top of the tag (the set we are + # about to publish) must itself be a descendant of the tag along + # first-parent — i.e. no sibling commits from master sneak in via a + # non-first-parent path. Enforce by requiring that the symmetric + # difference is empty in one direction: commits in source that are + # NOT first-parent-reachable from source starting at the tag. + # We do this by intersecting: + # A = commits reachable from source but not from tag (full DAG) + # B = commits on the first-parent chain from source down to tag + # and requiring A == B. + all_added="$(git rev-list "${LATEST_TAG_SHA}..${source_sha}" | sort)" + first_parent_added="$( + git rev-list --first-parent "${LATEST_TAG_SHA}..${source_sha}" | sort + )" + + if [[ "${all_added}" != "${first_parent_added}" ]]; then + echo "::error::Source branch '${SOURCE_BRANCH}' contains commits not on its first-parent chain from '${LATEST_TAG}'." + echo "::error::This usually means the branch was cut from master (not from the tag) or contains a merge from master." + echo "Commits reachable but not on first-parent chain:" + comm -23 <(printf '%s\n' "${all_added}") <(printf '%s\n' "${first_parent_added}") \ + | while read -r sha; do + echo " $(git log -1 --format='%h %s' "${sha}")" + done + exit 1 + fi + + added_count="$(printf '%s\n' "${all_added}" | grep -c . || true)" + echo "Source branch is cut directly from ${LATEST_TAG} with ${added_count} commit(s) on top." + + - name: Validate PR exists, is open, named correctly, has latest commit, and checks pass + env: + GH_TOKEN: ${{ steps.app-token.outputs.token }} + SOURCE_BRANCH: ${{ steps.resolve.outputs.source_branch }} + SOURCE_COMMIT: ${{ inputs.commit }} + NEW_VERSION: ${{ steps.latest.outputs.new_version }} + REPO: ${{ github.repository }} + run: | + set -euo pipefail + + expected_title="ComfyUI backport release ${NEW_VERSION}" + + # Find open PRs from this branch into master. The --state open filter + # is load-bearing: a closed/merged PR with passing checks must not be + # accepted as authorization for a new release. + pr_json="$( + gh pr list \ + --repo "${REPO}" \ + --state open \ + --head "${SOURCE_BRANCH}" \ + --base master \ + --json number,title,headRefOid,state \ + --limit 10 + )" + + pr_count="$(echo "${pr_json}" | jq 'length')" + if [[ "${pr_count}" -eq 0 ]]; then + echo "::error::No open PR found from '${SOURCE_BRANCH}' into 'master'. The PR must exist and be open." + exit 1 + fi + + # Pick the PR matching the expected title + pr_number="$(echo "${pr_json}" | jq -r --arg t "${expected_title}" ' + map(select(.title == $t)) | .[0].number // empty + ')" + pr_head_sha="$(echo "${pr_json}" | jq -r --arg t "${expected_title}" ' + map(select(.title == $t)) | .[0].headRefOid // empty + ')" + + if [[ -z "${pr_number}" ]]; then + echo "::error::No open PR from '${SOURCE_BRANCH}' into 'master' is titled '${expected_title}'." + echo "Found PRs:" + echo "${pr_json}" | jq -r '.[] | " #\(.number): \(.title)"' + exit 1 + fi + + # The PR's current head commit must equal the SHA the operator gave us. + # This is what closes the door on releasing stale code: if anyone has + # pushed to the branch since the operator validated tests passed, the + # PR head will have advanced past SOURCE_COMMIT and we abort. (The + # resolve step already proved the branch tip == SOURCE_COMMIT; this + # ties that same SHA to the PR that authorizes the release.) + if [[ "${pr_head_sha}" != "${SOURCE_COMMIT}" ]]; then + echo "::error::PR #${pr_number} head commit is ${pr_head_sha}, but the operator-provided commit is ${SOURCE_COMMIT}." + echo "::error::The PR has new commits since this release was authorized. Re-run with the new head SHA after verifying its checks." + exit 1 + fi + + echo "Found open PR #${pr_number} titled '${expected_title}' at head ${pr_head_sha} (matches operator-provided commit)." + + # Verify all check runs on the head commit have completed successfully. + # A check is considered passing if conclusion is success, neutral, or skipped. + checks_json="$( + gh api \ + --paginate \ + "repos/${REPO}/commits/${pr_head_sha}/check-runs" \ + --jq '.check_runs[] | {name: .name, status: .status, conclusion: .conclusion}' + )" + + if [[ -z "${checks_json}" ]]; then + echo "::error::No check runs found on PR head commit ${pr_head_sha}." + exit 1 + fi + + echo "Check runs on ${pr_head_sha}:" + echo "${checks_json}" | jq -s '.' + + failing="$(echo "${checks_json}" | jq -s ' + map(select( + .status != "completed" + or (.conclusion as $c + | ["success","neutral","skipped"] + | index($c) | not) + )) + ')" + + failing_count="$(echo "${failing}" | jq 'length')" + if [[ "${failing_count}" -gt 0 ]]; then + echo "::error::One or more checks have not passed on PR head commit ${pr_head_sha}:" + echo "${failing}" | jq -r '.[] | " - \(.name): status=\(.status) conclusion=\(.conclusion)"' + exit 1 + fi + + echo "All checks have passed on ${pr_head_sha}." + + - name: Prepare release branch + id: prepare + env: + GH_TOKEN: ${{ steps.app-token.outputs.token }} + REPO: ${{ github.repository }} + RELEASE_BRANCH: ${{ steps.latest.outputs.release_branch }} + LATEST_TAG: ${{ steps.latest.outputs.latest_tag }} + LATEST_TAG_SHA: ${{ steps.latest.outputs.latest_sha }} + PATCH: ${{ steps.latest.outputs.patch }} + run: | + set -euo pipefail + + # Try to fetch the release branch. If patch == 0, it shouldn't exist yet + # and we'll create it from the latest stable tag. If patch > 0, it must + # already exist and its tip must equal the latest stable tag commit (i.e. + # the previous patch release). + if git ls-remote --exit-code --heads origin "${RELEASE_BRANCH}" >/dev/null 2>&1; then + echo "Release branch '${RELEASE_BRANCH}' already exists on origin." + git fetch origin "refs/heads/${RELEASE_BRANCH}:refs/remotes/origin/${RELEASE_BRANCH}" + git checkout -B "${RELEASE_BRANCH}" "refs/remotes/origin/${RELEASE_BRANCH}" + + current_tip="$(git rev-parse HEAD)" + if [[ "${current_tip}" != "${LATEST_TAG_SHA}" ]]; then + echo "::error::Release branch '${RELEASE_BRANCH}' tip (${current_tip}) is not at the latest stable release '${LATEST_TAG}' (${LATEST_TAG_SHA})." + echo "::error::Refusing to release on top of a divergent branch." + exit 1 + fi + echo "branch_existed=true" >> "$GITHUB_OUTPUT" + else + if [[ "${PATCH}" != "0" ]]; then + echo "::error::Release branch '${RELEASE_BRANCH}' does not exist on origin, but the latest stable release '${LATEST_TAG}' has patch=${PATCH} (>0). This is inconsistent." + exit 1 + fi + echo "Release branch '${RELEASE_BRANCH}' does not exist. Creating from ${LATEST_TAG}." + git checkout -B "${RELEASE_BRANCH}" "refs/tags/${LATEST_TAG}" + echo "branch_existed=false" >> "$GITHUB_OUTPUT" + fi + + - name: Fast-forward merge source branch into release branch + env: + SOURCE_BRANCH: ${{ steps.resolve.outputs.source_branch }} + SOURCE_COMMIT: ${{ inputs.commit }} + RELEASE_BRANCH: ${{ steps.latest.outputs.release_branch }} + run: | + set -euo pipefail + + # --ff-only guarantees no merge commit is created. If a fast-forward is + # not possible (i.e. the release branch has commits the source branch + # doesn't), the merge will fail and we abort. Because we already validated + # that the source branch is rooted on the latest stable tag, and the + # release branch tip equals that same tag, this fast-forward should + # always succeed for a well-formed backport branch. + # + # We merge the operator-provided SHA, not the branch ref, so a push to + # the branch in the window between resolve and now cannot smuggle new + # commits into the release. + if ! git merge --ff-only "${SOURCE_COMMIT}"; then + echo "::error::Cannot fast-forward '${RELEASE_BRANCH}' to ${SOURCE_COMMIT} (tip of '${SOURCE_BRANCH}'). A merge commit would be required. Aborting." + exit 1 + fi + + echo "Fast-forwarded '${RELEASE_BRANCH}' to ${SOURCE_COMMIT} (tip of '${SOURCE_BRANCH}')." + + - name: Bump version files + env: + NEW_VERSION_NO_V: ${{ steps.latest.outputs.new_version_no_v }} + run: | + set -euo pipefail + + if [[ ! -f comfyui_version.py ]]; then + echo "::error::comfyui_version.py not found in repo root." + exit 1 + fi + if [[ ! -f pyproject.toml ]]; then + echo "::error::pyproject.toml not found in repo root." + exit 1 + fi + + # Replace the version string in comfyui_version.py. + # Expected format: __version__ = "X.Y.Z" + python3 - "$NEW_VERSION_NO_V" <<'PY' + import re, sys, pathlib + new = sys.argv[1] + + p = pathlib.Path("comfyui_version.py") + src = p.read_text() + new_src, n = re.subn( + r'(__version__\s*=\s*[\'"])[^\'"]+([\'"])', + lambda m: f'{m.group(1)}{new}{m.group(2)}', + src, + count=1, + ) + if n != 1: + sys.exit("Could not find __version__ assignment in comfyui_version.py") + p.write_text(new_src) + + p = pathlib.Path("pyproject.toml") + src = p.read_text() + # Replace the first `version = "..."` inside [project] or [tool.poetry]. + new_src, n = re.subn( + r'(?m)^(version\s*=\s*")[^"]+(")', + lambda m: f'{m.group(1)}{new}{m.group(2)}', + src, + count=1, + ) + if n != 1: + sys.exit("Could not find version assignment in pyproject.toml") + p.write_text(new_src) + PY + + echo "Updated version to ${NEW_VERSION_NO_V} in comfyui_version.py and pyproject.toml." + git --no-pager diff -- comfyui_version.py pyproject.toml + + - name: Commit version bump and tag release + env: + NEW_VERSION: ${{ steps.latest.outputs.new_version }} + run: | + set -euo pipefail + + git add comfyui_version.py pyproject.toml + git commit -m "ComfyUI ${NEW_VERSION}" + + if git rev-parse -q --verify "refs/tags/${NEW_VERSION}" >/dev/null; then + echo "::error::Tag ${NEW_VERSION} already exists locally." + exit 1 + fi + git tag "${NEW_VERSION}" + + - name: Verify tag does not already exist on origin + env: + NEW_VERSION: ${{ steps.latest.outputs.new_version }} + run: | + set -euo pipefail + if git ls-remote --exit-code --tags origin "refs/tags/${NEW_VERSION}" >/dev/null 2>&1; then + echo "::error::Tag ${NEW_VERSION} already exists on origin. Aborting." + exit 1 + fi + + - name: Push release branch and tag + env: + RELEASE_BRANCH: ${{ steps.latest.outputs.release_branch }} + NEW_VERSION: ${{ steps.latest.outputs.new_version }} + run: | + set -euo pipefail + + # Push the branch first, then the tag. Atomic-ish: if the branch push + # fails we never publish the tag. + git push origin "refs/heads/${RELEASE_BRANCH}:refs/heads/${RELEASE_BRANCH}" + git push origin "refs/tags/${NEW_VERSION}" + + echo "Released ${NEW_VERSION} on ${RELEASE_BRANCH}." + + - name: Delete remote source branch + env: + GH_TOKEN: ${{ steps.app-token.outputs.token }} + REPO: ${{ github.repository }} + SOURCE_BRANCH: ${{ steps.resolve.outputs.source_branch }} + SOURCE_COMMIT: ${{ inputs.commit }} + RELEASE_BRANCH: ${{ steps.latest.outputs.release_branch }} + DEFAULT_BRANCH: ${{ github.event.repository.default_branch }} + run: | + set -euo pipefail + + # Belt-and-braces: the resolve step already refuses the default branch, + # but never delete the default or the release branch under any + # circumstances. + if [[ "${SOURCE_BRANCH}" == "${DEFAULT_BRANCH}" || "${SOURCE_BRANCH}" == "${RELEASE_BRANCH}" ]]; then + echo "::error::Refusing to delete '${SOURCE_BRANCH}' (matches default or release branch)." + exit 1 + fi + + # Delete the source branch on origin, but only if its tip is still the + # SHA we released from. If someone pushed new commits to it after we + # resolved it, leave it alone — those commits would be silently lost. + current_tip="$(git ls-remote origin "refs/heads/${SOURCE_BRANCH}" | awk '{print $1}')" + if [[ -z "${current_tip}" ]]; then + echo "Source branch '${SOURCE_BRANCH}' no longer exists on origin; nothing to delete." + exit 0 + fi + if [[ "${current_tip}" != "${SOURCE_COMMIT}" ]]; then + echo "::warning::Source branch '${SOURCE_BRANCH}' tip (${current_tip}) no longer matches released commit (${SOURCE_COMMIT}). Leaving it in place." + exit 0 + fi + + git push origin --delete "refs/heads/${SOURCE_BRANCH}" + echo "Deleted remote branch '${SOURCE_BRANCH}'." + + - name: Summary + if: always() + env: + NEW_VERSION: ${{ steps.latest.outputs.new_version }} + RELEASE_BRANCH: ${{ steps.latest.outputs.release_branch }} + LATEST_TAG: ${{ steps.latest.outputs.latest_tag }} + SOURCE_BRANCH: ${{ steps.resolve.outputs.source_branch }} + SOURCE_COMMIT: ${{ inputs.commit }} + run: | + # SOURCE_BRANCH is empty if the resolve step never produced an output + # (e.g. the workflow failed in or before that step). Show a placeholder + # in that case so the summary table still renders cleanly. + source_branch_display="${SOURCE_BRANCH:-(unresolved)}" + { + echo "## Backport release" + echo "" + echo "| Field | Value |" + echo "|---|---|" + echo "| Source commit | \`${SOURCE_COMMIT}\` |" + echo "| Source branch | \`${source_branch_display}\` |" + echo "| Previous stable | \`${LATEST_TAG}\` |" + echo "| New version | \`${NEW_VERSION}\` |" + echo "| Release branch | \`${RELEASE_BRANCH}\` |" + } >> "$GITHUB_STEP_SUMMARY" diff --git a/.github/workflows/check-line-endings.yml b/.github/workflows/check-line-endings.yml index eeb594d6c..a69a24a87 100644 --- a/.github/workflows/check-line-endings.yml +++ b/.github/workflows/check-line-endings.yml @@ -17,7 +17,7 @@ jobs: - name: Check for Windows line endings (CRLF) run: | # Get the list of changed files in the PR - CHANGED_FILES=$(git diff --name-only ${{ github.event.pull_request.base.sha }}..${{ github.event.pull_request.head.sha }}) + CHANGED_FILES=$(git diff --name-only ${{ github.event.pull_request.base.sha }}..${{ github.event.pull_request.head.sha }} -- ':!.ci') # Flag to track if CRLF is found CRLF_FOUND=false diff --git a/.github/workflows/detect-unreviewed-merge.yml b/.github/workflows/detect-unreviewed-merge.yml new file mode 100644 index 000000000..4fabecb94 --- /dev/null +++ b/.github/workflows/detect-unreviewed-merge.yml @@ -0,0 +1,24 @@ +name: Detect Unreviewed Merge + +# SOC 2 compliance — reusable workflow lives in Comfy-Org/github-workflows, +# tracking issues are filed in Comfy-Org/unreviewed-merges. + +on: + push: + branches: [master] + +concurrency: + group: detect-unreviewed-merge-${{ github.sha }} + cancel-in-progress: false + +permissions: + contents: read + pull-requests: read + +jobs: + detect: + uses: Comfy-Org/github-workflows/.github/workflows/detect-unreviewed-merge.yml@4d9cb6b87f953bb7cd69954280e1465fb9bd2040 # v1 + with: + approval-mode: latest-per-reviewer + secrets: + UNREVIEWED_MERGES_TOKEN: ${{ secrets.UNREVIEWED_MERGES_TOKEN }} diff --git a/README.md b/README.md index 0eecd8a4b..dc2389266 100644 --- a/README.md +++ b/README.md @@ -20,7 +20,7 @@ [website-url]: https://www.comfy.org/ [discord-shield]: https://img.shields.io/badge/dynamic/json?url=https%3A%2F%2Fdiscord.com%2Fapi%2Finvites%2Fcomfyorg%3Fwith_counts%3Dtrue&query=%24.approximate_member_count&logo=discord&logoColor=white&label=Discord&color=green&suffix=%20total -[discord-url]: https://www.comfy.org/discord +[discord-url]: https://discord.com/invite/comfyorg [twitter-shield]: https://img.shields.io/twitter/follow/ComfyUI [twitter-url]: https://x.com/ComfyUI @@ -433,7 +433,7 @@ See also: [https://www.comfy.org/](https://www.comfy.org/) ## Frontend Development -As of August 15, 2024, we have transitioned to a new frontend, which is now hosted in a separate repository: [ComfyUI Frontend](https://github.com/Comfy-Org/ComfyUI_frontend). This repository now hosts the compiled JS (from TS/Vue) under the `web/` directory. +As of August 15, 2024, we have transitioned to a new frontend, which is now hosted in a separate repository: [ComfyUI Frontend](https://github.com/Comfy-Org/ComfyUI_frontend). The compiled JS files (from TS/Vue) are published to [pypi](https://pypi.org/project/comfyui-frontend-package) and installed as a dependency in ComfyUI. ### Reporting Issues and Requesting Features diff --git a/alembic_db/versions/0004_drop_tag_type.py b/alembic_db/versions/0004_drop_tag_type.py new file mode 100644 index 000000000..582bec4e8 --- /dev/null +++ b/alembic_db/versions/0004_drop_tag_type.py @@ -0,0 +1,39 @@ +""" +Drop the vestigial tags.tag_type column. + +tag_type was always "user" in practice — no code path ever set it to anything +else (no system/seeded classification was ever wired up) and nothing queried it. +The column, its index (ix_tags_tag_type), and the corresponding API field were +dead weight, so they are removed. + +Revision ID: 0004_drop_tag_type +Revises: 0003_add_metadata_job_id +Create Date: 2026-06-03 +""" + +from alembic import op +import sqlalchemy as sa + +revision = "0004_drop_tag_type" +down_revision = "0003_add_metadata_job_id" +branch_labels = None +depends_on = None + + +def upgrade() -> None: + with op.batch_alter_table("tags") as batch_op: + batch_op.drop_index("ix_tags_tag_type") + batch_op.drop_column("tag_type") + + +def downgrade() -> None: + with op.batch_alter_table("tags") as batch_op: + batch_op.add_column( + sa.Column( + "tag_type", + sa.String(length=32), + nullable=False, + server_default="user", + ) + ) + batch_op.create_index("ix_tags_tag_type", ["tag_type"]) diff --git a/app/assets/api/routes.py b/app/assets/api/routes.py index 68126b6a5..7ef462f5c 100644 --- a/app/assets/api/routes.py +++ b/app/assets/api/routes.py @@ -39,6 +39,7 @@ from app.assets.services import ( update_asset_metadata, upload_from_temp_path, ) +from app.assets.services.cursor import InvalidCursorError from app.assets.services.tagging import list_tag_histogram ROUTES = web.RouteTableDef() @@ -160,10 +161,12 @@ def _build_asset_response(result: schemas.AssetDetailResult | schemas.UploadResu preview_url = None else: preview_url = _build_preview_url_from_view(result.tags, result.ref.user_metadata) + asset_content_hash = result.asset.hash if result.asset else None return schemas_out.Asset( id=result.ref.id, name=result.ref.name, - asset_hash=result.asset.hash if result.asset else None, + hash=asset_content_hash, + asset_hash=asset_content_hash, size=int(result.asset.size_bytes) if result.asset else None, mime_type=result.asset.mime_type if result.asset else None, tags=result.tags, @@ -172,7 +175,7 @@ def _build_asset_response(result: schemas.AssetDetailResult | schemas.UploadResu user_metadata=result.ref.user_metadata or {}, metadata=result.ref.system_metadata, job_id=result.ref.job_id, - prompt_id=result.ref.job_id, # deprecated: mirrors job_id for cloud compat + prompt_id=result.ref.job_id, # deprecated alias of job_id, kept for compatibility created_at=result.ref.created_at, updated_at=result.ref.updated_at, last_access_time=result.ref.last_access_time, @@ -209,24 +212,37 @@ async def list_assets_route(request: web.Request) -> web.Response: order_candidate = (q.order or "desc").lower() order = order_candidate if order_candidate in {"asc", "desc"} else "desc" - result = list_assets_page( - owner_id=USER_MANAGER.get_request_user_id(request), - include_tags=q.include_tags, - exclude_tags=q.exclude_tags, - name_contains=q.name_contains, - metadata_filter=q.metadata_filter, - limit=q.limit, - offset=q.offset, - sort=sort, - order=order, - ) + try: + result = list_assets_page( + owner_id=USER_MANAGER.get_request_user_id(request), + include_tags=q.include_tags, + exclude_tags=q.exclude_tags, + name_contains=q.name_contains, + metadata_filter=q.metadata_filter, + limit=q.limit, + offset=q.offset, + sort=sort, + order=order, + after=q.after, + ) + except InvalidCursorError as e: + return _build_error_response(400, "INVALID_CURSOR", str(e)) summaries = [_build_asset_response(item) for item in result.items] + # has_more semantics differ by mode: + # - cursor mode: a non-empty next_cursor means there are more results. + # - offset mode: derived from total - (offset + page size). + if q.after is not None: + has_more = result.next_cursor is not None + else: + has_more = (q.offset + len(summaries)) < result.total + payload = schemas_out.AssetsList( assets=summaries, total=result.total, - has_more=(q.offset + len(summaries)) < result.total, + has_more=has_more, + next_cursor=result.next_cursor, ) return web.json_response(payload.model_dump(mode="json", exclude_none=True)) @@ -517,18 +533,14 @@ async def update_asset_route(request: web.Request) -> web.Response: @_require_assets_feature_enabled async def delete_asset_route(request: web.Request) -> web.Response: reference_id = str(uuid.UUID(request.match_info["id"])) - delete_content_param = request.query.get("delete_content") - delete_content = ( - False - if delete_content_param is None - else delete_content_param.lower() not in {"0", "false", "no"} - ) try: + # Deleting an asset is a soft delete of the reference; the underlying + # content is preserved (it may be shared with other references). deleted = delete_asset_reference( reference_id=reference_id, owner_id=USER_MANAGER.get_request_user_id(request), - delete_content_if_orphan=delete_content, + delete_content_if_orphan=False, ) except Exception: logging.exception( @@ -573,8 +585,8 @@ async def get_tags(request: web.Request) -> web.Response: ) tags = [ - schemas_out.TagUsage(name=name, count=count, type=tag_type) - for (name, tag_type, count) in rows + schemas_out.TagUsage(name=name, count=count) + for (name, count) in rows ] payload = schemas_out.TagsList( tags=tags, total=total, has_more=(query.offset + len(tags)) < total diff --git a/app/assets/api/schemas_in.py b/app/assets/api/schemas_in.py index 186a6ae1e..af666746d 100644 --- a/app/assets/api/schemas_in.py +++ b/app/assets/api/schemas_in.py @@ -59,6 +59,11 @@ class ListAssetsQuery(BaseModel): limit: conint(ge=1, le=500) = 20 offset: conint(ge=0) = 0 + # Opaque keyset cursor. When supplied, `offset` is ignored. Cursor pagination + # is supported for sort values `created_at`, `updated_at`, `name`, `size`. + # Supplying `after` together with `sort=last_access_time` returns + # 400 INVALID_CURSOR; that sort only supports offset/limit. + after: str | None = None sort: Literal["name", "created_at", "updated_at", "size", "last_access_time"] = ( "created_at" diff --git a/app/assets/api/schemas_out.py b/app/assets/api/schemas_out.py index d99b1098d..4e38e19d1 100644 --- a/app/assets/api/schemas_out.py +++ b/app/assets/api/schemas_out.py @@ -10,6 +10,7 @@ class Asset(BaseModel): id: str name: str + hash: str | None = None asset_hash: str | None = None size: int | None = None mime_type: str | None = None @@ -40,12 +41,13 @@ class AssetsList(BaseModel): assets: list[Asset] total: int has_more: bool + # Opaque cursor for the next page. Omitted when there are no more results. + next_cursor: str | None = None class TagUsage(BaseModel): name: str count: int - type: str class TagsList(BaseModel): diff --git a/app/assets/database/models.py b/app/assets/database/models.py index a3af8a192..9b61d309a 100644 --- a/app/assets/database/models.py +++ b/app/assets/database/models.py @@ -227,7 +227,6 @@ class Tag(Base): __tablename__ = "tags" name: Mapped[str] = mapped_column(String(512), primary_key=True) - tag_type: Mapped[str] = mapped_column(String(32), nullable=False, default="user") asset_reference_links: Mapped[list[AssetReferenceTag]] = relationship( back_populates="tag", @@ -240,7 +239,5 @@ class Tag(Base): overlaps="asset_reference_links,tag_links,tags,asset_reference", ) - __table_args__ = (Index("ix_tags_tag_type", "tag_type"),) - def __repr__(self) -> str: return f"" diff --git a/app/assets/database/queries/asset_reference.py b/app/assets/database/queries/asset_reference.py index 8b90ae511..792411800 100644 --- a/app/assets/database/queries/asset_reference.py +++ b/app/assets/database/queries/asset_reference.py @@ -266,9 +266,18 @@ def list_references_page( metadata_filter: dict | None = None, sort: str | None = None, order: str | None = None, + after_cursor_value: object | None = None, + after_cursor_id: str | None = None, ) -> tuple[list[AssetReference], dict[str, list[str]], int]: """List references with pagination, filtering, and sorting. + When ``after_cursor_value``/``after_cursor_id`` are supplied the query uses + keyset pagination — ``offset`` is ignored and a WHERE clause selects rows + strictly after the given ``(sort_col, id)`` position in the active sort + direction. The cursor value must already be typed for the column + (datetime for time sorts, int for size, str for name); the caller decodes + the opaque cursor string and resolves to the typed value. + Returns (references, tag_map, total_count). """ base = ( @@ -297,9 +306,31 @@ def list_references_page( "size": Asset.size_bytes, } sort_col = sort_map.get(sort, AssetReference.created_at) - sort_exp = sort_col.desc() if order == "desc" else sort_col.asc() + descending = order == "desc" - base = base.order_by(sort_exp).limit(limit).offset(offset) + # Keyset WHERE: (sort_col, id) strictly less-than / greater-than the cursor. + # Equivalent to: sort_col v OR (sort_col = v AND id cursor_id). + if after_cursor_value is not None and after_cursor_id is not None: + if descending: + keyset = sa.or_( + sort_col < after_cursor_value, + sa.and_(sort_col == after_cursor_value, AssetReference.id < after_cursor_id), + ) + else: + keyset = sa.or_( + sort_col > after_cursor_value, + sa.and_(sort_col == after_cursor_value, AssetReference.id > after_cursor_id), + ) + base = base.where(keyset) + + # Secondary ORDER BY id (matching the primary direction) gives the keyset + # comparison a deterministic tiebreaker on duplicate sort_col values. + id_exp = AssetReference.id.desc() if descending else AssetReference.id.asc() + sort_exp = sort_col.desc() if descending else sort_col.asc() + + base = base.order_by(sort_exp, id_exp).limit(limit) + if after_cursor_id is None: + base = base.offset(offset) count_stmt = ( select(sa.func.count()) diff --git a/app/assets/database/queries/tags.py b/app/assets/database/queries/tags.py index f4126dba8..d41d73a10 100644 --- a/app/assets/database/queries/tags.py +++ b/app/assets/database/queries/tags.py @@ -55,13 +55,11 @@ def validate_tags_exist(session: Session, tags: list[str]) -> None: raise ValueError(f"Unknown tags: {missing}") -def ensure_tags_exist( - session: Session, names: Iterable[str], tag_type: str = "user" -) -> None: +def ensure_tags_exist(session: Session, names: Iterable[str]) -> None: wanted = normalize_tags(list(names)) if not wanted: return - rows = [{"name": n, "tag_type": tag_type} for n in list(dict.fromkeys(wanted))] + rows = [{"name": n} for n in list(dict.fromkeys(wanted))] ins = ( sqlite.insert(Tag) .values(rows) @@ -97,7 +95,7 @@ def set_reference_tags( to_remove = [t for t in current if t not in desired] if to_add: - ensure_tags_exist(session, to_add, tag_type="user") + ensure_tags_exist(session, to_add) session.add_all( [ AssetReferenceTag( @@ -142,7 +140,7 @@ def add_tags_to_reference( return AddTagsResult(added=[], already_present=[], total_tags=total) if create_if_missing: - ensure_tags_exist(session, norm, tag_type="user") + ensure_tags_exist(session, norm) current = set(get_reference_tags(session, reference_id)) @@ -289,7 +287,6 @@ def list_tags_with_usage( q = ( select( Tag.name, - Tag.tag_type, func.coalesce(counts_sq.c.cnt, 0).label("count"), ) .select_from(Tag) @@ -331,7 +328,7 @@ def list_tags_with_usage( rows = (session.execute(q.limit(limit).offset(offset))).all() total = (session.execute(total_q)).scalar_one() - rows_norm = [(name, ttype, int(count or 0)) for (name, ttype, count) in rows] + rows_norm = [(name, int(count or 0)) for (name, count) in rows] return rows_norm, int(total or 0) diff --git a/app/assets/scanner.py b/app/assets/scanner.py index ebb6869af..2c1e97840 100644 --- a/app/assets/scanner.py +++ b/app/assets/scanner.py @@ -33,6 +33,7 @@ from app.assets.services.file_utils import ( verify_file_unchanged, ) from app.assets.services.hashing import HashCheckpoint, compute_blake3_hash +from app.assets.services.image_dimensions import extract_image_dimensions from app.assets.services.metadata_extract import extract_file_metadata from app.assets.services.path_utils import ( compute_relative_filename, @@ -354,7 +355,7 @@ def insert_asset_specs(specs: list[SeedAssetSpec], tag_pool: set[str]) -> int: return 0 with create_session() as sess: if tag_pool: - ensure_tags_exist(sess, tag_pool, tag_type="user") + ensure_tags_exist(sess, tag_pool) result = batch_insert_seed_assets(sess, specs=specs, owner_id="") sess.commit() return result.inserted_refs @@ -506,6 +507,10 @@ def enrich_asset( if extract_metadata and metadata: system_metadata = metadata.to_user_metadata() + if mime_type and mime_type.startswith("image/"): + dims = extract_image_dimensions(file_path, mime_type=mime_type) + if dims: + system_metadata.update(dims) set_reference_system_metadata(session, reference_id, system_metadata) if full_hash: diff --git a/app/assets/services/asset_management.py b/app/assets/services/asset_management.py index 5aefd9956..d4e4fc61c 100644 --- a/app/assets/services/asset_management.py +++ b/app/assets/services/asset_management.py @@ -1,8 +1,19 @@ import contextlib import mimetypes import os +from datetime import timezone from typing import Sequence +from app.assets.services.cursor import ( + CursorPayload, + InvalidCursorError, + decode_cursor, + decode_cursor_int, + decode_cursor_time, + encode_cursor, + encode_cursor_from_time, +) + from app.assets.database.models import Asset from app.assets.database.queries import ( @@ -149,6 +160,16 @@ def delete_asset_reference( owner_id: str, delete_content_if_orphan: bool = True, ) -> bool: + """Delete an asset reference. + + With ``delete_content_if_orphan=False`` (a soft delete), the reference is + hidden and the underlying content is preserved. With ``True``, the content + is also removed once it becomes orphaned. + + Note: the public DELETE /api/assets/{id} endpoint always soft-deletes + (passes ``False``); the orphan-reclamation path is intentionally + internal-only, retained for a future GC/admin caller. + """ with create_session() as session: if not delete_content_if_orphan: # Soft delete: mark the reference as deleted but keep everything @@ -242,6 +263,11 @@ def get_asset_by_hash(asset_hash: str) -> AssetData | None: return extract_asset_data(asset) +# Sort fields that support cursor pagination. `last_access_time` is not +# in this list — it falls back to offset/limit. +_CURSOR_SORT_FIELDS = ("created_at", "updated_at", "name", "size") + + def list_assets_page( owner_id: str = "", include_tags: Sequence[str] | None = None, @@ -252,7 +278,39 @@ def list_assets_page( offset: int = 0, sort: str = "created_at", order: str = "desc", + after: str | None = None, ) -> ListAssetsResult: + """List assets with optional cursor pagination. + + When ``after`` is supplied it overrides ``offset``. The cursor's sort field + must match ``sort`` and be in the cursor-supported allowlist; mismatches + raise InvalidCursorError so the handler can map to 400 INVALID_CURSOR. + """ + cursor_value: object | None = None + cursor_id: str | None = None + # Mint next_cursor on every page where the sort is cursor-supported, not + # only when the request itself arrived with a cursor. Otherwise a first + # request (no `after`) returns next_cursor=None and the client can never + # enter cursor mode. + mint_cursor = sort in _CURSOR_SORT_FIELDS + + if after is not None: + if sort not in _CURSOR_SORT_FIELDS: + raise InvalidCursorError( + f"cursor pagination is not supported for sort={sort!r}" + ) + payload = decode_cursor(after, _CURSOR_SORT_FIELDS, expected_order=order) + if payload.sort_field != sort: + raise InvalidCursorError( + f"cursor sort field {payload.sort_field!r} does not match request sort {sort!r}" + ) + cursor_value, cursor_id = _resolve_cursor_value(payload), payload.id + + # Over-fetch by one row so we can distinguish "exactly `limit` rows total + # remaining" from "more rows past this page" without a second query. Drop + # the sentinel before returning. + fetch_limit = limit + 1 if mint_cursor else limit + with create_session() as session: refs, tag_map, total = list_references_page( session, @@ -261,12 +319,22 @@ def list_assets_page( exclude_tags=exclude_tags, name_contains=name_contains, metadata_filter=metadata_filter, - limit=limit, + limit=fetch_limit, offset=offset, sort=sort, order=order, + after_cursor_value=cursor_value, + after_cursor_id=cursor_id, ) + next_cursor: str | None = None + if mint_cursor and len(refs) > limit: + # There's at least one more row past this page — mint a cursor from + # the last row of the page (i.e. index `limit - 1`, since we + # over-fetched), and drop the sentinel. + next_cursor = _encode_next_cursor(refs[limit - 1], sort, order) + refs = refs[:limit] + items: list[AssetSummaryData] = [] for ref in refs: items.append( @@ -277,7 +345,39 @@ def list_assets_page( ) ) - return ListAssetsResult(items=items, total=total) + return ListAssetsResult(items=items, total=total, next_cursor=next_cursor) + + +def _resolve_cursor_value(payload: CursorPayload) -> object: + """Map a decoded cursor payload to a column-typed Python value.""" + if payload.sort_field in ("created_at", "updated_at"): + # DB stores naive UTC; strip tzinfo so the comparison binds against a + # `TIMESTAMP WITHOUT TIME ZONE` column without an offset shift. + return decode_cursor_time(payload).replace(tzinfo=None) + if payload.sort_field == "size": + return decode_cursor_int(payload) + return payload.value # name, str-typed + + +def _encode_next_cursor(ref, sort: str, order: str) -> str | None: + """Mint a cursor pointing at *ref* for the given sort dimension. + + Returns None when the boundary row carries a NULL sort value (e.g. an asset + record whose size_bytes hasn't been backfilled). Continuing pagination + across a NULL boundary is undefined under keyset ordering — better to + truncate cleanly here than to mint a cursor that mis-positions. + """ + if sort == "name": + return encode_cursor("name", ref.name, ref.id, order=order) + if sort == "size": + if ref.asset is None or ref.asset.size_bytes is None: + return None + return encode_cursor("size", str(ref.asset.size_bytes), ref.id, order=order) + # created_at / updated_at — DB datetimes are naive UTC; attach tz before encoding. + value = ref.created_at if sort == "created_at" else ref.updated_at + if value is None: + return None + return encode_cursor_from_time(sort, value.replace(tzinfo=timezone.utc), ref.id, order=order) def resolve_hash_to_path( diff --git a/app/assets/services/cursor.py b/app/assets/services/cursor.py new file mode 100644 index 000000000..6c7791528 --- /dev/null +++ b/app/assets/services/cursor.py @@ -0,0 +1,213 @@ +"""Opaque keyset-pagination cursor for /api/assets. + +Payload JSON uses short keys to keep the encoded length small: + + {"s": , "v": , "id": , "o": } + +The `o` key binds the cursor to the sort direction it was minted under, +so replaying a `desc` cursor against an `asc` request fails with +``INVALID_CURSOR`` rather than silently walking the wrong direction. +`o` is mandatory on every payload — a cursor without it is rejected as +malformed. + +Encoding is base64url with no padding. Cursors are opaque tokens: the +payload format is internal to this server, and clients must treat a +cursor as a black box handed back via `next_cursor`. No byte-level +compatibility with any other implementation is required. + +Time values are serialized as Unix microseconds (UTC) — microsecond +precision is sufficient to round-trip the timestamps stored by the +database without rounding rows in the same millisecond bucket. +""" +from __future__ import annotations + +import base64 +import json +from dataclasses import dataclass +from datetime import datetime, timezone +from typing import Iterable, Optional + + +class InvalidCursorError(ValueError): + """Raised on a malformed, oversized, or unsupported-sort-field cursor. + + Map to a 400 response with code ``INVALID_CURSOR`` at the handler. + """ + + +# Wire-format length caps. Cursors are user-controlled, so caps protect the +# decode path from oversized allocations and downstream SQL predicates from +# unbounded strings. +# +# MAX_CURSOR_VALUE_LENGTH is 512 to fit the `AssetReference.name` column max +# (`String(512)`) — otherwise a long-named asset would mint a cursor the same +# server then refuses on the next request. +# +# MAX_ENCODED_CURSOR_LENGTH is the decode-path guard, sized comfortably above +# the largest cursor the per-field caps can produce. Worst case is value + id +# at their caps with every character JSON-escaping to the six-byte `\uXXXX` +# form (control characters), which is ~5.2 KB once base64url-encoded. At 8192 +# the encoder can never mint a cursor that exceeds it, so a freshly minted +# cursor always decodes on the next request and there is no user-visible +# "cursor too long" failure. +MAX_ENCODED_CURSOR_LENGTH = 8192 +MAX_CURSOR_VALUE_LENGTH = 512 +MAX_CURSOR_ID_LENGTH = 128 + + +@dataclass(frozen=True) +class CursorPayload: + sort_field: str + value: str + id: str + order: str + + +_VALID_ORDERS = ("asc", "desc") + + +def encode_cursor(sort_field: str, value: str, id: str, order: str = "desc") -> str: + """Encode a cursor payload as a base64url (no-padding) string. + + `order` binds the cursor to the sort direction it was minted under so a + later request with a flipped `order` query parameter is rejected with + ``INVALID_CURSOR`` rather than silently walking the wrong direction. + """ + if order not in _VALID_ORDERS: + raise InvalidCursorError(f"order must be one of {_VALID_ORDERS}, got {order!r}") + # Symmetric input validation: the encoder must reject anything the + # decoder rejects, or the same server will mint cursors it then 400s on + # the next request. + if not id: + raise InvalidCursorError("id must be non-empty") + if len(id) > MAX_CURSOR_ID_LENGTH: + raise InvalidCursorError("id exceeds maximum length") + if len(value) > MAX_CURSOR_VALUE_LENGTH: + raise InvalidCursorError("value exceeds maximum length") + payload = {"s": sort_field, "v": value, "id": id, "o": order} + raw = json.dumps(payload, separators=(",", ":"), ensure_ascii=False) + # No mint-time length guard is needed: the per-field caps above bound the + # encoded length well below MAX_ENCODED_CURSOR_LENGTH (see its definition), + # so the encoder can never produce a cursor the decode path would reject. + return base64.urlsafe_b64encode(raw.encode("utf-8")).rstrip(b"=").decode("ascii") + + +def encode_cursor_from_time(sort_field: str, t: datetime, id: str, order: str = "desc") -> str: + """Encode a time-typed cursor at Unix microsecond precision. + + Accepts an aware datetime (any timezone) and normalizes to UTC. Naive + datetimes are rejected so callers can't accidentally encode the local + wall-clock value of a UTC-stored timestamp. + """ + if t.tzinfo is None: + raise ValueError("encode_cursor_from_time requires an aware datetime") + micros = _datetime_to_unix_micros(t.astimezone(timezone.utc)) + return encode_cursor(sort_field, str(micros), id, order=order) + + +def decode_cursor( + cursor: str, + allowed_sort_fields: Iterable[str], + expected_order: str | None = None, +) -> CursorPayload: + """Parse an opaque cursor. + + ``allowed_sort_fields`` is the endpoint's accepted sort-field list — a + cursor carrying a field outside this set is rejected so a cursor minted + for one column can't be replayed against another (e.g. a ``created_at`` + timestamp string compared against a ``name`` column). + + ``expected_order`` (``"asc"``/``"desc"``), when supplied, must match the + payload's ``o`` field. ``o`` is required on every payload; a cursor + missing it is rejected as malformed. + + Passing no allowed fields rejects every cursor. + """ + if len(cursor) > MAX_ENCODED_CURSOR_LENGTH: + raise InvalidCursorError("cursor exceeds maximum length") + + try: + # urlsafe_b64decode requires correct padding; we strip on encode, so + # restore the trailing '=' pad here. + padding = "=" * (-len(cursor) % 4) + raw = base64.urlsafe_b64decode(cursor + padding) + except (ValueError, base64.binascii.Error) as e: + raise InvalidCursorError(f"encoding: {e}") from e + + try: + decoded = json.loads(raw) + except (json.JSONDecodeError, UnicodeDecodeError) as e: + raise InvalidCursorError(f"payload: {e}") from e + + if not isinstance(decoded, dict): + raise InvalidCursorError("payload: expected object") + + sort_field = decoded.get("s") + value = decoded.get("v") + id = decoded.get("id") + order = decoded.get("o") + + if not isinstance(sort_field, str) or not isinstance(value, str) or not isinstance(id, str): + raise InvalidCursorError("payload: missing or non-string s/v/id") + + if id == "": + raise InvalidCursorError("missing id") + if len(id) > MAX_CURSOR_ID_LENGTH: + raise InvalidCursorError("id exceeds maximum length") + if len(value) > MAX_CURSOR_VALUE_LENGTH: + raise InvalidCursorError("value exceeds maximum length") + + if sort_field not in allowed_sort_fields: + raise InvalidCursorError(f"unsupported sort field {sort_field!r}") + + if not isinstance(order, str): + raise InvalidCursorError("missing or non-string o") + if order not in _VALID_ORDERS: + raise InvalidCursorError(f"unsupported order {order!r}") + if expected_order is not None and order != expected_order: + raise InvalidCursorError( + f"cursor order {order!r} does not match request order {expected_order!r}" + ) + + return CursorPayload(sort_field=sort_field, value=value, id=id, order=order) + + +def decode_cursor_time(payload: Optional[CursorPayload]) -> datetime: + """Parse a time-typed cursor value as Unix microseconds, returning UTC.""" + if payload is None: + raise InvalidCursorError("nil cursor payload") + try: + micros = int(payload.value) + except ValueError as e: + raise InvalidCursorError(f"value is not a valid timestamp: {e}") from e + try: + return _unix_micros_to_datetime(micros) + except (OverflowError, OSError, ValueError) as e: + # Crafted out-of-range microseconds (e.g. > datetime.MAX_YEAR) blow up + # in fromtimestamp / datetime construction. Map to 400, not 500. + raise InvalidCursorError(f"value is out of representable range: {e}") from e + + +def decode_cursor_int(payload: Optional[CursorPayload]) -> int: + """Parse a cursor value as a base-10 integer.""" + if payload is None: + raise InvalidCursorError("nil cursor payload") + try: + return int(payload.value) + except ValueError as e: + raise InvalidCursorError(f"value is not a valid integer: {e}") from e + + +_EPOCH = datetime(1970, 1, 1, tzinfo=timezone.utc) + + +def _datetime_to_unix_micros(t: datetime) -> int: + """Convert an aware UTC datetime to Unix microseconds (integer math).""" + delta = t - _EPOCH + return (delta.days * 86_400 + delta.seconds) * 1_000_000 + delta.microseconds + + +def _unix_micros_to_datetime(micros: int) -> datetime: + """Convert Unix microseconds to a UTC datetime, preserving precision.""" + seconds, micro_remainder = divmod(micros, 1_000_000) + return datetime.fromtimestamp(seconds, tz=timezone.utc).replace(microsecond=micro_remainder) diff --git a/app/assets/services/image_dimensions.py b/app/assets/services/image_dimensions.py new file mode 100644 index 000000000..ccd97399a --- /dev/null +++ b/app/assets/services/image_dimensions.py @@ -0,0 +1,63 @@ +"""Image dimension extraction for asset ingest. + +Reads only the image header via Pillow to capture width/height cheaply, +without a full pixel decode. Returns a metadata dict suitable for merging +into ``AssetReference.system_metadata``. +""" +from __future__ import annotations + +import logging +from typing import Any + +logger = logging.getLogger(__name__) + + +def extract_image_dimensions( + file_path: str, mime_type: str | None = None +) -> dict[str, Any] | None: + """Extract image dimensions for the file at ``file_path``. + + Args: + file_path: Absolute path to a file on disk. + mime_type: Optional MIME type hint. When provided and not prefixed + with ``image/``, extraction is skipped without touching the file. + + Returns: + ``{"kind": "image", "width": W, "height": H}`` when the file is a + recognizable image with positive dimensions, otherwise ``None``. + + The dict shape is intended to be merged into ``system_metadata`` so the + asset response surfaces ``metadata.kind`` plus dimension fields for image + assets. Forward-compatible: future media kinds (e.g. ``"video"`` with + duration/fps) can extend this shape without schema changes. + """ + if mime_type is not None and not mime_type.startswith("image/"): + return None + + try: + from PIL import Image, UnidentifiedImageError + except ImportError: + logger.debug( + "Pillow not available; skipping image dimension extraction for %s", + file_path, + ) + return None + + try: + with Image.open(file_path) as img: + width, height = img.size + except (OSError, UnidentifiedImageError, ValueError) as exc: + logger.debug( + "Failed to read image dimensions from %s: %s", file_path, exc + ) + return None + + if ( + not isinstance(width, int) + or not isinstance(height, int) + or width <= 0 + or height <= 0 + ): + return None + + return {"kind": "image", "width": width, "height": height} diff --git a/app/assets/services/ingest.py b/app/assets/services/ingest.py index f0b070517..3b6dc237c 100644 --- a/app/assets/services/ingest.py +++ b/app/assets/services/ingest.py @@ -17,9 +17,11 @@ from app.assets.database.queries import ( get_reference_by_file_path, get_reference_tags, get_or_create_reference, + list_references_by_asset_id, reference_exists, remove_missing_tag_for_asset_id, set_reference_metadata, + set_reference_system_metadata, set_reference_tags, update_asset_hash_and_mime, upsert_asset, @@ -29,6 +31,7 @@ from app.assets.database.queries import ( from app.assets.helpers import get_utc_now, normalize_tags from app.assets.services.bulk_ingest import batch_insert_seed_assets from app.assets.services.file_utils import get_size_and_mtime_ns +from app.assets.services.image_dimensions import extract_image_dimensions from app.assets.services.path_utils import ( compute_relative_filename, get_name_and_tags_from_asset_path, @@ -118,6 +121,14 @@ def _ingest_file_from_path( user_metadata=user_metadata, ) + _maybe_store_image_dimensions( + session, + reference_id=reference_id, + file_path=locator, + mime_type=mime_type, + current_system_metadata=ref.system_metadata, + ) + try: remove_missing_tag_for_asset_id(session, asset_id=asset.id) except Exception: @@ -288,6 +299,13 @@ def _register_existing_asset( user_metadata=new_meta, ) + _backfill_image_dimensions_from_siblings( + session, + asset_id=asset.id, + new_reference_id=ref.id, + current_system_metadata=ref.system_metadata, + ) + if tags is not None: set_reference_tags( session, @@ -334,6 +352,87 @@ def _update_metadata_with_filename( ) +_IMAGE_DIMENSION_KEYS = ("kind", "width", "height") + + +def _maybe_store_image_dimensions( + session: Session, + reference_id: str, + file_path: str, + mime_type: str | None, + current_system_metadata: dict | None, +) -> None: + """Populate ``kind``/``width``/``height`` on system_metadata for image refs. + + Non-image MIME types are a no-op. Pre-existing keys (e.g. enricher-written + safetensors metadata, download provenance) are preserved by merge. + """ + if not mime_type or not mime_type.startswith("image/"): + return + + dims = extract_image_dimensions(file_path, mime_type=mime_type) + if not dims: + return + + current = current_system_metadata or {} + merged = dict(current) + merged.update(dims) + if merged != current: + set_reference_system_metadata( + session, + reference_id=reference_id, + system_metadata=merged, + ) + + +def _backfill_image_dimensions_from_siblings( + session: Session, + asset_id: str, + new_reference_id: str, + current_system_metadata: dict | None, +) -> None: + """Copy image dimension keys from any sibling reference of the same asset. + + The from-hash path doesn't read the file bytes, so dimensions can't be + extracted there directly. When another reference of the same asset already + carries image dimensions, copy them onto the new reference so consumers + see consistent metadata regardless of how the asset was registered. + + Best-effort: missing siblings, non-image siblings, or absent dimension + keys leave the target reference unchanged. + """ + current = current_system_metadata or {} + if current.get("kind") == "image" and "width" in current and "height" in current: + return + + for sibling in list_references_by_asset_id(session, asset_id): + if sibling.id == new_reference_id: + continue + meta = sibling.system_metadata or {} + if meta.get("kind") != "image": + continue + width = meta.get("width") + height = meta.get("height") + if ( + type(width) is not int + or type(height) is not int + or width <= 0 + or height <= 0 + ): + continue + merged = dict(current) + merged["kind"] = "image" + merged["width"] = width + merged["height"] = height + if merged != current: + set_reference_system_metadata( + session, + reference_id=new_reference_id, + system_metadata=merged, + ) + return + + def _sanitize_filename(name: str | None, fallback: str) -> str: n = os.path.basename((name or "").strip() or fallback) return n if n else fallback diff --git a/app/assets/services/metadata_extract.py b/app/assets/services/metadata_extract.py index a004929bc..bdfe60218 100644 --- a/app/assets/services/metadata_extract.py +++ b/app/assets/services/metadata_extract.py @@ -4,7 +4,6 @@ Tier 1: Filesystem metadata (zero parsing) Tier 2: Safetensors header metadata (fast JSON read only) """ -from __future__ import annotations import json import logging diff --git a/app/assets/services/schemas.py b/app/assets/services/schemas.py index 0eb128f58..4d2af8a02 100644 --- a/app/assets/services/schemas.py +++ b/app/assets/services/schemas.py @@ -56,7 +56,6 @@ class IngestResult: class TagUsage(NamedTuple): name: str - tag_type: str count: int @@ -71,6 +70,7 @@ class AssetSummaryData: class ListAssetsResult: items: list[AssetSummaryData] total: int + next_cursor: str | None = None @dataclass(frozen=True) diff --git a/app/assets/services/tagging.py b/app/assets/services/tagging.py index 37b612753..5fa39d26a 100644 --- a/app/assets/services/tagging.py +++ b/app/assets/services/tagging.py @@ -75,7 +75,7 @@ def list_tags( owner_id=owner_id, ) - return [TagUsage(name, tag_type, count) for name, tag_type, count in rows], total + return [TagUsage(name, count) for name, count in rows], total def list_tag_histogram( diff --git a/app/custom_node_manager.py b/app/custom_node_manager.py index 281febca9..738af2abd 100644 --- a/app/custom_node_manager.py +++ b/app/custom_node_manager.py @@ -1,5 +1,3 @@ -from __future__ import annotations - import os import folder_paths import glob diff --git a/app/frontend_management.py b/app/frontend_management.py index d0596b276..8e84e8dd9 100644 --- a/app/frontend_management.py +++ b/app/frontend_management.py @@ -1,4 +1,3 @@ -from __future__ import annotations import argparse import logging import os @@ -62,6 +61,8 @@ def get_comfy_package_versions(): def check_comfy_packages_versions(): """Warn for every comfy* package whose installed version is below requirements.txt.""" from packaging.version import InvalidVersion, parse as parse_pep440 + outdated_packages = [] + for pkg in get_comfy_package_versions(): installed_str = pkg["installed"] required_str = pkg["required"] @@ -73,19 +74,26 @@ def check_comfy_packages_versions(): logging.error(f"Failed to check {pkg['name']} version: {e}") continue if outdated: - app.logger.log_startup_warning( - f""" + outdated_packages.append((pkg["name"], installed_str, required_str)) + else: + logging.info("{} version: {}".format(pkg["name"], installed_str)) + + if outdated_packages: + package_warnings = "\n".join( + f"Installed {name} version {installed} is lower than the recommended version {required}." + for name, installed, required in outdated_packages + ) + app.logger.log_startup_warning( + f""" ________________________________________________________________________ WARNING WARNING WARNING WARNING WARNING -Installed {pkg["name"]} version {installed_str} is lower than the recommended version {required_str}. +{package_warnings} {get_missing_requirements_message()} ________________________________________________________________________ """.strip() - ) - else: - logging.info("{} version: {}".format(pkg["name"], installed_str)) + ) REQUEST_TIMEOUT = 10 # seconds diff --git a/app/logger.py b/app/logger.py index 3d26d98fe..bde815822 100644 --- a/app/logger.py +++ b/app/logger.py @@ -5,6 +5,40 @@ import logging import sys import threading +ANSI_NAMED_COLORS = { + 'black': '\033[30m', + 'red': '\033[31m', + 'green': '\033[32m', + 'yellow': '\033[33m', + 'blue': '\033[34m', + 'magenta': '\033[35m', + 'cyan': '\033[36m', + 'white': '\033[37m', +} + +ANSI_LEVEL_COLORS = { + 'DEBUG': ANSI_NAMED_COLORS['cyan'], + 'INFO': ANSI_NAMED_COLORS['green'], + 'WARNING': ANSI_NAMED_COLORS['yellow'], + 'ERROR': ANSI_NAMED_COLORS['red'], + 'CRITICAL': ANSI_NAMED_COLORS['magenta'], +} + +ANSI_RESET = '\033[0m' +ANSI_BOLD = '\033[1m' + + +class ColoredFormatter(logging.Formatter): + def format(self, record): + color = ANSI_LEVEL_COLORS.get(record.levelname, '') + bold = ANSI_BOLD if record.levelno >= logging.WARNING else '' + level_tag = f"{bold}{color}[{record.levelname}]{ANSI_RESET} " + message = super().format(record) + line_color = ANSI_NAMED_COLORS.get(getattr(record, 'color', ''), '') + if line_color: + return f"{level_tag}{line_color}{message}{ANSI_RESET}" + return level_tag + message + logs = None stdout_interceptor = None stderr_interceptor = None @@ -68,8 +102,10 @@ def setup_logger(log_level: str = 'INFO', capacity: int = 300, use_stdout: bool logger = logging.getLogger() logger.setLevel(log_level) + formatter = ColoredFormatter("%(message)s") + stream_handler = logging.StreamHandler() - stream_handler.setFormatter(logging.Formatter("%(message)s")) + stream_handler.setFormatter(formatter) if use_stdout: # Only errors and critical to stderr @@ -77,7 +113,7 @@ def setup_logger(log_level: str = 'INFO', capacity: int = 300, use_stdout: bool # Lesser to stdout stdout_handler = logging.StreamHandler(sys.stdout) - stdout_handler.setFormatter(logging.Formatter("%(message)s")) + stdout_handler.setFormatter(formatter) stdout_handler.addFilter(lambda record: record.levelno < logging.ERROR) logger.addHandler(stdout_handler) diff --git a/app/model_manager.py b/app/model_manager.py index f124d1117..8f6e34b33 100644 --- a/app/model_manager.py +++ b/app/model_manager.py @@ -1,5 +1,3 @@ -from __future__ import annotations - import os import base64 import json diff --git a/app/user_manager.py b/app/user_manager.py index 0517b3344..7b11e381c 100644 --- a/app/user_manager.py +++ b/app/user_manager.py @@ -1,4 +1,3 @@ -from __future__ import annotations import json import os import re diff --git a/blueprints/Audio Generation (Stable Audio 3 Medium Base).json b/blueprints/Audio Generation (Stable Audio 3 Medium Base).json new file mode 100644 index 000000000..e561fe634 --- /dev/null +++ b/blueprints/Audio Generation (Stable Audio 3 Medium Base).json @@ -0,0 +1,2091 @@ +{ + "revision": 0, + "last_node_id": 52, + "last_link_id": 0, + "nodes": [ + { + "id": 52, + "type": "8b66c757-fe2f-4184-91f3-479a19deb565", + "pos": [ + 370, + 1120 + ], + "size": [ + 420, + 450 + ], + "flags": { + "collapsed": false + }, + "order": 0, + "mode": 0, + "inputs": [ + { + "label": "user_input", + "name": "user_input", + "type": "STRING", + "widget": { + "name": "user_input" + }, + "link": null + }, + { + "label": "duration", + "name": "duration", + "type": "FLOAT", + "widget": { + "name": "duration" + }, + "link": null + }, + { + "label": "seed", + "name": "seed", + "type": "INT", + "widget": { + "name": "seed" + }, + "link": null + }, + { + "label": "use_reprompt", + "name": "use_reprompt", + "type": "BOOLEAN", + "widget": { + "name": "use_reprompt" + }, + "link": null + }, + { + "label": "reprompt_category", + "name": "category", + "type": "COMBO", + "widget": { + "name": "category" + }, + "link": null + }, + { + "label": "ckpt_name", + "name": "ckpt_name", + "type": "COMBO", + "widget": { + "name": "ckpt_name" + }, + "link": null + }, + { + "label": "sa_clip", + "name": "sa_clip", + "type": "COMBO", + "widget": { + "name": "sa_clip" + }, + "link": null + }, + { + "label": "qwen_clip", + "name": "qwen_clip", + "type": "COMBO", + "widget": { + "name": "qwen_clip" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "AUDIO", + "name": "AUDIO", + "type": "AUDIO", + "links": [] + } + ], + "properties": { + "proxyWidgets": [ + [ + "31", + "value" + ], + [ + "36", + "value" + ], + [ + "3", + "seed" + ], + [ + "35", + "value" + ], + [ + "43", + "choice" + ], + [ + "25", + "ckpt_name" + ], + [ + "26", + "clip_name" + ], + [ + "29", + "clip_name" + ] + ] + }, + "widgets_values": [], + "title": "Audio Generation (Stable Audio 3 Medium Base)" + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "8b66c757-fe2f-4184-91f3-479a19deb565", + "version": 1, + "state": { + "lastGroupId": 8, + "lastNodeId": 56, + "lastLinkId": 84, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Audio Generation (Stable Audio 3 Medium Base)", + "inputNode": { + "id": -10, + "bounding": [ + -810, + 400, + 155.953125, + 208 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + 1750, + 1041, + 128, + 68 + ] + }, + "inputs": [ + { + "id": "78ae2515-114b-494a-becc-43c7b6c2dc2f", + "name": "user_input", + "type": "STRING", + "linkIds": [ + 68 + ], + "label": "user_input", + "pos": [ + -678.046875, + 424 + ] + }, + { + "id": "5ca95030-aff4-4544-b545-f0d814e0e49a", + "name": "duration", + "type": "FLOAT", + "linkIds": [ + 82 + ], + "label": "duration", + "pos": [ + -678.046875, + 444 + ] + }, + { + "id": "718eb10f-da1a-4cea-a9c7-3040f98fe960", + "name": "seed", + "type": "INT", + "linkIds": [ + 76 + ], + "label": "seed", + "pos": [ + -678.046875, + 464 + ] + }, + { + "id": "dc020099-39e6-4009-9937-408409d71736", + "name": "use_reprompt", + "type": "BOOLEAN", + "linkIds": [ + 83 + ], + "label": "use_reprompt", + "pos": [ + -678.046875, + 484 + ] + }, + { + "id": "edae394c-6324-44d6-8ac5-d8caa5ae2169", + "name": "category", + "type": "COMBO", + "linkIds": [ + 78 + ], + "label": "reprompt_category", + "pos": [ + -678.046875, + 504 + ] + }, + { + "id": "be19b747-6a47-4028-9c30-d52f54a712ea", + "name": "ckpt_name", + "type": "COMBO", + "linkIds": [ + 79 + ], + "label": "ckpt_name", + "pos": [ + -678.046875, + 524 + ] + }, + { + "id": "bc9241a2-bc20-4c5d-8cb1-f2958f598642", + "name": "sa_clip", + "type": "COMBO", + "linkIds": [ + 80 + ], + "label": "sa_clip", + "pos": [ + -678.046875, + 544 + ] + }, + { + "id": "a33a2468-6d6d-4cb6-937c-3510bf16ebac", + "name": "qwen_clip", + "type": "COMBO", + "linkIds": [ + 81 + ], + "label": "qwen_clip", + "pos": [ + -678.046875, + 564 + ] + } + ], + "outputs": [ + { + "id": "bbe988dd-5c03-44fd-a965-c712f9204988", + "name": "AUDIO", + "type": "AUDIO", + "linkIds": [ + 27 + ], + "localized_name": "AUDIO", + "pos": [ + 1774, + 1065 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 7, + "type": "CLIPTextEncode", + "pos": [ + 620, + 420 + ], + "size": [ + 440, + 140 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 35 + }, + { + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "slot_index": 0, + "links": [ + 6 + ] + } + ], + "properties": { + "Node name for S&R": "CLIPTextEncode", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "" + ], + "color": "#223", + "bgcolor": "#335" + }, + { + "id": 12, + "type": "VAEDecodeAudio", + "pos": [ + 1450, + 110 + ], + "size": [ + 230, + 100 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "localized_name": "samples", + "name": "samples", + "type": "LATENT", + "link": 13 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 39 + } + ], + "outputs": [ + { + "localized_name": "AUDIO", + "name": "AUDIO", + "type": "AUDIO", + "slot_index": 0, + "links": [ + 27 + ] + } + ], + "properties": { + "Node name for S&R": "VAEDecodeAudio", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 11, + "type": "EmptyLatentAudio", + "pos": [ + 630, + 610 + ], + "size": [ + 430, + 140 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "seconds", + "name": "seconds", + "type": "FLOAT", + "widget": { + "name": "seconds" + }, + "link": 50 + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "widget": { + "name": "batch_size" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "LATENT", + "name": "LATENT", + "type": "LATENT", + "links": [ + 12 + ] + } + ], + "properties": { + "Node name for S&R": "EmptyLatentAudio", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 60, + 1 + ] + }, + { + "id": 3, + "type": "KSampler", + "pos": [ + 1100, + 100 + ], + "size": [ + 320, + 350 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 30 + }, + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "link": 4 + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "link": 6 + }, + { + "localized_name": "latent_image", + "name": "latent_image", + "type": "LATENT", + "link": 12 + }, + { + "localized_name": "seed", + "name": "seed", + "type": "INT", + "widget": { + "name": "seed" + }, + "link": 76 + }, + { + "localized_name": "steps", + "name": "steps", + "type": "INT", + "widget": { + "name": "steps" + }, + "link": null + }, + { + "localized_name": "cfg", + "name": "cfg", + "type": "FLOAT", + "widget": { + "name": "cfg" + }, + "link": null + }, + { + "localized_name": "sampler_name", + "name": "sampler_name", + "type": "COMBO", + "widget": { + "name": "sampler_name" + }, + "link": null + }, + { + "localized_name": "scheduler", + "name": "scheduler", + "type": "COMBO", + "widget": { + "name": "scheduler" + }, + "link": null + }, + { + "localized_name": "denoise", + "name": "denoise", + "type": "FLOAT", + "widget": { + "name": "denoise" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "LATENT", + "name": "LATENT", + "type": "LATENT", + "slot_index": 0, + "links": [ + 13 + ] + } + ], + "properties": { + "Node name for S&R": "KSampler", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 0, + "randomize", + 50, + 7, + "lcm", + "simple", + 1 + ] + }, + { + "id": 29, + "type": "CLIPLoader", + "pos": [ + 690, + 1580 + ], + "size": [ + 430, + 170 + ], + "flags": {}, + "order": 8, + "mode": 0, + "showAdvanced": false, + "inputs": [ + { + "localized_name": "clip_name", + "name": "clip_name", + "type": "COMBO", + "widget": { + "name": "clip_name" + }, + "link": 81 + }, + { + "localized_name": "type", + "name": "type", + "type": "COMBO", + "widget": { + "name": "type" + }, + "link": null + }, + { + "localized_name": "device", + "name": "device", + "shape": 7, + "type": "COMBO", + "widget": { + "name": "device" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "CLIP", + "name": "CLIP", + "type": "CLIP", + "links": [ + 40 + ] + } + ], + "properties": { + "Node name for S&R": "CLIPLoader", + "cnr_id": "comfy-core", + "ver": "0.18.1", + "models": [ + { + "name": "qwen3.5_2b_bf16.safetensors", + "url": "https://huggingface.co/Comfy-Org/Qwen3.5/resolve/main/text_encoders/qwen3.5_2b_bf16.safetensors", + "directory": "text_encoders" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "qwen3.5_2b_bf16.safetensors", + "stable_diffusion", + "default" + ] + }, + { + "id": 6, + "type": "CLIPTextEncode", + "pos": [ + 610, + 130 + ], + "size": [ + 450, + 240 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 34 + }, + { + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": 49 + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "slot_index": 0, + "links": [ + 4 + ] + } + ], + "properties": { + "Node name for S&R": "CLIPTextEncode", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "" + ], + "color": "#232", + "bgcolor": "#353" + }, + { + "id": 34, + "type": "ComfySwitchNode", + "pos": [ + 210, + 610 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 10, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 47 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 46 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 48 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 49 + ] + } + ], + "properties": { + "Node name for S&R": "ComfySwitchNode" + }, + "widgets_values": [ + false + ] + }, + { + "id": 41, + "type": "ComfyMathExpression", + "pos": [ + 1370, + 1360 + ], + "size": [ + 230, + 80 + ], + "flags": { + "collapsed": true + }, + "order": 16, + "mode": 0, + "inputs": [ + { + "label": "a", + "localized_name": "values.a", + "name": "values.a", + "type": "FLOAT,INT,BOOLEAN", + "link": 56 + }, + { + "label": "b", + "localized_name": "values.b", + "name": "values.b", + "shape": 7, + "type": "FLOAT,INT,BOOLEAN", + "link": null + }, + { + "localized_name": "expression", + "name": "expression", + "type": "STRING", + "widget": { + "name": "expression" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "FLOAT", + "name": "FLOAT", + "type": "FLOAT", + "links": null + }, + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 57 + ] + }, + { + "localized_name": "BOOL", + "name": "BOOL", + "type": "BOOLEAN", + "links": null + } + ], + "properties": { + "Node name for S&R": "ComfyMathExpression" + }, + "widgets_values": [ + "a" + ] + }, + { + "id": 42, + "type": "PreviewAny", + "pos": [ + 1370, + 1310 + ], + "size": [ + 230, + 40 + ], + "flags": { + "collapsed": true + }, + "order": 17, + "mode": 0, + "inputs": [ + { + "localized_name": "source", + "name": "source", + "type": "*", + "link": 57 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 58 + ] + } + ], + "properties": { + "Node name for S&R": "PreviewAny" + }, + "widgets_values": [ + null, + null, + null + ] + }, + { + "id": 39, + "type": "StringReplace", + "pos": [ + 1040, + 900 + ], + "size": [ + 270, + 280 + ], + "flags": {}, + "order": 14, + "mode": 0, + "inputs": [ + { + "localized_name": "string", + "name": "string", + "type": "STRING", + "widget": { + "name": "string" + }, + "link": 52 + }, + { + "localized_name": "find", + "name": "find", + "type": "STRING", + "widget": { + "name": "find" + }, + "link": null + }, + { + "localized_name": "replace", + "name": "replace", + "type": "STRING", + "widget": { + "name": "replace" + }, + "link": 53 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 59 + ] + } + ], + "title": "Text Replace (USER INPUT)", + "properties": { + "Node name for S&R": "StringReplace" + }, + "widgets_values": [ + "", + "USER_INPUT", + "" + ] + }, + { + "id": 28, + "type": "TextGenerate", + "pos": [ + 1200, + 1580 + ], + "size": [ + 430, + 420 + ], + "flags": {}, + "order": 7, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 40 + }, + { + "localized_name": "image", + "name": "image", + "shape": 7, + "type": "IMAGE", + "link": null + }, + { + "localized_name": "video", + "name": "video", + "shape": 7, + "type": "IMAGE", + "link": null + }, + { + "localized_name": "audio", + "name": "audio", + "shape": 7, + "type": "AUDIO", + "link": null + }, + { + "localized_name": "prompt", + "name": "prompt", + "type": "STRING", + "widget": { + "name": "prompt" + }, + "link": 60 + }, + { + "localized_name": "max_length", + "name": "max_length", + "type": "INT", + "widget": { + "name": "max_length" + }, + "link": null + }, + { + "localized_name": "sampling_mode", + "name": "sampling_mode", + "type": "COMFY_DYNAMICCOMBO_V3", + "widget": { + "name": "sampling_mode" + }, + "link": null + }, + { + "localized_name": "temperature", + "name": "sampling_mode.temperature", + "type": "FLOAT", + "widget": { + "name": "sampling_mode.temperature" + }, + "link": null + }, + { + "localized_name": "top_k", + "name": "sampling_mode.top_k", + "type": "INT", + "widget": { + "name": "sampling_mode.top_k" + }, + "link": null + }, + { + "localized_name": "top_p", + "name": "sampling_mode.top_p", + "type": "FLOAT", + "widget": { + "name": "sampling_mode.top_p" + }, + "link": null + }, + { + "localized_name": "min_p", + "name": "sampling_mode.min_p", + "type": "FLOAT", + "widget": { + "name": "sampling_mode.min_p" + }, + "link": null + }, + { + "localized_name": "repetition_penalty", + "name": "sampling_mode.repetition_penalty", + "type": "FLOAT", + "widget": { + "name": "sampling_mode.repetition_penalty" + }, + "link": null + }, + { + "localized_name": "seed", + "name": "sampling_mode.seed", + "type": "INT", + "widget": { + "name": "sampling_mode.seed" + }, + "link": null + }, + { + "localized_name": "presence_penalty", + "name": "sampling_mode.presence_penalty", + "shape": 7, + "type": "FLOAT", + "widget": { + "name": "sampling_mode.presence_penalty" + }, + "link": null + }, + { + "localized_name": "thinking", + "name": "thinking", + "shape": 7, + "type": "BOOLEAN", + "widget": { + "name": "thinking" + }, + "link": null + }, + { + "localized_name": "use_default_template", + "name": "use_default_template", + "shape": 7, + "type": "BOOLEAN", + "widget": { + "name": "use_default_template" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "generated_text", + "name": "generated_text", + "type": "STRING", + "links": [ + 46, + 84 + ] + } + ], + "properties": { + "Node name for S&R": "TextGenerate", + "cnr_id": "comfy-core", + "ver": "0.18.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "", + 256, + "on", + 0.7, + 64, + 0.95, + 0.05, + 1.05, + 0, + 0, + false, + true + ] + }, + { + "id": 31, + "type": "PrimitiveStringMultiline", + "pos": [ + -390, + 160 + ], + "size": [ + 400, + 200 + ], + "flags": {}, + "order": 9, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "STRING", + "widget": { + "name": "value" + }, + "link": 68 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 47, + 53 + ] + } + ], + "title": "User: short description (USER_INPUT in template)", + "properties": { + "Node name for S&R": "PrimitiveStringMultiline" + }, + "widgets_values": [ + "" + ] + }, + { + "id": 43, + "type": "CustomCombo", + "pos": [ + 140, + 910 + ], + "size": [ + 550, + 320 + ], + "flags": {}, + "order": 18, + "mode": 0, + "inputs": [ + { + "localized_name": "choice", + "name": "choice", + "type": "COMBO", + "widget": { + "name": "choice" + }, + "link": 78 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 65 + ] + }, + { + "localized_name": "INDEX", + "name": "INDEX", + "type": "INT", + "links": null + } + ], + "title": "Custom Combo (Category index)", + "properties": { + "Node name for S&R": "CustomCombo" + }, + "widgets_values": [ + "Music", + 0, + "Music", + "Instrument", + "SFX", + "One-shot", + "" + ] + }, + { + "id": 49, + "type": "JsonExtractString", + "pos": [ + 720, + 1200 + ], + "size": [ + 300, + 180 + ], + "flags": {}, + "order": 19, + "mode": 0, + "inputs": [ + { + "localized_name": "json_string", + "name": "json_string", + "type": "STRING", + "widget": { + "name": "json_string" + }, + "link": null + }, + { + "localized_name": "key", + "name": "key", + "type": "STRING", + "widget": { + "name": "key" + }, + "link": 65 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 66 + ] + } + ], + "properties": { + "Node name for S&R": "JsonExtractString" + }, + "widgets_values": [ + "{\n \"Music\": \"You are an expert musician and musicologist and prompt engineer. Transform the user's input into a detailed, vivid music prompt for a full instrumental track.\\n\\n1. Start with the genre or style and optional adjectives (e.g., upbeat, dreamy, aggressive).\\n2. List the main instruments that define the track.\\n3. Add supporting elements or layers such as pads, harmonics, effects, or field recordings.\\n4. Include rhythm or percussion elements like drums, hi-hats, congas, brushes, or polyrhythms.\\n5. Integrate mood and energy naturally in the sentence (e.g., \\\"creating suspenseful tension\\\" or \\\"bright and uplifting\\\").\\n6. Specify the BPM.\\n7. Specify the track length as an integer in seconds. Use ranges: energetic/dance 120-180s, pop/rock 180-210s, cinematic/ambient 240-300s.\\n8. Combine all elements into one natural, fluid sentence. Avoid semicolons.\\n\\nTemplate:\\nGenre/Style with main instruments, supporting instruments/layers, and rhythm/percussion creating mood/energy. BPM: X. Length: Y seconds\\n\\nExamples:\\n- Jazz ballad with smooth saxophone lead, piano chords, upright bass, brushed drums, and soft strings that swing gently for a warm and cozy evening. BPM: 85. Length: 180 seconds\\n- EDM festival track with pulsing synth leads, plucked arpeggios, layered pads, side-chained bass, punchy kick and snare, and hi-hat rolls creating bright, energetic, and uplifting dance energy. BPM: 128. Length: 150 seconds\\n- Lo-fi hip-hop chill track with mellow electric piano, soft vinyl crackle, subtle synth pads, low-pass filtered drums, percussion loops, and soft plucked bass for a relaxed, dreamy vibe. BPM: 75. Length: 150 seconds\\n- Heavy metal anthem with distorted electric guitars, bass guitar, double bass drums, and cymbal crashes with fast palm-muted riffs creating intense, aggressive energy. BPM: 160. Length: 180 seconds\\n- Melancholic piano piece with soft piano lead, string pads, subtle atmospheric synths, and minimal brush percussion evoking a reflective rainy-day feeling. BPM: 60. Length: 240 seconds\\n- Suspenseful electronic thriller with pulsing bass synth, arpeggiated lead synth, cinematic pads, glitchy percussion, and high string stabs creating dark and tense energy. BPM: 100. Length: 200 seconds\\n- Dreamy ambient soundscape with layered pads, soft bell textures, gentle drones, and wind and water field recordings for ethereal and spacious meditation. BPM: 40. Length: 300 seconds\\n- Fingerpicking acoustic guitar solo with harmonics, subtle reverb, occasional shaker and soft stomp percussion, and soft pad layers for warm intimate storytelling. BPM: 70. Length: 120 seconds\\n- Synthwave 80s retro track with arpeggiated synth leads, analog pads, electric bass, punchy electronic drums, gated reverb snares, and atmospheric FX for nostalgic and vibrant energy. BPM: 110. Length: 180 seconds\\n- Tribal percussion ensemble with congas, djembes, bongos, shakers, and frame drums layered with deep synthetic sub-bass in complex polyrhythms. BPM: 100. Length: 140 seconds\\n- 1920s swing jazz with brass section, upright bass, piano, brushed drums, banjo, clarinet, and soft strings that swing lively for energetic dance vibes. BPM: 110. Length: 180 seconds\\n- Futuristic electronic sci-fi track with pulsing bass synth, evolving lead synths, layered pads, glitch percussion, robotic FX, and sub-bass for tense cinematic energy. BPM: 125. Length: 200 seconds\\n- Ambient underwater soundscape with flowing water textures, soft piano motifs, synth drones, distant bells, and underwater reverb for spacious meditative immersion. BPM: 45. Length: 300 seconds\\n- Horror cinematic track with dissonant strings, eerie piano stabs, cinematic percussion including taiko and low toms, and synth FX producing suspenseful creepy tension. BPM: 90. Length: 240 seconds\\n- Reggae track with offbeat guitar, warm basslines, snare, kick, congas, and horn stabs giving laid-back groovy energy. BPM: 85. Length: 150 seconds\\n- Blues track with soulful electric guitar solos, walking bass, piano, and shuffle drums creating expressive and emotive storytelling. BPM: 90. Length: 180 seconds\\n- Latin salsa with congas, timbales, horns, piano montunos, bass, and layered percussion for vibrant danceable energy. BPM: 120. Length: 210 seconds\\n- Afrobeat track with electric guitar stabs, horns, layered percussion, congas, shakers, bass groove, and synth pads for vibrant rhythmic energy. BPM: 105. Length: 200 seconds\\n- Indie rock track with electric guitar riffs, bass, live drum kit, layered synths, and subtle strings for energetic yet emotional feel. BPM: 110. Length: 180 seconds\\n- Funk groove with slap bass, electric guitar chords, brass stabs, drums, congas, and rhythmic keyboards creating high-energy danceable rhythm. BPM: 105. Length: 180 seconds\\n- Drum and bass track with fast breakbeat drums, deep sub-bass, sharp synth leads, pads, and atmospheric FX for high-energy club motion. BPM: 175. Length: 150 seconds\\n- Dark ambient track with drones, distant bells, low rumbles, soft wind textures, and synth pads producing eerie immersive tension. BPM: 50. Length: 300 seconds\\n- Tropical house track with marimba, steel drums, soft synths, smooth bass, layered percussion, and light piano riffs for sunny chill dance vibes. BPM: 110. Length: 180 seconds\\n- Progressive rock track with electric guitar leads, organ, bass, drum kit, synth layers, and occasional strings for epic layered energy. BPM: 100. Length: 220 seconds\\n- Music box melody with delicate metallic tones and soft resonance, lullaby style, with gentle ambient reverb. BPM: 60. Length: 20 seconds\\n- Soft piano arpeggio with warm felted tone and slow attack, lullaby style, with intimate room ambience. BPM: 60. Length: 30 seconds\\n- Harp gentle plucked pattern with airy resonance, lullaby style, with dreamy reverb tail. BPM: 65. Length: 25 seconds\\n- Acoustic guitar fingerstyle pattern with warm nylon strings and soft dynamics, lullaby style, with subtle room resonance. BPM: 60. Length: 30 seconds\\n- Ambient synth pad with smooth evolving texture and soft harmonics, lullaby style, with wide stereo ambience. BPM: 50. Length: 40 seconds\\n- Early rock piano with walking left-hand bass line, shuffle rhythms, and blues scale improvisations in energetic 1950s boogie-woogie style. BPM: 160. Length: 180 seconds\\n- Trip Hop track with jazzy sampled vibraphone, mid-tempo breakbeat drums, harp, Latin ethnic percussion, and sweeping cinematic strings creating airy, relaxing, soulful lounge vibes. BPM: 90. Length: 180 seconds\\n- Country outlaw cinematic instrumental with blues pedal steel guitar, rustic mandolin, fiddle call-and-response, tape-driven rattly drum kit, autoharp, and soaring accordion solo for raw, emotional southern blues expression. BPM: 85. Length: 200 seconds\\n- Neo Classical track with sweeping string section, elegant horns, and delicate piano creating soothing, hypnotic, modern, soft, and classic mood. BPM: 70. Length: 180 seconds\\n- Art Rock desert track with desolate piano chords, western-themed rhythm guitars, unique lead guitars, rattly vintage drum kit, and supporting bass creating lonely, expansive, beautiful, and strange atmospheres. BPM: 95. Length: 180 seconds\\n- Cinematic Sci-Fi score with dramatic horn section, building marcato strings, gliding bassoon, thunderous cymbals, subdued timpani, and subtle synth drones producing awe-inspiring, uplifting, epic intergalactic energy. BPM: 100. Length: 220 seconds\\n- West Coast Hip Hop instrumental with cascading harp melodies, smooth Rhodes piano chops, vintage boom bap drums, and walking double bass producing raw, street, and soulful block-party vibes. BPM: 92. Length: 180 seconds\\n- Synthwave futuristic track with pulsating synth bass, exciting chords, soaring leads, and reverberating drum machine patterns creating gritty, pounding, and cool energy. BPM: 110. Length: 180 seconds\\n- Breakbeat track with complex percussion, intricate breakbeats, gritty synths, lush pads, and 808 bassline producing fresh, modern, futuristic, and rave-ready energy. BPM: 140. Length: 160 seconds\\n- Lounge Jazz 1960s smooth track with laid-back drums, piano chords, double bass, soft electric piano, subtle flute, and unique percussion creating beautiful, atmospheric, eclectic, retro, and chill vibes. BPM: 85. Length: 180 seconds\\n- Latin Jazz 1950s blissful track with laid-back Latin drums, euphoric piano chords, double bass, orchestral accompaniment, acoustic guitar, and vibraphone producing nostalgic, beautiful, atmospheric, cinematic, and chill mood. BPM: 95. Length: 180 seconds\\n- Acid Jazz 1970s summertime track with smooth electric piano, trippy synth leads, laid-back vintage drum kit, fuzzy electric bass, and uplifting violin producing retro, psychedelic, jazzy, relaxing energy. BPM: 100. Length: 180 seconds\\n- Progressive Soul 1970s track with feel-good piano, psychedelic organ, groovy vintage drum kit with percussion, fuzzy electric bass, and synth strings producing retro, raw, soulful, joyous atmosphere. BPM: 90. Length: 180 seconds\\n- Discotheque 1970s French-inspired track with sultry piano, psychedelic guitars, groovy drum kit, fuzzy electric bass, and melancholic organ producing retro, raw, laid-back, and relaxing mood. BPM: 105. Length: 180 seconds\\n- Soul Jazz 1970s track with expressive saxophone, smooth piano, groovy drum kit, rhythmic upright bass, sweeping strings, and minimal vibraphone producing retro, raw, laid-back, and epic energy. BPM: 95. Length: 180 seconds\\n- Vintage R&B 1970s live studio track with subtle brass, smooth piano, sweeping strings, and minimal drums producing retro, beautiful, uplifting, nostalgic mood. BPM: 85. Length: 180 seconds\\n- 50s Pop track with Latin influence, string section, bold brass, vibraphone, acoustic guitar, flute, ethnic percussion, and brushed drums creating sexy, epic, vintage, retro, melancholic, jazzy, dramatic energy. BPM: 100. Length: 180 seconds\\n- A piece of calm, quiet, mellow, serene music perfect for a peaceful film score, featuring soft modulating piano, ambient sfx and foley, beautiful vibraphone, and subtle synthesizer drones. The mood is cinematic, thoughtful, serene and nostalgic. BPM: 55. Length: 300 seconds\",\n \"Instrument\": \"You are a music metadata expert. Given an instrument, generate a descriptive prompt for a generative audio model.\\n\\n1. Identify the instrument.\\n2. Add playing style or technique.\\n3. Include details about material, timbre, or texture.\\n4. Add musical style or mood. Specify the genre, context, or emotional character.\\n5. Add spatial or production qualities.\\n6. Specify BPM: Always include a BPM appropriate to the style and context.\\n7. Specify length: Provide an integer in seconds (6–20 s for loops, 20–180 s for stems).\\n\\nExamples:\\n- Synth arpeggio loop with bright detuned oscillators. BPM: 120. Length: 8 seconds\\n- Chord stab loop with sharp percussive attack. BPM: 90. Length: 6 seconds\\n- Guitar muted strum loop with tight rhythmic feel. BPM: 100. Length: 8 seconds\\n- Pluck sequence loop with bright resonant tone. BPM: 128. Length: 10 seconds\\n- Marimba and vibraphone percussive loop with resonant wooden and metallic tones. BPM: 110. Length: 12 seconds\\n- Drum loop with deep muffled kick on beat one, snappy rimshot snare on beats two and four with rolling ghost note fills, and tight closed hi-hats with subtle open accents. BPM: 85. Length: 10 seconds\\n- Drum groove loop with brushed snare swinging on the ride, soft feathered kick on downbeats, and light closed hi-hat taps on the upbeats. BPM: 130. Length: 12 seconds\\n- Kick and hi-hat loop with four-on-the-floor punchy kick, tight closed hi-hats on every eighth note, and a sharp dry snare on beats two and four. BPM: 130. Length: 15 seconds\\n- Vinyl crackle drum loop with warm low-pass filtered kick, dusty snare with tape saturation, and shuffled closed hi-hats with subtle vinyl crackle ambiance. BPM: 80. Length: 10 seconds\\n- Ambient pad loop with evolving texture. BPM: 80. Length: 12 seconds\\n- Melodic synth bass groove loop with pumping sidechain feel. BPM: 122. Length: 10 seconds\\n- Melodic Bass slap and pop rhythm loop. BPM: 100. Length: 8 seconds\\n- Acoustic bass walking line loop with natural wooden resonance. BPM: 120. Length: 12 seconds\\n- String pizzicato motif loop, suspenseful, with tight string texture. BPM: 90. Length: 8 seconds\\n- Brass staccato riff loop with sharp bright attack. BPM: 130. Length: 10 seconds\\n- Flute airy melodic loop with wooden headjoint resonance. BPM: 100. Length: 6 seconds\\n- Pan flute ambient loop with breathy timbre. BPM: 75. Length: 8 seconds\\n- Clarinet riff loop with warm smooth reed tone. BPM: 120. Length: 10 seconds\\n- Oboe motif loop, orchestral, with rich double reed resonance. BPM: 80. Length: 8 seconds\\n- Recorder Renaissance motif loop with soft wooden timbre. BPM: 100. Length: 6 seconds\\n- Electric sitar riff loop with buzzing resonant tone. BPM: 90. Length: 10 seconds\\n- Koto plucked motif loop with resonant wooden strings. BPM: 90. Length: 8 seconds\\n- Shamisen folk melody loop with percussive twang. BPM: 100. Length: 8 seconds\\n- Banjo fingerpicking loop with metallic string resonance. BPM: 110. Length: 10 seconds\\n- Mandolin tremolo loop with crisp wooden body tone. BPM: 120. Length: 10 seconds\\n- Acoustic guitar chord vamp loop with natural room resonance. BPM: 110. Length: 12 seconds\\n- Nylon string guitar arpeggio loop with warm, soft timbre. BPM: 90. Length: 15 seconds\\n- Electric guitar riff loop with driven distorted tone. BPM: 130. Length: 10 seconds\\n- Slide guitar melody loop with warm resonant glide. BPM: 100. Length: 12 seconds\\n- Steel guitar slide loop with bright pedal steel tone. BPM: 95. Length: 12 seconds\\n- Harpsichord arpeggio loop with crisp plucked attack. BPM: 120. Length: 10 seconds\\n- Rhodes chord vamp loop with warm electric piano tone. BPM: 100. Length: 12 seconds\\n- Clavinet funky rhythm loop. BPM: 105. Length: 10 seconds\\n- Organ chord vamp loop with full drawbar warmth. BPM: 90. Length: 12 seconds\\n- Drum loop with booming 808 kick on beat one, crisp snare on beat three, and rapid triplet hi-hat rolls with open hat accents for aggressive high-energy feel. BPM: 140. Length: 8 seconds\\n- Breakbeat drum loop with chopped Amen-style snare flurries, driving kick on the one, fast sixteenth-note closed hi-hats, and syncopated open hat accents. BPM: 170. Length: 10 seconds\\n- Glitch percussion loop with stuttered kick transients, randomised snare hits processed with bit-crushing, and erratic hi-hat patterns with pitch-shifted metallic ticks. BPM: 120. Length: 12 seconds\\n- Metallic hits loop with distorted kick impacts, processed metal-plate snare slams, and grinding hi-hat noise bursts for aggressive mechanical texture. BPM: 120. Length: 10 seconds\\n- Timpani hits loop, cinematic, with deep resonant kick-like timpani strikes on beat one, rolling snare-style timpani fills, and no hi-hats for a grand orchestral feel. BPM: 70. Length: 8 seconds\\n- Snare roll loop, dramatic, with accelerating snare drum rolls building from soft to crashing, deep supporting kick pulses, and no hi-hats for maximum impact. BPM: 100. Length: 8 seconds\\n- Accordion motif loop with bright reedy bellows tone. BPM: 100. Length: 10 seconds\\n- Harmonica blues riff loop with expressive reed timbre. BPM: 90. Length: 10 seconds\\n- Trombone riff loop with warm sliding brass tone. BPM: 120. Length: 10 seconds\\n- French horn melodic loop, cinematic. BPM: 80. Length: 12 seconds\\n- Soprano sax ballad loop. BPM: 70. Length: 12 seconds\\n- Alto sax bebop riff loop. BPM: 200. Length: 10 seconds\\n- Electric violin melodic loop with reverb. BPM: 90. Length: 10 seconds\\n- String pad loop with cinematic texture. BPM: 70. Length: 15 seconds\\n- Granular synth evolving texture loop. BPM: 90. Length: 15 seconds\\n- Piano motif loop with soft felt hammer tone. BPM: 80. Length: 10 seconds\\n- Pad and synth loop with lush detuned shimmer. BPM: 85. Length: 12 seconds\\n- Synth lead loop with sidechain pumping compression. BPM: 128. Length: 10 seconds\\n- Analog synth bassline loop with deep warm low-end. BPM: 122. Length: 12 seconds\\n- FM synth lead motif loop with bright metallic shimmer. BPM: 110. Length: 10 seconds\\n- Bass groove loop with tight rhythmic two-bar pattern. BPM: 100. Length: 16 seconds\\n- Acoustic guitar fingerstyle motif loop with warm wood resonance. BPM: 90. Length: 45 seconds\\n- Sombre acoustic guitar motif loop with cavernous reverb, delicate fingerpicking, and expressive melancholic tone. BPM: 70. Length: 45 seconds\\n- Electric guitar rock riff motif loop. BPM: 130. Length: 40 seconds\\n- Vintage electric guitar motif loop, live-recorded in a vintage studio, with expressive and dynamic solo performance. BPM: 90. Length: 40 seconds\\n- Piano chord progression motif loop with rich harmonic movement. BPM: 120. Length: 60 seconds\\n- String ensemble cinematic motif loop with rich wooden resonance. BPM: 80. Length: 120 seconds\\n- Brass ensemble cinematic motif loop with bright metallic timbre. BPM: 90. Length: 90 seconds\\n- Ethnic percussion ensemble motif loop with deep resonant djembe kick tones, slapped snare-like rim hits on congas, and layered shakers and bells providing hi-hat-like rhythmic texture with polyrhythmic patterns. BPM: 100. Length: 90 seconds\\n- Synth ambient motif loop with evolving textures. BPM: 80. Length: 180 seconds\\n- Motif loop with warm dusty vinyl crackle and tape saturation. BPM: 80. Length: 60 seconds\\n- Synth lead and bass motif loop with bright punchy energy. BPM: 128. Length: 90 seconds\\n- Funk band motif loop: bass, drums, guitar. BPM: 100. Length: 90 seconds\\n- Ethnic flute motif for cinematic use. BPM: 80. Length: 30 seconds\\n- Steel drum melodic motif loop with bright metallic resonance. BPM: 110. Length: 20 seconds\\n- Marimba percussive motif loop with resonant wooden tone. BPM: 100. Length: 20 seconds\\n- Vibraphone melodic motif loop with metallic shimmer. BPM: 90. Length: 25 seconds\\n- Piano cinematic motif loop with resonant wooden tone. BPM: 80. Length: 30 seconds\\n- Violin expressive cinematic motif loop with rich wooden resonance. BPM: 75. Length: 25 seconds\\n- Cello expressive motif loop with deep wooden resonance. BPM: 70. Length: 30 seconds\\n- Trumpet expressive motif loop with brassy overtones. BPM: 100. Length: 25 seconds\\n- Sax expressive motif loop with warm reed timbre. BPM: 95. Length: 25 seconds\\n- Ethnic drum ensemble motif loop with booming natural-skin bass drum kicks, sharp hand-slap snare accents on djembes and talking drums, and layered wooden and metal percussion providing rhythmic hi-hat-like patterns. BPM: 95. Length: 30 seconds\\n- Ambient drone motif loop. BPM: 60. Length: 180 seconds\\n- Orchestral tension motif loop. BPM: 90. Length: 150 seconds\\n- Electronic track motif loop with drums, bass, synth. BPM: 128. Length: 180 seconds\",\n \"SFX\": \"You are a professional sound design expert. Convert the user's input into a precise, vivid sound effects description suitable for generative audio models.\\n\\nDescribe clearly:\\n- Sound source\\n- Physical character (texture, timbre, material: metal, wood, glass, concrete, etc.)\\n- Spatial qualities (indoor/outdoor, cave/open field/underwater, dry/reverberant, close-up/distant, echoing/muffled)\\n- Temporal evolution (attack, decay, movement, transitions over time)\\n- Include motion or spatial movement if applicable (passing, approaching, stereo movement)\\n\\nAudio length rules:\\n- Very short sounds (impacts, clicks, gunshots): 1–3 seconds\\n- Medium actions (footsteps, object movement, transitions): 3–6 seconds\\n- Ambience / environments: 6–15 seconds\\n- Always append: Length: X seconds (integer only, no decimals).\\n\\nOutput constraints:\\n- Length: 1–2 dense sentences maximum\\n- Output ONLY the final rewritten prompt\\n- No explanations, no formatting, no quotes\\n- Use concise but dense technical language\\n- Focus strictly on sound effects or ambience\\n- Always append: Length: X seconds (integer only, no decimals).\\n\\nQuality guidelines:\\n- Be specific and avoid vague terms\\n- Prioritize clarity and realism\\n- Combine elements into one coherent scene\\n- Avoid redundancy\\n\\nExamples:\\n- Heavy rain hitting a metal roof during a thunderstorm, distant thunder rumbles, stereo, realistic ambience. Length: 45 seconds\\n- Quiet forest at dawn with birds chirping, soft wind through leaves, distant stream flowing. Length: 60 seconds\\n- Busy city street at night, cars passing, muffled conversations, occasional horn, urban ambience. Length: 50 seconds\\n- Ocean waves crashing against rocky cliffs, strong wind, dramatic and cinematic. Length: 70 seconds\\n- Wooden door creaking open slowly in an old house, echoing interior, eerie tone. Length: 3 seconds\\n- Glass bottle shattering on concrete, sharp impact, scattered fragments. Length: 2 seconds\\n- Footsteps on gravel, steady walking pace, close perspective. Length: 8 seconds\\n- Typing rapidly on a mechanical keyboard, crisp tactile clicks. Length: 5 seconds\\n- Punch impact with deep bass hit, cinematic trailer style. Length: 2 seconds\\n- Car speeding past at high velocity, doppler effect, realistic whoosh. Length: 3 seconds\\n- Object falling from height and hitting ground with a heavy thud. Length: 2 seconds\\n- Sword swing whooshing through air, fast motion, clean metallic tone. Length: 2 seconds\\n- Futuristic laser blast, clean energy pulse, high-tech sound design. Length: 1 seconds\\n- Spaceship engine humming, low frequency rumble, interior perspective. Length: 90 seconds\\n- Magical spell casting, shimmering particles, rising tonal energy. Length: 8 seconds\\n- Teleportation effect, glitchy digital distortion with a soft whoosh. Length: 5 seconds\\n- Dark eerie drone with distant whispers, creepy, slow build tension. Length: 120 seconds\\n- Sudden horror jump scare sting, sharp violin hit, cinematic. Length: 1 second\\n- Metal scraping slowly in a dark tunnel, echoing and ominous. Length: 20 seconds\\n- Explosion with debris scattering, deep bass, cinematic realism. Length: 4 seconds\\n- Building collapsing, rumbling concrete, dust and debris falling. Length: 25 seconds\\n- Fire crackling intensely, wood burning, close-up detail. Length: 80 seconds\\n- Gunshot in a large empty warehouse, loud echo decay. Length: 2 seconds\\n- Retro arcade coin insert sound, 8-bit style. Length: 1 second\\n- Level up chime, bright, rewarding, fantasy RPG style. Length: 2 seconds\\n- Error buzzer, short, digital, UI feedback. Length: 1 second\\n- Menu navigation clicks, soft futuristic interface sounds. Length: 3 seconds\\n- Layered soundscape: rain, thunder, footsteps, and distant sirens all blending naturally. Length: 90 seconds\\n- Rapid sequence of three impacts: metal hit, glass break, wood crack, spaced evenly. Length: 4 seconds\\n- Sound moving from left to right stereo field: passing motorcycle. Length: 5 seconds\\n- Close vs far perspective transition: footsteps approaching then fading away. Length: 6 seconds\\n- Tape stop sub drop, a massive sub-bass note that mimics a vinyl record or tape machine being turned off, the pitch and speed drop simultaneously, causing the high-end harmonics to smear and thicken as the sound grinds to a halt at a sub-sonic frequency. Length: 11 seconds\\n- Gravel and leaves footsteps, the sound of a hard boot stepping onto dry leaves or gravel, crisp and natural with detailed texture. Length: 11 seconds\\n- Ghostship moan, a massive, deep wooden groan with a low-frequency moan, like heavy timber under immense structural tension, swaying slowly, processed with long, dark wooden room reverb for a sense of scale. Length: 11 seconds\\n- Bicycle chain, a continuous metallic whirring sound of a chain moving over sprockets, with individual teeth catching the links, processed with resonant band-pass filter to emphasize metallic singing. Length: 11 seconds\\n- Warp drive, a sound that starts with a massive suck-back of ambient noise, followed by a supersonic crack and high-pitched zing that disappears into the distance, giving the sense of stretching space-time. Length: 11 seconds\\n- Ice cubes, high-pitched musical clinking of hard ice hitting a thin glass, bright resonant ring with subtle liquid sloshing around the edges. Length: 11 seconds\\n- Paper shuffle, the sound of a thick stack of heavy bond paper being squared up on a desk, dry papery thud with a quick fanning sound as air moves between the pages. Length: 11 seconds\\n- Drawer slam, a blunt, powerful thud made by slamming a wooden desk drawer shut, pronounced low-mid body, slightly distorted for aggressive character. Length: 3 seconds\",\n \"One-shot\": \"You are a music metadata expert. Given an instrument or sound, generate a descriptive prompt for a short, isolated one-shot audio sample for music production.\\n\\n1. Identify the instrument or sound source.\\n2. Describe the playing technique or hit type (e.g., pluck, slam, tap, stab).\\n3. Include details about material, timbre, or texture.\\n4. Add spatial or production qualities (dry/wet, room, close-mic).\\n5. Specify length: short integer in seconds (1–11 s).\\n\\nExamples:\\n- Piano key hit with bright percussive attack and resonant wooden body. Length: 2 seconds\\n- Kick drum punchy low-end hit with warm skin resonance. Length: 2 seconds\\n- Snare drum rimshot accent with crisp snare wires. Length: 2 seconds\\n- Acoustic guitar fingerstyle note with warm spruce tone. Length: 3 seconds\\n- Bass pluck with jazzy tone and resonant wooden body. Length: 3 seconds\\n- Electric guitar power chord with distortion. Length: 3 seconds\\n- Metallic glitch percussion hit with sharp metallic texture. Length: 2 seconds\\n- Tabla resonant tone hit with natural skin timbre. Length: 2 seconds\\n- Djembe slap accent with dry wooden resonance. Length: 2 seconds\\n- Synth stab with reverb tail. Length: 3 seconds\\n- Violin expressive note with vibrato and rich wooden resonance. Length: 3 seconds\\n- Cello legato note, cinematic, with warm resonant body. Length: 3 seconds\\n- Trumpet bright accent with slightly brassy overtones. Length: 2 seconds\\n- Melodic saxophone jazz riff with smooth reed timbre and a slight vibrato bend. Length: 3 seconds\\n- Harp pluck with airy tone and resonant strings. Length: 2 seconds\\n- Glockenspiel bell-like note with bright metallic clarity. Length: 2 seconds\\n- Metallic clang sound design hit. Length: 2 seconds\\n- Granular texture hit. Length: 3 seconds\\n- Reversed piano hit. Length: 2 seconds\\n- Synth riser effect. Length: 6 seconds\\n- Percussion impact hit. Length: 2 seconds\\n- Cinematic hit. Length: 2 seconds\\n- Dry clap, a crisp, natural single hand clap recorded in a dead room with an extremely sharp transient and no room reflections. Length: 1 second\\n- Studio hat, a classic, natural recording of 14-inch hi-hats played tightly closed, zero ring, very fast decay. Length: 1 second\\n- Disco open hat, bright 14-inch open hi-hat with long, shimmering decay, perfect for disco or dance grooves. Length: 1 second\\n- Pillow kick, acoustic kick drum muffled with a heavy blanket, producing a short, dry \\\"thump\\\" with almost zero resonance. Length: 1 second\\n- Short 808, punchy 808 kick with sharp, distorted transient and fast-decaying sub-tail. Length: 1 second\\n- Egg shaker, classic plastic egg shaker recorded with a small-diaphragm condenser mic, producing a light, consistent \\\"tick\\\" with very short sustain. Length: 1 second\\n- African drums, dynamic African drums and percussion ensemble with natural acoustic textures. Length: 3 seconds\\n- Latin drums, dynamic Latin drums and percussion ensemble featuring authentic rhythmic patterns. Length: 3 seconds\\n- String quartet, euphoric string quartet with dynamic and emotional playing, full of expressive harmonies and movement. Length: 3 seconds\\n- Piano, nostalgic, atmospheric piano piece with dynamic and emotional performance, intimate and resonant. Length: 3 seconds\\n- Analogue drift pad, warm polyphonic pad with three detuned oscillators (saw + triangle), subtle pitch drift, and lush bucket-brigade chorus for wide, nostalgic stereo image. Length: 11 seconds\\n- Phase distortion bass, Casio CZ-style phase-distorted sine wave warped into a jagged sawtooth for retro synth bass tone. Length: 11 seconds\\n- Vibrato saxophone, bright lyrical alto sax with fast fluttery vibrato, reedy vintage tone, captured with ribbon mic for warm nostalgic sound. Length: 11 seconds\\n- Lofi upright bass, upright bass recorded with ribbon mic in a wooden room, natural air with slightly boxy resonance, tape-saturated for dusty 1950s jazz feel. Length: 2 seconds\"\n}", + "Music" + ] + }, + { + "id": 40, + "type": "StringReplace", + "pos": [ + 1350, + 900 + ], + "size": [ + 260, + 280 + ], + "flags": {}, + "order": 15, + "mode": 0, + "inputs": [ + { + "localized_name": "string", + "name": "string", + "type": "STRING", + "widget": { + "name": "string" + }, + "link": 59 + }, + { + "localized_name": "find", + "name": "find", + "type": "STRING", + "widget": { + "name": "find" + }, + "link": null + }, + { + "localized_name": "replace", + "name": "replace", + "type": "STRING", + "widget": { + "name": "replace" + }, + "link": 58 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 60 + ] + } + ], + "title": "Text Replace (AUDIO LENGTH)", + "properties": { + "Node name for S&R": "StringReplace" + }, + "widgets_values": [ + "", + "AUDIO_LENGTH", + "" + ] + }, + { + "id": 38, + "type": "StringReplace", + "pos": [ + 720, + 900 + ], + "size": [ + 290, + 280 + ], + "flags": {}, + "order": 13, + "mode": 0, + "inputs": [ + { + "localized_name": "string", + "name": "string", + "type": "STRING", + "widget": { + "name": "string" + }, + "link": null + }, + { + "localized_name": "find", + "name": "find", + "type": "STRING", + "widget": { + "name": "find" + }, + "link": null + }, + { + "localized_name": "replace", + "name": "replace", + "type": "STRING", + "widget": { + "name": "replace" + }, + "link": 66 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 52 + ] + } + ], + "title": "Text Replace (PROMPT TEMPLATE)", + "properties": { + "Node name for S&R": "StringReplace" + }, + "widgets_values": [ + "SYSTEM_PROMPTS\n\nInput: USER_INPUT\nTarget audio length: AUDIO_LENGTH seconds.\nOutput:", + "SYSTEM_PROMPTS", + "" + ] + }, + { + "id": 35, + "type": "PrimitiveBoolean", + "pos": [ + -390, + 570 + ], + "size": [ + 400, + 100 + ], + "flags": {}, + "order": 11, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "BOOLEAN", + "widget": { + "name": "value" + }, + "link": 83 + } + ], + "outputs": [ + { + "localized_name": "BOOLEAN", + "name": "BOOLEAN", + "type": "BOOLEAN", + "links": [ + 48 + ] + } + ], + "title": "Boolean (Enable_Reprompt)", + "properties": { + "Node name for S&R": "PrimitiveBoolean" + }, + "widgets_values": [ + true + ] + }, + { + "id": 36, + "type": "PrimitiveFloat", + "pos": [ + -390, + 410 + ], + "size": [ + 400, + 110 + ], + "flags": {}, + "order": 12, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "FLOAT", + "widget": { + "name": "value" + }, + "link": 82 + } + ], + "outputs": [ + { + "localized_name": "FLOAT", + "name": "FLOAT", + "type": "FLOAT", + "links": [ + 50, + 56 + ] + } + ], + "title": "Float (Duration)", + "properties": { + "Node name for S&R": "PrimitiveFloat" + }, + "widgets_values": [ + 150 + ] + }, + { + "id": 25, + "type": "CheckpointLoaderSimple", + "pos": [ + 100, + 130 + ], + "size": [ + 440, + 190 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "localized_name": "ckpt_name", + "name": "ckpt_name", + "type": "COMBO", + "widget": { + "name": "ckpt_name" + }, + "link": 79 + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "links": [ + 30 + ] + }, + { + "localized_name": "CLIP", + "name": "CLIP", + "type": "CLIP", + "links": [] + }, + { + "localized_name": "VAE", + "name": "VAE", + "type": "VAE", + "links": [ + 39 + ] + } + ], + "properties": { + "Node name for S&R": "CheckpointLoaderSimple", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "models": [ + { + "name": "stable_audio_3_medium_base.safetensors", + "url": "https://huggingface.co/Comfy-Org/stable-audio-3/resolve/main/checkpoints/stable_audio_3_medium_base.safetensors", + "directory": "checkpoints" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "stable_audio_3_medium_base.safetensors" + ] + }, + { + "id": 26, + "type": "CLIPLoader", + "pos": [ + 100, + 390 + ], + "size": [ + 440, + 170 + ], + "flags": {}, + "order": 6, + "mode": 0, + "inputs": [ + { + "localized_name": "clip_name", + "name": "clip_name", + "type": "COMBO", + "widget": { + "name": "clip_name" + }, + "link": 80 + }, + { + "localized_name": "type", + "name": "type", + "type": "COMBO", + "widget": { + "name": "type" + }, + "link": null + }, + { + "localized_name": "device", + "name": "device", + "shape": 7, + "type": "COMBO", + "widget": { + "name": "device" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "CLIP", + "name": "CLIP", + "type": "CLIP", + "links": [ + 34, + 35 + ] + } + ], + "properties": { + "Node name for S&R": "CLIPLoader", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "models": [ + { + "name": "t5gemma_b_b_ul2.safetensors", + "url": "https://huggingface.co/Comfy-Org/stable-audio-3/resolve/main/text_encoders/t5gemma_b_b_ul2.safetensors", + "directory": "text_encoders" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "t5gemma_b_b_ul2.safetensors", + "stable_audio", + "default" + ] + }, + { + "id": 54, + "type": "PreviewAny", + "pos": [ + 1720, + 1580 + ], + "size": [ + 420, + 550 + ], + "flags": {}, + "order": 20, + "mode": 4, + "inputs": [ + { + "localized_name": "source", + "name": "source", + "type": "*", + "link": 84 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": null + } + ], + "properties": { + "Node name for S&R": "PreviewAny" + }, + "widgets_values": [ + null, + null, + null + ] + } + ], + "groups": [ + { + "id": 1, + "title": "Loaders: checkpoint & CLIP", + "bounding": [ + 80, + 50, + 485.721654232725, + 527.2848777754299 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 2, + "title": "CLIP encode: conditioning", + "bounding": [ + 600, + 60, + 470, + 510 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 3, + "title": "User inputs: prompt & duration", + "bounding": [ + -400, + 10, + 430, + 740 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 7, + "title": "Reprompt: full branch (template + LLM)", + "bounding": [ + 60, + 780, + 1630, + 1360 + ], + "color": "#444", + "flags": {} + }, + { + "id": 4, + "title": "Reprompt: JSON extract & template fills", + "bounding": [ + 120, + 820, + 1520, + 650 + ], + "color": "#444", + "flags": {} + }, + { + "id": 5, + "title": "Helpers: duration to string", + "bounding": [ + 1340, + 1180, + 280, + 250 + ], + "color": "#444", + "flags": {} + }, + { + "id": 6, + "title": "Reprompt: Qwen TextGenerate", + "bounding": [ + 680, + 1510, + 960, + 614.65625 + ], + "color": "#444", + "flags": {} + }, + { + "id": 8, + "title": "Audio generation: Stable Audio", + "bounding": [ + 60, + 10, + 1627.3616782294932, + 737.0545987464304 + ], + "color": "#3f789e", + "flags": {} + } + ], + "links": [ + { + "id": 35, + "origin_id": 26, + "origin_slot": 0, + "target_id": 7, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 13, + "origin_id": 3, + "origin_slot": 0, + "target_id": 12, + "target_slot": 0, + "type": "LATENT" + }, + { + "id": 39, + "origin_id": 25, + "origin_slot": 2, + "target_id": 12, + "target_slot": 1, + "type": "VAE" + }, + { + "id": 50, + "origin_id": 36, + "origin_slot": 0, + "target_id": 11, + "target_slot": 0, + "type": "FLOAT" + }, + { + "id": 30, + "origin_id": 25, + "origin_slot": 0, + "target_id": 3, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 4, + "origin_id": 6, + "origin_slot": 0, + "target_id": 3, + "target_slot": 1, + "type": "CONDITIONING" + }, + { + "id": 6, + "origin_id": 7, + "origin_slot": 0, + "target_id": 3, + "target_slot": 2, + "type": "CONDITIONING" + }, + { + "id": 12, + "origin_id": 11, + "origin_slot": 0, + "target_id": 3, + "target_slot": 3, + "type": "LATENT" + }, + { + "id": 34, + "origin_id": 26, + "origin_slot": 0, + "target_id": 6, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 49, + "origin_id": 34, + "origin_slot": 0, + "target_id": 6, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 47, + "origin_id": 31, + "origin_slot": 0, + "target_id": 34, + "target_slot": 0, + "type": "STRING" + }, + { + "id": 46, + "origin_id": 28, + "origin_slot": 0, + "target_id": 34, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 48, + "origin_id": 35, + "origin_slot": 0, + "target_id": 34, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 56, + "origin_id": 36, + "origin_slot": 0, + "target_id": 41, + "target_slot": 0, + "type": "FLOAT" + }, + { + "id": 57, + "origin_id": 41, + "origin_slot": 1, + "target_id": 42, + "target_slot": 0, + "type": "INT" + }, + { + "id": 52, + "origin_id": 38, + "origin_slot": 0, + "target_id": 39, + "target_slot": 0, + "type": "STRING" + }, + { + "id": 53, + "origin_id": 31, + "origin_slot": 0, + "target_id": 39, + "target_slot": 2, + "type": "STRING" + }, + { + "id": 40, + "origin_id": 29, + "origin_slot": 0, + "target_id": 28, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 60, + "origin_id": 40, + "origin_slot": 0, + "target_id": 28, + "target_slot": 4, + "type": "STRING" + }, + { + "id": 65, + "origin_id": 43, + "origin_slot": 0, + "target_id": 49, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 59, + "origin_id": 39, + "origin_slot": 0, + "target_id": 40, + "target_slot": 0, + "type": "STRING" + }, + { + "id": 58, + "origin_id": 42, + "origin_slot": 0, + "target_id": 40, + "target_slot": 2, + "type": "STRING" + }, + { + "id": 66, + "origin_id": 49, + "origin_slot": 0, + "target_id": 38, + "target_slot": 2, + "type": "STRING" + }, + { + "id": 27, + "origin_id": 12, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "AUDIO" + }, + { + "id": 68, + "origin_id": -10, + "origin_slot": 0, + "target_id": 31, + "target_slot": 0, + "type": "STRING" + }, + { + "id": 76, + "origin_id": -10, + "origin_slot": 2, + "target_id": 3, + "target_slot": 4, + "type": "INT" + }, + { + "id": 78, + "origin_id": -10, + "origin_slot": 4, + "target_id": 43, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 79, + "origin_id": -10, + "origin_slot": 5, + "target_id": 25, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 80, + "origin_id": -10, + "origin_slot": 6, + "target_id": 26, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 81, + "origin_id": -10, + "origin_slot": 7, + "target_id": 29, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 82, + "origin_id": -10, + "origin_slot": 1, + "target_id": 36, + "target_slot": 0, + "type": "FLOAT" + }, + { + "id": 83, + "origin_id": -10, + "origin_slot": 3, + "target_id": 35, + "target_slot": 0, + "type": "BOOLEAN" + }, + { + "id": 84, + "origin_id": 28, + "origin_slot": 0, + "target_id": 54, + "target_slot": 0, + "type": "STRING" + } + ], + "extra": {}, + "category": "Audio/Music generation", + "description": "Generates music, instrument loops, sound effects, and one-shots from text using the Stable Audio 3 Medium base checkpoint, with optional Qwen 3.5 category-based prompt expansion (Music, Instrument, SFX, One-shot)." + } + ] + }, + "extra": {} +} \ No newline at end of file diff --git a/blueprints/Audio Generation (Stable Audio 3 Medium).json b/blueprints/Audio Generation (Stable Audio 3 Medium).json new file mode 100644 index 000000000..30add5b05 --- /dev/null +++ b/blueprints/Audio Generation (Stable Audio 3 Medium).json @@ -0,0 +1,2091 @@ +{ + "revision": 0, + "last_node_id": 52, + "last_link_id": 0, + "nodes": [ + { + "id": 52, + "type": "8b66c757-fe2f-4184-91f3-479a19deb565", + "pos": [ + 370, + 1120 + ], + "size": [ + 420, + 450 + ], + "flags": { + "collapsed": false + }, + "order": 0, + "mode": 0, + "inputs": [ + { + "label": "user_input", + "name": "user_input", + "type": "STRING", + "widget": { + "name": "user_input" + }, + "link": null + }, + { + "label": "duration", + "name": "duration", + "type": "FLOAT", + "widget": { + "name": "duration" + }, + "link": null + }, + { + "label": "seed", + "name": "seed", + "type": "INT", + "widget": { + "name": "seed" + }, + "link": null + }, + { + "label": "use_reprompt", + "name": "use_reprompt", + "type": "BOOLEAN", + "widget": { + "name": "use_reprompt" + }, + "link": null + }, + { + "label": "reprompt_category", + "name": "category", + "type": "COMBO", + "widget": { + "name": "category" + }, + "link": null + }, + { + "label": "ckpt_name", + "name": "ckpt_name", + "type": "COMBO", + "widget": { + "name": "ckpt_name" + }, + "link": null + }, + { + "label": "sa_clip", + "name": "sa_clip", + "type": "COMBO", + "widget": { + "name": "sa_clip" + }, + "link": null + }, + { + "label": "qwen_clip", + "name": "qwen_clip", + "type": "COMBO", + "widget": { + "name": "qwen_clip" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "AUDIO", + "name": "AUDIO", + "type": "AUDIO", + "links": [] + } + ], + "title": "Audio Generation (Stable Audio 3 Medium)", + "properties": { + "proxyWidgets": [ + [ + "31", + "value" + ], + [ + "36", + "value" + ], + [ + "3", + "seed" + ], + [ + "35", + "value" + ], + [ + "43", + "choice" + ], + [ + "25", + "ckpt_name" + ], + [ + "26", + "clip_name" + ], + [ + "29", + "clip_name" + ] + ] + }, + "widgets_values": [] + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "8b66c757-fe2f-4184-91f3-479a19deb565", + "version": 1, + "state": { + "lastGroupId": 8, + "lastNodeId": 56, + "lastLinkId": 84, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Audio Generation (Stable Audio 3 Medium)", + "inputNode": { + "id": -10, + "bounding": [ + -810, + 400, + 155.953125, + 208 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + 1750, + 1041, + 128, + 68 + ] + }, + "inputs": [ + { + "id": "78ae2515-114b-494a-becc-43c7b6c2dc2f", + "name": "user_input", + "type": "STRING", + "linkIds": [ + 68 + ], + "label": "user_input", + "pos": [ + -678.046875, + 424 + ] + }, + { + "id": "5ca95030-aff4-4544-b545-f0d814e0e49a", + "name": "duration", + "type": "FLOAT", + "linkIds": [ + 82 + ], + "label": "duration", + "pos": [ + -678.046875, + 444 + ] + }, + { + "id": "718eb10f-da1a-4cea-a9c7-3040f98fe960", + "name": "seed", + "type": "INT", + "linkIds": [ + 76 + ], + "label": "seed", + "pos": [ + -678.046875, + 464 + ] + }, + { + "id": "dc020099-39e6-4009-9937-408409d71736", + "name": "use_reprompt", + "type": "BOOLEAN", + "linkIds": [ + 83 + ], + "label": "use_reprompt", + "pos": [ + -678.046875, + 484 + ] + }, + { + "id": "edae394c-6324-44d6-8ac5-d8caa5ae2169", + "name": "category", + "type": "COMBO", + "linkIds": [ + 78 + ], + "label": "reprompt_category", + "pos": [ + -678.046875, + 504 + ] + }, + { + "id": "be19b747-6a47-4028-9c30-d52f54a712ea", + "name": "ckpt_name", + "type": "COMBO", + "linkIds": [ + 79 + ], + "label": "ckpt_name", + "pos": [ + -678.046875, + 524 + ] + }, + { + "id": "bc9241a2-bc20-4c5d-8cb1-f2958f598642", + "name": "sa_clip", + "type": "COMBO", + "linkIds": [ + 80 + ], + "label": "sa_clip", + "pos": [ + -678.046875, + 544 + ] + }, + { + "id": "a33a2468-6d6d-4cb6-937c-3510bf16ebac", + "name": "qwen_clip", + "type": "COMBO", + "linkIds": [ + 81 + ], + "label": "qwen_clip", + "pos": [ + -678.046875, + 564 + ] + } + ], + "outputs": [ + { + "id": "bbe988dd-5c03-44fd-a965-c712f9204988", + "name": "AUDIO", + "type": "AUDIO", + "linkIds": [ + 27 + ], + "localized_name": "AUDIO", + "pos": [ + 1774, + 1065 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 7, + "type": "CLIPTextEncode", + "pos": [ + 620, + 420 + ], + "size": [ + 440, + 140 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 35 + }, + { + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "slot_index": 0, + "links": [ + 6 + ] + } + ], + "properties": { + "Node name for S&R": "CLIPTextEncode", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "" + ], + "color": "#223", + "bgcolor": "#335" + }, + { + "id": 12, + "type": "VAEDecodeAudio", + "pos": [ + 1450, + 110 + ], + "size": [ + 230, + 100 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "localized_name": "samples", + "name": "samples", + "type": "LATENT", + "link": 13 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 39 + } + ], + "outputs": [ + { + "localized_name": "AUDIO", + "name": "AUDIO", + "type": "AUDIO", + "slot_index": 0, + "links": [ + 27 + ] + } + ], + "properties": { + "Node name for S&R": "VAEDecodeAudio", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 11, + "type": "EmptyLatentAudio", + "pos": [ + 630, + 610 + ], + "size": [ + 430, + 140 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "seconds", + "name": "seconds", + "type": "FLOAT", + "widget": { + "name": "seconds" + }, + "link": 50 + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "widget": { + "name": "batch_size" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "LATENT", + "name": "LATENT", + "type": "LATENT", + "links": [ + 12 + ] + } + ], + "properties": { + "Node name for S&R": "EmptyLatentAudio", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 60, + 1 + ] + }, + { + "id": 3, + "type": "KSampler", + "pos": [ + 1100, + 100 + ], + "size": [ + 320, + 350 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 30 + }, + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "link": 4 + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "link": 6 + }, + { + "localized_name": "latent_image", + "name": "latent_image", + "type": "LATENT", + "link": 12 + }, + { + "localized_name": "seed", + "name": "seed", + "type": "INT", + "widget": { + "name": "seed" + }, + "link": 76 + }, + { + "localized_name": "steps", + "name": "steps", + "type": "INT", + "widget": { + "name": "steps" + }, + "link": null + }, + { + "localized_name": "cfg", + "name": "cfg", + "type": "FLOAT", + "widget": { + "name": "cfg" + }, + "link": null + }, + { + "localized_name": "sampler_name", + "name": "sampler_name", + "type": "COMBO", + "widget": { + "name": "sampler_name" + }, + "link": null + }, + { + "localized_name": "scheduler", + "name": "scheduler", + "type": "COMBO", + "widget": { + "name": "scheduler" + }, + "link": null + }, + { + "localized_name": "denoise", + "name": "denoise", + "type": "FLOAT", + "widget": { + "name": "denoise" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "LATENT", + "name": "LATENT", + "type": "LATENT", + "slot_index": 0, + "links": [ + 13 + ] + } + ], + "properties": { + "Node name for S&R": "KSampler", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 0, + "randomize", + 8, + 1, + "lcm", + "simple", + 1 + ] + }, + { + "id": 29, + "type": "CLIPLoader", + "pos": [ + 690, + 1580 + ], + "size": [ + 430, + 170 + ], + "flags": {}, + "order": 8, + "mode": 0, + "showAdvanced": false, + "inputs": [ + { + "localized_name": "clip_name", + "name": "clip_name", + "type": "COMBO", + "widget": { + "name": "clip_name" + }, + "link": 81 + }, + { + "localized_name": "type", + "name": "type", + "type": "COMBO", + "widget": { + "name": "type" + }, + "link": null + }, + { + "localized_name": "device", + "name": "device", + "shape": 7, + "type": "COMBO", + "widget": { + "name": "device" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "CLIP", + "name": "CLIP", + "type": "CLIP", + "links": [ + 40 + ] + } + ], + "properties": { + "Node name for S&R": "CLIPLoader", + "cnr_id": "comfy-core", + "ver": "0.18.1", + "models": [ + { + "name": "qwen3.5_2b_bf16.safetensors", + "url": "https://huggingface.co/Comfy-Org/Qwen3.5/resolve/main/text_encoders/qwen3.5_2b_bf16.safetensors", + "directory": "text_encoders" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "qwen3.5_2b_bf16.safetensors", + "stable_diffusion", + "default" + ] + }, + { + "id": 6, + "type": "CLIPTextEncode", + "pos": [ + 610, + 130 + ], + "size": [ + 450, + 240 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 34 + }, + { + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": 49 + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "slot_index": 0, + "links": [ + 4 + ] + } + ], + "properties": { + "Node name for S&R": "CLIPTextEncode", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "" + ], + "color": "#232", + "bgcolor": "#353" + }, + { + "id": 34, + "type": "ComfySwitchNode", + "pos": [ + 210, + 610 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 10, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 47 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 46 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 48 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 49 + ] + } + ], + "properties": { + "Node name for S&R": "ComfySwitchNode" + }, + "widgets_values": [ + false + ] + }, + { + "id": 41, + "type": "ComfyMathExpression", + "pos": [ + 1370, + 1360 + ], + "size": [ + 230, + 80 + ], + "flags": { + "collapsed": true + }, + "order": 16, + "mode": 0, + "inputs": [ + { + "label": "a", + "localized_name": "values.a", + "name": "values.a", + "type": "FLOAT,INT,BOOLEAN", + "link": 56 + }, + { + "label": "b", + "localized_name": "values.b", + "name": "values.b", + "shape": 7, + "type": "FLOAT,INT,BOOLEAN", + "link": null + }, + { + "localized_name": "expression", + "name": "expression", + "type": "STRING", + "widget": { + "name": "expression" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "FLOAT", + "name": "FLOAT", + "type": "FLOAT", + "links": null + }, + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 57 + ] + }, + { + "localized_name": "BOOL", + "name": "BOOL", + "type": "BOOLEAN", + "links": null + } + ], + "properties": { + "Node name for S&R": "ComfyMathExpression" + }, + "widgets_values": [ + "a" + ] + }, + { + "id": 42, + "type": "PreviewAny", + "pos": [ + 1370, + 1310 + ], + "size": [ + 230, + 40 + ], + "flags": { + "collapsed": true + }, + "order": 17, + "mode": 0, + "inputs": [ + { + "localized_name": "source", + "name": "source", + "type": "*", + "link": 57 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 58 + ] + } + ], + "properties": { + "Node name for S&R": "PreviewAny" + }, + "widgets_values": [ + null, + null, + null + ] + }, + { + "id": 39, + "type": "StringReplace", + "pos": [ + 1040, + 900 + ], + "size": [ + 270, + 280 + ], + "flags": {}, + "order": 14, + "mode": 0, + "inputs": [ + { + "localized_name": "string", + "name": "string", + "type": "STRING", + "widget": { + "name": "string" + }, + "link": 52 + }, + { + "localized_name": "find", + "name": "find", + "type": "STRING", + "widget": { + "name": "find" + }, + "link": null + }, + { + "localized_name": "replace", + "name": "replace", + "type": "STRING", + "widget": { + "name": "replace" + }, + "link": 53 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 59 + ] + } + ], + "title": "Text Replace (USER INPUT)", + "properties": { + "Node name for S&R": "StringReplace" + }, + "widgets_values": [ + "", + "USER_INPUT", + "" + ] + }, + { + "id": 28, + "type": "TextGenerate", + "pos": [ + 1200, + 1580 + ], + "size": [ + 430, + 420 + ], + "flags": {}, + "order": 7, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 40 + }, + { + "localized_name": "image", + "name": "image", + "shape": 7, + "type": "IMAGE", + "link": null + }, + { + "localized_name": "video", + "name": "video", + "shape": 7, + "type": "IMAGE", + "link": null + }, + { + "localized_name": "audio", + "name": "audio", + "shape": 7, + "type": "AUDIO", + "link": null + }, + { + "localized_name": "prompt", + "name": "prompt", + "type": "STRING", + "widget": { + "name": "prompt" + }, + "link": 60 + }, + { + "localized_name": "max_length", + "name": "max_length", + "type": "INT", + "widget": { + "name": "max_length" + }, + "link": null + }, + { + "localized_name": "sampling_mode", + "name": "sampling_mode", + "type": "COMFY_DYNAMICCOMBO_V3", + "widget": { + "name": "sampling_mode" + }, + "link": null + }, + { + "localized_name": "temperature", + "name": "sampling_mode.temperature", + "type": "FLOAT", + "widget": { + "name": "sampling_mode.temperature" + }, + "link": null + }, + { + "localized_name": "top_k", + "name": "sampling_mode.top_k", + "type": "INT", + "widget": { + "name": "sampling_mode.top_k" + }, + "link": null + }, + { + "localized_name": "top_p", + "name": "sampling_mode.top_p", + "type": "FLOAT", + "widget": { + "name": "sampling_mode.top_p" + }, + "link": null + }, + { + "localized_name": "min_p", + "name": "sampling_mode.min_p", + "type": "FLOAT", + "widget": { + "name": "sampling_mode.min_p" + }, + "link": null + }, + { + "localized_name": "repetition_penalty", + "name": "sampling_mode.repetition_penalty", + "type": "FLOAT", + "widget": { + "name": "sampling_mode.repetition_penalty" + }, + "link": null + }, + { + "localized_name": "seed", + "name": "sampling_mode.seed", + "type": "INT", + "widget": { + "name": "sampling_mode.seed" + }, + "link": null + }, + { + "localized_name": "presence_penalty", + "name": "sampling_mode.presence_penalty", + "shape": 7, + "type": "FLOAT", + "widget": { + "name": "sampling_mode.presence_penalty" + }, + "link": null + }, + { + "localized_name": "thinking", + "name": "thinking", + "shape": 7, + "type": "BOOLEAN", + "widget": { + "name": "thinking" + }, + "link": null + }, + { + "localized_name": "use_default_template", + "name": "use_default_template", + "shape": 7, + "type": "BOOLEAN", + "widget": { + "name": "use_default_template" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "generated_text", + "name": "generated_text", + "type": "STRING", + "links": [ + 46, + 84 + ] + } + ], + "properties": { + "Node name for S&R": "TextGenerate", + "cnr_id": "comfy-core", + "ver": "0.18.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "", + 256, + "on", + 0.7, + 64, + 0.95, + 0.05, + 1.05, + 0, + 0, + false, + true + ] + }, + { + "id": 31, + "type": "PrimitiveStringMultiline", + "pos": [ + -390, + 160 + ], + "size": [ + 400, + 200 + ], + "flags": {}, + "order": 9, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "STRING", + "widget": { + "name": "value" + }, + "link": 68 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 47, + 53 + ] + } + ], + "title": "User: short description (USER_INPUT in template)", + "properties": { + "Node name for S&R": "PrimitiveStringMultiline" + }, + "widgets_values": [ + "" + ] + }, + { + "id": 43, + "type": "CustomCombo", + "pos": [ + 140, + 910 + ], + "size": [ + 550, + 320 + ], + "flags": {}, + "order": 18, + "mode": 0, + "inputs": [ + { + "localized_name": "choice", + "name": "choice", + "type": "COMBO", + "widget": { + "name": "choice" + }, + "link": 78 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 65 + ] + }, + { + "localized_name": "INDEX", + "name": "INDEX", + "type": "INT", + "links": null + } + ], + "title": "Custom Combo (Category index)", + "properties": { + "Node name for S&R": "CustomCombo" + }, + "widgets_values": [ + "Music", + 0, + "Music", + "Instrument", + "SFX", + "One-shot", + "" + ] + }, + { + "id": 49, + "type": "JsonExtractString", + "pos": [ + 720, + 1200 + ], + "size": [ + 300, + 180 + ], + "flags": {}, + "order": 19, + "mode": 0, + "inputs": [ + { + "localized_name": "json_string", + "name": "json_string", + "type": "STRING", + "widget": { + "name": "json_string" + }, + "link": null + }, + { + "localized_name": "key", + "name": "key", + "type": "STRING", + "widget": { + "name": "key" + }, + "link": 65 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 66 + ] + } + ], + "properties": { + "Node name for S&R": "JsonExtractString" + }, + "widgets_values": [ + "{\n \"Music\": \"You are an expert musician and musicologist and prompt engineer. Transform the user's input into a detailed, vivid music prompt for a full instrumental track.\\n\\n1. Start with the genre or style and optional adjectives (e.g., upbeat, dreamy, aggressive).\\n2. List the main instruments that define the track.\\n3. Add supporting elements or layers such as pads, harmonics, effects, or field recordings.\\n4. Include rhythm or percussion elements like drums, hi-hats, congas, brushes, or polyrhythms.\\n5. Integrate mood and energy naturally in the sentence (e.g., \\\"creating suspenseful tension\\\" or \\\"bright and uplifting\\\").\\n6. Specify the BPM.\\n7. Specify the track length as an integer in seconds. Use ranges: energetic/dance 120-180s, pop/rock 180-210s, cinematic/ambient 240-300s.\\n8. Combine all elements into one natural, fluid sentence. Avoid semicolons.\\n\\nTemplate:\\nGenre/Style with main instruments, supporting instruments/layers, and rhythm/percussion creating mood/energy. BPM: X. Length: Y seconds\\n\\nExamples:\\n- Jazz ballad with smooth saxophone lead, piano chords, upright bass, brushed drums, and soft strings that swing gently for a warm and cozy evening. BPM: 85. Length: 180 seconds\\n- EDM festival track with pulsing synth leads, plucked arpeggios, layered pads, side-chained bass, punchy kick and snare, and hi-hat rolls creating bright, energetic, and uplifting dance energy. BPM: 128. Length: 150 seconds\\n- Lo-fi hip-hop chill track with mellow electric piano, soft vinyl crackle, subtle synth pads, low-pass filtered drums, percussion loops, and soft plucked bass for a relaxed, dreamy vibe. BPM: 75. Length: 150 seconds\\n- Heavy metal anthem with distorted electric guitars, bass guitar, double bass drums, and cymbal crashes with fast palm-muted riffs creating intense, aggressive energy. BPM: 160. Length: 180 seconds\\n- Melancholic piano piece with soft piano lead, string pads, subtle atmospheric synths, and minimal brush percussion evoking a reflective rainy-day feeling. BPM: 60. Length: 240 seconds\\n- Suspenseful electronic thriller with pulsing bass synth, arpeggiated lead synth, cinematic pads, glitchy percussion, and high string stabs creating dark and tense energy. BPM: 100. Length: 200 seconds\\n- Dreamy ambient soundscape with layered pads, soft bell textures, gentle drones, and wind and water field recordings for ethereal and spacious meditation. BPM: 40. Length: 300 seconds\\n- Fingerpicking acoustic guitar solo with harmonics, subtle reverb, occasional shaker and soft stomp percussion, and soft pad layers for warm intimate storytelling. BPM: 70. Length: 120 seconds\\n- Synthwave 80s retro track with arpeggiated synth leads, analog pads, electric bass, punchy electronic drums, gated reverb snares, and atmospheric FX for nostalgic and vibrant energy. BPM: 110. Length: 180 seconds\\n- Tribal percussion ensemble with congas, djembes, bongos, shakers, and frame drums layered with deep synthetic sub-bass in complex polyrhythms. BPM: 100. Length: 140 seconds\\n- 1920s swing jazz with brass section, upright bass, piano, brushed drums, banjo, clarinet, and soft strings that swing lively for energetic dance vibes. BPM: 110. Length: 180 seconds\\n- Futuristic electronic sci-fi track with pulsing bass synth, evolving lead synths, layered pads, glitch percussion, robotic FX, and sub-bass for tense cinematic energy. BPM: 125. Length: 200 seconds\\n- Ambient underwater soundscape with flowing water textures, soft piano motifs, synth drones, distant bells, and underwater reverb for spacious meditative immersion. BPM: 45. Length: 300 seconds\\n- Horror cinematic track with dissonant strings, eerie piano stabs, cinematic percussion including taiko and low toms, and synth FX producing suspenseful creepy tension. BPM: 90. Length: 240 seconds\\n- Reggae track with offbeat guitar, warm basslines, snare, kick, congas, and horn stabs giving laid-back groovy energy. BPM: 85. Length: 150 seconds\\n- Blues track with soulful electric guitar solos, walking bass, piano, and shuffle drums creating expressive and emotive storytelling. BPM: 90. Length: 180 seconds\\n- Latin salsa with congas, timbales, horns, piano montunos, bass, and layered percussion for vibrant danceable energy. BPM: 120. Length: 210 seconds\\n- Afrobeat track with electric guitar stabs, horns, layered percussion, congas, shakers, bass groove, and synth pads for vibrant rhythmic energy. BPM: 105. Length: 200 seconds\\n- Indie rock track with electric guitar riffs, bass, live drum kit, layered synths, and subtle strings for energetic yet emotional feel. BPM: 110. Length: 180 seconds\\n- Funk groove with slap bass, electric guitar chords, brass stabs, drums, congas, and rhythmic keyboards creating high-energy danceable rhythm. BPM: 105. Length: 180 seconds\\n- Drum and bass track with fast breakbeat drums, deep sub-bass, sharp synth leads, pads, and atmospheric FX for high-energy club motion. BPM: 175. Length: 150 seconds\\n- Dark ambient track with drones, distant bells, low rumbles, soft wind textures, and synth pads producing eerie immersive tension. BPM: 50. Length: 300 seconds\\n- Tropical house track with marimba, steel drums, soft synths, smooth bass, layered percussion, and light piano riffs for sunny chill dance vibes. BPM: 110. Length: 180 seconds\\n- Progressive rock track with electric guitar leads, organ, bass, drum kit, synth layers, and occasional strings for epic layered energy. BPM: 100. Length: 220 seconds\\n- Music box melody with delicate metallic tones and soft resonance, lullaby style, with gentle ambient reverb. BPM: 60. Length: 20 seconds\\n- Soft piano arpeggio with warm felted tone and slow attack, lullaby style, with intimate room ambience. BPM: 60. Length: 30 seconds\\n- Harp gentle plucked pattern with airy resonance, lullaby style, with dreamy reverb tail. BPM: 65. Length: 25 seconds\\n- Acoustic guitar fingerstyle pattern with warm nylon strings and soft dynamics, lullaby style, with subtle room resonance. BPM: 60. Length: 30 seconds\\n- Ambient synth pad with smooth evolving texture and soft harmonics, lullaby style, with wide stereo ambience. BPM: 50. Length: 40 seconds\\n- Early rock piano with walking left-hand bass line, shuffle rhythms, and blues scale improvisations in energetic 1950s boogie-woogie style. BPM: 160. Length: 180 seconds\\n- Trip Hop track with jazzy sampled vibraphone, mid-tempo breakbeat drums, harp, Latin ethnic percussion, and sweeping cinematic strings creating airy, relaxing, soulful lounge vibes. BPM: 90. Length: 180 seconds\\n- Country outlaw cinematic instrumental with blues pedal steel guitar, rustic mandolin, fiddle call-and-response, tape-driven rattly drum kit, autoharp, and soaring accordion solo for raw, emotional southern blues expression. BPM: 85. Length: 200 seconds\\n- Neo Classical track with sweeping string section, elegant horns, and delicate piano creating soothing, hypnotic, modern, soft, and classic mood. BPM: 70. Length: 180 seconds\\n- Art Rock desert track with desolate piano chords, western-themed rhythm guitars, unique lead guitars, rattly vintage drum kit, and supporting bass creating lonely, expansive, beautiful, and strange atmospheres. BPM: 95. Length: 180 seconds\\n- Cinematic Sci-Fi score with dramatic horn section, building marcato strings, gliding bassoon, thunderous cymbals, subdued timpani, and subtle synth drones producing awe-inspiring, uplifting, epic intergalactic energy. BPM: 100. Length: 220 seconds\\n- West Coast Hip Hop instrumental with cascading harp melodies, smooth Rhodes piano chops, vintage boom bap drums, and walking double bass producing raw, street, and soulful block-party vibes. BPM: 92. Length: 180 seconds\\n- Synthwave futuristic track with pulsating synth bass, exciting chords, soaring leads, and reverberating drum machine patterns creating gritty, pounding, and cool energy. BPM: 110. Length: 180 seconds\\n- Breakbeat track with complex percussion, intricate breakbeats, gritty synths, lush pads, and 808 bassline producing fresh, modern, futuristic, and rave-ready energy. BPM: 140. Length: 160 seconds\\n- Lounge Jazz 1960s smooth track with laid-back drums, piano chords, double bass, soft electric piano, subtle flute, and unique percussion creating beautiful, atmospheric, eclectic, retro, and chill vibes. BPM: 85. Length: 180 seconds\\n- Latin Jazz 1950s blissful track with laid-back Latin drums, euphoric piano chords, double bass, orchestral accompaniment, acoustic guitar, and vibraphone producing nostalgic, beautiful, atmospheric, cinematic, and chill mood. BPM: 95. Length: 180 seconds\\n- Acid Jazz 1970s summertime track with smooth electric piano, trippy synth leads, laid-back vintage drum kit, fuzzy electric bass, and uplifting violin producing retro, psychedelic, jazzy, relaxing energy. BPM: 100. Length: 180 seconds\\n- Progressive Soul 1970s track with feel-good piano, psychedelic organ, groovy vintage drum kit with percussion, fuzzy electric bass, and synth strings producing retro, raw, soulful, joyous atmosphere. BPM: 90. Length: 180 seconds\\n- Discotheque 1970s French-inspired track with sultry piano, psychedelic guitars, groovy drum kit, fuzzy electric bass, and melancholic organ producing retro, raw, laid-back, and relaxing mood. BPM: 105. Length: 180 seconds\\n- Soul Jazz 1970s track with expressive saxophone, smooth piano, groovy drum kit, rhythmic upright bass, sweeping strings, and minimal vibraphone producing retro, raw, laid-back, and epic energy. BPM: 95. Length: 180 seconds\\n- Vintage R&B 1970s live studio track with subtle brass, smooth piano, sweeping strings, and minimal drums producing retro, beautiful, uplifting, nostalgic mood. BPM: 85. Length: 180 seconds\\n- 50s Pop track with Latin influence, string section, bold brass, vibraphone, acoustic guitar, flute, ethnic percussion, and brushed drums creating sexy, epic, vintage, retro, melancholic, jazzy, dramatic energy. BPM: 100. Length: 180 seconds\\n- A piece of calm, quiet, mellow, serene music perfect for a peaceful film score, featuring soft modulating piano, ambient sfx and foley, beautiful vibraphone, and subtle synthesizer drones. The mood is cinematic, thoughtful, serene and nostalgic. BPM: 55. Length: 300 seconds\",\n \"Instrument\": \"You are a music metadata expert. Given an instrument, generate a descriptive prompt for a generative audio model.\\n\\n1. Identify the instrument.\\n2. Add playing style or technique.\\n3. Include details about material, timbre, or texture.\\n4. Add musical style or mood. Specify the genre, context, or emotional character.\\n5. Add spatial or production qualities.\\n6. Specify BPM: Always include a BPM appropriate to the style and context.\\n7. Specify length: Provide an integer in seconds (6–20 s for loops, 20–180 s for stems).\\n\\nExamples:\\n- Synth arpeggio loop with bright detuned oscillators. BPM: 120. Length: 8 seconds\\n- Chord stab loop with sharp percussive attack. BPM: 90. Length: 6 seconds\\n- Guitar muted strum loop with tight rhythmic feel. BPM: 100. Length: 8 seconds\\n- Pluck sequence loop with bright resonant tone. BPM: 128. Length: 10 seconds\\n- Marimba and vibraphone percussive loop with resonant wooden and metallic tones. BPM: 110. Length: 12 seconds\\n- Drum loop with deep muffled kick on beat one, snappy rimshot snare on beats two and four with rolling ghost note fills, and tight closed hi-hats with subtle open accents. BPM: 85. Length: 10 seconds\\n- Drum groove loop with brushed snare swinging on the ride, soft feathered kick on downbeats, and light closed hi-hat taps on the upbeats. BPM: 130. Length: 12 seconds\\n- Kick and hi-hat loop with four-on-the-floor punchy kick, tight closed hi-hats on every eighth note, and a sharp dry snare on beats two and four. BPM: 130. Length: 15 seconds\\n- Vinyl crackle drum loop with warm low-pass filtered kick, dusty snare with tape saturation, and shuffled closed hi-hats with subtle vinyl crackle ambiance. BPM: 80. Length: 10 seconds\\n- Ambient pad loop with evolving texture. BPM: 80. Length: 12 seconds\\n- Melodic synth bass groove loop with pumping sidechain feel. BPM: 122. Length: 10 seconds\\n- Melodic Bass slap and pop rhythm loop. BPM: 100. Length: 8 seconds\\n- Acoustic bass walking line loop with natural wooden resonance. BPM: 120. Length: 12 seconds\\n- String pizzicato motif loop, suspenseful, with tight string texture. BPM: 90. Length: 8 seconds\\n- Brass staccato riff loop with sharp bright attack. BPM: 130. Length: 10 seconds\\n- Flute airy melodic loop with wooden headjoint resonance. BPM: 100. Length: 6 seconds\\n- Pan flute ambient loop with breathy timbre. BPM: 75. Length: 8 seconds\\n- Clarinet riff loop with warm smooth reed tone. BPM: 120. Length: 10 seconds\\n- Oboe motif loop, orchestral, with rich double reed resonance. BPM: 80. Length: 8 seconds\\n- Recorder Renaissance motif loop with soft wooden timbre. BPM: 100. Length: 6 seconds\\n- Electric sitar riff loop with buzzing resonant tone. BPM: 90. Length: 10 seconds\\n- Koto plucked motif loop with resonant wooden strings. BPM: 90. Length: 8 seconds\\n- Shamisen folk melody loop with percussive twang. BPM: 100. Length: 8 seconds\\n- Banjo fingerpicking loop with metallic string resonance. BPM: 110. Length: 10 seconds\\n- Mandolin tremolo loop with crisp wooden body tone. BPM: 120. Length: 10 seconds\\n- Acoustic guitar chord vamp loop with natural room resonance. BPM: 110. Length: 12 seconds\\n- Nylon string guitar arpeggio loop with warm, soft timbre. BPM: 90. Length: 15 seconds\\n- Electric guitar riff loop with driven distorted tone. BPM: 130. Length: 10 seconds\\n- Slide guitar melody loop with warm resonant glide. BPM: 100. Length: 12 seconds\\n- Steel guitar slide loop with bright pedal steel tone. BPM: 95. Length: 12 seconds\\n- Harpsichord arpeggio loop with crisp plucked attack. BPM: 120. Length: 10 seconds\\n- Rhodes chord vamp loop with warm electric piano tone. BPM: 100. Length: 12 seconds\\n- Clavinet funky rhythm loop. BPM: 105. Length: 10 seconds\\n- Organ chord vamp loop with full drawbar warmth. BPM: 90. Length: 12 seconds\\n- Drum loop with booming 808 kick on beat one, crisp snare on beat three, and rapid triplet hi-hat rolls with open hat accents for aggressive high-energy feel. BPM: 140. Length: 8 seconds\\n- Breakbeat drum loop with chopped Amen-style snare flurries, driving kick on the one, fast sixteenth-note closed hi-hats, and syncopated open hat accents. BPM: 170. Length: 10 seconds\\n- Glitch percussion loop with stuttered kick transients, randomised snare hits processed with bit-crushing, and erratic hi-hat patterns with pitch-shifted metallic ticks. BPM: 120. Length: 12 seconds\\n- Metallic hits loop with distorted kick impacts, processed metal-plate snare slams, and grinding hi-hat noise bursts for aggressive mechanical texture. BPM: 120. Length: 10 seconds\\n- Timpani hits loop, cinematic, with deep resonant kick-like timpani strikes on beat one, rolling snare-style timpani fills, and no hi-hats for a grand orchestral feel. BPM: 70. Length: 8 seconds\\n- Snare roll loop, dramatic, with accelerating snare drum rolls building from soft to crashing, deep supporting kick pulses, and no hi-hats for maximum impact. BPM: 100. Length: 8 seconds\\n- Accordion motif loop with bright reedy bellows tone. BPM: 100. Length: 10 seconds\\n- Harmonica blues riff loop with expressive reed timbre. BPM: 90. Length: 10 seconds\\n- Trombone riff loop with warm sliding brass tone. BPM: 120. Length: 10 seconds\\n- French horn melodic loop, cinematic. BPM: 80. Length: 12 seconds\\n- Soprano sax ballad loop. BPM: 70. Length: 12 seconds\\n- Alto sax bebop riff loop. BPM: 200. Length: 10 seconds\\n- Electric violin melodic loop with reverb. BPM: 90. Length: 10 seconds\\n- String pad loop with cinematic texture. BPM: 70. Length: 15 seconds\\n- Granular synth evolving texture loop. BPM: 90. Length: 15 seconds\\n- Piano motif loop with soft felt hammer tone. BPM: 80. Length: 10 seconds\\n- Pad and synth loop with lush detuned shimmer. BPM: 85. Length: 12 seconds\\n- Synth lead loop with sidechain pumping compression. BPM: 128. Length: 10 seconds\\n- Analog synth bassline loop with deep warm low-end. BPM: 122. Length: 12 seconds\\n- FM synth lead motif loop with bright metallic shimmer. BPM: 110. Length: 10 seconds\\n- Bass groove loop with tight rhythmic two-bar pattern. BPM: 100. Length: 16 seconds\\n- Acoustic guitar fingerstyle motif loop with warm wood resonance. BPM: 90. Length: 45 seconds\\n- Sombre acoustic guitar motif loop with cavernous reverb, delicate fingerpicking, and expressive melancholic tone. BPM: 70. Length: 45 seconds\\n- Electric guitar rock riff motif loop. BPM: 130. Length: 40 seconds\\n- Vintage electric guitar motif loop, live-recorded in a vintage studio, with expressive and dynamic solo performance. BPM: 90. Length: 40 seconds\\n- Piano chord progression motif loop with rich harmonic movement. BPM: 120. Length: 60 seconds\\n- String ensemble cinematic motif loop with rich wooden resonance. BPM: 80. Length: 120 seconds\\n- Brass ensemble cinematic motif loop with bright metallic timbre. BPM: 90. Length: 90 seconds\\n- Ethnic percussion ensemble motif loop with deep resonant djembe kick tones, slapped snare-like rim hits on congas, and layered shakers and bells providing hi-hat-like rhythmic texture with polyrhythmic patterns. BPM: 100. Length: 90 seconds\\n- Synth ambient motif loop with evolving textures. BPM: 80. Length: 180 seconds\\n- Motif loop with warm dusty vinyl crackle and tape saturation. BPM: 80. Length: 60 seconds\\n- Synth lead and bass motif loop with bright punchy energy. BPM: 128. Length: 90 seconds\\n- Funk band motif loop: bass, drums, guitar. BPM: 100. Length: 90 seconds\\n- Ethnic flute motif for cinematic use. BPM: 80. Length: 30 seconds\\n- Steel drum melodic motif loop with bright metallic resonance. BPM: 110. Length: 20 seconds\\n- Marimba percussive motif loop with resonant wooden tone. BPM: 100. Length: 20 seconds\\n- Vibraphone melodic motif loop with metallic shimmer. BPM: 90. Length: 25 seconds\\n- Piano cinematic motif loop with resonant wooden tone. BPM: 80. Length: 30 seconds\\n- Violin expressive cinematic motif loop with rich wooden resonance. BPM: 75. Length: 25 seconds\\n- Cello expressive motif loop with deep wooden resonance. BPM: 70. Length: 30 seconds\\n- Trumpet expressive motif loop with brassy overtones. BPM: 100. Length: 25 seconds\\n- Sax expressive motif loop with warm reed timbre. BPM: 95. Length: 25 seconds\\n- Ethnic drum ensemble motif loop with booming natural-skin bass drum kicks, sharp hand-slap snare accents on djembes and talking drums, and layered wooden and metal percussion providing rhythmic hi-hat-like patterns. BPM: 95. Length: 30 seconds\\n- Ambient drone motif loop. BPM: 60. Length: 180 seconds\\n- Orchestral tension motif loop. BPM: 90. Length: 150 seconds\\n- Electronic track motif loop with drums, bass, synth. BPM: 128. Length: 180 seconds\",\n \"SFX\": \"You are a professional sound design expert. Convert the user's input into a precise, vivid sound effects description suitable for generative audio models.\\n\\nDescribe clearly:\\n- Sound source\\n- Physical character (texture, timbre, material: metal, wood, glass, concrete, etc.)\\n- Spatial qualities (indoor/outdoor, cave/open field/underwater, dry/reverberant, close-up/distant, echoing/muffled)\\n- Temporal evolution (attack, decay, movement, transitions over time)\\n- Include motion or spatial movement if applicable (passing, approaching, stereo movement)\\n\\nAudio length rules:\\n- Very short sounds (impacts, clicks, gunshots): 1–3 seconds\\n- Medium actions (footsteps, object movement, transitions): 3–6 seconds\\n- Ambience / environments: 6–15 seconds\\n- Always append: Length: X seconds (integer only, no decimals).\\n\\nOutput constraints:\\n- Length: 1–2 dense sentences maximum\\n- Output ONLY the final rewritten prompt\\n- No explanations, no formatting, no quotes\\n- Use concise but dense technical language\\n- Focus strictly on sound effects or ambience\\n- Always append: Length: X seconds (integer only, no decimals).\\n\\nQuality guidelines:\\n- Be specific and avoid vague terms\\n- Prioritize clarity and realism\\n- Combine elements into one coherent scene\\n- Avoid redundancy\\n\\nExamples:\\n- Heavy rain hitting a metal roof during a thunderstorm, distant thunder rumbles, stereo, realistic ambience. Length: 45 seconds\\n- Quiet forest at dawn with birds chirping, soft wind through leaves, distant stream flowing. Length: 60 seconds\\n- Busy city street at night, cars passing, muffled conversations, occasional horn, urban ambience. Length: 50 seconds\\n- Ocean waves crashing against rocky cliffs, strong wind, dramatic and cinematic. Length: 70 seconds\\n- Wooden door creaking open slowly in an old house, echoing interior, eerie tone. Length: 3 seconds\\n- Glass bottle shattering on concrete, sharp impact, scattered fragments. Length: 2 seconds\\n- Footsteps on gravel, steady walking pace, close perspective. Length: 8 seconds\\n- Typing rapidly on a mechanical keyboard, crisp tactile clicks. Length: 5 seconds\\n- Punch impact with deep bass hit, cinematic trailer style. Length: 2 seconds\\n- Car speeding past at high velocity, doppler effect, realistic whoosh. Length: 3 seconds\\n- Object falling from height and hitting ground with a heavy thud. Length: 2 seconds\\n- Sword swing whooshing through air, fast motion, clean metallic tone. Length: 2 seconds\\n- Futuristic laser blast, clean energy pulse, high-tech sound design. Length: 1 seconds\\n- Spaceship engine humming, low frequency rumble, interior perspective. Length: 90 seconds\\n- Magical spell casting, shimmering particles, rising tonal energy. Length: 8 seconds\\n- Teleportation effect, glitchy digital distortion with a soft whoosh. Length: 5 seconds\\n- Dark eerie drone with distant whispers, creepy, slow build tension. Length: 120 seconds\\n- Sudden horror jump scare sting, sharp violin hit, cinematic. Length: 1 second\\n- Metal scraping slowly in a dark tunnel, echoing and ominous. Length: 20 seconds\\n- Explosion with debris scattering, deep bass, cinematic realism. Length: 4 seconds\\n- Building collapsing, rumbling concrete, dust and debris falling. Length: 25 seconds\\n- Fire crackling intensely, wood burning, close-up detail. Length: 80 seconds\\n- Gunshot in a large empty warehouse, loud echo decay. Length: 2 seconds\\n- Retro arcade coin insert sound, 8-bit style. Length: 1 second\\n- Level up chime, bright, rewarding, fantasy RPG style. Length: 2 seconds\\n- Error buzzer, short, digital, UI feedback. Length: 1 second\\n- Menu navigation clicks, soft futuristic interface sounds. Length: 3 seconds\\n- Layered soundscape: rain, thunder, footsteps, and distant sirens all blending naturally. Length: 90 seconds\\n- Rapid sequence of three impacts: metal hit, glass break, wood crack, spaced evenly. Length: 4 seconds\\n- Sound moving from left to right stereo field: passing motorcycle. Length: 5 seconds\\n- Close vs far perspective transition: footsteps approaching then fading away. Length: 6 seconds\\n- Tape stop sub drop, a massive sub-bass note that mimics a vinyl record or tape machine being turned off, the pitch and speed drop simultaneously, causing the high-end harmonics to smear and thicken as the sound grinds to a halt at a sub-sonic frequency. Length: 11 seconds\\n- Gravel and leaves footsteps, the sound of a hard boot stepping onto dry leaves or gravel, crisp and natural with detailed texture. Length: 11 seconds\\n- Ghostship moan, a massive, deep wooden groan with a low-frequency moan, like heavy timber under immense structural tension, swaying slowly, processed with long, dark wooden room reverb for a sense of scale. Length: 11 seconds\\n- Bicycle chain, a continuous metallic whirring sound of a chain moving over sprockets, with individual teeth catching the links, processed with resonant band-pass filter to emphasize metallic singing. Length: 11 seconds\\n- Warp drive, a sound that starts with a massive suck-back of ambient noise, followed by a supersonic crack and high-pitched zing that disappears into the distance, giving the sense of stretching space-time. Length: 11 seconds\\n- Ice cubes, high-pitched musical clinking of hard ice hitting a thin glass, bright resonant ring with subtle liquid sloshing around the edges. Length: 11 seconds\\n- Paper shuffle, the sound of a thick stack of heavy bond paper being squared up on a desk, dry papery thud with a quick fanning sound as air moves between the pages. Length: 11 seconds\\n- Drawer slam, a blunt, powerful thud made by slamming a wooden desk drawer shut, pronounced low-mid body, slightly distorted for aggressive character. Length: 3 seconds\",\n \"One-shot\": \"You are a music metadata expert. Given an instrument or sound, generate a descriptive prompt for a short, isolated one-shot audio sample for music production.\\n\\n1. Identify the instrument or sound source.\\n2. Describe the playing technique or hit type (e.g., pluck, slam, tap, stab).\\n3. Include details about material, timbre, or texture.\\n4. Add spatial or production qualities (dry/wet, room, close-mic).\\n5. Specify length: short integer in seconds (1–11 s).\\n\\nExamples:\\n- Piano key hit with bright percussive attack and resonant wooden body. Length: 2 seconds\\n- Kick drum punchy low-end hit with warm skin resonance. Length: 2 seconds\\n- Snare drum rimshot accent with crisp snare wires. Length: 2 seconds\\n- Acoustic guitar fingerstyle note with warm spruce tone. Length: 3 seconds\\n- Bass pluck with jazzy tone and resonant wooden body. Length: 3 seconds\\n- Electric guitar power chord with distortion. Length: 3 seconds\\n- Metallic glitch percussion hit with sharp metallic texture. Length: 2 seconds\\n- Tabla resonant tone hit with natural skin timbre. Length: 2 seconds\\n- Djembe slap accent with dry wooden resonance. Length: 2 seconds\\n- Synth stab with reverb tail. Length: 3 seconds\\n- Violin expressive note with vibrato and rich wooden resonance. Length: 3 seconds\\n- Cello legato note, cinematic, with warm resonant body. Length: 3 seconds\\n- Trumpet bright accent with slightly brassy overtones. Length: 2 seconds\\n- Melodic saxophone jazz riff with smooth reed timbre and a slight vibrato bend. Length: 3 seconds\\n- Harp pluck with airy tone and resonant strings. Length: 2 seconds\\n- Glockenspiel bell-like note with bright metallic clarity. Length: 2 seconds\\n- Metallic clang sound design hit. Length: 2 seconds\\n- Granular texture hit. Length: 3 seconds\\n- Reversed piano hit. Length: 2 seconds\\n- Synth riser effect. Length: 6 seconds\\n- Percussion impact hit. Length: 2 seconds\\n- Cinematic hit. Length: 2 seconds\\n- Dry clap, a crisp, natural single hand clap recorded in a dead room with an extremely sharp transient and no room reflections. Length: 1 second\\n- Studio hat, a classic, natural recording of 14-inch hi-hats played tightly closed, zero ring, very fast decay. Length: 1 second\\n- Disco open hat, bright 14-inch open hi-hat with long, shimmering decay, perfect for disco or dance grooves. Length: 1 second\\n- Pillow kick, acoustic kick drum muffled with a heavy blanket, producing a short, dry \\\"thump\\\" with almost zero resonance. Length: 1 second\\n- Short 808, punchy 808 kick with sharp, distorted transient and fast-decaying sub-tail. Length: 1 second\\n- Egg shaker, classic plastic egg shaker recorded with a small-diaphragm condenser mic, producing a light, consistent \\\"tick\\\" with very short sustain. Length: 1 second\\n- African drums, dynamic African drums and percussion ensemble with natural acoustic textures. Length: 3 seconds\\n- Latin drums, dynamic Latin drums and percussion ensemble featuring authentic rhythmic patterns. Length: 3 seconds\\n- String quartet, euphoric string quartet with dynamic and emotional playing, full of expressive harmonies and movement. Length: 3 seconds\\n- Piano, nostalgic, atmospheric piano piece with dynamic and emotional performance, intimate and resonant. Length: 3 seconds\\n- Analogue drift pad, warm polyphonic pad with three detuned oscillators (saw + triangle), subtle pitch drift, and lush bucket-brigade chorus for wide, nostalgic stereo image. Length: 11 seconds\\n- Phase distortion bass, Casio CZ-style phase-distorted sine wave warped into a jagged sawtooth for retro synth bass tone. Length: 11 seconds\\n- Vibrato saxophone, bright lyrical alto sax with fast fluttery vibrato, reedy vintage tone, captured with ribbon mic for warm nostalgic sound. Length: 11 seconds\\n- Lofi upright bass, upright bass recorded with ribbon mic in a wooden room, natural air with slightly boxy resonance, tape-saturated for dusty 1950s jazz feel. Length: 2 seconds\"\n}", + "Music" + ] + }, + { + "id": 40, + "type": "StringReplace", + "pos": [ + 1350, + 900 + ], + "size": [ + 260, + 280 + ], + "flags": {}, + "order": 15, + "mode": 0, + "inputs": [ + { + "localized_name": "string", + "name": "string", + "type": "STRING", + "widget": { + "name": "string" + }, + "link": 59 + }, + { + "localized_name": "find", + "name": "find", + "type": "STRING", + "widget": { + "name": "find" + }, + "link": null + }, + { + "localized_name": "replace", + "name": "replace", + "type": "STRING", + "widget": { + "name": "replace" + }, + "link": 58 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 60 + ] + } + ], + "title": "Text Replace (AUDIO LENGTH)", + "properties": { + "Node name for S&R": "StringReplace" + }, + "widgets_values": [ + "", + "AUDIO_LENGTH", + "" + ] + }, + { + "id": 38, + "type": "StringReplace", + "pos": [ + 720, + 900 + ], + "size": [ + 290, + 280 + ], + "flags": {}, + "order": 13, + "mode": 0, + "inputs": [ + { + "localized_name": "string", + "name": "string", + "type": "STRING", + "widget": { + "name": "string" + }, + "link": null + }, + { + "localized_name": "find", + "name": "find", + "type": "STRING", + "widget": { + "name": "find" + }, + "link": null + }, + { + "localized_name": "replace", + "name": "replace", + "type": "STRING", + "widget": { + "name": "replace" + }, + "link": 66 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 52 + ] + } + ], + "title": "Text Replace (PROMPT TEMPLATE)", + "properties": { + "Node name for S&R": "StringReplace" + }, + "widgets_values": [ + "SYSTEM_PROMPTS\n\nInput: USER_INPUT\nTarget audio length: AUDIO_LENGTH seconds.\nOutput:", + "SYSTEM_PROMPTS", + "" + ] + }, + { + "id": 35, + "type": "PrimitiveBoolean", + "pos": [ + -390, + 570 + ], + "size": [ + 400, + 100 + ], + "flags": {}, + "order": 11, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "BOOLEAN", + "widget": { + "name": "value" + }, + "link": 83 + } + ], + "outputs": [ + { + "localized_name": "BOOLEAN", + "name": "BOOLEAN", + "type": "BOOLEAN", + "links": [ + 48 + ] + } + ], + "title": "Boolean (Enable_Reprompt)", + "properties": { + "Node name for S&R": "PrimitiveBoolean" + }, + "widgets_values": [ + true + ] + }, + { + "id": 36, + "type": "PrimitiveFloat", + "pos": [ + -390, + 410 + ], + "size": [ + 400, + 110 + ], + "flags": {}, + "order": 12, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "FLOAT", + "widget": { + "name": "value" + }, + "link": 82 + } + ], + "outputs": [ + { + "localized_name": "FLOAT", + "name": "FLOAT", + "type": "FLOAT", + "links": [ + 50, + 56 + ] + } + ], + "title": "Float (Duration)", + "properties": { + "Node name for S&R": "PrimitiveFloat" + }, + "widgets_values": [ + 150 + ] + }, + { + "id": 25, + "type": "CheckpointLoaderSimple", + "pos": [ + 100, + 130 + ], + "size": [ + 440, + 190 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "localized_name": "ckpt_name", + "name": "ckpt_name", + "type": "COMBO", + "widget": { + "name": "ckpt_name" + }, + "link": 79 + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "links": [ + 30 + ] + }, + { + "localized_name": "CLIP", + "name": "CLIP", + "type": "CLIP", + "links": [] + }, + { + "localized_name": "VAE", + "name": "VAE", + "type": "VAE", + "links": [ + 39 + ] + } + ], + "properties": { + "Node name for S&R": "CheckpointLoaderSimple", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "models": [ + { + "name": "stable_audio_3_medium.safetensors", + "url": "https://huggingface.co/Comfy-Org/stable-audio-3/resolve/main/checkpoints/stable_audio_3_medium.safetensors", + "directory": "checkpoints" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "stable_audio_3_medium.safetensors" + ] + }, + { + "id": 26, + "type": "CLIPLoader", + "pos": [ + 100, + 390 + ], + "size": [ + 440, + 170 + ], + "flags": {}, + "order": 6, + "mode": 0, + "inputs": [ + { + "localized_name": "clip_name", + "name": "clip_name", + "type": "COMBO", + "widget": { + "name": "clip_name" + }, + "link": 80 + }, + { + "localized_name": "type", + "name": "type", + "type": "COMBO", + "widget": { + "name": "type" + }, + "link": null + }, + { + "localized_name": "device", + "name": "device", + "shape": 7, + "type": "COMBO", + "widget": { + "name": "device" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "CLIP", + "name": "CLIP", + "type": "CLIP", + "links": [ + 34, + 35 + ] + } + ], + "properties": { + "Node name for S&R": "CLIPLoader", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "models": [ + { + "name": "t5gemma_b_b_ul2.safetensors", + "url": "https://huggingface.co/Comfy-Org/stable-audio-3/resolve/main/text_encoders/t5gemma_b_b_ul2.safetensors", + "directory": "text_encoders" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "t5gemma_b_b_ul2.safetensors", + "stable_audio", + "default" + ] + }, + { + "id": 54, + "type": "PreviewAny", + "pos": [ + 1720, + 1580 + ], + "size": [ + 420, + 550 + ], + "flags": {}, + "order": 20, + "mode": 4, + "inputs": [ + { + "localized_name": "source", + "name": "source", + "type": "*", + "link": 84 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": null + } + ], + "properties": { + "Node name for S&R": "PreviewAny" + }, + "widgets_values": [ + null, + null, + null + ] + } + ], + "groups": [ + { + "id": 1, + "title": "Loaders: checkpoint & CLIP", + "bounding": [ + 80, + 50, + 485.721654232725, + 527.2848777754299 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 2, + "title": "CLIP encode: conditioning", + "bounding": [ + 600, + 60, + 470, + 510 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 3, + "title": "User inputs: prompt & duration", + "bounding": [ + -400, + 10, + 430, + 740 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 7, + "title": "Reprompt: full branch (template + LLM)", + "bounding": [ + 60, + 780, + 1630, + 1360 + ], + "color": "#444", + "flags": {} + }, + { + "id": 4, + "title": "Reprompt: JSON extract & template fills", + "bounding": [ + 120, + 820, + 1520, + 650 + ], + "color": "#444", + "flags": {} + }, + { + "id": 5, + "title": "Helpers: duration to string", + "bounding": [ + 1340, + 1180, + 280, + 250 + ], + "color": "#444", + "flags": {} + }, + { + "id": 6, + "title": "Reprompt: Qwen TextGenerate", + "bounding": [ + 680, + 1510, + 960, + 614.65625 + ], + "color": "#444", + "flags": {} + }, + { + "id": 8, + "title": "Audio generation: Stable Audio", + "bounding": [ + 60, + 10, + 1627.3616782294932, + 737.0545987464304 + ], + "color": "#3f789e", + "flags": {} + } + ], + "links": [ + { + "id": 35, + "origin_id": 26, + "origin_slot": 0, + "target_id": 7, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 13, + "origin_id": 3, + "origin_slot": 0, + "target_id": 12, + "target_slot": 0, + "type": "LATENT" + }, + { + "id": 39, + "origin_id": 25, + "origin_slot": 2, + "target_id": 12, + "target_slot": 1, + "type": "VAE" + }, + { + "id": 50, + "origin_id": 36, + "origin_slot": 0, + "target_id": 11, + "target_slot": 0, + "type": "FLOAT" + }, + { + "id": 30, + "origin_id": 25, + "origin_slot": 0, + "target_id": 3, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 4, + "origin_id": 6, + "origin_slot": 0, + "target_id": 3, + "target_slot": 1, + "type": "CONDITIONING" + }, + { + "id": 6, + "origin_id": 7, + "origin_slot": 0, + "target_id": 3, + "target_slot": 2, + "type": "CONDITIONING" + }, + { + "id": 12, + "origin_id": 11, + "origin_slot": 0, + "target_id": 3, + "target_slot": 3, + "type": "LATENT" + }, + { + "id": 34, + "origin_id": 26, + "origin_slot": 0, + "target_id": 6, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 49, + "origin_id": 34, + "origin_slot": 0, + "target_id": 6, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 47, + "origin_id": 31, + "origin_slot": 0, + "target_id": 34, + "target_slot": 0, + "type": "STRING" + }, + { + "id": 46, + "origin_id": 28, + "origin_slot": 0, + "target_id": 34, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 48, + "origin_id": 35, + "origin_slot": 0, + "target_id": 34, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 56, + "origin_id": 36, + "origin_slot": 0, + "target_id": 41, + "target_slot": 0, + "type": "FLOAT" + }, + { + "id": 57, + "origin_id": 41, + "origin_slot": 1, + "target_id": 42, + "target_slot": 0, + "type": "INT" + }, + { + "id": 52, + "origin_id": 38, + "origin_slot": 0, + "target_id": 39, + "target_slot": 0, + "type": "STRING" + }, + { + "id": 53, + "origin_id": 31, + "origin_slot": 0, + "target_id": 39, + "target_slot": 2, + "type": "STRING" + }, + { + "id": 40, + "origin_id": 29, + "origin_slot": 0, + "target_id": 28, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 60, + "origin_id": 40, + "origin_slot": 0, + "target_id": 28, + "target_slot": 4, + "type": "STRING" + }, + { + "id": 65, + "origin_id": 43, + "origin_slot": 0, + "target_id": 49, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 59, + "origin_id": 39, + "origin_slot": 0, + "target_id": 40, + "target_slot": 0, + "type": "STRING" + }, + { + "id": 58, + "origin_id": 42, + "origin_slot": 0, + "target_id": 40, + "target_slot": 2, + "type": "STRING" + }, + { + "id": 66, + "origin_id": 49, + "origin_slot": 0, + "target_id": 38, + "target_slot": 2, + "type": "STRING" + }, + { + "id": 27, + "origin_id": 12, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "AUDIO" + }, + { + "id": 68, + "origin_id": -10, + "origin_slot": 0, + "target_id": 31, + "target_slot": 0, + "type": "STRING" + }, + { + "id": 76, + "origin_id": -10, + "origin_slot": 2, + "target_id": 3, + "target_slot": 4, + "type": "INT" + }, + { + "id": 78, + "origin_id": -10, + "origin_slot": 4, + "target_id": 43, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 79, + "origin_id": -10, + "origin_slot": 5, + "target_id": 25, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 80, + "origin_id": -10, + "origin_slot": 6, + "target_id": 26, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 81, + "origin_id": -10, + "origin_slot": 7, + "target_id": 29, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 82, + "origin_id": -10, + "origin_slot": 1, + "target_id": 36, + "target_slot": 0, + "type": "FLOAT" + }, + { + "id": 83, + "origin_id": -10, + "origin_slot": 3, + "target_id": 35, + "target_slot": 0, + "type": "BOOLEAN" + }, + { + "id": 84, + "origin_id": 28, + "origin_slot": 0, + "target_id": 54, + "target_slot": 0, + "type": "STRING" + } + ], + "extra": {}, + "category": "Audio/Music generation", + "description": "Generates music, instrument loops, sound effects, and one-shots from text using Stable Audio 3 Medium, with optional Qwen 3.5 category-based prompt expansion (Music, Instrument, SFX, One-shot)." + } + ] + }, + "extra": {} +} \ No newline at end of file diff --git a/blueprints/Canny to Image (Z-Image-Turbo).json b/blueprints/Canny to Image (Z-Image-Turbo).json index 14deb64cc..903d372b1 100644 --- a/blueprints/Canny to Image (Z-Image-Turbo).json +++ b/blueprints/Canny to Image (Z-Image-Turbo).json @@ -1553,7 +1553,7 @@ "VHS_MetadataImage": true, "VHS_KeepIntermediate": true }, - "category": "Image generation and editing/Canny to image", + "category": "Image generation and editing/Conditioned", "description": "Generates an image from a Canny edge map using Z-Image-Turbo, with text conditioning." } ] diff --git a/blueprints/Canny to Video (LTX 2.0).json b/blueprints/Canny to Video (LTX 2.0).json index a9682c8a4..ed602b521 100644 --- a/blueprints/Canny to Video (LTX 2.0).json +++ b/blueprints/Canny to Video (LTX 2.0).json @@ -3600,7 +3600,7 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Video generation and editing/Canny to video", + "category": "Video generation and editing/Conditioned", "description": "Generates video from Canny edge maps using LTX-2, with optional synchronized audio." } ] diff --git a/blueprints/ControlNet (Z-Image-Turbo).json b/blueprints/ControlNet (Z-Image-Turbo).json index fbec95a97..160ee11e2 100644 --- a/blueprints/ControlNet (Z-Image-Turbo).json +++ b/blueprints/ControlNet (Z-Image-Turbo).json @@ -1401,7 +1401,7 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image generation and editing/ControlNet", + "category": "Image generation and editing/Conditioned", "description": "Generates images from a text prompt and ControlNet conditioning (e.g. depth, canny) using Z-Image-Turbo." } ] diff --git a/blueprints/Depth to Image (Z-Image-Turbo).json b/blueprints/Depth to Image (Z-Image-Turbo).json index fe9ef0f72..2790827a3 100644 --- a/blueprints/Depth to Image (Z-Image-Turbo).json +++ b/blueprints/Depth to Image (Z-Image-Turbo).json @@ -1579,7 +1579,7 @@ "VHS_MetadataImage": true, "VHS_KeepIntermediate": true }, - "category": "Image generation and editing/Depth to image", + "category": "Image generation and editing/Conditioned", "description": "Generates an image from a depth map using Z-Image-Turbo with text conditioning." }, { diff --git a/blueprints/Depth to Video (ltx 2.0).json b/blueprints/Depth to Video (ltx 2.0).json index bd51e4476..56912de51 100644 --- a/blueprints/Depth to Video (ltx 2.0).json +++ b/blueprints/Depth to Video (ltx 2.0).json @@ -4233,7 +4233,7 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Video generation and editing/Depth to video", + "category": "Video generation and editing/Conditioned", "description": "Generates depth-controlled video with LTX-2: motion and structure follow a depth-reference video alongside text prompting, optional first-frame image conditioning, with optional synchronized audio." }, { diff --git a/blueprints/First-Last-Frame to Video (LTX-2.3).json b/blueprints/First-Last-Frame to Video (LTX-2.3).json index f509aefe0..4cae2dc24 100644 --- a/blueprints/First-Last-Frame to Video (LTX-2.3).json +++ b/blueprints/First-Last-Frame to Video (LTX-2.3).json @@ -3350,7 +3350,7 @@ } ], "extra": {}, - "category": "Video generation and editing/First-Last-Frame to Video", + "category": "Video generation and editing/Conditioned", "description": "Generates a video interpolating between first and last keyframes using LTX-2.3." } ] diff --git a/blueprints/First-Last-Frame to Video.json b/blueprints/First-Last-Frame to Video.json index 84dfafbcd..d76e1e045 100644 --- a/blueprints/First-Last-Frame to Video.json +++ b/blueprints/First-Last-Frame to Video.json @@ -3350,7 +3350,7 @@ } ], "extra": {}, - "category": "Video generation and editing/First-Last-Frame to Video", + "category": "Video generation and editing/FLF2V", "description": "Generates a video that interpolates between the first and last keyframes using LTX-2.3, including optional audio." } ] diff --git a/blueprints/Geometry Estimation (MoGe).json b/blueprints/Geometry Estimation (MoGe).json new file mode 100644 index 000000000..e6f08bf71 --- /dev/null +++ b/blueprints/Geometry Estimation (MoGe).json @@ -0,0 +1,1266 @@ +{ + "revision": 0, + "last_node_id": 67, + "last_link_id": 0, + "nodes": [ + { + "id": 67, + "type": "936dfaf2-575a-48b5-9e0c-df391319d11f", + "pos": [ + -3950, + 5000 + ], + "size": [ + 430, + 480 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "source_image", + "name": "source_image", + "type": "IMAGE", + "link": null + }, + { + "localized_name": "inference_resolution", + "name": "inference_resolution", + "type": "INT", + "widget": { + "name": "inference_resolution" + }, + "link": null + }, + { + "localized_name": "inference_batch_size", + "name": "inference_batch_size", + "type": "INT", + "widget": { + "name": "inference_batch_size" + }, + "link": null + }, + { + "localized_name": "mesh_frame_index", + "name": "mesh_frame_index", + "type": "INT", + "widget": { + "name": "mesh_frame_index" + }, + "link": null + }, + { + "localized_name": "mesh_decimation", + "name": "mesh_decimation", + "type": "INT", + "widget": { + "name": "mesh_decimation" + }, + "link": null + }, + { + "localized_name": "mesh_gap_threshold", + "name": "mesh_gap_threshold", + "type": "FLOAT", + "widget": { + "name": "mesh_gap_threshold" + }, + "link": null + }, + { + "localized_name": "mesh_texture", + "name": "mesh_texture", + "type": "BOOLEAN", + "widget": { + "name": "mesh_texture" + }, + "link": null + }, + { + "localized_name": "moge_model", + "name": "moge_model", + "type": "COMBO", + "widget": { + "name": "moge_model" + }, + "link": null + }, + { + "label": "auto_resize_input", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "mesh", + "name": "mesh", + "type": "MESH", + "links": [] + }, + { + "localized_name": "normal_opengl", + "name": "normal_opengl", + "type": "IMAGE", + "links": [] + }, + { + "localized_name": "normal_directx", + "name": "normal_directx", + "type": "IMAGE", + "links": [] + } + ], + "properties": { + "proxyWidgets": [ + [ + "55", + "resolution_level" + ], + [ + "55", + "batch_size" + ], + [ + "54", + "batch_index" + ], + [ + "54", + "decimation" + ], + [ + "54", + "discontinuity_threshold" + ], + [ + "54", + "texture" + ], + [ + "58", + "model_name" + ], + [ + "66", + "switch" + ] + ], + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [], + "title": "Geometry Estimation (MoGe)" + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "936dfaf2-575a-48b5-9e0c-df391319d11f", + "version": 1, + "state": { + "lastGroupId": 1, + "lastNodeId": 69, + "lastLinkId": 91, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Geometry Estimation (MoGe)", + "inputNode": { + "id": -10, + "bounding": [ + -5130, + 5320, + 167.337890625, + 228 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + -3090, + 4966, + 131.51953125, + 108 + ] + }, + "inputs": [ + { + "id": "cc8ce79d-ba20-4a25-a51c-c2afcd35e520", + "name": "source_image", + "type": "IMAGE", + "linkIds": [ + 48, + 55, + 56, + 82 + ], + "localized_name": "source_image", + "pos": [ + -4986.662109375, + 5344 + ] + }, + { + "id": "06eefa21-8e60-49f3-9a34-35b081f4ae52", + "name": "inference_resolution", + "type": "INT", + "linkIds": [ + 73 + ], + "localized_name": "inference_resolution", + "pos": [ + -4986.662109375, + 5364 + ] + }, + { + "id": "616638fe-f603-4d10-bae9-fc87c134380f", + "name": "inference_batch_size", + "type": "INT", + "linkIds": [ + 74 + ], + "localized_name": "inference_batch_size", + "pos": [ + -4986.662109375, + 5384 + ] + }, + { + "id": "fcacfca9-7927-4c38-94da-8ab22256325f", + "name": "mesh_frame_index", + "type": "INT", + "linkIds": [ + 75 + ], + "localized_name": "mesh_frame_index", + "pos": [ + -4986.662109375, + 5404 + ] + }, + { + "id": "acbfe7f9-1b69-42c1-8614-4ccf54b28d4e", + "name": "mesh_decimation", + "type": "INT", + "linkIds": [ + 76 + ], + "localized_name": "mesh_decimation", + "pos": [ + -4986.662109375, + 5424 + ] + }, + { + "id": "cd20f9a7-3a0a-4c4c-98d7-96f423867b87", + "name": "mesh_gap_threshold", + "type": "FLOAT", + "linkIds": [ + 77 + ], + "localized_name": "mesh_gap_threshold", + "pos": [ + -4986.662109375, + 5444 + ] + }, + { + "id": "6f5c15f7-7f77-4fc9-b47b-3514467b06b6", + "name": "mesh_texture", + "type": "BOOLEAN", + "linkIds": [ + 78 + ], + "localized_name": "mesh_texture", + "pos": [ + -4986.662109375, + 5464 + ] + }, + { + "id": "65694805-186e-4181-a721-df8b5af49d31", + "name": "moge_model", + "type": "COMBO", + "linkIds": [ + 79 + ], + "localized_name": "moge_model", + "pos": [ + -4986.662109375, + 5484 + ] + }, + { + "id": "badf1be1-53c6-4fc1-b5cd-79ad3daf1674", + "name": "switch", + "type": "BOOLEAN", + "linkIds": [ + 83 + ], + "label": "auto_resize_input", + "pos": [ + -4986.662109375, + 5504 + ] + } + ], + "outputs": [ + { + "id": "3c616ea0-9a4c-4cff-a405-662320229df0", + "name": "mesh", + "type": "MESH", + "linkIds": [ + 34 + ], + "localized_name": "mesh", + "pos": [ + -3066, + 4990 + ] + }, + { + "id": "ff85a763-b7f7-4bcc-9b1d-a4eaf55ad2f9", + "name": "normal_opengl", + "type": "IMAGE", + "linkIds": [ + 62 + ], + "localized_name": "normal_opengl", + "pos": [ + -3066, + 5010 + ] + }, + { + "id": "26b3f88a-0ba0-4d4d-9c7d-0ad76106c844", + "name": "normal_directx", + "type": "IMAGE", + "linkIds": [ + 63 + ], + "localized_name": "normal_directx", + "pos": [ + -3066, + 5030 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 54, + "type": "MoGePointMapToMesh", + "pos": [ + -3440, + 5220 + ], + "size": [ + 290, + 200 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "moge_geometry", + "name": "moge_geometry", + "type": "MOGE_GEOMETRY", + "link": 33 + }, + { + "localized_name": "batch_index", + "name": "batch_index", + "type": "INT", + "widget": { + "name": "batch_index" + }, + "link": 75 + }, + { + "localized_name": "decimation", + "name": "decimation", + "type": "INT", + "widget": { + "name": "decimation" + }, + "link": 76 + }, + { + "localized_name": "discontinuity_threshold", + "name": "discontinuity_threshold", + "type": "FLOAT", + "widget": { + "name": "discontinuity_threshold" + }, + "link": 77 + }, + { + "localized_name": "texture", + "name": "texture", + "type": "BOOLEAN", + "widget": { + "name": "texture" + }, + "link": 78 + } + ], + "outputs": [ + { + "localized_name": "MESH", + "name": "MESH", + "type": "MESH", + "links": [ + 34 + ] + } + ], + "properties": { + "Node name for S&R": "MoGePointMapToMesh", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 0, + 1, + 0.04, + true + ] + }, + { + "id": 55, + "type": "MoGeInference", + "pos": [ + -3790, + 5180 + ], + "size": [ + 270, + 230 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "moge_model", + "name": "moge_model", + "type": "MOGE_MODEL", + "link": 58 + }, + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 81 + }, + { + "localized_name": "resolution_level", + "name": "resolution_level", + "type": "INT", + "widget": { + "name": "resolution_level" + }, + "link": 73 + }, + { + "localized_name": "fov_x_degrees", + "name": "fov_x_degrees", + "type": "FLOAT", + "widget": { + "name": "fov_x_degrees" + }, + "link": null + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "widget": { + "name": "batch_size" + }, + "link": 74 + }, + { + "localized_name": "force_projection", + "name": "force_projection", + "type": "BOOLEAN", + "widget": { + "name": "force_projection" + }, + "link": null + }, + { + "localized_name": "apply_mask", + "name": "apply_mask", + "type": "BOOLEAN", + "widget": { + "name": "apply_mask" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "moge_geometry", + "name": "moge_geometry", + "type": "MOGE_GEOMETRY", + "links": [ + 33, + 59, + 60 + ] + } + ], + "properties": { + "Node name for S&R": "MoGeInference", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 9, + 0, + 4, + true, + true + ] + }, + { + "id": 58, + "type": "LoadMoGeModel", + "pos": [ + -4180, + 4910 + ], + "size": [ + 270, + 140 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "model_name", + "name": "model_name", + "type": "COMBO", + "widget": { + "name": "model_name" + }, + "link": 79 + } + ], + "outputs": [ + { + "localized_name": "MOGE_MODEL", + "name": "MOGE_MODEL", + "type": "MOGE_MODEL", + "links": [ + 58 + ] + } + ], + "properties": { + "Node name for S&R": "LoadMoGeModel", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "models": [ + { + "name": "moge_2_vitl_normal_fp16.safetensors", + "url": "https://huggingface.co/Comfy-Org/MoGe/resolve/main/geometry_estimation/moge_2_vitl_normal_fp16.safetensors", + "directory": "geometry_estimation" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "moge_2_vitl_normal_fp16.safetensors" + ] + }, + { + "id": 59, + "type": "ComfyMathExpression", + "pos": [ + -4720, + 4910 + ], + "size": [ + 400, + 200 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "label": "a", + "localized_name": "values.a", + "name": "values.a", + "type": "FLOAT,INT,BOOLEAN", + "link": 49 + }, + { + "label": "b", + "localized_name": "values.b", + "name": "values.b", + "shape": 7, + "type": "FLOAT,INT,BOOLEAN", + "link": null + }, + { + "localized_name": "expression", + "name": "expression", + "type": "STRING", + "widget": { + "name": "expression" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "FLOAT", + "name": "FLOAT", + "type": "FLOAT", + "links": null + }, + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": null + }, + { + "localized_name": "BOOL", + "name": "BOOL", + "type": "BOOLEAN", + "links": [ + 53 + ] + } + ], + "properties": { + "Node name for S&R": "ComfyMathExpression", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "a > 2048" + ] + }, + { + "id": 60, + "type": "GetImageSize", + "pos": [ + -4980, + 4910 + ], + "size": [ + 230, + 160 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 48 + } + ], + "outputs": [ + { + "localized_name": "width", + "name": "width", + "type": "INT", + "links": [ + 49 + ] + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "links": null + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "links": null + } + ], + "properties": { + "Node name for S&R": "GetImageSize", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 61, + "type": "ResizeImagesByLongerEdge", + "pos": [ + -4650, + 5210 + ], + "size": [ + 310, + 110 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "link": 55 + }, + { + "localized_name": "longer_edge", + "name": "longer_edge", + "type": "INT", + "widget": { + "name": "longer_edge" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "links": [ + 54 + ] + } + ], + "properties": { + "Node name for S&R": "ResizeImagesByLongerEdge", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 2048 + ] + }, + { + "id": 62, + "type": "ComfySwitchNode", + "pos": [ + -4180, + 5120 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 6, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 56 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 54 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 53 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 80 + ] + } + ], + "properties": { + "Node name for S&R": "ComfySwitchNode", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + false + ] + }, + { + "id": 63, + "type": "MoGeRender", + "pos": [ + -3430, + 4890 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 7, + "mode": 0, + "inputs": [ + { + "localized_name": "moge_geometry", + "name": "moge_geometry", + "type": "MOGE_GEOMETRY", + "link": 59 + }, + { + "localized_name": "output", + "name": "output", + "type": "COMBO", + "widget": { + "name": "output" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 62 + ] + } + ], + "properties": { + "Node name for S&R": "MoGeRender", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "normal_opengl" + ] + }, + { + "id": 64, + "type": "MoGeRender", + "pos": [ + -3430, + 5050 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 8, + "mode": 0, + "inputs": [ + { + "localized_name": "moge_geometry", + "name": "moge_geometry", + "type": "MOGE_GEOMETRY", + "link": 60 + }, + { + "localized_name": "output", + "name": "output", + "type": "COMBO", + "widget": { + "name": "output" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 63 + ] + } + ], + "properties": { + "Node name for S&R": "MoGeRender", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "normal_directx" + ] + }, + { + "id": 66, + "type": "ComfySwitchNode", + "pos": [ + -4160, + 5340 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 9, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 82 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 80 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 83 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 81 + ] + } + ], + "properties": { + "Node name for S&R": "ComfySwitchNode", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + true + ] + } + ], + "groups": [ + { + "id": 1, + "title": "auto_resize_if_width_gt_2048", + "bounding": [ + -5000, + 4840, + 690, + 280 + ], + "color": "#3f789e", + "flags": {} + } + ], + "links": [ + { + "id": 33, + "origin_id": 55, + "origin_slot": 0, + "target_id": 54, + "target_slot": 0, + "type": "MOGE_GEOMETRY" + }, + { + "id": 58, + "origin_id": 58, + "origin_slot": 0, + "target_id": 55, + "target_slot": 0, + "type": "MOGE_MODEL" + }, + { + "id": 49, + "origin_id": 60, + "origin_slot": 0, + "target_id": 59, + "target_slot": 0, + "type": "INT" + }, + { + "id": 54, + "origin_id": 61, + "origin_slot": 0, + "target_id": 62, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 53, + "origin_id": 59, + "origin_slot": 2, + "target_id": 62, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 59, + "origin_id": 55, + "origin_slot": 0, + "target_id": 63, + "target_slot": 0, + "type": "MOGE_GEOMETRY" + }, + { + "id": 60, + "origin_id": 55, + "origin_slot": 0, + "target_id": 64, + "target_slot": 0, + "type": "MOGE_GEOMETRY" + }, + { + "id": 48, + "origin_id": -10, + "origin_slot": 0, + "target_id": 60, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 55, + "origin_id": -10, + "origin_slot": 0, + "target_id": 61, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 56, + "origin_id": -10, + "origin_slot": 0, + "target_id": 62, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 34, + "origin_id": 54, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "MESH" + }, + { + "id": 62, + "origin_id": 63, + "origin_slot": 0, + "target_id": -20, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 63, + "origin_id": 64, + "origin_slot": 0, + "target_id": -20, + "target_slot": 2, + "type": "IMAGE" + }, + { + "id": 73, + "origin_id": -10, + "origin_slot": 1, + "target_id": 55, + "target_slot": 2, + "type": "INT" + }, + { + "id": 74, + "origin_id": -10, + "origin_slot": 2, + "target_id": 55, + "target_slot": 4, + "type": "INT" + }, + { + "id": 75, + "origin_id": -10, + "origin_slot": 3, + "target_id": 54, + "target_slot": 1, + "type": "INT" + }, + { + "id": 76, + "origin_id": -10, + "origin_slot": 4, + "target_id": 54, + "target_slot": 2, + "type": "INT" + }, + { + "id": 77, + "origin_id": -10, + "origin_slot": 5, + "target_id": 54, + "target_slot": 3, + "type": "FLOAT" + }, + { + "id": 78, + "origin_id": -10, + "origin_slot": 6, + "target_id": 54, + "target_slot": 4, + "type": "BOOLEAN" + }, + { + "id": 79, + "origin_id": -10, + "origin_slot": 7, + "target_id": 58, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 80, + "origin_id": 62, + "origin_slot": 0, + "target_id": 66, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 81, + "origin_id": 66, + "origin_slot": 0, + "target_id": 55, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 82, + "origin_id": -10, + "origin_slot": 0, + "target_id": 66, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 83, + "origin_id": -10, + "origin_slot": 8, + "target_id": 66, + "target_slot": 2, + "type": "BOOLEAN" + } + ], + "category": "3D/Geometry Estimation", + "description": "Estimates 3D scene geometry from an input image using MoGe, outputting a mesh plus OpenGL and DirectX normal maps.", + "extra": {} + } + ] + }, + "extra": {} +} \ No newline at end of file diff --git a/blueprints/Image Captioning (gemini).json b/blueprints/Image Captioning (gemini).json index 2fc5d6746..9005e5191 100644 --- a/blueprints/Image Captioning (gemini).json +++ b/blueprints/Image Captioning (gemini).json @@ -310,9 +310,9 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Text generation/Image Captioning", + "category": "Image Tools", "description": "Generates descriptive captions for images using Google's Gemini multimodal LLM." } ] } -} +} \ No newline at end of file diff --git a/blueprints/Image to Depth Map (Lotus).json b/blueprints/Image Depth Estimation (Lotus Depth).json similarity index 92% rename from blueprints/Image to Depth Map (Lotus).json rename to blueprints/Image Depth Estimation (Lotus Depth).json index 12f10ba5b..8aa338d0d 100644 --- a/blueprints/Image to Depth Map (Lotus).json +++ b/blueprints/Image Depth Estimation (Lotus Depth).json @@ -1,19 +1,18 @@ { - "id": "6af0a6c1-0161-4528-8685-65776e838d44", "revision": 0, - "last_node_id": 75, - "last_link_id": 245, + "last_node_id": 76, + "last_link_id": 0, "nodes": [ { - "id": 75, - "type": "488652fd-6edf-4d06-8f9f-4d84d3a34eaf", + "id": 76, + "type": "96338968-1242-4f02-b6a1-d496af4bcffe", "pos": [ - 600, - 830 + 670, + 1280 ], "size": [ 400, - 110 + 201.3125 ], "flags": {}, "order": 0, @@ -59,47 +58,44 @@ "links": [] } ], + "title": "Image Depth Estimation (Lotus Depth)", "properties": { "proxyWidgets": [ [ - "-1", + "28", "sigma" ], [ - "-1", + "10", "unet_name" ], [ - "-1", + "14", "vae_name" ] ], "cnr_id": "comfy-core", "ver": "0.14.1" }, - "widgets_values": [ - 999.0000000000002, - "lotus-depth-d-v1-1.safetensors", - "vae-ft-mse-840000-ema-pruned.safetensors" - ] + "widgets_values": [] } ], "links": [], - "groups": [], + "version": 0.4, "definitions": { "subgraphs": [ { - "id": "488652fd-6edf-4d06-8f9f-4d84d3a34eaf", + "id": "96338968-1242-4f02-b6a1-d496af4bcffe", "version": 1, "state": { "lastGroupId": 1, - "lastNodeId": 75, + "lastNodeId": 76, "lastLinkId": 245, "lastRerouteId": 0 }, "revision": 0, "config": {}, - "name": "Image to Depth Map (Lotus)", + "name": "Image Depth Estimation (Lotus Depth)", "inputNode": { "id": -10, "bounding": [ @@ -191,12 +187,12 @@ "id": 10, "type": "UNETLoader", "pos": [ - 108.05555555555557, - -253.05555555555557 + 110, + -250 ], "size": [ - 254.93706597222226, - 82 + 260, + 90 ], "flags": {}, "order": 4, @@ -234,9 +230,9 @@ } ], "properties": { + "Node name for S&R": "UNETLoader", "cnr_id": "comfy-core", "ver": "0.3.34", - "Node name for S&R": "UNETLoader", "models": [ { "name": "lotus-depth-d-v1-1.safetensors", @@ -255,12 +251,12 @@ "id": 18, "type": "DisableNoise", "pos": [ - 607.0641494069639, - -268.33337840371513 + 610, + -270 ], "size": [ - 175, - 33.333333333333336 + 180, + 40 ], "flags": {}, "order": 0, @@ -278,26 +274,25 @@ } ], "properties": { + "Node name for S&R": "DisableNoise", "cnr_id": "comfy-core", "ver": "0.3.34", - "Node name for S&R": "DisableNoise", "widget_ue_connectable": {} - }, - "widgets_values": [] + } }, { - "id": 23, + "id": 74, "type": "VAEEncode", "pos": [ 620, 160 ], "size": [ - 175, + 180, 50 ], "flags": {}, - "order": 10, + "order": 11, "mode": 0, "inputs": [ { @@ -325,12 +320,11 @@ } ], "properties": { + "Node name for S&R": "VAEEncode", "cnr_id": "comfy-core", "ver": "0.3.34", - "Node name for S&R": "VAEEncode", "widget_ue_connectable": {} - }, - "widgets_values": [] + } }, { "id": 21, @@ -341,7 +335,7 @@ ], "size": [ 210, - 58 + 60 ], "flags": {}, "order": 1, @@ -369,9 +363,9 @@ } ], "properties": { + "Node name for S&R": "KSamplerSelect", "cnr_id": "comfy-core", "ver": "0.3.34", - "Node name for S&R": "KSamplerSelect", "widget_ue_connectable": {} }, "widgets_values": [ @@ -386,7 +380,7 @@ -170 ], "size": [ - 175, + 180, 50 ], "flags": {}, @@ -418,12 +412,11 @@ } ], "properties": { + "Node name for S&R": "BasicGuider", "cnr_id": "comfy-core", "ver": "0.3.34", - "Node name for S&R": "BasicGuider", "widget_ue_connectable": {} - }, - "widgets_values": [] + } }, { "id": 16, @@ -433,8 +426,8 @@ -130 ], "size": [ - 295.99609375, - 271.65798611111114 + 300, + 280 ], "flags": {}, "order": 6, @@ -490,12 +483,11 @@ } ], "properties": { + "Node name for S&R": "SamplerCustomAdvanced", "cnr_id": "comfy-core", "ver": "0.3.34", - "Node name for S&R": "SamplerCustomAdvanced", "widget_ue_connectable": {} - }, - "widgets_values": [] + } }, { "id": 28, @@ -506,10 +498,10 @@ ], "size": [ 210, - 58 + 60 ], "flags": {}, - "order": 11, + "order": 10, "mode": 0, "inputs": [ { @@ -540,9 +532,9 @@ } ], "properties": { + "Node name for S&R": "SetFirstSigma", "cnr_id": "comfy-core", "ver": "0.3.34", - "Node name for S&R": "SetFirstSigma", "widget_ue_connectable": {} }, "widgets_values": [ @@ -557,7 +549,7 @@ -120 ], "size": [ - 175, + 180, 50 ], "flags": {}, @@ -589,12 +581,11 @@ } ], "properties": { + "Node name for S&R": "VAEDecode", "cnr_id": "comfy-core", "ver": "0.3.34", - "Node name for S&R": "VAEDecode", "widget_ue_connectable": {} - }, - "widgets_values": [] + } }, { "id": 22, @@ -604,8 +595,8 @@ -220 ], "size": [ - 175, - 33.333333333333336 + 180, + 40 ], "flags": {}, "order": 9, @@ -630,12 +621,11 @@ } ], "properties": { + "Node name for S&R": "ImageInvert", "cnr_id": "comfy-core", "ver": "0.3.34", - "Node name for S&R": "ImageInvert", "widget_ue_connectable": {} - }, - "widgets_values": [] + } }, { "id": 14, @@ -645,8 +635,8 @@ -90 ], "size": [ - 254.93706597222226, - 58 + 260, + 60 ], "flags": {}, "order": 5, @@ -675,9 +665,9 @@ } ], "properties": { + "Node name for S&R": "VAELoader", "cnr_id": "comfy-core", "ver": "0.3.34", - "Node name for S&R": "VAELoader", "models": [ { "name": "vae-ft-mse-840000-ema-pruned.safetensors", @@ -692,15 +682,15 @@ ] }, { - "id": 68, + "id": 75, "type": "LotusConditioning", "pos": [ 400, -150 ], "size": [ - 175, - 33.333333333333336 + 180, + 40 ], "flags": {}, "order": 2, @@ -718,12 +708,11 @@ } ], "properties": { + "Node name for S&R": "LotusConditioning", "cnr_id": "comfy-core", "ver": "0.3.34", - "Node name for S&R": "LotusConditioning", "widget_ue_connectable": {} - }, - "widgets_values": [] + } }, { "id": 20, @@ -734,7 +723,7 @@ ], "size": [ 210, - 106 + 110 ], "flags": {}, "order": 8, @@ -786,9 +775,9 @@ } ], "properties": { + "Node name for S&R": "BasicScheduler", "cnr_id": "comfy-core", "ver": "0.3.34", - "Node name for S&R": "BasicScheduler", "widget_ue_connectable": {} }, "widgets_values": [ @@ -850,7 +839,7 @@ }, { "id": 201, - "origin_id": 23, + "origin_id": 74, "origin_slot": 0, "target_id": 16, "target_slot": 4, @@ -866,7 +855,7 @@ }, { "id": 238, - "origin_id": 68, + "origin_id": 75, "origin_slot": 0, "target_id": 19, "target_slot": 1, @@ -892,7 +881,7 @@ "id": 38, "origin_id": 14, "origin_slot": 0, - "target_id": 23, + "target_id": 74, "target_slot": 1, "type": "VAE" }, @@ -908,7 +897,7 @@ "id": 37, "origin_id": -10, "origin_slot": 0, - "target_id": 23, + "target_id": 74, "target_slot": 0, "type": "IMAGE" }, @@ -948,12 +937,11 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image generation and editing/Depth to image", + "category": "Conditioning & Preprocessors/Depth", "description": "Estimates a monocular depth map from an input image using the Lotus depth estimation model." } ] }, - "config": {}, "extra": { "ds": { "scale": 1.3589709866044692, @@ -961,8 +949,6 @@ -138.53613935617864, -786.0629126022195 ] - }, - "workflowRendererVersion": "LG" - }, - "version": 0.4 + } + } } \ No newline at end of file diff --git a/blueprints/Image Depth Estimation (MoGe).json b/blueprints/Image Depth Estimation (MoGe).json new file mode 100644 index 000000000..e2d5d1298 --- /dev/null +++ b/blueprints/Image Depth Estimation (MoGe).json @@ -0,0 +1,1154 @@ +{ + "revision": 0, + "last_node_id": 49, + "last_link_id": 0, + "nodes": [ + { + "id": 49, + "type": "ca1fac5f-abe5-4729-b7fe-2299f6630a65", + "pos": [ + -3970, + 5000 + ], + "size": [ + 430, + 330 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "source_image", + "name": "source_image", + "type": "IMAGE", + "link": null + }, + { + "localized_name": "inference_resolution", + "name": "inference_resolution", + "type": "INT", + "widget": { + "name": "inference_resolution" + }, + "link": null + }, + { + "localized_name": "inference_batch_size", + "name": "inference_batch_size", + "type": "INT", + "widget": { + "name": "inference_batch_size" + }, + "link": null + }, + { + "localized_name": "moge_model", + "name": "moge_model", + "type": "COMBO", + "widget": { + "name": "moge_model" + }, + "link": null + }, + { + "label": "auto_resize_input", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "depth_colored", + "name": "depth_colored", + "type": "IMAGE", + "links": [] + }, + { + "localized_name": "depth", + "name": "depth", + "type": "IMAGE", + "links": [] + }, + { + "name": "MASK", + "type": "MASK", + "links": [] + } + ], + "title": "Image Depth Estimation (MoGe)", + "properties": { + "proxyWidgets": [ + [ + "13", + "resolution_level" + ], + [ + "13", + "batch_size" + ], + [ + "32", + "model_name" + ], + [ + "53", + "switch" + ] + ], + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [] + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "ca1fac5f-abe5-4729-b7fe-2299f6630a65", + "version": 1, + "state": { + "lastGroupId": 1, + "lastNodeId": 69, + "lastLinkId": 90, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Image Depth Estimation (MoGe)", + "description": "Estimates monocular depth from an input image using MoGe, outputting both raw and colorized depth maps plus a mask.", + "inputNode": { + "id": -10, + "bounding": [ + -5130, + 5320, + 167.337890625, + 148 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + -3090, + 4966, + 129, + 108 + ] + }, + "inputs": [ + { + "id": "cc8ce79d-ba20-4a25-a51c-c2afcd35e520", + "name": "source_image", + "type": "IMAGE", + "linkIds": [ + 48, + 55, + 56, + 82 + ], + "localized_name": "source_image", + "pos": [ + -4986.662109375, + 5344 + ] + }, + { + "id": "06eefa21-8e60-49f3-9a34-35b081f4ae52", + "name": "inference_resolution", + "type": "INT", + "linkIds": [ + 73 + ], + "localized_name": "inference_resolution", + "pos": [ + -4986.662109375, + 5364 + ] + }, + { + "id": "616638fe-f603-4d10-bae9-fc87c134380f", + "name": "inference_batch_size", + "type": "INT", + "linkIds": [ + 74 + ], + "localized_name": "inference_batch_size", + "pos": [ + -4986.662109375, + 5384 + ] + }, + { + "id": "65694805-186e-4181-a721-df8b5af49d31", + "name": "moge_model", + "type": "COMBO", + "linkIds": [ + 79 + ], + "localized_name": "moge_model", + "pos": [ + -4986.662109375, + 5404 + ] + }, + { + "id": "badf1be1-53c6-4fc1-b5cd-79ad3daf1674", + "name": "switch", + "type": "BOOLEAN", + "linkIds": [ + 83 + ], + "label": "auto_resize_input", + "pos": [ + -4986.662109375, + 5424 + ] + } + ], + "outputs": [ + { + "id": "59c37b52-074f-49fc-9731-483f899c12c4", + "name": "depth_colored", + "type": "IMAGE", + "linkIds": [ + 36 + ], + "localized_name": "depth_colored", + "pos": [ + -3066, + 4990 + ] + }, + { + "id": "f583e936-da5c-4630-9901-391fa605c1f8", + "name": "depth", + "type": "IMAGE", + "linkIds": [ + 40 + ], + "localized_name": "depth", + "pos": [ + -3066, + 5010 + ] + }, + { + "id": "6845b6a1-1980-454a-9451-314f24495c1d", + "name": "MASK", + "type": "MASK", + "linkIds": [ + 86 + ], + "pos": [ + -3066, + 5030 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 13, + "type": "MoGeInference", + "pos": [ + -3790, + 5180 + ], + "size": [ + 270, + 230 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "moge_model", + "name": "moge_model", + "type": "MOGE_MODEL", + "link": 58 + }, + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 81 + }, + { + "localized_name": "resolution_level", + "name": "resolution_level", + "type": "INT", + "widget": { + "name": "resolution_level" + }, + "link": 73 + }, + { + "localized_name": "fov_x_degrees", + "name": "fov_x_degrees", + "type": "FLOAT", + "widget": { + "name": "fov_x_degrees" + }, + "link": null + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "widget": { + "name": "batch_size" + }, + "link": 74 + }, + { + "localized_name": "force_projection", + "name": "force_projection", + "type": "BOOLEAN", + "widget": { + "name": "force_projection" + }, + "link": null + }, + { + "localized_name": "apply_mask", + "name": "apply_mask", + "type": "BOOLEAN", + "widget": { + "name": "apply_mask" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "moge_geometry", + "name": "moge_geometry", + "type": "MOGE_GEOMETRY", + "links": [ + 35, + 39, + 61 + ] + } + ], + "properties": { + "Node name for S&R": "MoGeInference", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 3, + 0, + 4, + true, + true + ] + }, + { + "id": 23, + "type": "MoGeRender", + "pos": [ + -3430, + 4870 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "moge_geometry", + "name": "moge_geometry", + "type": "MOGE_GEOMETRY", + "link": 35 + }, + { + "localized_name": "output", + "name": "output", + "type": "COMBO", + "widget": { + "name": "output" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 36 + ] + } + ], + "properties": { + "Node name for S&R": "MoGeRender", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "depth_colored" + ] + }, + { + "id": 25, + "type": "MoGeRender", + "pos": [ + -3430, + 5030 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "moge_geometry", + "name": "moge_geometry", + "type": "MOGE_GEOMETRY", + "link": 39 + }, + { + "localized_name": "output", + "name": "output", + "type": "COMBO", + "widget": { + "name": "output" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 40 + ] + } + ], + "properties": { + "Node name for S&R": "MoGeRender", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "depth" + ] + }, + { + "id": 32, + "type": "LoadMoGeModel", + "pos": [ + -4180, + 4880 + ], + "size": [ + 270, + 140 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "model_name", + "name": "model_name", + "type": "COMBO", + "widget": { + "name": "model_name" + }, + "link": 79 + } + ], + "outputs": [ + { + "localized_name": "MOGE_MODEL", + "name": "MOGE_MODEL", + "type": "MOGE_MODEL", + "links": [ + 58 + ] + } + ], + "properties": { + "Node name for S&R": "LoadMoGeModel", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "models": [ + { + "name": "moge_2_vitl_normal_fp16.safetensors", + "url": "https://huggingface.co/Comfy-Org/MoGe/resolve/main/geometry_estimation/moge_2_vitl_normal_fp16.safetensors", + "directory": "geometry_estimation" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "moge_2_vitl_normal_fp16.safetensors" + ] + }, + { + "id": 36, + "type": "ComfyMathExpression", + "pos": [ + -4720, + 4910 + ], + "size": [ + 400, + 200 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "label": "a", + "localized_name": "values.a", + "name": "values.a", + "type": "FLOAT,INT,BOOLEAN", + "link": 49 + }, + { + "label": "b", + "localized_name": "values.b", + "name": "values.b", + "shape": 7, + "type": "FLOAT,INT,BOOLEAN", + "link": null + }, + { + "localized_name": "expression", + "name": "expression", + "type": "STRING", + "widget": { + "name": "expression" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "FLOAT", + "name": "FLOAT", + "type": "FLOAT", + "links": null + }, + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": null + }, + { + "localized_name": "BOOL", + "name": "BOOL", + "type": "BOOLEAN", + "links": [ + 53 + ] + } + ], + "properties": { + "Node name for S&R": "ComfyMathExpression", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "a > 2048" + ] + }, + { + "id": 37, + "type": "GetImageSize", + "pos": [ + -4980, + 4910 + ], + "size": [ + 230, + 160 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 48 + } + ], + "outputs": [ + { + "localized_name": "width", + "name": "width", + "type": "INT", + "links": [ + 49 + ] + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "links": null + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "links": null + } + ], + "properties": { + "Node name for S&R": "GetImageSize", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 40, + "type": "ResizeImagesByLongerEdge", + "pos": [ + -4650, + 5210 + ], + "size": [ + 310, + 110 + ], + "flags": {}, + "order": 6, + "mode": 0, + "inputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "link": 55 + }, + { + "localized_name": "longer_edge", + "name": "longer_edge", + "type": "INT", + "widget": { + "name": "longer_edge" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "links": [ + 54 + ] + } + ], + "properties": { + "Node name for S&R": "ResizeImagesByLongerEdge", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 2048 + ] + }, + { + "id": 42, + "type": "ComfySwitchNode", + "pos": [ + -4180, + 5060 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 7, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 56 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 54 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 53 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 80 + ] + } + ], + "properties": { + "Node name for S&R": "ComfySwitchNode", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + false + ] + }, + { + "id": 45, + "type": "MoGeRender", + "pos": [ + -3430, + 5200 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 8, + "mode": 0, + "inputs": [ + { + "localized_name": "moge_geometry", + "name": "moge_geometry", + "type": "MOGE_GEOMETRY", + "link": 61 + }, + { + "localized_name": "output", + "name": "output", + "type": "COMBO", + "widget": { + "name": "output" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 85 + ] + } + ], + "properties": { + "Node name for S&R": "MoGeRender", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "mask" + ] + }, + { + "id": 53, + "type": "ComfySwitchNode", + "pos": [ + -4160, + 5340 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 9, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 82 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 80 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 83 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 81 + ] + } + ], + "properties": { + "Node name for S&R": "ComfySwitchNode", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + true + ] + }, + { + "id": 68, + "type": "ImageToMask", + "pos": [ + -3420, + 5360 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 10, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 85 + }, + { + "localized_name": "channel", + "name": "channel", + "type": "COMBO", + "widget": { + "name": "channel" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MASK", + "name": "MASK", + "type": "MASK", + "links": [ + 86 + ] + } + ], + "properties": { + "Node name for S&R": "ImageToMask" + }, + "widgets_values": [ + "red" + ] + } + ], + "groups": [ + { + "id": 1, + "title": "auto_resize_if_width_gt_2048", + "bounding": [ + -5000, + 4840, + 690, + 280 + ], + "color": "#3f789e", + "flags": {} + } + ], + "links": [ + { + "id": 58, + "origin_id": 32, + "origin_slot": 0, + "target_id": 13, + "target_slot": 0, + "type": "MOGE_MODEL" + }, + { + "id": 35, + "origin_id": 13, + "origin_slot": 0, + "target_id": 23, + "target_slot": 0, + "type": "MOGE_GEOMETRY" + }, + { + "id": 39, + "origin_id": 13, + "origin_slot": 0, + "target_id": 25, + "target_slot": 0, + "type": "MOGE_GEOMETRY" + }, + { + "id": 49, + "origin_id": 37, + "origin_slot": 0, + "target_id": 36, + "target_slot": 0, + "type": "INT" + }, + { + "id": 54, + "origin_id": 40, + "origin_slot": 0, + "target_id": 42, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 53, + "origin_id": 36, + "origin_slot": 2, + "target_id": 42, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 61, + "origin_id": 13, + "origin_slot": 0, + "target_id": 45, + "target_slot": 0, + "type": "MOGE_GEOMETRY" + }, + { + "id": 48, + "origin_id": -10, + "origin_slot": 0, + "target_id": 37, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 55, + "origin_id": -10, + "origin_slot": 0, + "target_id": 40, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 56, + "origin_id": -10, + "origin_slot": 0, + "target_id": 42, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 36, + "origin_id": 23, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 40, + "origin_id": 25, + "origin_slot": 0, + "target_id": -20, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 73, + "origin_id": -10, + "origin_slot": 1, + "target_id": 13, + "target_slot": 2, + "type": "INT" + }, + { + "id": 74, + "origin_id": -10, + "origin_slot": 2, + "target_id": 13, + "target_slot": 4, + "type": "INT" + }, + { + "id": 79, + "origin_id": -10, + "origin_slot": 3, + "target_id": 32, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 80, + "origin_id": 42, + "origin_slot": 0, + "target_id": 53, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 81, + "origin_id": 53, + "origin_slot": 0, + "target_id": 13, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 82, + "origin_id": -10, + "origin_slot": 0, + "target_id": 53, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 83, + "origin_id": -10, + "origin_slot": 4, + "target_id": 53, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 85, + "origin_id": 45, + "origin_slot": 0, + "target_id": 68, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 86, + "origin_id": 68, + "origin_slot": 0, + "target_id": -20, + "target_slot": 2, + "type": "MASK" + } + ], + "extra": {}, + "category": "Conditioning & Preprocessors/Depth" + } + ] + }, + "extra": {} +} \ No newline at end of file diff --git a/blueprints/Image Face Detection (Mediapipe).json b/blueprints/Image Face Detection (Mediapipe).json new file mode 100644 index 000000000..e2548d485 --- /dev/null +++ b/blueprints/Image Face Detection (Mediapipe).json @@ -0,0 +1,779 @@ +{ + "revision": 0, + "last_node_id": 33, + "last_link_id": 0, + "nodes": [ + { + "id": 33, + "type": "6062babb-b649-4a71-be9e-20ebce567744", + "pos": [ + -450, + 4240 + ], + "size": [ + 420, + 400 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": null + }, + { + "name": "face_landmarker", + "type": "FACE_LANDMARKER", + "link": null + }, + { + "name": "detector_variant", + "type": "COMBO", + "widget": { + "name": "detector_variant" + }, + "link": null + }, + { + "name": "num_faces", + "type": "INT", + "widget": { + "name": "num_faces" + }, + "link": null + }, + { + "label": "custom_face_oval", + "name": "regions.face_oval", + "type": "BOOLEAN", + "widget": { + "name": "regions.face_oval" + }, + "link": null + }, + { + "label": "custom_lips", + "name": "regions.lips", + "type": "BOOLEAN", + "widget": { + "name": "regions.lips" + }, + "link": null + }, + { + "label": "custom_left_eye", + "name": "regions.left_eye", + "type": "BOOLEAN", + "widget": { + "name": "regions.left_eye" + }, + "link": null + }, + { + "label": "custom_right_eye", + "name": "regions.right_eye", + "type": "BOOLEAN", + "widget": { + "name": "regions.right_eye" + }, + "link": null + }, + { + "label": "custom_irises", + "name": "regions.irises", + "type": "BOOLEAN", + "widget": { + "name": "regions.irises" + }, + "link": null + }, + { + "name": "model_name", + "type": "COMBO", + "widget": { + "name": "model_name" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "face_landmarks", + "name": "face_landmarks", + "type": "FACE_LANDMARKS", + "links": [] + }, + { + "localized_name": "bboxes", + "name": "bboxes", + "type": "BOUNDING_BOX", + "links": [] + }, + { + "label": "mask", + "name": "MASK_1", + "type": "MASK", + "links": [] + } + ], + "title": "Image Face Detection (Mediapipe)", + "properties": { + "proxyWidgets": [ + [ + "11", + "detector_variant" + ], + [ + "11", + "num_faces" + ], + [ + "20", + "regions.face_oval" + ], + [ + "20", + "regions.lips" + ], + [ + "20", + "regions.left_eye" + ], + [ + "20", + "regions.right_eye" + ], + [ + "20", + "regions.irises" + ], + [ + "2", + "model_name" + ] + ], + "cnr_id": "comfy-core", + "ver": "0.22.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [] + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "6062babb-b649-4a71-be9e-20ebce567744", + "version": 1, + "state": { + "lastGroupId": 2, + "lastNodeId": 158, + "lastLinkId": 140, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Image Face Detection (Mediapipe)", + "description": "Detects facial landmarks from an image using MediaPipe, outputting landmark data, face bounding boxes, and an optional face-region mask.", + "inputNode": { + "id": -10, + "bounding": [ + -710, + 4300, + 148.880859375, + 248 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + 140, + 4480, + 137.677734375, + 108 + ] + }, + "inputs": [ + { + "id": "705dc1ae-6dc9-4155-92df-52f816ad451e", + "name": "image", + "type": "IMAGE", + "linkIds": [ + 60 + ], + "localized_name": "image", + "pos": [ + -585.119140625, + 4324 + ] + }, + { + "id": "d6277190-732c-4604-b7cd-d3a9588bf761", + "name": "face_landmarker", + "type": "FACE_LANDMARKER", + "linkIds": [ + 74 + ], + "pos": [ + -585.119140625, + 4344 + ] + }, + { + "id": "ac473a08-6a86-42a7-b460-e70c6c5e1e2b", + "name": "detector_variant", + "type": "COMBO", + "linkIds": [ + 75 + ], + "pos": [ + -585.119140625, + 4364 + ] + }, + { + "id": "1bec2252-ca2d-496e-8a33-33a61d21f897", + "name": "num_faces", + "type": "INT", + "linkIds": [ + 76 + ], + "pos": [ + -585.119140625, + 4384 + ] + }, + { + "id": "17994fa2-0ea0-4c9b-a70a-19789c459c80", + "name": "regions.face_oval", + "type": "BOOLEAN", + "linkIds": [ + 77 + ], + "label": "custom_face_oval", + "pos": [ + -585.119140625, + 4404 + ] + }, + { + "id": "1c6c5893-2aee-4c37-b702-15ef2e20d863", + "name": "regions.lips", + "type": "BOOLEAN", + "linkIds": [ + 78 + ], + "label": "custom_lips", + "pos": [ + -585.119140625, + 4424 + ] + }, + { + "id": "f353fcea-4b6f-42a1-8fdd-32b3aa1e1f09", + "name": "regions.left_eye", + "type": "BOOLEAN", + "linkIds": [ + 79 + ], + "label": "custom_left_eye", + "pos": [ + -585.119140625, + 4444 + ] + }, + { + "id": "1387e121-c1fb-4522-8f0d-43459e11dd86", + "name": "regions.right_eye", + "type": "BOOLEAN", + "linkIds": [ + 80 + ], + "label": "custom_right_eye", + "pos": [ + -585.119140625, + 4464 + ] + }, + { + "id": "14acb0a0-d1f4-48f3-ba31-811b26236ef9", + "name": "regions.irises", + "type": "BOOLEAN", + "linkIds": [ + 81 + ], + "label": "custom_irises", + "pos": [ + -585.119140625, + 4484 + ] + }, + { + "id": "25a82859-87de-42c8-8431-09948665546e", + "name": "model_name", + "type": "COMBO", + "linkIds": [ + 86 + ], + "pos": [ + -585.119140625, + 4504 + ] + } + ], + "outputs": [ + { + "id": "d2ba3f92-e8b1-49c3-9590-cfad56c54cf4", + "name": "face_landmarks", + "type": "FACE_LANDMARKS", + "linkIds": [ + 44 + ], + "localized_name": "face_landmarks", + "pos": [ + 164, + 4504 + ] + }, + { + "id": "4f356bb0-d4c4-4f93-b4cf-0845a65c4e6d", + "name": "bboxes", + "type": "BOUNDING_BOX", + "linkIds": [ + 25 + ], + "localized_name": "bboxes", + "pos": [ + 164, + 4524 + ] + }, + { + "id": "f6309e1d-6397-4363-b38f-778a122abc51", + "name": "MASK_1", + "type": "MASK", + "linkIds": [ + 83 + ], + "label": "mask", + "pos": [ + 164, + 4544 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 11, + "type": "MediaPipeFaceLandmarker", + "pos": [ + -280, + 4280 + ], + "size": [ + 350, + 220 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "face_detection_model", + "name": "face_detection_model", + "type": "FACE_DETECTION_MODEL", + "link": 66 + }, + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 60 + }, + { + "localized_name": "detector_variant", + "name": "detector_variant", + "type": "COMBO", + "widget": { + "name": "detector_variant" + }, + "link": 75 + }, + { + "localized_name": "num_faces", + "name": "num_faces", + "type": "INT", + "widget": { + "name": "num_faces" + }, + "link": 76 + }, + { + "localized_name": "min_confidence", + "name": "min_confidence", + "type": "FLOAT", + "widget": { + "name": "min_confidence" + }, + "link": null + }, + { + "localized_name": "missing_frame_fallback", + "name": "missing_frame_fallback", + "type": "COMBO", + "widget": { + "name": "missing_frame_fallback" + }, + "link": null + }, + { + "name": "face_landmarker", + "type": "FACE_LANDMARKER", + "link": 74 + } + ], + "outputs": [ + { + "localized_name": "face_landmarks", + "name": "face_landmarks", + "type": "FACE_LANDMARKS", + "links": [ + 44, + 46 + ] + }, + { + "localized_name": "bboxes", + "name": "bboxes", + "type": "BOUNDING_BOX", + "links": [ + 25 + ] + } + ], + "properties": { + "Node name for S&R": "MediaPipeFaceLandmarker", + "cnr_id": "comfy-core", + "ver": "0.22.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "full", + 0, + 0.5, + "empty" + ] + }, + { + "id": 2, + "type": "LoadMediaPipeFaceLandmarker", + "pos": [ + -290, + 4060 + ], + "size": [ + 350, + 140 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "model_name", + "name": "model_name", + "type": "COMBO", + "widget": { + "name": "model_name" + }, + "link": 86 + } + ], + "outputs": [ + { + "localized_name": "FACE_DETECTION_MODEL", + "name": "FACE_DETECTION_MODEL", + "type": "FACE_DETECTION_MODEL", + "links": [ + 66 + ] + } + ], + "properties": { + "Node name for S&R": "LoadMediaPipeFaceLandmarker", + "cnr_id": "comfy-core", + "ver": "0.22.0", + "models": [ + { + "name": "mediapipe_face_fp32.safetensors", + "url": "https://huggingface.co/Comfy-Org/mediapipe/resolve/main/detection/mediapipe_face_fp32.safetensors", + "directory": "detection" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "mediapipe_face_fp32.safetensors" + ] + }, + { + "id": 20, + "type": "MediaPipeFaceMask", + "pos": [ + -290, + 4560 + ], + "size": [ + 360, + 180 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "face_landmarks", + "name": "face_landmarks", + "type": "FACE_LANDMARKS", + "link": 46 + }, + { + "localized_name": "regions", + "name": "regions", + "type": "COMFY_DYNAMICCOMBO_V3", + "widget": { + "name": "regions" + }, + "link": null + }, + { + "localized_name": "regions.face_oval", + "name": "regions.face_oval", + "type": "BOOLEAN", + "widget": { + "name": "regions.face_oval" + }, + "link": 77 + }, + { + "localized_name": "regions.lips", + "name": "regions.lips", + "type": "BOOLEAN", + "widget": { + "name": "regions.lips" + }, + "link": 78 + }, + { + "localized_name": "regions.left_eye", + "name": "regions.left_eye", + "type": "BOOLEAN", + "widget": { + "name": "regions.left_eye" + }, + "link": 79 + }, + { + "localized_name": "regions.right_eye", + "name": "regions.right_eye", + "type": "BOOLEAN", + "widget": { + "name": "regions.right_eye" + }, + "link": 80 + }, + { + "localized_name": "regions.irises", + "name": "regions.irises", + "type": "BOOLEAN", + "widget": { + "name": "regions.irises" + }, + "link": 81 + } + ], + "outputs": [ + { + "localized_name": "MASK", + "name": "MASK", + "type": "MASK", + "links": [ + 83 + ] + } + ], + "properties": { + "Node name for S&R": "MediaPipeFaceMask", + "cnr_id": "comfy-core", + "ver": "0.22.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "custom", + true, + false, + false, + false, + false + ] + } + ], + "groups": [], + "links": [ + { + "id": 66, + "origin_id": 2, + "origin_slot": 0, + "target_id": 11, + "target_slot": 0, + "type": "FACE_DETECTION_MODEL" + }, + { + "id": 46, + "origin_id": 11, + "origin_slot": 0, + "target_id": 20, + "target_slot": 0, + "type": "FACE_LANDMARKS" + }, + { + "id": 60, + "origin_id": -10, + "origin_slot": 0, + "target_id": 11, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 44, + "origin_id": 11, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "FACE_LANDMARKS" + }, + { + "id": 25, + "origin_id": 11, + "origin_slot": 1, + "target_id": -20, + "target_slot": 1, + "type": "BOUNDING_BOX" + }, + { + "id": 74, + "origin_id": -10, + "origin_slot": 1, + "target_id": 11, + "target_slot": 6, + "type": "FACE_LANDMARKER" + }, + { + "id": 75, + "origin_id": -10, + "origin_slot": 2, + "target_id": 11, + "target_slot": 2, + "type": "COMBO" + }, + { + "id": 76, + "origin_id": -10, + "origin_slot": 3, + "target_id": 11, + "target_slot": 3, + "type": "INT" + }, + { + "id": 77, + "origin_id": -10, + "origin_slot": 4, + "target_id": 20, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 78, + "origin_id": -10, + "origin_slot": 5, + "target_id": 20, + "target_slot": 3, + "type": "BOOLEAN" + }, + { + "id": 79, + "origin_id": -10, + "origin_slot": 6, + "target_id": 20, + "target_slot": 4, + "type": "BOOLEAN" + }, + { + "id": 80, + "origin_id": -10, + "origin_slot": 7, + "target_id": 20, + "target_slot": 5, + "type": "BOOLEAN" + }, + { + "id": 81, + "origin_id": -10, + "origin_slot": 8, + "target_id": 20, + "target_slot": 6, + "type": "BOOLEAN" + }, + { + "id": 83, + "origin_id": 20, + "origin_slot": 0, + "target_id": -20, + "target_slot": 2, + "type": "MASK" + }, + { + "id": 86, + "origin_id": -10, + "origin_slot": 9, + "target_id": 2, + "target_slot": 0, + "type": "COMBO" + } + ], + "extra": {}, + "category": "Conditioning & Preprocessors/Face Detection" + } + ] + }, + "extra": {} +} \ No newline at end of file diff --git a/blueprints/Image Segmentation (SAM3).json b/blueprints/Image Segmentation (SAM3).json index b405bf623..a2ef40ac8 100644 --- a/blueprints/Image Segmentation (SAM3).json +++ b/blueprints/Image Segmentation (SAM3).json @@ -703,7 +703,7 @@ } ], "extra": {}, - "category": "Image Tools/Image Segmentation", + "category": "Conditioning & Preprocessors/Segmentation & Mask", "description": "Segments images into masks using Meta SAM3 from text prompts, points, or boxes." } ] diff --git a/blueprints/Image Upscale(Z-image-Turbo).json b/blueprints/Image Upscale(Z-image-Turbo).json index bd803a0b1..25d2838a8 100644 --- a/blueprints/Image Upscale(Z-image-Turbo).json +++ b/blueprints/Image Upscale(Z-image-Turbo).json @@ -1302,7 +1302,7 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image generation and editing/Enhance", + "category": "Image generation and editing/Upscale", "description": "Upscales images to higher resolution using Z-Image-Turbo." } ] @@ -1312,4 +1312,4 @@ "workflowRendererVersion": "LG" }, "version": 0.4 -} +} \ No newline at end of file diff --git a/blueprints/Image to Pose Map (SDPose Multi-Person).json b/blueprints/Image to Pose Map (SDPose Multi-Person).json new file mode 100644 index 000000000..38df20775 --- /dev/null +++ b/blueprints/Image to Pose Map (SDPose Multi-Person).json @@ -0,0 +1,1206 @@ +{ + "revision": 0, + "last_node_id": 675, + "last_link_id": 0, + "nodes": [ + { + "id": 675, + "type": "01b6a731-fb78-4070-9a38-c87146da9604", + "pos": [ + -2480, + 3400 + ], + "size": [ + 370, + 590.625 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "localized_name": "input", + "name": "input", + "type": "IMAGE,MASK", + "link": null + }, + { + "label": "resize_target_longer_size", + "name": "resize_type.longer_size", + "type": "INT", + "widget": { + "name": "resize_type.longer_size" + }, + "link": null + }, + { + "name": "scale_method", + "type": "COMBO", + "widget": { + "name": "scale_method" + }, + "link": null + }, + { + "name": "draw_body", + "type": "BOOLEAN", + "widget": { + "name": "draw_body" + }, + "link": null + }, + { + "name": "draw_hands", + "type": "BOOLEAN", + "widget": { + "name": "draw_hands" + }, + "link": null + }, + { + "name": "draw_face", + "type": "BOOLEAN", + "widget": { + "name": "draw_face" + }, + "link": null + }, + { + "name": "draw_feet", + "type": "BOOLEAN", + "widget": { + "name": "draw_feet" + }, + "link": null + }, + { + "name": "stick_width", + "type": "INT", + "widget": { + "name": "stick_width" + }, + "link": null + }, + { + "name": "face_point_size", + "type": "INT", + "widget": { + "name": "face_point_size" + }, + "link": null + }, + { + "name": "score_threshold", + "type": "FLOAT", + "widget": { + "name": "score_threshold" + }, + "link": null + }, + { + "label": "detect_threshold", + "name": "threshold", + "type": "FLOAT", + "widget": { + "name": "threshold" + }, + "link": null + }, + { + "label": "detect_class", + "name": "class_name", + "type": "COMBO", + "widget": { + "name": "class_name" + }, + "link": null + }, + { + "name": "max_detections", + "type": "INT", + "widget": { + "name": "max_detections" + }, + "link": null + }, + { + "name": "ckpt_name", + "type": "COMBO", + "widget": { + "name": "ckpt_name" + }, + "link": null + }, + { + "name": "unet_name", + "type": "COMBO", + "widget": { + "name": "unet_name" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [] + }, + { + "name": "keypoints", + "type": "POSE_KEYPOINT", + "links": null + }, + { + "name": "bboxes", + "type": "BOUNDING_BOX", + "links": [] + } + ], + "properties": { + "proxyWidgets": [ + [ + "674", + "resize_type.longer_size" + ], + [ + "674", + "scale_method" + ], + [ + "672", + "draw_body" + ], + [ + "672", + "draw_hands" + ], + [ + "672", + "draw_face" + ], + [ + "672", + "draw_feet" + ], + [ + "672", + "stick_width" + ], + [ + "672", + "face_point_size" + ], + [ + "672", + "score_threshold" + ], + [ + "678", + "threshold" + ], + [ + "678", + "class_name" + ], + [ + "678", + "max_detections" + ], + [ + "673", + "ckpt_name" + ], + [ + "677", + "unet_name" + ] + ], + "cnr_id": "comfy-core", + "ver": "0.15.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [], + "title": "Image to Pose Map (SDPose Multi-Person)" + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "01b6a731-fb78-4070-9a38-c87146da9604", + "version": 1, + "state": { + "lastGroupId": 2, + "lastNodeId": 691, + "lastLinkId": 1740, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Image to Pose Map (SDPose Multi-Person)", + "inputNode": { + "id": -10, + "bounding": [ + -3350, + 3410, + 190.8984375, + 348 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + -1840, + 3570, + 128, + 108 + ] + }, + "inputs": [ + { + "id": "e24699c3-1356-4634-9eb4-19bb58e5c0b0", + "name": "input", + "type": "IMAGE,MASK", + "linkIds": [ + 1700 + ], + "localized_name": "input", + "pos": [ + -3183.1015625, + 3434 + ] + }, + { + "id": "088eefc1-cd8a-4573-993f-9e4da008a12d", + "name": "resize_type.longer_size", + "type": "INT", + "linkIds": [ + 1704 + ], + "label": "resize_target_longer_size", + "pos": [ + -3183.1015625, + 3454 + ] + }, + { + "id": "b6449bd3-73d4-41c8-b81f-cf8d33f76a2e", + "name": "scale_method", + "type": "COMBO", + "linkIds": [ + 1705 + ], + "pos": [ + -3183.1015625, + 3474 + ] + }, + { + "id": "4cff52ad-ed07-4c97-8803-fcbd89554fd0", + "name": "draw_body", + "type": "BOOLEAN", + "linkIds": [ + 1706 + ], + "pos": [ + -3183.1015625, + 3494 + ] + }, + { + "id": "7af63dce-f7df-4d7e-8215-d7c7f60bf81c", + "name": "draw_hands", + "type": "BOOLEAN", + "linkIds": [ + 1707 + ], + "pos": [ + -3183.1015625, + 3514 + ] + }, + { + "id": "af3a9bce-61f9-4aca-b530-9f65e028b35e", + "name": "draw_face", + "type": "BOOLEAN", + "linkIds": [ + 1708 + ], + "pos": [ + -3183.1015625, + 3534 + ] + }, + { + "id": "4620f6a3-2c85-4b79-ad8f-35d0326b568f", + "name": "draw_feet", + "type": "BOOLEAN", + "linkIds": [ + 1709 + ], + "pos": [ + -3183.1015625, + 3554 + ] + }, + { + "id": "fee5d0c9-8d4b-4934-81d8-ba2206dc56cb", + "name": "stick_width", + "type": "INT", + "linkIds": [ + 1710 + ], + "pos": [ + -3183.1015625, + 3574 + ] + }, + { + "id": "aafdd060-ba81-4324-a9cc-b656e1ebc133", + "name": "face_point_size", + "type": "INT", + "linkIds": [ + 1711 + ], + "pos": [ + -3183.1015625, + 3594 + ] + }, + { + "id": "514c5503-f9e6-4d23-b1ae-1d3291acb2a3", + "name": "score_threshold", + "type": "FLOAT", + "linkIds": [ + 1712 + ], + "pos": [ + -3183.1015625, + 3614 + ] + }, + { + "id": "4eb3e4ea-7a36-4511-8483-0d12aadd32f7", + "name": "threshold", + "type": "FLOAT", + "linkIds": [ + 1718 + ], + "label": "detect_threshold", + "pos": [ + -3183.1015625, + 3634 + ] + }, + { + "id": "c76a7a05-81e6-4b17-a9e0-85f47a5844f2", + "name": "class_name", + "type": "COMBO", + "linkIds": [ + 1719 + ], + "label": "detect_class", + "pos": [ + -3183.1015625, + 3654 + ] + }, + { + "id": "4417e988-6e80-4236-be31-4c179037f5a2", + "name": "max_detections", + "type": "INT", + "linkIds": [ + 1720 + ], + "pos": [ + -3183.1015625, + 3674 + ] + }, + { + "id": "7d7c4a0b-0d1b-4c98-942b-f90548d2a492", + "name": "ckpt_name", + "type": "COMBO", + "linkIds": [ + 1721 + ], + "pos": [ + -3183.1015625, + 3694 + ] + }, + { + "id": "4d75122c-2c14-452a-98fe-d1545d3e012a", + "name": "unet_name", + "type": "COMBO", + "linkIds": [ + 1722 + ], + "pos": [ + -3183.1015625, + 3714 + ] + } + ], + "outputs": [ + { + "id": "f05ed8cc-9403-4f14-8085-4364b06f8a48", + "name": "IMAGE", + "type": "IMAGE", + "linkIds": [ + 1701 + ], + "localized_name": "IMAGE", + "pos": [ + -1816, + 3594 + ] + }, + { + "id": "4b64118e-3cef-4eeb-9dad-4cd09cfd63a2", + "name": "keypoints", + "type": "POSE_KEYPOINT", + "linkIds": [ + 1725 + ], + "pos": [ + -1816, + 3614 + ] + }, + { + "id": "a27f7e34-dcbc-4fb0-a4e1-2c5fc423ca5f", + "name": "bboxes", + "type": "BOUNDING_BOX", + "linkIds": [ + 1726 + ], + "pos": [ + -1816, + 3634 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 671, + "type": "SDPoseKeypointExtractor", + "pos": [ + -2550, + 3080 + ], + "size": [ + 270, + 180 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 1696 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 1697 + }, + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 1698 + }, + { + "localized_name": "bboxes", + "name": "bboxes", + "shape": 7, + "type": "BOUNDING_BOX", + "link": 1717 + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "widget": { + "name": "batch_size" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "keypoints", + "name": "keypoints", + "type": "POSE_KEYPOINT", + "links": [ + 1699, + 1725 + ] + } + ], + "properties": { + "Node name for S&R": "SDPoseKeypointExtractor", + "cnr_id": "comfy-core", + "ver": "0.15.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 16 + ] + }, + { + "id": 674, + "type": "ResizeImageMaskNode", + "pos": [ + -2970, + 3580 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "input", + "name": "input", + "type": "IMAGE,MASK", + "link": 1700 + }, + { + "localized_name": "resize_type", + "name": "resize_type", + "type": "COMFY_DYNAMICCOMBO_V3", + "widget": { + "name": "resize_type" + }, + "link": null + }, + { + "localized_name": "resize_type.longer_size", + "name": "resize_type.longer_size", + "type": "INT", + "widget": { + "name": "resize_type.longer_size" + }, + "link": 1704 + }, + { + "localized_name": "scale_method", + "name": "scale_method", + "type": "COMBO", + "widget": { + "name": "scale_method" + }, + "link": 1705 + } + ], + "outputs": [ + { + "localized_name": "resized", + "name": "resized", + "type": "*", + "links": [ + 1698, + 1716 + ] + } + ], + "properties": { + "Node name for S&R": "ResizeImageMaskNode", + "cnr_id": "comfy-core", + "ver": "0.15.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "scale longer dimension", + 1024, + "lanczos" + ] + }, + { + "id": 672, + "type": "SDPoseDrawKeypoints", + "pos": [ + -2540, + 3590 + ], + "size": [ + 270, + 280 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "keypoints", + "name": "keypoints", + "type": "POSE_KEYPOINT", + "link": 1699 + }, + { + "localized_name": "draw_body", + "name": "draw_body", + "type": "BOOLEAN", + "widget": { + "name": "draw_body" + }, + "link": 1706 + }, + { + "localized_name": "draw_hands", + "name": "draw_hands", + "type": "BOOLEAN", + "widget": { + "name": "draw_hands" + }, + "link": 1707 + }, + { + "localized_name": "draw_face", + "name": "draw_face", + "type": "BOOLEAN", + "widget": { + "name": "draw_face" + }, + "link": 1708 + }, + { + "localized_name": "draw_feet", + "name": "draw_feet", + "type": "BOOLEAN", + "widget": { + "name": "draw_feet" + }, + "link": 1709 + }, + { + "localized_name": "stick_width", + "name": "stick_width", + "type": "INT", + "widget": { + "name": "stick_width" + }, + "link": 1710 + }, + { + "localized_name": "face_point_size", + "name": "face_point_size", + "type": "INT", + "widget": { + "name": "face_point_size" + }, + "link": 1711 + }, + { + "localized_name": "score_threshold", + "name": "score_threshold", + "type": "FLOAT", + "widget": { + "name": "score_threshold" + }, + "link": 1712 + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 1701 + ] + } + ], + "properties": { + "Node name for S&R": "SDPoseDrawKeypoints", + "cnr_id": "comfy-core", + "ver": "0.15.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + true, + true, + true, + true, + 4, + 2, + 0.5 + ] + }, + { + "id": 673, + "type": "CheckpointLoaderSimple", + "pos": [ + -3040, + 3080 + ], + "size": [ + 390, + 190 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "ckpt_name", + "name": "ckpt_name", + "type": "COMBO", + "widget": { + "name": "ckpt_name" + }, + "link": 1721 + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "links": [ + 1696 + ] + }, + { + "localized_name": "CLIP", + "name": "CLIP", + "type": "CLIP", + "links": [] + }, + { + "localized_name": "VAE", + "name": "VAE", + "type": "VAE", + "links": [ + 1697 + ] + } + ], + "properties": { + "Node name for S&R": "CheckpointLoaderSimple", + "cnr_id": "comfy-core", + "ver": "0.15.0", + "models": [ + { + "name": "sdpose_wholebody_fp16.safetensors", + "url": "https://huggingface.co/Comfy-Org/SDPose/resolve/main/checkpoints/sdpose_wholebody_fp16.safetensors", + "directory": "checkpoints" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "sdpose_wholebody_fp16.safetensors" + ] + }, + { + "id": 677, + "type": "UNETLoader", + "pos": [ + -3030, + 3330 + ], + "size": [ + 370, + 140 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "localized_name": "unet_name", + "name": "unet_name", + "type": "COMBO", + "widget": { + "name": "unet_name" + }, + "link": 1722 + }, + { + "localized_name": "weight_dtype", + "name": "weight_dtype", + "type": "COMBO", + "widget": { + "name": "weight_dtype" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "links": [ + 1715 + ] + } + ], + "properties": { + "Node name for S&R": "UNETLoader", + "cnr_id": "comfy-core", + "ver": "0.14.1", + "models": [ + { + "name": "rt_detr_v4-x-hgnet_fp16.safetensors", + "url": "https://huggingface.co/Comfy-Org/SDPose/resolve/main/diffusion_models/rt_detr_v4-x-hgnet_fp16.safetensors", + "directory": "diffusion_models" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "rt_detr_v4-x-hgnet_fp16.safetensors", + "default" + ] + }, + { + "id": 678, + "type": "RTDETR_detect", + "pos": [ + -2540, + 3320 + ], + "size": [ + 270, + 200 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "label": "model", + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 1715 + }, + { + "label": "image", + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 1716 + }, + { + "localized_name": "threshold", + "name": "threshold", + "type": "FLOAT", + "widget": { + "name": "threshold" + }, + "link": 1718 + }, + { + "localized_name": "class_name", + "name": "class_name", + "type": "COMBO", + "widget": { + "name": "class_name" + }, + "link": 1719 + }, + { + "localized_name": "max_detections", + "name": "max_detections", + "type": "INT", + "widget": { + "name": "max_detections" + }, + "link": 1720 + } + ], + "outputs": [ + { + "localized_name": "bboxes", + "name": "bboxes", + "type": "BOUNDING_BOX", + "links": [ + 1717, + 1726 + ] + } + ], + "properties": { + "Node name for S&R": "RTDETR_detect", + "cnr_id": "comfy-core", + "ver": "0.15.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 0.5, + "person", + 1 + ] + } + ], + "groups": [], + "links": [ + { + "id": 1696, + "origin_id": 673, + "origin_slot": 0, + "target_id": 671, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 1697, + "origin_id": 673, + "origin_slot": 2, + "target_id": 671, + "target_slot": 1, + "type": "VAE" + }, + { + "id": 1698, + "origin_id": 674, + "origin_slot": 0, + "target_id": 671, + "target_slot": 2, + "type": "IMAGE" + }, + { + "id": 1699, + "origin_id": 671, + "origin_slot": 0, + "target_id": 672, + "target_slot": 0, + "type": "POSE_KEYPOINT" + }, + { + "id": 1700, + "origin_id": -10, + "origin_slot": 0, + "target_id": 674, + "target_slot": 0, + "type": "IMAGE,MASK" + }, + { + "id": 1701, + "origin_id": 672, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 1704, + "origin_id": -10, + "origin_slot": 1, + "target_id": 674, + "target_slot": 2, + "type": "INT" + }, + { + "id": 1705, + "origin_id": -10, + "origin_slot": 2, + "target_id": 674, + "target_slot": 3, + "type": "COMBO" + }, + { + "id": 1706, + "origin_id": -10, + "origin_slot": 3, + "target_id": 672, + "target_slot": 1, + "type": "BOOLEAN" + }, + { + "id": 1707, + "origin_id": -10, + "origin_slot": 4, + "target_id": 672, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 1708, + "origin_id": -10, + "origin_slot": 5, + "target_id": 672, + "target_slot": 3, + "type": "BOOLEAN" + }, + { + "id": 1709, + "origin_id": -10, + "origin_slot": 6, + "target_id": 672, + "target_slot": 4, + "type": "BOOLEAN" + }, + { + "id": 1710, + "origin_id": -10, + "origin_slot": 7, + "target_id": 672, + "target_slot": 5, + "type": "INT" + }, + { + "id": 1711, + "origin_id": -10, + "origin_slot": 8, + "target_id": 672, + "target_slot": 6, + "type": "INT" + }, + { + "id": 1712, + "origin_id": -10, + "origin_slot": 9, + "target_id": 672, + "target_slot": 7, + "type": "FLOAT" + }, + { + "id": 1715, + "origin_id": 677, + "origin_slot": 0, + "target_id": 678, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 1716, + "origin_id": 674, + "origin_slot": 0, + "target_id": 678, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 1717, + "origin_id": 678, + "origin_slot": 0, + "target_id": 671, + "target_slot": 3, + "type": "BOUNDING_BOX" + }, + { + "id": 1718, + "origin_id": -10, + "origin_slot": 10, + "target_id": 678, + "target_slot": 2, + "type": "FLOAT" + }, + { + "id": 1719, + "origin_id": -10, + "origin_slot": 11, + "target_id": 678, + "target_slot": 3, + "type": "COMBO" + }, + { + "id": 1720, + "origin_id": -10, + "origin_slot": 12, + "target_id": 678, + "target_slot": 4, + "type": "INT" + }, + { + "id": 1721, + "origin_id": -10, + "origin_slot": 13, + "target_id": 673, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 1722, + "origin_id": -10, + "origin_slot": 14, + "target_id": 677, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 1725, + "origin_id": 671, + "origin_slot": 0, + "target_id": -20, + "target_slot": 1, + "type": "POSE_KEYPOINT" + }, + { + "id": 1726, + "origin_id": 678, + "origin_slot": 0, + "target_id": -20, + "target_slot": 2, + "type": "BOUNDING_BOX" + } + ], + "extra": { + "workflowRendererVersion": "LG" + }, + "category": "Conditioning & Preprocessors/Pose", + "description": "Detects multiple people in an image and outputs per-person pose keypoints, skeleton renders, and bounding boxes using SDPose." + } + ] + }, + "extra": {} +} \ No newline at end of file diff --git a/blueprints/Image to Pose Map (SDPose-OOD).json b/blueprints/Image to Pose Map (SDPose-OOD).json new file mode 100644 index 000000000..76ee9ff4e --- /dev/null +++ b/blueprints/Image to Pose Map (SDPose-OOD).json @@ -0,0 +1,888 @@ +{ + "revision": 0, + "last_node_id": 675, + "last_link_id": 0, + "nodes": [ + { + "id": 675, + "type": "01b6a731-fb78-4070-9a38-c87146da9604", + "pos": [ + -2480, + 3400 + ], + "size": [ + 360, + 433.3125 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "input", + "name": "input", + "type": "IMAGE,MASK", + "link": null + }, + { + "label": "resize_target_longer_size", + "name": "resize_type.longer_size", + "type": "INT", + "widget": { + "name": "resize_type.longer_size" + }, + "link": null + }, + { + "name": "scale_method", + "type": "COMBO", + "widget": { + "name": "scale_method" + }, + "link": null + }, + { + "name": "draw_body", + "type": "BOOLEAN", + "widget": { + "name": "draw_body" + }, + "link": null + }, + { + "name": "draw_hands", + "type": "BOOLEAN", + "widget": { + "name": "draw_hands" + }, + "link": null + }, + { + "name": "draw_face", + "type": "BOOLEAN", + "widget": { + "name": "draw_face" + }, + "link": null + }, + { + "name": "draw_feet", + "type": "BOOLEAN", + "widget": { + "name": "draw_feet" + }, + "link": null + }, + { + "name": "stick_width", + "type": "INT", + "widget": { + "name": "stick_width" + }, + "link": null + }, + { + "name": "face_point_size", + "type": "INT", + "widget": { + "name": "face_point_size" + }, + "link": null + }, + { + "name": "score_threshold", + "type": "FLOAT", + "widget": { + "name": "score_threshold" + }, + "link": null + }, + { + "name": "ckpt_name", + "type": "COMBO", + "widget": { + "name": "ckpt_name" + }, + "link": null + }, + { + "name": "bboxes", + "shape": 7, + "type": "BOUNDING_BOX", + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [] + }, + { + "name": "keypoints", + "type": "POSE_KEYPOINT", + "links": null + } + ], + "properties": { + "proxyWidgets": [ + [ + "674", + "resize_type.longer_size" + ], + [ + "674", + "scale_method" + ], + [ + "672", + "draw_body" + ], + [ + "672", + "draw_hands" + ], + [ + "672", + "draw_face" + ], + [ + "672", + "draw_feet" + ], + [ + "672", + "stick_width" + ], + [ + "672", + "face_point_size" + ], + [ + "672", + "score_threshold" + ], + [ + "673", + "ckpt_name" + ] + ], + "cnr_id": "comfy-core", + "ver": "0.15.1", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + } + }, + "widgets_values": [], + "title": "Image to Pose Map (SDPose-OOD)" + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "01b6a731-fb78-4070-9a38-c87146da9604", + "version": 1, + "state": { + "lastGroupId": 0, + "lastNodeId": 676, + "lastLinkId": 1715, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Image to Pose Map (SDPose-OOD)", + "inputNode": { + "id": -10, + "bounding": [ + -3290, + 3590, + 190.8984375, + 288 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + -1756.2451602089645, + 3366, + 128, + 88 + ] + }, + "inputs": [ + { + "id": "e24699c3-1356-4634-9eb4-19bb58e5c0b0", + "name": "input", + "type": "IMAGE,MASK", + "linkIds": [ + 1700 + ], + "localized_name": "input", + "pos": [ + -3123.1015625, + 3614 + ] + }, + { + "id": "088eefc1-cd8a-4573-993f-9e4da008a12d", + "name": "resize_type.longer_size", + "type": "INT", + "linkIds": [ + 1704 + ], + "label": "resize_target_longer_size", + "pos": [ + -3123.1015625, + 3634 + ] + }, + { + "id": "b6449bd3-73d4-41c8-b81f-cf8d33f76a2e", + "name": "scale_method", + "type": "COMBO", + "linkIds": [ + 1705 + ], + "pos": [ + -3123.1015625, + 3654 + ] + }, + { + "id": "4cff52ad-ed07-4c97-8803-fcbd89554fd0", + "name": "draw_body", + "type": "BOOLEAN", + "linkIds": [ + 1706 + ], + "pos": [ + -3123.1015625, + 3674 + ] + }, + { + "id": "7af63dce-f7df-4d7e-8215-d7c7f60bf81c", + "name": "draw_hands", + "type": "BOOLEAN", + "linkIds": [ + 1707 + ], + "pos": [ + -3123.1015625, + 3694 + ] + }, + { + "id": "af3a9bce-61f9-4aca-b530-9f65e028b35e", + "name": "draw_face", + "type": "BOOLEAN", + "linkIds": [ + 1708 + ], + "pos": [ + -3123.1015625, + 3714 + ] + }, + { + "id": "4620f6a3-2c85-4b79-ad8f-35d0326b568f", + "name": "draw_feet", + "type": "BOOLEAN", + "linkIds": [ + 1709 + ], + "pos": [ + -3123.1015625, + 3734 + ] + }, + { + "id": "fee5d0c9-8d4b-4934-81d8-ba2206dc56cb", + "name": "stick_width", + "type": "INT", + "linkIds": [ + 1710 + ], + "pos": [ + -3123.1015625, + 3754 + ] + }, + { + "id": "aafdd060-ba81-4324-a9cc-b656e1ebc133", + "name": "face_point_size", + "type": "INT", + "linkIds": [ + 1711 + ], + "pos": [ + -3123.1015625, + 3774 + ] + }, + { + "id": "514c5503-f9e6-4d23-b1ae-1d3291acb2a3", + "name": "score_threshold", + "type": "FLOAT", + "linkIds": [ + 1712 + ], + "pos": [ + -3123.1015625, + 3794 + ] + }, + { + "id": "ae46de61-2cc6-483e-8ee9-87e4144a2ffa", + "name": "ckpt_name", + "type": "COMBO", + "linkIds": [ + 1713 + ], + "pos": [ + -3123.1015625, + 3814 + ] + }, + { + "id": "41bec0c6-dffa-4c78-9289-ee678715ae54", + "name": "bboxes", + "type": "BOUNDING_BOX", + "linkIds": [ + 1714 + ], + "pos": [ + -3123.1015625, + 3834 + ] + } + ], + "outputs": [ + { + "id": "f05ed8cc-9403-4f14-8085-4364b06f8a48", + "name": "IMAGE", + "type": "IMAGE", + "linkIds": [ + 1701 + ], + "localized_name": "IMAGE", + "pos": [ + -1732.2451602089645, + 3390 + ] + }, + { + "id": "29a6584e-4685-4986-8ffd-e6d8539953fd", + "name": "keypoints", + "type": "POSE_KEYPOINT", + "linkIds": [ + 1715 + ], + "pos": [ + -1732.2451602089645, + 3410 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 671, + "type": "SDPoseKeypointExtractor", + "pos": [ + -2470, + 3250 + ], + "size": [ + 270, + 180 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 1696 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 1697 + }, + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 1698 + }, + { + "localized_name": "bboxes", + "name": "bboxes", + "shape": 7, + "type": "BOUNDING_BOX", + "link": 1714 + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "widget": { + "name": "batch_size" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "keypoints", + "name": "keypoints", + "type": "POSE_KEYPOINT", + "links": [ + 1699, + 1715 + ] + } + ], + "properties": { + "Node name for S&R": "SDPoseKeypointExtractor", + "cnr_id": "comfy-core", + "ver": "0.15.0", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + } + }, + "widgets_values": [ + 16 + ] + }, + { + "id": 674, + "type": "ResizeImageMaskNode", + "pos": [ + -2960, + 3490 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "input", + "name": "input", + "type": "IMAGE,MASK", + "link": 1700 + }, + { + "localized_name": "resize_type", + "name": "resize_type", + "type": "COMFY_DYNAMICCOMBO_V3", + "widget": { + "name": "resize_type" + }, + "link": null + }, + { + "localized_name": "resize_type.longer_size", + "name": "resize_type.longer_size", + "type": "INT", + "widget": { + "name": "resize_type.longer_size" + }, + "link": 1704 + }, + { + "localized_name": "scale_method", + "name": "scale_method", + "type": "COMBO", + "widget": { + "name": "scale_method" + }, + "link": 1705 + } + ], + "outputs": [ + { + "localized_name": "resized", + "name": "resized", + "type": "*", + "links": [ + 1698 + ] + } + ], + "properties": { + "Node name for S&R": "ResizeImageMaskNode", + "cnr_id": "comfy-core", + "ver": "0.15.0", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + } + }, + "widgets_values": [ + "scale longer dimension", + 1024, + "area" + ] + }, + { + "id": 672, + "type": "SDPoseDrawKeypoints", + "pos": [ + -2120, + 3260 + ], + "size": [ + 270, + 280 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "keypoints", + "name": "keypoints", + "type": "POSE_KEYPOINT", + "link": 1699 + }, + { + "localized_name": "draw_body", + "name": "draw_body", + "type": "BOOLEAN", + "widget": { + "name": "draw_body" + }, + "link": 1706 + }, + { + "localized_name": "draw_hands", + "name": "draw_hands", + "type": "BOOLEAN", + "widget": { + "name": "draw_hands" + }, + "link": 1707 + }, + { + "localized_name": "draw_face", + "name": "draw_face", + "type": "BOOLEAN", + "widget": { + "name": "draw_face" + }, + "link": 1708 + }, + { + "localized_name": "draw_feet", + "name": "draw_feet", + "type": "BOOLEAN", + "widget": { + "name": "draw_feet" + }, + "link": 1709 + }, + { + "localized_name": "stick_width", + "name": "stick_width", + "type": "INT", + "widget": { + "name": "stick_width" + }, + "link": 1710 + }, + { + "localized_name": "face_point_size", + "name": "face_point_size", + "type": "INT", + "widget": { + "name": "face_point_size" + }, + "link": 1711 + }, + { + "localized_name": "score_threshold", + "name": "score_threshold", + "type": "FLOAT", + "widget": { + "name": "score_threshold" + }, + "link": 1712 + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 1701 + ] + } + ], + "properties": { + "Node name for S&R": "SDPoseDrawKeypoints", + "cnr_id": "comfy-core", + "ver": "0.15.0", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + } + }, + "widgets_values": [ + true, + true, + true, + true, + 4, + 2, + 0.5 + ] + }, + { + "id": 673, + "type": "CheckpointLoaderSimple", + "pos": [ + -2960, + 3250 + ], + "size": [ + 390, + 190 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "ckpt_name", + "name": "ckpt_name", + "type": "COMBO", + "widget": { + "name": "ckpt_name" + }, + "link": 1713 + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "links": [ + 1696 + ] + }, + { + "localized_name": "CLIP", + "name": "CLIP", + "type": "CLIP", + "links": [] + }, + { + "localized_name": "VAE", + "name": "VAE", + "type": "VAE", + "links": [ + 1697 + ] + } + ], + "properties": { + "Node name for S&R": "CheckpointLoaderSimple", + "cnr_id": "comfy-core", + "ver": "0.15.0", + "models": [ + { + "name": "sdpose_wholebody_fp16.safetensors", + "url": "https://huggingface.co/Comfy-Org/SDPose/resolve/main/checkpoints/sdpose_wholebody_fp16.safetensors", + "directory": "checkpoints" + } + ], + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + } + }, + "widgets_values": [ + "sdpose_wholebody_fp16.safetensors" + ] + } + ], + "groups": [], + "links": [ + { + "id": 1696, + "origin_id": 673, + "origin_slot": 0, + "target_id": 671, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 1697, + "origin_id": 673, + "origin_slot": 2, + "target_id": 671, + "target_slot": 1, + "type": "VAE" + }, + { + "id": 1698, + "origin_id": 674, + "origin_slot": 0, + "target_id": 671, + "target_slot": 2, + "type": "IMAGE" + }, + { + "id": 1699, + "origin_id": 671, + "origin_slot": 0, + "target_id": 672, + "target_slot": 0, + "type": "POSE_KEYPOINT" + }, + { + "id": 1700, + "origin_id": -10, + "origin_slot": 0, + "target_id": 674, + "target_slot": 0, + "type": "IMAGE,MASK" + }, + { + "id": 1701, + "origin_id": 672, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 1704, + "origin_id": -10, + "origin_slot": 1, + "target_id": 674, + "target_slot": 2, + "type": "INT" + }, + { + "id": 1705, + "origin_id": -10, + "origin_slot": 2, + "target_id": 674, + "target_slot": 3, + "type": "COMBO" + }, + { + "id": 1706, + "origin_id": -10, + "origin_slot": 3, + "target_id": 672, + "target_slot": 1, + "type": "BOOLEAN" + }, + { + "id": 1707, + "origin_id": -10, + "origin_slot": 4, + "target_id": 672, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 1708, + "origin_id": -10, + "origin_slot": 5, + "target_id": 672, + "target_slot": 3, + "type": "BOOLEAN" + }, + { + "id": 1709, + "origin_id": -10, + "origin_slot": 6, + "target_id": 672, + "target_slot": 4, + "type": "BOOLEAN" + }, + { + "id": 1710, + "origin_id": -10, + "origin_slot": 7, + "target_id": 672, + "target_slot": 5, + "type": "INT" + }, + { + "id": 1711, + "origin_id": -10, + "origin_slot": 8, + "target_id": 672, + "target_slot": 6, + "type": "INT" + }, + { + "id": 1712, + "origin_id": -10, + "origin_slot": 9, + "target_id": 672, + "target_slot": 7, + "type": "FLOAT" + }, + { + "id": 1713, + "origin_id": -10, + "origin_slot": 10, + "target_id": 673, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 1714, + "origin_id": -10, + "origin_slot": 11, + "target_id": 671, + "target_slot": 3, + "type": "BOUNDING_BOX" + }, + { + "id": 1715, + "origin_id": 671, + "origin_slot": 0, + "target_id": -20, + "target_slot": 1, + "type": "POSE_KEYPOINT" + } + ], + "extra": { + "workflowRendererVersion": "LG" + }, + "category": "Conditioning & Preprocessors/Pose", + "description": "Extracts human pose keypoints and stick-figure visuals from an image using SDPose-OOD, with optional bounding-box input per subject." + } + ] + }, + "extra": { + "ue_links": [] + } +} \ No newline at end of file diff --git a/blueprints/Merge Videos.json b/blueprints/Merge Videos.json new file mode 100644 index 000000000..689e6ec16 --- /dev/null +++ b/blueprints/Merge Videos.json @@ -0,0 +1,1219 @@ +{ + "revision": 0, + "last_node_id": 26, + "last_link_id": 0, + "nodes": [ + { + "id": 26, + "type": "32e6dbcc-e2d7-45c0-a245-fc74b8271dfb", + "pos": [ + -980, + 480 + ], + "size": [ + 290, + 190 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "label": "base_video", + "localized_name": "clip_to_resize", + "name": "clip_to_resize", + "type": "VIDEO", + "link": null + }, + { + "label": "second_video", + "localized_name": "base_video", + "name": "base_video", + "type": "VIDEO", + "link": null + }, + { + "label": "pad_second_video", + "localized_name": "pad_second_video", + "name": "pad_second_video", + "type": "BOOLEAN", + "widget": { + "name": "pad_second_video" + }, + "link": null + }, + { + "name": "interpolation", + "type": "COMBO", + "widget": { + "name": "interpolation" + }, + "link": null + }, + { + "name": "padding_color", + "type": "COMBO", + "widget": { + "name": "padding_color" + }, + "link": null + }, + { + "label": "drop_audio", + "localized_name": "drop_audio", + "name": "drop_audio", + "type": "BOOLEAN", + "widget": { + "name": "drop_audio" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "merged_video", + "name": "merged_video", + "type": "VIDEO", + "links": [] + } + ], + "properties": { + "proxyWidgets": [ + [ + "28", + "value" + ], + [ + "6", + "interpolation" + ], + [ + "6", + "padding_color" + ], + [ + "11", + "value" + ] + ], + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [], + "title": "Merge Videos" + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "32e6dbcc-e2d7-45c0-a245-fc74b8271dfb", + "version": 1, + "state": { + "lastGroupId": 2, + "lastNodeId": 34, + "lastLinkId": 75, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Merge Videos", + "inputNode": { + "id": -10, + "bounding": [ + -1990, + 700, + 152.5546875, + 168 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + 1210, + 614, + 128, + 68 + ] + }, + "inputs": [ + { + "id": "2fb09e41-c5fa-4654-b9d2-569b59626ec4", + "name": "clip_to_resize", + "type": "VIDEO", + "linkIds": [ + 50 + ], + "localized_name": "clip_to_resize", + "label": "base_video", + "pos": [ + -1861.4453125, + 724 + ] + }, + { + "id": "017f8d09-7900-4dc9-b95c-0cab31bcde7d", + "name": "base_video", + "type": "VIDEO", + "linkIds": [ + 51 + ], + "localized_name": "base_video", + "label": "second_video", + "pos": [ + -1861.4453125, + 744 + ] + }, + { + "id": "a39894ce-1785-4037-b39c-b40d2e470c43", + "name": "pad_second_video", + "type": "BOOLEAN", + "linkIds": [ + 59 + ], + "localized_name": "pad_second_video", + "label": "pad_second_video", + "pos": [ + -1861.4453125, + 764 + ] + }, + { + "id": "b4fb86cb-8d87-4193-8533-88a57df50e18", + "name": "interpolation", + "type": "COMBO", + "linkIds": [ + 60 + ], + "pos": [ + -1861.4453125, + 784 + ] + }, + { + "id": "2413a2e2-cfdc-4d1d-9e2e-81e7acdf35e3", + "name": "padding_color", + "type": "COMBO", + "linkIds": [ + 62 + ], + "pos": [ + -1861.4453125, + 804 + ] + }, + { + "id": "338b1e09-0efb-424f-949b-e730a0aa8527", + "name": "drop_audio", + "type": "BOOLEAN", + "linkIds": [ + 63 + ], + "localized_name": "drop_audio", + "label": "drop_audio", + "pos": [ + -1861.4453125, + 824 + ] + } + ], + "outputs": [ + { + "id": "be99efc6-7fb3-4059-93d0-136dc8cc8faf", + "name": "merged_video", + "type": "VIDEO", + "linkIds": [ + 16 + ], + "localized_name": "merged_video", + "pos": [ + 1234, + 638 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 11, + "type": "PrimitiveBoolean", + "pos": [ + -990, + 1230 + ], + "size": [ + 270, + 80 + ], + "flags": {}, + "order": 8, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "BOOLEAN", + "widget": { + "name": "value" + }, + "link": 63 + } + ], + "outputs": [ + { + "localized_name": "BOOLEAN", + "name": "BOOLEAN", + "type": "BOOLEAN", + "links": [ + 14 + ] + } + ], + "properties": { + "Node name for S&R": "PrimitiveBoolean", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + false + ] + }, + { + "id": 10, + "type": "EmptyAudio", + "pos": [ + -990, + 1060 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "duration", + "name": "duration", + "type": "FLOAT", + "widget": { + "name": "duration" + }, + "link": null + }, + { + "localized_name": "sample_rate", + "name": "sample_rate", + "type": "INT", + "widget": { + "name": "sample_rate" + }, + "link": null + }, + { + "localized_name": "channels", + "name": "channels", + "type": "INT", + "widget": { + "name": "channels" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "AUDIO", + "name": "AUDIO", + "type": "AUDIO", + "links": [ + 22 + ] + } + ], + "properties": { + "Node name for S&R": "EmptyAudio", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 60, + 44100, + 2 + ] + }, + { + "id": 3, + "type": "ComfySwitchNode", + "pos": [ + -370, + 1010 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 21 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 22 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 14 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 12 + ] + } + ], + "properties": { + "Node name for S&R": "ComfySwitchNode", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + false + ] + }, + { + "id": 6, + "type": "ResizeAndPadImage", + "pos": [ + -400, + 440 + ], + "size": [ + 270, + 210 + ], + "flags": {}, + "order": 4, + "mode": 0, + "showAdvanced": true, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 39 + }, + { + "localized_name": "target_width", + "name": "target_width", + "type": "INT", + "widget": { + "name": "target_width" + }, + "link": 4 + }, + { + "localized_name": "target_height", + "name": "target_height", + "type": "INT", + "widget": { + "name": "target_height" + }, + "link": 5 + }, + { + "localized_name": "padding_color", + "name": "padding_color", + "type": "COMBO", + "widget": { + "name": "padding_color" + }, + "link": 62 + }, + { + "localized_name": "interpolation", + "name": "interpolation", + "type": "COMBO", + "widget": { + "name": "interpolation" + }, + "link": 60 + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 75 + ] + } + ], + "properties": { + "Node name for S&R": "ResizeAndPadImage", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 512, + 512, + "white", + "lanczos" + ] + }, + { + "id": 8, + "type": "CreateVideo", + "pos": [ + 880, + 280 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 6, + "mode": 0, + "inputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "link": 19 + }, + { + "localized_name": "audio", + "name": "audio", + "shape": 7, + "type": "AUDIO", + "link": 12 + }, + { + "localized_name": "fps", + "name": "fps", + "type": "FLOAT", + "widget": { + "name": "fps" + }, + "link": 15 + } + ], + "outputs": [ + { + "localized_name": "VIDEO", + "name": "VIDEO", + "type": "VIDEO", + "links": [ + 16 + ] + } + ], + "properties": { + "Node name for S&R": "CreateVideo", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 30 + ] + }, + { + "id": 9, + "type": "AudioMerge", + "pos": [ + -990, + 890 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 7, + "mode": 0, + "inputs": [ + { + "localized_name": "audio1", + "name": "audio1", + "type": "AUDIO", + "link": 9 + }, + { + "localized_name": "audio2", + "name": "audio2", + "type": "AUDIO", + "link": 10 + }, + { + "localized_name": "merge_method", + "name": "merge_method", + "type": "COMBO", + "widget": { + "name": "merge_method" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "AUDIO", + "name": "AUDIO", + "type": "AUDIO", + "links": [ + 21 + ] + } + ], + "properties": { + "Node name for S&R": "AudioMerge", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "add" + ] + }, + { + "id": 2, + "type": "GetVideoComponents", + "pos": [ + -1590, + 460 + ], + "size": [ + 230, + 100 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "video", + "name": "video", + "type": "VIDEO", + "link": 51 + } + ], + "outputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "links": [ + 39, + 54 + ] + }, + { + "localized_name": "audio", + "name": "audio", + "type": "AUDIO", + "links": [ + 9 + ] + }, + { + "localized_name": "fps", + "name": "fps", + "type": "FLOAT", + "links": null + } + ], + "properties": { + "Node name for S&R": "GetVideoComponents", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 27, + "type": "ComfySwitchNode", + "pos": [ + 60, + 70 + ], + "size": [ + 280, + 130 + ], + "flags": {}, + "order": 10, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 54 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 75 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 56 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 55 + ] + } + ], + "properties": { + "Node name for S&R": "ComfySwitchNode", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + false + ] + }, + { + "id": 1, + "type": "GetVideoComponents", + "pos": [ + -1600, + 30 + ], + "size": [ + 230, + 100 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "video", + "name": "video", + "type": "VIDEO", + "link": 50 + } + ], + "outputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "links": [ + 3, + 17 + ] + }, + { + "localized_name": "audio", + "name": "audio", + "type": "AUDIO", + "links": [ + 10 + ] + }, + { + "localized_name": "fps", + "name": "fps", + "type": "FLOAT", + "links": [ + 15 + ] + } + ], + "properties": { + "Node name for S&R": "GetVideoComponents", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 7, + "type": "GetImageSize", + "pos": [ + -1000, + 480 + ], + "size": [ + 260, + 110 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 3 + } + ], + "outputs": [ + { + "localized_name": "width", + "name": "width", + "type": "INT", + "links": [ + 4 + ] + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "links": [ + 5 + ] + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "links": null + } + ], + "properties": { + "Node name for S&R": "GetImageSize", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 28, + "type": "PrimitiveBoolean", + "pos": [ + -1590, + 190 + ], + "size": [ + 270, + 80 + ], + "flags": {}, + "order": 11, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "BOOLEAN", + "widget": { + "name": "value" + }, + "link": 59 + } + ], + "outputs": [ + { + "localized_name": "BOOLEAN", + "name": "BOOLEAN", + "type": "BOOLEAN", + "links": [ + 56 + ] + } + ], + "properties": { + "Node name for S&R": "PrimitiveBoolean", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + false + ] + }, + { + "id": 13, + "type": "BatchImagesNode", + "pos": [ + 530, + 10 + ], + "size": [ + 230, + 120 + ], + "flags": {}, + "order": 9, + "mode": 0, + "inputs": [ + { + "label": "image0", + "localized_name": "images.image0", + "name": "images.image0", + "type": "IMAGE", + "link": 17 + }, + { + "label": "image1", + "localized_name": "images.image1", + "name": "images.image1", + "shape": 7, + "type": "IMAGE", + "link": 55 + }, + { + "label": "image2", + "localized_name": "images.image2", + "name": "images.image2", + "shape": 7, + "type": "IMAGE", + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 19 + ] + } + ], + "properties": { + "Node name for S&R": "BatchImagesNode", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + } + ], + "groups": [ + { + "id": 1, + "title": "Audio", + "bounding": [ + -1000, + 820, + 915, + 496 + ], + "color": "#3f789e", + "flags": {} + } + ], + "links": [ + { + "id": 21, + "origin_id": 9, + "origin_slot": 0, + "target_id": 3, + "target_slot": 0, + "type": "AUDIO" + }, + { + "id": 22, + "origin_id": 10, + "origin_slot": 0, + "target_id": 3, + "target_slot": 1, + "type": "AUDIO" + }, + { + "id": 14, + "origin_id": 11, + "origin_slot": 0, + "target_id": 3, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 9, + "origin_id": 2, + "origin_slot": 1, + "target_id": 9, + "target_slot": 0, + "type": "AUDIO" + }, + { + "id": 10, + "origin_id": 1, + "origin_slot": 1, + "target_id": 9, + "target_slot": 1, + "type": "AUDIO" + }, + { + "id": 39, + "origin_id": 2, + "origin_slot": 0, + "target_id": 6, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 4, + "origin_id": 7, + "origin_slot": 0, + "target_id": 6, + "target_slot": 1, + "type": "INT" + }, + { + "id": 5, + "origin_id": 7, + "origin_slot": 1, + "target_id": 6, + "target_slot": 2, + "type": "INT" + }, + { + "id": 3, + "origin_id": 1, + "origin_slot": 0, + "target_id": 7, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 17, + "origin_id": 1, + "origin_slot": 0, + "target_id": 13, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 19, + "origin_id": 13, + "origin_slot": 0, + "target_id": 8, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 12, + "origin_id": 3, + "origin_slot": 0, + "target_id": 8, + "target_slot": 1, + "type": "AUDIO" + }, + { + "id": 15, + "origin_id": 1, + "origin_slot": 2, + "target_id": 8, + "target_slot": 2, + "type": "FLOAT" + }, + { + "id": 16, + "origin_id": 8, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "VIDEO" + }, + { + "id": 50, + "origin_id": -10, + "origin_slot": 0, + "target_id": 1, + "target_slot": 0, + "type": "VIDEO" + }, + { + "id": 51, + "origin_id": -10, + "origin_slot": 1, + "target_id": 2, + "target_slot": 0, + "type": "VIDEO" + }, + { + "id": 54, + "origin_id": 2, + "origin_slot": 0, + "target_id": 27, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 55, + "origin_id": 27, + "origin_slot": 0, + "target_id": 13, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 56, + "origin_id": 28, + "origin_slot": 0, + "target_id": 27, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 59, + "origin_id": -10, + "origin_slot": 2, + "target_id": 28, + "target_slot": 0, + "type": "BOOLEAN" + }, + { + "id": 60, + "origin_id": -10, + "origin_slot": 3, + "target_id": 6, + "target_slot": 4, + "type": "COMBO" + }, + { + "id": 62, + "origin_id": -10, + "origin_slot": 4, + "target_id": 6, + "target_slot": 3, + "type": "COMBO" + }, + { + "id": 63, + "origin_id": -10, + "origin_slot": 5, + "target_id": 11, + "target_slot": 0, + "type": "BOOLEAN" + }, + { + "id": 75, + "origin_id": 6, + "origin_slot": 0, + "target_id": 27, + "target_slot": 1, + "type": "IMAGE" + } + ], + "extra": {}, + "category": "Video Tools", + "description": "Concatenates two videos end-to-end with optional resize, letterbox padding, and audio merge or drop." + } + ] + }, + "extra": {} +} \ No newline at end of file diff --git a/blueprints/Pose to Image (Z-Image-Turbo).json b/blueprints/Pose to Image (Z-Image-Turbo).json index 5c2749efe..92ee80907 100644 --- a/blueprints/Pose to Image (Z-Image-Turbo).json +++ b/blueprints/Pose to Image (Z-Image-Turbo).json @@ -1298,7 +1298,7 @@ "VHS_MetadataImage": true, "VHS_KeepIntermediate": true }, - "category": "Image generation and editing/Pose to image", + "category": "Image generation and editing/Conditioned", "description": "Generates an image from pose keypoints using Z-Image-Turbo with text conditioning." } ] diff --git a/blueprints/Pose to Video (LTX 2.0).json b/blueprints/Pose to Video (LTX 2.0).json index 1ce49351a..04eb69972 100644 --- a/blueprints/Pose to Video (LTX 2.0).json +++ b/blueprints/Pose to Video (LTX 2.0).json @@ -3870,7 +3870,7 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Video generation and editing/Pose to video", + "category": "Video generation and editing/Conditioned", "description": "Generates video from pose reference frames using LTX-2, with optional synchronized audio." } ] diff --git a/blueprints/Prompt Enhance.json b/blueprints/Prompt Enhance.json index e260b1203..e3a77a73b 100644 --- a/blueprints/Prompt Enhance.json +++ b/blueprints/Prompt Enhance.json @@ -270,7 +270,7 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Text generation/Prompt enhance", + "category": "Text Tools", "description": "Expands short text prompts into detailed descriptions using a text generation model for better generation quality." } ] diff --git a/blueprints/Remove Background (BiRefNet).json b/blueprints/Remove Background (BiRefNet).json index 732a4adc4..9ec441e51 100644 --- a/blueprints/Remove Background (BiRefNet).json +++ b/blueprints/Remove Background (BiRefNet).json @@ -389,7 +389,7 @@ } ], "extra": {}, - "category": "Image generation and editing/Background Removal" + "category": "Image Tools/Background Removal" } ] }, diff --git a/blueprints/Select Per-Line Text by Index.json b/blueprints/Select Per-Line Text by Index.json new file mode 100644 index 000000000..8a4020d50 --- /dev/null +++ b/blueprints/Select Per-Line Text by Index.json @@ -0,0 +1,485 @@ +{ + "revision": 0, + "last_node_id": 10, + "last_link_id": 0, + "nodes": [ + { + "id": 10, + "type": "3fb7557a-470d-4983-9d8c-6d5caa9788f0", + "pos": [ + -250, + 8590 + ], + "size": [ + 280, + 360 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "text_per_line", + "name": "text_per_line", + "type": "STRING", + "widget": { + "name": "text_per_line" + }, + "link": null + }, + { + "localized_name": "index", + "name": "index", + "type": "INT", + "widget": { + "name": "index" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "selected_line", + "name": "selected_line", + "type": "STRING", + "links": [] + } + ], + "properties": { + "proxyWidgets": [ + [ + "2", + "string" + ], + [ + "3", + "value" + ] + ], + "cnr_id": "comfy-core", + "ver": "0.19.0", + "ue_properties": { + "widget_ue_connectable": {}, + "input_ue_unconnectable": {} + } + }, + "widgets_values": [], + "title": "Select Per-Line Text by Index" + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "3fb7557a-470d-4983-9d8c-6d5caa9788f0", + "version": 1, + "state": { + "lastGroupId": 0, + "lastNodeId": 10, + "lastLinkId": 14, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Select Per-Line Text by Index", + "inputNode": { + "id": -10, + "bounding": [ + -990, + 8595, + 128, + 88 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + 710, + 8585, + 128, + 68 + ] + }, + "inputs": [ + { + "id": "75417d82-a934-4ac9-b667-d8dcd5a3bfb3", + "name": "text_per_line", + "type": "STRING", + "linkIds": [ + 13 + ], + "localized_name": "text_per_line", + "pos": [ + -886, + 8619 + ] + }, + { + "id": "46e69a73-1804-4ca6-9175-31445bf0be96", + "name": "index", + "type": "INT", + "linkIds": [ + 14 + ], + "localized_name": "index", + "pos": [ + -886, + 8639 + ] + } + ], + "outputs": [ + { + "id": "e34e8ad1-84d2-4bd2-a460-eb7de6067c10", + "name": "selected_line", + "type": "STRING", + "linkIds": [ + 10 + ], + "localized_name": "selected_line", + "pos": [ + 734, + 8609 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 1, + "type": "PreviewAny", + "pos": [ + -500, + 8400 + ], + "size": [ + 230, + 180 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "source", + "name": "source", + "type": "*", + "link": 1 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 6 + ] + } + ], + "properties": { + "Node name for S&R": "PreviewAny", + "cnr_id": "comfy-core", + "ver": "0.19.0", + "ue_properties": { + "widget_ue_connectable": {}, + "input_ue_unconnectable": {} + } + }, + "widgets_values": [ + null, + null, + null + ] + }, + { + "id": 2, + "type": "RegexExtract", + "pos": [ + -240, + 8740 + ], + "size": [ + 470, + 460 + ], + "flags": {}, + "order": 1, + "mode": 0, + "showAdvanced": false, + "inputs": [ + { + "localized_name": "string", + "name": "string", + "type": "STRING", + "widget": { + "name": "string" + }, + "link": 13 + }, + { + "localized_name": "regex_pattern", + "name": "regex_pattern", + "type": "STRING", + "widget": { + "name": "regex_pattern" + }, + "link": 9 + }, + { + "localized_name": "mode", + "name": "mode", + "type": "COMBO", + "widget": { + "name": "mode" + }, + "link": null + }, + { + "localized_name": "case_insensitive", + "name": "case_insensitive", + "type": "BOOLEAN", + "widget": { + "name": "case_insensitive" + }, + "link": null + }, + { + "localized_name": "multiline", + "name": "multiline", + "type": "BOOLEAN", + "widget": { + "name": "multiline" + }, + "link": null + }, + { + "localized_name": "dotall", + "name": "dotall", + "type": "BOOLEAN", + "widget": { + "name": "dotall" + }, + "link": null + }, + { + "localized_name": "group_index", + "name": "group_index", + "type": "INT", + "widget": { + "name": "group_index" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 10 + ] + } + ], + "properties": { + "Node name for S&R": "RegexExtract", + "cnr_id": "comfy-core", + "ver": "0.19.0", + "ue_properties": { + "widget_ue_connectable": {}, + "input_ue_unconnectable": {} + } + }, + "widgets_values": [ + "", + "", + "First Group", + false, + false, + false, + 1 + ] + }, + { + "id": 3, + "type": "PrimitiveInt", + "pos": [ + -810, + 8400 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "INT", + "widget": { + "name": "value" + }, + "link": 14 + } + ], + "outputs": [ + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 1 + ] + } + ], + "title": "Int (line index)", + "properties": { + "Node name for S&R": "Int (line index)", + "cnr_id": "comfy-core", + "ver": "0.19.0", + "ue_properties": { + "widget_ue_connectable": {}, + "input_ue_unconnectable": {} + } + }, + "widgets_values": [ + 0, + "fixed" + ] + }, + { + "id": 8, + "type": "StringReplace", + "pos": [ + -240, + 8400 + ], + "size": [ + 400, + 280 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "string", + "name": "string", + "type": "STRING", + "widget": { + "name": "string" + }, + "link": null + }, + { + "localized_name": "find", + "name": "find", + "type": "STRING", + "widget": { + "name": "find" + }, + "link": null + }, + { + "localized_name": "replace", + "name": "replace", + "type": "STRING", + "widget": { + "name": "replace" + }, + "link": 6 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 9 + ] + } + ], + "properties": { + "Node name for S&R": "StringReplace", + "cnr_id": "comfy-core", + "ver": "0.19.0", + "ue_properties": { + "widget_ue_connectable": {}, + "input_ue_unconnectable": {} + } + }, + "widgets_values": [ + "^(?:[^\\n]*\\n){index}([^\\n]*)(?:\\n|$)", + "index", + "" + ] + } + ], + "groups": [], + "links": [ + { + "id": 1, + "origin_id": 3, + "origin_slot": 0, + "target_id": 1, + "target_slot": 0, + "type": "INT" + }, + { + "id": 9, + "origin_id": 8, + "origin_slot": 0, + "target_id": 2, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 6, + "origin_id": 1, + "origin_slot": 0, + "target_id": 8, + "target_slot": 2, + "type": "STRING" + }, + { + "id": 10, + "origin_id": 2, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "STRING" + }, + { + "id": 13, + "origin_id": -10, + "origin_slot": 0, + "target_id": 2, + "target_slot": 0, + "type": "STRING" + }, + { + "id": 14, + "origin_id": -10, + "origin_slot": 1, + "target_id": 3, + "target_slot": 0, + "type": "INT" + } + ], + "extra": {}, + "category": "Text Tools", + "description": "Selects one line from multiline text by zero-based index for batch or list-driven prompt workflows." + } + ] + }, + "extra": { + "ue_links": [], + "links_added_by_ue": [] + } +} \ No newline at end of file diff --git a/blueprints/Split Image Grid to Tiles.json b/blueprints/Split Image Grid to Tiles.json new file mode 100644 index 000000000..d1f6e40ef --- /dev/null +++ b/blueprints/Split Image Grid to Tiles.json @@ -0,0 +1,714 @@ +{ + "revision": 0, + "last_node_id": 251, + "last_link_id": 0, + "nodes": [ + { + "id": 251, + "type": "609e1fd1-b731-4b78-89ac-d19b1156b025", + "pos": [ + -1490, + 130 + ], + "size": [ + 230, + 164 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "source_image", + "name": "source_image", + "type": "IMAGE", + "link": null + }, + { + "localized_name": "columns", + "name": "columns", + "type": "INT", + "widget": { + "name": "columns" + }, + "link": null + }, + { + "localized_name": "rows", + "name": "rows", + "type": "INT", + "widget": { + "name": "rows" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "tiles", + "name": "tiles", + "type": "IMAGE", + "links": [] + } + ], + "properties": { + "proxyWidgets": [ + [ + "228", + "value" + ], + [ + "252", + "value" + ] + ], + "cnr_id": "comfy-core", + "ver": "0.20.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [], + "title": "Split Image Grid to Tiles" + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "609e1fd1-b731-4b78-89ac-d19b1156b025", + "version": 1, + "state": { + "lastGroupId": 9, + "lastNodeId": 252, + "lastLinkId": 429, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Split Image Grid to Tiles", + "inputNode": { + "id": -10, + "bounding": [ + -1690, + 260, + 128, + 108 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + -510, + 590, + 128, + 68 + ] + }, + "inputs": [ + { + "id": "866ac798-cfbc-450a-b755-e704f86404d9", + "name": "source_image", + "type": "IMAGE", + "linkIds": [ + 386, + 389 + ], + "localized_name": "source_image", + "pos": [ + -1586, + 284 + ] + }, + { + "id": "bc37b1f8-8ab2-4f19-bd00-75d4fbc4feb3", + "name": "columns", + "type": "INT", + "linkIds": [ + 427 + ], + "localized_name": "columns", + "pos": [ + -1586, + 304 + ] + }, + { + "id": "d45915da-e848-43dd-9ccc-e3161e9c99d9", + "name": "rows", + "type": "INT", + "linkIds": [ + 428 + ], + "localized_name": "rows", + "pos": [ + -1586, + 324 + ] + } + ], + "outputs": [ + { + "id": "18bc780f-064b-4038-87c6-67dba71deb08", + "name": "tiles", + "type": "IMAGE", + "linkIds": [ + 394 + ], + "localized_name": "tiles", + "shape": 6, + "pos": [ + -486, + 614 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 225, + "type": "SplitImageToTileList", + "pos": [ + -1010, + 620 + ], + "size": [ + 290, + 170 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 386 + }, + { + "localized_name": "tile_width", + "name": "tile_width", + "type": "INT", + "widget": { + "name": "tile_width" + }, + "link": 403 + }, + { + "localized_name": "tile_height", + "name": "tile_height", + "type": "INT", + "widget": { + "name": "tile_height" + }, + "link": 404 + }, + { + "localized_name": "overlap", + "name": "overlap", + "type": "INT", + "widget": { + "name": "overlap" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "shape": 6, + "type": "IMAGE", + "links": [ + 394 + ] + } + ], + "properties": { + "Node name for S&R": "SplitImageToTileList", + "cnr_id": "comfy-core", + "ver": "0.20.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 1024, + 1024, + 0 + ] + }, + { + "id": 231, + "type": "ComfyMathExpression", + "pos": [ + -1080, + 330 + ], + "size": [ + 370, + 190 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "label": "a", + "localized_name": "values.a", + "name": "values.a", + "type": "FLOAT,INT,BOOLEAN", + "link": 390 + }, + { + "label": "b", + "localized_name": "values.b", + "name": "values.b", + "shape": 7, + "type": "FLOAT,INT,BOOLEAN", + "link": 429 + }, + { + "label": "c", + "localized_name": "values.c", + "name": "values.c", + "shape": 7, + "type": "FLOAT,INT,BOOLEAN", + "link": null + }, + { + "localized_name": "expression", + "name": "expression", + "type": "STRING", + "widget": { + "name": "expression" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "FLOAT", + "name": "FLOAT", + "type": "FLOAT", + "links": null + }, + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 404 + ] + }, + { + "localized_name": "BOOL", + "name": "BOOL", + "type": "BOOLEAN", + "links": null + } + ], + "title": "Math Expression （Height）", + "properties": { + "Node name for S&R": "ComfyMathExpression", + "cnr_id": "comfy-core", + "ver": "0.18.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + } + }, + "widgets_values": [ + "max(1, (int(a) + int(b) - 1) // int(b))" + ] + }, + { + "id": 229, + "type": "ComfyMathExpression", + "pos": [ + -1090, + -30 + ], + "size": [ + 370, + 190 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "label": "a", + "localized_name": "values.a", + "name": "values.a", + "type": "FLOAT,INT,BOOLEAN", + "link": 387 + }, + { + "label": "b", + "localized_name": "values.b", + "name": "values.b", + "shape": 7, + "type": "FLOAT,INT,BOOLEAN", + "link": 388 + }, + { + "label": "c", + "localized_name": "values.c", + "name": "values.c", + "shape": 7, + "type": "FLOAT,INT,BOOLEAN", + "link": null + }, + { + "localized_name": "expression", + "name": "expression", + "type": "STRING", + "widget": { + "name": "expression" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "FLOAT", + "name": "FLOAT", + "type": "FLOAT", + "links": null + }, + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 403 + ] + }, + { + "localized_name": "BOOL", + "name": "BOOL", + "type": "BOOLEAN", + "links": null + } + ], + "title": "Math Expression （Width）", + "properties": { + "Node name for S&R": "ComfyMathExpression", + "cnr_id": "comfy-core", + "ver": "0.18.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + } + }, + "widgets_values": [ + "max(1, (int(a) + int(b) - 1) // int(b))" + ] + }, + { + "id": 228, + "type": "PrimitiveInt", + "pos": [ + -1380, + 90 + ], + "size": [ + 230, + 110 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "INT", + "widget": { + "name": "value" + }, + "link": 427 + } + ], + "outputs": [ + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 388 + ] + } + ], + "title": "Int (grid columns)", + "properties": { + "Node name for S&R": "Int (grid columns)", + "cnr_id": "comfy-core", + "ver": "0.18.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + } + }, + "widgets_values": [ + 2, + "fixed" + ] + }, + { + "id": 230, + "type": "GetImageSize", + "pos": [ + -1380, + 290 + ], + "size": [ + 230, + 100 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 389 + } + ], + "outputs": [ + { + "localized_name": "width", + "name": "width", + "type": "INT", + "links": [ + 387 + ] + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "links": [ + 390 + ] + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "links": null + } + ], + "properties": { + "Node name for S&R": "GetImageSize", + "cnr_id": "comfy-core", + "ver": "0.18.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + } + } + }, + { + "id": 252, + "type": "PrimitiveInt", + "pos": [ + -1380, + 470 + ], + "size": [ + 230, + 110 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "INT", + "widget": { + "name": "value" + }, + "link": 428 + } + ], + "outputs": [ + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 429 + ] + } + ], + "title": "Int (grid rows)", + "properties": { + "Node name for S&R": "Int (grid rows)", + "cnr_id": "comfy-core", + "ver": "0.18.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + } + }, + "widgets_values": [ + 3, + "fixed" + ] + } + ], + "groups": [], + "links": [ + { + "id": 403, + "origin_id": 229, + "origin_slot": 1, + "target_id": 225, + "target_slot": 1, + "type": "INT" + }, + { + "id": 404, + "origin_id": 231, + "origin_slot": 1, + "target_id": 225, + "target_slot": 2, + "type": "INT" + }, + { + "id": 390, + "origin_id": 230, + "origin_slot": 1, + "target_id": 231, + "target_slot": 0, + "type": "INT" + }, + { + "id": 387, + "origin_id": 230, + "origin_slot": 0, + "target_id": 229, + "target_slot": 0, + "type": "INT" + }, + { + "id": 388, + "origin_id": 228, + "origin_slot": 0, + "target_id": 229, + "target_slot": 1, + "type": "INT" + }, + { + "id": 386, + "origin_id": -10, + "origin_slot": 0, + "target_id": 225, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 389, + "origin_id": -10, + "origin_slot": 0, + "target_id": 230, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 394, + "origin_id": 225, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 427, + "origin_id": -10, + "origin_slot": 1, + "target_id": 228, + "target_slot": 0, + "type": "INT" + }, + { + "id": 428, + "origin_id": -10, + "origin_slot": 2, + "target_id": 252, + "target_slot": 0, + "type": "INT" + }, + { + "id": 429, + "origin_id": 252, + "origin_slot": 0, + "target_id": 231, + "target_slot": 1, + "type": "INT" + } + ], + "extra": {}, + "category": "Image Tools/Crop", + "description": "Splits an image into a configurable columns×rows grid of equal tiles for tiled generation or processing." + } + ] + }, + "extra": {} +} \ No newline at end of file diff --git a/blueprints/Text to Image (Anima).json b/blueprints/Text to Image (Anima).json new file mode 100644 index 000000000..787908ca9 --- /dev/null +++ b/blueprints/Text to Image (Anima).json @@ -0,0 +1,1085 @@ +{ + "revision": 0, + "last_node_id": 60, + "last_link_id": 0, + "nodes": [ + { + "id": 60, + "type": "a3c0dab6-b250-4585-a0f9-8fb8b074fb2f", + "pos": [ + -10, + 130 + ], + "size": [ + 500, + 640 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "label": "prompt", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": null + }, + { + "name": "width", + "type": "INT", + "widget": { + "name": "width" + }, + "link": null + }, + { + "name": "height", + "type": "INT", + "widget": { + "name": "height" + }, + "link": null + }, + { + "name": "steps", + "type": "INT", + "widget": { + "name": "steps" + }, + "link": null + }, + { + "name": "cfg", + "type": "FLOAT", + "widget": { + "name": "cfg" + }, + "link": null + }, + { + "name": "seed", + "type": "INT", + "widget": { + "name": "seed" + }, + "link": null + }, + { + "name": "unet_name", + "type": "COMBO", + "widget": { + "name": "unet_name" + }, + "link": null + }, + { + "name": "clip_name", + "type": "COMBO", + "widget": { + "name": "clip_name" + }, + "link": null + }, + { + "name": "vae_name", + "type": "COMBO", + "widget": { + "name": "vae_name" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [] + } + ], + "properties": { + "proxyWidgets": [ + [ + "11", + "text" + ], + [ + "28", + "width" + ], + [ + "28", + "height" + ], + [ + "19", + "steps" + ], + [ + "19", + "cfg" + ], + [ + "19", + "seed" + ], + [ + "44", + "unet_name" + ], + [ + "45", + "clip_name" + ], + [ + "15", + "vae_name" + ] + ], + "cnr_id": "comfy-core", + "ver": "0.18.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [], + "title": "Text to Image (Anima)" + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "a3c0dab6-b250-4585-a0f9-8fb8b074fb2f", + "version": 1, + "state": { + "lastGroupId": 3, + "lastNodeId": 70, + "lastLinkId": 104, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Text to Image (Anima)", + "inputNode": { + "id": -10, + "bounding": [ + -330, + 530, + 120, + 220 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + 1229.9999873482075, + 505, + 120, + 60 + ] + }, + "inputs": [ + { + "id": "4693f350-6ba0-446d-80d4-3038c661d26c", + "name": "text", + "type": "STRING", + "linkIds": [ + 95 + ], + "label": "prompt", + "pos": [ + -230, + 550 + ] + }, + { + "id": "4a7886a9-4ed7-49bb-afc2-977bb78a303d", + "name": "width", + "type": "INT", + "linkIds": [ + 96 + ], + "pos": [ + -230, + 570 + ] + }, + { + "id": "f6c04461-d29e-49e3-8790-07bb662bbbfe", + "name": "height", + "type": "INT", + "linkIds": [ + 97 + ], + "pos": [ + -230, + 590 + ] + }, + { + "id": "7a24f998-3808-4837-8bff-52304ad09fb6", + "name": "steps", + "type": "INT", + "linkIds": [ + 98 + ], + "pos": [ + -230, + 610 + ] + }, + { + "id": "aaa99698-b222-40fe-b946-28067528a85c", + "name": "cfg", + "type": "FLOAT", + "linkIds": [ + 99 + ], + "pos": [ + -230, + 630 + ] + }, + { + "id": "053df9ae-7311-4816-aa23-7fa13c656ced", + "name": "seed", + "type": "INT", + "linkIds": [ + 100 + ], + "pos": [ + -230, + 650 + ] + }, + { + "id": "c59194ea-015c-41a7-8edd-ae7ffc220b63", + "name": "unet_name", + "type": "COMBO", + "linkIds": [ + 101 + ], + "pos": [ + -230, + 670 + ] + }, + { + "id": "e655aa3b-2db7-4e25-9ea2-61550fa7ae2d", + "name": "clip_name", + "type": "COMBO", + "linkIds": [ + 102 + ], + "pos": [ + -230, + 690 + ] + }, + { + "id": "94965a7a-74dd-4f5a-87e3-9f87995d554f", + "name": "vae_name", + "type": "COMBO", + "linkIds": [ + 103 + ], + "pos": [ + -230, + 710 + ] + } + ], + "outputs": [ + { + "id": "ef85ac0a-2152-4232-bfa1-929cfc913718", + "name": "IMAGE", + "type": "IMAGE", + "linkIds": [ + 82 + ], + "localized_name": "IMAGE", + "pos": [ + 1249.9999873482075, + 525 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 45, + "type": "CLIPLoader", + "pos": [ + -60, + 380 + ], + "size": [ + 310, + 150 + ], + "flags": {}, + "order": 7, + "mode": 0, + "inputs": [ + { + "localized_name": "clip_name", + "name": "clip_name", + "type": "COMBO", + "widget": { + "name": "clip_name" + }, + "link": 102 + }, + { + "localized_name": "type", + "name": "type", + "type": "COMBO", + "widget": { + "name": "type" + }, + "link": null + }, + { + "localized_name": "device", + "name": "device", + "shape": 7, + "type": "COMBO", + "widget": { + "name": "device" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "CLIP", + "name": "CLIP", + "type": "CLIP", + "links": [ + 80, + 81 + ] + } + ], + "properties": { + "Node name for S&R": "CLIPLoader", + "cnr_id": "comfy-core", + "ver": "0.11.0", + "models": [ + { + "name": "qwen_3_06b_base.safetensors", + "url": "https://huggingface.co/circlestone-labs/Anima/resolve/main/split_files/text_encoders/qwen_3_06b_base.safetensors", + "directory": "text_encoders" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "qwen_3_06b_base.safetensors", + "stable_diffusion", + "default" + ] + }, + { + "id": 15, + "type": "VAELoader", + "pos": [ + -50, + 610 + ], + "size": [ + 310, + 100 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "vae_name", + "name": "vae_name", + "type": "COMBO", + "widget": { + "name": "vae_name" + }, + "link": 103 + } + ], + "outputs": [ + { + "localized_name": "VAE", + "name": "VAE", + "type": "VAE", + "links": [ + 11 + ] + } + ], + "properties": { + "Node name for S&R": "VAELoader", + "cnr_id": "comfy-core", + "ver": "0.3.40", + "models": [ + { + "name": "qwen_image_vae.safetensors", + "url": "https://huggingface.co/circlestone-labs/Anima/resolve/main/split_files/vae/qwen_image_vae.safetensors", + "directory": "vae" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "qwen_image_vae.safetensors" + ] + }, + { + "id": 8, + "type": "VAEDecode", + "pos": [ + 880, + 840 + ], + "size": [ + 230, + 90 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "samples", + "name": "samples", + "type": "LATENT", + "link": 10 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 11 + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "slot_index": 0, + "links": [ + 82 + ] + } + ], + "properties": { + "Node name for S&R": "VAEDecode", + "cnr_id": "comfy-core", + "ver": "0.3.40", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 28, + "type": "EmptyLatentImage", + "pos": [ + -50, + 830 + ], + "size": [ + 310, + 150 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "localized_name": "width", + "name": "width", + "type": "INT", + "widget": { + "name": "width" + }, + "link": 96 + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "widget": { + "name": "height" + }, + "link": 97 + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "widget": { + "name": "batch_size" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "LATENT", + "name": "LATENT", + "type": "LATENT", + "links": [ + 78 + ] + } + ], + "properties": { + "Node name for S&R": "EmptyLatentImage", + "cnr_id": "comfy-core", + "ver": "0.3.40", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 1024, + 1024, + 1 + ] + }, + { + "id": 12, + "type": "CLIPTextEncode", + "pos": [ + 330, + 830 + ], + "size": [ + 490, + 140 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 81 + }, + { + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "links": [ + 40 + ] + } + ], + "title": "CLIP Text Encode (Negative Prompt)", + "properties": { + "Node name for S&R": "CLIPTextEncode", + "cnr_id": "comfy-core", + "ver": "0.3.65", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "worst quality, low quality, score_1, score_2, score_3, blurry, jpeg artifacts, sepia" + ], + "color": "#223", + "bgcolor": "#335" + }, + { + "id": 19, + "type": "KSampler", + "pos": [ + 870, + 120 + ], + "size": [ + 300, + 620 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 79 + }, + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "link": 39 + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "link": 40 + }, + { + "localized_name": "latent_image", + "name": "latent_image", + "type": "LATENT", + "link": 78 + }, + { + "localized_name": "seed", + "name": "seed", + "type": "INT", + "widget": { + "name": "seed" + }, + "link": 100 + }, + { + "localized_name": "steps", + "name": "steps", + "type": "INT", + "widget": { + "name": "steps" + }, + "link": 98 + }, + { + "localized_name": "cfg", + "name": "cfg", + "type": "FLOAT", + "widget": { + "name": "cfg" + }, + "link": 99 + }, + { + "localized_name": "sampler_name", + "name": "sampler_name", + "type": "COMBO", + "widget": { + "name": "sampler_name" + }, + "link": null + }, + { + "localized_name": "scheduler", + "name": "scheduler", + "type": "COMBO", + "widget": { + "name": "scheduler" + }, + "link": null + }, + { + "localized_name": "denoise", + "name": "denoise", + "type": "FLOAT", + "widget": { + "name": "denoise" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "LATENT", + "name": "LATENT", + "type": "LATENT", + "slot_index": 0, + "links": [ + 10 + ] + } + ], + "properties": { + "Node name for S&R": "KSampler", + "cnr_id": "comfy-core", + "ver": "0.3.40", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 0, + "fixed", + 30, + 4, + "er_sde", + "simple", + 1 + ] + }, + { + "id": 11, + "type": "CLIPTextEncode", + "pos": [ + 320, + 170 + ], + "size": [ + 490, + 610 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 80 + }, + { + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": 95 + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "links": [ + 39 + ] + } + ], + "title": "CLIP Text Encode (Positive Prompt)", + "properties": { + "Node name for S&R": "CLIPTextEncode", + "cnr_id": "comfy-core", + "ver": "0.3.65", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "" + ], + "color": "#232", + "bgcolor": "#353" + }, + { + "id": 44, + "type": "UNETLoader", + "pos": [ + -50, + 170 + ], + "size": [ + 310, + 130 + ], + "flags": {}, + "order": 6, + "mode": 0, + "inputs": [ + { + "localized_name": "unet_name", + "name": "unet_name", + "type": "COMBO", + "widget": { + "name": "unet_name" + }, + "link": 101 + }, + { + "localized_name": "weight_dtype", + "name": "weight_dtype", + "type": "COMBO", + "widget": { + "name": "weight_dtype" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "links": [ + 79 + ] + } + ], + "properties": { + "Node name for S&R": "UNETLoader", + "cnr_id": "comfy-core", + "ver": "0.11.0", + "models": [ + { + "name": "anima-base-v1.0.safetensors", + "url": "https://huggingface.co/circlestone-labs/Anima/resolve/main/split_files/diffusion_models/anima-base-v1.0.safetensors", + "directory": "diffusion_models" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "anima-base-v1.0.safetensors", + "default" + ] + } + ], + "groups": [ + { + "id": 1, + "title": "Model", + "bounding": [ + -80, + 80, + 360, + 640 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 2, + "title": "Image Size(1MP)", + "bounding": [ + -80, + 750, + 360, + 240 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 3, + "title": "Prompt", + "bounding": [ + 300, + 80, + 530, + 910 + ], + "color": "#3f789e", + "flags": {} + } + ], + "links": [ + { + "id": 10, + "origin_id": 19, + "origin_slot": 0, + "target_id": 8, + "target_slot": 0, + "type": "LATENT" + }, + { + "id": 11, + "origin_id": 15, + "origin_slot": 0, + "target_id": 8, + "target_slot": 1, + "type": "VAE" + }, + { + "id": 81, + "origin_id": 45, + "origin_slot": 0, + "target_id": 12, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 79, + "origin_id": 44, + "origin_slot": 0, + "target_id": 19, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 39, + "origin_id": 11, + "origin_slot": 0, + "target_id": 19, + "target_slot": 1, + "type": "CONDITIONING" + }, + { + "id": 40, + "origin_id": 12, + "origin_slot": 0, + "target_id": 19, + "target_slot": 2, + "type": "CONDITIONING" + }, + { + "id": 78, + "origin_id": 28, + "origin_slot": 0, + "target_id": 19, + "target_slot": 3, + "type": "LATENT" + }, + { + "id": 80, + "origin_id": 45, + "origin_slot": 0, + "target_id": 11, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 82, + "origin_id": 8, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 95, + "origin_id": -10, + "origin_slot": 0, + "target_id": 11, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 96, + "origin_id": -10, + "origin_slot": 1, + "target_id": 28, + "target_slot": 0, + "type": "INT" + }, + { + "id": 97, + "origin_id": -10, + "origin_slot": 2, + "target_id": 28, + "target_slot": 1, + "type": "INT" + }, + { + "id": 98, + "origin_id": -10, + "origin_slot": 3, + "target_id": 19, + "target_slot": 5, + "type": "INT" + }, + { + "id": 99, + "origin_id": -10, + "origin_slot": 4, + "target_id": 19, + "target_slot": 6, + "type": "FLOAT" + }, + { + "id": 100, + "origin_id": -10, + "origin_slot": 5, + "target_id": 19, + "target_slot": 4, + "type": "INT" + }, + { + "id": 101, + "origin_id": -10, + "origin_slot": 6, + "target_id": 44, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 102, + "origin_id": -10, + "origin_slot": 7, + "target_id": 45, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 103, + "origin_id": -10, + "origin_slot": 8, + "target_id": 15, + "target_slot": 0, + "type": "COMBO" + } + ], + "extra": {}, + "category": "Image generation and editing/Text to image" + } + ] + }, + "extra": {} +} \ No newline at end of file diff --git a/blueprints/Video Captioning (Gemini).json b/blueprints/Video Captioning (Gemini).json index 7642b23c1..54a7d6e78 100644 --- a/blueprints/Video Captioning (Gemini).json +++ b/blueprints/Video Captioning (Gemini).json @@ -307,9 +307,9 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Text generation/Video Captioning", + "category": "Video Tools", "description": "Generates descriptive captions for video input using Google's Gemini multimodal LLM." } ] } -} +} \ No newline at end of file diff --git a/blueprints/Video Depth Estimation (MoGe).json b/blueprints/Video Depth Estimation (MoGe).json new file mode 100644 index 000000000..025e20cda --- /dev/null +++ b/blueprints/Video Depth Estimation (MoGe).json @@ -0,0 +1,1226 @@ +{ + "revision": 0, + "last_node_id": 72, + "last_link_id": 0, + "nodes": [ + { + "id": 72, + "type": "7ff83f68-6848-47a8-aa43-9036ca6c46e8", + "pos": [ + -4440, + 4550 + ], + "size": [ + 430, + 330 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "inference_resolution", + "name": "inference_resolution", + "type": "INT", + "widget": { + "name": "inference_resolution" + }, + "link": null + }, + { + "localized_name": "inference_batch_size", + "name": "inference_batch_size", + "type": "INT", + "widget": { + "name": "inference_batch_size" + }, + "link": null + }, + { + "localized_name": "moge_model", + "name": "moge_model", + "type": "COMBO", + "widget": { + "name": "moge_model" + }, + "link": null + }, + { + "label": "auto_resize_input", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": null + }, + { + "name": "video", + "type": "VIDEO", + "link": null + } + ], + "outputs": [ + { + "localized_name": "depth_colored", + "name": "depth_colored", + "type": "IMAGE", + "links": [] + }, + { + "localized_name": "depth", + "name": "depth", + "type": "IMAGE", + "links": [] + }, + { + "name": "MASK", + "type": "MASK", + "links": [] + } + ], + "properties": { + "proxyWidgets": [ + [ + "13", + "resolution_level" + ], + [ + "13", + "batch_size" + ], + [ + "32", + "model_name" + ], + [ + "53", + "switch" + ] + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.21.1" + }, + "widgets_values": [], + "title": "Video Depth Estimation (MoGe)" + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "7ff83f68-6848-47a8-aa43-9036ca6c46e8", + "version": 1, + "state": { + "lastGroupId": 1, + "lastNodeId": 72, + "lastLinkId": 96, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Video Depth Estimation (MoGe)", + "inputNode": { + "id": -10, + "bounding": [ + -5320, + 5320, + 167.337890625, + 148 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + -3090, + 4966, + 129, + 108 + ] + }, + "inputs": [ + { + "id": "06eefa21-8e60-49f3-9a34-35b081f4ae52", + "name": "inference_resolution", + "type": "INT", + "linkIds": [ + 73 + ], + "localized_name": "inference_resolution", + "pos": [ + -5176.662109375, + 5344 + ] + }, + { + "id": "616638fe-f603-4d10-bae9-fc87c134380f", + "name": "inference_batch_size", + "type": "INT", + "linkIds": [ + 74 + ], + "localized_name": "inference_batch_size", + "pos": [ + -5176.662109375, + 5364 + ] + }, + { + "id": "65694805-186e-4181-a721-df8b5af49d31", + "name": "moge_model", + "type": "COMBO", + "linkIds": [ + 79 + ], + "localized_name": "moge_model", + "pos": [ + -5176.662109375, + 5384 + ] + }, + { + "id": "badf1be1-53c6-4fc1-b5cd-79ad3daf1674", + "name": "switch", + "type": "BOOLEAN", + "linkIds": [ + 83 + ], + "label": "auto_resize_input", + "pos": [ + -5176.662109375, + 5404 + ] + }, + { + "id": "749bad18-d00a-4ec4-a5ff-e45b1d0cf089", + "name": "video", + "type": "VIDEO", + "linkIds": [ + 91 + ], + "pos": [ + -5176.662109375, + 5424 + ] + } + ], + "outputs": [ + { + "id": "59c37b52-074f-49fc-9731-483f899c12c4", + "name": "depth_colored", + "type": "IMAGE", + "linkIds": [ + 36 + ], + "localized_name": "depth_colored", + "pos": [ + -3066, + 4990 + ] + }, + { + "id": "f583e936-da5c-4630-9901-391fa605c1f8", + "name": "depth", + "type": "IMAGE", + "linkIds": [ + 40 + ], + "localized_name": "depth", + "pos": [ + -3066, + 5010 + ] + }, + { + "id": "6845b6a1-1980-454a-9451-314f24495c1d", + "name": "MASK", + "type": "MASK", + "linkIds": [ + 86 + ], + "pos": [ + -3066, + 5030 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 13, + "type": "MoGeInference", + "pos": [ + -3790, + 5180 + ], + "size": [ + 270, + 230 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "moge_model", + "name": "moge_model", + "type": "MOGE_MODEL", + "link": 58 + }, + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 81 + }, + { + "localized_name": "resolution_level", + "name": "resolution_level", + "type": "INT", + "widget": { + "name": "resolution_level" + }, + "link": 73 + }, + { + "localized_name": "fov_x_degrees", + "name": "fov_x_degrees", + "type": "FLOAT", + "widget": { + "name": "fov_x_degrees" + }, + "link": null + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "widget": { + "name": "batch_size" + }, + "link": 74 + }, + { + "localized_name": "force_projection", + "name": "force_projection", + "type": "BOOLEAN", + "widget": { + "name": "force_projection" + }, + "link": null + }, + { + "localized_name": "apply_mask", + "name": "apply_mask", + "type": "BOOLEAN", + "widget": { + "name": "apply_mask" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "moge_geometry", + "name": "moge_geometry", + "type": "MOGE_GEOMETRY", + "links": [ + 35, + 39, + 61 + ] + } + ], + "properties": { + "Node name for S&R": "MoGeInference", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.21.1" + }, + "widgets_values": [ + 9, + 0, + 4, + true, + true + ] + }, + { + "id": 23, + "type": "MoGeRender", + "pos": [ + -3430, + 4870 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "moge_geometry", + "name": "moge_geometry", + "type": "MOGE_GEOMETRY", + "link": 35 + }, + { + "localized_name": "output", + "name": "output", + "type": "COMBO", + "widget": { + "name": "output" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 36 + ] + } + ], + "properties": { + "Node name for S&R": "MoGeRender", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.21.1" + }, + "widgets_values": [ + "depth_colored" + ] + }, + { + "id": 25, + "type": "MoGeRender", + "pos": [ + -3430, + 5030 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "moge_geometry", + "name": "moge_geometry", + "type": "MOGE_GEOMETRY", + "link": 39 + }, + { + "localized_name": "output", + "name": "output", + "type": "COMBO", + "widget": { + "name": "output" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 40 + ] + } + ], + "properties": { + "Node name for S&R": "MoGeRender", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.21.1" + }, + "widgets_values": [ + "depth" + ] + }, + { + "id": 32, + "type": "LoadMoGeModel", + "pos": [ + -4180, + 4880 + ], + "size": [ + 270, + 140 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "model_name", + "name": "model_name", + "type": "COMBO", + "widget": { + "name": "model_name" + }, + "link": 79 + } + ], + "outputs": [ + { + "localized_name": "MOGE_MODEL", + "name": "MOGE_MODEL", + "type": "MOGE_MODEL", + "links": [ + 58 + ] + } + ], + "properties": { + "Node name for S&R": "LoadMoGeModel", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.21.1", + "models": [ + { + "name": "moge_2_vitl_normal_fp16.safetensors", + "url": "https://huggingface.co/Comfy-Org/MoGe/resolve/main/geometry_estimation/moge_2_vitl_normal_fp16.safetensors", + "directory": "geometry_estimation" + } + ] + }, + "widgets_values": [ + "moge_2_vitl_normal_fp16.safetensors" + ] + }, + { + "id": 36, + "type": "ComfyMathExpression", + "pos": [ + -4720, + 4910 + ], + "size": [ + 400, + 200 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "label": "a", + "localized_name": "values.a", + "name": "values.a", + "type": "FLOAT,INT,BOOLEAN", + "link": 49 + }, + { + "label": "b", + "localized_name": "values.b", + "name": "values.b", + "shape": 7, + "type": "FLOAT,INT,BOOLEAN", + "link": null + }, + { + "localized_name": "expression", + "name": "expression", + "type": "STRING", + "widget": { + "name": "expression" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "FLOAT", + "name": "FLOAT", + "type": "FLOAT", + "links": null + }, + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": null + }, + { + "localized_name": "BOOL", + "name": "BOOL", + "type": "BOOLEAN", + "links": [ + 53 + ] + } + ], + "properties": { + "Node name for S&R": "ComfyMathExpression", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.21.1" + }, + "widgets_values": [ + "a > 2048" + ] + }, + { + "id": 37, + "type": "GetImageSize", + "pos": [ + -4980, + 4910 + ], + "size": [ + 230, + 160 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 92 + } + ], + "outputs": [ + { + "localized_name": "width", + "name": "width", + "type": "INT", + "links": [ + 49 + ] + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "links": null + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "links": null + } + ], + "properties": { + "Node name for S&R": "GetImageSize", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.21.1" + } + }, + { + "id": 40, + "type": "ResizeImagesByLongerEdge", + "pos": [ + -4650, + 5210 + ], + "size": [ + 310, + 110 + ], + "flags": {}, + "order": 6, + "mode": 0, + "inputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "link": 93 + }, + { + "localized_name": "longer_edge", + "name": "longer_edge", + "type": "INT", + "widget": { + "name": "longer_edge" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "links": [ + 54 + ] + } + ], + "properties": { + "Node name for S&R": "ResizeImagesByLongerEdge", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.21.1" + }, + "widgets_values": [ + 2048 + ] + }, + { + "id": 42, + "type": "ComfySwitchNode", + "pos": [ + -4180, + 5060 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 7, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 94 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 54 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 53 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 80 + ] + } + ], + "properties": { + "Node name for S&R": "ComfySwitchNode", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.21.1" + }, + "widgets_values": [ + false + ] + }, + { + "id": 45, + "type": "MoGeRender", + "pos": [ + -3430, + 5200 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 8, + "mode": 0, + "inputs": [ + { + "localized_name": "moge_geometry", + "name": "moge_geometry", + "type": "MOGE_GEOMETRY", + "link": 61 + }, + { + "localized_name": "output", + "name": "output", + "type": "COMBO", + "widget": { + "name": "output" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 85 + ] + } + ], + "properties": { + "Node name for S&R": "MoGeRender", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.21.1" + }, + "widgets_values": [ + "mask" + ] + }, + { + "id": 53, + "type": "ComfySwitchNode", + "pos": [ + -4160, + 5340 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 9, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 95 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 80 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 83 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 81 + ] + } + ], + "properties": { + "Node name for S&R": "ComfySwitchNode", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.21.1" + }, + "widgets_values": [ + true + ] + }, + { + "id": 68, + "type": "ImageToMask", + "pos": [ + -3420, + 5360 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 10, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 85 + }, + { + "localized_name": "channel", + "name": "channel", + "type": "COMBO", + "widget": { + "name": "channel" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MASK", + "name": "MASK", + "type": "MASK", + "links": [ + 86 + ] + } + ], + "properties": { + "Node name for S&R": "ImageToMask", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.22.0" + }, + "widgets_values": [ + "red" + ] + }, + { + "id": 70, + "type": "GetVideoComponents", + "pos": [ + -4920, + 5490 + ], + "size": [ + 230, + 120 + ], + "flags": {}, + "order": 11, + "mode": 0, + "inputs": [ + { + "localized_name": "video", + "name": "video", + "type": "VIDEO", + "link": 91 + } + ], + "outputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "links": [ + 92, + 93, + 94, + 95 + ] + }, + { + "localized_name": "audio", + "name": "audio", + "type": "AUDIO", + "links": null + }, + { + "localized_name": "fps", + "name": "fps", + "type": "FLOAT", + "links": null + } + ], + "properties": { + "Node name for S&R": "GetVideoComponents", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.22.0" + } + } + ], + "groups": [ + { + "id": 1, + "title": "auto_resize_if_width_gt_2048", + "bounding": [ + -5000, + 4840, + 690, + 280 + ], + "color": "#3f789e", + "flags": {} + } + ], + "links": [ + { + "id": 58, + "origin_id": 32, + "origin_slot": 0, + "target_id": 13, + "target_slot": 0, + "type": "MOGE_MODEL" + }, + { + "id": 35, + "origin_id": 13, + "origin_slot": 0, + "target_id": 23, + "target_slot": 0, + "type": "MOGE_GEOMETRY" + }, + { + "id": 39, + "origin_id": 13, + "origin_slot": 0, + "target_id": 25, + "target_slot": 0, + "type": "MOGE_GEOMETRY" + }, + { + "id": 49, + "origin_id": 37, + "origin_slot": 0, + "target_id": 36, + "target_slot": 0, + "type": "INT" + }, + { + "id": 54, + "origin_id": 40, + "origin_slot": 0, + "target_id": 42, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 53, + "origin_id": 36, + "origin_slot": 2, + "target_id": 42, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 61, + "origin_id": 13, + "origin_slot": 0, + "target_id": 45, + "target_slot": 0, + "type": "MOGE_GEOMETRY" + }, + { + "id": 36, + "origin_id": 23, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 40, + "origin_id": 25, + "origin_slot": 0, + "target_id": -20, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 73, + "origin_id": -10, + "origin_slot": 0, + "target_id": 13, + "target_slot": 2, + "type": "INT" + }, + { + "id": 74, + "origin_id": -10, + "origin_slot": 1, + "target_id": 13, + "target_slot": 4, + "type": "INT" + }, + { + "id": 79, + "origin_id": -10, + "origin_slot": 2, + "target_id": 32, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 80, + "origin_id": 42, + "origin_slot": 0, + "target_id": 53, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 81, + "origin_id": 53, + "origin_slot": 0, + "target_id": 13, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 83, + "origin_id": -10, + "origin_slot": 3, + "target_id": 53, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 85, + "origin_id": 45, + "origin_slot": 0, + "target_id": 68, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 86, + "origin_id": 68, + "origin_slot": 0, + "target_id": -20, + "target_slot": 2, + "type": "MASK" + }, + { + "id": 91, + "origin_id": -10, + "origin_slot": 4, + "target_id": 70, + "target_slot": 0, + "type": "VIDEO" + }, + { + "id": 92, + "origin_id": 70, + "origin_slot": 0, + "target_id": 37, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 93, + "origin_id": 70, + "origin_slot": 0, + "target_id": 40, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 94, + "origin_id": 70, + "origin_slot": 0, + "target_id": 42, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 95, + "origin_id": 70, + "origin_slot": 0, + "target_id": 53, + "target_slot": 0, + "type": "IMAGE" + } + ], + "extra": {}, + "category": "Conditioning & Preprocessors/Depth", + "description": "Estimates monocular depth from an input video using MoGe, outputting both raw and colorized depth maps plus a mask." + } + ] + }, + "extra": {} +} \ No newline at end of file diff --git a/blueprints/Video Face Detection (Mediapipe).json b/blueprints/Video Face Detection (Mediapipe).json new file mode 100644 index 000000000..c70352481 --- /dev/null +++ b/blueprints/Video Face Detection (Mediapipe).json @@ -0,0 +1,1109 @@ +{ + "revision": 0, + "last_node_id": 167, + "last_link_id": 0, + "nodes": [ + { + "id": 167, + "type": "ca14b151-8f5e-4386-aab7-d2ec84eaf43c", + "pos": [ + -3410, + 6100 + ], + "size": [ + 420, + 481.3125 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "name": "video", + "type": "VIDEO", + "link": null + }, + { + "label": "trim_audio", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": null + }, + { + "name": "start_time", + "type": "FLOAT", + "widget": { + "name": "start_time" + }, + "link": null + }, + { + "name": "duration", + "type": "FLOAT", + "widget": { + "name": "duration" + }, + "link": null + }, + { + "label": "face_landmarker", + "name": "face_landmarker_1", + "type": "FACE_LANDMARKER", + "link": null + }, + { + "label": "detector_variant", + "name": "detector_variant_1", + "type": "COMBO", + "widget": { + "name": "detector_variant_1" + }, + "link": null + }, + { + "label": "num_faces", + "name": "num_faces_1", + "type": "INT", + "widget": { + "name": "num_faces_1" + }, + "link": null + }, + { + "label": "face_oval", + "name": "regions.face_oval", + "type": "BOOLEAN", + "widget": { + "name": "regions.face_oval" + }, + "link": null + }, + { + "label": "face_lips", + "name": "regions.lips", + "type": "BOOLEAN", + "widget": { + "name": "regions.lips" + }, + "link": null + }, + { + "label": "left_eye", + "name": "regions.left_eye", + "type": "BOOLEAN", + "widget": { + "name": "regions.left_eye" + }, + "link": null + }, + { + "label": "right_eye", + "name": "regions.right_eye_1", + "type": "BOOLEAN", + "widget": { + "name": "regions.right_eye_1" + }, + "link": null + }, + { + "label": "irises", + "name": "regions.irises_1", + "type": "BOOLEAN", + "widget": { + "name": "regions.irises_1" + }, + "link": null + }, + { + "name": "model_name", + "type": "COMBO", + "widget": { + "name": "model_name" + }, + "link": null + } + ], + "outputs": [ + { + "label": "mask", + "name": "MASK_1", + "type": "MASK", + "links": [] + }, + { + "label": "bboxes", + "name": "bboxes_1", + "type": "BOUNDING_BOX", + "links": null + }, + { + "name": "face_landmarks", + "type": "FACE_LANDMARKS", + "links": null + } + ], + "title": "Video Face Detection (Mediapipe)", + "properties": { + "proxyWidgets": [ + [ + "165", + "switch" + ], + [ + "164", + "start_time" + ], + [ + "164", + "duration" + ], + [ + "11", + "detector_variant" + ], + [ + "11", + "num_faces" + ], + [ + "20", + "regions.face_oval" + ], + [ + "20", + "regions.lips" + ], + [ + "20", + "regions.left_eye" + ], + [ + "20", + "regions.right_eye" + ], + [ + "20", + "regions.irises" + ], + [ + "2", + "model_name" + ] + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.22.0" + }, + "widgets_values": [] + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "ca14b151-8f5e-4386-aab7-d2ec84eaf43c", + "version": 1, + "state": { + "lastGroupId": 2, + "lastNodeId": 167, + "lastLinkId": 168, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Video Face Detection (Mediapipe)", + "description": "Detects facial landmarks from a video using MediaPipe, outputting landmark data, face bounding boxes, and an optional face-region mask.", + "inputNode": { + "id": -10, + "bounding": [ + -1060, + 4350, + 142.587890625, + 308 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + 470, + 4460, + 137.677734375, + 108 + ] + }, + "inputs": [ + { + "id": "16e5a20f-22bc-4960-a67b-e32c64409c49", + "name": "video", + "type": "VIDEO", + "linkIds": [ + 150, + 153 + ], + "pos": [ + -941.412109375, + 4374 + ] + }, + { + "id": "cc7fc7d4-24ec-4c00-878e-1af1b6809b4b", + "name": "switch", + "type": "BOOLEAN", + "linkIds": [ + 154 + ], + "label": "trim_audio", + "pos": [ + -941.412109375, + 4394 + ] + }, + { + "id": "efa9ab9f-ca70-449c-be43-5ca60c7f0d59", + "name": "start_time", + "type": "FLOAT", + "linkIds": [ + 155 + ], + "pos": [ + -941.412109375, + 4414 + ] + }, + { + "id": "45050127-4089-4b85-bf81-73b725196c2e", + "name": "duration", + "type": "FLOAT", + "linkIds": [ + 156 + ], + "pos": [ + -941.412109375, + 4434 + ] + }, + { + "id": "239fcd3b-6324-4824-8255-98199ae58914", + "name": "face_landmarker_1", + "type": "FACE_LANDMARKER", + "linkIds": [ + 157 + ], + "label": "face_landmarker", + "pos": [ + -941.412109375, + 4454 + ] + }, + { + "id": "f79f67b9-5bcb-4cab-9101-8b9dee461bca", + "name": "detector_variant_1", + "type": "COMBO", + "linkIds": [ + 158 + ], + "label": "detector_variant", + "pos": [ + -941.412109375, + 4474 + ] + }, + { + "id": "3369790b-e730-41bf-b5b2-dc1f5fafbe11", + "name": "num_faces_1", + "type": "INT", + "linkIds": [ + 159 + ], + "label": "num_faces", + "pos": [ + -941.412109375, + 4494 + ] + }, + { + "id": "964f6b5f-44ac-456e-ba3a-a3039dfe0729", + "name": "regions.face_oval", + "type": "BOOLEAN", + "linkIds": [ + 160 + ], + "label": "face_oval", + "pos": [ + -941.412109375, + 4514 + ] + }, + { + "id": "d6e89b51-65a2-4f37-a561-8cec3a5040fd", + "name": "regions.lips", + "type": "BOOLEAN", + "linkIds": [ + 161 + ], + "label": "face_lips", + "pos": [ + -941.412109375, + 4534 + ] + }, + { + "id": "49f02319-ea4a-4a69-88f8-589d2ef7c97a", + "name": "regions.left_eye", + "type": "BOOLEAN", + "linkIds": [ + 162 + ], + "label": "left_eye", + "pos": [ + -941.412109375, + 4554 + ] + }, + { + "id": "89179a19-aca6-4469-a0b9-2a4bd21bceea", + "name": "regions.right_eye_1", + "type": "BOOLEAN", + "linkIds": [ + 163 + ], + "label": "right_eye", + "pos": [ + -941.412109375, + 4574 + ] + }, + { + "id": "f5667690-24b5-4df9-9210-b8610c68ff5f", + "name": "regions.irises_1", + "type": "BOOLEAN", + "linkIds": [ + 164 + ], + "label": "irises", + "pos": [ + -941.412109375, + 4594 + ] + }, + { + "id": "66c805f6-6ccd-41f9-8a77-fc934b7f4713", + "name": "model_name", + "type": "COMBO", + "linkIds": [ + 165 + ], + "pos": [ + -941.412109375, + 4614 + ] + } + ], + "outputs": [ + { + "id": "f6309e1d-6397-4363-b38f-778a122abc51", + "name": "MASK_1", + "type": "MASK", + "linkIds": [ + 83 + ], + "label": "mask", + "pos": [ + 494, + 4484 + ] + }, + { + "id": "59669f0a-b4b2-49d1-85f8-fc2a88059b1a", + "name": "bboxes_1", + "type": "BOUNDING_BOX", + "linkIds": [ + 166 + ], + "label": "bboxes", + "pos": [ + 494, + 4504 + ] + }, + { + "id": "57f66731-e106-4f8b-a0a0-aed3c620b37b", + "name": "face_landmarks", + "type": "FACE_LANDMARKS", + "linkIds": [ + 167 + ], + "pos": [ + 494, + 4524 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 11, + "type": "MediaPipeFaceLandmarker", + "pos": [ + -60, + 4380 + ], + "size": [ + 350, + 220 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "face_detection_model", + "name": "face_detection_model", + "type": "FACE_DETECTION_MODEL", + "link": 66 + }, + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 149 + }, + { + "localized_name": "detector_variant", + "name": "detector_variant", + "type": "COMBO", + "widget": { + "name": "detector_variant" + }, + "link": 158 + }, + { + "localized_name": "num_faces", + "name": "num_faces", + "type": "INT", + "widget": { + "name": "num_faces" + }, + "link": 159 + }, + { + "localized_name": "min_confidence", + "name": "min_confidence", + "type": "FLOAT", + "widget": { + "name": "min_confidence" + }, + "link": null + }, + { + "localized_name": "missing_frame_fallback", + "name": "missing_frame_fallback", + "type": "COMBO", + "widget": { + "name": "missing_frame_fallback" + }, + "link": null + }, + { + "name": "face_landmarker", + "type": "FACE_LANDMARKER", + "link": 157 + } + ], + "outputs": [ + { + "localized_name": "face_landmarks", + "name": "face_landmarks", + "type": "FACE_LANDMARKS", + "links": [ + 46, + 167 + ] + }, + { + "localized_name": "bboxes", + "name": "bboxes", + "type": "BOUNDING_BOX", + "links": [ + 166 + ] + } + ], + "properties": { + "Node name for S&R": "MediaPipeFaceLandmarker", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.22.0" + }, + "widgets_values": [ + "full", + 0, + 0.5, + "empty" + ] + }, + { + "id": 2, + "type": "LoadMediaPipeFaceLandmarker", + "pos": [ + -70, + 4160 + ], + "size": [ + 350, + 140 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "model_name", + "name": "model_name", + "type": "COMBO", + "widget": { + "name": "model_name" + }, + "link": 165 + } + ], + "outputs": [ + { + "localized_name": "FACE_DETECTION_MODEL", + "name": "FACE_DETECTION_MODEL", + "type": "FACE_DETECTION_MODEL", + "links": [ + 66 + ] + } + ], + "properties": { + "Node name for S&R": "LoadMediaPipeFaceLandmarker", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.22.0", + "models": [ + { + "name": "mediapipe_face_fp32.safetensors", + "url": "https://huggingface.co/Comfy-Org/mediapipe/resolve/main/detection/mediapipe_face_fp32.safetensors", + "directory": "detection" + } + ] + }, + "widgets_values": [ + "mediapipe_face_fp32.safetensors" + ] + }, + { + "id": 20, + "type": "MediaPipeFaceMask", + "pos": [ + -70, + 4660 + ], + "size": [ + 360, + 180 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "face_landmarks", + "name": "face_landmarks", + "type": "FACE_LANDMARKS", + "link": 46 + }, + { + "localized_name": "regions", + "name": "regions", + "type": "COMFY_DYNAMICCOMBO_V3", + "widget": { + "name": "regions" + }, + "link": null + }, + { + "localized_name": "regions.face_oval", + "name": "regions.face_oval", + "type": "BOOLEAN", + "widget": { + "name": "regions.face_oval" + }, + "link": 160 + }, + { + "localized_name": "regions.lips", + "name": "regions.lips", + "type": "BOOLEAN", + "widget": { + "name": "regions.lips" + }, + "link": 161 + }, + { + "localized_name": "regions.left_eye", + "name": "regions.left_eye", + "type": "BOOLEAN", + "widget": { + "name": "regions.left_eye" + }, + "link": 162 + }, + { + "localized_name": "regions.right_eye", + "name": "regions.right_eye", + "type": "BOOLEAN", + "widget": { + "name": "regions.right_eye" + }, + "link": 163 + }, + { + "localized_name": "regions.irises", + "name": "regions.irises", + "type": "BOOLEAN", + "widget": { + "name": "regions.irises" + }, + "link": 164 + } + ], + "outputs": [ + { + "localized_name": "MASK", + "name": "MASK", + "type": "MASK", + "links": [ + 83 + ] + } + ], + "properties": { + "Node name for S&R": "MediaPipeFaceMask", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.22.0" + }, + "widgets_values": [ + "custom", + true, + false, + false, + false, + false + ] + }, + { + "id": 160, + "type": "GetVideoComponents", + "pos": [ + -420, + 4360 + ], + "size": [ + 230, + 120 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "video", + "name": "video", + "type": "VIDEO", + "link": 152 + } + ], + "outputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "links": [ + 149 + ] + }, + { + "localized_name": "audio", + "name": "audio", + "type": "AUDIO", + "links": null + }, + { + "localized_name": "fps", + "name": "fps", + "type": "FLOAT", + "links": null + } + ], + "properties": { + "Node name for S&R": "GetVideoComponents", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.22.0" + } + }, + { + "id": 164, + "type": "Video Slice", + "pos": [ + -780, + 4330 + ], + "size": [ + 270, + 170 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "localized_name": "video", + "name": "video", + "type": "VIDEO", + "link": 150 + }, + { + "localized_name": "start_time", + "name": "start_time", + "type": "FLOAT", + "widget": { + "name": "start_time" + }, + "link": 155 + }, + { + "localized_name": "duration", + "name": "duration", + "type": "FLOAT", + "widget": { + "name": "duration" + }, + "link": 156 + }, + { + "localized_name": "strict_duration", + "name": "strict_duration", + "type": "BOOLEAN", + "widget": { + "name": "strict_duration" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "VIDEO", + "name": "VIDEO", + "type": "VIDEO", + "links": [ + 151 + ] + } + ], + "properties": { + "Node name for S&R": "Video Slice", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.22.0" + }, + "widgets_values": [ + 0, + 0, + false + ] + }, + { + "id": 165, + "type": "ComfySwitchNode", + "pos": [ + -420, + 4590 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 153 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 151 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 154 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 152 + ] + } + ], + "properties": { + "Node name for S&R": "ComfySwitchNode", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.22.0" + }, + "widgets_values": [ + false + ] + } + ], + "groups": [], + "links": [ + { + "id": 66, + "origin_id": 2, + "origin_slot": 0, + "target_id": 11, + "target_slot": 0, + "type": "FACE_DETECTION_MODEL" + }, + { + "id": 46, + "origin_id": 11, + "origin_slot": 0, + "target_id": 20, + "target_slot": 0, + "type": "FACE_LANDMARKS" + }, + { + "id": 83, + "origin_id": 20, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "MASK" + }, + { + "id": 149, + "origin_id": 160, + "origin_slot": 0, + "target_id": 11, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 150, + "origin_id": -10, + "origin_slot": 0, + "target_id": 164, + "target_slot": 0, + "type": "VIDEO" + }, + { + "id": 151, + "origin_id": 164, + "origin_slot": 0, + "target_id": 165, + "target_slot": 1, + "type": "VIDEO" + }, + { + "id": 152, + "origin_id": 165, + "origin_slot": 0, + "target_id": 160, + "target_slot": 0, + "type": "VIDEO" + }, + { + "id": 153, + "origin_id": -10, + "origin_slot": 0, + "target_id": 165, + "target_slot": 0, + "type": "VIDEO" + }, + { + "id": 154, + "origin_id": -10, + "origin_slot": 1, + "target_id": 165, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 155, + "origin_id": -10, + "origin_slot": 2, + "target_id": 164, + "target_slot": 1, + "type": "FLOAT" + }, + { + "id": 156, + "origin_id": -10, + "origin_slot": 3, + "target_id": 164, + "target_slot": 2, + "type": "FLOAT" + }, + { + "id": 157, + "origin_id": -10, + "origin_slot": 4, + "target_id": 11, + "target_slot": 6, + "type": "FACE_LANDMARKER" + }, + { + "id": 158, + "origin_id": -10, + "origin_slot": 5, + "target_id": 11, + "target_slot": 2, + "type": "COMBO" + }, + { + "id": 159, + "origin_id": -10, + "origin_slot": 6, + "target_id": 11, + "target_slot": 3, + "type": "INT" + }, + { + "id": 160, + "origin_id": -10, + "origin_slot": 7, + "target_id": 20, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 161, + "origin_id": -10, + "origin_slot": 8, + "target_id": 20, + "target_slot": 3, + "type": "BOOLEAN" + }, + { + "id": 162, + "origin_id": -10, + "origin_slot": 9, + "target_id": 20, + "target_slot": 4, + "type": "BOOLEAN" + }, + { + "id": 163, + "origin_id": -10, + "origin_slot": 10, + "target_id": 20, + "target_slot": 5, + "type": "BOOLEAN" + }, + { + "id": 164, + "origin_id": -10, + "origin_slot": 11, + "target_id": 20, + "target_slot": 6, + "type": "BOOLEAN" + }, + { + "id": 165, + "origin_id": -10, + "origin_slot": 12, + "target_id": 2, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 166, + "origin_id": 11, + "origin_slot": 1, + "target_id": -20, + "target_slot": 1, + "type": "BOUNDING_BOX" + }, + { + "id": 167, + "origin_id": 11, + "origin_slot": 0, + "target_id": -20, + "target_slot": 2, + "type": "FACE_LANDMARKS" + } + ], + "extra": {}, + "category": "Conditioning & Preprocessors/Face Detection" + } + ] + }, + "extra": {} +} \ No newline at end of file diff --git a/blueprints/Video Inpaint (VOID).json b/blueprints/Video Inpaint (VOID).json new file mode 100644 index 000000000..a7cc806b5 --- /dev/null +++ b/blueprints/Video Inpaint (VOID).json @@ -0,0 +1,4340 @@ +{ + "revision": 0, + "last_node_id": 167, + "last_link_id": 0, + "nodes": [ + { + "id": 167, + "type": "c3157b75-484a-459e-b8de-57823bef5130", + "pos": [ + -430, + 690 + ], + "size": [ + 590, + 723.9375 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "label": "Source video", + "localized_name": "source_video", + "name": "source_video", + "type": "VIDEO", + "link": null + }, + { + "label": "Positive prompt (inpaint fill)", + "localized_name": "positive_prompt", + "name": "positive_prompt", + "type": "STRING", + "widget": { + "name": "positive_prompt" + }, + "link": null + }, + { + "label": "Negative prompt", + "localized_name": "negative_prompt", + "name": "negative_prompt", + "type": "STRING", + "widget": { + "name": "negative_prompt" + }, + "link": null + }, + { + "label": "SAM3 object mask prompt", + "localized_name": "sam3_text_prompt", + "name": "sam3_text_prompt", + "type": "STRING", + "widget": { + "name": "sam3_text_prompt" + }, + "link": null + }, + { + "label": "Start frame index", + "localized_name": "start_frame_index", + "name": "start_frame_index", + "type": "INT", + "widget": { + "name": "start_frame_index" + }, + "link": null + }, + { + "label": "Clip duration (seconds)", + "localized_name": "duration_seconds", + "name": "duration_seconds", + "type": "INT", + "widget": { + "name": "duration_seconds" + }, + "link": null + }, + { + "label": "Width (pass 2)", + "localized_name": "latent_width", + "name": "latent_width", + "type": "INT", + "widget": { + "name": "latent_width" + }, + "link": null + }, + { + "label": "Height (pass 2)", + "localized_name": "latent_height", + "name": "latent_height", + "type": "INT", + "widget": { + "name": "latent_height" + }, + "link": null + }, + { + "label": "Skip pass 2 (reuse pass 1)", + "localized_name": "skip_pass_2", + "name": "skip_pass_2", + "type": "BOOLEAN", + "widget": { + "name": "skip_pass_2" + }, + "link": null + }, + { + "label": "Noise seed", + "localized_name": "noise_seed", + "name": "noise_seed", + "type": "INT", + "widget": { + "name": "noise_seed" + }, + "link": null + }, + { + "label": "SAM3 checkpoint", + "localized_name": "sam3_checkpoint", + "name": "sam3_checkpoint", + "type": "COMBO", + "widget": { + "name": "sam3_checkpoint" + }, + "link": null + }, + { + "label": "VOID UNet — pass 1", + "localized_name": "void_unet_pass1", + "name": "void_unet_pass1", + "type": "COMBO", + "widget": { + "name": "void_unet_pass1" + }, + "link": null + }, + { + "label": "VOID UNet — pass 2", + "localized_name": "void_unet_pass2", + "name": "void_unet_pass2", + "type": "COMBO", + "widget": { + "name": "void_unet_pass2" + }, + "link": null + }, + { + "label": "Optical flow model", + "localized_name": "optical_flow_model", + "name": "optical_flow_model", + "type": "COMBO", + "widget": { + "name": "optical_flow_model" + }, + "link": null + }, + { + "label": "CLIP / T5 weights", + "localized_name": "clip_name", + "name": "clip_name", + "type": "COMBO", + "widget": { + "name": "clip_name" + }, + "link": null + }, + { + "label": "VAE weights", + "localized_name": "vae_name", + "name": "vae_name", + "type": "COMBO", + "widget": { + "name": "vae_name" + }, + "link": null + } + ], + "outputs": [ + { + "label": "Pass 1 (intermediate)", + "localized_name": "pass_1_video", + "name": "pass_1_video", + "type": "VIDEO", + "links": [] + }, + { + "label": "Pass 2 (final)", + "localized_name": "final_pass_2_video", + "name": "final_pass_2_video", + "type": "VIDEO", + "links": [] + } + ], + "properties": { + "proxyWidgets": [ + [ + "6", + "text" + ], + [ + "7", + "text" + ], + [ + "149", + "text" + ], + [ + "168", + "value" + ], + [ + "163", + "value" + ], + [ + "147", + "value" + ], + [ + "148", + "value" + ], + [ + "153", + "value" + ], + [ + "141", + "noise_seed" + ], + [ + "149", + "ckpt_name" + ], + [ + "144", + "unet_name" + ], + [ + "143", + "unet_name" + ], + [ + "142", + "model_name" + ], + [ + "2", + "clip_name" + ], + [ + "3", + "vae_name" + ] + ] + }, + "widgets_values": [], + "title": "Video Inpaint (VOID)" + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "c3157b75-484a-459e-b8de-57823bef5130", + "version": 1, + "state": { + "lastGroupId": 13, + "lastNodeId": 171, + "lastLinkId": 406, + "lastRerouteId": 0 + }, + "revision": 5, + "config": {}, + "name": "Video Inpaint (VOID)", + "inputNode": { + "id": -10, + "bounding": [ + -1530, + 800, + 203.1796875, + 368 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + 2030, + 710, + 166.130859375, + 88 + ] + }, + "inputs": [ + { + "id": "1865ea29-14b1-4471-b5e0-d35bba595b9c", + "name": "source_video", + "type": "VIDEO", + "linkIds": [ + 373 + ], + "localized_name": "source_video", + "label": "Source video", + "pos": [ + -1350.8203125, + 824 + ] + }, + { + "id": "f1b2b2c4-bc2e-4e72-b16c-7e560e58d2d6", + "name": "positive_prompt", + "type": "STRING", + "linkIds": [ + 377 + ], + "localized_name": "positive_prompt", + "label": "Positive prompt (inpaint fill)", + "pos": [ + -1350.8203125, + 844 + ] + }, + { + "id": "931ac4dd-3cb6-4555-a1f0-619be81d64f6", + "name": "negative_prompt", + "type": "STRING", + "linkIds": [ + 387 + ], + "localized_name": "negative_prompt", + "label": "Negative prompt", + "pos": [ + -1350.8203125, + 864 + ] + }, + { + "id": "7a0963c3-bf2f-464d-80c2-6a6c90569883", + "name": "sam3_text_prompt", + "type": "STRING", + "linkIds": [ + 388 + ], + "localized_name": "sam3_text_prompt", + "label": "SAM3 object mask prompt", + "pos": [ + -1350.8203125, + 884 + ] + }, + { + "id": "f53f340f-2031-401d-b613-157622ef336f", + "name": "start_frame_index", + "type": "INT", + "linkIds": [ + 389 + ], + "localized_name": "start_frame_index", + "label": "Start frame index", + "pos": [ + -1350.8203125, + 904 + ] + }, + { + "id": "d5b8704b-7c8c-4cf0-87cd-26b293f65f83", + "name": "duration_seconds", + "type": "INT", + "linkIds": [ + 390 + ], + "localized_name": "duration_seconds", + "label": "Clip duration (seconds)", + "pos": [ + -1350.8203125, + 924 + ] + }, + { + "id": "7140209f-5058-4933-ae06-438256f77f23", + "name": "latent_width", + "type": "INT", + "linkIds": [ + 391 + ], + "localized_name": "latent_width", + "label": "Width (pass 2)", + "pos": [ + -1350.8203125, + 944 + ] + }, + { + "id": "084a140a-6fa9-4676-9483-ad30e0b14947", + "name": "latent_height", + "type": "INT", + "linkIds": [ + 392 + ], + "localized_name": "latent_height", + "label": "Height (pass 2)", + "pos": [ + -1350.8203125, + 964 + ] + }, + { + "id": "a8109321-e101-4ed8-b6f3-8ad1c815f35c", + "name": "skip_pass_2", + "type": "BOOLEAN", + "linkIds": [ + 393 + ], + "localized_name": "skip_pass_2", + "label": "Skip pass 2 (reuse pass 1)", + "pos": [ + -1350.8203125, + 984 + ] + }, + { + "id": "6964ab42-0662-47f2-9c2a-96782fdcb883", + "name": "noise_seed", + "type": "INT", + "linkIds": [ + 400 + ], + "localized_name": "noise_seed", + "label": "Noise seed", + "pos": [ + -1350.8203125, + 1004 + ] + }, + { + "id": "dccde360-461d-417e-b3f5-e1a4d6cece39", + "name": "sam3_checkpoint", + "type": "COMBO", + "linkIds": [ + 401 + ], + "localized_name": "sam3_checkpoint", + "label": "SAM3 checkpoint", + "pos": [ + -1350.8203125, + 1024 + ] + }, + { + "id": "5ce0d036-be08-4539-9ec6-e923fcdb8825", + "name": "void_unet_pass1", + "type": "COMBO", + "linkIds": [ + 402 + ], + "localized_name": "void_unet_pass1", + "label": "VOID UNet — pass 1", + "pos": [ + -1350.8203125, + 1044 + ] + }, + { + "id": "c1de695a-a08a-40bc-b9e4-d156fef73cd0", + "name": "void_unet_pass2", + "type": "COMBO", + "linkIds": [ + 403 + ], + "localized_name": "void_unet_pass2", + "label": "VOID UNet — pass 2", + "pos": [ + -1350.8203125, + 1064 + ] + }, + { + "id": "99da50bc-db57-4a21-9831-0f77b3c4fe99", + "name": "optical_flow_model", + "type": "COMBO", + "linkIds": [ + 404 + ], + "localized_name": "optical_flow_model", + "label": "Optical flow model", + "pos": [ + -1350.8203125, + 1084 + ] + }, + { + "id": "c756ce20-cfa6-4fe0-9eb0-543d56781cb7", + "name": "clip_name", + "type": "COMBO", + "linkIds": [ + 405 + ], + "localized_name": "clip_name", + "label": "CLIP / T5 weights", + "pos": [ + -1350.8203125, + 1104 + ] + }, + { + "id": "d8eb12ad-a805-42d9-86b4-6f2c2cc5a231", + "name": "vae_name", + "type": "COMBO", + "linkIds": [ + 406 + ], + "localized_name": "vae_name", + "label": "VAE weights", + "pos": [ + -1350.8203125, + 1124 + ] + } + ], + "outputs": [ + { + "id": "a21e83df-8c95-43a3-bd73-feeea67e90cd", + "name": "pass_1_video", + "type": "VIDEO", + "linkIds": [ + 77 + ], + "localized_name": "pass_1_video", + "label": "Pass 1 (intermediate)", + "pos": [ + 2054, + 734 + ] + }, + { + "id": "02c265f3-012f-499f-a4e8-a6d6aaf72885", + "name": "final_pass_2_video", + "type": "VIDEO", + "linkIds": [ + 362 + ], + "localized_name": "final_pass_2_video", + "label": "Pass 2 (final)", + "pos": [ + 2054, + 754 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 2, + "type": "CLIPLoader", + "pos": [ + -710, + 30 + ], + "size": [ + 320, + 150 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "clip_name", + "name": "clip_name", + "type": "COMBO", + "widget": { + "name": "clip_name" + }, + "link": 405 + }, + { + "localized_name": "type", + "name": "type", + "type": "COMBO", + "widget": { + "name": "type" + }, + "link": null + }, + { + "localized_name": "device", + "name": "device", + "shape": 7, + "type": "COMBO", + "widget": { + "name": "device" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "CLIP", + "name": "CLIP", + "type": "CLIP", + "slot_index": 0, + "links": [ + 2, + 3 + ] + } + ], + "properties": { + "Node name for S&R": "CLIPLoader", + "cnr_id": "comfy-core", + "ver": "0.18.1", + "models": [ + { + "name": "t5xxl_fp16.safetensors", + "url": "https://huggingface.co/comfyanonymous/flux_text_encoders/resolve/main/t5xxl_fp16.safetensors", + "directory": "text_encoders" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "t5xxl_fp16.safetensors", + "cogvideox", + "default" + ] + }, + { + "id": 3, + "type": "VAELoader", + "pos": [ + -710, + 220 + ], + "size": [ + 320, + 90 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "vae_name", + "name": "vae_name", + "type": "COMBO", + "widget": { + "name": "vae_name" + }, + "link": 406 + } + ], + "outputs": [ + { + "localized_name": "VAE", + "name": "VAE", + "type": "VAE", + "slot_index": 0, + "links": [ + 4, + 45, + 70 + ] + } + ], + "properties": { + "Node name for S&R": "VAELoader", + "cnr_id": "comfy-core", + "ver": "0.18.1", + "models": [ + { + "name": "cogvideox_vae.safetensors", + "url": "https://huggingface.co/Comfy-Org/void-model/resolve/main/vae/cogvideox_vae.safetensors", + "directory": "vae" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "cogvideox_vae.safetensors" + ] + }, + { + "id": 7, + "type": "CLIPTextEncode", + "pos": [ + -260, + 200 + ], + "size": [ + 590, + 180 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 3 + }, + { + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": 387 + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "slot_index": 0, + "links": [ + 9 + ] + } + ], + "title": "Negative Prompt", + "properties": { + "Node name for S&R": "CLIPTextEncode", + "cnr_id": "comfy-core", + "ver": "0.18.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "" + ], + "color": "#223", + "bgcolor": "#335" + }, + { + "id": 136, + "type": "CFGGuider", + "pos": [ + 410, + 1640 + ], + "size": [ + 300, + 130 + ], + "flags": {}, + "order": 16, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 322 + }, + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "link": 309 + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "link": 310 + }, + { + "localized_name": "cfg", + "name": "cfg", + "type": "FLOAT", + "widget": { + "name": "cfg" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "GUIDER", + "name": "GUIDER", + "type": "GUIDER", + "links": [ + 311 + ] + } + ], + "title": "CFGGuider (Pass 2 cfg=6)", + "properties": { + "Node name for S&R": "CFGGuider", + "cnr_id": "comfy-core", + "ver": "0.21.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 6 + ] + }, + { + "id": 138, + "type": "BasicScheduler", + "pos": [ + 410, + 160 + ], + "size": [ + 270, + 150 + ], + "flags": {}, + "order": 18, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 324 + }, + { + "localized_name": "scheduler", + "name": "scheduler", + "type": "COMBO", + "widget": { + "name": "scheduler" + }, + "link": null + }, + { + "localized_name": "steps", + "name": "steps", + "type": "INT", + "widget": { + "name": "steps" + }, + "link": null + }, + { + "localized_name": "denoise", + "name": "denoise", + "type": "FLOAT", + "widget": { + "name": "denoise" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "SIGMAS", + "name": "SIGMAS", + "type": "SIGMAS", + "links": [ + 315 + ] + } + ], + "properties": { + "Node name for S&R": "BasicScheduler", + "cnr_id": "comfy-core", + "ver": "0.21.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "simple", + 30, + 1 + ] + }, + { + "id": 140, + "type": "CFGGuider", + "pos": [ + 410, + -30 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 19, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 325 + }, + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "link": 317 + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "link": 318 + }, + { + "localized_name": "cfg", + "name": "cfg", + "type": "FLOAT", + "widget": { + "name": "cfg" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "GUIDER", + "name": "GUIDER", + "type": "GUIDER", + "links": [ + 319 + ] + } + ], + "properties": { + "Node name for S&R": "CFGGuider", + "cnr_id": "comfy-core", + "ver": "0.21.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 6 + ] + }, + { + "id": 141, + "type": "RandomNoise", + "pos": [ + 410, + -180 + ], + "size": [ + 270, + 90 + ], + "flags": {}, + "order": 20, + "mode": 0, + "inputs": [ + { + "localized_name": "noise_seed", + "name": "noise_seed", + "type": "INT", + "widget": { + "name": "noise_seed" + }, + "link": 400 + } + ], + "outputs": [ + { + "localized_name": "NOISE", + "name": "NOISE", + "type": "NOISE", + "links": [ + 320 + ] + } + ], + "properties": { + "Node name for S&R": "RandomNoise", + "cnr_id": "comfy-core", + "ver": "0.21.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 43, + "fixed" + ] + }, + { + "id": 31, + "type": "VOIDWarpedNoise", + "pos": [ + 410, + 1090 + ], + "size": [ + 300, + 200 + ], + "flags": {}, + "order": 7, + "mode": 0, + "inputs": [ + { + "localized_name": "optical_flow", + "name": "optical_flow", + "type": "OPTICAL_FLOW", + "link": 321 + }, + { + "localized_name": "video", + "name": "video", + "type": "IMAGE", + "link": 72 + }, + { + "localized_name": "width", + "name": "width", + "type": "INT", + "widget": { + "name": "width" + }, + "link": 333 + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "widget": { + "name": "height" + }, + "link": 335 + }, + { + "localized_name": "length", + "name": "length", + "type": "INT", + "widget": { + "name": "length" + }, + "link": 67 + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "widget": { + "name": "batch_size" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "warped_noise", + "name": "warped_noise", + "type": "LATENT", + "slot_index": 0, + "links": [ + 53 + ] + } + ], + "title": "Warped Noise (from Pass 1 output)", + "properties": { + "Node name for S&R": "VOIDWarpedNoise", + "cnr_id": "comfy-core", + "ver": "0.18.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 672, + 384, + 45, + 1 + ] + }, + { + "id": 35, + "type": "SamplerCustomAdvanced", + "pos": [ + 870, + 1110 + ], + "size": [ + 250, + 170 + ], + "flags": {}, + "order": 9, + "mode": 0, + "inputs": [ + { + "localized_name": "noise", + "name": "noise", + "type": "NOISE", + "link": 54 + }, + { + "localized_name": "guider", + "name": "guider", + "type": "GUIDER", + "link": 311 + }, + { + "localized_name": "sampler", + "name": "sampler", + "type": "SAMPLER", + "link": 305 + }, + { + "localized_name": "sigmas", + "name": "sigmas", + "type": "SIGMAS", + "link": 313 + }, + { + "localized_name": "latent_image", + "name": "latent_image", + "type": "LATENT", + "link": 48 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "LATENT", + "slot_index": 0, + "links": [ + 49 + ] + }, + { + "localized_name": "denoised_output", + "name": "denoised_output", + "type": "LATENT", + "slot_index": 1, + "links": [] + } + ], + "title": "Pass 2 Sample", + "properties": { + "Node name for S&R": "SamplerCustomAdvanced", + "cnr_id": "comfy-core", + "ver": "0.18.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 132, + "type": "MaskPreview", + "pos": [ + 390, + 560 + ], + "size": [ + 790, + 430 + ], + "flags": {}, + "order": 15, + "mode": 4, + "inputs": [ + { + "localized_name": "mask", + "name": "mask", + "type": "MASK", + "link": 340 + } + ], + "outputs": [], + "properties": { + "Node name for S&R": "MaskPreview", + "cnr_id": "comfy-core", + "ver": "0.21.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 142, + "type": "OpticalFlowLoader", + "pos": [ + -710, + 410 + ], + "size": [ + 320, + 90 + ], + "flags": {}, + "order": 21, + "mode": 0, + "inputs": [ + { + "localized_name": "model_name", + "name": "model_name", + "type": "COMBO", + "widget": { + "name": "model_name" + }, + "link": 404 + } + ], + "outputs": [ + { + "localized_name": "OPTICAL_FLOW", + "name": "OPTICAL_FLOW", + "type": "OPTICAL_FLOW", + "links": [ + 321 + ] + } + ], + "properties": { + "Node name for S&R": "OpticalFlowLoader", + "cnr_id": "comfy-core", + "ver": "0.21.0", + "models": [ + { + "name": "raft_large_C_T_SKHT_V2-ff5fadd5.safetensors", + "url": "https://huggingface.co/Comfy-Org/void-model/resolve/main/optical_flow/raft_large_C_T_SKHT_V2-ff5fadd5.safetensors", + "directory": "optical_flow" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "raft_large_C_T_SKHT_V2-ff5fadd5.safetensors" + ] + }, + { + "id": 10, + "type": "VOIDInpaintConditioning", + "pos": [ + -110, + 430 + ], + "size": [ + 300, + 280 + ], + "flags": {}, + "order": 6, + "mode": 0, + "inputs": [ + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "link": 8 + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "link": 9 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 4 + }, + { + "localized_name": "video", + "name": "video", + "type": "IMAGE", + "link": 326 + }, + { + "localized_name": "quadmask", + "name": "quadmask", + "type": "MASK", + "link": 339 + }, + { + "localized_name": "width", + "name": "width", + "type": "INT", + "widget": { + "name": "width" + }, + "link": 332 + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "widget": { + "name": "height" + }, + "link": 334 + }, + { + "localized_name": "length", + "name": "length", + "type": "INT", + "widget": { + "name": "length" + }, + "link": 63 + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "widget": { + "name": "batch_size" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "slot_index": 0, + "links": [ + 309, + 317 + ] + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "slot_index": 1, + "links": [ + 310, + 318 + ] + }, + { + "localized_name": "latent", + "name": "latent", + "type": "LATENT", + "slot_index": 2, + "links": [ + 48, + 82 + ] + } + ], + "properties": { + "Node name for S&R": "VOIDInpaintConditioning", + "cnr_id": "comfy-core", + "ver": "0.18.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 672, + 384, + 45, + 1 + ] + }, + { + "id": 32, + "type": "VOIDWarpedNoiseSource", + "pos": [ + 410, + 1350 + ], + "size": [ + 300, + 50 + ], + "flags": {}, + "order": 8, + "mode": 0, + "inputs": [ + { + "localized_name": "warped_noise", + "name": "warped_noise", + "type": "LATENT", + "link": 53 + } + ], + "outputs": [ + { + "localized_name": "NOISE", + "name": "NOISE", + "type": "NOISE", + "slot_index": 0, + "links": [ + 54 + ] + } + ], + "title": "Warped Noise → NOISE", + "properties": { + "Node name for S&R": "VOIDWarpedNoiseSource", + "cnr_id": "comfy-core", + "ver": "0.18.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 137, + "type": "BasicScheduler", + "pos": [ + 410, + 1470 + ], + "size": [ + 300, + 150 + ], + "flags": {}, + "order": 17, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 323 + }, + { + "localized_name": "scheduler", + "name": "scheduler", + "type": "COMBO", + "widget": { + "name": "scheduler" + }, + "link": null + }, + { + "localized_name": "steps", + "name": "steps", + "type": "INT", + "widget": { + "name": "steps" + }, + "link": null + }, + { + "localized_name": "denoise", + "name": "denoise", + "type": "FLOAT", + "widget": { + "name": "denoise" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "SIGMAS", + "name": "SIGMAS", + "type": "SIGMAS", + "links": [ + 313 + ] + } + ], + "properties": { + "Node name for S&R": "BasicScheduler", + "cnr_id": "comfy-core", + "ver": "0.21.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "simple", + 30, + 1 + ] + }, + { + "id": 134, + "type": "VOIDSampler", + "pos": [ + 410, + 1800 + ], + "size": [ + 300, + 50 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [], + "outputs": [ + { + "localized_name": "SAMPLER", + "name": "SAMPLER", + "type": "SAMPLER", + "links": [ + 305 + ] + } + ], + "properties": { + "Node name for S&R": "VOIDSampler", + "cnr_id": "comfy-core", + "ver": "0.21.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 143, + "type": "UNETLoader", + "pos": [ + -710, + 550 + ], + "size": [ + 320, + 120 + ], + "flags": {}, + "order": 22, + "mode": 0, + "inputs": [ + { + "localized_name": "unet_name", + "name": "unet_name", + "type": "COMBO", + "widget": { + "name": "unet_name" + }, + "link": 403 + }, + { + "localized_name": "weight_dtype", + "name": "weight_dtype", + "type": "COMBO", + "widget": { + "name": "weight_dtype" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "links": [ + 322, + 323 + ] + } + ], + "properties": { + "Node name for S&R": "UNETLoader", + "cnr_id": "comfy-core", + "ver": "0.21.0", + "models": [ + { + "name": "void_pass2.safetensors", + "url": "https://huggingface.co/Comfy-Org/void-model/resolve/main/diffusion_models/void_pass2.safetensors", + "directory": "diffusion_models" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "void_pass2.safetensors", + "default" + ] + }, + { + "id": 144, + "type": "UNETLoader", + "pos": [ + -720, + -150 + ], + "size": [ + 320, + 120 + ], + "flags": {}, + "order": 23, + "mode": 0, + "inputs": [ + { + "localized_name": "unet_name", + "name": "unet_name", + "type": "COMBO", + "widget": { + "name": "unet_name" + }, + "link": 402 + }, + { + "localized_name": "weight_dtype", + "name": "weight_dtype", + "type": "COMBO", + "widget": { + "name": "weight_dtype" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "links": [ + 324, + 325 + ] + } + ], + "properties": { + "Node name for S&R": "UNETLoader", + "cnr_id": "comfy-core", + "ver": "0.21.0", + "models": [ + { + "name": "void_pass1.safetensors", + "url": "https://huggingface.co/Comfy-Org/void-model/resolve/main/diffusion_models/void_pass1.safetensors", + "directory": "diffusion_models" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "void_pass1.safetensors", + "default" + ] + }, + { + "id": 46, + "type": "CreateVideo", + "pos": [ + 1230, + -20 + ], + "size": [ + 240, + 110 + ], + "flags": {}, + "order": 13, + "mode": 0, + "inputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "link": 73 + }, + { + "localized_name": "audio", + "name": "audio", + "shape": 7, + "type": "AUDIO", + "link": 355 + }, + { + "localized_name": "fps", + "name": "fps", + "type": "FLOAT", + "widget": { + "name": "fps" + }, + "link": 368 + } + ], + "outputs": [ + { + "localized_name": "VIDEO", + "name": "VIDEO", + "type": "VIDEO", + "links": [ + 77 + ] + } + ], + "properties": { + "Node name for S&R": "CreateVideo", + "cnr_id": "comfy-core", + "ver": "0.21.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 30 + ] + }, + { + "id": 133, + "type": "VOIDSampler", + "pos": [ + 410, + 370 + ], + "size": [ + 280, + 50 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [], + "outputs": [ + { + "localized_name": "SAMPLER", + "name": "SAMPLER", + "type": "SAMPLER", + "links": [ + 304 + ] + } + ], + "properties": { + "Node name for S&R": "VOIDSampler", + "cnr_id": "comfy-core", + "ver": "0.21.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 49, + "type": "SamplerCustomAdvanced", + "pos": [ + 880, + -180 + ], + "size": [ + 250, + 270 + ], + "flags": {}, + "order": 14, + "mode": 0, + "inputs": [ + { + "localized_name": "noise", + "name": "noise", + "type": "NOISE", + "link": 320 + }, + { + "localized_name": "guider", + "name": "guider", + "type": "GUIDER", + "link": 319 + }, + { + "localized_name": "sampler", + "name": "sampler", + "type": "SAMPLER", + "link": 304 + }, + { + "localized_name": "sigmas", + "name": "sigmas", + "type": "SIGMAS", + "link": 315 + }, + { + "localized_name": "latent_image", + "name": "latent_image", + "type": "LATENT", + "link": 82 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "LATENT", + "links": [ + 83 + ] + }, + { + "localized_name": "denoised_output", + "name": "denoised_output", + "type": "LATENT", + "links": null + } + ], + "properties": { + "Node name for S&R": "SamplerCustomAdvanced", + "cnr_id": "comfy-core", + "ver": "0.21.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 45, + "type": "VAEDecode", + "pos": [ + 1230, + -180 + ], + "size": [ + 230, + 80 + ], + "flags": {}, + "order": 12, + "mode": 0, + "inputs": [ + { + "localized_name": "samples", + "name": "samples", + "type": "LATENT", + "link": 83 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 70 + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 72, + 73, + 342 + ] + } + ], + "properties": { + "Node name for S&R": "VAEDecode", + "cnr_id": "comfy-core", + "ver": "0.21.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 6, + "type": "CLIPTextEncode", + "pos": [ + -260, + -180 + ], + "size": [ + 580, + 310 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 2 + }, + { + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": 377 + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "slot_index": 0, + "links": [ + 8 + ] + } + ], + "title": "Positive Prompt", + "properties": { + "Node name for S&R": "CLIPTextEncode", + "cnr_id": "comfy-core", + "ver": "0.18.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "" + ], + "color": "#232", + "bgcolor": "#353" + }, + { + "id": 145, + "type": "ImageFromBatch", + "pos": [ + -410, + 850 + ], + "size": [ + 230, + 120 + ], + "flags": {}, + "order": 24, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 366 + }, + { + "localized_name": "batch_index", + "name": "batch_index", + "type": "INT", + "widget": { + "name": "batch_index" + }, + "link": 384 + }, + { + "localized_name": "length", + "name": "length", + "type": "INT", + "widget": { + "name": "length" + }, + "link": 361 + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 326, + 327, + 336 + ] + } + ], + "properties": { + "Node name for S&R": "ImageFromBatch", + "cnr_id": "comfy-core", + "ver": "0.21.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 0, + 197 + ] + }, + { + "id": 36, + "type": "VAEDecode", + "pos": [ + 1220, + 1110 + ], + "size": [ + 230, + 80 + ], + "flags": {}, + "order": 10, + "mode": 0, + "inputs": [ + { + "localized_name": "samples", + "name": "samples", + "type": "LATENT", + "link": 49 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 45 + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "slot_index": 0, + "links": [ + 341 + ] + } + ], + "title": "Pass 2 VAE Decode", + "properties": { + "Node name for S&R": "VAEDecode", + "cnr_id": "comfy-core", + "ver": "0.18.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 149, + "type": "c3e0d783-9aa3-4e75-a94d-19937968ef86", + "pos": [ + -20, + 840 + ], + "size": [ + 290, + 370 + ], + "flags": {}, + "order": 27, + "mode": 0, + "inputs": [ + { + "label": "image", + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 336 + }, + { + "label": "object", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": 388 + }, + { + "name": "bboxes", + "shape": 7, + "type": "BOUNDING_BOX", + "link": null + }, + { + "name": "positive_coords", + "shape": 7, + "type": "STRING", + "link": null + }, + { + "name": "negative_coords", + "shape": 7, + "type": "STRING", + "link": null + }, + { + "name": "threshold", + "type": "FLOAT", + "widget": { + "name": "threshold" + }, + "link": null + }, + { + "name": "refine_iterations", + "type": "INT", + "widget": { + "name": "refine_iterations" + }, + "link": null + }, + { + "name": "individual_masks", + "type": "BOOLEAN", + "widget": { + "name": "individual_masks" + }, + "link": null + }, + { + "name": "ckpt_name", + "type": "COMBO", + "widget": { + "name": "ckpt_name" + }, + "link": 401 + } + ], + "outputs": [ + { + "localized_name": "masks", + "name": "masks", + "type": "MASK", + "links": [ + 339, + 340 + ] + }, + { + "localized_name": "bboxes", + "name": "bboxes", + "type": "BOUNDING_BOX", + "links": [] + } + ], + "properties": { + "proxyWidgets": [ + [ + "78", + "text" + ], + [ + "75", + "threshold" + ], + [ + "75", + "refine_iterations" + ], + [ + "75", + "individual_masks" + ], + [ + "77", + "ckpt_name" + ] + ], + "cnr_id": "comfy-core", + "ver": "0.19.3", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "ue_properties": { + "widget_ue_connectable": { + "text": true + }, + "version": "7.7", + "input_ue_unconnectable": {} + } + }, + "widgets_values": [] + }, + { + "id": 43, + "type": "GetImageSize", + "pos": [ + -410, + 1140 + ], + "size": [ + 230, + 160 + ], + "flags": { + "collapsed": false + }, + "order": 11, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 327 + } + ], + "outputs": [ + { + "localized_name": "width", + "name": "width", + "type": "INT", + "links": null + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "links": null + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "links": [ + 63, + 67 + ] + } + ], + "properties": { + "Node name for S&R": "GetImageSize", + "cnr_id": "comfy-core", + "ver": "0.20.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 147, + "type": "PrimitiveInt", + "pos": [ + -570, + 1660 + ], + "size": [ + 270, + 90 + ], + "flags": {}, + "order": 25, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "INT", + "widget": { + "name": "value" + }, + "link": 391 + } + ], + "outputs": [ + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 332, + 333 + ] + } + ], + "title": "Int (Width)", + "properties": { + "Node name for S&R": "PrimitiveInt", + "cnr_id": "comfy-core", + "ver": "0.21.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 672, + "fixed" + ] + }, + { + "id": 148, + "type": "PrimitiveInt", + "pos": [ + -570, + 1790 + ], + "size": [ + 270, + 90 + ], + "flags": {}, + "order": 26, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "INT", + "widget": { + "name": "value" + }, + "link": 392 + } + ], + "outputs": [ + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 334, + 335 + ] + } + ], + "title": "Int (Height)", + "properties": { + "Node name for S&R": "PrimitiveInt", + "cnr_id": "comfy-core", + "ver": "0.21.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 384, + "fixed" + ] + }, + { + "id": 150, + "type": "ComfySwitchNode", + "pos": [ + 1510, + 1080 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 28, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 342 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 341 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 346 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 363 + ] + } + ], + "properties": { + "Node name for S&R": "ComfySwitchNode", + "cnr_id": "comfy-core", + "ver": "0.21.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + false + ] + }, + { + "id": 153, + "type": "PrimitiveBoolean", + "pos": [ + -580, + 1440 + ], + "size": [ + 270, + 80 + ], + "flags": {}, + "order": 29, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "BOOLEAN", + "widget": { + "name": "value" + }, + "link": 393 + } + ], + "outputs": [ + { + "localized_name": "BOOLEAN", + "name": "BOOLEAN", + "type": "BOOLEAN", + "links": [ + 346 + ] + } + ], + "title": "Boolean (Skip Pass 2?)", + "properties": { + "Node name for S&R": "PrimitiveBoolean", + "cnr_id": "comfy-core", + "ver": "0.21.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + false + ] + }, + { + "id": 158, + "type": "TrimAudioDuration", + "pos": [ + -10, + 1580 + ], + "size": [ + 270, + 120 + ], + "flags": {}, + "order": 30, + "mode": 0, + "inputs": [ + { + "localized_name": "audio", + "name": "audio", + "type": "AUDIO", + "link": 367 + }, + { + "localized_name": "start_index", + "name": "start_index", + "type": "FLOAT", + "widget": { + "name": "start_index" + }, + "link": 386 + }, + { + "localized_name": "duration", + "name": "duration", + "type": "FLOAT", + "widget": { + "name": "duration" + }, + "link": 385 + } + ], + "outputs": [ + { + "localized_name": "AUDIO", + "name": "AUDIO", + "type": "AUDIO", + "links": [ + 355, + 364 + ] + } + ], + "properties": { + "Node name for S&R": "TrimAudioDuration", + "cnr_id": "comfy-core", + "ver": "0.21.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 0, + 60 + ] + }, + { + "id": 163, + "type": "PrimitiveInt", + "pos": [ + -740, + 1170 + ], + "size": [ + 230, + 90 + ], + "flags": {}, + "order": 31, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "INT", + "widget": { + "name": "value" + }, + "link": 390 + } + ], + "outputs": [ + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 360 + ] + } + ], + "title": "Int (Video duration)", + "properties": { + "Node name for S&R": "PrimitiveInt", + "cnr_id": "comfy-core", + "ver": "0.21.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 5, + "fixed" + ] + }, + { + "id": 164, + "type": "ComfyMathExpression", + "pos": [ + -740, + 1300 + ], + "size": [ + 230, + 100 + ], + "flags": { + "collapsed": true + }, + "order": 32, + "mode": 0, + "inputs": [ + { + "label": "a", + "localized_name": "values.a", + "name": "values.a", + "type": "FLOAT,INT,BOOLEAN", + "link": 360 + }, + { + "label": "b", + "localized_name": "values.b", + "name": "values.b", + "shape": 7, + "type": "FLOAT,INT,BOOLEAN", + "link": 371 + }, + { + "label": "c", + "localized_name": "values.c", + "name": "values.c", + "shape": 7, + "type": "FLOAT,INT,BOOLEAN", + "link": null + }, + { + "localized_name": "expression", + "name": "expression", + "type": "STRING", + "widget": { + "name": "expression" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "FLOAT", + "name": "FLOAT", + "type": "FLOAT", + "links": [ + 385 + ] + }, + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 361 + ] + }, + { + "localized_name": "BOOL", + "name": "BOOL", + "type": "BOOLEAN", + "links": null + } + ], + "properties": { + "Node name for S&R": "ComfyMathExpression" + }, + "widgets_values": [ + "a * b" + ] + }, + { + "id": 165, + "type": "CreateVideo", + "pos": [ + 1510, + 1270 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 33, + "mode": 0, + "inputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "link": 363 + }, + { + "localized_name": "audio", + "name": "audio", + "shape": 7, + "type": "AUDIO", + "link": 364 + }, + { + "localized_name": "fps", + "name": "fps", + "type": "FLOAT", + "widget": { + "name": "fps" + }, + "link": 372 + } + ], + "outputs": [ + { + "localized_name": "VIDEO", + "name": "VIDEO", + "type": "VIDEO", + "links": [ + 362 + ] + } + ], + "properties": { + "Node name for S&R": "CreateVideo" + }, + "widgets_values": [ + 24 + ] + }, + { + "id": 166, + "type": "GetVideoComponents", + "pos": [ + -740, + 840 + ], + "size": [ + 230, + 100 + ], + "flags": {}, + "order": 34, + "mode": 0, + "inputs": [ + { + "localized_name": "video", + "name": "video", + "type": "VIDEO", + "link": 373 + } + ], + "outputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "links": [ + 366 + ] + }, + { + "localized_name": "audio", + "name": "audio", + "type": "AUDIO", + "links": [ + 367 + ] + }, + { + "localized_name": "fps", + "name": "fps", + "type": "FLOAT", + "links": [ + 368, + 371, + 372, + 383 + ] + } + ], + "properties": { + "Node name for S&R": "GetVideoComponents" + } + }, + { + "id": 168, + "type": "PrimitiveInt", + "pos": [ + -740, + 980 + ], + "size": [ + 230, + 90 + ], + "flags": {}, + "order": 35, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "INT", + "widget": { + "name": "value" + }, + "link": 389 + } + ], + "outputs": [ + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 382 + ] + } + ], + "title": "Int (Index)", + "properties": { + "Node name for S&R": "PrimitiveInt", + "cnr_id": "comfy-core", + "ver": "0.21.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 0, + "fixed" + ] + }, + { + "id": 169, + "type": "ComfyMathExpression", + "pos": [ + -740, + 1110 + ], + "size": [ + 230, + 100 + ], + "flags": { + "collapsed": true + }, + "order": 36, + "mode": 0, + "inputs": [ + { + "label": "a", + "localized_name": "values.a", + "name": "values.a", + "type": "FLOAT,INT,BOOLEAN", + "link": 382 + }, + { + "label": "b", + "localized_name": "values.b", + "name": "values.b", + "shape": 7, + "type": "FLOAT,INT,BOOLEAN", + "link": 383 + }, + { + "label": "c", + "localized_name": "values.c", + "name": "values.c", + "shape": 7, + "type": "FLOAT,INT,BOOLEAN", + "link": null + }, + { + "localized_name": "expression", + "name": "expression", + "type": "STRING", + "widget": { + "name": "expression" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "FLOAT", + "name": "FLOAT", + "type": "FLOAT", + "links": [ + 386 + ] + }, + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 384 + ] + }, + { + "localized_name": "BOOL", + "name": "BOOL", + "type": "BOOLEAN", + "links": null + } + ], + "properties": { + "Node name for S&R": "ComfyMathExpression" + }, + "widgets_values": [ + "a * b" + ] + } + ], + "groups": [ + { + "id": 1, + "title": "Models", + "bounding": [ + -790, + -260, + 470, + 990 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 2, + "title": "Input videos (place files in ComfyUI/input/)", + "bounding": [ + -790, + 760, + 660, + 560 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 3, + "title": "Shared: Text & Mask Conditioning", + "bounding": [ + -290, + -260, + 640, + 990 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 4, + "title": "Pass 1: Sample (Random Noise → DDIM)", + "bounding": [ + 380, + -260, + 810, + 750 + ], + "color": "#8A8", + "flags": {} + }, + { + "id": 6, + "title": "Pass 2: Sample (Warped Noise → DDIM)", + "bounding": [ + 380, + 1020, + 810, + 880 + ], + "color": "#8A8", + "flags": {} + }, + { + "id": 8, + "title": "Create Mask", + "bounding": [ + -100, + 760, + 450, + 560 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 9, + "title": "Pass 1", + "bounding": [ + -730, + -220, + 360, + 210 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 10, + "title": "Pass 2", + "bounding": [ + -720, + 340, + 340, + 340 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 11, + "title": "Output Video Size", + "bounding": [ + -790, + 1580, + 660, + 320 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 12, + "title": "Skip Pass 2", + "bounding": [ + -790, + 1350, + 660, + 200 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 13, + "title": "Trim Audio", + "bounding": [ + -100, + 1350, + 450, + 550 + ], + "color": "#3f789e", + "flags": {} + } + ], + "links": [ + { + "id": 3, + "origin_id": 2, + "origin_slot": 0, + "target_id": 7, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 322, + "origin_id": 143, + "origin_slot": 0, + "target_id": 136, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 309, + "origin_id": 10, + "origin_slot": 0, + "target_id": 136, + "target_slot": 1, + "type": "CONDITIONING" + }, + { + "id": 310, + "origin_id": 10, + "origin_slot": 1, + "target_id": 136, + "target_slot": 2, + "type": "CONDITIONING" + }, + { + "id": 324, + "origin_id": 144, + "origin_slot": 0, + "target_id": 138, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 325, + "origin_id": 144, + "origin_slot": 0, + "target_id": 140, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 317, + "origin_id": 10, + "origin_slot": 0, + "target_id": 140, + "target_slot": 1, + "type": "CONDITIONING" + }, + { + "id": 318, + "origin_id": 10, + "origin_slot": 1, + "target_id": 140, + "target_slot": 2, + "type": "CONDITIONING" + }, + { + "id": 321, + "origin_id": 142, + "origin_slot": 0, + "target_id": 31, + "target_slot": 0, + "type": "OPTICAL_FLOW" + }, + { + "id": 72, + "origin_id": 45, + "origin_slot": 0, + "target_id": 31, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 333, + "origin_id": 147, + "origin_slot": 0, + "target_id": 31, + "target_slot": 2, + "type": "INT" + }, + { + "id": 335, + "origin_id": 148, + "origin_slot": 0, + "target_id": 31, + "target_slot": 3, + "type": "INT" + }, + { + "id": 67, + "origin_id": 43, + "origin_slot": 2, + "target_id": 31, + "target_slot": 4, + "type": "INT" + }, + { + "id": 54, + "origin_id": 32, + "origin_slot": 0, + "target_id": 35, + "target_slot": 0, + "type": "NOISE" + }, + { + "id": 311, + "origin_id": 136, + "origin_slot": 0, + "target_id": 35, + "target_slot": 1, + "type": "GUIDER" + }, + { + "id": 305, + "origin_id": 134, + "origin_slot": 0, + "target_id": 35, + "target_slot": 2, + "type": "SAMPLER" + }, + { + "id": 313, + "origin_id": 137, + "origin_slot": 0, + "target_id": 35, + "target_slot": 3, + "type": "SIGMAS" + }, + { + "id": 48, + "origin_id": 10, + "origin_slot": 2, + "target_id": 35, + "target_slot": 4, + "type": "LATENT" + }, + { + "id": 340, + "origin_id": 149, + "origin_slot": 0, + "target_id": 132, + "target_slot": 0, + "type": "MASK" + }, + { + "id": 8, + "origin_id": 6, + "origin_slot": 0, + "target_id": 10, + "target_slot": 0, + "type": "CONDITIONING" + }, + { + "id": 9, + "origin_id": 7, + "origin_slot": 0, + "target_id": 10, + "target_slot": 1, + "type": "CONDITIONING" + }, + { + "id": 4, + "origin_id": 3, + "origin_slot": 0, + "target_id": 10, + "target_slot": 2, + "type": "VAE" + }, + { + "id": 326, + "origin_id": 145, + "origin_slot": 0, + "target_id": 10, + "target_slot": 3, + "type": "IMAGE" + }, + { + "id": 339, + "origin_id": 149, + "origin_slot": 0, + "target_id": 10, + "target_slot": 4, + "type": "MASK" + }, + { + "id": 332, + "origin_id": 147, + "origin_slot": 0, + "target_id": 10, + "target_slot": 5, + "type": "INT" + }, + { + "id": 334, + "origin_id": 148, + "origin_slot": 0, + "target_id": 10, + "target_slot": 6, + "type": "INT" + }, + { + "id": 63, + "origin_id": 43, + "origin_slot": 2, + "target_id": 10, + "target_slot": 7, + "type": "INT" + }, + { + "id": 53, + "origin_id": 31, + "origin_slot": 0, + "target_id": 32, + "target_slot": 0, + "type": "LATENT" + }, + { + "id": 323, + "origin_id": 143, + "origin_slot": 0, + "target_id": 137, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 73, + "origin_id": 45, + "origin_slot": 0, + "target_id": 46, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 355, + "origin_id": 158, + "origin_slot": 0, + "target_id": 46, + "target_slot": 1, + "type": "AUDIO" + }, + { + "id": 368, + "origin_id": 166, + "origin_slot": 2, + "target_id": 46, + "target_slot": 2, + "type": "FLOAT" + }, + { + "id": 320, + "origin_id": 141, + "origin_slot": 0, + "target_id": 49, + "target_slot": 0, + "type": "NOISE" + }, + { + "id": 319, + "origin_id": 140, + "origin_slot": 0, + "target_id": 49, + "target_slot": 1, + "type": "GUIDER" + }, + { + "id": 304, + "origin_id": 133, + "origin_slot": 0, + "target_id": 49, + "target_slot": 2, + "type": "SAMPLER" + }, + { + "id": 315, + "origin_id": 138, + "origin_slot": 0, + "target_id": 49, + "target_slot": 3, + "type": "SIGMAS" + }, + { + "id": 82, + "origin_id": 10, + "origin_slot": 2, + "target_id": 49, + "target_slot": 4, + "type": "LATENT" + }, + { + "id": 83, + "origin_id": 49, + "origin_slot": 0, + "target_id": 45, + "target_slot": 0, + "type": "LATENT" + }, + { + "id": 70, + "origin_id": 3, + "origin_slot": 0, + "target_id": 45, + "target_slot": 1, + "type": "VAE" + }, + { + "id": 2, + "origin_id": 2, + "origin_slot": 0, + "target_id": 6, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 366, + "origin_id": 166, + "origin_slot": 0, + "target_id": 145, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 361, + "origin_id": 164, + "origin_slot": 1, + "target_id": 145, + "target_slot": 2, + "type": "INT" + }, + { + "id": 49, + "origin_id": 35, + "origin_slot": 0, + "target_id": 36, + "target_slot": 0, + "type": "LATENT" + }, + { + "id": 45, + "origin_id": 3, + "origin_slot": 0, + "target_id": 36, + "target_slot": 1, + "type": "VAE" + }, + { + "id": 336, + "origin_id": 145, + "origin_slot": 0, + "target_id": 149, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 327, + "origin_id": 145, + "origin_slot": 0, + "target_id": 43, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 342, + "origin_id": 45, + "origin_slot": 0, + "target_id": 150, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 341, + "origin_id": 36, + "origin_slot": 0, + "target_id": 150, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 346, + "origin_id": 153, + "origin_slot": 0, + "target_id": 150, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 367, + "origin_id": 166, + "origin_slot": 1, + "target_id": 158, + "target_slot": 0, + "type": "AUDIO" + }, + { + "id": 360, + "origin_id": 163, + "origin_slot": 0, + "target_id": 164, + "target_slot": 0, + "type": "INT" + }, + { + "id": 371, + "origin_id": 166, + "origin_slot": 2, + "target_id": 164, + "target_slot": 1, + "type": "FLOAT" + }, + { + "id": 363, + "origin_id": 150, + "origin_slot": 0, + "target_id": 165, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 364, + "origin_id": 158, + "origin_slot": 0, + "target_id": 165, + "target_slot": 1, + "type": "AUDIO" + }, + { + "id": 372, + "origin_id": 166, + "origin_slot": 2, + "target_id": 165, + "target_slot": 2, + "type": "FLOAT" + }, + { + "id": 373, + "origin_id": -10, + "origin_slot": 0, + "target_id": 166, + "target_slot": 0, + "type": "VIDEO" + }, + { + "id": 77, + "origin_id": 46, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "VIDEO" + }, + { + "id": 362, + "origin_id": 165, + "origin_slot": 0, + "target_id": -20, + "target_slot": 1, + "type": "VIDEO" + }, + { + "id": 377, + "origin_id": -10, + "origin_slot": 1, + "target_id": 6, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 382, + "origin_id": 168, + "origin_slot": 0, + "target_id": 169, + "target_slot": 0, + "type": "INT" + }, + { + "id": 383, + "origin_id": 166, + "origin_slot": 2, + "target_id": 169, + "target_slot": 1, + "type": "FLOAT" + }, + { + "id": 384, + "origin_id": 169, + "origin_slot": 1, + "target_id": 145, + "target_slot": 1, + "type": "INT" + }, + { + "id": 385, + "origin_id": 164, + "origin_slot": 0, + "target_id": 158, + "target_slot": 2, + "type": "FLOAT" + }, + { + "id": 386, + "origin_id": 169, + "origin_slot": 0, + "target_id": 158, + "target_slot": 1, + "type": "FLOAT" + }, + { + "id": 387, + "origin_id": -10, + "origin_slot": 2, + "target_id": 7, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 388, + "origin_id": -10, + "origin_slot": 3, + "target_id": 149, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 389, + "origin_id": -10, + "origin_slot": 4, + "target_id": 168, + "target_slot": 0, + "type": "INT" + }, + { + "id": 390, + "origin_id": -10, + "origin_slot": 5, + "target_id": 163, + "target_slot": 0, + "type": "INT" + }, + { + "id": 391, + "origin_id": -10, + "origin_slot": 6, + "target_id": 147, + "target_slot": 0, + "type": "INT" + }, + { + "id": 392, + "origin_id": -10, + "origin_slot": 7, + "target_id": 148, + "target_slot": 0, + "type": "INT" + }, + { + "id": 393, + "origin_id": -10, + "origin_slot": 8, + "target_id": 153, + "target_slot": 0, + "type": "BOOLEAN" + }, + { + "id": 400, + "origin_id": -10, + "origin_slot": 9, + "target_id": 141, + "target_slot": 0, + "type": "INT" + }, + { + "id": 401, + "origin_id": -10, + "origin_slot": 10, + "target_id": 149, + "target_slot": 8, + "type": "COMBO" + }, + { + "id": 402, + "origin_id": -10, + "origin_slot": 11, + "target_id": 144, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 403, + "origin_id": -10, + "origin_slot": 12, + "target_id": 143, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 404, + "origin_id": -10, + "origin_slot": 13, + "target_id": 142, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 405, + "origin_id": -10, + "origin_slot": 14, + "target_id": 2, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 406, + "origin_id": -10, + "origin_slot": 15, + "target_id": 3, + "target_slot": 0, + "type": "COMBO" + } + ], + "extra": {}, + "category": "Video generation and editing/Inpaint video", + "description": "Removes objects from video by inpainting masked regions using VOID (CogVideoX), with SAM3 text-guided segmentation and optional two-pass optical-flow refinement." + }, + { + "id": "c3e0d783-9aa3-4e75-a94d-19937968ef86", + "version": 1, + "state": { + "lastGroupId": 13, + "lastNodeId": 171, + "lastLinkId": 406, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Image Segmentation (SAM3)", + "description": "Segments images into masks using Meta SAM3 from text prompts, points, or boxes.", + "inputNode": { + "id": -10, + "bounding": [ + -2260, + -3450, + 144.369140625, + 228 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + -1130, + -3305, + 128, + 88 + ] + }, + "inputs": [ + { + "id": "a6e75fa2-162a-4af0-a2fd-1e9c899a5ab6", + "name": "image", + "type": "IMAGE", + "linkIds": [ + 264 + ], + "localized_name": "image", + "label": "image", + "pos": [ + -2139.630859375, + -3426 + ] + }, + { + "id": "3cefd304-7631-4ff6-a5a0-5a0ffb120745", + "name": "text", + "type": "STRING", + "linkIds": [ + 265 + ], + "label": "object", + "pos": [ + -2139.630859375, + -3406 + ] + }, + { + "id": "1aec91c5-d8d2-441c-928c-49c14e7e80ed", + "name": "bboxes", + "type": "BOUNDING_BOX", + "linkIds": [ + 266 + ], + "pos": [ + -2139.630859375, + -3386 + ] + }, + { + "id": "1ec7ce1a-8257-4719-8a81-60ebc8a98899", + "name": "positive_coords", + "type": "STRING", + "linkIds": [ + 267 + ], + "pos": [ + -2139.630859375, + -3366 + ] + }, + { + "id": "c65f8b87-9bd7-48be-9fc2-823431e95019", + "name": "negative_coords", + "type": "STRING", + "linkIds": [ + 268 + ], + "pos": [ + -2139.630859375, + -3346 + ] + }, + { + "id": "bb4ba35a-ccfe-4c37-98e5-d9b0d69585fb", + "name": "threshold", + "type": "FLOAT", + "linkIds": [ + 269 + ], + "pos": [ + -2139.630859375, + -3326 + ] + }, + { + "id": "b1439668-b050-490b-a5dc-fc4052c55666", + "name": "refine_iterations", + "type": "INT", + "linkIds": [ + 270 + ], + "pos": [ + -2139.630859375, + -3306 + ] + }, + { + "id": "86e239e5-c098-4302-b54d-d42a38bc0f89", + "name": "individual_masks", + "type": "BOOLEAN", + "linkIds": [ + 271 + ], + "pos": [ + -2139.630859375, + -3286 + ] + }, + { + "id": "f9e0b9d4-b2f1-4907-a4a5-305656576706", + "name": "ckpt_name", + "type": "COMBO", + "linkIds": [ + 272 + ], + "pos": [ + -2139.630859375, + -3266 + ] + } + ], + "outputs": [ + { + "id": "ff50da09-1e59-4a58-9b7f-be1a00aa5913", + "name": "masks", + "type": "MASK", + "linkIds": [ + 231 + ], + "localized_name": "masks", + "pos": [ + -1106, + -3281 + ] + }, + { + "id": "8f622e40-8528-4078-b7d3-147e9f872194", + "name": "bboxes", + "type": "BOUNDING_BOX", + "linkIds": [ + 232 + ], + "localized_name": "bboxes", + "pos": [ + -1106, + -3261 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 75, + "type": "SAM3_Detect", + "pos": [ + -1470, + -3460 + ], + "size": [ + 270, + 260 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "label": "model", + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 237 + }, + { + "label": "image", + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 264 + }, + { + "label": "conditioning", + "localized_name": "conditioning", + "name": "conditioning", + "shape": 7, + "type": "CONDITIONING", + "link": 200 + }, + { + "label": "bboxes", + "localized_name": "bboxes", + "name": "bboxes", + "shape": 7, + "type": "BOUNDING_BOX", + "link": 266 + }, + { + "label": "positive_coords", + "localized_name": "positive_coords", + "name": "positive_coords", + "shape": 7, + "type": "STRING", + "link": 267 + }, + { + "label": "negative_coords", + "localized_name": "negative_coords", + "name": "negative_coords", + "shape": 7, + "type": "STRING", + "link": 268 + }, + { + "localized_name": "threshold", + "name": "threshold", + "type": "FLOAT", + "widget": { + "name": "threshold" + }, + "link": 269 + }, + { + "localized_name": "refine_iterations", + "name": "refine_iterations", + "type": "INT", + "widget": { + "name": "refine_iterations" + }, + "link": 270 + }, + { + "localized_name": "individual_masks", + "name": "individual_masks", + "type": "BOOLEAN", + "widget": { + "name": "individual_masks" + }, + "link": 271 + } + ], + "outputs": [ + { + "localized_name": "masks", + "name": "masks", + "type": "MASK", + "links": [ + 231 + ] + }, + { + "localized_name": "bboxes", + "name": "bboxes", + "type": "BOUNDING_BOX", + "links": [ + 232 + ] + } + ], + "properties": { + "Node name for S&R": "SAM3_Detect", + "cnr_id": "comfy-core", + "ver": "0.19.3", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + } + }, + "widgets_values": [ + 0.5, + 2, + false + ] + }, + { + "id": 77, + "type": "CheckpointLoaderSimple", + "pos": [ + -1970, + -3200 + ], + "size": [ + 330, + 140 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "ckpt_name", + "name": "ckpt_name", + "type": "COMBO", + "widget": { + "name": "ckpt_name" + }, + "link": 272 + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "links": [ + 237 + ] + }, + { + "localized_name": "CLIP", + "name": "CLIP", + "type": "CLIP", + "links": [ + 240 + ] + }, + { + "localized_name": "VAE", + "name": "VAE", + "type": "VAE", + "links": null + } + ], + "properties": { + "Node name for S&R": "CheckpointLoaderSimple", + "cnr_id": "comfy-core", + "ver": "0.19.3", + "models": [ + { + "name": "sam3.1_multiplex_fp16.safetensors", + "url": "https://huggingface.co/Comfy-Org/sam3.1/resolve/main/checkpoints/sam3.1_multiplex_fp16.safetensors", + "directory": "checkpoints" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + } + }, + "widgets_values": [ + "sam3.1_multiplex_fp16.safetensors" + ] + }, + { + "id": 78, + "type": "CLIPTextEncode", + "pos": [ + -2000, + -3000 + ], + "size": [ + 400, + 200 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 240 + }, + { + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": 265 + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "links": [ + 200 + ] + } + ], + "properties": { + "Node name for S&R": "CLIPTextEncode", + "cnr_id": "comfy-core", + "ver": "0.19.3", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + } + }, + "widgets_values": [ + "" + ] + } + ], + "groups": [], + "links": [ + { + "id": 237, + "origin_id": 77, + "origin_slot": 0, + "target_id": 75, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 200, + "origin_id": 78, + "origin_slot": 0, + "target_id": 75, + "target_slot": 2, + "type": "CONDITIONING" + }, + { + "id": 240, + "origin_id": 77, + "origin_slot": 1, + "target_id": 78, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 231, + "origin_id": 75, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "MASK" + }, + { + "id": 232, + "origin_id": 75, + "origin_slot": 1, + "target_id": -20, + "target_slot": 1, + "type": "BOUNDING_BOX" + }, + { + "id": 264, + "origin_id": -10, + "origin_slot": 0, + "target_id": 75, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 265, + "origin_id": -10, + "origin_slot": 1, + "target_id": 78, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 266, + "origin_id": -10, + "origin_slot": 2, + "target_id": 75, + "target_slot": 3, + "type": "BOUNDING_BOX" + }, + { + "id": 267, + "origin_id": -10, + "origin_slot": 3, + "target_id": 75, + "target_slot": 4, + "type": "STRING" + }, + { + "id": 268, + "origin_id": -10, + "origin_slot": 4, + "target_id": 75, + "target_slot": 5, + "type": "STRING" + }, + { + "id": 269, + "origin_id": -10, + "origin_slot": 5, + "target_id": 75, + "target_slot": 6, + "type": "FLOAT" + }, + { + "id": 270, + "origin_id": -10, + "origin_slot": 6, + "target_id": 75, + "target_slot": 7, + "type": "INT" + }, + { + "id": 271, + "origin_id": -10, + "origin_slot": 7, + "target_id": 75, + "target_slot": 8, + "type": "BOOLEAN" + }, + { + "id": 272, + "origin_id": -10, + "origin_slot": 8, + "target_id": 77, + "target_slot": 0, + "type": "COMBO" + } + ], + "extra": { + "ue_links": [] + } + } + ] + }, + "extra": {} +} \ No newline at end of file diff --git a/blueprints/Video Inpaint(Wan2.1 VACE).json b/blueprints/Video Inpaint(Wan2.1 VACE).json deleted file mode 100644 index a658be5f8..000000000 --- a/blueprints/Video Inpaint(Wan2.1 VACE).json +++ /dev/null @@ -1,2388 +0,0 @@ -{ - "id": "2f429c60-2e03-4117-908b-31e1fab04bba", - "revision": 0, - "last_node_id": 229, - "last_link_id": 366, - "nodes": [ - { - "id": 229, - "type": "53a657f3-c9eb-40f2-9ebd-1ed77d25ed67", - "pos": [ - -230, - 160 - ], - "size": [ - 400, - 480 - ], - "flags": {}, - "order": 0, - "mode": 0, - "inputs": [ - { - "label": "video mask", - "localized_name": "mask", - "name": "mask", - "type": "MASK", - "link": null - }, - { - "localized_name": "video", - "name": "video", - "type": "VIDEO", - "link": null - }, - { - "name": "width", - "type": "INT", - "widget": { - "name": "width" - }, - "link": null - }, - { - "name": "height", - "type": "INT", - "widget": { - "name": "height" - }, - "link": null - }, - { - "label": "reference image", - "name": "reference_image_1", - "type": "IMAGE", - "link": null - }, - { - "name": "unet_name", - "type": "COMBO", - "widget": { - "name": "unet_name" - }, - "link": null - }, - { - "name": "lora_name", - "type": "COMBO", - "widget": { - "name": "lora_name" - }, - "link": null - }, - { - "name": "clip_name", - "type": "COMBO", - "widget": { - "name": "clip_name" - }, - "link": null - }, - { - "name": "vae_name", - "type": "COMBO", - "widget": { - "name": "vae_name" - }, - "link": null - } - ], - "outputs": [ - { - "localized_name": "VIDEO", - "name": "VIDEO", - "type": "VIDEO", - "links": [] - } - ], - "properties": { - "proxyWidgets": [ - [ - "6", - "text" - ], - [ - "-1", - "width" - ], - [ - "-1", - "height" - ], - [ - "3", - "seed" - ], - [ - "3", - "control_after_generate" - ], - [ - "-1", - "unet_name" - ], - [ - "-1", - "lora_name" - ], - [ - "-1", - "clip_name" - ], - [ - "-1", - "vae_name" - ] - ], - "cnr_id": "comfy-core", - "ver": "0.13.0" - }, - "widgets_values": [ - null, - 720, - 720, - null, - null, - "wan2.1_vace_14B_fp16.safetensors", - "Wan21_CausVid_14B_T2V_lora_rank32.safetensors", - "umt5_xxl_fp8_e4m3fn_scaled.safetensors", - "wan_2.1_vae.safetensors" - ] - } - ], - "links": [], - "groups": [], - "definitions": { - "subgraphs": [ - { - "id": "53a657f3-c9eb-40f2-9ebd-1ed77d25ed67", - "version": 1, - "state": { - "lastGroupId": 25, - "lastNodeId": 229, - "lastLinkId": 366, - "lastRerouteId": 0 - }, - "revision": 0, - "config": {}, - "name": "Video Inpaint (Wan 2.1 VACE)", - "inputNode": { - "id": -10, - "bounding": [ - -970, - 800, - 132.54296875, - 220 - ] - }, - "outputNode": { - "id": -20, - "bounding": [ - 1480, - 535, - 120, - 60 - ] - }, - "inputs": [ - { - "id": "9fdda38d-6aa7-48ad-b425-f493d8aa585c", - "name": "mask", - "type": "MASK", - "linkIds": [ - 351, - 335, - 345 - ], - "localized_name": "mask", - "label": "video mask", - "pos": [ - -857.45703125, - 820 - ] - }, - { - "id": "8b1788cc-46d2-4f40-8b33-70fd56b4cb24", - "name": "video", - "type": "VIDEO", - "linkIds": [ - 336 - ], - "localized_name": "video", - "pos": [ - -857.45703125, - 840 - ] - }, - { - "id": "09393f21-257e-4476-bb02-54899a8252b8", - "name": "width", - "type": "INT", - "linkIds": [ - 355 - ], - "pos": [ - -857.45703125, - 860 - ] - }, - { - "id": "07a030f7-7eac-4b3f-b8f3-f00ee87b191d", - "name": "height", - "type": "INT", - "linkIds": [ - 356 - ], - "pos": [ - -857.45703125, - 880 - ] - }, - { - "id": "255908d3-6cc9-48fc-b76b-ab9fb72695bc", - "name": "reference_image_1", - "type": "IMAGE", - "linkIds": [ - 361 - ], - "label": "reference image", - "pos": [ - -857.45703125, - 900 - ] - }, - { - "id": "18a5d241-523c-433d-ae05-25b6e69d1e29", - "name": "unet_name", - "type": "COMBO", - "linkIds": [ - 363 - ], - "pos": [ - -857.45703125, - 920 - ] - }, - { - "id": "d7576e1b-da5f-402f-81b2-d37f838b1f8f", - "name": "lora_name", - "type": "COMBO", - "linkIds": [ - 364 - ], - "pos": [ - -857.45703125, - 940 - ] - }, - { - "id": "41676a3e-c710-4723-821e-f651ad3784b1", - "name": "clip_name", - "type": "COMBO", - "linkIds": [ - 365 - ], - "pos": [ - -857.45703125, - 960 - ] - }, - { - "id": "41fc878c-9aa6-4c12-bef3-ceda6b094b7c", - "name": "vae_name", - "type": "COMBO", - "linkIds": [ - 366 - ], - "pos": [ - -857.45703125, - 980 - ] - } - ], - "outputs": [ - { - "id": "d4861f39-1011-49dc-80fd-ee318b614a8d", - "name": "VIDEO", - "type": "VIDEO", - "linkIds": [ - 129 - ], - "localized_name": "VIDEO", - "pos": [ - 1500, - 555 - ] - } - ], - "widgets": [], - "nodes": [ - { - "id": 58, - "type": "TrimVideoLatent", - "pos": [ - 760, - 390 - ], - "size": [ - 315, - 60 - ], - "flags": { - "collapsed": false - }, - "order": 13, - "mode": 0, - "inputs": [ - { - "localized_name": "samples", - "name": "samples", - "type": "LATENT", - "link": 116 - }, - { - "localized_name": "trim_amount", - "name": "trim_amount", - "type": "INT", - "widget": { - "name": "trim_amount" - }, - "link": 115 - } - ], - "outputs": [ - { - "localized_name": "LATENT", - "name": "LATENT", - "type": "LATENT", - "links": [ - 117 - ] - } - ], - "properties": { - "cnr_id": "comfy-core", - "ver": "0.3.34", - "Node name for S&R": "TrimVideoLatent", - "enableTabs": false, - "tabWidth": 65, - "tabXOffset": 10, - "hasSecondTab": false, - "secondTabText": "Send Back", - "secondTabOffset": 80, - "secondTabWidth": 65, - "widget_ue_connectable": { - "trim_amount": true - } - }, - "widgets_values": [ - 0 - ] - }, - { - "id": 8, - "type": "VAEDecode", - "pos": [ - 770, - 500 - ], - "size": [ - 315, - 46 - ], - "flags": { - "collapsed": false - }, - "order": 11, - "mode": 0, - "inputs": [ - { - "localized_name": "samples", - "name": "samples", - "type": "LATENT", - "link": 117 - }, - { - "localized_name": "vae", - "name": "vae", - "type": "VAE", - "link": 76 - } - ], - "outputs": [ - { - "localized_name": "IMAGE", - "name": "IMAGE", - "type": "IMAGE", - "slot_index": 0, - "links": [ - 139 - ] - } - ], - "properties": { - "cnr_id": "comfy-core", - "ver": "0.3.34", - "Node name for S&R": "VAEDecode", - "enableTabs": false, - "tabWidth": 65, - "tabXOffset": 10, - "hasSecondTab": false, - "secondTabText": "Send Back", - "secondTabOffset": 80, - "secondTabWidth": 65, - "widget_ue_connectable": {} - }, - "widgets_values": [] - }, - { - "id": 48, - "type": "ModelSamplingSD3", - "pos": [ - 400, - 50 - ], - "size": [ - 315, - 58 - ], - "flags": {}, - "order": 9, - "mode": 0, - "inputs": [ - { - "localized_name": "model", - "name": "model", - "type": "MODEL", - "link": 279 - }, - { - "localized_name": "shift", - "name": "shift", - "type": "FLOAT", - "widget": { - "name": "shift" - }, - "link": null - } - ], - "outputs": [ - { - "localized_name": "MODEL", - "name": "MODEL", - "type": "MODEL", - "slot_index": 0, - "links": [ - 280 - ] - } - ], - "properties": { - "cnr_id": "comfy-core", - "ver": "0.3.34", - "Node name for S&R": "ModelSamplingSD3", - "enableTabs": false, - "tabWidth": 65, - "tabXOffset": 10, - "hasSecondTab": false, - "secondTabText": "Send Back", - "secondTabOffset": 80, - "secondTabWidth": 65, - "widget_ue_connectable": {} - }, - "widgets_values": [ - 5 - ] - }, - { - "id": 219, - "type": "InvertMask", - "pos": [ - 400, - 990 - ], - "size": [ - 140, - 26 - ], - "flags": {}, - "order": 24, - "mode": 0, - "inputs": [ - { - "localized_name": "mask", - "name": "mask", - "type": "MASK", - "link": 351 - } - ], - "outputs": [ - { - "localized_name": "MASK", - "name": "MASK", - "type": "MASK", - "links": [ - 352 - ] - } - ], - "properties": { - "cnr_id": "comfy-core", - "ver": "0.3.40", - "Node name for S&R": "InvertMask" - }, - "widgets_values": [] - }, - { - "id": 216, - "type": "MaskToImage", - "pos": [ - 560, - 990 - ], - "size": [ - 193.2779296875, - 26 - ], - "flags": {}, - "order": 23, - "mode": 0, - "inputs": [ - { - "localized_name": "mask", - "name": "mask", - "type": "MASK", - "link": 352 - } - ], - "outputs": [ - { - "localized_name": "IMAGE", - "name": "IMAGE", - "type": "IMAGE", - "links": [ - 334 - ] - } - ], - "properties": { - "cnr_id": "comfy-core", - "ver": "0.3.40", - "Node name for S&R": "MaskToImage" - }, - "widgets_values": [] - }, - { - "id": 213, - "type": "RebatchImages", - "pos": [ - 410, - 690 - ], - "size": [ - 230, - 60 - ], - "flags": {}, - "order": 21, - "mode": 0, - "inputs": [ - { - "localized_name": "images", - "name": "images", - "type": "IMAGE", - "link": 360 - }, - { - "localized_name": "batch_size", - "name": "batch_size", - "type": "INT", - "widget": { - "name": "batch_size" - }, - "link": 340 - } - ], - "outputs": [ - { - "localized_name": "IMAGE", - "name": "IMAGE", - "shape": 6, - "type": "IMAGE", - "links": [ - 333 - ] - } - ], - "properties": { - "cnr_id": "comfy-core", - "ver": "0.3.40", - "Node name for S&R": "RebatchImages" - }, - "widgets_values": [ - 1 - ] - }, - { - "id": 68, - "type": "CreateVideo", - "pos": [ - 1150, - 50 - ], - "size": [ - 270, - 78 - ], - "flags": { - "collapsed": false - }, - "order": 14, - "mode": 0, - "inputs": [ - { - "localized_name": "images", - "name": "images", - "type": "IMAGE", - "link": 139 - }, - { - "localized_name": "audio", - "name": "audio", - "shape": 7, - "type": "AUDIO", - "link": 362 - }, - { - "localized_name": "fps", - "name": "fps", - "type": "FLOAT", - "widget": { - "name": "fps" - }, - "link": 353 - } - ], - "outputs": [ - { - "localized_name": "VIDEO", - "name": "VIDEO", - "type": "VIDEO", - "links": [ - 129 - ] - } - ], - "properties": { - "cnr_id": "comfy-core", - "ver": "0.3.34", - "Node name for S&R": "CreateVideo", - "enableTabs": false, - "tabWidth": 65, - "tabXOffset": 10, - "hasSecondTab": false, - "secondTabText": "Send Back", - "secondTabOffset": 80, - "secondTabWidth": 65, - "widget_ue_connectable": {} - }, - "widgets_values": [ - 16 - ] - }, - { - "id": 208, - "type": "ImageCompositeMasked", - "pos": [ - 410, - 790 - ], - "size": [ - 230, - 146 - ], - "flags": {}, - "order": 18, - "mode": 0, - "inputs": [ - { - "localized_name": "destination", - "name": "destination", - "type": "IMAGE", - "link": 333 - }, - { - "localized_name": "source", - "name": "source", - "type": "IMAGE", - "link": 334 - }, - { - "localized_name": "mask", - "name": "mask", - "shape": 7, - "type": "MASK", - "link": 335 - }, - { - "localized_name": "x", - "name": "x", - "type": "INT", - "widget": { - "name": "x" - }, - "link": null - }, - { - "localized_name": "y", - "name": "y", - "type": "INT", - "widget": { - "name": "y" - }, - "link": null - }, - { - "localized_name": "resize_source", - "name": "resize_source", - "type": "BOOLEAN", - "widget": { - "name": "resize_source" - }, - "link": null - } - ], - "outputs": [ - { - "localized_name": "IMAGE", - "name": "IMAGE", - "type": "IMAGE", - "links": [ - 341, - 344 - ] - } - ], - "properties": { - "cnr_id": "comfy-core", - "ver": "0.3.40", - "Node name for S&R": "ImageCompositeMasked" - }, - "widgets_values": [ - 0, - 0, - true - ] - }, - { - "id": 214, - "type": "PreviewImage", - "pos": [ - 760, - 690 - ], - "size": [ - 300, - 300 - ], - "flags": {}, - "order": 22, - "mode": 0, - "inputs": [ - { - "localized_name": "images", - "name": "images", - "type": "IMAGE", - "link": 341 - } - ], - "outputs": [], - "properties": { - "cnr_id": "comfy-core", - "ver": "0.3.40", - "Node name for S&R": "PreviewImage" - }, - "widgets_values": [] - }, - { - "id": 111, - "type": "MaskToImage", - "pos": [ - 20, - 1270 - ], - "size": [ - 240, - 26 - ], - "flags": {}, - "order": 15, - "mode": 0, - "inputs": [ - { - "localized_name": "mask", - "name": "mask", - "type": "MASK", - "link": 345 - } - ], - "outputs": [ - { - "localized_name": "IMAGE", - "name": "IMAGE", - "type": "IMAGE", - "links": [ - 201 - ] - } - ], - "properties": { - "cnr_id": "comfy-core", - "ver": "0.3.34", - "Node name for S&R": "MaskToImage", - "enableTabs": false, - "tabWidth": 65, - "tabXOffset": 10, - "hasSecondTab": false, - "secondTabText": "Send Back", - "secondTabOffset": 80, - "secondTabWidth": 65, - "widget_ue_connectable": {} - }, - "widgets_values": [] - }, - { - "id": 129, - "type": "RepeatImageBatch", - "pos": [ - 20, - 1160 - ], - "size": [ - 240, - 60 - ], - "flags": {}, - "order": 16, - "mode": 0, - "inputs": [ - { - "localized_name": "image", - "name": "image", - "type": "IMAGE", - "link": 201 - }, - { - "localized_name": "amount", - "name": "amount", - "type": "INT", - "widget": { - "name": "amount" - }, - "link": 346 - } - ], - "outputs": [ - { - "localized_name": "IMAGE", - "name": "IMAGE", - "type": "IMAGE", - "links": [ - 202 - ] - } - ], - "properties": { - "cnr_id": "comfy-core", - "ver": "0.3.34", - "Node name for S&R": "RepeatImageBatch", - "enableTabs": false, - "tabWidth": 65, - "tabXOffset": 10, - "hasSecondTab": false, - "secondTabText": "Send Back", - "secondTabOffset": 80, - "secondTabWidth": 65, - "widget_ue_connectable": { - "amount": true - } - }, - "widgets_values": [ - 17 - ] - }, - { - "id": 130, - "type": "ImageToMask", - "pos": [ - 20, - 1050 - ], - "size": [ - 240, - 60 - ], - "flags": {}, - "order": 17, - "mode": 0, - "inputs": [ - { - "localized_name": "image", - "name": "image", - "type": "IMAGE", - "link": 202 - }, - { - "localized_name": "channel", - "name": "channel", - "type": "COMBO", - "widget": { - "name": "channel" - }, - "link": null - } - ], - "outputs": [ - { - "localized_name": "MASK", - "name": "MASK", - "type": "MASK", - "links": [ - 349 - ] - } - ], - "properties": { - "cnr_id": "comfy-core", - "ver": "0.3.34", - "Node name for S&R": "ImageToMask", - "enableTabs": false, - "tabWidth": 65, - "tabXOffset": 10, - "hasSecondTab": false, - "secondTabText": "Send Back", - "secondTabOffset": 80, - "secondTabWidth": 65, - "widget_ue_connectable": {} - }, - "widgets_values": [ - "red" - ] - }, - { - "id": 3, - "type": "KSampler", - "pos": [ - 770, - 50 - ], - "size": [ - 315, - 262 - ], - "flags": {}, - "order": 10, - "mode": 0, - "inputs": [ - { - "localized_name": "model", - "name": "model", - "type": "MODEL", - "link": 280 - }, - { - "localized_name": "positive", - "name": "positive", - "type": "CONDITIONING", - "link": 98 - }, - { - "localized_name": "negative", - "name": "negative", - "type": "CONDITIONING", - "link": 99 - }, - { - "localized_name": "latent_image", - "name": "latent_image", - "type": "LATENT", - "link": 160 - }, - { - "localized_name": "seed", - "name": "seed", - "type": "INT", - "widget": { - "name": "seed" - }, - "link": null - }, - { - "localized_name": "steps", - "name": "steps", - "type": "INT", - "widget": { - "name": "steps" - }, - "link": null - }, - { - "localized_name": "cfg", - "name": "cfg", - "type": "FLOAT", - "widget": { - "name": "cfg" - }, - "link": null - }, - { - "localized_name": "sampler_name", - "name": "sampler_name", - "type": "COMBO", - "widget": { - "name": "sampler_name" - }, - "link": null - }, - { - "localized_name": "scheduler", - "name": "scheduler", - "type": "COMBO", - "widget": { - "name": "scheduler" - }, - "link": null - }, - { - "localized_name": "denoise", - "name": "denoise", - "type": "FLOAT", - "widget": { - "name": "denoise" - }, - "link": null - } - ], - "outputs": [ - { - "localized_name": "LATENT", - "name": "LATENT", - "type": "LATENT", - "slot_index": 0, - "links": [ - 116 - ] - } - ], - "properties": { - "cnr_id": "comfy-core", - "ver": "0.3.34", - "Node name for S&R": "KSampler", - "enableTabs": false, - "tabWidth": 65, - "tabXOffset": 10, - "hasSecondTab": false, - "secondTabText": "Send Back", - "secondTabOffset": 80, - "secondTabWidth": 65, - "widget_ue_connectable": {} - }, - "widgets_values": [ - 584027519362099, - "randomize", - 4, - 1, - "uni_pc", - "simple", - 1 - ] - }, - { - "id": 224, - "type": "MarkdownNote", - "pos": [ - 420, - -160 - ], - "size": [ - 310, - 110 - ], - "flags": {}, - "order": 0, - "mode": 0, - "inputs": [], - "outputs": [], - "title": "About Video Size", - "properties": {}, - "widgets_values": [ - "| Model | 480P | 720P |\n| ------------------------------------------------------------ | ---- | ---- |\n| [VACE-1.3B](https://huggingface.co/Wan-AI/Wan2.1-VACE-1.3B) | ✅ | ❌ |\n| [VACE-14B](https://huggingface.co/Wan-AI/Wan2.1-VACE-14B) | ✅ | ✅ |" - ], - "color": "#432", - "bgcolor": "#000" - }, - { - "id": 223, - "type": "MarkdownNote", - "pos": [ - 770, - -210 - ], - "size": [ - 303.90106201171875, - 158.5415802001953 - ], - "flags": {}, - "order": 1, - "mode": 0, - "inputs": [], - "outputs": [], - "title": "KSampler Setting", - "properties": {}, - "widgets_values": [ - "## Default\n\n- steps:20\n- cfg:6.0\n\n## For CausVid LoRA\n\n- steps: 2-4\n- cfg: 1.0\n\n" - ], - "color": "#432", - "bgcolor": "#000" - }, - { - "id": 6, - "type": "CLIPTextEncode", - "pos": [ - -80, - 60 - ], - "size": [ - 420, - 280 - ], - "flags": {}, - "order": 7, - "mode": 0, - "inputs": [ - { - "localized_name": "clip", - "name": "clip", - "type": "CLIP", - "link": 74 - }, - { - "localized_name": "text", - "name": "text", - "type": "STRING", - "widget": { - "name": "text" - }, - "link": null - } - ], - "outputs": [ - { - "localized_name": "CONDITIONING", - "name": "CONDITIONING", - "type": "CONDITIONING", - "slot_index": 0, - "links": [ - 96 - ] - } - ], - "title": "CLIP Text Encode (Positive Prompt)", - "properties": { - "cnr_id": "comfy-core", - "ver": "0.3.34", - "Node name for S&R": "CLIPTextEncode", - "enableTabs": false, - "tabWidth": 65, - "tabXOffset": 10, - "hasSecondTab": false, - "secondTabText": "Send Back", - "secondTabOffset": 80, - "secondTabWidth": 65, - "widget_ue_connectable": {} - }, - "widgets_values": [ - "" - ], - "color": "#232", - "bgcolor": "#353" - }, - { - "id": 140, - "type": "UNETLoader", - "pos": [ - -505.8336486816406, - 88.22794342041016 - ], - "size": [ - 360, - 82 - ], - "flags": {}, - "order": 2, - "mode": 0, - "inputs": [ - { - "localized_name": "unet_name", - "name": "unet_name", - "type": "COMBO", - "widget": { - "name": "unet_name" - }, - "link": 363 - }, - { - "localized_name": "weight_dtype", - "name": "weight_dtype", - "type": "COMBO", - "widget": { - "name": "weight_dtype" - }, - "link": null - } - ], - "outputs": [ - { - "localized_name": "MODEL", - "name": "MODEL", - "type": "MODEL", - "slot_index": 0, - "links": [ - 248 - ] - } - ], - "properties": { - "cnr_id": "comfy-core", - "ver": "0.3.34", - "Node name for S&R": "UNETLoader", - "models": [ - { - "name": "wan2.1_vace_14B_fp16.safetensors", - "url": "https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/resolve/main/split_files/diffusion_models/wan2.1_vace_14B_fp16.safetensors", - "directory": "diffusion_models" - } - ], - "enableTabs": false, - "tabWidth": 65, - "tabXOffset": 10, - "hasSecondTab": false, - "secondTabText": "Send Back", - "secondTabOffset": 80, - "secondTabWidth": 65, - "widget_ue_connectable": {} - }, - "widgets_values": [ - "wan2.1_vace_14B_fp16.safetensors", - "fp8_e4m3fn_fast" - ] - }, - { - "id": 154, - "type": "LoraLoaderModelOnly", - "pos": [ - -505.8336486816406, - 228.2279510498047 - ], - "size": [ - 360, - 85.11004638671875 - ], - "flags": {}, - "order": 6, - "mode": 0, - "inputs": [ - { - "localized_name": "model", - "name": "model", - "type": "MODEL", - "link": 248 - }, - { - "localized_name": "lora_name", - "name": "lora_name", - "type": "COMBO", - "widget": { - "name": "lora_name" - }, - "link": 364 - }, - { - "localized_name": "strength_model", - "name": "strength_model", - "type": "FLOAT", - "widget": { - "name": "strength_model" - }, - "link": null - } - ], - "outputs": [ - { - "localized_name": "MODEL", - "name": "MODEL", - "type": "MODEL", - "links": [ - 279 - ] - } - ], - "properties": { - "cnr_id": "comfy-core", - "ver": "0.3.34", - "Node name for S&R": "LoraLoaderModelOnly", - "models": [ - { - "name": "Wan21_CausVid_14B_T2V_lora_rank32.safetensors", - "url": "https://huggingface.co/Kijai/WanVideo_comfy/resolve/main/Wan21_CausVid_14B_T2V_lora_rank32.safetensors", - "directory": "loras" - } - ], - "enableTabs": false, - "tabWidth": 65, - "tabXOffset": 10, - "hasSecondTab": false, - "secondTabText": "Send Back", - "secondTabOffset": 80, - "secondTabWidth": 65, - "widget_ue_connectable": {} - }, - "widgets_values": [ - "Wan21_CausVid_14B_T2V_lora_rank32.safetensors", - 0.30000000000000004 - ] - }, - { - "id": 38, - "type": "CLIPLoader", - "pos": [ - -499.14141845703125, - 368.0911865234375 - ], - "size": [ - 360, - 106 - ], - "flags": {}, - "order": 3, - "mode": 0, - "inputs": [ - { - "localized_name": "clip_name", - "name": "clip_name", - "type": "COMBO", - "widget": { - "name": "clip_name" - }, - "link": 365 - }, - { - "localized_name": "type", - "name": "type", - "type": "COMBO", - "widget": { - "name": "type" - }, - "link": null - }, - { - "localized_name": "device", - "name": "device", - "shape": 7, - "type": "COMBO", - "widget": { - "name": "device" - }, - "link": null - } - ], - "outputs": [ - { - "localized_name": "CLIP", - "name": "CLIP", - "type": "CLIP", - "slot_index": 0, - "links": [ - 74, - 75 - ] - } - ], - "properties": { - "cnr_id": "comfy-core", - "ver": "0.3.34", - "Node name for S&R": "CLIPLoader", - "models": [ - { - "name": "umt5_xxl_fp8_e4m3fn_scaled.safetensors", - "url": "https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/resolve/main/split_files/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors?download=true", - "directory": "text_encoders" - } - ], - "enableTabs": false, - "tabWidth": 65, - "tabXOffset": 10, - "hasSecondTab": false, - "secondTabText": "Send Back", - "secondTabOffset": 80, - "secondTabWidth": 65, - "widget_ue_connectable": {} - }, - "widgets_values": [ - "umt5_xxl_fp8_e4m3fn_scaled.safetensors", - "wan", - "default" - ] - }, - { - "id": 39, - "type": "VAELoader", - "pos": [ - -498.5298156738281, - 517.2576293945312 - ], - "size": [ - 360, - 60 - ], - "flags": {}, - "order": 4, - "mode": 0, - "inputs": [ - { - "localized_name": "vae_name", - "name": "vae_name", - "type": "COMBO", - "widget": { - "name": "vae_name" - }, - "link": 366 - } - ], - "outputs": [ - { - "localized_name": "VAE", - "name": "VAE", - "type": "VAE", - "slot_index": 0, - "links": [ - 76, - 101 - ] - } - ], - "properties": { - "cnr_id": "comfy-core", - "ver": "0.3.34", - "Node name for S&R": "VAELoader", - "models": [ - { - "name": "wan_2.1_vae.safetensors", - "url": "https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/resolve/main/split_files/vae/wan_2.1_vae.safetensors", - "directory": "vae" - } - ], - "enableTabs": false, - "tabWidth": 65, - "tabXOffset": 10, - "hasSecondTab": false, - "secondTabText": "Send Back", - "secondTabOffset": 80, - "secondTabWidth": 65, - "widget_ue_connectable": {} - }, - "widgets_values": [ - "wan_2.1_vae.safetensors" - ] - }, - { - "id": 221, - "type": "MarkdownNote", - "pos": [ - 380, - 1090 - ], - "size": [ - 480, - 170 - ], - "flags": {}, - "order": 5, - "mode": 0, - "inputs": [], - "outputs": [], - "title": "[EN] About video mask", - "properties": { - "widget_ue_connectable": {} - }, - "widgets_values": [ - "Currently, it's difficult to perfectly draw dynamic masks for different frames using only core nodes. However, to avoid requiring users to install additional custom nodes, our templates only use core nodes. You can refer to this implementation idea to achieve video inpainting.\n\nYou can use KJNode’s Points Editor and Sam2Segmentation to create some dynamic mask functions.\n\nCustom node links:\n- [ComfyUI-KJNodes](https://github.com/kijai/ComfyUI-KJNodes)\n- [ComfyUI-segment-anything-2](https://github.com/kijai/ComfyUI-segment-anything-2)" - ], - "color": "#432", - "bgcolor": "#000" - }, - { - "id": 7, - "type": "CLIPTextEncode", - "pos": [ - -80, - 390 - ], - "size": [ - 425.27801513671875, - 180.6060791015625 - ], - "flags": {}, - "order": 8, - "mode": 0, - "inputs": [ - { - "localized_name": "clip", - "name": "clip", - "type": "CLIP", - "link": 75 - }, - { - "localized_name": "text", - "name": "text", - "type": "STRING", - "widget": { - "name": "text" - }, - "link": null - } - ], - "outputs": [ - { - "localized_name": "CONDITIONING", - "name": "CONDITIONING", - "type": "CONDITIONING", - "slot_index": 0, - "links": [ - 97 - ] - } - ], - "title": "CLIP Text Encode (Negative Prompt)", - "properties": { - "cnr_id": "comfy-core", - "ver": "0.3.34", - "Node name for S&R": "CLIPTextEncode", - "enableTabs": false, - "tabWidth": 65, - "tabXOffset": 10, - "hasSecondTab": false, - "secondTabText": "Send Back", - "secondTabOffset": 80, - "secondTabWidth": 65, - "widget_ue_connectable": {} - }, - "widgets_values": [ - "过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走,过曝，" - ], - "color": "#223", - "bgcolor": "#335" - }, - { - "id": 229, - "type": "ImageFromBatch", - "pos": [ - -510, - 800 - ], - "size": [ - 270, - 82 - ], - "flags": {}, - "order": 25, - "mode": 0, - "inputs": [ - { - "localized_name": "image", - "name": "image", - "type": "IMAGE", - "link": 358 - }, - { - "localized_name": "batch_index", - "name": "batch_index", - "type": "INT", - "widget": { - "name": "batch_index" - }, - "link": null - }, - { - "localized_name": "length", - "name": "length", - "type": "INT", - "widget": { - "name": "length" - }, - "link": null - } - ], - "outputs": [ - { - "localized_name": "IMAGE", - "name": "IMAGE", - "type": "IMAGE", - "links": [ - 359, - 360 - ] - } - ], - "properties": { - "cnr_id": "comfy-core", - "ver": "0.13.0", - "Node name for S&R": "ImageFromBatch" - }, - "widgets_values": [ - 0, - 81 - ] - }, - { - "id": 49, - "type": "WanVaceToVideo", - "pos": [ - 400, - 200 - ], - "size": [ - 315, - 254 - ], - "flags": {}, - "order": 12, - "mode": 0, - "inputs": [ - { - "localized_name": "positive", - "name": "positive", - "type": "CONDITIONING", - "link": 96 - }, - { - "localized_name": "negative", - "name": "negative", - "type": "CONDITIONING", - "link": 97 - }, - { - "localized_name": "vae", - "name": "vae", - "type": "VAE", - "link": 101 - }, - { - "localized_name": "control_video", - "name": "control_video", - "shape": 7, - "type": "IMAGE", - "link": 344 - }, - { - "localized_name": "control_masks", - "name": "control_masks", - "shape": 7, - "type": "MASK", - "link": 349 - }, - { - "localized_name": "reference_image", - "name": "reference_image", - "shape": 7, - "type": "IMAGE", - "link": 361 - }, - { - "localized_name": "width", - "name": "width", - "type": "INT", - "widget": { - "name": "width" - }, - "link": 355 - }, - { - "localized_name": "height", - "name": "height", - "type": "INT", - "widget": { - "name": "height" - }, - "link": 356 - }, - { - "localized_name": "length", - "name": "length", - "type": "INT", - "widget": { - "name": "length" - }, - "link": null - }, - { - "localized_name": "batch_size", - "name": "batch_size", - "type": "INT", - "widget": { - "name": "batch_size" - }, - "link": null - }, - { - "localized_name": "strength", - "name": "strength", - "type": "FLOAT", - "widget": { - "name": "strength" - }, - "link": null - } - ], - "outputs": [ - { - "localized_name": "positive", - "name": "positive", - "type": "CONDITIONING", - "links": [ - 98 - ] - }, - { - "localized_name": "negative", - "name": "negative", - "type": "CONDITIONING", - "links": [ - 99 - ] - }, - { - "localized_name": "latent", - "name": "latent", - "type": "LATENT", - "links": [ - 160 - ] - }, - { - "localized_name": "trim_latent", - "name": "trim_latent", - "type": "INT", - "links": [ - 115 - ] - } - ], - "properties": { - "cnr_id": "comfy-core", - "ver": "0.3.34", - "Node name for S&R": "WanVaceToVideo", - "enableTabs": false, - "tabWidth": 65, - "tabXOffset": 10, - "hasSecondTab": false, - "secondTabText": "Send Back", - "secondTabOffset": 80, - "secondTabWidth": 65, - "widget_ue_connectable": { - "width": true, - "height": true, - "length": true - } - }, - "widgets_values": [ - 720, - 720, - 81, - 1, - 1 - ] - }, - { - "id": 211, - "type": "GetImageSize", - "pos": [ - 70, - 800 - ], - "size": [ - 190, - 66 - ], - "flags": { - "collapsed": false - }, - "order": 20, - "mode": 0, - "inputs": [ - { - "localized_name": "image", - "name": "image", - "type": "IMAGE", - "link": 359 - } - ], - "outputs": [ - { - "localized_name": "width", - "name": "width", - "type": "INT", - "links": null - }, - { - "localized_name": "height", - "name": "height", - "type": "INT", - "links": null - }, - { - "localized_name": "batch_size", - "name": "batch_size", - "type": "INT", - "links": [ - 340, - 346 - ] - } - ], - "properties": { - "cnr_id": "comfy-core", - "ver": "0.3.40", - "Node name for S&R": "GetImageSize" - }, - "widgets_values": [] - }, - { - "id": 210, - "type": "GetVideoComponents", - "pos": [ - -510, - 690 - ], - "size": [ - 193.530859375, - 66 - ], - "flags": {}, - "order": 19, - "mode": 0, - "inputs": [ - { - "localized_name": "video", - "name": "video", - "type": "VIDEO", - "link": 336 - } - ], - "outputs": [ - { - "localized_name": "images", - "name": "images", - "type": "IMAGE", - "links": [ - 358 - ] - }, - { - "localized_name": "audio", - "name": "audio", - "type": "AUDIO", - "links": [ - 362 - ] - }, - { - "localized_name": "fps", - "name": "fps", - "type": "FLOAT", - "links": [ - 353 - ] - } - ], - "properties": { - "cnr_id": "comfy-core", - "ver": "0.3.40", - "Node name for S&R": "GetVideoComponents" - }, - "widgets_values": [] - } - ], - "groups": [ - { - "id": 1, - "title": "Step1 - Load models here", - "bounding": [ - -540, - -30, - 430, - 620 - ], - "color": "#3f789e", - "font_size": 24, - "flags": {} - }, - { - "id": 2, - "title": "Prompt", - "bounding": [ - -90, - -30, - 450, - 620 - ], - "color": "#3f789e", - "font_size": 24, - "flags": {} - }, - { - "id": 3, - "title": "Sampling & Decoding", - "bounding": [ - 380, - -30, - 720, - 620 - ], - "color": "#3f789e", - "font_size": 24, - "flags": {} - }, - { - "id": 10, - "title": "Repeat Mask Batch", - "bounding": [ - -90, - 910, - 450, - 460 - ], - "color": "#3f789e", - "font_size": 24, - "flags": {} - }, - { - "id": 21, - "title": "Get video info", - "bounding": [ - -540, - 610, - 900, - 290 - ], - "color": "#3f789e", - "font_size": 24, - "flags": {} - }, - { - "id": 22, - "title": "Composite video & masks", - "bounding": [ - 380, - 610, - 720, - 420 - ], - "color": "#3f789e", - "font_size": 24, - "flags": {} - }, - { - "id": 23, - "title": "Step4 - Set video size & length", - "bounding": [ - 390, - 130, - 360, - 340 - ], - "color": "#A88", - "font_size": 24, - "flags": {} - }, - { - "id": 25, - "title": "14B", - "bounding": [ - -520, - 10, - 380, - 308.7100524902344 - ], - "color": "#3f789e", - "font_size": 24, - "flags": {} - } - ], - "links": [ - { - "id": 116, - "origin_id": 3, - "origin_slot": 0, - "target_id": 58, - "target_slot": 0, - "type": "LATENT" - }, - { - "id": 115, - "origin_id": 49, - "origin_slot": 3, - "target_id": 58, - "target_slot": 1, - "type": "INT" - }, - { - "id": 117, - "origin_id": 58, - "origin_slot": 0, - "target_id": 8, - "target_slot": 0, - "type": "LATENT" - }, - { - "id": 76, - "origin_id": 39, - "origin_slot": 0, - "target_id": 8, - "target_slot": 1, - "type": "VAE" - }, - { - "id": 279, - "origin_id": 154, - "origin_slot": 0, - "target_id": 48, - "target_slot": 0, - "type": "MODEL" - }, - { - "id": 352, - "origin_id": 219, - "origin_slot": 0, - "target_id": 216, - "target_slot": 0, - "type": "MASK" - }, - { - "id": 340, - "origin_id": 211, - "origin_slot": 2, - "target_id": 213, - "target_slot": 1, - "type": "INT" - }, - { - "id": 96, - "origin_id": 6, - "origin_slot": 0, - "target_id": 49, - "target_slot": 0, - "type": "CONDITIONING" - }, - { - "id": 97, - "origin_id": 7, - "origin_slot": 0, - "target_id": 49, - "target_slot": 1, - "type": "CONDITIONING" - }, - { - "id": 101, - "origin_id": 39, - "origin_slot": 0, - "target_id": 49, - "target_slot": 2, - "type": "VAE" - }, - { - "id": 344, - "origin_id": 208, - "origin_slot": 0, - "target_id": 49, - "target_slot": 3, - "type": "IMAGE" - }, - { - "id": 349, - "origin_id": 130, - "origin_slot": 0, - "target_id": 49, - "target_slot": 4, - "type": "MASK" - }, - { - "id": 139, - "origin_id": 8, - "origin_slot": 0, - "target_id": 68, - "target_slot": 0, - "type": "IMAGE" - }, - { - "id": 353, - "origin_id": 210, - "origin_slot": 2, - "target_id": 68, - "target_slot": 2, - "type": "FLOAT" - }, - { - "id": 333, - "origin_id": 213, - "origin_slot": 0, - "target_id": 208, - "target_slot": 0, - "type": "IMAGE" - }, - { - "id": 334, - "origin_id": 216, - "origin_slot": 0, - "target_id": 208, - "target_slot": 1, - "type": "IMAGE" - }, - { - "id": 341, - "origin_id": 208, - "origin_slot": 0, - "target_id": 214, - "target_slot": 0, - "type": "IMAGE" - }, - { - "id": 201, - "origin_id": 111, - "origin_slot": 0, - "target_id": 129, - "target_slot": 0, - "type": "IMAGE" - }, - { - "id": 346, - "origin_id": 211, - "origin_slot": 2, - "target_id": 129, - "target_slot": 1, - "type": "INT" - }, - { - "id": 202, - "origin_id": 129, - "origin_slot": 0, - "target_id": 130, - "target_slot": 0, - "type": "IMAGE" - }, - { - "id": 280, - "origin_id": 48, - "origin_slot": 0, - "target_id": 3, - "target_slot": 0, - "type": "MODEL" - }, - { - "id": 98, - "origin_id": 49, - "origin_slot": 0, - "target_id": 3, - "target_slot": 1, - "type": "CONDITIONING" - }, - { - "id": 99, - "origin_id": 49, - "origin_slot": 1, - "target_id": 3, - "target_slot": 2, - "type": "CONDITIONING" - }, - { - "id": 160, - "origin_id": 49, - "origin_slot": 2, - "target_id": 3, - "target_slot": 3, - "type": "LATENT" - }, - { - "id": 74, - "origin_id": 38, - "origin_slot": 0, - "target_id": 6, - "target_slot": 0, - "type": "CLIP" - }, - { - "id": 248, - "origin_id": 140, - "origin_slot": 0, - "target_id": 154, - "target_slot": 0, - "type": "MODEL" - }, - { - "id": 75, - "origin_id": 38, - "origin_slot": 0, - "target_id": 7, - "target_slot": 0, - "type": "CLIP" - }, - { - "id": 351, - "origin_id": -10, - "origin_slot": 0, - "target_id": 219, - "target_slot": 0, - "type": "MASK" - }, - { - "id": 335, - "origin_id": -10, - "origin_slot": 0, - "target_id": 208, - "target_slot": 2, - "type": "MASK" - }, - { - "id": 345, - "origin_id": -10, - "origin_slot": 0, - "target_id": 111, - "target_slot": 0, - "type": "MASK" - }, - { - "id": 336, - "origin_id": -10, - "origin_slot": 1, - "target_id": 210, - "target_slot": 0, - "type": "VIDEO" - }, - { - "id": 129, - "origin_id": 68, - "origin_slot": 0, - "target_id": -20, - "target_slot": 0, - "type": "VIDEO" - }, - { - "id": 355, - "origin_id": -10, - "origin_slot": 2, - "target_id": 49, - "target_slot": 6, - "type": "INT" - }, - { - "id": 356, - "origin_id": -10, - "origin_slot": 3, - "target_id": 49, - "target_slot": 7, - "type": "INT" - }, - { - "id": 358, - "origin_id": 210, - "origin_slot": 0, - "target_id": 229, - "target_slot": 0, - "type": "IMAGE" - }, - { - "id": 359, - "origin_id": 229, - "origin_slot": 0, - "target_id": 211, - "target_slot": 0, - "type": "IMAGE" - }, - { - "id": 360, - "origin_id": 229, - "origin_slot": 0, - "target_id": 213, - "target_slot": 0, - "type": "IMAGE" - }, - { - "id": 361, - "origin_id": -10, - "origin_slot": 4, - "target_id": 49, - "target_slot": 5, - "type": "IMAGE" - }, - { - "id": 362, - "origin_id": 210, - "origin_slot": 1, - "target_id": 68, - "target_slot": 1, - "type": "AUDIO" - }, - { - "id": 363, - "origin_id": -10, - "origin_slot": 5, - "target_id": 140, - "target_slot": 0, - "type": "COMBO" - }, - { - "id": 364, - "origin_id": -10, - "origin_slot": 6, - "target_id": 154, - "target_slot": 1, - "type": "COMBO" - }, - { - "id": 365, - "origin_id": -10, - "origin_slot": 7, - "target_id": 38, - "target_slot": 0, - "type": "COMBO" - }, - { - "id": 366, - "origin_id": -10, - "origin_slot": 8, - "target_id": 39, - "target_slot": 0, - "type": "COMBO" - } - ], - "extra": { - "workflowRendererVersion": "LG" - }, - "category": "Video generation and editing/Inpaint video", - "description": "Inpaints masked regions in video frames using Wan 2.1 VACE." - } - ] - }, - "config": {}, - "extra": { - "workflowRendererVersion": "LG", - "ds": { - "scale": 0.8183828377358485, - "offset": [ - 1215.8643989712405, - 178.87024992690183 - ] - } - }, - "version": 0.4 -} diff --git a/blueprints/Video Inpainting (Wan2.1 VACE).json b/blueprints/Video Inpainting (Wan2.1 VACE).json new file mode 100644 index 000000000..7460f3d44 --- /dev/null +++ b/blueprints/Video Inpainting (Wan2.1 VACE).json @@ -0,0 +1,4196 @@ +{ + "revision": 0, + "last_node_id": 306, + "last_link_id": 0, + "nodes": [ + { + "id": 306, + "type": "bd7f73a0-ec67-4f46-8671-17088d8e31b7", + "pos": [ + -2950, + -410 + ], + "size": [ + 440, + 650 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "label": "source_video", + "localized_name": "video", + "name": "video", + "type": "VIDEO", + "link": null + }, + { + "label": "reference_image", + "name": "reference_image_1", + "shape": 7, + "type": "IMAGE", + "link": null + }, + { + "label": "prompt", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": null + }, + { + "label": "width", + "name": "value", + "type": "INT", + "widget": { + "name": "value" + }, + "link": null + }, + { + "label": "height", + "name": "value_1", + "type": "INT", + "widget": { + "name": "value_1" + }, + "link": null + }, + { + "label": "frame_counts", + "name": "length", + "type": "INT", + "widget": { + "name": "length" + }, + "link": null + }, + { + "name": "seed", + "type": "INT", + "widget": { + "name": "seed" + }, + "link": null + }, + { + "label": "wan_vace_model", + "name": "unet_name", + "type": "COMBO", + "widget": { + "name": "unet_name" + }, + "link": null + }, + { + "label": "clip_model", + "name": "clip_name", + "type": "COMBO", + "widget": { + "name": "clip_name" + }, + "link": null + }, + { + "label": "vae_model", + "name": "vae_name", + "type": "COMBO", + "widget": { + "name": "vae_name" + }, + "link": null + }, + { + "label": "enable_turbo_mode", + "name": "value_2", + "type": "BOOLEAN", + "widget": { + "name": "value_2" + }, + "link": null + }, + { + "label": "lightning_lora", + "name": "lora_name", + "type": "COMBO", + "widget": { + "name": "lora_name" + }, + "link": null + }, + { + "label": "sam3_mask_object", + "name": "text_1", + "type": "STRING", + "widget": { + "name": "text_1" + }, + "link": null + }, + { + "label": "mask_expand", + "name": "expand", + "type": "INT", + "widget": { + "name": "expand" + }, + "link": null + }, + { + "label": "sam3_model", + "name": "ckpt_name", + "type": "COMBO", + "widget": { + "name": "ckpt_name" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "VIDEO", + "name": "VIDEO", + "type": "VIDEO", + "links": [] + } + ], + "properties": { + "proxyWidgets": [ + [ + "280", + "text" + ], + [ + "297", + "value" + ], + [ + "290", + "value" + ], + [ + "289", + "length" + ], + [ + "288", + "seed" + ], + [ + "299", + "unet_name" + ], + [ + "277", + "clip_name" + ], + [ + "278", + "vae_name" + ], + [ + "300", + "value" + ], + [ + "272", + "lora_name" + ], + [ + "268", + "text" + ], + [ + "269", + "expand" + ], + [ + "268", + "ckpt_name" + ], + [ + "312", + "$$canvas-image-preview" + ] + ], + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [], + "title": "Video Inpainting (Wan2.1 VACE)" + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "bd7f73a0-ec67-4f46-8671-17088d8e31b7", + "version": 1, + "state": { + "lastGroupId": 31, + "lastNodeId": 315, + "lastLinkId": 499, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Video Inpainting (Wan2.1 VACE)", + "inputNode": { + "id": -10, + "bounding": [ + -3450, + 3170, + 159.744140625, + 348 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + 900, + 2840, + 128, + 68 + ] + }, + "inputs": [ + { + "id": "a636746e-5b9f-4b91-96f0-7f2657415b93", + "name": "video", + "type": "VIDEO", + "linkIds": [ + 473 + ], + "localized_name": "video", + "label": "source_video", + "pos": [ + -3314.255859375, + 3194 + ] + }, + { + "id": "46275350-98b8-4d7c-8ca4-c452dc40a6bd", + "name": "reference_image_1", + "type": "IMAGE", + "linkIds": [ + 478 + ], + "label": "reference_image", + "pos": [ + -3314.255859375, + 3214 + ] + }, + { + "id": "0f5bee71-3485-4e10-81a7-2b9f85851353", + "name": "text", + "type": "STRING", + "linkIds": [ + 479 + ], + "label": "prompt", + "pos": [ + -3314.255859375, + 3234 + ] + }, + { + "id": "16675512-c229-43ed-944e-190a7f61b571", + "name": "value", + "type": "INT", + "linkIds": [ + 480 + ], + "label": "width", + "pos": [ + -3314.255859375, + 3254 + ] + }, + { + "id": "84330129-a0c7-44cd-91fe-c033946749db", + "name": "value_1", + "type": "INT", + "linkIds": [ + 481 + ], + "label": "height", + "pos": [ + -3314.255859375, + 3274 + ] + }, + { + "id": "3bd895e6-cba9-477b-bf6e-8c77dd56bb4a", + "name": "length", + "type": "INT", + "linkIds": [ + 494 + ], + "label": "frame_counts", + "pos": [ + -3314.255859375, + 3294 + ] + }, + { + "id": "dbc2e9c5-f86a-48ba-874a-2991c75d1ae7", + "name": "seed", + "type": "INT", + "linkIds": [ + 483 + ], + "pos": [ + -3314.255859375, + 3314 + ] + }, + { + "id": "572db94d-e64d-464f-bf3c-23a23aeb79f1", + "name": "unet_name", + "type": "COMBO", + "linkIds": [ + 485 + ], + "label": "wan_vace_model", + "pos": [ + -3314.255859375, + 3334 + ] + }, + { + "id": "32185180-f627-47c2-971b-6ef3007e9455", + "name": "clip_name", + "type": "COMBO", + "linkIds": [ + 486 + ], + "label": "clip_model", + "pos": [ + -3314.255859375, + 3354 + ] + }, + { + "id": "2af354d3-108a-42a9-acfc-7bad158715aa", + "name": "vae_name", + "type": "COMBO", + "linkIds": [ + 487 + ], + "label": "vae_model", + "pos": [ + -3314.255859375, + 3374 + ] + }, + { + "id": "c9777a8c-267f-4c5e-b4d5-e9727d822e50", + "name": "value_2", + "type": "BOOLEAN", + "linkIds": [ + 489 + ], + "label": "enable_turbo_mode", + "pos": [ + -3314.255859375, + 3394 + ] + }, + { + "id": "84a258a3-4f25-4edb-9f50-6fcd8411394e", + "name": "lora_name", + "type": "COMBO", + "linkIds": [ + 490 + ], + "label": "lightning_lora", + "pos": [ + -3314.255859375, + 3414 + ] + }, + { + "id": "9c5fb6f8-407b-4a13-94d8-cbbba546a082", + "name": "text_1", + "type": "STRING", + "linkIds": [ + 491 + ], + "label": "sam3_mask_object", + "pos": [ + -3314.255859375, + 3434 + ] + }, + { + "id": "598323c9-2256-44bd-9745-492a74628300", + "name": "expand", + "type": "INT", + "linkIds": [ + 496 + ], + "label": "mask_expand", + "pos": [ + -3314.255859375, + 3454 + ] + }, + { + "id": "856c1937-8caa-4d85-9d8a-6a900234d6d6", + "name": "ckpt_name", + "type": "COMBO", + "linkIds": [ + 497 + ], + "label": "sam3_model", + "pos": [ + -3314.255859375, + 3474 + ] + } + ], + "outputs": [ + { + "id": "be46c9d5-ced7-445b-996f-fff59d9b684d", + "name": "VIDEO", + "type": "VIDEO", + "linkIds": [ + 474 + ], + "localized_name": "VIDEO", + "pos": [ + 924, + 2864 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 266, + "type": "ModelSamplingSD3", + "pos": [ + -560, + 1940 + ], + "size": [ + 320, + 110 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 422 + }, + { + "localized_name": "shift", + "name": "shift", + "type": "FLOAT", + "widget": { + "name": "shift" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "slot_index": 0, + "links": [ + 454 + ] + } + ], + "properties": { + "Node name for S&R": "ModelSamplingSD3", + "cnr_id": "comfy-core", + "ver": "0.3.34", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "widget_ue_connectable": {} + }, + "widgets_values": [ + 5 + ] + }, + { + "id": 267, + "type": "CreateVideo", + "pos": [ + 530, + 2590 + ], + "size": [ + 310, + 130 + ], + "flags": { + "collapsed": false + }, + "order": 6, + "mode": 0, + "inputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "link": 423 + }, + { + "localized_name": "audio", + "name": "audio", + "shape": 7, + "type": "AUDIO", + "link": 424 + }, + { + "localized_name": "fps", + "name": "fps", + "type": "FLOAT", + "widget": { + "name": "fps" + }, + "link": 425 + } + ], + "outputs": [ + { + "localized_name": "VIDEO", + "name": "VIDEO", + "type": "VIDEO", + "links": [ + 474 + ] + } + ], + "properties": { + "Node name for S&R": "CreateVideo", + "cnr_id": "comfy-core", + "ver": "0.3.34", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "widget_ue_connectable": {} + }, + "widgets_values": [ + 16 + ] + }, + { + "id": 268, + "type": "17df2eeb-d89e-46ee-9480-a4ca2494b207", + "pos": [ + -1960, + 3220 + ], + "size": [ + 290, + 370 + ], + "flags": {}, + "order": 7, + "mode": 0, + "inputs": [ + { + "label": "image", + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 426 + }, + { + "label": "object", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": 491 + }, + { + "name": "bboxes", + "shape": 7, + "type": "BOUNDING_BOX", + "link": null + }, + { + "name": "positive_coords", + "shape": 7, + "type": "STRING", + "link": null + }, + { + "name": "negative_coords", + "shape": 7, + "type": "STRING", + "link": null + }, + { + "name": "threshold", + "type": "FLOAT", + "widget": { + "name": "threshold" + }, + "link": null + }, + { + "name": "refine_iterations", + "type": "INT", + "widget": { + "name": "refine_iterations" + }, + "link": null + }, + { + "name": "individual_masks", + "type": "BOOLEAN", + "widget": { + "name": "individual_masks" + }, + "link": null + }, + { + "name": "ckpt_name", + "type": "COMBO", + "widget": { + "name": "ckpt_name" + }, + "link": 497 + } + ], + "outputs": [ + { + "localized_name": "masks", + "name": "masks", + "type": "MASK", + "links": [ + 427 + ] + }, + { + "localized_name": "bboxes", + "name": "bboxes", + "type": "BOUNDING_BOX", + "links": [] + } + ], + "properties": { + "proxyWidgets": [ + [ + "237", + "text" + ], + [ + "75", + "threshold" + ], + [ + "75", + "refine_iterations" + ], + [ + "75", + "individual_masks" + ], + [ + "236", + "ckpt_name" + ] + ], + "cnr_id": "comfy-core", + "ver": "0.19.3", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "ue_properties": { + "widget_ue_connectable": { + "text": true + }, + "version": "7.7", + "input_ue_unconnectable": {} + } + }, + "widgets_values": [] + }, + { + "id": 269, + "type": "GrowMask", + "pos": [ + -1530, + 3220 + ], + "size": [ + 270, + 140 + ], + "flags": {}, + "order": 8, + "mode": 0, + "inputs": [ + { + "localized_name": "mask", + "name": "mask", + "type": "MASK", + "link": 427 + }, + { + "localized_name": "expand", + "name": "expand", + "type": "INT", + "widget": { + "name": "expand" + }, + "link": 496 + }, + { + "localized_name": "tapered_corners", + "name": "tapered_corners", + "type": "BOOLEAN", + "widget": { + "name": "tapered_corners" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MASK", + "name": "MASK", + "type": "MASK", + "links": [ + 441, + 445, + 449, + 498 + ] + } + ], + "properties": { + "Node name for S&R": "GrowMask", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 20, + true + ] + }, + { + "id": 270, + "type": "PrimitiveInt", + "pos": [ + -1350, + 1980 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "INT", + "widget": { + "name": "value" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 466 + ] + } + ], + "title": "Int （Steps）", + "properties": { + "Node name for S&R": "PrimitiveInt", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 20, + "fixed" + ] + }, + { + "id": 271, + "type": "PrimitiveFloat", + "pos": [ + -1340, + 2140 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "FLOAT", + "widget": { + "name": "value" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "FLOAT", + "name": "FLOAT", + "type": "FLOAT", + "links": [ + 432 + ] + } + ], + "title": "Float (CFG)", + "properties": { + "Node name for S&R": "PrimitiveFloat", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 6 + ] + }, + { + "id": 272, + "type": "LoraLoaderModelOnly", + "pos": [ + -1380, + 2390 + ], + "size": [ + 350, + 140 + ], + "flags": {}, + "order": 9, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 428 + }, + { + "localized_name": "lora_name", + "name": "lora_name", + "type": "COMBO", + "widget": { + "name": "lora_name" + }, + "link": 490 + }, + { + "localized_name": "strength_model", + "name": "strength_model", + "type": "FLOAT", + "widget": { + "name": "strength_model" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "links": [ + 430 + ] + } + ], + "properties": { + "Node name for S&R": "LoraLoaderModelOnly", + "cnr_id": "comfy-core", + "ver": "0.3.34", + "models": [ + { + "name": "Wan21_CausVid_14B_T2V_lora_rank32.safetensors", + "url": "https://huggingface.co/Kijai/WanVideo_comfy/resolve/main/Wan21_CausVid_14B_T2V_lora_rank32.safetensors", + "directory": "loras" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "widget_ue_connectable": {} + }, + "widgets_values": [ + "Wan21_CausVid_14B_T2V_lora_rank32.safetensors", + 0.30000000000000004 + ] + }, + { + "id": 273, + "type": "PrimitiveInt", + "pos": [ + -1340, + 2600 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "INT", + "widget": { + "name": "value" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 467 + ] + } + ], + "title": "Int (Steps)", + "properties": { + "Node name for S&R": "PrimitiveInt", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 6, + "fixed" + ] + }, + { + "id": 274, + "type": "PrimitiveFloat", + "pos": [ + -1340, + 2760 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "FLOAT", + "widget": { + "name": "value" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "FLOAT", + "name": "FLOAT", + "type": "FLOAT", + "links": [ + 433 + ] + } + ], + "title": "Float (CFG)", + "properties": { + "Node name for S&R": "PrimitiveFloat", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 1 + ] + }, + { + "id": 275, + "type": "ComfySwitchNode", + "pos": [ + -960, + 2530 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 10, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 429 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 430 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 431 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 422 + ] + } + ], + "properties": { + "Node name for S&R": "ComfySwitchNode", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + false + ] + }, + { + "id": 276, + "type": "ComfySwitchNode", + "pos": [ + -960, + 2340 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 11, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 432 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 433 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 434 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 459 + ] + } + ], + "properties": { + "Node name for S&R": "ComfySwitchNode", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + false + ] + }, + { + "id": 277, + "type": "CLIPLoader", + "pos": [ + -2710, + 2210 + ], + "size": [ + 360, + 170 + ], + "flags": {}, + "order": 12, + "mode": 0, + "inputs": [ + { + "localized_name": "clip_name", + "name": "clip_name", + "type": "COMBO", + "widget": { + "name": "clip_name" + }, + "link": 486 + }, + { + "localized_name": "type", + "name": "type", + "type": "COMBO", + "widget": { + "name": "type" + }, + "link": null + }, + { + "localized_name": "device", + "name": "device", + "shape": 7, + "type": "COMBO", + "widget": { + "name": "device" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "CLIP", + "name": "CLIP", + "type": "CLIP", + "slot_index": 0, + "links": [ + 435, + 436 + ] + } + ], + "properties": { + "Node name for S&R": "CLIPLoader", + "cnr_id": "comfy-core", + "ver": "0.3.34", + "models": [ + { + "name": "umt5_xxl_fp8_e4m3fn_scaled.safetensors", + "url": "https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/resolve/main/split_files/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors?download=true", + "directory": "text_encoders" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "widget_ue_connectable": {} + }, + "widgets_values": [ + "umt5_xxl_fp8_e4m3fn_scaled.safetensors", + "wan", + "default" + ] + }, + { + "id": 278, + "type": "VAELoader", + "pos": [ + -2700, + 2500 + ], + "size": [ + 360, + 110 + ], + "flags": {}, + "order": 13, + "mode": 0, + "inputs": [ + { + "localized_name": "vae_name", + "name": "vae_name", + "type": "COMBO", + "widget": { + "name": "vae_name" + }, + "link": 487 + } + ], + "outputs": [ + { + "localized_name": "VAE", + "name": "VAE", + "type": "VAE", + "slot_index": 0, + "links": [ + 439, + 471 + ] + } + ], + "properties": { + "Node name for S&R": "VAELoader", + "cnr_id": "comfy-core", + "ver": "0.3.34", + "models": [ + { + "name": "wan_2.1_vae.safetensors", + "url": "https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/resolve/main/split_files/vae/wan_2.1_vae.safetensors", + "directory": "vae" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "widget_ue_connectable": {} + }, + "widgets_values": [ + "wan_2.1_vae.safetensors" + ] + }, + { + "id": 279, + "type": "CLIPTextEncode", + "pos": [ + -2280, + 2410 + ], + "size": [ + 430, + 190 + ], + "flags": {}, + "order": 14, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 435 + }, + { + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "slot_index": 0, + "links": [ + 438 + ] + } + ], + "title": "CLIP Text Encode (Negative Prompt)", + "properties": { + "Node name for S&R": "CLIPTextEncode", + "cnr_id": "comfy-core", + "ver": "0.3.34", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "widget_ue_connectable": {} + }, + "widgets_values": [ + "过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走,过曝，" + ], + "color": "#223", + "bgcolor": "#335" + }, + { + "id": 280, + "type": "CLIPTextEncode", + "pos": [ + -2270, + 1940 + ], + "size": [ + 420, + 420 + ], + "flags": {}, + "order": 15, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 436 + }, + { + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": 479 + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "slot_index": 0, + "links": [ + 437 + ] + } + ], + "title": "CLIP Text Encode (Positive Prompt)", + "properties": { + "Node name for S&R": "CLIPTextEncode", + "cnr_id": "comfy-core", + "ver": "0.3.34", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "widget_ue_connectable": {} + }, + "widgets_values": [ + "" + ], + "color": "#232", + "bgcolor": "#353" + }, + { + "id": 281, + "type": "WanVaceToVideo", + "pos": [ + -1780, + 1940 + ], + "size": [ + 320, + 360 + ], + "flags": {}, + "order": 16, + "mode": 0, + "inputs": [ + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "link": 437 + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "link": 438 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 439 + }, + { + "localized_name": "control_video", + "name": "control_video", + "shape": 7, + "type": "IMAGE", + "link": 440 + }, + { + "localized_name": "control_masks", + "name": "control_masks", + "shape": 7, + "type": "MASK", + "link": 441 + }, + { + "localized_name": "reference_image", + "name": "reference_image", + "shape": 7, + "type": "IMAGE", + "link": 478 + }, + { + "localized_name": "width", + "name": "width", + "type": "INT", + "widget": { + "name": "width" + }, + "link": 442 + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "widget": { + "name": "height" + }, + "link": 443 + }, + { + "localized_name": "length", + "name": "length", + "type": "INT", + "widget": { + "name": "length" + }, + "link": 444 + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "widget": { + "name": "batch_size" + }, + "link": null + }, + { + "localized_name": "strength", + "name": "strength", + "type": "FLOAT", + "widget": { + "name": "strength" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "links": [ + 455 + ] + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "links": [ + 456 + ] + }, + { + "localized_name": "latent", + "name": "latent", + "type": "LATENT", + "links": [ + 457 + ] + }, + { + "localized_name": "trim_latent", + "name": "trim_latent", + "type": "INT", + "links": [ + 453 + ] + } + ], + "properties": { + "Node name for S&R": "WanVaceToVideo", + "cnr_id": "comfy-core", + "ver": "0.3.34", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "widget_ue_connectable": { + "width": true, + "height": true, + "length": true + } + }, + "widgets_values": [ + 720, + 720, + 81, + 1, + 1 + ] + }, + { + "id": 282, + "type": "InvertMask", + "pos": [ + -1510, + 3410 + ], + "size": [ + 230, + 80 + ], + "flags": {}, + "order": 17, + "mode": 0, + "inputs": [ + { + "localized_name": "mask", + "name": "mask", + "type": "MASK", + "link": 445 + } + ], + "outputs": [ + { + "localized_name": "MASK", + "name": "MASK", + "type": "MASK", + "links": [ + 446 + ] + } + ], + "properties": { + "Node name for S&R": "InvertMask", + "cnr_id": "comfy-core", + "ver": "0.3.40", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 283, + "type": "MaskToImage", + "pos": [ + -1510, + 3550 + ], + "size": [ + 230, + 80 + ], + "flags": {}, + "order": 18, + "mode": 0, + "inputs": [ + { + "localized_name": "mask", + "name": "mask", + "type": "MASK", + "link": 446 + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 448 + ] + } + ], + "properties": { + "Node name for S&R": "MaskToImage", + "cnr_id": "comfy-core", + "ver": "0.3.40", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 284, + "type": "ImageCompositeMasked", + "pos": [ + -1210, + 3210 + ], + "size": [ + 230, + 220 + ], + "flags": {}, + "order": 19, + "mode": 0, + "inputs": [ + { + "localized_name": "destination", + "name": "destination", + "type": "IMAGE", + "link": 447 + }, + { + "localized_name": "source", + "name": "source", + "type": "IMAGE", + "link": 448 + }, + { + "localized_name": "mask", + "name": "mask", + "shape": 7, + "type": "MASK", + "link": 449 + }, + { + "localized_name": "x", + "name": "x", + "type": "INT", + "widget": { + "name": "x" + }, + "link": null + }, + { + "localized_name": "y", + "name": "y", + "type": "INT", + "widget": { + "name": "y" + }, + "link": null + }, + { + "localized_name": "resize_source", + "name": "resize_source", + "type": "BOOLEAN", + "widget": { + "name": "resize_source" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 440, + 499 + ] + } + ], + "properties": { + "Node name for S&R": "ImageCompositeMasked", + "cnr_id": "comfy-core", + "ver": "0.3.40", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 0, + 0, + true + ] + }, + { + "id": 287, + "type": "TrimVideoLatent", + "pos": [ + -220, + 1950 + ], + "size": [ + 320, + 110 + ], + "flags": { + "collapsed": false + }, + "order": 20, + "mode": 0, + "inputs": [ + { + "localized_name": "samples", + "name": "samples", + "type": "LATENT", + "link": 452 + }, + { + "localized_name": "trim_amount", + "name": "trim_amount", + "type": "INT", + "widget": { + "name": "trim_amount" + }, + "link": 453 + } + ], + "outputs": [ + { + "localized_name": "LATENT", + "name": "LATENT", + "type": "LATENT", + "links": [ + 470 + ] + } + ], + "properties": { + "Node name for S&R": "TrimVideoLatent", + "cnr_id": "comfy-core", + "ver": "0.3.34", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "widget_ue_connectable": { + "trim_amount": true + } + }, + "widgets_values": [ + 0 + ] + }, + { + "id": 288, + "type": "KSampler", + "pos": [ + -560, + 2120 + ], + "size": [ + 320, + 350 + ], + "flags": {}, + "order": 21, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 454 + }, + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "link": 455 + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "link": 456 + }, + { + "localized_name": "latent_image", + "name": "latent_image", + "type": "LATENT", + "link": 457 + }, + { + "localized_name": "seed", + "name": "seed", + "type": "INT", + "widget": { + "name": "seed" + }, + "link": 483 + }, + { + "localized_name": "steps", + "name": "steps", + "type": "INT", + "widget": { + "name": "steps" + }, + "link": 458 + }, + { + "localized_name": "cfg", + "name": "cfg", + "type": "FLOAT", + "widget": { + "name": "cfg" + }, + "link": 459 + }, + { + "localized_name": "sampler_name", + "name": "sampler_name", + "type": "COMBO", + "widget": { + "name": "sampler_name" + }, + "link": null + }, + { + "localized_name": "scheduler", + "name": "scheduler", + "type": "COMBO", + "widget": { + "name": "scheduler" + }, + "link": null + }, + { + "localized_name": "denoise", + "name": "denoise", + "type": "FLOAT", + "widget": { + "name": "denoise" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "LATENT", + "name": "LATENT", + "type": "LATENT", + "slot_index": 0, + "links": [ + 452 + ] + } + ], + "properties": { + "Node name for S&R": "KSampler", + "cnr_id": "comfy-core", + "ver": "0.3.34", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "widget_ue_connectable": {} + }, + "widgets_values": [ + 832378512055965, + "fixed", + 4, + 1, + "uni_pc", + "simple", + 1 + ] + }, + { + "id": 289, + "type": "ImageFromBatch", + "pos": [ + -2360, + 3410 + ], + "size": [ + 270, + 140 + ], + "flags": {}, + "order": 22, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 460 + }, + { + "localized_name": "batch_index", + "name": "batch_index", + "type": "INT", + "widget": { + "name": "batch_index" + }, + "link": null + }, + { + "localized_name": "length", + "name": "length", + "type": "INT", + "widget": { + "name": "length" + }, + "link": 494 + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 463 + ] + } + ], + "properties": { + "Node name for S&R": "ImageFromBatch", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 0, + 81 + ] + }, + { + "id": 290, + "type": "PrimitiveInt", + "pos": [ + -2690, + 3540 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 23, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "INT", + "widget": { + "name": "value" + }, + "link": 481 + } + ], + "outputs": [ + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 461 + ] + } + ], + "title": "Int (Height)", + "properties": { + "Node name for S&R": "PrimitiveInt", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 720, + "fixed" + ] + }, + { + "id": 291, + "type": "ComfyMathExpression", + "pos": [ + -2650, + 3700 + ], + "size": [ + 230, + 80 + ], + "flags": { + "collapsed": true + }, + "order": 24, + "mode": 0, + "inputs": [ + { + "label": "a", + "localized_name": "values.a", + "name": "values.a", + "type": "FLOAT,INT,BOOLEAN", + "link": 461 + }, + { + "label": "b", + "localized_name": "values.b", + "name": "values.b", + "shape": 7, + "type": "FLOAT,INT,BOOLEAN", + "link": null + }, + { + "localized_name": "expression", + "name": "expression", + "type": "STRING", + "widget": { + "name": "expression" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "FLOAT", + "name": "FLOAT", + "type": "FLOAT", + "links": [] + }, + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 465 + ] + }, + { + "localized_name": "BOOL", + "name": "BOOL", + "type": "BOOLEAN", + "links": [] + } + ], + "properties": { + "Node name for S&R": "ComfyMathExpression", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "floor(a/16)*16" + ] + }, + { + "id": 292, + "type": "ComfyMathExpression", + "pos": [ + -2650, + 3500 + ], + "size": [ + 230, + 80 + ], + "flags": { + "collapsed": true + }, + "order": 25, + "mode": 0, + "inputs": [ + { + "label": "a", + "localized_name": "values.a", + "name": "values.a", + "type": "FLOAT,INT,BOOLEAN", + "link": 462 + }, + { + "label": "b", + "localized_name": "values.b", + "name": "values.b", + "shape": 7, + "type": "FLOAT,INT,BOOLEAN", + "link": null + }, + { + "localized_name": "expression", + "name": "expression", + "type": "STRING", + "widget": { + "name": "expression" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "FLOAT", + "name": "FLOAT", + "type": "FLOAT", + "links": [] + }, + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 464 + ] + }, + { + "localized_name": "BOOL", + "name": "BOOL", + "type": "BOOLEAN", + "links": [] + } + ], + "properties": { + "Node name for S&R": "ComfyMathExpression", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "floor(a/16)*16" + ] + }, + { + "id": 293, + "type": "ResizeImageMaskNode", + "pos": [ + -2360, + 3590 + ], + "size": [ + 280, + 160 + ], + "flags": {}, + "order": 26, + "mode": 0, + "inputs": [ + { + "localized_name": "input", + "name": "input", + "type": "IMAGE,MASK", + "link": 463 + }, + { + "localized_name": "resize_type", + "name": "resize_type", + "type": "COMFY_DYNAMICCOMBO_V3", + "widget": { + "name": "resize_type" + }, + "link": null + }, + { + "localized_name": "width", + "name": "resize_type.width", + "type": "INT", + "widget": { + "name": "resize_type.width" + }, + "link": 464 + }, + { + "localized_name": "height", + "name": "resize_type.height", + "type": "INT", + "widget": { + "name": "resize_type.height" + }, + "link": 465 + }, + { + "localized_name": "crop", + "name": "resize_type.crop", + "type": "COMBO", + "widget": { + "name": "resize_type.crop" + }, + "link": null + }, + { + "localized_name": "scale_method", + "name": "scale_method", + "type": "COMBO", + "widget": { + "name": "scale_method" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "resized", + "name": "resized", + "type": "*", + "links": [ + 426, + 447, + 469 + ] + } + ], + "properties": { + "Node name for S&R": "ResizeImageMaskNode", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "scale dimensions", + 512, + 512, + "center", + "area" + ] + }, + { + "id": 294, + "type": "ComfySwitchNode", + "pos": [ + -960, + 2150 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 27, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 466 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 467 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 468 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 458 + ] + } + ], + "properties": { + "Node name for S&R": "ComfySwitchNode", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + false + ] + }, + { + "id": 295, + "type": "GetImageSize", + "pos": [ + -2010, + 2920 + ], + "size": [ + 230, + 120 + ], + "flags": {}, + "order": 28, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 469 + } + ], + "outputs": [ + { + "localized_name": "width", + "name": "width", + "type": "INT", + "links": [ + 442 + ] + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "links": [ + 443 + ] + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "links": [ + 444 + ] + } + ], + "properties": { + "Node name for S&R": "GetImageSize", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 296, + "type": "VAEDecode", + "pos": [ + 520, + 2450 + ], + "size": [ + 320, + 100 + ], + "flags": { + "collapsed": false + }, + "order": 29, + "mode": 0, + "inputs": [ + { + "localized_name": "samples", + "name": "samples", + "type": "LATENT", + "link": 470 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 471 + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "slot_index": 0, + "links": [ + 423 + ] + } + ], + "properties": { + "Node name for S&R": "VAEDecode", + "cnr_id": "comfy-core", + "ver": "0.3.34", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "widget_ue_connectable": {} + } + }, + { + "id": 297, + "type": "PrimitiveInt", + "pos": [ + -2690, + 3350 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 30, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "INT", + "widget": { + "name": "value" + }, + "link": 480 + } + ], + "outputs": [ + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 462 + ] + } + ], + "title": "Int (Width)", + "properties": { + "Node name for S&R": "PrimitiveInt", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 720, + "fixed" + ] + }, + { + "id": 298, + "type": "GetVideoComponents", + "pos": [ + -2330, + 3210 + ], + "size": [ + 230, + 120 + ], + "flags": { + "collapsed": false + }, + "order": 31, + "mode": 0, + "inputs": [ + { + "localized_name": "video", + "name": "video", + "type": "VIDEO", + "link": 473 + } + ], + "outputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "links": [ + 460 + ] + }, + { + "localized_name": "audio", + "name": "audio", + "type": "AUDIO", + "links": [ + 424 + ] + }, + { + "localized_name": "fps", + "name": "fps", + "type": "FLOAT", + "links": [ + 425 + ] + } + ], + "properties": { + "Node name for S&R": "GetVideoComponents", + "cnr_id": "comfy-core", + "ver": "0.3.40", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 299, + "type": "UNETLoader", + "pos": [ + -2720, + 1980 + ], + "size": [ + 370, + 140 + ], + "flags": {}, + "order": 32, + "mode": 0, + "inputs": [ + { + "localized_name": "unet_name", + "name": "unet_name", + "type": "COMBO", + "widget": { + "name": "unet_name" + }, + "link": 485 + }, + { + "localized_name": "weight_dtype", + "name": "weight_dtype", + "type": "COMBO", + "widget": { + "name": "weight_dtype" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "slot_index": 0, + "links": [ + 428, + 429 + ] + } + ], + "properties": { + "Node name for S&R": "UNETLoader", + "cnr_id": "comfy-core", + "ver": "0.3.34", + "models": [ + { + "name": "wan2.1_vace_14B_fp16.safetensors", + "url": "https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/resolve/main/split_files/diffusion_models/wan2.1_vace_14B_fp16.safetensors", + "directory": "diffusion_models" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "widget_ue_connectable": {} + }, + "widgets_values": [ + "wan2.1_vace_14B_fp16.safetensors", + "fp8_e4m3fn_fast" + ] + }, + { + "id": 300, + "type": "PrimitiveBoolean", + "pos": [ + -1390, + 2980 + ], + "size": [ + 270, + 100 + ], + "flags": {}, + "order": 33, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "BOOLEAN", + "widget": { + "name": "value" + }, + "link": 489 + } + ], + "outputs": [ + { + "localized_name": "BOOLEAN", + "name": "BOOLEAN", + "type": "BOOLEAN", + "links": [ + 431, + 434, + 468 + ] + } + ], + "title": "Boolean (Enable Lightning LoRA)", + "properties": { + "Node name for S&R": "PrimitiveBoolean", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + true + ] + }, + { + "id": 308, + "type": "ImageFromBatch", + "pos": [ + -2360, + 3410 + ], + "size": [ + 270, + 140 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": null + }, + { + "localized_name": "batch_index", + "name": "batch_index", + "type": "INT", + "widget": { + "name": "batch_index" + }, + "link": null + }, + { + "localized_name": "length", + "name": "length", + "type": "INT", + "widget": { + "name": "length" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": null + } + ], + "properties": { + "Node name for S&R": "ImageFromBatch", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 0, + 1 + ] + }, + { + "id": 310, + "type": "MaskPreview", + "pos": [ + -900, + 3230 + ], + "size": [ + 230, + 80 + ], + "flags": {}, + "order": 34, + "mode": 4, + "inputs": [ + { + "localized_name": "mask", + "name": "mask", + "type": "MASK", + "link": 498 + } + ], + "outputs": [], + "properties": { + "Node name for S&R": "MaskPreview", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 312, + "type": "PreviewImage", + "pos": [ + -520, + 3230 + ], + "size": [ + 230, + 80 + ], + "flags": {}, + "order": 35, + "mode": 4, + "inputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "link": 499 + } + ], + "outputs": [], + "properties": { + "Node name for S&R": "PreviewImage", + "cnr_id": "comfy-core", + "ver": "0.21.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + } + ], + "groups": [ + { + "id": 1, + "title": "Models", + "bounding": [ + -2750, + 1860, + 430, + 770 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 2, + "title": "Prompt", + "bounding": [ + -2290, + 1860, + 460, + 770 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 3, + "title": "Sampling", + "bounding": [ + -590, + 1860, + 700, + 620 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 20, + "title": "Create Video Mask", + "bounding": [ + -2030, + 3110, + 440, + 550 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 23, + "title": "Conditioning", + "bounding": [ + -1800, + 1860, + 370, + 450 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 26, + "title": "Apply Mask to Video", + "bounding": [ + -1560, + 3110, + 1320, + 550 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 29, + "title": "Swtich Logic", + "bounding": [ + -1400, + 1860, + 780, + 1060 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 27, + "title": "Lightning LoRA", + "bounding": [ + -1390, + 2290, + 370, + 620 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 28, + "title": "Original", + "bounding": [ + -1390, + 1900, + 370, + 370 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 31, + "title": "Video Size Preprocessing", + "bounding": [ + -2740, + 3110, + 680, + 770 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 30, + "title": "Size", + "bounding": [ + -2710, + 3270, + 330, + 470 + ], + "color": "#3f789e", + "flags": {} + } + ], + "links": [ + { + "id": 422, + "origin_id": 275, + "origin_slot": 0, + "target_id": 266, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 423, + "origin_id": 296, + "origin_slot": 0, + "target_id": 267, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 424, + "origin_id": 298, + "origin_slot": 1, + "target_id": 267, + "target_slot": 1, + "type": "AUDIO" + }, + { + "id": 425, + "origin_id": 298, + "origin_slot": 2, + "target_id": 267, + "target_slot": 2, + "type": "FLOAT" + }, + { + "id": 426, + "origin_id": 293, + "origin_slot": 0, + "target_id": 268, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 427, + "origin_id": 268, + "origin_slot": 0, + "target_id": 269, + "target_slot": 0, + "type": "MASK" + }, + { + "id": 428, + "origin_id": 299, + "origin_slot": 0, + "target_id": 272, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 429, + "origin_id": 299, + "origin_slot": 0, + "target_id": 275, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 430, + "origin_id": 272, + "origin_slot": 0, + "target_id": 275, + "target_slot": 1, + "type": "MODEL" + }, + { + "id": 431, + "origin_id": 300, + "origin_slot": 0, + "target_id": 275, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 432, + "origin_id": 271, + "origin_slot": 0, + "target_id": 276, + "target_slot": 0, + "type": "FLOAT" + }, + { + "id": 433, + "origin_id": 274, + "origin_slot": 0, + "target_id": 276, + "target_slot": 1, + "type": "FLOAT" + }, + { + "id": 434, + "origin_id": 300, + "origin_slot": 0, + "target_id": 276, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 435, + "origin_id": 277, + "origin_slot": 0, + "target_id": 279, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 436, + "origin_id": 277, + "origin_slot": 0, + "target_id": 280, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 437, + "origin_id": 280, + "origin_slot": 0, + "target_id": 281, + "target_slot": 0, + "type": "CONDITIONING" + }, + { + "id": 438, + "origin_id": 279, + "origin_slot": 0, + "target_id": 281, + "target_slot": 1, + "type": "CONDITIONING" + }, + { + "id": 439, + "origin_id": 278, + "origin_slot": 0, + "target_id": 281, + "target_slot": 2, + "type": "VAE" + }, + { + "id": 440, + "origin_id": 284, + "origin_slot": 0, + "target_id": 281, + "target_slot": 3, + "type": "IMAGE" + }, + { + "id": 441, + "origin_id": 269, + "origin_slot": 0, + "target_id": 281, + "target_slot": 4, + "type": "MASK" + }, + { + "id": 442, + "origin_id": 295, + "origin_slot": 0, + "target_id": 281, + "target_slot": 6, + "type": "INT" + }, + { + "id": 443, + "origin_id": 295, + "origin_slot": 1, + "target_id": 281, + "target_slot": 7, + "type": "INT" + }, + { + "id": 444, + "origin_id": 295, + "origin_slot": 2, + "target_id": 281, + "target_slot": 8, + "type": "INT" + }, + { + "id": 445, + "origin_id": 269, + "origin_slot": 0, + "target_id": 282, + "target_slot": 0, + "type": "MASK" + }, + { + "id": 446, + "origin_id": 282, + "origin_slot": 0, + "target_id": 283, + "target_slot": 0, + "type": "MASK" + }, + { + "id": 447, + "origin_id": 293, + "origin_slot": 0, + "target_id": 284, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 448, + "origin_id": 283, + "origin_slot": 0, + "target_id": 284, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 449, + "origin_id": 269, + "origin_slot": 0, + "target_id": 284, + "target_slot": 2, + "type": "MASK" + }, + { + "id": 452, + "origin_id": 288, + "origin_slot": 0, + "target_id": 287, + "target_slot": 0, + "type": "LATENT" + }, + { + "id": 453, + "origin_id": 281, + "origin_slot": 3, + "target_id": 287, + "target_slot": 1, + "type": "INT" + }, + { + "id": 454, + "origin_id": 266, + "origin_slot": 0, + "target_id": 288, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 455, + "origin_id": 281, + "origin_slot": 0, + "target_id": 288, + "target_slot": 1, + "type": "CONDITIONING" + }, + { + "id": 456, + "origin_id": 281, + "origin_slot": 1, + "target_id": 288, + "target_slot": 2, + "type": "CONDITIONING" + }, + { + "id": 457, + "origin_id": 281, + "origin_slot": 2, + "target_id": 288, + "target_slot": 3, + "type": "LATENT" + }, + { + "id": 458, + "origin_id": 294, + "origin_slot": 0, + "target_id": 288, + "target_slot": 5, + "type": "INT" + }, + { + "id": 459, + "origin_id": 276, + "origin_slot": 0, + "target_id": 288, + "target_slot": 6, + "type": "FLOAT" + }, + { + "id": 460, + "origin_id": 298, + "origin_slot": 0, + "target_id": 289, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 461, + "origin_id": 290, + "origin_slot": 0, + "target_id": 291, + "target_slot": 0, + "type": "INT" + }, + { + "id": 462, + "origin_id": 297, + "origin_slot": 0, + "target_id": 292, + "target_slot": 0, + "type": "INT" + }, + { + "id": 463, + "origin_id": 289, + "origin_slot": 0, + "target_id": 293, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 464, + "origin_id": 292, + "origin_slot": 1, + "target_id": 293, + "target_slot": 2, + "type": "INT" + }, + { + "id": 465, + "origin_id": 291, + "origin_slot": 1, + "target_id": 293, + "target_slot": 3, + "type": "INT" + }, + { + "id": 466, + "origin_id": 270, + "origin_slot": 0, + "target_id": 294, + "target_slot": 0, + "type": "INT" + }, + { + "id": 467, + "origin_id": 273, + "origin_slot": 0, + "target_id": 294, + "target_slot": 1, + "type": "INT" + }, + { + "id": 468, + "origin_id": 300, + "origin_slot": 0, + "target_id": 294, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 469, + "origin_id": 293, + "origin_slot": 0, + "target_id": 295, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 470, + "origin_id": 287, + "origin_slot": 0, + "target_id": 296, + "target_slot": 0, + "type": "LATENT" + }, + { + "id": 471, + "origin_id": 278, + "origin_slot": 0, + "target_id": 296, + "target_slot": 1, + "type": "VAE" + }, + { + "id": 473, + "origin_id": -10, + "origin_slot": 0, + "target_id": 298, + "target_slot": 0, + "type": "VIDEO" + }, + { + "id": 474, + "origin_id": 267, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "VIDEO" + }, + { + "id": 478, + "origin_id": -10, + "origin_slot": 1, + "target_id": 281, + "target_slot": 5, + "type": "IMAGE" + }, + { + "id": 479, + "origin_id": -10, + "origin_slot": 2, + "target_id": 280, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 480, + "origin_id": -10, + "origin_slot": 3, + "target_id": 297, + "target_slot": 0, + "type": "INT" + }, + { + "id": 481, + "origin_id": -10, + "origin_slot": 4, + "target_id": 290, + "target_slot": 0, + "type": "INT" + }, + { + "id": 494, + "origin_id": -10, + "origin_slot": 5, + "target_id": 289, + "target_slot": 2, + "type": "INT" + }, + { + "id": 483, + "origin_id": -10, + "origin_slot": 6, + "target_id": 288, + "target_slot": 4, + "type": "INT" + }, + { + "id": 485, + "origin_id": -10, + "origin_slot": 7, + "target_id": 299, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 486, + "origin_id": -10, + "origin_slot": 8, + "target_id": 277, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 487, + "origin_id": -10, + "origin_slot": 9, + "target_id": 278, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 489, + "origin_id": -10, + "origin_slot": 10, + "target_id": 300, + "target_slot": 0, + "type": "BOOLEAN" + }, + { + "id": 490, + "origin_id": -10, + "origin_slot": 11, + "target_id": 272, + "target_slot": 1, + "type": "COMBO" + }, + { + "id": 491, + "origin_id": -10, + "origin_slot": 12, + "target_id": 268, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 496, + "origin_id": -10, + "origin_slot": 13, + "target_id": 269, + "target_slot": 1, + "type": "INT" + }, + { + "id": 497, + "origin_id": -10, + "origin_slot": 14, + "target_id": 268, + "target_slot": 8, + "type": "COMBO" + }, + { + "id": 498, + "origin_id": 269, + "origin_slot": 0, + "target_id": 310, + "target_slot": 0, + "type": "MASK" + }, + { + "id": 499, + "origin_id": 284, + "origin_slot": 0, + "target_id": 312, + "target_slot": 0, + "type": "IMAGE" + } + ], + "extra": {}, + "category": "Video generation and editing/Inpaint video", + "description": "Removes objects from video by inpainting masked regions using Wan 2.1 VACE, with SAM3 text-guided segmentation and optional Lightning LoRA turbo mode." + }, + { + "id": "17df2eeb-d89e-46ee-9480-a4ca2494b207", + "version": 1, + "state": { + "lastGroupId": 31, + "lastNodeId": 315, + "lastLinkId": 499, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Image Segmentation (SAM3)", + "description": "Segments images into masks using Meta SAM3 from text prompts, points, or boxes.", + "inputNode": { + "id": -10, + "bounding": [ + -2260, + -3450, + 136.369140625, + 220 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + -1130, + -3305, + 120, + 80 + ] + }, + "inputs": [ + { + "id": "a6e75fa2-162a-4af0-a2fd-1e9c899a5ab6", + "name": "image", + "type": "IMAGE", + "linkIds": [ + 264 + ], + "localized_name": "image", + "label": "image", + "pos": [ + -2143.630859375, + -3430 + ] + }, + { + "id": "3cefd304-7631-4ff6-a5a0-5a0ffb120745", + "name": "text", + "type": "STRING", + "linkIds": [ + 265 + ], + "label": "object", + "pos": [ + -2143.630859375, + -3410 + ] + }, + { + "id": "1aec91c5-d8d2-441c-928c-49c14e7e80ed", + "name": "bboxes", + "type": "BOUNDING_BOX", + "linkIds": [ + 266 + ], + "pos": [ + -2143.630859375, + -3390 + ] + }, + { + "id": "1ec7ce1a-8257-4719-8a81-60ebc8a98899", + "name": "positive_coords", + "type": "STRING", + "linkIds": [ + 267 + ], + "pos": [ + -2143.630859375, + -3370 + ] + }, + { + "id": "c65f8b87-9bd7-48be-9fc2-823431e95019", + "name": "negative_coords", + "type": "STRING", + "linkIds": [ + 268 + ], + "pos": [ + -2143.630859375, + -3350 + ] + }, + { + "id": "bb4ba35a-ccfe-4c37-98e5-d9b0d69585fb", + "name": "threshold", + "type": "FLOAT", + "linkIds": [ + 269 + ], + "pos": [ + -2143.630859375, + -3330 + ] + }, + { + "id": "b1439668-b050-490b-a5dc-fc4052c55666", + "name": "refine_iterations", + "type": "INT", + "linkIds": [ + 270 + ], + "pos": [ + -2143.630859375, + -3310 + ] + }, + { + "id": "86e239e5-c098-4302-b54d-d42a38bc0f89", + "name": "individual_masks", + "type": "BOOLEAN", + "linkIds": [ + 271 + ], + "pos": [ + -2143.630859375, + -3290 + ] + }, + { + "id": "f9e0b9d4-b2f1-4907-a4a5-305656576706", + "name": "ckpt_name", + "type": "COMBO", + "linkIds": [ + 272 + ], + "pos": [ + -2143.630859375, + -3270 + ] + } + ], + "outputs": [ + { + "id": "ff50da09-1e59-4a58-9b7f-be1a00aa5913", + "name": "masks", + "type": "MASK", + "linkIds": [ + 231 + ], + "localized_name": "masks", + "pos": [ + -1110, + -3285 + ] + }, + { + "id": "8f622e40-8528-4078-b7d3-147e9f872194", + "name": "bboxes", + "type": "BOUNDING_BOX", + "linkIds": [ + 232 + ], + "localized_name": "bboxes", + "pos": [ + -1110, + -3265 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 75, + "type": "SAM3_Detect", + "pos": [ + -1470, + -3460 + ], + "size": [ + 270, + 260 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "label": "model", + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 237 + }, + { + "label": "image", + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 264 + }, + { + "label": "conditioning", + "localized_name": "conditioning", + "name": "conditioning", + "shape": 7, + "type": "CONDITIONING", + "link": 200 + }, + { + "label": "bboxes", + "localized_name": "bboxes", + "name": "bboxes", + "shape": 7, + "type": "BOUNDING_BOX", + "link": 266 + }, + { + "label": "positive_coords", + "localized_name": "positive_coords", + "name": "positive_coords", + "shape": 7, + "type": "STRING", + "link": 267 + }, + { + "label": "negative_coords", + "localized_name": "negative_coords", + "name": "negative_coords", + "shape": 7, + "type": "STRING", + "link": 268 + }, + { + "localized_name": "threshold", + "name": "threshold", + "type": "FLOAT", + "widget": { + "name": "threshold" + }, + "link": 269 + }, + { + "localized_name": "refine_iterations", + "name": "refine_iterations", + "type": "INT", + "widget": { + "name": "refine_iterations" + }, + "link": 270 + }, + { + "localized_name": "individual_masks", + "name": "individual_masks", + "type": "BOOLEAN", + "widget": { + "name": "individual_masks" + }, + "link": 271 + } + ], + "outputs": [ + { + "localized_name": "masks", + "name": "masks", + "type": "MASK", + "links": [ + 231 + ] + }, + { + "localized_name": "bboxes", + "name": "bboxes", + "type": "BOUNDING_BOX", + "links": [ + 232 + ] + } + ], + "properties": { + "Node name for S&R": "SAM3_Detect", + "cnr_id": "comfy-core", + "ver": "0.19.3", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + } + }, + "widgets_values": [ + 0.5, + 2, + false + ] + }, + { + "id": 236, + "type": "CheckpointLoaderSimple", + "pos": [ + -1970, + -3200 + ], + "size": [ + 330, + 140 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "ckpt_name", + "name": "ckpt_name", + "type": "COMBO", + "widget": { + "name": "ckpt_name" + }, + "link": 272 + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "links": [ + 237 + ] + }, + { + "localized_name": "CLIP", + "name": "CLIP", + "type": "CLIP", + "links": [ + 240 + ] + }, + { + "localized_name": "VAE", + "name": "VAE", + "type": "VAE", + "links": null + } + ], + "properties": { + "Node name for S&R": "CheckpointLoaderSimple", + "cnr_id": "comfy-core", + "ver": "0.19.3", + "models": [ + { + "name": "sam3.1_multiplex_fp16.safetensors", + "url": "https://huggingface.co/Comfy-Org/sam3.1/resolve/main/checkpoints/sam3.1_multiplex_fp16.safetensors", + "directory": "checkpoints" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + } + }, + "widgets_values": [ + "sam3.1_multiplex_fp16.safetensors" + ] + }, + { + "id": 237, + "type": "CLIPTextEncode", + "pos": [ + -2000, + -3000 + ], + "size": [ + 400, + 200 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 240 + }, + { + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": 265 + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "links": [ + 200 + ] + } + ], + "properties": { + "Node name for S&R": "CLIPTextEncode", + "cnr_id": "comfy-core", + "ver": "0.19.3", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + } + }, + "widgets_values": [ + "" + ] + } + ], + "groups": [], + "links": [ + { + "id": 237, + "origin_id": 236, + "origin_slot": 0, + "target_id": 75, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 200, + "origin_id": 237, + "origin_slot": 0, + "target_id": 75, + "target_slot": 2, + "type": "CONDITIONING" + }, + { + "id": 240, + "origin_id": 236, + "origin_slot": 1, + "target_id": 237, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 231, + "origin_id": 75, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "MASK" + }, + { + "id": 232, + "origin_id": 75, + "origin_slot": 1, + "target_id": -20, + "target_slot": 1, + "type": "BOUNDING_BOX" + }, + { + "id": 264, + "origin_id": -10, + "origin_slot": 0, + "target_id": 75, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 265, + "origin_id": -10, + "origin_slot": 1, + "target_id": 237, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 266, + "origin_id": -10, + "origin_slot": 2, + "target_id": 75, + "target_slot": 3, + "type": "BOUNDING_BOX" + }, + { + "id": 267, + "origin_id": -10, + "origin_slot": 3, + "target_id": 75, + "target_slot": 4, + "type": "STRING" + }, + { + "id": 268, + "origin_id": -10, + "origin_slot": 4, + "target_id": 75, + "target_slot": 5, + "type": "STRING" + }, + { + "id": 269, + "origin_id": -10, + "origin_slot": 5, + "target_id": 75, + "target_slot": 6, + "type": "FLOAT" + }, + { + "id": 270, + "origin_id": -10, + "origin_slot": 6, + "target_id": 75, + "target_slot": 7, + "type": "INT" + }, + { + "id": 271, + "origin_id": -10, + "origin_slot": 7, + "target_id": 75, + "target_slot": 8, + "type": "BOOLEAN" + }, + { + "id": 272, + "origin_id": -10, + "origin_slot": 8, + "target_id": 236, + "target_slot": 0, + "type": "COMBO" + } + ], + "extra": { + "ue_links": [] + } + } + ] + }, + "extra": {} +} \ No newline at end of file diff --git a/blueprints/Video Segmentation (SAM3).json b/blueprints/Video Segmentation (SAM3).json index 4d9a13412..4c7253869 100644 --- a/blueprints/Video Segmentation (SAM3).json +++ b/blueprints/Video Segmentation (SAM3).json @@ -818,7 +818,7 @@ } ], "extra": {}, - "category": "Video Tools", + "category": "Conditioning & Preprocessors/Segmentation & Mask", "description": "Segments video into temporally consistent masks using Meta SAM3 from text or interactive prompts." } ] diff --git a/blueprints/Video Upscale(GAN x4).json b/blueprints/Video Upscale(GAN x4).json index 73476e36b..fc291ac41 100644 --- a/blueprints/Video Upscale(GAN x4).json +++ b/blueprints/Video Upscale(GAN x4).json @@ -412,7 +412,7 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Video generation and editing/Enhance video", + "category": "Video generation and editing/Upscale", "description": "Upscales video to 4× resolution using a GAN-based upscaling model." } ] diff --git a/blueprints/Video to Pose Map (SDPose Multi-Person).json b/blueprints/Video to Pose Map (SDPose Multi-Person).json new file mode 100644 index 000000000..64ef6e524 --- /dev/null +++ b/blueprints/Video to Pose Map (SDPose Multi-Person).json @@ -0,0 +1,1323 @@ +{ + "revision": 0, + "last_node_id": 675, + "last_link_id": 0, + "nodes": [ + { + "id": 675, + "type": "01b6a731-fb78-4070-9a38-c87146da9604", + "pos": [ + -2480, + 3400 + ], + "size": [ + 370, + 638.625 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "label": "resize_target_longer_size", + "name": "resize_type.longer_size", + "type": "INT", + "widget": { + "name": "resize_type.longer_size" + }, + "link": null + }, + { + "name": "scale_method", + "type": "COMBO", + "widget": { + "name": "scale_method" + }, + "link": null + }, + { + "name": "draw_body", + "type": "BOOLEAN", + "widget": { + "name": "draw_body" + }, + "link": null + }, + { + "name": "draw_hands", + "type": "BOOLEAN", + "widget": { + "name": "draw_hands" + }, + "link": null + }, + { + "name": "draw_face", + "type": "BOOLEAN", + "widget": { + "name": "draw_face" + }, + "link": null + }, + { + "name": "draw_feet", + "type": "BOOLEAN", + "widget": { + "name": "draw_feet" + }, + "link": null + }, + { + "name": "stick_width", + "type": "INT", + "widget": { + "name": "stick_width" + }, + "link": null + }, + { + "name": "face_point_size", + "type": "INT", + "widget": { + "name": "face_point_size" + }, + "link": null + }, + { + "name": "score_threshold", + "type": "FLOAT", + "widget": { + "name": "score_threshold" + }, + "link": null + }, + { + "label": "detect_threshold", + "name": "threshold", + "type": "FLOAT", + "widget": { + "name": "threshold" + }, + "link": null + }, + { + "label": "detect_class", + "name": "class_name", + "type": "COMBO", + "widget": { + "name": "class_name" + }, + "link": null + }, + { + "name": "max_detections", + "type": "INT", + "widget": { + "name": "max_detections" + }, + "link": null + }, + { + "name": "ckpt_name", + "type": "COMBO", + "widget": { + "name": "ckpt_name" + }, + "link": null + }, + { + "name": "unet_name", + "type": "COMBO", + "widget": { + "name": "unet_name" + }, + "link": null + }, + { + "name": "video", + "type": "VIDEO", + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [] + }, + { + "name": "keypoints", + "type": "POSE_KEYPOINT", + "links": null + }, + { + "name": "bboxes", + "type": "BOUNDING_BOX", + "links": [] + }, + { + "name": "audio", + "type": "AUDIO", + "links": [] + }, + { + "name": "fps", + "type": "FLOAT", + "links": [] + } + ], + "properties": { + "proxyWidgets": [ + [ + "674", + "resize_type.longer_size" + ], + [ + "674", + "scale_method" + ], + [ + "672", + "draw_body" + ], + [ + "672", + "draw_hands" + ], + [ + "672", + "draw_face" + ], + [ + "672", + "draw_feet" + ], + [ + "672", + "stick_width" + ], + [ + "672", + "face_point_size" + ], + [ + "672", + "score_threshold" + ], + [ + "678", + "threshold" + ], + [ + "678", + "class_name" + ], + [ + "678", + "max_detections" + ], + [ + "673", + "ckpt_name" + ], + [ + "677", + "unet_name" + ] + ], + "cnr_id": "comfy-core", + "ver": "0.15.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [], + "title": "Video to Pose Map (SDPose Multi-Person)" + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "01b6a731-fb78-4070-9a38-c87146da9604", + "version": 1, + "state": { + "lastGroupId": 2, + "lastNodeId": 699, + "lastLinkId": 1754, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Video to Pose Map (SDPose Multi-Person)", + "inputNode": { + "id": -10, + "bounding": [ + -3570, + 3300, + 182.8984375, + 340 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + -1890, + 3730, + 120, + 140 + ] + }, + "inputs": [ + { + "id": "088eefc1-cd8a-4573-993f-9e4da008a12d", + "name": "resize_type.longer_size", + "type": "INT", + "linkIds": [ + 1704 + ], + "label": "resize_target_longer_size", + "pos": [ + -3407.1015625, + 3320 + ] + }, + { + "id": "b6449bd3-73d4-41c8-b81f-cf8d33f76a2e", + "name": "scale_method", + "type": "COMBO", + "linkIds": [ + 1705 + ], + "pos": [ + -3407.1015625, + 3340 + ] + }, + { + "id": "4cff52ad-ed07-4c97-8803-fcbd89554fd0", + "name": "draw_body", + "type": "BOOLEAN", + "linkIds": [ + 1706 + ], + "pos": [ + -3407.1015625, + 3360 + ] + }, + { + "id": "7af63dce-f7df-4d7e-8215-d7c7f60bf81c", + "name": "draw_hands", + "type": "BOOLEAN", + "linkIds": [ + 1707 + ], + "pos": [ + -3407.1015625, + 3380 + ] + }, + { + "id": "af3a9bce-61f9-4aca-b530-9f65e028b35e", + "name": "draw_face", + "type": "BOOLEAN", + "linkIds": [ + 1708 + ], + "pos": [ + -3407.1015625, + 3400 + ] + }, + { + "id": "4620f6a3-2c85-4b79-ad8f-35d0326b568f", + "name": "draw_feet", + "type": "BOOLEAN", + "linkIds": [ + 1709 + ], + "pos": [ + -3407.1015625, + 3420 + ] + }, + { + "id": "fee5d0c9-8d4b-4934-81d8-ba2206dc56cb", + "name": "stick_width", + "type": "INT", + "linkIds": [ + 1710 + ], + "pos": [ + -3407.1015625, + 3440 + ] + }, + { + "id": "aafdd060-ba81-4324-a9cc-b656e1ebc133", + "name": "face_point_size", + "type": "INT", + "linkIds": [ + 1711 + ], + "pos": [ + -3407.1015625, + 3460 + ] + }, + { + "id": "514c5503-f9e6-4d23-b1ae-1d3291acb2a3", + "name": "score_threshold", + "type": "FLOAT", + "linkIds": [ + 1712 + ], + "pos": [ + -3407.1015625, + 3480 + ] + }, + { + "id": "4eb3e4ea-7a36-4511-8483-0d12aadd32f7", + "name": "threshold", + "type": "FLOAT", + "linkIds": [ + 1718 + ], + "label": "detect_threshold", + "pos": [ + -3407.1015625, + 3500 + ] + }, + { + "id": "c76a7a05-81e6-4b17-a9e0-85f47a5844f2", + "name": "class_name", + "type": "COMBO", + "linkIds": [ + 1719 + ], + "label": "detect_class", + "pos": [ + -3407.1015625, + 3520 + ] + }, + { + "id": "4417e988-6e80-4236-be31-4c179037f5a2", + "name": "max_detections", + "type": "INT", + "linkIds": [ + 1720 + ], + "pos": [ + -3407.1015625, + 3540 + ] + }, + { + "id": "7d7c4a0b-0d1b-4c98-942b-f90548d2a492", + "name": "ckpt_name", + "type": "COMBO", + "linkIds": [ + 1721 + ], + "pos": [ + -3407.1015625, + 3560 + ] + }, + { + "id": "4d75122c-2c14-452a-98fe-d1545d3e012a", + "name": "unet_name", + "type": "COMBO", + "linkIds": [ + 1722 + ], + "pos": [ + -3407.1015625, + 3580 + ] + }, + { + "id": "6c46c988-4dd1-41a2-957e-03caf60d7657", + "name": "video", + "type": "VIDEO", + "linkIds": [ + 1741 + ], + "pos": [ + -3407.1015625, + 3600 + ] + } + ], + "outputs": [ + { + "id": "f05ed8cc-9403-4f14-8085-4364b06f8a48", + "name": "IMAGE", + "type": "IMAGE", + "linkIds": [ + 1701 + ], + "localized_name": "IMAGE", + "pos": [ + -1870, + 3750 + ] + }, + { + "id": "4b64118e-3cef-4eeb-9dad-4cd09cfd63a2", + "name": "keypoints", + "type": "POSE_KEYPOINT", + "linkIds": [ + 1725 + ], + "pos": [ + -1870, + 3770 + ] + }, + { + "id": "a27f7e34-dcbc-4fb0-a4e1-2c5fc423ca5f", + "name": "bboxes", + "type": "BOUNDING_BOX", + "linkIds": [ + 1726 + ], + "pos": [ + -1870, + 3790 + ] + }, + { + "id": "b7fe351d-2b38-41ea-9f4d-3be1a0aad275", + "name": "audio", + "type": "AUDIO", + "linkIds": [ + 1743 + ], + "pos": [ + -1870, + 3810 + ] + }, + { + "id": "ae187b6f-c9ca-4487-b5c1-3ad775fe945e", + "name": "fps", + "type": "FLOAT", + "linkIds": [ + 1744 + ], + "pos": [ + -1870, + 3830 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 671, + "type": "SDPoseKeypointExtractor", + "pos": [ + -2550, + 3080 + ], + "size": [ + 270, + 180 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 1696 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 1697 + }, + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 1698 + }, + { + "localized_name": "bboxes", + "name": "bboxes", + "shape": 7, + "type": "BOUNDING_BOX", + "link": 1717 + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "widget": { + "name": "batch_size" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "keypoints", + "name": "keypoints", + "type": "POSE_KEYPOINT", + "links": [ + 1699, + 1725 + ] + } + ], + "properties": { + "Node name for S&R": "SDPoseKeypointExtractor", + "cnr_id": "comfy-core", + "ver": "0.15.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 16 + ] + }, + { + "id": 674, + "type": "ResizeImageMaskNode", + "pos": [ + -3010, + 3880 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "input", + "name": "input", + "type": "IMAGE,MASK", + "link": 1742 + }, + { + "localized_name": "resize_type", + "name": "resize_type", + "type": "COMFY_DYNAMICCOMBO_V3", + "widget": { + "name": "resize_type" + }, + "link": null + }, + { + "localized_name": "resize_type.longer_size", + "name": "resize_type.longer_size", + "type": "INT", + "widget": { + "name": "resize_type.longer_size" + }, + "link": 1704 + }, + { + "localized_name": "scale_method", + "name": "scale_method", + "type": "COMBO", + "widget": { + "name": "scale_method" + }, + "link": 1705 + } + ], + "outputs": [ + { + "localized_name": "resized", + "name": "resized", + "type": "*", + "links": [ + 1698, + 1716 + ] + } + ], + "properties": { + "Node name for S&R": "ResizeImageMaskNode", + "cnr_id": "comfy-core", + "ver": "0.15.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "scale longer dimension", + 1024, + "lanczos" + ] + }, + { + "id": 672, + "type": "SDPoseDrawKeypoints", + "pos": [ + -2540, + 3590 + ], + "size": [ + 270, + 280 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "keypoints", + "name": "keypoints", + "type": "POSE_KEYPOINT", + "link": 1699 + }, + { + "localized_name": "draw_body", + "name": "draw_body", + "type": "BOOLEAN", + "widget": { + "name": "draw_body" + }, + "link": 1706 + }, + { + "localized_name": "draw_hands", + "name": "draw_hands", + "type": "BOOLEAN", + "widget": { + "name": "draw_hands" + }, + "link": 1707 + }, + { + "localized_name": "draw_face", + "name": "draw_face", + "type": "BOOLEAN", + "widget": { + "name": "draw_face" + }, + "link": 1708 + }, + { + "localized_name": "draw_feet", + "name": "draw_feet", + "type": "BOOLEAN", + "widget": { + "name": "draw_feet" + }, + "link": 1709 + }, + { + "localized_name": "stick_width", + "name": "stick_width", + "type": "INT", + "widget": { + "name": "stick_width" + }, + "link": 1710 + }, + { + "localized_name": "face_point_size", + "name": "face_point_size", + "type": "INT", + "widget": { + "name": "face_point_size" + }, + "link": 1711 + }, + { + "localized_name": "score_threshold", + "name": "score_threshold", + "type": "FLOAT", + "widget": { + "name": "score_threshold" + }, + "link": 1712 + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 1701 + ] + } + ], + "properties": { + "Node name for S&R": "SDPoseDrawKeypoints", + "cnr_id": "comfy-core", + "ver": "0.15.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + true, + true, + true, + true, + 4, + 2, + 0.5 + ] + }, + { + "id": 673, + "type": "CheckpointLoaderSimple", + "pos": [ + -3040, + 3080 + ], + "size": [ + 390, + 160 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "ckpt_name", + "name": "ckpt_name", + "type": "COMBO", + "widget": { + "name": "ckpt_name" + }, + "link": 1721 + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "links": [ + 1696 + ] + }, + { + "localized_name": "CLIP", + "name": "CLIP", + "type": "CLIP", + "links": [] + }, + { + "localized_name": "VAE", + "name": "VAE", + "type": "VAE", + "links": [ + 1697 + ] + } + ], + "properties": { + "Node name for S&R": "CheckpointLoaderSimple", + "cnr_id": "comfy-core", + "ver": "0.15.0", + "models": [ + { + "name": "sdpose_wholebody_fp16.safetensors", + "url": "https://huggingface.co/Comfy-Org/SDPose/resolve/main/checkpoints/sdpose_wholebody_fp16.safetensors", + "directory": "checkpoints" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "sdpose_wholebody_fp16.safetensors" + ] + }, + { + "id": 677, + "type": "UNETLoader", + "pos": [ + -3030, + 3300 + ], + "size": [ + 370, + 110 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "localized_name": "unet_name", + "name": "unet_name", + "type": "COMBO", + "widget": { + "name": "unet_name" + }, + "link": 1722 + }, + { + "localized_name": "weight_dtype", + "name": "weight_dtype", + "type": "COMBO", + "widget": { + "name": "weight_dtype" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "links": [ + 1715 + ] + } + ], + "properties": { + "Node name for S&R": "UNETLoader", + "cnr_id": "comfy-core", + "ver": "0.14.1", + "models": [ + { + "name": "rt_detr_v4-x-hgnet_fp16.safetensors", + "url": "https://huggingface.co/Comfy-Org/SDPose/resolve/main/diffusion_models/rt_detr_v4-x-hgnet_fp16.safetensors", + "directory": "diffusion_models" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "rt_detr_v4-x-hgnet_fp16.safetensors", + "default" + ] + }, + { + "id": 678, + "type": "RTDETR_detect", + "pos": [ + -2540, + 3320 + ], + "size": [ + 270, + 200 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "label": "model", + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 1715 + }, + { + "label": "image", + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 1716 + }, + { + "localized_name": "threshold", + "name": "threshold", + "type": "FLOAT", + "widget": { + "name": "threshold" + }, + "link": 1718 + }, + { + "localized_name": "class_name", + "name": "class_name", + "type": "COMBO", + "widget": { + "name": "class_name" + }, + "link": 1719 + }, + { + "localized_name": "max_detections", + "name": "max_detections", + "type": "INT", + "widget": { + "name": "max_detections" + }, + "link": 1720 + } + ], + "outputs": [ + { + "localized_name": "bboxes", + "name": "bboxes", + "type": "BOUNDING_BOX", + "links": [ + 1717, + 1726 + ] + } + ], + "properties": { + "Node name for S&R": "RTDETR_detect", + "cnr_id": "comfy-core", + "ver": "0.15.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 0.5, + "person", + 2 + ] + }, + { + "id": 692, + "type": "GetVideoComponents", + "pos": [ + -3010, + 4100 + ], + "size": [ + 230, + 120 + ], + "flags": {}, + "order": 6, + "mode": 0, + "inputs": [ + { + "localized_name": "video", + "name": "video", + "type": "VIDEO", + "link": 1741 + } + ], + "outputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "links": [ + 1742 + ] + }, + { + "localized_name": "audio", + "name": "audio", + "type": "AUDIO", + "links": [ + 1743 + ] + }, + { + "localized_name": "fps", + "name": "fps", + "type": "FLOAT", + "links": [ + 1744 + ] + } + ], + "properties": { + "Node name for S&R": "GetVideoComponents", + "cnr_id": "comfy-core", + "ver": "0.18.1", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + } + ], + "groups": [], + "links": [ + { + "id": 1696, + "origin_id": 673, + "origin_slot": 0, + "target_id": 671, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 1697, + "origin_id": 673, + "origin_slot": 2, + "target_id": 671, + "target_slot": 1, + "type": "VAE" + }, + { + "id": 1698, + "origin_id": 674, + "origin_slot": 0, + "target_id": 671, + "target_slot": 2, + "type": "IMAGE" + }, + { + "id": 1699, + "origin_id": 671, + "origin_slot": 0, + "target_id": 672, + "target_slot": 0, + "type": "POSE_KEYPOINT" + }, + { + "id": 1701, + "origin_id": 672, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 1704, + "origin_id": -10, + "origin_slot": 0, + "target_id": 674, + "target_slot": 2, + "type": "INT" + }, + { + "id": 1705, + "origin_id": -10, + "origin_slot": 1, + "target_id": 674, + "target_slot": 3, + "type": "COMBO" + }, + { + "id": 1706, + "origin_id": -10, + "origin_slot": 2, + "target_id": 672, + "target_slot": 1, + "type": "BOOLEAN" + }, + { + "id": 1707, + "origin_id": -10, + "origin_slot": 3, + "target_id": 672, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 1708, + "origin_id": -10, + "origin_slot": 4, + "target_id": 672, + "target_slot": 3, + "type": "BOOLEAN" + }, + { + "id": 1709, + "origin_id": -10, + "origin_slot": 5, + "target_id": 672, + "target_slot": 4, + "type": "BOOLEAN" + }, + { + "id": 1710, + "origin_id": -10, + "origin_slot": 6, + "target_id": 672, + "target_slot": 5, + "type": "INT" + }, + { + "id": 1711, + "origin_id": -10, + "origin_slot": 7, + "target_id": 672, + "target_slot": 6, + "type": "INT" + }, + { + "id": 1712, + "origin_id": -10, + "origin_slot": 8, + "target_id": 672, + "target_slot": 7, + "type": "FLOAT" + }, + { + "id": 1715, + "origin_id": 677, + "origin_slot": 0, + "target_id": 678, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 1716, + "origin_id": 674, + "origin_slot": 0, + "target_id": 678, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 1717, + "origin_id": 678, + "origin_slot": 0, + "target_id": 671, + "target_slot": 3, + "type": "BOUNDING_BOX" + }, + { + "id": 1718, + "origin_id": -10, + "origin_slot": 9, + "target_id": 678, + "target_slot": 2, + "type": "FLOAT" + }, + { + "id": 1719, + "origin_id": -10, + "origin_slot": 10, + "target_id": 678, + "target_slot": 3, + "type": "COMBO" + }, + { + "id": 1720, + "origin_id": -10, + "origin_slot": 11, + "target_id": 678, + "target_slot": 4, + "type": "INT" + }, + { + "id": 1721, + "origin_id": -10, + "origin_slot": 12, + "target_id": 673, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 1722, + "origin_id": -10, + "origin_slot": 13, + "target_id": 677, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 1725, + "origin_id": 671, + "origin_slot": 0, + "target_id": -20, + "target_slot": 1, + "type": "POSE_KEYPOINT" + }, + { + "id": 1726, + "origin_id": 678, + "origin_slot": 0, + "target_id": -20, + "target_slot": 2, + "type": "BOUNDING_BOX" + }, + { + "id": 1741, + "origin_id": -10, + "origin_slot": 14, + "target_id": 692, + "target_slot": 0, + "type": "VIDEO" + }, + { + "id": 1742, + "origin_id": 692, + "origin_slot": 0, + "target_id": 674, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 1743, + "origin_id": 692, + "origin_slot": 1, + "target_id": -20, + "target_slot": 3, + "type": "AUDIO" + }, + { + "id": 1744, + "origin_id": 692, + "origin_slot": 2, + "target_id": -20, + "target_slot": 4, + "type": "FLOAT" + } + ], + "extra": { + "workflowRendererVersion": "LG" + }, + "category": "Conditioning & Preprocessors/Pose", + "description": "Extracts multi-person pose keypoints and skeleton frame sequences from video using SDPose with built-in person detection." + } + ] + }, + "extra": {} +} \ No newline at end of file diff --git a/comfy/background_removal/birefnet.py b/comfy/background_removal/birefnet.py index df54b2b90..78a80246e 100644 --- a/comfy/background_removal/birefnet.py +++ b/comfy/background_removal/birefnet.py @@ -105,7 +105,7 @@ class WindowAttention(nn.Module): relative_position_bias = self.relative_position_bias_table[self.relative_position_index.long().view(-1)].view( self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1) # Wh*Ww,Wh*Ww,nH - relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww + relative_position_bias = comfy.ops.cast_to_input(relative_position_bias.permute(2, 0, 1).contiguous(), attn) # nH, Wh*Ww, Wh*Ww attn = attn + relative_position_bias.unsqueeze(0) if mask is not None: diff --git a/comfy/bg_removal_model.py b/comfy/bg_removal_model.py index 6dec65e63..c772c5f6a 100644 --- a/comfy/bg_removal_model.py +++ b/comfy/bg_removal_model.py @@ -55,12 +55,7 @@ class BackgroundRemovalModel(): out = torch.nn.functional.interpolate(out, size=(H, W), mode="bicubic", antialias=False) mask = out.sigmoid().to(device=comfy.model_management.intermediate_device(), dtype=comfy.model_management.intermediate_dtype()) - if mask.ndim == 3: - mask = mask.unsqueeze(0) - if mask.shape[1] != 1: - mask = mask.movedim(-1, 1) - - return mask + return mask.squeeze(1) # (B, 1, H, W) -> (B, H, W) def load_background_removal_model(sd): diff --git a/comfy/cli_args.py b/comfy/cli_args.py index 9d88c8517..cba0dfa34 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -49,7 +49,7 @@ parser.add_argument("--temp-directory", type=str, default=None, help="Set the Co parser.add_argument("--input-directory", type=str, default=None, help="Set the ComfyUI input directory. Overrides --base-directory.") parser.add_argument("--auto-launch", action="store_true", help="Automatically launch ComfyUI in the default browser.") parser.add_argument("--disable-auto-launch", action="store_true", help="Disable auto launching the browser.") -parser.add_argument("--cuda-device", type=int, default=None, metavar="DEVICE_ID", help="Set the id of the cuda device this instance will use. All other devices will not be visible.") +parser.add_argument("--cuda-device", type=str, default=None, metavar="DEVICE_ID", help="Set the ids of cuda devices this instance will use, as a comma-separated list (e.g. '0' or '0,1'). All other devices will not be visible.") parser.add_argument("--default-device", type=int, default=None, metavar="DEFAULT_DEVICE_ID", help="Set the id of the default device, all other devices will stay visible.") cm_group = parser.add_mutually_exclusive_group() cm_group.add_argument("--cuda-malloc", action="store_true", help="Enable cudaMallocAsync (enabled by default for torch 2.0 and up).") @@ -111,7 +111,7 @@ parser.add_argument("--preview-method", type=LatentPreviewMethod, default=Latent parser.add_argument("--preview-size", type=int, default=512, help="Sets the maximum preview size for sampler nodes.") cache_group = parser.add_mutually_exclusive_group() -cache_group.add_argument("--cache-ram", nargs='*', type=float, default=[], metavar="GB", help="Use RAM pressure caching with the specified headroom thresholds. This is the default caching mode. The first value sets the active-cache threshold; the optional second value sets the inactive-cache/pin threshold. Defaults when no values are provided: active 25%% of system RAM (min 4GB, max 32GB), inactive 75%% of system RAM (min 12GB, max 96GB).") +cache_group.add_argument("--cache-ram", nargs='*', type=float, default=[], metavar="GB", help="Use RAM pressure caching with the specified headroom thresholds. This is the default caching mode. The first value sets the active-cache threshold; the optional second value sets the inactive-cache/pin threshold. Defaults when no values are provided: active 10%% of system RAM (min 2GB, max 10GB), inactive 100%% of system RAM (max 96GB).") cache_group.add_argument("--cache-classic", action="store_true", help="Use the old style (aggressive) caching.") cache_group.add_argument("--cache-lru", type=int, default=0, help="Use LRU caching with a maximum of N node results cached. May use more RAM/VRAM.") cache_group.add_argument("--cache-none", action="store_true", help="Reduced RAM/VRAM usage at the expense of executing every node for each run.") @@ -149,6 +149,7 @@ parser.add_argument("--async-offload", nargs='?', const=2, type=int, default=Non parser.add_argument("--disable-async-offload", action="store_true", help="Disable async weight offloading.") parser.add_argument("--disable-dynamic-vram", action="store_true", help="Disable dynamic VRAM and use estimate based model loading.") parser.add_argument("--enable-dynamic-vram", action="store_true", help="Enable dynamic VRAM on systems where it's not enabled by default.") +parser.add_argument("--fast-disk", action="store_true", help="Prefer disk-backed dynamic loading and offload over unpinned RAM. Can be faster for users with fast NVME disks.") parser.add_argument("--force-non-blocking", action="store_true", help="Force ComfyUI to use non-blocking operations for all applicable tensors. This may improve performance on some non-Nvidia systems but can cause issues with some workflows.") @@ -165,6 +166,8 @@ class PerformanceFeature(enum.Enum): parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. This is used to test new features so using it might crash your comfyui. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: {}".format(" ".join(map(lambda c: c.value, PerformanceFeature)))) +parser.add_argument("--debug-hang", action="store_true", help="Enable stack trace dumps on Ctrl-C for debugging hangs.") + parser.add_argument("--disable-pinned-memory", action="store_true", help="Disable pinned memory use.") parser.add_argument("--mmap-torch-files", action="store_true", help="Use mmap when loading ckpt/pt files.") diff --git a/comfy/clip_vision.py b/comfy/clip_vision.py index 71f2200b7..f8f5bc269 100644 --- a/comfy/clip_vision.py +++ b/comfy/clip_vision.py @@ -24,13 +24,16 @@ IMAGE_ENCODERS = { "siglip_vision_model": comfy.clip_model.CLIPVisionModelProjection, "siglip2_vision_model": comfy.clip_model.CLIPVisionModelProjection, "dinov2": comfy.image_encoders.dino2.Dinov2Model, - "dinov3": comfy.image_encoders.dino3.DINOv3ViTModel + "dinov3": comfy.image_encoders.dino3.DINOv3ViTModel, } class ClipVisionModel(): def __init__(self, json_config): - with open(json_config) as f: - config = json.load(f) + if isinstance(json_config, dict): + config = json_config + else: + with open(json_config) as f: + config = json.load(f) self.image_size = config.get("image_size", 224) self.image_mean = config.get("image_mean", [0.48145466, 0.4578275, 0.40821073]) @@ -136,8 +139,10 @@ def load_clipvision_from_sd(sd, prefix="", convert_keys=False): json_config = os.path.join(os.path.join(os.path.dirname(os.path.realpath(__file__)), "image_encoders"), "dino2_giant.json") elif 'encoder.layer.23.layer_scale2.lambda1' in sd: json_config = os.path.join(os.path.join(os.path.dirname(os.path.realpath(__file__)), "image_encoders"), "dino2_large.json") - elif 'layer.9.attention.o_proj.bias' in sd: # dinov3 + elif 'layer.9.attention.o_proj.bias' in sd: # dinov3 large (24 layers) json_config = os.path.join(os.path.join(os.path.dirname(os.path.realpath(__file__)), "image_encoders"), "dino3_large.json") + elif 'layer.0.mlp.gate_proj.weight' in sd and 'layer.31.norm1.weight' in sd: # Dinov3 ViT-H/16+ (SwiGLU gated MLP, 32 layers) + json_config = comfy.image_encoders.dino3.DINOV3_VITH_CONFIG else: return None diff --git a/comfy/comfy_types/node_typing.py b/comfy/comfy_types/node_typing.py index 57126fa4a..bb21eb1d1 100644 --- a/comfy/comfy_types/node_typing.py +++ b/comfy/comfy_types/node_typing.py @@ -1,6 +1,5 @@ """Comfy-specific type hinting""" -from __future__ import annotations from typing import Literal, TypedDict, Optional from typing_extensions import NotRequired from abc import ABC, abstractmethod diff --git a/comfy/controlnet.py b/comfy/controlnet.py index ba670b16d..6dbbaa959 100644 --- a/comfy/controlnet.py +++ b/comfy/controlnet.py @@ -15,13 +15,14 @@ You should have received a copy of the GNU General Public License along with this program. If not, see . """ - +from __future__ import annotations import torch from enum import Enum import math import os import logging +import copy import comfy.utils import comfy.model_management import comfy.model_detection @@ -38,7 +39,7 @@ import comfy.ldm.hydit.controlnet import comfy.ldm.flux.controlnet import comfy.ldm.qwen_image.controlnet import comfy.cldm.dit_embedder -from typing import TYPE_CHECKING +from typing import TYPE_CHECKING, Union if TYPE_CHECKING: from comfy.hooks import HookGroup @@ -64,6 +65,18 @@ class StrengthType(Enum): CONSTANT = 1 LINEAR_UP = 2 +class ControlIsolation: + '''Temporarily set a ControlBase object's previous_controlnet to None to prevent cascading calls.''' + def __init__(self, control: ControlBase): + self.control = control + self.orig_previous_controlnet = control.previous_controlnet + + def __enter__(self): + self.control.previous_controlnet = None + + def __exit__(self, *args): + self.control.previous_controlnet = self.orig_previous_controlnet + class ControlBase: def __init__(self): self.cond_hint_original = None @@ -77,7 +90,7 @@ class ControlBase: self.compression_ratio = 8 self.upscale_algorithm = 'nearest-exact' self.extra_args = {} - self.previous_controlnet = None + self.previous_controlnet: Union[ControlBase, None] = None self.extra_conds = [] self.strength_type = StrengthType.CONSTANT self.concat_mask = False @@ -85,6 +98,7 @@ class ControlBase: self.extra_concat = None self.extra_hooks: HookGroup = None self.preprocess_image = lambda a: a + self.multigpu_clones: dict[torch.device, ControlBase] = {} def set_cond_hint(self, cond_hint, strength=1.0, timestep_percent_range=(0.0, 1.0), vae=None, extra_concat=[]): self.cond_hint_original = cond_hint @@ -111,17 +125,38 @@ class ControlBase: def cleanup(self): if self.previous_controlnet is not None: self.previous_controlnet.cleanup() - + for device_cnet in self.multigpu_clones.values(): + with ControlIsolation(device_cnet): + device_cnet.cleanup() self.cond_hint = None self.extra_concat = None self.timestep_range = None def get_models(self): out = [] + for device_cnet in self.multigpu_clones.values(): + out += device_cnet.get_models_only_self() if self.previous_controlnet is not None: out += self.previous_controlnet.get_models() return out + def get_models_only_self(self): + 'Calls get_models, but temporarily sets previous_controlnet to None.' + with ControlIsolation(self): + return self.get_models() + + def get_instance_for_device(self, device): + 'Returns instance of this Control object intended for selected device.' + return self.multigpu_clones.get(device, self) + + def deepclone_multigpu(self, load_device, autoregister=False): + ''' + Create deep clone of Control object where model(s) is set to other devices. + + When autoregister is set to True, the deep clone is also added to multigpu_clones dict. + ''' + raise NotImplementedError("Classes inheriting from ControlBase should define their own deepclone_multigpu funtion.") + def get_extra_hooks(self): out = [] if self.extra_hooks is not None: @@ -130,7 +165,7 @@ class ControlBase: out += self.previous_controlnet.get_extra_hooks() return out - def copy_to(self, c): + def copy_to(self, c: ControlBase): c.cond_hint_original = self.cond_hint_original c.strength = self.strength c.timestep_percent_range = self.timestep_percent_range @@ -284,6 +319,14 @@ class ControlNet(ControlBase): self.copy_to(c) return c + def deepclone_multigpu(self, load_device, autoregister=False): + c = self.copy() + c.control_model = copy.deepcopy(c.control_model) + c.control_model_wrapped = comfy.model_patcher.ModelPatcher(c.control_model, load_device=load_device, offload_device=comfy.model_management.unet_offload_device()) + if autoregister: + self.multigpu_clones[load_device] = c + return c + def get_models(self): out = super().get_models() out.append(self.control_model_wrapped) @@ -314,6 +357,10 @@ class QwenFunControlNet(ControlNet): super().pre_run(model, percent_to_timestep_function) self.set_extra_arg("base_model", model.diffusion_model) + def cleanup(self): + self.extra_args.pop("base_model", None) + super().cleanup() + def copy(self): c = QwenFunControlNet(None, global_average_pooling=self.global_average_pooling, load_device=self.load_device, manual_cast_dtype=self.manual_cast_dtype) c.control_model = self.control_model @@ -906,6 +953,14 @@ class T2IAdapter(ControlBase): self.copy_to(c) return c + def deepclone_multigpu(self, load_device, autoregister=False): + c = self.copy() + c.t2i_model = copy.deepcopy(c.t2i_model) + c.device = load_device + if autoregister: + self.multigpu_clones[load_device] = c + return c + def load_t2i_adapter(t2i_data, model_options={}): #TODO: model_options compression_ratio = 8 upscale_algorithm = 'nearest-exact' diff --git a/comfy/float.py b/comfy/float.py index 184b3d6d0..3c82d6359 100644 --- a/comfy/float.py +++ b/comfy/float.py @@ -1,5 +1,20 @@ +import logging + import torch +_CK_STOCHASTIC_ROUNDING_AVAILABLE = False +try: + import comfy_kitchen as ck + _ck_stochastic_rounding_fp8 = ck.stochastic_rounding_fp8 + _CK_STOCHASTIC_ROUNDING_AVAILABLE = True +except (AttributeError, ImportError): + logging.warning("comfy_kitchen does not support stochastic FP8 rounding, please update comfy_kitchen.") + +if not _CK_STOCHASTIC_ROUNDING_AVAILABLE: + def _ck_stochastic_rounding_fp8(value, rng, dtype): + raise NotImplementedError("comfy_kitchen does not support stochastic FP8 rounding") + + def calc_mantissa(abs_x, exponent, normal_mask, MANTISSA_BITS, EXPONENT_BIAS, generator=None): mantissa_scaled = torch.where( normal_mask, @@ -57,6 +72,10 @@ def stochastic_rounding(value, dtype, seed=0): if dtype == torch.float8_e4m3fn or dtype == torch.float8_e5m2: generator = torch.Generator(device=value.device) generator.manual_seed(seed) + if _CK_STOCHASTIC_ROUNDING_AVAILABLE: + rng = torch.randint(0, 256, value.size(), dtype=torch.uint8, layout=value.layout, device=value.device, generator=generator) + return _ck_stochastic_rounding_fp8(value, rng, dtype) + output = torch.empty_like(value, dtype=dtype) num_slices = max(1, (value.numel() / (4096 * 4096))) slice_size = max(1, round(value.shape[0] / num_slices)) diff --git a/comfy/image_encoders/dino2.py b/comfy/image_encoders/dino2.py index ee86f8309..53e4fdb6c 100644 --- a/comfy/image_encoders/dino2.py +++ b/comfy/image_encoders/dino2.py @@ -1,7 +1,13 @@ import torch +import torch.nn.functional as F + from comfy.text_encoders.bert import BertAttention import comfy.model_management from comfy.ldm.modules.attention import optimized_attention_for_device +from comfy.ldm.depth_anything_3.reference_view_selector import ( + select_reference_view, reorder_by_reference, restore_original_order, + THRESH_FOR_REF_SELECTION, +) class Dino2AttentionOutput(torch.nn.Module): @@ -14,13 +20,41 @@ class Dino2AttentionOutput(torch.nn.Module): class Dino2AttentionBlock(torch.nn.Module): - def __init__(self, embed_dim, heads, layer_norm_eps, dtype, device, operations): + def __init__(self, embed_dim, heads, layer_norm_eps, dtype, device, operations, + qk_norm=False): super().__init__() + self.heads = heads + self.head_dim = embed_dim // heads self.attention = BertAttention(embed_dim, heads, dtype, device, operations) self.output = Dino2AttentionOutput(embed_dim, embed_dim, layer_norm_eps, dtype, device, operations) + if qk_norm: + self.q_norm = operations.LayerNorm(self.head_dim, dtype=dtype, device=device) + self.k_norm = operations.LayerNorm(self.head_dim, dtype=dtype, device=device) + else: + self.q_norm = None + self.k_norm = None - def forward(self, x, mask, optimized_attention): - return self.output(self.attention(x, mask, optimized_attention)) + def forward(self, x, mask, optimized_attention, pos=None, rope=None): + # Fast path used by the existing CLIP-vision DINOv2 (no DA3 extensions). + if self.q_norm is None and rope is None: + return self.output(self.attention(x, mask, optimized_attention)) + + # DA3 path: do QKV manually so we can apply per-head QK-norm and 2D RoPE. + attn = self.attention + B, N, C = x.shape + h = self.heads + d = self.head_dim + q = attn.query(x).view(B, N, h, d).transpose(1, 2) + k = attn.key(x).view(B, N, h, d).transpose(1, 2) + v = attn.value(x).view(B, N, h, d).transpose(1, 2) + if self.q_norm is not None: + q = self.q_norm(q) + k = self.k_norm(k) + if rope is not None and pos is not None: + q = rope(q, pos) + k = rope(k, pos) + out = optimized_attention(q, k, v, h, mask=mask, skip_reshape=True) + return self.output(out) class LayerScale(torch.nn.Module): @@ -64,9 +98,11 @@ class SwiGLUFFN(torch.nn.Module): class Dino2Block(torch.nn.Module): - def __init__(self, dim, num_heads, layer_norm_eps, dtype, device, operations, use_swiglu_ffn): + def __init__(self, dim, num_heads, layer_norm_eps, dtype, device, operations, use_swiglu_ffn, + qk_norm=False): super().__init__() - self.attention = Dino2AttentionBlock(dim, num_heads, layer_norm_eps, dtype, device, operations) + self.attention = Dino2AttentionBlock(dim, num_heads, layer_norm_eps, dtype, device, operations, + qk_norm=qk_norm) self.layer_scale1 = LayerScale(dim, dtype, device, operations) self.layer_scale2 = LayerScale(dim, dtype, device, operations) if use_swiglu_ffn: @@ -76,19 +112,90 @@ class Dino2Block(torch.nn.Module): self.norm1 = operations.LayerNorm(dim, eps=layer_norm_eps, dtype=dtype, device=device) self.norm2 = operations.LayerNorm(dim, eps=layer_norm_eps, dtype=dtype, device=device) - def forward(self, x, optimized_attention): - x = x + self.layer_scale1(self.attention(self.norm1(x), None, optimized_attention)) + def forward(self, x, optimized_attention, pos=None, rope=None, attn_mask=None): + x = x + self.layer_scale1(self.attention(self.norm1(x), attn_mask, optimized_attention, + pos=pos, rope=rope)) x = x + self.layer_scale2(self.mlp(self.norm2(x))) return x -class Dino2Encoder(torch.nn.Module): - def __init__(self, dim, num_heads, layer_norm_eps, num_layers, dtype, device, operations, use_swiglu_ffn): +# ----------------------------------------------------------------------------- +# 2D Rotary position embedding (DA3 extension) +# ----------------------------------------------------------------------------- + + +class _PositionGetter: + """Cache (h, w) -> flat (y, x) position grid used to feed ``rope``.""" + + def __init__(self): + self._cache: dict = {} + + def __call__(self, batch_size: int, height: int, width: int, device) -> torch.Tensor: + key = (height, width, device) + if key not in self._cache: + y = torch.arange(height, device=device) + x = torch.arange(width, device=device) + self._cache[key] = torch.cartesian_prod(y, x) + cached = self._cache[key] + return cached.view(1, height * width, 2).expand(batch_size, -1, -1).clone() + + +class RotaryPositionEmbedding2D(torch.nn.Module): + """2D RoPE used by DA3-Small/Base. No learnable parameters.""" + + def __init__(self, frequency: float = 100.0): super().__init__() - self.layer = torch.nn.ModuleList([Dino2Block(dim, num_heads, layer_norm_eps, dtype, device, operations, use_swiglu_ffn = use_swiglu_ffn) - for _ in range(num_layers)]) + self.base_frequency = frequency + self._freq_cache: dict = {} + + def _components(self, dim: int, seq_len: int, device, dtype): + key = (dim, seq_len, device, dtype) + if key not in self._freq_cache: + exp = torch.arange(0, dim, 2, device=device).float() / dim + inv_freq = 1.0 / (self.base_frequency ** exp) + pos = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + ang = torch.einsum("i,j->ij", pos, inv_freq) + ang = ang.to(dtype) + ang = torch.cat((ang, ang), dim=-1) + self._freq_cache[key] = (ang.cos().to(dtype), ang.sin().to(dtype)) + return self._freq_cache[key] + + @staticmethod + def _rotate(x: torch.Tensor) -> torch.Tensor: + d = x.shape[-1] + x1, x2 = x[..., : d // 2], x[..., d // 2:] + return torch.cat((-x2, x1), dim=-1) + + def _apply_1d(self, tokens, positions, cos_c, sin_c): + cos = F.embedding(positions, cos_c)[:, None, :, :] + sin = F.embedding(positions, sin_c)[:, None, :, :] + return (tokens * cos) + (self._rotate(tokens) * sin) + + def forward(self, tokens: torch.Tensor, positions: torch.Tensor) -> torch.Tensor: + feature_dim = tokens.size(-1) // 2 + max_pos = int(positions.max()) + 1 + cos_c, sin_c = self._components(feature_dim, max_pos, tokens.device, tokens.dtype) + v, h = tokens.chunk(2, dim=-1) + v = self._apply_1d(v, positions[..., 0], cos_c, sin_c) + h = self._apply_1d(h, positions[..., 1], cos_c, sin_c) + return torch.cat((v, h), dim=-1) + + +class Dino2Encoder(torch.nn.Module): + def __init__(self, dim, num_heads, layer_norm_eps, num_layers, dtype, device, operations, use_swiglu_ffn, + qknorm_start: int = -1): + super().__init__() + self.layer = torch.nn.ModuleList([ + Dino2Block( + dim, num_heads, layer_norm_eps, dtype, device, operations, + use_swiglu_ffn=use_swiglu_ffn, + qk_norm=(qknorm_start != -1 and i >= qknorm_start), + ) + for i in range(num_layers) + ]) def forward(self, x, intermediate_output=None): + # Backward-compat path used by ``ClipVisionModel`` (no DA3 extensions). optimized_attention = optimized_attention_for_device(x.device, False, small_input=True) if intermediate_output is not None: @@ -122,16 +229,27 @@ class Dino2PatchEmbeddings(torch.nn.Module): class Dino2Embeddings(torch.nn.Module): - def __init__(self, dim, dtype, device, operations): + def __init__(self, dim, dtype, device, operations, + patch_size: int = 14, image_size: int = 518, + use_mask_token: bool = True, + num_camera_tokens: int = 0): super().__init__() - patch_size = 14 - image_size = 518 self.patch_size = patch_size + self.image_size = image_size self.patch_embeddings = Dino2PatchEmbeddings(dim, patch_size=patch_size, image_size=image_size, dtype=dtype, device=device, operations=operations) self.position_embeddings = torch.nn.Parameter(torch.empty(1, (image_size // patch_size) ** 2 + 1, dim, dtype=dtype, device=device)) self.cls_token = torch.nn.Parameter(torch.empty(1, 1, dim, dtype=dtype, device=device)) # mask_token is a pre-training param, kept only so strict loading accepts the key. - self.mask_token = torch.nn.Parameter(torch.empty(1, dim, dtype=dtype, device=device)) + if use_mask_token: + self.mask_token = torch.nn.Parameter(torch.empty(1, dim, dtype=dtype, device=device)) + else: + self.mask_token = None + if num_camera_tokens > 0: + # DA3 stores (ref_token, src_token) pairs that get injected at the + # alt-attn boundary; see ``Dinov2Model._inject_camera_token``. + self.camera_token = torch.nn.Parameter(torch.empty(1, num_camera_tokens, dim, dtype=dtype, device=device)) + else: + self.camera_token = None def interpolate_pos_encoding(self, x, h_pixels, w_pixels): pos_embed = comfy.model_management.cast_to_device(self.position_embeddings, x.device, torch.float32) @@ -140,12 +258,22 @@ class Dino2Embeddings(torch.nn.Module): patch_pos = pos_embed[:, 1:] N = patch_pos.shape[1] M = int(N ** 0.5) + assert N == M * M, f"DINOv2 position grid must be square, got N={N} patches (sqrt={M})" h0 = h_pixels // self.patch_size w0 = w_pixels // self.patch_size - scale_factor = ((h0 + 0.1) / M, (w0 + 0.1) / M) # +0.1 matches upstream DINOv2's FP-rounding workaround so the interpolate output size lands on (h0, w0). + # +0.1 matches upstream DINOv2's FP-rounding workaround so the interpolate output size lands on (h0, w0). + # scale_factor is (height_scale, width_scale) -- height MUST come first; + # swapping these only happens to work for square inputs and breaks + # non-square paths like DA3-Small / DA3-Base multi-view. + scale_factor = ((h0 + 0.1) / M, (w0 + 0.1) / M) patch_pos = patch_pos.reshape(1, M, M, -1).permute(0, 3, 1, 2) patch_pos = torch.nn.functional.interpolate(patch_pos, scale_factor=scale_factor, mode="bicubic", antialias=False) + assert (h0, w0) == patch_pos.shape[-2:], ( + f"Interpolated pos-embed grid {tuple(patch_pos.shape[-2:])} does not match " + f"target patch grid ({h0}, {w0}) for input {h_pixels}x{w_pixels} (patch_size={self.patch_size}); " + f"check scale_factor axis order and +0.1 rounding workaround" + ) patch_pos = patch_pos.permute(0, 2, 3, 1).flatten(1, 2) return torch.cat((class_pos, patch_pos), dim=1).to(x.dtype) @@ -168,12 +296,51 @@ class Dinov2Model(torch.nn.Module): heads = config_dict["num_attention_heads"] layer_norm_eps = config_dict["layer_norm_eps"] use_swiglu_ffn = config_dict["use_swiglu_ffn"] + patch_size = config_dict.get("patch_size", 14) + image_size = config_dict.get("image_size", 518) + use_mask_token = config_dict.get("use_mask_token", True) - self.embeddings = Dino2Embeddings(dim, dtype, device, operations) - self.encoder = Dino2Encoder(dim, heads, layer_norm_eps, num_layers, dtype, device, operations, use_swiglu_ffn = use_swiglu_ffn) + # DA3 extensions (all default to disabled). + self.alt_start = config_dict.get("alt_start", -1) + self.qknorm_start = config_dict.get("qknorm_start", -1) + self.rope_start = config_dict.get("rope_start", -1) + self.cat_token = config_dict.get("cat_token", False) + rope_freq = config_dict.get("rope_freq", 100.0) + + self.embed_dim = dim + self.patch_size = patch_size + self.num_register_tokens = 0 + self.patch_start_idx = 1 + + if self.rope_start != -1 and rope_freq > 0: + self.rope = RotaryPositionEmbedding2D(frequency=rope_freq) + self._position_getter = _PositionGetter() + else: + self.rope = None + self._position_getter = None + + # camera_token shape: (1, 2, dim) -> (ref_token, src_token). + num_cam_tokens = 2 if self.alt_start != -1 else 0 + + self.embeddings = Dino2Embeddings( + dim, dtype, device, operations, + patch_size=patch_size, image_size=image_size, + use_mask_token=use_mask_token, num_camera_tokens=num_cam_tokens, + ) + self.encoder = Dino2Encoder( + dim, heads, layer_norm_eps, num_layers, dtype, device, operations, + use_swiglu_ffn=use_swiglu_ffn, + qknorm_start=self.qknorm_start, + ) self.layernorm = operations.LayerNorm(dim, eps=layer_norm_eps, dtype=dtype, device=device) def forward(self, pixel_values, attention_mask=None, intermediate_output=None): + if self.alt_start != -1: + raise RuntimeError( + "Dinov2Model.forward() is the backward-compatible CLIP-vision path and does not " + "apply DA3 extensions (RoPE, alternating attention, camera-token injection). " + "Use get_intermediate_layers_da3() for Depth Anything 3 models." + ) x = self.embeddings(pixel_values) x, i = self.encoder(x, intermediate_output=intermediate_output) x = self.layernorm(x) @@ -181,6 +348,7 @@ class Dinov2Model(torch.nn.Module): return x, i, pooled_output, None def get_intermediate_layers(self, pixel_values, indices, apply_norm=True): + """Single-view multi-layer feature extraction.""" x = self.embeddings(pixel_values) optimized_attention = optimized_attention_for_device(x.device, False, small_input=True) n_layers = len(self.encoder.layer) @@ -197,3 +365,132 @@ class Dinov2Model(torch.nn.Module): if i >= max_idx: break return [cache[i] for i in resolved] + + # ------------------------------------------------------------------ + # Depth Anything 3 forward + # ------------------------------------------------------------------ + def _prepare_rope_positions(self, B, S, H, W, device): + if self.rope is None: + return None, None + ph, pw = H // self.patch_size, W // self.patch_size + pos = self._position_getter(B * S, ph, pw, device=device) + # Shift so the cls/cam token at position 0 is reserved for "no diff". + pos = pos + 1 + cls_pos = torch.zeros(B * S, self.patch_start_idx, 2, device=device, dtype=pos.dtype) + # Per-view local: real grid positions for patches, 0 for cls token. + pos_local = torch.cat([cls_pos, pos], dim=1) + # Global (across views): same grid positions; cls token still at 0, + # but patches share the same positions in every view. + pos_global = torch.cat([cls_pos, torch.zeros_like(pos) + 1], dim=1) + return pos_local, pos_global + + def _inject_camera_token(self, x: torch.Tensor, B: int, S: int, cam_token: "torch.Tensor | None") -> torch.Tensor: + # x: (B, S, N, C). Replace token at index 0 with the camera token. + if cam_token is not None: + inj = cam_token + else: + ct = comfy.model_management.cast_to_device(self.embeddings.camera_token, x.device, x.dtype) + ref_token = ct[:, :1].expand(B, -1, -1) + src_token = ct[:, 1:].expand(B, max(S - 1, 0), -1) + inj = torch.cat([ref_token, src_token], dim=1) + x = x.clone() + x[:, :, 0] = inj + return x + + def get_intermediate_layers_da3(self, pixel_values, out_layers, cam_token=None, ref_view_strategy="saddle_balanced", export_feat_layers=None): + """Multi-view multi-layer feature extraction used by Depth Anything 3.""" + if pixel_values.ndim == 4: + pixel_values = pixel_values.unsqueeze(1) + assert pixel_values.ndim == 5 and pixel_values.shape[2] == 3, \ + f"expected (B,3,H,W) or (B,S,3,H,W); got {tuple(pixel_values.shape)}" + B, S, _, H, W = pixel_values.shape + + # Patch + cls + (interpolated) pos embed for each view. + x = pixel_values.reshape(B * S, 3, H, W) + x = self.embeddings(x) # (B*S, 1+N, C) + x = x.reshape(B, S, x.shape[-2], x.shape[-1]) # (B, S, 1+N, C) + + pos_local, pos_global = self._prepare_rope_positions(B, S, H, W, x.device) + # optimized_attention is only used by blocks without QK-norm/RoPE + # (vanilla DINOv2 path); enabling-aware blocks fall through to SDPA. + optimized_attention = optimized_attention_for_device(x.device, False, small_input=True) + + out_set = set(out_layers) + export_set = set(export_feat_layers) if export_feat_layers else set() + outputs: list[torch.Tensor] = [] + aux_outputs: list[torch.Tensor] = [] + local_x = x + b_idx = None + + + for i, blk in enumerate(self.encoder.layer): + apply_rope = self.rope is not None and i >= self.rope_start + block_rope = self.rope if apply_rope else None + l_pos = pos_local if apply_rope else None + g_pos = pos_global if apply_rope else None + + # Reference-view selection threshold: matches the upstream constant + # THRESH_FOR_REF_SELECTION = 3. Skipped when a user-supplied + # cam_token is provided (camera info already pins the geometry). + if (self.alt_start != -1 and i == self.alt_start - 1 and S >= THRESH_FOR_REF_SELECTION and cam_token is None): + b_idx = select_reference_view(x, strategy=ref_view_strategy) + x = reorder_by_reference(x, b_idx) + local_x = reorder_by_reference(local_x, b_idx) + + if self.alt_start != -1 and i == self.alt_start: + x = self._inject_camera_token(x, B, S, cam_token) + + if self.alt_start != -1 and i >= self.alt_start and (i % 2 == 1): + # Global attention across views: flatten S into the seq dim. + t = x.reshape(B, S * x.shape[-2], x.shape[-1]) + p = g_pos.reshape(B, S * g_pos.shape[-2], g_pos.shape[-1]) if g_pos is not None else None + t = blk(t, optimized_attention=optimized_attention, pos=p, rope=block_rope) + x = t.reshape(B, S, x.shape[-2], x.shape[-1]) + else: + # Per-view local attention. + t = x.reshape(B * S, x.shape[-2], x.shape[-1]) + p = l_pos.reshape(B * S, l_pos.shape[-2], l_pos.shape[-1]) if l_pos is not None else None + t = blk(t, optimized_attention=optimized_attention, pos=p, rope=block_rope) + x = t.reshape(B, S, x.shape[-2], x.shape[-1]) + local_x = x + + if i in out_set: + if self.cat_token: + out_x = torch.cat([local_x, x], dim=-1) + else: + out_x = x + # Restore original view order on the way out so heads see views + # in the user's expected order. + if b_idx is not None and self.alt_start != -1: + out_x = restore_original_order(out_x, b_idx) + outputs.append(out_x) + + if i in export_set: + aux = x + if b_idx is not None and self.alt_start != -1: + aux = restore_original_order(aux, b_idx) + aux_outputs.append(aux) + + # Apply final norm. When cat_token is set, only the right half + # ("global" features) is normalised; the left half is left as-is to + # match the upstream DA3 head signature. + normed: list[torch.Tensor] = [] + cls_tokens: list[torch.Tensor] = [] + for out_x in outputs: + cls_tokens.append(out_x[:, :, 0]) + if out_x.shape[-1] == self.embed_dim: + normed.append(self.layernorm(out_x)) + elif out_x.shape[-1] == self.embed_dim * 2: + left = out_x[..., :self.embed_dim] + right = self.layernorm(out_x[..., self.embed_dim:]) + normed.append(torch.cat([left, right], dim=-1)) + else: + raise ValueError(f"Unexpected token width: {out_x.shape[-1]}") + + # Drop cls/cam token from the patch sequence. + normed = [o[..., 1 + self.num_register_tokens:, :] for o in normed] + + # Final layernorm + drop cls token from auxiliary features too. + aux_normed = [self.layernorm(o)[..., 1 + self.num_register_tokens:, :] + for o in aux_outputs] + return list(zip(normed, cls_tokens)), aux_normed diff --git a/comfy/image_encoders/dino3.py b/comfy/image_encoders/dino3.py index 2835641f7..ad29b06f8 100644 --- a/comfy/image_encoders/dino3.py +++ b/comfy/image_encoders/dino3.py @@ -3,10 +3,31 @@ import torch import torch.nn as nn import torch.nn.functional as F -import comfy.model_management +import comfy.ops from comfy.ldm.modules.attention import optimized_attention_for_device from comfy.image_encoders.dino2 import LayerScale as DINOv3ViTLayerScale + +# DINOv3 ViT-H/16+ (SwiGLU) +DINOV3_VITH_CONFIG = { + "model_type": "dinov3", + "num_hidden_layers": 32, + "hidden_size": 1280, + "num_attention_heads": 20, + "num_register_tokens": 4, + "intermediate_size": 5120, + "layer_norm_eps": 1e-5, + "num_channels": 3, + "patch_size": 16, + "rope_theta": 100.0, + "use_gated_mlp": True, + "gated_mlp_act": "silu", + "image_size": 1024, + "image_mean": [0.485, 0.456, 0.406], + "image_std": [0.229, 0.224, 0.225], +} + + class DINOv3ViTMLP(nn.Module): def __init__(self, hidden_size, intermediate_size, mlp_bias, device, dtype, operations): super().__init__() @@ -19,10 +40,13 @@ class DINOv3ViTMLP(nn.Module): def forward(self, x): return self.down_proj(self.act_fn(self.up_proj(x))) + def rotate_half(x): x1 = x[..., : x.shape[-1] // 2] x2 = x[..., x.shape[-1] // 2 :] return torch.cat((-x2, x1), dim=-1) + + def apply_rotary_pos_emb(q, k, cos, sin, **kwargs): num_tokens = q.shape[-2] num_patches = sin.shape[-2] @@ -39,6 +63,7 @@ def apply_rotary_pos_emb(q, k, cos, sin, **kwargs): return q, k + class DINOv3ViTAttention(nn.Module): def __init__(self, hidden_size, num_attention_heads, device, dtype, operations): super().__init__() @@ -46,20 +71,12 @@ class DINOv3ViTAttention(nn.Module): self.num_heads = num_attention_heads self.head_dim = self.embed_dim // self.num_heads - self.k_proj = operations.Linear(self.embed_dim, self.embed_dim, bias=False, device=device, dtype=dtype) # key_bias = False + self.k_proj = operations.Linear(self.embed_dim, self.embed_dim, bias=False, device=device, dtype=dtype) # key_bias = False self.v_proj = operations.Linear(self.embed_dim, self.embed_dim, bias=True, device=device, dtype=dtype) - self.q_proj = operations.Linear(self.embed_dim, self.embed_dim, bias=True, device=device, dtype=dtype) self.o_proj = operations.Linear(self.embed_dim, self.embed_dim, bias=True, device=device, dtype=dtype) - def forward( - self, - hidden_states: torch.Tensor, - attention_mask: torch.Tensor | None = None, - position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None, - **kwargs, - ) -> tuple[torch.Tensor, torch.Tensor | None]: - + def forward(self, hidden_states, attention_mask=None, position_embeddings=None, **kwargs): batch_size, patches, _ = hidden_states.size() query_states = self.q_proj(hidden_states) @@ -75,7 +92,6 @@ class DINOv3ViTAttention(nn.Module): query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) attn = optimized_attention_for_device(query_states.device, mask=False) - attn_output = attn( query_states, key_states, value_states, self.num_heads, attention_mask, skip_reshape=True, skip_output_reshape=True, low_precision_attention=False, @@ -84,27 +100,24 @@ class DINOv3ViTAttention(nn.Module): attn_output = attn_output.transpose(1, 2) attn_output = attn_output.reshape(batch_size, patches, -1).contiguous() attn_output = self.o_proj(attn_output) - return attn_output + class DINOv3ViTGatedMLP(nn.Module): - def __init__(self, hidden_size, intermediate_size, mlp_bias, device, dtype, operations): + def __init__(self, hidden_size, intermediate_size, mlp_bias, device, dtype, operations, act="silu"): super().__init__() self.hidden_size = hidden_size self.intermediate_size = intermediate_size self.gate_proj = operations.Linear(self.hidden_size, self.intermediate_size, bias=mlp_bias, device=device, dtype=dtype) self.up_proj = operations.Linear(self.hidden_size, self.intermediate_size, bias=mlp_bias, device=device, dtype=dtype) self.down_proj = operations.Linear(self.intermediate_size, self.hidden_size, bias=mlp_bias, device=device, dtype=dtype) - self.act_fn = torch.nn.GELU() + self.act_fn = torch.nn.SiLU() if act == "silu" else torch.nn.GELU() def forward(self, x): - down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x)) - return down_proj + return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x)) -def get_patches_center_coordinates( - num_patches_h: int, num_patches_w: int, dtype: torch.dtype, device: torch.device -) -> torch.Tensor: +def get_patches_center_coordinates(num_patches_h, num_patches_w, dtype, device): coords_h = torch.arange(0.5, num_patches_h, dtype=dtype, device=device) coords_w = torch.arange(0.5, num_patches_w, dtype=dtype, device=device) coords_h = coords_h / num_patches_h @@ -114,105 +127,79 @@ def get_patches_center_coordinates( coords = 2.0 * coords - 1.0 return coords + class DINOv3ViTRopePositionEmbedding(nn.Module): inv_freq: torch.Tensor - def __init__(self, rope_theta, hidden_size, num_attention_heads, image_size, patch_size, device, dtype): + def __init__(self, rope_theta, hidden_size, num_attention_heads, patch_size, device, dtype): super().__init__() self.base = rope_theta self.head_dim = hidden_size // num_attention_heads - self.num_patches_h = image_size // patch_size - self.num_patches_w = image_size // patch_size self.patch_size = patch_size inv_freq = 1 / self.base ** torch.arange(0, 1, 4 / self.head_dim, dtype=torch.float32, device=device) self.register_buffer("inv_freq", inv_freq, persistent=False) - def forward(self, pixel_values: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: + def forward(self, pixel_values): _, _, height, width = pixel_values.shape num_patches_h = height // self.patch_size num_patches_w = width // self.patch_size - device = pixel_values.device - device_type = device.type if isinstance(device.type, str) and device.type != "mps" else "cpu" - with torch.amp.autocast(device_type = device_type, enabled=False): - patch_coords = get_patches_center_coordinates( - num_patches_h, num_patches_w, dtype=torch.float32, device=device - ) - - self.inv_freq = self.inv_freq.to(device) - angles = 2 * math.pi * patch_coords[:, :, None] * self.inv_freq[None, None, :] - angles = angles.flatten(1, 2) - angles = angles.tile(2) - - cos = torch.cos(angles) - sin = torch.sin(angles) - - dtype = pixel_values.dtype - return cos.to(dtype=dtype), sin.to(dtype=dtype) + patch_coords = get_patches_center_coordinates(num_patches_h, num_patches_w, dtype=torch.float32, device=pixel_values.device) + self.inv_freq = self.inv_freq.to(pixel_values.device) + angles = 2 * math.pi * patch_coords[:, :, None] * self.inv_freq[None, None, :] + angles = angles.flatten(1, 2) + angles = angles.tile(2) + cos = torch.cos(angles).to(dtype=pixel_values.dtype) + sin = torch.sin(angles).to(dtype=pixel_values.dtype) + return cos, sin class DINOv3ViTEmbeddings(nn.Module): def __init__(self, hidden_size, num_register_tokens, num_channels, patch_size, dtype, device, operations): super().__init__() - self.cls_token = nn.Parameter(torch.randn(1, 1, hidden_size, device=device, dtype=dtype)) - self.mask_token = nn.Parameter(torch.zeros(1, 1, hidden_size, device=device, dtype=dtype)) + self.cls_token = nn.Parameter(torch.empty(1, 1, hidden_size, device=device, dtype=dtype)) + self.mask_token = nn.Parameter(torch.empty(1, 1, hidden_size, device=device, dtype=dtype)) self.register_tokens = nn.Parameter(torch.empty(1, num_register_tokens, hidden_size, device=device, dtype=dtype)) self.patch_embeddings = operations.Conv2d( num_channels, hidden_size, kernel_size=patch_size, stride=patch_size, device=device, dtype=dtype ) - def forward(self, pixel_values: torch.Tensor, bool_masked_pos: torch.Tensor | None = None): + def forward(self, pixel_values, bool_masked_pos=None): batch_size = pixel_values.shape[0] - target_dtype = self.patch_embeddings.weight.dtype - patch_embeddings = self.patch_embeddings(pixel_values.to(dtype=target_dtype)) + patch_embeddings = self.patch_embeddings(pixel_values) patch_embeddings = patch_embeddings.flatten(2).transpose(1, 2) if bool_masked_pos is not None: - mask_token = self.mask_token.to(patch_embeddings.dtype) + mask_token = comfy.ops.cast_to_input(self.mask_token, patch_embeddings) patch_embeddings = torch.where(bool_masked_pos.unsqueeze(-1), mask_token, patch_embeddings) - cls_token = self.cls_token.expand(batch_size, -1, -1) - register_tokens = self.register_tokens.expand(batch_size, -1, -1) - device = patch_embeddings.device - cls_token = cls_token.to(device) - register_tokens = register_tokens.to(device) + cls_token = comfy.ops.cast_to_input(self.cls_token.expand(batch_size, -1, -1), patch_embeddings) + register_tokens = comfy.ops.cast_to_input(self.register_tokens.expand(batch_size, -1, -1), patch_embeddings) embeddings = torch.cat([cls_token, register_tokens, patch_embeddings], dim=1) - return embeddings + class DINOv3ViTLayer(nn.Module): - - def __init__(self, hidden_size, layer_norm_eps, use_gated_mlp, mlp_bias, intermediate_size, num_attention_heads, - device, dtype, operations): + def __init__(self, hidden_size, layer_norm_eps, use_gated_mlp, mlp_bias, intermediate_size, + num_attention_heads, device, dtype, operations, gated_mlp_act="silu"): super().__init__() - self.norm1 = operations.LayerNorm(hidden_size, eps=layer_norm_eps, device=device, dtype=dtype) self.attention = DINOv3ViTAttention(hidden_size, num_attention_heads, device=device, dtype=dtype, operations=operations) self.layer_scale1 = DINOv3ViTLayerScale(hidden_size, device=device, dtype=dtype, operations=None) self.norm2 = operations.LayerNorm(hidden_size, eps=layer_norm_eps, device=device, dtype=dtype) - if use_gated_mlp: - self.mlp = DINOv3ViTGatedMLP(hidden_size, intermediate_size, mlp_bias, device=device, dtype=dtype, operations=operations) + self.mlp = DINOv3ViTGatedMLP(hidden_size, intermediate_size, mlp_bias, device=device, dtype=dtype, operations=operations, act=gated_mlp_act) else: self.mlp = DINOv3ViTMLP(hidden_size, intermediate_size=intermediate_size, mlp_bias=mlp_bias, device=device, dtype=dtype, operations=operations) self.layer_scale2 = DINOv3ViTLayerScale(hidden_size, device=device, dtype=dtype, operations=None) - def forward( - self, - hidden_states: torch.Tensor, - attention_mask: torch.Tensor | None = None, - position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None, - ) -> torch.Tensor: + def forward(self, hidden_states, attention_mask=None, position_embeddings=None): residual = hidden_states hidden_states = self.norm1(hidden_states) - hidden_states = self.attention( - hidden_states, - attention_mask=attention_mask, - position_embeddings=position_embeddings, - ) + hidden_states = self.attention(hidden_states, attention_mask=attention_mask, position_embeddings=position_embeddings) hidden_states = self.layer_scale1(hidden_states) hidden_states = hidden_states + residual @@ -221,18 +208,12 @@ class DINOv3ViTLayer(nn.Module): hidden_states = self.mlp(hidden_states) hidden_states = self.layer_scale2(hidden_states) hidden_states = hidden_states + residual - return hidden_states class DINOv3ViTModel(nn.Module): def __init__(self, config, dtype, device, operations): super().__init__() - use_bf16 = comfy.model_management.should_use_bf16(device, prioritize_performance=True) - if dtype == torch.float16 and use_bf16: - dtype = torch.bfloat16 - elif dtype == torch.float16 and not use_bf16: - dtype = torch.float32 num_hidden_layers = config["num_hidden_layers"] hidden_size = config["hidden_size"] num_attention_heads = config["num_attention_heads"] @@ -242,45 +223,37 @@ class DINOv3ViTModel(nn.Module): num_channels = config["num_channels"] patch_size = config["patch_size"] rope_theta = config["rope_theta"] + use_gated_mlp = config.get("use_gated_mlp", False) + gated_mlp_act = config.get("gated_mlp_act", "silu") self.embeddings = DINOv3ViTEmbeddings( - hidden_size, num_register_tokens, num_channels=num_channels, patch_size=patch_size, dtype=dtype, device=device, operations=operations + hidden_size, num_register_tokens, num_channels=num_channels, patch_size=patch_size, + dtype=dtype, device=device, operations=operations ) self.rope_embeddings = DINOv3ViTRopePositionEmbedding( - rope_theta, hidden_size, num_attention_heads, image_size=512, patch_size=patch_size, dtype=dtype, device=device + rope_theta, hidden_size, num_attention_heads, patch_size=patch_size, dtype=dtype, device=device ) - self.layer = nn.ModuleList( - [DINOv3ViTLayer(hidden_size, layer_norm_eps, use_gated_mlp=False, mlp_bias=True, - intermediate_size=intermediate_size,num_attention_heads = num_attention_heads, - dtype=dtype, device=device, operations=operations) + self.layer = nn.ModuleList([ + DINOv3ViTLayer(hidden_size, layer_norm_eps, use_gated_mlp=use_gated_mlp, mlp_bias=True, + intermediate_size=intermediate_size, num_attention_heads=num_attention_heads, + dtype=dtype, device=device, operations=operations, gated_mlp_act=gated_mlp_act) for _ in range(num_hidden_layers)]) self.norm = operations.LayerNorm(hidden_size, eps=layer_norm_eps, dtype=dtype, device=device) def get_input_embeddings(self): return self.embeddings.patch_embeddings - def forward( - self, - pixel_values: torch.Tensor, - bool_masked_pos: torch.Tensor | None = None, - **kwargs, - ): - - pixel_values = pixel_values.to(self.embeddings.patch_embeddings.weight.dtype) + def forward(self, pixel_values, bool_masked_pos=None, **kwargs): hidden_states = self.embeddings(pixel_values, bool_masked_pos=bool_masked_pos) position_embeddings = self.rope_embeddings(pixel_values) - for i, layer_module in enumerate(self.layer): - hidden_states = layer_module( - hidden_states, - position_embeddings=position_embeddings, - ) + for layer_module in self.layer: + hidden_states = layer_module(hidden_states, position_embeddings=position_embeddings) if kwargs.get("skip_norm_elementwise", False): - sequence_output= F.layer_norm(hidden_states, hidden_states.shape[-1:]) + sequence_output = F.layer_norm(hidden_states, hidden_states.shape[-1:]) else: norm = self.norm.to(hidden_states.device) sequence_output = norm(hidden_states) pooled_output = sequence_output[:, 0, :] - return sequence_output, None, pooled_output, None diff --git a/comfy/latent_formats.py b/comfy/latent_formats.py index 986218a58..66eb2e0d2 100644 --- a/comfy/latent_formats.py +++ b/comfy/latent_formats.py @@ -239,6 +239,16 @@ class Flux2(LatentFormat): def process_out(self, latent): return latent +class TripoSplat(LatentFormat): + # Sequence latent (B, 8192, 16) the camera token rides alongside as a second nested latent + latent_channels = 16 + + def process_in(self, latent): + return latent + + def process_out(self, latent): + return latent + class Mochi(LatentFormat): latent_channels = 12 latent_dimensions = 3 @@ -802,13 +812,15 @@ class ZImagePixelSpace(ChromaRadiance): """ pass - class HiDreamO1Pixel(ChromaRadiance): """Pixel-space latent format for HiDream-O1. No VAE — model patches/unpatches raw RGB internally with patch_size=32. """ pass +class PixelDiTPixel(ChromaRadiance): + pass + class CogVideoX(LatentFormat): """Latent format for CogVideoX-2b (THUDM/CogVideoX-2b). diff --git a/comfy/ldm/audio/dit.py b/comfy/ldm/audio/dit.py index a6258b755..c28be5b49 100644 --- a/comfy/ldm/audio/dit.py +++ b/comfy/ldm/audio/dit.py @@ -433,11 +433,11 @@ class Attention(nn.Module): if self.differential: q, q_diff = q.unbind(dim=1) k, k_diff = k.unbind(dim=1) - out = optimized_attention(q, k, v, h, skip_reshape=True, transformer_options=transformer_options) - out_diff = optimized_attention(q_diff, k_diff, v, h, skip_reshape=True, transformer_options=transformer_options) + out = optimized_attention(q, k, v, h, skip_reshape=True, low_precision_attention=False, transformer_options=transformer_options) + out_diff = optimized_attention(q_diff, k_diff, v, h, skip_reshape=True, low_precision_attention=False, transformer_options=transformer_options) out = out - out_diff else: - out = optimized_attention(q, k, v, h, skip_reshape=True, transformer_options=transformer_options) + out = optimized_attention(q, k, v, h, skip_reshape=True, low_precision_attention=False, transformer_options=transformer_options) out = self.to_out(out) diff --git a/comfy/ldm/audio/vae_sa3.py b/comfy/ldm/audio/vae_sa3.py index 276846444..8be36d6ee 100644 --- a/comfy/ldm/audio/vae_sa3.py +++ b/comfy/ldm/audio/vae_sa3.py @@ -138,11 +138,11 @@ class Attention(nn.Module): k_diff = _apply_rotary_pos_emb(k_diff.float(), freqs).to(k_dtype) if self.differential: - out = (optimized_attention(q, k, v, h, mask=mask, skip_reshape=True) - - optimized_attention(q_diff, k_diff, v, h, mask=mask, skip_reshape=True)) + out = (optimized_attention(q, k, v, h, mask=mask, skip_reshape=True, low_precision_attention=False) + - optimized_attention(q_diff, k_diff, v, h, mask=mask, skip_reshape=True, low_precision_attention=False)) del q, k, v, q_diff, k_diff else: - out = optimized_attention(q, k, v, h, mask=mask, skip_reshape=True) + out = optimized_attention(q, k, v, h, mask=mask, skip_reshape=True, low_precision_attention=False) del q, k, v return self.to_out(out) diff --git a/comfy/ldm/chroma_radiance/model.py b/comfy/ldm/chroma_radiance/model.py index 4fb56165e..86af98d36 100644 --- a/comfy/ldm/chroma_radiance/model.py +++ b/comfy/ldm/chroma_radiance/model.py @@ -38,6 +38,8 @@ class ChromaRadianceParams(ChromaParams): # None means use the same dtype as the model. nerf_embedder_dtype: Optional[torch.dtype] use_x0: bool + # Use sequential txt_ids instead of zeros + use_sequential_txt_ids: bool class ChromaRadiance(Chroma): """ @@ -162,6 +164,9 @@ class ChromaRadiance(Chroma): if params.use_x0: self.register_buffer("__x0__", torch.tensor([])) + if params.use_sequential_txt_ids: + self.register_buffer("__sequential__", torch.tensor([])) + @property def _nerf_final_layer(self) -> nn.Module: if self.params.nerf_final_head_type == "linear": @@ -313,6 +318,9 @@ class ChromaRadiance(Chroma): img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0) img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs) txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype) + # Radiance after 2026-05-22 uses sequential txt_ids instead of zeros + if params.use_sequential_txt_ids: + txt_ids[:, :, 0] = torch.arange(context.shape[1], device=x.device, dtype=x.dtype).unsqueeze(0).expand(bs, -1) img_out = self.forward_orig( img, diff --git a/comfy/ldm/colormap.py b/comfy/ldm/colormap.py new file mode 100644 index 000000000..1f4d88bd9 --- /dev/null +++ b/comfy/ldm/colormap.py @@ -0,0 +1,25 @@ +"""Colormap utilities for depth and geometry visualisation.""" + +from __future__ import annotations + +import torch + + +def turbo(x: torch.Tensor) -> torch.Tensor: + """Anton Mikhailov polynomial approximation of the Turbo colormap. + + Args: + x: Float tensor with values in [0, 1]. + + Returns: + RGB tensor of the same shape as ``x`` with a trailing size-3 dimension. + """ + x = x.clamp(0.0, 1.0) + x2 = x * x + x3 = x2 * x + x4 = x2 * x2 + x5 = x4 * x + r = 0.13572138 + 4.61539260*x - 42.66032258*x2 + 132.13108234*x3 - 152.94239396*x4 + 59.28637943*x5 + g = 0.09140261 + 2.19418839*x + 4.84296658*x2 - 14.18503333*x3 + 4.27729857*x4 + 2.82956604*x5 + b = 0.10667330 + 12.64194608*x - 60.58204836*x2 + 110.36276771*x3 - 89.90310912*x4 + 27.34824973*x5 + return torch.stack([r, g, b], dim=-1).clamp(0.0, 1.0) diff --git a/comfy/ldm/cosmos/predict2.py b/comfy/ldm/cosmos/predict2.py index 2268bff38..671fe834d 100644 --- a/comfy/ldm/cosmos/predict2.py +++ b/comfy/ldm/cosmos/predict2.py @@ -14,15 +14,7 @@ from torchvision import transforms import comfy.patcher_extension from comfy.ldm.modules.attention import optimized_attention import comfy.ldm.common_dit - -def apply_rotary_pos_emb( - t: torch.Tensor, - freqs: torch.Tensor, -) -> torch.Tensor: - t_ = t.reshape(*t.shape[:-1], 2, -1).movedim(-2, -1).unsqueeze(-2).float() - t_out = freqs[..., 0] * t_[..., 0] + freqs[..., 1] * t_[..., 1] - t_out = t_out.movedim(-1, -2).reshape(*t.shape).type_as(t) - return t_out +import comfy.quant_ops # ---------------------- Feed Forward Network ----------------------- @@ -173,8 +165,7 @@ class Attention(nn.Module): k = self.k_norm(k) v = self.v_norm(v) if self.is_selfattn and rope_emb is not None: # only apply to self-attention! - q = apply_rotary_pos_emb(q, rope_emb) - k = apply_rotary_pos_emb(k, rope_emb) + q, k = comfy.quant_ops.ck.apply_rope_split_half(q, k, rope_emb) return q, k, v q, k, v = apply_norm_and_rotary_pos_emb(q, k, v, rope_emb) diff --git a/comfy/ldm/depth_anything_3/camera.py b/comfy/ldm/depth_anything_3/camera.py new file mode 100644 index 000000000..65a57d66f --- /dev/null +++ b/comfy/ldm/depth_anything_3/camera.py @@ -0,0 +1,177 @@ +"""Camera-token encoder and decoder for Depth Anything 3.""" + +from __future__ import annotations + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from comfy.ldm.modules.attention import optimized_attention_for_device +from .transform import affine_inverse, extri_intri_to_pose_encoding + + +# ----------------------------------------------------------------------- +# Building blocks (mirror depth_anything_3.model.utils.{attention,block}) +# ----------------------------------------------------------------------- + + +class _Mlp(nn.Module): + """Standard 2-layer MLP with GELU. Matches upstream ``utils.attention.Mlp``.""" + + def __init__(self, in_features, hidden_features=None, out_features=None, *, device=None, dtype=None, operations=None): + super().__init__() + out_features = out_features or in_features + hidden_features = hidden_features or in_features + self.fc1 = operations.Linear(in_features, hidden_features, bias=True, device=device, dtype=dtype) + self.fc2 = operations.Linear(hidden_features, out_features, bias=True, device=device, dtype=dtype) + + def forward(self, x): + return self.fc2(F.gelu(self.fc1(x))) + + +class _LayerScale(nn.Module): + """Per-channel learnable scaling. Matches upstream LayerScale.""" + + def __init__(self, dim, *, device=None, dtype=None): + super().__init__() + self.gamma = nn.Parameter(torch.empty(dim, device=device, dtype=dtype)) + + def forward(self, x): + return x * self.gamma.to(dtype=x.dtype, device=x.device) + + +class _Attention(nn.Module): + """ Self-attention with fused QKV projection. Mirrors upstream utils.attention.Attention; + Layout matches the HF safetensors (attn.qkv.{weight,bias} and attn.proj.{weight,bias}).""" + + def __init__(self, dim, num_heads, *, device=None, dtype=None, operations=None): + super().__init__() + assert dim % num_heads == 0 + self.num_heads = num_heads + self.head_dim = dim // num_heads + self.qkv = operations.Linear(dim, dim * 3, bias=True, device=device, dtype=dtype) + self.proj = operations.Linear(dim, dim, bias=True, device=device, dtype=dtype) + + def forward(self, x): + B, N, C = x.shape + qkv = self.qkv(x).reshape(B, N, 3, C) + q, k, v = qkv.unbind(2) # each (B, N, C) + attn_fn = optimized_attention_for_device(x.device, small_input=True) + out = attn_fn(q, k, v, heads=self.num_heads) + return self.proj(out) + + +class _Block(nn.Module): + """Pre-norm transformer block with LayerScale. Used by :class:CameraEnc. Layout follows upstream utils.block.Block.""" + + def __init__(self, dim, num_heads, mlp_ratio=4, init_values=0.01, *, device=None, dtype=None, operations=None): + super().__init__() + self.norm1 = operations.LayerNorm(dim, device=device, dtype=dtype) + self.attn = _Attention(dim, num_heads, device=device, dtype=dtype, operations=operations) + self.ls1 = _LayerScale(dim, device=device, dtype=dtype) if init_values else nn.Identity() + self.norm2 = operations.LayerNorm(dim, device=device, dtype=dtype) + self.mlp = _Mlp(in_features=dim, hidden_features=int(dim * mlp_ratio), device=device, dtype=dtype, operations=operations) + self.ls2 = _LayerScale(dim, device=device, dtype=dtype) if init_values else nn.Identity() + + def forward(self, x): + x = x + self.ls1(self.attn(self.norm1(x))) + x = x + self.ls2(self.mlp(self.norm2(x))) + return x + + +class CameraEnc(nn.Module): + """Encode per-view (extrinsics, intrinsics) into a camera token. + + Maps a 9-D pose-encoding vector through a small MLP up to the backbone's + ``embed_dim``, then runs ``trunk_depth`` transformer blocks. The output + has shape ``(B, S, embed_dim)`` and is injected at block ``alt_start`` + of the DINOv2 backbone in place of the cls token. + + Parameters mirror the upstream ``cam_enc.py`` so HF weights load directly. + """ + + def __init__( + self, + dim_out: int = 1024, + dim_in: int = 9, + trunk_depth: int = 4, + target_dim: int = 9, + num_heads: int = 16, + mlp_ratio: int = 4, + init_values: float = 0.01, + *, + device=None, dtype=None, operations=None, + **_kwargs, + ): + super().__init__() + self.target_dim = target_dim + self.trunk_depth = trunk_depth + self.trunk = nn.Sequential(*[ + _Block(dim_out, num_heads=num_heads, mlp_ratio=mlp_ratio, + init_values=init_values, + device=device, dtype=dtype, operations=operations) + for _ in range(trunk_depth) + ]) + self.token_norm = operations.LayerNorm(dim_out, device=device, dtype=dtype) + self.trunk_norm = operations.LayerNorm(dim_out, device=device, dtype=dtype) + self.pose_branch = _Mlp( + in_features=dim_in, + hidden_features=dim_out // 2, + out_features=dim_out, + device=device, dtype=dtype, operations=operations, + ) + + def forward(self, extrinsics: torch.Tensor, intrinsics: torch.Tensor, + image_size_hw) -> torch.Tensor: + """Encode camera parameters into ``(B, S, dim_out)`` tokens.""" + c2ws = affine_inverse(extrinsics) + pose_encoding = extri_intri_to_pose_encoding(c2ws, intrinsics, image_size_hw) + tokens = self.pose_branch(pose_encoding.to(self.pose_branch.fc1.weight.dtype)) + tokens = self.token_norm(tokens) + tokens = self.trunk(tokens) + tokens = self.trunk_norm(tokens) + return tokens + + +class CameraDec(nn.Module): + """Decode the final cam token into a 9-D pose encoding. + + Output layout: ``[T(3), quat_xyzw(4), fov_h, fov_w]``. The translation is + always predicted by the network; the quaternion and FoV can either be + predicted or supplied via ``camera_encoding`` (used at training time + when GT cameras are available -- not exercised at inference here). + + Parameters mirror the upstream ``cam_dec.py`` so HF weights load directly. + """ + + def __init__(self, dim_in: int = 1536, + *, device=None, dtype=None, operations=None, **_kwargs): + super().__init__() + d = dim_in + self.backbone = nn.Sequential( + operations.Linear(d, d, device=device, dtype=dtype), + nn.ReLU(), + operations.Linear(d, d, device=device, dtype=dtype), + nn.ReLU(), + ) + self.fc_t = operations.Linear(d, 3, device=device, dtype=dtype) + self.fc_qvec = operations.Linear(d, 4, device=device, dtype=dtype) + self.fc_fov = nn.Sequential( + operations.Linear(d, 2, device=device, dtype=dtype), + nn.ReLU(), + ) + + def forward(self, feat: torch.Tensor, + camera_encoding: "torch.Tensor | None" = None) -> torch.Tensor: + """Decode ``(B, N, dim_in)`` cam tokens into ``(B, N, 9)`` pose enc.""" + B, N = feat.shape[:2] + feat = feat.reshape(B * N, -1) + feat = self.backbone(feat) + out_t = self.fc_t(feat.float()).reshape(B, N, 3) + if camera_encoding is None: + out_qvec = self.fc_qvec(feat.float()).reshape(B, N, 4) + out_fov = self.fc_fov(feat.float()).reshape(B, N, 2) + else: + out_qvec = camera_encoding[..., 3:7] + out_fov = camera_encoding[..., -2:] + return torch.cat([out_t, out_qvec, out_fov], dim=-1) diff --git a/comfy/ldm/depth_anything_3/dpt.py b/comfy/ldm/depth_anything_3/dpt.py new file mode 100644 index 000000000..fb940873b --- /dev/null +++ b/comfy/ldm/depth_anything_3/dpt.py @@ -0,0 +1,489 @@ +"""DPT / DualDPT heads for Depth Anything 3.""" + +from __future__ import annotations + +from typing import List, Optional, Sequence, Tuple + +import torch +import torch.nn as nn +import torch.nn.functional as F + + +class Permute(nn.Module): + def __init__(self, dims: Tuple[int, ...]): + super().__init__() + self.dims = dims + + def forward(self, x: torch.Tensor) -> torch.Tensor: + return x.permute(*self.dims) + + +def _custom_interpolate( + x: torch.Tensor, + size: Optional[Tuple[int, int]] = None, + scale_factor: Optional[float] = None, + mode: str = "bilinear", + align_corners: bool = True, +) -> torch.Tensor: + if size is None: + assert scale_factor is not None + size = (int(x.shape[-2] * scale_factor), int(x.shape[-1] * scale_factor)) + INT_MAX = 1610612736 + total = size[0] * size[1] * x.shape[0] * x.shape[1] + if total > INT_MAX: + chunks = torch.chunk(x, chunks=(total // INT_MAX) + 1, dim=0) + outs = [F.interpolate(c, size=size, mode=mode, align_corners=align_corners) for c in chunks] + return torch.cat(outs, dim=0).contiguous() + return F.interpolate(x, size=size, mode=mode, align_corners=align_corners) + + +def _create_uv_grid(width: int, height: int, aspect_ratio: float, dtype, device) -> torch.Tensor: + """Normalised UV grid spanning (-x_span, -y_span)..(x_span, y_span).""" + diag_factor = (aspect_ratio ** 2 + 1.0) ** 0.5 + span_x = aspect_ratio / diag_factor + span_y = 1.0 / diag_factor + left_x = -span_x * (width - 1) / width + right_x = span_x * (width - 1) / width + top_y = -span_y * (height - 1) / height + bottom_y = span_y * (height - 1) / height + x_coords = torch.linspace(left_x, right_x, steps=width, dtype=dtype, device=device) + y_coords = torch.linspace(top_y, bottom_y, steps=height, dtype=dtype, device=device) + uu, vv = torch.meshgrid(x_coords, y_coords, indexing="xy") + return torch.stack((uu, vv), dim=-1) # (H, W, 2) + + +def _make_sincos_pos_embed(embed_dim: int, pos: torch.Tensor, omega_0: float = 100.0) -> torch.Tensor: + omega = torch.arange(embed_dim // 2, dtype=torch.float32, device=pos.device) + omega = 1.0 / omega_0 ** (omega / (embed_dim / 2.0)) + pos = pos.reshape(-1) + out = torch.einsum("m,d->md", pos, omega) + return torch.cat([out.sin(), out.cos()], dim=1).float() + + +def _position_grid_to_embed(pos_grid: torch.Tensor, embed_dim: int, omega_0: float = 100.0) -> torch.Tensor: + H, W, _ = pos_grid.shape + pos_flat = pos_grid.reshape(-1, 2) + emb_x = _make_sincos_pos_embed(embed_dim // 2, pos_flat[:, 0], omega_0=omega_0) + emb_y = _make_sincos_pos_embed(embed_dim // 2, pos_flat[:, 1], omega_0=omega_0) + emb = torch.cat([emb_x, emb_y], dim=-1) + return emb.view(H, W, embed_dim) + + +def _add_pos_embed(x: torch.Tensor, W: int, H: int, ratio: float = 0.1) -> torch.Tensor: + """Stateless UV positional embedding added to a feature map (B, C, h, w).""" + pw, ph = x.shape[-1], x.shape[-2] + pe = _create_uv_grid(pw, ph, aspect_ratio=W / H, dtype=x.dtype, device=x.device) + pe = _position_grid_to_embed(pe, x.shape[1]) * ratio + pe = pe.permute(2, 0, 1)[None].expand(x.shape[0], -1, -1, -1).to(dtype=x.dtype) + return x + pe + + +def _apply_activation(x: torch.Tensor, activation: str) -> torch.Tensor: + act = (activation or "linear").lower() + if act == "exp": + return torch.exp(x) + if act == "expp1": + return torch.exp(x) + 1 + if act == "expm1": + return torch.expm1(x) + if act == "relu": + return torch.relu(x) + if act == "sigmoid": + return torch.sigmoid(x) + if act == "softplus": + return F.softplus(x) + if act == "tanh": + return torch.tanh(x) + return x + + +# ----------------------------------------------------------------------------- +# Fusion building blocks +# ----------------------------------------------------------------------------- + + +class ResidualConvUnit(nn.Module): + def __init__(self, features: int, device=None, dtype=None, operations=None): + super().__init__() + self.conv1 = operations.Conv2d(features, features, 3, 1, 1, bias=True, device=device, dtype=dtype) + self.conv2 = operations.Conv2d(features, features, 3, 1, 1, bias=True, device=device, dtype=dtype) + self.activation = nn.ReLU(inplace=False) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + out = self.activation(x) + out = self.conv1(out) + out = self.activation(out) + out = self.conv2(out) + return out + x + + +class FeatureFusionBlock(nn.Module): + def __init__(self, features: int, has_residual: bool = True, align_corners: bool = True, device=None, dtype=None, operations=None): + super().__init__() + self.align_corners = align_corners + self.has_residual = has_residual + if has_residual: + self.resConfUnit1 = ResidualConvUnit(features, device=device, dtype=dtype, operations=operations) + else: + self.resConfUnit1 = None + self.resConfUnit2 = ResidualConvUnit(features, device=device, dtype=dtype, operations=operations) + self.out_conv = operations.Conv2d(features, features, 1, 1, 0, bias=True, device=device, dtype=dtype) + + def forward(self, *xs: torch.Tensor, size: Optional[Tuple[int, int]] = None) -> torch.Tensor: + y = xs[0] + if self.has_residual and len(xs) > 1 and self.resConfUnit1 is not None: + y = y + self.resConfUnit1(xs[1]) + y = self.resConfUnit2(y) + if size is None: + up_kwargs = {"scale_factor": 2.0} + else: + up_kwargs = {"size": size} + y = _custom_interpolate(y, **up_kwargs, mode="bilinear", align_corners=self.align_corners) + y = self.out_conv(y) + return y + + +class _Scratch(nn.Module): + """Container that mirrors upstream ``scratch`` attribute layout.""" + + +def _make_scratch(in_shape: List[int], out_shape: int, device=None, dtype=None, operations=None) -> _Scratch: + scratch = _Scratch() + scratch.layer1_rn = operations.Conv2d(in_shape[0], out_shape, 3, 1, 1, bias=False, device=device, dtype=dtype) + scratch.layer2_rn = operations.Conv2d(in_shape[1], out_shape, 3, 1, 1, bias=False, device=device, dtype=dtype) + scratch.layer3_rn = operations.Conv2d(in_shape[2], out_shape, 3, 1, 1, bias=False, device=device, dtype=dtype) + scratch.layer4_rn = operations.Conv2d(in_shape[3], out_shape, 3, 1, 1, bias=False, device=device, dtype=dtype) + return scratch + + +def _make_fusion_block(features: int, has_residual: bool = True, device=None, dtype=None, operations=None) -> FeatureFusionBlock: + return FeatureFusionBlock(features, has_residual=has_residual, align_corners=True, device=device, dtype=dtype, operations=operations) + + +# ----------------------------------------------------------------------------- +# DPT (single head + optional sky head) -- used by DA3Mono/Metric +# ----------------------------------------------------------------------------- + + +class DPT(nn.Module): + """Single-head DPT used by DA3Mono-Large and DA3Metric-Large.""" + + def __init__( + self, + dim_in: int, + patch_size: int = 14, + output_dim: int = 1, + activation: str = "exp", + conf_activation: str = "expp1", + features: int = 256, + out_channels: Sequence[int] = (256, 512, 1024, 1024), + pos_embed: bool = False, + down_ratio: int = 1, + head_name: str = "depth", + use_sky_head: bool = True, + sky_name: str = "sky", + sky_activation: str = "relu", + norm_type: str = "idt", + device=None, dtype=None, operations=None, + ): + super().__init__() + self.patch_size = patch_size + self.activation = activation + self.conf_activation = conf_activation + self.pos_embed = pos_embed + self.down_ratio = down_ratio + self.head_main = head_name + self.sky_name = sky_name + self.out_dim = output_dim + self.has_conf = output_dim > 1 + self.use_sky_head = use_sky_head + self.sky_activation = sky_activation + self.intermediate_layer_idx: Tuple[int, int, int, int] = (0, 1, 2, 3) + + if norm_type == "layer": + self.norm = operations.LayerNorm(dim_in, device=device, dtype=dtype) + else: + self.norm = nn.Identity() + + out_channels = list(out_channels) + self.projects = nn.ModuleList([ + operations.Conv2d(dim_in, oc, kernel_size=1, stride=1, padding=0, device=device, dtype=dtype) + for oc in out_channels + ]) + self.resize_layers = nn.ModuleList([ + operations.ConvTranspose2d(out_channels[0], out_channels[0], kernel_size=4, stride=4, padding=0, device=device, dtype=dtype), + operations.ConvTranspose2d(out_channels[1], out_channels[1], kernel_size=2, stride=2, padding=0, device=device, dtype=dtype), + nn.Identity(), + operations.Conv2d(out_channels[3], out_channels[3], kernel_size=3, stride=2, padding=1, device=device, dtype=dtype), + ]) + + self.scratch = _make_scratch(out_channels, features, device=device, dtype=dtype, operations=operations) + self.scratch.refinenet1 = _make_fusion_block(features, device=device, dtype=dtype, operations=operations) + self.scratch.refinenet2 = _make_fusion_block(features, device=device, dtype=dtype, operations=operations) + self.scratch.refinenet3 = _make_fusion_block(features, device=device, dtype=dtype, operations=operations) + self.scratch.refinenet4 = _make_fusion_block(features, has_residual=False, device=device, dtype=dtype, operations=operations) + + head_features_1 = features + head_features_2 = 32 + self.scratch.output_conv1 = operations.Conv2d( + head_features_1, head_features_1 // 2, kernel_size=3, stride=1, padding=1, + device=device, dtype=dtype, + ) + self.scratch.output_conv2 = nn.Sequential( + operations.Conv2d(head_features_1 // 2, head_features_2, kernel_size=3, stride=1, padding=1, device=device, dtype=dtype), + nn.ReLU(inplace=False), + operations.Conv2d(head_features_2, output_dim, kernel_size=1, stride=1, padding=0, device=device, dtype=dtype), + ) + + if self.use_sky_head: + self.scratch.sky_output_conv2 = nn.Sequential( + operations.Conv2d(head_features_1 // 2, head_features_2, kernel_size=3, stride=1, padding=1, device=device, dtype=dtype), + nn.ReLU(inplace=False), + operations.Conv2d(head_features_2, 1, kernel_size=1, stride=1, padding=0, device=device, dtype=dtype), + ) + + def forward(self, feats: List[torch.Tensor], H: int, W: int, patch_start_idx: int = 0, **_kwargs) -> dict: + # feats[i][0] is the patch-token tensor with shape (B, S, N_patch, C) + B, S, N, C = feats[0][0].shape + feats_flat = [feat[0].reshape(B * S, N, C) for feat in feats] + + ph, pw = H // self.patch_size, W // self.patch_size + resized = [] + for stage_idx, take_idx in enumerate(self.intermediate_layer_idx): + x = feats_flat[take_idx][:, patch_start_idx:] + x = self.norm(x) + x = x.permute(0, 2, 1).contiguous().reshape(B * S, C, ph, pw) + x = self.projects[stage_idx](x) + if self.pos_embed: + x = _add_pos_embed(x, W, H) + x = self.resize_layers[stage_idx](x) + resized.append(x) + + l1_rn = self.scratch.layer1_rn(resized[0]) + l2_rn = self.scratch.layer2_rn(resized[1]) + l3_rn = self.scratch.layer3_rn(resized[2]) + l4_rn = self.scratch.layer4_rn(resized[3]) + + out = self.scratch.refinenet4(l4_rn, size=l3_rn.shape[2:]) + out = self.scratch.refinenet3(out, l3_rn, size=l2_rn.shape[2:]) + out = self.scratch.refinenet2(out, l2_rn, size=l1_rn.shape[2:]) + out = self.scratch.refinenet1(out, l1_rn) + + h_out = int(ph * self.patch_size / self.down_ratio) + w_out = int(pw * self.patch_size / self.down_ratio) + + fused = self.scratch.output_conv1(out) + fused = _custom_interpolate(fused, (h_out, w_out), mode="bilinear", align_corners=True) + if self.pos_embed: + fused = _add_pos_embed(fused, W, H) + feat = fused + + main_logits = self.scratch.output_conv2(feat) + outs = {} + if self.has_conf: + fmap = main_logits.permute(0, 2, 3, 1) + pred = _apply_activation(fmap[..., :-1], self.activation) + conf = _apply_activation(fmap[..., -1], self.conf_activation) + outs[self.head_main] = pred.squeeze(-1).view(B, S, *pred.shape[1:-1]) + outs[f"{self.head_main}_conf"] = conf.view(B, S, *conf.shape[1:]) + else: + pred = _apply_activation(main_logits, self.activation) + outs[self.head_main] = pred.squeeze(1).view(B, S, *pred.shape[2:]) + + if self.use_sky_head: + sky_logits = self.scratch.sky_output_conv2(feat) + if self.sky_activation.lower() == "sigmoid": + sky = torch.sigmoid(sky_logits) + elif self.sky_activation.lower() == "relu": + sky = F.relu(sky_logits) + else: + sky = sky_logits + outs[self.sky_name] = sky.squeeze(1).view(B, S, *sky.shape[2:]) + + return outs + + +# ----------------------------------------------------------------------------- +# DualDPT (depth + auxiliary "ray" head) -- used by DA3-Small / DA3-Base +# ----------------------------------------------------------------------------- + + +class DualDPT(nn.Module): + """Two-head DPT used by DA3-Small / DA3-Base.""" + + def __init__( + self, + dim_in: int, + patch_size: int = 14, + output_dim: int = 2, + activation: str = "exp", + conf_activation: str = "expp1", + features: int = 256, + out_channels: Sequence[int] = (256, 512, 1024, 1024), + pos_embed: bool = True, + down_ratio: int = 1, + aux_pyramid_levels: int = 4, + aux_out1_conv_num: int = 5, + head_names: Tuple[str, str] = ("depth", "ray"), + device=None, dtype=None, operations=None, + ): + super().__init__() + self.patch_size = patch_size + self.activation = activation + self.conf_activation = conf_activation + self.pos_embed = pos_embed + self.down_ratio = down_ratio + self.aux_levels = aux_pyramid_levels + self.aux_out1_conv_num = aux_out1_conv_num + self.head_main, self.head_aux = head_names + self.intermediate_layer_idx: Tuple[int, int, int, int] = (0, 1, 2, 3) + # Toggle the auxiliary ray branch at runtime. Default off (mono path). + # DepthAnything3Net flips this on when running multi-view + ray-pose. + self.enable_aux: bool = False + + self.norm = operations.LayerNorm(dim_in, device=device, dtype=dtype) + out_channels = list(out_channels) + self.projects = nn.ModuleList([ + operations.Conv2d(dim_in, oc, kernel_size=1, stride=1, padding=0, device=device, dtype=dtype) + for oc in out_channels + ]) + self.resize_layers = nn.ModuleList([ + operations.ConvTranspose2d(out_channels[0], out_channels[0], kernel_size=4, stride=4, padding=0, device=device, dtype=dtype), + operations.ConvTranspose2d(out_channels[1], out_channels[1], kernel_size=2, stride=2, padding=0, device=device, dtype=dtype), + nn.Identity(), + operations.Conv2d(out_channels[3], out_channels[3], kernel_size=3, stride=2, padding=1, device=device, dtype=dtype), + ]) + + self.scratch = _make_scratch(out_channels, features, device=device, dtype=dtype, operations=operations) + # Main fusion chain + self.scratch.refinenet1 = _make_fusion_block(features, device=device, dtype=dtype, operations=operations) + self.scratch.refinenet2 = _make_fusion_block(features, device=device, dtype=dtype, operations=operations) + self.scratch.refinenet3 = _make_fusion_block(features, device=device, dtype=dtype, operations=operations) + self.scratch.refinenet4 = _make_fusion_block(features, has_residual=False, device=device, dtype=dtype, operations=operations) + # Auxiliary fusion chain (separate copies) + self.scratch.refinenet1_aux = _make_fusion_block(features, device=device, dtype=dtype, operations=operations) + self.scratch.refinenet2_aux = _make_fusion_block(features, device=device, dtype=dtype, operations=operations) + self.scratch.refinenet3_aux = _make_fusion_block(features, device=device, dtype=dtype, operations=operations) + self.scratch.refinenet4_aux = _make_fusion_block(features, has_residual=False, device=device, dtype=dtype, operations=operations) + + head_features_1 = features + head_features_2 = 32 + + # Main head neck + final projection + self.scratch.output_conv1 = operations.Conv2d( + head_features_1, head_features_1 // 2, kernel_size=3, stride=1, padding=1, + device=device, dtype=dtype, + ) + self.scratch.output_conv2 = nn.Sequential( + operations.Conv2d(head_features_1 // 2, head_features_2, kernel_size=3, stride=1, padding=1, device=device, dtype=dtype), + nn.ReLU(inplace=False), + operations.Conv2d(head_features_2, output_dim, kernel_size=1, stride=1, padding=0, device=device, dtype=dtype), + ) + + # Aux pre-head per level (multi-level pyramid) + self.scratch.output_conv1_aux = nn.ModuleList([ + self._make_aux_out1_block(head_features_1, device=device, dtype=dtype, operations=operations) + for _ in range(self.aux_levels) + ]) + + # Aux final projection per level (includes LayerNorm permute path). + ln_seq = [Permute((0, 2, 3, 1)), + operations.LayerNorm(head_features_2, device=device, dtype=dtype), + Permute((0, 3, 1, 2))] + self.scratch.output_conv2_aux = nn.ModuleList([ + nn.Sequential( + operations.Conv2d(head_features_1 // 2, head_features_2, kernel_size=3, stride=1, padding=1, device=device, dtype=dtype), + *ln_seq, + nn.ReLU(inplace=False), + operations.Conv2d(head_features_2, 7, kernel_size=1, stride=1, padding=0, device=device, dtype=dtype), + ) + for _ in range(self.aux_levels) + ]) + + @staticmethod + def _make_aux_out1_block(in_ch: int, *, device=None, dtype=None, operations=None) -> nn.Sequential: + # aux_out1_conv_num=5 in all Apache-2.0 variants. + return nn.Sequential( + operations.Conv2d(in_ch, in_ch // 2, 3, 1, 1, device=device, dtype=dtype), + operations.Conv2d(in_ch // 2, in_ch, 3, 1, 1, device=device, dtype=dtype), + operations.Conv2d(in_ch, in_ch // 2, 3, 1, 1, device=device, dtype=dtype), + operations.Conv2d(in_ch // 2, in_ch, 3, 1, 1, device=device, dtype=dtype), + operations.Conv2d(in_ch, in_ch // 2, 3, 1, 1, device=device, dtype=dtype), + ) + + def forward(self, feats: List[torch.Tensor], H: int, W: int, patch_start_idx: int = 0, **_kwargs) -> dict: + B, S, N, C = feats[0][0].shape + feats_flat = [feat[0].reshape(B * S, N, C) for feat in feats] + + ph, pw = H // self.patch_size, W // self.patch_size + resized = [] + for stage_idx, take_idx in enumerate(self.intermediate_layer_idx): + x = feats_flat[take_idx][:, patch_start_idx:] + x = self.norm(x) + x = x.permute(0, 2, 1).contiguous().reshape(B * S, C, ph, pw) + x = self.projects[stage_idx](x) + if self.pos_embed: + x = _add_pos_embed(x, W, H) + x = self.resize_layers[stage_idx](x) + resized.append(x) + + l1_rn = self.scratch.layer1_rn(resized[0]) + l2_rn = self.scratch.layer2_rn(resized[1]) + l3_rn = self.scratch.layer3_rn(resized[2]) + l4_rn = self.scratch.layer4_rn(resized[3]) + + # Main pyramid (output_conv1 is applied inside the upstream `_fuse`, + # before interpolation -- replicate that order here). + m = self.scratch.refinenet4(l4_rn, size=l3_rn.shape[2:]) + if self.enable_aux: + a4 = self.scratch.refinenet4_aux(l4_rn, size=l3_rn.shape[2:]) + aux_pyr = [a4] + m = self.scratch.refinenet3(m, l3_rn, size=l2_rn.shape[2:]) + if self.enable_aux: + aux_pyr.append(self.scratch.refinenet3_aux(aux_pyr[-1], l3_rn, size=l2_rn.shape[2:])) + m = self.scratch.refinenet2(m, l2_rn, size=l1_rn.shape[2:]) + if self.enable_aux: + aux_pyr.append(self.scratch.refinenet2_aux(aux_pyr[-1], l2_rn, size=l1_rn.shape[2:])) + m = self.scratch.refinenet1(m, l1_rn) + if self.enable_aux: + aux_pyr.append(self.scratch.refinenet1_aux(aux_pyr[-1], l1_rn)) + m = self.scratch.output_conv1(m) + + h_out = int(ph * self.patch_size / self.down_ratio) + w_out = int(pw * self.patch_size / self.down_ratio) + + m = _custom_interpolate(m, (h_out, w_out), mode="bilinear", align_corners=True) + if self.pos_embed: + m = _add_pos_embed(m, W, H) + main_logits = self.scratch.output_conv2(m) + fmap = main_logits.permute(0, 2, 3, 1) + depth_pred = _apply_activation(fmap[..., :-1], self.activation) + depth_conf = _apply_activation(fmap[..., -1], self.conf_activation) + + outs = { + self.head_main: depth_pred.squeeze(-1).view(B, S, *depth_pred.shape[1:-1]), + f"{self.head_main}_conf": depth_conf.view(B, S, *depth_conf.shape[1:]), + } + + if self.enable_aux: + # Auxiliary "ray" head (multi-level inside) -- only the last level + # is returned. Mirrors upstream ``DualDPT._fuse`` + ``_forward_impl``: + # each aux pyramid level goes through ``output_conv1_aux[i]`` + # (5-layer conv stack that ends at ``features // 2`` channels), + # then the last level optionally gets a pos-embed and finally + # ``output_conv2_aux[-1]``. + aux_processed = [ + self.scratch.output_conv1_aux[i](a) for i, a in enumerate(aux_pyr) + ] + last_aux = aux_processed[-1] + if self.pos_embed: + last_aux = _add_pos_embed(last_aux, W, H) + last_aux_logits = self.scratch.output_conv2_aux[-1](last_aux) + fmap_last = last_aux_logits.permute(0, 2, 3, 1) + # Channels: [ray(6), ray_conf(1)]; ray uses 'linear' activation. + aux_pred = fmap_last[..., :-1] + aux_conf = _apply_activation(fmap_last[..., -1], self.conf_activation) + outs[self.head_aux] = aux_pred.view(B, S, *aux_pred.shape[1:]) + outs[f"{self.head_aux}_conf"] = aux_conf.view(B, S, *aux_conf.shape[1:]) + + return outs diff --git a/comfy/ldm/depth_anything_3/model.py b/comfy/ldm/depth_anything_3/model.py new file mode 100644 index 000000000..f3c8a5ee3 --- /dev/null +++ b/comfy/ldm/depth_anything_3/model.py @@ -0,0 +1,236 @@ +from __future__ import annotations + +from typing import Dict, Optional, Sequence + +import torch +import torch.nn as nn + +from comfy.image_encoders.dino2 import Dinov2Model + +from .camera import CameraDec, CameraEnc +from .dpt import DPT, DualDPT +from .ray_pose import get_extrinsic_from_camray +from .transform import affine_inverse, pose_encoding_to_extri_intri + + +_HEAD_REGISTRY = { + "dpt": DPT, + "dualdpt": DualDPT, +} + + +# Backbone presets (mirror the upstream DINOv2 ViT variants). +_BACKBONE_PRESETS = { + "vits": dict(hidden_size=384, num_hidden_layers=12, num_attention_heads=6, use_swiglu_ffn=False), + "vitb": dict(hidden_size=768, num_hidden_layers=12, num_attention_heads=12, use_swiglu_ffn=False), + "vitl": dict(hidden_size=1024, num_hidden_layers=24, num_attention_heads=16, use_swiglu_ffn=False), + "vitg": dict(hidden_size=1536, num_hidden_layers=40, num_attention_heads=24, use_swiglu_ffn=True), +} + + +def _build_backbone_config( + backbone_name: str, + *, + alt_start: int, + qknorm_start: int, + rope_start: int, + cat_token: bool, +) -> dict: + if backbone_name not in _BACKBONE_PRESETS: + raise ValueError(f"Unknown DINOv2 backbone variant: {backbone_name!r}") + cfg = dict(_BACKBONE_PRESETS[backbone_name]) + cfg.update(dict( + layer_norm_eps=1e-6, + patch_size=14, + image_size=518, + # No mask_token in DA3 weights; omit param to avoid load warnings. + use_mask_token=False, + alt_start=alt_start, + qknorm_start=qknorm_start, + rope_start=rope_start, + cat_token=cat_token, + rope_freq=100.0, + )) + return cfg + + +class DepthAnything3Net(nn.Module): + + PATCH_SIZE = 14 + + def __init__( + self, + # --- Backbone --- + backbone_name: str = "vitl", + out_layers: Sequence[int] = (4, 11, 17, 23), + alt_start: int = -1, + qknorm_start: int = -1, + rope_start: int = -1, + cat_token: bool = False, + # --- Head --- + head_type: str = "dpt", # dpt or dualdpt + head_dim_in: int = 1024, + head_output_dim: int = 1, # 1 = depth only, 2 = depth+conf + head_features: int = 256, + head_out_channels: Sequence[int] = (256, 512, 1024, 1024), + head_use_sky_head: bool = True, # ignored by DualDPT + head_pos_embed: Optional[bool] = None, # default: True for DualDPT, False for DPT + # --- Camera (multi-view) --- + has_cam_enc: bool = False, + has_cam_dec: bool = False, + cam_dim_out: Optional[int] = None, # CameraEnc dim_out (defaults to embed_dim) + cam_dec_dim_in: Optional[int] = None, # CameraDec dim_in (defaults to 2*embed_dim with cat_token) + # ComfyUI plumbing + device=None, dtype=None, operations=None, + **_ignored, + ): + super().__init__() + head_cls = _HEAD_REGISTRY[head_type.lower()] + self.head_type = head_type.lower() + self.has_sky = (self.head_type == "dpt") and head_use_sky_head + self.has_conf = head_output_dim > 1 + self.out_layers = list(out_layers) + + backbone_cfg = _build_backbone_config( + backbone_name, + alt_start=alt_start, + qknorm_start=qknorm_start, + rope_start=rope_start, + cat_token=cat_token, + ) + self.backbone = Dinov2Model(backbone_cfg, dtype, device, operations) + + head_kwargs = dict( + dim_in=head_dim_in, + patch_size=self.PATCH_SIZE, + output_dim=head_output_dim, + features=head_features, + out_channels=tuple(head_out_channels), + device=device, dtype=dtype, operations=operations, + ) + if self.head_type == "dpt": + head_kwargs.update( + use_sky_head=head_use_sky_head, + pos_embed=(False if head_pos_embed is None else head_pos_embed), + ) + else: # dualdpt + head_kwargs.update( + pos_embed=(True if head_pos_embed is None else head_pos_embed), + ) + self.head = head_cls(**head_kwargs) + + # Built only if checkpoint has weights; cam_enc output dim == embed_dim. + embed_dim = backbone_cfg["hidden_size"] + if has_cam_enc: + self.cam_enc = CameraEnc( + dim_out=cam_dim_out if cam_dim_out is not None else embed_dim, + num_heads=max(1, embed_dim // 64), + device=device, dtype=dtype, operations=operations, + ) + else: + self.cam_enc = None + if has_cam_dec: + default_dim = embed_dim * (2 if cat_token else 1) + self.cam_dec = CameraDec( + dim_in=cam_dec_dim_in if cam_dec_dim_in is not None else default_dim, + device=device, dtype=dtype, operations=operations, + ) + else: + self.cam_dec = None + + self.dtype = dtype + + def forward( + self, + image: torch.Tensor, + extrinsics: Optional[torch.Tensor] = None, + intrinsics: Optional[torch.Tensor] = None, + *, + use_ray_pose: bool = False, + ref_view_strategy: str = "saddle_balanced", + export_feat_layers: Optional[Sequence[int]] = None, + **_unused, + ) -> Dict[str, torch.Tensor]: + """Run depth and optionally pose prediction.""" + if image.ndim == 4: + image = image.unsqueeze(1) # (B, 1, 3, H, W) + assert image.ndim == 5 and image.shape[2] == 3, \ + f"image must be (B,3,H,W) or (B,S,3,H,W); got {tuple(image.shape)}" + + B, S, _, H, W = image.shape + assert H % self.PATCH_SIZE == 0 and W % self.PATCH_SIZE == 0, \ + f"image H,W must be multiples of {self.PATCH_SIZE}; got {(H, W)}" + + # Camera-token preparation (multi-view path). + cam_token = None + if extrinsics is not None and intrinsics is not None and self.cam_enc is not None: + cam_token = self.cam_enc(extrinsics, intrinsics, (H, W)) + + # Toggle aux ray output on/off depending on what the caller asked for. + if isinstance(self.head, DualDPT): + self.head.enable_aux = bool(use_ray_pose) + + feats, aux_feats = self.backbone.get_intermediate_layers_da3( + image, self.out_layers, cam_token=cam_token, + ref_view_strategy=ref_view_strategy, + export_feat_layers=export_feat_layers, + ) + head_out = self.head(feats, H=H, W=W, patch_start_idx=0) + + # Pose prediction. + out: Dict[str, torch.Tensor] = {} + if use_ray_pose and "ray" in head_out and "ray_conf" in head_out: + ray = head_out["ray"] + ray_conf = head_out["ray_conf"] + extr_c2w, focal, pp = get_extrinsic_from_camray( + ray, ray_conf, ray.shape[-3], ray.shape[-2], + ) + # Match the upstream output: w2c, drop the homogeneous row. + extr_w2c = affine_inverse(extr_c2w)[:, :, :3, :] + # Build pixel-space intrinsics from the normalised focal/pp output. + intr = torch.eye(3, device=ray.device, dtype=ray.dtype) + intr = intr[None, None].expand(extr_c2w.shape[0], extr_c2w.shape[1], 3, 3).clone() + intr[:, :, 0, 0] = focal[:, :, 0] / 2 * W + intr[:, :, 1, 1] = focal[:, :, 1] / 2 * H + intr[:, :, 0, 2] = pp[:, :, 0] * W * 0.5 + intr[:, :, 1, 2] = pp[:, :, 1] * H * 0.5 + out["extrinsics"] = extr_w2c + out["intrinsics"] = intr + elif self.cam_dec is not None and S > 1: + # Decode the cam-token of the final out_layer into a pose encoding. + cam_feat = feats[-1][1] # (B, S, dim_in_to_cam_dec) + pose_enc = self.cam_dec(cam_feat) + c2w_3x4, intr = pose_encoding_to_extri_intri(pose_enc, (H, W)) + # Match the upstream output convention: w2c (world->camera), 3x4. + c2w_4x4 = torch.cat([ + c2w_3x4, + torch.tensor([0, 0, 0, 1], device=c2w_3x4.device, dtype=c2w_3x4.dtype) + .view(1, 1, 1, 4).expand(B, S, 1, 4), + ], dim=-2) + out["extrinsics"] = affine_inverse(c2w_4x4)[:, :, :3, :] + out["intrinsics"] = intr + + # Flatten the views axis for per-pixel outputs (depth/conf/sky) so the + # per-image consumer keeps its (B*S, H, W) interface. + for k, v in head_out.items(): + if k in ("ray", "ray_conf"): + # Keep multi-view shape for downstream pose work. + out[k] = v + elif v.ndim >= 3 and v.shape[0] == B and v.shape[1] == S: + out[k] = v.reshape(B * S, *v.shape[2:]) + else: + out[k] = v + + if export_feat_layers: + out["aux_features"] = self._reshape_aux_features(aux_feats, H, W) + return out + + def _reshape_aux_features(self, aux_feats, H: int, W: int): + """Reshape (B, S, N, C) aux features into (B, S, h_p, w_p, C).""" + ph, pw = H // self.PATCH_SIZE, W // self.PATCH_SIZE + out = [] + for f in aux_feats: + B, S, N, C = f.shape + assert N == ph * pw, f"aux feature seq mismatch: {N} != {ph}*{pw}" + out.append(f.reshape(B, S, ph, pw, C)) + return out diff --git a/comfy/ldm/depth_anything_3/preprocess.py b/comfy/ldm/depth_anything_3/preprocess.py new file mode 100644 index 000000000..2238bd0d6 --- /dev/null +++ b/comfy/ldm/depth_anything_3/preprocess.py @@ -0,0 +1,128 @@ +"""Input/output preprocessing helpers for Depth Anything 3.""" + +from __future__ import annotations + +from typing import Tuple + +import torch + +import comfy.utils + +PATCH_SIZE = 14 + +# ImageNet normalization constants used during DA3 training. +_IMAGENET_MEAN = torch.tensor([0.485, 0.456, 0.406]) +_IMAGENET_STD = torch.tensor([0.229, 0.224, 0.225]) + + +def _round_to_patch(x: int, patch: int = PATCH_SIZE) -> int: + down = (x // patch) * patch + up = down + patch + return up if abs(up - x) <= abs(x - down) else down + + +def compute_target_size(orig_h: int, orig_w: int, process_res: int, method: str = "upper_bound_resize") -> Tuple[int, int]: + """Compute (target_h, target_w) for a single image. + upper_bound_resize: scale longest side to process_res, then round each dim to nearest multiple of 14 (default upstream method). + lower_bound_resize: scale shortest side to process_res, then round.""" + + if method == "upper_bound_resize": + longest = max(orig_h, orig_w) + scale = process_res / float(longest) + elif method == "lower_bound_resize": + shortest = min(orig_h, orig_w) + scale = process_res / float(shortest) + else: + raise ValueError(f"Unsupported process_res_method: {method}") + + new_w = max(1, _round_to_patch(int(round(orig_w * scale)))) + new_h = max(1, _round_to_patch(int(round(orig_h * scale)))) + return new_h, new_w + + +def preprocess_image(image: torch.Tensor, process_res: int = 504, method: str = "upper_bound_resize") -> torch.Tensor: + assert image.ndim == 4 and image.shape[-1] == 3, f"expected (B,H,W,3) IMAGE; got {tuple(image.shape)}" + B, H, W, _ = image.shape + target_h, target_w = compute_target_size(H, W, process_res, method) + + # (B, H, W, 3) -> (B, 3, H, W) + x = image.movedim(-1, 1).contiguous() + if (target_h, target_w) != (H, W): + # Upstream uses cv2 INTER_CUBIC (upscale) / INTER_AREA (downscale). + # Lanczos in ``common_upscale`` is anti-aliased and produces the + # closest pixel-wise match in a sweep across {bilinear, bicubic, + # area, lanczos, bislerp}. Used in both directions for simplicity. + x = comfy.utils.common_upscale(x.float(), target_w, target_h, "lanczos", "disabled",) + x = x.clamp(0.0, 1.0) + + mean = _IMAGENET_MEAN.to(device=x.device, dtype=x.dtype).view(1, 3, 1, 1) + std = _IMAGENET_STD.to(device=x.device, dtype=x.dtype).view(1, 3, 1, 1) + x = (x - mean) / std + return x + + +# ----------------------------------------------------------------------------- +# Output post-processing (sky-aware clipping for Mono/Metric variants) +# ----------------------------------------------------------------------------- + + +def compute_non_sky_mask(sky_prediction: torch.Tensor, threshold: float = 0.3) -> torch.Tensor: + """Boolean mask: True for non-sky pixels (sky probability < threshold).""" + return sky_prediction < threshold + + +def apply_sky_aware_clip(depth: torch.Tensor, sky: torch.Tensor, threshold: float = 0.3, quantile: float = 0.99) -> torch.Tensor: + """Clips sky regions to the 99th percentile of non-sky depth. Returns a new depth tensor.""" + non_sky = compute_non_sky_mask(sky, threshold=threshold) + if non_sky.sum() <= 10 or (~non_sky).sum() <= 10: + return depth.clone() + + non_sky_depth = depth[non_sky] + if non_sky_depth.numel() > 100_000: + idx = torch.randint(0, non_sky_depth.numel(), (100_000,), device=non_sky_depth.device) + sampled = non_sky_depth[idx] + else: + sampled = non_sky_depth + + max_depth = torch.quantile(sampled, quantile) + out = depth.clone() + out[~non_sky] = max_depth + return out + + +def normalize_depth_v2_style(depth: torch.Tensor, sky: torch.Tensor | None = None, low_quantile: float = 0.01, high_quantile: float = 0.99) -> torch.Tensor: + """V2-style normalization computes percentile bounds over non-sky pixels (when available), then maps depth into [0, 1] with near = white (1.0).""" + if sky is not None: + mask = compute_non_sky_mask(sky) + if mask.any(): + valid = depth[mask] + else: + valid = depth.flatten() + else: + valid = depth.flatten() + + if valid.numel() > 100_000: + idx = torch.randint(0, valid.numel(), (100_000,), device=valid.device) + sample = valid[idx] + else: + sample = valid + + lo = torch.quantile(sample, low_quantile) + hi = torch.quantile(sample, high_quantile) + rng = (hi - lo).clamp(min=1e-6) + norm = ((depth - lo) / rng).clamp(0.0, 1.0) + # Nearer pixels are brighter (1.0) + norm = 1.0 - norm + if sky is not None: + # Sky pixels become black (far / unknown) + sky_mask = ~compute_non_sky_mask(sky) + norm = torch.where(sky_mask, torch.zeros_like(norm), norm) + return norm + + +def normalize_depth_min_max(depth: torch.Tensor) -> torch.Tensor: + """Simple per-frame min/max normalization with near=1.0 convention.""" + lo = depth.amin(dim=(-2, -1), keepdim=True) + hi = depth.amax(dim=(-2, -1), keepdim=True) + rng = (hi - lo).clamp(min=1e-6) + return 1.0 - ((depth - lo) / rng).clamp(0.0, 1.0) diff --git a/comfy/ldm/depth_anything_3/ray_pose.py b/comfy/ldm/depth_anything_3/ray_pose.py new file mode 100644 index 000000000..90890f1da --- /dev/null +++ b/comfy/ldm/depth_anything_3/ray_pose.py @@ -0,0 +1,272 @@ +"""Ray-to-pose conversion for the multi-view path of Depth Anything 3.""" + +from __future__ import annotations + +from typing import Optional, Tuple + +import torch + + +# qr/svd use fp32: CUDA often has no fp16/bf16 kernels for these ops. + + +def _ql_decomposition(A: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: + """Decompose A = Q @ L with Q orthogonal and L lower-triangular. + Implemented in terms of QR by reversing the columns/rows; the standard + trick from the upstream reference. Inputs A are (3, 3).""" + P = torch.tensor([[0, 0, 1], [0, 1, 0], [1, 0, 0]], device=A.device, dtype=A.dtype) + A_tilde = A @ P + # CUDA QR is not implemented for fp16/bf16; upcast just for this call. + Q_tilde, R_tilde = torch.linalg.qr(A_tilde.float()) + Q_tilde = Q_tilde.to(A.dtype) + R_tilde = R_tilde.to(A.dtype) + Q = Q_tilde @ P + L = P @ R_tilde @ P + d = torch.diag(L) + sign = torch.sign(d) + Q = Q * sign[None, :] # scale columns of Q + L = L * sign[:, None] # scale rows of L + return Q, L + + +def _homogenize_points(points: torch.Tensor) -> torch.Tensor: + return torch.cat([points, torch.ones_like(points[..., :1])], dim=-1) + + +# ----------------------------------------------------------------------------- +# Weighted-LSQ + RANSAC homography (batched) +# ----------------------------------------------------------------------------- + + +def _find_homography_weighted_lsq(src_pts: torch.Tensor, dst_pts: torch.Tensor, confident_weight: torch.Tensor,) -> torch.Tensor: + """Solve a single H with weighted least-squares (DLT).""" + N = src_pts.shape[0] + if N < 4: + raise ValueError("At least 4 points are required to compute a homography.") + w = confident_weight.sqrt().unsqueeze(1) # (N, 1) + x = src_pts[:, 0:1] + y = src_pts[:, 1:2] + u = dst_pts[:, 0:1] + v = dst_pts[:, 1:2] + zeros = torch.zeros_like(x) + A1 = torch.cat([-x * w, -y * w, -w, zeros, zeros, zeros, x * u * w, y * u * w, u * w], dim=1) + A2 = torch.cat([zeros, zeros, zeros, -x * w, -y * w, -w, x * v * w, y * v * w, v * w], dim=1) + A = torch.cat([A1, A2], dim=0) # (2N, 9) + # CUDA SVD is not implemented for fp16/bf16; upcast just for this call. + _, _, Vh = torch.linalg.svd(A.float()) + Vh = Vh.to(A.dtype) + H = Vh[-1].reshape(3, 3) + return H / H[-1, -1] + + +def _find_homography_weighted_lsq_batched(src_pts_batch: torch.Tensor, dst_pts_batch: torch.Tensor, confident_weight_batch: torch.Tensor) -> torch.Tensor: + """Batched DLT solver. Inputs (B, K, 2) / (B, K); output (B, 3, 3).""" + B, K, _ = src_pts_batch.shape + w = confident_weight_batch.sqrt().unsqueeze(2) + x = src_pts_batch[:, :, 0:1] + y = src_pts_batch[:, :, 1:2] + u = dst_pts_batch[:, :, 0:1] + v = dst_pts_batch[:, :, 1:2] + zeros = torch.zeros_like(x) + A1 = torch.cat([-x * w, -y * w, -w, zeros, zeros, zeros, x * u * w, y * u * w, u * w], dim=2) + A2 = torch.cat([zeros, zeros, zeros, -x * w, -y * w, -w, x * v * w, y * v * w, v * w], dim=2) + A = torch.cat([A1, A2], dim=1) # (B, 2K, 9) + # CUDA SVD is not implemented for fp16/bf16; upcast just for this call. + _, _, Vh = torch.linalg.svd(A.float()) + Vh = Vh.to(A.dtype) + H = Vh[:, -1].reshape(B, 3, 3) + return H / H[:, 2:3, 2:3] + + +def _ransac_find_homography_weighted_batched( + src_pts: torch.Tensor, # (B, N, 2) + dst_pts: torch.Tensor, # (B, N, 2) + confident_weight: torch.Tensor, # (B, N) + n_sample: int, + n_iter: int = 100, + reproj_threshold: float = 3.0, + num_sample_for_ransac: int = 8, + max_inlier_num: int = 10000, + rand_sample_iters_idx: Optional[torch.Tensor] = None, +) -> torch.Tensor: + """Batched weighted-RANSAC homography estimator. Returns (B, 3, 3) homography matrices.""" + B, N, _ = src_pts.shape + assert N >= 4 + device = src_pts.device + + sorted_idx = torch.argsort(confident_weight, descending=True, dim=1) + candidate_idx = sorted_idx[:, :n_sample] # (B, n_sample) + + if rand_sample_iters_idx is None: + rand_sample_iters_idx = torch.stack( + [torch.randperm(n_sample, device=device)[:num_sample_for_ransac] + for _ in range(n_iter)], + dim=0, + ) + + rand_idx = candidate_idx[:, rand_sample_iters_idx] # (B, n_iter, k) + b_idx = ( + torch.arange(B, device=device) + .view(B, 1, 1) + .expand(B, n_iter, num_sample_for_ransac) + ) + src_b = src_pts[b_idx, rand_idx] + dst_b = dst_pts[b_idx, rand_idx] + w_b = confident_weight[b_idx, rand_idx] + + cB, cN = src_b.shape[:2] + H_batch = _find_homography_weighted_lsq_batched( + src_b.flatten(0, 1), dst_b.flatten(0, 1), w_b.flatten(0, 1), + ).unflatten(0, (cB, cN)) # (B, n_iter, 3, 3) + + src_homo = torch.cat([src_pts, torch.ones(B, N, 1, device=device, dtype=src_pts.dtype)], dim=2) + proj = torch.bmm( + src_homo.unsqueeze(1).expand(B, n_iter, N, 3).reshape(-1, N, 3), + H_batch.reshape(-1, 3, 3).transpose(1, 2), + ) # (B*n_iter, N, 3) + proj_xy = (proj[:, :, :2] / proj[:, :, 2:3]).reshape(B, n_iter, N, 2) + err = ((proj_xy - dst_pts.unsqueeze(1)) ** 2).sum(-1).sqrt() # (B, n_iter, N) + inlier_mask = err < reproj_threshold + score = (inlier_mask * confident_weight.unsqueeze(1)).sum(dim=2) + best_idx = torch.argmax(score, dim=1) + best_inlier_mask = inlier_mask[torch.arange(B, device=device), best_idx] + + # Refit with the inlier set (per-batch, since the inlier counts vary). + H_inlier_list = [] + for b in range(B): + mask = best_inlier_mask[b] + in_src = src_pts[b][mask] + in_dst = dst_pts[b][mask] + in_w = confident_weight[b][mask] + if in_src.shape[0] < 4: + # Fall back to identity when RANSAC fails to find enough inliers. + H_inlier_list.append(torch.eye(3, device=device, dtype=src_pts.dtype)) + continue + sorted_w = torch.argsort(in_w, descending=True) + if len(sorted_w) > max_inlier_num: + keep = max(int(len(sorted_w) * 0.95), max_inlier_num) + sorted_w = sorted_w[:keep][torch.randperm(keep, device=device)[:max_inlier_num]] + H_inlier_list.append( + _find_homography_weighted_lsq(in_src[sorted_w], in_dst[sorted_w], in_w[sorted_w]) + ) + return torch.stack(H_inlier_list, dim=0) + + +# ----------------------------------------------------------------------------- +# Camera-ray utilities +# ----------------------------------------------------------------------------- + + +def _unproject_identity(num_y: int, num_x: int, B: int, S: int, device, dtype) -> torch.Tensor: + """Camera-space unit rays for an identity intrinsic on a 2x2 image plane.""" + dx = 1.0 / num_x + dy = 1.0 / num_y + # Centered camera-space coords directly (skip the K^-1 step since it's + # just a translation by -1 on x and y when K is identity-with-center=1). + y = torch.linspace(-(1 - dy), (1 - dy), num_y, device=device, dtype=dtype) + x = torch.linspace(-(1 - dx), (1 - dx), num_x, device=device, dtype=dtype) + yy, xx = torch.meshgrid(y, x, indexing="ij") + grid = torch.stack((xx, yy), dim=-1) # (h, w, 2) + grid = grid.unsqueeze(0).unsqueeze(0).expand(B, S, num_y, num_x, 2) + return torch.cat([grid, torch.ones_like(grid[..., :1])], dim=-1) + + +def _camray_to_caminfo( + camray: torch.Tensor, # (B, S, h, w, 6) + confidence: Optional[torch.Tensor] = None, # (B, S, h, w) + reproj_threshold: float = 0.2, +) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + """Convert per-pixel camera rays to per-view (R, T, focal, principal).""" + if confidence is None: + confidence = torch.ones_like(camray[..., 0]) + B, S, h, w, _ = camray.shape + device = camray.device + dtype = camray.dtype + + rays_target = camray[..., :3] # (B, S, h, w, 3) + rays_origin = _unproject_identity(h, w, B, S, device, dtype) + + # Flatten (B*S, h*w, *) for the RANSAC routine. + rays_target = rays_target.flatten(0, 1).flatten(1, 2) + rays_origin = rays_origin.flatten(0, 1).flatten(1, 2) + weights = confidence.flatten(0, 1).flatten(1, 2).clone() + + # Project to 2D in homogeneous form (the upstream calls this "perspective division"). + z_thresh = 1e-4 + mask = (rays_target[:, :, 2].abs() > z_thresh) & (rays_origin[:, :, 2].abs() > z_thresh) + weights = torch.where(mask, weights, torch.zeros_like(weights)) + src = rays_origin.clone() + dst = rays_target.clone() + src[..., 0] = torch.where(mask, src[..., 0] / src[..., 2], src[..., 0]) + src[..., 1] = torch.where(mask, src[..., 1] / src[..., 2], src[..., 1]) + dst[..., 0] = torch.where(mask, dst[..., 0] / dst[..., 2], dst[..., 0]) + dst[..., 1] = torch.where(mask, dst[..., 1] / dst[..., 2], dst[..., 1]) + src = src[..., :2] + dst = dst[..., :2] + + N = src.shape[1] + n_iter = 100 + sample_ratio = 0.3 + num_sample_for_ransac = 8 + n_sample = max(num_sample_for_ransac, int(N * sample_ratio)) + rand_idx = torch.stack( + [torch.randperm(n_sample, device=device)[:num_sample_for_ransac] for _ in range(n_iter)], + dim=0, + ) + + # Chunk along the view axis to keep peak memory predictable. + chunk = 2 + A_list = [] + for i in range(0, src.shape[0], chunk): + A = _ransac_find_homography_weighted_batched( + src[i:i + chunk], dst[i:i + chunk], weights[i:i + chunk], + n_sample=n_sample, n_iter=n_iter, + num_sample_for_ransac=num_sample_for_ransac, + reproj_threshold=reproj_threshold, + rand_sample_iters_idx=rand_idx, + max_inlier_num=8000, + ) + # Flip sign on dets that come out < 0 (so that the QL produces a + # right-handed rotation). ``det`` lacks fp16/bf16 CUDA kernels, so + # do the comparison in fp32. + flip = torch.linalg.det(A.float()) < 0 + A = torch.where(flip[:, None, None], -A, A) + A_list.append(A) + A = torch.cat(A_list, dim=0) # (B*S, 3, 3) + + R_list, f_list, pp_list = [], [], [] + for i in range(A.shape[0]): + R, L = _ql_decomposition(A[i]) + L = L / L[2][2] + f_list.append(torch.stack((L[0][0], L[1][1]))) + pp_list.append(torch.stack((L[2][0], L[2][1]))) + R_list.append(R) + R = torch.stack(R_list).reshape(B, S, 3, 3) + focal = torch.stack(f_list).reshape(B, S, 2) + pp = torch.stack(pp_list).reshape(B, S, 2) + + # Translation: confidence-weighted average of camray direction(s). + cf = confidence.flatten(0, 1).flatten(1, 2) + T = (camray.flatten(0, 1).flatten(1, 2)[..., 3:] * cf.unsqueeze(-1)).sum(dim=1) + T = T / cf.sum(dim=-1, keepdim=True) + T = T.reshape(B, S, 3) + + # Match upstream output convention: focal -> 1/focal, pp + 1. + return R, T, 1.0 / focal, pp + 1.0 + + +def get_extrinsic_from_camray( + camray: torch.Tensor, # (B, S, h, w, 6) + conf: torch.Tensor, # (B, S, h, w, 1) or (B, S, h, w) + patch_size_y: int, + patch_size_x: int, +) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """Wrap a 4x4 extrinsic + per-view focal + principal-point output.""" + if conf.ndim == 5 and conf.shape[-1] == 1: + conf = conf.squeeze(-1) + R, T, focal, pp = _camray_to_caminfo(camray, confidence=conf) + extr = torch.cat([R, T.unsqueeze(-1)], dim=-1) # (B, S, 3, 4) + homo_row = torch.tensor([0, 0, 0, 1], dtype=R.dtype, device=R.device) + homo_row = homo_row.view(1, 1, 1, 4).expand(R.shape[0], R.shape[1], 1, 4) + extr = torch.cat([extr, homo_row], dim=-2) # (B, S, 4, 4) + return extr, focal, pp diff --git a/comfy/ldm/depth_anything_3/reference_view_selector.py b/comfy/ldm/depth_anything_3/reference_view_selector.py new file mode 100644 index 000000000..90f00be92 --- /dev/null +++ b/comfy/ldm/depth_anything_3/reference_view_selector.py @@ -0,0 +1,87 @@ +"""Reference-view selection for the multi-view path of Depth Anything 3.""" + +from __future__ import annotations + +from typing import Literal + +import torch + + +RefViewStrategy = Literal["first", "middle", "saddle_balanced", "saddle_sim_range"] + + +# Per the upstream constants module: ``THRESH_FOR_REF_SELECTION = 3``. +# Reference selection only runs when there are at least this many views. +THRESH_FOR_REF_SELECTION: int = 3 + + +def select_reference_view(x: torch.Tensor, strategy: RefViewStrategy = "saddle_balanced") -> torch.Tensor: + """Pick a reference view index per batch element.""" + B, S, _, _ = x.shape + if S <= 1: + return torch.zeros(B, dtype=torch.long, device=x.device) + if strategy == "first": + return torch.zeros(B, dtype=torch.long, device=x.device) + if strategy == "middle": + return torch.full((B,), S // 2, dtype=torch.long, device=x.device) + + # Feature-based strategies: normalised cls/cam token per view. + img_class_feat = x[:, :, 0] / x[:, :, 0].norm(dim=-1, keepdim=True) # (B,S,C) + + if strategy == "saddle_balanced": + sim = torch.matmul(img_class_feat, img_class_feat.transpose(1, 2)) # (B,S,S) + sim_no_diag = sim - torch.eye(S, device=sim.device).unsqueeze(0) + sim_score = sim_no_diag.sum(dim=-1) / (S - 1) # (B,S) + feat_norm = x[:, :, 0].norm(dim=-1) # (B,S) + feat_var = img_class_feat.var(dim=-1) # (B,S) + + def _normalize(metric): + mn = metric.min(dim=1, keepdim=True).values + mx = metric.max(dim=1, keepdim=True).values + return (metric - mn) / (mx - mn + 1e-8) + + sim_n, norm_n, var_n = _normalize(sim_score), _normalize(feat_norm), _normalize(feat_var) + balance = (sim_n - 0.5).abs() + (norm_n - 0.5).abs() + (var_n - 0.5).abs() + return balance.argmin(dim=1) + + if strategy == "saddle_sim_range": + sim = torch.matmul(img_class_feat, img_class_feat.transpose(1, 2)) + sim_no_diag = sim - torch.eye(S, device=sim.device).unsqueeze(0) + sim_max = sim_no_diag.max(dim=-1).values + sim_min = sim_no_diag.min(dim=-1).values + return (sim_max - sim_min).argmax(dim=1) + + raise ValueError( + f"Unknown reference view selection strategy: {strategy!r}. " + f"Must be one of: 'first', 'middle', 'saddle_balanced', 'saddle_sim_range'" + ) + + +def reorder_by_reference(x: torch.Tensor, b_idx: torch.Tensor) -> torch.Tensor: + """Reorder x so the reference view is at position 0 in axis S.""" + B, S = x.shape[0], x.shape[1] + if S <= 1: + return x + positions = torch.arange(S, device=x.device).unsqueeze(0).expand(B, -1) + b_idx_exp = b_idx.unsqueeze(1) + reorder = torch.where( + (positions > 0) & (positions <= b_idx_exp), + positions - 1, + positions, + ) + reorder[:, 0] = b_idx + batch = torch.arange(B, device=x.device).unsqueeze(1) + return x[batch, reorder] + + +def restore_original_order(x: torch.Tensor, b_idx: torch.Tensor) -> torch.Tensor: + """Inverse of reorder_by_reference.""" + B, S = x.shape[0], x.shape[1] + if S <= 1: + return x + target_positions = torch.arange(S, device=x.device).unsqueeze(0).expand(B, -1) + b_idx_exp = b_idx.unsqueeze(1) + restore = torch.where(target_positions < b_idx_exp, target_positions + 1, target_positions) + restore = torch.scatter(restore, dim=1, index=b_idx_exp, src=torch.zeros_like(b_idx_exp)) + batch = torch.arange(B, device=x.device).unsqueeze(1) + return x[batch, restore] diff --git a/comfy/ldm/depth_anything_3/transform.py b/comfy/ldm/depth_anything_3/transform.py new file mode 100644 index 000000000..b735d7bec --- /dev/null +++ b/comfy/ldm/depth_anything_3/transform.py @@ -0,0 +1,160 @@ +"""Geometry / camera transform helpers for Depth Anything 3.""" + +from __future__ import annotations + +from typing import Tuple + +import torch +import torch.nn.functional as F + + +# ----------------------------------------------------------------------------- +# Affine 4x4 helpers +# ----------------------------------------------------------------------------- + + +def as_homogeneous(ext: torch.Tensor) -> torch.Tensor: + """Promote (...,3,4) extrinsics to (...,4,4) homogeneous form. No-op when the input is already ``(...,4,4)``.""" + if ext.shape[-2:] == (4, 4): + return ext + if ext.shape[-2:] == (3, 4): + ones = torch.zeros_like(ext[..., :1, :4]) + ones[..., 0, 3] = 1.0 + return torch.cat([ext, ones], dim=-2) + raise ValueError(f"Invalid affine shape: {ext.shape}") + + +def affine_inverse(A: torch.Tensor) -> torch.Tensor: + """Inverse of an affine matrix ``[R|T; 0 0 0 1]``.""" + R = A[..., :3, :3] + T = A[..., :3, 3:] + P = A[..., 3:, :] + return torch.cat([torch.cat([R.mT, -R.mT @ T], dim=-1), P], dim=-2) + + +# ----------------------------------------------------------------------------- +# Quaternion <-> rotation matrix (xyzw / scalar-last) +# ----------------------------------------------------------------------------- + + +def _sqrt_positive_part(x: torch.Tensor) -> torch.Tensor: + """sqrt(max(0, x)) with a zero subgradient where x == 0.""" + ret = torch.zeros_like(x) + positive_mask = x > 0 + if torch.is_grad_enabled(): + ret[positive_mask] = torch.sqrt(x[positive_mask]) + else: + ret = torch.where(positive_mask, torch.sqrt(x), ret) + return ret + + +def standardize_quaternion(quaternions: torch.Tensor) -> torch.Tensor: + """Force the real part of a unit quaternion (xyzw) to be non-negative.""" + return torch.where(quaternions[..., 3:4] < 0, -quaternions, quaternions) + + +def quat_to_mat(quaternions: torch.Tensor) -> torch.Tensor: + """Convert quaternions (xyzw) to (...,3,3) rotation matrices.""" + i, j, k, r = torch.unbind(quaternions, -1) + two_s = 2.0 / (quaternions * quaternions).sum(-1) + o = torch.stack( + ( + 1 - two_s * (j * j + k * k), + two_s * (i * j - k * r), + two_s * (i * k + j * r), + two_s * (i * j + k * r), + 1 - two_s * (i * i + k * k), + two_s * (j * k - i * r), + two_s * (i * k - j * r), + two_s * (j * k + i * r), + 1 - two_s * (i * i + j * j), + ), + -1, + ) + return o.reshape(quaternions.shape[:-1] + (3, 3)) + + +def mat_to_quat(matrix: torch.Tensor) -> torch.Tensor: + """Convert (...,3,3) rotation matrices to quaternions (xyzw).""" + if matrix.size(-1) != 3 or matrix.size(-2) != 3: + raise ValueError(f"Invalid rotation matrix shape {matrix.shape}.") + + batch_dim = matrix.shape[:-2] + m00, m01, m02, m10, m11, m12, m20, m21, m22 = torch.unbind( + matrix.reshape(batch_dim + (9,)), dim=-1 + ) + + q_abs = _sqrt_positive_part( + torch.stack( + [ + 1.0 + m00 + m11 + m22, + 1.0 + m00 - m11 - m22, + 1.0 - m00 + m11 - m22, + 1.0 - m00 - m11 + m22, + ], + dim=-1, + ) + ) + + quat_by_rijk = torch.stack( + [ + torch.stack([q_abs[..., 0] ** 2, m21 - m12, m02 - m20, m10 - m01], dim=-1), + torch.stack([m21 - m12, q_abs[..., 1] ** 2, m10 + m01, m02 + m20], dim=-1), + torch.stack([m02 - m20, m10 + m01, q_abs[..., 2] ** 2, m12 + m21], dim=-1), + torch.stack([m10 - m01, m20 + m02, m21 + m12, q_abs[..., 3] ** 2], dim=-1), + ], + dim=-2, + ) + + flr = torch.tensor(0.1).to(dtype=q_abs.dtype, device=q_abs.device) + quat_candidates = quat_by_rijk / (2.0 * q_abs[..., None].max(flr)) + + out = quat_candidates[F.one_hot(q_abs.argmax(dim=-1), num_classes=4) > 0.5, :].reshape( + batch_dim + (4,) + ) + # Reorder rijk -> xyzw (i.e. ijkr). + out = out[..., [1, 2, 3, 0]] + return standardize_quaternion(out) + + +# ----------------------------------------------------------------------------- +# Pose-encoding <-> extrinsics + intrinsics +# ----------------------------------------------------------------------------- + + +def extri_intri_to_pose_encoding(extrinsics: torch.Tensor, intrinsics: torch.Tensor, image_size_hw: Tuple[int, int]) -> torch.Tensor: + """Pack (extr, intr, image_size) into the 9-D pose-encoding vector. + extrinsics: camera-to-world (c2w) (B,S,4,4) matrices, + intrinsics: pixel-space (B,S,3,3) matrices, + image_size_hw: is a (H, W) pair. + """ + R = extrinsics[..., :3, :3] + T = extrinsics[..., :3, 3] + quat = mat_to_quat(R) + H, W = image_size_hw + fov_h = 2 * torch.atan((H / 2) / intrinsics[..., 1, 1]) + fov_w = 2 * torch.atan((W / 2) / intrinsics[..., 0, 0]) + return torch.cat([T, quat, fov_h[..., None], fov_w[..., None]], dim=-1).float() + + +def pose_encoding_to_extri_intri(pose_encoding: torch.Tensor, image_size_hw: Tuple[int, int]) -> Tuple[torch.Tensor, torch.Tensor]: + """Inverse of extri_intri_to_pose_encoding.""" + T = pose_encoding[..., :3] + quat = pose_encoding[..., 3:7] + fov_h = pose_encoding[..., 7] + fov_w = pose_encoding[..., 8] + # Normalize to unit quaternion. CameraDec outputs raw values; a near-zero + # quaternion causes two_s = 2/norm² → inf in quat_to_mat → NaN extrinsics. + quat = quat / quat.norm(dim=-1, keepdim=True).clamp(min=1e-6) + R = quat_to_mat(quat) + extrinsics = torch.cat([R, T[..., None]], dim=-1) + H, W = image_size_hw + fy = (H / 2.0) / torch.clamp(torch.tan(fov_h / 2.0), 1e-6) + fx = (W / 2.0) / torch.clamp(torch.tan(fov_w / 2.0), 1e-6) + intrinsics = torch.zeros(pose_encoding.shape[:2] + (3, 3), device=pose_encoding.device, dtype=pose_encoding.dtype) + intrinsics[..., 0, 0] = fx + intrinsics[..., 1, 1] = fy + intrinsics[..., 0, 2] = W / 2 + intrinsics[..., 1, 2] = H / 2 + intrinsics[..., 2, 2] = 1.0 + return extrinsics, intrinsics diff --git a/comfy/ldm/ernie/model.py b/comfy/ldm/ernie/model.py index eba661aec..f158ca1d2 100644 --- a/comfy/ldm/ernie/model.py +++ b/comfy/ldm/ernie/model.py @@ -5,6 +5,7 @@ import torch.nn.functional as F from comfy.ldm.modules.attention import optimized_attention import comfy.model_management +import comfy.quant_ops def rope(pos: torch.Tensor, dim: int, theta: int) -> torch.Tensor: assert dim % 2 == 0 @@ -19,15 +20,6 @@ def rope(pos: torch.Tensor, dim: int, theta: int) -> torch.Tensor: out = torch.stack([torch.cos(out), torch.sin(out)], dim=0) return out.to(dtype=torch.float32, device=pos.device) -def apply_rotary_emb(x_in: torch.Tensor, freqs_cis: torch.Tensor) -> torch.Tensor: - rot_dim = freqs_cis.shape[-1] - x, x_pass = x_in[..., :rot_dim], x_in[..., rot_dim:] - cos_ = freqs_cis[0] - sin_ = freqs_cis[1] - x1, x2 = x.chunk(2, dim=-1) - x_rotated = torch.cat((-x2, x1), dim=-1) - return torch.cat((x * cos_ + x_rotated * sin_, x_pass), dim=-1) - class ErnieImageEmbedND3(nn.Module): def __init__(self, dim: int, theta: int, axes_dim: tuple): super().__init__() @@ -37,8 +29,16 @@ class ErnieImageEmbedND3(nn.Module): def forward(self, ids: torch.Tensor) -> torch.Tensor: emb = torch.cat([rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(3)], dim=-1) - emb = emb.unsqueeze(3) # [2, B, S, 1, head_dim//2] - return torch.stack([emb, emb], dim=-1).reshape(*emb.shape[:-1], -1) # [B, S, 1, head_dim] + cos_ = emb[0] + sin_ = emb[1] + N = cos_.shape[-1] + half = N // 2 + cos_top = cos_[..., :half].repeat_interleave(2, dim=-1) + sin_top = sin_[..., :half].repeat_interleave(2, dim=-1) + cos_bot = cos_[..., half:].repeat_interleave(2, dim=-1) + sin_bot = sin_[..., half:].repeat_interleave(2, dim=-1) + rot = torch.stack([cos_top, -sin_top, sin_bot, cos_bot], dim=-1) + return rot.reshape(*rot.shape[:-1], 2, 2).unsqueeze(2) class ErnieImagePatchEmbedDynamic(nn.Module): def __init__(self, in_channels: int, embed_dim: int, patch_size: int, operations, device=None, dtype=None): @@ -115,8 +115,7 @@ class ErnieImageAttention(nn.Module): key = self.norm_k(key) if image_rotary_emb is not None: - query = apply_rotary_emb(query, image_rotary_emb) - key = apply_rotary_emb(key, image_rotary_emb) + query, key = comfy.quant_ops.ck.apply_rope_split_half(query, key, image_rotary_emb) q_flat = query.reshape(B, S, -1) k_flat = key.reshape(B, S, -1) @@ -274,7 +273,7 @@ class ErnieImageModel(nn.Module): image_ids = image_ids.view(1, N_img, 3).expand(B, -1, -1) - rotary_pos_emb = self.pos_embed(torch.cat([image_ids, text_ids], dim=1)).to(x.dtype) + rotary_pos_emb = self.pos_embed(torch.cat([image_ids, text_ids], dim=1)) del image_ids, text_ids sample = self.time_proj(timesteps).to(dtype) diff --git a/comfy/ldm/flux/math.py b/comfy/ldm/flux/math.py index 6d0aed827..891dea7dd 100644 --- a/comfy/ldm/flux/math.py +++ b/comfy/ldm/flux/math.py @@ -4,7 +4,7 @@ from torch import Tensor from comfy.ldm.modules.attention import optimized_attention import comfy.model_management -import logging +import comfy.quant_ops def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transformer_options={}) -> Tensor: @@ -44,21 +44,15 @@ def _apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor): return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis) -try: - import comfy.quant_ops - q_apply_rope = comfy.quant_ops.ck.apply_rope - q_apply_rope1 = comfy.quant_ops.ck.apply_rope1 - def apply_rope(xq, xk, freqs_cis): - if comfy.model_management.in_training: - return _apply_rope(xq, xk, freqs_cis) - else: - return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis) - def apply_rope1(x, freqs_cis): - if comfy.model_management.in_training: - return _apply_rope1(x, freqs_cis) - else: - return q_apply_rope1(x, freqs_cis) -except: - logging.warning("No comfy kitchen, using old apply_rope functions.") - apply_rope = _apply_rope - apply_rope1 = _apply_rope1 +def apply_rope(xq, xk, freqs_cis): + if comfy.model_management.in_training: + return _apply_rope(xq, xk, freqs_cis) + else: + return comfy.quant_ops.ck.apply_rope(xq, xk, freqs_cis) + + +def apply_rope1(x, freqs_cis): + if comfy.model_management.in_training: + return _apply_rope1(x, freqs_cis) + else: + return comfy.quant_ops.ck.apply_rope1(x, freqs_cis) diff --git a/comfy/ldm/hunyuan3dv2_1/hunyuandit.py b/comfy/ldm/hunyuan3dv2_1/hunyuandit.py index bc36b8998..4e4819fe3 100644 --- a/comfy/ldm/hunyuan3dv2_1/hunyuandit.py +++ b/comfy/ldm/hunyuan3dv2_1/hunyuandit.py @@ -607,9 +607,13 @@ class HunYuanDiTPlain(nn.Module): def forward(self, x, t, context, transformer_options = {}, **kwargs): x = x.movedim(-1, -2) - if context.shape[0] >= 2: - uncond_emb, cond_emb = context.chunk(2, dim = 0) - context = torch.cat([cond_emb, uncond_emb], dim = 0) + + swap_cfg_halves = context.shape[0] >= 2 + + if swap_cfg_halves: + first_half, second_half = context.chunk(2, dim = 0) + context = torch.cat([second_half, first_half], dim = 0) + main_condition = context t = 1.0 - t @@ -657,8 +661,8 @@ class HunYuanDiTPlain(nn.Module): output = self.final_layer(combined) output = output.movedim(-2, -1) * (-1.0) - if output.shape[0] >= 2: - cond_emb, uncond_emb = output.chunk(2, dim = 0) - return torch.cat([uncond_emb, cond_emb]) - else: - return output + if swap_cfg_halves: + first_half, second_half = output.chunk(2, dim = 0) + output = torch.cat([second_half, first_half], dim = 0) + + return output diff --git a/comfy/ldm/ideogram4/model.py b/comfy/ldm/ideogram4/model.py new file mode 100644 index 000000000..b86c65bf0 --- /dev/null +++ b/comfy/ldm/ideogram4/model.py @@ -0,0 +1,297 @@ +""" +The Ideogram 4 transformer is a NextDiT/Lumina2-family single-stream model +consumes Qwen3-VL hidden-state features (concatenated from 13 layers -> 53248 dims) +packs ``[text tokens, image tokens]`` into one sequence with block-diagonal segment attention and 3D interleaved MRoPE. +""" + +from __future__ import annotations + +import math + +import torch +import torch.nn as nn +import torch.nn.functional as F + +import comfy.patcher_extension +from comfy.ldm.lumina.model import FeedForward +from comfy.ldm.modules.attention import optimized_attention_masked +from comfy.text_encoders.llama import apply_rope, precompute_freqs_cis + +# Per-token role indicators +SEQUENCE_PADDING_INDICATOR = -1 +OUTPUT_IMAGE_INDICATOR = 2 +LLM_TOKEN_INDICATOR = 3 +# Image grid coordinates are offset so they never collide with text positions +IMAGE_POSITION_OFFSET = 65536 + + +class Ideogram4Attention(nn.Module): + def __init__(self, hidden_size, num_heads, eps=1e-5, dtype=None, device=None, operations=None): + super().__init__() + self.num_heads = num_heads + self.head_dim = hidden_size // num_heads + self.hidden_size = hidden_size + + self.qkv = operations.Linear(hidden_size, hidden_size * 3, bias=False, dtype=dtype, device=device) + self.norm_q = operations.RMSNorm(self.head_dim, eps=eps, elementwise_affine=True, dtype=dtype, device=device) + self.norm_k = operations.RMSNorm(self.head_dim, eps=eps, elementwise_affine=True, dtype=dtype, device=device) + self.o = operations.Linear(hidden_size, hidden_size, bias=False, dtype=dtype, device=device) + + def forward(self, x, attn_mask, freqs_cis, transformer_options={}): + batch_size, seq_len, _ = x.shape + qkv = self.qkv(x).view(batch_size, seq_len, 3, self.num_heads, self.head_dim) + q, k, v = qkv.unbind(dim=2) + + q = self.norm_q(q) + k = self.norm_k(k) + + # (B, heads, L, head_dim) + q = q.transpose(1, 2) + k = k.transpose(1, 2) + v = v.transpose(1, 2) + + q, k = apply_rope(q, k, freqs_cis) + + out = optimized_attention_masked(q, k, v, self.num_heads, attn_mask, skip_reshape=True, transformer_options=transformer_options) + return self.o(out) + + +class Ideogram4TransformerBlock(nn.Module): + def __init__(self, hidden_size, intermediate_size, num_heads, norm_eps, adaln_dim, dtype=None, device=None, operations=None): + super().__init__() + self.attention = Ideogram4Attention(hidden_size, num_heads, eps=1e-5, dtype=dtype, device=device, operations=operations) + self.feed_forward = FeedForward( + dim=hidden_size, hidden_dim=intermediate_size, multiple_of=1, ffn_dim_multiplier=None, + operation_settings={"operations": operations, "dtype": dtype, "device": device}, + ) + + self.attention_norm1 = operations.RMSNorm(hidden_size, eps=norm_eps, elementwise_affine=True, dtype=dtype, device=device) + self.ffn_norm1 = operations.RMSNorm(hidden_size, eps=norm_eps, elementwise_affine=True, dtype=dtype, device=device) + self.attention_norm2 = operations.RMSNorm(hidden_size, eps=norm_eps, elementwise_affine=True, dtype=dtype, device=device) + self.ffn_norm2 = operations.RMSNorm(hidden_size, eps=norm_eps, elementwise_affine=True, dtype=dtype, device=device) + + self.adaln_modulation = operations.Linear(adaln_dim, 4 * hidden_size, bias=True, dtype=dtype, device=device) + + def forward(self, x, attn_mask, freqs_cis, adaln_input, transformer_options={}): + mod = self.adaln_modulation(adaln_input) + scale_msa, gate_msa, scale_mlp, gate_mlp = mod.chunk(4, dim=-1) + gate_msa = torch.tanh(gate_msa) + gate_mlp = torch.tanh(gate_mlp) + scale_msa = 1.0 + scale_msa + scale_mlp = 1.0 + scale_mlp + + attn_out = self.attention(self.attention_norm1(x) * scale_msa, attn_mask, freqs_cis, transformer_options=transformer_options) + x = x + gate_msa * self.attention_norm2(attn_out) + x = x + gate_mlp * self.ffn_norm2(self.feed_forward(self.ffn_norm1(x) * scale_mlp)) + return x + + +def _sinusoidal_embedding(t, dim, scale=1e4): + t = t.to(torch.float32) + half = dim // 2 + freq = math.log(scale) / (half - 1) + freq = torch.exp(torch.arange(half, dtype=torch.float32, device=t.device) * -freq) + emb = t.unsqueeze(-1) * freq + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1) + if dim % 2 == 1: + emb = F.pad(emb, (0, 1)) + return emb + + +class Ideogram4EmbedScalar(nn.Module): + def __init__(self, dim, input_range=(0.0, 1.0), dtype=None, device=None, operations=None): + super().__init__() + self.dim = dim + self.range_min, self.range_max = input_range + self.mlp_in = operations.Linear(dim, dim, bias=True, dtype=dtype, device=device) + self.mlp_out = operations.Linear(dim, dim, bias=True, dtype=dtype, device=device) + + def forward(self, x): + x = x.to(torch.float32) + scaled = 1e4 * (x - self.range_min) / (self.range_max - self.range_min) + emb = _sinusoidal_embedding(scaled, self.dim) + emb = emb.to(self.mlp_in.weight.dtype) + emb = F.silu(self.mlp_in(emb)) + return self.mlp_out(emb) + + +class Ideogram4FinalLayer(nn.Module): + def __init__(self, hidden_size, out_channels, adaln_dim, dtype=None, device=None, operations=None): + super().__init__() + self.norm_final = operations.LayerNorm(hidden_size, eps=1e-6, elementwise_affine=False, dtype=dtype, device=device) + self.linear = operations.Linear(hidden_size, out_channels, bias=True, dtype=dtype, device=device) + self.adaln_modulation = operations.Linear(adaln_dim, hidden_size, bias=True, dtype=dtype, device=device) + + def forward(self, x, c): + scale = 1.0 + self.adaln_modulation(F.silu(c)) + return self.linear(self.norm_final(x) * scale) + + +class Ideogram4Transformer(nn.Module): + """A single Ideogram 4 backbone operating on a packed token sequence.""" + + def __init__(self, emb_dim, num_layers, num_heads, intermediate_size, adaln_dim, + in_channels, llm_features_dim, rope_theta, mrope_section, norm_eps, + dtype=None, device=None, operations=None): + super().__init__() + self.head_dim = emb_dim // num_heads + self.rope_theta = rope_theta + self.mrope_section = tuple(mrope_section) + + self.input_proj = operations.Linear(in_channels, emb_dim, bias=True, dtype=dtype, device=device) + self.llm_cond_norm = operations.RMSNorm(llm_features_dim, eps=1e-6, elementwise_affine=True, dtype=dtype, device=device) + self.llm_cond_proj = operations.Linear(llm_features_dim, emb_dim, bias=True, dtype=dtype, device=device) + self.t_embedding = Ideogram4EmbedScalar(emb_dim, input_range=(0.0, 1.0), dtype=dtype, device=device, operations=operations) + self.adaln_proj = operations.Linear(emb_dim, adaln_dim, bias=True, dtype=dtype, device=device) + + self.embed_image_indicator = operations.Embedding(2, emb_dim, dtype=dtype, device=device) + + self.layers = nn.ModuleList([ + Ideogram4TransformerBlock(emb_dim, intermediate_size, num_heads, norm_eps, adaln_dim, + dtype=dtype, device=device, operations=operations) + for _ in range(num_layers) + ]) + + self.final_layer = Ideogram4FinalLayer(emb_dim, in_channels, adaln_dim, dtype=dtype, device=device, operations=operations) + + def _backbone(self, llm_features, x, t, position_ids, attn_mask, indicator, transformer_options={}): + indicator = indicator.to(torch.long) + output_image_mask = (indicator == OUTPUT_IMAGE_INDICATOR).to(x.dtype).unsqueeze(-1) + + x = x * output_image_mask + h = self.input_proj(x) * output_image_mask + + t_cond = self.t_embedding(t) + if t.dim() == 1: + t_cond = t_cond.unsqueeze(1) + adaln_input = F.silu(self.adaln_proj(t_cond)) + + # h is zero on the text rows (content lives only on image rows), add writes the text features in place + if llm_features is not None: + L_text = llm_features.shape[1] + text_mask = (indicator[:, :L_text] == LLM_TOKEN_INDICATOR).to(x.dtype).unsqueeze(-1) + llm = self.llm_cond_norm(llm_features * text_mask) + llm = self.llm_cond_proj(llm) * text_mask + h[:, :L_text] = h[:, :L_text] + llm + + h = h + self.embed_image_indicator((indicator == OUTPUT_IMAGE_INDICATOR).to(torch.long), out_dtype=h.dtype) + + # Qwen3-VL interleaved MRoPE; position_ids (B, L, 3) -> (3, L) (same across batch). + freqs_cis = precompute_freqs_cis( + self.head_dim, position_ids[0].transpose(0, 1), self.rope_theta, + rope_dims=self.mrope_section, interleaved_mrope=True, device=position_ids.device, + ) + + if attn_mask is not None and attn_mask.dtype == torch.bool: + attn_mask = torch.zeros_like(attn_mask, dtype=h.dtype).masked_fill_(~attn_mask, -torch.finfo(h.dtype).max) + + for layer in self.layers: + h = layer(h, attn_mask, freqs_cis, adaln_input, transformer_options=transformer_options) + + return self.final_layer(h, adaln_input) + + +class Ideogram4Transformer2DModel(Ideogram4Transformer): + """Ideogram 4 single-stream DiT. + + Runs a packed ``[text, image]`` sequence when text context is supplied, or an image-only sequence when ``context is None``. + """ + + def __init__(self, image_model=None, in_channels=128, num_layers=34, num_attention_heads=18, attention_head_dim=256, intermediate_size=12288, + adaln_dim=512, llm_features_dim=53248, rope_theta=5000000, mrope_section=(24, 20, 20), norm_eps=1e-5, + dtype=None, device=None, operations=None, **kwargs): + emb_dim = num_attention_heads * attention_head_dim + super().__init__( + emb_dim=emb_dim, num_layers=num_layers, num_heads=num_attention_heads, + intermediate_size=intermediate_size, adaln_dim=adaln_dim, in_channels=in_channels, + llm_features_dim=llm_features_dim, rope_theta=rope_theta, mrope_section=mrope_section, + norm_eps=norm_eps, dtype=dtype, device=device, operations=operations) + self.dtype = dtype + self.in_channels = in_channels + self.out_channels = in_channels + # 128-dim token = patch (2x2) * ae_channels (32). + self.patch_size = 2 + self.ae_channels = in_channels // (self.patch_size * self.patch_size) + + def _img_to_tokens(self, x): + B, C, gh, gw = x.shape + x = x.view(B, self.ae_channels, self.patch_size, self.patch_size, gh, gw) + x = x.permute(0, 4, 5, 2, 3, 1) # (B, gh, gw, pi, pj, c) + return x.reshape(B, gh * gw, C) + + def _tokens_to_img(self, tokens, gh, gw): + B = tokens.shape[0] + C = tokens.shape[-1] + x = tokens.reshape(B, gh, gw, self.patch_size, self.patch_size, self.ae_channels) + x = x.permute(0, 5, 3, 4, 1, 2) # (B, c, pi, pj, gh, gw) + return x.reshape(B, C, gh, gw) + + def _image_position_ids(self, gh, gw, device): + h_idx = torch.arange(gh, device=device).view(-1, 1).expand(gh, gw).reshape(-1) + w_idx = torch.arange(gw, device=device).view(1, -1).expand(gh, gw).reshape(-1) + t_idx = torch.zeros_like(h_idx) + return torch.stack([t_idx, h_idx, w_idx], dim=1) + IMAGE_POSITION_OFFSET # (L_img, 3) + + def _run_conditional(self, x_chunk, context_chunk, attn_mask_chunk, t_chunk, gh, gw, transformer_options): + B = x_chunk.shape[0] + device = x_chunk.device + img_tokens = self._img_to_tokens(x_chunk) + L_img = img_tokens.shape[1] + L_text = context_chunk.shape[1] + L = L_text + L_img + latent_dim = img_tokens.shape[-1] + + x_full = torch.zeros(B, L, latent_dim, dtype=img_tokens.dtype, device=device) + x_full[:, L_text:] = img_tokens + + text_pos = torch.arange(L_text, device=device).view(-1, 1).expand(L_text, 3) + img_pos = self._image_position_ids(gh, gw, device) + position_ids = torch.cat([text_pos, img_pos], dim=0).unsqueeze(0).expand(B, L, 3) + + indicator = torch.empty(B, L, dtype=torch.long, device=device) + indicator[:, :L_text] = LLM_TOKEN_INDICATOR + indicator[:, L_text:] = OUTPUT_IMAGE_INDICATOR + + attn_mask = None + if attn_mask_chunk is not None: + segment_ids = torch.ones(B, L, dtype=torch.long, device=device) + pad = (attn_mask_chunk == 0) + segment_ids[:, :L_text][pad] = SEQUENCE_PADDING_INDICATOR + indicator[:, :L_text][pad] = 0 + # Block-diagonal mask from segment ids: (B, 1, L, L), True = attend. + attn_mask = (segment_ids.unsqueeze(2) == segment_ids.unsqueeze(1)).unsqueeze(1) + + out = self._backbone(context_chunk, x_full, t_chunk, position_ids, attn_mask, indicator, + transformer_options=transformer_options) + return self._tokens_to_img(out[:, L_text:], gh, gw) + + def _run_image_only(self, x_chunk, t_chunk, gh, gw, transformer_options): + B = x_chunk.shape[0] + device = x_chunk.device + img_tokens = self._img_to_tokens(x_chunk) + L_img = img_tokens.shape[1] + + position_ids = self._image_position_ids(gh, gw, device).unsqueeze(0).expand(B, L_img, 3) + indicator = torch.full((B, L_img), OUTPUT_IMAGE_INDICATOR, dtype=torch.long, device=device) + + # Image-only sequence is a single segment -> no mask, full attention, no LLM context. + out = self._backbone(None, img_tokens, t_chunk, position_ids, None, indicator, transformer_options=transformer_options) + return self._tokens_to_img(out, gh, gw) + + def forward(self, x, timesteps, context=None, attention_mask=None, transformer_options={}, **kwargs): + return comfy.patcher_extension.WrapperExecutor.new_class_executor( + self._forward, + self, + comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options), + ).execute(x, timesteps, context, attention_mask, transformer_options, **kwargs) + + def _forward(self, x, timesteps, context=None, attention_mask=None, transformer_options={}, **kwargs): + bs, c, gh, gw = x.shape + + timesteps = 1.0 - timesteps + + # unconditional pass + if context is None: + return -self._run_image_only(x, timesteps, gh, gw, transformer_options) + + return -self._run_conditional(x, context, attention_mask, timesteps, gh, gw, transformer_options) diff --git a/comfy/ldm/lens/model.py b/comfy/ldm/lens/model.py new file mode 100644 index 000000000..cd5015ddc --- /dev/null +++ b/comfy/ldm/lens/model.py @@ -0,0 +1,510 @@ +"""Lens denoising transformer (DiT)""" + +from __future__ import annotations + +from typing import Any, Dict, Optional, Tuple + +import torch +import torch.nn as nn +import torch.nn.functional as F + +import comfy.ldm.flux.layers +import comfy.patcher_extension +from comfy.ldm.flux.layers import EmbedND +from comfy.ldm.flux.math import apply_rope +from comfy.ldm.modules.attention import optimized_attention + + +def _lens_time_proj(t: torch.Tensor, dim: int = 256) -> torch.Tensor: + return comfy.ldm.flux.layers.timestep_embedding(t, dim) + + +def _lens_position_ids( + frame: int, height: int, width: int, text_seq_len: int, + scale_rope: bool = True, device=None, +) -> torch.Tensor: + """Lens axial (frame, h, w) position ids for joint image + text sequence. + + With ``scale_rope=True`` h/w are centered around 0 (negative + positive + halves) and text starts at ``max(h//2, w//2)``. Result shape ``[seq, 3]``; + caller adds a batch dim for ``EmbedND``. + """ + if scale_rope: + h_pos = torch.cat([torch.arange(-(height - height // 2), 0, device=device), + torch.arange(0, height // 2, device=device)]) + w_pos = torch.cat([torch.arange(-(width - width // 2), 0, device=device), + torch.arange(0, width // 2, device=device)]) + text_start = max(height // 2, width // 2) + else: + h_pos = torch.arange(height, device=device) + w_pos = torch.arange(width, device=device) + text_start = max(height, width) + + f_pos = torch.arange(frame, device=device) + img_ids = torch.zeros(frame, height, width, 3, device=device) + img_ids[..., 0] = f_pos[:, None, None] + img_ids[..., 1] = h_pos[None, :, None] + img_ids[..., 2] = w_pos[None, None, :] + img_ids = img_ids.reshape(-1, 3) + + # Text positions replicate across all 3 axes (matches original packing). + txt_pos = torch.arange(text_start, text_start + text_seq_len, device=device).float() + txt_ids = txt_pos[:, None].expand(text_seq_len, 3) + + return torch.cat([img_ids, txt_ids], dim=0) + + +class _TimestepEmbedder(nn.Module): + def __init__(self, in_channels: int, time_embed_dim: int, dtype=None, device=None, operations=None) -> None: + super().__init__() + self.linear_1 = operations.Linear(in_channels, time_embed_dim, dtype=dtype, device=device) + self.linear_2 = operations.Linear(time_embed_dim, time_embed_dim, dtype=dtype, device=device) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + x = self.linear_1(x) + x = F.silu(x) + return self.linear_2(x) + + +class LensTimestepProjEmbeddings(nn.Module): + def __init__(self, embedding_dim: int, dtype=None, device=None, operations=None) -> None: + super().__init__() + self.timestep_embedder = _TimestepEmbedder(256, embedding_dim, dtype=dtype, device=device, operations=operations) + + def forward(self, timestep: torch.Tensor, hidden_states: torch.Tensor) -> torch.Tensor: + proj = _lens_time_proj(timestep, 256) + return self.timestep_embedder(proj.to(dtype=hidden_states.dtype)) + + +class GateMLP(nn.Module): + """SwiGLU MLP.""" + + def __init__(self, dim: int, hidden_dim: int, dtype=None, device=None, operations=None) -> None: + super().__init__() + self.w1 = operations.Linear(dim, hidden_dim, bias=False, dtype=dtype, device=device) + self.w2 = operations.Linear(hidden_dim, dim, bias=False, dtype=dtype, device=device) + self.w3 = operations.Linear(dim, hidden_dim, bias=False, dtype=dtype, device=device) + + def forward(self, x): + return self.w2(F.silu(self.w1(x), inplace=True).mul_(self.w3(x))) + + +class LensJointAttention(nn.Module): + """Joint image+text attention with fused QKV per stream.""" + + def __init__( + self, + query_dim: int, + added_kv_proj_dim: int, + dim_head: int = 64, + heads: int = 8, + out_dim: Optional[int] = None, + eps: float = 1e-5, + dtype=None, + device=None, + operations=None, + ) -> None: + super().__init__() + self.inner_dim = out_dim if out_dim is not None else dim_head * heads + self.heads = self.inner_dim // dim_head + self.dim_head = dim_head + self.out_dim = out_dim if out_dim is not None else query_dim + + self.norm_q = operations.RMSNorm(dim_head, eps=eps, dtype=dtype, device=device) + self.norm_k = operations.RMSNorm(dim_head, eps=eps, dtype=dtype, device=device) + self.norm_added_q = operations.RMSNorm(dim_head, eps=eps, dtype=dtype, device=device) + self.norm_added_k = operations.RMSNorm(dim_head, eps=eps, dtype=dtype, device=device) + + self.img_qkv = operations.Linear(query_dim, 3 * self.inner_dim, bias=True, dtype=dtype, device=device) + self.txt_qkv = operations.Linear(added_kv_proj_dim, 3 * self.inner_dim, bias=True, dtype=dtype, device=device) + + # ModuleList([Linear, Identity]) for state-dict key compatibility. + self.to_out = nn.ModuleList([ + operations.Linear(self.inner_dim, self.out_dim, bias=True, dtype=dtype, device=device), + nn.Identity(), + ]) + self.to_add_out = operations.Linear(self.inner_dim, query_dim, bias=True, dtype=dtype, device=device) + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: torch.Tensor, + freqs_cis: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + transformer_options: Optional[Dict[str, Any]] = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + bsz, seq_img, _ = hidden_states.shape + seq_txt = encoder_hidden_states.shape[1] + + # image stream + img_qkv = self.img_qkv(hidden_states).view(bsz, seq_img, 3, self.heads, self.dim_head) + img_q, img_k, img_v = img_qkv.unbind(dim=2) + img_q = self.norm_q(img_q) + img_k = self.norm_k(img_k) + del img_qkv + + # text stream + txt_qkv = self.txt_qkv(encoder_hidden_states).view(bsz, seq_txt, 3, self.heads, self.dim_head) + txt_q, txt_k, txt_v = txt_qkv.unbind(dim=2) + txt_q = self.norm_added_q(txt_q) + txt_k = self.norm_added_k(txt_k) + + # [B, S, H, D] → [B, H, S, D] for attention, dels to avoid VRAM peaks + q = torch.cat([img_q, txt_q], dim=1).transpose(1, 2) + del img_q, txt_q + k = torch.cat([img_k, txt_k], dim=1).transpose(1, 2) + del img_k, txt_k + v = torch.cat([img_v, txt_v], dim=1).transpose(1, 2) + del img_v, txt_v + + q, k = apply_rope(q, k, freqs_cis) + + if attention_mask is not None: + expected = (bsz, 1, 1, seq_img + seq_txt) + if attention_mask.shape != expected: + raise ValueError( + f"attention_mask must be {expected}, got {tuple(attention_mask.shape)}" + ) + attention_mask = attention_mask.to(q.dtype) + + out = optimized_attention( + q, k, v, self.heads, mask=attention_mask, skip_reshape=True, + transformer_options=transformer_options, + ) + + img_out = self.to_out[1](self.to_out[0](out[:, :seq_img, :])) + txt_out = self.to_add_out(out[:, seq_img:, :]) + return img_out, txt_out + + +class LensTransformerBlock(nn.Module): + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + eps: float = 1e-6, + rms_norm: bool = True, + dtype=None, + device=None, + operations=None, + ) -> None: + super().__init__() + + self.attn = LensJointAttention( + query_dim=dim, + added_kv_proj_dim=dim, + dim_head=attention_head_dim, + heads=num_attention_heads, + out_dim=dim, + eps=1e-5, + dtype=dtype, + device=device, + operations=operations, + ) + + if rms_norm: + NormCls = operations.RMSNorm + norm_kwargs = {} + else: + NormCls = operations.LayerNorm + norm_kwargs = {"elementwise_affine": False} + + mlp_hidden = int(dim / 3 * 8) + + # Sequential(SiLU, Linear) so state-dict lands at img_mod.1.{weight,bias}. + self.img_mod = nn.Sequential( + nn.SiLU(), + operations.Linear(dim, 6 * dim, bias=True, dtype=dtype, device=device), + ) + self.img_norm1 = NormCls(dim, eps=eps, dtype=dtype, device=device, **norm_kwargs) + self.img_norm2 = NormCls(dim, eps=eps, dtype=dtype, device=device, **norm_kwargs) + self.img_mlp = GateMLP(dim, mlp_hidden, dtype=dtype, device=device, operations=operations) + + self.txt_mod = nn.Sequential( + nn.SiLU(), + operations.Linear(dim, 6 * dim, bias=True, dtype=dtype, device=device), + ) + self.txt_norm1 = NormCls(dim, eps=eps, dtype=dtype, device=device, **norm_kwargs) + self.txt_norm2 = NormCls(dim, eps=eps, dtype=dtype, device=device, **norm_kwargs) + self.txt_mlp = GateMLP(dim, mlp_hidden, dtype=dtype, device=device, operations=operations) + + @staticmethod + def _modulate(x: torch.Tensor, mod_params: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: + shift, scale, gate = mod_params.chunk(3, dim=-1) + return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1), gate.unsqueeze(1) + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: torch.Tensor, + temb: torch.Tensor, + freqs_cis: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + transformer_options: Optional[Dict[str, Any]] = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + img_mod1, img_mod2 = self.img_mod(temb).chunk(2, dim=-1) + txt_mod1, txt_mod2 = self.txt_mod(temb).chunk(2, dim=-1) + + img_modulated, img_gate1 = self._modulate(self.img_norm1(hidden_states), img_mod1) + txt_modulated, txt_gate1 = self._modulate(self.txt_norm1(encoder_hidden_states), txt_mod1) + + img_attn, txt_attn = self.attn( + hidden_states=img_modulated, + encoder_hidden_states=txt_modulated, + freqs_cis=freqs_cis, + attention_mask=attention_mask, + transformer_options=transformer_options, + ) + + hidden_states = hidden_states + img_gate1 * img_attn + encoder_hidden_states = encoder_hidden_states + txt_gate1 * txt_attn + + img_modulated2, img_gate2 = self._modulate(self.img_norm2(hidden_states), img_mod2) + hidden_states = hidden_states + img_gate2 * self.img_mlp(img_modulated2) + + txt_modulated2, txt_gate2 = self._modulate(self.txt_norm2(encoder_hidden_states), txt_mod2) + encoder_hidden_states = encoder_hidden_states + txt_gate2 * self.txt_mlp(txt_modulated2) + + return encoder_hidden_states, hidden_states + + +class _AdaLayerNormContinuousNoAffine(nn.Module): + """AdaLayerNormContinuous(elementwise_affine=False). + + The reference uses ``scale, shift = chunk(2)`` (scale first) — opposite + to Flux's ``LastLayer``. + """ + + def __init__(self, embedding_dim: int, conditioning_embedding_dim: int, eps: float = 1e-6, + dtype=None, device=None, operations=None) -> None: + super().__init__() + self.linear = operations.Linear( + conditioning_embedding_dim, embedding_dim * 2, bias=True, dtype=dtype, device=device + ) + self.eps = eps + self.embedding_dim = embedding_dim + + def forward(self, x: torch.Tensor, conditioning: torch.Tensor) -> torch.Tensor: + emb = self.linear(F.silu(conditioning)) + scale, shift = torch.chunk(emb, 2, dim=-1) + x = F.layer_norm(x, (self.embedding_dim,), None, None, self.eps) + return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1) + + +class LensTransformer2DModel(nn.Module): + """Lens dual-stream MMDiT (48 blocks, inner_dim=1536, multi-layer text).""" + + def __init__( + self, + patch_size: int = 2, + in_channels: int = 128, + out_channels: Optional[int] = 32, + num_layers: int = 48, + attention_head_dim: int = 64, + num_attention_heads: int = 24, + enc_hidden_dim: int = 2880, + axes_dims_rope: Tuple[int, int, int] = (8, 28, 28), + rms_norm: bool = True, + multi_layer_encoder_feature: bool = True, + selected_layer_index: Tuple[int, ...] = (5, 11, 17, 23), + image_model=None, # unused; accepted for detection-side configs. + dtype=None, + device=None, + operations=None, + ) -> None: + super().__init__() + self.patch_size = patch_size + self.in_channels = in_channels + self.out_channels = out_channels if out_channels is not None else in_channels + self.inner_dim = num_attention_heads * attention_head_dim + self.multi_layer_encoder_feature = multi_layer_encoder_feature + self.selected_layer_index = list(selected_layer_index) + self.dtype = dtype + + self.pos_embed = EmbedND(dim=attention_head_dim, theta=10000, axes_dim=list(axes_dims_rope)) + self.time_text_embed = LensTimestepProjEmbeddings( + embedding_dim=self.inner_dim, dtype=dtype, device=device, operations=operations + ) + + if self.multi_layer_encoder_feature: + self.txt_norm = nn.ModuleList( + [operations.RMSNorm(enc_hidden_dim, eps=1e-5, dtype=dtype, device=device) + for _ in self.selected_layer_index] + ) + self.txt_in = operations.Linear( + enc_hidden_dim * len(self.selected_layer_index), + self.inner_dim, bias=True, dtype=dtype, device=device, + ) + else: + self.txt_norm = operations.RMSNorm(enc_hidden_dim, eps=1e-5, dtype=dtype, device=device) + self.txt_in = operations.Linear(enc_hidden_dim, self.inner_dim, bias=True, dtype=dtype, device=device) + + self.img_in = operations.Linear(in_channels, self.inner_dim, bias=True, dtype=dtype, device=device) + + self.transformer_blocks = nn.ModuleList([ + LensTransformerBlock( + dim=self.inner_dim, + num_attention_heads=num_attention_heads, + attention_head_dim=attention_head_dim, + eps=1e-6, + rms_norm=rms_norm, + dtype=dtype, device=device, operations=operations, + ) + for _ in range(num_layers) + ]) + + self.norm_out = _AdaLayerNormContinuousNoAffine( + self.inner_dim, self.inner_dim, eps=1e-6, + dtype=dtype, device=device, operations=operations, + ) + self.proj_out = operations.Linear( + self.inner_dim, patch_size * patch_size * self.out_channels, bias=True, + dtype=dtype, device=device, + ) + + def forward(self, x: torch.Tensor, timestep: torch.Tensor, context: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, + transformer_options: Optional[Dict[str, Any]] = None, **kwargs) -> torch.Tensor: + if transformer_options is None: + transformer_options = {} + return comfy.patcher_extension.WrapperExecutor.new_class_executor( + self._forward, self, + comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options), + ).execute(x, timestep, context, attention_mask, transformer_options, **kwargs) + + def _forward( + self, + x: torch.Tensor, + timestep: torch.Tensor, + context: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + transformer_options: Optional[Dict[str, Any]] = None, + control: Optional[Dict[str, Any]] = None, + **kwargs, + ) -> torch.Tensor: + """ComfyUI bridge: ``(x[B,128,h,w], t[B], context[B,S,L*H], mask[B,S])``.""" + if transformer_options is None: + transformer_options = {} + transformer_options = transformer_options.copy() + patches = transformer_options.get("patches", {}) + patches_replace = transformer_options.get("patches_replace", {}) + blocks_replace = patches_replace.get("dit", {}) + + B, C, h, w = x.shape + hidden_states = x.permute(0, 2, 3, 1).reshape(B, h * w, C) + + if self.multi_layer_encoder_feature: + L = len(self.selected_layer_index) + enc_dim = context.shape[-1] // L + encoder_hidden_states = list( + context.reshape(B, -1, L, enc_dim).unbind(dim=2) + ) + text_seq_len = encoder_hidden_states[0].shape[1] + else: + encoder_hidden_states = context + text_seq_len = context.shape[1] + + if attention_mask is None: + attention_mask = torch.ones( + (B, text_seq_len), dtype=torch.bool, device=x.device + ) + + img_len = h * w + joint_mask = self._build_joint_attention_mask(attention_mask, img_len) + + hidden_states = self.img_in(hidden_states) + timestep = timestep.to(hidden_states.dtype) + + if self.multi_layer_encoder_feature: + normed = [self.txt_norm[i](encoder_hidden_states[i]) for i in range(L)] + encoder_hidden_states = torch.cat(normed, dim=-1) + else: + encoder_hidden_states = self.txt_norm(encoder_hidden_states) + encoder_hidden_states = self.txt_in(encoder_hidden_states) + + if "post_input" in patches: + for p in patches["post_input"]: + out = p({ + "img": hidden_states, + "txt": encoder_hidden_states, + "transformer_options": transformer_options, + }) + hidden_states = out["img"] + encoder_hidden_states = out["txt"] + + temb = self.time_text_embed(timestep, hidden_states) + ids = _lens_position_ids(1, h, w, text_seq_len, device=hidden_states.device).unsqueeze(0) + freqs_cis = self.pos_embed(ids) + + transformer_options["total_blocks"] = len(self.transformer_blocks) + transformer_options["block_type"] = "double" + for i, block in enumerate(self.transformer_blocks): + transformer_options["block_index"] = i + if ("double_block", i) in blocks_replace: + def block_wrap(args): + out = {} + out["txt"], out["img"] = block( + hidden_states=args["img"], + encoder_hidden_states=args["txt"], + temb=args["vec"], + freqs_cis=args["pe"], + attention_mask=args.get("attn_mask"), + transformer_options=args.get("transformer_options"), + ) + return out + out = blocks_replace[("double_block", i)]( + { + "img": hidden_states, + "txt": encoder_hidden_states, + "vec": temb, + "pe": freqs_cis, + "attn_mask": joint_mask, + "transformer_options": transformer_options, + }, + {"original_block": block_wrap}, + ) + encoder_hidden_states = out["txt"] + hidden_states = out["img"] + else: + encoder_hidden_states, hidden_states = block( + hidden_states=hidden_states, + encoder_hidden_states=encoder_hidden_states, + temb=temb, + freqs_cis=freqs_cis, + attention_mask=joint_mask, + transformer_options=transformer_options, + ) + + if "double_block" in patches: + for p in patches["double_block"]: + out = p({ + "img": hidden_states, + "txt": encoder_hidden_states, + "x": x, + "block_index": i, + "transformer_options": transformer_options, + }) + hidden_states = out["img"] + encoder_hidden_states = out["txt"] + + if control is not None: + control_i = control.get("input") + if control_i is not None and i < len(control_i): + add = control_i[i] + if add is not None: + hidden_states[:, :add.shape[1]] += add + + hidden_states = self.norm_out(hidden_states, temb) + out = self.proj_out(hidden_states) + return out.reshape(B, h, w, C).permute(0, 3, 1, 2).contiguous() + + @staticmethod + def _build_joint_attention_mask(text_mask: torch.Tensor, img_len: int) -> torch.Tensor: + if text_mask.dtype != torch.bool: + text_mask = text_mask.bool() + bsz = text_mask.shape[0] + img_ones = torch.ones((bsz, img_len), dtype=torch.bool, device=text_mask.device) + joint = torch.cat([img_ones, text_mask], dim=1) + additive = torch.zeros_like(joint, dtype=torch.float32) + additive.masked_fill_(~joint, torch.finfo(torch.float32).min) + return additive[:, None, None, :] diff --git a/comfy/ldm/lightricks/av_model.py b/comfy/ldm/lightricks/av_model.py index bc09fb77e..ef9938465 100644 --- a/comfy/ldm/lightricks/av_model.py +++ b/comfy/ldm/lightricks/av_model.py @@ -767,25 +767,25 @@ class LTXAVModel(LTXVModel): # Cross-attention timesteps - compress these too av_ca_audio_scale_shift_timestep, _ = self.av_ca_audio_scale_shift_adaln_single( - timestep.max().expand_as(a_timestep_flat), + a_timestep_flat, {"resolution": None, "aspect_ratio": None}, batch_size=batch_size, hidden_dtype=hidden_dtype, ) av_ca_video_scale_shift_timestep, _ = self.av_ca_video_scale_shift_adaln_single( - a_timestep.max().expand_as(timestep_flat), + timestep_flat, {"resolution": None, "aspect_ratio": None}, batch_size=batch_size, hidden_dtype=hidden_dtype, ) av_ca_a2v_gate_noise_timestep, _ = self.av_ca_a2v_gate_adaln_single( - a_timestep.max().expand_as(timestep_flat) * av_ca_factor, + a_timestep_scaled.max().expand_as(timestep_flat) * av_ca_factor, {"resolution": None, "aspect_ratio": None}, batch_size=batch_size, hidden_dtype=hidden_dtype, ) av_ca_v2a_gate_noise_timestep, _ = self.av_ca_v2a_gate_adaln_single( - timestep.max().expand_as(a_timestep_flat) * av_ca_factor, + timestep_scaled.max().expand_as(a_timestep_flat) * av_ca_factor, {"resolution": None, "aspect_ratio": None}, batch_size=batch_size, hidden_dtype=hidden_dtype, diff --git a/comfy/ldm/lightricks/vae/causal_audio_autoencoder.py b/comfy/ldm/lightricks/vae/causal_audio_autoencoder.py index b556b128f..58b67d45a 100644 --- a/comfy/ldm/lightricks/vae/causal_audio_autoencoder.py +++ b/comfy/ldm/lightricks/vae/causal_audio_autoencoder.py @@ -1,4 +1,3 @@ -from __future__ import annotations import torch from torch import nn from torch.nn import functional as F diff --git a/comfy/ldm/lightricks/vae/causal_video_autoencoder.py b/comfy/ldm/lightricks/vae/causal_video_autoencoder.py index 998122c85..5975015e2 100644 --- a/comfy/ldm/lightricks/vae/causal_video_autoencoder.py +++ b/comfy/ldm/lightricks/vae/causal_video_autoencoder.py @@ -1,4 +1,3 @@ -from __future__ import annotations import threading import torch from torch import nn diff --git a/comfy/ldm/lumina/model.py b/comfy/ldm/lumina/model.py index 9e432d5c0..d0ee97d33 100644 --- a/comfy/ldm/lumina/model.py +++ b/comfy/ldm/lumina/model.py @@ -1,5 +1,4 @@ # Code from: https://github.com/Alpha-VLLM/Lumina-Image-2.0/blob/main/models/model.py -from __future__ import annotations from typing import List, Optional, Tuple diff --git a/comfy/ldm/modules/attention.py b/comfy/ldm/modules/attention.py index d4f038a63..b75a76f77 100644 --- a/comfy/ldm/modules/attention.py +++ b/comfy/ldm/modules/attention.py @@ -810,12 +810,12 @@ optimized_attention = attention_basic if model_management.sage_attention_enabled(): logging.info("Using sage attention") optimized_attention = attention_sage -elif model_management.xformers_enabled(): - logging.info("Using xformers attention") - optimized_attention = attention_xformers elif model_management.flash_attention_enabled(): logging.info("Using Flash Attention") optimized_attention = attention_flash +elif model_management.xformers_enabled(): + logging.info("Using xformers attention") + optimized_attention = attention_xformers elif model_management.pytorch_attention_enabled(): logging.info("Using pytorch attention") optimized_attention = attention_pytorch diff --git a/comfy/ldm/modules/diffusionmodules/mmdit.py b/comfy/ldm/modules/diffusionmodules/mmdit.py index 0dc8fe789..9ab3c463c 100644 --- a/comfy/ldm/modules/diffusionmodules/mmdit.py +++ b/comfy/ldm/modules/diffusionmodules/mmdit.py @@ -211,7 +211,7 @@ class TimestepEmbedder(nn.Module): Embeds scalar timesteps into vector representations. """ - def __init__(self, hidden_size, frequency_embedding_size=256, output_size=None, dtype=None, device=None, operations=None): + def __init__(self, hidden_size, frequency_embedding_size=256, output_size=None, dtype=None, device=None, operations=None, max_period=10000): super().__init__() if output_size is None: output_size = hidden_size @@ -221,9 +221,10 @@ class TimestepEmbedder(nn.Module): operations.Linear(hidden_size, output_size, bias=True, dtype=dtype, device=device), ) self.frequency_embedding_size = frequency_embedding_size + self.max_period = max_period def forward(self, t, dtype, **kwargs): - t_freq = timestep_embedding(t, self.frequency_embedding_size).to(dtype) + t_freq = timestep_embedding(t, self.frequency_embedding_size, max_period=self.max_period).to(dtype) t_emb = self.mlp(t_freq) return t_emb diff --git a/comfy/ldm/moge/geometry.py b/comfy/ldm/moge/geometry.py index 7fdc97871..d1a1e445f 100644 --- a/comfy/ldm/moge/geometry.py +++ b/comfy/ldm/moge/geometry.py @@ -1,6 +1,5 @@ """Pure-torch + scipy geometry helpers for MoGe inference and mesh export.""" -from __future__ import annotations from typing import Optional, Tuple diff --git a/comfy/ldm/moge/model.py b/comfy/ldm/moge/model.py index 6876c4af2..1695626bc 100644 --- a/comfy/ldm/moge/model.py +++ b/comfy/ldm/moge/model.py @@ -4,7 +4,6 @@ V1: DINOv2 backbone + multi-output head (points, mask). V2: DINOv2 encoder + neck + per-output heads (points, mask, normal, optional metric-scale MLP). """ -from __future__ import annotations from numbers import Number from typing import Any, Dict, List, Optional, Tuple, Union diff --git a/comfy/ldm/moge/modules.py b/comfy/ldm/moge/modules.py index 235a59212..f6443d65a 100644 --- a/comfy/ldm/moge/modules.py +++ b/comfy/ldm/moge/modules.py @@ -1,6 +1,5 @@ """Building blocks for MoGe: residual conv stack, resamplers, MLP, DINOv2 encoder, v1 head.""" -from __future__ import annotations from typing import List, Optional, Sequence, Tuple, Union diff --git a/comfy/ldm/moge/panorama.py b/comfy/ldm/moge/panorama.py index de53ebe68..18d0cb665 100644 --- a/comfy/ldm/moge/panorama.py +++ b/comfy/ldm/moge/panorama.py @@ -6,7 +6,6 @@ equirect distance map via a multi-scale Poisson + gradient sparse solve. Image sampling uses F.grid_sample (GPU); the sparse solve uses lsmr (CPU). """ -from __future__ import annotations from typing import Callable, List, Optional, Tuple diff --git a/comfy/ldm/pixeldit/model.py b/comfy/ldm/pixeldit/model.py new file mode 100644 index 000000000..b044b9b29 --- /dev/null +++ b/comfy/ldm/pixeldit/model.py @@ -0,0 +1,239 @@ +import torch +import torch.nn as nn +import torch.nn.functional as F + +import comfy.ldm.common_dit +import comfy.patcher_extension +from comfy.ldm.flux.math import apply_rope, rope +from comfy.ldm.hidream.model import FeedForwardSwiGLU +from comfy.ldm.modules.attention import optimized_attention +from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder + +from .modules import ( + FinalLayer, + PatchTokenEmbedder, + PiTBlock, + PixelTokenEmbedder, + apply_adaln_, + precompute_freqs_cis_2d, +) + + +class MMDiTJointAttention(nn.Module): + """Joint MMDiT attention with separate Q/K/V/proj for image and text streams. + + RoPE is applied to each stream before concatenation so each stream uses its own + 2D/1D positional encoding. Concat order is [text, image] (text first). + """ + def __init__(self, dim, num_heads=8, qkv_bias=False, dtype=None, device=None, operations=None): + super().__init__() + assert dim % num_heads == 0 + self.num_heads = num_heads + self.head_dim = dim // num_heads + + self.qkv_x = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device) + self.qkv_y = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device) + + self.q_norm_x = operations.RMSNorm(self.head_dim, eps=1e-6, dtype=dtype, device=device) + self.k_norm_x = operations.RMSNorm(self.head_dim, eps=1e-6, dtype=dtype, device=device) + self.q_norm_y = operations.RMSNorm(self.head_dim, eps=1e-6, dtype=dtype, device=device) + self.k_norm_y = operations.RMSNorm(self.head_dim, eps=1e-6, dtype=dtype, device=device) + + self.proj_x = operations.Linear(dim, dim, dtype=dtype, device=device) + self.proj_y = operations.Linear(dim, dim, dtype=dtype, device=device) + + def forward(self, x, y, pos_img, pos_txt=None, attn_mask=None, transformer_options={}): + B, Nx, _ = x.shape + _, Ny, _ = y.shape + H = self.num_heads + D = self.head_dim + + qkv_x = self.qkv_x(x).reshape(B, Nx, 3, H, D).permute(2, 0, 3, 1, 4) + qx, kx, vx = qkv_x.unbind(0) + qx = self.q_norm_x(qx) + kx = self.k_norm_x(kx) + + qkv_y = self.qkv_y(y).reshape(B, Ny, 3, H, D).permute(2, 0, 3, 1, 4) + qy, ky, vy = qkv_y.unbind(0) + qy = self.q_norm_y(qy) + ky = self.k_norm_y(ky) + + qx, kx = apply_rope(qx, kx, pos_img[None, None]) + if pos_txt is not None: + qy, ky = apply_rope(qy, ky, pos_txt[None, None]) + + q_joint = torch.cat([qy, qx], dim=2) + k_joint = torch.cat([ky, kx], dim=2) + v_joint = torch.cat([vy, vx], dim=2) + + out_joint = optimized_attention( + q_joint, k_joint, v_joint, H, + mask=attn_mask, skip_reshape=True, skip_output_reshape=True, + transformer_options=transformer_options, + ) + + out_y = out_joint[:, :, :Ny, :].transpose(1, 2).reshape(B, Ny, H * D) + out_x = out_joint[:, :, Ny:, :].transpose(1, 2).reshape(B, Nx, H * D) + + return self.proj_x(out_x), self.proj_y(out_y) + + +class MMDiTBlockT2I(nn.Module): + def __init__(self, hidden_size, groups, mlp_ratio=4.0, dtype=None, device=None, operations=None): + super().__init__() + self.norm_x1 = operations.RMSNorm(hidden_size, eps=1e-6, dtype=dtype, device=device) + self.norm_y1 = operations.RMSNorm(hidden_size, eps=1e-6, dtype=dtype, device=device) + self.attn = MMDiTJointAttention(hidden_size, num_heads=groups, qkv_bias=False, dtype=dtype, device=device, operations=operations) + self.norm_x2 = operations.RMSNorm(hidden_size, eps=1e-6, dtype=dtype, device=device) + self.norm_y2 = operations.RMSNorm(hidden_size, eps=1e-6, dtype=dtype, device=device) + mlp_hidden_dim = int(hidden_size * mlp_ratio) + self.mlp_x = FeedForwardSwiGLU(hidden_size, mlp_hidden_dim, multiple_of=1, dtype=dtype, device=device, operations=operations) + self.mlp_y = FeedForwardSwiGLU(hidden_size, mlp_hidden_dim, multiple_of=1, dtype=dtype, device=device, operations=operations) + self.adaLN_modulation_img = nn.Sequential(operations.Linear(hidden_size, 6 * hidden_size, bias=True, dtype=dtype, device=device)) + self.adaLN_modulation_txt = nn.Sequential(operations.Linear(hidden_size, 6 * hidden_size, bias=True, dtype=dtype, device=device)) + + def forward(self, x, y, c, pos_img, pos_txt=None, attn_mask=None, transformer_options={}): + shift_msa_x, scale_msa_x, gate_msa_x, shift_mlp_x, scale_mlp_x, gate_mlp_x = self.adaLN_modulation_img(c).chunk(6, dim=-1) + shift_msa_y, scale_msa_y, gate_msa_y, shift_mlp_y, scale_mlp_y, gate_mlp_y = self.adaLN_modulation_txt(c).chunk(6, dim=-1) + + x_norm = apply_adaln_(self.norm_x1(x), shift_msa_x, scale_msa_x) + y_norm = apply_adaln_(self.norm_y1(y), shift_msa_y, scale_msa_y) + attn_x, attn_y = self.attn(x_norm, y_norm, pos_img, pos_txt, attn_mask, transformer_options=transformer_options) + x = torch.addcmul(x, gate_msa_x, attn_x) + y = torch.addcmul(y, gate_msa_y, attn_y) + + x = torch.addcmul(x, gate_mlp_x, self.mlp_x(apply_adaln_(self.norm_x2(x), shift_mlp_x, scale_mlp_x))) + y = torch.addcmul(y, gate_mlp_y, self.mlp_y(apply_adaln_(self.norm_y2(y), shift_mlp_y, scale_mlp_y))) + return x, y + + +class PixDiT_T2I(nn.Module): + """PixelDiT T2I model. Hardcoded for the released 1024px Stage-3 checkpoint + (also runs at 512px when fed the appropriate latent size and flow_shift). + + Forward: + x: [B, 3, H, W] pixel-space input (no VAE) + timesteps:[B] in [0, 1000] (ComfyUI flow sampling convention) + context: [B, Ltxt, 2304] Gemma-2-2b-it hidden states (chi_prompt prepended) + Returns flow-matching velocity [B, 3, H, W]. + """ + def __init__( + self, + in_channels=3, + num_groups=24, + hidden_size=1536, + pixel_hidden_size=16, + pixel_attn_hidden_size=1152, + pixel_num_groups=16, + patch_depth=14, + pixel_depth=2, + patch_size=16, + txt_embed_dim=2304, + txt_max_length=300, + use_text_rope=True, + text_rope_theta=10000.0, + image_model=None, + dtype=None, + device=None, + operations=None, + pixel_mlp_chunks=2, + ): + super().__init__() + self.dtype = dtype + self.in_channels = in_channels + self.out_channels = in_channels + self.hidden_size = hidden_size + self.num_groups = num_groups + self.patch_depth = patch_depth + self.pixel_depth = pixel_depth + self.patch_size = patch_size + self.pixel_hidden_size = pixel_hidden_size + self.pixel_attn_hidden_size = pixel_attn_hidden_size + self.pixel_num_groups = pixel_num_groups + self.txt_embed_dim = txt_embed_dim + self.txt_max_length = txt_max_length + self.use_text_rope = use_text_rope + self.text_rope_theta = text_rope_theta + + self.pixel_embedder = PixelTokenEmbedder(self.in_channels, self.pixel_hidden_size, dtype=dtype, device=device, operations=operations) + self.s_embedder = PatchTokenEmbedder(self.in_channels * self.patch_size ** 2, self.hidden_size, bias=True, dtype=dtype, device=device, operations=operations) + self.t_embedder = TimestepEmbedder(self.hidden_size, dtype=dtype, device=device, operations=operations, max_period=10) + self.y_embedder = PatchTokenEmbedder(self.txt_embed_dim, self.hidden_size, bias=True, use_norm=True, dtype=dtype, device=device, operations=operations) + self.y_pos_embedding = nn.Parameter(torch.empty(1, self.txt_max_length, self.hidden_size, dtype=dtype, device=device)) + + self.patch_blocks = nn.ModuleList([ + MMDiTBlockT2I(self.hidden_size, self.num_groups, + dtype=dtype, device=device, operations=operations) + for _ in range(self.patch_depth) + ]) + self.pixel_blocks = nn.ModuleList([ + PiTBlock( + self.pixel_hidden_size, + self.hidden_size, + patch_size=self.patch_size, + num_heads=self.num_groups, + attn_hidden_size=self.pixel_attn_hidden_size, + attn_num_heads=self.pixel_num_groups, + dtype=dtype, device=device, operations=operations, + mlp_chunks=pixel_mlp_chunks, + ) + for _ in range(self.pixel_depth) + ]) + + self.final_layer = FinalLayer(self.pixel_hidden_size, self.out_channels, dtype=dtype, device=device, operations=operations) + + def _fetch_patch_pos(self, height, width, device, dtype, **rope_opts): + return precompute_freqs_cis_2d(self.hidden_size // self.num_groups, height, width, device=device, dtype=dtype, **rope_opts) + + def _fetch_text_pos(self, length, device, dtype): + return rope(torch.arange(length, dtype=torch.float32, device=device).reshape(1, -1), self.hidden_size // self.num_groups, self.text_rope_theta).squeeze(0).to(dtype=dtype) + + def forward(self, x, timesteps, context=None, attention_mask=None, transformer_options={}, **kwargs): + return comfy.patcher_extension.WrapperExecutor.new_class_executor( + self._forward, self, comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options), + ).execute(x, timesteps, context, attention_mask, transformer_options, **kwargs) + + def _pre_patch_block(self, s, i, **kwargs): + """Hook for subclasses to inject per-block state into the patch stream (e.g. PiD's LQ gate).""" + return s + + def _forward(self, x, timesteps, context=None, attention_mask=None, transformer_options={}, **kwargs): + H_orig, W_orig = x.shape[2], x.shape[3] + x = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size)) + B, _, H, W = x.shape + Hs = H // self.patch_size + Ws = W // self.patch_size + L = Hs * Ws + + pos_img = self._fetch_patch_pos(Hs, Ws, x.device, x.dtype, **(transformer_options.get("rope_options") or {})) + x_patches = F.unfold(x, kernel_size=self.patch_size, stride=self.patch_size).transpose(1, 2) + + t_emb = self.t_embedder(timesteps.view(-1), x.dtype).view(B, -1, self.hidden_size) + + if context is None or context.dim() != 3: + raise ValueError("PixDiT_T2I requires context (text embeddings) of shape [B, L, D]") + Ltxt = min(context.shape[1], self.txt_max_length) + y = context[:, :Ltxt, :] + y_emb = self.y_embedder(y).view(B, Ltxt, self.hidden_size) + y_emb = y_emb + self.y_pos_embedding[:, :Ltxt, :].to(y_emb) # y_pos_embedding is a raw nn.Parameter + + condition = F.silu(t_emb) + pos_txt = self._fetch_text_pos(Ltxt, x.device, x.dtype) if self.use_text_rope else None + + s = self.s_embedder(x_patches) + for i, blk in enumerate(self.patch_blocks): + s = self._pre_patch_block(s, i, **kwargs) + s, y_emb = blk(s, y_emb, condition, pos_img, pos_txt, None, transformer_options=transformer_options) + s = F.silu(t_emb + s) + + s_cond = s.view(B * L, self.hidden_size) + x_pixels = self.pixel_embedder(x, patch_size=self.patch_size) + for blk in self.pixel_blocks: + x_pixels = blk(x_pixels, s_cond, H, W, self.patch_size, mask=None, transformer_options=transformer_options) + + x_pixels = self.final_layer(x_pixels) + C_out = self.out_channels + P2 = self.patch_size * self.patch_size + x_pixels = x_pixels.view(B, L, P2, C_out).permute(0, 3, 2, 1).reshape(B, C_out * P2, L) + out = F.fold(x_pixels, (H, W), kernel_size=self.patch_size, stride=self.patch_size) + return out[:, :, :H_orig, :W_orig] diff --git a/comfy/ldm/pixeldit/modules.py b/comfy/ldm/pixeldit/modules.py new file mode 100644 index 000000000..4b1e538c7 --- /dev/null +++ b/comfy/ldm/pixeldit/modules.py @@ -0,0 +1,187 @@ +import torch +import torch.nn as nn + +from comfy.ldm.flux.math import apply_rope, rope +from comfy.ldm.modules.attention import optimized_attention +from comfy.ldm.modules.diffusionmodules.mmdit import Mlp, get_1d_sincos_pos_embed_from_grid_torch + + +def apply_adaln_(x, shift, scale): + return x.addcmul_(x, scale).add_(shift) + + +def precompute_freqs_cis_2d(dim, height, width, theta=10000.0, scale=16.0, + ref_grid_h=None, ref_grid_w=None, + scale_x=1.0, scale_y=1.0, shift_x=0.0, shift_y=0.0, + device=None, dtype=torch.float32, **kwargs): + """2D RoPE with x/y axis frequencies interleaved at stride 2 across head dim. + + rope_options: + scale_x / scale_y multiply the position range (RoPE extrapolation). + shift_x / shift_y offset the position origin (tiled / regional inference). + With ref_grid_h/w set, also applies NTK-aware per-axis theta scaling + (rope_mode='ntk_aware'): theta_axis = theta * (current/ref)^(dim_axis/(dim_axis-2)). + Returns Flux-format rotation matrices of shape [H*W, dim/2, 2, 2]. + Layout of head-dim pairs: [x_0, y_0, x_1, y_1, ..., x_{dim/4-1}, y_{dim/4-1}]. + """ + dim_axis = dim // 2 + if ref_grid_h is not None and dim_axis > 2: + h_ntk = (height / ref_grid_h) ** (dim_axis / (dim_axis - 2)) + w_ntk = (width / ref_grid_w) ** (dim_axis / (dim_axis - 2)) + else: + h_ntk = w_ntk = 1.0 + + x_lin = torch.linspace(shift_x, scale * scale_x + shift_x, width, device=device) + y_lin = torch.linspace(shift_y, scale * scale_y + shift_y, height, device=device) + y_grid, x_grid = torch.meshgrid(y_lin, x_lin, indexing="ij") + x_rope = rope(x_grid.reshape(1, -1), dim_axis, theta * w_ntk).squeeze(0) + y_rope = rope(y_grid.reshape(1, -1), dim_axis, theta * h_ntk).squeeze(0) + out = torch.stack([x_rope, y_rope], dim=2).reshape(height * width, dim // 2, 2, 2) + return out.to(dtype=dtype) + + +def get_2d_sincos_pos_embed(embed_dim, height, width, device=None, dtype=torch.float32): + """Standard 2D sin/cos absolute positional embedding (ViT-style). + + first half encodes W-coordinates, second half H. + """ + assert embed_dim % 4 == 0 + grid_h = torch.arange(height, dtype=torch.float32, device=device) + grid_w = torch.arange(width, dtype=torch.float32, device=device) + grid_y, grid_x = torch.meshgrid(grid_h, grid_w, indexing="ij") + emb_w = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_x.reshape(-1), device=device) + emb_h = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_y.reshape(-1), device=device) + return torch.cat([emb_w, emb_h], dim=1).to(dtype=dtype) + + +class RotaryAttention(nn.Module): + """Single-stream self-attention with rotary positional encoding (used inside PiTBlock).""" + def __init__(self, dim, num_heads=8, qkv_bias=False, dtype=None, device=None, operations=None): + super().__init__() + assert dim % num_heads == 0 + self.num_heads = num_heads + self.head_dim = dim // num_heads + self.qkv = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device) + self.q_norm = operations.RMSNorm(self.head_dim, eps=1e-6, dtype=dtype, device=device) + self.k_norm = operations.RMSNorm(self.head_dim, eps=1e-6, dtype=dtype, device=device) + self.proj = operations.Linear(dim, dim, dtype=dtype, device=device) + + def forward(self, x, pos, mask=None, transformer_options={}): + B, N, C = x.shape + H = self.num_heads + D = self.head_dim + qkv = self.qkv(x).reshape(B, N, 3, H, D).permute(2, 0, 3, 1, 4) + q, k, v = qkv.unbind(0) + q, k = apply_rope(self.q_norm(q), self.k_norm(k), pos[None, None]) + x = optimized_attention(q, k, v, H, mask=mask, skip_reshape=True, transformer_options=transformer_options) + return self.proj(x) + + +class FinalLayer(nn.Module): + def __init__(self, hidden_size, out_channels, dtype=None, device=None, operations=None): + super().__init__() + self.norm = operations.RMSNorm(hidden_size, eps=1e-6, dtype=dtype, device=device) + self.linear = operations.Linear(hidden_size, out_channels, bias=True, dtype=dtype, device=device) + + def forward(self, x): + return self.linear(self.norm(x)) + + +class PatchTokenEmbedder(nn.Module): + """Linear projection used both for patchified-image tokens and text-feature tokens.""" + def __init__(self, in_chans, embed_dim, use_norm=False, bias=True, dtype=None, device=None, operations=None): + super().__init__() + self.proj = operations.Linear(in_chans, embed_dim, bias=bias, dtype=dtype, device=device) + self.norm = operations.RMSNorm(embed_dim, eps=1e-6, dtype=dtype, device=device) if use_norm else nn.Identity() + + def forward(self, x): + return self.norm(self.proj(x)) + + +class PixelTokenEmbedder(nn.Module): + """Pixel-level embedder: lifts each RGB pixel to hidden_size and packs into per-patch sequences.""" + def __init__(self, in_channels, hidden_size_output, dtype=None, device=None, operations=None): + super().__init__() + self.in_channels = in_channels + self.hidden_size_output = hidden_size_output + self.proj = operations.Linear(self.in_channels, self.hidden_size_output, bias=True, dtype=dtype, device=device) + + def forward(self, inputs, patch_size): + B, _, H, W = inputs.shape + Hs, Ws = H // patch_size, W // patch_size + P2 = patch_size * patch_size + x = inputs.permute(0, 2, 3, 1).contiguous() + x = self.proj(x) + pos_full = get_2d_sincos_pos_embed(self.hidden_size_output, H, W, device=x.device, dtype=x.dtype).view(H, W, self.hidden_size_output) + x = x + pos_full.unsqueeze(0) + x = x.view(B, Hs, patch_size, Ws, patch_size, self.hidden_size_output) + return x.permute(0, 1, 3, 2, 4, 5).reshape(B * Hs * Ws, P2, self.hidden_size_output) + + +class PiTBlock(nn.Module): + """Pixel-level transformer block. + + Compresses each patch's P^2 pixel tokens → 1 attention token via a linear, + runs global self-attention across patches with 2D RoPE, then expands back to P^2 tokens. + Conditioning is per-pixel adaLN from the patch-level features. + """ + def __init__(self, pixel_hidden_size, patch_hidden_size, patch_size, num_heads, mlp_ratio=4.0, + attn_hidden_size=None, attn_num_heads=None, dtype=None, device=None, operations=None, mlp_chunks=1): + super().__init__() + self.pixel_dim = pixel_hidden_size + self.context_dim = patch_hidden_size + self.attn_dim = attn_hidden_size if attn_hidden_size is not None else patch_hidden_size + self.num_heads = attn_num_heads if attn_num_heads is not None else num_heads + assert self.attn_dim % self.num_heads == 0 + + p2 = patch_size * patch_size + self.compress_to_attn = operations.Linear(p2 * self.pixel_dim, self.attn_dim, bias=True, dtype=dtype, device=device) + self.expand_from_attn = operations.Linear(self.attn_dim, p2 * self.pixel_dim, bias=True, dtype=dtype, device=device) + + self.norm1 = operations.RMSNorm(self.pixel_dim, eps=1e-6, dtype=dtype, device=device) + self.attn = RotaryAttention(self.attn_dim, num_heads=self.num_heads, qkv_bias=False, dtype=dtype, device=device, operations=operations) + self.norm2 = operations.RMSNorm(self.pixel_dim, eps=1e-6, dtype=dtype, device=device) + self.mlp = Mlp(self.pixel_dim, hidden_features=int(self.pixel_dim * mlp_ratio), dtype=dtype, device=device, operations=operations) + + self.adaLN_modulation_msa = operations.Linear(self.context_dim, 3 * self.pixel_dim * p2, bias=True, dtype=dtype, device=device) + self.adaLN_modulation_mlp = operations.Linear(self.context_dim, 3 * self.pixel_dim * p2, bias=True, dtype=dtype, device=device) + + self._rope_fn = precompute_freqs_cis_2d + self.mlp_chunks = max(1, int(mlp_chunks)) + + def _fetch_pos(self, height, width, device, dtype, **rope_opts): + return self._rope_fn(self.attn_dim // self.num_heads, height, width, device=device, dtype=dtype, **rope_opts) + + def forward(self, x, s_cond, image_height, image_width, patch_size, mask=None, transformer_options={}): + BL, P2, _ = x.shape + Hs, Ws = image_height // patch_size, image_width // patch_size + L = Hs * Ws + B = BL // L + + # Attention path uses only msa params; compute, use, free before mlp params allocate. + msa_params = self.adaLN_modulation_msa(s_cond).view(BL, P2, 3 * self.pixel_dim) + shift_msa, scale_msa, gate_msa = msa_params.chunk(3, dim=-1) + + x_norm = apply_adaln_(self.norm1(x), shift_msa, scale_msa) + x_flat = x_norm.view(BL, P2 * self.pixel_dim) + + x_comp = self.compress_to_attn(x_flat).view(B, L, self.attn_dim) + pos_comp = self._fetch_pos(Hs, Ws, x.device, x.dtype, **(transformer_options.get("rope_options") or {})) + attn_out = self.attn(x_comp, pos_comp, mask=mask, transformer_options=transformer_options) + attn_flat = self.expand_from_attn(attn_out.view(B * L, self.attn_dim)) + attn_exp = attn_flat.view(BL, P2, self.pixel_dim) + x = torch.addcmul(x, gate_msa, attn_exp) + del msa_params, shift_msa, scale_msa, gate_msa + + mlp_params = self.adaLN_modulation_mlp(s_cond).view(BL, P2, 3 * self.pixel_dim) + shift_mlp, scale_mlp, gate_mlp = mlp_params.chunk(3, dim=-1) + gate_mlp = gate_mlp.contiguous() # detach from mlp_params so the del below frees shift+scale storage before the MLP + mlp_input = apply_adaln_(self.norm2(x), shift_mlp, scale_mlp) + del mlp_params, shift_mlp, scale_mlp + + # MLP in chunks since the peak memory usage is huge here + chunk_size = (BL + self.mlp_chunks - 1) // self.mlp_chunks + for s in range(0, BL, chunk_size): + e = min(s + chunk_size, BL) + x[s:e].addcmul_(gate_mlp[s:e], self.mlp(mlp_input[s:e])) + return x diff --git a/comfy/ldm/pixeldit/pid.py b/comfy/ldm/pixeldit/pid.py new file mode 100644 index 000000000..21b73907a --- /dev/null +++ b/comfy/ldm/pixeldit/pid.py @@ -0,0 +1,227 @@ +"""PiD — Pixel Diffusion Decoder. Decodes a Flux/SD3/Flux2/Z-Image latent +directly to a 4x-upscaled image in 4 distilled flow-matching steps. PixDiT_T2I +body + LQ projection branch injected before each MMDiT patch block. +""" + +from typing import List + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from .model import PixDiT_T2I +from .modules import precompute_freqs_cis_2d + + +class SigmaAwareGatePerTokenPerDim(nn.Module): + """gate = sigmoid(content_proj(cat[x, lq]) - exp(log_alpha) * sigma); out = x + gate * lq. + + Trained init gives ~0.88 gate at sigma=0, ~0.05 at sigma=1. + """ + + def __init__(self, dim: int, dtype=None, device=None, operations=None): + super().__init__() + self.content_proj = operations.Linear(dim * 2, dim, dtype=dtype, device=device) + self.log_alpha = nn.Parameter(torch.empty((), dtype=dtype, device=device)) + + def forward(self, x: torch.Tensor, lq: torch.Tensor, sigma: torch.Tensor) -> torch.Tensor: + content_logit = self.content_proj(torch.cat([x, lq], dim=-1)) + # log_alpha is a raw nn.Parameter -> doesn't auto-cast under dynamic VRAM. + log_alpha = self.log_alpha.to(device=x.device, dtype=torch.float32) + sigma_offset = -log_alpha.exp() * sigma.float().view(-1, 1, 1) + gate = torch.sigmoid(content_logit + sigma_offset) + return x + (gate * lq).to(x.dtype) + + +class ResBlock(nn.Module): + """Pre-activation ResNet block: GN -> SiLU -> Conv -> GN -> SiLU -> Conv + skip.""" + + def __init__(self, channels: int, num_groups: int = 4, dtype=None, device=None, operations=None): + super().__init__() + self.block = nn.Sequential( + operations.GroupNorm(num_groups, channels, dtype=dtype, device=device), + nn.SiLU(), + operations.Conv2d(channels, channels, kernel_size=3, padding=1, dtype=dtype, device=device), + operations.GroupNorm(num_groups, channels, dtype=dtype, device=device), + nn.SiLU(), + operations.Conv2d(channels, channels, kernel_size=3, padding=1, dtype=dtype, device=device), + ) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + return x + self.block(x) + + +class LQProjection2D(nn.Module): + """LQ latent -> per-block patch-aligned features for controlnet-style injection.""" + + def __init__( + self, + latent_channels: int, + hidden_dim: int = 512, + out_dim: int = 1536, + patch_size: int = 16, + sr_scale: int = 4, + latent_spatial_down_factor: int = 8, + num_res_blocks: int = 4, + num_outputs: int = 7, + interval: int = 2, + dtype=None, device=None, operations=None, + ): + super().__init__() + self.latent_channels = latent_channels + self.hidden_dim = hidden_dim + self.out_dim = out_dim + self.patch_size = patch_size + self.sr_scale = sr_scale + self.latent_spatial_down_factor = latent_spatial_down_factor + self.num_outputs = num_outputs + self.interval = interval + + z_to_patch_ratio = (sr_scale * latent_spatial_down_factor) / patch_size + self.z_to_patch_ratio = z_to_patch_ratio + if z_to_patch_ratio >= 1: + self.latent_fold_factor = 0 + latent_proj_in_ch = latent_channels + else: + fold_factor = int(1 / z_to_patch_ratio) + assert fold_factor * z_to_patch_ratio == 1.0 + self.latent_fold_factor = fold_factor + latent_proj_in_ch = latent_channels * fold_factor * fold_factor + + layers = [ + operations.Conv2d(latent_proj_in_ch, hidden_dim, kernel_size=3, padding=1, dtype=dtype, device=device), + nn.SiLU(), + operations.Conv2d(hidden_dim, hidden_dim, kernel_size=3, padding=1, dtype=dtype, device=device), + ] + for _ in range(num_res_blocks): + layers.append(ResBlock(hidden_dim, dtype=dtype, device=device, operations=operations)) + self.latent_proj = nn.Sequential(*layers) + + self.output_heads = nn.ModuleList( + [operations.Linear(hidden_dim, out_dim, dtype=dtype, device=device) for _ in range(num_outputs)] + ) + self.gate_modules = nn.ModuleList( + [SigmaAwareGatePerTokenPerDim(out_dim, dtype=dtype, device=device, operations=operations) + for _ in range(num_outputs)] + ) + + def is_gate_active(self, block_idx: int) -> bool: + return block_idx % self.interval == 0 + + def output_index(self, block_idx: int) -> int: + return block_idx // self.interval + + def gate(self, x: torch.Tensor, lq_feature: torch.Tensor, sigma: torch.Tensor, out_idx: int) -> torch.Tensor: + return self.gate_modules[out_idx](x, lq_feature, sigma) + + def _align_latent_to_patch_grid(self, lq_latent: torch.Tensor, pH: int, pW: int) -> torch.Tensor: + B, z_dim = lq_latent.shape[:2] + if self.z_to_patch_ratio >= 1: + if lq_latent.shape[2] != pH or lq_latent.shape[3] != pW: + z_aligned = F.interpolate(lq_latent, size=(pH, pW), mode="nearest") + else: + z_aligned = lq_latent + else: + f = self.latent_fold_factor + zH_expected, zW_expected = pH * f, pW * f + if lq_latent.shape[2] != zH_expected or lq_latent.shape[3] != zW_expected: + lq_latent = F.interpolate(lq_latent, size=(zH_expected, zW_expected), mode="nearest") + z_aligned = lq_latent.reshape(B, z_dim, pH, f, pW, f).permute(0, 1, 3, 5, 2, 4) + z_aligned = z_aligned.reshape(B, z_dim * f * f, pH, pW) + return self.latent_proj(z_aligned) + + def forward(self, lq_latent: torch.Tensor, target_pH: int, target_pW: int) -> List[torch.Tensor]: + feat = self._align_latent_to_patch_grid(lq_latent, target_pH, target_pW) + B, C, H, W = feat.shape + tokens = feat.permute(0, 2, 3, 1).contiguous().view(B, H * W, C) + return [head(tokens) for head in self.output_heads] + + +class PidNet(PixDiT_T2I): + """PixDiT_T2I + LQ injection (one sigma-gated feature inserted before each patch block).""" + + def __init__( + self, + lq_latent_channels: int = 16, + lq_hidden_dim: int = 512, + lq_num_res_blocks: int = 4, + lq_interval: int = 2, + sr_scale: int = 4, + latent_spatial_down_factor: int = 8, + rope_ref_h: int = 1024, # NTK ref resolution in PIXEL units: 1024px / patch=16 -> grid_ref=64. + rope_ref_w: int = 1024, + image_model=None, + dtype=None, device=None, operations=None, + **pixdit_kwargs, + ): + super().__init__(dtype=dtype, device=device, operations=operations, **pixdit_kwargs) + + self.rope_ref_grid_h = rope_ref_h // self.patch_size + self.rope_ref_grid_w = rope_ref_w // self.patch_size + + # Parent's PiTBlocks were built with plain RoPE — swap in NTK-aware. + def _pit_rope_fn(head_dim, h, w, device=None, dtype=torch.float32, **rope_opts): + return precompute_freqs_cis_2d(head_dim, h, w, ref_grid_h=self.rope_ref_grid_h, ref_grid_w=self.rope_ref_grid_w, device=device, dtype=dtype, **rope_opts) + for blk in self.pixel_blocks: + blk._rope_fn = _pit_rope_fn + + num_lq_outputs = (self.patch_depth + lq_interval - 1) // lq_interval + self.lq_proj = LQProjection2D( + latent_channels=lq_latent_channels, + hidden_dim=lq_hidden_dim, + out_dim=self.hidden_size, + patch_size=self.patch_size, + sr_scale=sr_scale, + latent_spatial_down_factor=latent_spatial_down_factor, + num_res_blocks=lq_num_res_blocks, + num_outputs=num_lq_outputs, + interval=lq_interval, + dtype=dtype, + device=device, + operations=operations, + ) + + def _fetch_patch_pos(self, height, width, device, dtype, **rope_opts): + return precompute_freqs_cis_2d( + self.hidden_size // self.num_groups, + height, width, + ref_grid_h=self.rope_ref_grid_h, ref_grid_w=self.rope_ref_grid_w, + device=device, dtype=dtype, **rope_opts, + ) + + def _pre_patch_block(self, s, i, pid_lq_features, pid_degrade_sigma, **kwargs): + if not self.lq_proj.is_gate_active(i): + return s + out_idx = self.lq_proj.output_index(i) + if out_idx >= len(pid_lq_features): + return s + return self.lq_proj.gate(s, pid_lq_features[out_idx], pid_degrade_sigma, out_idx) + + def _forward(self, x, timesteps, context=None, attention_mask=None, transformer_options={}, lq_latent=None, degrade_sigma=None, **kwargs): + if lq_latent is None: + raise ValueError("PidNet requires lq_latent — attach via PiDConditioning") + expected_c = self.lq_proj.latent_channels + if lq_latent.shape[1] != expected_c: + raise ValueError( + f"Input latent has {lq_latent.shape[1]} channels, this model variant expects {expected_c}. " + f"Flux1/SD3 = 16 channels, Flux2 = 128 channels." + ) + B = x.shape[0] + # Match the backbone's pad_to_patch_size (round up) so the LQ grid lines up with the patch stream. + Hs = -(-x.shape[2] // self.patch_size) + Ws = -(-x.shape[3] // self.patch_size) + + degrade_sigma = degrade_sigma.to(device=x.device, dtype=torch.float32).reshape(-1) + if degrade_sigma.numel() == 1 and B > 1: + degrade_sigma = degrade_sigma.expand(B).contiguous() + + lq_features = self.lq_proj(lq_latent=lq_latent.to(x), target_pH=Hs, target_pW=Ws) + + return super()._forward( + x, timesteps, + context=context, attention_mask=attention_mask, + transformer_options=transformer_options, + pid_lq_features=lq_features, + pid_degrade_sigma=degrade_sigma, + **kwargs, + ) diff --git a/comfy/ldm/qwen_image/model.py b/comfy/ldm/qwen_image/model.py index 0862f72f7..e49886dd9 100644 --- a/comfy/ldm/qwen_image/model.py +++ b/comfy/ldm/qwen_image/model.py @@ -51,6 +51,9 @@ class FeedForward(nn.Module): return hidden_states +# Addin this back because Nunchaku custom nodes rely on it, see comment here: +# https://github.com/Comfy-Org/ComfyUI/pull/14178#issuecomment-4640475161 +# TODO: Eventually remove this once we natively support SVDQuants def apply_rotary_emb(x, freqs_cis): if x.shape[1] == 0: return x diff --git a/comfy/ldm/triposplat/gaussian.py b/comfy/ldm/triposplat/gaussian.py new file mode 100644 index 000000000..a4cd2f62f --- /dev/null +++ b/comfy/ldm/triposplat/gaussian.py @@ -0,0 +1,199 @@ +# TripoSplat 3D gaussian container. Operates on already-decoded +# tensors and exposes them as render-ready tensors (render_tensors) for the generic SPLAT type. +import torch +import torch.nn.functional as F + +import comfy.model_management + + +class GaussianModel: + def __init__(self, aabb: list, sh_degree: int = 0, mininum_kernel_size: float = 0.0, + scaling_bias: float = 0.01, opacity_bias: float = 0.1, + scaling_activation: str = "exp", device=None): + self.sh_degree = sh_degree + self.mininum_kernel_size = mininum_kernel_size + self.scaling_bias = scaling_bias + self.opacity_bias = opacity_bias + self.device = device + self.aabb = torch.tensor(aabb, dtype=torch.float32, device=device) + + if scaling_activation == "exp": + self._scaling_activation = torch.exp + self._inverse_scaling_activation = torch.log + elif scaling_activation == "softplus": + self._scaling_activation = F.softplus + self._inverse_scaling_activation = lambda x: x + torch.log(-torch.expm1(-x)) + + self._opacity_activation = torch.sigmoid + self._inverse_opacity_activation = lambda x: torch.log(x / (1 - x)) + + self.scale_bias = self._inverse_scaling_activation(torch.tensor(self.scaling_bias)).to(self.device) + self.rots_bias = torch.zeros(4, device=self.device) + self.rots_bias[0] = 1 + self.opacity_bias_val = self._inverse_opacity_activation(torch.tensor(self.opacity_bias)).to(self.device) + + self._storage = {} + + def _get_store(self, name): + return self._storage.get(name) + + def _set_store(self, name, value): + self._storage[name] = value + + @property + def _xyz(self): + return self._get_store("_xyz") + @_xyz.setter + def _xyz(self, value): + if value is None: + self._set_store("_xyz", None) + self._set_store("xyz", None) + return + self._set_store("_xyz", value) + self._set_store("xyz", value * self.aabb[None, 3:] + self.aabb[None, :3]) + + @property + def get_xyz(self): + return self._get_store("xyz") + + @property + def _features_dc(self): + return self._get_store("_features_dc") + @_features_dc.setter + def _features_dc(self, value): + self._set_store("_features_dc", value) + + @property + def _opacity(self): + return self._get_store("_opacity") + @_opacity.setter + def _opacity(self, value): + if value is None: + self._set_store("_opacity", None) + self._set_store("opacity", None) + return + self._set_store("_opacity", value) + self._set_store("opacity", self._opacity_activation(value + self.opacity_bias_val)) + + @property + def get_opacity(self): + return self._get_store("opacity") + + @property + def _scaling(self): + return self._get_store("_scaling") + @_scaling.setter + def _scaling(self, value): + if value is None: + self._set_store("_scaling", None) + self._set_store("scaling", None) + return + self._set_store("_scaling", value) + s = self._scaling_activation(value + self.scale_bias) + s = torch.square(s) + self.mininum_kernel_size ** 2 + self._set_store("scaling", torch.sqrt(s)) + + @property + def get_scaling(self): + return self._get_store("scaling") + + @property + def _rotation(self): + return self._get_store("_rotation") + @_rotation.setter + def _rotation(self, value): + self._set_store("_rotation", value) + + _DEFAULT_TRANSFORM = [[1, 0, 0], [0, 0, -1], [0, 1, 0]] + + def render_tensors(self): + # Render-ready (activated, world-space) tensors for the generic SPLAT type. The axis transform + # (a 3x3 rotation, object frame -> viewer Y-up) is baked into positions and rotations. + # Returns float tensors on the intermediate device: positions (N,3), scales (N,3) linear, + # rotations (N,4) wxyz, opacities (N,1) in [0,1], sh (N,K,3) coefficients. + xyz = self.get_xyz.float() + scaling = self.get_scaling.float() + opacity = self.get_opacity.float() + rotation = (self._rotation + self.rots_bias[None, :]).float() + sh = self._features_dc.float() # (N, K, 3) + T = torch.as_tensor(self._DEFAULT_TRANSFORM, dtype=torch.float32, device=xyz.device) + xyz = xyz @ T.T + rotation = _matrix_to_quat(torch.matmul(T, _quat_to_matrix(rotation))) + rotation = rotation / torch.linalg.norm(rotation, dim=-1, keepdim=True) + out_device = comfy.model_management.intermediate_device() + return ( + xyz.to(out_device).contiguous(), scaling.to(out_device).contiguous(), + rotation.to(out_device).contiguous(), opacity.to(out_device).contiguous(), + sh.to(out_device).contiguous(), + ) + + +def _quat_to_matrix(q): + q = q / torch.linalg.norm(q, dim=-1, keepdim=True) + w, x, y, z = q[:, 0], q[:, 1], q[:, 2], q[:, 3] + R = torch.stack([ + 1 - 2*(y*y + z*z), 2*(x*y - w*z), 2*(x*z + w*y), + 2*(x*y + w*z), 1 - 2*(x*x + z*z), 2*(y*z - w*x), + 2*(x*z - w*y), 2*(y*z + w*x), 1 - 2*(x*x + y*y), + ], dim=-1).reshape(-1, 3, 3) + return R + + +def _matrix_to_quat(R): + trace = R[:, 0, 0] + R[:, 1, 1] + R[:, 2, 2] + q = torch.zeros((R.shape[0], 4), dtype=R.dtype, device=R.device) + s = torch.sqrt(torch.clamp(trace + 1, min=0)) * 2 + q[:, 0] = 0.25 * s + denom = torch.where(s != 0, s, torch.ones_like(s)) + q[:, 1] = (R[:, 2, 1] - R[:, 1, 2]) / denom + q[:, 2] = (R[:, 0, 2] - R[:, 2, 0]) / denom + q[:, 3] = (R[:, 1, 0] - R[:, 0, 1]) / denom + m01 = (R[:, 0, 0] >= R[:, 1, 1]) & (R[:, 0, 0] >= R[:, 2, 2]) & (s == 0) + s1 = torch.sqrt(torch.clamp(1 + R[:, 0, 0] - R[:, 1, 1] - R[:, 2, 2], min=0)) * 2 + q[m01, 0] = (R[m01, 2, 1] - R[m01, 1, 2]) / s1[m01] + q[m01, 1] = 0.25 * s1[m01] + q[m01, 2] = (R[m01, 0, 1] + R[m01, 1, 0]) / s1[m01] + q[m01, 3] = (R[m01, 0, 2] + R[m01, 2, 0]) / s1[m01] + m11 = (R[:, 1, 1] > R[:, 0, 0]) & (R[:, 1, 1] >= R[:, 2, 2]) & (s == 0) + s2 = torch.sqrt(torch.clamp(1 + R[:, 1, 1] - R[:, 0, 0] - R[:, 2, 2], min=0)) * 2 + q[m11, 0] = (R[m11, 0, 2] - R[m11, 2, 0]) / s2[m11] + q[m11, 1] = (R[m11, 0, 1] + R[m11, 1, 0]) / s2[m11] + q[m11, 2] = 0.25 * s2[m11] + q[m11, 3] = (R[m11, 1, 2] + R[m11, 2, 1]) / s2[m11] + m21 = (R[:, 2, 2] > R[:, 0, 0]) & (R[:, 2, 2] > R[:, 1, 1]) & (s == 0) + s3 = torch.sqrt(torch.clamp(1 + R[:, 2, 2] - R[:, 0, 0] - R[:, 1, 1], min=0)) * 2 + q[m21, 0] = (R[m21, 1, 0] - R[m21, 0, 1]) / s3[m21] + q[m21, 1] = (R[m21, 0, 2] + R[m21, 2, 0]) / s3[m21] + q[m21, 2] = (R[m21, 1, 2] + R[m21, 2, 1]) / s3[m21] + q[m21, 3] = 0.25 * s3[m21] + return q / torch.linalg.norm(q, dim=-1, keepdim=True) + + +def build_gaussian_models(decoder, points_pred: dict, pred: dict): + # Assemble GaussianModels from the elastic decoder layout. decoder is the ElasticGaussianFixedlenDecoder + # (carries layout / rep_config / _get_offset) + x = points_pred + offset = decoder._get_offset(pred['features']) + h = pred["features"] + ret = [] + for i in range(h.shape[0]): + g = GaussianModel( + sh_degree=0, + aabb=[-0.5, -0.5, -0.5, 1.0, 1.0, 1.0], + mininum_kernel_size=decoder.rep_config['filter_kernel_size_3d'], + scaling_bias=decoder.rep_config['scaling_bias'], + opacity_bias=decoder.rep_config['opacity_bias'], + scaling_activation=decoder.rep_config['scaling_activation'], + device=h.device, + ) + _x = x["points"][i, :, None, :] + for k, v in decoder.layout.items(): + if k == '_xyz': + setattr(g, k, (offset[i] + _x).flatten(0, 1)) + elif k in ('_xyz_center', '_offset_scale'): + continue + else: + feats = h[i][:, v['range'][0]:v['range'][1]].reshape(-1, *v['shape']).flatten(0, 1) + setattr(g, k, feats * decoder.rep_config['lr'][k]) + ret.append(g) + return ret diff --git a/comfy/ldm/triposplat/model.py b/comfy/ldm/triposplat/model.py new file mode 100644 index 000000000..d8a531772 --- /dev/null +++ b/comfy/ldm/triposplat/model.py @@ -0,0 +1,326 @@ +# TripoSplat flow-matching denoiser (LatentSeqMMFlowModel). Registered as a ModelType.FLOW arch and +# driven by the standard KSampler; jointly denoises the (B, 8192, 16) latent and a (B, 1, 5) camera token +# carried as a 2-element nested latent. +import numpy as np +import torch +import torch.nn as nn +import torch.nn.functional as F + +import comfy.model_management +import comfy.patcher_extension +import comfy.rmsnorm +from comfy.ldm.modules.attention import optimized_attention +from comfy.ldm.flux.math import apply_rope + + +class MultiHeadRMSNorm(nn.Module): + def __init__(self, dim, heads, dtype=None, device=None): + super().__init__() + self.gamma = nn.Parameter(torch.empty(heads, dim, dtype=dtype, device=device)) + + def forward(self, x): + x = comfy.rmsnorm.rms_norm(x) + return x * comfy.model_management.cast_to(self.gamma, x.dtype, x.device) + + +# Positional embeddings + +class RePo3DRotaryEmbedding(nn.Module): + def __init__(self, model_channels, num_heads, head_dim, repo_hidden_ratio=0.125, max_freq=16.0, + dtype=None, device=None, operations=None): + super().__init__() + self.num_heads = num_heads + self.head_dim = head_dim + repo_hidden_size = int(model_channels * repo_hidden_ratio) + self.norm = operations.LayerNorm(model_channels, dtype=dtype, device=device) + self.gate_map = operations.Linear(model_channels, repo_hidden_size, bias=False, dtype=dtype, device=device) + self.content_map = operations.Linear(model_channels, repo_hidden_size, bias=False, dtype=dtype, device=device) + self.act = nn.SiLU() + self.final_map = operations.Linear(repo_hidden_size, 3 * num_heads, bias=False, dtype=dtype, device=device) + self.dim_0 = 2 * (head_dim // 6) + self.dim_1 = 2 * (head_dim // 6) + self.dim_2 = head_dim - self.dim_0 - self.dim_1 + dims = [self.dim_0, self.dim_1, self.dim_2] + freqs_list = [] + for d in dims: + freq_dim = d // 2 + freqs_list.append(torch.linspace(1.0, float(max_freq), steps=freq_dim, dtype=torch.float32)) + self.freqs_0 = nn.Parameter(freqs_list[0]) + self.freqs_1 = nn.Parameter(freqs_list[1]) + self.freqs_2 = nn.Parameter(freqs_list[2]) + + def forward(self, hidden_states): + h = self.norm(hidden_states) + feat = self.act(self.gate_map(h)) * self.content_map(h) + out = self.final_map(feat) + B, L, _ = out.shape + delta_pos = out.reshape(B, L, self.num_heads, 3) + f0 = comfy.model_management.cast_to(self.freqs_0, torch.float32, out.device) + f1 = comfy.model_management.cast_to(self.freqs_1, torch.float32, out.device) + f2 = comfy.model_management.cast_to(self.freqs_2, torch.float32, out.device) + ang_0 = delta_pos[..., 0].unsqueeze(-1) * f0 * torch.pi + ang_1 = delta_pos[..., 1].unsqueeze(-1) * f1 * torch.pi + ang_2 = delta_pos[..., 2].unsqueeze(-1) * f2 * torch.pi + ang = torch.cat([ang_0, ang_1, ang_2], dim=-1).float() # (B, L, heads, head_dim/2) + cos, sin = ang.cos(), ang.sin() + return torch.stack([cos, -sin, sin, cos], dim=-1).reshape(*ang.shape, 2, 2) + + +class PcdAbsolutePositionEmbedder(nn.Module): + # Sinusoidal absolute position embedding. Two fixed schedules are used in TripoSplat: + # "pow2" (flow-model latent anchors) and "log2" (octree / gaussian decoders). + def __init__(self, channels: int, in_channels: int = 3, max_res: int = 16, schedule: str = "pow2"): + super().__init__() + self.channels = channels + self.in_channels = in_channels + self.max_res = max_res + self.schedule = schedule + self.freq_dim = channels // in_channels // 2 + + def _freqs(self, device): + if self.schedule == "pow2": + freqs_2exp = torch.arange(self.max_res, dtype=torch.float32, device=device) + res_dim = max(0, self.freq_dim - self.max_res) + freqs_res = (torch.arange(res_dim, dtype=torch.float32, device=device) / max(res_dim, 1) * self.max_res + if res_dim > 0 else torch.empty(0, device=device)) + freqs = torch.cat([freqs_2exp, freqs_res], dim=0)[:self.freq_dim] + return torch.pow(2.0, freqs) * 2.0 # *2 folds this schedule's 2*pi into the shared *pi below + logs = torch.linspace(0.0, float(self.max_res), steps=self.freq_dim, dtype=torch.float32, device=device) + return torch.pow(2.0, logs) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + orig_dtype = x.dtype + x = x.float() + *dims, D = x.shape + out = torch.outer(x.reshape(-1), self._freqs(x.device)) * torch.pi + out = torch.cat([out.sin(), out.cos()], dim=-1).reshape(*dims, -1) + if out.shape[-1] < self.channels: + out = torch.cat([out, torch.zeros(*dims, self.channels - out.shape[-1], + device=out.device, dtype=out.dtype)], dim=-1) + return out.to(orig_dtype) + + +def attention(q, k, v, transformer_options=None): + # q, k, v: (B, L, heads, dim) -> (B, L, heads, dim). Shared optimized_attention call convention. + out = optimized_attention(q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), heads=q.shape[2], + skip_reshape=True, skip_output_reshape=True, low_precision_attention=False, + transformer_options=transformer_options) + return out.transpose(1, 2) + + +# Transformer building blocks + +class MLP(nn.Module): + def __init__(self, in_channels, hidden_channels, out_channels, dtype=None, device=None, operations=None): + super().__init__() + self.mlp = nn.Sequential( + operations.Linear(in_channels, hidden_channels, dtype=dtype, device=device), + nn.GELU(approximate="tanh"), + operations.Linear(hidden_channels, out_channels, dtype=dtype, device=device), + ) + + def forward(self, x): + return self.mlp(x) + + +class RopeMultiHeadAttention(nn.Module): + def __init__(self, channels, num_heads, qkv_bias=True, qk_rms_norm=False, use_rope=False, + dtype=None, device=None, operations=None): + super().__init__() + self.channels = channels + self.num_heads = num_heads + self.head_dim = channels // num_heads + self.qk_rms_norm = qk_rms_norm + self.use_rope = use_rope + self.qkv = operations.Linear(channels, channels * 3, bias=qkv_bias, dtype=dtype, device=device) + if self.qk_rms_norm: + self.q_norm = MultiHeadRMSNorm(self.head_dim, num_heads, dtype=dtype, device=device) + self.k_norm = MultiHeadRMSNorm(self.head_dim, num_heads, dtype=dtype, device=device) + self.out = operations.Linear(channels, channels, dtype=dtype, device=device) + + def forward(self, x, rope_emb=None, transformer_options=None): + B, L, C = x.shape + qkv = self.qkv(x).reshape(B, L, 3, self.num_heads, self.head_dim) + q, k, v = qkv.unbind(2) + if self.use_rope: + q, k = apply_rope(q, k, rope_emb) + if self.qk_rms_norm: + q = self.q_norm(q) + k = self.k_norm(k) + h = attention(q, k, v, transformer_options) # (B, L, heads, dim) + return self.out(h.reshape(B, L, C)) + + +class UnifiedTransformerBlock(nn.Module): + def __init__(self, channels, num_heads, mlp_ratio=4.0, + use_rope=False, qk_rms_norm=False, qkv_bias=True, + modulation=True, share_mod=False, + dtype=None, device=None, operations=None): + super().__init__() + self.modulation = modulation + self.share_mod = share_mod + self.norm1 = operations.LayerNorm(channels, elementwise_affine=not modulation, eps=1e-6, dtype=dtype, device=device) + self.norm2 = operations.LayerNorm(channels, elementwise_affine=not modulation, eps=1e-6, dtype=dtype, device=device) + self.attn = RopeMultiHeadAttention(channels, num_heads=num_heads, + qkv_bias=qkv_bias, use_rope=use_rope, qk_rms_norm=qk_rms_norm, + dtype=dtype, device=device, operations=operations) + self.mlp = MLP(channels, int(channels * mlp_ratio), channels, dtype=dtype, device=device, operations=operations) + if modulation: + if not share_mod: + self.adaLN_modulation = nn.Sequential( + nn.SiLU(), operations.Linear(channels, 6 * channels, bias=True, dtype=dtype, device=device)) + self.shift_table = nn.Parameter(torch.empty(1, 6 * channels, dtype=dtype, device=device)) + + def forward(self, x, mod=None, rotary_emb=None, transformer_options=None): + if self.modulation: + if not self.share_mod: + mod = self.adaLN_modulation(mod) + mod = mod + comfy.model_management.cast_to(self.shift_table, mod.dtype, mod.device) + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = mod.chunk(6, dim=1) + h = torch.addcmul(shift_msa.unsqueeze(1), self.norm1(x), 1 + scale_msa.unsqueeze(1)) + x = torch.addcmul(x, self.attn(h, rope_emb=rotary_emb, transformer_options=transformer_options), gate_msa.unsqueeze(1)) + h = torch.addcmul(shift_mlp.unsqueeze(1), self.norm2(x), 1 + scale_mlp.unsqueeze(1)) + x = torch.addcmul(x, self.mlp(h), gate_mlp.unsqueeze(1)) + else: + x = x + self.attn(self.norm1(x), rope_emb=rotary_emb, transformer_options=transformer_options) + x = x + self.mlp(self.norm2(x)) + return x + + +class TimestepEmbedder(nn.Module): + def __init__(self, hidden_size, frequency_embedding_size=256, dtype=None, device=None, operations=None): + super().__init__() + self.mlp = nn.Sequential( + operations.Linear(frequency_embedding_size, hidden_size, bias=True, dtype=dtype, device=device), + nn.SiLU(), + operations.Linear(hidden_size, hidden_size, bias=True, dtype=dtype, device=device), + ) + self.frequency_embedding_size = frequency_embedding_size + + @staticmethod + def timestep_embedding(t, dim, max_period=10000): + half = dim // 2 + freqs = torch.exp(-np.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(device=t.device) + args = t[:, None].float() * freqs[None] + embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1) + if dim % 2: + embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1) + return embedding + + def forward(self, t): + emb = self.timestep_embedding(t, self.frequency_embedding_size) + return self.mlp(emb.to(self.mlp[0].weight.dtype)) + + +class LatentSeqMMFlowModel(nn.Module): + def __init__(self, image_model=None, q_token_length=8192, in_channels=16, model_channels=1024, + cond_channels=1280, out_channels=16, num_blocks=24, num_refiner_blocks=2, + num_heads=None, num_head_channels=64, cam_channels=5, cond2_channels=128, + mlp_ratio=4, share_mod=True, qk_rms_norm=True, + dtype=None, device=None, operations=None, **kwargs): + super().__init__() + self.dtype = dtype + self.q_token_length = q_token_length + self.in_channels = in_channels + self.cam_channels = cam_channels + self.model_channels = model_channels + self.cond_channels = cond_channels + self.cond2_channels = cond2_channels + self.out_channels = out_channels + self.num_blocks = num_blocks + self.num_refiner_blocks = num_refiner_blocks + self.num_heads = num_heads or model_channels // num_head_channels + self.mlp_ratio = mlp_ratio + self.share_mod = share_mod + self.qk_rms_norm = qk_rms_norm + + factory_kwargs = dict(dtype=dtype, device=device) + op_kwargs = dict(operations=operations, **factory_kwargs) + + self.t_embedder = TimestepEmbedder(model_channels, **op_kwargs) + if share_mod: + self.adaLN_modulation = nn.Sequential(nn.SiLU(), operations.Linear(model_channels, 6 * model_channels, bias=True, **factory_kwargs)) + + self.input_layer = operations.Linear(in_channels, model_channels, **factory_kwargs) + self.cond_embedder = operations.Linear(cond_channels, model_channels, **factory_kwargs) + self.cond_embedder2 = operations.Linear(cond2_channels, model_channels, **factory_kwargs) if cond2_channels is not None else None + + # Fixed Sobol (low-discrepancy) 3D anchor positions for the latent tokens, used as positional encoding. + # The embedder is parameter-free and the anchors are fixed, precompute once. + sobol_seq = torch.quasirandom.SobolEngine(dimension=3, scramble=True, seed=123).draw(q_token_length) + pos_emb = PcdAbsolutePositionEmbedder(model_channels)(sobol_seq.unsqueeze(0)) + self.register_buffer("pos_emb", pos_emb, persistent=False) + + # RePo3DRotaryEmbedding layers for the refiner and main blocks + repo_kwargs = dict(num_heads=self.num_heads, head_dim=num_head_channels, **op_kwargs) + self.noise_repo_layers = nn.ModuleList( + [RePo3DRotaryEmbedding(model_channels, **repo_kwargs) for _ in range(num_refiner_blocks)]) + self.context_repo_layers = nn.ModuleList( + [RePo3DRotaryEmbedding(model_channels, **repo_kwargs) for _ in range(num_refiner_blocks)]) + self.repo_layers = nn.ModuleList( + [RePo3DRotaryEmbedding(model_channels, **repo_kwargs) for _ in range(num_blocks)]) + + # Refiner blocks + block_kwargs = dict(num_heads=self.num_heads, mlp_ratio=self.mlp_ratio, use_rope=True, qk_rms_norm=self.qk_rms_norm, **op_kwargs) + self.noise_refiner = nn.ModuleList( + [UnifiedTransformerBlock(model_channels, modulation=True, share_mod=self.share_mod, **block_kwargs) for _ in range(num_refiner_blocks)]) + self.context_refiner = nn.ModuleList( + [UnifiedTransformerBlock(model_channels, modulation=False, **block_kwargs) for _ in range(num_refiner_blocks)]) + + self.cam_refiner = MLP(self.cam_channels, model_channels, model_channels, **op_kwargs) + + self.blocks = nn.ModuleList( + [UnifiedTransformerBlock(model_channels, modulation=True, share_mod=self.share_mod, **block_kwargs) for _ in range(num_blocks)]) + + self.shift_table = nn.Parameter(torch.empty(1, 2, model_channels, **factory_kwargs)) + self.out_layer = operations.Linear(model_channels, out_channels, **factory_kwargs) + self.cam_out_layer = operations.Linear(model_channels, cam_channels, **factory_kwargs) + + def forward(self, x, t, context=None, ref_latents=None, transformer_options={}, **kwargs): + return comfy.patcher_extension.WrapperExecutor.new_class_executor( + self._forward, + self, + comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options) + ).execute(x, t, context, ref_latents, transformer_options, **kwargs) + + def _forward(self, x, t, context=None, ref_latents=None, transformer_options={}, **kwargs): + # x is the unpacked nested latent: [latent (B,8192,in_channels), camera (B,1,cam_channels)]. + # context == feature1. + z, camera = x[0], x[1] + feat1 = context + + h_x = self.input_layer(z) + h_cond = self.cond_embedder(feat1) + if ref_latents is not None and self.cond_embedder2 is not None: + # Flatten the Flux2 VAE latent (B,128,h,w) to a token sequence and front-pad to feat1's length + # (the pad count = feat1's prefix tokens: DINOv3 cls + registers), then add to the context. + feat2 = ref_latents[0].flatten(2).transpose(1, 2) + feat2 = F.pad(feat2, (0, 0, feat1.shape[1] - feat2.shape[1], 0)) + h_cond = h_cond + self.cond_embedder2(feat2.to(h_cond.dtype)) + t_emb = self.t_embedder(t) + t_mod = self.adaLN_modulation(t_emb) if self.share_mod else t_emb + + h_x = h_x + self.pos_emb.to(z) + + for i, block in enumerate(self.noise_refiner): + h_x = block(h_x, mod=t_mod, rotary_emb=self.noise_repo_layers[i](h_x), transformer_options=transformer_options) + + for i, block in enumerate(self.context_refiner): + h_cond = block(h_cond, mod=None, rotary_emb=self.context_repo_layers[i](h_cond), transformer_options=transformer_options) + + cam = camera.to(z) + h_cam = self.cam_refiner(cam) + h = torch.cat([h_x, h_cond, h_cam], dim=1) + + for i, block in enumerate(self.blocks): + h = block(h, mod=t_mod, rotary_emb=self.repo_layers[i](h), transformer_options=transformer_options) + + h_x = F.layer_norm(h[:, :z.shape[1]].float(), h.shape[-1:]).to(z) + h_cam = F.layer_norm(h[:, -cam.shape[1]:].float(), h.shape[-1:]).to(z) + + shift, scale = (comfy.model_management.cast_to(self.shift_table, t_emb.dtype, t_emb.device) + t_emb.unsqueeze(1)).chunk(2, dim=1) + scale = 1 + scale + h_x = torch.addcmul(shift, h_x, scale) + h_cam = torch.addcmul(shift, h_cam, scale) + + return self.out_layer(h_x), self.cam_out_layer(h_cam) diff --git a/comfy/ldm/triposplat/preview.py b/comfy/ldm/triposplat/preview.py new file mode 100644 index 000000000..6a942bb53 --- /dev/null +++ b/comfy/ldm/triposplat/preview.py @@ -0,0 +1,91 @@ +# Live preview for TripoSplat: decode an x0 estimate into a coarse gaussian splat and render it with a perspective orbit camera. +import numpy as np +from PIL import Image + +_C0 = 0.28209479177387814 +_LATENT_TOKENS = 8192 # q_token_length +_LATENT_CH = 16 # in_channels +_OBJECT_TO_VIEWER = np.array([[1, 0, 0], [0, 0, -1], [0, 1, 0]], np.float32) # object frame -> viewer Y-up frame + + +def _view_matrix(yaw_deg, pitch_deg): + y, p = np.radians(yaw_deg), np.radians(pitch_deg) + Ry = np.array([[np.cos(y), 0, np.sin(y)], [0, 1, 0], [-np.sin(y), 0, np.cos(y)]], np.float32) + Rx = np.array([[1, 0, 0], [0, np.cos(p), -np.sin(p)], [0, np.sin(p), np.cos(p)]], np.float32) + return Rx @ Ry + + +def render_splat(xyz, rgb, scale, opacity=None, yaw=35.0, pitch=30.0, size=320, min_px=2, gain=1.0, + max_px=9, min_opacity=0.0, fov=35.0, dist=2.2): + # Project gaussian centers with a perspective camera and paint each as a filled disk whose screen + # radius follows the gaussian's world-space scale, composited with a nearest-wins z-buffer. + # gain scales the footprint (≈ std spanned), `min_px`/`max_px` clamp the on-screen radius. + + pts = xyz.astype(np.float32) @ _OBJECT_TO_VIEWER.T + v = pts @ _view_matrix(yaw, pitch).T + zc = v[:, 2] + dist + keep = zc > 1e-2 + if opacity is not None and min_opacity > 0.0: # culls gaussians with very low opacity + keep = keep & (opacity > min_opacity) + v, zc, scale = v[keep], zc[keep], scale[keep] + col = (np.clip(rgb, 0, 1)[:, :3] * 255).astype(np.uint8)[keep] + if v.shape[0] == 0: + return Image.fromarray(np.zeros((size, size, 3), np.uint8)) + f = (size / 2) / np.tan(np.radians(fov) / 2) + cx = size / 2 + f * v[:, 0] / zc + cy = size / 2 + f * v[:, 1] / zc + radius = np.clip(np.round(f * scale / zc * gain), min_px, max_px).astype(np.int32) + + # Expand each splat to its disk pixels, bucketed by integer radius so it stays vectorized. + px, py, pz, pc = [], [], [], [] + for r in range(int(radius.min()), int(radius.max()) + 1): + m = radius == r + if not m.any(): + continue + dy, dx = np.mgrid[-r:r + 1, -r:r + 1] + disk = (dx * dx + dy * dy) <= r * r + ox, oy = dx[disk], dy[disk] + px.append((cx[m, None] + ox).ravel()) + py.append((cy[m, None] + oy).ravel()) + pz.append(np.repeat(zc[m], ox.size)) + pc.append(np.repeat(col[m], ox.size, axis=0)) + px, py = np.concatenate(px), np.concatenate(py) + pz, pc = np.concatenate(pz), np.concatenate(pc) + xi = np.clip(px, 0, size - 1).astype(np.int64) + yi = np.clip(py, 0, size - 1).astype(np.int64) + + # Nearest-wins z-buffer: pack (quantized depth, source index), per-pixel min picks the closest + # splat, then decode the winning index back to its color. + pid = yi * size + xi + q = np.clip((pz * 1024.0).astype(np.int64), 0, (1 << 20) - 1) # near = small + key = (q << 32) | np.arange(pid.size, dtype=np.int64) + buf = np.full(size * size, 1 << 62, np.int64) + np.minimum.at(buf, pid, key) + img = np.zeros((size * size, 3), np.uint8) + hit = buf < (1 << 62) + img[hit] = pc[buf[hit] & 0xFFFFFFFF] + return Image.fromarray(img.reshape(size, size, 3)) + + +def _extract_latent(x0): + # x0 from the sampler callback is the nested latent packed to (B, 1, TOKENS*CH + 1*5); + # the plain single-latent case is (B, TOKENS, CH). Return the (B, TOKENS, CH) latent stream. + if x0.ndim == 3 and x0.shape[1] == _LATENT_TOKENS and x0.shape[2] == _LATENT_CH: + return x0 + flat = x0.reshape(x0.shape[0], -1) + return flat[:, :_LATENT_TOKENS * _LATENT_CH].reshape(x0.shape[0], _LATENT_TOKENS, _LATENT_CH) + + +def decode_x0_to_image(decoder, x0, cfg): + # Decode x0 at a coarse octree level / few gaussians and render a preview image. + latent = _extract_latent(x0) + fsm = decoder.first_stage_model + gaussian = fsm.decode(latent.to(decoder.device, decoder.vae_dtype), + num_gaussians=cfg.get("gaussians", 16384), level=cfg.get("level", 5))[0] + xyz = gaussian.get_xyz.float().cpu().numpy() + rgb = gaussian._features_dc.float().cpu().numpy()[:, 0, :] * _C0 + 0.5 + scale = gaussian.get_scaling.float().cpu().numpy().max(axis=1) # per-splat world radius (largest axis) + opacity = gaussian.get_opacity.float().cpu().numpy()[:, 0] + return render_splat(xyz, rgb, scale, opacity=opacity, yaw=cfg.get("yaw", 35.0), pitch=cfg.get("pitch", 30.0), + size=cfg.get("size", 320), min_px=1, gain=1.0, max_px=cfg.get("point_size", 3), + min_opacity=0.01) diff --git a/comfy/ldm/triposplat/vae.py b/comfy/ldm/triposplat/vae.py new file mode 100644 index 000000000..e5ed9fd36 --- /dev/null +++ b/comfy/ldm/triposplat/vae.py @@ -0,0 +1,382 @@ +# TripoSplat gaussian decoder ("VAE"): an octree probability decoder picks point coords, then an +# elastic-gaussian decoder predicts per-point gaussian params. OctreeGaussianDecoder.decode() returns +# a Gaussian. The octree sampler uses the global torch RNG (no generator) like upstream, so seed it for repeatable decodes. +import numpy as np +import torch +import torch.nn as nn +import torch.nn.functional as F + +import comfy.model_management +import comfy.ops +from .gaussian import build_gaussian_models +from .model import MultiHeadRMSNorm, MLP, PcdAbsolutePositionEmbedder, attention + + +# Quasi-random sampling utilities (pure functions, dtype/device-agnostic) + +PRIMES = [2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31, 37, 41, 43, 47, 53] + + +def radical_inverse(base, n): + val = 0 + inv_base = 1.0 / base + inv_base_n = inv_base + while n > 0: + digit = n % base + val += digit * inv_base_n + n //= base + inv_base_n *= inv_base + return val + + +def halton_sequence(dim, n): + return [radical_inverse(PRIMES[i], n) for i in range(dim)] + + +def hammersley_sequence(dim, n, num_samples): + return [n / num_samples] + halton_sequence(dim - 1, n) + + +def sample_probs(probs, counts, generator=None): + # Systematic resampling: distribute counts[r] draws across the P bins of row r + batch_shape = counts.shape + R = counts.numel() + P = probs.size(-1) + device = probs.device + probs = probs.reshape(R, P).to(torch.float32).clamp_min(0) + counts = counts.reshape(R).to(device=device, dtype=torch.long) + + row_sums = probs.sum(1, keepdim=True) + probs = torch.where(row_sums == 0, probs.new_tensor(1.0 / P), probs / row_sums.clamp_min(1)) + cdf = probs.cumsum(dim=1).clamp(max=1.0 - 1e-12) + + Nmax = int(counts.max()) + if Nmax == 0: + return counts.new_zeros(*batch_shape, P) + cnt = counts.clamp_min(1).float().unsqueeze(1) # (R, 1) + grid = torch.arange(Nmax, device=device, dtype=torch.float32).unsqueeze(0) # (1, Nmax) + u = (torch.rand(R, 1, generator=generator).to(device) + grid) / cnt # (R, Nmax) systematic samples (CPU-seeded) + idx = torch.searchsorted(cdf, u.clamp(max=1.0 - 1e-12)).clamp_max(P - 1) + weight = (grid < counts.unsqueeze(1)).to(cdf.dtype) # mask out j >= counts[r] + out = torch.zeros(R, P, dtype=torch.float32, device=device) + out.scatter_add_(1, idx, weight) + return out.to(torch.long).view(*batch_shape, P) + + +class MultiHeadAttention(nn.Module): + def __init__(self, channels, num_heads, ctx_channels=None, type="self", qkv_bias=True, qk_rms_norm=False, + dtype=None, device=None, operations=None): + super().__init__() + assert channels % num_heads == 0 + self.channels = channels + self.head_dim = channels // num_heads + self.ctx_channels = ctx_channels if ctx_channels is not None else channels + self.num_heads = num_heads + self._type = type + self.qk_rms_norm = qk_rms_norm + if self._type == "self": + self.to_qkv = operations.Linear(channels, channels * 3, bias=qkv_bias, dtype=dtype, device=device) + else: + self.to_q = operations.Linear(channels, channels, bias=qkv_bias, dtype=dtype, device=device) + self.to_kv = operations.Linear(self.ctx_channels, channels * 2, bias=qkv_bias, dtype=dtype, device=device) + if self.qk_rms_norm: + self.q_rms_norm = MultiHeadRMSNorm(self.head_dim, num_heads, dtype=dtype, device=device) + self.k_rms_norm = MultiHeadRMSNorm(self.head_dim, num_heads, dtype=dtype, device=device) + self.to_out = operations.Linear(channels, channels, dtype=dtype, device=device) + + def forward(self, x, context=None): + B, L, C = x.shape + if self._type == "self": + q, k, v = self.to_qkv(x).reshape(B, L, 3, self.num_heads, -1).unbind(dim=2) + else: + Lkv = context.shape[1] + q = self.to_q(x).reshape(B, L, self.num_heads, -1) + k, v = self.to_kv(context).reshape(B, Lkv, 2, self.num_heads, -1).unbind(dim=2) + if self.qk_rms_norm: + q = self.q_rms_norm(q) + k = self.k_rms_norm(k) + h = attention(q, k, v) + return self.to_out(h.reshape(B, L, -1)) + + +# Octree probability decoder + +class LevelEmbedder(nn.Module): + def __init__(self, hidden_size, frequency_embedding_size=256, max_period=1024, + dtype=None, device=None, operations=None): + super().__init__() + self.mlp = nn.Sequential( + operations.Linear(frequency_embedding_size, hidden_size, bias=True, dtype=dtype, device=device), + nn.SiLU(), + operations.Linear(hidden_size, hidden_size, bias=True, dtype=dtype, device=device), + ) + self.frequency_embedding_size = frequency_embedding_size + self.max_period = max_period + + @staticmethod + def level_embedding(t, dim, max_period=1024): + half = dim // 2 + freqs = torch.exp(-np.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(device=t.device) + args = t[:, None].float() * freqs[None] * 2 * torch.pi + embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1) + if dim % 2: + embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1) + return embedding + + def forward(self, t): + emb = self.level_embedding(t, self.frequency_embedding_size, self.max_period) + return self.mlp(emb.to(self.mlp[0].weight.dtype)) + + +class ModulatedTransformerCrossOnlyBlock(nn.Module): + def __init__(self, channels, ctx_channels, num_heads, mlp_ratio=4.0, share_mod=False, + qk_rms_norm_cross=True, qkv_bias=True, dtype=None, device=None, operations=None): + super().__init__() + self.share_mod = share_mod + self.norm1 = operations.LayerNorm(channels, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) + self.norm2 = operations.LayerNorm(channels, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) + self.cross_attn = MultiHeadAttention(channels, ctx_channels=ctx_channels, num_heads=num_heads, + type="cross", qkv_bias=qkv_bias, + qk_rms_norm=qk_rms_norm_cross, dtype=dtype, device=device, operations=operations) + self.mlp = MLP(channels, int(channels * mlp_ratio), channels, dtype=dtype, device=device, operations=operations) + if not share_mod: + self.adaLN_modulation = nn.Sequential( + nn.SiLU(), operations.Linear(channels, 6 * channels, bias=True, dtype=dtype, device=device)) + + def forward(self, x, mod, context): + if self.share_mod: + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = mod.chunk(6, dim=1) + else: + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(mod).chunk(6, dim=1) + h = torch.addcmul(shift_msa.unsqueeze(1), self.norm1(x), 1 + scale_msa.unsqueeze(1)) + x = torch.addcmul(x, self.cross_attn(h, context), gate_msa.unsqueeze(1)) + h = torch.addcmul(shift_mlp.unsqueeze(1), self.norm2(x), 1 + scale_mlp.unsqueeze(1)) + x = torch.addcmul(x, self.mlp(h), gate_mlp.unsqueeze(1)) + return x + + +class OctreeProbabilityFixedlenDecoder(nn.Module): + # Cross-attention transformer over octree coords -> per-node 8-way child occupancy logits. + def __init__(self, model_channels=1024, cond_channels=16, num_blocks=4, num_heads=16, + num_head_channels=64, mlp_ratio=4.0, share_mod=True, + qk_rms_norm_cross=True, dtype=None, device=None, operations=None): + super().__init__() + self.model_channels = model_channels + self.cond_channels = cond_channels + self.num_blocks = num_blocks + self.num_heads = num_heads or model_channels // num_head_channels + self.mlp_ratio = mlp_ratio + self.share_mod = share_mod + self.qk_rms_norm_cross = qk_rms_norm_cross + self.input_layer = operations.Linear(model_channels, model_channels, dtype=dtype, device=device) + self.l_embedder = LevelEmbedder(model_channels, dtype=dtype, device=device, operations=operations) + if share_mod: + self.adaLN_modulation = nn.Sequential( + nn.SiLU(), operations.Linear(model_channels, 6 * model_channels, bias=True, dtype=dtype, device=device)) + if cond_channels is not None: + self.blocks = nn.ModuleList([ + ModulatedTransformerCrossOnlyBlock( + model_channels, ctx_channels=cond_channels, num_heads=self.num_heads, + mlp_ratio=self.mlp_ratio, qk_rms_norm_cross=self.qk_rms_norm_cross, + share_mod=self.share_mod, dtype=dtype, device=device, operations=operations) + for _ in range(num_blocks) + ]) + self.out_proj = operations.Linear(model_channels, 8, dtype=dtype, device=device) + self.in_proj = operations.Linear(3, model_channels, dtype=dtype, device=device) + self.pos_embedder = PcdAbsolutePositionEmbedder(channels=model_channels, in_channels=3, max_res=10, schedule="log2") + + def forward(self, x, l, cond): + d = next(self.parameters()).dtype + B, L, _ = x.shape + h = self.in_proj(x.to(d)) + self.pos_embedder(x.reshape(-1, 3)).reshape(B, L, -1).to(d) + h = self.input_layer(h) + l_emb = self.l_embedder(l) + if self.share_mod: + l_emb = self.adaLN_modulation(l_emb) + cond = cond.to(d) + for block in self.blocks: + h = block(h, l_emb, cond) + h = F.layer_norm(h.float(), h.shape[-1:]).to(d) + logits = self.out_proj(h) + return {"logits": logits, "probs": torch.softmax(logits, dim=-1)} + + @staticmethod + def sample(model, cond, num_points, level, temperature=1.0, generator=None): + B = cond.shape[0] + device = cond.device + child_offset = torch.tensor([[i, j, k] for k in [0, 1] for j in [0, 1] for i in [0, 1]], + dtype=torch.long, device=device) + prev_coords_int = torch.zeros(B, 1, 3, dtype=torch.long, device=device) + prev_counts = torch.full((B, 1), num_points, dtype=torch.long, device=device) + prev_log_probs = torch.zeros(B, 1, dtype=torch.float32, device=device) + batch_indices_range = torch.arange(B, device=device).unsqueeze(1) + + for lv in range(1, level + 1): + res_p = 1 << (lv - 1) + res = 1 << lv + parent_coords_norm = (prev_coords_int.to(torch.float32) + 0.5) / res_p + res_tensor = torch.full((B,), res, dtype=torch.long, device=device) + pred_logits = model(parent_coords_norm, res_tensor, cond)["logits"] / temperature + pred_probs = torch.softmax(pred_logits, dim=-1) + pred_log_probs = torch.log_softmax(pred_logits, dim=-1) + sampled = sample_probs(pred_probs, prev_counts, generator=generator).flatten(1, 2) + pred_log_probs = pred_log_probs.flatten(1, 2) + prev_log_probs_expanded = prev_log_probs.repeat_interleave(8, dim=1) + child_coords_int = (prev_coords_int[:, :, None, :] * 2 + child_offset[None, None, :, :]).flatten(1, 2) + mask = sampled > 0 + max_valid = mask.sum(dim=1).max().item() + scatter_indices = mask.cumsum(dim=1) - 1 + valid_scatter_indices = scatter_indices[mask] + valid_batch_indices = batch_indices_range.expand_as(mask)[mask] + next_prev_coords_int = torch.zeros(B, max_valid, 3, dtype=child_coords_int.dtype, device=device) + next_prev_coords_int[valid_batch_indices, valid_scatter_indices] = child_coords_int[mask] + next_prev_counts = torch.zeros(B, max_valid, dtype=sampled.dtype, device=device) + next_prev_counts[valid_batch_indices, valid_scatter_indices] = sampled[mask] + next_prev_log_probs = torch.zeros(B, max_valid, dtype=prev_log_probs.dtype, device=device) + next_prev_log_probs[valid_batch_indices, valid_scatter_indices] = (prev_log_probs_expanded + pred_log_probs)[mask] + prev_coords_int = next_prev_coords_int + prev_counts = next_prev_counts + prev_log_probs = next_prev_log_probs + + res = 1 << level + prev_log_probs = torch.repeat_interleave(prev_log_probs.flatten(0, 1), prev_counts.flatten(0, 1), dim=0).reshape(B, num_points) + coords_int = torch.repeat_interleave(prev_coords_int.flatten(0, 1), prev_counts.flatten(0, 1), dim=0).reshape(B, num_points, -1) + rand = torch.rand(coords_int.shape, dtype=torch.float32, generator=generator).to(device) + coords_norm = (coords_int.to(torch.float32) + rand) / res + return {"points": coords_norm, "log_probs": prev_log_probs} + + +# Elastic gaussian decoder + +class TransformerCrossBlock(nn.Module): + def __init__(self, channels, ctx_channels, num_heads, mlp_ratio=4.0, + qk_rms_norm=True, qk_rms_norm_cross=True, qkv_bias=True, + dtype=None, device=None, operations=None): + super().__init__() + self.norm1 = operations.LayerNorm(channels, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) + self.norm2 = operations.LayerNorm(channels, elementwise_affine=True, eps=1e-6, dtype=dtype, device=device) + self.norm3 = operations.LayerNorm(channels, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) + self.self_attn = MultiHeadAttention(channels, num_heads=num_heads, type="self", qkv_bias=qkv_bias, + qk_rms_norm=qk_rms_norm, dtype=dtype, device=device, operations=operations) + self.cross_attn = MultiHeadAttention(channels, ctx_channels=ctx_channels, num_heads=num_heads, type="cross", + qkv_bias=qkv_bias, qk_rms_norm=qk_rms_norm_cross, dtype=dtype, device=device, operations=operations) + self.mlp = MLP(channels, int(channels * mlp_ratio), channels, dtype=dtype, device=device, operations=operations) + + def forward(self, x, context): + x = x + self.self_attn(self.norm1(x)) + x = x + self.cross_attn(self.norm2(x), context) + x = x + self.mlp(self.norm3(x)) + return x + + +class ElasticGaussianFixedlenDecoder(nn.Module): + # Cross-attention transformer over sampled octree points -> per-point gaussian params. + def __init__(self, in_channels=3, model_channels=1024, cond_channels=16, num_blocks=16, num_heads=16, + num_head_channels=64, mlp_ratio=4.0, *, representation_config=None, + qk_rms_norm=True, qk_rms_norm_cross=True, dtype=None, device=None, operations=None): + super().__init__() + self.rep_config = representation_config or dict( + lr=dict(_xyz=1.0, _features_dc=1.0, _opacity=1.0, _scaling=1.0, _rotation=0.1), + perturb_offset=True, perturbe_size=1.5, offset_scale=0.05, num_gaussians=32, + filter_kernel_size_3d=0.0009, scaling_bias=0.004, opacity_bias=0.1, + scaling_activation="softplus", + ) + self.out_channels = self._calc_layout() + self.model_channels = model_channels + self.cond_channels = cond_channels + self.num_blocks = num_blocks + self.num_heads = num_heads or model_channels // num_head_channels + self.mlp_ratio = mlp_ratio + self.input_layer = operations.Linear(model_channels, model_channels, dtype=dtype, device=device) + if cond_channels is not None: + self.blocks = nn.ModuleList([ + TransformerCrossBlock(model_channels, ctx_channels=cond_channels, + num_heads=self.num_heads, mlp_ratio=self.mlp_ratio, + qk_rms_norm=qk_rms_norm, qk_rms_norm_cross=qk_rms_norm_cross, + dtype=dtype, device=device, operations=operations) + for _ in range(num_blocks) + ]) + self.in_proj = operations.Linear(in_channels, model_channels, dtype=dtype, device=device) + self.pos_embedder = PcdAbsolutePositionEmbedder(channels=model_channels, in_channels=3, max_res=10, schedule="log2") + self.out_proj = operations.Linear(model_channels, self.out_channels, dtype=dtype, device=device) + self._build_perturbation() + + def _calc_layout(self): + ng = self.rep_config['num_gaussians'] + self.layout = { + '_xyz': {'shape': (ng, 3), 'size': ng * 3}, + '_features_dc': {'shape': (ng, 1, 3), 'size': ng * 3}, + '_scaling': {'shape': (ng, 3), 'size': ng * 3}, + '_rotation': {'shape': (ng, 4), 'size': ng * 4}, + '_opacity': {'shape': (ng, 1), 'size': ng}, + } + self.layout['_offset_scale'] = {'shape': (ng, 1), 'size': ng} + start = 0 + for k, v in self.layout.items(): + v['range'] = (start, start + v['size']) + start += v['size'] + return start + + def _build_perturbation(self): + ng = self.rep_config['num_gaussians'] + perturbation = torch.tensor([hammersley_sequence(3, i, ng) for i in range(ng)]).float() + perturbation = torch.atanh((perturbation * 2 - 1) / self.rep_config['perturbe_size']) + self.register_buffer('points_offset_perturbation', perturbation) + base = torch.tensor(self.rep_config['offset_scale']) + self.register_buffer('base_offset_scale', torch.log(torch.exp(base) - 1.0)) + + def _get_offset(self, h): + B = h.shape[0] + r = self.layout['_offset_scale']['range'] + _offset_scale = F.softplus( + h[:, :, r[0]:r[1]].reshape(B, -1, *self.layout['_offset_scale']['shape']) + + comfy.model_management.cast_to(self.base_offset_scale, h.dtype, h.device)) + + r = self.layout['_xyz']['range'] + offset = h[:, :, r[0]:r[1]].reshape(B, -1, *self.layout['_xyz']['shape']) + offset = offset * self.rep_config['lr']['_xyz'] + if self.rep_config['perturb_offset']: + offset = offset + comfy.model_management.cast_to(self.points_offset_perturbation, offset.dtype, offset.device) + offset = torch.tanh(offset) * 0.5 * self.rep_config['perturbe_size'] + offset = offset * _offset_scale + return offset + + def forward(self, x=None, cond=None): + pcd = x["points"] + d = next(self.parameters()).dtype + B, L, _ = pcd.shape + h = self.in_proj(pcd.to(d)) + self.pos_embedder(pcd.reshape(-1, 3)).reshape(B, L, -1).to(d) + h = self.input_layer(h) + cond = cond.to(d) + for block in self.blocks: + h = block(h, cond) + h = F.layer_norm(h.float(), h.shape[-1:]).to(h.dtype) + return {"features": self.out_proj(h)} + + +# Combined octree gaussian decoder (comfy first-stage model) + +class OctreeGaussianDecoder(nn.Module): + _MAX_VOXEL_LEVEL = 8 + + def __init__(self, dtype=None, device=None, operations=None): + super().__init__() + if operations is None: + operations = comfy.ops.disable_weight_init + self.octree = OctreeProbabilityFixedlenDecoder(dtype=dtype, device=device, operations=operations) + self.gs = ElasticGaussianFixedlenDecoder(dtype=dtype, device=device, operations=operations) + + @property + def gaussians_per_point(self) -> int: + return self.gs.rep_config['num_gaussians'] + + def decode(self, latent: torch.Tensor, num_gaussians: int, level: int = None, generator=None): + # level defaults to the full octree depth, a lower level is cheaper (coarser) for live previews. + # generator (a CPU torch.Generator) makes the octree sampling reproducible without touching global RNG. + level = self._MAX_VOXEL_LEVEL if level is None else level + num_decoder_tokens = max(1, num_gaussians // self.gaussians_per_point) + points_pred = OctreeProbabilityFixedlenDecoder.sample( + self.octree, latent, num_points=num_decoder_tokens, level=level, temperature=1.0, generator=generator, + ) + pred = self.gs(x=points_pred, cond=latent) + return build_gaussian_models(self.gs, points_pred, pred) # one GaussianModel per batch item diff --git a/comfy/ldm/wan/model.py b/comfy/ldm/wan/model.py index 70dfe7b16..282408891 100644 --- a/comfy/ldm/wan/model.py +++ b/comfy/ldm/wan/model.py @@ -8,7 +8,7 @@ from einops import rearrange from comfy.ldm.modules.attention import optimized_attention from comfy.ldm.flux.layers import EmbedND -from comfy.ldm.flux.math import apply_rope1 +from comfy.ldm.flux.math import apply_rope1, rope import comfy.ldm.common_dit import comfy.model_management import comfy.patcher_extension @@ -570,6 +570,14 @@ class WanModel(torch.nn.Module): full_ref = self.ref_conv(full_ref).flatten(2).transpose(1, 2) x = torch.concat((full_ref, x), dim=1) + # In-context reference (Bernini) + context_latents = kwargs.get("context_latents", None) + main_len = x.shape[1] + if context_latents is not None: + for lat in context_latents: + cl = self.patch_embedding(lat.float().to(x.device)).to(x.dtype).flatten(2).transpose(1, 2) + x = torch.cat([x, cl], dim=1) + # context context = self.text_embedding(context) @@ -599,6 +607,9 @@ class WanModel(torch.nn.Module): # head x = self.head(x, e) + if context_latents is not None: + x = x[:, :main_len] + if full_ref is not None: x = x[:, full_ref.shape[1]:] @@ -606,7 +617,7 @@ class WanModel(torch.nn.Module): x = self.unpatchify(x, grid_sizes) return x - def rope_encode(self, t, h, w, t_start=0, steps_t=None, steps_h=None, steps_w=None, device=None, dtype=None, transformer_options={}): + def rope_encode(self, t, h, w, t_start=0, steps_t=None, steps_h=None, steps_w=None, device=None, dtype=None, transformer_options={}, source_id=0): patch_size = self.patch_size t_len = ((t + (patch_size[0] // 2)) // patch_size[0]) h_len = ((h + (patch_size[1] // 2)) // patch_size[1]) @@ -638,6 +649,13 @@ class WanModel(torch.nn.Module): img_ids = img_ids.reshape(1, -1, img_ids.shape[-1]) freqs = self.rope_embedder(img_ids).movedim(1, 2) + + # In-context reference: a non-zero source_id composes an extra rotation into the spatial rope + if source_id: + d = self.dim // self.num_heads + pos = torch.tensor([[float(source_id)]], device=freqs.device, dtype=torch.float32) + id_rot = rope(pos, d, self.rope_embedder.theta).reshape(1, 1, 1, d // 2, 2, 2).to(freqs.dtype) + freqs = torch.einsum('...ij,...jk->...ik', freqs, id_rot) return freqs def forward(self, x, timestep, context, clip_fea=None, time_dim_concat=None, transformer_options={}, **kwargs): @@ -661,6 +679,15 @@ class WanModel(torch.nn.Module): t_len += 1 freqs = self.rope_encode(t_len, h, w, device=x.device, dtype=x.dtype, transformer_options=transformer_options) + + # In-context reference: one rope block per stream, each with it's own source_id (1, 2, ...) to distinguish from the target (id 0). + context_latents = kwargs.get("context_latents", None) + if context_latents is not None: + context_latents = [comfy.ldm.common_dit.pad_to_patch_size(lat, self.patch_size) for lat in context_latents] + for i, lat in enumerate(context_latents): + freqs = torch.cat([freqs, self.rope_encode(lat.shape[-3], lat.shape[-2], lat.shape[-1], device=x.device, dtype=x.dtype, transformer_options=transformer_options, source_id=i + 1)], dim=1) + kwargs = {**kwargs, "context_latents": context_latents} + return self.forward_orig(x, timestep, context, clip_fea=clip_fea, freqs=freqs, transformer_options=transformer_options, **kwargs)[:, :, :t, :h, :w] def unpatchify(self, x, grid_sizes): @@ -1631,13 +1658,15 @@ class SCAILWanModel(WanModel): self.patch_embedding_pose = operations.Conv3d(in_dim, dim, kernel_size=patch_size, stride=patch_size, device=device, dtype=torch.float32) - def forward_orig(self, x, t, context, clip_fea=None, freqs=None, transformer_options={}, pose_latents=None, reference_latent=None, **kwargs): + def forward_orig(self, x, t, context, clip_fea=None, freqs=None, transformer_options={}, pose_latents=None, reference_latent=None, ref_mask_latents=None, sam_latents=None, **kwargs): if reference_latent is not None: x = torch.cat((reference_latent, x), dim=2) # embeddings x = self.patch_embedding(x.float()).to(x.dtype) + if ref_mask_latents is not None: # SCAIL-2 additive mask stream + x = x + self.patch_embedding_mask(ref_mask_latents.float()).to(x.dtype) grid_sizes = x.shape[2:] transformer_options["grid_sizes"] = grid_sizes x = x.flatten(2).transpose(1, 2) @@ -1645,6 +1674,8 @@ class SCAILWanModel(WanModel): scail_pose_seq_len = 0 if pose_latents is not None: scail_x = self.patch_embedding_pose(pose_latents.float()).to(x.dtype) + if sam_latents is not None: # SCAIL-2 additive mask stream + scail_x = scail_x + self.patch_embedding_mask(sam_latents.float()).to(x.dtype) scail_x = scail_x.flatten(2).transpose(1, 2) scail_pose_seq_len = scail_x.shape[1] x = torch.cat([x, scail_x], dim=1) @@ -1695,7 +1726,36 @@ class SCAILWanModel(WanModel): return x - def rope_encode(self, t, h, w, t_start=0, steps_t=None, steps_h=None, steps_w=None, device=None, dtype=None, pose_latents=None, reference_latent=None, transformer_options={}): + # ref_mask_flag is a scalar bool (CONDConstant, SCAIL-2 only). False => replacement mode, + # which places ref/pose via H/W rope shifts instead of the animation-mode temporal offset. + def rope_encode(self, t, h, w, t_start=0, steps_t=None, steps_h=None, steps_w=None, device=None, dtype=None, pose_latents=None, reference_latent=None, ref_mask_flag=None, transformer_options={}): + if ref_mask_flag is not None and not bool(ref_mask_flag): + REF_ROPE_H = 120.0 + POSE_ROPE_W = 120.0 + + ref_t_patches = 0 + if reference_latent is not None: + ref_t_patches = (reference_latent.shape[2] + (self.patch_size[0] // 2)) // self.patch_size[0] + main_t_patches = t - ref_t_patches + + parts = [] + if ref_t_patches > 0: + ref_tf = {"rope_options": {"shift_y": REF_ROPE_H, "shift_x": 0.0, "scale_y": 1.0, "scale_x": 1.0}} + parts.append(super().rope_encode(ref_t_patches, h, w, t_start=0, device=device, dtype=dtype, transformer_options=ref_tf)) + if main_t_patches > 0: + parts.append(super().rope_encode(main_t_patches, h, w, t_start=0, device=device, dtype=dtype, transformer_options=transformer_options)) + + if pose_latents is not None: + F_pose, H_pose, W_pose = pose_latents.shape[-3], pose_latents.shape[-2], pose_latents.shape[-1] + h_scale = h / H_pose + w_scale = w / W_pose + h_shift = (h_scale - 1) / 2 + w_shift = (w_scale - 1) / 2 + pose_tf = {"rope_options": {"shift_y": h_shift, "shift_x": POSE_ROPE_W + w_shift, "scale_y": h_scale, "scale_x": w_scale}} + parts.append(super().rope_encode(F_pose, H_pose, W_pose, t_start=0, device=device, dtype=dtype, transformer_options=pose_tf)) + + return torch.cat(parts, dim=1) + main_freqs = super().rope_encode(t, h, w, t_start=t_start, steps_t=steps_t, steps_h=steps_h, steps_w=steps_w, device=device, dtype=dtype, transformer_options=transformer_options) if pose_latents is None: @@ -1719,12 +1779,16 @@ class SCAILWanModel(WanModel): return torch.cat([main_freqs, pose_freqs], dim=1) - def _forward(self, x, timestep, context, clip_fea=None, time_dim_concat=None, transformer_options={}, pose_latents=None, **kwargs): + def _forward(self, x, timestep, context, clip_fea=None, time_dim_concat=None, transformer_options={}, pose_latents=None, ref_mask_latents=None, sam_latents=None, **kwargs): bs, c, t, h, w = x.shape x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size) if pose_latents is not None: pose_latents = comfy.ldm.common_dit.pad_to_patch_size(pose_latents, self.patch_size) + if ref_mask_latents is not None: # SCAIL-2 + ref_mask_latents = comfy.ldm.common_dit.pad_to_patch_size(ref_mask_latents, self.patch_size) + if sam_latents is not None: # SCAIL-2 + sam_latents = comfy.ldm.common_dit.pad_to_patch_size(sam_latents, self.patch_size) t_len = t if time_dim_concat is not None: @@ -1737,5 +1801,15 @@ class SCAILWanModel(WanModel): reference_latent = comfy.ldm.common_dit.pad_to_patch_size(kwargs.pop("reference_latent"), self.patch_size) t_len += reference_latent.shape[2] - freqs = self.rope_encode(t_len, h, w, device=x.device, dtype=x.dtype, transformer_options=transformer_options, pose_latents=pose_latents, reference_latent=reference_latent) - return self.forward_orig(x, timestep, context, clip_fea=clip_fea, freqs=freqs, transformer_options=transformer_options, pose_latents=pose_latents, reference_latent=reference_latent, **kwargs)[:, :, :t, :h, :w] + ref_mask_flag = kwargs.pop("ref_mask_flag", None) # SCAIL-2 + + freqs = self.rope_encode(t_len, h, w, device=x.device, dtype=x.dtype, transformer_options=transformer_options, pose_latents=pose_latents, reference_latent=reference_latent, ref_mask_flag=ref_mask_flag) + return self.forward_orig(x, timestep, context, clip_fea=clip_fea, freqs=freqs, transformer_options=transformer_options, pose_latents=pose_latents, reference_latent=reference_latent, ref_mask_latents=ref_mask_latents, sam_latents=sam_latents, **kwargs)[:, :, :t, :h, :w] + + +class SCAIL2WanModel(SCAILWanModel): + """SCAIL-2: SCAIL-Preview + an additive binary multi-identity mask stream.""" + + def __init__(self, model_type="scail2", patch_size=(1, 2, 2), in_dim=20, mask_in_dim=28, dim=5120, operations=None, device=None, dtype=None, **kwargs): + super().__init__(model_type=model_type, patch_size=patch_size, in_dim=in_dim, dim=dim, operations=operations, device=device, dtype=dtype, **kwargs) + self.patch_embedding_mask = operations.Conv3d(mask_in_dim, dim, kernel_size=patch_size, stride=patch_size, device=device, dtype=torch.float32) diff --git a/comfy/lora.py b/comfy/lora.py index c0e8b865c..2c8d0f0bf 100644 --- a/comfy/lora.py +++ b/comfy/lora.py @@ -16,7 +16,6 @@ along with this program. If not, see . """ -from __future__ import annotations import comfy.memory_management import comfy.utils import comfy.model_management @@ -358,6 +357,12 @@ def model_lora_keys_unet(model, key_map={}): key_lora = k[len("diffusion_model."):-len(".weight")] key_map["transformer.{}".format(key_lora)] = k + if isinstance(model, (comfy.model_base.LTXV, comfy.model_base.LTXAV)): + for k in sdk: + if k.startswith("diffusion_model.") and k.endswith(".weight"): + key_lora = k[len("diffusion_model."):-len(".weight")] + key_map["{}".format(key_lora)] = k + return key_map diff --git a/comfy/memory_management.py b/comfy/memory_management.py index c43f0c4a2..e032b7dcd 100644 --- a/comfy/memory_management.py +++ b/comfy/memory_management.py @@ -1,16 +1,16 @@ import math import ctypes -import threading import dataclasses import torch from typing import NamedTuple +import comfy_aimdo.host_buffer from comfy.quant_ops import QuantizedTensor class TensorFileSlice(NamedTuple): file_ref: object - thread_id: int + lock: object offset: int size: int @@ -18,21 +18,18 @@ class TensorFileSlice(NamedTuple): def read_tensor_file_slice_into(tensor, destination, stream=None, destination2=None): if isinstance(tensor, QuantizedTensor): - if not isinstance(destination, QuantizedTensor): - return False - if tensor._layout_cls != destination._layout_cls: - return False - - if not read_tensor_file_slice_into(tensor._qdata, destination._qdata, stream=stream, + if not read_tensor_file_slice_into(tensor._qdata, + destination._qdata if destination is not None else None, stream=stream, destination2=(destination2._qdata if destination2 is not None else None)): return False - dst_orig_dtype = destination._params.orig_dtype - destination._params.copy_from(tensor._params, non_blocking=False) - destination._params = dataclasses.replace(destination._params, orig_dtype=dst_orig_dtype) + if destination is not None: + dst_orig_dtype = destination._params.orig_dtype + destination._params.copy_from(tensor._params, non_blocking=False) + destination._params = dataclasses.replace(destination._params, orig_dtype=dst_orig_dtype) if destination2 is not None: dst_orig_dtype = destination2._params.orig_dtype - destination2._params.copy_from(destination._params, non_blocking=True) + destination2._params.copy_from(destination._params if destination is not None else tensor._params, non_blocking=True) destination2._params = dataclasses.replace(destination2._params, orig_dtype=dst_orig_dtype) return True @@ -40,11 +37,15 @@ def read_tensor_file_slice_into(tensor, destination, stream=None, destination2=N if info is None: return False + if destination is not None and destination.device.type != "cpu" and destination2 is None: + destination2 = destination + destination = None + file_obj = info.file_ref - if (destination.device.type != "cpu" - or file_obj is None - or threading.get_ident() != info.thread_id - or destination.numel() * destination.element_size() < info.size + if (file_obj is None + or (destination is None and destination2 is None) + or (destination is not None and (destination.device.type != "cpu" or destination.numel() * destination.element_size() < info.size)) + or (destination2 is not None and (destination2.device.type == "cpu" or destination2.numel() * destination2.element_size() < info.size)) or tensor.numel() * tensor.element_size() != info.size or tensor.storage_offset() != 0 or not tensor.is_contiguous()): @@ -53,31 +54,44 @@ def read_tensor_file_slice_into(tensor, destination, stream=None, destination2=N if info.size == 0: return True + if destination is None: + stream_ptr = getattr(stream, "cuda_stream", 0) if stream is not None else 0 + comfy_aimdo.host_buffer.read_file_to_device(file_obj, info.offset, info.size, + stream_ptr, destination2.data_ptr(), + destination2.device.index, + mark_cold=False) + return True + hostbuf = getattr(destination.untyped_storage(), "_comfy_hostbuf", None) if hostbuf is not None: stream_ptr = getattr(stream, "cuda_stream", 0) if stream is not None else 0 device_ptr = destination2.data_ptr() if destination2 is not None else 0 - hostbuf.read_file_slice(file_obj, info.offset, info.size, - offset=destination.data_ptr() - hostbuf.get_raw_address(), - stream=stream_ptr, - device_ptr=device_ptr, - device=None if destination2 is None else destination2.device.index) + with info.lock: + hostbuf.read_file_slice(file_obj, info.offset, info.size, + offset=destination.data_ptr() - hostbuf.get_raw_address(), + stream=stream_ptr, + device_ptr=device_ptr, + device=None if destination2 is None else destination2.device.index) return True + if not hasattr(file_obj, "seek") or not hasattr(file_obj, "readinto"): + return False + buf_type = ctypes.c_ubyte * info.size view = memoryview(buf_type.from_address(destination.data_ptr())) try: - file_obj.seek(info.offset) - done = 0 - while done < info.size: - try: - n = file_obj.readinto(view[done:]) - except OSError: - return False - if n <= 0: - return False - done += n + with info.lock: + file_obj.seek(info.offset) + done = 0 + while done < info.size: + try: + n = file_obj.readinto(view[done:]) + except OSError: + return False + if n <= 0: + return False + done += n return True finally: view.release() diff --git a/comfy/model_base.py b/comfy/model_base.py index 142ec530a..0e29cb9bb 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -35,6 +35,7 @@ import comfy.ldm.hydit.models import comfy.ldm.audio.dit import comfy.ldm.audio.embedders import comfy.ldm.flux.model +import comfy.ldm.lens.model import comfy.ldm.lightricks.model import comfy.ldm.hunyuan_video.model import comfy.ldm.cosmos.model @@ -45,12 +46,16 @@ import comfy.ldm.wan.model_animate import comfy.ldm.wan.ar_model import comfy.ldm.wan.model_wandancer import comfy.ldm.hunyuan3d.model +import comfy.ldm.triposplat.model import comfy.ldm.hidream.model import comfy.ldm.chroma.model import comfy.ldm.chroma_radiance.model +import comfy.ldm.pixeldit.model +import comfy.ldm.pixeldit.pid import comfy.ldm.ace.model import comfy.ldm.omnigen.omnigen2 import comfy.ldm.qwen_image.model +import comfy.ldm.ideogram4.model import comfy.ldm.kandinsky5.model import comfy.ldm.anima.model import comfy.ldm.trellis2.model @@ -61,6 +66,7 @@ import comfy.ldm.ernie.model import comfy.ldm.sam3.detector import comfy.ldm.hidream_o1.model from comfy.ldm.hidream_o1.conditioning import build_extra_conds +import comfy.ldm.depth_anything_3.model import comfy.model_management import comfy.patcher_extension @@ -1059,6 +1065,27 @@ class Flux2(Flux): out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) return out + +class Lens(BaseModel): + def __init__(self, model_config, model_type=ModelType.FLUX, device=None): + super().__init__( + model_config, model_type, device=device, + unet_model=comfy.ldm.lens.model.LensTransformer2DModel, + ) + + def encode_adm(self, **kwargs): + return None # Lens has no pooled/ADM conditioning. + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + cross_attn = kwargs.get("cross_attn", None) + if cross_attn is not None: + out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) + attention_mask = kwargs.get("attention_mask", None) + if attention_mask is not None: + out['attention_mask'] = comfy.conds.CONDRegular(attention_mask) + return out + class GenmoMochi(BaseModel): def __init__(self, model_config, model_type=ModelType.FLOW, device=None): super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.genmo.joint_model.asymm_models_joint.AsymmDiTJoint) @@ -1376,6 +1403,53 @@ class ZImagePixelSpace(Lumina2): BaseModel.__init__(self, model_config, model_type, device=device, unet_model=comfy.ldm.lumina.model.NextDiTPixelSpace) self.memory_usage_factor_conds = ("ref_latents",) + +class PixelDiTT2I(BaseModel): + def __init__(self, model_config, model_type=ModelType.FLOW, device=None): + super().__init__(model_config, model_type, device=device, + unet_model=comfy.ldm.pixeldit.model.PixDiT_T2I) + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + attention_mask = kwargs.get("attention_mask", None) + if attention_mask is not None: + out["attention_mask"] = comfy.conds.CONDRegular(attention_mask) + return out + + +class PiD(PixelDiTT2I): + def __init__(self, model_config, model_type=ModelType.FLOW, device=None): + BaseModel.__init__(self, model_config, model_type, device=device, + unet_model=comfy.ldm.pixeldit.pid.PidNet) + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + lq_latent = kwargs.get("lq_latent", None) + if lq_latent is not None: + out["lq_latent"] = comfy.conds.CONDRegular(lq_latent) + degrade_sigma = kwargs.get("degrade_sigma", None) + if degrade_sigma is not None: + out["degrade_sigma"] = comfy.conds.CONDRegular(degrade_sigma) + return out + + def resize_cond_for_context_window(self, cond_key, cond_value, window, x_in, device, retain_index_list=[]): + if cond_key == "lq_latent" and hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor): + lq = cond_value.cond + dim = window.dim + if dim >= lq.ndim: + return None + lq_proj = self.diffusion_model.lq_proj + ratio = lq_proj.sr_scale * lq_proj.latent_spatial_down_factor + # Map x window indices -> lq indices (deduplicated, sorted, in-bounds). + lq_size = lq.size(dim) + lq_indices = sorted({i // ratio for i in window.index_list if 0 <= i // ratio < lq_size}) + if not lq_indices: + return None + idx = tuple([slice(None)] * dim + [lq_indices]) + return cond_value._copy_with(lq[idx].to(device)) + return super().resize_cond_for_context_window(cond_key, cond_value, window, x_in, device, retain_index_list=retain_index_list) + + class WAN21(BaseModel): def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None): super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel) @@ -1446,8 +1520,26 @@ class WAN21(BaseModel): if reference_latents is not None: out['reference_latent'] = comfy.conds.CONDRegular(self.process_latent_in(reference_latents[-1])[:, :, 0]) + # In-context reference conditioning (Bernini) + context_latents = kwargs.get("context_latents", None) + if context_latents is not None: + out['context_latents'] = comfy.conds.CONDList([self.process_latent_in(l) for l in context_latents]) + return out + def resize_cond_for_context_window(self, cond_key, cond_value, window, x_in, device, retain_index_list=[]): + # In-context cond slicing (Bernini) + if cond_key == "context_latents" and isinstance(getattr(cond_value, "cond", None), list): + dim = window.dim + out = [] + for lat in cond_value.cond: + if lat.ndim > dim and lat.shape[dim] > 1 and lat.shape[dim] == x_in.shape[dim]: + out.append(window.get_tensor(lat, device, dim=dim, retain_index_list=retain_index_list)) + else: + out.append(lat.to(device)) + return cond_value._copy_with(out) + return super().resize_cond_for_context_window(cond_key, cond_value, window, x_in, device, retain_index_list=retain_index_list) + class WAN21_CausalAR(WAN21): def __init__(self, model_config, model_type=ModelType.FLOW, device=None): @@ -1702,6 +1794,80 @@ class WAN21_SCAIL(WAN21): out['pose_latents'] = [pose_latents.shape[0], 20, *pose_latents.shape[2:]] return out +class WAN21_SCAIL2(WAN21_SCAIL): + """SCAIL-2: SCAIL-Preview + an additive binary multi-identity mask stream.""" + + def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None): + super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.SCAIL2WanModel) + self.memory_usage_factor_conds = ("reference_latent", "pose_latents", "ref_mask_latents", "sam_latents") + self.memory_usage_shape_process = { + "pose_latents": lambda shape: [shape[0], shape[1], 1.5, shape[-2], shape[-1]], + "sam_latents": lambda shape: [shape[0], shape[1], 1.5, shape[-2], shape[-1]], + } + self.image_to_video = image_to_video + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + + driving_mask_28ch = kwargs.get("driving_mask_28ch", None) + if driving_mask_28ch is not None: + out['sam_latents'] = comfy.conds.CONDRegular(driving_mask_28ch.movedim(1, 2).contiguous()) + + ref_mask_28ch = kwargs.get("ref_mask_28ch", None) + if ref_mask_28ch is not None: + out['ref_mask_latents'] = comfy.conds.CONDRegular(ref_mask_28ch.movedim(1, 2).contiguous()) + + ref_mask_flag = kwargs.get("ref_mask_flag", None) + if ref_mask_flag is not None: + out['ref_mask_flag'] = comfy.conds.CONDConstant(ref_mask_flag) + + return out + + def extra_conds_shapes(self, **kwargs): + out = super().extra_conds_shapes(**kwargs) + driving_mask_28ch = kwargs.get("driving_mask_28ch", None) + if driving_mask_28ch is not None: + s = driving_mask_28ch.shape + out['sam_latents'] = [s[0], 28, s[1], s[3], s[4]] + ref_mask_28ch = kwargs.get("ref_mask_28ch", None) + if ref_mask_28ch is not None: + s = ref_mask_28ch.shape + out['ref_mask_latents'] = [s[0], 28, s[1], s[3], s[4]] + return out + + def resize_cond_for_context_window(self, cond_key, cond_value, window, x_in, device, retain_index_list=[]): + if cond_key in ("sam_latents", "pose_latents"): + return comfy.context_windows.slice_cond(cond_value, window, x_in, device, temporal_dim=2, temporal_offset=1) + return super().resize_cond_for_context_window(cond_key, cond_value, window, x_in, device, retain_index_list=retain_index_list) + + def concat_cond(self, **kwargs): + # The 4 extra channels are the history_mask (1 at clean-anchor frames). + noise = kwargs.get("noise", None) + extra_channels = self.diffusion_model.patch_embedding.weight.shape[1] - noise.shape[1] + if extra_channels != 4: + return super().concat_cond(**kwargs) + + mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None)) + if mask is None: + return torch.zeros_like(noise)[:, :4] + + device = kwargs["device"] + if mask.shape[1] != 4: + mask = torch.mean(mask, dim=1, keepdim=True) + mask = 1.0 - mask + mask = utils.common_upscale(mask.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center") + if mask.shape[-3] < noise.shape[-3]: + mask = torch.nn.functional.pad(mask, (0, 0, 0, 0, 0, noise.shape[-3] - mask.shape[-3]), mode='constant', value=0) + if mask.shape[1] == 1: + mask = mask.repeat(1, 4, 1, 1, 1) + mask = utils.resize_to_batch_size(mask, noise.shape[0]) + return mask + + def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs): + # Hold anchor constant across all sigmas instead of base sigma*noise + (1-sigma)*latent_image. + return latent_image + + class WAN22_WanDancer(WAN21): def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=True, device=None): super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model_wandancer.WanDancerModel) @@ -1756,6 +1922,24 @@ class Hunyuan3Dv2_1(BaseModel): out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([guidance])) return out +class TripoSplat(BaseModel): + def __init__(self, model_config, model_type=ModelType.FLOW, device=None): + super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.triposplat.model.LatentSeqMMFlowModel) + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + cross_attn = kwargs.get("cross_attn", None) # DINOv3 token sequence -> cross-attention context. + if cross_attn is not None: + out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) + ref_latents = kwargs.get("reference_latents", None) # Flux2 VAE image latent -> additive second conditioning. + if ref_latents is not None: + out['ref_latents'] = comfy.conds.CONDList(list(ref_latents)) + latent_shapes = kwargs.get("latent_shapes", None) # {latent, camera} nested latent + if latent_shapes is not None: + out['latent_shapes'] = comfy.conds.CONDConstant(latent_shapes) + return out + + class HiDream(BaseModel): def __init__(self, model_config, model_type=ModelType.FLOW, device=None): super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hidream.model.HiDreamImageTransformer2DModel) @@ -1949,6 +2133,21 @@ class QwenImage(BaseModel): out['ref_latents'] = list([1, 16, sum(map(lambda a: math.prod(a.size()), ref_latents)) // 16]) return out +class Ideogram4(BaseModel): + def __init__(self, model_config, model_type=ModelType.FLOW, device=None): + super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.ideogram4.model.Ideogram4Transformer2DModel) + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + attention_mask = kwargs.get("attention_mask", None) + if attention_mask is not None: + if torch.numel(attention_mask) != attention_mask.sum(): + out['attention_mask'] = comfy.conds.CONDRegular(attention_mask) + cross_attn = kwargs.get("cross_attn", None) + if cross_attn is not None: + out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) + return out + class HunyuanImage21(BaseModel): def __init__(self, model_config, model_type=ModelType.FLOW, device=None): super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hunyuan_video.model.HunyuanVideo) @@ -2142,6 +2341,12 @@ class RT_DETR_v4(BaseModel): def __init__(self, model_config, model_type=ModelType.FLOW, device=None): super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.rt_detr.rtdetr_v4.RTv4) + +class DepthAnything3(BaseModel): + def __init__(self, model_config, model_type=ModelType.FLOW, device=None): + super().__init__(model_config, model_type, device=device, + unet_model=comfy.ldm.depth_anything_3.model.DepthAnything3Net) + class ErnieImage(BaseModel): def __init__(self, model_config, model_type=ModelType.FLOW, device=None): super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.ernie.model.ErnieImageModel) diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 327134d68..6c5d85fc8 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -351,6 +351,10 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["use_x0"] = True else: dit_config["use_x0"] = False + if "{}__sequential__".format(key_prefix) in state_dict_keys: # sequential txt_ids + dit_config["use_sequential_txt_ids"] = True + else: + dit_config["use_sequential_txt_ids"] = False else: dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys dit_config["yak_mlp"] = '{}double_blocks.0.img_mlp.gate_proj.weight'.format(key_prefix) in state_dict_keys @@ -501,6 +505,23 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["extra_per_block_abs_pos_emb_type"] = "learnable" return dit_config + # PiD (Pixel Diffusion Decoder). Must check BEFORE plain PixelDiT_T2I. + _lq_w_key = '{}lq_proj.latent_proj.0.weight'.format(key_prefix) + if _lq_w_key in state_dict_keys: + in_ch = int(state_dict[_lq_w_key].shape[1]) + _gate_prefix = '{}lq_proj.gate_modules.'.format(key_prefix) + num_gates = len({k[len(_gate_prefix):].split('.')[0] + for k in state_dict_keys if k.startswith(_gate_prefix)}) + dit_config = {"image_model": "pid", + "lq_latent_channels": in_ch, + "latent_spatial_down_factor": 16 if in_ch >= 64 else 8} + if num_gates > 0: + dit_config["lq_interval"] = (14 + num_gates - 1) // num_gates + return dit_config + + if '{}core.pixel_embedder.proj.weight'.format(key_prefix) in state_dict_keys: # PixelDiT T2I + return {"image_model": "pixeldit_t2i"} + if '{}cap_embedder.1.weight'.format(key_prefix) in state_dict_keys and '{}noise_refiner.0.attention.k_norm.weight'.format(key_prefix) in state_dict_keys: # Lumina 2 dit_config = {} dit_config["image_model"] = "lumina2" @@ -647,6 +668,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["model_type"] = "humo" elif '{}face_adapter.fuser_blocks.0.k_norm.weight'.format(key_prefix) in state_dict_keys: dit_config["model_type"] = "animate" + elif '{}patch_embedding_mask.weight'.format(key_prefix) in state_dict_keys: + dit_config["model_type"] = "scail2" elif '{}patch_embedding_pose.weight'.format(key_prefix) in state_dict_keys: dit_config["model_type"] = "scail" elif '{}patch_embedding_global.weight'.format(key_prefix) in state_dict_keys: @@ -697,6 +720,9 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["guidance_cond_proj_dim"] = None#f"{key_prefix}t_embedder.cond_proj.weight" in state_dict_keys return dit_config + if '{}cam_out_layer.weight'.format(key_prefix) in state_dict_keys and '{}repo_layers.0.final_map.weight'.format(key_prefix) in state_dict_keys: # TripoSplat + return {"image_model": "triposplat"} + if '{}t_embedder1.mlp.0.weight'.format(key_prefix) in state_dict_keys and '{}x_embedder.proj1.weight'.format(key_prefix) in state_dict_keys: # HiDream-O1 return {"image_model": "hidream_o1"} @@ -793,6 +819,30 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["timestep_scale"] = 1000.0 return dit_config + if '{}transformer_blocks.0.attn.norm_added_q.weight'.format(key_prefix) in state_dict_keys \ + and '{}transformer_blocks.0.img_mlp.w1.weight'.format(key_prefix) in state_dict_keys: # Lens + img_in_w = state_dict['{}img_in.weight'.format(key_prefix)] + proj_out_w = state_dict['{}proj_out.weight'.format(key_prefix)] + multi_layer = '{}txt_norm.0.weight'.format(key_prefix) in state_dict_keys + if multi_layer: + enc_hidden_dim = state_dict['{}txt_norm.0.weight'.format(key_prefix)].shape[0] + # Indices are TE-side; the DiT just consumes L layers in order. + selected_layer_index = tuple(range(count_blocks(state_dict_keys, '{}txt_norm.'.format(key_prefix) + '{}.'))) + else: + enc_hidden_dim = state_dict['{}txt_norm.weight'.format(key_prefix)].shape[0] + selected_layer_index = (0,) + + return { + "image_model": "lens", + "in_channels": img_in_w.shape[1], + "out_channels": proj_out_w.shape[0] // 4, # patch_size ** 2 (=2² default) + "num_layers": count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.'), + "num_attention_heads": img_in_w.shape[0] // 64, # // attention_head_dim default + "enc_hidden_dim": enc_hidden_dim, + "multi_layer_encoder_feature": multi_layer, + "selected_layer_index": selected_layer_index, + } + if '{}txt_norm.weight'.format(key_prefix) in state_dict_keys: # Qwen Image dit_config = {} dit_config["image_model"] = "qwen_image" @@ -805,6 +855,13 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["default_ref_method"] = "negative_index" return dit_config + if '{}embed_image_indicator.weight'.format(key_prefix) in state_dict_keys: # Ideogram 4 + dit_config = {} + dit_config["image_model"] = "ideogram4" + dit_config["in_channels"] = state_dict['{}input_proj.weight'.format(key_prefix)].shape[1] + dit_config["num_layers"] = count_blocks(state_dict_keys, '{}layers.'.format(key_prefix) + '{}.') + return dit_config + if '{}visual_transformer_blocks.0.cross_attention.key_norm.weight'.format(key_prefix) in state_dict_keys: # Kandinsky 5 dit_config = {} model_dim = state_dict['{}visual_embeddings.in_layer.bias'.format(key_prefix)].shape[0] @@ -843,6 +900,95 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["enc_h"] = state_dict['{}encoder.pan_blocks.1.cv4.conv.weight'.format(key_prefix)].shape[0] return dit_config + # Depth Anything 3 (repackaged to ComfyUI's native Dinov2Model layout via scripts/convert_da3.py) + if '{}backbone.embeddings.patch_embeddings.projection.weight'.format(key_prefix) in state_dict_keys: + dit_config = {} + dit_config["image_model"] = "DepthAnything3" + + patch_w = state_dict['{}backbone.embeddings.patch_embeddings.projection.weight'.format(key_prefix)] + embed_dim = patch_w.shape[0] + depth = count_blocks(state_dict_keys, '{}backbone.encoder.layer.'.format(key_prefix) + '{}.') + + # Backbone preset is determined by embed_dim (matches vits/vitb/vitl/vitg). + backbone_name = {384: "vits", 768: "vitb", 1024: "vitl", 1536: "vitg"}.get(embed_dim) + if backbone_name is None: + return None + dit_config["backbone_name"] = backbone_name + + # Detect DA3 extensions on top of vanilla DINOv2. + has_camera_token = '{}backbone.embeddings.camera_token'.format(key_prefix) in state_dict_keys + # qk-norm shows up as `attention.q_norm.weight` on enabled blocks. + qknorm_indices = [ + i for i in range(depth) + if '{}backbone.encoder.layer.{}.attention.q_norm.weight'.format(key_prefix, i) in state_dict_keys + ] + qknorm_start = qknorm_indices[0] if qknorm_indices else -1 + + # The DA3 main-series configs always set alt_start == qknorm_start == rope_start. + # cat_token=True is implied by the presence of camera_token. + if has_camera_token: + dit_config["alt_start"] = qknorm_start + dit_config["rope_start"] = qknorm_start + dit_config["qknorm_start"] = qknorm_start + dit_config["cat_token"] = True + else: + dit_config["alt_start"] = -1 + dit_config["rope_start"] = -1 + dit_config["qknorm_start"] = -1 + dit_config["cat_token"] = False + + # Detect head type and config. + has_aux = '{}head.scratch.refinenet1_aux.out_conv.weight'.format(key_prefix) in state_dict_keys + dit_config["head_dim_in"] = state_dict['{}head.projects.0.weight'.format(key_prefix)].shape[1] + dit_config["head_features"] = state_dict['{}head.scratch.refinenet1.out_conv.weight'.format(key_prefix)].shape[0] + dit_config["head_out_channels"] = [ + state_dict['{}head.projects.{}.weight'.format(key_prefix, i)].shape[0] + for i in range(4) + ] + if has_aux: + # DualDPT: dim_in = 2 * embed_dim (because cat_token doubles token width). + dit_config["head_type"] = "dualdpt" + dit_config["head_output_dim"] = 2 + dit_config["head_use_sky_head"] = False + else: + dit_config["head_type"] = "dpt" + dit_config["head_output_dim"] = state_dict[ + '{}head.scratch.output_conv2.2.weight'.format(key_prefix) + ].shape[0] + dit_config["head_use_sky_head"] = ( + '{}head.scratch.sky_output_conv2.0.weight'.format(key_prefix) in state_dict_keys + ) + + # out_layers: hard-coded per upstream YAML config (depth-aware default). + if depth >= 24: + # vitl: depths used vary between DA3-Large (DualDPT) and Mono/Metric (DPT). + if has_aux: + dit_config["out_layers"] = [11, 15, 19, 23] + else: + dit_config["out_layers"] = [4, 11, 17, 23] + else: + # vits/vitb: 12 blocks + dit_config["out_layers"] = [5, 7, 9, 11] + + # Camera encoder/decoder presence (multi-view + pose path). + has_cam_enc = '{}cam_enc.token_norm.weight'.format(key_prefix) in state_dict_keys + has_cam_dec = '{}cam_dec.fc_t.weight'.format(key_prefix) in state_dict_keys + dit_config["has_cam_enc"] = has_cam_enc + dit_config["has_cam_dec"] = has_cam_dec + if has_cam_enc: + cam_enc_w = state_dict.get( + '{}cam_enc.pose_branch.fc2.weight'.format(key_prefix) + ) + if cam_enc_w is not None: + dit_config["cam_dim_out"] = cam_enc_w.shape[0] + if has_cam_dec: + cam_dec_w = state_dict.get( + '{}cam_dec.fc_t.weight'.format(key_prefix) + ) + if cam_dec_w is not None: + dit_config["cam_dec_dim_in"] = cam_dec_w.shape[1] + return dit_config + if '{}layers.0.mlp.linear_fc2.weight'.format(key_prefix) in state_dict_keys: # Ernie Image dit_config = {} dit_config["image_model"] = "ernie" diff --git a/comfy/model_management.py b/comfy/model_management.py index 3894dfa9c..9dc0a4e13 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -15,6 +15,7 @@ You should have received a copy of the GNU General Public License along with this program. If not, see . """ +from __future__ import annotations import psutil import logging @@ -27,13 +28,18 @@ import platform import weakref import gc import os -from contextlib import nullcontext +from contextlib import contextmanager, nullcontext import comfy.memory_management import comfy.utils import comfy.quant_ops import comfy_aimdo.host_buffer import comfy_aimdo.vram_buffer +from typing import TYPE_CHECKING +if TYPE_CHECKING: + from comfy.model_patcher import ModelPatcher + + class VRAMState(Enum): DISABLED = 0 #No vram present: no need to move models to vram NO_VRAM = 1 #Very low vram: enable all the options to save vram @@ -204,6 +210,107 @@ def get_torch_device(): else: return torch.device(torch.cuda.current_device()) +def get_all_torch_devices(exclude_current=False): + global cpu_state + devices = [] + if cpu_state == CPUState.GPU: + # NVIDIA + AMD/ROCm both expose their GPUs through torch.cuda.*; + # without the AMD arm, single-GPU ROCm users get an empty list + # which silently turns unload_all_models() into a no-op. + if is_nvidia() or is_amd(): + for i in range(torch.cuda.device_count()): + devices.append(torch.device("cuda", i)) + elif is_intel_xpu(): + for i in range(torch.xpu.device_count()): + devices.append(torch.device("xpu", i)) + elif is_ascend_npu(): + for i in range(torch.npu.device_count()): + devices.append(torch.device("npu", i)) + elif is_mlu(): + for i in range(torch.mlu.device_count()): + devices.append(torch.device("mlu", i)) + else: + # Fallback for unhandled GPU backends (e.g. DirectML): at least + # report the current device so callers like unload_all_models() + # do not silently no-op. + devices.append(get_torch_device()) + else: + devices.append(get_torch_device()) + if exclude_current: + current = get_torch_device() + if current in devices: + devices.remove(current) + return devices + +def get_gpu_device_options(): + """Return list of device option strings for node widgets. + + Always includes "default" and "cpu". When multiple GPUs are present, + adds "gpu:0", "gpu:1", etc. (vendor-agnostic labels). + """ + options = ["default", "cpu"] + devices = get_all_torch_devices() + if len(devices) > 1: + for i in range(len(devices)): + options.append(f"gpu:{i}") + return options + +def get_gpu_device_options_no_cpu(): + """Variant of get_gpu_device_options that omits "cpu". + + Intended for components like the VAE selector where running on CPU + is impractical and should not be offered as a choice. + """ + return [o for o in get_gpu_device_options() if o != "cpu"] + +def resolve_gpu_device_option(option: str): + """Resolve a device option string to a torch.device. + + Returns None for "default" (let the caller use its normal default). + Returns torch.device("cpu") for "cpu". + For "gpu:N", returns the Nth torch device. Returns None if the + index is out of range, the option string is malformed, or + unrecognized (callers are expected to log their own context-rich + message before falling back to the default device). + """ + if option is None or option == "default": + return None + if option == "cpu": + return torch.device("cpu") + if option.startswith("gpu:"): + try: + idx = int(option[4:]) + except ValueError: + return None + devices = get_all_torch_devices() + if 0 <= idx < len(devices): + return devices[idx] + return None + +@contextmanager +def cuda_device_context(device): + """Context manager that sets torch.cuda.current_device to match *device*. + + Used when running operations on a non-default CUDA device so that custom + CUDA kernels (e.g. comfy_kitchen fp8 quantization) pick up the correct + device index. The previous device is restored on exit. + + No-op when *device* is not CUDA, has no explicit index, or already matches + the current device. + """ + prev = None + if device.type == "cuda" and device.index is not None: + prev = torch.cuda.current_device() + if prev != device.index: + torch.cuda.set_device(device) + else: + prev = None + try: + yield + finally: + if prev is not None: + torch.cuda.set_device(prev) + def get_total_memory(dev=None, torch_total_too=False): global directml_enabled if dev is None: @@ -492,9 +599,13 @@ try: logging.info("Device: {}".format(get_torch_device_name(get_torch_device()))) except: logging.warning("Could not pick default device.") +try: + for device in get_all_torch_devices(exclude_current=True): + logging.info("Device: {}".format(get_torch_device_name(device))) +except: + pass - -current_loaded_models = [] +current_loaded_models: list[LoadedModel] = [] DIRTY_MMAPS = set() @@ -530,15 +641,17 @@ def free_pins(size, evict_active=False): return freed_total def ensure_pin_budget(size, evict_active=False): - shortfall = size + comfy.memory_management.RAM_CACHE_HEADROOM / 2 - psutil.virtual_memory().available + if args.fast_disk: + shortfall = TOTAL_PINNED_MEMORY + size - MAX_PINNED_MEMORY + else: + shortfall = size + max(comfy.memory_management.RAM_CACHE_HEADROOM / 2, 2048 * 1024 ** 2) - psutil.virtual_memory().available if shortfall <= 0: return True to_free = shortfall + PIN_PRESSURE_HYSTERESIS return free_pins(to_free, evict_active=evict_active) >= shortfall -def ensure_pin_registerable(size, evict_active=False): - shortfall = TOTAL_PINNED_MEMORY + size - MAX_PINNED_MEMORY +def free_registrations(shortfall, evict_active=True): if MAX_PINNED_MEMORY <= 0: return False if shortfall <= 0: @@ -547,14 +660,24 @@ def ensure_pin_registerable(size, evict_active=False): shortfall += REGISTERABLE_PIN_HYSTERESIS for loaded_model in reversed(current_loaded_models): model = loaded_model.model - if model is not None and model.is_dynamic() and (evict_active or not model.model.dynamic_pins[model.load_device]["active"]): + if model is not None and model.is_dynamic() and not model.model.dynamic_pins[model.load_device]["active"]: shortfall -= model.unregister_inactive_pins(shortfall) if shortfall <= 0: return True + if evict_active: + for loaded_model in current_loaded_models: + model = loaded_model.model + if model is not None and model.is_dynamic() and model.model.dynamic_pins[model.load_device]["active"]: + shortfall -= model.unregister_inactive_pins(shortfall) + if shortfall <= 0: + return True return shortfall <= REGISTERABLE_PIN_HYSTERESIS +def ensure_pin_registerable(size, evict_active=True): + return free_registrations(TOTAL_PINNED_MEMORY + size - MAX_PINNED_MEMORY, evict_active=evict_active) + class LoadedModel: - def __init__(self, model): + def __init__(self, model: ModelPatcher): self._set_model(model) self.device = model.load_device self.real_model = None @@ -562,7 +685,7 @@ class LoadedModel: self.model_finalizer = None self._patcher_finalizer = None - def _set_model(self, model): + def _set_model(self, model: ModelPatcher): self._model = weakref.ref(model) if model.parent is not None: self._parent_model = weakref.ref(model.parent) @@ -573,6 +696,7 @@ class LoadedModel: model = self._parent_model() if model is not None: self._set_model(model) + self.device = model.load_device @property def model(self): @@ -691,9 +815,9 @@ def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, pins for x in can_unload_sorted: i = x[-1] memory_to_free = 1e32 - if current_loaded_models[i].model.is_dynamic() and (not DISABLE_SMART_MEMORY or device is None): + if not DISABLE_SMART_MEMORY or device is None: memory_to_free = 0 if device is None else memory_required - get_free_memory(device) - if for_dynamic: + if current_loaded_models[i].model.is_dynamic() and for_dynamic: #don't actually unload dynamic models for the sake of other dynamic models #as that works on-demand. memory_required -= current_loaded_models[i].model.loaded_size() @@ -705,6 +829,10 @@ def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, pins for i in sorted(unloaded_model, reverse=True): unloaded_models.append(current_loaded_models.pop(i)) + if not for_dynamic and pins_required > 0: + ensure_pin_budget(pins_required) + ensure_pin_registerable(pins_required) + if len(unloaded_model) > 0: soft_empty_cache() elif device is not None: @@ -767,15 +895,19 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu model_to_unload.model_finalizer.detach() total_memory_required = {} + total_pins_required = {} for loaded_model in models_to_load: device = loaded_model.device total_memory_required[device] = total_memory_required.get(device, 0) + loaded_model.model_memory_required(device) + if not loaded_model.model.is_dynamic(): + total_pins_required[device] = total_pins_required.get(device, 0) + loaded_model.model_memory() for device in total_memory_required: if device != torch.device("cpu"): free_memory(total_memory_required[device] * 1.1 + extra_mem, device, - for_dynamic=free_for_dynamic) + for_dynamic=free_for_dynamic, + pins_required=total_pins_required.get(device, 0)) for device in total_memory_required: if device != torch.device("cpu"): @@ -826,8 +958,6 @@ def loaded_models(only_currently_used=False): def cleanup_models_gc(): do_gc = False - reset_cast_buffers() - for i in range(len(current_loaded_models)): cur = current_loaded_models[i] if cur.is_dead(): @@ -1171,7 +1301,6 @@ STREAM_CAST_BUFFERS = {} LARGEST_CASTED_WEIGHT = (None, 0) STREAM_AIMDO_CAST_BUFFERS = {} LARGEST_AIMDO_CASTED_WEIGHT = (None, 0) -STREAM_PIN_BUFFERS = {} DEFAULT_AIMDO_CAST_BUFFER_RESERVATION_SIZE = 16 * 1024 ** 3 @@ -1214,42 +1343,13 @@ def get_aimdo_cast_buffer(offload_stream, device): STREAM_AIMDO_CAST_BUFFERS[offload_stream] = cast_buffer return cast_buffer -def get_pin_buffer(offload_stream): - pin_buffer = STREAM_PIN_BUFFERS.get(offload_stream, None) - if pin_buffer is None: - pin_buffer = comfy_aimdo.host_buffer.HostBuffer(0, 0, pinned_hostbuf_size(8 * 1024**3)) - STREAM_PIN_BUFFERS[offload_stream] = pin_buffer - elif offload_stream is not None: - event = getattr(pin_buffer, "_comfy_event", None) - if event is not None: - event.synchronize() - delattr(pin_buffer, "_comfy_event") - return pin_buffer - -def resize_pin_buffer(pin_buffer, size): - global TOTAL_PINNED_MEMORY - old_size = pin_buffer.size - if size <= old_size: - return True - growth = size - old_size - comfy.memory_management.extra_ram_release(comfy.memory_management.RAM_CACHE_HEADROOM) - ensure_pin_budget(growth, evict_active=True) - ensure_pin_registerable(growth, evict_active=True) - try: - pin_buffer.extend(size=size, reallocate=True) - except RuntimeError: - return False - TOTAL_PINNED_MEMORY += pin_buffer.size - old_size - return True - def reset_cast_buffers(): - global TOTAL_PINNED_MEMORY global LARGEST_CASTED_WEIGHT global LARGEST_AIMDO_CASTED_WEIGHT LARGEST_CASTED_WEIGHT = (None, 0) LARGEST_AIMDO_CASTED_WEIGHT = (None, 0) - for offload_stream in set(STREAM_CAST_BUFFERS) | set(STREAM_AIMDO_CAST_BUFFERS) | set(STREAM_PIN_BUFFERS): + for offload_stream in set(STREAM_CAST_BUFFERS) | set(STREAM_AIMDO_CAST_BUFFERS): if offload_stream is not None: offload_stream.synchronize() synchronize() @@ -1258,20 +1358,24 @@ def reset_cast_buffers(): mmap_obj.bounce() DIRTY_MMAPS.clear() - for pin_buffer in STREAM_PIN_BUFFERS.values(): - TOTAL_PINNED_MEMORY -= pin_buffer.size - TOTAL_PINNED_MEMORY = max(0, TOTAL_PINNED_MEMORY) - for loaded_model in current_loaded_models: model = loaded_model.model if model is not None and model.is_dynamic(): - model.model.dynamic_pins[model.load_device]["active"] = False + pin_state = model.model.dynamic_pins[model.load_device] + + if pin_state["active"]: + *_, buckets = pin_state["weights"] + for size, bucket in list(buckets.items()): + bucket[:] = [ entry for entry in bucket if entry[-1] is not None ] + if not bucket: + del buckets[size] + + pin_state["active"] = False model.partially_unload_ram(1e30, subsets=[ "patches" ]) - model.model.dynamic_pins[model.load_device]["patches"] = (comfy_aimdo.host_buffer.HostBuffer(0, 8 * 1024 * 1024, pinned_hostbuf_size(model.model_size())), [], [-1], [0]) + model.model.dynamic_pins[model.load_device]["patches"] = (comfy_aimdo.host_buffer.HostBuffer(0, 8 * 1024 * 1024, pinned_hostbuf_size(model.model_size())), [], [-1], [0], [0], {}) STREAM_CAST_BUFFERS.clear() STREAM_AIMDO_CAST_BUFFERS.clear() - STREAM_PIN_BUFFERS.clear() soft_empty_cache() def get_offload_stream(device): @@ -1324,7 +1428,7 @@ def cast_to_gathered(tensors, r, non_blocking=False, stream=None, r2=None): if hasattr(wf_context, "as_context"): wf_context = wf_context.as_context(stream) - dest_views = comfy.memory_management.interpret_gathered_like(tensors, r) + dest_views = comfy.memory_management.interpret_gathered_like(tensors, r) if r is not None else [None] * len(tensors) dest2_views = comfy.memory_management.interpret_gathered_like(tensors, r2) if r2 is not None else None with wf_context: for tensor in tensors: @@ -1336,9 +1440,10 @@ def cast_to_gathered(tensors, r, non_blocking=False, stream=None, r2=None): continue storage = tensor._qdata.untyped_storage() if isinstance(tensor, comfy.quant_ops.QuantizedTensor) else tensor.untyped_storage() mark_mmap_dirty(storage) - dest_view.copy_(tensor, non_blocking=non_blocking) + if dest_view is not None: + dest_view.copy_(tensor, non_blocking=non_blocking) if dest2_view is not None: - dest2_view.copy_(dest_view, non_blocking=non_blocking) + dest2_view.copy_(tensor if dest_view is None else dest_view, non_blocking=non_blocking) def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, stream=None, r=None): @@ -1611,6 +1716,13 @@ def is_device_xpu(device): def is_device_cuda(device): return is_device_type(device, 'cuda') +def set_torch_device(device): + """Set the current device for the given torch device. Supports CUDA and XPU.""" + if is_device_cuda(device): + torch.cuda.set_device(device) + elif is_device_xpu(device): + torch.xpu.set_device(device) + def is_directml_enabled(): global directml_enabled if directml_enabled: @@ -1848,7 +1960,34 @@ def soft_empty_cache(force=False): torch.cuda.ipc_collect() def unload_all_models(): - free_memory(1e30, get_torch_device()) + for device in get_all_torch_devices(): + free_memory(1e30, device) + +def unload_model_and_clones(model: ModelPatcher, unload_additional_models=True, all_devices=False): + 'Unload only model and its clones - primarily for multigpu cloning purposes.' + initial_keep_loaded: list[LoadedModel] = current_loaded_models.copy() + additional_models = [] + if unload_additional_models: + additional_models = model.get_nested_additional_models() + keep_loaded = [] + for loaded_model in initial_keep_loaded: + if loaded_model.model is not None: + if model.clone_base_uuid == loaded_model.model.clone_base_uuid: + continue + # check additional models if they are a match + skip = False + for add_model in additional_models: + if add_model.clone_base_uuid == loaded_model.model.clone_base_uuid: + skip = True + break + if skip: + continue + keep_loaded.append(loaded_model) + if not all_devices: + free_memory(1e30, get_torch_device(), keep_loaded) + else: + for device in get_all_torch_devices(): + free_memory(1e30, device, keep_loaded) def debug_memory_summary(): if is_amd() or is_nvidia(): diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index c8ed02e70..d70b42bf8 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -78,12 +78,15 @@ def set_model_options_pre_cfg_function(model_options, pre_cfg_function, disable_ def create_model_options_clone(orig_model_options: dict): return comfy.patcher_extension.copy_nested_dicts(orig_model_options) -def create_hook_patches_clone(orig_hook_patches): +def create_hook_patches_clone(orig_hook_patches, copy_tuples=False): new_hook_patches = {} for hook_ref in orig_hook_patches: new_hook_patches[hook_ref] = {} for k in orig_hook_patches[hook_ref]: new_hook_patches[hook_ref][k] = orig_hook_patches[hook_ref][k][:] + if copy_tuples: + for i in range(len(new_hook_patches[hook_ref][k])): + new_hook_patches[hook_ref][k][i] = tuple(new_hook_patches[hook_ref][k][i]) return new_hook_patches def wipe_lowvram_weight(m): @@ -329,7 +332,10 @@ class ModelPatcher: self.is_clip = False self.hook_mode = comfy.hooks.EnumHookMode.MaxSpeed - self.cached_patcher_init: tuple[Callable, tuple] | None = None + self.cached_patcher_init: tuple[Callable, tuple] | tuple[Callable, tuple, int] | None = None + self.is_multigpu_base_clone = False + self.clone_base_uuid = uuid.uuid4() + if not hasattr(self.model, 'model_loaded_weight_memory'): self.model.model_loaded_weight_memory = 0 @@ -366,20 +372,24 @@ class ModelPatcher: #than pays for CFG. So return everything both torch and Aimdo could give us aimdo_mem = 0 if comfy.memory_management.aimdo_enabled: - aimdo_mem = comfy_aimdo.model_vbar.vbars_analyze() + aimdo_device = device.index if getattr(device, "type", None) == "cuda" else None + aimdo_mem = comfy_aimdo.model_vbar.vbars_analyze(aimdo_device) return comfy.model_management.get_free_memory(device) + aimdo_mem def get_clone_model_override(self): return self.model, (self.backup, self.backup_buffers, self.object_patches_backup, self.pinned) - def clone(self, disable_dynamic=False, model_override=None): + def clone(self, disable_dynamic=False, model_override=None, force_deepcopy=False): class_ = self.__class__ - if self.is_dynamic() and disable_dynamic: - class_ = ModelPatcher + if self.is_dynamic() and disable_dynamic or force_deepcopy: + if self.is_dynamic() and disable_dynamic: + class_ = ModelPatcher if model_override is None: if self.cached_patcher_init is None: raise RuntimeError("Cannot create non-dynamic delegate: cached_patcher_init is not initialized.") temp_model_patcher = self.cached_patcher_init[0](*self.cached_patcher_init[1], disable_dynamic=True) + if len(self.cached_patcher_init) > 2: + temp_model_patcher = temp_model_patcher[self.cached_patcher_init[2]] model_override = temp_model_patcher.get_clone_model_override() if model_override is None: model_override = self.get_clone_model_override() @@ -438,19 +448,113 @@ class ModelPatcher: n.hook_mode = self.hook_mode n.cached_patcher_init = self.cached_patcher_init + n.is_multigpu_base_clone = self.is_multigpu_base_clone + n.clone_base_uuid = self.clone_base_uuid for callback in self.get_all_callbacks(CallbacksMP.ON_CLONE): callback(self, n) return n + def deepclone_multigpu(self, new_load_device=None, models_cache: dict[uuid.UUID,ModelPatcher]=None): + logging.info(f"Creating deepclone of {self.model.__class__.__name__} for {new_load_device if new_load_device else self.load_device}.") + if self.cached_patcher_init is None: + raise RuntimeError( + f"Cannot create multigpu deepclone of {self.model.__class__.__name__}: " + "the loader that produced this model does not support multigpu " + "(cached_patcher_init is not initialized). Use a core loader " + "(CheckpointLoaderSimple, UNETLoader, CLIPLoader/DualCLIPLoader, VAELoader), " + "or have the custom loader register a cached_patcher_init factory." + ) + comfy.model_management.unload_model_and_clones(self) + # Produce a freshly-loaded patcher from the loader factory so the multigpu + # clone owns its own untainted model weights (rather than relying on + # copy.deepcopy of an already-patched/already-loaded module). + temp_model_patcher: ModelPatcher | list[ModelPatcher] = self.cached_patcher_init[0](*self.cached_patcher_init[1]) + if len(self.cached_patcher_init) > 2: + temp_model_patcher = temp_model_patcher[self.cached_patcher_init[2]] + # Override clone()'s normal "share self.model + share backup containers" with + # the pristine model from temp_model_patcher plus empty backup containers -- + # the fresh model has no patches applied, so any deepcopy of self's stale + # backup/object_patches_backup/pinned would just propagate dead state that + # no longer corresponds to anything in n.model. + model_override = (temp_model_patcher.model, ({}, {}, {}, set())) + n = self.clone(model_override=model_override) + # clone() copies hook_backup by reference from self; reset since model is pristine. + n.hook_backup = {} + # set load device, if present + if new_load_device is not None: + n.load_device = new_load_device + # Ensure any per-device bookkeeping (e.g. ModelPatcherDynamic.dynamic_pins) + # has an entry for n.load_device on the freshly-loaded n.model. temp_model_patcher's + # __init__ only registered its own (default) load_device. + if hasattr(n, "register_load_device"): + n.register_load_device(n.load_device) + # multigpu clone should not have multigpu additional_models entry + n.remove_additional_models("multigpu") + # multigpu_clone all stored additional_models; make sure circular references are properly handled + if models_cache is None: + models_cache = {} + for key, model_list in n.additional_models.items(): + for i in range(len(model_list)): + add_model = n.additional_models[key][i] + if add_model.clone_base_uuid not in models_cache: + models_cache[add_model.clone_base_uuid] = add_model.deepclone_multigpu(new_load_device=new_load_device, models_cache=models_cache) + n.additional_models[key][i] = models_cache[add_model.clone_base_uuid] + for callback in self.get_all_callbacks(CallbacksMP.ON_DEEPCLONE_MULTIGPU): + callback(self, n) + return n + + def match_multigpu_clones(self): + multigpu_models = self.get_additional_models_with_key("multigpu") + if len(multigpu_models) > 0: + new_multigpu_models = [] + for mm in multigpu_models: + # clone main model, but bring over relevant props from existing multigpu clone + n = self.clone() + n.load_device = mm.load_device + n.backup = mm.backup + n.object_patches_backup = mm.object_patches_backup + n.hook_backup = mm.hook_backup + n.model = mm.model + n.is_multigpu_base_clone = mm.is_multigpu_base_clone + n.remove_additional_models("multigpu") + orig_additional_models: dict[str, list[ModelPatcher]] = comfy.patcher_extension.copy_nested_dicts(n.additional_models) + n.additional_models = comfy.patcher_extension.copy_nested_dicts(mm.additional_models) + # figure out which additional models are not present in multigpu clone + models_cache = {} + for mm_add_model in mm.get_additional_models(): + models_cache[mm_add_model.clone_base_uuid] = mm_add_model + remove_models_uuids = set(list(models_cache.keys())) + for key, model_list in orig_additional_models.items(): + for orig_add_model in model_list: + if orig_add_model.clone_base_uuid not in models_cache: + models_cache[orig_add_model.clone_base_uuid] = orig_add_model.deepclone_multigpu(new_load_device=n.load_device, models_cache=models_cache) + existing_list = n.get_additional_models_with_key(key) + existing_list.append(models_cache[orig_add_model.clone_base_uuid]) + n.set_additional_models(key, existing_list) + if orig_add_model.clone_base_uuid in remove_models_uuids: + remove_models_uuids.remove(orig_add_model.clone_base_uuid) + # remove duplicate additional models + for key, model_list in n.additional_models.items(): + new_model_list = [x for x in model_list if x.clone_base_uuid not in remove_models_uuids] + n.set_additional_models(key, new_model_list) + for callback in self.get_all_callbacks(CallbacksMP.ON_MATCH_MULTIGPU_CLONES): + callback(self, n) + new_multigpu_models.append(n) + self.set_additional_models("multigpu", new_multigpu_models) + def is_clone(self, other): if hasattr(other, 'model') and self.model is other.model: return True return False - def clone_has_same_weights(self, clone: 'ModelPatcher'): - if not self.is_clone(clone): - return False + def clone_has_same_weights(self, clone: ModelPatcher, allow_multigpu=False): + if allow_multigpu: + if self.clone_base_uuid != clone.clone_base_uuid: + return False + else: + if not self.is_clone(clone): + return False if self.current_hooks != clone.current_hooks: return False @@ -1232,7 +1336,7 @@ class ModelPatcher: return self.additional_models.get(key, []) def get_additional_models(self): - all_models = [] + all_models: list[ModelPatcher] = [] for models in self.additional_models.values(): all_models.extend(models) return all_models @@ -1286,9 +1390,18 @@ class ModelPatcher: for callback in self.get_all_callbacks(CallbacksMP.ON_PRE_RUN): callback(self) - def prepare_state(self, timestep): + def prepare_state(self, timestep, model_options): + ignore_multigpu = model_options.get("ignore_multigpu", False) for callback in self.get_all_callbacks(CallbacksMP.ON_PREPARE_STATE): - callback(self, timestep) + callback(self, timestep, model_options) + if not ignore_multigpu and "multigpu_clones" in model_options: + model_options["ignore_multigpu"] = True + try: + for p in model_options["multigpu_clones"].values(): + p: ModelPatcher + p.prepare_state(timestep, model_options) + finally: + model_options.pop("ignore_multigpu", None) def restore_hook_patches(self): if self.hook_patches_backup is not None: @@ -1301,12 +1414,18 @@ class ModelPatcher: def prepare_hook_patches_current_keyframe(self, t: torch.Tensor, hook_group: comfy.hooks.HookGroup, model_options: dict[str]): curr_t = t[0] reset_current_hooks = False + multigpu_kf_changed_cache = None transformer_options = model_options.get("transformer_options", {}) for hook in hook_group.hooks: changed = hook.hook_keyframe.prepare_current_keyframe(curr_t=curr_t, transformer_options=transformer_options) # if keyframe changed, remove any cached HookGroups that contain hook with the same hook_ref; # this will cause the weights to be recalculated when sampling if changed: + # cache changed for multigpu usage + if "multigpu_clones" in model_options: + if multigpu_kf_changed_cache is None: + multigpu_kf_changed_cache = [] + multigpu_kf_changed_cache.append(hook) # reset current_hooks if contains hook that changed if self.current_hooks is not None: for current_hook in self.current_hooks.hooks: @@ -1318,6 +1437,28 @@ class ModelPatcher: self.cached_hook_patches.pop(cached_group) if reset_current_hooks: self.patch_hooks(None) + if "multigpu_clones" in model_options: + for p in model_options["multigpu_clones"].values(): + p: ModelPatcher + p._handle_changed_hook_keyframes(multigpu_kf_changed_cache) + + def _handle_changed_hook_keyframes(self, kf_changed_cache: list[comfy.hooks.Hook]): + 'Used to handle multigpu behavior inside prepare_hook_patches_current_keyframe.' + if kf_changed_cache is None: + return + reset_current_hooks = False + # reset current_hooks if contains hook that changed + for hook in kf_changed_cache: + if self.current_hooks is not None: + for current_hook in self.current_hooks.hooks: + if current_hook == hook: + reset_current_hooks = True + break + for cached_group in list(self.cached_hook_patches.keys()): + if cached_group.contains(hook): + self.cached_hook_patches.pop(cached_group) + if reset_current_hooks: + self.patch_hooks(None) def register_all_hook_patches(self, hooks: comfy.hooks.HookGroup, target_dict: dict[str], model_options: dict=None, registered: comfy.hooks.HookGroup = None): @@ -1566,16 +1707,27 @@ class ModelPatcherDynamic(ModelPatcher): self.model.dynamic_vbars = {} if not hasattr(self.model, "dynamic_pins"): self.model.dynamic_pins = {} - if self.load_device not in self.model.dynamic_pins: - self.model.dynamic_pins[self.load_device] = { - "weights": (comfy_aimdo.host_buffer.HostBuffer(0, 0, 0), [], [-1], [0]), - "patches": (comfy_aimdo.host_buffer.HostBuffer(0, 0, 0), [], [-1], [0]), + self.register_load_device(self.load_device) + self.non_dynamic_delegate_model = None + assert load_device is not None + + def register_load_device(self, device): + """Ensure dynamic_pins has an entry for *device*. + + Called from __init__ and also from any code that retargets an + already-constructed patcher to a new load_device (e.g. the + Select{Model,CLIP,VAE}Device selector nodes); without this entry + partially_unload_ram() raises KeyError when it tries to read the + per-device pin state. + """ + if device not in self.model.dynamic_pins: + self.model.dynamic_pins[device] = { + "weights": (comfy_aimdo.host_buffer.HostBuffer(0, 0, 0), [], [-1], [0], [0], {}), + "patches": (comfy_aimdo.host_buffer.HostBuffer(0, 0, 0), [], [-1], [0], [0], {}), "hostbufs_initialized": False, "failed": False, "active": False, } - self.non_dynamic_delegate_model = None - assert load_device is not None def is_dynamic(self): return True @@ -1613,6 +1765,16 @@ class ModelPatcherDynamic(ModelPatcher): #use all ModelPatcherDynamic this is ignored and its all done dynamically. return super().memory_required(input_shape=input_shape) * 1.3 + (1024 ** 3) + def restore_loaded_backups(self): + restored = self.model.model_loaded_weight_memory + for key in list(self.backup.keys()): + bk = self.backup.pop(key) + comfy.utils.set_attr_param(self.model, key, bk.weight) + for key in list(self.backup_buffers.keys()): + comfy.utils.set_attr_buffer(self.model, key, self.backup_buffers.pop(key)) + self.model.model_loaded_weight_memory = 0 + return restored + def load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False, full_load=False, dirty=False): @@ -1629,7 +1791,7 @@ class ModelPatcherDynamic(ModelPatcher): num_patches = 0 allocated_size = 0 - self.model.model_loaded_weight_memory = 0 + self.restore_loaded_backups() with self.use_ejected(): self.unpatch_hooks() @@ -1638,8 +1800,8 @@ class ModelPatcherDynamic(ModelPatcher): pin_state = self.model.dynamic_pins[self.load_device] if not pin_state["hostbufs_initialized"]: hostbuf_size = comfy.model_management.pinned_hostbuf_size(self.model_size()) - pin_state["weights"] = (comfy_aimdo.host_buffer.HostBuffer(0, 64 * 1024 * 1024, hostbuf_size), [], [-1], [0]) - pin_state["patches"] = (comfy_aimdo.host_buffer.HostBuffer(0, 8 * 1024 * 1024, hostbuf_size), [], [-1], [0]) + pin_state["weights"] = (comfy_aimdo.host_buffer.HostBuffer(0, 64 * 1024 * 1024, hostbuf_size), [], [-1], [0], [0], {}) + pin_state["patches"] = (comfy_aimdo.host_buffer.HostBuffer(0, 8 * 1024 * 1024, hostbuf_size), [], [-1], [0], [0], {}) pin_state["hostbufs_initialized"] = True pin_state["failed"] = False pin_state["active"] = True @@ -1716,6 +1878,9 @@ class ModelPatcherDynamic(ModelPatcher): force_load=True if force_load: + if hasattr(m, "_v"): + comfy_aimdo.model_vbar.vbar_unpin(m._v) + delattr(m, "_v") force_load_param(self, "weight", device_to) force_load_param(self, "bias", device_to) else: @@ -1773,29 +1938,21 @@ class ModelPatcherDynamic(ModelPatcher): freed = 0 if vbar is None else vbar.free_memory(memory_to_free) if freed < memory_to_free: - for key in list(self.backup.keys()): - bk = self.backup.pop(key) - comfy.utils.set_attr_param(self.model, key, bk.weight) - for key in list(self.backup_buffers.keys()): - comfy.utils.set_attr_buffer(self.model, key, self.backup_buffers.pop(key)) - freed += self.model.model_loaded_weight_memory - self.model.model_loaded_weight_memory = 0 + freed += self.restore_loaded_backups() return freed def loaded_ram_size(self): - return (self.model.dynamic_pins[self.load_device]["weights"][0].size + - self.model.dynamic_pins[self.load_device]["patches"][0].size) + return (self.model.dynamic_pins[self.load_device]["weights"][0].size) def pinned_memory_size(self): - return (self.model.dynamic_pins[self.load_device]["weights"][3][0] + - self.model.dynamic_pins[self.load_device]["patches"][3][0]) + return (self.model.dynamic_pins[self.load_device]["weights"][3][0]) def unregister_inactive_pins(self, ram_to_unload, subsets=[ "weights", "patches" ]): freed = 0 pin_state = self.model.dynamic_pins[self.load_device] for subset in subsets: - hostbuf, stack, stack_split, pinned_size = pin_state[subset] + hostbuf, stack, stack_split, pinned_size, *_ = pin_state[subset] split = stack_split[0] while split >= 0: module, offset = stack[split] @@ -1820,10 +1977,12 @@ class ModelPatcherDynamic(ModelPatcher): freed = 0 pin_state = self.model.dynamic_pins[self.load_device] for subset in subsets: - hostbuf, stack, stack_split, pinned_size = pin_state[subset] + hostbuf, stack, stack_split, pinned_size, *_ = pin_state[subset] while len(stack) > 0: module, offset = stack.pop() size = module._pin.numel() * module._pin.element_size() + module._pin_balancer_entry[-1] = None + del module._pin_balancer_entry del module._pin hostbuf.truncate(offset, do_unregister=module._pin_registered) stack_split[0] = min(stack_split[0], len(stack) - 1) diff --git a/comfy/model_prefetch.py b/comfy/model_prefetch.py index 72e11dec6..aa6d22d77 100644 --- a/comfy/model_prefetch.py +++ b/comfy/model_prefetch.py @@ -1,4 +1,5 @@ import comfy_aimdo.model_vbar +import comfy.memory_management import comfy.model_management import comfy.ops @@ -50,7 +51,17 @@ def prefetch_queue_pop(queue, device, module): if hasattr(s, "_v"): comfy_modules.append(s) + registerable_size = 0 + for s in comfy_modules: + registerable_size += comfy.memory_management.vram_aligned_size([s.weight, s.bias]) + for param_key in ("weight", "bias"): + lowvram_fn = getattr(s, param_key + "_lowvram_function", None) + if lowvram_fn is not None: + registerable_size += lowvram_fn.memory_required() + offload_stream = comfy.ops.cast_modules_with_vbar(comfy_modules, None, device, None, True) + if not comfy.model_management.args.fast_disk: + comfy.model_management.ensure_pin_registerable(registerable_size) comfy.model_management.sync_stream(device, offload_stream) queue[0] = (offload_stream, (prefetch, comfy_modules)) diff --git a/comfy/multigpu.py b/comfy/multigpu.py new file mode 100644 index 000000000..2b6d8260d --- /dev/null +++ b/comfy/multigpu.py @@ -0,0 +1,250 @@ +from __future__ import annotations +import queue +import threading +import torch +import logging + +from collections import namedtuple +from typing import TYPE_CHECKING +if TYPE_CHECKING: + from comfy.model_patcher import ModelPatcher +import comfy.utils +import comfy.patcher_extension +import comfy.model_management + + +class MultiGPUThreadPool: + """Persistent thread pool for multi-GPU work distribution. + + Maintains one worker thread per extra GPU device. Each thread calls + set_torch_device() once at startup so that compiled kernel caches + (inductor/triton) stay warm across diffusion steps. + """ + + def __init__(self, devices: list[torch.device]): + self._workers: list[threading.Thread] = [] + self._work_queues: dict[torch.device, queue.Queue] = {} + self._result_queues: dict[torch.device, queue.Queue] = {} + + for device in devices: + wq = queue.Queue() + rq = queue.Queue() + self._work_queues[device] = wq + self._result_queues[device] = rq + t = threading.Thread(target=self._worker_loop, args=(device, wq, rq), daemon=True) + t.start() + self._workers.append(t) + + def _worker_loop(self, device: torch.device, work_q: queue.Queue, result_q: queue.Queue): + try: + comfy.model_management.set_torch_device(device) + except Exception as e: + logging.error(f"MultiGPUThreadPool: failed to set device {device}: {e}") + while True: + item = work_q.get() + if item is None: + return + result_q.put((None, e)) + return + while True: + item = work_q.get() + if item is None: + break + fn, args, kwargs = item + try: + result = fn(*args, **kwargs) + result_q.put((result, None)) + except comfy.model_management.InterruptProcessingException as e: + result_q.put((None, e)) + except Exception as e: + result_q.put((None, e)) + + def submit(self, device: torch.device, fn, *args, **kwargs): + self._work_queues[device].put((fn, args, kwargs)) + + def get_result(self, device: torch.device): + return self._result_queues[device].get() + + @property + def devices(self) -> list[torch.device]: + return list(self._work_queues.keys()) + + def shutdown(self): + for wq in self._work_queues.values(): + wq.put(None) # sentinel + for t in self._workers: + t.join(timeout=5.0) + + +class GPUOptions: + def __init__(self, device_index: int, relative_speed: float): + self.device_index = device_index + self.relative_speed = relative_speed + + def clone(self): + return GPUOptions(self.device_index, self.relative_speed) + + def create_dict(self): + return { + "relative_speed": self.relative_speed + } + +class GPUOptionsGroup: + def __init__(self): + self.options: dict[int, GPUOptions] = {} + + def add(self, info: GPUOptions): + self.options[info.device_index] = info + + def clone(self): + c = GPUOptionsGroup() + for opt in self.options.values(): + c.add(opt) + return c + + def register(self, model: ModelPatcher): + opts_dict = {} + # get devices that are valid for this model + devices: list[torch.device] = [model.load_device] + for extra_model in model.get_additional_models_with_key("multigpu"): + extra_model: ModelPatcher + devices.append(extra_model.load_device) + # create dictionary with actual device mapped to its GPUOptions + device_opts_list: list[GPUOptions] = [] + for device in devices: + device_opts = self.options.get(device.index, GPUOptions(device_index=device.index, relative_speed=1.0)) + opts_dict[device] = device_opts.create_dict() + device_opts_list.append(device_opts) + # make relative_speed relative to 1.0 + min_speed = min([x.relative_speed for x in device_opts_list]) + for value in opts_dict.values(): + value['relative_speed'] /= min_speed + model.model_options['multigpu_options'] = opts_dict + + +def create_multigpu_deepclones(model: ModelPatcher, max_gpus: int, gpu_options: GPUOptionsGroup=None, reuse_loaded=False): + 'Prepare ModelPatcher to contain deepclones of its BaseModel and related properties.' + model = model.clone() + # check if multigpu is already prepared - get the load devices from them if possible to exclude + skip_devices = set() + multigpu_models = model.get_additional_models_with_key("multigpu") + if len(multigpu_models) > 0: + for mm in multigpu_models: + skip_devices.add(mm.load_device) + skip_devices = list(skip_devices) + + # Exclude the primary model's actual device, not the global current device: + # after SelectModelDevice(gpu:N) the primary may not live on the process's + # current CUDA device, and excluding the wrong device picks bad extras. + all_devices = comfy.model_management.get_all_torch_devices(exclude_current=False) + full_extra_devices = [d for d in all_devices if d != model.load_device] + limit_extra_devices = full_extra_devices[:max_gpus-1] + extra_devices = limit_extra_devices.copy() + # exclude skipped devices + for skip in skip_devices: + if skip in extra_devices: + extra_devices.remove(skip) + # create new deepclones + if len(extra_devices) > 0: + for device in extra_devices: + device_patcher = None + if reuse_loaded: + # Only reuse a previously-loaded MultiGPU clone. A SelectModelDevice + # patcher on the same device shares clone_base_uuid but has + # is_multigpu_base_clone=False, which would later be filtered out by + # prepare_model_patcher_multigpu_clones() and silently shrink the + # work split back to one GPU. + loaded_models: list[ModelPatcher] = comfy.model_management.loaded_models() + for lm in loaded_models: + if lm.model is None: + continue + if lm.load_device != device: + continue + if lm.clone_base_uuid != model.clone_base_uuid: + continue + if not getattr(lm, "is_multigpu_base_clone", False): + continue + device_patcher = lm.clone() + logging.info(f"Reusing loaded multigpu deepclone of {device_patcher.model.__class__.__name__} for {device}") + break + if device_patcher is None: + device_patcher = model.deepclone_multigpu(new_load_device=device) + # Always flag the clone; whether reused or freshly deepcloned, it must + # advertise itself as a MultiGPU base clone so the cond scheduler picks + # it up in prepare_model_patcher_multigpu_clones(). + device_patcher.is_multigpu_base_clone = True + multigpu_models = model.get_additional_models_with_key("multigpu") + multigpu_models.append(device_patcher) + model.set_additional_models("multigpu", multigpu_models) + model.match_multigpu_clones() + if gpu_options is None: + gpu_options = GPUOptionsGroup() + gpu_options.register(model) + else: + logging.info("No extra torch devices need initialization, skipping initializing MultiGPU Work Units.") + # only keep model clones that don't go 'past' the intended max_gpu count; + # this prunes any inherited multigpu clones whose load_device is no longer allowed + # when max_gpus is lowered between runs. + allowed_devices = set(limit_extra_devices) + allowed_devices.add(model.load_device) + multigpu_models = model.get_additional_models_with_key("multigpu") + new_multigpu_models = [m for m in multigpu_models if m.load_device in allowed_devices] + if len(new_multigpu_models) != len(multigpu_models): + model.set_additional_models("multigpu", new_multigpu_models) + model.match_multigpu_clones() + return model + + +LoadBalance = namedtuple('LoadBalance', ['work_per_device', 'idle_time']) +def load_balance_devices(model_options: dict[str], total_work: int, return_idle_time=False, work_normalized: int=None): + 'Optimize work assigned to different devices, accounting for their relative speeds and splittable work.' + opts_dict = model_options['multigpu_options'] + devices = list(model_options['multigpu_clones'].keys()) + speed_per_device = [] + work_per_device = [] + # get sum of each device's relative_speed + total_speed = 0.0 + for opts in opts_dict.values(): + total_speed += opts['relative_speed'] + # get relative work for each device; + # obtained by w = (W*r)/R + for device in devices: + relative_speed = opts_dict[device]['relative_speed'] + relative_work = (total_work*relative_speed) / total_speed + speed_per_device.append(relative_speed) + work_per_device.append(relative_work) + # relative work must be expressed in whole numbers, but likely is a decimal; + # perform rounding while maintaining total sum equal to total work (sum of relative works) + work_per_device = round_preserved(work_per_device) + dict_work_per_device = {} + for device, relative_work in zip(devices, work_per_device): + dict_work_per_device[device] = relative_work + if not return_idle_time: + return LoadBalance(dict_work_per_device, None) + # divide relative work by relative speed to get estimated completion time of said work by each device; + # time here is relative and does not correspond to real-world units + completion_time = [w/r for w,r in zip(work_per_device, speed_per_device)] + # calculate relative time spent by the devices waiting on each other after their work is completed + idle_time = abs(min(completion_time) - max(completion_time)) + # if need to compare work idle time, need to normalize to a common total work + if work_normalized: + idle_time *= (work_normalized/total_work) + + return LoadBalance(dict_work_per_device, idle_time) + +def round_preserved(values: list[float]): + 'Round all values in a list, preserving the combined sum of values.' + # get floor of values; casting to int does it too + floored = [int(x) for x in values] + total_floored = sum(floored) + # get remainder to distribute + remainder = round(sum(values)) - total_floored + # pair values with fractional portions + fractional = [(i, x-floored[i]) for i, x in enumerate(values)] + # sort by fractional part in descending order + fractional.sort(key=lambda x: x[1], reverse=True) + # distribute the remainder + for i in range(remainder): + index = fractional[i][0] + floored[index] += 1 + return floored diff --git a/comfy/ops.py b/comfy/ops.py index 9bcd6c900..119177c37 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -18,6 +18,7 @@ import torch import logging +import contextlib import comfy.model_management from comfy.cli_args import args, PerformanceFeature import comfy.float @@ -75,8 +76,6 @@ except: cast_to = comfy.model_management.cast_to #TODO: remove once no more references -STREAM_PIN_BUFFER_HEADROOM = 8 * 1024 * 1024 - def cast_to_input(weight, input, non_blocking=False, copy=True): return comfy.model_management.cast_to(weight, input.dtype, input.device, non_blocking=non_blocking, copy=copy) @@ -93,9 +92,6 @@ def cast_modules_with_vbar(comfy_modules, dtype, device, bias_dtype, non_blockin offload_stream = None cast_buffer = None cast_buffer_offset = 0 - stream_pin_hostbuf = None - stream_pin_offset = 0 - stream_pin_queue = [] def ensure_offload_stream(module, required_size, check_largest): nonlocal offload_stream @@ -129,22 +125,6 @@ def cast_modules_with_vbar(comfy_modules, dtype, device, bias_dtype, non_blockin cast_buffer_offset += buffer_size return buffer - def get_stream_pin_buffer_offset(buffer_size): - nonlocal stream_pin_hostbuf - nonlocal stream_pin_offset - - if buffer_size == 0 or offload_stream is None: - return None - - if stream_pin_hostbuf is None: - stream_pin_hostbuf = comfy.model_management.get_pin_buffer(offload_stream) - if stream_pin_hostbuf is None: - return None - - offset = stream_pin_offset - stream_pin_offset += buffer_size - return offset - for s in comfy_modules: signature = comfy_aimdo.model_vbar.vbar_fault(s._v) resident = comfy_aimdo.model_vbar.vbar_signature_compare(signature, s._v_signature) @@ -183,12 +163,18 @@ def cast_modules_with_vbar(comfy_modules, dtype, device, bias_dtype, non_blockin if xfer_dest is None: xfer_dest = get_cast_buffer(dest_size) - def cast_maybe_lowvram_patch(xfer_source, xfer_dest, stream): + def cast_maybe_lowvram_patch(xfer_source, xfer_dest, stream, xfer_dest2=None): if xfer_source is not None: if getattr(xfer_source, "is_lowvram_patch", False): - xfer_source.prepare(xfer_dest, stream, copy=True, commit=False) - else: - comfy.model_management.cast_to_gathered(xfer_source, xfer_dest, non_blocking=non_blocking, stream=stream) + if xfer_dest is not None: + xfer_source.prepare(xfer_dest, stream, copy=True, commit=False) + xfer_source = [ xfer_dest ] + xfer_dest = xfer_dest2 + xfer_dest2 = None + elif xfer_dest2 is not None: + xfer_source.prepare(xfer_dest2, stream, copy=True, commit=False) + return + comfy.model_management.cast_to_gathered(xfer_source, xfer_dest, non_blocking=non_blocking, stream=stream, r2=xfer_dest2) def handle_pin(m, pin, source, dest, subset="weights", size=None): if pin is not None: @@ -197,19 +183,7 @@ def cast_modules_with_vbar(comfy_modules, dtype, device, bias_dtype, non_blockin if signature is None: comfy.pinned_memory.pin_memory(m, subset=subset, size=size) pin = comfy.pinned_memory.get_pin(m, subset=subset) - if pin is not None: - if isinstance(source, list): - comfy.model_management.cast_to_gathered(source, pin, non_blocking=non_blocking, stream=offload_stream, r2=dest) - else: - cast_maybe_lowvram_patch(source, pin, None) - cast_maybe_lowvram_patch([ pin ], dest, offload_stream) - return - if pin is None: - pin_offset = get_stream_pin_buffer_offset(size) - if pin_offset is not None: - stream_pin_queue.append((source, pin_offset, size, dest)) - return - cast_maybe_lowvram_patch(source, dest, offload_stream) + cast_maybe_lowvram_patch(source, pin, offload_stream, xfer_dest2=dest) handle_pin(s, pin, xfer_source, xfer_dest, size=dest_size) @@ -231,23 +205,6 @@ def cast_modules_with_vbar(comfy_modules, dtype, device, bias_dtype, non_blockin prefetch["needs_cast"] = needs_cast s._prefetch = prefetch - if stream_pin_offset > 0: - if stream_pin_hostbuf.size < stream_pin_offset: - if not comfy.model_management.resize_pin_buffer(stream_pin_hostbuf, stream_pin_offset + STREAM_PIN_BUFFER_HEADROOM): - for xfer_source, _, _, xfer_dest in stream_pin_queue: - cast_maybe_lowvram_patch(xfer_source, xfer_dest, offload_stream) - return offload_stream - stream_pin_tensor = comfy_aimdo.torch.hostbuf_to_tensor(stream_pin_hostbuf) - stream_pin_tensor.untyped_storage()._comfy_hostbuf = stream_pin_hostbuf - for xfer_source, pin_offset, pin_size, xfer_dest in stream_pin_queue: - pin = stream_pin_tensor[pin_offset:pin_offset + pin_size] - if isinstance(xfer_source, list): - comfy.model_management.cast_to_gathered(xfer_source, pin, non_blocking=non_blocking, stream=offload_stream, r2=xfer_dest) - else: - cast_maybe_lowvram_patch(xfer_source, pin, None) - comfy.model_management.cast_to_gathered([ pin ], xfer_dest, non_blocking=non_blocking, stream=offload_stream) - stream_pin_hostbuf._comfy_event = offload_stream.record_event() - return offload_stream @@ -1047,6 +1004,144 @@ class QuantLinearFunc(torch.autograd.Function): return grad_input, grad_weight, grad_bias, None, None, None +# Quantized-weight module helpers + +def _quantized_apply(module, fn, recurse=True): + """Re-wrap Parameters after fn so .to()/.cuda() propagate through QuantizedTensor weights.""" + if recurse: + for child in module.children(): + child._apply(fn) + for key, param in module._parameters.items(): + if param is None: + continue + p = fn(param) + if (not torch.is_inference_mode_enabled()) and p.is_inference(): + p = p.clone() + module.register_parameter(key, torch.nn.Parameter(p, requires_grad=False)) + for key, buf in module._buffers.items(): + if buf is not None: + module._buffers[key] = fn(buf) + return module + + +def _load_quantized_module(module, super_load, state_dict, prefix, local_metadata, strict, + missing_keys, unexpected_keys, error_msgs, load_extra_params=False): + """Shared _load_from_state_dict body for quantized-weight modules. + + Pops weight (+ scales, +/- extras), populates module.weight as a Parameter + or Parameter-wrapped QuantizedTensor, then calls super_load and strips + consumed keys from missing_keys. Reads compute_dtype from factory_kwargs + and disabled formats from module._disabled_formats. + """ + device = module.factory_kwargs["device"] + compute_dtype = module.factory_kwargs["dtype"] + disabled_formats = module._disabled_formats + layer_name = prefix.rstrip('.') + + weight = state_dict.pop(f"{prefix}weight", None) + if weight is None: + logging.warning(f"Missing weight for layer {layer_name}") + module.weight = None + return + manually_loaded_keys = [f"{prefix}weight"] + + def pop_scale(name, dtype=None): + key = f"{prefix}{name}" + v = state_dict.pop(key, None) + if v is not None: + v = v.to(device=device) + if dtype is not None: + v = v.view(dtype=dtype) + manually_loaded_keys.append(key) + return v + + layer_conf = state_dict.pop(f"{prefix}comfy_quant", None) + if layer_conf is not None: + layer_conf = json.loads(layer_conf.numpy().tobytes()) + + if layer_conf is None: + module.weight = torch.nn.Parameter(weight.to(device=device, dtype=compute_dtype), requires_grad=False) + else: + module.quant_format = layer_conf.get("format", None) + module._full_precision_mm_config = layer_conf.get("full_precision_matrix_mult", False) + if not module._full_precision_mm: + module._full_precision_mm = module._full_precision_mm_config + if module.quant_format in disabled_formats: + module._full_precision_mm = True + if module.quant_format is None: + raise ValueError(f"Unknown quantization format for layer {layer_name}") + + qconfig = QUANT_ALGOS[module.quant_format] + module.layout_type = qconfig["comfy_tensor_layout"] + layout_cls = get_layout_class(module.layout_type) + + # Per-format scales; fp8 dtype views handle both legacy uint8-on-disk and native fp8. + if module.quant_format in ("float8_e4m3fn", "float8_e5m2"): + scales = {"scale": pop_scale("weight_scale")} + elif module.quant_format == "mxfp8": + bs = pop_scale("weight_scale", torch.float8_e8m0fnu) + if bs is None: + raise ValueError(f"Missing MXFP8 block scales for layer {layer_name}") + scales = {"scale": bs} + elif module.quant_format == "nvfp4": + ts = pop_scale("weight_scale_2") + bs = pop_scale("weight_scale", torch.float8_e4m3fn) + if ts is None or bs is None: + raise ValueError(f"Missing NVFP4 scales for layer {layer_name}") + scales = {"scale": ts, "block_scale": bs} + else: + raise ValueError(f"Unsupported quantization format: {module.quant_format}") + + params = layout_cls.Params(**scales, orig_dtype=compute_dtype, orig_shape=module._orig_shape) + module.weight = torch.nn.Parameter( + QuantizedTensor(weight.to(device=device, dtype=qconfig["storage_t"]), module.layout_type, params), + requires_grad=False, + ) + + if load_extra_params: + for param_name in qconfig["parameters"]: + if param_name in {"weight_scale", "weight_scale_2"}: + continue + param_key = f"{prefix}{param_name}" + _v = state_dict.pop(param_key, None) + if _v is None: + continue + module.register_parameter(param_name, torch.nn.Parameter(_v.to(device=device), requires_grad=False)) + manually_loaded_keys.append(param_key) + + super_load(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) + for key in manually_loaded_keys: + if key in missing_keys: + missing_keys.remove(key) + + +def _quantized_weight_state_dict(module, sd, prefix, extra_quant_conf=None, extra_quant_params=()): + """Shared state_dict body. extra_quant_conf merges into the comfy_quant JSON; + extra_quant_params names attributes written as additional top-level keys.""" + if not hasattr(module, 'weight'): + logging.warning(f"Warning: state dict on uninitialized op {prefix}") + return sd + bias = getattr(module, 'bias', None) + if bias is not None: + sd[f"{prefix}bias"] = bias + if module.weight is None: + return sd + if isinstance(module.weight, QuantizedTensor): + sd.update(module.weight.state_dict(f"{prefix}weight")) + quant_conf = {"format": module.quant_format} + if getattr(module, '_full_precision_mm_config', False): + quant_conf["full_precision_matrix_mult"] = True + if extra_quant_conf: + quant_conf.update(extra_quant_conf) + sd[f"{prefix}comfy_quant"] = torch.tensor(list(json.dumps(quant_conf).encode("utf-8")), dtype=torch.uint8) + for name in extra_quant_params: + value = getattr(module, name, None) + if value is not None: + sd[f"{prefix}{name}"] = value + else: + sd[f"{prefix}weight"] = module.weight + return sd + def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_precision_mm=False, disabled=[]): class MixedPrecisionOps(manual_cast): @@ -1056,21 +1151,16 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec _disabled = disabled class Linear(torch.nn.Module, CastWeightBiasOp): - def __init__( - self, - in_features: int, - out_features: int, - bias: bool = True, - device=None, - dtype=None, - ) -> None: + _disabled_formats = disabled + + def __init__(self, in_features: int, out_features: int, bias: bool = True, device=None, dtype=None): super().__init__() self.factory_kwargs = {"device": device, "dtype": MixedPrecisionOps._compute_dtype} - # self.factory_kwargs = {"device": device, "dtype": dtype} self.in_features = in_features self.out_features = out_features + self._orig_shape = (out_features, in_features) if bias: self.bias = torch.nn.Parameter(torch.empty(out_features, **self.factory_kwargs)) else: @@ -1083,151 +1173,12 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec def reset_parameters(self): return None - def _load_scale_param(self, state_dict, prefix, param_name, device, manually_loaded_keys, dtype=None): - key = f"{prefix}{param_name}" - value = state_dict.pop(key, None) - if value is not None: - value = value.to(device=device) - if dtype is not None: - value = value.view(dtype=dtype) - manually_loaded_keys.append(key) - return value - - def _load_from_state_dict(self, state_dict, prefix, local_metadata, - strict, missing_keys, unexpected_keys, error_msgs): - - device = self.factory_kwargs["device"] - layer_name = prefix.rstrip('.') - weight_key = f"{prefix}weight" - weight = state_dict.pop(weight_key, None) - if weight is None: - logging.warning(f"Missing weight for layer {layer_name}") - self.weight = None - return - - manually_loaded_keys = [weight_key] - - layer_conf = state_dict.pop(f"{prefix}comfy_quant", None) - if layer_conf is not None: - layer_conf = json.loads(layer_conf.numpy().tobytes()) - - if layer_conf is None: - self.weight = torch.nn.Parameter(weight.to(device=device, dtype=MixedPrecisionOps._compute_dtype), requires_grad=False) - else: - self.quant_format = layer_conf.get("format", None) - self._full_precision_mm_config = layer_conf.get("full_precision_matrix_mult", False) - if not self._full_precision_mm: - self._full_precision_mm = self._full_precision_mm_config - - if self.quant_format in MixedPrecisionOps._disabled: - self._full_precision_mm = True - - if self.quant_format is None: - raise ValueError(f"Unknown quantization format for layer {layer_name}") - - qconfig = QUANT_ALGOS[self.quant_format] - self.layout_type = qconfig["comfy_tensor_layout"] - layout_cls = get_layout_class(self.layout_type) - - # Load format-specific parameters - if self.quant_format in ["float8_e4m3fn", "float8_e5m2"]: - # FP8: single tensor scale - scale = self._load_scale_param(state_dict, prefix, "weight_scale", device, manually_loaded_keys) - - params = layout_cls.Params( - scale=scale, - orig_dtype=MixedPrecisionOps._compute_dtype, - orig_shape=(self.out_features, self.in_features), - ) - - elif self.quant_format == "mxfp8": - # MXFP8: E8M0 block scales stored as uint8 in safetensors - block_scale = self._load_scale_param(state_dict, prefix, "weight_scale", device, manually_loaded_keys, - dtype=torch.uint8) - - if block_scale is None: - raise ValueError(f"Missing MXFP8 block scales for layer {layer_name}") - - block_scale = block_scale.view(torch.float8_e8m0fnu) - - params = layout_cls.Params( - scale=block_scale, - orig_dtype=MixedPrecisionOps._compute_dtype, - orig_shape=(self.out_features, self.in_features), - ) - - elif self.quant_format == "nvfp4": - # NVFP4: tensor_scale (weight_scale_2) + block_scale (weight_scale) - tensor_scale = self._load_scale_param(state_dict, prefix, "weight_scale_2", device, manually_loaded_keys) - block_scale = self._load_scale_param(state_dict, prefix, "weight_scale", device, manually_loaded_keys, - dtype=torch.float8_e4m3fn) - - if tensor_scale is None or block_scale is None: - raise ValueError(f"Missing NVFP4 scales for layer {layer_name}") - - params = layout_cls.Params( - scale=tensor_scale, - block_scale=block_scale, - orig_dtype=MixedPrecisionOps._compute_dtype, - orig_shape=(self.out_features, self.in_features), - ) - else: - raise ValueError(f"Unsupported quantization format: {self.quant_format}") - - self.weight = torch.nn.Parameter( - QuantizedTensor(weight.to(device=device, dtype=qconfig["storage_t"]), self.layout_type, params), - requires_grad=False - ) - - for param_name in qconfig["parameters"]: - if param_name in {"weight_scale", "weight_scale_2"}: - continue # Already handled above - - param_key = f"{prefix}{param_name}" - _v = state_dict.pop(param_key, None) - if _v is None: - continue - self.register_parameter(param_name, torch.nn.Parameter(_v.to(device=device), requires_grad=False)) - manually_loaded_keys.append(param_key) - - super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) - - for key in manually_loaded_keys: - if key in missing_keys: - missing_keys.remove(key) + def _load_from_state_dict(self, *args): + _load_quantized_module(self, super()._load_from_state_dict, *args, load_extra_params=True) def state_dict(self, *args, destination=None, prefix="", **kwargs): - if destination is not None: - sd = destination - else: - sd = {} - - if not hasattr(self, 'weight'): - logging.warning("Warning: state dict on uninitialized op {}".format(prefix)) - return sd - - if self.bias is not None: - sd["{}bias".format(prefix)] = self.bias - - if self.weight is None: - return sd - - if isinstance(self.weight, QuantizedTensor): - sd_out = self.weight.state_dict("{}weight".format(prefix)) - for k in sd_out: - sd[k] = sd_out[k] - - quant_conf = {"format": self.quant_format} - if self._full_precision_mm_config: - quant_conf["full_precision_matrix_mult"] = True - sd["{}comfy_quant".format(prefix)] = torch.tensor(list(json.dumps(quant_conf).encode('utf-8')), dtype=torch.uint8) - - input_scale = getattr(self, 'input_scale', None) - if input_scale is not None: - sd["{}input_scale".format(prefix)] = input_scale - else: - sd["{}weight".format(prefix)] = self.weight - return sd + sd = destination if destination is not None else {} + return _quantized_weight_state_dict(self, sd, prefix, extra_quant_params=("input_scale",)) def _forward(self, input, weight, bias): return torch.nn.functional.linear(input, weight, bias) @@ -1317,25 +1268,126 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec self.weight = torch.nn.Parameter(weight, requires_grad=False) def _apply(self, fn, recurse=True): # This is to get torch.compile + moving weights to another device working - if recurse: - for module in self.children(): - module._apply(fn) + return _quantized_apply(self, fn, recurse) - for key, param in self._parameters.items(): - if param is None: - continue - p = fn(param) - if (not torch.is_inference_mode_enabled()) and p.is_inference(): - p = p.clone() - self.register_parameter(key, torch.nn.Parameter(p, requires_grad=False)) - for key, buf in self._buffers.items(): - if buf is not None: - self._buffers[key] = fn(buf) - return self + class MoEExperts(torch.nn.Module, CastWeightBiasOp): + """Container for E quantized expert weights, indexed via expert_weight(i). + + The bank lives on self.weight as a single 3D tensor — either a + compute_dtype Parameter or a Parameter wrapping a QuantizedTensor + with leading expert dim. + + State-dict layout matches mixed_precision_ops.Linear with a leading + expert dim: + {prefix}.weight quant data (storage_t), leading dim = E + {prefix}.weight_scale block / per-tensor scale + {prefix}.weight_scale_2 [E] or scalar NVFP4 only + {prefix}.bias [E, out_features] optional, compute_dtype + {prefix}.comfy_quant json -> {{"format": "...", "num_experts": E}} + + Without comfy_quant the weight loads as a plain compute_dtype 3D Parameter [E, out, in]. + """ + + _disabled_formats = disabled + + def __init__(self, num_experts: int, in_features: int, out_features: int, bias: bool = True, device=None, dtype=None): + super().__init__() + self.num_experts = num_experts + self.in_features = in_features + self.out_features = out_features + self._orig_shape = (num_experts, out_features, in_features) + self.factory_kwargs = {"device": device, "dtype": MixedPrecisionOps._compute_dtype} + if bias: + self.bias = torch.nn.Parameter(torch.empty(num_experts, out_features, **self.factory_kwargs)) + else: + self.register_parameter("bias", None) + + # Populated by _load_from_state_dict: + self.weight = None + self.quant_format = None + self.layout_type = None + self._full_precision_mm = MixedPrecisionOps._full_precision_mm + self._full_precision_mm_config = False + self._resident_bank = None + + def reset_parameters(self): + return None + + def _apply(self, fn, recurse=True): + return _quantized_apply(self, fn, recurse) + + def _load_from_state_dict(self, *args): + _load_quantized_module(self, super()._load_from_state_dict, *args, load_extra_params=False) + + def expert_weight(self, i: int): + """Expert i's weight (Tensor or per-expert QuantizedTensor view).""" + if isinstance(self.weight, QuantizedTensor): + return self._expert_qt_from(self.weight, i) + return self.weight[i] + + @contextlib.contextmanager + def bank_resident(self, input): + """Cast the whole bank once; expert_linear inside reuses the cast. + Not re-entrant — do not nest calls on the same instance. + """ + weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True) + self._resident_bank = (weight, bias) + try: + yield self + finally: + self._resident_bank = None + uncast_bias_weight(self, weight, bias, offload_stream) + + def expert_linear(self, input: torch.Tensor, i: int) -> torch.Tensor: + """Linear against expert i's weight (with optional bias).""" + resident = getattr(self, "_resident_bank", None) + if resident is not None: + weight, bias = resident + return self._expert_linear_impl(input, weight, bias, i) + weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True) + try: + return self._expert_linear_impl(input, weight, bias, i) + finally: + uncast_bias_weight(self, weight, bias, offload_stream) + + def _expert_linear_impl(self, input, weight, bias, i): + if isinstance(weight, QuantizedTensor): + qw = self._expert_qt_from(weight, i) + else: + qw = weight[i] + b = cast_to_input(bias[i], input, copy=False) if bias is not None else None + + if isinstance(qw, QuantizedTensor): + use_fast = ( + not self._full_precision_mm + and qw.layout_cls.supports_fast_matmul() + and input.dim() == 2 + ) + if use_fast: + qin = QuantizedTensor.from_float(input, self.layout_type) + return torch.nn.functional.linear(qin, qw, b) + out = input @ qw.dequantize().t() + return out + b if b is not None else out + return torch.nn.functional.linear(input, qw, b) + + def _expert_qt_from(self, weight: QuantizedTensor, i: int) -> QuantizedTensor: + """Build a per-expert QuantizedTensor by indexing into a resident bank.""" + params = weight._params + kwargs = { + "scale": params.scale[i] if params.scale.dim() else params.scale, + "orig_dtype": params.orig_dtype, + "orig_shape": (self.out_features, self.in_features), + } + if hasattr(params, "block_scale"): # NVFP4 + kwargs["block_scale"] = params.block_scale[i] + return QuantizedTensor(weight._qdata[i], weight._layout_cls, type(params)(**kwargs)) + + def state_dict(self, *args, destination=None, prefix="", **kwargs): + sd = destination if destination is not None else {} + return _quantized_weight_state_dict(self, sd, prefix, extra_quant_conf={"num_experts": self.num_experts}) class Embedding(manual_cast.Embedding): - def _load_from_state_dict(self, state_dict, prefix, local_metadata, - strict, missing_keys, unexpected_keys, error_msgs): + def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs): weight_key = f"{prefix}weight" layer_conf = state_dict.pop(f"{prefix}comfy_quant", None) if layer_conf is not None: @@ -1343,14 +1395,16 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec # Only fp8 makes sense for embeddings (per-row dequant via index select). # Block-scaled formats (NVFP4, MXFP8) can't do per-row lookup efficiently. - quant_format = layer_conf.get("format", None) if layer_conf is not None else None - if quant_format in ["float8_e4m3fn", "float8_e5m2"] and weight_key in state_dict: + quant_format = layer_conf.get("format") if layer_conf is not None else None + manually_loaded_keys = [] + + if quant_format in ("float8_e4m3fn", "float8_e5m2") and weight_key in state_dict: self.quant_format = quant_format qconfig = QUANT_ALGOS[quant_format] self.layout_type = qconfig["comfy_tensor_layout"] layout_cls = get_layout_class(self.layout_type) weight = state_dict.pop(weight_key) - manually_loaded_keys = [weight_key] + manually_loaded_keys.append(weight_key) scale_key = f"{prefix}weight_scale" scale = state_dict.pop(scale_key, None) @@ -1366,35 +1420,19 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec self.weight = torch.nn.Parameter( QuantizedTensor(weight.to(dtype=qconfig["storage_t"]), qconfig["comfy_tensor_layout"], params), requires_grad=False) + elif layer_conf is not None: + # Unsupported format — restore the marker so it round-trips; fall through to default load. + state_dict[f"{prefix}comfy_quant"] = torch.tensor( + list(json.dumps(layer_conf).encode('utf-8')), dtype=torch.uint8) - super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) - for k in manually_loaded_keys: - if k in missing_keys: - missing_keys.remove(k) - else: - if layer_conf is not None: - state_dict[f"{prefix}comfy_quant"] = torch.tensor(list(json.dumps(layer_conf).encode('utf-8')), dtype=torch.uint8) - super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) + super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) + for k in manually_loaded_keys: + if k in missing_keys: + missing_keys.remove(k) def state_dict(self, *args, destination=None, prefix="", **kwargs): - if destination is not None: - sd = destination - else: - sd = {} - - if not hasattr(self, 'weight') or self.weight is None: - return sd - - if isinstance(self.weight, QuantizedTensor): - sd_out = self.weight.state_dict("{}weight".format(prefix)) - for k in sd_out: - sd[k] = sd_out[k] - - quant_conf = {"format": self.quant_format} - sd["{}comfy_quant".format(prefix)] = torch.tensor(list(json.dumps(quant_conf).encode('utf-8')), dtype=torch.uint8) - else: - sd["{}weight".format(prefix)] = self.weight - return sd + sd = destination if destination is not None else {} + return _quantized_weight_state_dict(self, sd, prefix) def forward_comfy_cast_weights(self, input, out_dtype=None): weight = self.weight diff --git a/comfy/patcher_extension.py b/comfy/patcher_extension.py index 5ee4d5ee5..189ee84ca 100644 --- a/comfy/patcher_extension.py +++ b/comfy/patcher_extension.py @@ -1,8 +1,9 @@ -from __future__ import annotations from typing import Callable class CallbacksMP: ON_CLONE = "on_clone" + ON_DEEPCLONE_MULTIGPU = "on_deepclone_multigpu" + ON_MATCH_MULTIGPU_CLONES = "on_match_multigpu_clones" ON_LOAD = "on_load_after" ON_DETACH = "on_detach_after" ON_CLEANUP = "on_cleanup" diff --git a/comfy/pinned_memory.py b/comfy/pinned_memory.py index 0e8f573ba..cb77c517a 100644 --- a/comfy/pinned_memory.py +++ b/comfy/pinned_memory.py @@ -1,17 +1,55 @@ +import bisect + import comfy.model_management import comfy.memory_management +import comfy.utils import comfy_aimdo.host_buffer import comfy_aimdo.torch import torch from comfy.cli_args import args +def _add_to_bucket(module, buckets, size, priority): + bucket = buckets.setdefault(size, []) + entry = [-priority, 0, module] + entry[1] = id(entry) + bisect.insort(bucket, entry) + module._pin_balancer_entry = entry + +def _steal_pin(module, stack, buckets, size, priority): + bucket = buckets.get(size) + if bucket is None: + return False + + while bucket and bucket[-1][-1] is None: + bucket.pop() + if not bucket: + del buckets[size] + return False + + if priority <= -bucket[-1][0]: + return False + + *_, victim = bucket.pop() + module._pin = victim._pin + module._pin_registered = victim._pin_registered + module._pin_stack_index = victim._pin_stack_index + stack[module._pin_stack_index] = (module, stack[module._pin_stack_index][1]) + + victim._pin_registered = False + del victim._pin + del victim._pin_stack_index + del victim._pin_balancer_entry + + _add_to_bucket(module, buckets, size, priority) + return True + def get_pin(module, subset="weights"): pin = getattr(module, "_pin", None) if pin is None or module._pin_registered or args.disable_pinned_memory: return pin - _, _, stack_split, pinned_size = module._pin_state[subset] + _, _, stack_split, pinned_size, *_ = module._pin_state[subset] size = pin.nbytes comfy.model_management.ensure_pin_registerable(size) @@ -31,33 +69,51 @@ def pin_memory(module, subset="weights", size=None): return pin = get_pin(module, subset) - if pin is not None or pin_state["failed"]: + if pin is not None: return - hostbuf, stack, stack_split, pinned_size = pin_state[subset] + hostbuf, stack, stack_split, pinned_size, counter, buckets = pin_state[subset] if size is None: size = comfy.memory_management.vram_aligned_size([ module.weight, module.bias ]) offset = hostbuf.size - registerable_size = size + max(0, hostbuf.size - pinned_size[0]) + registerable_size = size + priority = getattr(module, "_pin_balancer_priority", None) + + if priority is None: + priority = comfy.utils.bit_reverse_range(counter[0], 16) + counter[0] += 1 + module._pin_balancer_priority = priority comfy.memory_management.extra_ram_release(comfy.memory_management.RAM_CACHE_HEADROOM) if (not comfy.model_management.ensure_pin_budget(size) or not comfy.model_management.ensure_pin_registerable(registerable_size)): - pin_state["failed"] = True - return False + return _steal_pin(module, stack, buckets, size, priority) + extended = False try: - hostbuf.extend(size=size) + hostbuf.extend(size=size, register=False) + extended = True + pin = comfy_aimdo.torch.hostbuf_to_tensor(hostbuf)[offset:offset + size] + pin.untyped_storage()._comfy_hostbuf = hostbuf + if torch.cuda.cudart().cudaHostRegister(pin.data_ptr(), size, 1) != 0: + comfy.model_management.discard_cuda_async_error() + comfy.model_management.free_registrations(size) + if torch.cuda.cudart().cudaHostRegister(pin.data_ptr(), size, 1) != 0: + comfy.model_management.discard_cuda_async_error() + del pin + hostbuf.truncate(offset, do_unregister=False) + return _steal_pin(module, stack, buckets, size, priority) except RuntimeError: - pin_state["failed"] = True - return False + if extended: + hostbuf.truncate(offset, do_unregister=False) + return _steal_pin(module, stack, buckets, size, priority) - module._pin = comfy_aimdo.torch.hostbuf_to_tensor(hostbuf)[offset:offset + size] - module._pin.untyped_storage()._comfy_hostbuf = hostbuf + module._pin = pin stack.append((module, offset)) module._pin_registered = True module._pin_stack_index = len(stack) - 1 stack_split[0] = max(stack_split[0], module._pin_stack_index) comfy.model_management.TOTAL_PINNED_MEMORY += size pinned_size[0] += size + _add_to_bucket(module, buckets, size, priority) return True diff --git a/comfy/sampler_helpers.py b/comfy/sampler_helpers.py index 3782fd2d5..bdce2f2d8 100644 --- a/comfy/sampler_helpers.py +++ b/comfy/sampler_helpers.py @@ -1,16 +1,18 @@ from __future__ import annotations +import torch import uuid import math import collections import comfy.model_management import comfy.conds +import comfy.model_patcher import comfy.utils import comfy.hooks import comfy.patcher_extension from typing import TYPE_CHECKING if TYPE_CHECKING: - from comfy.model_patcher import ModelPatcher from comfy.model_base import BaseModel + from comfy.model_patcher import ModelPatcher from comfy.controlnet import ControlBase def prepare_mask(noise_mask, shape, device): @@ -119,6 +121,47 @@ def cleanup_additional_models(models): if hasattr(m, 'cleanup'): m.cleanup() +def preprocess_multigpu_conds(conds: dict[str, list[dict[str]]], model: ModelPatcher, model_options: dict[str]): + '''If multigpu acceleration required, creates deepclones of ControlNets and GLIGEN per device.''' + multigpu_models: list[ModelPatcher] = model.get_additional_models_with_key("multigpu") + if len(multigpu_models) == 0: + return + extra_devices = [x.load_device for x in multigpu_models] + # handle controlnets + controlnets: set[ControlBase] = set() + for k in conds: + for kk in conds[k]: + if 'control' in kk: + controlnets.add(kk['control']) + if len(controlnets) > 0: + # first, unload all controlnet clones + for cnet in list(controlnets): + cnet_models = cnet.get_models() + for cm in cnet_models: + comfy.model_management.unload_model_and_clones(cm, unload_additional_models=True) + + # next, make sure each controlnet has a deepclone for all relevant devices + for cnet in controlnets: + curr_cnet = cnet + while curr_cnet is not None: + for device in extra_devices: + if device not in curr_cnet.multigpu_clones: + curr_cnet.deepclone_multigpu(device, autoregister=True) + curr_cnet = curr_cnet.previous_controlnet + # since all device clones are now present, recreate the linked list for cloned cnets per device + for cnet in controlnets: + curr_cnet = cnet + while curr_cnet is not None: + prev_cnet = curr_cnet.previous_controlnet + for device in extra_devices: + device_cnet = curr_cnet.get_instance_for_device(device) + prev_device_cnet = None + if prev_cnet is not None: + prev_device_cnet = prev_cnet.get_instance_for_device(device) + device_cnet.set_previous_controlnet(prev_device_cnet) + curr_cnet = prev_cnet + # potentially handle gligen - since not widely used, ignored for now + def estimate_memory(model, noise_shape, conds): cond_shapes = collections.defaultdict(list) cond_shapes_min = {} @@ -143,7 +186,8 @@ def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None return executor.execute(model, noise_shape, conds, model_options=model_options, force_full_load=force_full_load, force_offload=force_offload) def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False, force_offload=False): - real_model: BaseModel = None + model.match_multigpu_clones() + preprocess_multigpu_conds(conds, model, model_options) models, inference_memory = get_additional_models(conds, model.model_dtype()) models += get_additional_models_from_model_options(model_options) models += model.get_nested_additional_models() # TODO: does this require inference_memory update? @@ -155,7 +199,7 @@ def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=Non memory_required += inference_memory minimum_memory_required += inference_memory comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required, minimum_memory_required=minimum_memory_required, force_full_load=force_full_load) - real_model = model.model + real_model: BaseModel = model.model return real_model, conds, models @@ -201,3 +245,18 @@ def prepare_model_patcher(model: ModelPatcher, conds, model_options: dict): comfy.patcher_extension.merge_nested_dicts(to_load_options.setdefault(wc_name, {}), model_options["transformer_options"][wc_name], copy_dict1=False) return to_load_options + +def prepare_model_patcher_multigpu_clones(model_patcher: ModelPatcher, loaded_models: list[ModelPatcher], model_options: dict): + ''' + In case multigpu acceleration is enabled, prep ModelPatchers for each device. + ''' + multigpu_patchers: list[ModelPatcher] = [x for x in loaded_models if x.is_multigpu_base_clone] + if len(multigpu_patchers) > 0: + multigpu_dict: dict[torch.device, ModelPatcher] = {} + multigpu_dict[model_patcher.load_device] = model_patcher + for x in multigpu_patchers: + x.hook_patches = comfy.model_patcher.create_hook_patches_clone(model_patcher.hook_patches, copy_tuples=True) + x.hook_mode = model_patcher.hook_mode # match main model's hook_mode + multigpu_dict[x.load_device] = x + model_options["multigpu_clones"] = multigpu_dict + return multigpu_patchers diff --git a/comfy/samplers.py b/comfy/samplers.py index 0a4d062db..25c5a855f 100755 --- a/comfy/samplers.py +++ b/comfy/samplers.py @@ -1,7 +1,9 @@ from __future__ import annotations + +import comfy.model_management from .k_diffusion import sampling as k_diffusion_sampling from .extra_samplers import uni_pc -from typing import TYPE_CHECKING, Callable, NamedTuple +from typing import TYPE_CHECKING, Callable, NamedTuple, Any if TYPE_CHECKING: from comfy.model_patcher import ModelPatcher from comfy.model_base import BaseModel @@ -16,6 +18,7 @@ import comfy.model_patcher import comfy.patcher_extension import comfy.hooks import comfy.context_windows +import comfy.multigpu import comfy.utils import scipy.stats import numpy @@ -141,7 +144,7 @@ def can_concat_cond(c1, c2): return cond_equal_size(c1.conditioning, c2.conditioning) -def cond_cat(c_list): +def cond_cat(c_list, device=None): temp = {} for x in c_list: for k in x: @@ -153,6 +156,8 @@ def cond_cat(c_list): for k in temp: conds = temp[k] out[k] = conds[0].concat(conds[1:]) + if device is not None and hasattr(out[k], 'to'): + out[k] = out[k].to(device) return out @@ -212,7 +217,12 @@ def _calc_cond_batch_outer(model: BaseModel, conds: list[list[dict]], x_in: torc ) return executor.execute(model, conds, x_in, timestep, model_options) -def _calc_cond_batch(model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options): +def _calc_cond_batch(model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]): + # NOTE: keep in sync with _calc_cond_batch_multigpu below. Shared logic + # (hooked_to_run accumulation, memory-fit batching, per-chunk output + # aggregation) is duplicated there with per-device scheduling layered on top. + if 'multigpu_clones' in model_options: + return _calc_cond_batch_multigpu(model, conds, x_in, timestep, model_options) out_conds = [] out_counts = [] # separate conds by matching hooks @@ -244,7 +254,7 @@ def _calc_cond_batch(model: BaseModel, conds: list[list[dict]], x_in: torch.Tens if has_default_conds: finalize_default_conds(model, hooked_to_run, default_conds, x_in, timestep, model_options) - model.current_patcher.prepare_state(timestep) + model.current_patcher.prepare_state(timestep, model_options) # run every hooked_to_run separately for hooks, to_run in hooked_to_run.items(): @@ -265,7 +275,6 @@ def _calc_cond_batch(model: BaseModel, conds: list[list[dict]], x_in: torch.Tens input_shape = [len(batch_amount) * first_shape[0]] + list(first_shape)[1:] cond_shapes = collections.defaultdict(list) for tt in batch_amount: - cond = {k: v.size() for k, v in to_run[tt][0].conditioning.items()} for k, v in to_run[tt][0].conditioning.items(): cond_shapes[k].append(v.size()) @@ -345,6 +354,236 @@ def _calc_cond_batch(model: BaseModel, conds: list[list[dict]], x_in: torch.Tens return out_conds +def _calc_cond_batch_multigpu(model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]): + # NOTE: keep in sync with _calc_cond_batch above. Same conds-by-hooks + # accumulation, memory-fit batching, and output aggregation, but adds a + # per-device scheduler, per-device patcher/control lookup, tensor .to(device) + # placement, and MultiGPUThreadPool dispatch around the inner loop. + out_conds = [] + out_counts = [] + # separate conds by matching hooks + hooked_to_run: dict[comfy.hooks.HookGroup,list[tuple[tuple,int]]] = {} + default_conds = [] + has_default_conds = False + + output_device = x_in.device + + for i in range(len(conds)): + out_conds.append(torch.zeros_like(x_in)) + out_counts.append(torch.ones_like(x_in) * 1e-37) + + cond = conds[i] + default_c = [] + if cond is not None: + for x in cond: + if 'default' in x: + default_c.append(x) + has_default_conds = True + continue + p = get_area_and_mult(x, x_in, timestep) + if p is None: + continue + if p.hooks is not None: + model.current_patcher.prepare_hook_patches_current_keyframe(timestep, p.hooks, model_options) + hooked_to_run.setdefault(p.hooks, list()) + hooked_to_run[p.hooks] += [(p, i)] + default_conds.append(default_c) + + if has_default_conds: + finalize_default_conds(model, hooked_to_run, default_conds, x_in, timestep, model_options) + + model.current_patcher.prepare_state(timestep, model_options) + + devices = list(model_options['multigpu_clones'].keys()) + device_batched_hooked_to_run: dict[torch.device, list[tuple[comfy.hooks.HookGroup, tuple]]] = {} + # Track conds currently scheduled per device; single source of truth for capacity checks. + device_load: dict[torch.device, int] = {d: 0 for d in devices} + + total_conds = sum(len(to_run) for to_run in hooked_to_run.values()) + conds_per_device = max(1, math.ceil(total_conds / len(devices))) + + def next_available_device(start: int) -> tuple[int, torch.device]: + """Return (index, device) for the next device with remaining capacity, starting at `start`. + + Scans at most len(devices) positions, so this always terminates. Raises if no device + has remaining capacity, which would indicate a bug in conds_per_device accounting. + """ + for offset in range(len(devices)): + i = (start + offset) % len(devices) + if device_load[devices[i]] < conds_per_device: + return i, devices[i] + raise RuntimeError( + f"MultiGPU scheduler: all {len(devices)} devices at capacity " + f"({conds_per_device}) but conds remain to schedule" + ) + + # run every hooked_to_run separately + index_device = 0 + for hooks, to_run in hooked_to_run.items(): + while len(to_run) > 0: + index_device, current_device = next_available_device(index_device) + remaining_capacity = conds_per_device - device_load[current_device] + + first = to_run[0] + first_shape = first[0][0].shape + # collect candidate indices that can be concatenated with `first`, up to remaining capacity + to_batch_temp = [] + for x in range(len(to_run)): + if can_concat_cond(to_run[x][0], first[0]) and len(to_batch_temp) < remaining_capacity: + to_batch_temp += [x] + + to_batch_temp.reverse() + to_batch = to_batch_temp[:1] + + free_memory = comfy.model_management.get_free_memory(current_device) + for i in range(1, len(to_batch_temp) + 1): + batch_amount = to_batch_temp[:len(to_batch_temp)//i] + input_shape = [len(batch_amount) * first_shape[0]] + list(first_shape)[1:] + cond_shapes = collections.defaultdict(list) + for tt in batch_amount: + for k, v in to_run[tt][0].conditioning.items(): + cond_shapes[k].append(v.size()) + if model.memory_required(input_shape, cond_shapes=cond_shapes) * 1.5 < free_memory: + to_batch = batch_amount + break + + conds_to_batch = [to_run.pop(x) for x in to_batch] + device_load[current_device] += len(conds_to_batch) + device_batched_hooked_to_run.setdefault(current_device, []).append((hooks, conds_to_batch)) + + if device_load[current_device] >= conds_per_device: + index_device += 1 + + class thread_result(NamedTuple): + output: Any + mult: Any + area: Any + batch_chunks: int + cond_or_uncond: Any + error: Exception = None + + def _handle_batch(device: torch.device, batch_tuple: tuple[comfy.hooks.HookGroup, tuple], results: list[thread_result]): + try: + comfy.model_management.set_torch_device(device) + model_current: BaseModel = model_options["multigpu_clones"][device].model + # run every hooked_to_run separately + with torch.no_grad(): + for hooks, to_batch in batch_tuple: + input_x = [] + mult = [] + c = [] + cond_or_uncond = [] + uuids = [] + area = [] + control: ControlBase = None + patches = None + for x in to_batch: + o = x + p = o[0] + input_x.append(p.input_x) + mult.append(p.mult) + c.append(p.conditioning) + area.append(p.area) + cond_or_uncond.append(o[1]) + uuids.append(p.uuid) + control = p.control + patches = p.patches + + batch_chunks = len(cond_or_uncond) + input_x = torch.cat(input_x).to(device) + c = cond_cat(c, device=device) + timestep_ = torch.cat([timestep.to(device)] * batch_chunks) + + transformer_options = model_current.current_patcher.apply_hooks(hooks=hooks) + if 'transformer_options' in model_options: + transformer_options = comfy.patcher_extension.merge_nested_dicts(transformer_options, + model_options['transformer_options'], + copy_dict1=False) + + if patches is not None: + transformer_options["patches"] = comfy.patcher_extension.merge_nested_dicts( + transformer_options.get("patches", {}), + patches + ) + + transformer_options["cond_or_uncond"] = cond_or_uncond[:] + transformer_options["uuids"] = uuids[:] + transformer_options["sigmas"] = timestep.to(device) + transformer_options["sample_sigmas"] = transformer_options["sample_sigmas"].to(device) + transformer_options["multigpu_thread_device"] = device + + cast_transformer_options(transformer_options, device=device) + c['transformer_options'] = transformer_options + + if control is not None: + device_control = control.get_instance_for_device(device) + c['control'] = device_control.get_control(input_x, timestep_, c, len(cond_or_uncond), transformer_options) + + if 'model_function_wrapper' in model_options: + output = model_options['model_function_wrapper'](model_current.apply_model, {"input": input_x, "timestep": timestep_, "c": c, "cond_or_uncond": cond_or_uncond}).to(output_device).chunk(batch_chunks) + else: + output = model_current.apply_model(input_x, timestep_, **c).to(output_device).chunk(batch_chunks) + # TODO: non-NVIDIA support -- the `.to(output_device)` copies + # above are async on CUDA, so the main thread's aggregation + # could race with in-flight transfers. CUDA-only QA has not + # surfaced this in practice, but before extending multigpu + # beyond NVIDIA add a `torch.cuda.synchronize(output_device)` + # here (guarded by `output_device.type == "cuda"`). + results.append(thread_result(output, mult, area, batch_chunks, cond_or_uncond)) + except Exception as e: + results.append(thread_result(None, None, None, None, None, error=e)) + raise + + + def _handle_batch_pooled(device, batch_tuple): + worker_results = [] + _handle_batch(device, batch_tuple, worker_results) + return worker_results + + results: list[thread_result] = [] + thread_pool: comfy.multigpu.MultiGPUThreadPool = model_options.get("multigpu_thread_pool") + + # Submit all GPU work to pool threads + pool_devices = [] + for device, batch_tuple in device_batched_hooked_to_run.items(): + if thread_pool is not None: + thread_pool.submit(device, _handle_batch_pooled, device, batch_tuple) + pool_devices.append(device) + else: + # Fallback: no pool, run everything on main thread + _handle_batch(device, batch_tuple, results) + + # Collect results from pool workers + for device in pool_devices: + worker_results, error = thread_pool.get_result(device) + if error is not None: + raise error + results.extend(worker_results) + + for output, mult, area, batch_chunks, cond_or_uncond, error in results: + if error is not None: + raise error + for o in range(batch_chunks): + cond_index = cond_or_uncond[o] + a = area[o] + if a is None: + out_conds[cond_index] += output[o] * mult[o] + out_counts[cond_index] += mult[o] + else: + out_c = out_conds[cond_index] + out_cts = out_counts[cond_index] + dims = len(a) // 2 + for i in range(dims): + out_c = out_c.narrow(i + 2, a[i + dims], a[i]) + out_cts = out_cts.narrow(i + 2, a[i + dims], a[i]) + out_c += output[o] * mult[o] + out_cts += mult[o] + + for i in range(len(out_conds)): + out_conds[i] /= out_counts[i] + + return out_conds + def calc_cond_uncond_batch(model, cond, uncond, x_in, timestep, model_options): #TODO: remove logging.warning("WARNING: The comfy.samplers.calc_cond_uncond_batch function is deprecated please use the calc_cond_batch one instead.") return tuple(calc_cond_batch(model, [cond, uncond], x_in, timestep, model_options)) @@ -643,12 +882,21 @@ def calculate_start_end_timesteps(model, conds): def pre_run_control(model, conds): s = model.model_sampling + # Per-device model lookup so multigpu control clones get the matching + # diffusion_model (e.g. QwenFunControlNet stashes it into extra_args). + device_models: dict = {} + patcher = getattr(model, "current_patcher", None) + if patcher is not None: + for p in patcher.get_additional_models_with_key("multigpu"): + device_models[p.load_device] = p.model for t in range(len(conds)): x = conds[t] percent_to_timestep_function = lambda a: s.percent_to_sigma(a) if 'control' in x: x['control'].pre_run(model, percent_to_timestep_function) + for device, device_cnet in x['control'].multigpu_clones.items(): + device_cnet.pre_run(device_models.get(device, model), percent_to_timestep_function) def apply_empty_x_to_equal_area(conds, uncond, name, uncond_fill_func): cond_cnets = [] @@ -891,7 +1139,9 @@ def cast_to_load_options(model_options: dict[str], device=None, dtype=None): to_load_options = model_options.get("to_load_options", None) if to_load_options is None: return + cast_transformer_options(to_load_options, device, dtype) +def cast_transformer_options(transformer_options: dict[str], device=None, dtype=None): casts = [] if device is not None: casts.append(device) @@ -900,18 +1150,17 @@ def cast_to_load_options(model_options: dict[str], device=None, dtype=None): # if nothing to apply, do nothing if len(casts) == 0: return - # try to call .to on patches - if "patches" in to_load_options: - patches = to_load_options["patches"] + if "patches" in transformer_options: + patches = transformer_options["patches"] for name in patches: patch_list = patches[name] for i in range(len(patch_list)): if hasattr(patch_list[i], "to"): for cast in casts: patch_list[i] = patch_list[i].to(cast) - if "patches_replace" in to_load_options: - patches = to_load_options["patches_replace"] + if "patches_replace" in transformer_options: + patches = transformer_options["patches_replace"] for name in patches: patch_list = patches[name] for k in patch_list: @@ -921,8 +1170,8 @@ def cast_to_load_options(model_options: dict[str], device=None, dtype=None): # try to call .to on any wrappers/callbacks wrappers_and_callbacks = ["wrappers", "callbacks"] for wc_name in wrappers_and_callbacks: - if wc_name in to_load_options: - wc: dict[str, list] = to_load_options[wc_name] + if wc_name in transformer_options: + wc: dict[str, list] = transformer_options[wc_name] for wc_dict in wc.values(): for wc_list in wc_dict.values(): for i in range(len(wc_list)): @@ -930,7 +1179,6 @@ def cast_to_load_options(model_options: dict[str], device=None, dtype=None): for cast in casts: wc_list[i] = wc_list[i].to(cast) - class CFGGuider: def __init__(self, model_patcher: ModelPatcher): self.model_patcher = model_patcher @@ -985,16 +1233,32 @@ class CFGGuider: self.inner_model, self.conds, self.loaded_models = comfy.sampler_helpers.prepare_sampling(self.model_patcher, noise.shape, self.conds, self.model_options) device = self.model_patcher.load_device - noise = noise.to(device=device, dtype=torch.float32) - latent_image = latent_image.to(device=device, dtype=torch.float32) - sigmas = sigmas.to(device) - cast_to_load_options(self.model_options, device=device, dtype=self.model_patcher.model_dtype()) + multigpu_patchers = comfy.sampler_helpers.prepare_model_patcher_multigpu_clones(self.model_patcher, self.loaded_models, self.model_options) - try: - self.model_patcher.pre_run() - output = self.inner_sample(noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=latent_shapes) - finally: - self.model_patcher.cleanup() + # Create persistent thread pool for all GPU devices (main + extras) + if multigpu_patchers: + extra_devices = [p.load_device for p in multigpu_patchers] + all_devices = [device] + extra_devices + self.model_options["multigpu_thread_pool"] = comfy.multigpu.MultiGPUThreadPool(all_devices) + + with comfy.model_management.cuda_device_context(device): + try: + noise = noise.to(device=device, dtype=torch.float32) + latent_image = latent_image.to(device=device, dtype=torch.float32) + sigmas = sigmas.to(device) + cast_to_load_options(self.model_options, device=device, dtype=self.model_patcher.model_dtype()) + + self.model_patcher.pre_run() + for multigpu_patcher in multigpu_patchers: + multigpu_patcher.pre_run() + output = self.inner_sample(noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=latent_shapes) + finally: + thread_pool = self.model_options.pop("multigpu_thread_pool", None) + if thread_pool is not None: + thread_pool.shutdown() + self.model_patcher.cleanup() + for multigpu_patcher in multigpu_patchers: + multigpu_patcher.cleanup() comfy.sampler_helpers.cleanup_models(self.conds, self.loaded_models) del self.inner_model diff --git a/comfy/sd.py b/comfy/sd.py index 309fb7763..bce0d0bf8 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -1,4 +1,3 @@ -from __future__ import annotations import json import torch from enum import Enum @@ -18,6 +17,7 @@ import comfy.ldm.wan.vae import comfy.ldm.trellis2.vae import comfy.ldm.wan.vae2_2 import comfy.ldm.hunyuan3d.vae +import comfy.ldm.triposplat.vae import comfy.ldm.ace.vae.music_dcae_pipeline import comfy.ldm.cogvideo.vae import comfy.ldm.hunyuan_video.vae @@ -51,6 +51,7 @@ import comfy.text_encoders.lt import comfy.text_encoders.hunyuan_video import comfy.text_encoders.cosmos import comfy.text_encoders.lumina2 +import comfy.text_encoders.pixeldit import comfy.text_encoders.wan import comfy.text_encoders.hidream import comfy.text_encoders.ace @@ -58,6 +59,7 @@ import comfy.text_encoders.omnigen2 import comfy.text_encoders.qwen_image import comfy.text_encoders.hunyuan_image import comfy.text_encoders.z_image +import comfy.text_encoders.ideogram4 import comfy.text_encoders.ovis import comfy.text_encoders.kandinsky5 import comfy.text_encoders.jina_clip_2 @@ -70,6 +72,7 @@ import comfy.text_encoders.ernie import comfy.text_encoders.gemma4 import comfy.text_encoders.cogvideo import comfy.text_encoders.sa3 +import comfy.text_encoders.gpt_oss import comfy.model_patcher import comfy.lora @@ -336,41 +339,43 @@ class CLIP: self.cond_stage_model.set_clip_options({"projected_pooled": False}) self.load_model(tokens) - self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device}) + device = self.patcher.load_device + self.cond_stage_model.set_clip_options({"execution_device": device}) all_hooks.reset() self.patcher.patch_hooks(None) if show_pbar: pbar = ProgressBar(len(scheduled_keyframes)) - for scheduled_opts in scheduled_keyframes: - t_range = scheduled_opts[0] - # don't bother encoding any conds outside of start_percent and end_percent bounds - if "start_percent" in add_dict: - if t_range[1] < add_dict["start_percent"]: - continue - if "end_percent" in add_dict: - if t_range[0] > add_dict["end_percent"]: - continue - hooks_keyframes = scheduled_opts[1] - for hook, keyframe in hooks_keyframes: - hook.hook_keyframe._current_keyframe = keyframe - # apply appropriate hooks with values that match new hook_keyframe - self.patcher.patch_hooks(all_hooks) - # perform encoding as normal - o = self.cond_stage_model.encode_token_weights(tokens) - cond, pooled = o[:2] - pooled_dict = {"pooled_output": pooled} - # add clip_start_percent and clip_end_percent in pooled - pooled_dict["clip_start_percent"] = t_range[0] - pooled_dict["clip_end_percent"] = t_range[1] - # add/update any keys with the provided add_dict - pooled_dict.update(add_dict) - # add hooks stored on clip - self.add_hooks_to_dict(pooled_dict) - all_cond_pooled.append([cond, pooled_dict]) - if show_pbar: - pbar.update(1) - model_management.throw_exception_if_processing_interrupted() + with model_management.cuda_device_context(device): + for scheduled_opts in scheduled_keyframes: + t_range = scheduled_opts[0] + # don't bother encoding any conds outside of start_percent and end_percent bounds + if "start_percent" in add_dict: + if t_range[1] < add_dict["start_percent"]: + continue + if "end_percent" in add_dict: + if t_range[0] > add_dict["end_percent"]: + continue + hooks_keyframes = scheduled_opts[1] + for hook, keyframe in hooks_keyframes: + hook.hook_keyframe._current_keyframe = keyframe + # apply appropriate hooks with values that match new hook_keyframe + self.patcher.patch_hooks(all_hooks) + # perform encoding as normal + o = self.cond_stage_model.encode_token_weights(tokens) + cond, pooled = o[:2] + pooled_dict = {"pooled_output": pooled} + # add clip_start_percent and clip_end_percent in pooled + pooled_dict["clip_start_percent"] = t_range[0] + pooled_dict["clip_end_percent"] = t_range[1] + # add/update any keys with the provided add_dict + pooled_dict.update(add_dict) + # add hooks stored on clip + self.add_hooks_to_dict(pooled_dict) + all_cond_pooled.append([cond, pooled_dict]) + if show_pbar: + pbar.update(1) + model_management.throw_exception_if_processing_interrupted() all_hooks.reset() return all_cond_pooled @@ -384,8 +389,12 @@ class CLIP: self.cond_stage_model.set_clip_options({"projected_pooled": False}) self.load_model(tokens) - self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device}) - o = self.cond_stage_model.encode_token_weights(tokens) + device = self.patcher.load_device + self.cond_stage_model.set_clip_options({"execution_device": device}) + + with model_management.cuda_device_context(device): + o = self.cond_stage_model.encode_token_weights(tokens) + cond, pooled = o[:2] if return_dict: out = {"cond": cond, "pooled_output": pooled} @@ -447,9 +456,12 @@ class CLIP: self.cond_stage_model.reset_clip_options() self.load_model(tokens) + device = self.patcher.load_device self.cond_stage_model.set_clip_options({"layer": None}) - self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device}) - return self.cond_stage_model.generate(tokens, do_sample=do_sample, max_length=max_length, temperature=temperature, top_k=top_k, top_p=top_p, min_p=min_p, repetition_penalty=repetition_penalty, seed=seed, presence_penalty=presence_penalty) + self.cond_stage_model.set_clip_options({"execution_device": device}) + + with model_management.cuda_device_context(device): + return self.cond_stage_model.generate(tokens, do_sample=do_sample, max_length=max_length, temperature=temperature, top_k=top_k, top_p=top_p, min_p=min_p, repetition_penalty=repetition_penalty, seed=seed, presence_penalty=presence_penalty) def decode(self, token_ids, skip_special_tokens=True): return self.tokenizer.decode(token_ids, skip_special_tokens=skip_special_tokens) @@ -897,6 +909,16 @@ class VAE: #Force cast it for --disable-dynamic-vram users until there is a true core fix. if not comfy.memory_management.aimdo_enabled: self.disable_offload = True + elif "gs.base_offset_scale" in sd and "octree.out_proj.weight" in sd: # TripoSplat octree gaussian decoder + self.first_stage_model = comfy.ldm.triposplat.vae.OctreeGaussianDecoder() + self.latent_channels = 16 + self.latent_dim = 1 + self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32] + # The generic VAE.encode/decode path isn't used: VAEDecodeTripoSplat calls the gaussian + # decoder directly (structured GaussianSplat objects, not a tensor and reserves VRAM itself from num_gaussians. + def _no_generic_io(*args, **kwargs): + raise RuntimeError("TripoSplat gaussian decoder: use the 'TripoSplat Decode' (VAEDecodeTripoSplat)") + self.memory_used_encode = self.memory_used_decode = _no_generic_io else: logging.warning("WARNING: No VAE weights detected, VAE not initalized.") self.first_stage_model = None @@ -1039,50 +1061,52 @@ class VAE: do_tile = False if self.latent_dim == 2 and samples_in.ndim == 5: samples_in = samples_in[:, :, 0] - try: - memory_used = self.memory_used_decode(samples_in.shape, self.vae_dtype) - model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload) - free_memory = self.patcher.get_free_memory(self.device) - batch_number = int(free_memory / memory_used) - batch_number = max(1, batch_number) - # Pre-allocate output for VAEs that support direct buffer writes - preallocated = False - if getattr(self.first_stage_model, 'comfy_has_chunked_io', False): - pixel_samples = torch.empty(self.first_stage_model.decode_output_shape(samples_in.shape), device=self.output_device, dtype=self.vae_output_dtype()) - preallocated = True + with model_management.cuda_device_context(self.device): + try: + memory_used = self.memory_used_decode(samples_in.shape, self.vae_dtype) + model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload) + free_memory = self.patcher.get_free_memory(self.device) + batch_number = int(free_memory / memory_used) + batch_number = max(1, batch_number) - for x in range(0, samples_in.shape[0], batch_number): - samples = samples_in[x:x + batch_number].to(device=self.device, dtype=self.vae_dtype) - if preallocated: - self.first_stage_model.decode(samples, output_buffer=pixel_samples[x:x+batch_number], **vae_options) - else: - out = self.first_stage_model.decode(samples, **vae_options).to(device=self.output_device, dtype=self.vae_output_dtype(), copy=True) - if pixel_samples is None: - pixel_samples = torch.empty((samples_in.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype()) - pixel_samples[x:x+batch_number].copy_(out) - del out - self.process_output(pixel_samples[x:x+batch_number]) - except Exception as e: - model_management.raise_non_oom(e) - logging.warning("Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.") - #NOTE: We don't know what tensors were allocated to stack variables at the time of the - #exception and the exception itself refs them all until we get out of this except block. - #So we just set a flag for tiler fallback so that tensor gc can happen once the - #exception is fully off the books. - do_tile = True + # Pre-allocate output for VAEs that support direct buffer writes + preallocated = False + if getattr(self.first_stage_model, 'comfy_has_chunked_io', False): + pixel_samples = torch.empty(self.first_stage_model.decode_output_shape(samples_in.shape), device=self.output_device, dtype=self.vae_output_dtype()) + preallocated = True - if do_tile: - comfy.model_management.soft_empty_cache() - dims = samples_in.ndim - 2 - if dims == 1 or self.extra_1d_channel is not None: - pixel_samples = self.decode_tiled_1d(samples_in) - elif dims == 2: - pixel_samples = self.decode_tiled_(samples_in) - elif dims == 3: - tile = 256 // self.spacial_compression_decode() - overlap = tile // 4 - pixel_samples = self.decode_tiled_3d(samples_in, tile_x=tile, tile_y=tile, overlap=(1, overlap, overlap)) + for x in range(0, samples_in.shape[0], batch_number): + samples = samples_in[x:x + batch_number].to(device=self.device, dtype=self.vae_dtype) + if preallocated: + self.first_stage_model.decode(samples, output_buffer=pixel_samples[x:x+batch_number], **vae_options) + else: + out = self.first_stage_model.decode(samples, **vae_options).to(device=self.output_device, dtype=self.vae_output_dtype(), copy=True) + if pixel_samples is None: + pixel_samples = torch.empty((samples_in.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype()) + pixel_samples[x:x+batch_number].copy_(out) + del out + self.process_output(pixel_samples[x:x+batch_number]) + except Exception as e: + model_management.raise_non_oom(e) + logging.warning("Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.") + #NOTE: We don't know what tensors were allocated to stack variables at the time of the + #exception and the exception itself refs them all until we get out of this except block. + #So we just set a flag for tiler fallback so that tensor gc can happen once the + #exception is fully off the books. + do_tile = True + + if do_tile: + comfy.model_management.soft_empty_cache() + dims = samples_in.ndim - 2 + if dims == 1 or self.extra_1d_channel is not None: + pixel_samples = self.decode_tiled_1d(samples_in) + elif dims == 2: + pixel_samples = self.decode_tiled_(samples_in) + elif dims == 3: + tile = 256 // self.spacial_compression_decode() + overlap = tile // 4 + pixel_samples = self.decode_tiled_3d(samples_in, tile_x=tile, tile_y=tile, overlap=(1, overlap, overlap)) pixel_samples = pixel_samples.to(self.output_device).movedim(1,-1) return pixel_samples @@ -1100,20 +1124,21 @@ class VAE: if overlap is not None: args["overlap"] = overlap - if dims == 1 or self.extra_1d_channel is not None: - args.pop("tile_y") - output = self.decode_tiled_1d(samples, **args) - elif dims == 2: - output = self.decode_tiled_(samples, **args) - elif dims == 3: - if overlap_t is None: - args["overlap"] = (1, overlap, overlap) - else: - args["overlap"] = (max(1, overlap_t), overlap, overlap) - if tile_t is not None: - args["tile_t"] = max(2, tile_t) + with model_management.cuda_device_context(self.device): + if dims == 1 or self.extra_1d_channel is not None: + args.pop("tile_y") + output = self.decode_tiled_1d(samples, **args) + elif dims == 2: + output = self.decode_tiled_(samples, **args) + elif dims == 3: + if overlap_t is None: + args["overlap"] = (1, overlap, overlap) + else: + args["overlap"] = (max(1, overlap_t), overlap, overlap) + if tile_t is not None: + args["tile_t"] = max(2, tile_t) - output = self.decode_tiled_3d(samples, **args) + output = self.decode_tiled_3d(samples, **args) return output.movedim(1, -1) def encode(self, pixel_samples): @@ -1126,44 +1151,46 @@ class VAE: pixel_samples = pixel_samples.movedim(1, 0).unsqueeze(0) else: pixel_samples = pixel_samples.unsqueeze(2) - try: - memory_used = self.memory_used_encode(pixel_samples.shape, self.vae_dtype) - model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload) - free_memory = self.patcher.get_free_memory(self.device) - batch_number = int(free_memory / max(1, memory_used)) - batch_number = max(1, batch_number) - samples = None - for x in range(0, pixel_samples.shape[0], batch_number): - pixels_in = self.process_input(pixel_samples[x:x + batch_number]).to(self.vae_dtype) - if getattr(self.first_stage_model, 'comfy_has_chunked_io', False): - out = self.first_stage_model.encode(pixels_in, device=self.device) + + with model_management.cuda_device_context(self.device): + try: + memory_used = self.memory_used_encode(pixel_samples.shape, self.vae_dtype) + model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload) + free_memory = self.patcher.get_free_memory(self.device) + batch_number = int(free_memory / max(1, memory_used)) + batch_number = max(1, batch_number) + samples = None + for x in range(0, pixel_samples.shape[0], batch_number): + pixels_in = self.process_input(pixel_samples[x:x + batch_number]).to(self.vae_dtype) + if getattr(self.first_stage_model, 'comfy_has_chunked_io', False): + out = self.first_stage_model.encode(pixels_in, device=self.device) + else: + pixels_in = pixels_in.to(self.device) + out = self.first_stage_model.encode(pixels_in) + out = out.to(self.output_device).to(dtype=self.vae_output_dtype()) + if samples is None: + samples = torch.empty((pixel_samples.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype()) + samples[x:x + batch_number] = out + + except Exception as e: + model_management.raise_non_oom(e) + logging.warning("Warning: Ran out of memory when regular VAE encoding, retrying with tiled VAE encoding.") + #NOTE: We don't know what tensors were allocated to stack variables at the time of the + #exception and the exception itself refs them all until we get out of this except block. + #So we just set a flag for tiler fallback so that tensor gc can happen once the + #exception is fully off the books. + do_tile = True + + if do_tile: + comfy.model_management.soft_empty_cache() + if self.latent_dim == 3: + tile = 256 + overlap = tile // 4 + samples = self.encode_tiled_3d(pixel_samples, tile_x=tile, tile_y=tile, overlap=(1, overlap, overlap)) + elif self.latent_dim == 1 or self.extra_1d_channel is not None: + samples = self.encode_tiled_1d(pixel_samples) else: - pixels_in = pixels_in.to(self.device) - out = self.first_stage_model.encode(pixels_in) - out = out.to(self.output_device).to(dtype=self.vae_output_dtype()) - if samples is None: - samples = torch.empty((pixel_samples.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype()) - samples[x:x + batch_number] = out - - except Exception as e: - model_management.raise_non_oom(e) - logging.warning("Warning: Ran out of memory when regular VAE encoding, retrying with tiled VAE encoding.") - #NOTE: We don't know what tensors were allocated to stack variables at the time of the - #exception and the exception itself refs them all until we get out of this except block. - #So we just set a flag for tiler fallback so that tensor gc can happen once the - #exception is fully off the books. - do_tile = True - - if do_tile: - comfy.model_management.soft_empty_cache() - if self.latent_dim == 3: - tile = 256 - overlap = tile // 4 - samples = self.encode_tiled_3d(pixel_samples, tile_x=tile, tile_y=tile, overlap=(1, overlap, overlap)) - elif self.latent_dim == 1 or self.extra_1d_channel is not None: - samples = self.encode_tiled_1d(pixel_samples) - else: - samples = self.encode_tiled_(pixel_samples) + samples = self.encode_tiled_(pixel_samples) return samples @@ -1189,26 +1216,27 @@ class VAE: if overlap is not None: args["overlap"] = overlap - if dims == 1: - args.pop("tile_y") - samples = self.encode_tiled_1d(pixel_samples, **args) - elif dims == 2: - samples = self.encode_tiled_(pixel_samples, **args) - elif dims == 3: - if tile_t is not None: - tile_t_latent = max(2, self.downscale_ratio[0](tile_t)) - else: - tile_t_latent = 9999 - args["tile_t"] = self.upscale_ratio[0](tile_t_latent) + with model_management.cuda_device_context(self.device): + if dims == 1: + args.pop("tile_y") + samples = self.encode_tiled_1d(pixel_samples, **args) + elif dims == 2: + samples = self.encode_tiled_(pixel_samples, **args) + elif dims == 3: + if tile_t is not None: + tile_t_latent = max(2, self.downscale_ratio[0](tile_t)) + else: + tile_t_latent = 9999 + args["tile_t"] = self.upscale_ratio[0](tile_t_latent) - if overlap_t is None: - args["overlap"] = (1, overlap, overlap) - else: - args["overlap"] = (self.upscale_ratio[0](max(1, min(tile_t_latent // 2, self.downscale_ratio[0](overlap_t)))), overlap, overlap) - maximum = pixel_samples.shape[2] - maximum = self.upscale_ratio[0](self.downscale_ratio[0](maximum)) + if overlap_t is None: + args["overlap"] = (1, overlap, overlap) + else: + args["overlap"] = (self.upscale_ratio[0](max(1, min(tile_t_latent // 2, self.downscale_ratio[0](overlap_t)))), overlap, overlap) + maximum = pixel_samples.shape[2] + maximum = self.upscale_ratio[0](self.downscale_ratio[0](maximum)) - samples = self.encode_tiled_3d(pixel_samples[:,:,:maximum], **args) + samples = self.encode_tiled_3d(pixel_samples[:,:,:maximum], **args) return samples @@ -1282,6 +1310,9 @@ class CLIPType(Enum): FLUX2 = 25 LONGCAT_IMAGE = 26 COGVIDEOX = 27 + LENS = 28 + PIXELDIT = 29 + IDEOGRAM4 = 30 @@ -1334,6 +1365,7 @@ class TEModel(Enum): GEMMA_4_E2B = 30 GEMMA_4_31B = 31 T5_GEMMA = 32 + GPT_OSS_20B = 33 def detect_te_model(sd): @@ -1375,6 +1407,9 @@ def detect_te_model(sd): else: return TEModel.GEMMA_3_4B return TEModel.GEMMA_2_2B + # Must precede the Qwen2.5-7B k_proj.bias=512 check (GPT-OSS also has 8*64=512). + if "layers.0.self_attn.sinks" in sd and "layers.0.mlp.experts.gate_up_proj.weight" in sd: + return TEModel.GPT_OSS_20B if 'model.layers.0.self_attn.k_proj.bias' in sd: weight = sd['model.layers.0.self_attn.k_proj.bias'] if weight.shape[0] == 256: @@ -1521,8 +1556,12 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip clip_target.tokenizer = variant.tokenizer tokenizer_data["tokenizer_json"] = clip_data[0].get("tokenizer_json", None) elif te_model == TEModel.GEMMA_2_2B: - clip_target.clip = comfy.text_encoders.lumina2.te(**llama_detect(clip_data)) - clip_target.tokenizer = comfy.text_encoders.lumina2.LuminaTokenizer + if clip_type == CLIPType.PIXELDIT: + clip_target.clip = comfy.text_encoders.pixeldit.pixeldit_te(**llama_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.pixeldit.PixelDiTGemma2Tokenizer + else: + clip_target.clip = comfy.text_encoders.lumina2.te(**llama_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.lumina2.LuminaTokenizer tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None) elif te_model == TEModel.GEMMA_3_4B: clip_target.clip = comfy.text_encoders.lumina2.te(**llama_detect(clip_data), model_type="gemma3_4b") @@ -1557,6 +1596,10 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip clip_target.clip = comfy.text_encoders.flux.flux2_te(**llama_detect(clip_data), pruned=te_model == TEModel.MISTRAL3_24B_PRUNED_FLUX2) clip_target.tokenizer = comfy.text_encoders.flux.Flux2Tokenizer tokenizer_data["tekken_model"] = clip_data[0].get("tekken_model", None) + elif te_model == TEModel.GPT_OSS_20B: + clip_target.clip = comfy.text_encoders.gpt_oss.lens_te(**llama_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.gpt_oss.LensTokenizer + tokenizer_data["tokenizer_json"] = clip_data[0].get("tokenizer_json", None) elif te_model == TEModel.QWEN3_4B: if clip_type == CLIPType.FLUX or clip_type == CLIPType.FLUX2: clip_target.clip = comfy.text_encoders.flux.klein_te(**llama_detect(clip_data), model_type="qwen3_4b") @@ -1568,8 +1611,12 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip clip_target.clip = comfy.text_encoders.ovis.te(**llama_detect(clip_data)) clip_target.tokenizer = comfy.text_encoders.ovis.OvisTokenizer elif te_model == TEModel.QWEN3_8B: - clip_target.clip = comfy.text_encoders.flux.klein_te(**llama_detect(clip_data), model_type="qwen3_8b") - clip_target.tokenizer = comfy.text_encoders.flux.KleinTokenizer8B + if clip_type == CLIPType.IDEOGRAM4: + clip_target.clip = comfy.text_encoders.ideogram4.te(**llama_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.ideogram4.Ideogram4Tokenizer + else: + clip_target.clip = comfy.text_encoders.flux.klein_te(**llama_detect(clip_data), model_type="qwen3_8b") + clip_target.tokenizer = comfy.text_encoders.flux.KleinTokenizer8B elif te_model == TEModel.JINA_CLIP_2: clip_target.clip = comfy.text_encoders.jina_clip_2.JinaClip2TextModelWrapper clip_target.tokenizer = comfy.text_encoders.jina_clip_2.JinaClip2TokenizerWrapper @@ -1723,12 +1770,52 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o out = load_state_dict_guess_config(sd, output_vae, output_clip, output_clipvision, embedding_directory, output_model, model_options, te_model_options=te_model_options, metadata=metadata, disable_dynamic=disable_dynamic) if out is None: raise RuntimeError("ERROR: Could not detect model type of: {}\n{}".format(ckpt_path, model_detection_error_hint(ckpt_path, sd))) - if output_model and out[0] is not None: - out[0].cached_patcher_init = (load_checkpoint_guess_config_model_only, (ckpt_path, embedding_directory, model_options, te_model_options)) - if output_clip and out[1] is not None: - out[1].patcher.cached_patcher_init = (load_checkpoint_guess_config_clip_only, (ckpt_path, embedding_directory, model_options, te_model_options)) + if out[0] is not None: + out[0].cached_patcher_init = (load_checkpoint_guess_config, (ckpt_path, False, False, False, embedding_directory, output_model, model_options, te_model_options), 0) + # Register reload factories for the CLIP and VAE produced by the same checkpoint so + # ModelPatcher.deepclone_multigpu can spawn per-device copies (Select{CLIP,VAE}Device, + # MultiGPU work-units, etc.) without falling back to copy.deepcopy of an + # already-loaded module. + if out[1] is not None and getattr(out[1], "patcher", None) is not None: + out[1].patcher.cached_patcher_init = (load_checkpoint_clip_patcher, (ckpt_path, embedding_directory, model_options, te_model_options)) + if out[2] is not None and getattr(out[2], "patcher", None) is not None: + out[2].patcher.cached_patcher_init = (load_checkpoint_vae_patcher, (ckpt_path, embedding_directory, model_options, te_model_options)) return out + +def load_checkpoint_clip_patcher(ckpt_path, embedding_directory=None, model_options={}, te_model_options={}, disable_dynamic=False): + """Reload only the CLIP patcher from a checkpoint. Used as the cached_patcher_init + factory for the CLIP returned by load_checkpoint_guess_config.""" + _, clip, _, _ = load_checkpoint_guess_config( + ckpt_path, + output_vae=False, + output_clip=True, + output_clipvision=False, + embedding_directory=embedding_directory, + output_model=False, + model_options=model_options, + te_model_options=te_model_options, + disable_dynamic=disable_dynamic, + ) + return clip.patcher + + +def load_checkpoint_vae_patcher(ckpt_path, embedding_directory=None, model_options={}, te_model_options={}, disable_dynamic=False): + """Reload only the VAE patcher from a checkpoint. Used as the cached_patcher_init + factory for the VAE returned by load_checkpoint_guess_config.""" + _, _, vae, _ = load_checkpoint_guess_config( + ckpt_path, + output_vae=True, + output_clip=False, + output_clipvision=False, + embedding_directory=embedding_directory, + output_model=False, + model_options=model_options, + te_model_options=te_model_options, + disable_dynamic=disable_dynamic, + ) + return vae.patcher + def load_checkpoint_guess_config_model_only(ckpt_path, embedding_directory=None, model_options={}, te_model_options={}, disable_dynamic=False): model, *_ = load_checkpoint_guess_config(ckpt_path, False, False, False, embedding_directory=embedding_directory, @@ -1755,7 +1842,7 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c diffusion_model_prefix = model_detection.unet_prefix_from_state_dict(sd) parameters = comfy.utils.calculate_parameters(sd, diffusion_model_prefix) weight_dtype = comfy.utils.weight_dtype(sd, diffusion_model_prefix) - load_device = model_management.get_torch_device() + load_device = model_options.get("load_device", model_management.get_torch_device()) custom_operations = model_options.get("custom_operations", None) if custom_operations is None: @@ -1795,13 +1882,15 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c inital_load_device = model_management.unet_inital_load_device(parameters, unet_dtype) model = model_config.get_model(sd, diffusion_model_prefix, device=inital_load_device) ModelPatcher = comfy.model_patcher.ModelPatcher if disable_dynamic else comfy.model_patcher.CoreModelPatcher - model_patcher = ModelPatcher(model, load_device=load_device, offload_device=model_management.unet_offload_device()) + offload_device = model_options.get("offload_device", model_management.unet_offload_device()) + model_patcher = ModelPatcher(model, load_device=load_device, offload_device=offload_device) model.load_model_weights(sd, diffusion_model_prefix, assign=model_patcher.is_dynamic()) if output_vae: vae_sd = comfy.utils.state_dict_prefix_replace(sd, {k: "" for k in model_config.vae_key_prefix}, filter_keys=True) vae_sd = model_config.process_vae_state_dict(vae_sd) - vae = VAE(sd=vae_sd, metadata=metadata) + vae_device = model_options.get("load_device", None) + vae = VAE(sd=vae_sd, metadata=metadata, device=vae_device) if output_clip: if te_model_options.get("custom_operations", None) is None: @@ -1885,7 +1974,7 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None, disable parameters = comfy.utils.calculate_parameters(sd) weight_dtype = comfy.utils.weight_dtype(sd) - load_device = model_management.get_torch_device() + load_device = model_options.get("load_device", model_management.get_torch_device()) model_config = model_detection.model_config_from_unet(sd, "", metadata=metadata) if model_config is not None: @@ -1910,7 +1999,7 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None, disable else: logging.warning("{} {}".format(diffusers_keys[k], k)) - offload_device = model_management.unet_offload_device() + offload_device = model_options.get("offload_device", model_management.unet_offload_device()) unet_weight_dtype = list(model_config.supported_inference_dtypes) if model_config.quant_config is not None: weight_dtype = None @@ -1952,6 +2041,26 @@ def load_diffusion_model(unet_path, model_options={}, disable_dynamic=False): model.cached_patcher_init = (load_diffusion_model, (unet_path, model_options)) return model + +def load_vae_patcher(vae_path, metadata=None, device=None, disable_dynamic=False): + """Reload a disk-backed VAE from ``vae_path`` and return its patcher. + + Used as the ``cached_patcher_init`` factory on ``VAE.patcher`` so + :meth:`comfy.model_patcher.ModelPatcher.deepclone_multigpu` can produce a + fresh, untainted VAE patcher (no inherited per-device load state, no + in-place quantization fallout) for multigpu work-units and the + SelectVAEDevice node. The optional ``device`` matches the source loader's + VAE initialization path; the deepclone's ``load_device`` still controls + where the cloned patcher is targeted. + """ + if metadata is None: + sd, metadata = comfy.utils.load_torch_file(vae_path, return_metadata=True) + else: + sd = comfy.utils.load_torch_file(vae_path) + vae = VAE(sd=sd, metadata=metadata, device=device) + vae.throw_exception_if_invalid() + return vae.patcher + def load_unet(unet_path, dtype=None): logging.warning("The load_unet function has been deprecated and will be removed please switch to: load_diffusion_model") return load_diffusion_model(unet_path, model_options={"dtype": dtype}) diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 34b8b0d95..e0297708c 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -24,12 +24,14 @@ import comfy.text_encoders.qwen_image import comfy.text_encoders.hunyuan_image import comfy.text_encoders.kandinsky5 import comfy.text_encoders.z_image +import comfy.text_encoders.ideogram4 import comfy.text_encoders.anima import comfy.text_encoders.ace15 import comfy.text_encoders.longcat_image import comfy.text_encoders.ernie import comfy.text_encoders.cogvideo import comfy.text_encoders.hidream_o1 +import comfy.text_encoders.pixeldit from . import supported_models_base from . import latent_formats @@ -829,6 +831,50 @@ class Flux2(Flux): return None + +class Lens(supported_models_base.BASE): + """Microsoft Lens (3.8B dual-stream MMDiT, GPT-OSS-20B text features, Flux2 VAE).""" + + unet_config = { + "image_model": "lens", + } + + sampling_settings = { + "shift": 1.829, # Default mu for 1440x1440 (and any seq_len > 4300 + } + + unet_extra_config = {} + latent_format = latent_formats.Flux2 + + memory_usage_factor = 4.0 + + supported_inference_dtypes = [torch.bfloat16, torch.float32] # fp16 causes NaNs + + vae_key_prefix = ["vae."] + text_encoder_key_prefix = ["text_encoders."] + + def __init__(self, unet_config): + super().__init__(unet_config) + + def get_model(self, state_dict, prefix="", device=None): + return model_base.Lens(self, model_type=model_base.ModelType.FLUX, device=device) + + def clip_target(self, state_dict={}): + pref = self.text_encoder_key_prefix[0] + for hint in ("gpt_oss.transformer.", ""): + full_prefix = "{}{}".format(pref, hint) + if "{}layers.0.self_attn.sinks".format(full_prefix) in state_dict: + detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, full_prefix) + return supported_models_base.ClipTarget( + comfy.text_encoders.gpt_oss.LensTokenizer, + comfy.text_encoders.gpt_oss.lens_te(**detect), + ) + return supported_models_base.ClipTarget( + comfy.text_encoders.gpt_oss.LensTokenizer, + comfy.text_encoders.gpt_oss.lens_te(), + ) + + class GenmoMochi(supported_models_base.BASE): unet_config = { "image_model": "mochi_preview", @@ -1159,6 +1205,72 @@ class ZImagePixelSpace(ZImage): def get_model(self, state_dict, prefix="", device=None): return model_base.ZImagePixelSpace(self, device=device) +class PixelDiTT2I(supported_models_base.BASE): + unet_config = { + "image_model": "pixeldit_t2i", + } + + unet_extra_config = {} + + sampling_settings = { + "shift": 4.0, # 1024px stage 3 default; 2.0 for 512px + } + + latent_format = latent_formats.PixelDiTPixel + memory_usage_factor = 0.04 + supported_inference_dtypes = [torch.bfloat16, torch.float32] + + vae_key_prefix = ["vae."] + text_encoder_key_prefix = ["text_encoders."] + + def get_model(self, state_dict, prefix="", device=None): + return model_base.PixelDiTT2I(self, device=device) + + def process_unet_state_dict(self, state_dict): + # pixel_dim from pixel_embedder.proj.weight = (pixel_dim, in_channels); p2 derived per-weight from total // (6 * pixel_dim). + pixel_dim = next(v for k, v in state_dict.items() if k.endswith("pixel_embedder.proj.weight")).shape[0] + + out = {} + marker = ".adaLN_modulation.0." + for k, v in state_dict.items(): + if k.startswith("_repa_projector") or k.startswith("net_ema."): + continue + if k.startswith("core."): + k = k[len("core."):] + elif k.startswith("net."): + k = k[len("net."):] + if "pixel_blocks." in k and marker in k: + # Split into msa (chunks 0-2) and mlp (chunks 3-5) for the two-Linear PiTBlock to reduce peak VRAM + p2 = v.shape[0] // (6 * pixel_dim) + trail = v.shape[1:] # () for bias, (in_dim,) for weight + vv = v.view(p2, 6, pixel_dim, *trail) + base, suffix = k.split(marker) + out[f"{base}.adaLN_modulation_msa.{suffix}"] = vv[:, 0:3].reshape(3 * p2 * pixel_dim, *trail).contiguous() + out[f"{base}.adaLN_modulation_mlp.{suffix}"] = vv[:, 3:6].reshape(3 * p2 * pixel_dim, *trail).contiguous() + else: + out[k] = v + return out + + def clip_target(self, state_dict={}): + return supported_models_base.ClipTarget( + comfy.text_encoders.pixeldit.PixelDiTGemma2Tokenizer, + comfy.text_encoders.pixeldit.PixelDiTGemma2TE, + ) + +class PiD(PixelDiTT2I): + unet_config = { + "image_model": "pid", + } + + sampling_settings = { + "shift": 1.5, # close approximation of the original distill 4 steps [0.999, 0.866, 0.634, 0.342, 0] + } + + memory_usage_factor = 0.04 + + def get_model(self, state_dict, prefix="", device=None): + return model_base.PiD(self, device=device) + class WAN21_T2V(supported_models_base.BASE): unet_config = { "image_model": "wan2.1", @@ -1362,6 +1474,17 @@ class WAN21_SCAIL(WAN21_T2V): out = model_base.WAN21_SCAIL(self, image_to_video=False, device=device) return out + +class WAN21_SCAIL2(WAN21_T2V): + unet_config = { + "image_model": "wan2.1", + "model_type": "scail2", + } + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.WAN21_SCAIL2(self, image_to_video=False, device=device) + return out + class WAN22_WanDancer(WAN21_T2V): unet_config = { "image_model": "wan2.1", @@ -1451,6 +1574,30 @@ class Hunyuan3Dv2mini(Hunyuan3Dv2): latent_format = latent_formats.Hunyuan3Dv2mini +class TripoSplat(supported_models_base.BASE): + # Image -> 3D gaussian splat flow denoiser + unet_config = { + "image_model": "triposplat", + } + + unet_extra_config = {} + + sampling_settings = { + "shift": 3.0, + } + + memory_usage_factor = 0.6 + + latent_format = latent_formats.TripoSplat + + supported_inference_dtypes = [torch.float16, torch.bfloat16, torch.float32] + + def get_model(self, state_dict, prefix="", device=None): + return model_base.TripoSplat(self, device=device) + + def clip_target(self, state_dict={}): + return None + class HiDream(supported_models_base.BASE): unet_config = { "image_model": "hidream", @@ -1635,6 +1782,44 @@ class Omnigen2(supported_models_base.BASE): hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_3b.transformer.".format(pref)) return supported_models_base.ClipTarget(comfy.text_encoders.omnigen2.Omnigen2Tokenizer, comfy.text_encoders.omnigen2.te(**hunyuan_detect)) +class Ideogram4(supported_models_base.BASE): + unet_config = { + "image_model": "ideogram4", + } + + sampling_settings = { + "multiplier": 1.0, + "shift": 1.0, + } + + memory_usage_factor = 11.6 + + unet_extra_config = { + "num_attention_heads": 18, + "attention_head_dim": 256, + "intermediate_size": 12288, + "adaln_dim": 512, + "llm_features_dim": 53248, + "rope_theta": 5000000, + "mrope_section": [24, 20, 20], + "norm_eps": 1e-5, + } + latent_format = latent_formats.Flux2 + + supported_inference_dtypes = [torch.bfloat16, torch.float32] + + vae_key_prefix = ["vae."] + text_encoder_key_prefix = ["text_encoders."] + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.Ideogram4(self, device=device) + return out + + def clip_target(self, state_dict={}): + pref = self.text_encoder_key_prefix[0] + hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3vl_8b.transformer.".format(pref)) + return supported_models_base.ClipTarget(comfy.text_encoders.ideogram4.Ideogram4Tokenizer, comfy.text_encoders.ideogram4.te(**hunyuan_detect)) + class QwenImage(supported_models_base.BASE): unet_config = { "image_model": "qwen_image", @@ -1895,6 +2080,23 @@ class RT_DETR_v4(supported_models_base.BASE): return None +class DepthAnything3(supported_models_base.BASE): + unet_config = { + "image_model": "DepthAnything3", + } + + # Mono path: no num_heads / num_head_channels needed. + unet_extra_config = {} + + supported_inference_dtypes = [torch.float16, torch.bfloat16, torch.float32] + + def get_model(self, state_dict, prefix="", device=None): + return model_base.DepthAnything3(self, device=device) + + def clip_target(self, state_dict={}): + return None + + class ErnieImage(supported_models_base.BASE): unet_config = { "image_model": "ernie", @@ -2093,6 +2295,8 @@ models = [ CosmosI2VPredict2, ZImagePixelSpace, ZImage, + PiD, + PixelDiTT2I, Lumina2, WAN22_T2V, WAN21_CausalAR_T2V, @@ -2107,10 +2311,12 @@ models = [ WAN22_Animate, WAN21_FlowRVS, WAN21_SCAIL, + WAN21_SCAIL2, WAN22_WanDancer, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, + TripoSplat, HiDream, HiDreamO1, Chroma, @@ -2119,7 +2325,9 @@ models = [ ACEStep15, Omnigen2, QwenImage, + Ideogram4, Flux2, + Lens, Kandinsky5Image, Kandinsky5, Anima, @@ -2131,5 +2339,6 @@ models = [ CogVideoX_I2V, CogVideoX_T2V, SVD_img2vid, - Trellis2 + Trellis2, + DepthAnything3, ] diff --git a/comfy/text_encoders/gpt_oss.py b/comfy/text_encoders/gpt_oss.py new file mode 100644 index 000000000..d596ef9a0 --- /dev/null +++ b/comfy/text_encoders/gpt_oss.py @@ -0,0 +1,600 @@ +"""GPT-OSS text encoder for Lens.""" + +from __future__ import annotations + +import math +from dataclasses import dataclass +from typing import Any, List, Optional, Sequence + +import torch +import torch.nn as nn +import torch.nn.functional as F + +import comfy.ops +from comfy import sd1_clip +from comfy.ldm.modules.attention import TORCH_HAS_GQA, optimized_attention_for_device +from comfy.text_encoders.llama import RMSNorm, apply_rope + + +@dataclass +class GptOss20BConfig: + vocab_size: int = 201088 + hidden_size: int = 2880 + intermediate_size: int = 2880 + num_hidden_layers: int = 24 + num_attention_heads: int = 64 + num_key_value_heads: int = 8 + head_dim: int = 64 + num_local_experts: int = 32 + num_experts_per_tok: int = 4 + sliding_window: int = 128 + original_max_position_embeddings: int = 4096 + rope_theta: float = 150000.0 + rope_factor: float = 32.0 + rope_beta_fast: float = 32.0 + rope_beta_slow: float = 1.0 + rope_truncate: bool = False + rms_norm_eps: float = 1e-5 + attention_bias: bool = True + layer_types: Optional[List[str]] = None + moe_alpha: float = 1.702 + moe_limit: float = 7.0 + + def __post_init__(self): + if self.layer_types is None: + self.layer_types = [ + "sliding_attention" if (i + 1) % 2 else "full_attention" + for i in range(self.num_hidden_layers) + ] + + +def _yarn_inv_freq(head_dim: int, base: float, factor: float, beta_fast: float, beta_slow: float, + original_max_position_embeddings: int, truncate: bool, device=None) -> tuple[torch.Tensor, float]: + """YARN inv_freq + attention scaling (matches transformers).""" + dim = head_dim + + def find_correction_dim(num_rotations: float) -> float: + return (dim * math.log(original_max_position_embeddings / (num_rotations * 2 * math.pi))) / ( + 2 * math.log(base) + ) + + def find_correction_range() -> tuple[float, float]: + low = find_correction_dim(beta_fast) + high = find_correction_dim(beta_slow) + if truncate: + low = math.floor(low) + high = math.ceil(high) + return max(low, 0), min(high, dim - 1) + + def linear_ramp_factor(min_: float, max_: float, n: int) -> torch.Tensor: + if min_ == max_: + max_ += 0.001 + linear = (torch.arange(n, dtype=torch.float32, device=device) - min_) / (max_ - min_) + return torch.clamp(linear, 0, 1) + + def get_mscale(scale: float) -> float: + if scale <= 1: + return 1.0 + return 0.1 * math.log(scale) + 1.0 + + attention_scaling = get_mscale(factor) + + pos_freqs = base ** (torch.arange(0, dim, 2, dtype=torch.float32, device=device) / dim) + inv_freq_extrapolation = 1.0 / pos_freqs + inv_freq_interpolation = 1.0 / (factor * pos_freqs) + + low, high = find_correction_range() + extrap_factor = 1 - linear_ramp_factor(low, high, dim // 2) + inv_freq = inv_freq_interpolation * (1 - extrap_factor) + inv_freq_extrapolation * extrap_factor + return inv_freq, attention_scaling + + +def _build_freqs_cis(inv_freq: torch.Tensor, attention_scaling: float, position_ids: torch.Tensor, dtype: torch.dtype, +) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + inv_freq_e = inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1) + pos_e = position_ids[:, None, :].float() + freqs = (inv_freq_e @ pos_e).transpose(1, 2) + emb = torch.cat((freqs, freqs), dim=-1) + cos = (emb.cos() * attention_scaling).to(dtype).unsqueeze(1) + sin = (emb.sin() * attention_scaling).to(dtype).unsqueeze(1) + sin_split = sin.shape[-1] // 2 + return cos, sin[..., :sin_split], -sin[..., sin_split:] + + +def _attention_with_sinks(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, sinks: torch.Tensor, + attention_mask: Optional[torch.Tensor], num_heads: int, num_kv_groups: int) -> torch.Tensor: + """Attention with per-head sinks. + + Sinks add a learned term to each row's softmax denominator but contribute + nothing to the output. We fake this by appending one zero k/v position and + putting the sink logit in the mask at that column. + """ + + if num_kv_groups > 1 and not TORCH_HAS_GQA: + k = k.repeat_interleave(num_kv_groups, dim=1) + v = v.repeat_interleave(num_kv_groups, dim=1) + + B, _, S_q, D = q.shape + H_kv = k.shape[1] + S_kv = k.shape[-2] + + k = torch.cat([k, k.new_zeros(B, H_kv, 1, D)], dim=-2) + v = torch.cat([v, v.new_zeros(B, H_kv, 1, D)], dim=-2) + + sinks_col = sinks.to(q.dtype).view(1, num_heads, 1, 1).expand(B, num_heads, S_q, 1) + if attention_mask is not None: + mask_left = attention_mask[..., :S_kv].expand(B, num_heads, S_q, S_kv) + else: + mask_left = q.new_zeros(B, num_heads, S_q, S_kv) + mask = torch.cat([mask_left, sinks_col], dim=-1) + + op = optimized_attention_for_device(q.device, mask=True, small_input=True) + return op(q, k, v, num_heads, mask=mask, skip_reshape=True, enable_gqa=True) + + +class GptOssAttention(nn.Module): + def __init__(self, config: GptOss20BConfig, layer_idx: int, device=None, dtype=None, ops: Any = None): + super().__init__() + self.layer_idx = layer_idx + self.layer_type = config.layer_types[layer_idx] + self.num_heads = config.num_attention_heads + self.num_kv_heads = config.num_key_value_heads + self.num_kv_groups = self.num_heads // self.num_kv_heads + self.head_dim = config.head_dim + self.hidden_size = config.hidden_size + self.sliding_window = config.sliding_window if self.layer_type == "sliding_attention" else None + + bias = config.attention_bias + self.q_proj = ops.Linear(config.hidden_size, self.num_heads * self.head_dim, bias=bias, device=device, dtype=dtype) + self.k_proj = ops.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=bias, device=device, dtype=dtype) + self.v_proj = ops.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=bias, device=device, dtype=dtype) + self.o_proj = ops.Linear(self.num_heads * self.head_dim, config.hidden_size, bias=bias, device=device, dtype=dtype) + self.sinks = nn.Parameter(torch.empty(self.num_heads, device=device, dtype=dtype)) + + def forward(self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor], freqs_cis) -> torch.Tensor: + B, S, _ = hidden_states.shape + + q = self.q_proj(hidden_states).view(B, S, self.num_heads, self.head_dim).transpose(1, 2) + k = self.k_proj(hidden_states).view(B, S, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = self.v_proj(hidden_states).view(B, S, self.num_kv_heads, self.head_dim).transpose(1, 2) + + q, k = apply_rope(q, k, freqs_cis) + + out = _attention_with_sinks(q, k, v, self.sinks, attention_mask, self.num_heads, self.num_kv_groups) + return self.o_proj(out) + + +# Mixture of Experts + +class GptOssTopKRouter(nn.Module): + def __init__(self, config: GptOss20BConfig, device=None, dtype=None): + super().__init__() + self.top_k = config.num_experts_per_tok + self.num_experts = config.num_local_experts + self.weight = nn.Parameter(torch.empty(config.num_local_experts, config.hidden_size, device=device, dtype=dtype)) + self.bias = nn.Parameter(torch.empty(config.num_local_experts, device=device, dtype=dtype)) + + def forward(self, hidden_states: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: + weight = comfy.ops.cast_to_input(self.weight, hidden_states, copy=False) + bias = comfy.ops.cast_to_input(self.bias, hidden_states, copy=False) + logits = F.linear(hidden_states, weight, bias) + top_vals, top_idx = torch.topk(logits, self.top_k, dim=-1) + # Softmax over top-k slice only + scores = F.softmax(top_vals, dim=-1, dtype=top_vals.dtype) + return scores, top_idx + + +class GptOssExperts(nn.Module): + def __init__(self, config: GptOss20BConfig, device=None, dtype=None, ops: Any = None): + super().__init__() + self.num_experts = config.num_local_experts + self.hidden_size = config.hidden_size + self.intermediate_size = config.intermediate_size + self.alpha = config.moe_alpha + self.limit = config.moe_limit + + E = self.num_experts + H = self.hidden_size + I = self.intermediate_size + + self.gate_up_proj = ops.MoEExperts(num_experts=E, in_features=H, out_features=2 * I, bias=True, device=device, dtype=dtype) + self.down_proj = ops.MoEExperts(num_experts=E, in_features=I, out_features=H, bias=True, device=device, dtype=dtype) + + def _apply_gate(self, gate_up: torch.Tensor) -> torch.Tensor: + gate = gate_up[..., ::2] + up = gate_up[..., 1::2] + gate = gate.clamp(max=self.limit) + up = up.clamp(min=-self.limit, max=self.limit) + glu = gate * torch.sigmoid(gate * self.alpha) + return torch.addcmul(glu, up, glu) + + def forward(self, hidden_states: torch.Tensor, router_indices: torch.Tensor, routing_weights: torch.Tensor) -> torch.Tensor: + N = hidden_states.shape[0] + top_k = router_indices.shape[-1] + H = hidden_states.shape[-1] + + per_pair = torch.zeros((N * top_k, H), dtype=hidden_states.dtype, device=hidden_states.device) + + expert_mask = F.one_hot(router_indices, num_classes=self.num_experts).permute(2, 1, 0) + expert_hit = torch.greater(expert_mask.sum(dim=(-1, -2)), 0).nonzero() + + with self.gate_up_proj.bank_resident(hidden_states) as gate_up_bank, \ + self.down_proj.bank_resident(hidden_states) as down_bank: + for ei in expert_hit: + expert_idx = int(ei.item()) + top_k_pos, token_idx = torch.where(expert_mask[expert_idx]) + current = hidden_states[token_idx] + + gate_up = gate_up_bank.expert_linear(current, expert_idx) + gated = self._apply_gate(gate_up) + expert_out = down_bank.expert_linear(gated, expert_idx) + + weighted = expert_out * routing_weights[token_idx, top_k_pos, None] + + flat_idx = token_idx * top_k + top_k_pos + per_pair[flat_idx] = weighted.to(per_pair.dtype) + + return per_pair.view(N, top_k, H).sum(dim=1) + + +class GptOssMLP(nn.Module): + def __init__(self, config: GptOss20BConfig, device=None, dtype=None, ops: Any = None): + super().__init__() + self.router = GptOssTopKRouter(config, device=device, dtype=dtype) + self.experts = GptOssExperts(config, device=device, dtype=dtype, ops=ops) + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + B, S, H = hidden_states.shape + flat = hidden_states.reshape(-1, H) + scores, idx = self.router(flat) + out = self.experts(flat, idx, scores) + return out.reshape(B, S, H) + + +# Decoder layer + model + +class GptOssDecoderLayer(nn.Module): + def __init__(self, config: GptOss20BConfig, layer_idx: int, device=None, dtype=None, ops: Any = None): + super().__init__() + self.self_attn = GptOssAttention(config, layer_idx, device=device, dtype=dtype, ops=ops) + self.mlp = GptOssMLP(config, device=device, dtype=dtype, ops=ops) + self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, device=device, dtype=dtype) + self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, device=device, dtype=dtype) + self.layer_type = config.layer_types[layer_idx] + + def forward(self, x: torch.Tensor, attention_masks: dict[str, Optional[torch.Tensor]], freqs_cis) -> torch.Tensor: + residual = x + x = self.input_layernorm(x) + x = self.self_attn(x, attention_masks[self.layer_type], freqs_cis) + x = residual + x + + residual = x + x = self.post_attention_layernorm(x) + x = self.mlp(x) + x = residual + x + return x + + +def _make_full_causal_mask(B: int, S: int, key_padding_mask: Optional[torch.Tensor], dtype, device): + neg = torch.finfo(dtype).min + mask = torch.full((S, S), neg, dtype=dtype, device=device).triu_(1) + mask = mask.unsqueeze(0).unsqueeze(0).expand(B, 1, S, S).contiguous() + if key_padding_mask is not None: + kp = key_padding_mask.to(dtype=dtype) + kp = (1.0 - kp).reshape(B, 1, 1, S) * neg + mask = mask + kp + return mask + + +def _make_sliding_causal_mask(B: int, S: int, window: int, key_padding_mask: Optional[torch.Tensor], dtype, device): + neg = torch.finfo(dtype).min + i = torch.arange(S, device=device).view(-1, 1) + j = torch.arange(S, device=device).view(1, -1) + keep = (j <= i) & (j > i - window) + mask = torch.where(keep, torch.zeros((), dtype=dtype, device=device), torch.full((), neg, dtype=dtype, device=device)) + mask = mask.unsqueeze(0).unsqueeze(0).expand(B, 1, S, S).contiguous() + if key_padding_mask is not None: + kp = key_padding_mask.to(dtype=dtype) + kp = (1.0 - kp).reshape(B, 1, 1, S) * neg + mask = mask + kp + return mask + + +class GptOssModel(nn.Module): + """GPT-OSS decoder with multi-layer hidden-state capture + early exit.""" + + def __init__(self, config: GptOss20BConfig, device=None, dtype=None, ops: Any = None): + super().__init__() + self.config = config + self.dtype = dtype + self.embed_tokens = ops.Embedding(config.vocab_size, config.hidden_size, device=device, dtype=dtype) + self.layers = nn.ModuleList( + [ + GptOssDecoderLayer(config, i, device=device, dtype=dtype, ops=ops) + for i in range(config.num_hidden_layers) + ] + ) + self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, device=device, dtype=dtype) + + # Always build on CPU so the buffer survives meta-device construction. + inv_freq, attn_scaling = _yarn_inv_freq( + head_dim=config.head_dim, + base=config.rope_theta, + factor=config.rope_factor, + beta_fast=config.rope_beta_fast, + beta_slow=config.rope_beta_slow, + original_max_position_embeddings=config.original_max_position_embeddings, + truncate=config.rope_truncate, + device=torch.device("cpu"), + ) + self.register_buffer("rope_inv_freq", inv_freq, persistent=False) + self.rope_attention_scaling = float(attn_scaling) + + @property + def num_layers(self) -> int: + return self.config.num_hidden_layers + + def get_input_embeddings(self): + return self.embed_tokens + + def _build_attention_masks(self, B: int, S: int, attention_mask: Optional[torch.Tensor], dtype: torch.dtype, device, + ) -> dict[str, torch.Tensor]: + full = _make_full_causal_mask(B, S, attention_mask, dtype, device) + masks = {"full_attention": full} + if any(t == "sliding_attention" for t in self.config.layer_types): + masks["sliding_attention"] = _make_sliding_causal_mask( + B, S, self.config.sliding_window, attention_mask, dtype, device + ) + return masks + + def forward(self, input_ids: torch.LongTensor, attention_mask: Optional[torch.Tensor] = None, + capture_layers: Optional[Sequence[int]] = None) -> dict[str, Any]: + B, S = input_ids.shape + device = input_ids.device + dtype = self.dtype + + hidden_states = self.embed_tokens(input_ids, out_dtype=dtype) + + position_ids = torch.arange(S, device=device).unsqueeze(0).expand(B, -1) + freqs_cis = _build_freqs_cis(self.rope_inv_freq.to(device=device), self.rope_attention_scaling, position_ids, dtype) + + attn_masks = self._build_attention_masks(B, S, attention_mask, dtype, device) + + capture_layers = list(capture_layers) if capture_layers else None + if capture_layers: + max_layer = max(capture_layers) + wanted = {idx: pos for pos, idx in enumerate(capture_layers)} + captured: List[Optional[torch.Tensor]] = [None] * len(capture_layers) + else: + max_layer = self.config.num_hidden_layers - 1 + wanted = None + captured = None + + for i, layer in enumerate(self.layers): + hidden_states = layer(hidden_states, attn_masks, freqs_cis) + if wanted is not None and i in wanted: + captured[wanted[i]] = hidden_states + if i >= max_layer: + break + + if captured is not None: + return {"hidden_states": captured} + return {"last_hidden_state": self.norm(hidden_states)} + + +# Lens chat-template constants (verbatim from the reference pipeline). +_LENS_CHAT_SYSTEM = ( + "Describe the image by detailing the color, shape, size, texture, " + "quantity, text, spatial relationships of the objects and background." +) +_LENS_CHAT_ASSISTANT_THINKING = "Need to generate one image according to the description." +LENS_TXT_OFFSET = 97 +LENS_SELECTED_LAYERS = (5, 11, 17, 23) +LENS_MAX_TOKENS = 512 + + +# The reference GPT-OSS Harmony template injects today's date here +_LENS_CHAT_DATE = "2026-05-23" + + +def _lens_render_chat(prompt: str) -> str: + """Render the Lens prompt in GPT-OSS Harmony format.""" + return ( + f"<|start|>system<|message|>" + f"You are ChatGPT, a large language model trained by OpenAI.\n" + f"Knowledge cutoff: 2024-06\n" + f"Current date: {_LENS_CHAT_DATE}\n\n" + f"Reasoning: medium\n\n" + f"# Valid channels: analysis, commentary, final. " + f"Channel must be included for every message.<|end|>" + f"<|start|>developer<|message|># Instructions\n\n" + f"{_LENS_CHAT_SYSTEM}\n\n<|end|>" + f"<|start|>user<|message|>{prompt}<|end|>" + f"<|start|>assistant<|channel|>analysis<|message|>" + f"{_LENS_CHAT_ASSISTANT_THINKING}<|end|>" + f"<|start|>assistant<|channel|>final<|message|>" + ) + + +# GPT-OSS-20B fixed token IDs (from the tokenizer's added-tokens table). +_LENS_PAD_TOKEN_ID = 199999 # <|endoftext|> + + +class _GptOssRawTokenizer: + """Raw ``tokenizers.Tokenizer`` wrapper. + + The tokenizer JSON ships as a byte tensor inside the encoder checkpoint + (``tokenizer_json`` key) rather than as a committed file. Extracted + it in ``sd.py`` and passes it here via ``tokenizer_data``. + """ + + def __init__(self, tokenizer_json_bytes=None, **kwargs): + from tokenizers import Tokenizer + if isinstance(tokenizer_json_bytes, torch.Tensor): + tokenizer_json_bytes = bytes(tokenizer_json_bytes.tolist()) + if tokenizer_json_bytes is None: + raise ValueError( + "Lens tokenizer requires the ``tokenizer_json`` byte tensor in the " + "encoder state dict. Re-bundle the encoder via bundle_te.py so it " + "embeds the tokenizer." + ) + self.tokenizer = Tokenizer.from_str(tokenizer_json_bytes.decode("utf-8")) + + @classmethod + def from_pretrained(cls, tokenizer_data, **kwargs): + return cls(tokenizer_json_bytes=tokenizer_data, **kwargs) + + def __call__(self, text): + return {"input_ids": self.tokenizer.encode(text, add_special_tokens=False).ids} + + def get_vocab(self): + return self.tokenizer.get_vocab() + + def convert_tokens_to_ids(self, tokens): + return [self.tokenizer.token_to_id(t) for t in tokens] + + def decode(self, ids, **kwargs): + return self.tokenizer.decode(ids, skip_special_tokens=kwargs.get("skip_special_tokens", False)) + + +class LensGptOssTokenizer(sd1_clip.SDTokenizer): + tokenizer_json_data = None + + def __init__(self, embedding_directory=None, tokenizer_data={}): + tokenizer_json = tokenizer_data.get("tokenizer_json", None) + self.tokenizer_json_data = tokenizer_json + super().__init__( + tokenizer_json, + embedding_directory=embedding_directory, + pad_with_end=False, + embedding_size=2880, + embedding_key="gpt_oss", + tokenizer_class=_GptOssRawTokenizer, + has_start_token=False, + has_end_token=False, + pad_to_max_length=False, + max_length=99999999, + min_length=1, + pad_left=False, + disable_weights=True, + tokenizer_data=tokenizer_data, + ) + self.pad_token_id = _LENS_PAD_TOKEN_ID + + def tokenize_with_weights(self, text: str, return_word_ids=False, **kwargs): + # Empty prompt -> empty list; encode_token_weights returns zeros (uncond). + if not text or not text.strip(): + return [[]] + rendered = _lens_render_chat(text) + ids = self.tokenizer(rendered)["input_ids"] + if len(ids) > LENS_MAX_TOKENS: + ids = ids[:LENS_MAX_TOKENS] + return [[(int(t), 1.0) for t in ids]] + + def state_dict(self): + if self.tokenizer_json_data is not None: + return {"tokenizer_json": self.tokenizer_json_data} + return {} + + +class LensTokenizer(sd1_clip.SD1Tokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__( + embedding_directory=embedding_directory, + tokenizer_data=tokenizer_data, + name="gpt_oss", + tokenizer=LensGptOssTokenizer, + ) + + +class LensGptOssClipModel(nn.Module): + """SDClipModel-shaped Lens GPT-OSS encoder (multi-layer feature extractor).""" + + def __init__(self, device="cpu", dtype=None, model_options=None, **kwargs): + super().__init__() + model_options = dict(model_options or {}) + + operations = model_options.get("custom_operations") + if operations is None: + quant_config = model_options.get("quantization_metadata") or {} + operations = comfy.ops.mixed_precision_ops(quant_config, dtype, full_precision_mm=True) + self.operations = operations + + cfg_overrides = model_options.get("gpt_oss_config", {}) + self.config = GptOss20BConfig(**cfg_overrides) + self.selected_layers = tuple(model_options.get("selected_layers", LENS_SELECTED_LAYERS)) + self.txt_offset = int(model_options.get("txt_offset", LENS_TXT_OFFSET)) + + self.transformer = GptOssModel(self.config, device=device, dtype=dtype, ops=operations) + self.num_layers = self.config.num_hidden_layers + self.dtype = dtype + self.execution_device = None + self._pad_token_id = _LENS_PAD_TOKEN_ID + + def set_clip_options(self, options): + self.execution_device = options.get("execution_device", self.execution_device) + + def reset_clip_options(self): + self.execution_device = None + + def _gather_tokens(self, token_weight_pairs): + ids_list = [[int(t[0]) for t in batch] for batch in token_weight_pairs] + pad_id = self._pad_token_id + max_len = max(len(x) for x in ids_list) + device = self.execution_device + ids = torch.full((len(ids_list), max_len), pad_id, dtype=torch.long, device=device) + mask = torch.zeros((len(ids_list), max_len), dtype=torch.long, device=device) + for i, x in enumerate(ids_list): + ids[i, : len(x)] = torch.tensor(x, dtype=torch.long, device=device) + mask[i, : len(x)] = 1 + return ids, mask + + def encode_token_weights(self, token_weight_pairs): + # Empty negative: emit zero-length features + zero mask + if all(len(batch) == 0 for batch in token_weight_pairs): + device = self.execution_device + B = len(token_weight_pairs) + L = len(self.selected_layers) + H = self.config.hidden_size + flat = torch.zeros(B, 0, L * H, dtype=self.dtype, device=device) + mask = torch.zeros(B, 0, dtype=torch.long, device=device) + return flat, None, {"attention_mask": mask, "num_layers_stacked": L} + + input_ids, attn_mask = self._gather_tokens(token_weight_pairs) + out = self.transformer(input_ids, attention_mask=attn_mask, capture_layers=self.selected_layers) + layers = out["hidden_states"] # list of L × [B, S, H] + stacked = torch.stack(layers, dim=2) # [B, S, L, H] + + offset = self.txt_offset + if stacked.shape[1] > offset: + stacked = stacked[:, offset:].contiguous() + mask_trim = attn_mask[:, offset:] + else: + stacked = stacked[:, :0] + mask_trim = attn_mask[:, :0] + + B, S, L, H = stacked.shape + flat = stacked.reshape(B, S, L * H) + extra = {"attention_mask": mask_trim, "num_layers_stacked": L} + return flat, None, extra + + def load_sd(self, sd): + return self.transformer.load_state_dict(sd, strict=False, assign=True) + + +class LensTEModel(sd1_clip.SD1ClipModel): + def __init__(self, device="cpu", dtype=None, model_options=None): + super().__init__(device=device, dtype=dtype, name="gpt_oss", clip_model=LensGptOssClipModel, model_options=model_options or {}) + + +def lens_te(dtype_llama=None, llama_quantization_metadata=None): + class LensTEModel_(LensTEModel): + def __init__(self, device="cpu", dtype=None, model_options=None): + mo = dict(model_options or {}) + if llama_quantization_metadata is not None: + mo["quantization_metadata"] = llama_quantization_metadata + if dtype is None and dtype_llama is not None: + dtype = dtype_llama + super().__init__(device=device, dtype=dtype, model_options=mo) + + return LensTEModel_ diff --git a/comfy/text_encoders/ideogram4.py b/comfy/text_encoders/ideogram4.py new file mode 100644 index 000000000..84243772d --- /dev/null +++ b/comfy/text_encoders/ideogram4.py @@ -0,0 +1,79 @@ +"""Ideogram 4 text encoder: Qwen3-VL-8B language model, 13-layer tap. + +Ideogram 4 conditions on the concatenation of hidden states from 13 layers of +Qwen3-VL (layers 0,3,...,33,35), giving a 4096*13 = 53248-dim feature per token. +""" + +import os + +from transformers import Qwen2Tokenizer + +import comfy.text_encoders.llama +from comfy import sd1_clip + +# Reference taps outputs of layers (0,3,...,35); comfy captures layer inputs, offset by +1. +IDEOGRAM4_TAP_LAYERS = [1, 4, 7, 10, 13, 16, 19, 22, 25, 28, 31, 34, 36] + + +class Qwen3VLTokenizer(sd1_clip.SDTokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer") + super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, + embedding_size=4096, embedding_key='qwen3vl_8b', tokenizer_class=Qwen2Tokenizer, + has_start_token=False, has_end_token=False, pad_to_max_length=False, + max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data) + + +class Ideogram4Tokenizer(sd1_clip.SD1Tokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, + name="qwen3vl_8b", tokenizer=Qwen3VLTokenizer) + + self.llama_template = "<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n" + + def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, **kwargs): + if text.startswith('<|im_start|>'): + llama_text = text + elif llama_template is None: + llama_text = self.llama_template.format(text) + else: + llama_text = llama_template.format(text) + return super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, disable_weights=True, **kwargs) + + +# Qwen3-VL-8B = 5e6 (vs plain Qwen3-8B's 1e6) +# final_norm/lm_head off -> Ideogram only reads raw tapped hidden states +QWEN3VL_8B_CONFIG = {"rope_theta": 5000000.0, "final_norm": False, "lm_head": False} + + +class Qwen3VL8BModel(sd1_clip.SDClipModel): + def __init__(self, device="cpu", layer="hidden", layer_idx=None, dtype=None, attention_mask=True, model_options={}): + super().__init__(device=device, layer=IDEOGRAM4_TAP_LAYERS, layer_idx=None, + textmodel_json_config=dict(QWEN3VL_8B_CONFIG), + dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, + model_class=comfy.text_encoders.llama.Qwen3_8B, + enable_attention_masks=attention_mask, return_attention_masks=attention_mask, + model_options=model_options) + + +class Ideogram4TEModel(sd1_clip.SD1ClipModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + super().__init__(device=device, dtype=dtype, name="qwen3vl_8b", clip_model=Qwen3VL8BModel, model_options=model_options) + + def encode_token_weights(self, token_weight_pairs): + out, pooled, extra = super().encode_token_weights(token_weight_pairs) + b, n, seq, h = out.shape # (B, n_taps=13, seq, 4096) stacked in ascending layer order. + out = out.permute(0, 2, 3, 1).reshape(b, seq, h * n) # (B, seq, 4096*13). permute -> (B, seq, H, taps). + return out, pooled, extra + + +def te(dtype_llama=None, llama_quantization_metadata=None): + class Ideogram4TEModel_(Ideogram4TEModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + if dtype_llama is not None: + dtype = dtype_llama + if llama_quantization_metadata is not None: + model_options = model_options.copy() + model_options["quantization_metadata"] = llama_quantization_metadata + super().__init__(device=device, dtype=dtype, model_options=model_options) + return Ideogram4TEModel_ diff --git a/comfy/text_encoders/pixeldit.py b/comfy/text_encoders/pixeldit.py new file mode 100644 index 000000000..3539711e4 --- /dev/null +++ b/comfy/text_encoders/pixeldit.py @@ -0,0 +1,104 @@ +import torch + +from comfy import sd1_clip +from .lumina2 import Gemma2BTokenizer, LuminaModel +import comfy.text_encoders.llama + + +class PixelDiTGemma2_2BModel(sd1_clip.SDClipModel): + def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}): + llama_quantization_metadata = model_options.get("llama_quantization_metadata", None) + if llama_quantization_metadata is not None: + model_options = model_options.copy() + model_options["quantization_metadata"] = llama_quantization_metadata + + super().__init__( + device=device, layer=layer, layer_idx=layer_idx, + textmodel_json_config={}, dtype=dtype, + special_tokens={"start": 2, "pad": 0}, + layer_norm_hidden_state=False, + model_class=comfy.text_encoders.llama.Gemma2_2B, + enable_attention_masks=attention_mask, + return_attention_masks=attention_mask, + model_options=model_options, + ) + + +_PIXELDIT_CHI_PROMPT = ( + 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual descriptions ' + "suitable for image generation. Evaluate the level of detail in the user prompt:\n" + "- If the prompt is simple, focus on adding specifics about colors, shapes, sizes, textures, " + "and spatial relationships to create vivid and concrete scenes.\n" + "- If the prompt is already detailed, refine and enhance the existing details slightly without " + "overcomplicating.\n" + "Here are examples of how to transform or refine prompts:\n" + "- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up in a round shape, " + "sleeping peacefully on a warm sunny windowsill, surrounded by pots of blooming red flowers.\n" + "- User Prompt: A busy city street -> Enhanced: A bustling city street scene at dusk, featuring " + "glowing street lamps, a diverse crowd of people in colorful clothing, and a double-decker bus " + "passing by towering glass skyscrapers.\n" + "Please generate only the enhanced description for the prompt below and avoid including any " + "additional commentary or evaluations:\n" + "User Prompt: " +) + +_PIXELDIT_MAX_LENGTH = 300 +_PIXELDIT_CHI_PROMPT_DETECT_PREFIX = 'Given a user prompt, generate an "Enhanced prompt"' + + +class PixelDiTGemma2Tokenizer(sd1_clip.SD1Tokenizer): + def __init__(self, embedding_directory=None, tokenizer_data=None): + if tokenizer_data is None: + tokenizer_data = {} + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, + name="gemma2_2b", tokenizer=Gemma2BTokenizer) + + def tokenize_with_weights(self, text, return_word_ids=False, **kwargs): + if not text.strip(): + return super().tokenize_with_weights("", return_word_ids=return_word_ids, disable_weights=True, min_length=_PIXELDIT_MAX_LENGTH) + + chi_token_count = len(self.gemma2_2b.tokenizer(_PIXELDIT_CHI_PROMPT)["input_ids"]) + combined = text if text.startswith(_PIXELDIT_CHI_PROMPT_DETECT_PREFIX) else _PIXELDIT_CHI_PROMPT + text + max_length_all = chi_token_count + _PIXELDIT_MAX_LENGTH - 2 + out = super().tokenize_with_weights(combined, return_word_ids=return_word_ids, + disable_weights=True, min_length=max_length_all) + out["gemma2_2b"] = [out["gemma2_2b"][0][:max_length_all]] + return out + + def untokenize(self, token_weight_pair): + return self.gemma2_2b.untokenize(token_weight_pair) + + def state_dict(self): + return self.gemma2_2b.state_dict() + + +class PixelDiTGemma2TE(LuminaModel): + # PixelDiT's select_index: keep BOS + last 299 embeddings of the padded sequence. + def __init__(self, device="cpu", dtype=None, model_options={}): + super().__init__(device=device, dtype=dtype, name="gemma2_2b", + clip_model=PixelDiTGemma2_2BModel, model_options=model_options) + + def encode_token_weights(self, token_weight_pairs): + result = super().encode_token_weights(token_weight_pairs) + cond, pooled = result[0], result[1] + extra = result[2] if len(result) > 2 else None + if cond.shape[1] > _PIXELDIT_MAX_LENGTH: + cond = torch.cat([cond[:, :1], cond[:, -(_PIXELDIT_MAX_LENGTH - 1):]], dim=1) + if extra is not None and "attention_mask" in extra: + am = extra["attention_mask"] + extra["attention_mask"] = torch.cat([am[..., :1], am[..., -(_PIXELDIT_MAX_LENGTH - 1):]], dim=-1) + if extra is not None: + return cond, pooled, extra + return cond, pooled + + +def pixeldit_te(dtype_llama=None, llama_quantization_metadata=None): + class PixelDiTTE_(PixelDiTGemma2TE): + def __init__(self, device="cpu", dtype=None, model_options={}): + if llama_quantization_metadata is not None: + model_options = model_options.copy() + model_options["llama_quantization_metadata"] = llama_quantization_metadata + if dtype_llama is not None: + dtype = dtype_llama + super().__init__(device=device, dtype=dtype, model_options=model_options) + return PixelDiTTE_ diff --git a/comfy/utils.py b/comfy/utils.py index 00e382fac..09d783fff 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -85,8 +85,9 @@ _TYPES = { def load_safetensors(ckpt): import comfy_aimdo.model_mmap - f = open(ckpt, "rb", buffering=0) + file_lock = threading.Lock() model_mmap = comfy_aimdo.model_mmap.ModelMMAP(ckpt) + f = model_mmap.get_file_handle() file_size = os.path.getsize(ckpt) mv = memoryview((ctypes.c_uint8 * file_size).from_address(model_mmap.get())) @@ -111,7 +112,7 @@ def load_safetensors(ckpt): storage = tensor.untyped_storage() setattr(storage, "_comfy_tensor_file_slice", - comfy.memory_management.TensorFileSlice(f, threading.get_ident(), data_base_offset + start, end - start)) + comfy.memory_management.TensorFileSlice(f, file_lock, data_base_offset + start, end - start)) setattr(storage, "_comfy_tensor_mmap_refs", (model_mmap, mv)) sd[name] = tensor @@ -1019,10 +1020,11 @@ def bislerp(samples, width, height): def lanczos(samples, width, height): #the below API is strict and expects grayscale to be squeezed - samples = samples.squeeze(1) if samples.shape[1] == 1 else samples.movedim(1, -1) + if samples.ndim == 4: + samples = samples.squeeze(1) if samples.shape[1] == 1 else samples.movedim(1, -1) images = [Image.fromarray(np.clip(255. * image.cpu().numpy(), 0, 255).astype(np.uint8)) for image in samples] images = [image.resize((width, height), resample=Image.Resampling.LANCZOS) for image in images] - images = [torch.from_numpy(np.array(image).astype(np.float32) / 255.0).movedim(-1, 0) for image in images] + images = [torch.from_numpy(t).movedim(-1, 0) if (t := np.array(image).astype(np.float32) / 255.0).ndim == 3 else torch.from_numpy(t) for image in images] result = torch.stack(images) return result.to(samples.device, samples.dtype) @@ -1450,3 +1452,10 @@ def deepcopy_list_dict(obj, memo=None): memo[obj_id] = res return res + +def bit_reverse_range(index, bits): + result = 0 + for _ in range(bits): + result = (result << 1) | (index & 1) + index >>= 1 + return result diff --git a/comfy_api/latest/__init__.py b/comfy_api/latest/__init__.py index 04973fea0..294ad425e 100644 --- a/comfy_api/latest/__init__.py +++ b/comfy_api/latest/__init__.py @@ -1,5 +1,3 @@ -from __future__ import annotations - from abc import ABC, abstractmethod from typing import TYPE_CHECKING from comfy_api.internal import ComfyAPIBase @@ -7,7 +5,7 @@ from comfy_api.internal.singleton import ProxiedSingleton from comfy_api.internal.async_to_sync import create_sync_class from ._input import ImageInput, AudioInput, MaskInput, LatentInput, VideoInput from ._input_impl import VideoFromFile, VideoFromComponents -from ._util import VideoCodec, VideoContainer, VideoComponents, MESH, VOXEL, File3D +from ._util import VideoCodec, VideoContainer, VideoComponents, MESH, VOXEL, SPLAT, File3D from . import _io_public as io from . import _ui_public as ui from comfy_execution.utils import get_executing_context @@ -145,6 +143,7 @@ class Types: VideoComponents = VideoComponents MESH = MESH VOXEL = VOXEL + SPLAT = SPLAT File3D = File3D diff --git a/comfy_api/latest/_input/video_types.py b/comfy_api/latest/_input/video_types.py index 451e9526e..8fff52c16 100644 --- a/comfy_api/latest/_input/video_types.py +++ b/comfy_api/latest/_input/video_types.py @@ -65,6 +65,12 @@ class VideoInput(ABC): buffer.seek(0) return buffer + def get_active_trim_window(self) -> tuple[float, float]: + """Return the active trim as ``(start_time, duration)`` in seconds (start_time normalized + to ``>= 0``; ``duration == 0`` means "until the end"). Default: no trim; trimmable subclasses override. + """ + return 0.0, 0.0 + # Provide a default implementation, but subclasses can provide optimized versions # if possible. def get_dimensions(self) -> tuple[int, int]: diff --git a/comfy_api/latest/_input_impl/video_types.py b/comfy_api/latest/_input_impl/video_types.py index 942278d88..4a12ff9c1 100644 --- a/comfy_api/latest/_input_impl/video_types.py +++ b/comfy_api/latest/_input_impl/video_types.py @@ -1,4 +1,3 @@ -from __future__ import annotations from av.container import InputContainer from av.subtitles.stream import SubtitleStream from fractions import Fraction @@ -76,6 +75,12 @@ class VideoFromFile(VideoInput): self.__file.seek(0) return self.__file + def get_active_trim_window(self) -> tuple[float, float]: + start_time = self.__start_time + if start_time < 0: + start_time = max(self._get_raw_duration() + start_time, 0.0) + return float(start_time), float(self.__duration) + def get_dimensions(self) -> tuple[int, int]: """ Returns the dimensions of the video input. diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index 5ed968960..37614a4c3 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -28,7 +28,7 @@ if TYPE_CHECKING: from comfy_api.internal import (_ComfyNodeInternal, _NodeOutputInternal, classproperty, copy_class, first_real_override, is_class, prune_dict, shallow_clone_class) from comfy_execution.graph_utils import ExecutionBlocker -from ._util import MESH, VOXEL, SVG as _SVG, File3D +from ._util import MESH, VOXEL, SPLAT, SVG as _SVG, File3D class FolderType(str, Enum): @@ -684,6 +684,10 @@ class Voxel(ComfyTypeIO): class Mesh(ComfyTypeIO): Type = MESH +@comfytype(io_type="SPLAT") +class Splat(ComfyTypeIO): + Type = SPLAT + @comfytype(io_type="FILE_3D") class File3DAny(ComfyTypeIO): @@ -727,6 +731,42 @@ class File3DUSDZ(ComfyTypeIO): Type = File3D +@comfytype(io_type="FILE_3D_PLY") +class File3DPLY(ComfyTypeIO): + """PLY format 3D file - point cloud or Gaussian splat.""" + Type = File3D + + +@comfytype(io_type="FILE_3D_SPLAT") +class File3DSPLAT(ComfyTypeIO): + """SPLAT format 3D file - 3D Gaussian splat.""" + Type = File3D + + +@comfytype(io_type="FILE_3D_SPZ") +class File3DSPZ(ComfyTypeIO): + """SPZ format 3D file - compressed 3D Gaussian splat.""" + Type = File3D + + +@comfytype(io_type="FILE_3D_KSPLAT") +class File3DKSPLAT(ComfyTypeIO): + """KSPLAT format 3D file - 3D Gaussian splat.""" + Type = File3D + + +@comfytype(io_type="FILE_3D_SPLAT_ANY") +class File3DSplatAny(ComfyTypeIO): + """General 3D Gaussian splat file type - accepts any supported splat container (.ply / .spz / .splat / .ksplat).""" + Type = File3D + + +@comfytype(io_type="FILE_3D_POINT_CLOUD_ANY") +class File3DPointCloudAny(ComfyTypeIO): + """General point cloud file type - accepts any supported point cloud container (currently .ply).""" + Type = File3D + + @comfytype(io_type="HOOKS") class Hooks(ComfyTypeIO): if TYPE_CHECKING: @@ -762,14 +802,32 @@ class Accumulation(ComfyTypeIO): @comfytype(io_type="LOAD3D_CAMERA") class Load3DCamera(ComfyTypeIO): class CameraInfo(TypedDict): - position: dict[str, float | int] - target: dict[str, float | int] - zoom: int - cameraType: str + # Coordinate system: right-handed, Y-up, camera looks down -Z + position: dict[str, float | int] # scene units + target: dict[str, float | int] # scene units; OrbitControls focus point + zoom: float | int # dimensionless, 1 = 100% + cameraType: str # 'perspective' | 'orthographic' + quaternion: NotRequired[dict[str, float | int]] # normalized, dimensionless; camera world rotation + fov: NotRequired[float | int] # degrees, vertical FOV (perspective only) + aspect: NotRequired[float | int] # width / height (perspective only) + near: NotRequired[float | int] # scene units + far: NotRequired[float | int] # scene units + frustum: NotRequired[dict[str, float | int]] # orthographic only: {left, right, top, bottom} in scene units Type = CameraInfo +@comfytype(io_type="LOAD3D_MODEL_INFO") +class Load3DModelInfo(ComfyTypeIO): + class Model3DTransform(TypedDict): + # Coordinate system: right-handed, Y-up, world space + position: dict[str, float | int] # scene units + quaternion: dict[str, float | int] # normalized, dimensionless; world rotation + scale: dict[str, float | int] # dimensionless multiplier + + Type = list[Model3DTransform] + + @comfytype(io_type="LOAD_3D") class Load3D(ComfyTypeIO): """3D models are stored as a dictionary.""" @@ -779,6 +837,7 @@ class Load3D(ComfyTypeIO): normal: str camera_info: Load3DCamera.CameraInfo recording: NotRequired[str] + model_3d_info: NotRequired[list[Load3DModelInfo.Model3DTransform]] Type = Model3DDict @@ -2277,6 +2336,7 @@ __all__ = [ "LossMap", "Voxel", "Mesh", + "Splat", "File3DAny", "File3DGLB", "File3DGLTF", @@ -2284,6 +2344,12 @@ __all__ = [ "File3DOBJ", "File3DSTL", "File3DUSDZ", + "File3DPLY", + "File3DSPLAT", + "File3DSPZ", + "File3DKSPLAT", + "File3DSplatAny", + "File3DPointCloudAny", "Hooks", "HookKeyframes", "TimestepsRange", @@ -2291,6 +2357,7 @@ __all__ = [ "FlowControl", "Accumulation", "Load3DCamera", + "Load3DModelInfo", "Load3D", "Load3DAnimation", "Photomaker", diff --git a/comfy_api/latest/_ui.py b/comfy_api/latest/_ui.py index e238cdf3c..b48713d41 100644 --- a/comfy_api/latest/_ui.py +++ b/comfy_api/latest/_ui.py @@ -285,7 +285,7 @@ class AudioSaveHelper: results = [] for batch_number, waveform in enumerate(audio["waveform"].cpu()): filename_with_batch_num = filename.replace("%batch_num%", str(batch_number)) - file = f"{filename_with_batch_num}_{counter:05}_.{format}" + file = f"{filename_with_batch_num}_{counter:05}.{format}" output_path = os.path.join(full_output_folder, file) # Use original sample rate initially @@ -452,6 +452,16 @@ class PreviewUI3D(_UIOutput): return {"result": [self.model_file, self.camera_info, self.bg_image_path]} +class PreviewUI3DAdvanced(_UIOutput): + def __init__(self, model_file, camera_info, model_3d_info): + self.model_file = model_file + self.camera_info = camera_info + self.model_3d_info = model_3d_info + + def as_dict(self): + return {"result": [self.model_file, self.camera_info, self.model_3d_info]} + + class PreviewText(_UIOutput): def __init__(self, value: str, **kwargs): self.value = value @@ -471,5 +481,6 @@ __all__ = [ "PreviewAudio", "PreviewVideo", "PreviewUI3D", + "PreviewUI3DAdvanced", "PreviewText", ] diff --git a/comfy_api/latest/_util/__init__.py b/comfy_api/latest/_util/__init__.py index 115baf392..b27f5a97e 100644 --- a/comfy_api/latest/_util/__init__.py +++ b/comfy_api/latest/_util/__init__.py @@ -1,5 +1,5 @@ from .video_types import VideoContainer, VideoCodec, VideoComponents -from .geometry_types import VOXEL, MESH, File3D +from .geometry_types import VOXEL, MESH, SPLAT, File3D from .image_types import SVG __all__ = [ @@ -9,6 +9,7 @@ __all__ = [ "VideoComponents", "VOXEL", "MESH", + "SPLAT", "File3D", "SVG", ] diff --git a/comfy_api/latest/_util/geometry_types.py b/comfy_api/latest/_util/geometry_types.py index 93f697119..825425128 100644 --- a/comfy_api/latest/_util/geometry_types.py +++ b/comfy_api/latest/_util/geometry_types.py @@ -12,6 +12,24 @@ class VOXEL: self.voxel_colors = voxel_colors self.resolution = resolution # each 3d model has its own resolution +class SPLAT: + """A batch of 3D Gaussian splats in render-ready (activated, world-space) form. + + Tensors are (B, N, ...) and zero-padded to a common N across the batch; `counts` (B,) holds the + real per-item lengths (None when rows are uniform and no slicing is needed). SH coefficients are + stored as (B, N, K, 3) with K = (sh_degree + 1)**2; the DC (diffuse) term is sh[..., 0, :]. + """ + + def __init__(self, positions: torch.Tensor, scales: torch.Tensor, rotations: torch.Tensor, + opacities: torch.Tensor, sh: torch.Tensor, counts: torch.Tensor | None = None): + self.positions = positions # (B, N, 3) world-space centers + self.scales = scales # (B, N, 3) linear (positive) per-axis std + self.rotations = rotations # (B, N, 4) quaternion wxyz (normalized) + self.opacities = opacities # (B, N, 1) in [0, 1] + self.sh = sh # (B, N, K, 3) spherical-harmonic color coefficients + self.counts = counts # (B,) real lengths, or None + + class MESH: def __init__(self, vertices: torch.Tensor, faces: torch.Tensor, uvs: torch.Tensor | None = None, @@ -19,7 +37,8 @@ class MESH: texture: torch.Tensor | None = None, metallic_roughness: torch.Tensor | None = None, vertex_counts: torch.Tensor | None = None, - face_counts: torch.Tensor | None = None): + face_counts: torch.Tensor | None = None, + unlit: bool = False): assert (vertex_counts is None) == (face_counts is None), \ "vertex_counts and face_counts must be provided together (both or neither)" @@ -34,6 +53,8 @@ class MESH: # these hold the real per-item lengths (B,). None means rows are uniform and no slicing is needed. self.vertex_counts = vertex_counts self.face_counts = face_counts + # Render flat / emissive (no scene lighting) when saved, e.g. for gaussian-splat-derived meshes. + self.unlit = unlit class File3D: diff --git a/comfy_api/latest/_util/video_types.py b/comfy_api/latest/_util/video_types.py index c92477f08..6c9d6a526 100644 --- a/comfy_api/latest/_util/video_types.py +++ b/comfy_api/latest/_util/video_types.py @@ -1,4 +1,3 @@ -from __future__ import annotations from dataclasses import dataclass from enum import Enum from fractions import Fraction diff --git a/comfy_api_nodes/apis/__init__.py b/comfy_api_nodes/apis/__init__.py index 46a583b5e..9c4cfb9b6 100644 --- a/comfy_api_nodes/apis/__init__.py +++ b/comfy_api_nodes/apis/__init__.py @@ -3,7 +3,6 @@ # timestamp: 2025-07-30T08:54:00+00:00 # pylint: disable -from __future__ import annotations from datetime import date, datetime from enum import Enum diff --git a/comfy_api_nodes/apis/anthropic.py b/comfy_api_nodes/apis/anthropic.py index 6cac537ea..46a5bb428 100644 --- a/comfy_api_nodes/apis/anthropic.py +++ b/comfy_api_nodes/apis/anthropic.py @@ -35,6 +35,19 @@ class AnthropicMessage(BaseModel): content: list[AnthropicTextContent | AnthropicImageContent] = Field(...) +class AnthropicThinkingConfig(BaseModel): + type: Literal["enabled", "disabled", "adaptive"] = Field(...) + budget_tokens: int | None = Field( + None, ge=1024, + description="Reasoning budget in tokens. Used when type is 'enabled'. Must be less than max_tokens.", + ) + + +class AnthropicOutputConfig(BaseModel): + """Used with `thinking.type='adaptive'` on models like Opus 4.7.""" + effort: Literal["low", "medium", "high"] | None = Field(None) + + class AnthropicMessagesRequest(BaseModel): model: str = Field(...) messages: list[AnthropicMessage] = Field(...) @@ -44,6 +57,8 @@ class AnthropicMessagesRequest(BaseModel): top_p: float | None = Field(None, ge=0.0, le=1.0) top_k: int | None = Field(None, ge=0) stop_sequences: list[str] | None = Field(None) + thinking: AnthropicThinkingConfig | None = Field(None) + output_config: AnthropicOutputConfig | None = Field(None) class AnthropicResponseTextBlock(BaseModel): @@ -51,6 +66,14 @@ class AnthropicResponseTextBlock(BaseModel): text: str = Field(...) +class AnthropicResponseThinkingBlock(BaseModel): + type: Literal["thinking"] = "thinking" + thinking: str = Field(...) + + +AnthropicResponseBlock = AnthropicResponseTextBlock | AnthropicResponseThinkingBlock + + class AnthropicCacheCreationUsage(BaseModel): ephemeral_5m_input_tokens: int | None = Field(None) ephemeral_1h_input_tokens: int | None = Field(None) @@ -69,7 +92,7 @@ class AnthropicMessagesResponse(BaseModel): type: str | None = Field(None) role: str | None = Field(None) model: str | None = Field(None) - content: list[AnthropicResponseTextBlock] | None = Field(None) + content: list[AnthropicResponseBlock] | None = Field(None) stop_reason: str | None = Field(None) stop_sequence: str | None = Field(None) usage: AnthropicMessagesUsage | None = Field(None) diff --git a/comfy_api_nodes/apis/beeble.py b/comfy_api_nodes/apis/beeble.py new file mode 100644 index 000000000..90175b214 --- /dev/null +++ b/comfy_api_nodes/apis/beeble.py @@ -0,0 +1,32 @@ +from pydantic import BaseModel, Field + + +class CreateSwitchXRequest(BaseModel): + generation_type: str = Field(...) + source_uri: str = Field(...) + alpha_mode: str = Field(...) + prompt: str | None = Field(None, max_length=2000) + reference_image_uri: str | None = Field(None) + alpha_uri: str | None = Field(None) + max_resolution: int = Field(1080) + callback_url: str | None = Field(None) + idempotency_key: str | None = Field(None, max_length=256, min_length=1) + + +class SwitchXOutputUrls(BaseModel): + render: str | None = Field(None) + source: str | None = Field(None) + alpha: str | None = Field(None) + + +class SwitchXStatusResponse(BaseModel): + id: str = Field(...) + status: str = Field(...) + progress: int | None = Field(None) + generation_type: str | None = Field(None) + alpha_mode: str | None = Field(None) + output: SwitchXOutputUrls | None = Field(None) + error: str | None = Field(None) + created_at: str | None = Field(None) + modified_at: str | None = Field(None) + completed_at: str | None = Field(None) diff --git a/comfy_api_nodes/apis/bfl.py b/comfy_api_nodes/apis/bfl.py index d8d3557b3..4c950da84 100644 --- a/comfy_api_nodes/apis/bfl.py +++ b/comfy_api_nodes/apis/bfl.py @@ -1,73 +1,72 @@ -from __future__ import annotations - from enum import Enum -from typing import Any, Dict, Optional +from typing import Any -from pydantic import BaseModel, Field, confloat, conint - - -class BFLOutputFormat(str, Enum): - png = 'png' - jpeg = 'jpeg' +from pydantic import BaseModel, Field class BFLFluxExpandImageRequest(BaseModel): - prompt: str = Field(..., description='The description of the changes you want to make. This text guides the expansion process, allowing you to specify features, styles, or modifications for the expanded areas.') - prompt_upsampling: Optional[bool] = Field( - None, description='Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation.' - ) - seed: Optional[int] = Field(None, description='The seed value for reproducibility.') - top: conint(ge=0, le=2048) = Field(..., description='Number of pixels to expand at the top of the image') - bottom: conint(ge=0, le=2048) = Field(..., description='Number of pixels to expand at the bottom of the image') - left: conint(ge=0, le=2048) = Field(..., description='Number of pixels to expand at the left side of the image') - right: conint(ge=0, le=2048) = Field(..., description='Number of pixels to expand at the right side of the image') - steps: conint(ge=15, le=50) = Field(..., description='Number of steps for the image generation process') - guidance: confloat(ge=1.5, le=100) = Field(..., description='Guidance strength for the image generation process') - safety_tolerance: Optional[conint(ge=0, le=6)] = Field( - 6, description='Tolerance level for input and output moderation. Between 0 and 6, 0 being most strict, 6 being least strict. Defaults to 2.' - ) - output_format: Optional[BFLOutputFormat] = Field( - BFLOutputFormat.png, description="Output format for the generated image. Can be 'jpeg' or 'png'.", examples=['png'] - ) - image: str = Field(None, description='A Base64-encoded string representing the image you wish to expand') + prompt: str = Field(...) + prompt_upsampling: bool | None = Field(None) + seed: int | None = Field(None) + top: int = Field(...) + bottom: int = Field(...) + left: int = Field(...) + right: int = Field(...) + steps: int = Field(...) + guidance: float = Field(...) + safety_tolerance: int = Field(6) + output_format: str = Field("png") + image: str = Field(None, description="A Base64-encoded string representing the image you wish to expand") class BFLFluxFillImageRequest(BaseModel): - prompt: str = Field(..., description='The description of the changes you want to make. This text guides the expansion process, allowing you to specify features, styles, or modifications for the expanded areas.') - prompt_upsampling: Optional[bool] = Field( - None, description='Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation.' + prompt: str = Field(...) + prompt_upsampling: bool | None = Field(None) + seed: int | None = Field(None) + steps: int = Field(...) + guidance: float = Field(...) + safety_tolerance: int = Field(6) + output_format: str = Field("png") + image: str = Field( + None, description="Base64-encoded string representing the image to modify. Can contain alpha mask if desired.", ) - seed: Optional[int] = Field(None, description='The seed value for reproducibility.') - steps: conint(ge=15, le=50) = Field(..., description='Number of steps for the image generation process') - guidance: confloat(ge=1.5, le=100) = Field(..., description='Guidance strength for the image generation process') - safety_tolerance: Optional[conint(ge=0, le=6)] = Field( - 6, description='Tolerance level for input and output moderation. Between 0 and 6, 0 being most strict, 6 being least strict. Defaults to 2.' + mask: str = Field( + None, description="Base64-encoded string representing the mask of the areas you wish to modify." ) - output_format: Optional[BFLOutputFormat] = Field( - BFLOutputFormat.png, description="Output format for the generated image. Can be 'jpeg' or 'png'.", examples=['png'] + + +class BFLFluxEraseRequest(BaseModel): + image: str = Field(..., description="A Base64-encoded string representing the image to erase from.") + mask: str = Field( + ..., + description="A Base64-encoded black/white mask matching the input dimensions; " + "white (255) marks areas to remove, black (0) marks areas to preserve.", ) - image: str = Field(None, description='A Base64-encoded string representing the image you wish to modify. Can contain alpha mask if desired.') - mask: str = Field(None, description='A Base64-encoded string representing the mask of the areas you with to modify.') + dilate_pixels: int = Field(10) + seed: int | None = Field(None) + output_format: str = Field("png") + + +class BFLFluxVTORequest(BaseModel): + prompt: str = Field( + ..., description="Natural-language styling instruction. Required field, but may be an empty string." + ) + person: str = Field(..., description="A Base64-encoded string representing the person image.") + garment: str = Field(..., description="A Base64-encoded string representing the garment reference image.") + seed: int | None = Field(None) + safety_tolerance: int = Field(5) + output_format: str = Field("png") class BFLFluxProGenerateRequest(BaseModel): - prompt: str = Field(..., description='The text prompt for image generation.') - prompt_upsampling: Optional[bool] = Field( - None, description='Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation.' - ) - seed: Optional[int] = Field(None, description='The seed value for reproducibility.') - width: conint(ge=256, le=1440) = Field(1024, description='Width of the generated image in pixels. Must be a multiple of 32.') - height: conint(ge=256, le=1440) = Field(768, description='Height of the generated image in pixels. Must be a multiple of 32.') - safety_tolerance: Optional[conint(ge=0, le=6)] = Field( - 6, description='Tolerance level for input and output moderation. Between 0 and 6, 0 being most strict, 6 being least strict. Defaults to 2.' - ) - output_format: Optional[BFLOutputFormat] = Field( - BFLOutputFormat.png, description="Output format for the generated image. Can be 'jpeg' or 'png'.", examples=['png'] - ) - image_prompt: Optional[str] = Field(None, description='Optional image to remix in base64 format') - # image_prompt_strength: Optional[confloat(ge=0.0, le=1.0)] = Field( - # None, description='Blend between the prompt and the image prompt.' - # ) + prompt: str = Field(...) + prompt_upsampling: bool | None = Field(None) + seed: int | None = Field(None) + width: int = Field(1024, description="Must be a multiple of 32.") + height: int = Field(768, description="Must be a multiple of 32.") + safety_tolerance: int = Field(6) + output_format: str = Field("png") + image_prompt: str | None = Field(None, description="Optional image to remix in base64 format") class Flux2ProGenerateRequest(BaseModel): @@ -85,55 +84,37 @@ class Flux2ProGenerateRequest(BaseModel): input_image_7: str | None = Field(None, description="Base64 encoded image for image-to-image generation") input_image_8: str | None = Field(None, description="Base64 encoded image for image-to-image generation") input_image_9: str | None = Field(None, description="Base64 encoded image for image-to-image generation") - safety_tolerance: int | None = Field( - 5, description="Tolerance level for input and output moderation. Value 0 being most strict.", ge=0, le=5 - ) - output_format: str | None = Field( - "png", description="Output format for the generated image. Can be 'jpeg' or 'png'." - ) + safety_tolerance: int = Field(5) + output_format: str = Field("png") class BFLFluxKontextProGenerateRequest(BaseModel): - prompt: str = Field(..., description='The text prompt for what you wannt to edit.') - input_image: Optional[str] = Field(None, description='Image to edit in base64 format') - seed: Optional[int] = Field(None, description='The seed value for reproducibility.') - guidance: confloat(ge=0.1, le=99.0) = Field(..., description='Guidance strength for the image generation process') - steps: conint(ge=1, le=150) = Field(..., description='Number of steps for the image generation process') - safety_tolerance: Optional[conint(ge=0, le=2)] = Field( - 2, description='Tolerance level for input and output moderation. Between 0 and 2, 0 being most strict, 6 being least strict. Defaults to 2.' - ) - output_format: Optional[BFLOutputFormat] = Field( - BFLOutputFormat.png, description="Output format for the generated image. Can be 'jpeg' or 'png'.", examples=['png'] - ) - aspect_ratio: Optional[str] = Field(None, description='Aspect ratio of the image between 21:9 and 9:21.') - prompt_upsampling: Optional[bool] = Field( - None, description='Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation.' - ) + prompt: str = Field(...) + input_image: str | None = Field(None, description="Image to edit in base64 format") + seed: int | None = Field(None) + guidance: float = Field(...) + steps: int = Field(...) + safety_tolerance: int = Field(2) + output_format: str = Field("png") + aspect_ratio: str | None = Field(None) + prompt_upsampling: bool | None = Field(None) class BFLFluxProUltraGenerateRequest(BaseModel): - prompt: str = Field(..., description='The text prompt for image generation.') - prompt_upsampling: Optional[bool] = Field( - None, description='Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation.' - ) - seed: Optional[int] = Field(None, description='The seed value for reproducibility.') - aspect_ratio: Optional[str] = Field(None, description='Aspect ratio of the image between 21:9 and 9:21.') - safety_tolerance: Optional[conint(ge=0, le=6)] = Field( - 6, description='Tolerance level for input and output moderation. Between 0 and 6, 0 being most strict, 6 being least strict. Defaults to 2.' - ) - output_format: Optional[BFLOutputFormat] = Field( - BFLOutputFormat.png, description="Output format for the generated image. Can be 'jpeg' or 'png'.", examples=['png'] - ) - raw: Optional[bool] = Field(None, description='Generate less processed, more natural-looking images.') - image_prompt: Optional[str] = Field(None, description='Optional image to remix in base64 format') - image_prompt_strength: Optional[confloat(ge=0.0, le=1.0)] = Field( - None, description='Blend between the prompt and the image prompt.' - ) + prompt: str = Field(...) + prompt_upsampling: bool | None = Field(None) + seed: int | None = Field(None) + aspect_ratio: str | None = Field(None) + safety_tolerance: int = Field(6) + output_format: str = Field("png") + raw: bool | None = Field(None) + image_prompt: str | None = Field(None, description="Optional image to remix in base64 format") + image_prompt_strength: float | None = Field(None) class BFLFluxProGenerateResponse(BaseModel): - id: str = Field(..., description="The unique identifier for the generation task.") - polling_url: str = Field(..., description="URL to poll for the generation result.") + id: str = Field(...) + polling_url: str = Field(...) cost: float | None = Field(None, description="Price in cents") @@ -147,7 +128,7 @@ class BFLStatus(str, Enum): class BFLFluxStatusResponse(BaseModel): - id: str = Field(..., description="The unique identifier for the generation task.") - status: BFLStatus = Field(..., description="The status of the task.") - result: Optional[Dict[str, Any]] = Field(None, description="The result of the task (null if not completed).") - progress: Optional[float] = Field(None, description="The progress of the task (0.0 to 1.0).", ge=0.0, le=1.0) + id: str = Field(...) + status: BFLStatus = Field(...) + result: dict[str, Any] | None = Field(None) + progress: float | None = Field(None, ge=0.0, le=1.0) diff --git a/comfy_api_nodes/apis/bria.py b/comfy_api_nodes/apis/bria.py index e08a519a8..7a98428c3 100644 --- a/comfy_api_nodes/apis/bria.py +++ b/comfy_api_nodes/apis/bria.py @@ -97,3 +97,28 @@ class BriaRemoveVideoBackgroundResult(BaseModel): class BriaRemoveVideoBackgroundResponse(BaseModel): status: str = Field(...) result: BriaRemoveVideoBackgroundResult | None = Field(None) + + +class BriaVideoGreenScreenRequest(BaseModel): + video: str = Field(..., description="Publicly accessible URL of the input video.") + green_shade: str = Field( + default="broadcast_green", + description="Solid chroma-key shade applied behind the foreground " + "(broadcast_green, chroma_green, or blue_screen).", + ) + output_container_and_codec: str = Field(...) + preserve_audio: bool = Field(True) + seed: int = Field(...) + + +class BriaVideoReplaceBackgroundRequest(BaseModel): + video: str = Field(..., description="Publicly accessible URL of the input (foreground) video.") + background_url: str = Field( + ..., + description="Publicly accessible URL of the background image or video to composite behind " + "the foreground. Stretched to the foreground frame; match its aspect ratio for " + "undistorted results.", + ) + output_container_and_codec: str = Field(...) + preserve_audio: bool = Field(True) + seed: int = Field(...) diff --git a/comfy_api_nodes/apis/bytedance.py b/comfy_api_nodes/apis/bytedance.py index 03f4c445b..47f24586c 100644 --- a/comfy_api_nodes/apis/bytedance.py +++ b/comfy_api_nodes/apis/bytedance.py @@ -158,8 +158,9 @@ class SeedanceCreateAssetResponse(BaseModel): class SeedanceVirtualLibraryCreateAssetRequest(BaseModel): - url: str = Field(..., description="Publicly accessible URL of the image asset to upload.") + url: str = Field(..., description="Publicly accessible URL of the asset to upload.") hash: str = Field(..., description="Dedup key. Re-submitting the same hash returns the existing asset id.") + asset_type: str | None = Field(None, description="BytePlus asset type. Defaults to Image server-side when omitted.") # Dollars per 1K tokens, keyed by (model_id, has_video_input). diff --git a/comfy_api_nodes/apis/gemini.py b/comfy_api_nodes/apis/gemini.py index 22879fe18..caaba8f36 100644 --- a/comfy_api_nodes/apis/gemini.py +++ b/comfy_api_nodes/apis/gemini.py @@ -108,13 +108,19 @@ class GeminiVideoMetadata(BaseModel): startOffset: GeminiOffset | None = Field(None) +class GeminiThinkingConfig(BaseModel): + includeThoughts: bool | None = Field(None) + thinkingLevel: str = Field(...) + + class GeminiGenerationConfig(BaseModel): - maxOutputTokens: int | None = Field(None, ge=16, le=8192) + maxOutputTokens: int | None = Field(None, ge=16, le=65536) seed: int | None = Field(None) stopSequences: list[str] | None = Field(None) temperature: float | None = Field(None, ge=0.0, le=2.0) topK: int | None = Field(None, ge=1) topP: float | None = Field(None, ge=0.0, le=1.0) + thinkingConfig: GeminiThinkingConfig | None = Field(None) class GeminiImageOutputOptions(BaseModel): @@ -128,11 +134,6 @@ class GeminiImageConfig(BaseModel): imageOutputOptions: GeminiImageOutputOptions = Field(default_factory=GeminiImageOutputOptions) -class GeminiThinkingConfig(BaseModel): - includeThoughts: bool | None = Field(None) - thinkingLevel: str = Field(...) - - class GeminiImageGenerationConfig(GeminiGenerationConfig): responseModalities: list[str] | None = Field(None) imageConfig: GeminiImageConfig | None = Field(None) diff --git a/comfy_api_nodes/apis/ideogram.py b/comfy_api_nodes/apis/ideogram.py index 737e18e3b..c5ad9559f 100644 --- a/comfy_api_nodes/apis/ideogram.py +++ b/comfy_api_nodes/apis/ideogram.py @@ -290,3 +290,19 @@ class IdeogramV3Request(BaseModel): None, description='Optional masks for character reference images. When provided, must match the number of character_reference_images. Each mask should be a grayscale image of the same dimensions as the corresponding character reference image. The images should be in JPEG, PNG or WebP format.' ) + + +class IdeogramV4Request(BaseModel): + text_prompt: str | None = Field( + None, + description="Natural-language prompt; Magic Prompt is applied automatically. " + "Supply exactly one of text_prompt or json_prompt.", + ) + json_prompt: dict[str, Any] | None = Field( + None, + description="Structured V4 prompt object consumed directly (disables Magic Prompt). " + "Supply exactly one of text_prompt or json_prompt.", + ) + resolution: str | None = Field(None, description="Output resolution in WIDTHxHEIGHT (e.g. '2048x2048').") + rendering_speed: str | None = Field(None, description="Rendering speed: 'TURBO', 'DEFAULT', or 'QUALITY'.") + enable_copyright_detection: bool | None = Field(None, description="Opt into post-generation copyright detection.") diff --git a/comfy_api_nodes/apis/krea.py b/comfy_api_nodes/apis/krea.py new file mode 100644 index 000000000..6e294a3b7 --- /dev/null +++ b/comfy_api_nodes/apis/krea.py @@ -0,0 +1,46 @@ +"""Pydantic models for the Krea image-generation API.""" + +from pydantic import BaseModel, Field + + +class KreaMoodboard(BaseModel): + id: str = Field(...) + strength: float = Field(default=0.35, ge=-0.5, le=1.5) + + +class KreaImageStyleReference(BaseModel): + strength: float = Field(..., ge=-2.0, le=2.0) + url: str | None = Field(default=None) + + +class KreaGenerateImageRequest(BaseModel): + prompt: str = Field(...) + aspect_ratio: str = Field(...) + resolution: str = Field(...) + seed: int | None = Field(default=None) + creativity: str = Field(default="medium") + moodboards: list[KreaMoodboard] | None = Field(default=None) + image_style_references: list[KreaImageStyleReference] | None = Field(default=None) + + +class KreaJobResult(BaseModel): + urls: list[str] | None = Field(default=None) + style_id: str | None = Field(default=None) + + +class KreaJob(BaseModel): + job_id: str = Field(...) + status: str = Field(...) + created_at: str = Field(...) + completed_at: str | None = Field(default=None) + result: KreaJobResult | None = Field(default=None) + + +class KreaAssetResponse(BaseModel): + id: str = Field(...) + image_url: str = Field(...) + uploaded_at: str = Field(...) + width: float | None = Field(default=None) + height: float | None = Field(default=None) + size_bytes: float | None = Field(default=None) + mime_type: str | None = Field(default=None) diff --git a/comfy_api_nodes/apis/openrouter.py b/comfy_api_nodes/apis/openrouter.py new file mode 100644 index 000000000..e30d9bcfb --- /dev/null +++ b/comfy_api_nodes/apis/openrouter.py @@ -0,0 +1,93 @@ +"""Pydantic models for the OpenRouter chat completions API. + +See: https://openrouter.ai/docs/api/api-reference/chat/send-chat-completion-request +""" + +from typing import Literal + +from pydantic import BaseModel, Field + + +class OpenRouterTextContent(BaseModel): + type: Literal["text"] = "text" + text: str = Field(...) + + +class OpenRouterImageUrl(BaseModel): + url: str = Field(...) + + +class OpenRouterImageContent(BaseModel): + type: Literal["image_url"] = "image_url" + image_url: OpenRouterImageUrl = Field(...) + + +class OpenRouterVideoUrl(BaseModel): + url: str = Field(...) + + +class OpenRouterVideoContent(BaseModel): + type: Literal["video_url"] = "video_url" + video_url: OpenRouterVideoUrl = Field(...) + + +OpenRouterContentBlock = OpenRouterTextContent | OpenRouterImageContent | OpenRouterVideoContent + + +class OpenRouterMessage(BaseModel): + role: Literal["system", "user", "assistant"] = Field(...) + content: str | list[OpenRouterContentBlock] = Field(...) + + +class OpenRouterReasoningConfig(BaseModel): + effort: str | None = Field(None) + exclude: bool | None = Field(None, description="If true, model reasons but reasoning is excluded from response.") + + +class OpenRouterWebSearchOptions(BaseModel): + search_context_size: str | None = Field(None) + + +class OpenRouterChatRequest(BaseModel): + model: str = Field(...) + messages: list[OpenRouterMessage] = Field(...) + seed: int | None = Field(None) + reasoning: OpenRouterReasoningConfig | None = Field(None) + web_search_options: OpenRouterWebSearchOptions | None = Field(None) + stream: bool = Field(False) + + +class OpenRouterUsage(BaseModel): + prompt_tokens: int | None = Field(None) + completion_tokens: int | None = Field(None) + total_tokens: int | None = Field(None) + cost: float | None = Field(None, description="Server-side authoritative USD cost of the call.") + + +class OpenRouterResponseMessage(BaseModel): + role: str | None = Field(None) + content: str | None = Field(None) + reasoning: str | None = Field(None) + refusal: str | None = Field(None) + + +class OpenRouterChoice(BaseModel): + index: int | None = Field(None) + message: OpenRouterResponseMessage | None = Field(None) + finish_reason: str | None = Field(None) + + +class OpenRouterError(BaseModel): + code: int | str | None = Field(None) + message: str | None = Field(None) + metadata: dict | None = Field(None) + + +class OpenRouterChatResponse(BaseModel): + id: str | None = Field(None) + model: str | None = Field(None) + object: str | None = Field(None) + provider: str | None = Field(None) + choices: list[OpenRouterChoice] | None = Field(None) + usage: OpenRouterUsage | None = Field(None) + error: OpenRouterError | None = Field(None) diff --git a/comfy_api_nodes/apis/rodin.py b/comfy_api_nodes/apis/rodin.py index fc26a6e73..24524d642 100644 --- a/comfy_api_nodes/apis/rodin.py +++ b/comfy_api_nodes/apis/rodin.py @@ -1,7 +1,5 @@ -from __future__ import annotations - from enum import Enum -from typing import Optional, List + from pydantic import BaseModel, Field @@ -11,44 +9,76 @@ class Rodin3DGenerateRequest(BaseModel): material: str = Field(..., description="The material type.") quality_override: int = Field(..., description="The poly count of the mesh.") mesh_mode: str = Field(..., description="It controls the type of faces of generated models.") - TAPose: Optional[bool] = Field(None, description="") + TAPose: bool | None = Field(None, description="") + + +class Rodin3DGen25Request(BaseModel): + + tier: str = Field(..., description="Gen-2.5 tier (e.g. Gen-2.5-High).") + prompt: str | None = Field(None, description="Required for Text-to-3D; ignored otherwise.") + seed: int | None = Field(None, description="0-65535.") + material: str | None = Field(None, description="PBR | Shaded | All | None.") + geometry_file_format: str | None = Field(None, description="glb | usdz | fbx | obj | stl.") + texture_mode: str | None = Field(None, description="legacy | extreme-low | low | medium | high.") + mesh_mode: str | None = Field(None, description="Raw (triangular) | Quad.") + quality_override: int | None = Field(None, description="Mesh face count override.") + geometry_instruct_mode: str | None = Field(None, description="faithful | creative.") + bbox_condition: list[int] | None = Field(None, description="Bounding box [Width(Y), Height(Z), Length(X)] in cm.") + height: int | None = Field(None, description="Approximate model height in cm.") + TAPose: bool | None = Field(None, description="T/A pose for human-like models.") + hd_texture: bool | None = Field(None, description="Enhanced texture quality.") + texture_delight: bool | None = Field(None, description="Remove baked lighting from textures.") + is_micro: bool | None = Field(None, description="Micro detail (Extreme-High only).") + use_original_alpha: bool | None = Field(None, description="Preserve image transparency.") + preview_render: bool | None = Field(None, description="Generate high-quality preview render.") + addons: list[str] | None = Field(None, description='Optional addons, e.g. ["HighPack"].') + class GenerateJobsData(BaseModel): - uuids: List[str] = Field(..., description="str LIST") + uuids: list[str] = Field(..., description="str LIST") subscription_key: str = Field(..., description="subscription key") + class Rodin3DGenerateResponse(BaseModel): - message: Optional[str] = Field(None, description="Return message.") - prompt: Optional[str] = Field(None, description="Generated Prompt from image.") - submit_time: Optional[str] = Field(None, description="Submit Time") - uuid: Optional[str] = Field(None, description="Task str") - jobs: Optional[GenerateJobsData] = Field(None, description="Details of jobs") + message: str | None = Field(None, description="Return message.") + prompt: str | None = Field(None, description="Generated Prompt from image.") + submit_time: str | None = Field(None, description="Submit Time") + uuid: str | None = Field(None, description="Task str") + jobs: GenerateJobsData | None = Field(None, description="Details of jobs") + class JobStatus(str, Enum): """ Status for jobs """ + Done = "Done" Failed = "Failed" Generating = "Generating" Waiting = "Waiting" + class Rodin3DCheckStatusRequest(BaseModel): subscription_key: str = Field(..., description="subscription from generate endpoint") + class JobItem(BaseModel): uuid: str = Field(..., description="uuid") - status: JobStatus = Field(...,description="Status Currently") + status: JobStatus = Field(..., description="Status Currently") + class Rodin3DCheckStatusResponse(BaseModel): - jobs: List[JobItem] = Field(..., description="Job status List") + jobs: list[JobItem] = Field(..., description="Job status List") + class Rodin3DDownloadRequest(BaseModel): task_uuid: str = Field(..., description="Task str") + class RodinResourceItem(BaseModel): url: str = Field(..., description="Download Url") name: str = Field(..., description="File name with ext") + class Rodin3DDownloadResponse(BaseModel): - list: List[RodinResourceItem] = Field(..., description="Source List") + items: list[RodinResourceItem] = Field(..., alias="list", description="Source List") diff --git a/comfy_api_nodes/apis/stability.py b/comfy_api_nodes/apis/stability.py index 718360187..5b9b5ac7d 100644 --- a/comfy_api_nodes/apis/stability.py +++ b/comfy_api_nodes/apis/stability.py @@ -1,5 +1,3 @@ -from __future__ import annotations - from enum import Enum from typing import Optional diff --git a/comfy_api_nodes/apis/tripo.py b/comfy_api_nodes/apis/tripo.py index bce6b0e89..7ac81d42c 100644 --- a/comfy_api_nodes/apis/tripo.py +++ b/comfy_api_nodes/apis/tripo.py @@ -1,25 +1,25 @@ from enum import Enum -from typing import Optional, Any +from typing import Any from pydantic import BaseModel, Field, RootModel class TripoModelVersion(str, Enum): - v3_1_20260211 = 'v3.1-20260211' - v3_0_20250812 = 'v3.0-20250812' - v2_5_20250123 = 'v2.5-20250123' - v2_0_20240919 = 'v2.0-20240919' - v1_4_20240625 = 'v1.4-20240625' + v3_1_20260211 = "v3.1-20260211" + v3_0_20250812 = "v3.0-20250812" + v2_5_20250123 = "v2.5-20250123" + v2_0_20240919 = "v2.0-20240919" + v1_4_20240625 = "v1.4-20240625" class TripoGeometryQuality(str, Enum): - standard = 'standard' - detailed = 'detailed' + standard = "standard" + detailed = "detailed" class TripoTextureQuality(str, Enum): - standard = 'standard' - detailed = 'detailed' + standard = "standard" + detailed = "detailed" class TripoStyle(str, Enum): @@ -33,6 +33,7 @@ class TripoStyle(str, Enum): ANCIENT_BRONZE = "ancient_bronze" NONE = "None" + class TripoTaskType(str, Enum): TEXT_TO_MODEL = "text_to_model" IMAGE_TO_MODEL = "image_to_model" @@ -45,26 +46,27 @@ class TripoTaskType(str, Enum): STYLIZE_MODEL = "stylize_model" CONVERT_MODEL = "convert_model" + class TripoTextureAlignment(str, Enum): ORIGINAL_IMAGE = "original_image" GEOMETRY = "geometry" + class TripoOrientation(str, Enum): ALIGN_IMAGE = "align_image" DEFAULT = "default" + class TripoOutFormat(str, Enum): GLB = "glb" FBX = "fbx" -class TripoTopology(str, Enum): - BIP = "bip" - QUAD = "quad" class TripoSpec(str, Enum): MIXAMO = "mixamo" TRIPO = "tripo" + class TripoAnimation(str, Enum): IDLE = "preset:idle" WALK = "preset:walk" @@ -83,11 +85,6 @@ class TripoAnimation(str, Enum): SERPENTINE_MARCH = "preset:serpentine:march" AQUATIC_MARCH = "preset:aquatic:march" -class TripoStylizeStyle(str, Enum): - LEGO = "lego" - VOXEL = "voxel" - VORONOI = "voronoi" - MINECRAFT = "minecraft" class TripoConvertFormat(str, Enum): GLTF = "GLTF" @@ -97,6 +94,7 @@ class TripoConvertFormat(str, Enum): STL = "STL" _3MF = "3MF" + class TripoTextureFormat(str, Enum): BMP = "BMP" DPX = "DPX" @@ -108,6 +106,7 @@ class TripoTextureFormat(str, Enum): TIFF = "TIFF" WEBP = "WEBP" + class TripoTaskStatus(str, Enum): QUEUED = "queued" RUNNING = "running" @@ -118,183 +117,223 @@ class TripoTaskStatus(str, Enum): BANNED = "banned" EXPIRED = "expired" + class TripoFbxPreset(str, Enum): BLENDER = "blender" MIXAMO = "mixamo" _3DSMAX = "3dsmax" + class TripoFileTokenReference(BaseModel): - type: Optional[str] = Field(None, description='The type of the reference') + type: str | None = Field(None, description="The type of the reference") file_token: str + class TripoUrlReference(BaseModel): - type: Optional[str] = Field(None, description='The type of the reference') + type: str | None = Field(None, description="The type of the reference") url: str + class TripoObjectStorage(BaseModel): bucket: str key: str + class TripoObjectReference(BaseModel): type: str object: TripoObjectStorage + class TripoFileEmptyReference(BaseModel): pass + class TripoFileReference(RootModel): root: TripoFileTokenReference | TripoUrlReference | TripoObjectReference | TripoFileEmptyReference -class TripoGetStsTokenRequest(BaseModel): - format: str = Field(..., description='The format of the image') class TripoTextToModelRequest(BaseModel): - type: TripoTaskType = Field(TripoTaskType.TEXT_TO_MODEL, description='Type of task') - prompt: str = Field(..., description='The text prompt describing the model to generate', max_length=1024) - negative_prompt: Optional[str] = Field(None, description='The negative text prompt', max_length=1024) - model_version: Optional[TripoModelVersion] = TripoModelVersion.v2_5_20250123 - face_limit: Optional[int] = Field(None, description='The number of faces to limit the generation to') - texture: Optional[bool] = Field(True, description='Whether to apply texture to the generated model') - pbr: Optional[bool] = Field(True, description='Whether to apply PBR to the generated model') - image_seed: Optional[int] = Field(None, description='The seed for the text') - model_seed: Optional[int] = Field(None, description='The seed for the model') - texture_seed: Optional[int] = Field(None, description='The seed for the texture') - texture_quality: Optional[TripoTextureQuality] = TripoTextureQuality.standard - geometry_quality: Optional[TripoGeometryQuality] = TripoGeometryQuality.standard - style: Optional[TripoStyle] = None - auto_size: Optional[bool] = Field(False, description='Whether to auto-size the model') - quad: Optional[bool] = Field(False, description='Whether to apply quad to the generated model') + type: TripoTaskType = Field(TripoTaskType.TEXT_TO_MODEL, description="Type of task") + prompt: str = Field(..., description="The text prompt describing the model to generate", max_length=1024) + negative_prompt: str | None = Field(None, description="The negative text prompt", max_length=1024) + model_version: TripoModelVersion | None = TripoModelVersion.v2_5_20250123 + face_limit: int | None = Field(None, description="The number of faces to limit the generation to") + texture: bool | None = Field(True, description="Whether to apply texture to the generated model") + pbr: bool | None = Field(True, description="Whether to apply PBR to the generated model") + image_seed: int | None = Field(None, description="The seed for the text") + model_seed: int | None = Field(None, description="The seed for the model") + texture_seed: int | None = Field(None, description="The seed for the texture") + texture_quality: TripoTextureQuality | None = TripoTextureQuality.standard + geometry_quality: TripoGeometryQuality | None = TripoGeometryQuality.standard + style: TripoStyle | None = None + auto_size: bool | None = Field(False, description="Whether to auto-size the model") + quad: bool | None = Field(False, description="Whether to apply quad to the generated model") + class TripoImageToModelRequest(BaseModel): - type: TripoTaskType = Field(TripoTaskType.IMAGE_TO_MODEL, description='Type of task') - file: TripoFileReference = Field(..., description='The file reference to convert to a model') - model_version: Optional[TripoModelVersion] = Field(None, description='The model version to use for generation') - face_limit: Optional[int] = Field(None, description='The number of faces to limit the generation to') - texture: Optional[bool] = Field(True, description='Whether to apply texture to the generated model') - pbr: Optional[bool] = Field(True, description='Whether to apply PBR to the generated model') - model_seed: Optional[int] = Field(None, description='The seed for the model') - texture_seed: Optional[int] = Field(None, description='The seed for the texture') - texture_quality: Optional[TripoTextureQuality] = TripoTextureQuality.standard - geometry_quality: Optional[TripoGeometryQuality] = TripoGeometryQuality.standard - texture_alignment: Optional[TripoTextureAlignment] = Field(TripoTextureAlignment.ORIGINAL_IMAGE, description='The texture alignment method') - style: Optional[TripoStyle] = Field(None, description='The style to apply to the generated model') - auto_size: Optional[bool] = Field(False, description='Whether to auto-size the model') - orientation: Optional[TripoOrientation] = TripoOrientation.DEFAULT - quad: Optional[bool] = Field(False, description='Whether to apply quad to the generated model') + type: TripoTaskType = Field(TripoTaskType.IMAGE_TO_MODEL, description="Type of task") + file: TripoFileReference = Field(..., description="The file reference to convert to a model") + model_version: TripoModelVersion | None = Field(None, description="The model version to use for generation") + face_limit: int | None = Field(None, description="The number of faces to limit the generation to") + texture: bool | None = Field(True, description="Whether to apply texture to the generated model") + pbr: bool | None = Field(True, description="Whether to apply PBR to the generated model") + model_seed: int | None = Field(None, description="The seed for the model") + texture_seed: int | None = Field(None, description="The seed for the texture") + texture_quality: TripoTextureQuality | None = TripoTextureQuality.standard + geometry_quality: TripoGeometryQuality | None = TripoGeometryQuality.standard + texture_alignment: TripoTextureAlignment | None = Field( + TripoTextureAlignment.ORIGINAL_IMAGE, description="The texture alignment method" + ) + style: TripoStyle | None = Field(None, description="The style to apply to the generated model") + auto_size: bool | None = Field(False, description="Whether to auto-size the model") + orientation: TripoOrientation | None = TripoOrientation.DEFAULT + quad: bool | None = Field(False, description="Whether to apply quad to the generated model") + class TripoMultiviewToModelRequest(BaseModel): type: TripoTaskType = TripoTaskType.MULTIVIEW_TO_MODEL - files: list[TripoFileReference] = Field(..., description='The file references to convert to a model') - model_version: Optional[TripoModelVersion] = Field(None, description='The model version to use for generation') - orthographic_projection: Optional[bool] = Field(False, description='Whether to use orthographic projection') - face_limit: Optional[int] = Field(None, description='The number of faces to limit the generation to') - texture: Optional[bool] = Field(True, description='Whether to apply texture to the generated model') - pbr: Optional[bool] = Field(True, description='Whether to apply PBR to the generated model') - model_seed: Optional[int] = Field(None, description='The seed for the model') - texture_seed: Optional[int] = Field(None, description='The seed for the texture') - texture_quality: Optional[TripoTextureQuality] = TripoTextureQuality.standard - geometry_quality: Optional[TripoGeometryQuality] = TripoGeometryQuality.standard - texture_alignment: Optional[TripoTextureAlignment] = TripoTextureAlignment.ORIGINAL_IMAGE - auto_size: Optional[bool] = Field(False, description='Whether to auto-size the model') - orientation: Optional[TripoOrientation] = Field(TripoOrientation.DEFAULT, description='The orientation for the model') - quad: Optional[bool] = Field(False, description='Whether to apply quad to the generated model') + files: list[TripoFileReference] = Field(..., description="The file references to convert to a model") + model_version: TripoModelVersion | None = Field(None, description="The model version to use for generation") + orthographic_projection: bool | None = Field(False, description="Whether to use orthographic projection") + face_limit: int | None = Field(None, description="The number of faces to limit the generation to") + texture: bool | None = Field(True, description="Whether to apply texture to the generated model") + pbr: bool | None = Field(True, description="Whether to apply PBR to the generated model") + model_seed: int | None = Field(None, description="The seed for the model") + texture_seed: int | None = Field(None, description="The seed for the texture") + texture_quality: TripoTextureQuality | None = TripoTextureQuality.standard + geometry_quality: TripoGeometryQuality | None = TripoGeometryQuality.standard + texture_alignment: TripoTextureAlignment | None = TripoTextureAlignment.ORIGINAL_IMAGE + auto_size: bool | None = Field(False, description="Whether to auto-size the model") + orientation: TripoOrientation | None = Field(TripoOrientation.DEFAULT, description="The orientation for the model") + quad: bool | None = Field(False, description="Whether to apply quad to the generated model") + class TripoTextureModelRequest(BaseModel): - type: TripoTaskType = Field(TripoTaskType.TEXTURE_MODEL, description='Type of task') - original_model_task_id: str = Field(..., description='The task ID of the original model') - texture: Optional[bool] = Field(True, description='Whether to apply texture to the model') - pbr: Optional[bool] = Field(True, description='Whether to apply PBR to the model') - model_seed: Optional[int] = Field(None, description='The seed for the model') - texture_seed: Optional[int] = Field(None, description='The seed for the texture') - texture_quality: Optional[TripoTextureQuality] = Field(None, description='The quality of the texture') - texture_alignment: Optional[TripoTextureAlignment] = Field(TripoTextureAlignment.ORIGINAL_IMAGE, description='The texture alignment method') + type: TripoTaskType = Field(TripoTaskType.TEXTURE_MODEL, description="Type of task") + original_model_task_id: str = Field(..., description="The task ID of the original model") + texture: bool | None = Field(True, description="Whether to apply texture to the model") + pbr: bool | None = Field(True, description="Whether to apply PBR to the model") + model_seed: int | None = Field(None, description="The seed for the model") + texture_seed: int | None = Field(None, description="The seed for the texture") + texture_quality: TripoTextureQuality | None = Field(None, description="The quality of the texture") + texture_alignment: TripoTextureAlignment | None = Field( + TripoTextureAlignment.ORIGINAL_IMAGE, description="The texture alignment method" + ) + class TripoRefineModelRequest(BaseModel): - type: TripoTaskType = Field(TripoTaskType.REFINE_MODEL, description='Type of task') - draft_model_task_id: str = Field(..., description='The task ID of the draft model') + type: TripoTaskType = Field(TripoTaskType.REFINE_MODEL, description="Type of task") + draft_model_task_id: str = Field(..., description="The task ID of the draft model") -class TripoAnimatePrerigcheckRequest(BaseModel): - type: TripoTaskType = Field(TripoTaskType.ANIMATE_PRERIGCHECK, description='Type of task') - original_model_task_id: str = Field(..., description='The task ID of the original model') class TripoAnimateRigRequest(BaseModel): - type: TripoTaskType = Field(TripoTaskType.ANIMATE_RIG, description='Type of task') - original_model_task_id: str = Field(..., description='The task ID of the original model') - out_format: Optional[TripoOutFormat] = Field(TripoOutFormat.GLB, description='The output format') - spec: Optional[TripoSpec] = Field(TripoSpec.TRIPO, description='The specification for rigging') + type: TripoTaskType = Field(TripoTaskType.ANIMATE_RIG, description="Type of task") + original_model_task_id: str = Field(..., description="The task ID of the original model") + out_format: TripoOutFormat | None = Field(TripoOutFormat.GLB, description="The output format") + spec: TripoSpec | None = Field(TripoSpec.TRIPO, description="The specification for rigging") + class TripoAnimateRetargetRequest(BaseModel): - type: TripoTaskType = Field(TripoTaskType.ANIMATE_RETARGET, description='Type of task') - original_model_task_id: str = Field(..., description='The task ID of the original model') - animation: TripoAnimation = Field(..., description='The animation to apply') - out_format: Optional[TripoOutFormat] = Field(TripoOutFormat.GLB, description='The output format') - bake_animation: Optional[bool] = Field(True, description='Whether to bake the animation') + type: TripoTaskType = Field(TripoTaskType.ANIMATE_RETARGET, description="Type of task") + original_model_task_id: str = Field(..., description="The task ID of the original model") + animation: TripoAnimation = Field(..., description="The animation to apply") + out_format: TripoOutFormat | None = Field(TripoOutFormat.GLB, description="The output format") + bake_animation: bool | None = Field(True, description="Whether to bake the animation") -class TripoStylizeModelRequest(BaseModel): - type: TripoTaskType = Field(TripoTaskType.STYLIZE_MODEL, description='Type of task') - style: TripoStylizeStyle = Field(..., description='The style to apply to the model') - original_model_task_id: str = Field(..., description='The task ID of the original model') - block_size: Optional[int] = Field(80, description='The block size for stylization') class TripoConvertModelRequest(BaseModel): - type: TripoTaskType = Field(TripoTaskType.CONVERT_MODEL, description='Type of task') - format: TripoConvertFormat = Field(..., description='The format to convert to') - original_model_task_id: str = Field(..., description='The task ID of the original model') - quad: Optional[bool] = Field(None, description='Whether to apply quad to the model') - force_symmetry: Optional[bool] = Field(None, description='Whether to force symmetry') - face_limit: Optional[int] = Field(None, description='The number of faces to limit the conversion to') - flatten_bottom: Optional[bool] = Field(None, description='Whether to flatten the bottom of the model') - flatten_bottom_threshold: Optional[float] = Field(None, description='The threshold for flattening the bottom') - texture_size: Optional[int] = Field(None, description='The size of the texture') - texture_format: Optional[TripoTextureFormat] = Field(TripoTextureFormat.JPEG, description='The format of the texture') - pivot_to_center_bottom: Optional[bool] = Field(None, description='Whether to pivot to the center bottom') - scale_factor: Optional[float] = Field(None, description='The scale factor for the model') - with_animation: Optional[bool] = Field(None, description='Whether to include animations') - pack_uv: Optional[bool] = Field(None, description='Whether to pack the UVs') - bake: Optional[bool] = Field(None, description='Whether to bake the model') - part_names: Optional[list[str]] = Field(None, description='The names of the parts to include') - fbx_preset: Optional[TripoFbxPreset] = Field(None, description='The preset for the FBX export') - export_vertex_colors: Optional[bool] = Field(None, description='Whether to export the vertex colors') - export_orientation: Optional[TripoOrientation] = Field(None, description='The orientation for the export') - animate_in_place: Optional[bool] = Field(None, description='Whether to animate in place') + type: TripoTaskType = Field(TripoTaskType.CONVERT_MODEL, description="Type of task") + format: TripoConvertFormat = Field(..., description="The format to convert to") + original_model_task_id: str = Field(..., description="The task ID of the original model") + quad: bool | None = Field(None, description="Whether to apply quad to the model") + force_symmetry: bool | None = Field(None, description="Whether to force symmetry") + face_limit: int | None = Field(None, description="The number of faces to limit the conversion to") + flatten_bottom: bool | None = Field(None, description="Whether to flatten the bottom of the model") + flatten_bottom_threshold: float | None = Field(None, description="The threshold for flattening the bottom") + texture_size: int | None = Field(None, description="The size of the texture") + texture_format: TripoTextureFormat | None = Field(TripoTextureFormat.JPEG, description="The format of the texture") + pivot_to_center_bottom: bool | None = Field(None, description="Whether to pivot to the center bottom") + scale_factor: float | None = Field(None, description="The scale factor for the model") + with_animation: bool | None = Field(None, description="Whether to include animations") + pack_uv: bool | None = Field(None, description="Whether to pack the UVs") + bake: bool | None = Field(None, description="Whether to bake the model") + part_names: list[str] | None = Field(None, description="The names of the parts to include") + fbx_preset: TripoFbxPreset | None = Field(None, description="The preset for the FBX export") + export_vertex_colors: bool | None = Field(None, description="Whether to export the vertex colors") + export_orientation: TripoOrientation | None = Field(None, description="The orientation for the export") + animate_in_place: bool | None = Field(None, description="Whether to animate in place") + + +class TripoP1CommonRequest(BaseModel): + """Fields supported by Tripo P1 across all input types.""" + + model_version: str = Field("P1-20260311") + model_seed: int | None = Field(None, description="Random seed for geometry generation") + face_limit: int | None = Field(None, ge=48, le=20000, description="Target face count (48-20000)") + texture: bool | None = Field(None, description="Enable texturing; pbr=True forces this true") + pbr: bool | None = Field(None, description="Enable PBR maps; when true, texture is also enabled") + texture_seed: int | None = Field(None, description="Random seed for texture generation") + texture_quality: str | None = Field(None, description='"standard" or "detailed"') + auto_size: bool | None = Field(None, description="Scale to real-world meters") + compress: str | None = Field(None, description='Only "geometry" is supported') + export_uv: bool | None = Field(None, description="Perform UV unwrapping during generation") + + +class TripoP1TextToModelRequest(TripoP1CommonRequest): + type: str = "text_to_model" + prompt: str = Field(..., max_length=1024) + negative_prompt: str | None = Field(None, max_length=255) + image_seed: int | None = None + + +class TripoP1ImageToModelRequest(TripoP1CommonRequest): + type: str = "image_to_model" + file: TripoFileReference + enable_image_autofix: bool | None = None + texture_alignment: str | None = Field(None, description='"original_image" or "geometry"') + orientation: str | None = Field(None, description='"default" or "align_image"; needs texture=true') + + +class TripoP1MultiviewToModelRequest(TripoP1CommonRequest): + """P1 multiview generation. + + Tripo requires `files` to be exactly four entries in [front, left, back, right] order with `{}` + (TripoFileEmptyReference) for omitted slots; front is required and at least two images total must be provided. + """ + + type: str = "multiview_to_model" + files: list[TripoFileReference] + texture_alignment: str | None = None + orientation: str | None = None class TripoTaskOutput(BaseModel): - model: Optional[str] = Field(None, description='URL to the model') - base_model: Optional[str] = Field(None, description='URL to the base model') - pbr_model: Optional[str] = Field(None, description='URL to the PBR model') - rendered_image: Optional[str] = Field(None, description='URL to the rendered image') - riggable: Optional[bool] = Field(None, description='Whether the model is riggable') + model: str | None = Field(None, description="URL to the model") + base_model: str | None = Field(None, description="URL to the base model") + pbr_model: str | None = Field(None, description="URL to the PBR model") + rendered_image: str | None = Field(None, description="URL to the rendered image") + riggable: bool | None = Field(None, description="Whether the model is riggable") + class TripoTask(BaseModel): - task_id: str = Field(..., description='The task ID') - type: Optional[str] = Field(None, description='The type of task') - status: Optional[TripoTaskStatus] = Field(None, description='The status of the task') - input: Optional[dict[str, Any]] = Field(None, description='The input parameters for the task') - output: Optional[TripoTaskOutput] = Field(None, description='The output of the task') - progress: Optional[int] = Field(None, description='The progress of the task', ge=0, le=100) - create_time: Optional[int] = Field(None, description='The creation time of the task') - running_left_time: Optional[int] = Field(None, description='The estimated time left for the task') - queue_position: Optional[int] = Field(None, description='The position in the queue') + task_id: str = Field(..., description="The task ID") + type: str | None = Field(None, description="The type of task") + status: TripoTaskStatus | None = Field(None, description="The status of the task") + input: dict[str, Any] | None = Field(None, description="The input parameters for the task") + output: TripoTaskOutput | None = Field(None, description="The output of the task") + progress: int | None = Field(None, description="The progress of the task", ge=0, le=100) + create_time: int | None = Field(None, description="The creation time of the task") + running_left_time: int | None = Field(None, description="The estimated time left for the task") + queue_position: int | None = Field(None, description="The position in the queue") consumed_credit: int | None = Field(None) + class TripoTaskResponse(BaseModel): - code: int = Field(0, description='The response code') - data: TripoTask = Field(..., description='The task data') + code: int = Field(0, description="The response code") + data: TripoTask = Field(..., description="The task data") -class TripoGeneralResponse(BaseModel): - code: int = Field(0, description='The response code') - data: dict[str, str] = Field(..., description='The task ID data') - -class TripoBalanceData(BaseModel): - balance: float = Field(..., description='The account balance') - frozen: float = Field(..., description='The frozen balance') - -class TripoBalanceResponse(BaseModel): - code: int = Field(0, description='The response code') - data: TripoBalanceData = Field(..., description='The balance data') class TripoErrorResponse(BaseModel): - code: int = Field(..., description='The error code') - message: str = Field(..., description='The error message') - suggestion: str = Field(..., description='The suggestion for fixing the error') + code: int = Field(..., description="The error code") + message: str = Field(..., description="The error message") + suggestion: str = Field(..., description="The suggestion for fixing the error") diff --git a/comfy_api_nodes/nodes_anthropic.py b/comfy_api_nodes/nodes_anthropic.py index 28dd70d4e..87a870553 100644 --- a/comfy_api_nodes/nodes_anthropic.py +++ b/comfy_api_nodes/nodes_anthropic.py @@ -9,8 +9,11 @@ from comfy_api_nodes.apis.anthropic import ( AnthropicMessage, AnthropicMessagesRequest, AnthropicMessagesResponse, + AnthropicOutputConfig, + AnthropicResponseTextBlock, AnthropicRole, AnthropicTextContent, + AnthropicThinkingConfig, ) from comfy_api_nodes.util import ( ApiEndpoint, @@ -32,15 +35,29 @@ CLAUDE_MODELS: dict[str, str] = { "Haiku 4.5": "claude-haiku-4-5-20251001", } +_THINKING_UNSUPPORTED = {"Haiku 4.5"} +# Models that use the newer "adaptive" thinking mode (Opus 4.7 requires it; older models keep the explicit budget API). +# Anthropic decides the actual budget when adaptive is used, based on the `output_config.effort` hint. +_ADAPTIVE_THINKING_MODELS = {"Opus 4.7", "Opus 4.6", "Sonnet 4.6"} -def _claude_model_inputs(): - return [ +# Budget mode (Sonnet 4.5): effort -> reasoning budget in tokens. Must be < max_tokens. +# Sized so even the "high" budget fits comfortably under the default max_tokens=32768. +_REASONING_BUDGET: dict[str, int] = { + "low": 2048, + "medium": 8192, + "high": 16384, +} +_REASONING_EFFORTS = ["off", "low", "medium", "high"] + + +def _claude_model_inputs(model_label: str): + inputs: list = [ IO.Int.Input( "max_tokens", - default=16000, - min=32, - max=32000, - tooltip="Maximum number of tokens to generate before stopping.", + default=32768, + min=4096, + max=64000, + tooltip="Maximum number of tokens to generate (includes reasoning tokens when enabled).", advanced=True, ), IO.Float.Input( @@ -49,10 +66,24 @@ def _claude_model_inputs(): min=0.0, max=1.0, step=0.01, - tooltip="Controls randomness. 0.0 is deterministic, 1.0 is most random. Ignored for Opus 4.7.", + tooltip=( + "Controls randomness. 0.0 is deterministic, 1.0 is most random. " + "Ignored for Opus 4.7 and any model when reasoning_effort is set." + ), advanced=True, ), ] + if model_label not in _THINKING_UNSUPPORTED: + inputs.append( + IO.Combo.Input( + "reasoning_effort", + options=_REASONING_EFFORTS, + default="off", + tooltip="Extended thinking effort. 'off' disables reasoning.", + advanced=True, + ) + ) + return inputs def _model_price_per_million(model: str) -> tuple[float, float] | None: @@ -95,7 +126,11 @@ def calculate_tokens_price(response: AnthropicMessagesResponse) -> float | None: def _get_text_from_response(response: AnthropicMessagesResponse) -> str: if not response.content: return "" - return "\n".join(block.text for block in response.content if block.text) + # Thinking blocks are silently dropped — we never want reasoning in the output. + return "\n".join( + block.text for block in response.content + if isinstance(block, AnthropicResponseTextBlock) and block.text + ) async def _build_image_content_blocks( @@ -120,7 +155,7 @@ class ClaudeNode(IO.ComfyNode): return IO.Schema( node_id="ClaudeNode", display_name="Anthropic Claude", - category="api node/text/Anthropic", + category="partner/text/Anthropic", essentials_category="Text Generation", description="Generate text responses with Anthropic's Claude models. " "Provide a text prompt and optionally one or more images for multimodal context.", @@ -133,7 +168,10 @@ class ClaudeNode(IO.ComfyNode): ), IO.DynamicCombo.Input( "model", - options=[IO.DynamicCombo.Option(label, _claude_model_inputs()) for label in CLAUDE_MODELS], + options=[ + IO.DynamicCombo.Option(label, _claude_model_inputs(label)) + for label in CLAUDE_MODELS + ], tooltip="The Claude model used to generate the response.", ), IO.Int.Input( @@ -207,8 +245,29 @@ class ClaudeNode(IO.ComfyNode): ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=True, min_length=1) model_label = model["model"] - max_tokens = model["max_tokens"] - temperature = None if model_label == "Opus 4.7" else model["temperature"] + max_tokens = model.get("max_tokens", 32768) + reasoning_effort = model.get("reasoning_effort", "off") + thinking_enabled = reasoning_effort not in ("off", None) and model_label not in _THINKING_UNSUPPORTED + + # Anthropic requires temperature to be unset (defaults to 1.0) when thinking is enabled. + # Opus 4.7 also rejects user-supplied temperature. + if thinking_enabled or model_label == "Opus 4.7": + temperature = None + else: + temperature = model.get("temperature", 1.0) + + thinking_cfg: AnthropicThinkingConfig | None = None + output_cfg: AnthropicOutputConfig | None = None + if thinking_enabled: + if model_label in _ADAPTIVE_THINKING_MODELS: + # Adaptive mode - Anthropic chooses the budget based on effort hint + thinking_cfg = AnthropicThinkingConfig(type="adaptive") + output_cfg = AnthropicOutputConfig(effort=reasoning_effort) + else: + # Budget mode (Sonnet 4.5). Leave at least 1024 tokens for the actual response + budget = _REASONING_BUDGET[reasoning_effort] + budget = min(budget, max(1024, max_tokens - 1024)) + thinking_cfg = AnthropicThinkingConfig(type="enabled", budget_tokens=budget) image_tensors: list[Input.Image] = [t for t in (images or {}).values() if t is not None] if sum(get_number_of_images(t) for t in image_tensors) > CLAUDE_MAX_IMAGES: @@ -229,6 +288,8 @@ class ClaudeNode(IO.ComfyNode): messages=[AnthropicMessage(role=AnthropicRole.user, content=content)], system=system_prompt or None, temperature=temperature, + thinking=thinking_cfg, + output_config=output_cfg, ), price_extractor=calculate_tokens_price, ) diff --git a/comfy_api_nodes/nodes_beeble.py b/comfy_api_nodes/nodes_beeble.py new file mode 100644 index 000000000..d863c2130 --- /dev/null +++ b/comfy_api_nodes/nodes_beeble.py @@ -0,0 +1,404 @@ +from fractions import Fraction + +from typing_extensions import override + +from comfy_api.latest import IO, ComfyExtension, Input, InputImpl, Types +from comfy_api_nodes.apis.beeble import ( + CreateSwitchXRequest, + SwitchXStatusResponse, +) +from comfy_api_nodes.util import ( + ApiEndpoint, + bytesio_to_image_tensor, + convert_mask_to_image, + download_url_as_bytesio, + download_url_to_image_tensor, + download_url_to_video_output, + downscale_image_tensor, + downscale_video_to_max_pixels, + poll_op, + sync_op, + upload_image_to_comfyapi, + upload_video_to_comfyapi, + validate_string, + validate_video_frame_count, +) + +_MAX_PIXELS = 2_770_000 +_MAX_FRAMES = 240 +_MAX_PROMPT_LEN = 2000 + + +def _validate_inputs(prompt: str | None, reference_image: Input.Image | None) -> str | None: + """Beeble requires at least one of prompt or reference_image. Returns the cleaned prompt.""" + cleaned = prompt.strip() if prompt else "" + if not cleaned and reference_image is None: + raise ValueError("At least one of 'prompt' or 'reference_image' must be provided.") + if cleaned: + validate_string(cleaned, strip_whitespace=False, max_length=_MAX_PROMPT_LEN) + return cleaned or None + + +async def _upload_mask_as_image( + cls: type[IO.ComfyNode], + mask: Input.Image, + *, + wait_label: str, +) -> str: + """Encode a single-frame MASK (H, W) or (1, H, W) as a PNG and upload.""" + if mask.dim() == 2: + mask = mask.unsqueeze(0) + image = convert_mask_to_image(mask[:1]) + return await upload_image_to_comfyapi( + cls, + image, + mime_type="image/png", + wait_label=wait_label, + total_pixels=_MAX_PIXELS, + ) + + +async def _upload_mask_batch_as_video( + cls: type[IO.ComfyNode], + mask: Input.Image, + *, + frame_rate: Fraction, + source_frame_count: int, + wait_label: str, +) -> str: + """Encode a MASK batch (N, H, W) as a grayscale H.264 MP4 at frame_rate and upload. + + The matte is always downscaled to the pixel budget so it stays within Beeble's limit and + keeps the same dimensions as the (similarly downscaled) source — both use the same algorithm + from the same starting dimensions, and downscaling is a no-op when already within budget. + """ + if mask.dim() == 2: + mask = mask.unsqueeze(0) + if mask.shape[0] != source_frame_count: + raise ValueError( + f"Custom alpha video frame count ({mask.shape[0]}) does not match the " + f"source video frame count ({source_frame_count}). The Beeble API requires " + "one mask per source frame." + ) + images = downscale_image_tensor(convert_mask_to_image(mask), _MAX_PIXELS) + alpha_video = InputImpl.VideoFromComponents(Types.VideoComponents(images=images, audio=None, frame_rate=frame_rate)) + return await upload_video_to_comfyapi(cls, alpha_video, wait_label=wait_label) + + +def _alpha_mode_input(*, video: bool) -> IO.DynamicCombo.Input: + """Build the alpha_mode DynamicCombo with mode-specific extra inputs.""" + select_keyframe_tooltip = ( + "First-frame keyframe mask. Beeble propagates this across the video." if video else "Grayscale keyframe mask." + ) + custom_tooltip = ( + "Per-frame grayscale mask covering the entire video. " + "Must have the same frame count as the source. " + "Connect a MASK output from SAM3_TrackToMask or similar." + if video + else "Grayscale mask to apply." + ) + return IO.DynamicCombo.Input( + "alpha_mode", + tooltip=( + "Controls how SwitchX decides what to keep vs. regenerate. " + "'auto' isolates the main subject automatically. " + "'fill' regenerates the entire frame while preserving geometry. " + "'select' propagates a first-frame keyframe across the clip. " + "'custom' uses a per-frame alpha matte you provide." + ), + options=[ + IO.DynamicCombo.Option("auto", []), + IO.DynamicCombo.Option("fill", []), + IO.DynamicCombo.Option( + "select", + [IO.Mask.Input("alpha_keyframe", tooltip=select_keyframe_tooltip)], + ), + IO.DynamicCombo.Option( + "custom", + [IO.Mask.Input("alpha_mask", tooltip=custom_tooltip)], + ), + ], + ) + + +def _common_inputs(*, source: IO.Input, video: bool) -> list[IO.Input]: + return [ + source, + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip=( + "Text description of the desired output (max 2000 chars). " + "At least one of 'prompt' or 'reference_image' is required." + ), + ), + IO.Image.Input( + "reference_image", + optional=True, + tooltip=( + "Reference image whose look (background, lighting, costume) the result " + "should adopt. At least one of 'reference_image' or 'prompt' is required." + ), + ), + _alpha_mode_input(video=video), + IO.Combo.Input( + "max_resolution", + options=["1080p", "720p"], + default="1080p", + tooltip="Maximum output resolution.", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + control_after_generate=True, + tooltip=( + "Seed controls whether the node should re-run; " "results are non-deterministic regardless of seed." + ), + ), + ] + + +async def _submit_and_poll( + cls: type[IO.ComfyNode], + request: CreateSwitchXRequest, +) -> SwitchXStatusResponse: + initial = await sync_op( + cls, + ApiEndpoint(path="/proxy/beeble/v1/switchx/generations", method="POST"), + response_model=SwitchXStatusResponse, + data=request, + ) + return await poll_op( + cls, + ApiEndpoint(path=f"/proxy/beeble/v1/switchx/generations/{initial.id}"), + response_model=SwitchXStatusResponse, + status_extractor=lambda r: r.status, + progress_extractor=lambda r: r.progress, + ) + + +def _require_output_url(response: SwitchXStatusResponse, name: str) -> str: + if response.output is None or getattr(response.output, name) is None: + raise RuntimeError(f"Beeble job {response.id} completed without a {name!r} output URL.") + return getattr(response.output, name) + + +def _alpha_url(response: SwitchXStatusResponse, mode: str) -> str | None: + """URL of the alpha matte, or None when the mode produces no separate matte. + + 'fill' selects the whole frame, so Beeble writes no alpha asset even though the status + response still returns a (dangling) signed URL for it — fetching it 403s with S3 + AccessDenied. The other three modes ('auto', 'custom', 'select') all produce a real, + downloadable matte. + """ + if mode == "fill" or response.output is None: + return None + return response.output.alpha + + +class BeebleSwitchXVideoEdit(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="BeebleSwitchXVideoEdit", + display_name="Beeble SwitchX Video Edit", + category="partner/video/Beeble", + description=( + "Edit a video with Beeble SwitchX. Switches anything in the scene (background, " + "lighting, costume) while preserving the original subject's pixels and motion. " + "Provide a reference image and/or text prompt to describe the new look. " + "Max 240 frames, max ~2.77MP per frame." + ), + inputs=_common_inputs(source=IO.Video.Input("video"), video=True), + outputs=[ + IO.Video.Output(display_name="video"), + IO.Video.Output( + display_name="alpha", + tooltip="The alpha matte Beeble used. Empty for 'fill' mode, which has no separate matte.", + ), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["max_resolution"]), + expr=""" + ( + $rate := widgets.max_resolution = "1080p" ? 0.429 : 0.143; + {"type":"usd","usd": $rate, "format":{"suffix":"/30 frames"}} + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + video: Input.Video, + prompt: str, + alpha_mode: dict, + max_resolution: str, + seed: int, + reference_image: Input.Image | None = None, + ) -> IO.NodeOutput: + cleaned_prompt = _validate_inputs(prompt, reference_image) + + validate_video_frame_count(video, max_frame_count=_MAX_FRAMES) + video = downscale_video_to_max_pixels(video, _MAX_PIXELS) + + mode = alpha_mode["alpha_mode"] + alpha_uri: str | None = None + if mode == "select": + alpha_uri = await _upload_mask_as_image(cls, alpha_mode["alpha_keyframe"], wait_label="Uploading keyframe") + elif mode == "custom": + alpha_uri = await _upload_mask_batch_as_video( + cls, + alpha_mode["alpha_mask"], + frame_rate=video.get_frame_rate(), + source_frame_count=video.get_frame_count(), + wait_label="Uploading alpha video", + ) + + source_uri = await upload_video_to_comfyapi(cls, video, wait_label="Uploading source") + reference_uri: str | None = None + if reference_image is not None: + reference_uri = await upload_image_to_comfyapi( + cls, + reference_image, + mime_type="image/png", + wait_label="Uploading reference", + total_pixels=_MAX_PIXELS, + ) + + request = CreateSwitchXRequest( + generation_type="video", + source_uri=source_uri, + alpha_mode=mode, + prompt=cleaned_prompt, + reference_image_uri=reference_uri, + alpha_uri=alpha_uri, + max_resolution=1080 if max_resolution == "1080p" else 720, + ) + response = await _submit_and_poll(cls, request) + + render = await download_url_to_video_output(_require_output_url(response, "render")) + alpha = None + if (alpha_url := _alpha_url(response, mode)) is not None: + alpha = await download_url_to_video_output(alpha_url) + return IO.NodeOutput(render, alpha) + + +class BeebleSwitchXImageEdit(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="BeebleSwitchXImageEdit", + display_name="Beeble SwitchX Image Edit", + category="partner/image/Beeble", + description=( + "Edit a single image with Beeble SwitchX. Switches anything in the scene " + "(background, lighting, costume) while preserving the original subject's pixels. " + "Provide a reference image and/or text prompt to describe the new look. " + "Max ~2.77MP." + ), + inputs=_common_inputs(source=IO.Image.Input("image"), video=False), + outputs=[ + IO.Image.Output(display_name="image"), + IO.Mask.Output( + display_name="alpha", + tooltip="The alpha matte Beeble used. Empty for 'fill' mode, which has no separate matte.", + ), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["max_resolution"]), + expr=""" + ( + $rate := widgets.max_resolution = "1080p" ? 0.429 : 0.143; + {"type":"usd","usd": $rate} + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + image: Input.Image, + prompt: str, + alpha_mode: dict, + max_resolution: str, + seed: int, + reference_image: Input.Image | None = None, + ) -> IO.NodeOutput: + cleaned_prompt = _validate_inputs(prompt, reference_image) + + image = downscale_image_tensor(image, _MAX_PIXELS) + + mode = alpha_mode["alpha_mode"] + alpha_uri: str | None = None + if mode == "select": + alpha_uri = await _upload_mask_as_image(cls, alpha_mode["alpha_keyframe"], wait_label="Uploading keyframe") + elif mode == "custom": + alpha_uri = await _upload_mask_as_image(cls, alpha_mode["alpha_mask"], wait_label="Uploading alpha") + + source_uri = await upload_image_to_comfyapi( + cls, + image, + mime_type="image/png", + wait_label="Uploading source", + total_pixels=None, + ) + reference_uri: str | None = None + if reference_image is not None: + reference_uri = await upload_image_to_comfyapi( + cls, + reference_image, + mime_type="image/png", + wait_label="Uploading reference", + total_pixels=_MAX_PIXELS, + ) + + request = CreateSwitchXRequest( + generation_type="image", + source_uri=source_uri, + alpha_mode=mode, + prompt=cleaned_prompt, + reference_image_uri=reference_uri, + alpha_uri=alpha_uri, + max_resolution=1080 if max_resolution == "1080p" else 720, + ) + response = await _submit_and_poll(cls, request) + + render = await download_url_to_image_tensor(_require_output_url(response, "render")) + alpha_mask = None + if (alpha_url := _alpha_url(response, mode)) is not None: + alpha_image = bytesio_to_image_tensor(await download_url_as_bytesio(alpha_url), mode="L") + alpha_mask = alpha_image.squeeze(-1) if alpha_image.dim() == 4 else alpha_image + return IO.NodeOutput(render, alpha_mask) + + +class BeebleExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + BeebleSwitchXVideoEdit, + BeebleSwitchXImageEdit, + ] + + +async def comfy_entrypoint() -> BeebleExtension: + return BeebleExtension() diff --git a/comfy_api_nodes/nodes_bfl.py b/comfy_api_nodes/nodes_bfl.py index 3f0ce29d8..259c54ef9 100644 --- a/comfy_api_nodes/nodes_bfl.py +++ b/comfy_api_nodes/nodes_bfl.py @@ -4,17 +4,20 @@ from typing_extensions import override from comfy_api.latest import IO, ComfyExtension, Input from comfy_api_nodes.apis.bfl import ( + BFLFluxEraseRequest, BFLFluxExpandImageRequest, BFLFluxFillImageRequest, BFLFluxKontextProGenerateRequest, BFLFluxProGenerateResponse, BFLFluxProUltraGenerateRequest, BFLFluxStatusResponse, + BFLFluxVTORequest, BFLStatus, Flux2ProGenerateRequest, ) from comfy_api_nodes.util import ( ApiEndpoint, + convert_mask_to_image, download_url_to_image_tensor, get_number_of_images, poll_op, @@ -22,19 +25,11 @@ from comfy_api_nodes.util import ( sync_op, tensor_to_base64_string, validate_aspect_ratio_string, + validate_image_dimensions, validate_string, ) -def convert_mask_to_image(mask: Input.Image): - """ - Make mask have the expected amount of dims (4) and channels (3) to be recognized as an image. - """ - mask = mask.unsqueeze(-1) - mask = torch.cat([mask] * 3, dim=-1) - return mask - - class FluxProUltraImageNode(IO.ComfyNode): @classmethod @@ -42,7 +37,7 @@ class FluxProUltraImageNode(IO.ComfyNode): return IO.Schema( node_id="FluxProUltraImageNode", display_name="Flux 1.1 [pro] Ultra Image", - category="api node/image/BFL", + category="partner/image/BFL", description="Generates images using Flux Pro 1.1 Ultra via api based on prompt and resolution.", inputs=[ IO.String.Input( @@ -160,7 +155,7 @@ class FluxKontextProImageNode(IO.ComfyNode): return IO.Schema( node_id=cls.NODE_ID, display_name=cls.DISPLAY_NAME, - category="api node/image/BFL", + category="partner/image/BFL", description="Edits images using Flux.1 Kontext [pro] via api based on prompt and aspect ratio.", inputs=[ IO.String.Input( @@ -282,7 +277,7 @@ class FluxProExpandNode(IO.ComfyNode): return IO.Schema( node_id="FluxProExpandNode", display_name="Flux.1 Expand Image", - category="api node/image/BFL", + category="partner/image/BFL", description="Outpaints image based on prompt.", inputs=[ IO.Image.Input("image"), @@ -419,7 +414,7 @@ class FluxProFillNode(IO.ComfyNode): return IO.Schema( node_id="FluxProFillNode", display_name="Flux.1 Fill Image", - category="api node/image/BFL", + category="partner/image/BFL", description="Inpaints image based on mask and prompt.", inputs=[ IO.Image.Input("image"), @@ -519,6 +514,174 @@ class FluxProFillNode(IO.ComfyNode): return IO.NodeOutput(await download_url_to_image_tensor(response.result["sample"])) +class FluxEraseNode(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="FluxEraseNode", + display_name="Flux Erase Image", + category="partner/image/BFL", + description="Removes the masked object from an image and reconstructs the background. " + "Paint the mask over what you want to erase.", + inputs=[ + IO.Image.Input("image"), + IO.Mask.Input("mask", tooltip="White areas are removed; black areas are preserved."), + IO.Int.Input( + "dilate_pixels", + default=10, + min=0, + max=25, + tooltip="Expands the mask boundaries to ensure clean coverage of the object's edges.", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + control_after_generate=True, + tooltip="The random seed used for creating the noise.", + optional=True, + ), + ], + outputs=[IO.Image.Output()], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"range_usd","min_usd":0.03,"max_usd":0.06,"format":{"approximate":true}}""", + ), + ) + + @classmethod + async def execute( + cls, + image: Input.Image, + mask: Input.Image, + dilate_pixels: int = 10, + seed: int = 0, + ) -> IO.NodeOutput: + validate_image_dimensions(image, min_width=256, min_height=256) + mask = resize_mask_to_image(mask, image) + mask = tensor_to_base64_string(convert_mask_to_image(mask)) + initial_response = await sync_op( + cls, + ApiEndpoint(path="/proxy/bfl/v1/flux-tools/erase-v1", method="POST"), + response_model=BFLFluxProGenerateResponse, + data=BFLFluxEraseRequest( + image=tensor_to_base64_string(image[:, :, :, :3]), # make sure image will have alpha channel removed + mask=mask, + dilate_pixels=dilate_pixels, + seed=seed, + ), + ) + + def price_extractor(_r: BaseModel) -> float | None: + return None if initial_response.cost is None else initial_response.cost / 100 + + response = await poll_op( + cls, + ApiEndpoint(initial_response.polling_url), + response_model=BFLFluxStatusResponse, + status_extractor=lambda r: r.status, + progress_extractor=lambda r: r.progress, + price_extractor=price_extractor, + completed_statuses=[BFLStatus.ready], + failed_statuses=[ + BFLStatus.request_moderated, + BFLStatus.content_moderated, + BFLStatus.error, + BFLStatus.task_not_found, + ], + queued_statuses=[], + ) + return IO.NodeOutput(await download_url_to_image_tensor(response.result["sample"])) + + +class FluxVTONode(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="FluxVTONode", + display_name="Flux Virtual Try-On", + category="partner/image/BFL", + description="Virtual try-on: dresses the person in the provided garment.", + inputs=[ + IO.Image.Input("person", tooltip="Image of the person to dress."), + IO.Image.Input("garment", tooltip="Image of the garment to apply."), + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Optional natural-language styling instruction (e.g. how the garment should fit).", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=0xFFFFFFFFFFFFFFFF, + control_after_generate=True, + tooltip="The random seed used for creating the noise.", + ), + ], + outputs=[IO.Image.Output()], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"range_usd","min_usd":0.0375,"max_usd":0.075,"format":{"approximate":true}}""", + ), + ) + + @classmethod + async def execute( + cls, + person: Input.Image, + garment: Input.Image, + prompt: str = "", + seed: int = 0, + ) -> IO.NodeOutput: + initial_response = await sync_op( + cls, + ApiEndpoint(path="/proxy/bfl/v1/flux-tools/vto-v1", method="POST"), + response_model=BFLFluxProGenerateResponse, + data=BFLFluxVTORequest( + prompt=prompt, + person=tensor_to_base64_string(person[:, :, :, :3]), + garment=tensor_to_base64_string(garment[:, :, :, :3]), + seed=seed, + ), + ) + + def price_extractor(_r: BaseModel) -> float | None: + return None if initial_response.cost is None else initial_response.cost / 100 + + response = await poll_op( + cls, + ApiEndpoint(initial_response.polling_url), + response_model=BFLFluxStatusResponse, + status_extractor=lambda r: r.status, + progress_extractor=lambda r: r.progress, + price_extractor=price_extractor, + completed_statuses=[BFLStatus.ready], + failed_statuses=[ + BFLStatus.request_moderated, + BFLStatus.content_moderated, + BFLStatus.error, + BFLStatus.task_not_found, + ], + queued_statuses=[], + ) + return IO.NodeOutput(await download_url_to_image_tensor(response.result["sample"])) + + class Flux2ProImageNode(IO.ComfyNode): NODE_ID = "Flux2ProImageNode" @@ -545,7 +708,7 @@ class Flux2ProImageNode(IO.ComfyNode): return IO.Schema( node_id=cls.NODE_ID, display_name=cls.DISPLAY_NAME, - category="api node/image/BFL", + category="partner/image/BFL", description="Generates images synchronously based on prompt and resolution.", inputs=[ IO.String.Input( @@ -716,7 +879,7 @@ class Flux2ImageNode(IO.ComfyNode): return IO.Schema( node_id="Flux2ImageNode", display_name="Flux.2 Image", - category="api node/image/BFL", + category="partner/image/BFL", description="Generate images via Flux.2 [pro] or Flux.2 [max] from a prompt and optional reference images.", inputs=[ IO.String.Input( @@ -853,6 +1016,8 @@ class BFLExtension(ComfyExtension): FluxKontextMaxImageNode, FluxProExpandNode, FluxProFillNode, + FluxEraseNode, + FluxVTONode, Flux2ProImageNode, Flux2MaxImageNode, Flux2ImageNode, diff --git a/comfy_api_nodes/nodes_bria.py b/comfy_api_nodes/nodes_bria.py index 4044ee3ea..e138fafa9 100644 --- a/comfy_api_nodes/nodes_bria.py +++ b/comfy_api_nodes/nodes_bria.py @@ -1,14 +1,19 @@ +import av +import torch +from av.codec import CodecContext from typing_extensions import override from comfy_api.latest import IO, ComfyExtension, Input from comfy_api_nodes.apis.bria import ( BriaEditImageRequest, + BriaImageEditResponse, BriaRemoveBackgroundRequest, BriaRemoveBackgroundResponse, BriaRemoveVideoBackgroundRequest, BriaRemoveVideoBackgroundResponse, - BriaImageEditResponse, BriaStatusResponse, + BriaVideoGreenScreenRequest, + BriaVideoReplaceBackgroundRequest, InputModerationSettings, ) from comfy_api_nodes.util import ( @@ -31,7 +36,7 @@ class BriaImageEditNode(IO.ComfyNode): return IO.Schema( node_id="BriaImageEditNode", display_name="Bria FIBO Image Edit", - category="api node/image/Bria", + category="partner/image/Bria", description="Edit images using Bria latest model", inputs=[ IO.Combo.Input("model", options=["FIBO"]), @@ -169,7 +174,7 @@ class BriaRemoveImageBackground(IO.ComfyNode): return IO.Schema( node_id="BriaRemoveImageBackground", display_name="Bria Remove Image Background", - category="api node/image/Bria", + category="partner/image/Bria", description="Remove the background from an image using Bria RMBG 2.0.", inputs=[ IO.Image.Input("image"), @@ -245,7 +250,7 @@ class BriaRemoveVideoBackground(IO.ComfyNode): return IO.Schema( node_id="BriaRemoveVideoBackground", display_name="Bria Remove Video Background", - category="api node/video/Bria", + category="partner/video/Bria", description="Remove the background from a video using Bria. ", inputs=[ IO.Video.Input("video"), @@ -316,6 +321,248 @@ class BriaRemoveVideoBackground(IO.ComfyNode): return IO.NodeOutput(await download_url_to_video_output(response.result.video_url)) +class BriaVideoGreenScreen(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="BriaVideoGreenScreen", + display_name="Bria Video Green Screen", + category="partner/video/Bria", + description="Replace a video's background with a solid chroma-key screen using Bria.", + inputs=[ + IO.Video.Input("video"), + IO.Combo.Input( + "green_shade", + options=["broadcast_green", "chroma_green", "blue_screen"], + tooltip="Solid chroma-key shade applied behind the foreground: " + "broadcast_green (#00B140), chroma_green (#00FF00), or blue_screen (#0000FF).", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed controls whether the node should re-run; " + "results are non-deterministic regardless of seed.", + ), + ], + outputs=[IO.Video.Output()], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.14,"format":{"suffix":"/second"}}""", + ), + ) + + @classmethod + async def execute( + cls, + video: Input.Video, + green_shade: str, + seed: int, + ) -> IO.NodeOutput: + validate_video_duration(video, max_duration=60.0) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/bria/v2/video/edit/green_screen", method="POST"), + data=BriaVideoGreenScreenRequest( + video=await upload_video_to_comfyapi(cls, video), + green_shade=green_shade, + output_container_and_codec="mp4_h264", + seed=seed, + ), + response_model=BriaStatusResponse, + ) + response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/bria/v2/status/{response.request_id}"), + status_extractor=lambda r: r.status, + response_model=BriaRemoveVideoBackgroundResponse, + ) + return IO.NodeOutput(await download_url_to_video_output(response.result.video_url)) + + +class BriaVideoReplaceBackground(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="BriaVideoReplaceBackground", + display_name="Bria Video Replace Background", + category="partner/video/Bria", + description="Replace a video's background with a supplied image or video using Bria. " + "The output keeps the foreground's resolution and frame rate; a background with a " + "different aspect ratio is stretched to fit, so match it for undistorted results.", + inputs=[ + IO.Video.Input("video", tooltip="Foreground video whose background is replaced."), + IO.Image.Input( + "background_image", + optional=True, + tooltip="Background image to composite behind the foreground. " + "Provide either a background image or a background video, not both.", + ), + IO.Video.Input( + "background_video", + optional=True, + tooltip="Background video to composite behind the foreground. " + "Provide either a background image or a background video, not both.", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed controls whether the node should re-run; " + "results are non-deterministic regardless of seed.", + ), + ], + outputs=[IO.Video.Output()], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.14,"format":{"suffix":"/second"}}""", + ), + ) + + @classmethod + async def execute( + cls, + video: Input.Video, + seed: int, + background_image: Input.Image | None = None, + background_video: Input.Video | None = None, + ) -> IO.NodeOutput: + if (background_image is None) == (background_video is None): + raise ValueError("Provide either a background image or a background video, not both.") + validate_video_duration(video, max_duration=60.0) + if background_video is not None: + validate_video_duration(background_video, max_duration=60.0) + background_url = await upload_video_to_comfyapi(cls, background_video, wait_label="Uploading background") + else: + background_url = await upload_image_to_comfyapi(cls, background_image, wait_label="Uploading background") + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/bria/v2/video/edit/replace_background", method="POST"), + data=BriaVideoReplaceBackgroundRequest( + video=await upload_video_to_comfyapi(cls, video), + background_url=background_url, + output_container_and_codec="mp4_h264", + seed=seed, + ), + response_model=BriaStatusResponse, + ) + response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/bria/v2/status/{response.request_id}"), + status_extractor=lambda r: r.status, + response_model=BriaRemoveVideoBackgroundResponse, + ) + return IO.NodeOutput(await download_url_to_video_output(response.result.video_url)) + + +def _video_to_images_and_mask(video: Input.Video) -> tuple[Input.Image, Input.Mask]: + """Decode a transparent webm (VP9 + alpha) into image frames and an alpha mask. + + VP9 keeps its alpha in a side layer that PyAV's default vp9 decoder drops, so the frames + are decoded with libvpx-vp9. Returns RGB images [B,H,W,3] in 0..1 and a mask [B,H,W] + following the Load Image convention (1 = transparent) for compositing or Save WEBM. + """ + rgb_frames: list[torch.Tensor] = [] + alpha_frames: list[torch.Tensor] = [] + with av.open(video.get_stream_source(), mode="r") as container: + stream = container.streams.video[0] + decoder = CodecContext.create("libvpx-vp9", "r") if stream.codec_context.name == "vp9" else None + for packet in container.demux(stream): + for frame in (decoder.decode(packet) if decoder is not None else packet.decode()): + rgba = torch.from_numpy(frame.to_ndarray(format="rgba")).float() / 255.0 + rgb_frames.append(rgba[..., :3]) + alpha_frames.append(rgba[..., 3]) + images = torch.stack(rgb_frames) if rgb_frames else torch.zeros(0, 0, 0, 3) + mask = (1.0 - torch.stack(alpha_frames)) if alpha_frames else torch.zeros((images.shape[0], 64, 64)) + return images, mask + + +class BriaTransparentVideoBackground(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="BriaTransparentVideoBackground", + display_name="Bria Remove Video Background (Transparent)", + category="partner/video/Bria", + description="Remove the background from a video using Bria and return the cut-out frames " + "plus an alpha mask. Connect both to a compositing node, or feed them to Save WEBM to " + "write a transparent video.", + inputs=[ + IO.Video.Input("video"), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed controls whether the node should re-run; " + "results are non-deterministic regardless of seed.", + ), + ], + outputs=[ + IO.Image.Output(display_name="images"), + IO.Mask.Output(display_name="mask"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.14,"format":{"suffix":"/second"}}""", + ), + ) + + @classmethod + async def execute( + cls, + video: Input.Video, + seed: int, + ) -> IO.NodeOutput: + validate_video_duration(video, max_duration=60.0) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/bria/v2/video/edit/remove_background", method="POST"), + data=BriaRemoveVideoBackgroundRequest( + video=await upload_video_to_comfyapi(cls, video), + background_color="Transparent", + output_container_and_codec="webm_vp9", + seed=seed, + ), + response_model=BriaStatusResponse, + ) + response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/bria/v2/status/{response.request_id}"), + status_extractor=lambda r: r.status, + response_model=BriaRemoveVideoBackgroundResponse, + ) + video_out = await download_url_to_video_output(response.result.video_url) + images, mask = _video_to_images_and_mask(video_out) + return IO.NodeOutput(images, mask) + + class BriaExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: @@ -323,6 +570,9 @@ class BriaExtension(ComfyExtension): BriaImageEditNode, BriaRemoveImageBackground, BriaRemoveVideoBackground, + BriaVideoGreenScreen, + # BriaVideoReplaceBackground, # server returns Status 500 when we pass background video + BriaTransparentVideoBackground, ] diff --git a/comfy_api_nodes/nodes_bytedance.py b/comfy_api_nodes/nodes_bytedance.py index d6b479336..c30ddc446 100644 --- a/comfy_api_nodes/nodes_bytedance.py +++ b/comfy_api_nodes/nodes_bytedance.py @@ -2,11 +2,13 @@ import hashlib import logging import math import re +from io import BytesIO import torch from typing_extensions import override -from comfy_api.latest import IO, ComfyExtension, Input +from comfy.utils import common_upscale +from comfy_api.latest import IO, ComfyExtension, Input, Types from comfy_api_nodes.apis.bytedance import ( RECOMMENDED_PRESETS, RECOMMENDED_PRESETS_SEEDREAM_4, @@ -43,15 +45,17 @@ from comfy_api_nodes.util import ( ApiEndpoint, download_url_to_image_tensor, download_url_to_video_output, + downscale_image_tensor_by_max_side, + downscale_video_to_max_pixels, get_number_of_images, image_tensor_pair_to_batch, poll_op, - resize_video_to_pixel_budget, sync_op, upload_audio_to_comfyapi, upload_image_to_comfyapi, upload_images_to_comfyapi, upload_video_to_comfyapi, + upscale_video_to_min_pixels, validate_image_aspect_ratio, validate_image_dimensions, validate_string, @@ -110,15 +114,62 @@ def _validate_ref_video_pixels(video: Input.Video, model_id: str, resolution: st max_px = limits.get("max") if min_px and pixels < min_px: raise ValueError( - f"Reference video {index} is too small: {w}x{h} = {pixels:,}px. " f"Minimum is {min_px:,}px for this model." + f"Reference video {index} is too small: {w}x{h} = {pixels:,} total pixels. " + f"Minimum for this model is {min_px:,} total pixels." ) if max_px and pixels > max_px: raise ValueError( - f"Reference video {index} is too large: {w}x{h} = {pixels:,}px. " - f"Maximum is {max_px:,}px for this model. Try downscaling the video." + f"Reference video {index} is too large: {w}x{h} = {pixels:,} total pixels. " + f"Maximum for this model is {max_px:,} total pixels. Try downscaling the video." ) +def _prepare_seedance_image(image: Input.Image) -> Input.Image: + """Auto-downscale a Seedance image input to the per-side limits, then validate it.""" + validate_image_aspect_ratio(image, (2, 5), (5, 2), strict=False) # 0.4 to 2.5 + image = downscale_image_tensor_by_max_side(image, max_side=6000) + validate_image_dimensions(image, min_width=300, min_height=300, max_width=6000, max_height=6000) + return image + + +# Supported output aspect ratios, used to pre-size FLF frames to matching pixel pair to avoid the 1080p stretch jump. +SEEDANCE2_RATIO_WH = { + "16:9": (16, 9), + "4:3": (4, 3), + "1:1": (1, 1), + "3:4": (3, 4), + "9:16": (9, 16), + "21:9": (21, 9), +} +SEEDANCE2_RES_SHORT_SIDE = {"480p": 480, "720p": 720, "1080p": 1080} + + +def _seedance2_target_dims(resolution: str, ratio: str, image: torch.Tensor) -> tuple[int, int]: + """Exact supported output (width, height) for (resolution, ratio). + + The shorter side equals the resolution number (e.g. 1080p 16:9 -> 1920x1080). For ratio + "adaptive" (or any unexpected value) the ratio is derived from the image's own aspect, snapped + to the nearest supported ratio, so the output keeps the frame's orientation. + """ + short = SEEDANCE2_RES_SHORT_SIDE[resolution] + if ratio not in SEEDANCE2_RATIO_WH: + aspect = image.shape[-2] / image.shape[-3] # W / H; tensor is (B, H, W, C) + ratio = min(SEEDANCE2_RATIO_WH, key=lambda k: abs(SEEDANCE2_RATIO_WH[k][0] / SEEDANCE2_RATIO_WH[k][1] - aspect)) + rw, rh = SEEDANCE2_RATIO_WH[ratio] + if rw >= rh: # landscape or square: shorter side is the height + out_w, out_h = round(short * rw / rh), short + else: # portrait: shorter side is the width + out_w, out_h = short, round(short * rh / rw) + return out_w - out_w % 2, out_h - out_h % 2 + + +def _resize_to_exact(image: torch.Tensor, width: int, height: int) -> torch.Tensor: + """Center-crop to the target aspect and resize to exactly width x height (lanczos).""" + samples = image.movedim(-1, 1) # (B, H, W, C) -> (B, C, H, W) + resized = common_upscale(samples, width, height, "lanczos", "center") + return resized.movedim(1, -1) + + async def _resolve_reference_assets( cls: type[IO.ComfyNode], asset_ids: list[str], @@ -306,6 +357,26 @@ async def _seedance_virtual_library_upload_image_asset( return f"asset://{create_resp.asset_id}" +async def _seedance_virtual_library_upload_video_asset( + cls: type[IO.ComfyNode], + video: Input.Video, + *, + wait_label: str = "Uploading video", +) -> str: + buf = BytesIO() + video.save_to(buf, format=Types.VideoContainer.MP4, codec=Types.VideoCodec.H264) + video_hash = hashlib.sha256(buf.getbuffer()).hexdigest() + public_url = await upload_video_to_comfyapi(cls, video, wait_label=wait_label) + create_resp = await sync_op( + cls, + ApiEndpoint(path="/proxy/seedance/virtual-library/assets", method="POST"), + response_model=SeedanceCreateAssetResponse, + data=SeedanceVirtualLibraryCreateAssetRequest(url=public_url, hash=video_hash, asset_type="Video"), + ) + await _wait_for_asset_active(cls, create_resp.asset_id, group_id="virtual-library") + return f"asset://{create_resp.asset_id}" + + def _seedance2_price_extractor(model_id: str, has_video_input: bool): """Returns a price_extractor closure for Seedance 2.0 poll_op.""" rate = SEEDANCE2_PRICE_PER_1K_TOKENS.get((model_id, has_video_input)) @@ -336,7 +407,7 @@ class ByteDanceImageNode(IO.ComfyNode): return IO.Schema( node_id="ByteDanceImageNode", display_name="ByteDance Image", - category="api node/image/ByteDance", + category="partner/image/ByteDance", description="Generate images using ByteDance models via api based on prompt", inputs=[ IO.Combo.Input("model", options=["seedream-3-0-t2i-250415"]), @@ -460,7 +531,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode): return IO.Schema( node_id="ByteDanceSeedreamNode", display_name="ByteDance Seedream 4.5 & 5.0", - category="api node/image/ByteDance", + category="partner/image/ByteDance", description="Unified text-to-image generation and precise single-sentence editing at up to 4K resolution.", inputs=[ IO.Combo.Input( @@ -722,7 +793,7 @@ class ByteDanceSeedreamNodeV2(IO.ComfyNode): return IO.Schema( node_id="ByteDanceSeedreamNodeV2", display_name="ByteDance Seedream 4.5 & 5.0", - category="api node/image/ByteDance", + category="partner/image/ByteDance", description="Unified text-to-image generation and precise single-sentence editing at up to 4K resolution.", inputs=[ IO.String.Input( @@ -888,7 +959,7 @@ class ByteDanceTextToVideoNode(IO.ComfyNode): return IO.Schema( node_id="ByteDanceTextToVideoNode", display_name="ByteDance Text to Video", - category="api node/video/ByteDance", + category="partner/video/ByteDance", description="Generate video using ByteDance models via api based on prompt", inputs=[ IO.Combo.Input( @@ -1016,7 +1087,7 @@ class ByteDanceImageToVideoNode(IO.ComfyNode): return IO.Schema( node_id="ByteDanceImageToVideoNode", display_name="ByteDance Image to Video", - category="api node/video/ByteDance", + category="partner/video/ByteDance", description="Generate video using ByteDance models via api based on image and prompt", inputs=[ IO.Combo.Input( @@ -1153,7 +1224,7 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode): return IO.Schema( node_id="ByteDanceFirstLastFrameNode", display_name="ByteDance First-Last-Frame to Video", - category="api node/video/ByteDance", + category="partner/video/ByteDance", description="Generate video using prompt and first and last frames.", inputs=[ IO.Combo.Input( @@ -1301,7 +1372,7 @@ class ByteDanceImageReferenceNode(IO.ComfyNode): return IO.Schema( node_id="ByteDanceImageReferenceNode", display_name="ByteDance Reference Images to Video", - category="api node/video/ByteDance", + category="partner/video/ByteDance", description="Generate video using prompt and reference images.", inputs=[ IO.Combo.Input( @@ -1544,7 +1615,7 @@ class ByteDance2TextToVideoNode(IO.ComfyNode): return IO.Schema( node_id="ByteDance2TextToVideoNode", display_name="ByteDance Seedance 2.0 Text to Video", - category="api node/video/ByteDance", + category="partner/video/ByteDance", description="Generate video using Seedance 2.0 models based on a text prompt.", inputs=[ IO.DynamicCombo.Input( @@ -1645,7 +1716,7 @@ class ByteDance2FirstLastFrameNode(IO.ComfyNode): return IO.Schema( node_id="ByteDance2FirstLastFrameNode", display_name="ByteDance Seedance 2.0 First-Last-Frame to Video", - category="api node/video/ByteDance", + category="partner/video/ByteDance", description="Generate video using Seedance 2.0 from a first frame image and optional last frame image.", inputs=[ IO.DynamicCombo.Input( @@ -1676,14 +1747,14 @@ class ByteDance2FirstLastFrameNode(IO.ComfyNode): "first_frame_asset_id", default="", tooltip="Seedance asset_id to use as the first frame. " - "Mutually exclusive with the first_frame image input.", + "Mutually exclusive with the first_frame image input.", optional=True, ), IO.String.Input( "last_frame_asset_id", default="", tooltip="Seedance asset_id to use as the last frame. " - "Mutually exclusive with the last_frame image input.", + "Mutually exclusive with the last_frame image input.", optional=True, ), IO.Int.Input( @@ -1758,6 +1829,29 @@ class ByteDance2FirstLastFrameNode(IO.ComfyNode): if last_frame is not None and last_frame_asset_id: raise ValueError("Provide only one of last_frame or last_frame_asset_id, not both.") + request_ratio = model["ratio"] + if first_frame_asset_id or last_frame_asset_id: + if first_frame is not None: + first_frame = _prepare_seedance_image(first_frame) + if last_frame is not None: + last_frame = _prepare_seedance_image(last_frame) + else: + # The 1080p FLF stretch fix (pre-size frames to a supported pixel pair + submit ratio="adaptive") + # only applies to local image inputs we can resize. + request_ratio = "adaptive" + target_dims: tuple[int, int] | None = None + if first_frame is not None: + validate_image_aspect_ratio(first_frame, (2, 5), (5, 2), strict=False) # 0.4 to 2.5 + validate_image_dimensions(first_frame, min_width=300, min_height=300) + target_dims = _seedance2_target_dims(model["resolution"], model["ratio"], first_frame) + first_frame = _resize_to_exact(first_frame, *target_dims) + if last_frame is not None: + validate_image_aspect_ratio(last_frame, (2, 5), (5, 2), strict=False) # 0.4 to 2.5 + validate_image_dimensions(last_frame, min_width=300, min_height=300) + if target_dims is None: + target_dims = _seedance2_target_dims(model["resolution"], model["ratio"], last_frame) + last_frame = _resize_to_exact(last_frame, *target_dims) + asset_ids_to_resolve = [a for a in (first_frame_asset_id, last_frame_asset_id) if a] image_assets: dict[str, str] = {} if asset_ids_to_resolve: @@ -1807,7 +1901,7 @@ class ByteDance2FirstLastFrameNode(IO.ComfyNode): content=content, generate_audio=model["generate_audio"], resolution=model["resolution"], - ratio=model["ratio"], + ratio=request_ratio, duration=model["duration"], seed=seed, watermark=watermark, @@ -1864,12 +1958,21 @@ def _seedance2_reference_inputs(resolutions: list[str], default_ratio: str = "16 ), IO.Boolean.Input( "auto_downscale", - default=False, - advanced=True, + default=True, optional=True, tooltip="Automatically downscale reference videos that exceed the model's pixel budget " "for the selected resolution. Aspect ratio is preserved; videos already within limits are untouched.", ), + IO.Boolean.Input( + "auto_upscale", + default=False, + advanced=True, + optional=True, + tooltip="Automatically upscale reference videos that are below the model's minimum pixel count " + "for the selected resolution. Aspect ratio is preserved; videos already meeting the minimum are " + "untouched. Note: upscaling a low-resolution source does not add real detail and may produce " + "lower-quality generations.", + ), IO.Autogrow.Input( "reference_assets", template=IO.Autogrow.TemplateNames( @@ -1898,7 +2001,7 @@ class ByteDance2ReferenceNode(IO.ComfyNode): return IO.Schema( node_id="ByteDance2ReferenceNode", display_name="ByteDance Seedance 2.0 Reference to Video", - category="api node/video/ByteDance", + category="partner/video/ByteDance", description="Generate, edit, or extend video using Seedance 2.0 with reference images, " "videos, and audio. Supports multimodal reference, video editing, and video extension.", inputs=[ @@ -2023,6 +2126,9 @@ class ByteDance2ReferenceNode(IO.ComfyNode): f"(audios={len(reference_audios)}, audio assets={len(reference_audio_assets)}). Maximum is 3." ) + for key in reference_images: + reference_images[key] = _prepare_seedance_image(reference_images[key]) + model_id = SEEDANCE_MODELS[model["model"]] has_video_input = total_videos > 0 @@ -2030,7 +2136,13 @@ class ByteDance2ReferenceNode(IO.ComfyNode): max_px = SEEDANCE2_REF_VIDEO_PIXEL_LIMITS.get(model_id, {}).get(model["resolution"], {}).get("max") if max_px: for key in reference_videos: - reference_videos[key] = resize_video_to_pixel_budget(reference_videos[key], max_px) + reference_videos[key] = downscale_video_to_max_pixels(reference_videos[key], max_px) + + if model.get("auto_upscale") and reference_videos: + min_px = SEEDANCE2_REF_VIDEO_PIXEL_LIMITS.get(model_id, {}).get(model["resolution"], {}).get("min") + if min_px: + for key in reference_videos: + reference_videos[key] = upscale_video_to_min_pixels(reference_videos[key], min_px) total_video_duration = 0.0 for i, key in enumerate(reference_videos, 1): @@ -2089,7 +2201,7 @@ class ByteDance2ReferenceNode(IO.ComfyNode): content.append( TaskVideoContent( video_url=TaskVideoContentUrl( - url=await upload_video_to_comfyapi( + url=await _seedance_virtual_library_upload_video_asset( cls, reference_videos[key], wait_label=f"Uploading video {i}", @@ -2186,7 +2298,7 @@ class ByteDanceCreateImageAsset(IO.ComfyNode): return IO.Schema( node_id="ByteDanceCreateImageAsset", display_name="ByteDance Create Image Asset", - category="api node/image/ByteDance", + category="partner/image/ByteDance", description=( "Create a Seedance 2.0 personal image asset. Uploads the input image and " "registers it in the given asset group. If group_id is empty, runs a real-person " @@ -2253,7 +2365,7 @@ class ByteDanceCreateVideoAsset(IO.ComfyNode): return IO.Schema( node_id="ByteDanceCreateVideoAsset", display_name="ByteDance Create Video Asset", - category="api node/video/ByteDance", + category="partner/video/ByteDance", description=( "Create a Seedance 2.0 personal video asset. Uploads the input video and " "registers it in the given asset group. If group_id is empty, runs a real-person " diff --git a/comfy_api_nodes/nodes_bytedance_llm.py b/comfy_api_nodes/nodes_bytedance_llm.py index fa7fe370a..cb41defa0 100644 --- a/comfy_api_nodes/nodes_bytedance_llm.py +++ b/comfy_api_nodes/nodes_bytedance_llm.py @@ -144,7 +144,7 @@ class ByteDanceSeedNode(IO.ComfyNode): return IO.Schema( node_id="ByteDanceSeedNode", display_name="ByteDance Seed", - category="api node/text/ByteDance", + category="partner/text/ByteDance", essentials_category="Text Generation", description="Generate text responses with ByteDance's Seed 2.0 models. " "Provide a text prompt and optionally one or more images or videos for multimodal context.", diff --git a/comfy_api_nodes/nodes_elevenlabs.py b/comfy_api_nodes/nodes_elevenlabs.py index e452daf77..eba578a45 100644 --- a/comfy_api_nodes/nodes_elevenlabs.py +++ b/comfy_api_nodes/nodes_elevenlabs.py @@ -69,7 +69,7 @@ class ElevenLabsSpeechToText(IO.ComfyNode): return IO.Schema( node_id="ElevenLabsSpeechToText", display_name="ElevenLabs Speech to Text", - category="api node/audio/ElevenLabs", + category="partner/audio/ElevenLabs", description="Transcribe audio to text. " "Supports automatic language detection, speaker diarization, and audio event tagging.", inputs=[ @@ -210,7 +210,7 @@ class ElevenLabsVoiceSelector(IO.ComfyNode): return IO.Schema( node_id="ElevenLabsVoiceSelector", display_name="ElevenLabs Voice Selector", - category="api node/audio/ElevenLabs", + category="partner/audio/ElevenLabs", description="Select a predefined ElevenLabs voice for text-to-speech generation.", inputs=[ IO.Combo.Input( @@ -239,7 +239,7 @@ class ElevenLabsTextToSpeech(IO.ComfyNode): return IO.Schema( node_id="ElevenLabsTextToSpeech", display_name="ElevenLabs Text to Speech", - category="api node/audio/ElevenLabs", + category="partner/audio/ElevenLabs", description="Convert text to speech.", inputs=[ IO.Custom(ELEVENLABS_VOICE).Input( @@ -414,7 +414,7 @@ class ElevenLabsAudioIsolation(IO.ComfyNode): return IO.Schema( node_id="ElevenLabsAudioIsolation", display_name="ElevenLabs Voice Isolation", - category="api node/audio/ElevenLabs", + category="partner/audio/ElevenLabs", description="Remove background noise from audio, isolating vocals or speech.", inputs=[ IO.Audio.Input( @@ -459,7 +459,7 @@ class ElevenLabsTextToSoundEffects(IO.ComfyNode): return IO.Schema( node_id="ElevenLabsTextToSoundEffects", display_name="ElevenLabs Text to Sound Effects", - category="api node/audio/ElevenLabs", + category="partner/audio/ElevenLabs", description="Generate sound effects from text descriptions.", inputs=[ IO.String.Input( @@ -555,7 +555,7 @@ class ElevenLabsInstantVoiceClone(IO.ComfyNode): return IO.Schema( node_id="ElevenLabsInstantVoiceClone", display_name="ElevenLabs Instant Voice Clone", - category="api node/audio/ElevenLabs", + category="partner/audio/ElevenLabs", description="Create a cloned voice from audio samples. " "Provide 1-8 audio recordings of the voice to clone.", inputs=[ @@ -658,7 +658,7 @@ class ElevenLabsSpeechToSpeech(IO.ComfyNode): return IO.Schema( node_id="ElevenLabsSpeechToSpeech", display_name="ElevenLabs Speech to Speech", - category="api node/audio/ElevenLabs", + category="partner/audio/ElevenLabs", description="Transform speech from one voice to another while preserving the original content and emotion.", inputs=[ IO.Custom(ELEVENLABS_VOICE).Input( @@ -793,7 +793,7 @@ class ElevenLabsTextToDialogue(IO.ComfyNode): return IO.Schema( node_id="ElevenLabsTextToDialogue", display_name="ElevenLabs Text to Dialogue", - category="api node/audio/ElevenLabs", + category="partner/audio/ElevenLabs", description="Generate multi-speaker dialogue from text. Each dialogue entry has its own text and voice.", inputs=[ IO.Float.Input( diff --git a/comfy_api_nodes/nodes_gemini.py b/comfy_api_nodes/nodes_gemini.py index d18c958a8..3d4be6065 100644 --- a/comfy_api_nodes/nodes_gemini.py +++ b/comfy_api_nodes/nodes_gemini.py @@ -8,7 +8,7 @@ import os from enum import Enum from fnmatch import fnmatch from io import BytesIO -from typing import Literal +from typing import Any, Literal import torch from typing_extensions import override @@ -19,6 +19,7 @@ from comfy_api_nodes.apis.gemini import ( GeminiContent, GeminiFileData, GeminiGenerateContentRequest, + GeminiGenerationConfig, GeminiGenerateContentResponse, GeminiImageConfig, GeminiImageGenerateContentRequest, @@ -40,13 +41,18 @@ from comfy_api_nodes.util import ( get_number_of_images, sync_op, tensor_to_base64_string, + upload_audio_to_comfyapi, + upload_image_to_comfyapi, upload_images_to_comfyapi, + upload_video_to_comfyapi, validate_string, video_to_base64_string, ) GEMINI_BASE_ENDPOINT = "/proxy/vertexai/gemini" GEMINI_MAX_INPUT_FILE_SIZE = 20 * 1024 * 1024 # 20 MB +GEMINI_URL_INPUT_BUDGET = 10 +GEMINI_MAX_INLINE_BYTES = 18 * 1024 * 1024 GEMINI_IMAGE_SYS_PROMPT = ( "You are an expert image-generation engine. You must ALWAYS produce an image.\n" "Interpret all user input—regardless of " @@ -285,6 +291,140 @@ def calculate_tokens_price(response: GeminiGenerateContentResponse) -> float | N return final_price / 1_000_000.0 +def create_video_parts(video_input: Input.Video) -> list[GeminiPart]: + """Convert a single video input to Gemini API compatible parts (inline MP4/H.264).""" + base_64_string = video_to_base64_string( + video_input, container_format=Types.VideoContainer.MP4, codec=Types.VideoCodec.H264 + ) + return [ + GeminiPart( + inlineData=GeminiInlineData( + mimeType=GeminiMimeType.video_mp4, + data=base_64_string, + ) + ) + ] + + +def create_audio_parts(audio_input: Input.Audio) -> list[GeminiPart]: + """Convert an audio input to Gemini API compatible parts (one inline MP3 part per batch item).""" + audio_parts: list[GeminiPart] = [] + for batch_index in range(audio_input["waveform"].shape[0]): + # Recreate an IO.AUDIO object for the given batch dimension index + audio_at_index = Input.Audio( + waveform=audio_input["waveform"][batch_index].unsqueeze(0), + sample_rate=audio_input["sample_rate"], + ) + # Convert to MP3 format for compatibility with Gemini API + audio_bytes = audio_to_base64_string( + audio_at_index, + container_format="mp3", + codec_name="libmp3lame", + ) + audio_parts.append( + GeminiPart( + inlineData=GeminiInlineData( + mimeType=GeminiMimeType.audio_mp3, + data=audio_bytes, + ) + ) + ) + return audio_parts + + +def _flatten_images(images: list[Input.Image]) -> list[torch.Tensor]: + """Expand any batched image tensors into individual (H, W, C) frames, preserving order.""" + frames: list[torch.Tensor] = [] + for img in images: + if len(img.shape) == 4: + frames.extend(img[i] for i in range(img.shape[0])) + else: + frames.append(img) + return frames + + +def _flatten_audio(audios: list[Input.Audio]) -> list[Input.Audio]: + """Expand any batched audio inputs into individual single-clip audio inputs, preserving order.""" + clips: list[Input.Audio] = [] + for audio in audios: + waveform = audio["waveform"] + for i in range(waveform.shape[0]): + clips.append(Input.Audio(waveform=waveform[i].unsqueeze(0), sample_rate=audio["sample_rate"])) + return clips + + +async def _media_url_part(cls: type[IO.ComfyNode], kind: str, payload: Any) -> GeminiPart: + """Upload a single media unit to ComfyAPI storage and return a fileData (URL) part.""" + if kind == "image": + url = await upload_image_to_comfyapi(cls, payload, mime_type="image/png", wait_label="Uploading image") + return GeminiPart(fileData=GeminiFileData(mimeType=GeminiMimeType.image_png, fileUri=url)) + if kind == "audio": + url = await upload_audio_to_comfyapi( + cls, payload, container_format="mp3", codec_name="libmp3lame", mime_type="audio/mp3" + ) + return GeminiPart(fileData=GeminiFileData(mimeType=GeminiMimeType.audio_mp3, fileUri=url)) + url = await upload_video_to_comfyapi(cls, payload, wait_label="Uploading video") + return GeminiPart(fileData=GeminiFileData(mimeType=GeminiMimeType.video_mp4, fileUri=url)) + + +def _media_inline_part(kind: str, payload: Any) -> tuple[GeminiPart, int]: + """Encode a single media unit as an inline base64 part; returns (part, base64_length).""" + if kind == "image": + data = tensor_to_base64_string(payload, mime_type="image/webp") + mime = GeminiMimeType.image_webp + elif kind == "audio": + data = audio_to_base64_string(payload, container_format="mp3", codec_name="libmp3lame") + mime = GeminiMimeType.audio_mp3 + else: + data = video_to_base64_string( + payload, container_format=Types.VideoContainer.MP4, codec=Types.VideoCodec.H264 + ) + mime = GeminiMimeType.video_mp4 + return GeminiPart(inlineData=GeminiInlineData(mimeType=mime, data=data)), len(data) + + +async def build_gemini_media_parts( + cls: type[IO.ComfyNode], + images: list[Input.Image], + audios: list[Input.Audio], + videos: list[Input.Video], + *, + url_budget: int = GEMINI_URL_INPUT_BUDGET, + max_inline_bytes: int = GEMINI_MAX_INLINE_BYTES, +) -> list[GeminiPart]: + """Build Gemini parts for multimodal inputs (images, audio, video). + + fileData URLs are preferred for every media type: the upload is fetched directly by the + model, keeping the request body tiny regardless of media size. The URL budget is shared + across all media and assigned largest-first (video, then audio, then images), so that if it + is ever exhausted the inline-base64 overflow is limited to the smallest items. Total inline + payload is capped by `max_inline_bytes`. + """ + units: list[tuple[str, Any]] = ( + [("video", v) for v in videos] + + [("audio", a) for a in _flatten_audio(audios)] + + [("image", f) for f in _flatten_images(images)] + ) + + parts: list[GeminiPart] = [] + url_used = 0 + inline_bytes = 0 + for kind, payload in units: + if url_used < url_budget: + parts.append(await _media_url_part(cls, kind, payload)) + url_used += 1 + continue + part, nbytes = _media_inline_part(kind, payload) + inline_bytes += nbytes + if inline_bytes > max_inline_bytes: + raise ValueError( + f"Too much media to send inline (over {max_inline_bytes // (1024 * 1024)}MB after the first " + f"{url_budget} inputs are uploaded as URLs). Reduce the number or size of attached media." + ) + parts.append(part) + return parts + + class GeminiNode(IO.ComfyNode): """ Node to generate text responses from a Gemini model. @@ -300,7 +440,7 @@ class GeminiNode(IO.ComfyNode): return IO.Schema( node_id="GeminiNode", display_name="Google Gemini", - category="api node/text/Gemini", + category="partner/text/Gemini", description="Generate text responses with Google's Gemini AI model. " "You can provide multiple types of inputs (text, images, audio, video) " "as context for generating more relevant and meaningful responses.", @@ -407,58 +547,9 @@ class GeminiNode(IO.ComfyNode): ) """, ), + is_deprecated=True, ) - @classmethod - def create_video_parts(cls, video_input: Input.Video) -> list[GeminiPart]: - """Convert video input to Gemini API compatible parts.""" - - base_64_string = video_to_base64_string( - video_input, container_format=Types.VideoContainer.MP4, codec=Types.VideoCodec.H264 - ) - return [ - GeminiPart( - inlineData=GeminiInlineData( - mimeType=GeminiMimeType.video_mp4, - data=base_64_string, - ) - ) - ] - - @classmethod - def create_audio_parts(cls, audio_input: Input.Audio) -> list[GeminiPart]: - """ - Convert audio input to Gemini API compatible parts. - - Args: - audio_input: Audio input from ComfyUI, containing waveform tensor and sample rate. - - Returns: - List of GeminiPart objects containing the encoded audio. - """ - audio_parts: list[GeminiPart] = [] - for batch_index in range(audio_input["waveform"].shape[0]): - # Recreate an IO.AUDIO object for the given batch dimension index - audio_at_index = Input.Audio( - waveform=audio_input["waveform"][batch_index].unsqueeze(0), - sample_rate=audio_input["sample_rate"], - ) - # Convert to MP3 format for compatibility with Gemini API - audio_bytes = audio_to_base64_string( - audio_at_index, - container_format="mp3", - codec_name="libmp3lame", - ) - audio_parts.append( - GeminiPart( - inlineData=GeminiInlineData( - mimeType=GeminiMimeType.audio_mp3, - data=audio_bytes, - ) - ) - ) - return audio_parts - @classmethod async def execute( cls, @@ -482,9 +573,9 @@ class GeminiNode(IO.ComfyNode): if images is not None: parts.extend(await create_image_parts(cls, images)) if audio is not None: - parts.extend(cls.create_audio_parts(audio)) + parts.extend(create_audio_parts(audio)) if video is not None: - parts.extend(cls.create_video_parts(video)) + parts.extend(create_video_parts(video)) if files is not None: parts.extend(files) @@ -512,6 +603,210 @@ class GeminiNode(IO.ComfyNode): return IO.NodeOutput(output_text or "Empty response from Gemini model...") +GEMINI_V2_MODELS: dict[str, str] = { + "Gemini 3.1 Pro": "gemini-3.1-pro-preview", + "Gemini 3.1 Flash-Lite": "gemini-3.1-flash-lite-preview", +} + + +def _gemini_text_model_inputs(thinking_default: str) -> list[Input]: + """Per-model inputs revealed by the model DynamicCombo (shared media + sampling controls).""" + return [ + IO.Autogrow.Input( + "images", + template=IO.Autogrow.TemplateNames( + IO.Image.Input("image"), + names=[f"image_{i}" for i in range(1, 17)], + min=0, + ), + tooltip="Optional image(s) to use as context for the model. Up to 16 images.", + ), + IO.Autogrow.Input( + "audio", + template=IO.Autogrow.TemplateNames( + IO.Audio.Input("audio"), + names=["audio_1"], + min=0, + ), + tooltip="Optional audio clip to use as context for the model.", + ), + IO.Autogrow.Input( + "video", + template=IO.Autogrow.TemplateNames( + IO.Video.Input("video"), + names=["video_1"], + min=0, + ), + tooltip="Optional video clip to use as context for the model.", + ), + IO.Custom("GEMINI_INPUT_FILES").Input( + "files", + optional=True, + tooltip="Optional file(s) to use as context for the model. " + "Accepts inputs from the Gemini Input Files node.", + ), + IO.Combo.Input( + "thinking_level", + options=["LOW", "HIGH"], + default=thinking_default, + tooltip="How hard the model reasons internally before answering. " + "HIGH improves quality on difficult tasks but costs more (thinking) tokens and is slower.", + ), + IO.Float.Input( + "temperature", + default=1.0, + min=0.0, + max=2.0, + step=0.01, + tooltip="Controls randomness. Lower is more focused/deterministic, higher is more creative.", + advanced=True, + ), + IO.Float.Input( + "top_p", + default=0.95, + min=0.0, + max=1.0, + step=0.01, + tooltip="Nucleus sampling: sample from the smallest token set whose cumulative probability reaches top_p.", + advanced=True, + ), + IO.Int.Input( + "max_output_tokens", + default=32768, + min=16, + max=65536, + tooltip="Maximum tokens to generate, including the model's internal thinking. " + "With thinking_level HIGH, a low value can leave no room for the answer; raise this if " + "responses come back empty or truncated. The model stops early when finished, so a higher " + "cap costs nothing extra for short replies.", + advanced=True, + ), + ] + + +class GeminiNodeV2(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="GeminiNodeV2", + display_name="Google Gemini", + category="partner/text/Gemini", + essentials_category="Text Generation", + description="Generate text responses with Google's Gemini models. Provide a text prompt and, " + "optionally, one or more images, audio clips, videos, or files as multimodal context.", + inputs=[ + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Text input to the model. Include detailed instructions, questions, or context.", + ), + IO.DynamicCombo.Input( + "model", + options=[ + IO.DynamicCombo.Option("Gemini 3.1 Pro", _gemini_text_model_inputs("HIGH")), + IO.DynamicCombo.Option("Gemini 3.1 Flash-Lite", _gemini_text_model_inputs("LOW")), + ], + tooltip="The Gemini model used to generate the response.", + ), + IO.Int.Input( + "seed", + default=42, + min=0, + max=2147483647, + control_after_generate=True, + tooltip="Seed for sampling. Set to 0 for a random seed. Deterministic output isn't guaranteed.", + ), + IO.String.Input( + "system_prompt", + multiline=True, + default="", + optional=True, + advanced=True, + tooltip="Foundational instructions that dictate the model's behavior.", + ), + ], + outputs=[ + IO.String.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model"]), + expr=""" + ( + $m := widgets.model; + $contains($m, "lite") ? { + "type": "list_usd", + "usd": [0.00025, 0.0015], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } : { + "type": "list_usd", + "usd": [0.002, 0.012], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + prompt: str, + model: dict, + seed: int, + system_prompt: str = "", + ) -> IO.NodeOutput: + validate_string(prompt, strip_whitespace=True, min_length=1) + model_id = GEMINI_V2_MODELS[model["model"]] + + parts: list[GeminiPart] = [GeminiPart(text=prompt)] + images = [t for t in (model.get("images") or {}).values() if t is not None] + audios = [a for a in (model.get("audio") or {}).values() if a is not None] + videos = [v for v in (model.get("video") or {}).values() if v is not None] + if images or audios or videos: + parts.extend(await build_gemini_media_parts(cls, images, audios, videos)) + files = model.get("files") + if files is not None: + parts.extend(files) + + gemini_system_prompt = None + if system_prompt: + gemini_system_prompt = GeminiSystemInstructionContent(parts=[GeminiTextPart(text=system_prompt)], role=None) + + response = await sync_op( + cls, + endpoint=ApiEndpoint(path=f"{GEMINI_BASE_ENDPOINT}/{model_id}", method="POST"), + data=GeminiGenerateContentRequest( + contents=[ + GeminiContent( + role=GeminiRole.user, + parts=parts, + ) + ], + generationConfig=GeminiGenerationConfig( + temperature=model["temperature"], + topP=model["top_p"], + maxOutputTokens=model["max_output_tokens"], + seed=seed if seed > 0 else None, + thinkingConfig=GeminiThinkingConfig(thinkingLevel=model["thinking_level"]), + ), + systemInstruction=gemini_system_prompt, + ), + response_model=GeminiGenerateContentResponse, + price_extractor=calculate_tokens_price, + ) + + output_text = get_text_from_response(response) + return IO.NodeOutput(output_text or "Empty response from Gemini model...") + + class GeminiInputFiles(IO.ComfyNode): """ Loads and formats input files for use with the Gemini API. @@ -541,7 +836,7 @@ class GeminiInputFiles(IO.ComfyNode): return IO.Schema( node_id="GeminiInputFiles", display_name="Gemini Input Files", - category="api node/text/Gemini", + category="partner/text/Gemini", description="Loads and prepares input files to include as inputs for Gemini LLM nodes. " "The files will be read by the Gemini model when generating a response. " "The contents of the text file count toward the token limit. " @@ -598,7 +893,7 @@ class GeminiImage(IO.ComfyNode): return IO.Schema( node_id="GeminiImageNode", display_name="Nano Banana (Google Gemini Image)", - category="api node/image/Gemini", + category="partner/image/Gemini", description="Edit images synchronously via Google API.", inputs=[ IO.String.Input( @@ -731,7 +1026,7 @@ class GeminiImage2(IO.ComfyNode): return IO.Schema( node_id="GeminiImage2Node", display_name="Nano Banana Pro (Google Gemini Image)", - category="api node/image/Gemini", + category="partner/image/Gemini", description="Generate or edit images synchronously via Google Vertex API.", inputs=[ IO.String.Input( @@ -869,7 +1164,7 @@ class GeminiNanoBanana2(IO.ComfyNode): return IO.Schema( node_id="GeminiNanoBanana2", display_name="Nano Banana 2", - category="api node/image/Gemini", + category="partner/image/Gemini", description="Generate or edit images synchronously via Google Vertex API.", inputs=[ IO.String.Input( @@ -1085,7 +1380,7 @@ class GeminiNanoBanana2V2(IO.ComfyNode): return IO.Schema( node_id="GeminiNanoBanana2V2", display_name="Nano Banana 2", - category="api node/image/Gemini", + category="partner/image/Gemini", description="Generate or edit images synchronously via Google Vertex API.", inputs=[ IO.String.Input( @@ -1129,6 +1424,26 @@ class GeminiNanoBanana2V2(IO.ComfyNode): tooltip="Foundational instructions that dictate an AI's behavior.", advanced=True, ), + IO.Float.Input( + "temperature", + default=1.0, + min=0.0, + max=2.0, + step=0.01, + optional=True, + tooltip="Controls randomness in generation. Lower is more focused/deterministic.", + advanced=True, + ), + IO.Float.Input( + "top_p", + default=0.95, + min=0.0, + max=1.0, + step=0.01, + optional=True, + tooltip="Nucleus sampling threshold. Lower is more focused, higher more diverse.", + advanced=True, + ), ], outputs=[ IO.Image.Output(), @@ -1165,6 +1480,8 @@ class GeminiNanoBanana2V2(IO.ComfyNode): seed: int, response_modalities: str, system_prompt: str = "", + temperature: float = 1.0, + top_p: float = 0.95, ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=True, min_length=1) model_choice = model["model"] @@ -1204,6 +1521,8 @@ class GeminiNanoBanana2V2(IO.ComfyNode): responseModalities=(["IMAGE"] if response_modalities == "IMAGE" else ["TEXT", "IMAGE"]), imageConfig=image_config, thinkingConfig=GeminiThinkingConfig(thinkingLevel=model["thinking_level"]), + temperature=temperature, + topP=top_p, ), systemInstruction=gemini_system_prompt, ), @@ -1222,6 +1541,7 @@ class GeminiExtension(ComfyExtension): async def get_node_list(self) -> list[type[IO.ComfyNode]]: return [ GeminiNode, + GeminiNodeV2, GeminiImage, GeminiImage2, GeminiNanoBanana2, diff --git a/comfy_api_nodes/nodes_grok.py b/comfy_api_nodes/nodes_grok.py index a103f24ee..2ae529813 100644 --- a/comfy_api_nodes/nodes_grok.py +++ b/comfy_api_nodes/nodes_grok.py @@ -29,6 +29,11 @@ from comfy_api_nodes.util import ( ) +_GROK_VIDEO_MODEL_API_IDS = { + "grok-imagine-video-1.5": "grok-imagine-video-1.5-preview", +} + + def _extract_grok_price(response) -> float | None: if response.usage and response.usage.cost_in_usd_ticks is not None: return response.usage.cost_in_usd_ticks / 10_000_000_000 @@ -49,7 +54,7 @@ class GrokImageNode(IO.ComfyNode): return IO.Schema( node_id="GrokImageNode", display_name="Grok Image", - category="api node/image/Grok", + category="partner/image/Grok", description="Generate images using Grok based on a text prompt", inputs=[ IO.Combo.Input( @@ -58,7 +63,6 @@ class GrokImageNode(IO.ComfyNode): "grok-imagine-image-quality", "grok-imagine-image-pro", "grok-imagine-image", - "grok-imagine-image-beta", ], ), IO.String.Input( @@ -224,7 +228,7 @@ class GrokImageEditNode(IO.ComfyNode): return IO.Schema( node_id="GrokImageEditNode", display_name="Grok Image Edit", - category="api node/image/Grok", + category="partner/image/Grok", description="Modify an existing image based on a text prompt", inputs=[ IO.Combo.Input( @@ -233,7 +237,6 @@ class GrokImageEditNode(IO.ComfyNode): "grok-imagine-image-quality", "grok-imagine-image-pro", "grok-imagine-image", - "grok-imagine-image-beta", ], ), IO.Image.Input("image", display_name="images"), @@ -366,7 +369,7 @@ class GrokImageEditNodeV2(IO.ComfyNode): return IO.Schema( node_id="GrokImageEditNodeV2", display_name="Grok Image Edit", - category="api node/image/Grok", + category="partner/image/Grok", description="Modify an existing image based on a text prompt", inputs=[ IO.String.Input( @@ -503,10 +506,14 @@ class GrokVideoNode(IO.ComfyNode): return IO.Schema( node_id="GrokVideoNode", display_name="Grok Video", - category="api node/video/Grok", + category="partner/video/Grok", description="Generate video from a prompt or an image", inputs=[ - IO.Combo.Input("model", options=["grok-imagine-video", "grok-imagine-video-beta"]), + IO.Combo.Input( + "model", + options=["grok-imagine-video", "grok-imagine-video-1.5"], + tooltip="grok-imagine-video-1.5 currently always requires an input image.", + ), IO.String.Input( "prompt", multiline=True, @@ -542,7 +549,11 @@ class GrokVideoNode(IO.ComfyNode): tooltip="Seed to determine if node should re-run; " "actual results are nondeterministic regardless of seed.", ), - IO.Image.Input("image", optional=True), + IO.Image.Input( + "image", + optional=True, + tooltip="Optional starting image for grok-imagine-video. Required for grok-imagine-video-1.5.", + ), ], outputs=[ IO.Video.Output(), @@ -554,12 +565,16 @@ class GrokVideoNode(IO.ComfyNode): ], is_api_node=True, price_badge=IO.PriceBadge( - depends_on=IO.PriceBadgeDepends(widgets=["duration", "resolution"], inputs=["image"]), + depends_on=IO.PriceBadgeDepends(widgets=["model", "duration", "resolution"], inputs=["image"]), expr=""" ( - $rate := widgets.resolution = "720p" ? 0.07 : 0.05; + $is15 := $contains(widgets.model, "1.5"); + $rate := $is15 + ? (widgets.resolution = "720p" ? 0.2002 : 0.1144) + : (widgets.resolution = "720p" ? 0.07 : 0.05); + $imgCost := $is15 ? 0.0143 : 0.002; $base := $rate * widgets.duration; - {"type":"usd","usd": inputs.image.connected ? $base + 0.002 : $base} + {"type":"usd","usd": inputs.image.connected ? $base + $imgCost : $base} ) """, ), @@ -576,8 +591,8 @@ class GrokVideoNode(IO.ComfyNode): seed: int, image: Input.Image | None = None, ) -> IO.NodeOutput: - if model == "grok-imagine-video-beta": - model = "grok-imagine-video" + if image is None and model == "grok-imagine-video-1.5": + raise ValueError(f"The '{model}' model requires an input image; connect one to the 'image' input.") image_url = None if image is not None: if get_number_of_images(image) != 1: @@ -588,7 +603,7 @@ class GrokVideoNode(IO.ComfyNode): cls, ApiEndpoint(path="/proxy/xai/v1/videos/generations", method="POST"), data=VideoGenerationRequest( - model=model, + model=_GROK_VIDEO_MODEL_API_IDS.get(model, model), image=image_url, prompt=prompt, resolution=resolution, @@ -603,7 +618,7 @@ class GrokVideoNode(IO.ComfyNode): ApiEndpoint(path=f"/proxy/xai/v1/videos/{initial_response.request_id}"), status_extractor=lambda r: r.status if r.status is not None else "complete", response_model=VideoStatusResponse, - price_extractor=_extract_grok_price, + price_extractor=_extract_grok_video_price if model == "grok-imagine-video-1.5" else _extract_grok_price, ) return IO.NodeOutput(await download_url_to_video_output(response.video.url)) @@ -615,10 +630,10 @@ class GrokVideoEditNode(IO.ComfyNode): return IO.Schema( node_id="GrokVideoEditNode", display_name="Grok Video Edit", - category="api node/video/Grok", + category="partner/video/Grok", description="Edit an existing video based on a text prompt.", inputs=[ - IO.Combo.Input("model", options=["grok-imagine-video", "grok-imagine-video-beta"]), + IO.Combo.Input("model", options=["grok-imagine-video"]), IO.String.Input( "prompt", multiline=True, @@ -693,7 +708,7 @@ class GrokVideoReferenceNode(IO.ComfyNode): return IO.Schema( node_id="GrokVideoReferenceNode", display_name="Grok Reference-to-Video", - category="api node/video/Grok", + category="partner/video/Grok", description="Generate video guided by reference images as style and content references.", inputs=[ IO.String.Input( @@ -826,7 +841,7 @@ class GrokVideoExtendNode(IO.ComfyNode): return IO.Schema( node_id="GrokVideoExtendNode", display_name="Grok Video Extend", - category="api node/video/Grok", + category="partner/video/Grok", description="Extend an existing video with a seamless continuation based on a text prompt.", inputs=[ IO.String.Input( diff --git a/comfy_api_nodes/nodes_hitpaw.py b/comfy_api_nodes/nodes_hitpaw.py index bca5170e4..062d3cf1d 100644 --- a/comfy_api_nodes/nodes_hitpaw.py +++ b/comfy_api_nodes/nodes_hitpaw.py @@ -71,7 +71,7 @@ class HitPawGeneralImageEnhance(IO.ComfyNode): return IO.Schema( node_id="HitPawGeneralImageEnhance", display_name="HitPaw General Image Enhance", - category="api node/image/HitPaw", + category="partner/image/HitPaw", description="Upscale low-resolution images to super-resolution, eliminate artifacts and noise. " f"Maximum output: {MAX_MP_GENERATIVE} megapixels.", inputs=[ @@ -201,7 +201,7 @@ class HitPawVideoEnhance(IO.ComfyNode): return IO.Schema( node_id="HitPawVideoEnhance", display_name="HitPaw Video Enhance", - category="api node/video/HitPaw", + category="partner/video/HitPaw", description="Upscale low-resolution videos to high resolution, eliminate artifacts and noise. " "Prices shown are per second of video.", inputs=[ diff --git a/comfy_api_nodes/nodes_hunyuan3d.py b/comfy_api_nodes/nodes_hunyuan3d.py index 5fc31bccd..fcd27b7fb 100644 --- a/comfy_api_nodes/nodes_hunyuan3d.py +++ b/comfy_api_nodes/nodes_hunyuan3d.py @@ -123,7 +123,7 @@ class TencentTextToModelNode(IO.ComfyNode): return IO.Schema( node_id="TencentTextToModelNode", display_name="Hunyuan3D: Text to Model", - category="api node/3d/Tencent", + category="partner/3d/Tencent", essentials_category="3D", inputs=[ IO.Combo.Input( @@ -242,7 +242,7 @@ class TencentImageToModelNode(IO.ComfyNode): return IO.Schema( node_id="TencentImageToModelNode", display_name="Hunyuan3D: Image(s) to Model", - category="api node/3d/Tencent", + category="partner/3d/Tencent", essentials_category="3D", inputs=[ IO.Combo.Input( @@ -415,7 +415,7 @@ class TencentModelTo3DUVNode(IO.ComfyNode): return IO.Schema( node_id="TencentModelTo3DUVNode", display_name="Hunyuan3D: Model to UV", - category="api node/3d/Tencent", + category="partner/3d/Tencent", description="Perform UV unfolding on a 3D model to generate UV texture. " "Input model must have less than 30000 faces.", inputs=[ @@ -505,7 +505,7 @@ class Tencent3DTextureEditNode(IO.ComfyNode): return IO.Schema( node_id="Tencent3DTextureEditNode", display_name="Hunyuan3D: 3D Texture Edit", - category="api node/3d/Tencent", + category="partner/3d/Tencent", description="After inputting the 3D model, perform 3D model texture redrawing.", inputs=[ IO.MultiType.Input( @@ -594,7 +594,7 @@ class Tencent3DPartNode(IO.ComfyNode): return IO.Schema( node_id="Tencent3DPartNode", display_name="Hunyuan3D: 3D Part", - category="api node/3d/Tencent", + category="partner/3d/Tencent", description="Automatically perform component identification and generation based on the model structure.", inputs=[ IO.MultiType.Input( @@ -666,7 +666,7 @@ class TencentSmartTopologyNode(IO.ComfyNode): return IO.Schema( node_id="TencentSmartTopologyNode", display_name="Hunyuan3D: Smart Topology", - category="api node/3d/Tencent", + category="partner/3d/Tencent", description="Perform smart retopology on a 3D model. " "Supports GLB/OBJ formats; max 200MB; recommended for high-poly models.", inputs=[ diff --git a/comfy_api_nodes/nodes_ideogram.py b/comfy_api_nodes/nodes_ideogram.py index 97c3609bd..3b914a850 100644 --- a/comfy_api_nodes/nodes_ideogram.py +++ b/comfy_api_nodes/nodes_ideogram.py @@ -10,6 +10,7 @@ from comfy_api_nodes.apis.ideogram import ( ImageRequest, IdeogramV3Request, IdeogramV3EditRequest, + IdeogramV4Request, ) from comfy_api_nodes.util import ( ApiEndpoint, @@ -17,6 +18,7 @@ from comfy_api_nodes.util import ( download_url_as_bytesio, resize_mask_to_image, sync_op, + validate_string, ) V1_V1_RES_MAP = { @@ -234,7 +236,7 @@ class IdeogramV1(IO.ComfyNode): return IO.Schema( node_id="IdeogramV1", display_name="Ideogram V1", - category="api node/image/Ideogram", + category="partner/image/Ideogram", description="Generates images using the Ideogram V1 model.", inputs=[ IO.String.Input( @@ -360,7 +362,7 @@ class IdeogramV2(IO.ComfyNode): return IO.Schema( node_id="IdeogramV2", display_name="Ideogram V2", - category="api node/image/Ideogram", + category="partner/image/Ideogram", description="Generates images using the Ideogram V2 model.", inputs=[ IO.String.Input( @@ -526,7 +528,7 @@ class IdeogramV3(IO.ComfyNode): return IO.Schema( node_id="IdeogramV3", display_name="Ideogram V3", - category="api node/image/Ideogram", + category="partner/image/Ideogram", description="Generates images using the Ideogram V3 model. " "Supports both regular image generation from text prompts and image editing with mask.", inputs=[ @@ -798,6 +800,119 @@ class IdeogramV3(IO.ComfyNode): return IO.NodeOutput(await download_and_process_images(image_urls)) +class IdeogramV4(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="IdeogramV4", + display_name="Ideogram V4", + category="partner/image/Ideogram", + description="Generates images using the Ideogram 4.0 model from a text prompt.", + inputs=[ + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Text prompt for the image generation.", + ), + IO.Combo.Input( + "resolution", + options=[ + "Auto", + "2048x2048 (1:1)", + "1440x2880 (1:2)", + "2880x1440 (2:1)", + "1664x2496 (2:3)", + "2496x1664 (3:2)", + "1792x2240 (4:5)", + "2240x1792 (5:4)", + "1440x2560 (9:16)", + "2560x1440 (16:9)", + "1600x2560 (5:8)", + "2560x1600 (8:5)", + "1728x2304 (3:4)", + "2304x1728 (4:3)", + "1296x3168 (9:22)", + "3168x1296 (22:9)", + "1152x2944 (9:23)", + "2944x1152 (23:9)", + "1248x3328 (3:8)", + "3328x1248 (8:3)", + "1280x3072 (5:12)", + "3072x1280 (12:5)", + ], + default="Auto", + ), + IO.Combo.Input( + "rendering_speed", + options=["DEFAULT", "TURBO", "QUALITY"], + default="DEFAULT", + tooltip="Controls the trade-off between generation speed and quality.", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + step=1, + control_after_generate=True, + display_mode=IO.NumberDisplay.number, + ), + ], + outputs=[ + IO.Image.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["rendering_speed"]), + expr=""" + ( + $speed := widgets.rendering_speed; + $price := + $contains($speed,"turbo") ? 0.0429 : + $contains($speed,"quality") ? 0.143 : + 0.0858; + {"type":"usd","usd": $price} + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + prompt: str, + resolution: str, + rendering_speed: str, + seed: int, + ): + validate_string(prompt, strip_whitespace=True, min_length=1) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/ideogram/ideogram-v4/generate", method="POST"), + response_model=IdeogramGenerateResponse, + data=IdeogramV4Request( + text_prompt=prompt, + resolution=resolution.split(" ")[0] if resolution != "Auto" else None, + rendering_speed=rendering_speed, + ), + max_retries=1, + ) + + if not response.data or len(response.data) == 0: + raise Exception("No images were generated in the response") + image_urls = [image_data.url for image_data in response.data if image_data.url] + if not image_urls: + raise Exception("No image URLs were generated in the response") + return IO.NodeOutput(await download_and_process_images(image_urls)) + + class IdeogramExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: @@ -805,6 +920,7 @@ class IdeogramExtension(ComfyExtension): IdeogramV1, IdeogramV2, IdeogramV3, + IdeogramV4, ] diff --git a/comfy_api_nodes/nodes_kling.py b/comfy_api_nodes/nodes_kling.py index 7586f1816..d11e42540 100644 --- a/comfy_api_nodes/nodes_kling.py +++ b/comfy_api_nodes/nodes_kling.py @@ -642,7 +642,7 @@ class KlingCameraControls(IO.ComfyNode): return IO.Schema( node_id="KlingCameraControls", display_name="Kling Camera Controls", - category="api node/video/Kling", + category="partner/video/Kling", description="Allows specifying configuration options for Kling Camera Controls and motion control effects.", inputs=[ IO.Combo.Input("camera_control_type", options=KlingCameraControlType), @@ -762,7 +762,7 @@ class KlingTextToVideoNode(IO.ComfyNode): return IO.Schema( node_id="KlingTextToVideoNode", display_name="Kling Text to Video", - category="api node/video/Kling", + category="partner/video/Kling", description="Kling Text to Video Node", inputs=[ IO.String.Input("prompt", multiline=True, tooltip="Positive text prompt"), @@ -849,7 +849,7 @@ class OmniProTextToVideoNode(IO.ComfyNode): return IO.Schema( node_id="KlingOmniProTextToVideoNode", display_name="Kling 3.0 Omni Text to Video", - category="api node/video/Kling", + category="partner/video/Kling", description="Use text prompts to generate videos with the latest Kling model.", inputs=[ IO.Combo.Input("model_name", options=["kling-v3-omni", "kling-video-o1"]), @@ -998,7 +998,7 @@ class OmniProFirstLastFrameNode(IO.ComfyNode): return IO.Schema( node_id="KlingOmniProFirstLastFrameNode", display_name="Kling 3.0 Omni First-Last-Frame to Video", - category="api node/video/Kling", + category="partner/video/Kling", description="Use a start frame, an optional end frame, or reference images with the latest Kling model.", inputs=[ IO.Combo.Input("model_name", options=["kling-v3-omni", "kling-video-o1"]), @@ -1205,7 +1205,7 @@ class OmniProImageToVideoNode(IO.ComfyNode): return IO.Schema( node_id="KlingOmniProImageToVideoNode", display_name="Kling 3.0 Omni Image to Video", - category="api node/video/Kling", + category="partner/video/Kling", description="Use up to 7 reference images to generate a video with the latest Kling model.", inputs=[ IO.Combo.Input("model_name", options=["kling-v3-omni", "kling-video-o1"]), @@ -1374,7 +1374,7 @@ class OmniProVideoToVideoNode(IO.ComfyNode): return IO.Schema( node_id="KlingOmniProVideoToVideoNode", display_name="Kling 3.0 Omni Video to Video", - category="api node/video/Kling", + category="partner/video/Kling", description="Use a video and up to 4 reference images to generate a video with the latest Kling model.", inputs=[ IO.Combo.Input("model_name", options=["kling-v3-omni", "kling-video-o1"]), @@ -1485,7 +1485,7 @@ class OmniProEditVideoNode(IO.ComfyNode): return IO.Schema( node_id="KlingOmniProEditVideoNode", display_name="Kling 3.0 Omni Edit Video", - category="api node/video/Kling", + category="partner/video/Kling", essentials_category="Video Generation", description="Edit an existing video with the latest model from Kling.", inputs=[ @@ -1593,7 +1593,7 @@ class OmniProImageNode(IO.ComfyNode): return IO.Schema( node_id="KlingOmniProImageNode", display_name="Kling 3.0 Omni Image", - category="api node/image/Kling", + category="partner/image/Kling", description="Create or edit images with the latest model from Kling.", inputs=[ IO.Combo.Input("model_name", options=["kling-v3-omni", "kling-image-o1"]), @@ -1721,7 +1721,7 @@ class KlingCameraControlT2VNode(IO.ComfyNode): return IO.Schema( node_id="KlingCameraControlT2VNode", display_name="Kling Text to Video (Camera Control)", - category="api node/video/Kling", + category="partner/video/Kling", description="Transform text into cinematic videos with professional camera movements that simulate real-world cinematography. Control virtual camera actions including zoom, rotation, pan, tilt, and first-person view, while maintaining focus on your original text.", inputs=[ IO.String.Input("prompt", multiline=True, tooltip="Positive text prompt"), @@ -1783,7 +1783,7 @@ class KlingImage2VideoNode(IO.ComfyNode): return IO.Schema( node_id="KlingImage2VideoNode", display_name="Kling Image(First Frame) to Video", - category="api node/video/Kling", + category="partner/video/Kling", inputs=[ IO.Image.Input("start_frame", tooltip="The reference image used to generate the video."), IO.String.Input("prompt", multiline=True, tooltip="Positive text prompt"), @@ -1882,7 +1882,7 @@ class KlingCameraControlI2VNode(IO.ComfyNode): return IO.Schema( node_id="KlingCameraControlI2VNode", display_name="Kling Image to Video (Camera Control)", - category="api node/video/Kling", + category="partner/video/Kling", description="Transform still images into cinematic videos with professional camera movements that simulate real-world cinematography. Control virtual camera actions including zoom, rotation, pan, tilt, and first-person view, while maintaining focus on your original image.", inputs=[ IO.Image.Input( @@ -1953,7 +1953,7 @@ class KlingStartEndFrameNode(IO.ComfyNode): return IO.Schema( node_id="KlingStartEndFrameNode", display_name="Kling Start-End Frame to Video", - category="api node/video/Kling", + category="partner/video/Kling", description="Generate a video sequence that transitions between your provided start and end images. The node creates all frames in between, producing a smooth transformation from the first frame to the last.", inputs=[ IO.Image.Input( @@ -2047,7 +2047,7 @@ class KlingVideoExtendNode(IO.ComfyNode): return IO.Schema( node_id="KlingVideoExtendNode", display_name="Kling Video Extend", - category="api node/video/Kling", + category="partner/video/Kling", description="Kling Video Extend Node. Extend videos made by other Kling nodes. The video_id is created by using other Kling Nodes.", inputs=[ IO.String.Input( @@ -2128,7 +2128,7 @@ class KlingDualCharacterVideoEffectNode(IO.ComfyNode): return IO.Schema( node_id="KlingDualCharacterVideoEffectNode", display_name="Kling Dual Character Video Effects", - category="api node/video/Kling", + category="partner/video/Kling", description="Achieve different special effects when generating a video based on the effect_scene. First image will be positioned on left side, second on right side of the composite.", inputs=[ IO.Image.Input("image_left", tooltip="Left side image"), @@ -2218,7 +2218,7 @@ class KlingSingleImageVideoEffectNode(IO.ComfyNode): return IO.Schema( node_id="KlingSingleImageVideoEffectNode", display_name="Kling Video Effects", - category="api node/video/Kling", + category="partner/video/Kling", description="Achieve different special effects when generating a video based on the effect_scene.", inputs=[ IO.Image.Input( @@ -2291,7 +2291,7 @@ class KlingLipSyncAudioToVideoNode(IO.ComfyNode): return IO.Schema( node_id="KlingLipSyncAudioToVideoNode", display_name="Kling Lip Sync Video with Audio", - category="api node/video/Kling", + category="partner/video/Kling", essentials_category="Video Generation", description="Kling Lip Sync Audio to Video Node. Syncs mouth movements in a video file to the audio content of an audio file. When using, ensure that the audio contains clearly distinguishable vocals and that the video contains a distinct face. The audio file should not be larger than 5MB. The video file should not be larger than 100MB, should have height/width between 720px and 1920px, and should be between 2s and 10s in length.", inputs=[ @@ -2343,7 +2343,7 @@ class KlingLipSyncTextToVideoNode(IO.ComfyNode): return IO.Schema( node_id="KlingLipSyncTextToVideoNode", display_name="Kling Lip Sync Video with Text", - category="api node/video/Kling", + category="partner/video/Kling", description="Kling Lip Sync Text to Video Node. Syncs mouth movements in a video file to a text prompt. The video file should not be larger than 100MB, should have height/width between 720px and 1920px, and should be between 2s and 10s in length.", inputs=[ IO.Video.Input("video"), @@ -2411,7 +2411,7 @@ class KlingVirtualTryOnNode(IO.ComfyNode): return IO.Schema( node_id="KlingVirtualTryOnNode", display_name="Kling Virtual Try On", - category="api node/image/Kling", + category="partner/image/Kling", description="Kling Virtual Try On Node. Input a human image and a cloth image to try on the cloth on the human. You can merge multiple clothing item pictures into one image with a white background.", inputs=[ IO.Image.Input("human_image"), @@ -2478,7 +2478,7 @@ class KlingImageGenerationNode(IO.ComfyNode): return IO.Schema( node_id="KlingImageGenerationNode", display_name="Kling 3.0 Image", - category="api node/image/Kling", + category="partner/image/Kling", description="Kling Image Generation Node. Generate an image from a text prompt with an optional reference image.", inputs=[ IO.String.Input("prompt", multiline=True, tooltip="Positive text prompt"), @@ -2615,7 +2615,7 @@ class TextToVideoWithAudio(IO.ComfyNode): return IO.Schema( node_id="KlingTextToVideoWithAudio", display_name="Kling 2.6 Text to Video with Audio", - category="api node/video/Kling", + category="partner/video/Kling", inputs=[ IO.Combo.Input("model_name", options=["kling-v2-6"]), IO.String.Input("prompt", multiline=True, tooltip="Positive text prompt."), @@ -2683,7 +2683,7 @@ class ImageToVideoWithAudio(IO.ComfyNode): return IO.Schema( node_id="KlingImageToVideoWithAudio", display_name="Kling 2.6 Image(First Frame) to Video with Audio", - category="api node/video/Kling", + category="partner/video/Kling", inputs=[ IO.Combo.Input("model_name", options=["kling-v2-6"]), IO.Image.Input("start_frame"), @@ -2753,7 +2753,7 @@ class MotionControl(IO.ComfyNode): return IO.Schema( node_id="KlingMotionControl", display_name="Kling Motion Control", - category="api node/video/Kling", + category="partner/video/Kling", inputs=[ IO.String.Input("prompt", multiline=True), IO.Image.Input("reference_image"), @@ -2854,7 +2854,7 @@ class KlingVideoNode(IO.ComfyNode): return IO.Schema( node_id="KlingVideoNode", display_name="Kling 3.0 Video", - category="api node/video/Kling", + category="partner/video/Kling", description="Generate videos with Kling V3. " "Supports text-to-video and image-to-video with optional storyboard multi-prompt and audio generation.", inputs=[ @@ -3077,7 +3077,7 @@ class KlingFirstLastFrameNode(IO.ComfyNode): return IO.Schema( node_id="KlingFirstLastFrameNode", display_name="Kling 3.0 First-Last-Frame to Video", - category="api node/video/Kling", + category="partner/video/Kling", description="Generate videos with Kling V3 using first and last frames.", inputs=[ IO.String.Input("prompt", multiline=True, default=""), @@ -3202,7 +3202,7 @@ class KlingAvatarNode(IO.ComfyNode): return IO.Schema( node_id="KlingAvatarNode", display_name="Kling Avatar 2.0", - category="api node/video/Kling", + category="partner/video/Kling", description="Generate broadcast-style digital human videos from a single photo and an audio file.", inputs=[ IO.Image.Input( diff --git a/comfy_api_nodes/nodes_krea.py b/comfy_api_nodes/nodes_krea.py new file mode 100644 index 000000000..b9e6268f2 --- /dev/null +++ b/comfy_api_nodes/nodes_krea.py @@ -0,0 +1,294 @@ +"""Krea image-generation nodes.""" + +import re + +from typing_extensions import override + +from comfy_api.latest import IO, ComfyExtension, Input +from comfy_api_nodes.apis.krea import ( + KreaAssetResponse, + KreaGenerateImageRequest, + KreaImageStyleReference, + KreaJob, + KreaMoodboard, +) +from comfy_api_nodes.util import ( + ApiEndpoint, + download_url_to_image_tensor, + poll_op, + sync_op, + tensor_to_bytesio, + validate_string, +) + + +class KreaIO: + STYLE_REF = "KREA_STYLE_REF" + + +async def _upload_image_to_krea_assets(cls: type[IO.ComfyNode], image: Input.Image) -> str: + """Upload an image to Krea's /assets endpoint and return the Krea-hosted image URL.""" + img_io = tensor_to_bytesio(image, total_pixels=2048 * 2048, mime_type="image/png") + response = await sync_op( + cls, + endpoint=ApiEndpoint(path="/proxy/krea/assets", method="POST"), + response_model=KreaAssetResponse, + files=[("file", (img_io.name, img_io, "image/png"))], + content_type="multipart/form-data", + max_retries=1, + wait_label="Uploading reference", + ) + return response.image_url + + +_MODEL_MEDIUM = "Krea 2 Medium" +_MODEL_MEDIUM_TURBO = "Krea 2 Medium Turbo" +_MODEL_LARGE = "Krea 2 Large" +_MODEL_ENDPOINTS: dict[str, str] = { + _MODEL_MEDIUM: "/proxy/krea/generate/image/krea/krea-2/medium", + _MODEL_MEDIUM_TURBO: "/proxy/krea/generate/image/krea/krea-2/medium-turbo", + _MODEL_LARGE: "/proxy/krea/generate/image/krea/krea-2/large", +} + +_ASPECT_RATIOS = ["1:1", "4:3", "3:2", "16:9", "2.35:1", "4:5", "2:3", "9:16"] +_RESOLUTIONS = ["1K"] +_CREATIVITY_LEVELS = ["raw", "low", "medium", "high"] +_KREA_QUEUED_STATUSES = ["backlogged", "queued", "scheduled"] + +_UUID_RE = re.compile(r"^[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}$") + + +def _krea_model_inputs() -> list: + """Nested inputs shared by Krea 2 Medium, Medium Turbo and Large under the DynamicCombo.""" + return [ + IO.Combo.Input( + "aspect_ratio", + options=_ASPECT_RATIOS, + tooltip="Output aspect ratio.", + ), + IO.Combo.Input( + "resolution", + options=_RESOLUTIONS, + tooltip="Resolution scale.", + ), + IO.Combo.Input( + "creativity", + options=_CREATIVITY_LEVELS, + default="medium", + tooltip="Prompt interpretation strength: raw stays closest to the prompt; high is most creative.", + ), + IO.String.Input( + "moodboard_id", + default="", + tooltip="Optional Krea moodboard UUID (e.g. from the Krea website). " + "Leave empty to disable. Only one moodboard is supported per request.", + optional=True, + ), + IO.Float.Input( + "moodboard_strength", + default=0.35, + min=-0.5, + max=1.5, + step=0.05, + tooltip="Moodboard influence; ignored when moodboard_id is empty.", + optional=True, + ), + IO.Custom(KreaIO.STYLE_REF).Input( + "style_reference", + optional=True, + tooltip="Optional chain of style references (max 10) from Krea 2 Style Reference nodes.", + ), + ] + + +class Krea2ImageNode(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="Krea2ImageNode", + display_name="Krea 2 Image", + category="partner/image/Krea", + description=( + "Generate images via Krea 2 — pick Medium (expressive illustrations) or " + "Large (expressive photorealism). Supports an optional moodboard and up " + "to 10 chained image style references." + ), + inputs=[ + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Text prompt for the image.", + ), + IO.DynamicCombo.Input( + "model", + options=[ + IO.DynamicCombo.Option(_MODEL_MEDIUM, _krea_model_inputs()), + IO.DynamicCombo.Option(_MODEL_MEDIUM_TURBO, _krea_model_inputs()), + IO.DynamicCombo.Option(_MODEL_LARGE, _krea_model_inputs()), + ], + tooltip="Krea 2 Medium is best for expressive illustrations; " + "Krea 2 Large is best for expressive photorealism.", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + control_after_generate=True, + tooltip="Random seed for reproducibility.", + ), + ], + outputs=[IO.Image.Output()], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends( + widgets=["model", "model.moodboard_id"], + inputs=["model.style_reference"], + ), + expr=""" + ( + $rates := { + "krea 2 medium turbo": {"text": 0.015, "style": 0.0175, "moodboard": 0.02}, + "krea 2 medium": {"text": 0.03, "style": 0.035, "moodboard": 0.04}, + "krea 2 large": {"text": 0.06, "style": 0.065, "moodboard": 0.07} + }; + $r := $lookup($rates, widgets.model); + $hasMoodboard := $length($lookup(widgets, "model.moodboard_id")) > 0; + $hasStyle := $lookup(inputs, "model.style_reference").connected; + $usd := $hasMoodboard ? $r.moodboard : ($hasStyle ? $r.style : $r.text); + {"type":"usd","usd": $usd} + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + prompt: str, + model: dict, + seed: int, + ) -> IO.NodeOutput: + validate_string(prompt, strip_whitespace=False, min_length=1) + + model_choice = model["model"] + endpoint_path = _MODEL_ENDPOINTS.get(model_choice) + if endpoint_path is None: + raise ValueError(f"Unknown Krea 2 model: {model_choice!r}") + + moodboards: list[KreaMoodboard] | None = None + mb_id = (model.get("moodboard_id") or "").strip() + if mb_id: + if not _UUID_RE.match(mb_id): + raise ValueError(f"moodboard_id must be a UUID (received {mb_id!r}); copy it from the Krea website.") + mb_strength = model.get("moodboard_strength") + moodboards = [KreaMoodboard(id=mb_id, strength=0.35 if mb_strength is None else float(mb_strength))] + + style_reference = model.get("style_reference") + image_style_references: list[KreaImageStyleReference] | None = None + if style_reference: + if len(style_reference) > 10: + raise ValueError(f"Krea 2 accepts at most 10 image_style_references; received {len(style_reference)}.") + image_style_references = [ + KreaImageStyleReference(url=ref["url"], strength=float(ref["strength"])) for ref in style_reference + ] + initial = await sync_op( + cls, + ApiEndpoint(path=endpoint_path, method="POST"), + response_model=KreaJob, + data=KreaGenerateImageRequest( + prompt=prompt, + aspect_ratio=model["aspect_ratio"], + resolution=model["resolution"], + seed=seed, + creativity=model["creativity"], + moodboards=moodboards, + image_style_references=image_style_references, + ), + ) + job = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/krea/jobs/{initial.job_id}", method="GET"), + response_model=KreaJob, + status_extractor=lambda r: r.status, + queued_statuses=_KREA_QUEUED_STATUSES, + ) + if not job.result or not job.result.urls: + raise RuntimeError(f"Krea 2 job {job.job_id} completed without any image URLs.") + image = await download_url_to_image_tensor(job.result.urls[0]) + return IO.NodeOutput(image) + + +class Krea2StyleReferenceNode(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="Krea2StyleReferenceNode", + display_name="Krea 2 Style Reference", + category="partner/image/Krea", + description=( + "Add an image style reference to a Krea 2 generation. Chain multiple Krea 2 " + "Style Reference nodes (max 10) and feed the final `style_reference` output " + "into Krea 2 Image. Each image is uploaded to ComfyAPI storage and passed as URL." + ), + inputs=[ + IO.Image.Input( + "image", + tooltip="Reference image whose style influences the generation.", + ), + IO.Float.Input( + "strength", + default=1.0, + min=-2.0, + max=2.0, + step=0.05, + tooltip="Reference strength; negative values invert the style influence.", + ), + IO.Custom(KreaIO.STYLE_REF).Input( + "style_reference", + optional=True, + tooltip="Optional incoming chain of style references; this node appends one more.", + ), + ], + outputs=[IO.Custom(KreaIO.STYLE_REF).Output(display_name="style_reference")], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + ) + + @classmethod + async def execute( + cls, + image: Input.Image, + strength: float, + style_reference: list[dict] | None = None, + ) -> IO.NodeOutput: + chain: list[dict] = list(style_reference) if style_reference else [] + if len(chain) >= 10: + raise ValueError("Krea 2 accepts at most 10 image_style_references in one generation.") + url = await _upload_image_to_krea_assets(cls, image) + chain.append({"url": url, "strength": float(strength)}) + return IO.NodeOutput(chain) + + +class KreaExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + Krea2ImageNode, + Krea2StyleReferenceNode, + ] + + +async def comfy_entrypoint() -> KreaExtension: + return KreaExtension() diff --git a/comfy_api_nodes/nodes_ltxv.py b/comfy_api_nodes/nodes_ltxv.py index 0a219af96..878e04b4e 100644 --- a/comfy_api_nodes/nodes_ltxv.py +++ b/comfy_api_nodes/nodes_ltxv.py @@ -50,7 +50,7 @@ class TextToVideoNode(IO.ComfyNode): return IO.Schema( node_id="LtxvApiTextToVideo", display_name="LTXV Text To Video", - category="api node/video/LTXV", + category="partner/video/LTXV", description="Professional-quality videos with customizable duration and resolution.", inputs=[ IO.Combo.Input("model", options=list(MODELS_MAP.keys())), @@ -127,7 +127,7 @@ class ImageToVideoNode(IO.ComfyNode): return IO.Schema( node_id="LtxvApiImageToVideo", display_name="LTXV Image To Video", - category="api node/video/LTXV", + category="partner/video/LTXV", description="Professional-quality videos with customizable duration and resolution based on start image.", inputs=[ IO.Image.Input("image", tooltip="First frame to be used for the video."), diff --git a/comfy_api_nodes/nodes_luma.py b/comfy_api_nodes/nodes_luma.py index d92a7c382..0d31ac77e 100644 --- a/comfy_api_nodes/nodes_luma.py +++ b/comfy_api_nodes/nodes_luma.py @@ -46,7 +46,7 @@ class LumaReferenceNode(IO.ComfyNode): return IO.Schema( node_id="LumaReferenceNode", display_name="Luma Reference", - category="api node/image/Luma", + category="partner/image/Luma", description="Holds an image and weight for use with Luma Generate Image node.", inputs=[ IO.Image.Input( @@ -85,7 +85,7 @@ class LumaConceptsNode(IO.ComfyNode): return IO.Schema( node_id="LumaConceptsNode", display_name="Luma Concepts", - category="api node/video/Luma", + category="partner/video/Luma", description="Camera Concepts for use with Luma Text to Video and Luma Image to Video nodes.", inputs=[ IO.Combo.Input( @@ -134,7 +134,7 @@ class LumaImageGenerationNode(IO.ComfyNode): return IO.Schema( node_id="LumaImageNode", display_name="Luma Text to Image", - category="api node/image/Luma", + category="partner/image/Luma", description="Generates images synchronously based on prompt and aspect ratio.", inputs=[ IO.String.Input( @@ -278,7 +278,7 @@ class LumaImageModifyNode(IO.ComfyNode): return IO.Schema( node_id="LumaImageModifyNode", display_name="Luma Image to Image", - category="api node/image/Luma", + category="partner/image/Luma", description="Modifies images synchronously based on prompt and aspect ratio.", inputs=[ IO.Image.Input( @@ -371,7 +371,7 @@ class LumaTextToVideoGenerationNode(IO.ComfyNode): return IO.Schema( node_id="LumaVideoNode", display_name="Luma Text to Video", - category="api node/video/Luma", + category="partner/video/Luma", description="Generates videos synchronously based on prompt and output_size.", inputs=[ IO.String.Input( @@ -472,7 +472,7 @@ class LumaImageToVideoGenerationNode(IO.ComfyNode): return IO.Schema( node_id="LumaImageToVideoNode", display_name="Luma Image to Video", - category="api node/video/Luma", + category="partner/video/Luma", description="Generates videos synchronously based on prompt, input images, and output_size.", inputs=[ IO.String.Input( @@ -724,7 +724,7 @@ class LumaImageNode(IO.ComfyNode): return IO.Schema( node_id="LumaImageNode2", display_name="Luma UNI-1 Image", - category="api node/image/Luma", + category="partner/image/Luma", description="Generate images from text using the Luma UNI-1 model.", inputs=[ IO.String.Input( @@ -853,7 +853,7 @@ class LumaImageEditNode(IO.ComfyNode): return IO.Schema( node_id="LumaImageEditNode2", display_name="Luma UNI-1 Image Edit", - category="api node/image/Luma", + category="partner/image/Luma", description="Edit an existing image with a text prompt using the Luma UNI-1 model.", inputs=[ IO.Image.Input( diff --git a/comfy_api_nodes/nodes_magnific.py b/comfy_api_nodes/nodes_magnific.py index 38b881fea..4ce4735df 100644 --- a/comfy_api_nodes/nodes_magnific.py +++ b/comfy_api_nodes/nodes_magnific.py @@ -61,7 +61,7 @@ class MagnificImageUpscalerCreativeNode(IO.ComfyNode): return IO.Schema( node_id="MagnificImageUpscalerCreativeNode", display_name="Magnific Image Upscale (Creative)", - category="api node/image/Magnific", + category="partner/image/Magnific", description="Prompt‑guided enhancement, stylization, and 2x/4x/8x/16x upscaling. " "Maximum output: 25.3 megapixels.", inputs=[ @@ -240,7 +240,7 @@ class MagnificImageUpscalerPreciseV2Node(IO.ComfyNode): return IO.Schema( node_id="MagnificImageUpscalerPreciseV2Node", display_name="Magnific Image Upscale (Precise V2)", - category="api node/image/Magnific", + category="partner/image/Magnific", description="High-fidelity upscaling with fine control over sharpness, grain, and detail. " "Maximum output: 10060×10060 pixels.", inputs=[ @@ -400,7 +400,7 @@ class MagnificImageStyleTransferNode(IO.ComfyNode): return IO.Schema( node_id="MagnificImageStyleTransferNode", display_name="Magnific Image Style Transfer", - category="api node/image/Magnific", + category="partner/image/Magnific", description="Transfer the style from a reference image to your input image.", inputs=[ IO.Image.Input("image", tooltip="The image to apply style transfer to."), @@ -549,7 +549,7 @@ class MagnificImageRelightNode(IO.ComfyNode): return IO.Schema( node_id="MagnificImageRelightNode", display_name="Magnific Image Relight", - category="api node/image/Magnific", + category="partner/image/Magnific", description="Relight an image with lighting adjustments and optional reference-based light transfer.", inputs=[ IO.Image.Input("image", tooltip="The image to relight."), @@ -789,7 +789,7 @@ class MagnificImageSkinEnhancerNode(IO.ComfyNode): return IO.Schema( node_id="MagnificImageSkinEnhancerNode", display_name="Magnific Image Skin Enhancer", - category="api node/image/Magnific", + category="partner/image/Magnific", description="Skin enhancement for portraits with multiple processing modes.", inputs=[ IO.Image.Input("image", tooltip="The portrait image to enhance."), diff --git a/comfy_api_nodes/nodes_meshy.py b/comfy_api_nodes/nodes_meshy.py index 3cf577f4a..3a24f1095 100644 --- a/comfy_api_nodes/nodes_meshy.py +++ b/comfy_api_nodes/nodes_meshy.py @@ -33,7 +33,7 @@ class MeshyTextToModelNode(IO.ComfyNode): return IO.Schema( node_id="MeshyTextToModelNode", display_name="Meshy: Text to Model", - category="api node/3d/Meshy", + category="partner/3d/Meshy", inputs=[ IO.Combo.Input("model", options=["latest"]), IO.String.Input("prompt", multiline=True, default=""), @@ -145,7 +145,7 @@ class MeshyRefineNode(IO.ComfyNode): return IO.Schema( node_id="MeshyRefineNode", display_name="Meshy: Refine Draft Model", - category="api node/3d/Meshy", + category="partner/3d/Meshy", description="Refine a previously created draft model.", inputs=[ IO.Combo.Input("model", options=["latest"]), @@ -240,7 +240,7 @@ class MeshyImageToModelNode(IO.ComfyNode): return IO.Schema( node_id="MeshyImageToModelNode", display_name="Meshy: Image to Model", - category="api node/3d/Meshy", + category="partner/3d/Meshy", inputs=[ IO.Combo.Input("model", options=["latest"]), IO.Image.Input("image"), @@ -405,7 +405,7 @@ class MeshyMultiImageToModelNode(IO.ComfyNode): return IO.Schema( node_id="MeshyMultiImageToModelNode", display_name="Meshy: Multi-Image to Model", - category="api node/3d/Meshy", + category="partner/3d/Meshy", inputs=[ IO.Combo.Input("model", options=["latest"]), IO.Autogrow.Input( @@ -575,7 +575,7 @@ class MeshyRigModelNode(IO.ComfyNode): return IO.Schema( node_id="MeshyRigModelNode", display_name="Meshy: Rig Model", - category="api node/3d/Meshy", + category="partner/3d/Meshy", description="Provides a rigged character in standard formats. " "Auto-rigging is currently not suitable for untextured meshes, non-humanoid assets, " "or humanoid assets with unclear limb and body structure.", @@ -656,7 +656,7 @@ class MeshyAnimateModelNode(IO.ComfyNode): return IO.Schema( node_id="MeshyAnimateModelNode", display_name="Meshy: Animate Model", - category="api node/3d/Meshy", + category="partner/3d/Meshy", description="Apply a specific animation action to a previously rigged character.", inputs=[ IO.Custom("MESHY_RIGGED_TASK_ID").Input("rig_task_id"), @@ -722,7 +722,7 @@ class MeshyTextureNode(IO.ComfyNode): return IO.Schema( node_id="MeshyTextureNode", display_name="Meshy: Texture Model", - category="api node/3d/Meshy", + category="partner/3d/Meshy", inputs=[ IO.Combo.Input("model", options=["latest"]), IO.Custom("MESHY_TASK_ID").Input("meshy_task_id"), diff --git a/comfy_api_nodes/nodes_minimax.py b/comfy_api_nodes/nodes_minimax.py index b5d0b461f..6250af146 100644 --- a/comfy_api_nodes/nodes_minimax.py +++ b/comfy_api_nodes/nodes_minimax.py @@ -101,7 +101,7 @@ class MinimaxTextToVideoNode(IO.ComfyNode): return IO.Schema( node_id="MinimaxTextToVideoNode", display_name="MiniMax Text to Video", - category="api node/video/MiniMax", + category="partner/video/MiniMax", description="Generates videos synchronously based on a prompt, and optional parameters.", inputs=[ IO.String.Input( @@ -163,7 +163,7 @@ class MinimaxImageToVideoNode(IO.ComfyNode): return IO.Schema( node_id="MinimaxImageToVideoNode", display_name="MiniMax Image to Video", - category="api node/video/MiniMax", + category="partner/video/MiniMax", description="Generates videos synchronously based on an image and prompt, and optional parameters.", inputs=[ IO.Image.Input( @@ -230,7 +230,7 @@ class MinimaxSubjectToVideoNode(IO.ComfyNode): return IO.Schema( node_id="MinimaxSubjectToVideoNode", display_name="MiniMax Subject to Video", - category="api node/video/MiniMax", + category="partner/video/MiniMax", description="Generates videos synchronously based on an image and prompt, and optional parameters.", inputs=[ IO.Image.Input( @@ -294,7 +294,7 @@ class MinimaxHailuoVideoNode(IO.ComfyNode): return IO.Schema( node_id="MinimaxHailuoVideoNode", display_name="MiniMax Hailuo Video", - category="api node/video/MiniMax", + category="partner/video/MiniMax", description="Generates videos from prompt, with optional start frame using the new MiniMax Hailuo-02 model.", inputs=[ IO.String.Input( diff --git a/comfy_api_nodes/nodes_openai.py b/comfy_api_nodes/nodes_openai.py index a5a188634..0fe5fb9d0 100644 --- a/comfy_api_nodes/nodes_openai.py +++ b/comfy_api_nodes/nodes_openai.py @@ -99,7 +99,7 @@ class OpenAIDalle2(IO.ComfyNode): return IO.Schema( node_id="OpenAIDalle2", display_name="OpenAI DALL·E 2", - category="api node/image/OpenAI", + category="partner/image/OpenAI", description="Generates images synchronously via OpenAI's DALL·E 2 endpoint.", inputs=[ IO.String.Input( @@ -249,7 +249,7 @@ class OpenAIDalle3(IO.ComfyNode): return IO.Schema( node_id="OpenAIDalle3", display_name="OpenAI DALL·E 3", - category="api node/image/OpenAI", + category="partner/image/OpenAI", description="Generates images synchronously via OpenAI's DALL·E 3 endpoint.", inputs=[ IO.String.Input( @@ -371,7 +371,7 @@ class OpenAIGPTImage1(IO.ComfyNode): return IO.Schema( node_id="OpenAIGPTImage1", display_name="OpenAI GPT Image 2", - category="api node/image/OpenAI", + category="partner/image/OpenAI", description="Generates images synchronously via OpenAI's GPT Image endpoint.", is_deprecated=True, inputs=[ @@ -695,7 +695,7 @@ class OpenAIGPTImageNodeV2(IO.ComfyNode): return IO.Schema( node_id="OpenAIGPTImageNodeV2", display_name="OpenAI GPT Image 2", - category="api node/image/OpenAI", + category="partner/image/OpenAI", description="Generates images via OpenAI's GPT Image endpoint.", inputs=[ IO.String.Input( @@ -962,7 +962,7 @@ class OpenAIChatNode(IO.ComfyNode): return IO.Schema( node_id="OpenAIChatNode", display_name="OpenAI ChatGPT", - category="api node/text/OpenAI", + category="partner/text/OpenAI", essentials_category="Text Generation", description="Generate text responses from an OpenAI model.", inputs=[ @@ -1201,7 +1201,7 @@ class OpenAIInputFiles(IO.ComfyNode): return IO.Schema( node_id="OpenAIInputFiles", display_name="OpenAI ChatGPT Input Files", - category="api node/text/OpenAI", + category="partner/text/OpenAI", description="Loads and prepares input files (text, pdf, etc.) to include as inputs for the OpenAI Chat Node. The files will be read by the OpenAI model when generating a response. 🛈 TIP: Can be chained together with other OpenAI Input File nodes.", inputs=[ IO.Combo.Input( @@ -1248,7 +1248,7 @@ class OpenAIChatConfig(IO.ComfyNode): return IO.Schema( node_id="OpenAIChatConfig", display_name="OpenAI ChatGPT Advanced Options", - category="api node/text/OpenAI", + category="partner/text/OpenAI", description="Allows specifying advanced configuration options for the OpenAI Chat Nodes.", inputs=[ IO.Combo.Input( diff --git a/comfy_api_nodes/nodes_openrouter.py b/comfy_api_nodes/nodes_openrouter.py new file mode 100644 index 000000000..ba98133f0 --- /dev/null +++ b/comfy_api_nodes/nodes_openrouter.py @@ -0,0 +1,374 @@ +"""API Nodes for OpenRouter LLM chat completions.""" + +from dataclasses import dataclass +from typing import Literal + +from typing_extensions import override + +from comfy_api.latest import IO, ComfyExtension, Input +from comfy_api_nodes.apis.openrouter import ( + OpenRouterChatRequest, + OpenRouterChatResponse, + OpenRouterContentBlock, + OpenRouterImageContent, + OpenRouterImageUrl, + OpenRouterMessage, + OpenRouterReasoningConfig, + OpenRouterTextContent, + OpenRouterVideoContent, + OpenRouterVideoUrl, + OpenRouterWebSearchOptions, +) +from comfy_api_nodes.util import ( + ApiEndpoint, + get_number_of_images, + sync_op, + upload_images_to_comfyapi, + upload_video_to_comfyapi, + validate_string, +) + +OPENROUTER_CHAT_ENDPOINT = "/proxy/openrouter/api/v1/chat/completions" + + +Profile = Literal["standard", "reasoning", "frontier_reasoning", "perplexity", "perplexity_reasoning"] + + +@dataclass(frozen=True) +class _ModelSpec: + slug: str # exact OpenRouter model id + profile: Profile + price_in: float # USD per token (prompt) + price_out: float # USD per token (completion) + max_images: int = 0 # 0 = no image input; otherwise max URL-passed images supported + max_videos: int = 0 # 0 = no video input; otherwise max URL-passed videos supported + + +MODELS: list[_ModelSpec] = [ + _ModelSpec("anthropic/claude-opus-4.7", "frontier_reasoning", 0.000005, 0.000025, max_images=20), + _ModelSpec("openai/gpt-5.5-pro", "frontier_reasoning", 0.00003, 0.00018, max_images=20), + _ModelSpec("openai/gpt-5.5", "frontier_reasoning", 0.000005, 0.00003, max_images=20), + _ModelSpec("google/gemini-3.5-flash", "reasoning", 0.0000015, 0.000009, max_images=20, max_videos=4), + _ModelSpec("x-ai/grok-4.20", "reasoning", 0.00000125, 0.0000025, max_images=20), + _ModelSpec("x-ai/grok-4.3", "reasoning", 0.00000125, 0.0000025, max_images=20), + _ModelSpec("deepseek/deepseek-v4-pro", "reasoning", 0.000000435, 0.00000087), + _ModelSpec("deepseek/deepseek-v4-flash", "reasoning", 0.000000112, 0.000000224), + _ModelSpec("deepseek/deepseek-v3.2", "reasoning", 0.000000252, 0.000000378), + _ModelSpec("qwen/qwen3.6-max-preview", "reasoning", 0.00000104, 0.00000624), + _ModelSpec("qwen/qwen3.6-plus", "reasoning", 0.000000325, 0.00000195, max_images=10, max_videos=4), + _ModelSpec("qwen/qwen3.6-flash", "reasoning", 0.0000001875, 0.000001125, max_images=10, max_videos=4), + _ModelSpec("mistralai/mistral-large-2512", "standard", 0.0000005, 0.0000015, max_images=8), + _ModelSpec("mistralai/mistral-medium-3-5", "reasoning", 0.0000015, 0.0000075, max_images=8), + _ModelSpec("z-ai/glm-4.6", "reasoning", 0.00000043, 0.00000174), + _ModelSpec("z-ai/glm-5", "reasoning", 0.0000006, 0.00000192), + _ModelSpec("moonshotai/kimi-k2.6", "reasoning", 0.00000073, 0.00000349, max_images=10), + _ModelSpec("moonshotai/kimi-k2-thinking", "reasoning", 0.0000006, 0.0000025), + _ModelSpec("perplexity/sonar-pro", "perplexity", 0.000003, 0.000015), + _ModelSpec("perplexity/sonar-reasoning-pro", "perplexity_reasoning", 0.000002, 0.000008), + _ModelSpec("perplexity/sonar-deep-research", "perplexity_reasoning", 0.000002, 0.000008), +] + +_MODELS_BY_SLUG: dict[str, _ModelSpec] = {m.slug: m for m in MODELS} +_REASONING_EFFORTS = ["off", "low", "medium", "high"] +_SEARCH_CONTEXT_SIZES = ["low", "medium", "high"] + + +def _reasoning_extra_inputs() -> list: + return [ + IO.Combo.Input( + "reasoning_effort", + options=_REASONING_EFFORTS, + default="off", + tooltip="Reasoning effort. 'off' disables reasoning entirely.", + advanced=True, + ), + ] + + +def _perplexity_extra_inputs() -> list: + return [ + IO.Combo.Input( + "search_context_size", + options=_SEARCH_CONTEXT_SIZES, + default="medium", + tooltip="How much web search context to retrieve. Larger = more grounded but slower/pricier.", + advanced=True, + ), + ] + + +def _profile_inputs(profile: Profile) -> list: + if profile == "standard": + return [] + if profile in ("reasoning", "frontier_reasoning"): + return _reasoning_extra_inputs() + if profile == "perplexity": + return _perplexity_extra_inputs() + if profile == "perplexity_reasoning": + return _perplexity_extra_inputs() + _reasoning_extra_inputs() + raise ValueError(f"Unknown profile: {profile}") + + +def _media_inputs(spec: _ModelSpec) -> list: + extras: list = [] + if spec.max_images > 0: + extras.append( + IO.Autogrow.Input( + "images", + template=IO.Autogrow.TemplateNames( + IO.Image.Input("image"), + names=[f"image_{i}" for i in range(1, spec.max_images + 1)], + min=0, + ), + tooltip=f"Optional reference image(s) — up to {spec.max_images}. Sent as URLs.", + ) + ) + if spec.max_videos > 0: + extras.append( + IO.Autogrow.Input( + "videos", + template=IO.Autogrow.TemplateNames( + IO.Video.Input("video"), + names=[f"video_{i}" for i in range(1, spec.max_videos + 1)], + min=0, + ), + tooltip=f"Optional reference video(s) — up to {spec.max_videos}. Sent as URLs.", + ) + ) + return extras + + +def _inputs_for_model(spec: _ModelSpec) -> list: + return _profile_inputs(spec.profile) + _media_inputs(spec) + + +def _build_model_options() -> list[IO.DynamicCombo.Option]: + return [IO.DynamicCombo.Option(spec.slug, _inputs_for_model(spec)) for spec in MODELS] + + +def _calculate_price(response: OpenRouterChatResponse) -> float | None: + if response.usage and response.usage.cost is not None: + return float(response.usage.cost) + return None + + +def _price_badge_jsonata() -> str: + rates_pairs = [] + for spec in MODELS: + prompt_per_1k = spec.price_in * 1000 + completion_per_1k = spec.price_out * 1000 + rates_pairs.append(f' "{spec.slug}": [{prompt_per_1k:.8g}, {completion_per_1k:.8g}]') + rates_block = ",\n".join(rates_pairs) + return ( + "(\n" + " $rates := {\n" + f"{rates_block}\n" + " };\n" + " $r := $lookup($rates, widgets.model);\n" + " $r ? {\n" + ' "type": "list_usd",\n' + ' "usd": $r,\n' + ' "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" }\n' + ' } : {"type": "text", "text": "Token-based"}\n' + ")" + ) + + +async def _build_image_blocks( + cls: type[IO.ComfyNode], spec: _ModelSpec, images: list[Input.Image] +) -> list[OpenRouterImageContent]: + urls = await upload_images_to_comfyapi( + cls, + images, + max_images=spec.max_images, + total_pixels=2048 * 2048, + mime_type="image/png", + wait_label="Uploading reference images", + ) + return [OpenRouterImageContent(image_url=OpenRouterImageUrl(url=url)) for url in urls] + + +async def _build_video_blocks(cls: type[IO.ComfyNode], videos: list[Input.Video]) -> list[OpenRouterVideoContent]: + blocks: list[OpenRouterVideoContent] = [] + total = len(videos) + for idx, video in enumerate(videos): + label = "Uploading reference video" + if total > 1: + label = f"{label} ({idx + 1}/{total})" + url = await upload_video_to_comfyapi(cls, video, wait_label=label) + blocks.append(OpenRouterVideoContent(video_url=OpenRouterVideoUrl(url=url))) + return blocks + + +def _user_message(prompt: str, media_blocks: list[OpenRouterContentBlock]) -> OpenRouterMessage: + if not media_blocks: + return OpenRouterMessage(role="user", content=prompt) + blocks: list[OpenRouterContentBlock] = list(media_blocks) + blocks.append(OpenRouterTextContent(text=prompt)) + return OpenRouterMessage(role="user", content=blocks) + + +def _build_messages( + system_prompt: str, prompt: str, media_blocks: list[OpenRouterContentBlock] +) -> list[OpenRouterMessage]: + messages: list[OpenRouterMessage] = [] + if system_prompt: + messages.append(OpenRouterMessage(role="system", content=system_prompt)) + messages.append(_user_message(prompt, media_blocks)) + return messages + + +def _build_request( + slug: str, + system_prompt: str, + prompt: str, + media_blocks: list[OpenRouterContentBlock], + *, + seed: int, + reasoning_effort: str | None, + search_context_size: str | None, +) -> OpenRouterChatRequest: + reasoning_cfg: OpenRouterReasoningConfig | None = None + if reasoning_effort and reasoning_effort != "off": + # exclude=True asks providers to reason internally but not return the trace + reasoning_cfg = OpenRouterReasoningConfig(effort=reasoning_effort, exclude=True) + web_search_cfg: OpenRouterWebSearchOptions | None = None + if search_context_size: + web_search_cfg = OpenRouterWebSearchOptions(search_context_size=search_context_size) + return OpenRouterChatRequest( + model=slug, + messages=_build_messages(system_prompt, prompt, media_blocks), + seed=seed if seed > 0 else None, + reasoning=reasoning_cfg, + web_search_options=web_search_cfg, + ) + + +def _extract_text(response: OpenRouterChatResponse) -> str: + if response.error: + code = response.error.code if response.error.code is not None else "unknown" + raise ValueError(f"OpenRouter error ({code}): {response.error.message or 'no message'}") + if not response.choices: + raise ValueError("Empty response from OpenRouter (no choices).") + message = response.choices[0].message + if not message: + raise ValueError("Empty response from OpenRouter (no message).") + if message.refusal: + raise ValueError(f"Model refused to respond: {message.refusal}") + return message.content or "" + + +class OpenRouterLLMNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="OpenRouterLLMNode", + display_name="OpenRouter LLM", + category="partner/text/OpenRouter", + essentials_category="Text Generation", + description=( + "Generate text responses through OpenRouter. Routes to a curated set of popular " + "models from xAI, DeepSeek, Qwen, Mistral, Z.AI (GLM), Moonshot (Kimi), and " + "Perplexity Sonar." + ), + inputs=[ + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Text input to the model.", + ), + IO.DynamicCombo.Input( + "model", + options=_build_model_options(), + tooltip="The OpenRouter model used to generate the response.", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + control_after_generate=True, + tooltip="Seed for sampling. Set to 0 to omit. Most models treat this as a hint only.", + ), + IO.String.Input( + "system_prompt", + multiline=True, + default="", + optional=True, + advanced=True, + tooltip="Foundational instructions that dictate the model's behavior.", + ), + ], + outputs=[IO.String.Output()], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model"]), + expr=_price_badge_jsonata(), + ), + ) + + @classmethod + async def execute( + cls, + prompt: str, + model: dict, + seed: int, + system_prompt: str = "", + ) -> IO.NodeOutput: + validate_string(prompt, strip_whitespace=True, min_length=1) + slug: str = model["model"] + spec = _MODELS_BY_SLUG.get(slug) + if spec is None: + raise ValueError(f"Unknown OpenRouter model: {slug}") + + reasoning_effort: str | None = model.get("reasoning_effort") + search_context_size: str | None = model.get("search_context_size") + + image_tensors: list[Input.Image] = [t for t in (model.get("images") or {}).values() if t is not None] + if image_tensors and sum(get_number_of_images(t) for t in image_tensors) > spec.max_images: + raise ValueError(f"Up to {spec.max_images} images are supported for {slug}.") + video_inputs: list[Input.Video] = [v for v in (model.get("videos") or {}).values() if v is not None] + if video_inputs and len(video_inputs) > spec.max_videos: + raise ValueError(f"Up to {spec.max_videos} videos are supported for {slug}.") + + media_blocks: list[OpenRouterContentBlock] = [] + if image_tensors: + media_blocks.extend(await _build_image_blocks(cls, spec, image_tensors)) + if video_inputs: + media_blocks.extend(await _build_video_blocks(cls, video_inputs)) + + request = _build_request( + slug, + system_prompt, + prompt, + media_blocks, + seed=seed, + reasoning_effort=reasoning_effort, + search_context_size=search_context_size, + ) + + response = await sync_op( + cls, + ApiEndpoint(path=OPENROUTER_CHAT_ENDPOINT, method="POST"), + response_model=OpenRouterChatResponse, + data=request, + price_extractor=_calculate_price, + ) + return IO.NodeOutput(_extract_text(response)) + + +class OpenRouterExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [OpenRouterLLMNode] + + +async def comfy_entrypoint() -> OpenRouterExtension: + return OpenRouterExtension() diff --git a/comfy_api_nodes/nodes_pixverse.py b/comfy_api_nodes/nodes_pixverse.py index e17a24ae7..4c8b723b9 100644 --- a/comfy_api_nodes/nodes_pixverse.py +++ b/comfy_api_nodes/nodes_pixverse.py @@ -53,7 +53,7 @@ class PixverseTemplateNode(IO.ComfyNode): return IO.Schema( node_id="PixverseTemplateNode", display_name="PixVerse Template", - category="api node/video/PixVerse", + category="partner/video/PixVerse", inputs=[ IO.Combo.Input("template", options=list(pixverse_templates.keys())), ], @@ -74,7 +74,7 @@ class PixverseTextToVideoNode(IO.ComfyNode): return IO.Schema( node_id="PixverseTextToVideoNode", display_name="PixVerse Text to Video", - category="api node/video/PixVerse", + category="partner/video/PixVerse", description="Generates videos based on prompt and output_size.", inputs=[ IO.String.Input( @@ -192,7 +192,7 @@ class PixverseImageToVideoNode(IO.ComfyNode): return IO.Schema( node_id="PixverseImageToVideoNode", display_name="PixVerse Image to Video", - category="api node/video/PixVerse", + category="partner/video/PixVerse", description="Generates videos based on prompt and output_size.", inputs=[ IO.Image.Input("image"), @@ -310,7 +310,7 @@ class PixverseTransitionVideoNode(IO.ComfyNode): return IO.Schema( node_id="PixverseTransitionVideoNode", display_name="PixVerse Transition Video", - category="api node/video/PixVerse", + category="partner/video/PixVerse", description="Generates videos based on prompt and output_size.", inputs=[ IO.Image.Input("first_frame"), diff --git a/comfy_api_nodes/nodes_quiver.py b/comfy_api_nodes/nodes_quiver.py index 3269c0afe..34929fa0c 100644 --- a/comfy_api_nodes/nodes_quiver.py +++ b/comfy_api_nodes/nodes_quiver.py @@ -62,7 +62,7 @@ class QuiverTextToSVGNode(IO.ComfyNode): return IO.Schema( node_id="QuiverTextToSVGNode", display_name="Quiver Text to SVG", - category="api node/image/Quiver", + category="partner/image/Quiver", description="Generate an SVG from a text prompt using Quiver AI.", inputs=[ IO.String.Input( @@ -177,7 +177,7 @@ class QuiverImageToSVGNode(IO.ComfyNode): return IO.Schema( node_id="QuiverImageToSVGNode", display_name="Quiver Image to SVG", - category="api node/image/Quiver", + category="partner/image/Quiver", description="Vectorize a raster image into SVG using Quiver AI.", inputs=[ IO.Image.Input( diff --git a/comfy_api_nodes/nodes_recraft.py b/comfy_api_nodes/nodes_recraft.py index c60cfbc4a..c44942f50 100644 --- a/comfy_api_nodes/nodes_recraft.py +++ b/comfy_api_nodes/nodes_recraft.py @@ -178,7 +178,7 @@ class RecraftColorRGBNode(IO.ComfyNode): return IO.Schema( node_id="RecraftColorRGB", display_name="Recraft Color RGB", - category="api node/image/Recraft", + category="partner/image/Recraft", description="Create Recraft Color by choosing specific RGB values.", inputs=[ IO.Int.Input("r", default=0, min=0, max=255, tooltip="Red value of color."), @@ -204,7 +204,7 @@ class RecraftControlsNode(IO.ComfyNode): return IO.Schema( node_id="RecraftControls", display_name="Recraft Controls", - category="api node/image/Recraft", + category="partner/image/Recraft", description="Create Recraft Controls for customizing Recraft generation.", inputs=[ IO.Custom(RecraftIO.COLOR).Input("colors", optional=True), @@ -228,7 +228,7 @@ class RecraftStyleV3RealisticImageNode(IO.ComfyNode): return IO.Schema( node_id="RecraftStyleV3RealisticImage", display_name="Recraft Style - Realistic Image", - category="api node/image/Recraft", + category="partner/image/Recraft", description="Select realistic_image style and optional substyle.", inputs=[ IO.Combo.Input("substyle", options=get_v3_substyles(cls.RECRAFT_STYLE)), @@ -253,7 +253,7 @@ class RecraftStyleV3DigitalIllustrationNode(RecraftStyleV3RealisticImageNode): return IO.Schema( node_id="RecraftStyleV3DigitalIllustration", display_name="Recraft Style - Digital Illustration", - category="api node/image/Recraft", + category="partner/image/Recraft", description="Select realistic_image style and optional substyle.", inputs=[ IO.Combo.Input("substyle", options=get_v3_substyles(cls.RECRAFT_STYLE)), @@ -272,7 +272,7 @@ class RecraftStyleV3VectorIllustrationNode(RecraftStyleV3RealisticImageNode): return IO.Schema( node_id="RecraftStyleV3VectorIllustrationNode", display_name="Recraft Style - Realistic Image", - category="api node/image/Recraft", + category="partner/image/Recraft", description="Select realistic_image style and optional substyle.", inputs=[ IO.Combo.Input("substyle", options=get_v3_substyles(cls.RECRAFT_STYLE)), @@ -291,7 +291,7 @@ class RecraftStyleV3LogoRasterNode(RecraftStyleV3RealisticImageNode): return IO.Schema( node_id="RecraftStyleV3LogoRaster", display_name="Recraft Style - Logo Raster", - category="api node/image/Recraft", + category="partner/image/Recraft", description="Select realistic_image style and optional substyle.", inputs=[ IO.Combo.Input("substyle", options=get_v3_substyles(cls.RECRAFT_STYLE, include_none=False)), @@ -308,7 +308,7 @@ class RecraftStyleInfiniteStyleLibrary(IO.ComfyNode): return IO.Schema( node_id="RecraftStyleV3InfiniteStyleLibrary", display_name="Recraft Style - Infinite Style Library", - category="api node/image/Recraft", + category="partner/image/Recraft", description="Choose style based on preexisting UUID from Recraft's Infinite Style Library.", inputs=[ IO.String.Input("style_id", default="", tooltip="UUID of style from Infinite Style Library."), @@ -331,7 +331,7 @@ class RecraftCreateStyleNode(IO.ComfyNode): return IO.Schema( node_id="RecraftCreateStyleNode", display_name="Recraft Create Style", - category="api node/image/Recraft", + category="partner/image/Recraft", description="Create a custom style from reference images. " "Upload 1-5 images to use as style references. " "Total size of all images is limited to 5 MB.", @@ -400,7 +400,7 @@ class RecraftTextToImageNode(IO.ComfyNode): return IO.Schema( node_id="RecraftTextToImageNode", display_name="Recraft Text to Image", - category="api node/image/Recraft", + category="partner/image/Recraft", description="Generates images synchronously based on prompt and resolution.", inputs=[ IO.String.Input("prompt", multiline=True, default="", tooltip="Prompt for the image generation."), @@ -512,7 +512,7 @@ class RecraftImageToImageNode(IO.ComfyNode): return IO.Schema( node_id="RecraftImageToImageNode", display_name="Recraft Image to Image", - category="api node/image/Recraft", + category="partner/image/Recraft", description="Modify image based on prompt and strength.", inputs=[ IO.Image.Input("image"), @@ -630,7 +630,7 @@ class RecraftImageInpaintingNode(IO.ComfyNode): return IO.Schema( node_id="RecraftImageInpaintingNode", display_name="Recraft Image Inpainting", - category="api node/image/Recraft", + category="partner/image/Recraft", description="Modify image based on prompt and mask.", inputs=[ IO.Image.Input("image"), @@ -732,7 +732,7 @@ class RecraftTextToVectorNode(IO.ComfyNode): return IO.Schema( node_id="RecraftTextToVectorNode", display_name="Recraft Text to Vector", - category="api node/image/Recraft", + category="partner/image/Recraft", description="Generates SVG synchronously based on prompt and resolution.", inputs=[ IO.String.Input("prompt", default="", tooltip="Prompt for the image generation.", multiline=True), @@ -832,7 +832,7 @@ class RecraftVectorizeImageNode(IO.ComfyNode): return IO.Schema( node_id="RecraftVectorizeImageNode", display_name="Recraft Vectorize Image", - category="api node/image/Recraft", + category="partner/image/Recraft", essentials_category="Image Tools", description="Generates SVG synchronously from an input image.", inputs=[ @@ -876,7 +876,7 @@ class RecraftReplaceBackgroundNode(IO.ComfyNode): return IO.Schema( node_id="RecraftReplaceBackgroundNode", display_name="Recraft Replace Background", - category="api node/image/Recraft", + category="partner/image/Recraft", description="Replace background on image, based on provided prompt.", inputs=[ IO.Image.Input("image"), @@ -963,7 +963,7 @@ class RecraftRemoveBackgroundNode(IO.ComfyNode): return IO.Schema( node_id="RecraftRemoveBackgroundNode", display_name="Recraft Remove Background", - category="api node/image/Recraft", + category="partner/image/Recraft", essentials_category="Image Tools", description="Remove background from image, and return processed image and mask.", inputs=[ @@ -1012,7 +1012,7 @@ class RecraftCrispUpscaleNode(IO.ComfyNode): return IO.Schema( node_id="RecraftCrispUpscaleNode", display_name="Recraft Crisp Upscale Image", - category="api node/image/Recraft", + category="partner/image/Recraft", description="Upscale image synchronously.\n" "Enhances a given raster image using ‘crisp upscale’ tool, " "increasing image resolution, making the image sharper and cleaner.", @@ -1058,7 +1058,7 @@ class RecraftCreativeUpscaleNode(RecraftCrispUpscaleNode): return IO.Schema( node_id="RecraftCreativeUpscaleNode", display_name="Recraft Creative Upscale Image", - category="api node/image/Recraft", + category="partner/image/Recraft", description="Upscale image synchronously.\n" "Enhances a given raster image using ‘creative upscale’ tool, " "boosting resolution with a focus on refining small details and faces.", @@ -1086,7 +1086,7 @@ class RecraftV4TextToImageNode(IO.ComfyNode): return IO.Schema( node_id="RecraftV4TextToImageNode", display_name="Recraft V4 Text to Image", - category="api node/image/Recraft", + category="partner/image/Recraft", description="Generates images using Recraft V4 or V4 Pro models.", inputs=[ IO.String.Input( @@ -1210,7 +1210,7 @@ class RecraftV4TextToVectorNode(IO.ComfyNode): return IO.Schema( node_id="RecraftV4TextToVectorNode", display_name="Recraft V4 Text to Vector", - category="api node/image/Recraft", + category="partner/image/Recraft", description="Generates SVG using Recraft V4 or V4 Pro models.", inputs=[ IO.String.Input( diff --git a/comfy_api_nodes/nodes_reve.py b/comfy_api_nodes/nodes_reve.py index a87395394..177349a8b 100644 --- a/comfy_api_nodes/nodes_reve.py +++ b/comfy_api_nodes/nodes_reve.py @@ -109,7 +109,7 @@ class ReveImageCreateNode(IO.ComfyNode): return IO.Schema( node_id="ReveImageCreateNode", display_name="Reve Image Create", - category="api node/image/Reve", + category="partner/image/Reve", description="Generate images from text descriptions using Reve.", inputs=[ IO.String.Input( @@ -200,7 +200,7 @@ class ReveImageEditNode(IO.ComfyNode): return IO.Schema( node_id="ReveImageEditNode", display_name="Reve Image Edit", - category="api node/image/Reve", + category="partner/image/Reve", description="Edit images using natural language instructions with Reve.", inputs=[ IO.Image.Input("image", tooltip="The image to edit."), @@ -300,7 +300,7 @@ class ReveImageRemixNode(IO.ComfyNode): return IO.Schema( node_id="ReveImageRemixNode", display_name="Reve Image Remix", - category="api node/image/Reve", + category="partner/image/Reve", description="Combine reference images with text prompts to create new images using Reve.", inputs=[ IO.Autogrow.Input( diff --git a/comfy_api_nodes/nodes_rodin.py b/comfy_api_nodes/nodes_rodin.py index 2b829b8db..0375a2123 100644 --- a/comfy_api_nodes/nodes_rodin.py +++ b/comfy_api_nodes/nodes_rodin.py @@ -5,32 +5,37 @@ Rodin API docs: https://developer.hyper3d.ai/ """ -from inspect import cleandoc -import folder_paths as comfy_paths -import os import logging import math +import os +from inspect import cleandoc from io import BytesIO -from typing_extensions import override +from typing import Any + +import aiohttp from PIL import Image +from typing_extensions import override + +import folder_paths as comfy_paths +from comfy_api.latest import IO, ComfyExtension, Types from comfy_api_nodes.apis.rodin import ( - Rodin3DGenerateRequest, - Rodin3DGenerateResponse, + JobStatus, Rodin3DCheckStatusRequest, Rodin3DCheckStatusResponse, Rodin3DDownloadRequest, Rodin3DDownloadResponse, - JobStatus, + Rodin3DGen25Request, + Rodin3DGenerateRequest, + Rodin3DGenerateResponse, ) from comfy_api_nodes.util import ( - sync_op, - poll_op, ApiEndpoint, download_url_to_bytesio, download_url_to_file_3d, + poll_op, + sync_op, + validate_string, ) -from comfy_api.latest import ComfyExtension, IO, Types - COMMON_PARAMETERS = [ IO.Int.Input( @@ -51,40 +56,30 @@ COMMON_PARAMETERS = [ ] -def get_quality_mode(poly_count): - polycount = poly_count.split("-") - poly = polycount[1] - count = polycount[0] - if poly == "Triangle": - mesh_mode = "Raw" - elif poly == "Quad": - mesh_mode = "Quad" - else: - mesh_mode = "Quad" - - if count == "4K": - quality_override = 4000 - elif count == "8K": - quality_override = 8000 - elif count == "18K": - quality_override = 18000 - elif count == "50K": - quality_override = 50000 - elif count == "2K": - quality_override = 2000 - elif count == "20K": - quality_override = 20000 - elif count == "150K": - quality_override = 150000 - elif count == "500K": - quality_override = 500000 - else: - quality_override = 18000 - - return mesh_mode, quality_override +_QUALITY_MESH_OPTIONS: dict[str, tuple[str, int]] = { + "4K-Quad": ("Quad", 4000), + "8K-Quad": ("Quad", 8000), + "18K-Quad": ("Quad", 18000), + "50K-Quad": ("Quad", 50000), + "200K-Quad": ("Quad", 200000), + "2K-Triangle": ("Raw", 2000), + "20K-Triangle": ("Raw", 20000), + "150K-Triangle": ("Raw", 150000), + "200K-Triangle": ("Raw", 200000), + "500K-Triangle": ("Raw", 500000), + "1M-Triangle": ("Raw", 1000000), +} -def tensor_to_filelike(tensor, max_pixels: int = 2048*2048): +def get_quality_mode(poly_count: str) -> tuple[str, int]: + """Map a polygon-count preset like '18K-Quad' to (mesh_mode, quality_override). + + Falls back to ('Quad', 18000) for unknown labels; legacy parity. + """ + return _QUALITY_MESH_OPTIONS.get(poly_count, ("Quad", 18000)) + + +def tensor_to_filelike(tensor, max_pixels: int = 2048 * 2048): """ Converts a PyTorch tensor to a file-like object. @@ -96,8 +91,8 @@ def tensor_to_filelike(tensor, max_pixels: int = 2048*2048): - io.BytesIO: A file-like object containing the image data. """ array = tensor.cpu().numpy() - array = (array * 255).astype('uint8') - image = Image.fromarray(array, 'RGB') + array = (array * 255).astype("uint8") + image = Image.fromarray(array, "RGB") original_width, original_height = image.size original_pixels = original_width * original_height @@ -112,7 +107,7 @@ def tensor_to_filelike(tensor, max_pixels: int = 2048*2048): image = image.resize((new_width, new_height), Image.Resampling.LANCZOS) img_byte_arr = BytesIO() - image.save(img_byte_arr, format='PNG') # PNG is used for lossless compression + image.save(img_byte_arr, format="PNG") # PNG is used for lossless compression img_byte_arr.seek(0) return img_byte_arr @@ -145,11 +140,9 @@ async def create_generate_task( TAPose=ta_pose, ), files=[ - ( - "images", - open(image, "rb") if isinstance(image, str) else tensor_to_filelike(image) - ) - for image in images if image is not None + ("images", open(image, "rb") if isinstance(image, str) else tensor_to_filelike(image)) + for image in images + if image is not None ], content_type="multipart/form-data", ) @@ -177,6 +170,7 @@ def check_rodin_status(response: Rodin3DCheckStatusResponse) -> str: return "DONE" return "Generating" + def extract_progress(response: Rodin3DCheckStatusResponse) -> int | None: if not response.jobs: return None @@ -214,7 +208,7 @@ async def download_files(url_list, task_uuid: str) -> tuple[str | None, Types.Fi model_file_path = None file_3d = None - for i in url_list.list: + for i in url_list.items: file_path = os.path.join(save_path, i.name) if i.name.lower().endswith(".glb"): model_file_path = os.path.join(result_folder_name, i.name) @@ -236,7 +230,7 @@ class Rodin3D_Regular(IO.ComfyNode): return IO.Schema( node_id="Rodin3D_Regular", display_name="Rodin 3D Generate - Regular Generate", - category="api node/3d/Rodin", + category="partner/3d/Rodin", description=cleandoc(cls.__doc__ or ""), inputs=[ IO.Image.Input("Images"), @@ -295,7 +289,7 @@ class Rodin3D_Detail(IO.ComfyNode): return IO.Schema( node_id="Rodin3D_Detail", display_name="Rodin 3D Generate - Detail Generate", - category="api node/3d/Rodin", + category="partner/3d/Rodin", description=cleandoc(cls.__doc__ or ""), inputs=[ IO.Image.Input("Images"), @@ -354,7 +348,7 @@ class Rodin3D_Smooth(IO.ComfyNode): return IO.Schema( node_id="Rodin3D_Smooth", display_name="Rodin 3D Generate - Smooth Generate", - category="api node/3d/Rodin", + category="partner/3d/Rodin", description=cleandoc(cls.__doc__ or ""), inputs=[ IO.Image.Input("Images"), @@ -412,7 +406,7 @@ class Rodin3D_Sketch(IO.ComfyNode): return IO.Schema( node_id="Rodin3D_Sketch", display_name="Rodin 3D Generate - Sketch Generate", - category="api node/3d/Rodin", + category="partner/3d/Rodin", description=cleandoc(cls.__doc__ or ""), inputs=[ IO.Image.Input("Images"), @@ -474,7 +468,7 @@ class Rodin3D_Gen2(IO.ComfyNode): return IO.Schema( node_id="Rodin3D_Gen2", display_name="Rodin 3D Generate - Gen-2 Generate", - category="api node/3d/Rodin", + category="partner/3d/Rodin", description=cleandoc(cls.__doc__ or ""), inputs=[ IO.Image.Input("Images"), @@ -489,7 +483,16 @@ class Rodin3D_Gen2(IO.ComfyNode): IO.Combo.Input("Material_Type", options=["PBR", "Shaded"], default="PBR", optional=True), IO.Combo.Input( "Polygon_count", - options=["4K-Quad", "8K-Quad", "18K-Quad", "50K-Quad", "2K-Triangle", "20K-Triangle", "150K-Triangle", "500K-Triangle"], + options=[ + "4K-Quad", + "8K-Quad", + "18K-Quad", + "50K-Quad", + "2K-Triangle", + "20K-Triangle", + "150K-Triangle", + "500K-Triangle", + ], default="500K-Triangle", optional=True, ), @@ -542,6 +545,566 @@ class Rodin3D_Gen2(IO.ComfyNode): return IO.NodeOutput(model_path, file_3d) +def _rodin_multipart_parser(data: dict[str, Any]) -> aiohttp.FormData: + """Convert a Rodin request dict to an aiohttp form, fixing bool/list serialization. + + Booleans --> "true"/"false". Lists --> one field per element. + """ + form = aiohttp.FormData(default_to_multipart=True) + for key, value in data.items(): + if value is None: + continue + if isinstance(value, bool): + form.add_field(key, "true" if value else "false") + elif isinstance(value, list): + for item in value: + form.add_field(key, str(item)) + elif isinstance(value, (bytes, bytearray)): + form.add_field(key, value) + else: + form.add_field(key, str(value)) + return form + + +async def _create_gen25_task( + cls: type[IO.ComfyNode], + request: Rodin3DGen25Request, + images: list | None, +) -> tuple[str, str]: + """Submit a Gen-2.5 generate job; returns (task_uuid, subscription_key).""" + + if images is not None and len(images) > 5: + raise ValueError("Rodin Gen-2.5 supports at most 5 input images.") + + files = None + if images: + files = [ + ( + "images", + open(image, "rb") if isinstance(image, str) else tensor_to_filelike(image), + ) + for image in images + if image is not None + ] + + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/rodin/api/v2/rodin", method="POST"), + response_model=Rodin3DGenerateResponse, + data=request, + files=files, + content_type="multipart/form-data", + multipart_parser=_rodin_multipart_parser, + ) + + if not response.uuid or not response.jobs or not response.jobs.subscription_key: + raise RuntimeError(f"Rodin Gen-2.5 submit failed: message={response.message!r}") + return response.uuid, response.jobs.subscription_key + + +_PREVIEWABLE_3D_EXTS = {".glb", ".obj", ".fbx", ".stl", ".gltf"} + + +async def _download_gen25_files( + download_list: Rodin3DDownloadResponse, + task_uuid: str, + geometry_file_format: str, +) -> Types.File3D | None: + """Download every file in the list; return the File3D matching the chosen format.""" + + folder_name = f"Rodin3D_Gen25_{task_uuid}" + save_dir = os.path.join(comfy_paths.get_output_directory(), folder_name) + os.makedirs(save_dir, exist_ok=True) + + target_ext = f".{geometry_file_format.lower().lstrip('.')}" + file_3d: Types.File3D | None = None + + for item in download_list.items: + file_path = os.path.join(save_dir, item.name) + ext = os.path.splitext(item.name.lower())[1] + # Prefer the file matching the user's chosen format; fall back below. + if file_3d is None and ext == target_ext and ext in _PREVIEWABLE_3D_EXTS: + file_3d = await download_url_to_file_3d(item.url, target_ext.lstrip(".")) + with open(file_path, "wb") as f: + f.write(file_3d.get_bytes()) + continue + await download_url_to_bytesio(item.url, file_path) + + # If the chosen format wasn't found, surface any model file we did get. + if file_3d is None: + for item in download_list.items: + ext = os.path.splitext(item.name.lower())[1] + if ext in _PREVIEWABLE_3D_EXTS: + file_3d = await download_url_to_file_3d(item.url, ext.lstrip(".")) + break + return file_3d + + +_MODE_REGULAR = "Regular" +_MODE_FAST = "Fast" +_MODE_EXTREME_HIGH = "Extreme-High" + +_REGULAR_POLY_OPTIONS = [ + "Default", + "4K-Quad", + "8K-Quad", + "18K-Quad", + "50K-Quad", + "2K-Triangle", + "20K-Triangle", + "150K-Triangle", + "500K-Triangle", + "1M-Triangle", +] + +_TEXTURE_MODE_OPTIONS = ["Default", "legacy", "extreme-low", "low", "medium", "high"] +_GEOMETRY_FORMAT_OPTIONS = ["glb", "fbx", "obj", "stl"] +_MATERIAL_OPTIONS = ["PBR", "Shaded", "All", "None"] + + +def _build_mode_input(name: str = "mode") -> IO.DynamicCombo.Input: + return IO.DynamicCombo.Input( + name, + options=[ + IO.DynamicCombo.Option( + _MODE_REGULAR, + [ + IO.Combo.Input( + "tier", + options=["Gen-2.5-Low", "Gen-2.5-Medium", "Gen-2.5-High"], + default="Gen-2.5-High", + tooltip="Quality tier. Higher tiers produce higher-fidelity geometry.", + ), + IO.Combo.Input( + "polygon_count", + options=_REGULAR_POLY_OPTIONS, + default="Default", + tooltip="Preset face count. 'Default' uses the server's default for the selected tier.", + ), + IO.Boolean.Input( + "creative", + default=False, + tooltip="Creative mode (Medium/High only). Enhances generative robustness.", + ), + ], + ), + IO.DynamicCombo.Option( + _MODE_FAST, + [ + IO.Combo.Input( + "tier", + options=[ + "Gen-2.5-Extreme-Low", + "Gen-2.5-Low", + "Gen-2.5-Medium", + "Gen-2.5-High", + ], + default="Gen-2.5-Low", + ), + IO.Int.Input( + "mesh_faces", + default=20000, + min=1000, + max=20000, + display_mode=IO.NumberDisplay.number, + tooltip="Mesh face count (1K-20K in Fast mode).", + ), + ], + ), + IO.DynamicCombo.Option( + _MODE_EXTREME_HIGH, + [ + IO.Combo.Input("mesh_mode", options=["Raw", "Quad"], default="Raw"), + IO.Int.Input( + "mesh_faces", + default=1000000, + min=20000, + max=2000000, + display_mode=IO.NumberDisplay.number, + tooltip=( + "Mesh face count. Raw mode: 20K-2M. " + "Quad mode: keep under 200K (upstream may reject higher values)." + ), + ), + IO.Boolean.Input( + "is_micro", + default=False, + tooltip="Enable micro detail (Extreme-High only).", + ), + IO.Boolean.Input( + "creative", + default=False, + tooltip="Creative mode. Enhances generative robustness.", + ), + ], + ), + ], + tooltip=( + "Generation mode. Regular = balanced. Fast = 1K-20K faces for rapid prototyping. " + "Extreme-High = 20K-2M faces with optional micro details." + ), + ) + + +def _build_common_inputs(*, include_image_only: bool) -> list: + inputs: list = [ + IO.Combo.Input("material", options=_MATERIAL_OPTIONS, default="Shaded"), + IO.Combo.Input("geometry_file_format", options=_GEOMETRY_FORMAT_OPTIONS, default="glb"), + IO.Combo.Input( + "texture_mode", + options=_TEXTURE_MODE_OPTIONS, + default="Default", + optional=True, + tooltip="Texture quality preset. 'Default' uses the server's default for the selected tier.", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=65535, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + optional=True, + ), + IO.Boolean.Input( + "TAPose", default=False, optional=True, advanced=True, tooltip="T/A pose for human-like models." + ), + IO.Boolean.Input( + "hd_texture", default=False, optional=True, advanced=True, tooltip="High-quality texture enhancement." + ), + IO.Boolean.Input( + "texture_delight", + default=False, + optional=True, + advanced=True, + tooltip="Remove baked lighting from textures.", + ), + ] + if include_image_only: + inputs.append( + IO.Boolean.Input( + "use_original_alpha", + default=False, + optional=True, + advanced=True, + tooltip="Preserve image transparency.", + ) + ) + inputs.extend( + [ + IO.Boolean.Input( + "addon_highpack", + default=False, + optional=True, + advanced=True, + tooltip="HighPack addon: 4K textures and ~16x faces in Quad mode.", + ), + IO.Int.Input( + "bbox_width", + default=0, + min=0, + max=300, + display_mode=IO.NumberDisplay.number, + optional=True, + advanced=True, + tooltip="Bounding-box width (Y axis). Set to 0 with the others to skip bbox.", + ), + IO.Int.Input( + "bbox_height", + default=0, + min=0, + max=300, + display_mode=IO.NumberDisplay.number, + optional=True, + advanced=True, + tooltip="Bounding-box height (Z axis).", + ), + IO.Int.Input( + "bbox_length", + default=0, + min=0, + max=300, + display_mode=IO.NumberDisplay.number, + optional=True, + advanced=True, + tooltip="Bounding-box length (X axis).", + ), + IO.Int.Input( + "height_cm", + default=0, + min=0, + max=10000, + display_mode=IO.NumberDisplay.number, + optional=True, + advanced=True, + tooltip="Approximate model height in centimeters (0 to skip).", + ), + ] + ) + return inputs + + +_PRICE_EXPR = """ +( + $baseCredits := widgets.mode = "extreme-high" ? 1.0 : 0.5; + $addonCredits := widgets.addon_highpack ? 1.0 : 0.0; + $total := ($baseCredits * 1.5) + ($addonCredits * 0.8); + {"type":"usd","usd": $total} +) +""" + + +def _resolve_mode_params(mode_input: dict) -> dict: + """Translate the DynamicCombo `mode` payload into Gen-2.5 request fields. + + Returns a dict with: tier, quality_override, mesh_mode, geometry_instruct_mode, is_micro. + Missing keys mean "do not send" (so we don't override server defaults). + """ + selected = mode_input["mode"] + out: dict = {} + + if selected == _MODE_REGULAR: + out["tier"] = mode_input["tier"] + polygon = mode_input.get("polygon_count", "Default") + if polygon != "Default": + mesh_mode, faces = get_quality_mode(polygon) + out["mesh_mode"] = mesh_mode + out["quality_override"] = faces + if mode_input.get("creative"): + out["geometry_instruct_mode"] = "creative" + + elif selected == _MODE_FAST: + out["tier"] = mode_input["tier"] + out["mesh_mode"] = "Raw" + out["quality_override"] = int(mode_input["mesh_faces"]) + + elif selected == _MODE_EXTREME_HIGH: + out["tier"] = "Gen-2.5-Extreme-High" + out["mesh_mode"] = mode_input["mesh_mode"] + out["quality_override"] = int(mode_input["mesh_faces"]) + if mode_input.get("is_micro"): + out["is_micro"] = True + if mode_input.get("creative"): + out["geometry_instruct_mode"] = "creative" + return out + + +def _build_request( + *, + mode_input: dict, + material: str, + geometry_file_format: str, + texture_mode: str, + seed: int, + TAPose: bool, + hd_texture: bool, + texture_delight: bool, + addon_highpack: bool, + bbox_width: int, + bbox_height: int, + bbox_length: int, + height_cm: int, + prompt: str | None = None, + use_original_alpha: bool = False, +) -> Rodin3DGen25Request: + mode_params = _resolve_mode_params(mode_input) + + bbox = None + if bbox_width and bbox_height and bbox_length: + bbox = [bbox_width, bbox_height, bbox_length] + + return Rodin3DGen25Request( + tier=mode_params["tier"], + prompt=prompt or None, + seed=seed, + material=material, + geometry_file_format=geometry_file_format, + texture_mode=None if texture_mode == "Default" else texture_mode, + mesh_mode=mode_params.get("mesh_mode"), + quality_override=mode_params.get("quality_override"), + geometry_instruct_mode=mode_params.get("geometry_instruct_mode"), + bbox_condition=bbox, + height=height_cm or None, + TAPose=TAPose or None, + hd_texture=hd_texture or None, + texture_delight=texture_delight or None, + is_micro=mode_params.get("is_micro"), + use_original_alpha=use_original_alpha or None, + addons=["HighPack"] if addon_highpack else None, + ) + + +class Rodin3D_Gen25_Image(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="Rodin3D_Gen25_Image", + display_name="Rodin 3D Gen-2.5 - Image to 3D", + category="partner/3d/Rodin", + description=( + "Generate a 3D model from 1-5 reference images via Rodin Gen-2.5. " + "Pick a mode (Fast / Regular / Extreme-High) to tune quality vs. cost." + ), + inputs=[ + IO.Autogrow.Input( + "images", + template=IO.Autogrow.TemplatePrefix(IO.Image.Input("image"), prefix="image", min=1, max=5), + tooltip="1-5 images. The first image is used for materials when multi-view.", + ), + _build_mode_input(), + *_build_common_inputs(include_image_only=True), + ], + outputs=[IO.File3DAny.Output(display_name="model_file")], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["mode", "addon_highpack"]), + expr=_PRICE_EXPR, + ), + ) + + @classmethod + async def execute( + cls, + images: IO.Autogrow.Type, + mode: dict, + material: str, + geometry_file_format: str, + texture_mode: str, + seed: int, + TAPose: bool, + hd_texture: bool, + texture_delight: bool, + use_original_alpha: bool, + addon_highpack: bool, + bbox_width: int, + bbox_height: int, + bbox_length: int, + height_cm: int, + ) -> IO.NodeOutput: + image_tensors = [img for img in images.values() if img is not None] + if not image_tensors: + raise ValueError("Rodin Gen-2.5 Image-to-3D requires at least one image.") + + # Flatten multi-image tensors into individual frames; the API accepts each as a separate part. + flat_images: list = [] + for tensor in image_tensors: + if hasattr(tensor, "shape") and len(tensor.shape) == 4: + for i in range(tensor.shape[0]): + flat_images.append(tensor[i]) + else: + flat_images.append(tensor) + + if len(flat_images) > 5: + raise ValueError(f"Rodin Gen-2.5 accepts at most 5 images; received {len(flat_images)}.") + + request = _build_request( + mode_input=mode, + material=material, + geometry_file_format=geometry_file_format, + texture_mode=texture_mode, + seed=seed, + TAPose=TAPose, + hd_texture=hd_texture, + texture_delight=texture_delight, + addon_highpack=addon_highpack, + bbox_width=bbox_width, + bbox_height=bbox_height, + bbox_length=bbox_length, + height_cm=height_cm, + prompt=None, + use_original_alpha=use_original_alpha, + ) + + task_uuid, subscription_key = await _create_gen25_task(cls, request, flat_images) + await poll_for_task_status(subscription_key, cls) + download_list = await get_rodin_download_list(task_uuid, cls) + file_3d = await _download_gen25_files(download_list, task_uuid, geometry_file_format) + return IO.NodeOutput(file_3d) + + +class Rodin3D_Gen25_Text(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="Rodin3D_Gen25_Text", + display_name="Rodin 3D Gen-2.5 - Text to 3D", + category="partner/3d/Rodin", + description=( + "Generate a 3D model from a text prompt via Rodin Gen-2.5. " + "Pick a mode (Fast / Regular / Extreme-High) to tune quality vs. cost." + ), + inputs=[ + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Text prompt for the 3D model.", + ), + _build_mode_input(), + *_build_common_inputs(include_image_only=False), + ], + outputs=[IO.File3DAny.Output(display_name="model_file")], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["mode", "addon_highpack"]), + expr=_PRICE_EXPR, + ), + ) + + @classmethod + async def execute( + cls, + prompt: str, + mode: dict, + material: str, + geometry_file_format: str, + texture_mode: str, + seed: int, + TAPose: bool, + hd_texture: bool, + texture_delight: bool, + addon_highpack: bool, + bbox_width: int, + bbox_height: int, + bbox_length: int, + height_cm: int, + ) -> IO.NodeOutput: + validate_string(prompt, field_name="prompt", min_length=1, max_length=2500) + request = _build_request( + mode_input=mode, + material=material, + geometry_file_format=geometry_file_format, + texture_mode=texture_mode, + seed=seed, + TAPose=TAPose, + hd_texture=hd_texture, + texture_delight=texture_delight, + addon_highpack=addon_highpack, + bbox_width=bbox_width, + bbox_height=bbox_height, + bbox_length=bbox_length, + height_cm=height_cm, + prompt=prompt, + ) + task_uuid, subscription_key = await _create_gen25_task(cls, request, images=None) + await poll_for_task_status(subscription_key, cls) + download_list = await get_rodin_download_list(task_uuid, cls) + file_3d = await _download_gen25_files(download_list, task_uuid, geometry_file_format) + return IO.NodeOutput(file_3d) + + class Rodin3DExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: @@ -551,6 +1114,8 @@ class Rodin3DExtension(ComfyExtension): Rodin3D_Smooth, Rodin3D_Sketch, Rodin3D_Gen2, + Rodin3D_Gen25_Image, + Rodin3D_Gen25_Text, ] diff --git a/comfy_api_nodes/nodes_runway.py b/comfy_api_nodes/nodes_runway.py index 573170ba2..b9c5c81a1 100644 --- a/comfy_api_nodes/nodes_runway.py +++ b/comfy_api_nodes/nodes_runway.py @@ -140,7 +140,7 @@ class RunwayImageToVideoNodeGen3a(IO.ComfyNode): return IO.Schema( node_id="RunwayImageToVideoNodeGen3a", display_name="Runway Image to Video (Gen3a Turbo)", - category="api node/video/Runway", + category="partner/video/Runway", description="Generate a video from a single starting frame using Gen3a Turbo model. " "Before diving in, review these best practices to ensure that " "your input selections will set your generation up for success: " @@ -234,7 +234,7 @@ class RunwayImageToVideoNodeGen4(IO.ComfyNode): return IO.Schema( node_id="RunwayImageToVideoNodeGen4", display_name="Runway Image to Video (Gen4 Turbo)", - category="api node/video/Runway", + category="partner/video/Runway", description="Generate a video from a single starting frame using Gen4 Turbo model. " "Before diving in, review these best practices to ensure that " "your input selections will set your generation up for success: " @@ -329,7 +329,7 @@ class RunwayFirstLastFrameNode(IO.ComfyNode): return IO.Schema( node_id="RunwayFirstLastFrameNode", display_name="Runway First-Last-Frame to Video", - category="api node/video/Runway", + category="partner/video/Runway", description="Upload first and last keyframes, draft a prompt, and generate a video. " "More complex transitions, such as cases where the Last frame is completely different " "from the First frame, may benefit from the longer 10s duration. " @@ -440,7 +440,7 @@ class RunwayTextToImageNode(IO.ComfyNode): return IO.Schema( node_id="RunwayTextToImageNode", display_name="Runway Text to Image", - category="api node/image/Runway", + category="partner/image/Runway", description="Generate an image from a text prompt using Runway's Gen 4 model. " "You can also include reference image to guide the generation.", inputs=[ diff --git a/comfy_api_nodes/nodes_sonilo.py b/comfy_api_nodes/nodes_sonilo.py index 5518f5902..9ce896ed0 100644 --- a/comfy_api_nodes/nodes_sonilo.py +++ b/comfy_api_nodes/nodes_sonilo.py @@ -34,7 +34,7 @@ class SoniloVideoToMusic(IO.ComfyNode): return IO.Schema( node_id="SoniloVideoToMusic", display_name="Sonilo Video to Music", - category="api node/audio/Sonilo", + category="partner/audio/Sonilo", description="Generate music from video content using Sonilo's AI model. " "Analyzes the video and creates matching music.", inputs=[ @@ -99,7 +99,7 @@ class SoniloTextToMusic(IO.ComfyNode): return IO.Schema( node_id="SoniloTextToMusic", display_name="Sonilo Text to Music", - category="api node/audio/Sonilo", + category="partner/audio/Sonilo", description="Generate music from a text prompt using Sonilo's AI model. " "Leave duration at 0 to let the model infer it from the prompt.", inputs=[ diff --git a/comfy_api_nodes/nodes_sora.py b/comfy_api_nodes/nodes_sora.py index c1d485188..4ff1d649f 100644 --- a/comfy_api_nodes/nodes_sora.py +++ b/comfy_api_nodes/nodes_sora.py @@ -34,7 +34,7 @@ class OpenAIVideoSora2(IO.ComfyNode): return IO.Schema( node_id="OpenAIVideoSora2", display_name="OpenAI Sora - Video (DEPRECATED)", - category="api node/video/Sora", + category="partner/video/Sora", description=( "OpenAI video and audio generation.\n\n" "DEPRECATION NOTICE: OpenAI will stop serving the Sora v2 API in September 2026. " diff --git a/comfy_api_nodes/nodes_stability.py b/comfy_api_nodes/nodes_stability.py index 906d8ff35..9eaba173b 100644 --- a/comfy_api_nodes/nodes_stability.py +++ b/comfy_api_nodes/nodes_stability.py @@ -62,7 +62,7 @@ class StabilityStableImageUltraNode(IO.ComfyNode): return IO.Schema( node_id="StabilityStableImageUltraNode", display_name="Stability AI Stable Image Ultra", - category="api node/image/Stability AI", + category="partner/image/Stability AI", description=cleandoc(cls.__doc__ or ""), inputs=[ IO.String.Input( @@ -197,7 +197,7 @@ class StabilityStableImageSD_3_5Node(IO.ComfyNode): return IO.Schema( node_id="StabilityStableImageSD_3_5Node", display_name="Stability AI Stable Diffusion 3.5 Image", - category="api node/image/Stability AI", + category="partner/image/Stability AI", description=cleandoc(cls.__doc__ or ""), inputs=[ IO.String.Input( @@ -354,7 +354,7 @@ class StabilityUpscaleConservativeNode(IO.ComfyNode): return IO.Schema( node_id="StabilityUpscaleConservativeNode", display_name="Stability AI Upscale Conservative", - category="api node/image/Stability AI", + category="partner/image/Stability AI", description=cleandoc(cls.__doc__ or ""), inputs=[ IO.Image.Input("image"), @@ -457,7 +457,7 @@ class StabilityUpscaleCreativeNode(IO.ComfyNode): return IO.Schema( node_id="StabilityUpscaleCreativeNode", display_name="Stability AI Upscale Creative", - category="api node/image/Stability AI", + category="partner/image/Stability AI", description=cleandoc(cls.__doc__ or ""), inputs=[ IO.Image.Input("image"), @@ -578,7 +578,7 @@ class StabilityUpscaleFastNode(IO.ComfyNode): return IO.Schema( node_id="StabilityUpscaleFastNode", display_name="Stability AI Upscale Fast", - category="api node/image/Stability AI", + category="partner/image/Stability AI", description=cleandoc(cls.__doc__ or ""), inputs=[ IO.Image.Input("image"), @@ -630,7 +630,7 @@ class StabilityTextToAudio(IO.ComfyNode): return IO.Schema( node_id="StabilityTextToAudio", display_name="Stability AI Text To Audio", - category="api node/audio/Stability AI", + category="partner/audio/Stability AI", essentials_category="Audio", description=cleandoc(cls.__doc__ or ""), inputs=[ @@ -708,7 +708,7 @@ class StabilityAudioToAudio(IO.ComfyNode): return IO.Schema( node_id="StabilityAudioToAudio", display_name="Stability AI Audio To Audio", - category="api node/audio/Stability AI", + category="partner/audio/Stability AI", description=cleandoc(cls.__doc__ or ""), inputs=[ IO.Combo.Input( @@ -802,7 +802,7 @@ class StabilityAudioInpaint(IO.ComfyNode): return IO.Schema( node_id="StabilityAudioInpaint", display_name="Stability AI Audio Inpaint", - category="api node/audio/Stability AI", + category="partner/audio/Stability AI", description=cleandoc(cls.__doc__ or ""), inputs=[ IO.Combo.Input( diff --git a/comfy_api_nodes/nodes_topaz.py b/comfy_api_nodes/nodes_topaz.py index e79c16d3c..f7ef4cbf6 100644 --- a/comfy_api_nodes/nodes_topaz.py +++ b/comfy_api_nodes/nodes_topaz.py @@ -52,7 +52,7 @@ class TopazImageEnhance(IO.ComfyNode): return IO.Schema( node_id="TopazImageEnhance", display_name="Topaz Image Enhance", - category="api node/image/Topaz", + category="partner/image/Topaz", description="Industry-standard upscaling and image enhancement.", inputs=[ IO.Combo.Input("model", options=["Reimagine"]), @@ -235,7 +235,7 @@ class TopazVideoEnhance(IO.ComfyNode): return IO.Schema( node_id="TopazVideoEnhance", display_name="Topaz Video Enhance (Legacy)", - category="api node/video/Topaz", + category="partner/video/Topaz", description="Breathe new life into video with powerful upscaling and recovery technology.", inputs=[ IO.Video.Input("video"), @@ -475,7 +475,7 @@ class TopazVideoEnhanceV2(IO.ComfyNode): return IO.Schema( node_id="TopazVideoEnhanceV2", display_name="Topaz Video Enhance", - category="api node/video/Topaz", + category="partner/video/Topaz", description="Breathe new life into video with powerful upscaling and recovery technology.", inputs=[ IO.Video.Input("video"), diff --git a/comfy_api_nodes/nodes_tripo.py b/comfy_api_nodes/nodes_tripo.py index d6501dee4..a3f2cb053 100644 --- a/comfy_api_nodes/nodes_tripo.py +++ b/comfy_api_nodes/nodes_tripo.py @@ -11,6 +11,9 @@ from comfy_api_nodes.apis.tripo import ( TripoModelVersion, TripoMultiviewToModelRequest, TripoOrientation, + TripoP1ImageToModelRequest, + TripoP1MultiviewToModelRequest, + TripoP1TextToModelRequest, TripoRefineModelRequest, TripoStyle, TripoTaskResponse, @@ -80,7 +83,7 @@ class TripoTextToModelNode(IO.ComfyNode): return IO.Schema( node_id="TripoTextToModelNode", display_name="Tripo: Text to Model", - category="api node/3d/Tripo", + category="partner/3d/Tripo", inputs=[ IO.String.Input("prompt", multiline=True), IO.String.Input("negative_prompt", multiline=True, optional=True), @@ -93,10 +96,22 @@ class TripoTextToModelNode(IO.ComfyNode): IO.Int.Input("image_seed", default=42, optional=True, advanced=True), IO.Int.Input("model_seed", default=42, optional=True, advanced=True), IO.Int.Input("texture_seed", default=42, optional=True, advanced=True), - IO.Combo.Input("texture_quality", default="standard", options=["standard", "detailed"], optional=True, advanced=True), + IO.Combo.Input( + "texture_quality", + default="standard", + options=["standard", "detailed"], + optional=True, + advanced=True, + ), IO.Int.Input("face_limit", default=-1, min=-1, max=2000000, optional=True, advanced=True), IO.Boolean.Input("quad", default=False, optional=True, advanced=True), - IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True, advanced=True), + IO.Combo.Input( + "geometry_quality", + default="standard", + options=["standard", "detailed"], + optional=True, + advanced=True, + ), ], outputs=[ IO.String.Output(display_name="model_file"), # for backward compatibility only @@ -195,7 +210,7 @@ class TripoImageToModelNode(IO.ComfyNode): return IO.Schema( node_id="TripoImageToModelNode", display_name="Tripo: Image to Model", - category="api node/3d/Tripo", + category="partner/3d/Tripo", inputs=[ IO.Image.Input("image"), IO.Combo.Input( @@ -209,16 +224,36 @@ class TripoImageToModelNode(IO.ComfyNode): IO.Boolean.Input("pbr", default=True, optional=True), IO.Int.Input("model_seed", default=42, optional=True, advanced=True), IO.Combo.Input( - "orientation", options=TripoOrientation, default=TripoOrientation.DEFAULT, optional=True, advanced=True + "orientation", + options=TripoOrientation, + default=TripoOrientation.DEFAULT, + optional=True, + advanced=True, ), IO.Int.Input("texture_seed", default=42, optional=True, advanced=True), - IO.Combo.Input("texture_quality", default="standard", options=["standard", "detailed"], optional=True, advanced=True), IO.Combo.Input( - "texture_alignment", default="original_image", options=["original_image", "geometry"], optional=True, advanced=True + "texture_quality", + default="standard", + options=["standard", "detailed"], + optional=True, + advanced=True, + ), + IO.Combo.Input( + "texture_alignment", + default="original_image", + options=["original_image", "geometry"], + optional=True, + advanced=True, ), IO.Int.Input("face_limit", default=-1, min=-1, max=500000, optional=True, advanced=True), IO.Boolean.Input("quad", default=False, optional=True, advanced=True), - IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True, advanced=True), + IO.Combo.Input( + "geometry_quality", + default="standard", + options=["standard", "detailed"], + optional=True, + advanced=True, + ), ], outputs=[ IO.String.Output(display_name="model_file"), # for backward compatibility only @@ -323,7 +358,7 @@ class TripoMultiviewToModelNode(IO.ComfyNode): return IO.Schema( node_id="TripoMultiviewToModelNode", display_name="Tripo: Multiview to Model", - category="api node/3d/Tripo", + category="partner/3d/Tripo", inputs=[ IO.Image.Input("image"), IO.Image.Input("image_left", optional=True), @@ -346,13 +381,35 @@ class TripoMultiviewToModelNode(IO.ComfyNode): IO.Boolean.Input("pbr", default=True, optional=True), IO.Int.Input("model_seed", default=42, optional=True, advanced=True), IO.Int.Input("texture_seed", default=42, optional=True, advanced=True), - IO.Combo.Input("texture_quality", default="standard", options=["standard", "detailed"], optional=True, advanced=True), IO.Combo.Input( - "texture_alignment", default="original_image", options=["original_image", "geometry"], optional=True, advanced=True + "texture_quality", + default="standard", + options=["standard", "detailed"], + optional=True, + advanced=True, + ), + IO.Combo.Input( + "texture_alignment", + default="original_image", + options=["original_image", "geometry"], + optional=True, + advanced=True, ), IO.Int.Input("face_limit", default=-1, min=-1, max=500000, optional=True, advanced=True), - IO.Boolean.Input("quad", default=False, optional=True, advanced=True, tooltip="This parameter is deprecated and does nothing."), - IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True, advanced=True), + IO.Boolean.Input( + "quad", + default=False, + optional=True, + advanced=True, + tooltip="This parameter is deprecated and does nothing.", + ), + IO.Combo.Input( + "geometry_quality", + default="standard", + options=["standard", "detailed"], + optional=True, + advanced=True, + ), ], outputs=[ IO.String.Output(display_name="model_file"), # for backward compatibility only @@ -461,15 +518,25 @@ class TripoTextureNode(IO.ComfyNode): return IO.Schema( node_id="TripoTextureNode", display_name="Tripo: Texture model", - category="api node/3d/Tripo", + category="partner/3d/Tripo", inputs=[ IO.Custom("MODEL_TASK_ID").Input("model_task_id"), IO.Boolean.Input("texture", default=True, optional=True), IO.Boolean.Input("pbr", default=True, optional=True), IO.Int.Input("texture_seed", default=42, optional=True, advanced=True), - IO.Combo.Input("texture_quality", default="standard", options=["standard", "detailed"], optional=True, advanced=True), IO.Combo.Input( - "texture_alignment", default="original_image", options=["original_image", "geometry"], optional=True, advanced=True + "texture_quality", + default="standard", + options=["standard", "detailed"], + optional=True, + advanced=True, + ), + IO.Combo.Input( + "texture_alignment", + default="original_image", + options=["original_image", "geometry"], + optional=True, + advanced=True, ), ], outputs=[ @@ -528,7 +595,7 @@ class TripoRefineNode(IO.ComfyNode): return IO.Schema( node_id="TripoRefineNode", display_name="Tripo: Refine Draft model", - category="api node/3d/Tripo", + category="partner/3d/Tripo", description="Refine a draft model created by v1.4 Tripo models only.", inputs=[ IO.Custom("MODEL_TASK_ID").Input("model_task_id", tooltip="Must be a v1.4 Tripo model"), @@ -568,7 +635,7 @@ class TripoRigNode(IO.ComfyNode): return IO.Schema( node_id="TripoRigNode", display_name="Tripo: Rig model", - category="api node/3d/Tripo", + category="partner/3d/Tripo", inputs=[IO.Custom("MODEL_TASK_ID").Input("original_model_task_id")], outputs=[ IO.String.Output(display_name="model_file"), # for backward compatibility only @@ -605,7 +672,7 @@ class TripoRetargetNode(IO.ComfyNode): return IO.Schema( node_id="TripoRetargetNode", display_name="Tripo: Retarget rigged model", - category="api node/3d/Tripo", + category="partner/3d/Tripo", inputs=[ IO.Custom("RIG_TASK_ID").Input("original_model_task_id"), IO.Combo.Input( @@ -626,7 +693,7 @@ class TripoRetargetNode(IO.ComfyNode): "preset:hexapod:walk", "preset:octopod:walk", "preset:serpentine:march", - "preset:aquatic:march" + "preset:aquatic:march", ], ), ], @@ -670,7 +737,7 @@ class TripoConversionNode(IO.ComfyNode): return IO.Schema( node_id="TripoConversionNode", display_name="Tripo: Convert model", - category="api node/3d/Tripo", + category="partner/3d/Tripo", inputs=[ IO.Custom("MODEL_TASK_ID,RIG_TASK_ID,RETARGET_TASK_ID").Input("original_model_task_id"), IO.Combo.Input("format", options=["GLTF", "USDZ", "FBX", "OBJ", "STL", "3MF"]), @@ -817,7 +884,7 @@ class TripoConversionNode(IO.ComfyNode): # Parse part_names from comma-separated string to list part_names_list = None if part_names and part_names.strip(): - part_names_list = [name.strip() for name in part_names.split(',') if name.strip()] + part_names_list = [name.strip() for name in part_names.split(",") if name.strip()] response = await sync_op( cls, @@ -848,6 +915,373 @@ class TripoConversionNode(IO.ComfyNode): return await poll_until_finished(cls, response, average_duration=30) +def _p1_price_expr(*, geometry_credits: int, textured_credits: int, detailed_credits: int) -> str: + return ( + "(" + " $mode := widgets.output_mode;" + ' $detailed := $lookup(widgets, "output_mode.texture_quality") = "detailed";' + f' $credits := $mode = "geometry only" ? {geometry_credits} : ($detailed ? {detailed_credits} : {textured_credits});' + ' {"type":"usd","usd": $credits * 0.01, "format": {"approximate": true}}' + ")" + ) + + +def _p1_textured_inputs(*, include_image_alignment: bool) -> list: + """Inputs shown inside the 'Textured' branch of the P1 output_mode DynamicCombo.""" + inputs: list = [ + IO.Boolean.Input("pbr", default=True, tooltip="Include PBR maps. When on, base texture is forced on too."), + IO.Combo.Input("texture_quality", options=["standard", "detailed"], default="standard"), + ] + if include_image_alignment: + inputs.extend( + [ + IO.Combo.Input( + "texture_alignment", + options=["original_image", "geometry"], + default="original_image", + tooltip="Prioritize visual fidelity to the source image, or alignment to the mesh geometry.", + ), + IO.Combo.Input( + "orientation", + options=["default", "align_image"], + default="default", + tooltip="Rotate the output to match the source image. Only applies when textured.", + ), + ] + ) + inputs.append(IO.Int.Input("texture_seed", default=42, advanced=True)) + return inputs + + +def _build_p1_output_mode(*, include_image_alignment: bool) -> IO.DynamicCombo.Input: + return IO.DynamicCombo.Input( + "output_mode", + options=[ + IO.DynamicCombo.Option("Geometry only", []), + IO.DynamicCombo.Option("Textured", _p1_textured_inputs(include_image_alignment=include_image_alignment)), + ], + tooltip='"Geometry only" returns an untextured mesh. "Textured" adds color/PBR maps.', + ) + + +def _resolve_p1_texture_fields(output_mode: dict) -> dict: + """Translate the output_mode DynamicCombo payload into P1 request fields. + + pbr=true forces texture=true server-side, but we send both explicitly so the + intent is visible in the request body and logs. + """ + mode = output_mode["output_mode"] + if mode == "Geometry only": + return {"texture": False, "pbr": False} + out = { + "texture": True, + "pbr": bool(output_mode.get("pbr", True)), + "texture_quality": output_mode.get("texture_quality", "standard"), + "texture_seed": output_mode.get("texture_seed"), + } + if "texture_alignment" in output_mode: + out["texture_alignment"] = output_mode["texture_alignment"] + if "orientation" in output_mode: + out["orientation"] = output_mode["orientation"] + return out + + +def _p1_common_inputs() -> list: + """Inputs shared by all P1 nodes (placed after output_mode).""" + return [ + IO.Int.Input( + "face_limit", + default=-1, + min=-1, + max=20000, + optional=True, + advanced=True, + tooltip="Target face count, 48-20000. -1 lets Tripo pick adaptively.", + ), + IO.Int.Input("model_seed", default=42, optional=True, advanced=True), + IO.Boolean.Input( + "auto_size", + default=False, + optional=True, + advanced=True, + tooltip="Scale the output to approximate real-world meters.", + ), + IO.Boolean.Input( + "export_uv", + default=True, + optional=True, + advanced=True, + tooltip="UV unwrap during generation. Turn off for faster geometry-only runs.", + ), + IO.Boolean.Input( + "compress_geometry", + default=False, + optional=True, + advanced=True, + tooltip="Apply geometry-based compression. Decompress before editing.", + ), + ] + + +def _build_p1_request_kwargs( + *, + output_mode: dict, + face_limit: int, + model_seed: int, + auto_size: bool, + export_uv: bool, + compress_geometry: bool, +) -> dict: + """Common P1 request fields shared by all three node types.""" + kwargs: dict = { + "model_seed": model_seed, + "face_limit": face_limit if face_limit != -1 else None, + "auto_size": auto_size, + "export_uv": export_uv, + "compress": "geometry" if compress_geometry else None, + } + kwargs.update(_resolve_p1_texture_fields(output_mode)) + return kwargs + + +class TripoP1TextToModelNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="TripoP1TextToModelNode", + display_name="Tripo P1: Text to Model", + category="partner/3d/Tripo", + description="Tripo P1 text-to-3D. Optimized for low-poly, game-ready meshes with stable topology.", + inputs=[ + IO.String.Input("prompt", multiline=True, tooltip="Up to 1024 characters."), + IO.String.Input("negative_prompt", multiline=True, optional=True, tooltip="Up to 255 characters."), + _build_p1_output_mode(include_image_alignment=False), + IO.Int.Input("image_seed", default=42, optional=True, advanced=True), + *_p1_common_inputs(), + ], + outputs=[ + IO.String.Output(display_name="model_file"), # for backward compatibility only + IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"), + IO.File3DGLB.Output(display_name="GLB"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["output_mode", "output_mode.texture_quality"]), + expr=_p1_price_expr(geometry_credits=30, textured_credits=40, detailed_credits=50), + ), + ) + + @classmethod + async def execute( + cls, + prompt: str, + output_mode: dict, + negative_prompt: str | None = None, + image_seed: int | None = None, + face_limit: int = -1, + model_seed: int | None = None, + auto_size: bool = False, + export_uv: bool = True, + compress_geometry: bool = False, + ) -> IO.NodeOutput: + if not prompt: + raise RuntimeError("Prompt is required") + common = _build_p1_request_kwargs( + output_mode=output_mode, + face_limit=face_limit, + model_seed=model_seed, + auto_size=auto_size, + export_uv=export_uv, + compress_geometry=compress_geometry, + ) + request = TripoP1TextToModelRequest( + prompt=prompt, + negative_prompt=negative_prompt or None, + image_seed=image_seed, + **common, + ) + response = await sync_op( + cls, + endpoint=ApiEndpoint(path="/proxy/tripo/v2/openapi/task", method="POST"), + response_model=TripoTaskResponse, + data=request, + ) + return await poll_until_finished(cls, response, average_duration=60) + + +class TripoP1ImageToModelNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="TripoP1ImageToModelNode", + display_name="Tripo P1: Image to Model", + category="partner/3d/Tripo", + description="Tripo P1 image-to-3D. Optimized for low-poly, game-ready meshes.", + inputs=[ + IO.Image.Input("image"), + _build_p1_output_mode(include_image_alignment=True), + IO.Boolean.Input( + "enable_image_autofix", + default=False, + optional=True, + advanced=True, + tooltip="Pre-process the input image for better generation quality.", + ), + *_p1_common_inputs(), + ], + outputs=[ + IO.String.Output(display_name="model_file"), # for backward compatibility only + IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"), + IO.File3DGLB.Output(display_name="GLB"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["output_mode", "output_mode.texture_quality"]), + expr=_p1_price_expr(geometry_credits=40, textured_credits=50, detailed_credits=60), + ), + ) + + @classmethod + async def execute( + cls, + image: Input.Image, + output_mode: dict, + enable_image_autofix: bool = False, + face_limit: int = -1, + model_seed: int | None = None, + auto_size: bool = False, + export_uv: bool = True, + compress_geometry: bool = False, + ) -> IO.NodeOutput: + if image is None: + raise RuntimeError("Image is required") + tripo_file = TripoFileReference( + root=TripoUrlReference( + url=(await upload_images_to_comfyapi(cls, image, max_images=1))[0], + type="jpeg", + ) + ) + common = _build_p1_request_kwargs( + output_mode=output_mode, + face_limit=face_limit, + model_seed=model_seed, + auto_size=auto_size, + export_uv=export_uv, + compress_geometry=compress_geometry, + ) + request = TripoP1ImageToModelRequest( + file=tripo_file, + enable_image_autofix=enable_image_autofix, + **common, + ) + response = await sync_op( + cls, + endpoint=ApiEndpoint(path="/proxy/tripo/v2/openapi/task", method="POST"), + response_model=TripoTaskResponse, + data=request, + ) + return await poll_until_finished(cls, response, average_duration=60) + + +class TripoP1MultiviewToModelNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="TripoP1MultiviewToModelNode", + display_name="Tripo P1: Multiview to Model", + category="partner/3d/Tripo", + description="Tripo P1 multiview-to-3D from 2-4 reference images in [front, left, back, right] order. " + "Front is required; any combination of the other three may be omitted.", + inputs=[ + IO.Image.Input("image", tooltip="Front view (0°). Required."), + IO.Image.Input( + "image_left", + optional=True, + tooltip="Left view (90°), i.e. the subject's left side.", + ), + IO.Image.Input("image_back", optional=True, tooltip="Back view (180°)."), + IO.Image.Input( + "image_right", + optional=True, + tooltip="Right view (270°), i.e. the subject's right side.", + ), + _build_p1_output_mode(include_image_alignment=True), + *_p1_common_inputs(), + ], + outputs=[ + IO.String.Output(display_name="model_file"), # for backward compatibility only + IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"), + IO.File3DGLB.Output(display_name="GLB"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["output_mode", "output_mode.texture_quality"]), + expr=_p1_price_expr(geometry_credits=40, textured_credits=50, detailed_credits=60), + ), + ) + + @classmethod + async def execute( + cls, + image: Input.Image, + output_mode: dict, + image_left: Input.Image | None = None, + image_back: Input.Image | None = None, + image_right: Input.Image | None = None, + face_limit: int = -1, + model_seed: int | None = None, + auto_size: bool = False, + export_uv: bool = True, + compress_geometry: bool = False, + ) -> IO.NodeOutput: + views = [image, image_left, image_back, image_right] + if sum(1 for v in views if v is not None) < 2: + raise RuntimeError("Tripo P1 multiview requires at least 2 images (front plus one of left/back/right).") + + files: list[TripoFileReference] = [] + for view in views: + if view is None: + files.append(TripoFileReference(root=TripoFileEmptyReference())) + continue + url = (await upload_images_to_comfyapi(cls, view, max_images=1))[0] + files.append(TripoFileReference(root=TripoUrlReference(url=url, type="jpeg"))) + + common = _build_p1_request_kwargs( + output_mode=output_mode, + face_limit=face_limit, + model_seed=model_seed, + auto_size=auto_size, + export_uv=export_uv, + compress_geometry=compress_geometry, + ) + request = TripoP1MultiviewToModelRequest(files=files, **common) + response = await sync_op( + cls, + endpoint=ApiEndpoint(path="/proxy/tripo/v2/openapi/task", method="POST"), + response_model=TripoTaskResponse, + data=request, + ) + return await poll_until_finished(cls, response, average_duration=80) + + class TripoExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: @@ -855,6 +1289,9 @@ class TripoExtension(ComfyExtension): TripoTextToModelNode, TripoImageToModelNode, TripoMultiviewToModelNode, + TripoP1TextToModelNode, + TripoP1ImageToModelNode, + TripoP1MultiviewToModelNode, TripoTextureNode, TripoRefineNode, TripoRigNode, diff --git a/comfy_api_nodes/nodes_veo2.py b/comfy_api_nodes/nodes_veo2.py index 2ff75d9b2..ed34e928b 100644 --- a/comfy_api_nodes/nodes_veo2.py +++ b/comfy_api_nodes/nodes_veo2.py @@ -45,7 +45,7 @@ class VeoVideoGenerationNode(IO.ComfyNode): return IO.Schema( node_id="VeoVideoGenerationNode", display_name="Google Veo 2 Video Generation", - category="api node/video/Veo", + category="partner/video/Veo", description="Generates videos from text prompts using Google's Veo 2 API", inputs=[ IO.String.Input( @@ -256,7 +256,7 @@ class Veo3VideoGenerationNode(IO.ComfyNode): return IO.Schema( node_id="Veo3VideoGenerationNode", display_name="Google Veo 3 Video Generation", - category="api node/video/Veo", + category="partner/video/Veo", description="Generates videos from text prompts using Google's Veo 3 API", inputs=[ IO.String.Input( @@ -468,7 +468,7 @@ class Veo3FirstLastFrameNode(IO.ComfyNode): return IO.Schema( node_id="Veo3FirstLastFrameNode", display_name="Google Veo 3 First-Last-Frame to Video", - category="api node/video/Veo", + category="partner/video/Veo", description="Generate video using prompt and first and last frames.", inputs=[ IO.String.Input( diff --git a/comfy_api_nodes/nodes_vidu.py b/comfy_api_nodes/nodes_vidu.py index 8d90cefeb..8c5a43f5b 100644 --- a/comfy_api_nodes/nodes_vidu.py +++ b/comfy_api_nodes/nodes_vidu.py @@ -71,7 +71,7 @@ class ViduTextToVideoNode(IO.ComfyNode): return IO.Schema( node_id="ViduTextToVideoNode", display_name="Vidu Text To Video Generation", - category="api node/video/Vidu", + category="partner/video/Vidu", description="Generate video from a text prompt", inputs=[ IO.Combo.Input("model", options=["viduq1"], tooltip="Model name"), @@ -169,7 +169,7 @@ class ViduImageToVideoNode(IO.ComfyNode): return IO.Schema( node_id="ViduImageToVideoNode", display_name="Vidu Image To Video Generation", - category="api node/video/Vidu", + category="partner/video/Vidu", description="Generate video from image and optional prompt", inputs=[ IO.Combo.Input("model", options=["viduq1"], tooltip="Model name"), @@ -273,7 +273,7 @@ class ViduReferenceVideoNode(IO.ComfyNode): return IO.Schema( node_id="ViduReferenceVideoNode", display_name="Vidu Reference To Video Generation", - category="api node/video/Vidu", + category="partner/video/Vidu", description="Generate video from multiple images and a prompt", inputs=[ IO.Combo.Input("model", options=["viduq1"], tooltip="Model name"), @@ -388,7 +388,7 @@ class ViduStartEndToVideoNode(IO.ComfyNode): return IO.Schema( node_id="ViduStartEndToVideoNode", display_name="Vidu Start End To Video Generation", - category="api node/video/Vidu", + category="partner/video/Vidu", description="Generate a video from start and end frames and a prompt", inputs=[ IO.Combo.Input("model", options=["viduq1"], tooltip="Model name"), @@ -492,7 +492,7 @@ class Vidu2TextToVideoNode(IO.ComfyNode): return IO.Schema( node_id="Vidu2TextToVideoNode", display_name="Vidu2 Text-to-Video Generation", - category="api node/video/Vidu", + category="partner/video/Vidu", description="Generate video from a text prompt", inputs=[ IO.Combo.Input("model", options=["viduq2"]), @@ -584,7 +584,7 @@ class Vidu2ImageToVideoNode(IO.ComfyNode): return IO.Schema( node_id="Vidu2ImageToVideoNode", display_name="Vidu2 Image-to-Video Generation", - category="api node/video/Vidu", + category="partner/video/Vidu", description="Generate a video from an image and an optional prompt.", inputs=[ IO.Combo.Input("model", options=["viduq2-pro-fast", "viduq2-pro", "viduq2-turbo"]), @@ -714,7 +714,7 @@ class Vidu2ReferenceVideoNode(IO.ComfyNode): return IO.Schema( node_id="Vidu2ReferenceVideoNode", display_name="Vidu2 Reference-to-Video Generation", - category="api node/video/Vidu", + category="partner/video/Vidu", description="Generate a video from multiple reference images and a prompt.", inputs=[ IO.Combo.Input("model", options=["viduq2"]), @@ -849,7 +849,7 @@ class Vidu2StartEndToVideoNode(IO.ComfyNode): return IO.Schema( node_id="Vidu2StartEndToVideoNode", display_name="Vidu2 Start/End Frame-to-Video Generation", - category="api node/video/Vidu", + category="partner/video/Vidu", description="Generate a video from a start frame, an end frame, and a prompt.", inputs=[ IO.Combo.Input("model", options=["viduq2-pro-fast", "viduq2-pro", "viduq2-turbo"]), @@ -969,7 +969,7 @@ class ViduExtendVideoNode(IO.ComfyNode): return IO.Schema( node_id="ViduExtendVideoNode", display_name="Vidu Video Extension", - category="api node/video/Vidu", + category="partner/video/Vidu", description="Extend an existing video by generating additional frames.", inputs=[ IO.DynamicCombo.Input( @@ -1138,7 +1138,7 @@ class ViduMultiFrameVideoNode(IO.ComfyNode): return IO.Schema( node_id="ViduMultiFrameVideoNode", display_name="Vidu Multi-Frame Video Generation", - category="api node/video/Vidu", + category="partner/video/Vidu", description="Generate a video with multiple keyframe transitions.", inputs=[ IO.Combo.Input("model", options=["viduq2-pro", "viduq2-turbo"]), @@ -1284,7 +1284,7 @@ class Vidu3TextToVideoNode(IO.ComfyNode): return IO.Schema( node_id="Vidu3TextToVideoNode", display_name="Vidu Q3 Text-to-Video Generation", - category="api node/video/Vidu", + category="partner/video/Vidu", description="Generate video from a text prompt.", inputs=[ IO.DynamicCombo.Input( @@ -1429,7 +1429,7 @@ class Vidu3ImageToVideoNode(IO.ComfyNode): return IO.Schema( node_id="Vidu3ImageToVideoNode", display_name="Vidu Q3 Image-to-Video Generation", - category="api node/video/Vidu", + category="partner/video/Vidu", description="Generate a video from an image and an optional prompt.", inputs=[ IO.DynamicCombo.Input( @@ -1571,7 +1571,7 @@ class Vidu3StartEndToVideoNode(IO.ComfyNode): return IO.Schema( node_id="Vidu3StartEndToVideoNode", display_name="Vidu Q3 Start/End Frame-to-Video Generation", - category="api node/video/Vidu", + category="partner/video/Vidu", description="Generate a video from a start frame, an end frame, and a prompt.", inputs=[ IO.DynamicCombo.Input( diff --git a/comfy_api_nodes/nodes_wan.py b/comfy_api_nodes/nodes_wan.py index 68061bb5c..b7b97d70f 100644 --- a/comfy_api_nodes/nodes_wan.py +++ b/comfy_api_nodes/nodes_wan.py @@ -61,7 +61,7 @@ class WanTextToImageApi(IO.ComfyNode): return IO.Schema( node_id="WanTextToImageApi", display_name="Wan Text to Image", - category="api node/image/Wan", + category="partner/image/Wan", description="Generates an image based on a text prompt.", inputs=[ IO.Combo.Input( @@ -184,7 +184,7 @@ class WanImageToImageApi(IO.ComfyNode): return IO.Schema( node_id="WanImageToImageApi", display_name="Wan Image to Image", - category="api node/image/Wan", + category="partner/image/Wan", description="Generates an image from one or two input images and a text prompt. " "The output image is currently fixed at 1.6 MP, and its aspect ratio matches the input image(s).", inputs=[ @@ -312,7 +312,7 @@ class WanTextToVideoApi(IO.ComfyNode): return IO.Schema( node_id="WanTextToVideoApi", display_name="Wan Text to Video", - category="api node/video/Wan", + category="partner/video/Wan", description="Generates a video based on a text prompt.", inputs=[ IO.Combo.Input( @@ -495,7 +495,7 @@ class WanImageToVideoApi(IO.ComfyNode): return IO.Schema( node_id="WanImageToVideoApi", display_name="Wan Image to Video", - category="api node/video/Wan", + category="partner/video/Wan", description="Generates a video from the first frame and a text prompt.", inputs=[ IO.Combo.Input( @@ -674,7 +674,7 @@ class WanReferenceVideoApi(IO.ComfyNode): return IO.Schema( node_id="WanReferenceVideoApi", display_name="Wan Reference to Video", - category="api node/video/Wan", + category="partner/video/Wan", description="Use the character and voice from input videos, combined with a prompt, " "to generate a new video that maintains character consistency.", inputs=[ @@ -828,7 +828,7 @@ class Wan2TextToVideoApi(IO.ComfyNode): return IO.Schema( node_id="Wan2TextToVideoApi", display_name="Wan 2.7 Text to Video", - category="api node/video/Wan", + category="partner/video/Wan", description="Generates a video based on a text prompt using the Wan 2.7 model.", inputs=[ IO.DynamicCombo.Input( @@ -981,7 +981,7 @@ class Wan2ImageToVideoApi(IO.ComfyNode): return IO.Schema( node_id="Wan2ImageToVideoApi", display_name="Wan 2.7 Image to Video", - category="api node/video/Wan", + category="partner/video/Wan", description="Generate a video from a first-frame image, with optional last-frame image and audio.", inputs=[ IO.DynamicCombo.Input( @@ -1152,7 +1152,7 @@ class Wan2VideoContinuationApi(IO.ComfyNode): return IO.Schema( node_id="Wan2VideoContinuationApi", display_name="Wan 2.7 Video Continuation", - category="api node/video/Wan", + category="partner/video/Wan", description="Continue a video from where it left off, with optional last-frame control.", inputs=[ IO.DynamicCombo.Input( @@ -1319,7 +1319,7 @@ class Wan2VideoEditApi(IO.ComfyNode): return IO.Schema( node_id="Wan2VideoEditApi", display_name="Wan 2.7 Video Edit", - category="api node/video/Wan", + category="partner/video/Wan", description="Edit a video using text instructions, reference images, or style transfer.", inputs=[ IO.DynamicCombo.Input( @@ -1477,7 +1477,7 @@ class Wan2ReferenceVideoApi(IO.ComfyNode): return IO.Schema( node_id="Wan2ReferenceVideoApi", display_name="Wan 2.7 Reference to Video", - category="api node/video/Wan", + category="partner/video/Wan", description="Generate a video featuring a person or object from reference materials. " "Supports single-character performances and multi-character interactions.", inputs=[ @@ -1651,7 +1651,7 @@ class HappyHorseTextToVideoApi(IO.ComfyNode): return IO.Schema( node_id="HappyHorseTextToVideoApi", display_name="HappyHorse Text to Video", - category="api node/video/Wan", + category="partner/video/Wan", description="Generates a video based on a text prompt using the HappyHorse model.", inputs=[ IO.DynamicCombo.Input( @@ -1775,7 +1775,7 @@ class HappyHorseImageToVideoApi(IO.ComfyNode): return IO.Schema( node_id="HappyHorseImageToVideoApi", display_name="HappyHorse Image to Video", - category="api node/video/Wan", + category="partner/video/Wan", description="Generate a video from a first-frame image using the HappyHorse model.", inputs=[ IO.DynamicCombo.Input( @@ -1905,7 +1905,7 @@ class HappyHorseVideoEditApi(IO.ComfyNode): return IO.Schema( node_id="HappyHorseVideoEditApi", display_name="HappyHorse Video Edit", - category="api node/video/Wan", + category="partner/video/Wan", description="Edit a video using text instructions or reference images with the HappyHorse model. " "Output duration is 3-15s and matches the input video; inputs longer than 15s are truncated.", inputs=[ @@ -2046,7 +2046,7 @@ class HappyHorseReferenceVideoApi(IO.ComfyNode): return IO.Schema( node_id="HappyHorseReferenceVideoApi", display_name="HappyHorse Reference to Video", - category="api node/video/Wan", + category="partner/video/Wan", description="Generate a video featuring a person or object from reference materials with the HappyHorse " "model. Supports single-character performances and multi-character interactions.", inputs=[ diff --git a/comfy_api_nodes/nodes_wavespeed.py b/comfy_api_nodes/nodes_wavespeed.py index 65e45f60a..5839f9d37 100644 --- a/comfy_api_nodes/nodes_wavespeed.py +++ b/comfy_api_nodes/nodes_wavespeed.py @@ -27,7 +27,7 @@ class WavespeedFlashVSRNode(IO.ComfyNode): return IO.Schema( node_id="WavespeedFlashVSRNode", display_name="FlashVSR Video Upscale", - category="api node/video/WaveSpeed", + category="partner/video/WaveSpeed", description="Fast, high-quality video upscaler that " "boosts resolution and restores clarity for low-resolution or blurry footage.", inputs=[ @@ -98,7 +98,7 @@ class WavespeedImageUpscaleNode(IO.ComfyNode): return IO.Schema( node_id="WavespeedImageUpscaleNode", display_name="WaveSpeed Image Upscale", - category="api node/image/WaveSpeed", + category="partner/image/WaveSpeed", description="Boost image resolution and quality, upscaling photos to 4K or 8K for sharp, detailed results.", inputs=[ IO.Combo.Input("model", options=["SeedVR2", "Ultimate"]), diff --git a/comfy_api_nodes/util/__init__.py b/comfy_api_nodes/util/__init__.py index f3584aba9..25cb88869 100644 --- a/comfy_api_nodes/util/__init__.py +++ b/comfy_api_nodes/util/__init__.py @@ -16,16 +16,17 @@ from .conversions import ( convert_mask_to_image, downscale_image_tensor, downscale_image_tensor_by_max_side, + downscale_video_to_max_pixels, image_tensor_pair_to_batch, pil_to_bytesio, resize_mask_to_image, - resize_video_to_pixel_budget, tensor_to_base64_string, tensor_to_bytesio, tensor_to_pil, text_filepath_to_base64_string, text_filepath_to_data_uri, trim_video, + upscale_video_to_min_pixels, video_to_base64_string, ) from .download_helpers import ( @@ -88,16 +89,17 @@ __all__ = [ "convert_mask_to_image", "downscale_image_tensor", "downscale_image_tensor_by_max_side", + "downscale_video_to_max_pixels", "image_tensor_pair_to_batch", "pil_to_bytesio", "resize_mask_to_image", - "resize_video_to_pixel_budget", "tensor_to_base64_string", "tensor_to_bytesio", "tensor_to_pil", "text_filepath_to_base64_string", "text_filepath_to_data_uri", "trim_video", + "upscale_video_to_min_pixels", "video_to_base64_string", # Validation utilities "get_image_dimensions", diff --git a/comfy_api_nodes/util/client.py b/comfy_api_nodes/util/client.py index 052301c33..57c501724 100644 --- a/comfy_api_nodes/util/client.py +++ b/comfy_api_nodes/util/client.py @@ -86,7 +86,7 @@ class _PollUIState: _RETRY_STATUS = {408, 500, 502, 503, 504} # status 429 is handled separately COMPLETED_STATUSES = ["succeeded", "succeed", "success", "completed", "finished", "done", "complete"] FAILED_STATUSES = ["cancelled", "canceled", "canceling", "fail", "failed", "error"] -QUEUED_STATUSES = ["created", "queued", "queueing", "submitted", "initializing", "wait"] +QUEUED_STATUSES = ["created", "queued", "queueing", "submitted", "initializing", "wait", "in_queue"] async def sync_op( diff --git a/comfy_api_nodes/util/conversions.py b/comfy_api_nodes/util/conversions.py index be5d5719b..a1b5d599c 100644 --- a/comfy_api_nodes/util/conversions.py +++ b/comfy_api_nodes/util/conversions.py @@ -415,14 +415,48 @@ def trim_video(video: Input.Video, duration_sec: float) -> Input.Video: raise RuntimeError(f"Failed to trim video: {str(e)}") from e -def resize_video_to_pixel_budget(video: Input.Video, total_pixels: int) -> Input.Video: - """Downscale a video to fit within ``total_pixels`` (w * h), preserving aspect ratio. +def downscale_video_to_max_pixels(video: Input.Video, max_pixels: int) -> Input.Video: + """Downscale a video to fit within ``max_pixels`` (w * h), preserving aspect ratio. Returns the original video object untouched when it already fits. Preserves frame rate, duration, and audio. Aspect ratio is preserved up to a fraction of a percent (even-dim rounding). """ src_w, src_h = video.get_dimensions() - scale_dims = _compute_downscale_dims(src_w, src_h, total_pixels) + scale_dims = _compute_downscale_dims(src_w, src_h, max_pixels) + if scale_dims is None: + return video + return _apply_video_scale(video, scale_dims) + + +def _compute_upscale_dims(src_w: int, src_h: int, total_pixels: int) -> tuple[int, int] | None: + """Return upscaled (w, h) with even dims meeting at least ``total_pixels``, or None if already large enough. + + Source aspect ratio is preserved; output may drift by a fraction of a percent because both dimensions + are rounded up to even values (many codecs require divisible-by-2). The result is guaranteed to be at + least ``total_pixels``. + """ + pixels = src_w * src_h + if pixels >= total_pixels: + return None + scale = math.sqrt(total_pixels / pixels) + new_w = math.ceil(src_w * scale) + new_h = math.ceil(src_h * scale) + if new_w % 2: + new_w += 1 + if new_h % 2: + new_h += 1 + return new_w, new_h + + +def upscale_video_to_min_pixels(video: Input.Video, min_pixels: int) -> Input.Video: + """Upscale a video to meet at least ``min_pixels`` (w * h), preserving aspect ratio. + + Returns the original video object untouched when it already meets the minimum. Preserves frame rate, + duration, and audio. Aspect ratio is preserved up to a fraction of a percent (even-dim rounding). + Note: upscaling a low-resolution source does not add real detail; downstream model quality may suffer. + """ + src_w, src_h = video.get_dimensions() + scale_dims = _compute_upscale_dims(src_w, src_h, min_pixels) if scale_dims is None: return video return _apply_video_scale(video, scale_dims) @@ -435,6 +469,11 @@ def _apply_video_scale(video: Input.Video, scale_dims: tuple[int, int]) -> Input input_container = None output_container = None + # get_stream_source() is untrimmed, so apply the trim window in this same pass. + # start_time is normalized (>= 0); duration == 0 means "until the end". + start_time, duration = video.get_active_trim_window() + trimming = bool(start_time or duration) + try: input_source = video.get_stream_source() input_container = av.open(input_source, mode="r") @@ -453,16 +492,45 @@ def _apply_video_scale(video: Input.Video, scale_dims: tuple[int, int]) -> Input audio_stream.layout = stream.layout break + in_video = input_container.streams.video[0] + start_pts = int(start_time / in_video.time_base) if trimming else 0 + end_pts = int((start_time + duration) / in_video.time_base) if duration else None + if start_pts: + input_container.seek(start_pts, stream=in_video) + + encoded = 0 for frame in input_container.decode(video=0): + if trimming: + if frame.pts is None or frame.pts < start_pts: + continue + if end_pts is not None and frame.pts >= end_pts: + break frame = frame.reformat(width=out_w, height=out_h, format="yuv420p") + # Re-wrap as a fresh frame: dropping irregular source timestamps (VFR/AVI/GIF/...) + # lets the encoder assign clean ones and avoids mp4 muxer errors. + frame = av.VideoFrame.from_ndarray(frame.to_ndarray(format="yuv420p"), format="yuv420p") for packet in video_stream.encode(frame): output_container.mux(packet) + encoded += 1 for packet in video_stream.encode(): output_container.mux(packet) + if encoded == 0: + raise ValueError( + f"resize produced no frames (start_time={start_time}, duration={duration} " + "selected nothing from the source)" + ) + if audio_stream is not None: input_container.seek(0) for audio_frame in input_container.decode(audio=0): + if trimming: + if audio_frame.time is None or audio_frame.time < start_time: + continue + if duration and audio_frame.time > start_time + duration: + break + # Carry odd audio time bases the mp4 muxer rejects; reset pts, encoder assigns clean ones (MP3-in-AVI) + audio_frame.pts = None for packet in audio_stream.encode(audio_frame): output_container.mux(packet) for packet in audio_stream.encode(): diff --git a/comfy_execution/graph.py b/comfy_execution/graph.py index c47f3c79b..479ee8a53 100644 --- a/comfy_execution/graph.py +++ b/comfy_execution/graph.py @@ -1,4 +1,3 @@ -from __future__ import annotations from typing import Type, Literal import nodes diff --git a/comfy_execution/progress.py b/comfy_execution/progress.py index f951a3350..731b8dc66 100644 --- a/comfy_execution/progress.py +++ b/comfy_execution/progress.py @@ -1,5 +1,3 @@ -from __future__ import annotations - from typing import TypedDict, Dict, Optional, Tuple from typing_extensions import override from PIL import Image diff --git a/comfy_execution/validation.py b/comfy_execution/validation.py index e73624bd1..ae9a2376c 100644 --- a/comfy_execution/validation.py +++ b/comfy_execution/validation.py @@ -1,4 +1,3 @@ -from __future__ import annotations from comfy_api.latest import IO diff --git a/comfy_extras/mediapipe/face_geometry.py b/comfy_extras/mediapipe/face_geometry.py index 04b2b0557..4f3813430 100644 --- a/comfy_extras/mediapipe/face_geometry.py +++ b/comfy_extras/mediapipe/face_geometry.py @@ -2,7 +2,6 @@ + weighted Procrustes solver. Computes the 4x4 facial transformation matrix. """ -from __future__ import annotations import math import numpy as np diff --git a/comfy_extras/mediapipe/face_landmarker.py b/comfy_extras/mediapipe/face_landmarker.py index a792b6046..e6b463c4c 100644 --- a/comfy_extras/mediapipe/face_landmarker.py +++ b/comfy_extras/mediapipe/face_landmarker.py @@ -1,7 +1,6 @@ """Pure-PyTorch port of MediaPipe's face_landmarker_v2_with_blendshapes.task: BlazeFace detector → FaceMesh v2 → ARKit-52 blendshapes.""" -from __future__ import annotations import math from functools import lru_cache diff --git a/comfy_extras/nodes_ace.py b/comfy_extras/nodes_ace.py index 247d9ae8a..044077b18 100644 --- a/comfy_extras/nodes_ace.py +++ b/comfy_extras/nodes_ace.py @@ -11,7 +11,7 @@ class TextEncodeAceStepAudio(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="TextEncodeAceStepAudio", - category="conditioning", + category="model/conditioning", inputs=[ IO.Clip.Input("clip"), IO.String.Input("tags", multiline=True, dynamic_prompts=True), @@ -33,7 +33,7 @@ class TextEncodeAceStepAudio15(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="TextEncodeAceStepAudio1.5", - category="conditioning", + category="model/conditioning", inputs=[ IO.Clip.Input("clip"), IO.String.Input("tags", multiline=True, dynamic_prompts=True), @@ -67,7 +67,7 @@ class EmptyAceStepLatentAudio(IO.ComfyNode): return IO.Schema( node_id="EmptyAceStepLatentAudio", display_name="Empty Ace Step 1.0 Latent Audio", - category="latent/audio", + category="model/latent/audio", inputs=[ IO.Float.Input("seconds", default=120.0, min=1.0, max=1000.0, step=0.1), IO.Int.Input( @@ -90,7 +90,7 @@ class EmptyAceStep15LatentAudio(IO.ComfyNode): return IO.Schema( node_id="EmptyAceStep1.5LatentAudio", display_name="Empty Ace Step 1.5 Latent Audio", - category="latent/audio", + category="model/latent/audio", inputs=[ IO.Float.Input("seconds", default=120.0, min=1.0, max=1000.0, step=0.01), IO.Int.Input( diff --git a/comfy_extras/nodes_advanced_samplers.py b/comfy_extras/nodes_advanced_samplers.py index 20717ca38..77a561e30 100644 --- a/comfy_extras/nodes_advanced_samplers.py +++ b/comfy_extras/nodes_advanced_samplers.py @@ -45,7 +45,7 @@ class SamplerLCMUpscale(io.ComfyNode): def define_schema(cls) -> io.Schema: return io.Schema( node_id="SamplerLCMUpscale", - category="sampling/samplers", + category="model/sampling/samplers", inputs=[ io.Float.Input("scale_ratio", default=1.0, min=0.1, max=20.0, step=0.01, advanced=True), io.Int.Input("scale_steps", default=-1, min=-1, max=1000, step=1, advanced=True), @@ -91,7 +91,7 @@ class SamplerLCM(io.ComfyNode): def define_schema(cls) -> io.Schema: return io.Schema( node_id="SamplerLCM", - category="sampling/samplers", + category="model/sampling/samplers", description=("LCM sampler with tunable per-step noise. s_noise is a multiplier on the model's training noise scale"), inputs=[ io.Float.Input("s_noise", default=1.0, min=0.0, max=64.0, step=0.01, diff --git a/comfy_extras/nodes_align_your_steps.py b/comfy_extras/nodes_align_your_steps.py index 307f41337..f89a809bb 100644 --- a/comfy_extras/nodes_align_your_steps.py +++ b/comfy_extras/nodes_align_your_steps.py @@ -29,7 +29,7 @@ class AlignYourStepsScheduler(io.ComfyNode): return io.Schema( node_id="AlignYourStepsScheduler", search_aliases=["AYS scheduler"], - category="sampling/schedulers", + category="model/sampling/schedulers", inputs=[ io.Combo.Input("model_type", options=["SD1", "SDXL", "SVD"]), io.Int.Input("steps", default=10, min=1, max=10000), diff --git a/comfy_extras/nodes_apg.py b/comfy_extras/nodes_apg.py index fd561d360..4a352038a 100644 --- a/comfy_extras/nodes_apg.py +++ b/comfy_extras/nodes_apg.py @@ -16,7 +16,7 @@ class APG(io.ComfyNode): return io.Schema( node_id="APG", display_name="Adaptive Projected Guidance", - category="sampling/custom_sampling", + category="model/sampling/custom_sampling", inputs=[ io.Model.Input("model"), io.Float.Input( diff --git a/comfy_extras/nodes_ar_video.py b/comfy_extras/nodes_ar_video.py index 1a15facfa..c22359eb2 100644 --- a/comfy_extras/nodes_ar_video.py +++ b/comfy_extras/nodes_ar_video.py @@ -19,7 +19,7 @@ class EmptyARVideoLatent(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="EmptyARVideoLatent", - category="latent/video", + category="model/latent/video", inputs=[ io.Int.Input("width", default=832, min=16, max=8192, step=16), io.Int.Input("height", default=480, min=16, max=8192, step=16), @@ -53,7 +53,7 @@ class SamplerARVideo(io.ComfyNode): return io.Schema( node_id="SamplerARVideo", display_name="Sampler AR Video", - category="sampling/samplers", + category="model/sampling/samplers", inputs=[ io.Int.Input( "num_frame_per_block", @@ -85,7 +85,7 @@ class ARVideoI2V(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="ARVideoI2V", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Model.Input("model"), io.Vae.Input("vae"), diff --git a/comfy_extras/nodes_audio.py b/comfy_extras/nodes_audio.py index d5084497e..1dc97ecd7 100644 --- a/comfy_extras/nodes_audio.py +++ b/comfy_extras/nodes_audio.py @@ -1,5 +1,3 @@ -from __future__ import annotations - import av import torchaudio import torch @@ -18,7 +16,7 @@ class EmptyLatentAudio(IO.ComfyNode): return IO.Schema( node_id="EmptyLatentAudio", display_name="Empty Latent Audio", - category="latent/audio", + category="model/latent/audio", essentials_category="Audio", inputs=[ IO.Float.Input("seconds", default=47.6, min=1.0, max=1000.0, step=0.1), @@ -43,7 +41,7 @@ class ConditioningStableAudio(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="ConditioningStableAudio", - category="conditioning", + category="model/conditioning", inputs=[ IO.Conditioning.Input("positive"), IO.Conditioning.Input("negative"), @@ -72,7 +70,7 @@ class VAEEncodeAudio(IO.ComfyNode): node_id="VAEEncodeAudio", search_aliases=["audio to latent"], display_name="VAE Encode Audio", - category="latent/audio", + category="model/latent/audio", inputs=[ IO.Audio.Input("audio"), IO.Vae.Input("vae"), @@ -117,7 +115,7 @@ class VAEDecodeAudio(IO.ComfyNode): node_id="VAEDecodeAudio", search_aliases=["latent to audio"], display_name="VAE Decode Audio", - category="latent/audio", + category="model/latent/audio", inputs=[ IO.Latent.Input("samples"), IO.Vae.Input("vae"), @@ -139,7 +137,7 @@ class VAEDecodeAudioTiled(IO.ComfyNode): node_id="VAEDecodeAudioTiled", search_aliases=["latent to audio"], display_name="VAE Decode Audio (Tiled)", - category="latent/audio", + category="model/latent/audio", inputs=[ IO.Latent.Input("samples"), IO.Vae.Input("vae"), @@ -160,7 +158,7 @@ class SaveAudio(IO.ComfyNode): return IO.Schema( node_id="SaveAudio", search_aliases=["export flac"], - display_name="Save Audio (FLAC)", + display_name="Save Audio (FLAC) (Deprecated)", category="audio", essentials_category="Audio", inputs=[ @@ -169,6 +167,7 @@ class SaveAudio(IO.ComfyNode): ], hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo], is_output_node=True, + is_deprecated=True, ) @classmethod @@ -188,7 +187,7 @@ class SaveAudioMP3(IO.ComfyNode): return IO.Schema( node_id="SaveAudioMP3", search_aliases=["export mp3"], - display_name="Save Audio (MP3)", + display_name="Save Audio (MP3) (Deprecated)", category="audio", essentials_category="Audio", inputs=[ @@ -198,6 +197,7 @@ class SaveAudioMP3(IO.ComfyNode): ], hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo], is_output_node=True, + is_deprecated=True, ) @classmethod @@ -219,7 +219,7 @@ class SaveAudioOpus(IO.ComfyNode): return IO.Schema( node_id="SaveAudioOpus", search_aliases=["export opus"], - display_name="Save Audio (Opus)", + display_name="Save Audio (Opus) (Deprecated)", category="audio", inputs=[ IO.Audio.Input("audio"), @@ -228,6 +228,7 @@ class SaveAudioOpus(IO.ComfyNode): ], hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo], is_output_node=True, + is_deprecated=True, ) @classmethod @@ -243,6 +244,54 @@ class SaveAudioOpus(IO.ComfyNode): save_opus = execute # TODO: remove +class SaveAudioAdvanced(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="SaveAudioAdvanced", + search_aliases=["save audio", "export audio", "output audio", "write audio", "flac", "mp3", "opus"], + display_name="Save Audio (Advanced)", + description="Saves the input audio to your ComfyUI output directory.", + category="audio", + inputs=[ + IO.Audio.Input("audio", tooltip="The audio to save."), + IO.String.Input( + "filename_prefix", + default="audio/ComfyUI", + tooltip=( + "The prefix for the file to save. May include formatting tokens " + "such as %date:yyyy-MM-dd%." + ), + ), + IO.DynamicCombo.Input( + "format", + options=[ + IO.DynamicCombo.Option("flac", []), + IO.DynamicCombo.Option("mp3", [ + IO.Combo.Input("quality", options=["V0", "128k", "320k"], default="V0"), + ]), + IO.DynamicCombo.Option("opus", [ + IO.Combo.Input("quality", options=["64k", "96k", "128k", "192k", "320k"], default="128k"), + ]), + ], + tooltip="The file format in which to save the audio.", + ), + ], + hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo], + is_output_node=True, + ) + + @classmethod + def execute(cls, audio, filename_prefix: str, format: dict) -> IO.NodeOutput: + file_format = format.get("format", None) + quality = format.get("quality", None) + if quality: + ui=UI.AudioSaveHelper.get_save_audio_ui(audio, filename_prefix=filename_prefix, cls=cls, format=file_format, quality=quality) + else: + ui=UI.AudioSaveHelper.get_save_audio_ui(audio, filename_prefix=filename_prefix, cls=cls, format=file_format) + return IO.NodeOutput(ui=ui) + + class PreviewAudio(IO.ComfyNode): @classmethod def define_schema(cls): @@ -824,6 +873,7 @@ class AudioExtension(ComfyExtension): SaveAudio, SaveAudioMP3, SaveAudioOpus, + SaveAudioAdvanced, LoadAudio, PreviewAudio, ConditioningStableAudio, diff --git a/comfy_extras/nodes_audio_encoder.py b/comfy_extras/nodes_audio_encoder.py index 6a85da89b..2ae30d321 100644 --- a/comfy_extras/nodes_audio_encoder.py +++ b/comfy_extras/nodes_audio_encoder.py @@ -11,7 +11,7 @@ class AudioEncoderLoader(io.ComfyNode): return io.Schema( node_id="AudioEncoderLoader", display_name="Load Audio Encoder", - category="loaders", + category="model/loaders", inputs=[ io.Combo.Input( "audio_encoder_name", @@ -36,7 +36,7 @@ class AudioEncoderEncode(io.ComfyNode): def define_schema(cls) -> io.Schema: return io.Schema( node_id="AudioEncoderEncode", - category="conditioning", + category="model/conditioning", inputs=[ io.AudioEncoder.Input("audio_encoder"), io.Audio.Input("audio"), diff --git a/comfy_extras/nodes_bernini.py b/comfy_extras/nodes_bernini.py new file mode 100644 index 000000000..227fa5753 --- /dev/null +++ b/comfy_extras/nodes_bernini.py @@ -0,0 +1,115 @@ +import torch +from typing_extensions import override + +import comfy.model_management +import comfy.utils +import node_helpers +from comfy_api.latest import ComfyExtension, io + + +def _resize_long_edge(image, max_size, stride=16): + """Resize (preserve aspect) so the long edge <= max_size, then snap each side to `stride`""" + h, w = image.shape[1], image.shape[2] + scale = min(max_size / max(h, w), 1.0) + nh = max(stride, round(h * scale / stride) * stride) + nw = max(stride, round(w * scale / stride) * stride) + return comfy.utils.common_upscale(image[:, :, :, :3].movedim(-1, 1), nw, nh, "area", "disabled").movedim(1, -1) + + +class BerniniConditioning(io.ComfyNode): + """Bernini in-context conditioning for a Wan2.2-A14B model. + + Attaches the VAE-encoded source video / reference images to the conditioning + source video first, then each reference image + + The task is inferred from which inputs are connected: + (nothing) -> t2v (text-to-video) + source_video -> v2v (video-to-video) + source_video + ref_images -> rv2v (reference-guided video editing) + ref_images only -> r2v (reference-to-video) + source_video + ref_video -> ads2v (insert image/video into video) + + source_video is the edit base / canvas (resized to width x height). + reference_video is moving content to composite in. + Streams are ordered source_video, reference_video, then reference_images -> source_id (1, 2, 3, ...). + """ + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="BerniniConditioning", + display_name="Bernini Conditioning", + category="conditioning/video_models", + description="Conditioning node for Bernini in-context video/image conditioning. It can be used for the following tasks: t2v (text-to-video), v2v (video-to-video), rv2v (reference-guided video editing), r2v (reference-to-video), ads2v (insert image/video into video)." + "Reference images injected as in-context tokens (r2v, rv2v) are encoded independently at their own native aspect ratio (long edge capped at ref_max_size)", + inputs=[ + io.Conditioning.Input("positive"), + io.Conditioning.Input("negative"), + io.Vae.Input("vae"), + io.Int.Input("width", default=832, min=16, max=8192, step=16), + io.Int.Input("height", default=480, min=16, max=8192, step=16), + io.Int.Input("length", default=81, min=1, max=8192, step=4), + io.Int.Input("batch_size", default=1, min=1, max=4096), + io.Image.Input("source_video", optional=True, tooltip=( + "Source video to edit or restyle (v2v, rv2v). Resized to width/height and trimmed to length.")), + io.Image.Input("reference_video", optional=True, tooltip=( + "Video to insert into the source video (ads2v).")), + io.Autogrow.Input("reference_images", optional=True, + template=io.Autogrow.TemplatePrefix( + input=io.Image.Input("reference_image", tooltip=( + "Reference image injected as an in-context token (r2v, rv2v).")), + prefix="reference_image_", min=0, max=8)), + io.Int.Input("ref_max_size", default=848, min=16, max=8192, step=16, optional=True, tooltip=( + "Max size for the long edge of reference_video and reference_images. Resized with preserved aspect ratio and snapped to 16px.")), + ], + outputs=[ + io.Conditioning.Output(display_name="positive"), + io.Conditioning.Output(display_name="negative"), + io.Latent.Output(display_name="latent"), + ], + ) + + @classmethod + def execute(cls, positive, negative, vae, width, height, length, batch_size, + source_video=None, reference_video=None, reference_images=None, ref_max_size=848) -> io.NodeOutput: + latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], + device=comfy.model_management.intermediate_device()) + + # source_video (1), reference_video (2), reference_images (3, 4, ...). + context = [] + if source_video is not None: + vid = comfy.utils.common_upscale(source_video[:length, :, :, :3].movedim(-1, 1), width, height, "area", "center").movedim(1, -1) + context.append(vae.encode(vid[:, :, :, :3])) + + if reference_video is not None: + ref_vid = _resize_long_edge(reference_video[:length], ref_max_size) # moving content, native aspect + context.append(vae.encode(ref_vid[:, :, :, :3])) + + # reference_images is an autogrow dict {reference_image_0: IMAGE, ...}; each slot is a + # separate stream at its own native aspect (a multi-image batch in one slot -> one stream per frame). + if reference_images: + for name in sorted(reference_images): + imgs = reference_images[name] + if imgs is None: + continue + for i in range(imgs.shape[0]): + img = _resize_long_edge(imgs[i:i + 1], ref_max_size) # native aspect per ref + context.append(vae.encode(img[:, :, :, :3])) + + if context: + positive = node_helpers.conditioning_set_values(positive, {"context_latents": context}) + negative = node_helpers.conditioning_set_values(negative, {"context_latents": context}) + + return io.NodeOutput(positive, negative, {"samples": latent}) + + +class BerniniExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + BerniniConditioning, + ] + + +async def comfy_entrypoint() -> BerniniExtension: + return BerniniExtension() diff --git a/comfy_extras/nodes_bg_removal.py b/comfy_extras/nodes_bg_removal.py index 793fd802b..c7b33a821 100644 --- a/comfy_extras/nodes_bg_removal.py +++ b/comfy_extras/nodes_bg_removal.py @@ -11,7 +11,7 @@ class LoadBackgroundRemovalModel(IO.ComfyNode): return IO.Schema( node_id="LoadBackgroundRemovalModel", display_name="Load Background Removal Model", - category="loaders", + category="model/loaders", inputs=[ IO.Combo.Input("bg_removal_name", options=sorted(files), tooltip="The model used to remove backgrounds from images"), ], @@ -36,15 +36,15 @@ class RemoveBackground(IO.ComfyNode): category="image/background removal", description="Generates a foreground mask to remove the background from an image using a background removal model.", inputs=[ - IO.Image.Input("image", tooltip="Input image to remove the background from"), - IO.BackgroundRemoval.Input("bg_removal_model", tooltip="Background removal model used to generate the mask") + IO.BackgroundRemoval.Input("bg_removal_model", tooltip="Background removal model used to generate the mask"), + IO.Image.Input("image", tooltip="Input image to remove the background from") ], outputs=[ IO.Mask.Output("mask", tooltip="Generated foreground mask") ] ) @classmethod - def execute(cls, image, bg_removal_model): + def execute(cls, bg_removal_model, image): mask = bg_removal_model.encode_image(image) return IO.NodeOutput(mask) diff --git a/comfy_extras/nodes_camera_trajectory.py b/comfy_extras/nodes_camera_trajectory.py index 34b78e81b..13a1448f4 100644 --- a/comfy_extras/nodes_camera_trajectory.py +++ b/comfy_extras/nodes_camera_trajectory.py @@ -153,7 +153,7 @@ class WanCameraEmbedding(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="WanCameraEmbedding", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Combo.Input( "camera_pose", diff --git a/comfy_extras/nodes_cfg.py b/comfy_extras/nodes_cfg.py index 4ebb4b51e..b585c560f 100644 --- a/comfy_extras/nodes_cfg.py +++ b/comfy_extras/nodes_cfg.py @@ -57,24 +57,55 @@ class CFGNorm(io.ComfyNode): inputs=[ io.Model.Input("model"), io.Float.Input("strength", default=1.0, min=0.0, max=100.0, step=0.01), + io.Boolean.Input( + "pre_cfg", + default=False, + optional=True, + tooltip=( + "If true, rescale the combined noise BEFORE the sampler's CFG combine, " + "without clamping (can amplify). Matches the norm-scaled CFG used by " + "models like Lens. Default false keeps the original post-CFG x0-space " + "attenuate-only behavior." + ), + ), ], outputs=[io.Model.Output(display_name="patched_model")], is_experimental=True, ) @classmethod - def execute(cls, model, strength) -> io.NodeOutput: + def execute(cls, model, strength, pre_cfg=False) -> io.NodeOutput: m = model.clone() - def cfg_norm(args): - cond_p = args['cond_denoised'] - pred_text_ = args["denoised"] + if pre_cfg: + def cfg_norm_pre(args): + cond = args["cond"] + uncond = args["uncond"] + cond_scale = args["cond_scale"] + comb = uncond + cond_scale * (cond - uncond) + cond_norm = torch.linalg.vector_norm(cond, dim=1, keepdim=True) + comb_norm = torch.linalg.vector_norm(comb, dim=1, keepdim=True) + rescale = torch.where( + comb_norm > 0, + cond_norm / comb_norm.clamp_min(1e-12), + torch.ones_like(comb_norm), + ) + rescaled = comb * rescale + # strength blends back toward standard linear CFG (1.0 = full rescale). + if strength != 1.0: + rescaled = strength * rescaled + (1.0 - strength) * comb + return rescaled + m.set_model_sampler_cfg_function(cfg_norm_pre) + else: + def cfg_norm(args): + cond_p = args['cond_denoised'] + pred_text_ = args["denoised"] - norm_full_cond = torch.norm(cond_p, dim=1, keepdim=True) - norm_pred_text = torch.norm(pred_text_, dim=1, keepdim=True) - scale = (norm_full_cond / (norm_pred_text + 1e-8)).clamp(min=0.0, max=1.0) - return pred_text_ * scale * strength + norm_full_cond = torch.norm(cond_p, dim=1, keepdim=True) + norm_pred_text = torch.norm(pred_text_, dim=1, keepdim=True) + scale = (norm_full_cond / (norm_pred_text + 1e-8)).clamp(min=0.0, max=1.0) + return pred_text_ * scale * strength - m.set_model_sampler_post_cfg_function(cfg_norm) + m.set_model_sampler_post_cfg_function(cfg_norm) return io.NodeOutput(m) diff --git a/comfy_extras/nodes_chroma_radiance.py b/comfy_extras/nodes_chroma_radiance.py index 509436062..a4f673001 100644 --- a/comfy_extras/nodes_chroma_radiance.py +++ b/comfy_extras/nodes_chroma_radiance.py @@ -13,7 +13,7 @@ class EmptyChromaRadianceLatentImage(io.ComfyNode): def define_schema(cls) -> io.Schema: return io.Schema( node_id="EmptyChromaRadianceLatentImage", - category="latent/chroma_radiance", + category="model/latent/chroma_radiance", inputs=[ io.Int.Input(id="width", default=1024, min=16, max=nodes.MAX_RESOLUTION, step=16), io.Int.Input(id="height", default=1024, min=16, max=nodes.MAX_RESOLUTION, step=16), @@ -33,7 +33,7 @@ class ChromaRadianceOptions(io.ComfyNode): def define_schema(cls) -> io.Schema: return io.Schema( node_id="ChromaRadianceOptions", - category="model_patches/chroma_radiance", + category="model/patch/chroma_radiance", description="Allows setting advanced options for the Chroma Radiance model.", inputs=[ io.Model.Input(id="model"), @@ -65,6 +65,12 @@ class ChromaRadianceOptions(io.ComfyNode): tooltip="Allows overriding the default NeRF tile size. -1 means use the default (32). 0 means use non-tiling mode (may require a lot of VRAM).", advanced=True, ), + io.Boolean.Input( + id="force_sequential_txt_ids", + default=False, + tooltip="Force usage of sequential text token IDs instead of zeroes. Should be used for checkpoints from 2026-05-22 to 2026-06-01 that are trained in this way but do not contain the __sequential__ key in the state dict.", + advanced=True, + ), ], outputs=[io.Model.Output()], ) @@ -78,11 +84,15 @@ class ChromaRadianceOptions(io.ComfyNode): start_sigma: float, end_sigma: float, nerf_tile_size: int, + force_sequential_txt_ids: bool, ) -> io.NodeOutput: radiance_options = {} if nerf_tile_size >= 0: radiance_options["nerf_tile_size"] = nerf_tile_size + if force_sequential_txt_ids: + radiance_options["use_sequential_txt_ids"] = True + if not radiance_options: return io.NodeOutput(model) diff --git a/comfy_extras/nodes_color.py b/comfy_extras/nodes_color.py index 80ba121cd..688254e4e 100644 --- a/comfy_extras/nodes_color.py +++ b/comfy_extras/nodes_color.py @@ -7,29 +7,29 @@ class ColorToRGBInt(io.ComfyNode): def define_schema(cls) -> io.Schema: return io.Schema( node_id="ColorToRGBInt", - display_name="Color to RGB Int", - category="utils", - description="Convert a color to a RGB integer value.", + display_name="Color Picker", + category="utilities", + description="Return a color RGB integer value and hexadecimal representation.", inputs=[ io.Color.Input("color"), ], outputs=[ io.Int.Output(display_name="rgb_int"), + io.Color.Output(display_name="hex") ], ) @classmethod - def execute( - cls, - color: str, - ) -> io.NodeOutput: + def execute(cls, color: str) -> io.NodeOutput: # expect format #RRGGBB if len(color) != 7 or color[0] != "#": raise ValueError("Color must be in format #RRGGBB") r = int(color[1:3], 16) g = int(color[3:5], 16) b = int(color[5:7], 16) - return io.NodeOutput(r * 256 * 256 + g * 256 + b) + + rgb_int = r * 256 * 256 + g * 256 + b + return io.NodeOutput(rgb_int, color) class ColorExtension(ComfyExtension): diff --git a/comfy_extras/nodes_context_windows.py b/comfy_extras/nodes_context_windows.py index f7ca833dc..d9e32b9d9 100644 --- a/comfy_extras/nodes_context_windows.py +++ b/comfy_extras/nodes_context_windows.py @@ -1,4 +1,3 @@ -from __future__ import annotations from comfy_api.latest import ComfyExtension, io import comfy.context_windows import nodes @@ -10,7 +9,7 @@ class ContextWindowsManualNode(io.ComfyNode): return io.Schema( node_id="ContextWindowsManual", display_name="Context Windows (Manual)", - category="model_patches", + category="model/patch", description="Manually set context windows.", inputs=[ io.Model.Input("model", tooltip="The model to apply context windows to during sampling."), diff --git a/comfy_extras/nodes_controlnet.py b/comfy_extras/nodes_controlnet.py index 847cb0bdf..17d965405 100644 --- a/comfy_extras/nodes_controlnet.py +++ b/comfy_extras/nodes_controlnet.py @@ -9,7 +9,7 @@ class SetUnionControlNetType(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SetUnionControlNetType", - category="conditioning/controlnet", + category="model/conditioning/controlnet", inputs=[ io.ControlNet.Input("control_net"), io.Combo.Input("type", options=["auto"] + list(UNION_CONTROLNET_TYPES.keys())), @@ -39,7 +39,7 @@ class ControlNetInpaintingAliMamaApply(io.ComfyNode): return io.Schema( node_id="ControlNetInpaintingAliMamaApply", search_aliases=["masked controlnet"], - category="conditioning/controlnet", + category="model/conditioning/controlnet", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), diff --git a/comfy_extras/nodes_cosmos.py b/comfy_extras/nodes_cosmos.py index 7dd129d19..d754ab442 100644 --- a/comfy_extras/nodes_cosmos.py +++ b/comfy_extras/nodes_cosmos.py @@ -13,7 +13,7 @@ class EmptyCosmosLatentVideo(io.ComfyNode): def define_schema(cls) -> io.Schema: return io.Schema( node_id="EmptyCosmosLatentVideo", - category="latent/video", + category="model/latent/video", inputs=[ io.Int.Input("width", default=1280, min=16, max=nodes.MAX_RESOLUTION, step=16), io.Int.Input("height", default=704, min=16, max=nodes.MAX_RESOLUTION, step=16), @@ -45,7 +45,7 @@ class CosmosImageToVideoLatent(io.ComfyNode): def define_schema(cls) -> io.Schema: return io.Schema( node_id="CosmosImageToVideoLatent", - category="conditioning/inpaint", + category="model/conditioning/inpaint", inputs=[ io.Vae.Input("vae"), io.Int.Input("width", default=1280, min=16, max=nodes.MAX_RESOLUTION, step=16), @@ -88,7 +88,7 @@ class CosmosPredict2ImageToVideoLatent(io.ComfyNode): def define_schema(cls) -> io.Schema: return io.Schema( node_id="CosmosPredict2ImageToVideoLatent", - category="conditioning/inpaint", + category="model/conditioning/inpaint", inputs=[ io.Vae.Input("vae"), io.Int.Input("width", default=848, min=16, max=nodes.MAX_RESOLUTION, step=16), diff --git a/comfy_extras/nodes_curve.py b/comfy_extras/nodes_curve.py index 9803e8034..aa2d94bb6 100644 --- a/comfy_extras/nodes_curve.py +++ b/comfy_extras/nodes_curve.py @@ -1,5 +1,3 @@ -from __future__ import annotations - import numpy as np from comfy_api.latest import ComfyExtension, io @@ -13,7 +11,7 @@ class CurveEditor(io.ComfyNode): return io.Schema( node_id="CurveEditor", display_name="Curve Editor", - category="utils", + category="utilities", inputs=[ io.Curve.Input("curve"), io.Histogram.Input("histogram", optional=True), @@ -40,7 +38,7 @@ class ImageHistogram(io.ComfyNode): return io.Schema( node_id="ImageHistogram", display_name="Image Histogram", - category="utils", + category="utilities", inputs=[ io.Image.Input("image"), ], diff --git a/comfy_extras/nodes_custom_sampler.py b/comfy_extras/nodes_custom_sampler.py index 10b56b91c..3e97084a4 100644 --- a/comfy_extras/nodes_custom_sampler.py +++ b/comfy_extras/nodes_custom_sampler.py @@ -1,5 +1,7 @@ import math import comfy.samplers +import comfy.sampler_helpers +import comfy.patcher_extension import comfy.sample from comfy.k_diffusion import sampling as k_diffusion_sampling from comfy.k_diffusion import sa_solver @@ -17,7 +19,7 @@ class BasicScheduler(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="BasicScheduler", - category="sampling/schedulers", + category="model/sampling/schedulers", inputs=[ io.Model.Input("model"), io.Combo.Input("scheduler", options=comfy.samplers.SCHEDULER_NAMES), @@ -47,7 +49,7 @@ class KarrasScheduler(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="KarrasScheduler", - category="sampling/schedulers", + category="model/sampling/schedulers", inputs=[ io.Int.Input("steps", default=20, min=1, max=10000), io.Float.Input("sigma_max", default=14.614642, min=0.0, max=5000.0, step=0.01, round=False, advanced=True), @@ -69,7 +71,7 @@ class ExponentialScheduler(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="ExponentialScheduler", - category="sampling/schedulers", + category="model/sampling/schedulers", inputs=[ io.Int.Input("steps", default=20, min=1, max=10000), io.Float.Input("sigma_max", default=14.614642, min=0.0, max=5000.0, step=0.01, round=False, advanced=True), @@ -90,7 +92,7 @@ class PolyexponentialScheduler(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="PolyexponentialScheduler", - category="sampling/schedulers", + category="model/sampling/schedulers", inputs=[ io.Int.Input("steps", default=20, min=1, max=10000), io.Float.Input("sigma_max", default=14.614642, min=0.0, max=5000.0, step=0.01, round=False, advanced=True), @@ -112,7 +114,7 @@ class LaplaceScheduler(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LaplaceScheduler", - category="sampling/schedulers", + category="model/sampling/schedulers", inputs=[ io.Int.Input("steps", default=20, min=1, max=10000), io.Float.Input("sigma_max", default=14.614642, min=0.0, max=5000.0, step=0.01, round=False, advanced=True), @@ -136,7 +138,7 @@ class SDTurboScheduler(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SDTurboScheduler", - category="sampling/schedulers", + category="model/sampling/schedulers", inputs=[ io.Model.Input("model"), io.Int.Input("steps", default=1, min=1, max=10), @@ -160,7 +162,7 @@ class BetaSamplingScheduler(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="BetaSamplingScheduler", - category="sampling/schedulers", + category="model/sampling/schedulers", inputs=[ io.Model.Input("model"), io.Int.Input("steps", default=20, min=1, max=10000), @@ -182,7 +184,7 @@ class VPScheduler(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="VPScheduler", - category="sampling/schedulers", + category="model/sampling/schedulers", inputs=[ io.Int.Input("steps", default=20, min=1, max=10000), io.Float.Input("beta_d", default=19.9, min=0.0, max=5000.0, step=0.01, round=False, advanced=True), #TODO: fix default values @@ -204,7 +206,7 @@ class SplitSigmas(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SplitSigmas", - category="sampling/sigmas", + category="model/sampling/sigmas", inputs=[ io.Sigmas.Input("sigmas"), io.Int.Input("step", default=0, min=0, max=10000), @@ -228,7 +230,7 @@ class SplitSigmasDenoise(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SplitSigmasDenoise", - category="sampling/sigmas", + category="model/sampling/sigmas", inputs=[ io.Sigmas.Input("sigmas"), io.Float.Input("denoise", default=1.0, min=0.0, max=1.0, step=0.01), @@ -254,7 +256,7 @@ class FlipSigmas(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="FlipSigmas", - category="sampling/sigmas", + category="model/sampling/sigmas", inputs=[io.Sigmas.Input("sigmas")], outputs=[io.Sigmas.Output()] ) @@ -276,7 +278,7 @@ class SetFirstSigma(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SetFirstSigma", - category="sampling/sigmas", + category="model/sampling/sigmas", inputs=[ io.Sigmas.Input("sigmas"), io.Float.Input("sigma", default=136.0, min=0.0, max=20000.0, step=0.001, round=False), @@ -298,7 +300,7 @@ class ExtendIntermediateSigmas(io.ComfyNode): return io.Schema( node_id="ExtendIntermediateSigmas", search_aliases=["interpolate sigmas"], - category="sampling/sigmas", + category="model/sampling/sigmas", inputs=[ io.Sigmas.Input("sigmas"), io.Int.Input("steps", default=2, min=1, max=100), @@ -351,7 +353,7 @@ class SamplingPercentToSigma(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SamplingPercentToSigma", - category="sampling/sigmas", + category="model/sampling/sigmas", inputs=[ io.Model.Input("model"), io.Float.Input("sampling_percent", default=0.0, min=0.0, max=1.0, step=0.0001), @@ -379,7 +381,7 @@ class KSamplerSelect(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="KSamplerSelect", - category="sampling/samplers", + category="model/sampling/samplers", inputs=[io.Combo.Input("sampler_name", options=comfy.samplers.SAMPLER_NAMES)], outputs=[io.Sampler.Output()] ) @@ -396,7 +398,7 @@ class SamplerDPMPP_3M_SDE(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SamplerDPMPP_3M_SDE", - category="sampling/samplers", + category="model/sampling/samplers", inputs=[ io.Float.Input("eta", default=1.0, min=0.0, max=100.0, step=0.01, round=False, advanced=True), io.Float.Input("s_noise", default=1.0, min=0.0, max=100.0, step=0.01, round=False, advanced=True), @@ -421,7 +423,7 @@ class SamplerDPMPP_2M_SDE(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SamplerDPMPP_2M_SDE", - category="sampling/samplers", + category="model/sampling/samplers", inputs=[ io.Combo.Input("solver_type", options=['midpoint', 'heun']), io.Float.Input("eta", default=1.0, min=0.0, max=100.0, step=0.01, round=False, advanced=True), @@ -448,7 +450,7 @@ class SamplerDPMPP_SDE(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SamplerDPMPP_SDE", - category="sampling/samplers", + category="model/sampling/samplers", inputs=[ io.Float.Input("eta", default=1.0, min=0.0, max=100.0, step=0.01, round=False, advanced=True), io.Float.Input("s_noise", default=1.0, min=0.0, max=100.0, step=0.01, round=False, advanced=True), @@ -474,7 +476,7 @@ class SamplerDPMPP_2S_Ancestral(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SamplerDPMPP_2S_Ancestral", - category="sampling/samplers", + category="model/sampling/samplers", inputs=[ io.Float.Input("eta", default=1.0, min=0.0, max=100.0, step=0.01, round=False), io.Float.Input("s_noise", default=1.0, min=0.0, max=100.0, step=0.01, round=False), @@ -494,7 +496,7 @@ class SamplerEulerAncestral(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SamplerEulerAncestral", - category="sampling/samplers", + category="model/sampling/samplers", inputs=[ io.Float.Input("eta", default=1.0, min=0.0, max=100.0, step=0.01, round=False, advanced=True), io.Float.Input("s_noise", default=1.0, min=0.0, max=100.0, step=0.01, round=False, advanced=True), @@ -515,7 +517,7 @@ class SamplerEulerAncestralCFGPP(io.ComfyNode): return io.Schema( node_id="SamplerEulerAncestralCFGPP", display_name="SamplerEulerAncestralCFG++", - category="sampling/samplers", + category="model/sampling/samplers", inputs=[ io.Float.Input("eta", default=1.0, min=0.0, max=1.0, step=0.01, round=False), io.Float.Input("s_noise", default=1.0, min=0.0, max=10.0, step=0.01, round=False), @@ -537,7 +539,7 @@ class SamplerLMS(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SamplerLMS", - category="sampling/samplers", + category="model/sampling/samplers", inputs=[io.Int.Input("order", default=4, min=1, max=100, advanced=True)], outputs=[io.Sampler.Output()] ) @@ -554,7 +556,7 @@ class SamplerDPMAdaptative(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SamplerDPMAdaptative", - category="sampling/samplers", + category="model/sampling/samplers", inputs=[ io.Int.Input("order", default=3, min=2, max=3, advanced=True), io.Float.Input("rtol", default=0.05, min=0.0, max=100.0, step=0.01, round=False, advanced=True), @@ -585,7 +587,7 @@ class SamplerER_SDE(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SamplerER_SDE", - category="sampling/samplers", + category="model/sampling/samplers", inputs=[ io.Combo.Input("solver_type", options=["ER-SDE", "Reverse-time SDE", "ODE"]), io.Int.Input("max_stage", default=3, min=1, max=3, advanced=True), @@ -623,7 +625,7 @@ class SamplerSASolver(io.ComfyNode): return io.Schema( node_id="SamplerSASolver", search_aliases=["sde"], - category="sampling/samplers", + category="model/sampling/samplers", inputs=[ io.Model.Input("model"), io.Float.Input("eta", default=1.0, min=0.0, max=10.0, step=0.01, round=False, advanced=True), @@ -668,7 +670,7 @@ class SamplerSEEDS2(io.ComfyNode): return io.Schema( node_id="SamplerSEEDS2", search_aliases=["sde", "exp heun"], - category="sampling/samplers", + category="model/sampling/samplers", inputs=[ io.Combo.Input("solver_type", options=["phi_1", "phi_2"]), io.Float.Input("eta", default=1.0, min=0.0, max=100.0, step=0.01, round=False, tooltip="Stochastic strength", advanced=True), @@ -727,7 +729,7 @@ class SamplerCustom(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SamplerCustom", - category="sampling/custom_sampling", + category="model/sampling/custom_sampling", inputs=[ io.Model.Input("model"), io.Boolean.Input("add_noise", default=True, advanced=True), @@ -795,7 +797,7 @@ class BasicGuider(io.ComfyNode): return io.Schema( node_id="BasicGuider", display_name="Basic Guider", - category="sampling/guiders", + category="model/sampling/guiders", inputs=[ io.Model.Input("model"), io.Conditioning.Input("conditioning"), @@ -817,7 +819,7 @@ class CFGGuider(io.ComfyNode): return io.Schema( node_id="CFGGuider", display_name="CFG Guider", - category="sampling/guiders", + category="model/sampling/guiders", inputs=[ io.Model.Input("model"), io.Conditioning.Input("positive"), @@ -872,7 +874,7 @@ class DualCFGGuider(io.ComfyNode): node_id="DualCFGGuider", search_aliases=["dual prompt guidance"], display_name="Dual CFG Guider", - category="sampling/guiders", + category="model/sampling/guiders", inputs=[ io.Model.Input("model"), io.Conditioning.Input("cond1"), @@ -894,13 +896,92 @@ class DualCFGGuider(io.ComfyNode): get_guider = execute +class Guider_DualModel(comfy.samplers.CFGGuider): + # Runs the positive (cond) pass on the main model and the negative (uncond) pass on a separate model + def __init__(self, model_patcher, uncond_model_patcher): + super().__init__(model_patcher) + self.uncond_model_patcher = uncond_model_patcher + self.uncond_inner = None + + def outer_sample(self, noise, latent_image, sampler, sigmas, denoise_mask=None, callback=None, disable_pbar=False, seed=None, latent_shapes=None): + self.uncond_inner = None + self.uncond_loaded = [] + self._uncond_neg = None + # skip at cfg 1.0 + if not math.isclose(self.cfg, 1.0): + uc = {"negative": list(map(lambda a: a.copy(), self.conds["negative"]))} + self.uncond_inner, uc, self.uncond_loaded = comfy.sampler_helpers.prepare_sampling( + self.uncond_model_patcher, noise.shape, uc, self.uncond_model_patcher.model_options) + self._uncond_neg = uc["negative"] + self.uncond_model_patcher.pre_run() + try: + return super().outer_sample(noise, latent_image, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=latent_shapes) + finally: + if self.uncond_inner is not None: + self.uncond_model_patcher.cleanup() + comfy.sampler_helpers.cleanup_models({"negative": self._uncond_neg}, self.uncond_loaded) + self.uncond_inner = None + + def inner_sample(self, noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=None): + if self.uncond_inner is not None: + li = latent_image + if li is not None and torch.count_nonzero(li) > 0: + li = self.uncond_inner.process_latent_in(li) + self._uncond_conds = comfy.samplers.process_conds( + self.uncond_inner, noise, {"negative": self._uncond_neg}, device, li, denoise_mask, seed, latent_shapes=latent_shapes)["negative"] + return super().inner_sample(noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=latent_shapes) + + def predict_noise(self, x, timestep, model_options={}, seed=None): + positive = self.conds.get("positive", None) + cond = comfy.samplers.calc_cond_batch(self.inner_model, [positive], x, timestep, model_options)[0] + # uncond model not loaded (base cfg==1/no negative), or cfg driven to 1.0 this step -> single model, cond only + if self.uncond_inner is None or (math.isclose(self.cfg, 1.0) and not model_options.get("disable_cfg1_optimization", False)): + return cond + + uncond_model_options = model_options + if "multigpu_clones" in model_options: # TODO: support multigpu instead of just running uncond on a single GPU + uncond_model_options = {k: v for k, v in model_options.items() if k != "multigpu_clones"} + uncond = comfy.samplers.calc_cond_batch(self.uncond_inner, [self._uncond_conds], x, timestep, uncond_model_options)[0] + return comfy.samplers.cfg_function(self.inner_model, cond, uncond, self.cfg, x, timestep, + model_options=model_options, cond=positive, uncond=self._uncond_conds) + +class DualModelGuider(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="DualModelGuider", + display_name="Dual Model CFG Guider", + category="model/sampling/guiders", + is_experimental=True, + inputs=[ + io.Model.Input("model", tooltip="Model used for the positive (conditional) pass."), + io.Model.Input("model_negative", optional=True, tooltip="Model used for the negative (unconditional) pass. Use the same model for ordinary CFG."), + io.Conditioning.Input("positive"), + io.Float.Input("cfg", default=4.0, min=0.0, max=100.0, step=0.1, round=0.01), + io.Conditioning.Input("negative", optional=True, tooltip="Negative conditioning run on the negative model. Leave unconnected for a text-free (image-only) unconditional pass."), + ], + outputs=[io.Guider.Output()], + ) + + @classmethod + def execute(cls, model, positive, cfg, model_negative=None, negative=None) -> io.NodeOutput: + if negative is None: + negative = [[None, {}]] # null cond -> no cross_attn -> model runs image-only + + guider = Guider_DualModel(model, model_negative) if model_negative is not None else comfy.samplers.CFGGuider(model) + guider.set_conds(positive, negative) + guider.set_cfg(cfg) + return io.NodeOutput(guider) + + get_guider = execute + class DisableNoise(io.ComfyNode): @classmethod def define_schema(cls): return io.Schema( node_id="DisableNoise", search_aliases=["zero noise"], - category="sampling/noise", + category="model/sampling/noise", inputs=[], outputs=[io.Noise.Output()] ) @@ -917,7 +998,7 @@ class RandomNoise(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="RandomNoise", - category="sampling/noise", + category="model/sampling/noise", inputs=[io.Int.Input("noise_seed", default=0, min=0, max=0xffffffffffffffff, control_after_generate=True)], outputs=[io.Noise.Output()] ) @@ -934,7 +1015,7 @@ class SamplerCustomAdvanced(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SamplerCustomAdvanced", - category="sampling/custom_sampling", + category="model/sampling/custom_sampling", inputs=[ io.Noise.Input("noise"), io.Guider.Input("guider"), @@ -1054,11 +1135,53 @@ class ManualSigmas(io.ComfyNode): sigmas = torch.FloatTensor(sigmas) return io.NodeOutput(sigmas) +class CFGOverride(io.ComfyNode): + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="CFGOverride", + display_name="CFG Override", + description="Override cfg to a fixed value over a [start, end] percent (sigma) range. " + "With multiple overrides, the one nearest the sampler wins on overlap.", + category="sampling/custom_sampling", + inputs=[ + io.Model.Input("model"), + io.Float.Input("cfg", default=1.0, min=0.0, max=100.0, step=0.1, round=0.01), + io.Float.Input("start_percent", default=0.0, min=0.0, max=1.0, step=0.001), + io.Float.Input("end_percent", default=1.0, min=0.0, max=1.0, step=0.001), + ], + outputs=[io.Model.Output()], + ) + + @classmethod + def execute(cls, model, cfg, start_percent, end_percent) -> io.NodeOutput: + ms = model.get_model_object("model_sampling") + sigma_hi = ms.percent_to_sigma(start_percent) # percent->sigma decreasing, so hi >= lo + sigma_lo = ms.percent_to_sigma(end_percent) + + def predict_noise_wrapper(executor, *args, **kwargs): + sigma = float(args[1].flatten()[0]) # args = (x, timestep, model_options, seed) + if not (sigma_lo <= sigma <= sigma_hi): + return executor(*args, **kwargs) + guider = executor.class_obj # guider.cfg feeds cond_scale + saved = guider.cfg + guider.cfg = cfg + try: + return executor(*args, **kwargs) + finally: + guider.cfg = saved # restore for other steps/overrides + + m = model.clone() + m.add_wrapper(comfy.patcher_extension.WrappersMP.PREDICT_NOISE, predict_noise_wrapper) + return io.NodeOutput(m) + + class CustomSamplersExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[io.ComfyNode]]: return [ SamplerCustom, + CFGOverride, BasicScheduler, KarrasScheduler, ExponentialScheduler, @@ -1087,6 +1210,7 @@ class CustomSamplersExtension(ComfyExtension): SamplingPercentToSigma, CFGGuider, DualCFGGuider, + DualModelGuider, BasicGuider, RandomNoise, DisableNoise, diff --git a/comfy_extras/nodes_dataset.py b/comfy_extras/nodes_dataset.py index 22f5ff203..0253b4b4f 100644 --- a/comfy_extras/nodes_dataset.py +++ b/comfy_extras/nodes_dataset.py @@ -157,7 +157,7 @@ class LoadImageTextDataSetFromFolderNode(io.ComfyNode): return io.NodeOutput(output_tensor, captions) -def save_images_to_folder(image_list, output_dir, prefix="image"): +def save_images_to_folder(image_list, output_dir, prefix="image", overwrite=True): """Utility function to save a list of image tensors to disk. Args: @@ -197,7 +197,11 @@ def save_images_to_folder(image_list, output_dir, prefix="image"): raise ValueError(f"Expected torch.Tensor, got {type(img_tensor)}") # Save image - filename = f"{prefix}_{idx:05d}.png" + if overwrite: + filename = f"{prefix}_{idx:05d}.png" + else: + _, _, counter, _, resolved_prefix = folder_paths.get_save_image_path(prefix, output_dir) + filename = f"{resolved_prefix}_{counter:05}_{idx:05d}.png" filepath = os.path.join(output_dir, filename) img.save(filepath) saved_files.append(filename) @@ -230,19 +234,26 @@ class SaveImageDataSetToFolderNode(io.ComfyNode): tooltip="Prefix for saved image filenames.", advanced=True, ), + io.Combo.Input( + "mode", + default="overwrite", + options=["overwrite", "increment"], + tooltip="Whether to overwrite existing files or increment filenames to avoid overwriting." + ), ], outputs=[], is_deprecated=True, # This node is redundant and superseded by existing Save Image nodes where the target folder can be specified in the filename_prefix ) @classmethod - def execute(cls, images, folder_name, filename_prefix): + def execute(cls, images, folder_name, filename_prefix, mode): # Extract scalar values folder_name = folder_name[0] filename_prefix = filename_prefix[0] + mode = mode[0] output_dir = os.path.join(folder_paths.get_output_directory(), folder_name) - saved_files = save_images_to_folder(images, output_dir, filename_prefix) + saved_files = save_images_to_folder(images, output_dir, filename_prefix, mode=='overwrite') logging.info(f"Saved {len(saved_files)} images to {output_dir}.") return io.NodeOutput() @@ -278,18 +289,25 @@ class SaveImageTextDataSetToFolderNode(io.ComfyNode): tooltip="Prefix for saved image filenames.", advanced=True, ), + io.Combo.Input( + "mode", + default="overwrite", + options=["overwrite", "increment"], + tooltip="Whether to overwrite existing files or increment filenames to avoid overwriting." + ), ], outputs=[], ) @classmethod - def execute(cls, images, folder_name, filename_prefix, texts=None): + def execute(cls, images, folder_name, filename_prefix, mode, texts=None): # Extract scalar values folder_name = folder_name[0] filename_prefix = filename_prefix[0] + mode = mode[0] output_dir = os.path.join(folder_paths.get_output_directory(), folder_name) - saved_files = save_images_to_folder(images, output_dir, filename_prefix) + saved_files = save_images_to_folder(images, output_dir, filename_prefix, mode=='overwrite') # Save captions if texts: @@ -393,6 +411,21 @@ class ImageProcessingNode(io.ComfyNode): return has_group + @classmethod + def _ensure_image_list(cls, images): + """Normalize to a flat list of [1, H, W, C] tensors.""" + if isinstance(images, torch.Tensor): + if images.ndim != 4: + raise ValueError(f"Expected 4D image tensor, got shape {tuple(images.shape)}") + return [images[i:i+1] for i in range(images.shape[0])] + + flat = [] + for item in images: + if not isinstance(item, torch.Tensor) or item.ndim != 4: + raise ValueError(f"Expected 4D image tensor, got {type(item).__name__} shape {getattr(item, 'shape', None)}") + flat.extend([item[i:i+1] for i in range(item.shape[0])]) + return flat + @classmethod def define_schema(cls): if cls.node_id is None: @@ -440,6 +473,9 @@ class ImageProcessingNode(io.ComfyNode): """Execute the node. Routes to _process or _group_process based on mode.""" is_group = cls._detect_processing_mode() + if is_group: + images = cls._ensure_image_list(images) + # Extract scalar values from lists for parameters params = {} for k, v in kwargs.items(): @@ -574,7 +610,7 @@ class TextProcessingNode(io.ComfyNode): return io.Schema( node_id=cls.node_id, display_name=cls.display_name or cls.node_id, - category="dataset/text", + category="text", is_experimental=True, is_input_list=is_group, # True for group, False for individual inputs=inputs, @@ -1208,7 +1244,7 @@ class ResolutionBucket(io.ComfyNode): node_id="ResolutionBucket", search_aliases=["bucket by resolution", "group by resolution", "batch by resolution"], display_name="Resolution Bucket", - category="training", + category="model/training", description="Group latents and conditionings into buckets", is_experimental=True, is_input_list=True, @@ -1302,7 +1338,7 @@ class MakeTrainingDataset(io.ComfyNode): node_id="MakeTrainingDataset", search_aliases=["encode dataset"], display_name="Make Training Dataset", - category="training", + category="model/training", description="Encode images with VAE and texts with CLIP to create a training dataset of latents and conditionings.", is_experimental=True, is_input_list=True, # images and texts as lists @@ -1390,7 +1426,7 @@ class SaveTrainingDataset(io.ComfyNode): node_id="SaveTrainingDataset", search_aliases=["export dataset", "save dataset"], display_name="Save Training Dataset", - category="training", + category="model/training", description="Save encoded training dataset (latents + conditioning) to disk for efficient loading during training.", is_experimental=True, is_output_node=True, @@ -1493,7 +1529,7 @@ class LoadTrainingDataset(io.ComfyNode): node_id="LoadTrainingDataset", search_aliases=["import dataset", "training data"], display_name="Load Training Dataset", - category="training", + category="model/training", description="Load encoded training dataset (latents + conditioning) from disk for use in training.", is_experimental=True, inputs=[ diff --git a/comfy_extras/nodes_depth_anything_3.py b/comfy_extras/nodes_depth_anything_3.py new file mode 100644 index 000000000..020112515 --- /dev/null +++ b/comfy_extras/nodes_depth_anything_3.py @@ -0,0 +1,681 @@ +"""ComfyUI nodes for Depth Anything 3. +Model capability matrix: + +Variant head_type has_sky has_conf cam_dec +DA3-Small dualdpt False True yes +DA3-Base dualdpt False True yes +DA3-Mono-Large dpt True False no +DA3-Metric-Large dpt True False no (raw output is metres) +""" + +from __future__ import annotations + +import logging +from typing_extensions import override + +import torch + +import comfy.model_management as mm +import comfy.sd +import folder_paths +from comfy.ldm.colormap import turbo as _turbo +from comfy.ldm.depth_anything_3 import preprocess as da3_preprocess +from comfy_api.latest import ComfyExtension, Types, io +from comfy.ldm.moge.geometry import triangulate_grid_mesh + +DA3ModelType = io.Custom("DA3_MODEL") +DA3Geometry = io.Custom("DA3_GEOMETRY") +DA3PointCloud = io.Custom("DA3_POINT_CLOUD") + +# DA3_GEOMETRY is a dict with these optional keys (absent when the upstream model didn't produce them): +# +# Per-frame tensors - B = batch size in mono mode; B = S (number of views) in multi-view mode. +# "depth": torch.Tensor (B, H, W) -- raw model depth (always present; matches MoGe convention) +# "image": torch.Tensor (B, H, W, 3) -- source image in [0, 1], CPU (always present) +# "mode": str -- "mono" or "multiview" (always present) +# "sky": torch.Tensor (B, H, W) -- sky probability in [0, 1] (Mono/Metric variants only) +# "confidence": torch.Tensor (B, H, W) -- raw model confidence output (Small/Base variants only) +# +# Multi-view only - S = number of views; the leading 1 is the scene dimension from the model. +# "extrinsics": torch.Tensor (1, S, 3, 4) -- world-to-camera [R|t] matrices +# "intrinsics": torch.Tensor (1, S, 3, 3) -- pixel-space intrinsics +# +# DA3_POINT_CLOUD is a dict: +# "points": torch.Tensor (N, 3) -- 3-D coords in glTF convention (Y-up, Z-back) +# "colors": torch.Tensor (N, 3) -- RGB in [0, 1], or None +# "confidence": torch.Tensor (N,) -- raw confidence per point, or None + + +def _da3_unproject(depth: torch.Tensor, K: torch.Tensor) -> torch.Tensor: + """Pixel-space K⁻¹ unprojection: (H,W) depth → (H,W,3) point map in OpenCV space.""" + H, W = depth.shape + u = torch.arange(W, dtype=torch.float32, device=depth.device) + v = torch.arange(H, dtype=torch.float32, device=depth.device) + u, v = torch.meshgrid(u, v, indexing='xy') # both (H, W) + pix = torch.stack([u, v, torch.ones_like(u)], dim=-1) # (H, W, 3) + rays = torch.einsum('ij,hwj->hwi', torch.linalg.inv(K.to(depth.device)), pix) + return rays * depth.unsqueeze(-1) # (H, W, 3) + + +def _da3_default_K(H: int, W: int) -> torch.Tensor: + """Fallback ~60° FOV pinhole K for mono-mode DA3 (no intrinsics in geometry).""" + fx = fy = float(W) * 0.7 + return torch.tensor([[fx, 0.0, (W - 1) / 2.0], + [0.0, fy, (H - 1) / 2.0], + [0.0, 0.0, 1.0]], dtype=torch.float32) + + +def _da3_get_K(geometry: dict, b: int, H: int, W: int) -> torch.Tensor: + """Return pixel-space K for batch element b, falling back to a default estimate.""" + if "intrinsics" in geometry: + # shape (1, S, 3, 3) - leading scene dimension from the multiview head + return geometry["intrinsics"][0, b].float() + logging.getLogger("comfy").warning( + "DA3_GEOMETRY has no intrinsics (mono-mode model). " + "Using a ~60° FOV estimate; 3-D reconstruction may be inaccurate." + ) + return _da3_default_K(H, W) + + +def _da3_get_extrinsic(geometry: dict, b: int) -> torch.Tensor | None: + """Return the world-to-camera extrinsic for batch element b, or None in mono mode. + + The model outputs (1, S, 3, 4) [R|t] matrices; the fallback identity is (4, 4). + _da3_apply_extrinsic handles both shapes via [:3, :3] / [:3, 3] slicing. + """ + if "extrinsics" not in geometry: + return None + return geometry["extrinsics"][0, b].float() + + +def _da3_apply_extrinsic(points_cam: torch.Tensor, E: torch.Tensor) -> torch.Tensor: + """Transform (H,W,3) OpenCV camera-space points to world space.""" + E = E.to(points_cam.device).float() + if not torch.isfinite(E).all(): + logging.getLogger("comfy").warning( + "DA3 extrinsic matrix contains non-finite values (pose estimation may have failed). " + "Falling back to camera-space coordinates." + ) + return points_cam + H, W, _ = points_cam.shape + R = E[:3, :3] # (3, 3) rotation + t = E[:3, 3] # (3,) translation + R_inv = R.T # rotation inverse = transpose for orthogonal R + t_inv = -(R_inv @ t) # (3,) + pts = points_cam.reshape(-1, 3) # (N, 3) + pts_world = pts @ R_inv.T + t_inv # (N, 3) + return pts_world.reshape(H, W, 3) + + +def _normalize_confidence(conf: torch.Tensor) -> torch.Tensor: + """Map raw confidence to [0, 1] per image.""" + B = conf.shape[0] + out = [] + for i in range(B): + c = conf[i] + c_min, c_max = c.min(), c.max() + out.append((c - c_min) / (c_max - c_min) if c_max > c_min else torch.ones_like(c)) + return torch.stack(out, dim=0) + + +def _da3_build_mask(geometry: dict, b: int, H: int, W: int, confidence_threshold: float, use_sky_mask: bool) -> torch.Tensor: + """Build (H,W) bool keep-mask from sky probability and confidence.""" + mask = torch.ones(H, W, dtype=torch.bool) + if use_sky_mask and "sky" in geometry: + mask = mask & (geometry["sky"][b] < 0.5) + if "confidence" in geometry and confidence_threshold > 0.0: + conf_norm = _normalize_confidence(geometry["confidence"][b:b + 1])[0] + mask = mask & (conf_norm >= confidence_threshold) + return mask + + +class LoadDA3Model(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LoadDA3Model", + display_name="Load Depth Anything 3", + category="model/loaders", + inputs=[ + io.Combo.Input( + "model_name", + options=folder_paths.get_filename_list("geometry_estimation"), + ), + io.Combo.Input( + "weight_dtype", + options=["default", "fp16", "bf16", "fp32"], + default="default", + ), + ], + outputs=[DA3ModelType.Output()], + ) + + @classmethod + def execute(cls, model_name, weight_dtype) -> io.NodeOutput: + model_options = {} + if weight_dtype == "fp16": + model_options["dtype"] = torch.float16 + elif weight_dtype == "bf16": + model_options["dtype"] = torch.bfloat16 + elif weight_dtype == "fp32": + model_options["dtype"] = torch.float32 + + path = folder_paths.get_full_path_or_raise("geometry_estimation", model_name) + model = comfy.sd.load_diffusion_model(path, model_options=model_options) + return io.NodeOutput(model) + + +def _run_da3(model_patcher, image: torch.Tensor, process_res: int, method: str = "upper_bound_resize"): + """Run DA3 on (B,H,W,3), returns depth/conf/sky at original resolution (or None).""" + assert image.ndim == 4 and image.shape[-1] == 3, f"expected (B,H,W,3) IMAGE; got {tuple(image.shape)}" + + B, H, W, _ = image.shape + mm.load_model_gpu(model_patcher) + diffusion = model_patcher.model.diffusion_model + device = mm.get_torch_device() + dtype = diffusion.dtype if diffusion.dtype is not None else torch.float32 + + depths, confs, skies = [], [], [] + for i in range(B): + single = image[i:i + 1].to(device) + x = da3_preprocess.preprocess_image(single, process_res=process_res, method=method) + x = x.to(dtype=dtype) + with torch.no_grad(): + out = diffusion(x) + + depth_lr = out["depth"] + depth_full = torch.nn.functional.interpolate( + depth_lr.unsqueeze(1).float(), size=(H, W), + mode="bilinear", align_corners=False, + ).squeeze(1).cpu() + depths.append(depth_full) + + if "depth_conf" in out: + conf_full = torch.nn.functional.interpolate( + out["depth_conf"].unsqueeze(1).float(), size=(H, W), + mode="bilinear", align_corners=False, + ).squeeze(1).cpu() + confs.append(conf_full) + if "sky" in out: + sky_full = torch.nn.functional.interpolate( + out["sky"].unsqueeze(1).float(), size=(H, W), + mode="bilinear", align_corners=False, + ).squeeze(1).cpu() + skies.append(sky_full) + + depth = torch.cat(depths, dim=0) + confidence = torch.cat(confs, dim=0) if confs else None + sky = torch.cat(skies, dim=0) if skies else None + return depth, confidence, sky + + +class DA3Inference(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="DA3Inference", + search_aliases=["depth", "geometry", "da3", "depth anything", "monocular", "pointmap", "sky", "3d", "metric depth", "disparity"], + display_name="Run Depth Anything 3", + category="image/geometry estimation", + description="Run Depth Anything 3 on an image. In multi-view mode each image is treated as a separate view of the same scene.", + inputs=[ + DA3ModelType.Input("da3_model"), + io.Image.Input("image"), + io.Int.Input("resolution", default=504, min=140, max=2520, step=14, + tooltip="Resolution the model runs at (longest side, multiple of 14).\n" + "Lower = faster / less VRAM.\n" + "Higher = more detail.\n" + "Output is upsampled back to the original size."), + io.Combo.Input("resize_method", options=["upper_bound_resize", "lower_bound_resize"], default="upper_bound_resize", + tooltip="upper_bound_resize: scale so the longest side = resolution (caps memory, default).\n" + "lower_bound_resize: scale so the shortest side = resolution (preserves more detail on tall/wide images, uses more memory)."), + io.DynamicCombo.Input("mode", tooltip="mono: single view image (works with any model variant).\n" + "multiview: all images processed together for geometric consistency + camera pose (for Small/Base models only).", + options=[ + io.DynamicCombo.Option("mono", []), + io.DynamicCombo.Option("multiview", [ + io.Combo.Input("ref_view_strategy", options=["saddle_balanced", "saddle_sim_range", "first", "middle"], default="saddle_balanced", + tooltip="Which view acts as the geometric anchor.\n" + "- saddle_balanced: the view most 'average' across all others (best general choice).\n" + "- saddle_sim_range: the view most visually distinct from the others.\n" + "- first / middle: fixed positional picks."), + io.Combo.Input("pose_method", options=["cam_dec", "ray_pose"], default="cam_dec", + tooltip="How the camera field-of-view is estimated (for Small/Base models only).\n" + "- cam_dec: learned from image features.\n" + "- ray_pose: derived geometrically from the model's 3D ray output.\n" + "Affects perspective correctness of the 3D output. Try both if results look distorted."), + ]), + ]), + ], + outputs=[ + DA3Geometry.Output("da3_geometry", tooltip="Dictionary of non-normalized tensors.\n" + "Always has the keys: depth, image, mode.\n" + "Optional keys: sky (for Mono/Metric), confidence (for Small/Base), extrinsics + intrinsics (for multi-view)."), + ], + ) + + @classmethod + def execute(cls, da3_model, image, resolution, resize_method, mode) -> io.NodeOutput: + mode_val = mode["mode"] # "mono" or "multiview" + + if mode_val == "mono": + return cls._execute_mono(da3_model, image, resolution, resize_method) + + # Capability checks for multi-view mode. + diffusion = da3_model.model.diffusion_model + pose_method = mode["pose_method"] + ref_view_strategy = mode["ref_view_strategy"] + + has_cam_dec = diffusion.cam_dec is not None + has_dualdpt = diffusion.head_type == "dualdpt" + + if not has_cam_dec and not has_dualdpt: + raise ValueError( + "multi-view mode requires Small or Base model. The loaded model " + f"(head_type='{diffusion.head_type}') does not support cross-view " + "attention or camera pose estimation. Switch mode to 'mono', or " + "load Small or Base model for mult-view." + ) + + if pose_method == "cam_dec" and not has_cam_dec: + raise ValueError( + "pose_method='cam_dec' requires a camera decoder, but the loaded " + f"model (head_type='{diffusion.head_type}') does not have one. " + "Use pose_method='ray_pose' instead." + ) + if pose_method == "ray_pose" and not has_dualdpt: + raise ValueError( + "pose_method='ray_pose' requires a DualDPT head, but the loaded " + f"model has a '{diffusion.head_type}' head. " + "Use pose_method='cam_dec' instead." + ) + + return cls._execute_multiview( + da3_model, image, resolution, resize_method, + ref_view_strategy, pose_method, + ) + + @classmethod + def _execute_mono(cls, model, image, resolution, resize_method) -> io.NodeOutput: + depth, confidence, sky = _run_da3(model, image, resolution, method=resize_method) + + geometry: dict = { + "depth": depth.contiguous(), + "image": image[..., :3].cpu(), + "mode": "mono", + } + if sky is not None: + geometry["sky"] = sky.contiguous() + if confidence is not None: + geometry["confidence"] = confidence.contiguous() + return io.NodeOutput(geometry) + + @classmethod + def _execute_multiview(cls, model, image, resolution, resize_method, ref_view_strategy, pose_method) -> io.NodeOutput: + assert image.ndim == 4 and image.shape[-1] == 3, \ + f"expected (B,H,W,3) IMAGE; got {tuple(image.shape)}" + S, H, W, _ = image.shape + + mm.load_model_gpu(model) + diffusion = model.model.diffusion_model + device = mm.get_torch_device() + dtype = diffusion.dtype if diffusion.dtype is not None else torch.float32 + + # All views in a single forward pass: (1, S, 3, H', W'). + x = image.to(device) + x = da3_preprocess.preprocess_image(x, process_res=resolution, method=resize_method) + x = x.to(dtype=dtype).unsqueeze(0) + + use_ray_pose = (pose_method == "ray_pose") + with torch.no_grad(): + out = diffusion(x, use_ray_pose=use_ray_pose, ref_view_strategy=ref_view_strategy) + + depth = torch.nn.functional.interpolate( + out["depth"].float().unsqueeze(1), size=(H, W), + mode="bilinear", align_corners=False, + ).squeeze(1).cpu() + + sky = None + if "sky" in out: + sky = torch.nn.functional.interpolate( + out["sky"].unsqueeze(1).float(), size=(H, W), + mode="bilinear", align_corners=False, + ).squeeze(1).cpu() + + if "extrinsics" in out and "intrinsics" in out: + extrinsics = out["extrinsics"].float().cpu() + intrinsics = out["intrinsics"].float().cpu() + else: + extrinsics = torch.eye(4)[None, None].expand(1, S, 4, 4).clone() + intrinsics = torch.eye(3)[None, None].expand(1, S, 3, 3).clone() + + geometry: dict = { + "depth": depth.contiguous(), + "image": image[..., :3].cpu(), + "mode": "multiview", + "extrinsics": extrinsics.contiguous(), + "intrinsics": intrinsics.contiguous(), + } + if sky is not None: + geometry["sky"] = sky.contiguous() + if "depth_conf" in out: + conf = torch.nn.functional.interpolate( + out["depth_conf"].unsqueeze(1).float(), size=(H, W), + mode="bilinear", align_corners=False, + ).squeeze(1).cpu() + geometry["confidence"] = conf.contiguous() + return io.NodeOutput(geometry) + + +class DA3Render(io.ComfyNode): + """Render a visualization from a DA3_GEOMETRY packet.""" + + _DEPTH_RENDER_INPUTS = [ + io.Combo.Input("normalization", + options=["v2_style", "min_max", "raw"], + default="v2_style", + tooltip="- v2_style: mean/std normalisation for perceptually balanced results (default).\n" + "- min_max: stretches the full depth range to [0, 1] for maximum contrast.\n" + "- raw: no scaling,preserves metric units for Metric model."), + io.Boolean.Input("apply_sky_clip", default=False, + tooltip="Clip sky-region depth to the 99th percentile of foreground depth before normalisation. " + "Requires a sky key in the da3_geometry input (for Mono/Metric models only)."), + ] + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="DA3Render", + display_name="Render Depth Anything 3", + category="image/geometry estimation", + description="Render a depth map, confidence map, or sky mask from Depth Anything 3 geometry data.", + inputs=[ + DA3Geometry.Input("da3_geometry"), + io.DynamicCombo.Input("output", + tooltip="- depth: normalised greyscale depth image.\n" + "- depth_colored: depth mapped through the Turbo colormap.\n" + "- sky_mask: sky probability in [0, 1] (for Mono/Metric models only).\n" + "- confidence: normalised depth confidence (for Small/Base models only).", + options=[ + io.DynamicCombo.Option("depth", cls._DEPTH_RENDER_INPUTS), + io.DynamicCombo.Option("depth_colored", cls._DEPTH_RENDER_INPUTS), + io.DynamicCombo.Option("sky_mask", [ + io.Boolean.Input("colored", default=False, tooltip="Apply the Turbo colormap to the sky mask."), + ]), + io.DynamicCombo.Option("confidence", [ + io.Boolean.Input("colored", default=False, tooltip="Apply the Turbo colormap to the confidence map."), + ]), + ]), + ], + outputs=[io.Image.Output()], + ) + + @classmethod + def execute(cls, da3_geometry, output) -> io.NodeOutput: + output_val = output["output"] + + if output_val in ("depth", "depth_colored"): + normalization = output["normalization"] + apply_sky_clip = output["apply_sky_clip"] + if apply_sky_clip and "sky" not in da3_geometry: + raise ValueError( + "apply_sky_clip=True requires a sky tensor in the da3_geometry input, but none is present. " + "Run with Mono/Metric models or set apply_sky_clip=False." + ) + depth = da3_geometry["depth"] + sky = da3_geometry.get("sky") + if apply_sky_clip and sky is not None: + depth = torch.stack([ + da3_preprocess.apply_sky_aware_clip(depth[i], sky[i]) + for i in range(depth.shape[0]) + ], dim=0) + grey = cls._depth_to_image(depth, sky, normalization) # (B,H,W,3) greyscale + result = _turbo(grey[..., 0]) if output_val == "depth_colored" else grey + + elif output_val == "sky_mask": + if "sky" not in da3_geometry: + raise ValueError("geometry has no sky output; run with Mono/Metric models.") + sky = da3_geometry["sky"] + if output["colored"]: + result = _turbo(sky) + else: + result = sky.unsqueeze(-1).expand(*sky.shape, 3).contiguous() + + elif output_val == "confidence": + if "confidence" not in da3_geometry: + raise ValueError("da3_geometry has no confidence output; run with Small/Base models.") + conf = _normalize_confidence(da3_geometry["confidence"]) + if output["colored"]: + result = _turbo(conf) + else: + result = conf.unsqueeze(-1).expand(*conf.shape, 3).contiguous() + + else: + raise ValueError(f"Unknown output mode: {output_val}") + + return io.NodeOutput(result.float()) + + @staticmethod + def _depth_to_image(depth: torch.Tensor, sky_for_norm: torch.Tensor | None, normalization: str) -> torch.Tensor: + """Normalise depth and pack as an (B,H,W,3) image tensor.""" + + N = depth.shape[0] + if normalization == "v2_style": + norm = torch.stack([ + da3_preprocess.normalize_depth_v2_style( + depth[i], sky_for_norm[i] if sky_for_norm is not None else None) + for i in range(N) + ], dim=0) + elif normalization == "min_max": + norm = da3_preprocess.normalize_depth_min_max(depth) + else: + norm = depth + + out = norm.unsqueeze(-1).repeat(1, 1, 1, 3) + if normalization != "raw": + out = out.clamp(0.0, 1.0) + return out.contiguous() + + +class DA3GeometryToMesh(io.ComfyNode): + """Convert a DA3_GEOMETRY packet into a Types.MESH by unprojecting depth and triangulating.""" + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="DA3GeometryToMesh", + search_aliases=["da3", "depth anything", "mesh", "geometry", "3d", "triangulate"], + display_name="Convert DA3 Geometry to Mesh", + category="image/geometry estimation", + description="Convert a depth map into a triangulated 3D mesh.", + inputs=[ + DA3Geometry.Input("da3_geometry"), + io.Int.Input("batch_index", default=0, min=0, max=4096, tooltip="Which image of a batch to convert. Per-image vertex counts differ so batches cannot be stacked."), + io.Int.Input("decimation", default=1, min=1, max=8, tooltip="Vertex stride. 1 = full resolution, 2 = half, etc."), + io.Float.Input("discontinuity_threshold", default=0.04, min=0.0, max=1.0, step=0.01, tooltip="Drop triangles whose 3x3 depth span exceeds this fraction. 0 = off."), + io.Float.Input("confidence_threshold", default=0.1, min=0.0, max=1.0, step=0.01, + tooltip="Exclude pixels whose per-image normalised confidence is below this value (0 = keep all, 1 = keep only the single most confident pixel). " + "Used when the geometry has a confidence map (Small/Base models)."), + io.Boolean.Input("use_sky_mask", default=True, tooltip="Exclude sky-probability pixels (sky >= 0.5) from the mesh. Used when the geometry has a sky map (Mono/Metric models)."), + io.Boolean.Input("texture", default=True, tooltip="Use the source image as a base color texture."), + ], + outputs=[io.Mesh.Output()], + ) + + @classmethod + def execute(cls, da3_geometry, batch_index, decimation, discontinuity_threshold, confidence_threshold, use_sky_mask, texture) -> io.NodeOutput: + depth_all = da3_geometry["depth"] # (B, H, W) + B = depth_all.shape[0] + if batch_index >= B: + raise ValueError(f"batch_index {batch_index} is out of range; DA3_GEOMETRY has batch size {B}.") + + depth = depth_all[batch_index] # (H, W) + H, W = depth.shape + + # NaN/inf depth would propagate silently through unproject and produce an + # empty mesh; replace them with 0 here so those pixels are later excluded + # by the isfinite check inside triangulate_grid_mesh. + depth = depth.clone() + n_bad = (~torch.isfinite(depth)).sum().item() + if n_bad: + logging.getLogger("comfy").warning( + f"DA3GeometryToMesh: depth[{batch_index}] has {n_bad} non-finite pixels " + f"({100*n_bad/(H*W):.1f}%) - zeroed before unproject." + ) + depth[~torch.isfinite(depth)] = 0.0 + logging.getLogger("comfy").debug( + f"DA3GeometryToMesh: depth[{batch_index}] range " + f"[{depth.min():.4g}, {depth.max():.4g}], mean={depth.mean():.4g}" + ) + + K = _da3_get_K(da3_geometry, batch_index, H, W) + points = _da3_unproject(depth, K) # (H, W, 3) in OpenCV camera space + + # Apply world-to-camera inverse so multi-view frames share a common world frame. + E = _da3_get_extrinsic(da3_geometry, batch_index) + if E is not None: + points = _da3_apply_extrinsic(points, E) + + # Mask invalid pixels by setting them to inf so triangulate_grid_mesh skips them. + mask = _da3_build_mask(da3_geometry, batch_index, H, W, confidence_threshold, use_sky_mask) + # Also exclude pixels where depth was invalid. + mask = mask & (depth_all[batch_index] > 0) & torch.isfinite(depth_all[batch_index]) + points = points.clone() + points[~mask] = float('inf') + + verts, faces, uvs = triangulate_grid_mesh( + points, + decimation=decimation, + discontinuity_threshold=discontinuity_threshold, + depth=depth, + ) + if verts.shape[0] == 0 or faces.shape[0] == 0: + raise ValueError( + "DA3GeometryToMesh produced an empty mesh. " + "Try raising discontinuity_threshold, lowering confidence_threshold, " + "or disabling use_sky_mask." + ) + + # OpenCV (X right, Y down, Z forward) → glTF (X right, Y up, Z back). + # Same transform as MoGePointMapToMesh perspective branch. + verts = verts * torch.tensor([1.0, -1.0, -1.0], dtype=verts.dtype) + faces = faces[:, [0, 2, 1]].contiguous() + + tex = da3_geometry["image"][batch_index:batch_index + 1] if texture else None + mesh = Types.MESH( + vertices=verts.unsqueeze(0), + faces=faces.unsqueeze(0), + uvs=uvs.unsqueeze(0), + texture=tex, + ) + return io.NodeOutput(mesh) + + +class DA3GeometryToPointCloud(io.ComfyNode): + """Unproject a DA3_GEOMETRY depth map into a filtered DA3_POINT_CLOUD.""" + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="DA3GeometryToPointCloud", + search_aliases=["da3", "depth anything", "point cloud", "pointcloud", "3d", "geometry"], + display_name="Convert DA3 Geometry to Point Cloud", + category="image/geometry estimation", + description="Convert a depth map into a 3D point cloud.", + inputs=[ + DA3Geometry.Input("da3_geometry"), + io.Int.Input("batch_index", default=0, min=0, max=4096, tooltip="Which image of a batch to convert."), + io.Float.Input("confidence_threshold", default=0.1, min=0.0, max=1.0, step=0.01, + tooltip="Exclude pixels whose per-image normalised confidence is below this value (0 = keep all). Used when the geometry has a confidence map (Small/Base models)."), + io.Boolean.Input("use_sky_mask", default=True, + tooltip="Exclude sky-probability pixels (sky >= 0.5). Used when the geometry has a sky map (Mono/Metric models)."), + io.Int.Input("downsample", default=1, min=1, max=16, + tooltip="Take every Nth pixel (1 = full resolution). Higher values give fewer points and faster processing."), + ], + # TODO: add a proper PointCloud output type + outputs=[DA3PointCloud.Output(display_name="point_cloud")], + ) + + @classmethod + def execute(cls, da3_geometry, batch_index, confidence_threshold, use_sky_mask, downsample) -> io.NodeOutput: + depth_all = da3_geometry["depth"] # (B, H, W) + B = depth_all.shape[0] + if batch_index >= B: + raise ValueError(f"batch_index {batch_index} is out of range; DA3_GEOMETRY has batch size {B}.") + + depth = depth_all[batch_index].clone() # (H, W) + depth[~torch.isfinite(depth)] = 0.0 + H, W = depth.shape + + K = _da3_get_K(da3_geometry, batch_index, H, W) + + if downsample > 1: + depth = depth[::downsample, ::downsample].contiguous() + # Scale intrinsics to the downsampled grid. + K = K.clone() + K[0, :] /= downsample + K[1, :] /= downsample + + H_ds, W_ds = depth.shape + points = _da3_unproject(depth, K) # (H_ds, W_ds, 3) in OpenCV camera space + + # Apply world-to-camera inverse so multi-view frames share a common world frame. + E = _da3_get_extrinsic(da3_geometry, batch_index) + if E is not None: + points = _da3_apply_extrinsic(points, E) + + # Rebuild mask at downsampled resolution. + mask = _da3_build_mask(da3_geometry, batch_index, H, W, confidence_threshold, use_sky_mask) + if downsample > 1: + mask = mask[::downsample, ::downsample] + + mask = mask & torch.isfinite(depth) + + # OpenCV → glTF: flip Y and Z. + points_gltf = points.clone() + points_gltf[..., 1] *= -1.0 + points_gltf[..., 2] *= -1.0 + + pts_flat = points_gltf.reshape(-1, 3)[mask.reshape(-1)] + + colors_flat = None + if "image" in da3_geometry: + img = da3_geometry["image"][batch_index] # (H, W, 3) + if downsample > 1: + img = img[::downsample, ::downsample] + colors_flat = img.reshape(-1, 3)[mask.reshape(-1)] + + conf_flat = None + if "confidence" in da3_geometry: + conf = da3_geometry["confidence"][batch_index] # (H, W) + if downsample > 1: + conf = conf[::downsample, ::downsample] + conf_flat = conf.reshape(-1)[mask.reshape(-1)] + + if pts_flat.shape[0] == 0: + raise ValueError( + "DA3GeometryToPointCloud produced zero points after filtering. " + "Try lowering confidence_threshold or disabling use_sky_mask." + ) + + return io.NodeOutput({ + "points": pts_flat, + "colors": colors_flat, + "confidence": conf_flat, + }) + + +class DA3Extension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + LoadDA3Model, + DA3Inference, + DA3Render, + DA3GeometryToMesh, + # DA3GeometryToPointCloud, # Keep this commented out for now until we have a proper PointCloud output type + ] + + +async def comfy_entrypoint() -> DA3Extension: + return DA3Extension() diff --git a/comfy_extras/nodes_eps.py b/comfy_extras/nodes_eps.py index 0fb3871c8..8c397f132 100644 --- a/comfy_extras/nodes_eps.py +++ b/comfy_extras/nodes_eps.py @@ -18,7 +18,7 @@ class EpsilonScaling(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="Epsilon Scaling", - category="model_patches/unet", + category="model/patch/unet", inputs=[ io.Model.Input("model"), io.Float.Input( @@ -84,7 +84,7 @@ class TemporalScoreRescaling(io.ComfyNode): return io.Schema( node_id="TemporalScoreRescaling", display_name="TSR - Temporal Score Rescaling", - category="model_patches/unet", + category="model/patch/unet", inputs=[ io.Model.Input("model"), io.Float.Input( diff --git a/comfy_extras/nodes_flux.py b/comfy_extras/nodes_flux.py index 997f21c09..afc663b22 100644 --- a/comfy_extras/nodes_flux.py +++ b/comfy_extras/nodes_flux.py @@ -40,7 +40,7 @@ class EmptyFlux2LatentImage(io.ComfyNode): return io.Schema( node_id="EmptyFlux2LatentImage", display_name="Empty Flux 2 Latent", - category="latent", + category="model/latent", inputs=[ io.Int.Input("width", default=1024, min=16, max=nodes.MAX_RESOLUTION, step=16), io.Int.Input("height", default=1024, min=16, max=nodes.MAX_RESOLUTION, step=16), @@ -215,7 +215,7 @@ class Flux2Scheduler(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="Flux2Scheduler", - category="sampling/schedulers", + category="model/sampling/schedulers", inputs=[ io.Int.Input("steps", default=20, min=1, max=4096), io.Int.Input("width", default=1024, min=16, max=nodes.MAX_RESOLUTION, step=1), diff --git a/comfy_extras/nodes_frame_interpolation.py b/comfy_extras/nodes_frame_interpolation.py index 9dd34cfb8..4d5bca17e 100644 --- a/comfy_extras/nodes_frame_interpolation.py +++ b/comfy_extras/nodes_frame_interpolation.py @@ -19,7 +19,7 @@ class FrameInterpolationModelLoader(io.ComfyNode): return io.Schema( node_id="FrameInterpolationModelLoader", display_name="Load Frame Interpolation Model", - category="loaders", + category="model/loaders", inputs=[ io.Combo.Input("model_name", options=folder_paths.get_filename_list("frame_interpolation"), tooltip="Select a frame interpolation model to load. Models must be placed in the 'frame_interpolation' folder."), diff --git a/comfy_extras/nodes_freelunch.py b/comfy_extras/nodes_freelunch.py index 248efdef3..ccbd1fd90 100644 --- a/comfy_extras/nodes_freelunch.py +++ b/comfy_extras/nodes_freelunch.py @@ -29,7 +29,7 @@ class FreeU(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="FreeU", - category="model_patches/unet", + category="model/patch/unet", inputs=[ IO.Model.Input("model"), IO.Float.Input("b1", default=1.1, min=0.0, max=10.0, step=0.01, advanced=True), @@ -76,7 +76,7 @@ class FreeU_V2(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="FreeU_V2", - category="model_patches/unet", + category="model/patch/unet", inputs=[ IO.Model.Input("model"), IO.Float.Input("b1", default=1.3, min=0.0, max=10.0, step=0.01, advanced=True), diff --git a/comfy_extras/nodes_gaussian_splat.py b/comfy_extras/nodes_gaussian_splat.py new file mode 100644 index 000000000..116c14fde --- /dev/null +++ b/comfy_extras/nodes_gaussian_splat.py @@ -0,0 +1,1664 @@ +# Generic utility nodes for the SPLAT type (3D gaussian splats) + +import gzip +import logging +import math +import struct +from io import BytesIO + +import numpy as np +import torch +from typing_extensions import override +from scipy.ndimage import map_coordinates, minimum as _ndi_minimum, maximum as _ndi_maximum +from scipy.sparse import coo_matrix +from scipy.sparse.csgraph import connected_components + +import comfy.model_management +import comfy.utils +from comfy_api.latest import ComfyExtension, IO, Types +from comfy_extras.nodes_save_3d import pack_variable_mesh_batch +from server import PromptServer + +_C0 = 0.28209479177387814 # SH band-0 constant: DC coefficient -> base RGB + + +def _srgb_to_linear(c): + return torch.where(c <= 0.04045, c / 12.92, ((c.clamp_min(0) + 0.055) / 1.055) ** 2.4) + + +def _linear_to_srgb(c): + return torch.where(c <= 0.0031308, c * 12.92, 1.055 * c.clamp_min(0) ** (1 / 2.4) - 0.055) + + +def _real_len(g: Types.SPLAT, i: int) -> int: + # Real splat count of batch item i (honors variable-length `counts`). + return int(g.counts[i].item()) if g.counts is not None else g.positions.shape[1] + + +def _hex_to_rgb(h: str) -> tuple[float, float, float]: + # "#RRGGBB" -> (r,g,b) in [0,1]; falls back to black. + h = h.lstrip("#") + if len(h) != 6: + return (0.0, 0.0, 0.0) + return tuple(int(h[i:i + 2], 16) / 255.0 for i in (0, 2, 4)) + + +def _quantile(x, q): + # torch.quantile errors above 2**24 elements; stride-subsample large inputs for the estimate. + lim = 1 << 24 + if x.numel() > lim: + x = x[:: x.numel() // lim + 1] + return torch.quantile(x, q) + + +def _gaussian_ply_bytes(positions, scales, rotations, opacities, sh) -> bytes: + """Serialize render-ready gaussian tensors as a binary 3DGS .ply. + + positions (N,3) world; scales (N,3) linear; rotations (N,4) quat wxyz; opacities (N,1) in [0,1]; + sh (N,K,3) SH coefficients. Activated values are inverted to the standard 3D gaussian splat storage convention + (log scale, logit opacity). + """ + xyz = positions.cpu().numpy().astype(np.float32) + n = xyz.shape[0] + if n == 0: + raise ValueError("SplatToFile3D: gaussian is empty") + normals = np.zeros_like(xyz) + f = sh.cpu().numpy().astype(np.float32) # (N, K, 3) + f_dc = f[:, 0, :] # (N, 3) + f_rest = f[:, 1:, :].transpose(0, 2, 1).reshape(n, -1) # (N, 3*(K-1)) channel-major + op = opacities.cpu().numpy().astype(np.float32).reshape(n, 1).clip(1e-6, 1 - 1e-6) + op = np.log(op / (1.0 - op)) # inverse sigmoid (logit) + scale = np.log(scales.cpu().numpy().astype(np.float32).clip(min=1e-8)) + rot = rotations.cpu().numpy().astype(np.float32) # (N, 4) + + attrs = (['x', 'y', 'z', 'nx', 'ny', 'nz'] + + [f'f_dc_{i}' for i in range(3)] + + [f'f_rest_{i}' for i in range(f_rest.shape[1])] + + ['opacity'] + [f'scale_{i}' for i in range(3)] + [f'rot_{i}' for i in range(4)]) + elements = np.empty(n, dtype=[(a, 'f4') for a in attrs]) + elements[:] = list(map(tuple, np.concatenate([xyz, normals, f_dc, f_rest, op, scale, rot], axis=1))) + + header = "ply\nformat binary_little_endian 1.0\n" + f"element vertex {n}\n" + header += "".join(f"property float {a}\n" for a in attrs) + "end_header\n" + return header.encode('ascii') + elements.tobytes() + + +# .ksplat (mkkellogg SplatBuffer) level 0, SH degree 0: 4096-byte header, one 1024-byte section header, +# then N 44-byte records. Bucketing/quantization only exist at levels >= 1. See SplatBuffer.js. +_KSPLAT_HEADER_BYTES = 4096 +_KSPLAT_SECTION_HEADER_BYTES = 1024 +_KSPLAT_BYTES_PER_SPLAT = 44 # center 12 + scale 12 + rotation 16 + color(RGBA u8) 4 +_KSPLAT_VERSION = (0, 1) # SplatBuffer CurrentMajor/MinorVersion + + +def _gaussian_ksplat_bytes(positions, scales, rotations, opacities, sh) -> bytes: + """Serialize gaussian tensors as a level-0, SH degree-0 .ksplat (linear scale, opacity in color alpha). + + positions (N,3) world; scales (N,3) linear; rotations (N,4) wxyz; opacities (N,1) in [0,1]; sh (N,K,3). + """ + xyz = positions.cpu().numpy().astype(np.float32) + n = xyz.shape[0] + if n == 0: + raise ValueError("SplatToFile3D: gaussian is empty") + scale = scales.cpu().numpy().astype(np.float32) + rot = rotations.cpu().numpy().astype(np.float32) # wxyz, mirrors the .ply rot order + rot = rot / np.linalg.norm(rot, axis=1, keepdims=True).clip(1e-12) + rgb = np.clip(sh[:, 0, :].cpu().numpy().astype(np.float32) * _C0 + 0.5, 0, 1) + op = opacities.cpu().numpy().astype(np.float32).reshape(n, 1).clip(0, 1) + rgba = np.round(np.concatenate([rgb, op], axis=1) * 255.0).astype(np.uint8) # (N, 4) RGBA + + # 44-byte record: float center(3) + scale(3) + rot(4), then uint8 rgba(4). + floats = np.concatenate([xyz, scale, rot], axis=1).astype(' bytes: + """Serialize gaussian tensors as a gzip-compressed .spz (Niantic v2, SH degree 0, base color only). + + positions (N,3) world; scales (N,3) linear; rotations (N,4) wxyz; opacities (N,1) in [0,1]; sh (N,K,3). + """ + xyz = positions.cpu().numpy().astype(np.float32) + n = xyz.shape[0] + if n == 0: + raise ValueError("SplatToFile3D: gaussian is empty") + + # Positions: fixed point, masked to 24 bits, little-endian 3-byte words. + fixed = 1 << _SPZ_FRACTIONAL_BITS + qi = np.clip(np.round(xyz * fixed), -(1 << 23), (1 << 23) - 1).astype(np.int32) + qu = (qi & 0xFFFFFF).astype(np.uint32) + pos = np.stack([qu & 0xFF, (qu >> 8) & 0xFF, (qu >> 16) & 0xFF], axis=-1).reshape(n, 9).astype(np.uint8) + + alpha = np.round(opacities.cpu().numpy().astype(np.float32).reshape(n) * 255.0).clip(0, 255).astype(np.uint8) + + rgb = sh[:, 0, :].cpu().numpy().astype(np.float32) * _C0 + 0.5 + col = np.round(((rgb - 0.5) / _SPZ_COLOR_SCALE + 0.5) * 255.0).clip(0, 255).astype(np.uint8) # (N,3) + + sln = np.log(scales.cpu().numpy().astype(np.float32).clip(min=1e-9)) + scb = np.round((sln + 10.0) * 16.0).clip(0, 255).astype(np.uint8) # (N,3) inverts exp(b/16-10) + + rot = rotations.cpu().numpy().astype(np.float32) # wxyz + rot = rot / np.linalg.norm(rot, axis=1, keepdims=True).clip(1e-12) + rot[rot[:, 0] < 0] *= -1.0 # canonical w >= 0 (w dropped on decode) + rotb = np.round((rot[:, 1:4] + 1.0) * 127.5).clip(0, 255).astype(np.uint8) # (N,3) x,y,z + + header = bytearray(16) + struct.pack_into(' (positions, scales linear, rotations wxyz, opacities [0,1], sh (N,K,3)) ---- +# Inverse of the writers above and of spark's loaders. ksplat/splat/spz carry base color only (SH degree 0 +# -> K=1); .ply round-trips full SH. None of the formats flip axes, so import is the identity of export. +_PLY_DTYPES = {'char': 'i1', 'uchar': 'u1', 'short': 'i2', 'ushort': 'u2', 'int': 'i4', 'uint': 'u4', + 'float': 'f4', 'double': 'f8', 'int8': 'i1', 'uint8': 'u1', 'int16': 'i2', 'uint16': 'u2', + 'int32': 'i4', 'uint32': 'u4', 'float32': 'f4', 'float64': 'f8'} +_KSPLAT_COMPRESSION = { # level -> (bytesPerCenter, scale, rotation, color, shComponent, defaultScaleRange) + 0: (12, 12, 16, 4, 4, 1), 1: (6, 6, 8, 4, 2, 32767), 2: (6, 6, 8, 4, 1, 32767)} +_KSPLAT_SH_COMPONENTS = {0: 0, 1: 9, 2: 24, 3: 45} + + +def _rgb_to_sh_dc(rgb): + return ((np.asarray(rgb, np.float32) - 0.5) / _C0)[:, None, :] # (N,3) base color -> (N,1,3) SH DC + + +def _norm_quat(q): + return q / np.linalg.norm(q, axis=1, keepdims=True).clip(1e-12) + + +def _parse_ply_gaussian(data: bytes): + end = data.find(b'end_header') + if end < 0: + raise ValueError("File3DToSplat: not a PLY (missing end_header)") + header = data[:end].decode('ascii', 'replace') + body = end + len(b'end_header') + body += 2 if data[body:body + 2] == b'\r\n' else 1 + count, props, in_vertex = 0, [], False + for line in header.splitlines(): + p = line.split() + if not p: + continue + if p[0] == 'format' and p[1] != 'binary_little_endian': + raise ValueError(f"File3DToSplat: unsupported PLY format '{p[1]}' (need binary_little_endian)") + if p[0] == 'element': + in_vertex = p[1] == 'vertex' + if in_vertex: + count = int(p[2]) + elif p[0] == 'property' and in_vertex: + if p[1] == 'list': + raise ValueError("File3DToSplat: PLY vertex has list properties (unsupported)") + props.append((p[2], '<' + _PLY_DTYPES[p[1]])) + arr = np.frombuffer(data, np.dtype(props), count=count, offset=body) + names = arr.dtype.names + c = lambda k: arr[k].astype(np.float32) + n = count + + xyz = np.stack([c('x'), c('y'), c('z')], 1) + if 'scale_0' in names: + scale = np.exp(np.stack([c('scale_0'), c('scale_1'), c('scale_2')], 1)) # 3DGS stores log scale + else: + scale = np.full((n, 3), 0.01, np.float32) + if 'rot_0' in names: + rot = _norm_quat(np.stack([c('rot_0'), c('rot_1'), c('rot_2'), c('rot_3')], 1)) # wxyz + else: + rot = np.tile(np.array([1, 0, 0, 0], np.float32), (n, 1)) + opacity = 1.0 / (1.0 + np.exp(-c('opacity'))) if 'opacity' in names else np.ones(n, np.float32) + + if 'f_dc_0' in names: + dc = np.stack([c('f_dc_0'), c('f_dc_1'), c('f_dc_2')], 1) # (N,3) + rest = sorted((k for k in names if k.startswith('f_rest_')), key=lambda s: int(s.split('_')[-1])) + if rest: + r = np.stack([c(k) for k in rest], 1) # (N, 3*(K-1)) channel-major + kk = r.shape[1] // 3 + 1 + r = r.reshape(n, 3, kk - 1).transpose(0, 2, 1) # -> (N, K-1, 3) + sh = np.concatenate([dc[:, None, :], r], 1) + else: + sh = dc[:, None, :] + elif 'red' in names: + sh = _rgb_to_sh_dc(np.stack([c('red'), c('green'), c('blue')], 1) / 255.0) + else: + sh = np.zeros((n, 1, 3), np.float32) + return xyz, scale, rot, opacity, sh + + +def _parse_splat_gaussian(data: bytes): + # antimatter15 .splat: 32-byte records (f32 xyz, f32 scale, u8 rgba, u8 quat as (b-128)/128 wxyz). + if len(data) % 32 != 0: + raise ValueError("File3DToSplat: .splat size is not a multiple of 32 bytes") + rec = np.frombuffer(data, np.dtype([('xyz', ' 0: + ct, ft = (' full_splats: + lengths = np.frombuffer(data, '> 30) & 3 + q = np.zeros((n, 4), np.float32) # x,y,z,w + remaining, sumsq = combined.copy(), np.zeros(n, np.float64) + for comp in (3, 2, 1, 0): + active = comp != largest + value = (remaining & 0x1FF).astype(np.float64) + sign = (remaining >> 9) & 1 + remaining = np.where(active, remaining >> 10, remaining) + val = (1.0 / math.sqrt(2)) * (value / 0x1FF) + val = np.where(sign == 1, -val, val) + q[active, comp] = val[active] + sumsq += np.where(active, val * val, 0.0) + q[np.arange(n), largest] = np.sqrt(np.clip(1.0 - sumsq, 0, None)) + rot = _norm_quat(np.stack([q[:, 3], q[:, 0], q[:, 1], q[:, 2]], 1)) # xyzw -> wxyz + else: + qb = np.frombuffer(raw, np.uint8, count=n * 3, offset=off).reshape(n, 3).astype(np.float32) + xq = qb / 127.5 - 1.0 + w = np.sqrt(np.clip(1.0 - (xq ** 2).sum(1), 0, None)) + rot = _norm_quat(np.concatenate([w[:, None], xq], 1)) # wxyz + return xyz, scale, rot, alpha, _rgb_to_sh_dc(rgb) + + +_GAUSSIAN_PARSERS = {"ply": _parse_ply_gaussian, "splat": _parse_splat_gaussian, + "ksplat": _parse_ksplat_gaussian, "spz": _parse_spz_gaussian} + + +def _detect_splat_format(data: bytes) -> str: + if data[:3] == b'ply': + return "ply" + if data[:2] == b'\x1f\x8b': # gzip -> spz + return "spz" + if len(data) >= 2 and data[0] == 0 and data[1] >= 1: # ksplat version 0.x header + return "ksplat" + if len(data) % 32 == 0: + return "splat" + raise ValueError("File3DToSplat: could not determine splat format from contents") + + +def _gaussian_item(g: Types.SPLAT, i: int, device): + # Slice batch item i to its real length, as float32 torch tensors on `device` (SH DC -> base RGB). + end = _real_len(g, i) + to = lambda a: a.to(device=device, dtype=torch.float32) + xyz = to(g.positions[i, :end]) + rgb = (to(g.sh[i, :end, 0, :]) * _C0 + 0.5).clamp(0, 1) + opacity = to(g.opacities[i, :end]).reshape(-1) + scale = to(g.scales[i, :end]) + rot = to(g.rotations[i, :end]) + return xyz, rgb, opacity, scale, rot + + +def _quat_to_mat(q): + # q: (N, 4) wxyz, normalized -> (N, 3, 3) + q = q / q.norm(dim=-1, keepdim=True).clamp_min(1e-12) + w, x, y, z = q.unbind(-1) + return torch.stack([ + 1 - 2 * (y * y + z * z), 2 * (x * y - w * z), 2 * (x * z + w * y), + 2 * (x * y + w * z), 1 - 2 * (x * x + z * z), 2 * (y * z - w * x), + 2 * (x * z - w * y), 2 * (y * z + w * x), 1 - 2 * (x * x + y * y), + ], dim=-1).reshape(-1, 3, 3) + + +def _quat_mul(a, b): + # Hamilton product a (x) b, wxyz. + aw, ax, ay, az = a.unbind(-1) + bw, bx, by, bz = b.unbind(-1) + return torch.stack([ + aw * bw - ax * bx - ay * by - az * bz, + aw * bx + ax * bw + ay * bz - az * by, + aw * by - ax * bz + ay * bw + az * bx, + aw * bz + ax * by - ay * bx + az * bw, + ], dim=-1) + + +def _euler_to_quat(rx, ry, rz): + # Degrees, applied as Rz @ Ry @ Rx (rotate about X, then Y, then Z in world). Returns wxyz. + c, s = np.cos(np.radians([rx, ry, rz]) / 2.0), np.sin(np.radians([rx, ry, rz]) / 2.0) + qx = torch.tensor([c[0], s[0], 0.0, 0.0], dtype=torch.float32) + qy = torch.tensor([c[1], 0.0, s[1], 0.0], dtype=torch.float32) + qz = torch.tensor([c[2], 0.0, 0.0, s[2]], dtype=torch.float32) + return _quat_mul(_quat_mul(qz, qy), qx) + + +def _mat_to_quat(m): + # Rotation matrix (..., 3, 3) -> quaternion (..., 4) wxyz. Batched; builds the four candidate quaternions + # and keeps the one with the largest component (numerically stable across all rotations). + m00, m11, m22 = m[..., 0, 0], m[..., 1, 1], m[..., 2, 2] + m21, m12 = m[..., 2, 1], m[..., 1, 2] + m02, m20 = m[..., 0, 2], m[..., 2, 0] + m10, m01 = m[..., 1, 0], m[..., 0, 1] + q2 = torch.stack([1 + m00 + m11 + m22, 1 + m00 - m11 - m22, + 1 - m00 + m11 - m22, 1 - m00 - m11 + m22], -1) # 4 * (w^2, x^2, y^2, z^2) + cand = torch.stack([ + torch.stack([q2[..., 0], m21 - m12, m02 - m20, m10 - m01], -1), + torch.stack([m21 - m12, q2[..., 1], m10 + m01, m02 + m20], -1), + torch.stack([m02 - m20, m10 + m01, q2[..., 2], m12 + m21], -1), + torch.stack([m10 - m01, m02 + m20, m12 + m21, q2[..., 3]], -1), + ], -2) # (...,4,4) candidates, rows = wxyz + sel = q2.argmax(-1) + q = torch.gather(cand, -2, sel[..., None, None].expand(sel.shape + (1, 4)))[..., 0, :] + return q / q.norm(dim=-1, keepdim=True).clamp_min(1e-12) + + +class SplatToFile3D(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="SplatToFile3D", + display_name="Create 3D File (from Splat)", + search_aliases=["gaussian to ply", "splat to file", "export gaussian"], + category="3d/splat", + description="Serialize a gaussian splat to a File3D object for Save / Preview 3D nodes. " + "Supports one item per batch only.", + inputs=[ + IO.Splat.Input("splat"), + IO.Combo.Input("format", options=["ply", "ksplat", "spz"], # TODO: add "splat" when we have a writer for it + tooltip="ply: standard 3D Gaussian Splat with full spherical harmonics. " + "ksplat: mkkellogg SplatBuffer (level 0, uncompressed), base color only " + "spz: Niantic gzip-compressed (~10x smaller), base color only " + ), + ], + outputs=[IO.File3DSplatAny.Output(display_name="model_3d")], + ) + + @classmethod + def execute(cls, splat, format="ply") -> IO.NodeOutput: + if splat.positions.shape[0] > 1: + logging.warning("SplatToFile3D supports one item per batch only. Got %d; using first.", splat.positions.shape[0]) + end = _real_len(splat, 0) + writer = {"ksplat": _gaussian_ksplat_bytes, "spz": _gaussian_spz_bytes}.get(format, _gaussian_ply_bytes) + data = writer(splat.positions[0, :end], splat.scales[0, :end], + splat.rotations[0, :end], splat.opacities[0, :end], splat.sh[0, :end]) + return IO.NodeOutput(Types.File3D(BytesIO(data), file_format=format)) + + +class File3DToSplat(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="File3DToSplat", + display_name="Get Splat", + search_aliases=["load splat", "ply to splat", "import splat", "file to splat"], + category="3d/splat", + description="Parse a splat File3D into a gaussian splat. Inverse of Create 3D File (from Splat). " + "Supported format: PLY, SPLAT, KSPLAT, SPZ. PLY carries full spherical harmonics, " + "the other formats are base color only. Format is auto-detected from the file contents.", + inputs=[ + IO.MultiType.Input( + IO.File3DAny.Input("model_3d"), + types=[IO.File3DSplatAny, IO.File3DPLY, IO.File3DSPLAT, IO.File3DKSPLAT, IO.File3DSPZ], + tooltip="A gaussian splat 3D file", + ), + ], + outputs=[IO.Splat.Output(display_name="splat")], + ) + + @classmethod + def execute(cls, model_3d: Types.File3D) -> IO.NodeOutput: + data = model_3d.get_bytes() + fmt = (model_3d.format or "").lower() + parser = _GAUSSIAN_PARSERS.get(fmt) or _GAUSSIAN_PARSERS[_detect_splat_format(data)] + xyz, scale, rot, opacity, sh = parser(data) + + t = lambda a: torch.from_numpy(np.ascontiguousarray(a)).float() + splat = Types.SPLAT( + t(xyz)[None], # (1, N, 3) + t(scale)[None], # (1, N, 3) linear + t(rot)[None], # (1, N, 4) wxyz + t(opacity).reshape(1, -1, 1), # (1, N, 1) + t(sh)[None], # (1, N, K, 3) + ) + return IO.NodeOutput(splat) + + +def _view_matrix_t(yaw_deg, pitch_deg, device): + y, p = math.radians(yaw_deg), math.radians(pitch_deg) + cy, sy, cp, sp = math.cos(y), math.sin(y), math.cos(p), math.sin(p) + Ry = torch.tensor([[cy, 0, sy], [0, 1, 0], [-sy, 0, cy]], device=device) + Rx = torch.tensor([[1, 0, 0], [0, cp, -sp], [0, sp, cp]], device=device) + return Rx @ Ry + + +def _camera_basis(camera_info, dev): + # Look-at basis in the splat frame, named by their projection rows: right = image +x, up = image +y + # (down, since yflip=1), fwd = view/depth axis (eye -> scene). Load3D is three.js (right-handed, Y-up, + # camera looks down -Z); the splat is 3DGS (Y-down, Z-forward). World -> splat is a 180 deg rotation + # about X: (x, y, z) -> (x, -y, -z) (det +1, no mirror, no axis swap). + pos, tgt = camera_info.get("position", {}), camera_info.get("target", {}) + m = lambda d: torch.tensor([float(d.get("x", 0.0)), -float(d.get("y", 0.0)), -float(d.get("z", 0.0))], device=dev) + eye, target = m(pos), m(tgt) + mv = lambda v: torch.stack([v[0], -v[1], -v[2]]) # same world->splat map, for direction vectors + n = lambda v: v / v.norm().clamp_min(1e-8) + q = camera_info.get("quaternion") + if q: # exact camera world rotation (incl. roll) + qwxyz = torch.tensor([float(q.get("w", 1.0)), float(q.get("x", 0.0)), + float(q.get("y", 0.0)), float(q.get("z", 0.0))], device=dev) + R = _quat_to_mat(qwxyz[None])[0] # columns = camera world axes; looks down local -Z + right = n(mv(R[:, 0])) # camera +X -> image right + up = n(mv(-R[:, 1])) # camera +Y is image up; image-down row is its negative + fwd = n(mv(-R[:, 2])) # camera looks down local -Z -> view direction + return eye, target, right, up, fwd + fwd = n(target - eye) # no quaternion: orbit-consistent, roll-free + yaw = math.degrees(math.atan2(-float(fwd[0]), float(fwd[2]))) + pitch = math.degrees(math.asin(max(-1.0, min(1.0, float(fwd[1]))))) + W = _view_matrix_t(yaw, pitch, dev) + return eye, target, W[0], W[1], W[2] + + +def _lookat_quat_wxyz(position, target, dev): + # three.js lookAt in world frame: camera local +Z = (eye - target), up = world +Y. Returns wxyz. + z = position - target + z = z / z.norm().clamp_min(1e-8) + up0 = torch.tensor([0.0, 1.0, 0.0], device=dev) + if z.dot(up0).abs() > 0.999: # looking straight up/down + up0 = torch.tensor([0.0, 0.0, 1.0], device=dev) + x = torch.linalg.cross(up0, z) + x = x / x.norm().clamp_min(1e-8) + y = torch.linalg.cross(z, x) + R = torch.stack([x, y, z], dim=1) # columns = camera world axes + return _mat_to_quat(R[None])[0] + + +def _lookat_camera_info(position, target, fov, dev, zoom=1.0, camera_type="perspective", roll=0.0): + # Build a camera_info from a world-space (right-handed, Y-up) eye + look-at target; up = world +Y. + pos = torch.as_tensor(position, dtype=torch.float32, device=dev) + tgt = torch.as_tensor(target, dtype=torch.float32, device=dev) + q = _lookat_quat_wxyz(pos, tgt, dev) + if roll: # roll about the view axis (camera local Z) + a = math.radians(roll) + qz = torch.tensor([math.cos(a / 2), 0.0, 0.0, math.sin(a / 2)], device=dev) + q = _quat_mul(q[None], qz[None])[0] + xyz = lambda v: {"x": float(v[0]), "y": float(v[1]), "z": float(v[2])} + return {"position": xyz(pos), "target": xyz(tgt), + "quaternion": {"x": float(q[1]), "y": float(q[2]), "z": float(q[3]), "w": float(q[0])}, + "fov": float(fov), "cameraType": str(camera_type), "zoom": float(zoom)} + + +def _quat_camera_info(position, quat_xyzw, fov, dev, zoom=1.0, camera_type="perspective"): + # camera_info from an explicit world position + camera-rotation quaternion (three.js: looks down local -Z). + pos = torch.as_tensor(position, dtype=torch.float32, device=dev) + qx, qy, qz, qw = (float(c) for c in quat_xyzw) + qwxyz = torch.tensor([qw, qx, qy, qz], dtype=torch.float32, device=dev) + qwxyz = qwxyz / qwxyz.norm().clamp_min(1e-8) + R = _quat_to_mat(qwxyz[None])[0] + tgt = pos - R[:, 2] # look one unit down local -Z + xyz = lambda v: {"x": float(v[0]), "y": float(v[1]), "z": float(v[2])} + return {"position": xyz(pos), "target": xyz(tgt), + "quaternion": {"x": float(qwxyz[1]), "y": float(qwxyz[2]), "z": float(qwxyz[3]), "w": float(qwxyz[0])}, + "fov": float(fov), "cameraType": str(camera_type), "zoom": float(zoom)} + + +def _orbit_camera_info(yaw, pitch, distance, fov, pivot_splat, dev): + # Orbit helper for RenderSplat's default camera: yaw/pitch about `pivot_splat` (splat frame) at `distance`. + # World<->splat is the (x,-y,-z) map, so _camera_basis recovers exactly _view_matrix_t(yaw, pitch). + y, p = math.radians(yaw), math.radians(pitch) + cy, sy, cp, sp = math.cos(y), math.sin(y), math.cos(p), math.sin(p) + fwd_splat = torch.tensor([-cp * sy, sp, cp * cy], device=dev) # == _view_matrix_t(yaw, pitch)[2] + m = lambda v: torch.stack([v[0], -v[1], -v[2]]) # splat<->world (its own inverse) + return _lookat_camera_info(m(pivot_splat - distance * fwd_splat), m(pivot_splat), fov, dev) + + +def _orbit_camera_info_yaw(camera_info, angle_deg, dev): + # Turntable: rigidly rotate a camera_info about world +Y around its target by angle_deg. Returns a new dict. + a = math.radians(angle_deg) + ca, sa = math.cos(a), math.sin(a) + v = lambda d: torch.tensor([float(d.get("x", 0.0)), float(d.get("y", 0.0)), float(d.get("z", 0.0))], device=dev) + pos, tgt = v(camera_info.get("position", {})), v(camera_info.get("target", {})) + Ry = torch.tensor([[ca, 0.0, sa], [0.0, 1.0, 0.0], [-sa, 0.0, ca]], device=dev) + new_pos = tgt + Ry @ (pos - tgt) + q = camera_info.get("quaternion") or {} + qcur = torch.tensor([float(q.get("w", 1.0)), float(q.get("x", 0.0)), + float(q.get("y", 0.0)), float(q.get("z", 0.0))], device=dev) + qy = torch.tensor([math.cos(a / 2), 0.0, math.sin(a / 2), 0.0], device=dev) # world +Y rotation + qn = _quat_mul(qy[None], qcur[None])[0] + xyz = lambda t: {"x": float(t[0]), "y": float(t[1]), "z": float(t[2])} + return {**camera_info, "position": xyz(new_pos), + "quaternion": {"x": float(qn[1]), "y": float(qn[2]), "z": float(qn[3]), "w": float(qn[0])}} + + +def _gauss_blur(x, sigma, dev): + # Separable Gaussian blur of (1, C, H, W). Used to denoise the screen-space normal map. + r = max(1, int(round(3 * sigma))) + k = torch.exp(-0.5 * (torch.arange(-r, r + 1, device=dev, dtype=torch.float32) / sigma) ** 2) + k = k / k.sum() + c = x.shape[1] + x = torch.nn.functional.conv2d(x, k.view(1, 1, 1, -1).expand(c, 1, 1, -1), padding=(0, r), groups=c) + x = torch.nn.functional.conv2d(x, k.view(1, 1, -1, 1).expand(c, 1, -1, 1), padding=(r, 0), groups=c) + return x + + +def _render_gaussian(xyz, rgb, opacity, scale, rot, width, height, splat_scale, bg, camera_info, + sharpen=1.0, headlight_shading=0.0, render_style="color"): + # Perspective-correct anisotropic gaussian splat rasterizer. Each splat is weighted by its 3D Gaussian's + # peak along each pixel's ray (AAA / Hahlbohm), composited front-to-back across depth slabs. `render_style` + # selects the image: color / clay / depth / normal. Returns (image HxWx3, coverage mask HxW) on CPU. + dev = comfy.model_management.get_torch_device() + t = lambda a: torch.as_tensor(a, dtype=torch.float32, device=dev) + idev, idtype = comfy.model_management.intermediate_device(), comfy.model_management.intermediate_dtype() + xyz, rgb, opacity = t(xyz), t(rgb).clamp(0, 1), t(opacity).reshape(-1) + scale, rot = t(scale) * float(splat_scale), t(rot) + do_linear = render_style == "color" # colour blends in linear light, re-encoded at the end + if do_linear: + rgb = _srgb_to_linear(rgb) + flat = width * height + bg_t = t(bg) + bg_comp = _srgb_to_linear(bg_t) if do_linear else bg_t # background blended in the same space as the splats + need_depth = render_style == "depth" + need_normal = render_style in ("normal", "clay") or headlight_shading > 0 + + def background_only(): # no splats to rasterize -> just the background + empty mask + img = bg_t.expand(height, width, 3) if render_style == "color" else torch.zeros(height, width, 3, device=dev) + return img.to(idev, idtype), torch.zeros(height, width, device=idev, dtype=idtype) + + if xyz.shape[0] == 0: # empty input (e.g. all culled by opacity_threshold) + return background_only() + + eye, target, right, up, fwd = _camera_basis(camera_info, dev) # all camera state comes from camera_info + W = torch.stack([right, up, fwd], 0) # rows = camera axes (world -> camera) + cam = (xyz - eye) @ W.T + fov = float(camera_info.get("fov", 0) or 0) or 35.0 + zoom = float(camera_info.get("zoom", 1.0) or 1.0) # three.js digital zoom: scales the focal length + is_ortho = str(camera_info.get("cameraType", "")).lower().startswith("ortho") + xc, yc, zc = cam.unbind(-1) + + keep = zc > 1e-2 + xc, yc, zc, rgb, opacity, scale, rot = (a[keep] for a in (xc, yc, zc, rgb, opacity, scale, rot)) + if xc.shape[0] == 0: # nothing in front of the camera -> background only + return background_only() + if render_style == "clay": + rgb = torch.full_like(rgb, 0.75) # neutral albedo -> shading shows pure geometry + + f = (min(width, height) / 2) / math.tan(math.radians(fov) / 2) * zoom # fov over the smaller axis, x camera zoom + cx0, cy0 = width / 2, height / 2 + + # Camera-space 3D covariance per splat: Sigma = (W Rq) diag(scale^2) (W Rq)^T, plus a tiny relative + # regularizer for a stable inverse (a pixel-size Mip low-pass would over-thicken flat surfels and blur). + Mw = W[None] @ _quat_to_mat(rot) # (N,3,3) world -> camera + cam_cov = (Mw * scale.square()[:, None, :]) @ Mw.transpose(1, 2) + cam_cov = cam_cov + (cam_cov.diagonal(dim1=-2, dim2=-1).mean(-1) * 1e-3)[:, None, None] * torch.eye(3, device=dev) + + # Perspective-correct weighting: peak of the 3D Gaussian along each pixel ray. Precompute Si, Si@mu, mu^T Si mu. + mu = torch.stack([xc, yc, zc], -1) + si = torch.linalg.inv(cam_cov) + simu = (si @ mu[:, :, None])[:, :, 0] # (N,3) + musimu = (mu * simu).sum(-1) # (N,) + s00, s01, s02 = si[:, 0, 0], si[:, 0, 1], si[:, 0, 2] + s11, s12, s22 = si[:, 1, 1], si[:, 1, 2], si[:, 2, 2] + simu0, simu1, simu2 = simu.unbind(-1) + if need_normal: # surfel normal = thinnest axis, oriented toward camera + nrm = Mw[torch.arange(Mw.shape[0], device=dev), :, scale.argmin(-1)] # (N,3) camera-space normal + nrm = nrm * torch.where(nrm[:, 2:3] > 0, -1.0, 1.0) # flip so nz <= 0 (faces camera) + + # Screen centre (exact) + footprint radius from the affine 2D projection (used only to size the kernel). + # The image is +y-down, so the projection's y row is unflipped - it matches the splat frame's +Y. + jm = torch.zeros(xc.shape[0], 2, 3, device=dev) + if is_ortho: # parallel projection: screen = s * (xc, yc) + s = f / float((target - eye).norm().clamp_min(1e-6)) # pixels per world unit at the target plane + cx, cy = cx0 + s * xc, cy0 + s * yc + jm[:, 0, 0] = s + jm[:, 1, 1] = s + else: # perspective: screen = f * (xc, yc) / zc + invz = 1.0 / zc + cx, cy = cx0 + f * xc * invz, cy0 + f * yc * invz + jm[:, 0, 0], jm[:, 0, 2] = f * invz, -f * xc * invz.square() + jm[:, 1, 1], jm[:, 1, 2] = f * invz, -f * yc * invz.square() + cov2 = jm @ cam_cov @ jm.transpose(1, 2) + a, b, c = cov2[:, 0, 0], cov2[:, 0, 1], cov2[:, 1, 1] + max_eig = (a + c) * 0.5 + (((a - c) * 0.5).square() + b * b).clamp_min(0).sqrt() + radius = 3.0 * max_eig.clamp_min(1e-8).sqrt() + K = int(min(max(24, min(width, height) // 16), max(2, math.ceil(_quantile(radius, 0.995).item())))) + + # Per-splat kernel size: bucket splats by radius into a coarse ladder of window sizes (global K stays the cap) so + # small splats (the bulk of it) use a small window. + levels = [L for L in (16, 64, 256) if L < K] + [K] + levels_t = torch.tensor(levels, device=dev, dtype=torch.float32) + grids = [] + for L in levels: + rng = torch.arange(-L, L + 1, device=dev, dtype=torch.float32) + gy, gx = torch.meshgrid(rng, rng, indexing="ij") + grids.append((gx.reshape(-1), gy.reshape(-1))) + blevel = torch.bucketize(radius * (4.0 / 3.0), levels_t).clamp_(max=len(levels) - 1) # window >= ~4 sigma + + n = zc.shape[0] + ns = int(min(256, max(1, n // 1000))) # depth slabs: 1 per ~1000 splats, capped + nl = len(levels) + order = torch.argsort(zc) # front (small zc) -> back -> defines the slabs + bounds = torch.linspace(0, n, ns + 1, device=dev).round().long() + rank = torch.empty(n, dtype=torch.long, device=dev) + rank[order] = torch.arange(n, device=dev) # depth rank of each splat + slab_id = (torch.searchsorted(bounds, rank, right=True) - 1).clamp_(0, ns - 1) + key = slab_id * nl + blevel # group by slab, then kernel level (order-free within) + order = torch.argsort(key) + key = key[order] + + cxr, cyr = cx[order].round(), cy[order].round() + s00, s01, s02 = s00[order], s01[order], s02[order] + s11, s12, s22 = s11[order], s12[order], s22[order] + s01b, s02b, s12b = s01 * 2, s02 * 2, s12 * 2 # doubled cross terms for the fused quadratic forms + simu0, simu1, simu2, musimu = simu0[order], simu1[order], simu2[order], musimu[order] + opacity, rgb = opacity[order], rgb[order] + zc_o = zc[order] if need_depth else None + nrm_o = nrm[order] if need_normal else None + mux_o, muy_o, muz_o = (xc[order], yc[order], zc[order]) if is_ortho else (None, None, None) + + # Pack the per-splat scalars into one tensor so each chunk slices once + common = [cxr, cyr, s00, s11, s22, s01b, s02b, s12b, opacity] + pstack = torch.stack(common + ([s02, s12, mux_o, muy_o, muz_o] if is_ortho else [simu0, simu1, simu2, musimu])) + + # Precompute the (slab, level) run table on-GPU and pull it to the CPU once + starts = torch.cat([torch.zeros(1, dtype=torch.long, device=dev), (key[1:] != key[:-1]).nonzero().flatten() + 1]) + ks = key[starts] + run_lo = starts.tolist() + [n] + run_lev = (ks % nl).tolist() + run_slab = torch.div(ks, nl, rounding_mode="floor").tolist() + slab_runs = [[] for _ in range(ns)] + for r in range(len(run_lev)): + slab_runs[run_slab[r]].append((run_lo[r], run_lo[r + 1], run_lev[r])) + + def splat(lo, hi, ox, oy): # -> pixel idx (m,M), alpha (m,M); weight = 3D Gaussian peak along each pixel's ray + cols = pstack[:, lo:hi, None].unbind(0) + cxr_, cyr_, a00, a11, a22, b01, b02, b12, opa = cols[:9] # a* = Si components; b* = 2 * cross terms + px = cxr_ + ox[None, :] + py = cyr_ + oy[None, :] + valid = (px >= 0) & (px < width) & (py >= 0) & (py < height) + if is_ortho: # parallel ray (0,0,1) from screen point (X, Y, 0); rz constant per splat + c02, c12, mx, my, mz = cols[9:] + rx = (px - cx0) / s - mx + ry = (py - cy0) / s - my + rz = -mz + a22rz = a22 * rz + inx = torch.addcmul(b02 * rz, a00, rx).addcmul_(b01, ry) # a00 rx + b01 ry + b02 rz + rSr = torch.addcmul(a22rz * rz, rx, inx).addcmul_(ry, torch.addcmul(b12 * rz, a11, ry)) + dsr = torch.addcmul(a22rz, c02, rx).addcmul_(c12, ry) + q = torch.addcdiv(rSr, dsr * dsr, a22.clamp_min(1e-12), value=-1).clamp_min_(0) + else: # perspective ray (dx,dy,1) through the camera origin + su0, su1, su2, mus = cols[9:] + dx, dy = (px - cx0) / f, (py - cy0) / f + dsid = torch.addcmul(a22, dx, torch.addcmul(b02, a00, dx)) # a22 + dx*(a00 dx + b02) + dsid = dsid.addcmul_(dy, torch.addcmul(b12, a11, dy)) # + dy*(a11 dy + b12) + dsid = dsid.addcmul_(b01 * dx, dy) # + (2 s01) dx dy + dsimu = torch.addcmul(su2, dx, su0).addcmul_(dy, su1) + q = torch.addcdiv(mus, dsimu * dsimu, dsid.clamp_min(1e-12), value=-1).clamp_min_(0) + alpha = (opa * torch.exp(-0.5 * q) * valid).clamp_(0, 0.999) + idx = py.long().clamp(0, height - 1) * width + px.long().clamp(0, width - 1) + return idx, alpha + + # Front-to-back compositing over the depth slabs set up above. Within a slab the accumulation is a pure + # sum (order-independent), so splats are grouped by kernel level and each level uses its own tight window. + sharp = sharpen != 1.0 # winner-take-more colour blend: dominant splat shows more + cacc = torch.zeros((flat, 3), device=dev) + trans = torch.ones((flat,), device=dev) + a_buf = torch.zeros((flat,), device=dev) # sum alpha -> colour/depth/normal weight (alpha-weighted mean) + tau_buf = torch.zeros((flat,), device=dev) # sum -ln(1-alpha) -> slab opacity = 1-prod(1-alpha) + crgb = torch.zeros((flat, 3), device=dev) # sum alpha^p * rgb -> slab colour + wbuf = torch.zeros((flat,), device=dev) if sharp else None # sum alpha^p -> colour normalizer (sharp only) + dacc = torch.zeros((flat,), device=dev) if need_depth else None # front-weighted depth + nacc = torch.zeros((flat, 3), device=dev) if need_normal else None # front-weighted camera-space normal + zslab = torch.zeros((flat,), device=dev) if need_depth else None + nslab = torch.zeros((flat, 3), device=dev) if need_normal else None + stale = 0 # consecutive fully-occluded slabs -> early-out + for si in range(ns): + runs = slab_runs[si] + if not runs: + continue + a_buf.zero_() + tau_buf.zero_() + crgb.zero_() + if sharp: + wbuf.zero_() + if need_depth: + zslab.zero_() + if need_normal: + nslab.zero_() + for r_lo, r_hi, li in runs: # contiguous same-kernel-level runs in this slab + ox, oy = grids[li] + ch = max(2048, 10_000_000 // ox.shape[0]) # splats/chunk, bounded by this level's kernel size + for lo in range(r_lo, r_hi, ch): + hi = min(lo + ch, r_hi) + idx, alpha = splat(lo, hi, ox, oy) + idx, af = idx.reshape(-1), alpha.reshape(-1) + a_buf.index_add_(0, idx, af) + tau_buf.index_add_(0, idx, (-torch.log1p(-alpha)).reshape(-1)) # -ln(1-alpha), correct opacity merge + apw = alpha.pow(sharpen) if sharp else alpha # bias colour toward the highest-alpha splat + crgb.index_add_(0, idx, (apw[:, :, None] * rgb[lo:hi, None, :]).reshape(-1, 3)) + if sharp: + wbuf.index_add_(0, idx, apw.reshape(-1)) + if need_depth: + zslab.index_add_(0, idx, (alpha * zc_o[lo:hi, None]).reshape(-1)) + if need_normal: + nslab.index_add_(0, idx, (alpha[:, :, None] * nrm_o[lo:hi, None, :]).reshape(-1, 3)) + slab_a = 1 - torch.exp(-tau_buf) # 1 - prod(1-alpha): true opacity of the slab's splats + front = trans * slab_a + denom = wbuf if sharp else a_buf + cacc.addcmul_(front[:, None], crgb / denom.clamp_min(1e-8)[:, None]) # cacc += front * (crgb/denom) + if need_depth or need_normal: + ainv = a_buf.clamp_min(1e-8) # alpha-weighted-mean normalizer (depth/normal only) + if need_depth: + dacc.addcmul_(front, zslab / ainv) + if need_normal: + nacc.addcmul_(front[:, None], nslab / ainv[:, None]) + trans.mul_(1 - slab_a) + if si % 8 == 7: # checkpoint every 8 slabs (a per-slab GPU sync would cost more) + if float(front.max()) < 1e-3: # this checkpoint slab is fully occluded by what is in front + stale += 1 + if stale >= 2: # two occluded checkpoints running -> the rest are too -> stop + break + else: + stale = 0 + + cov = 1 - trans + covg = cov.reshape(height, width) + covm = covg > 0.5 if render_style in ("depth", "normal") else None # silhouette mask (depth/normal styles only) + depth_map = (dacc / cov.clamp_min(1e-6)).reshape(height, width) if need_depth else None + nrm_map = None + if need_normal: + # Per-splat surfel normals are jittery, so do a masked blur + nb = nacc.reshape(height, width, 3).permute(2, 0, 1)[None] + cb = cov.reshape(1, 1, height, width) + nb, cb = _gauss_blur(nb, 1.2, dev), _gauss_blur(cb, 1.2, dev) + normal = (nb / cb.clamp_min(1e-6))[0].permute(1, 2, 0) + nrm_map = normal / normal.norm(dim=-1, keepdim=True).clamp_min(1e-6) + + if render_style == "depth": # near = bright, far = dark, 0 off-object + d = torch.zeros(height, width, device=dev) + if bool(covm.any()): + lo, hi = depth_map[covm].min(), depth_map[covm].max() + d = torch.where(covm, ((hi - depth_map) / (hi - lo).clamp_min(1e-6)).clamp(0, 1), d) + img = d[:, :, None].expand(height, width, 3) + elif render_style == "normal": # OpenGL normal map: +X right, +Y up, +Z to viewer + enc = (nrm_map * t([1.0, -1.0, -1.0]) * 0.5 + 0.5).clamp(0, 1) + img = enc * covm[:, :, None] + else: # color / clay + img = cacc.reshape(height, width, 3) + if render_style == "clay": # studio key light + ambient -> sculpted matte look + kl = t([-0.4, -0.7, -0.6]) # key from screen upper-left, angled toward the viewer + kl = kl / kl.norm() + hl = (0.5 * (nrm_map * kl).sum(-1) + 0.5).clamp(0, 1) # half-Lambert: soft terminator, no harsh dark side + img = img * (0.35 + 0.65 * hl * hl)[:, :, None] # ambient floor + diffuse key + elif headlight_shading > 0: # camera headlight: darken faces turned from view + k = float(headlight_shading) + ndotl = (-nrm_map[:, :, 2]).clamp(0, 1) + img = img * (1 - 0.6 * k + 0.6 * k * ndotl)[:, :, None] + img = img.addcmul_(trans.reshape(height, width, 1), bg_comp) + if do_linear: # back to display space after linear compositing + img = _linear_to_srgb(img) + return img.clamp(0, 1).to(idev, idtype), covg.clamp(0, 1).to(idev, idtype) + + +class RenderSplat(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="RenderSplat", + display_name="Render Splat", + search_aliases=["splat to image", "render splat", "gaussian turntable"], + category="3d/splat", + description="Render a gaussian splat as an image with an anisotropic EWA rasterizer (oriented " + "elliptical splats, antialiased, depth-sorted front-to-back). The camera comes from a " + "camera_info input (Load / Preview 3D, or a Create Camera Info node); leave it empty to " + "auto-frame the splat. Set frames greater than 1 for a turntable batch of images to feed a Video node.", + inputs=[ + IO.Splat.Input("splat"), + IO.Int.Input("width", default=1024, min=64, max=2048, step=8), + IO.Int.Input("height", default=1024, min=64, max=2048, step=8), + IO.Int.Input("frames", default=1, min=-240, max=240, + tooltip="-1, 0, 1 = single still image; >1 = turntable, the camera orbits over a full " + "360 turn (works with any camera_info). Negative value orbits the other way."), + IO.Float.Input("splat_scale", default=1.0, min=0.1, max=5.0, step=0.05, advanced=True, + tooltip="Multiplier on each splat's projected footprint (lower = crisper points, " + "higher = softer/fuller surface)."), + IO.Float.Input("sharpen", default=2.0, min=1.0, max=8.0, step=0.5, + tooltip="Sharpen overlapping splats: 1.0 = physically-correct blend; higher biases " + "each pixel toward its dominant (nearest) splat for crisper texture, without " + "shrinking splats or opening gaps. Non-physical above 1."), + IO.Float.Input("headlight_shading", default=0.0, min=0.0, max=3.0, step=0.05, advanced=True, + tooltip="Diffuse shading from a light at the camera (headlight), using the splat surfel " + "normals: darkens surfaces that turn away from view to reveal form/curvature. " + "0 = flat albedo, 1 = strongest shading."), + IO.Float.Input("opacity_threshold", default=0.0, min=0.0, max=1.0, step=0.01, advanced=True, + tooltip="Cull gaussians with opacity below this (removes faint floaters)."), + IO.Combo.Input("render_style", options=["color", "clay", "depth", "normal"], + tooltip="What the image output shows: color, clay (neutral-albedo shaded), " + "depth (near=bright), normal (OpenGL normal map)."), + IO.Color.Input("background", default="#000000"), + IO.Image.Input("bg_image", optional=True, + tooltip="Optional background plate composited behind the splat (overrides the solid " + "background colour). Resized to the render size; a batch is used per frame, " + "a single image for all. color/clay only."), + IO.Load3DCamera.Input("camera_info", optional=True, + tooltip="Camera to render from - a Load3D / Preview3D camera or a Create Camera " + "Info node. If empty, the splat is auto-framed from a default 3/4 view."), + ], + outputs=[IO.Image.Output(display_name="image"), IO.Mask.Output(display_name="mask")], + ) + + @classmethod + def execute(cls, splat, width, height, frames, splat_scale, sharpen, headlight_shading, + opacity_threshold, background, render_style, camera_info=None, bg_image=None) -> IO.NodeOutput: + bg = _hex_to_rgb(background) + bg_imgs = None + if bg_image is not None: # resize the plate(s) to the render size: (B,H,W,3) + bi = bg_image[... , :3].movedim(-1, 1) # (B,3,H,W) + bi = comfy.utils.common_upscale(bi, width, height, "bicubic", "disabled") + bg_imgs = bi.movedim(1, -1).clamp(0, 1) + n_frames = abs(int(frames)) or 1 # magnitude = frame count (0 -> single still) + orbit_dir = -1.0 if frames < 0 else 1.0 # sign = orbit direction + imgs, masks = [], [] + device = comfy.model_management.get_torch_device() + total = splat.positions.shape[0] * n_frames + pbar = comfy.utils.ProgressBar(total) if total > 1 else None + k = 0 + for i in range(splat.positions.shape[0]): + xyz, rgb, opacity, scale, rot = _gaussian_item(splat, i, device) + if opacity_threshold > 0: + keep = opacity >= opacity_threshold + xyz, rgb, opacity, scale, rot = xyz[keep], rgb[keep], opacity[keep], scale[keep], rot[keep] + base_cam = camera_info + if base_cam is None: # no camera -> default 3/4 view, auto-framed on the splat + center = xyz.mean(0) if xyz.shape[0] else torch.zeros(3, device=device) + extent = (_quantile((xyz - center).norm(dim=-1), 0.99).clamp_min(1e-4) if xyz.shape[0] + else torch.tensor(1.0, device=device)) + dist = float(extent / (math.tan(math.radians(35.0) / 2) * 0.9)) + base_cam = _orbit_camera_info(35.0, 30.0, dist, 35.0, center, device) + for fr in range(n_frames): + cam_fr = (base_cam if n_frames == 1 + else _orbit_camera_info_yaw(base_cam, orbit_dir * 360.0 * fr / n_frames, device)) + bg_k = bg_imgs[k % bg_imgs.shape[0]] if bg_imgs is not None else bg # per-frame plate, or solid colour + img, mask = _render_gaussian(xyz, rgb, opacity, scale, rot, width, height, splat_scale, bg_k, cam_fr, + sharpen=sharpen, headlight_shading=headlight_shading, + render_style=render_style) + imgs.append(img) + masks.append(mask) + k += 1 + if pbar is not None: + pbar.update(1) + return IO.NodeOutput(torch.stack(imgs), torch.stack(masks)) + + +class CreateCameraInfo(IO.ComfyNode): # TODO: move to better file + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="CreateCameraInfo", + display_name="Create Camera Info", + search_aliases=["camera position", "make camera info", "orbit camera", "look at camera"], + category="3d", + description="Build a camera_info" + "Mode 'orbit' aims with yaw/pitch/distance around the target; " + "'look_at' places the camera at world position. Coordinates are the viewer's world space (right-handed,Y-up).", + inputs=[ + IO.DynamicCombo.Input("mode", options=[ + IO.DynamicCombo.Option("orbit", [ + IO.Float.Input("yaw", default=35.0, min=-360.0, max=360.0, step=1.0), + IO.Float.Input("pitch", default=30.0, min=-89.0, max=89.0, step=1.0), + IO.Float.Input("distance", default=4.0, min=0.01, max=1000.0, step=0.01, + tooltip="Camera distance from the target."), + ]), + IO.DynamicCombo.Option("look_at", [ + IO.Float.Input("position_x", default=4.0, min=-1000.0, max=1000.0, step=0.01, + tooltip="Camera position in world space (right-handed, Y-up)."), + IO.Float.Input("position_y", default=4.0, min=-1000.0, max=1000.0, step=0.01), + IO.Float.Input("position_z", default=4.0, min=-1000.0, max=1000.0, step=0.01), + ]), + IO.DynamicCombo.Option("quaternion", [ + IO.Float.Input("position_x", default=4.0, min=-1000.0, max=1000.0, step=0.01, + tooltip="Camera position in world space (right-handed, Y-up)."), + IO.Float.Input("position_y", default=4.0, min=-1000.0, max=1000.0, step=0.01), + IO.Float.Input("position_z", default=4.0, min=-1000.0, max=1000.0, step=0.01), + IO.Float.Input("quat_x", default=0.0, min=-1.0, max=1.0, step=0.001), + IO.Float.Input("quat_y", default=0.0, min=-1.0, max=1.0, step=0.001), + IO.Float.Input("quat_z", default=0.0, min=-1.0, max=1.0, step=0.001), + IO.Float.Input("quat_w", default=1.0, min=-1.0, max=1.0, step=0.001, + tooltip="Camera world-rotation quaternion (three.js: looks down local -Z). Normalized for you."), + ]), + ], tooltip="How to define the camera: orbit angles, an explicit position, or a position + quaternion."), + IO.Float.Input("target_x", default=0.0, min=-1000.0, max=1000.0, step=0.01, advanced=True, + tooltip="Look-at point (orbit pivot / aim). In orbit mode, move it to pan/translate the " + "whole camera. Ignored in quaternion mode. Defaults to the origin."), + IO.Float.Input("target_y", default=0.0, min=-1000.0, max=1000.0, step=0.01, advanced=True), + IO.Float.Input("target_z", default=0.0, min=-1000.0, max=1000.0, step=0.01, advanced=True), + IO.Float.Input("roll", default=0.0, min=-180.0, max=180.0, step=1.0, + tooltip="Camera roll about the view axis, degrees."), + IO.Float.Input("fov", default=35.0, min=1.0, max=120.0, step=1.0, + tooltip="Vertical field of view in degrees."), + IO.Float.Input("zoom", default=1.0, min=0.01, max=100.0, step=0.01, + tooltip="Digital zoom (focal-length multiplier). >1 zooms in without moving the camera."), + IO.Combo.Input("camera_type", options=["perspective", "orthographic"], + tooltip="Projection used by Render Splat: perspective (foreshortening) or orthographic (parallel)."), + ], + outputs=[IO.Load3DCamera.Output(display_name="camera_info")], + ) + + @classmethod + def execute(cls, mode, target_x, target_y, target_z, roll, fov, zoom=1.0, camera_type="perspective") -> IO.NodeOutput: + dev = comfy.model_management.get_torch_device() + kind = mode["mode"] + if kind == "quaternion": # explicit world position + camera rotation + position = [mode["position_x"], mode["position_y"], mode["position_z"]] + quat = [mode["quat_x"], mode["quat_y"], mode["quat_z"], mode["quat_w"]] + return IO.NodeOutput(_quat_camera_info(position, quat, fov, dev, zoom=zoom, camera_type=camera_type)) + target = [target_x, target_y, target_z] # orbit pivot / aim; move it to pan the whole camera + if kind == "orbit": # yaw/pitch/distance about the target (world Y-up) + y, p = math.radians(mode["yaw"]), math.radians(mode["pitch"]) + cy, sy, cp, sp = math.cos(y), math.sin(y), math.cos(p), math.sin(p) + d = mode["distance"] + position = [target_x + d * cp * sy, target_y + d * sp, target_z + d * cp * cy] + else: # look_at: explicit world-space camera position + position = [mode["position_x"], mode["position_y"], mode["position_z"]] + return IO.NodeOutput(_lookat_camera_info(position, target, fov, dev, zoom=zoom, camera_type=camera_type, roll=roll)) + + +class TransformSplat(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="TransformSplat", + display_name="Transform Splat", + search_aliases=["move splat", "rotate splat", "scale splat", "gaussian transform"], + category="3d/splat", + description="Translate, rotate, and scale a gaussian splat. " + "Non-uniform scale also reshapes every individual splat, slower process.", + inputs=[ + IO.Splat.Input("splat"), + IO.Float.Input("translate_x", default=0.0, min=-100.0, max=100.0, step=0.01), + IO.Float.Input("translate_y", default=0.0, min=-100.0, max=100.0, step=0.01), + IO.Float.Input("translate_z", default=0.0, min=-100.0, max=100.0, step=0.01), + IO.Float.Input("rotate_x", default=0.0, min=-360.0, max=360.0, step=1.0), + IO.Float.Input("rotate_y", default=0.0, min=-360.0, max=360.0, step=1.0), + IO.Float.Input("rotate_z", default=0.0, min=-360.0, max=360.0, step=1.0), + IO.Float.Input("scale_x", default=1.0, min=0.01, max=100.0, step=0.01), + IO.Float.Input("scale_y", default=1.0, min=0.01, max=100.0, step=0.01), + IO.Float.Input("scale_z", default=1.0, min=0.01, max=100.0, step=0.01), + ], + outputs=[IO.Splat.Output(display_name="splat")], + ) + + @classmethod + def execute(cls, splat, translate_x, translate_y, translate_z, + rotate_x, rotate_y, rotate_z, scale_x, scale_y, scale_z) -> IO.NodeOutput: + pos = splat.positions + dev, dt = pos.device, pos.dtype + q_rot = _euler_to_quat(rotate_x, rotate_y, rotate_z).to(device=dev, dtype=dt) + R = _quat_to_mat(q_rot[None])[0] # (3, 3) node rotation + D = torch.tensor([scale_x, scale_y, scale_z], dtype=dt, device=dev) + A = D[:, None] * R # diag(D) @ R: per-axis scale after rotation + t = torch.tensor([translate_x, translate_y, translate_z], dtype=dt, device=dev) + + positions = pos @ A.T + t # rotate, scale per-axis, then translate + if scale_x == scale_y == scale_z: # uniform: rotation/scale factor out cleanly + scales = splat.scales * scale_x + rotations = _quat_mul(q_rot.expand_as(splat.rotations), splat.rotations) + rotations = rotations / rotations.norm(dim=-1, keepdim=True).clamp_min(1e-12) + else: # non-uniform: transform Sigma = A R s^2 R^T A^T, re-extract + rg = _quat_to_mat(splat.rotations.reshape(-1, 4)) # (M,3,3) per-splat rotation + s2 = splat.scales.reshape(-1, 3).square() + cov = (rg * s2[:, None, :]) @ rg.transpose(-1, -2) # Sigma + cov = A @ cov @ A.T # A Sigma A^T (A broadcast over splats) + lam, V = torch.linalg.eigh(cov) # symmetric -> eigenvalues (asc), orthonormal axes + V = V * torch.where(torch.linalg.det(V) < 0, -1.0, 1.0)[..., None, None] # keep a proper rotation + scales = lam.clamp_min(0).sqrt().reshape(splat.scales.shape) + rotations = _mat_to_quat(V).reshape(splat.rotations.shape) + out = Types.SPLAT(positions, scales, rotations, splat.opacities, splat.sh, + counts=getattr(splat, "counts", None)) + return IO.NodeOutput(out) + + +class GetSplatCount(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="GetSplatCount", + display_name="Get Splat Count", + search_aliases=["splat count", "gaussian count", "number of splats", "splat info"], + category="3d/splat", + description="Returns the number of splats summed across the batch.", + inputs=[IO.Splat.Input("splat")], + outputs=[IO.Splat.Output(display_name="splat"), + IO.Int.Output(display_name="count"), + ], + hidden=[IO.Hidden.unique_id], + ) + + @classmethod + def execute(cls, splat) -> IO.NodeOutput: + count = sum(_real_len(splat, i) for i in range(splat.positions.shape[0])) + if cls.hidden.unique_id: # show the count inline on the node + PromptServer.instance.send_progress_text(f"{count:,} splats", cls.hidden.unique_id) + return IO.NodeOutput(splat, count) + + +def _pad_stack(items, n): + # Stack a list of (Lᵢ, *tail) tensors into (B, n, *tail), zero-padding each row up to n. + tail = items[0].shape[1:] + out = items[0].new_zeros((len(items), n, *tail)) + for i, t in enumerate(items): + out[i, :t.shape[0]] = t + return out + + +def _merge_gaussians(gaussians: list) -> Types.SPLAT: + # Concatenate SPLAT batches along the splat dimension (per item), padding SH to the highest degree. + gs = [g for g in gaussians if g is not None] + if not gs: + raise ValueError("MergeSplat: no gaussians to merge") + b = gs[0].positions.shape[0] + for g in gs: + if g.positions.shape[0] != b: + raise ValueError(f"MergeSplat: batch size mismatch ({b} vs {g.positions.shape[0]}).") + max_k = max(g.sh.shape[2] for g in gs) + + pos_b, scl_b, rot_b, op_b, sh_b, lengths = [], [], [], [], [], [] + for i in range(b): + pos_i, scl_i, rot_i, op_i, sh_i = [], [], [], [], [] + for g in gs: + end = _real_len(g, i) + pos_i.append(g.positions[i, :end]) + scl_i.append(g.scales[i, :end]) + rot_i.append(g.rotations[i, :end]) + op_i.append(g.opacities[i, :end]) + sh = g.sh[i, :end] # (end, K, 3) + if sh.shape[1] < max_k: # zero-pad lower-degree SH + sh = torch.cat([sh, sh.new_zeros(sh.shape[0], max_k - sh.shape[1], sh.shape[2])], dim=1) + sh_i.append(sh) + pos_b.append(torch.cat(pos_i)) + scl_b.append(torch.cat(scl_i)) + rot_b.append(torch.cat(rot_i)) + op_b.append(torch.cat(op_i)) + sh_b.append(torch.cat(sh_i)) + lengths.append(pos_b[-1].shape[0]) + + n = max(lengths) + counts = None + if len(set(lengths)) > 1: + counts = torch.tensor(lengths, device=gs[0].positions.device, dtype=torch.int64) + return Types.SPLAT(_pad_stack(pos_b, n), _pad_stack(scl_b, n), _pad_stack(rot_b, n), + _pad_stack(op_b, n), _pad_stack(sh_b, n), counts=counts) + + +class MergeSplat(IO.ComfyNode): + @classmethod + def define_schema(cls): + # Autogrow: a splat0/splat1/... input list that grows a fresh slot as you connect splats. + splats = IO.Autogrow.TemplatePrefix(IO.Splat.Input("splat"), prefix="splat", min=2, max=32) + return IO.Schema( + node_id="MergeSplat", + display_name="Merge Splats", + search_aliases=["union splat", "densify gaussian", "combine splat", "merge gaussian"], + category="3d/splat", + description="Concatenate any number of gaussian splats into one. Unioning several decodes of the same " + "latent at different seeds densifies the surface, this can improve surface quality when meshing.", + inputs=[IO.Autogrow.Input("splats", template=splats)], + outputs=[IO.Splat.Output(display_name="splat")], + ) + + @classmethod + def execute(cls, splats: IO.Autogrow.Type) -> IO.NodeOutput: + gs = [v for v in splats.values() if v is not None] + if not gs: + raise ValueError("MergeSplat: connect at least one splat.") + return IO.NodeOutput(_merge_gaussians(gs)) + + +def _inverse_covariance(scale, quat): + # Per-splat Sigma^-1 = R diag(1/s^2) R^T. scale (N,3) linear std, quat (N,4) wxyz -> (N,3,3). + q = quat / quat.norm(dim=1, keepdim=True).clamp_min(1e-12) + w, x, y, z = q.unbind(-1) + R = torch.stack([ + 1 - 2 * (y * y + z * z), 2 * (x * y - w * z), 2 * (x * z + w * y), + 2 * (x * y + w * z), 1 - 2 * (x * x + z * z), 2 * (y * z - w * x), + 2 * (x * z - w * y), 2 * (y * z + w * x), 1 - 2 * (x * x + y * y), + ], dim=1).reshape(-1, 3, 3) + inv_s2 = 1.0 / scale.clamp_min(1e-8) ** 2 # (N, 3) + return torch.einsum("nij,nj,nkj->nik", R, inv_s2, R) + + +def _splat_density(xyz, opacity, scale, quat, rgb, res, kernel, device, color_sharpen=1.0, chunk=4096, progress=None, + col_dtype=torch.float16): + # Splat each gaussian as its oriented-covariance disk (3-sigma, opacity-weighted) into a density grid, + # plus a colour volume. Each gaussian uses a voxel window sized to its OWN 3-sigma (capped at `kernel`). + # Colour is weighted by w^color_sharpen: >1 biases each voxel toward its dominant gaussian (crisper + # texture). Returns (density, colour numerator, colour normaliser, origin, voxel). + pad = 4.0 * scale.median() + lo = xyz.amin(0) - pad + hi = xyz.amax(0) + pad + voxel = ((hi - lo).max() / res).clamp_min(1e-8) + dx, dy, dz = (torch.ceil((hi - lo) / voxel).long() + 1).tolist() + + sinv = _inverse_covariance(scale, quat) + kreq = torch.ceil(3.0 * scale.amax(-1) / voxel).long().clamp(1, int(kernel)) # per-gaussian half-width + sharp = color_sharpen != 1.0 + vol = torch.zeros(dx * dy * dz, device=device) # Sum(w) density (surface) + colvol = torch.zeros(dx * dy * dz, 3, device=device, dtype=col_dtype) # Sum(w^p * rgb) colour numerator + wcol = torch.zeros(dx * dy * dz, device=device, dtype=col_dtype) if sharp else None # Sum(w^p) normaliser (p>1) + n, done = xyz.shape[0], 0 + for k in range(1, int(kernel) + 1): + sel = (kreq == k).nonzero(as_tuple=True)[0] + if sel.numel() == 0: + continue + rng = torch.arange(-k, k + 1, device=device, dtype=torch.float32) + off = torch.stack(torch.meshgrid(rng, rng, rng, indexing="ij"), -1).reshape(-1, 3) # (M, 3) + for st in range(0, sel.numel(), chunk): + gi = sel[st:st + chunk] + cc = xyz[gi] + idx = ((cc - lo) / voxel).round()[:, None, :] + off[None] # (b, M, 3) voxel coords + d = (lo + idx * voxel) - cc[:, None, :] # world offset to voxel center + quad = torch.einsum("bmi,bij,bmj->bm", d, sinv[gi], d) + wgt = opacity[gi, None] * torch.exp(-0.5 * quad) + wgt = torch.where(quad < 9.0, wgt, torch.zeros_like(wgt)) # clip beyond 3 sigma + ii = idx.long() + ix = ii[..., 0].clamp(0, dx - 1) + iy = ii[..., 1].clamp(0, dy - 1) + iz = ii[..., 2].clamp(0, dz - 1) + flat = (ix * (dy * dz) + iy * dz + iz).reshape(-1) + vol.index_add_(0, flat, wgt.reshape(-1)) + wp = wgt.pow(color_sharpen) if sharp else wgt # winner-take-more colour weight + colvol.index_add_(0, flat, (wp[..., None] * rgb[gi, None, :]).reshape(-1, 3).to(col_dtype)) + if sharp: + wcol.index_add_(0, flat, wp.reshape(-1).to(col_dtype)) + done += gi.numel() + if progress is not None: + progress(min(1.0, done / max(1, n))) + colnorm = (wcol if sharp else vol).reshape(dx, dy, dz) # p==1 -> Sum(w) == density + return vol.reshape(dx, dy, dz), colvol.reshape(dx, dy, dz, 3), colnorm, lo.cpu().numpy(), float(voxel) + + +def _connected_components_gpu(faces, nv): + # FastSV connected components: grandparent hooking + shortcutting, ~O(log nv) iterations. + # Returns per-vertex component labels (min node id, not densified). + a = torch.cat([faces[:, 0], faces[:, 1]]) # 2F edge endpoints: (v0,v1),(v1,v2) + b = torch.cat([faces[:, 1], faces[:, 2]]) + f = torch.arange(nv, device=faces.device) + while True: + gp = f[f] # grandparent + ga, gb = gp[a], gp[b] + new = f.clone() + new.scatter_reduce_(0, f[a], gb, "amin", include_self=True) # stochastic hooking onto roots + new.scatter_reduce_(0, f[b], ga, "amin", include_self=True) + new.scatter_reduce_(0, a, gb, "amin", include_self=True) # aggressive hooking, both directions + new.scatter_reduce_(0, b, ga, "amin", include_self=True) + new = new[new] # shortcut (path compression) + if torch.equal(new, f): + return f + f = new + + +def _clean_components_gpu(verts, faces, min_verts, device): + # GPU port of _clean_components: FastSV components + scatter reductions. Byte-identical to the numpy path + vt = torch.as_tensor(verts, device=device) + ft = torch.as_tensor(faces, device=device) + nv = vt.shape[0] + _, label = torch.unique(_connected_components_gpu(ft, nv), return_inverse=True) # dense 0..ncomp-1 + ncomp = int(label.max()) + 1 + flabel = label[ft[:, 0]] # component id per face + keep = torch.bincount(label, minlength=ncomp) >= min_verts # per-component vertex-count gate + if int(keep.sum()) > 1: + fcount = torch.bincount(flabel, minlength=ncomp) + largest = int(torch.where(keep, fcount, fcount.new_tensor(-1)).argmax()) + v0, v1, v2 = vt[ft[:, 0]], vt[ft[:, 1]], vt[ft[:, 2]] + cvol = torch.zeros(ncomp, device=device).scatter_add_(0, flabel, (v0 * torch.linalg.cross(v1, v2)).sum(-1)) + idx3 = label[:, None].expand(-1, 3) # per-component vertex bbox + cmin = torch.full((ncomp, 3), float("inf"), device=device).scatter_reduce_(0, idx3, vt, "amin", include_self=True) + cmax = torch.full((ncomp, 3), float("-inf"), device=device).scatter_reduce_(0, idx3, vt, "amax", include_self=True) + tol = 1e-4 * (cmax[largest] - cmin[largest]).max() + enclosed = (cmin >= cmin[largest] - tol).all(1) & (cmax <= cmax[largest] + tol).all(1) + inner = enclosed & (torch.sign(cvol) != torch.sign(cvol[largest])) & (torch.arange(ncomp, device=device) != largest) + keep &= ~inner + faces_k = ft[keep[flabel]] + if faces_k.shape[0] == 0: + return verts[:0], faces[:0] + used = torch.unique(faces_k) # sorted, matches np.unique + remap = torch.full((nv,), -1, dtype=torch.int64, device=device) + remap[used] = torch.arange(used.shape[0], device=device) + return vt[used].cpu().numpy(), remap[faces_k].cpu().numpy() + + +def _clean_components(verts, faces, min_verts, device=None): + # Drop floaters (components with < min_verts vertices) and inner shells - the surfel shell density + # extracts a double wall (outer + inner cavity surface). GPU path (FastSV CC + scatter reductions, ~13x + # faster) when an accelerator has headroom; else numpy/scipy. Both produce byte-identical output. + if device is not None and not comfy.model_management.is_device_cpu(device) and \ + comfy.model_management.get_free_memory(device) > 10 * faces.size * 8: # peak ~8.4x faces bytes + return _clean_components_gpu(verts, faces, min_verts, device) + nv = len(verts) + e = np.concatenate([faces[:, [0, 1]], faces[:, [1, 2]], faces[:, [0, 2]]], 0) + ncomp, label = connected_components(coo_matrix((np.ones(len(e)), (e[:, 0], e[:, 1])), shape=(nv, nv)), directed=False) + flabel = label[faces[:, 0]] # component id per face + keep = np.bincount(label, minlength=ncomp) >= min_verts # per-component vertex-count gate + if keep.sum() > 1: + fcount = np.bincount(flabel, minlength=ncomp) + largest = np.where(keep, fcount, -1).argmax() + v0, v1, v2 = verts[faces[:, 0]], verts[faces[:, 1]], verts[faces[:, 2]] + cvol = np.bincount(flabel, weights=np.einsum("ij,ij->i", v0, np.cross(v1, v2)), minlength=ncomp) # 6*signed vol + cidx = np.arange(ncomp) # per-component vertex bbox via ndimage (~6x faster than ufunc.at) + cmin = np.stack([_ndi_minimum(verts[:, a], label, cidx) for a in range(3)], 1) + cmax = np.stack([_ndi_maximum(verts[:, a], label, cidx) for a in range(3)], 1) + tol = 1e-4 * (cmax[largest] - cmin[largest]).max() + enclosed = (cmin >= cmin[largest] - tol).all(1) & (cmax <= cmax[largest] + tol).all(1) + inner = enclosed & (np.sign(cvol) != np.sign(cvol[largest])) & (np.arange(ncomp) != largest) + keep &= ~inner + faces = faces[keep[flabel]] + if len(faces) == 0: + return verts[:0], faces + used = np.unique(faces) + remap = np.full(nv, -1, np.int64) + remap[used] = np.arange(len(used)) + return verts[used], remap[faces] + + +def _surface_nets(vol, level, voxel, origin, device): + # Vectorized Surface Nets: one dual vertex per sign-changing cell at its edge-crossing mean, quads wound CCW-outward. + # Returns verts (V,3), faces (F,3). + vol = vol.to(device=device, dtype=torch.float32) + dx, dy, dz = vol.shape + origin_t = torch.as_tensor(origin, device=device, dtype=torch.float32) + empty = (np.zeros((0, 3), np.float32), np.zeros((0, 3), np.int64)) + if dx < 2 or dy < 2 or dz < 2: + return empty + + # Active = cells whose 8 corners aren't all in/all out. + inside = vol >= level # (dx,dy,dz) bool + cs8 = [inside[ox:ox + dx - 1, oy:oy + dy - 1, oz:oz + dz - 1] + for ox, oy, oz in ((0, 0, 0), (1, 0, 0), (0, 1, 0), (1, 1, 0), + (0, 0, 1), (1, 0, 1), (0, 1, 1), (1, 1, 1))] + any_in = cs8[0] | cs8[1] | cs8[2] | cs8[3] | cs8[4] | cs8[5] | cs8[6] | cs8[7] + all_in = cs8[0] & cs8[1] & cs8[2] & cs8[3] & cs8[4] & cs8[5] & cs8[6] & cs8[7] + active = any_in & ~all_in # (cx,cy,cz) straddling cells + nv = int(active.sum()) + if nv == 0: + return empty + + # Active cells only (a thin shell): each dual vertex = mean of its 12 edges' zero-crossings. + del any_in, all_in, cs8 # corner bool grids no longer needed + ac = active.nonzero(as_tuple=False) # (nv,3) cell min-corner indices + offs = torch.tensor([[0, 0, 0], [1, 0, 0], [0, 1, 0], [1, 1, 0], + [0, 0, 1], [1, 0, 1], [0, 1, 1], [1, 1, 1]], device=device) + offf = offs.to(torch.float32) + edges = torch.tensor([[0, 1], [0, 2], [0, 4], [1, 3], [1, 5], [2, 3], + [2, 6], [3, 7], [4, 5], [4, 6], [5, 7], [6, 7]], device=device) + e0, e1 = edges[:, 0], edges[:, 1] + oe0, oe1 = offf[e0], offf[e1] # (12,3) edge endpoints + + cstep = 1 << 18 # chunk to bound peak memory (CPU RAM too) + loc = [] + for st in range(0, nv, cstep): + ci = ac[st:st + cstep, None, :] + offs[None] # (m,8,3) + cval = vol[ci[..., 0], ci[..., 1], ci[..., 2]] # (m,8) corner values + csl = cval >= level + v0, v1 = cval[:, e0], cval[:, e1] # (m,12) + cross = (csl[:, e0] != csl[:, e1])[..., None].to(torch.float32) + denom = v1 - v0 + t = torch.where(denom.abs() > 1e-12, (level - v0) / denom, torch.full_like(denom, 0.5)).clamp(0, 1) + pts = torch.lerp(oe0, oe1, t[..., None]) # (m,12,3) local crossings (fused interp) + loc.append((pts * cross).sum(1) / cross.sum(1).clamp_min(1.0)) # (m,3) in [0,1] + local = torch.cat(loc, 0) if len(loc) > 1 else loc[0] # (nv,3) + verts = origin_t + (ac.to(torch.float32) + local) * voxel # world space + del loc, local, ac + + vid = torch.full((dx - 1, dy - 1, dz - 1), -1, dtype=torch.int32, device=device) + vid[active] = torch.arange(nv, dtype=torch.int32, device=device) + del active + + # Each straddling grid edge -> one quad from its 4 cells; `sol` (low-end sign) picks outward winding. + faces = [] + + def emit(cr, sol, a, b, d, c): + valid = cr & (a >= 0) & (b >= 0) & (c >= 0) & (d >= 0) + if not bool(valid.any()): + return + a, b, c, d, sol = a[valid], b[valid], c[valid], d[valid], sol[valid] + p2, p4 = torch.where(sol, b, c), torch.where(sol, c, b) # reverse quad winding where ~sol + faces.append(torch.stack([a, p2, d], 1)) + faces.append(torch.stack([a, d, p4], 1)) + + a = inside[0:dx - 1, 1:dy - 1, 1:dz - 1] + emit(a != inside[1:dx, 1:dy - 1, 1:dz - 1], a, + vid[:, 0:dy - 2, 0:dz - 2], vid[:, 1:dy - 1, 0:dz - 2], + vid[:, 1:dy - 1, 1:dz - 1], vid[:, 0:dy - 2, 1:dz - 1]) + a = inside[1:dx - 1, 0:dy - 1, 1:dz - 1] + emit(a != inside[1:dx - 1, 1:dy, 1:dz - 1], a, + vid[0:dx - 2, :, 0:dz - 2], vid[0:dx - 2, :, 1:dz - 1], + vid[1:dx - 1, :, 1:dz - 1], vid[1:dx - 1, :, 0:dz - 2]) + a = inside[1:dx - 1, 1:dy - 1, 0:dz - 1] + emit(a != inside[1:dx - 1, 1:dy - 1, 1:dz], a, + vid[0:dx - 2, 0:dy - 2, :], vid[1:dx - 1, 0:dy - 2, :], + vid[1:dx - 1, 1:dy - 1, :], vid[0:dx - 2, 1:dy - 1, :]) + + if not faces: + return empty + return verts.cpu().numpy().astype(np.float32), torch.cat(faces, 0).cpu().numpy().astype(np.int64) + + +def _otsu_level(values, bins=256): + # Otsu threshold: the density value that best splits inside/outside (max between-class variance). + hist, edges = np.histogram(values, bins=bins) + hist = hist.astype(np.float64) + centers = (edges[:-1] + edges[1:]) * 0.5 + w = np.cumsum(hist) # background-class weight at each split + mu = np.cumsum(hist * centers) + wf = w[-1] - w # foreground-class weight + mb = mu / np.where(w > 0, w, 1.0) + mf = (mu[-1] - mu) / np.where(wf > 0, wf, 1.0) + var_b = w * wf * (mb - mf) ** 2 # between-class variance + var_b[(w <= 0) | (wf <= 0)] = -1.0 + return float(centers[int(np.argmax(var_b))]) + + +def _taubin_smooth(verts, faces, iters, lam=0.5, mu=-0.53): + # Taubin lambda|mu smoothing: low-pass the mesh surface without the shrinkage of a Laplacian blur + # (the mu inflation pass cancels the lambda pass's volume loss). Uniform (umbrella) weights. + if iters <= 0 or len(verts) == 0 or len(faces) == 0: + return verts + nv = len(verts) + e = np.concatenate([faces[:, [0, 1]], faces[:, [1, 2]], faces[:, [0, 2]]], 0) + e = np.concatenate([e, e[:, ::-1]], 0) # symmetric adjacency + adj = coo_matrix((np.ones(len(e), np.float32), (e[:, 0], e[:, 1])), shape=(nv, nv)).tocsr() + adj.data[:] = 1.0 + deg = np.clip(np.asarray(adj.sum(1)).ravel(), 1.0, None).astype(np.float32)[:, None] + v = verts.astype(np.float32) # fp32 matvec: ~2x faster, sub-micron drift on unit-scale verts + for _ in range(int(iters)): + for fac in (lam, mu): + v = v + np.float32(fac) * ((adj @ v) / deg - v) # fac * (mean(neighbours) - v) + return np.ascontiguousarray(v) + + +def _sample_vertex_colours_gpu(colvol, colnorm, verts, origin, voxel, device): + # GPU trilinear sampling of the colour numerator (3ch) and normaliser (1ch) at vertex grid-coords + # reproduces scipy map_coordinates(order=1, mode='nearest'). Returns col (V,3) numpy. + dx, dy, dz = colnorm.shape + vt = torch.as_tensor(verts, device=device, dtype=torch.float32) + org = torch.as_tensor(origin, device=device, dtype=torch.float32) + gi = (vt - org) / voxel # (V,3) grid-index coords (x,y,z) + size = torch.tensor([dx, dy, dz], device=device, dtype=torch.float32) + g = 2.0 * gi / (size - 1).clamp_min(1.0) - 1.0 # -> [-1,1] (align_corners) + grid = torch.stack([g[:, 2], g[:, 1], g[:, 0]], -1)[None, None, None] # (1,1,1,V,3): grid_sample order (W=z,H=y,D=x) + + def samp(v): # (dx,dy,dz,C) cpu fp16 -> (C,V) fp32 on device + inp = v.to(device).permute(3, 0, 1, 2)[None].float() + o = torch.nn.functional.grid_sample(inp, grid, mode="bilinear", padding_mode="border", align_corners=True) + return o[0, :, 0, 0, :] + num = samp(colvol) # (3,V) + den = samp(colnorm[..., None]) # (1,V) + return (num / den.clamp_min(1e-8)).T.cpu().numpy() # (V,3) + + +def _gaussian_to_mesh(g: Types.SPLAT, i, res, kernel, taubin, level_bias, min_component, min_opacity, color_sharpen, device, progress=None): + # Mesh one splat: density + colour grids -> Surface Nets -> floater removal -> Taubin smoothing -> + # volume-sampled colours. Returns (verts, faces int64, colors in [0,1]), or None if no surface. + rep = progress if progress is not None else (lambda *_: None) + + end = _real_len(g, i) + xyz = g.positions[i, :end].to(device=device, dtype=torch.float32) + scale = g.scales[i, :end].to(device=device, dtype=torch.float32) + quat = g.rotations[i, :end].to(device=device, dtype=torch.float32) + opacity = g.opacities[i, :end].reshape(-1).to(device=device, dtype=torch.float32) + rgb = (g.sh[i, :end, 0, :].to(device=device, dtype=torch.float32) * _C0 + 0.5).clamp(0, 1) + + keep = opacity >= min_opacity + xyz, scale, quat, opacity, rgb = xyz[keep], scale[keep], quat[keep], opacity[keep], rgb[keep] + if xyz.shape[0] == 0: + return None + + vol, colvol, colnorm, origin, voxel = _splat_density(xyz, opacity, scale, quat, rgb, res, kernel, device, + color_sharpen=color_sharpen, + progress=lambda f: rep(0.25 * f)) # density build: 0 -> 25% + # Colour: sample on the GPU (grid_sample) when there's headroom + colour_gpu = not comfy.model_management.is_device_cpu(device) and comfy.model_management.get_free_memory(device) > 6 * vol.numel() * 4 + if colour_gpu: + colvol_cpu, colnorm_cpu = colvol.cpu(), colnorm.half().cpu() # park colours (fp16) off-GPU during meshing + colvol_np = colnorm_np = None + else: + colvol_np = colvol.cpu().numpy().astype(np.float32) # Sum(w^p * rgb) colour numerator (fp16 grid -> fp32) + colnorm_np = colnorm.cpu().numpy().astype(np.float32) # Sum(w^p) colour normaliser + del colvol, colnorm # free the colour grids before iso-surfacing + rep(0.40) + + vmin, vmax = float(vol.min()), float(vol.max()) + occ = vol[vol > vmax * 1e-3] # occupied voxels (skip the empty-space peak) + if occ.numel() == 0: + return None + # Otsu picks the inside/outside split principledly; `level_bias` nudges it (1.0 = auto). Clamp strictly + # inside the data range so a bias can't push the iso off the histogram. + level = min(max(_otsu_level(occ.cpu().numpy()) * level_bias, vmin + 1e-6 * (vmax - vmin)), + vmax - 1e-6 * (vmax - vmin)) + + # Iso-surface on the accelerator when there's headroom: ~15x faster than CPU, identical output. Chunked + # Surface Nets peaks at ~3-3.5x the density grid, so fall back to CPU for large grids / tight VRAM. + sn_dev = device + if not comfy.model_management.is_device_cpu(device) and comfy.model_management.get_free_memory(device) < 6 * vol.numel() * 4: + sn_dev = torch.device("cpu") + vol = vol.cpu() + verts, faces = _surface_nets(vol, level, voxel, origin, sn_dev) + del vol + rep(0.55) + if min_component > 0 and len(faces) > 0: + verts, faces = _clean_components(verts, faces, min_component, device) + if len(verts) == 0 or len(faces) == 0: + return None + + # Taubin smooths the blocky iso without shrinking it (unlike blurring the density, which rounds features). + verts = _taubin_smooth(verts, faces, taubin) + rep(0.7) + + # Colour each vertex from the co-splatted colour volume: trilinearly sample the numerator Sum(w^p*rgb) + # and normaliser Sum(w^p) separately, then divide. Normalising AFTER interpolation keeps zero-density + # edge voxels from pulling colours toward black, and matches the gaussians that formed the surface. + if colour_gpu: + col = _sample_vertex_colours_gpu(colvol_cpu, colnorm_cpu, verts, origin, voxel, device) + else: + coords = ((verts - origin) / voxel).T # (3, V) grid-index coords, matching volume axes + num = np.stack([map_coordinates(colvol_np[..., c], coords, order=1, mode="nearest") for c in range(3)], -1) + den = map_coordinates(colnorm_np, coords, order=1, mode="nearest") + col = num / np.clip(den, 1e-8, None)[:, None] + rep(1.0) + + # The unlit material's COLOR_0 is linear and the viewer sRGB-encodes it on output; the splat colours + # are display (sRGB) values, so convert sRGB -> linear here to land at the same brightness as the splat. + col = np.clip(col, 0, 1) + col = np.where(col <= 0.04045, col / 12.92, ((col + 0.055) / 1.055) ** 2.4).astype(np.float32) + + # Splat +Y is glTF's -Y: rotate 180 deg about X (negate Y,Z) to land upright. Proper rotation, so + # winding is kept; done after colouring (which works in the splat frame). + verts = np.ascontiguousarray(verts * np.array([1.0, -1.0, -1.0], dtype=np.float32)) + return (torch.from_numpy(verts), torch.from_numpy(faces), torch.from_numpy(col)) + + +class SplatToMesh(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="SplatToMesh", + display_name="Extract Mesh from Splat", + search_aliases=["splat to mesh", "gaussian surface nets", "splat surface", "mesh splat"], + category="3d/splat", + description="Extract a coloured mesh from a gaussian splat.", + inputs=[ + IO.Splat.Input("splat"), + IO.Int.Input("resolution", default=384, min=64, max=768, step=16, + tooltip="Density-grid resolution along the longest axis. Higher = finer surface, " + "more VRAM/time (grows with resolution^3)."), + IO.Int.Input("kernel", default=5, min=1, max=8, + tooltip="Max splat half-width in voxels. Each gaussian is rasterized over a window " + "sized to its own 3-sigma, capped here - small surfels stay cheap, large ones " + "aren't truncated. Raise if sparse splats leave gaps."), + IO.Int.Input("smooth", default=0, min=0, max=60, advanced = True, + tooltip="Taubin mesh-smoothing iterations. Smooths the surface without shrinking it " + "(volume-preserving), unlike blurring the density. 0 = raw surface."), + IO.Float.Input("level", default=0.4, min=0.0, max=2.0, step=0.01, + tooltip="Iso-surface level. Auto-picked by Otsu; this biases it (1.0 = auto, lower = " + "fatter/more-connected surface, higher = thinner/tighter)."), + IO.Int.Input("min_component", default=500, min=0, max=100000, step=50, advanced=True, + tooltip="Drop connected components smaller than this many vertices (0 = keep all). " + "Removes detached floater blobs and the inner shell of the double wall."), + IO.Float.Input("min_opacity", default=0.02, min=0.0, max=1.0, step=0.01, advanced=True, + tooltip="Ignore gaussians fainter than this before meshing."), + IO.Float.Input("color_sharpen", default=2.0, min=1.0, max=8.0, step=0.5, + tooltip="Crisp up the vertex texture: 1.0 = physically-correct blend; higher biases " + "each voxel's colour toward its dominant gaussian instead of averaging " + "neighbours (de-smears the texture). Colour only - geometry is unchanged."), + ], + outputs=[IO.Mesh.Output(display_name="mesh")], + ) + + @classmethod + def execute(cls, splat, resolution, kernel, smooth, level, min_component, min_opacity, color_sharpen) -> IO.NodeOutput: + device = comfy.model_management.get_torch_device() + b = splat.positions.shape[0] + prec = 1000 # each splat owns a 0..prec block of the bar; its callback advances within that block + pbar = comfy.utils.ProgressBar(b * prec) + + verts_l, faces_l, colors_l = [], [], [] + for i in range(b): + cb = lambda f, base=i * prec: pbar.update_absolute(base + int(min(max(f, 0.0), 1.0) * prec)) + res = _gaussian_to_mesh(splat, i, resolution, kernel, smooth, level, min_component, min_opacity, color_sharpen, device, cb) + if res is None: + logging.warning("SplatToMesh: splat %d produced no surface; emitting an empty mesh.", i) + v, f, c = torch.zeros((0, 3)), torch.zeros((0, 3), dtype=torch.int64), torch.zeros((0, 3)) + else: + v, f, c = res + verts_l.append(v) + faces_l.append(f) + colors_l.append(c) + pbar.update_absolute((i + 1) * prec) # snap to block end (covers empty / early-out splats) + # unlit: render flat (emissive-like) so SaveGLB matches the splat instead of lighting/washing it. + return IO.NodeOutput(pack_variable_mesh_batch(verts_l, faces_l, colors=colors_l, unlit=True)) + + +class GaussianExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [SplatToFile3D, File3DToSplat, RenderSplat, CreateCameraInfo, TransformSplat, + GetSplatCount, MergeSplat, SplatToMesh] + + +async def comfy_entrypoint() -> GaussianExtension: + return GaussianExtension() diff --git a/comfy_extras/nodes_gits.py b/comfy_extras/nodes_gits.py index 0b7666524..434a24387 100644 --- a/comfy_extras/nodes_gits.py +++ b/comfy_extras/nodes_gits.py @@ -340,7 +340,7 @@ class GITSScheduler(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="GITSScheduler", - category="sampling/schedulers", + category="model/sampling/schedulers", inputs=[ io.Float.Input("coeff", default=1.20, min=0.80, max=1.50, step=0.05, advanced=True), io.Int.Input("steps", default=10, min=2, max=1000), diff --git a/comfy_extras/nodes_hidream_o1.py b/comfy_extras/nodes_hidream_o1.py index f393745f6..8648d2e26 100644 --- a/comfy_extras/nodes_hidream_o1.py +++ b/comfy_extras/nodes_hidream_o1.py @@ -14,7 +14,7 @@ class EmptyHiDreamO1LatentImage(io.ComfyNode): return io.Schema( node_id="EmptyHiDreamO1LatentImage", display_name="Empty HiDream-O1 Latent Image", - category="latent/image", + category="model/latent/image", description=( "Empty pixel-space latent for HiDream-O1-Image. The model was " "trained at ~4 megapixels; lower resolutions go off-distribution " @@ -47,7 +47,7 @@ class HiDreamO1ReferenceImages(io.ComfyNode): return io.Schema( node_id="HiDreamO1ReferenceImages", display_name="HiDream-O1 Reference Images", - category="conditioning/image", + category="model/conditioning/image", description=( "Attach 1-10 reference images to conditioning, one for edit instruction" "or multiple for subject-driven personalization." diff --git a/comfy_extras/nodes_hunyuan.py b/comfy_extras/nodes_hunyuan.py index 9e4873be5..16fff12af 100644 --- a/comfy_extras/nodes_hunyuan.py +++ b/comfy_extras/nodes_hunyuan.py @@ -41,7 +41,7 @@ class EmptyHunyuanLatentVideo(io.ComfyNode): return io.Schema( node_id="EmptyHunyuanLatentVideo", display_name="Empty HunyuanVideo 1.0 Latent", - category="latent/video", + category="model/latent/video", inputs=[ io.Int.Input("width", default=848, min=16, max=nodes.MAX_RESOLUTION, step=16), io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16), @@ -81,7 +81,7 @@ class HunyuanVideo15ImageToVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="HunyuanVideo15ImageToVideo", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), @@ -132,7 +132,7 @@ class HunyuanVideo15SuperResolution(io.ComfyNode): return io.Schema( node_id="HunyuanVideo15SuperResolution", display_name="Hunyuan Video 1.5 Super Resolution", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), @@ -178,7 +178,7 @@ class LatentUpscaleModelLoader(io.ComfyNode): return io.Schema( node_id="LatentUpscaleModelLoader", display_name="Load Latent Upscale Model", - category="loaders", + category="model/loaders", inputs=[ io.Combo.Input("model_name", options=folder_paths.get_filename_list("latent_upscale_models")), ], @@ -227,7 +227,7 @@ class HunyuanVideo15LatentUpscaleWithModel(io.ComfyNode): return io.Schema( node_id="HunyuanVideo15LatentUpscaleWithModel", display_name="Hunyuan Video 15 Latent Upscale With Model", - category="latent", + category="model/latent", inputs=[ io.LatentUpscaleModel.Input("model"), io.Latent.Input("samples"), @@ -308,7 +308,7 @@ class HunyuanImageToVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="HunyuanImageToVideo", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Vae.Input("vae"), @@ -359,7 +359,7 @@ class EmptyHunyuanImageLatent(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="EmptyHunyuanImageLatent", - category="latent", + category="model/latent", inputs=[ io.Int.Input("width", default=2048, min=64, max=nodes.MAX_RESOLUTION, step=32), io.Int.Input("height", default=2048, min=64, max=nodes.MAX_RESOLUTION, step=32), @@ -384,7 +384,7 @@ class HunyuanRefinerLatent(io.ComfyNode): return io.Schema( node_id="HunyuanRefinerLatent", display_name="Hunyuan Latent Refiner", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), diff --git a/comfy_extras/nodes_hunyuan3d.py b/comfy_extras/nodes_hunyuan3d.py index bcd3f9198..60e530626 100644 --- a/comfy_extras/nodes_hunyuan3d.py +++ b/comfy_extras/nodes_hunyuan3d.py @@ -12,7 +12,7 @@ class EmptyLatentHunyuan3Dv2(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="EmptyLatentHunyuan3Dv2", - category="latent/3d", + category="model/latent/3d", inputs=[ IO.Int.Input("resolution", default=3072, min=1, max=8192), IO.Int.Input("batch_size", default=1, min=1, max=4096, tooltip="The number of latent images in the batch."), @@ -35,7 +35,7 @@ class Hunyuan3Dv2Conditioning(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="Hunyuan3Dv2Conditioning", - category="conditioning/3d_models", + category="model/conditioning/3d_models", inputs=[ IO.ClipVisionOutput.Input("clip_vision_output"), ], @@ -60,7 +60,7 @@ class Hunyuan3Dv2ConditioningMultiView(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="Hunyuan3Dv2ConditioningMultiView", - category="conditioning/3d_models", + category="model/conditioning/3d_models", inputs=[ IO.ClipVisionOutput.Input("front", optional=True), IO.ClipVisionOutput.Input("left", optional=True), @@ -97,7 +97,7 @@ class VAEDecodeHunyuan3D(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="VAEDecodeHunyuan3D", - category="latent/3d", + category="model/latent/3d", inputs=[ IO.Latent.Input("samples"), IO.Vae.Input("vae"), diff --git a/comfy_extras/nodes_hypernetwork.py b/comfy_extras/nodes_hypernetwork.py index 44a9c6f97..2d3f1bd05 100644 --- a/comfy_extras/nodes_hypernetwork.py +++ b/comfy_extras/nodes_hypernetwork.py @@ -103,7 +103,7 @@ class HypernetworkLoader(IO.ComfyNode): return IO.Schema( node_id="HypernetworkLoader", display_name="Load Hypernetwork", - category="loaders", + category="model/loaders", inputs=[ IO.Model.Input("model"), IO.Combo.Input("hypernetwork_name", options=folder_paths.get_filename_list("hypernetworks")), diff --git a/comfy_extras/nodes_hypertile.py b/comfy_extras/nodes_hypertile.py index 354d96db1..2a96416be 100644 --- a/comfy_extras/nodes_hypertile.py +++ b/comfy_extras/nodes_hypertile.py @@ -27,7 +27,7 @@ class HyperTile(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="HyperTile", - category="model_patches/unet", + category="model/patch/unet", inputs=[ io.Model.Input("model"), io.Int.Input("tile_size", default=256, min=1, max=2048, advanced=True), diff --git a/comfy_extras/nodes_ideogram4.py b/comfy_extras/nodes_ideogram4.py new file mode 100644 index 000000000..d5827db4f --- /dev/null +++ b/comfy_extras/nodes_ideogram4.py @@ -0,0 +1,64 @@ +"""Ideogram 4 sampling helper +""" + +import math + +import torch +from typing_extensions import override +from comfy_api.latest import ComfyExtension, io + +_LOGSNR_MIN = -15.0 +_LOGSNR_MAX = 18.0 + + +def _logit_normal_schedule(u, mean, std): + # Reference time (0=noise..1=clean) via the probit/ndtri quantile. + u = torch.as_tensor(u, dtype=torch.float64) + t = 1.0 - torch.special.expit(mean + std * torch.special.ndtri(u)) + t_min = 1.0 / (1.0 + math.exp(0.5 * _LOGSNR_MAX)) + t_max = 1.0 / (1.0 + math.exp(0.5 * _LOGSNR_MIN)) + return t.clamp(t_min, t_max) + + +def ideogram4_sigmas(num_steps, width, height, mu, std): + """Descending sigmas (len num_steps+1) for the reference schedule. + + mu + the resolution term form the logSNR shift; std is the spread. + """ + mean = mu + 0.5 * math.log((width * height) / (512 * 512)) + u = torch.linspace(0.0, 1.0, num_steps + 1, dtype=torch.float64) + sigmas = (1.0 - _logit_normal_schedule(u, mean, std)).flip(0) + sigmas[-1] = 0.0 # clamp leaves ~6e-4; force full denoise + return sigmas.to(torch.float32) + + +class Ideogram4Scheduler(io.ComfyNode): + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="Ideogram4Scheduler", + display_name="Ideogram 4 Scheduler", + category="sampling/custom_sampling/schedulers", + inputs=[ + io.Int.Input("steps", default=20, min=1, max=200), + io.Int.Input("width", default=1024, min=256, max=8192, step=16), + io.Int.Input("height", default=1024, min=256, max=8192, step=16), + io.Float.Input("mu", default=0.0, min=-10.0, max=10.0, step=0.05), + io.Float.Input("std", default=1.75, min=0.1, max=5.0, step=0.05), + ], + outputs=[io.Sigmas.Output()], + ) + + @classmethod + def execute(cls, steps, width, height, mu, std) -> io.NodeOutput: + return io.NodeOutput(ideogram4_sigmas(steps, width, height, mu, std)) + + +class Ideogram4Extension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [Ideogram4Scheduler] + + +async def comfy_entrypoint() -> Ideogram4Extension: + return Ideogram4Extension() diff --git a/comfy_extras/nodes_images.py b/comfy_extras/nodes_images.py index 4856346d7..469a7be55 100644 --- a/comfy_extras/nodes_images.py +++ b/comfy_extras/nodes_images.py @@ -1,17 +1,23 @@ -from __future__ import annotations - import nodes import folder_paths +import av import json + import os import re import math +import numpy as np +import struct import torch + +import zlib import comfy.utils +from fractions import Fraction from server import PromptServer from comfy_api.latest import ComfyExtension, IO, UI +from comfy.cli_args import args from typing_extensions import override SVG = IO.SVG.Type # TODO: temporary solution for backward compatibility, will be removed later. @@ -89,7 +95,7 @@ class BoundingBox(IO.ComfyNode): return IO.Schema( node_id="PrimitiveBoundingBox", display_name="Bounding Box", - category="utils/primitive", + category="utilities/primitive", inputs=[ IO.Int.Input("x", default=0, min=0, max=MAX_RESOLUTION), IO.Int.Input("y", default=0, min=0, max=MAX_RESOLUTION), @@ -835,6 +841,405 @@ class ImageMergeTileList(IO.ComfyNode): return IO.NodeOutput(merged_image) +# --------------------------------------------------------------------------- +# Format specifications +# --------------------------------------------------------------------------- + +# Maps (file_format, bit_depth, has_alpha) -> (numpy dtype scale, av pixel format, +# stream pix_fmt). Keeps the encode path declarative instead of branchy. +_FORMAT_SPECS = { + ("png", "8-bit", False): {"scale": 255.0, "dtype": np.uint8, "frame_fmt": "rgb24", "stream_fmt": "rgb24"}, + ("png", "8-bit", True): {"scale": 255.0, "dtype": np.uint8, "frame_fmt": "rgba", "stream_fmt": "rgba"}, + ("png", "16-bit", False): {"scale": 65535.0, "dtype": np.uint16, "frame_fmt": "rgb48le", "stream_fmt": "rgb48be"}, + ("png", "16-bit", True): {"scale": 65535.0, "dtype": np.uint16, "frame_fmt": "rgba64le", "stream_fmt": "rgba64be"}, + ("exr", "32-bit float", False): {"scale": 1.0, "dtype": np.float32, "frame_fmt": "gbrpf32le", "stream_fmt": "gbrpf32le"}, + ("exr", "32-bit float", True): {"scale": 1.0, "dtype": np.float32, "frame_fmt": "gbrapf32le", "stream_fmt": "gbrapf32le"}, +} + + +# --------------------------------------------------------------------------- +# Color transforms +# --------------------------------------------------------------------------- + +def srgb_to_linear(t: torch.Tensor) -> torch.Tensor: + """Inverse sRGB EOTF (IEC 61966-2-1). Operates on RGB channels only; + alpha (if present as the 4th channel) is passed through unchanged.""" + if t.shape[-1] == 4: + rgb, alpha = t[..., :3], t[..., 3:] + return torch.cat([srgb_to_linear(rgb), alpha], dim=-1) + + # Piecewise: linear toe below 0.04045, gamma curve above. + low = t / 12.92 + high = ((t.clamp(min=0.0) + 0.055) / 1.055) ** 2.4 + return torch.where(t <= 0.04045, low, high) + + +# HLG OETF constants from BT.2100 Table 5. +_HLG_A = 0.17883277 +_HLG_B = 0.28466892 +_HLG_C = 0.55991072928 # = 0.5 - a*ln(4*a) + + +def hlg_to_linear(t: torch.Tensor) -> torch.Tensor: + """Inverse HLG OETF (BT.2100). Maps a non-linear HLG signal in [0, 1] to + *scene*-linear light in [0, 1]. Per BT.2100 Note 5a, this is the correct + transform when converting HLG to a linear scene-light representation + (rather than display-light, which would also involve the HLG OOTF). + + Operates on RGB channels only; alpha is passed through unchanged.""" + if t.shape[-1] == 4: + rgb, alpha = t[..., :3], t[..., 3:] + return torch.cat([hlg_to_linear(rgb), alpha], dim=-1) + + # Piecewise: sqrt branch below 0.5, log branch above. + # Clamp inside the log branch so negative / out-of-range values don't blow up; + # values above 1.0 are allowed and extrapolate naturally. + low = (t ** 2) / 3.0 + high = (torch.exp((t.clamp(min=_HLG_C) - _HLG_C) / _HLG_A) + _HLG_B) / 12.0 + return torch.where(t <= 0.5, low, high) + + +# --------------------------------------------------------------------------- +# Metadata injection +# --------------------------------------------------------------------------- + +_PNG_SIGNATURE = b"\x89PNG\r\n\x1a\n" + + +def _png_chunk(chunk_type: bytes, data: bytes) -> bytes: + """Build a single PNG chunk: length | type | data | CRC32(type+data).""" + crc = zlib.crc32(chunk_type + data) & 0xFFFFFFFF + return struct.pack(">I", len(data)) + chunk_type + data + struct.pack(">I", crc) + + +def _png_text_chunk(keyword: str, text: str) -> bytes: + """tEXt chunk: latin-1 keyword + NUL + latin-1 text.""" + payload = keyword.encode("latin-1") + b"\x00" + text.encode("latin-1", errors="replace") + return _png_chunk(b"tEXt", payload) + + +def inject_png_metadata(png_bytes: bytes, prompt: dict | None, extra_pnginfo: dict | None) -> bytes: + """Insert ComfyUI prompt/workflow as tEXt chunks right after IHDR.""" + if not png_bytes.startswith(_PNG_SIGNATURE): + return png_bytes + + chunks: list[bytes] = [] + if prompt is not None: + chunks.append(_png_text_chunk("prompt", json.dumps(prompt))) + if extra_pnginfo: + for key, value in extra_pnginfo.items(): + chunks.append(_png_text_chunk(key, json.dumps(value))) + if not chunks: + return png_bytes + + # IHDR is always the first chunk; insert ours immediately after it. + ihdr_length = struct.unpack(">I", png_bytes[8:12])[0] + ihdr_end = 8 + 8 + ihdr_length + 4 # signature + (len+type) + data + crc + return png_bytes[:ihdr_end] + b"".join(chunks) + png_bytes[ihdr_end:] + + +# Standard chromaticities (CIE 1931 xy) for the colorspaces this node writes. +# Each tuple is (Rx, Ry, Gx, Gy, Bx, By, Wx, Wy). All share D65 white point. +_CHROMATICITIES = { + # ITU-R BT.709 / sRGB primaries + "Rec.709": (0.6400, 0.3300, 0.3000, 0.6000, 0.1500, 0.0600, 0.3127, 0.3290), + # ITU-R BT.2020 (UHDTV / wide-gamut HDR) primaries + "Rec.2020": (0.7080, 0.2920, 0.1700, 0.7970, 0.1310, 0.0460, 0.3127, 0.3290), +} + + +def _pack_chromaticities(primaries: tuple) -> bytes: + """Serialize 8 chromaticity floats into the EXR `chromaticities` payload.""" + return struct.pack("<8f", *primaries) + + +def _exr_attribute(name: str, attr_type: str, value: bytes) -> bytes: + """Serialize one EXR header attribute: name\\0 type\\0 size:int32 value.""" + return ( + name.encode("utf-8") + b"\x00" + + attr_type.encode("utf-8") + b"\x00" + + struct.pack(" bytes: + """Insert ComfyUI metadata and color-space info into an EXR header. + + Color: EXR pixels are linear by convention. The standard way to describe + their RGB→XYZ relationship is the `chromaticities` attribute. We pick the + primaries that match what the user told us their input was: + + colorspace="sRGB" → Rec. 709 / sRGB primaries (D65) + colorspace="HDR" → Rec. 2020 / BT.2100 primaries (D65) + + Pixels are always converted to linear scene light upstream (sRGB EOTF + inverse for sRGB; HLG OETF inverse for HDR), so the file content is + scene-linear in the indicated gamut. OpenEXR has no standard transfer- + function attribute (the OpenEXR TSC has discussed adding one but it + doesn't exist), so we don't invent one — `chromaticities` plus the EXR + linear-by-convention rule fully specifies the color. + + Prompt/workflow: written as plain `string` attributes using the same keys + (`prompt`, `workflow`, ...) that Comfy uses for PNG tEXt chunks, so the + same readers can pull them out symmetrically. + + Implementation note: the chunk-offset table that follows the header stores + *absolute* byte offsets into the file. Inserting N bytes into the header + means every offset must be incremented by N or the file becomes unreadable. + """ + if len(exr_bytes) < 8 or exr_bytes[:4] != b"\x76\x2f\x31\x01": + return exr_bytes + + new_blob = b"" + if prompt is not None: + new_blob += _exr_attribute("prompt", "string", json.dumps(prompt).encode("utf-8")) + if extra_pnginfo: + for key, value in extra_pnginfo.items(): + new_blob += _exr_attribute(key, "string", json.dumps(value).encode("utf-8")) + if colorspace is not None: + # Map each colorspace option to the RGB primaries the linear pixels + # are now in. "sRGB" and "linear" both produce Rec. 709 linear; "HDR" + # (HLG-encoded Rec. 2020 input) produces Rec. 2020 linear. + primaries_name = { + "sRGB": "Rec.709", + "linear": "Rec.709", + "HDR": "Rec.2020", + }.get(colorspace, "Rec.709") + new_blob += _exr_attribute( + "chromaticities", + "chromaticities", + _pack_chromaticities(_CHROMATICITIES[primaries_name]), + ) + if not new_blob: + return exr_bytes + + # Walk header attributes to find the terminating null byte, and pick up + # dataWindow + compression so we know how many chunks the offset table has. + pos = 8 # past magic (4) + version (4) + data_window = None + compression = 0 + while pos < len(exr_bytes) and exr_bytes[pos] != 0: + name_end = exr_bytes.index(b"\x00", pos) + attr_name = exr_bytes[pos:name_end].decode("latin-1", errors="replace") + type_end = exr_bytes.index(b"\x00", name_end + 1) + attr_type = exr_bytes[name_end + 1:type_end].decode("latin-1", errors="replace") + size = struct.unpack(" bytes: + """Encode a single HxWxC tensor to PNG or EXR bytes in memory. + + For EXR the input is interpreted according to `colorspace` and converted + to scene-linear (EXR's convention) before writing: + + "sRGB" → input is sRGB-encoded Rec. 709; apply inverse sRGB EOTF. + "HDR" → input is HLG-encoded Rec. 2020 (BT.2100); apply inverse HLG + OETF to get scene-linear, per BT.2100 Note 5a. + "linear" → input is already scene-linear (Rec. 709 primaries); write + through unchanged. Use this for renderer/compositor output. + + For PNG, colorspace selection does not modify pixels — PNG is delivered + sRGB-encoded and there is no PNG path for wide-gamut HDR in this node. + """ + height, width, num_channels = img_tensor.shape + has_alpha = num_channels == 4 + + spec = _FORMAT_SPECS[(file_format, bit_depth, has_alpha)] + + if spec["dtype"] == np.float32: + # EXR path: preserve full range, no clamp. + if colorspace == "sRGB": + img_tensor = srgb_to_linear(img_tensor) + elif colorspace == "HDR": + img_tensor = hlg_to_linear(img_tensor) + img_np = img_tensor.cpu().numpy().astype(np.float32) + else: + # PNG path: quantize to integer range. + scaled = (img_tensor * spec["scale"]).clamp(0, spec["scale"]) + img_np = scaled.to(torch.int32).cpu().numpy().astype(spec["dtype"]) + + # Encode directly via CodecContext. PyAV's `image2` muxer does NOT write to + # BytesIO (it expects a real file path), so we bypass the container entirely. + # For single-frame PNG/EXR the raw codec output IS the file. + codec = av.CodecContext.create(file_format, "w") + codec.width = width + codec.height = height + codec.pix_fmt = spec["stream_fmt"] + codec.time_base = Fraction(1, 1) + + frame = av.VideoFrame.from_ndarray(img_np, format=spec["frame_fmt"]) + if spec["frame_fmt"] != spec["stream_fmt"]: + frame = frame.reformat(format=spec["stream_fmt"]) + frame.pts = 0 + frame.time_base = codec.time_base + + packets = list(codec.encode(frame)) + list(codec.encode(None)) # flush with None + return b"".join(bytes(p) for p in packets) + + +# --------------------------------------------------------------------------- +# Node +# --------------------------------------------------------------------------- + +class SaveImageAdvanced(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="SaveImageAdvanced", + search_aliases=["save", "save image", "export image", "output image", "write image"], + display_name="Save Image (Advanced)", + description="Saves the input images to your ComfyUI output directory.", + category="image", + essentials_category="Basics", + inputs=[ + IO.Image.Input("images", tooltip="The images to save."), + IO.String.Input( + "filename_prefix", + default="ComfyUI", + tooltip=( + "The prefix for the file to save. May include formatting tokens " + "such as %date:yyyy-MM-dd% or %Empty Latent Image.width%." + ), + ), + IO.DynamicCombo.Input( + "format", + options=[ + IO.DynamicCombo.Option("png", [ + IO.Combo.Input("bit_depth", options=["8-bit", "16-bit"], + default="8-bit", advanced=True), + IO.Combo.Input("input_color_space", options=["sRGB"], + default="sRGB", advanced=True), + ]), + IO.DynamicCombo.Option("exr", [ + IO.Combo.Input("bit_depth", options=["32-bit float"], + default="32-bit float", advanced=True), + IO.Combo.Input( + "input_color_space", + options=["sRGB", "HDR", "linear"], + default="sRGB", + advanced=True, + tooltip=( + "Colorspace of the input tensor. The EXR is " + "always written as scene-linear in the matching " + "gamut.\n" + " 'sRGB' — input is sRGB-encoded Rec.709; " + "the inverse sRGB EOTF is applied.\n" + " 'HDR' — input is HLG-encoded Rec.2020 " + "(BT.2100); the inverse HLG OETF is applied " + "to get scene-linear light.\n" + " 'linear' — input is already scene-linear " + "(Rec.709 primaries); written through unchanged. " + "Use this for renderer/compositor output." + ), + ), + ]), + ], + tooltip="The file format in which to save the image.", + ), + ], + hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo], + is_output_node=True, + ) + + @classmethod + def execute(cls, images, filename_prefix: str, format: dict) -> IO.NodeOutput: + file_format = format["format"] + bit_depth = format["bit_depth"] + colorspace = format.get("input_color_space", "sRGB") + + output_dir = folder_paths.get_output_directory() + full_output_folder, filename, counter, subfolder, filename_prefix = ( + folder_paths.get_save_image_path( + filename_prefix, output_dir, images[0].shape[1], images[0].shape[0] + ) + ) + + prompt = cls.hidden.prompt + extra_pnginfo = cls.hidden.extra_pnginfo + write_metadata = not args.disable_metadata + + results = [] + for batch_number, image in enumerate(images): + encoded = _encode_image(image, file_format, bit_depth, colorspace) + + if write_metadata: + if file_format == "png": + encoded = inject_png_metadata(encoded, prompt, extra_pnginfo) + elif file_format == "exr": + encoded = inject_exr_metadata(encoded, prompt, extra_pnginfo, colorspace) + + name = filename.replace("%batch_num%", str(batch_number)) + file = f"{name}_{counter:05}.{file_format}" + with open(os.path.join(full_output_folder, file), "wb") as f: + f.write(encoded) + + results.append({"filename": file, "subfolder": subfolder, "type": "output"}) + counter += 1 + + return IO.NodeOutput(ui={"images": results}) + + class ImagesExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: @@ -847,6 +1252,7 @@ class ImagesExtension(ComfyExtension): ImageAddNoise, SaveAnimatedWEBP, SaveAnimatedPNG, + SaveImageAdvanced, SaveSVGNode, ImageStitch, ResizeAndPadImage, diff --git a/comfy_extras/nodes_ip2p.py b/comfy_extras/nodes_ip2p.py index 78f29915d..9c80834f0 100644 --- a/comfy_extras/nodes_ip2p.py +++ b/comfy_extras/nodes_ip2p.py @@ -9,7 +9,7 @@ class InstructPixToPixConditioning(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="InstructPixToPixConditioning", - category="conditioning/instructpix2pix", + category="model/conditioning/instructpix2pix", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), diff --git a/comfy_extras/nodes_kandinsky5.py b/comfy_extras/nodes_kandinsky5.py index 346c50cde..015965498 100644 --- a/comfy_extras/nodes_kandinsky5.py +++ b/comfy_extras/nodes_kandinsky5.py @@ -13,7 +13,7 @@ class Kandinsky5ImageToVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="Kandinsky5ImageToVideo", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), @@ -71,7 +71,7 @@ class NormalizeVideoLatentStart(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="NormalizeVideoLatentStart", - category="conditioning/video_models", + category="model/conditioning/video_models", description="Normalizes the initial frames of a video latent to match the mean and standard deviation of subsequent reference frames. Helps reduce differences between the starting frames and the rest of the video.", inputs=[ io.Latent.Input("latent"), diff --git a/comfy_extras/nodes_latent.py b/comfy_extras/nodes_latent.py index 8bb368dec..32da9e8ac 100644 --- a/comfy_extras/nodes_latent.py +++ b/comfy_extras/nodes_latent.py @@ -22,7 +22,7 @@ class LatentAdd(io.ComfyNode): return io.Schema( node_id="LatentAdd", search_aliases=["combine latents", "sum latents"], - category="latent/advanced", + category="model/latent/advanced", inputs=[ io.Latent.Input("samples1"), io.Latent.Input("samples2"), @@ -49,7 +49,7 @@ class LatentSubtract(io.ComfyNode): return io.Schema( node_id="LatentSubtract", search_aliases=["difference latent", "remove features"], - category="latent/advanced", + category="model/latent/advanced", inputs=[ io.Latent.Input("samples1"), io.Latent.Input("samples2"), @@ -76,7 +76,7 @@ class LatentMultiply(io.ComfyNode): return io.Schema( node_id="LatentMultiply", search_aliases=["scale latent", "amplify latent", "latent gain"], - category="latent/advanced", + category="model/latent/advanced", inputs=[ io.Latent.Input("samples"), io.Float.Input("multiplier", default=1.0, min=-10.0, max=10.0, step=0.01), @@ -100,7 +100,7 @@ class LatentInterpolate(io.ComfyNode): return io.Schema( node_id="LatentInterpolate", search_aliases=["blend latent", "mix latent", "lerp latent", "transition"], - category="latent/advanced", + category="model/latent/advanced", inputs=[ io.Latent.Input("samples1"), io.Latent.Input("samples2"), @@ -139,7 +139,7 @@ class LatentConcat(io.ComfyNode): return io.Schema( node_id="LatentConcat", search_aliases=["join latents", "stitch latents"], - category="latent/advanced", + category="model/latent/advanced", inputs=[ io.Latent.Input("samples1"), io.Latent.Input("samples2"), @@ -179,7 +179,7 @@ class LatentCut(io.ComfyNode): return io.Schema( node_id="LatentCut", search_aliases=["crop latent", "slice latent", "extract region"], - category="latent/advanced", + category="model/latent/advanced", inputs=[ io.Latent.Input("samples"), io.Combo.Input("dim", options=["x", "y", "t"]), @@ -220,7 +220,7 @@ class LatentCutToBatch(io.ComfyNode): return io.Schema( node_id="LatentCutToBatch", search_aliases=["slice to batch", "split latent", "tile latent"], - category="latent/advanced", + category="model/latent/advanced", inputs=[ io.Latent.Input("samples"), io.Combo.Input("dim", options=["t", "x", "y"]), @@ -262,7 +262,7 @@ class LatentBatch(io.ComfyNode): return io.Schema( node_id="LatentBatch", search_aliases=["combine latents", "merge latents", "join latents"], - category="latent/batch", + category="model/latent/batch", is_deprecated=True, inputs=[ io.Latent.Input("samples1"), @@ -290,7 +290,7 @@ class LatentBatchSeedBehavior(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LatentBatchSeedBehavior", - category="latent/advanced", + category="model/latent/advanced", inputs=[ io.Latent.Input("samples"), io.Combo.Input("seed_behavior", options=["random", "fixed"], default="fixed"), @@ -319,7 +319,7 @@ class LatentApplyOperation(io.ComfyNode): return io.Schema( node_id="LatentApplyOperation", search_aliases=["transform latent"], - category="latent/advanced/operations", + category="model/latent/advanced/operations", is_experimental=True, inputs=[ io.Latent.Input("samples"), @@ -343,7 +343,7 @@ class LatentApplyOperationCFG(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LatentApplyOperationCFG", - category="latent/advanced/operations", + category="model/latent/advanced/operations", is_experimental=True, inputs=[ io.Model.Input("model"), @@ -375,7 +375,7 @@ class LatentOperationTonemapReinhard(io.ComfyNode): return io.Schema( node_id="LatentOperationTonemapReinhard", search_aliases=["hdr latent"], - category="latent/advanced/operations", + category="model/latent/advanced/operations", is_experimental=True, inputs=[ io.Float.Input("multiplier", default=1.0, min=0.0, max=100.0, step=0.01), @@ -410,7 +410,7 @@ class LatentOperationSharpen(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LatentOperationSharpen", - category="latent/advanced/operations", + category="model/latent/advanced/operations", is_experimental=True, inputs=[ io.Int.Input("sharpen_radius", default=9, min=1, max=31, step=1, advanced=True), @@ -447,7 +447,7 @@ class ReplaceVideoLatentFrames(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="ReplaceVideoLatentFrames", - category="latent/batch", + category="model/latent/batch", inputs=[ io.Latent.Input("destination", tooltip="The destination latent where frames will be replaced."), io.Latent.Input("source", optional=True, tooltip="The source latent providing frames to insert into the destination latent. If not provided, the destination latent is returned unchanged."), diff --git a/comfy_extras/nodes_load_3d.py b/comfy_extras/nodes_load_3d.py index 9112bdd0a..455897859 100644 --- a/comfy_extras/nodes_load_3d.py +++ b/comfy_extras/nodes_load_3d.py @@ -34,7 +34,7 @@ class Load3D(IO.ComfyNode): essentials_category="Basics", is_experimental=True, inputs=[ - IO.Combo.Input("model_file", options=sorted(files), upload=IO.UploadType.model), + IO.Combo.Input("model_file", options=["none"] + sorted(files), upload=IO.UploadType.model), IO.Load3D.Input("image"), IO.Int.Input("width", default=1024, min=1, max=4096, step=1), IO.Int.Input("height", default=1024, min=1, max=4096, step=1), @@ -47,9 +47,18 @@ class Load3D(IO.ComfyNode): IO.Load3DCamera.Output(display_name="camera_info"), IO.Video.Output(display_name="recording_video"), IO.File3DAny.Output(display_name="model_3d"), + IO.Load3DModelInfo.Output(display_name="model_3d_info"), ], ) + @classmethod + def validate_inputs(cls, model_file, **kwargs) -> bool | str: + if not model_file or model_file == "none": + return True + if not folder_paths.exists_annotated_filepath(model_file): + return f"Invalid 3D model file: {model_file}" + return True + @classmethod def execute(cls, model_file, image, **kwargs) -> IO.NodeOutput: image_path = folder_paths.get_annotated_filepath(image['image']) @@ -68,8 +77,13 @@ class Load3D(IO.ComfyNode): video = InputImpl.VideoFromFile(recording_video_path) - file_3d = Types.File3D(folder_paths.get_annotated_filepath(model_file)) - return IO.NodeOutput(output_image, output_mask, model_file, normal_image, image['camera_info'], video, file_3d) + file_3d = None + mesh_path = "" + if model_file and model_file != "none": + file_3d = Types.File3D(folder_paths.get_annotated_filepath(model_file)) + mesh_path = model_file + model_3d_info = image.get('model_3d_info', []) + return IO.NodeOutput(output_image, output_mask, mesh_path, normal_image, image['camera_info'], video, file_3d, model_3d_info) process = execute # TODO: remove @@ -118,12 +132,200 @@ class Preview3D(IO.ComfyNode): process = execute # TODO: remove +class Preview3DAdvanced(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="Preview3DAdvanced", + display_name="Preview 3D (Advanced)", + search_aliases=["preview 3d", "3d viewer", "view mesh", "frame 3d", "3d camera output"], + category="3d", + is_experimental=True, + is_output_node=True, + inputs=[ + IO.MultiType.Input( + "model_3d", + types=[ + IO.File3DGLB, + IO.File3DGLTF, + IO.File3DFBX, + IO.File3DOBJ, + IO.File3DSTL, + IO.File3DUSDZ, + IO.File3DAny, + ], + tooltip="3D model file from an upstream 3D node.", + ), + IO.Load3DModelInfo.Input("model_3d_info", optional=True, advanced=True), + IO.Load3D.Input("viewport_state"), + IO.Load3DCamera.Input("camera_info", optional=True, advanced=True), + IO.Int.Input("width", default=1024, min=1, max=4096, step=1), + IO.Int.Input("height", default=1024, min=1, max=4096, step=1), + ], + outputs=[ + IO.File3DAny.Output(display_name="model_3d"), + IO.Load3DModelInfo.Output(display_name="model_3d_info"), + IO.Load3DCamera.Output(display_name="camera_info"), + IO.Int.Output(display_name="width"), + IO.Int.Output(display_name="height"), + ], + ) + + @classmethod + def execute(cls, model_3d: Types.File3D, viewport_state, width: int, height: int, **kwargs) -> IO.NodeOutput: + filename = f"preview3d_advanced_{uuid.uuid4().hex}.{model_3d.format}" + model_3d.save_to(os.path.join(folder_paths.get_temp_directory(), filename)) + + camera_info_input = kwargs.get("camera_info", None) + camera_info = camera_info_input if camera_info_input is not None else viewport_state['camera_info'] + model_3d_info_input = kwargs.get("model_3d_info", None) + model_3d_info = model_3d_info_input if model_3d_info_input is not None else viewport_state.get('model_3d_info', []) + return IO.NodeOutput( + model_3d, + model_3d_info, + camera_info, + width, + height, + ui=UI.PreviewUI3DAdvanced(filename, camera_info, model_3d_info), + ) + + +class PreviewGaussianSplat(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="PreviewGaussianSplat", + display_name="Preview Splat", + category="3d", + is_experimental=True, + is_output_node=True, + search_aliases=[ + "view splat", + "view gaussian", + "view gaussian splat", + "preview gaussian", + "preview gaussian splat", + "view 3dgs", + "preview 3dgs", + "preview ply", + "preview spz", + "preview splat", + "preview ksplat", + ], + inputs=[ + IO.MultiType.Input( + "model_3d", + types=[ + IO.File3DSplatAny, + IO.File3DPLY, + IO.File3DSPLAT, + IO.File3DSPZ, + IO.File3DKSPLAT, + ], + tooltip="A gaussian splat 3D file.", + ), + IO.Load3DModelInfo.Input("model_3d_info", optional=True, advanced=True), + IO.Load3D.Input("viewport_state"), + IO.Load3DCamera.Input("camera_info", optional=True, advanced=True), + IO.Int.Input("width", default=1024, min=1, max=4096, step=1), + IO.Int.Input("height", default=1024, min=1, max=4096, step=1), + ], + outputs=[ + IO.File3DSplatAny.Output(display_name="model_3d"), + IO.Load3DModelInfo.Output(display_name="model_3d_info"), + IO.Load3DCamera.Output(display_name="camera_info"), + IO.Int.Output(display_name="width"), + IO.Int.Output(display_name="height"), + ], + ) + + @classmethod + def execute(cls, model_3d: Types.File3D, viewport_state, width: int, height: int, **kwargs) -> IO.NodeOutput: + filename = f"preview_splat_{uuid.uuid4().hex}.{model_3d.format}" + model_3d.save_to(os.path.join(folder_paths.get_temp_directory(), filename)) + + camera_info_input = kwargs.get("camera_info", None) + camera_info = camera_info_input if camera_info_input is not None else viewport_state['camera_info'] + model_3d_info_input = kwargs.get("model_3d_info", None) + model_3d_info = model_3d_info_input if model_3d_info_input is not None else viewport_state.get('model_3d_info', []) + return IO.NodeOutput( + model_3d, + model_3d_info, + camera_info, + width, + height, + ui=UI.PreviewUI3DAdvanced(filename, camera_info, model_3d_info), + ) + + +class PreviewPointCloud(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="PreviewPointCloud", + display_name="Preview Point Cloud", + category="3d", + is_experimental=True, + is_output_node=True, + search_aliases=[ + "view point cloud", + "view pointcloud", + "preview point cloud", + "preview pointcloud", + "preview ply", + ], + inputs=[ + IO.MultiType.Input( + "model_3d", + types=[ + IO.File3DPointCloudAny, + IO.File3DPLY, + ], + tooltip="Point cloud file (.ply)", + ), + IO.Load3DModelInfo.Input("model_3d_info", optional=True, advanced=True), + IO.Load3D.Input("viewport_state"), + IO.Load3DCamera.Input("camera_info", optional=True, advanced=True), + IO.Int.Input("width", default=1024, min=1, max=4096, step=1), + IO.Int.Input("height", default=1024, min=1, max=4096, step=1), + ], + outputs=[ + IO.File3DPointCloudAny.Output(display_name="model_3d"), + IO.Load3DModelInfo.Output(display_name="model_3d_info"), + IO.Load3DCamera.Output(display_name="camera_info"), + IO.Int.Output(display_name="width"), + IO.Int.Output(display_name="height"), + ], + ) + + @classmethod + def execute(cls, model_3d: Types.File3D, viewport_state, width: int, height: int, **kwargs) -> IO.NodeOutput: + filename = f"preview_pointcloud_{uuid.uuid4().hex}.{model_3d.format}" + model_3d.save_to(os.path.join(folder_paths.get_temp_directory(), filename)) + + camera_info_input = kwargs.get("camera_info", None) + camera_info = camera_info_input if camera_info_input is not None else viewport_state['camera_info'] + model_3d_info_input = kwargs.get("model_3d_info", None) + model_3d_info = model_3d_info_input if model_3d_info_input is not None else viewport_state.get('model_3d_info', []) + return IO.NodeOutput( + model_3d, + model_3d_info, + camera_info, + width, + height, + ui=UI.PreviewUI3DAdvanced(filename, camera_info, model_3d_info), + ) + + class Load3DExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: return [ Load3D, Preview3D, + Preview3DAdvanced, + PreviewGaussianSplat, + PreviewPointCloud, ] diff --git a/comfy_extras/nodes_logic.py b/comfy_extras/nodes_logic.py index c066064ac..95f6ab848 100644 --- a/comfy_extras/nodes_logic.py +++ b/comfy_extras/nodes_logic.py @@ -1,4 +1,3 @@ -from __future__ import annotations from typing import TypedDict from typing_extensions import override from comfy_api.latest import ComfyExtension, io @@ -8,6 +7,82 @@ from comfy_api.latest import _io MISSING = object() +class NotNode(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="ComfyNotNode", + display_name="Not", + category="utilities/logic", + description="Logical NOT operation. Returns true if the value is falsy. Uses Python's rules for truthiness.", + search_aliases=["invert", "toggle", "negate", "flip boolean"], + inputs=[ + io.AnyType.Input("value"), + ], + outputs=[ + io.Boolean.Output(), + ], + ) + + @classmethod + def execute(cls, value) -> io.NodeOutput: + return io.NodeOutput(not value) + + +class AndNode(io.ComfyNode): + @classmethod + def define_schema(cls): + template = io.Autogrow.TemplatePrefix( + input=io.AnyType.Input("value"), + prefix="value", + min=1, + ) + return io.Schema( + node_id="ComfyAndNode", + display_name="And", + category="utilities/logic", + description="Logical AND operation. Returns true if all of the values are truthy. Uses Python's rules for truthiness.", + search_aliases=["all", "every"], + inputs=[ + io.Autogrow.Input("values", template=template), + ], + outputs=[ + io.Boolean.Output(), + ], + ) + + @classmethod + def execute(cls, values: io.Autogrow.Type) -> io.NodeOutput: + return io.NodeOutput(all(values.values())) + + +class OrNode(io.ComfyNode): + @classmethod + def define_schema(cls): + template = io.Autogrow.TemplatePrefix( + input=io.AnyType.Input("value"), + prefix="value", + min=1, + ) + return io.Schema( + node_id="ComfyOrNode", + display_name="Or", + category="utilities/logic", + description="Logical OR operation. Returns true if any of the values are truthy. Uses Python's rules for truthiness.", + search_aliases=["any", "some"], + inputs=[ + io.Autogrow.Input("values", template=template), + ], + outputs=[ + io.Boolean.Output(), + ], + ) + + @classmethod + def execute(cls, values: io.Autogrow.Type) -> io.NodeOutput: + return io.NodeOutput(any(values.values())) + + class SwitchNode(io.ComfyNode): @classmethod def define_schema(cls): @@ -15,7 +90,7 @@ class SwitchNode(io.ComfyNode): return io.Schema( node_id="ComfySwitchNode", display_name="Switch", - category="logic", + category="utilities/logic", is_experimental=True, inputs=[ io.Boolean.Input("switch"), @@ -46,7 +121,7 @@ class SoftSwitchNode(io.ComfyNode): return io.Schema( node_id="ComfySoftSwitchNode", display_name="Soft Switch", - category="logic", + category="utilities/logic", is_experimental=True, inputs=[ io.Boolean.Input("switch"), @@ -101,7 +176,7 @@ class CustomComboNode(io.ComfyNode): return io.Schema( node_id="CustomCombo", display_name="Custom Combo", - category="utils", + category="utilities", is_experimental=True, inputs=[io.Combo.Input("choice", options=[])], outputs=[ @@ -136,7 +211,7 @@ class DCTestNode(io.ComfyNode): return io.Schema( node_id="DCTestNode", display_name="DCTest", - category="logic", + category="utilities/logic", is_output_node=True, inputs=[io.DynamicCombo.Input("combo", options=[ io.DynamicCombo.Option("option1", [io.String.Input("string")]), @@ -174,7 +249,7 @@ class AutogrowNamesTestNode(io.ComfyNode): return io.Schema( node_id="AutogrowNamesTestNode", display_name="AutogrowNamesTest", - category="logic", + category="utilities/logic", inputs=[ _io.Autogrow.Input("autogrow", template=template) ], @@ -194,7 +269,7 @@ class AutogrowPrefixTestNode(io.ComfyNode): return io.Schema( node_id="AutogrowPrefixTestNode", display_name="AutogrowPrefixTest", - category="logic", + category="utilities/logic", inputs=[ _io.Autogrow.Input("autogrow", template=template) ], @@ -213,7 +288,7 @@ class ComboOutputTestNode(io.ComfyNode): return io.Schema( node_id="ComboOptionTestNode", display_name="ComboOptionTest", - category="logic", + category="utilities/logic", inputs=[io.Combo.Input("combo", options=["option1", "option2", "option3"]), io.Combo.Input("combo2", options=["option4", "option5", "option6"])], outputs=[io.Combo.Output(), io.Combo.Output()], @@ -230,7 +305,7 @@ class ConvertStringToComboNode(io.ComfyNode): node_id="ConvertStringToComboNode", search_aliases=["string to dropdown", "text to combo"], display_name="Convert String to Combo", - category="logic", + category="utilities/logic", inputs=[io.String.Input("string")], outputs=[io.Combo.Output()], ) @@ -246,7 +321,7 @@ class InvertBooleanNode(io.ComfyNode): node_id="InvertBooleanNode", search_aliases=["not", "toggle", "negate", "flip boolean"], display_name="Invert Boolean", - category="logic", + category="utilities/logic", inputs=[io.Boolean.Input("boolean")], outputs=[io.Boolean.Output()], ) @@ -261,6 +336,9 @@ class LogicExtension(ComfyExtension): return [ SwitchNode, CustomComboNode, + NotNode, + AndNode, + OrNode, # SoftSwitchNode, # ConvertStringToComboNode, # DCTestNode, diff --git a/comfy_extras/nodes_lora_debug.py b/comfy_extras/nodes_lora_debug.py index 937a0fbfb..3f68064e5 100644 --- a/comfy_extras/nodes_lora_debug.py +++ b/comfy_extras/nodes_lora_debug.py @@ -30,7 +30,7 @@ class LoraLoaderBypass: OUTPUT_TOOLTIPS = ("The modified diffusion model.", "The modified CLIP model.") FUNCTION = "load_lora" - CATEGORY = "loaders" + CATEGORY = "model/loaders" DESCRIPTION = "Apply LoRA in bypass mode. Unlike regular LoRA, this doesn't modify model weights - instead it injects the LoRA computation during forward pass. Useful for training scenarios." EXPERIMENTAL = True diff --git a/comfy_extras/nodes_lotus.py b/comfy_extras/nodes_lotus.py index 9f62ba2bf..9fe4c5c7b 100644 --- a/comfy_extras/nodes_lotus.py +++ b/comfy_extras/nodes_lotus.py @@ -10,7 +10,7 @@ class LotusConditioning(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LotusConditioning", - category="conditioning/lotus", + category="model/conditioning/lotus", inputs=[], outputs=[io.Conditioning.Output(display_name="conditioning")], ) diff --git a/comfy_extras/nodes_lt.py b/comfy_extras/nodes_lt.py index 51cf7951f..6d6078abe 100644 --- a/comfy_extras/nodes_lt.py +++ b/comfy_extras/nodes_lt.py @@ -25,7 +25,7 @@ class GetICLoRAParameters(io.ComfyNode): display_name="Get IC-LoRA Parameters", description="Extracts IC-LoRA parameters from the safetensors metadata of a LoRA-loaded " "model and outputs them for LTXVAddGuide (eg. reference_downscale_factor).", - category="conditioning/video_models", + category="model/conditioning/video_models", search_aliases=["ic-lora", "ic lora", "iclora", "downscale factor", "reference downscale"], inputs=[ io.Model.Input( @@ -62,7 +62,7 @@ class EmptyLTXVLatentVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="EmptyLTXVLatentVideo", - category="latent/video/ltxv", + category="model/latent/video/ltxv", inputs=[ io.Int.Input("width", default=768, min=64, max=nodes.MAX_RESOLUTION, step=32), io.Int.Input("height", default=512, min=64, max=nodes.MAX_RESOLUTION, step=32), @@ -86,7 +86,7 @@ class LTXVImgToVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LTXVImgToVideo", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), @@ -131,7 +131,7 @@ class LTXVImgToVideoInplace(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LTXVImgToVideoInplace", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Vae.Input("vae"), io.Image.Input("image"), @@ -226,10 +226,20 @@ def get_noise_mask(latent): noise_mask = noise_mask.clone() return noise_mask -def get_keyframe_idxs(cond): +def get_keyframe_idxs(cond, latent_shape=None): keyframe_idxs = conditioning_get_any_value(cond, "keyframe_idxs", None) if keyframe_idxs is None: return None, 0 + # Get number of keyframes from latent_shape or guide_attention_entries if available + if latent_shape is not None and len(latent_shape) == 5: + tokens_per_frame = latent_shape[-2] * latent_shape[-1] + num_keyframes = keyframe_idxs.shape[2] // tokens_per_frame + return keyframe_idxs, num_keyframes + entries = conditioning_get_any_value(cond, "guide_attention_entries", None) + if entries: + num_keyframes = sum(e["latent_shape"][0] for e in entries) + return keyframe_idxs, num_keyframes + # fallback, may under-count if keyframes share t-start # keyframe_idxs contains start/end positions (last dimension), checking for unqiue values only for start num_keyframes = torch.unique(keyframe_idxs[:, 0, :, 0]).shape[0] return keyframe_idxs, num_keyframes @@ -241,7 +251,7 @@ class LTXVAddGuide(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LTXVAddGuide", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), @@ -322,9 +332,9 @@ class LTXVAddGuide(io.ComfyNode): return factor @classmethod - def get_latent_index(cls, cond, latent_length, guide_length, frame_idx, scale_factors): + def get_latent_index(cls, cond, latent_length, guide_length, frame_idx, scale_factors, latent_shape=None): time_scale_factor, _, _ = scale_factors - _, num_keyframes = get_keyframe_idxs(cond) + _, num_keyframes = get_keyframe_idxs(cond, latent_shape) latent_count = latent_length - num_keyframes frame_idx = frame_idx if frame_idx >= 0 else max((latent_count - 1) * time_scale_factor + 1 + frame_idx, 0) if guide_length > 1 and frame_idx != 0: @@ -436,7 +446,7 @@ class LTXVAddGuide(io.ComfyNode): num_frames_to_keep = ((image.shape[0] - 1) // time_scale_factor) * time_scale_factor + 1 resolved_frame_idx = frame_idx if frame_idx < 0: - _, num_keyframes = get_keyframe_idxs(positive) + _, num_keyframes = get_keyframe_idxs(positive, latent_image.shape) resolved_frame_idx = max((latent_length - num_keyframes - 1) * time_scale_factor + 1 + frame_idx, 0) causal_fix = resolved_frame_idx == 0 or num_frames_to_keep == 1 @@ -454,7 +464,7 @@ class LTXVAddGuide(io.ComfyNode): if latent_downscale_factor > 1: t, guide_mask = cls.dilate_latent(t, latent_downscale_factor) - frame_idx, latent_idx = cls.get_latent_index(positive, latent_length, len(image), frame_idx, scale_factors) + frame_idx, latent_idx = cls.get_latent_index(positive, latent_length, len(image), frame_idx, scale_factors, latent_shape=latent_image.shape) assert latent_idx + t.shape[2] <= latent_length, "Conditioning frames exceed the length of the latent sequence." positive, negative, latent_image, noise_mask = cls.append_keyframe( @@ -488,7 +498,7 @@ class LTXVCropGuides(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LTXVCropGuides", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), @@ -506,7 +516,7 @@ class LTXVCropGuides(io.ComfyNode): latent_image = latent["samples"].clone() noise_mask = get_noise_mask(latent) - _, num_keyframes = get_keyframe_idxs(positive) + _, num_keyframes = get_keyframe_idxs(positive, latent_image.shape) if num_keyframes == 0: return io.NodeOutput(positive, negative, {"samples": latent_image, "noise_mask": noise_mask},) @@ -532,7 +542,7 @@ class LTXVConditioning(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LTXVConditioning", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), @@ -601,7 +611,7 @@ class LTXVScheduler(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LTXVScheduler", - category="sampling/schedulers", + category="model/sampling/schedulers", inputs=[ io.Int.Input("steps", default=20, min=1, max=10000), io.Float.Input("max_shift", default=2.05, min=0.0, max=100.0, step=0.01), @@ -736,7 +746,7 @@ class LTXVConcatAVLatent(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LTXVConcatAVLatent", - category="latent/video/ltxv", + category="model/latent/video/ltxv", inputs=[ io.Latent.Input("video_latent"), io.Latent.Input("audio_latent"), @@ -771,7 +781,7 @@ class LTXVSeparateAVLatent(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LTXVSeparateAVLatent", - category="latent/video/ltxv", + category="model/latent/video/ltxv", description="LTXV Separate AV Latent", inputs=[ io.Latent.Input("av_latent"), @@ -804,7 +814,7 @@ class LTXVReferenceAudio(io.ComfyNode): return io.Schema( node_id="LTXVReferenceAudio", display_name="LTXV Reference Audio (ID-LoRA)", - category="conditioning/audio", + category="model/conditioning/audio", description="Set reference audio for ID-LoRA speaker identity transfer. Encodes a reference audio clip into the conditioning and optionally patches the model with identity guidance (extra forward pass without reference, amplifying the speaker identity effect).", inputs=[ io.Model.Input("model"), diff --git a/comfy_extras/nodes_lt_audio.py b/comfy_extras/nodes_lt_audio.py index 51ddf584a..052186083 100644 --- a/comfy_extras/nodes_lt_audio.py +++ b/comfy_extras/nodes_lt_audio.py @@ -12,7 +12,7 @@ class LTXVAudioVAELoader(io.ComfyNode): return io.Schema( node_id="LTXVAudioVAELoader", display_name="Load LTXV Audio VAE", - category="loaders", + category="model/loaders", inputs=[ io.Combo.Input( "ckpt_name", @@ -40,7 +40,7 @@ class LTXVAudioVAEEncode(VAEEncodeAudio): return io.Schema( node_id="LTXVAudioVAEEncode", display_name="LTXV Audio VAE Encode", - category="latent/audio", + category="model/latent/audio", inputs=[ io.Audio.Input("audio", tooltip="The audio to be encoded."), io.Vae.Input( @@ -63,7 +63,7 @@ class LTXVAudioVAEDecode(io.ComfyNode): return io.Schema( node_id="LTXVAudioVAEDecode", display_name="LTXV Audio VAE Decode", - category="latent/audio", + category="model/latent/audio", inputs=[ io.Latent.Input("samples", tooltip="The latent to be decoded."), io.Vae.Input( @@ -96,7 +96,7 @@ class LTXVEmptyLatentAudio(io.ComfyNode): return io.Schema( node_id="LTXVEmptyLatentAudio", display_name="LTXV Empty Latent Audio", - category="latent/audio", + category="model/latent/audio", inputs=[ io.Int.Input( "frames_number", diff --git a/comfy_extras/nodes_lt_upsampler.py b/comfy_extras/nodes_lt_upsampler.py index f99ba13fb..be9a36e69 100644 --- a/comfy_extras/nodes_lt_upsampler.py +++ b/comfy_extras/nodes_lt_upsampler.py @@ -1,32 +1,32 @@ from comfy import model_management +from comfy_api.latest import ComfyExtension, IO +from typing_extensions import override import math -class LTXVLatentUpsampler: + +class LTXVLatentUpsampler(IO.ComfyNode): """ Upsamples a video latent by a factor of 2. """ @classmethod - def INPUT_TYPES(s): - return { - "required": { - "samples": ("LATENT",), - "upscale_model": ("LATENT_UPSCALE_MODEL",), - "vae": ("VAE",), - } - } + def define_schema(cls): + return IO.Schema( + node_id="LTXVLatentUpsampler", + category="model/latent/video", + is_experimental=True, + inputs=[ + IO.Latent.Input("samples"), + IO.LatentUpscaleModel.Input("upscale_model"), + IO.Vae.Input("vae"), + ], + outputs=[ + IO.Latent.Output(), + ], + ) - RETURN_TYPES = ("LATENT",) - FUNCTION = "upsample_latent" - CATEGORY = "latent/video" - EXPERIMENTAL = True - - def upsample_latent( - self, - samples: dict, - upscale_model, - vae, - ) -> tuple: + @classmethod + def execute(cls, samples, upscale_model, vae) -> IO.NodeOutput: """ Upsample the input latent using the provided model. @@ -34,7 +34,6 @@ class LTXVLatentUpsampler: samples (dict): Input latent samples upscale_model (LatentUpsampler): Loaded upscale model vae: VAE model for normalization - auto_tiling (bool): Whether to automatically tile the input for processing Returns: tuple: Tuple containing the upsampled latent @@ -67,9 +66,16 @@ class LTXVLatentUpsampler: return_dict = samples.copy() return_dict["samples"] = upsampled_latents return_dict.pop("noise_mask", None) - return (return_dict,) + return IO.NodeOutput(return_dict) + + upsample_latent = execute # TODO: remove -NODE_CLASS_MAPPINGS = { - "LTXVLatentUpsampler": LTXVLatentUpsampler, -} +class LTXVLatentUpsamplerExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [LTXVLatentUpsampler] + + +async def comfy_entrypoint() -> LTXVLatentUpsamplerExtension: + return LTXVLatentUpsamplerExtension() diff --git a/comfy_extras/nodes_lumina2.py b/comfy_extras/nodes_lumina2.py index b35ab8b7d..c060a86a0 100644 --- a/comfy_extras/nodes_lumina2.py +++ b/comfy_extras/nodes_lumina2.py @@ -81,7 +81,7 @@ class CLIPTextEncodeLumina2(io.ComfyNode): node_id="CLIPTextEncodeLumina2", search_aliases=["lumina prompt"], display_name="CLIP Text Encode for Lumina2", - category="conditioning", + category="model/conditioning", description="Encodes a system prompt and a user prompt using a CLIP model into an embedding " "that can be used to guide the diffusion model towards generating specific images.", inputs=[ diff --git a/comfy_extras/nodes_mask.py b/comfy_extras/nodes_mask.py index d15f1f4e7..52484697a 100644 --- a/comfy_extras/nodes_mask.py +++ b/comfy_extras/nodes_mask.py @@ -53,7 +53,7 @@ class LatentCompositeMasked(IO.ComfyNode): return IO.Schema( node_id="LatentCompositeMasked", search_aliases=["overlay latent", "layer latent", "paste latent", "inpaint latent"], - category="latent", + category="model/latent", inputs=[ IO.Latent.Input("destination"), IO.Latent.Input("source"), diff --git a/comfy_extras/nodes_math.py b/comfy_extras/nodes_math.py index 6030ee9d8..0883c65ac 100644 --- a/comfy_extras/nodes_math.py +++ b/comfy_extras/nodes_math.py @@ -4,7 +4,6 @@ Provides a ComfyMathExpression node that evaluates math expressions against dynamically-grown numeric inputs. """ -from __future__ import annotations import math import string @@ -70,7 +69,7 @@ class MathExpressionNode(io.ComfyNode): return io.Schema( node_id="ComfyMathExpression", display_name="Math Expression", - category="logic", + category="utilities", search_aliases=[ "expression", "formula", "calculate", "calculator", "eval", "math", @@ -103,11 +102,18 @@ class MathExpressionNode(io.ComfyNode): f"Math Expression '{expression}' must evaluate to a numeric result, " f"got {type(result).__name__}: {result!r}" ) - if not math.isfinite(result): + try: + float_result = float(result) + except OverflowError: + raise ValueError( + f"Math Expression '{expression}' produced a result too large to " + f"represent as a float: {result}" + ) from None + if not math.isfinite(float_result): raise ValueError( f"Math Expression '{expression}' produced a non-finite result: {result}" ) - return io.NodeOutput(float(result), int(result), bool(result)) + return io.NodeOutput(float_result, int(result), bool(result)) class MathExtension(ComfyExtension): diff --git a/comfy_extras/nodes_mediapipe.py b/comfy_extras/nodes_mediapipe.py index 6b7916aee..343d88dbb 100644 --- a/comfy_extras/nodes_mediapipe.py +++ b/comfy_extras/nodes_mediapipe.py @@ -10,7 +10,6 @@ Custom IO types: MediaPipeFaceLandmarker also emits the core BOUNDING_BOX type — pair with DrawBBoxes. """ -from __future__ import annotations import numpy as np import torch @@ -206,7 +205,7 @@ class LoadMediaPipeFaceLandmarker(io.ComfyNode): node_id="LoadMediaPipeFaceLandmarker", search_aliases=["face", "facial", "mediapipe", "face landmark", "face mesh", "blazeface", "face detection"], display_name="Load Face Detection Model (MediaPipe)", - category="loaders", + category="model/loaders", inputs=[ io.Combo.Input("model_name", options=folder_paths.get_filename_list("detection"), tooltip="Face detection model from models/detection/."), diff --git a/comfy_extras/nodes_mochi.py b/comfy_extras/nodes_mochi.py index d750194fc..3dcea6ab3 100644 --- a/comfy_extras/nodes_mochi.py +++ b/comfy_extras/nodes_mochi.py @@ -10,7 +10,7 @@ class EmptyMochiLatentVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="EmptyMochiLatentVideo", - category="latent/video", + category="model/latent/video", inputs=[ io.Int.Input("width", default=848, min=16, max=nodes.MAX_RESOLUTION, step=16), io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16), diff --git a/comfy_extras/nodes_model_downscale.py b/comfy_extras/nodes_model_downscale.py index 24d47a903..817542452 100644 --- a/comfy_extras/nodes_model_downscale.py +++ b/comfy_extras/nodes_model_downscale.py @@ -10,7 +10,7 @@ class PatchModelAddDownscale(io.ComfyNode): return io.Schema( node_id="PatchModelAddDownscale", display_name="PatchModelAddDownscale (Kohya Deep Shrink)", - category="model_patches/unet", + category="model/patch/unet", inputs=[ io.Model.Input("model"), io.Int.Input("block_number", default=3, min=1, max=32, step=1, advanced=True), diff --git a/comfy_extras/nodes_model_patch.py b/comfy_extras/nodes_model_patch.py index 748559a6b..bdccbf8c4 100644 --- a/comfy_extras/nodes_model_patch.py +++ b/comfy_extras/nodes_model_patch.py @@ -548,7 +548,7 @@ class USOStyleReference: FUNCTION = "apply_patch" EXPERIMENTAL = True - CATEGORY = "advanced/model_patches/flux" + CATEGORY = "model/patch/flux" def apply_patch(self, model, model_patch, clip_vision_output): encoded_image = torch.stack((clip_vision_output.all_hidden_states[:, -20], clip_vision_output.all_hidden_states[:, -11], clip_vision_output.penultimate_hidden_states)) @@ -594,7 +594,7 @@ class SUPIRApply(io.ComfyNode): def define_schema(cls) -> io.Schema: return io.Schema( node_id="SUPIRApply", - category="model_patches/supir", + category="model/patch/supir", is_experimental=True, inputs=[ io.Model.Input("model"), diff --git a/comfy_extras/nodes_moge.py b/comfy_extras/nodes_moge.py index 7968c6cda..c29abe53b 100644 --- a/comfy_extras/nodes_moge.py +++ b/comfy_extras/nodes_moge.py @@ -1,6 +1,5 @@ """ComfyUI nodes for the native MoGe (Monocular Geometry Estimation) integration.""" -from __future__ import annotations import torch @@ -9,6 +8,7 @@ import folder_paths from comfy_api.latest import ComfyExtension, Types, io from typing_extensions import override +from comfy.ldm.colormap import turbo as _turbo from comfy.ldm.moge.model import MoGeModel from comfy.ldm.moge.geometry import triangulate_grid_mesh from comfy.ldm.moge.panorama import get_panorama_cameras, split_panorama_image, merge_panorama_depth, spherical_uv_to_directions, _uv_grid @@ -28,19 +28,6 @@ MoGeGeometry = io.Custom("MOGE_GEOMETRY") # "image": torch.Tensor (B, H, W, 3) in [0, 1], CPU (always present) -def _turbo(x: torch.Tensor) -> torch.Tensor: - """Anton Mikhailov polynomial approximation of the turbo colormap.""" - x = x.clamp(0.0, 1.0) - x2 = x * x - x3 = x2 * x - x4 = x2 * x2 - x5 = x4 * x - r = 0.13572138 + 4.61539260*x - 42.66032258*x2 + 132.13108234*x3 - 152.94239396*x4 + 59.28637943*x5 - g = 0.09140261 + 2.19418839*x + 4.84296658*x2 - 14.18503333*x3 + 4.27729857*x4 + 2.82956604*x5 - b = 0.10667330 + 12.64194608*x - 60.58204836*x2 + 110.36276771*x3 - 89.90310912*x4 + 27.34824973*x5 - return torch.stack([r, g, b], dim=-1).clamp(0.0, 1.0) - - def _normals_from_points(points: torch.Tensor) -> torch.Tensor: """Camera-space surface normals from a (B, H, W, 3) point map (v1 fallback).""" finite = torch.isfinite(points).all(dim=-1) @@ -79,7 +66,7 @@ class LoadMoGeModel(io.ComfyNode): return io.Schema( node_id="LoadMoGeModel", display_name="Load MoGe Model", - category="loaders", + category="model/loaders", inputs=[ io.Combo.Input("model_name", options=folder_paths.get_filename_list("geometry_estimation")), ], @@ -105,7 +92,7 @@ class MoGePanoramaInference(io.ComfyNode): node_id="MoGePanoramaInference", search_aliases=["moge", "panorama", "depth", "geometry", "depth estimation", "geometry estimation"], display_name="Run MoGe Panorama Inference", - category="image/geometry_estimation", + category="image/geometry estimation", description="Run MoGe on an equirectangular panorama by splitting it into 12 perspective views, running inference on each, and merging the results into a single depth map.", inputs=[ MoGeModelType.Input("moge_model"), @@ -227,7 +214,7 @@ class MoGeInference(io.ComfyNode): search_aliases=["moge", "depth", "geometry", "depth estimation", "geometry estimation"], display_name="Run MoGe Inference", description="Run MoGe on a single image to estimate depth and geometry.", - category="image/geometry_estimation", + category="image/geometry estimation", inputs=[ MoGeModelType.Input("moge_model"), io.Image.Input("image"), @@ -284,7 +271,7 @@ class MoGeRender(io.ComfyNode): search_aliases=["moge", "render", "geometry", "depth", "normal"], display_name="Render MoGe Geometry", description="Render a depth map or normal map from geometry data", - category="image/geometry_estimation", + category="image/geometry estimation", inputs=[ MoGeGeometry.Input("moge_geometry"), io.Combo.Input("output", options=["depth", "depth_colored", "normal_opengl", "normal_directx", "mask"], default="depth", @@ -351,7 +338,7 @@ class MoGePointMapToMesh(io.ComfyNode): search_aliases=["moge", "mesh", "geometry", "point map"], display_name="Convert MoGe Point Map to Mesh", description="Convert a MoGe point map into a 3D mesh.", - category="image/geometry_estimation", + category="image/geometry estimation", inputs=[ MoGeGeometry.Input("moge_geometry"), io.Int.Input("batch_index", default=0, min=0, max=4096, diff --git a/comfy_extras/nodes_multigpu.py b/comfy_extras/nodes_multigpu.py new file mode 100644 index 000000000..d2f6fe67a --- /dev/null +++ b/comfy_extras/nodes_multigpu.py @@ -0,0 +1,408 @@ +from __future__ import annotations + +import copy +import logging +from inspect import cleandoc +from typing import TYPE_CHECKING +from typing_extensions import override + +from comfy_api.latest import ComfyExtension, io + +if TYPE_CHECKING: + from comfy.model_patcher import ModelPatcher + from comfy.sd import CLIP, VAE +import torch + +import comfy.model_management +import comfy.multigpu + + +class MultiGPUCFGSplitNode(io.ComfyNode): + """ + Prepares model to have sampling accelerated via splitting work units. + + Should be placed after nodes that modify the model object itself, such as compile or attention-switch nodes. + + Other than those exceptions, this node can be placed in any order. + """ + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="MultiGPU_WorkUnits", + display_name="MultiGPU CFG Split", + category="advanced/multigpu", + description=cleandoc(cls.__doc__), + inputs=[ + io.Model.Input("model"), + io.Int.Input("max_gpus", default=2, min=1, step=1), + ], + outputs=[ + io.Model.Output(), + ], + ) + + @classmethod + def execute(cls, model: ModelPatcher, max_gpus: int) -> io.NodeOutput: + model = comfy.multigpu.create_multigpu_deepclones(model, max_gpus, reuse_loaded=True) + return io.NodeOutput(model) + + +def _force_supported_compute_dtype(patcher: ModelPatcher, device: torch.device): + """Cast compute dtype to one the device supports; no-op if already supported.""" + weight_dtype = patcher.model_dtype() + cast_dtype = comfy.model_management.unet_manual_cast(weight_dtype, device) + if cast_dtype is None: + return + logging.info(f"Select Model Device: using {cast_dtype} compute dtype on {device} (model weight dtype was {weight_dtype}).") + patcher.set_model_compute_dtype(cast_dtype) + + +def _remember_base_devices(patcher: ModelPatcher): + """Stash the original load/offload device on the underlying model. + + Stored on patcher.model (which is shared with the input patcher), so + later "default" selections can recover the loader's original routing. + Only the first Select on a given chain writes these attrs; subsequent + deepclones inherit them onto their freshly-loaded model below. + """ + if not hasattr(patcher.model, "_select_base_load_device"): + patcher.model._select_base_load_device = patcher.load_device + patcher.model._select_base_offload_device = patcher.offload_device + + +def _propagate_base_devices(src_model, dst_model): + """Carry the loader-original device attrs onto the freshly-deepcloned model.""" + if hasattr(src_model, "_select_base_load_device") and not hasattr(dst_model, "_select_base_load_device"): + dst_model._select_base_load_device = src_model._select_base_load_device + dst_model._select_base_offload_device = src_model._select_base_offload_device + + +def _retarget_patcher(patcher: ModelPatcher, target_load_device, target_offload_device): + """Return a patcher whose actual model weights live on *target_load_device*. + + If *patcher* is already on *target_load_device* we just retarget the + (already-cloned) patcher's metadata in place. Otherwise we call + :meth:`ModelPatcher.deepclone_multigpu` to spawn a fresh model from + the loader's ``cached_patcher_init`` factory -- the only safe way to + move weights that may already be partially loaded onto another device. + + NOTE: reusing the input patcher's model when the requested device + matches its current load_device is a deliberate fast path. Anything + that has already mutated the original model (e.g. a prior KSampler + invocation on the same model) will be observed here. This is by + design and documented on the SelectXDeviceNode docstrings -- placing + Select X Device after a node that consumes the same model is not + recommended. + """ + if patcher.load_device == target_load_device: + # Fast path: weights already on the desired device, just update offload. + patcher.offload_device = target_offload_device + return patcher + src_model = patcher.model + patcher = patcher.deepclone_multigpu(new_load_device=target_load_device) + patcher.offload_device = target_offload_device + _propagate_base_devices(src_model, patcher.model) + if hasattr(patcher, "register_load_device"): + patcher.register_load_device(patcher.load_device) + return patcher + + +def _apply_patcher_device(patcher: ModelPatcher, resolved, base_offload_override=None): + """Resolve the requested device and produce a patcher routed there. + + For "default" we restore the loader's original load/offload pair. + For CPU we pin both load and offload to CPU (and, on a dynamic + patcher, downgrade to a plain ModelPatcher so the dynamic-only + code paths are bypassed). + For an explicit GPU we keep the loader's original offload but + target the requested load device; if that differs from the current + load device the patcher is deepcloned onto the new device. + """ + _remember_base_devices(patcher) + base_load = patcher.model._select_base_load_device + base_offload = base_offload_override if base_offload_override is not None else patcher.model._select_base_offload_device + + if resolved is None: + # "default" -> route back to the loader's original devices. + return _retarget_patcher(patcher, base_load, base_offload) + if resolved.type == "cpu": + if patcher.is_dynamic(): + # clone(disable_dynamic=True) requires cached_patcher_init; let the + # exception surface to the caller (Select*DeviceNode.execute), which + # will translate it into a passthrough+log so unsupported loaders + # don't hard-fail the workflow. + patcher = patcher.clone(disable_dynamic=True) + patcher.load_device = resolved + patcher.offload_device = resolved + return patcher + return _retarget_patcher(patcher, resolved, base_offload) + + +def _prune_multigpu_collision(model: ModelPatcher, primary_device): + """Drop any multigpu clone whose load_device matches *primary_device*. + + Without pruning, MultiGPU CFG Split would have stacked a clone on + the same device the primary now occupies (i.e. the workflow places + MultiGPU CFG Split before Select Model Device). Keeps the clone set + consistent with the new primary placement. + """ + multigpu_models = model.get_additional_models_with_key("multigpu") + if not multigpu_models: + return + filtered = [m for m in multigpu_models if m.load_device != primary_device] + if len(filtered) != len(multigpu_models): + logging.info(f"Select Model Device: pruning MultiGPU clone on {primary_device} that now collides with the primary model.") + model.set_additional_models("multigpu", filtered) + if hasattr(model, "match_multigpu_clones"): + model.match_multigpu_clones() + + +class SelectModelDeviceNode(io.ComfyNode): + """ + Place the diffusion model on a specific device (default / cpu / gpu:N). + + - "default" restores the device assigned by the loader (even after a + prior Select Model Device call). + - "cpu" pins both the load and offload device to CPU. + - "gpu:N" pins the load device to the Nth available GPU; the offload + device is restored to the loader's original choice. + + When the requested device differs from the device the input model is + already on, a fresh model is spawned via the loader's reload factory + (cached_patcher_init) so the new patcher owns independent weights on + the new device. Loaders that don't support multigpu (no factory) will + cause the node to pass through unchanged with a warning. + + If the workflow already has MultiGPU CFG Split applied and the chosen + GPU collides with one of the existing multigpu clones, that clone is + dropped so two patchers don't end up bound to the same device. + + When the selected device does not exist on the current machine + (e.g. a workflow built on a 2-GPU box opened on a 1-GPU box), + the node passes the model through unchanged and logs a message + instead of failing. + + NOTE: Placing Select Model Device *after* a node that has already + consumed the same model (e.g. a KSampler that ran on this model on + the original device) is not recommended -- any state the prior + consumer mutated on the original model will be observed when the + selected device matches the original (fast path). Place Select Model + Device before any consumer of the model. + """ + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="SelectModelDevice", + display_name="Select Model Device", + category="advanced/multigpu", + description=cleandoc(cls.__doc__), + inputs=[ + io.Model.Input("model"), + io.Combo.Input("device", options=comfy.model_management.get_gpu_device_options()), + ], + outputs=[ + io.Model.Output(), + ], + ) + + @classmethod + def validate_inputs(cls, device="default"): + # Allow unknown gpu:N values so portable workflows do not error + # at validation time; runtime fallback will handle them. + return True + + @classmethod + def execute(cls, model: ModelPatcher, device: str = "default") -> io.NodeOutput: + model = model.clone() + resolved = comfy.model_management.resolve_gpu_device_option(device) + if resolved is None and device not in (None, "default"): + logging.info(f"Select Model Device: requested device '{device}' not available, passing through unchanged.") + return io.NodeOutput(model) + try: + model = _apply_patcher_device(model, resolved) + except RuntimeError as e: + logging.warning(f"Select Model Device: cannot retarget model, passing through unchanged. ({e})") + return io.NodeOutput(model) + if resolved is not None: + _force_supported_compute_dtype(model, resolved) + _prune_multigpu_collision(model, model.load_device) + return io.NodeOutput(model) + + +class SelectCLIPDeviceNode(io.ComfyNode): + """ + Place the CLIP text encoder on a specific device (default / cpu / gpu:N). + + - "default" restores the device assigned by the loader. + - "cpu" pins both the load and offload device to CPU. + - "gpu:N" pins the load device to the Nth available GPU. + + When the selected device does not exist on the current machine + (e.g. a workflow built on a 2-GPU box opened on a 1-GPU box), + the node passes the CLIP through unchanged and logs a message + instead of failing. + """ + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="SelectCLIPDevice", + display_name="Select CLIP Device", + category="advanced/multigpu", + description=cleandoc(cls.__doc__), + inputs=[ + io.Clip.Input("clip"), + io.Combo.Input("device", options=comfy.model_management.get_gpu_device_options()), + ], + outputs=[ + io.Clip.Output(), + ], + ) + + @classmethod + def validate_inputs(cls, device="default"): + return True + + @classmethod + def execute(cls, clip: CLIP, device: str = "default") -> io.NodeOutput: + clip = clip.clone() + resolved = comfy.model_management.resolve_gpu_device_option(device) + if resolved is None and device not in (None, "default"): + logging.info(f"Select CLIP Device: requested device '{device}' not available, passing through unchanged.") + return io.NodeOutput(clip) + try: + clip.patcher = _apply_patcher_device(clip.patcher, resolved) + except RuntimeError as e: + logging.warning(f"Select CLIP Device: cannot retarget CLIP, passing through unchanged. ({e})") + return io.NodeOutput(clip) + + +class SelectVAEDeviceNode(io.ComfyNode): + """ + Place the VAE on a specific device (default / gpu:N). + + - "default" restores the device assigned by the loader. + - "gpu:N" pins the load device to the Nth available GPU; the offload + device is set to the standard VAE offload device. + + CPU is intentionally not exposed in the UI for the VAE; if a workflow + supplies "cpu" anyway (e.g. opened from another machine), the request + is dropped with a log message and the VAE is passed through unchanged. + + When the selected device does not exist on the current machine + (e.g. a workflow built on a 2-GPU box opened on a 1-GPU box), + the node passes the VAE through unchanged and logs a message + instead of failing. + """ + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="SelectVAEDevice", + display_name="Select VAE Device", + category="advanced/multigpu", + description=cleandoc(cls.__doc__), + inputs=[ + io.Vae.Input("vae"), + io.Combo.Input("device", options=comfy.model_management.get_gpu_device_options_no_cpu()), + ], + outputs=[ + io.Vae.Output(), + ], + ) + + @classmethod + def validate_inputs(cls, device="default"): + return True + + @classmethod + def execute(cls, vae: VAE, device: str = "default") -> io.NodeOutput: + # VAE has no .clone(); shallow-copy the wrapper and clone the patcher + # so we can retarget load/offload device without affecting the input VAE. + vae = copy.copy(vae) + vae.patcher = vae.patcher.clone() + resolved = comfy.model_management.resolve_gpu_device_option(device) + if resolved is None and device not in (None, "default"): + logging.info(f"Select VAE Device: requested device '{device}' not available, passing through unchanged.") + return io.NodeOutput(vae) + if resolved is not None and resolved.type == "cpu": + logging.info("Select VAE Device: CPU is not a supported choice, passing through unchanged.") + return io.NodeOutput(vae) + if not hasattr(vae, "_select_base_device"): + vae._select_base_device = vae.device + try: + vae.patcher = _apply_patcher_device( + vae.patcher, resolved, + base_offload_override=comfy.model_management.vae_offload_device(), + ) + except RuntimeError as e: + logging.warning(f"Select VAE Device: cannot retarget VAE, passing through unchanged. ({e})") + return io.NodeOutput(vae) + # Keep VAE wrapper in sync with whatever model the patcher now owns; + # deepclone_multigpu may have produced a fresh first_stage_model. + vae.first_stage_model = vae.patcher.model + vae.device = vae._select_base_device if resolved is None else resolved + return io.NodeOutput(vae) + + +class MultiGPUOptionsNode(io.ComfyNode): + """ + Select the relative speed of GPUs in the special case they have significantly different performance from one another. + + NOTE (not registered yet, see MultiGPUExtension.get_node_list below): + The output GPUOptionsGroup is plumbed through create_multigpu_deepclones() and stored on + model.model_options['multigpu_options'] via GPUOptionsGroup.register(), but the cond + scheduler in comfy/samplers.py (calc_cond_batch_outer_multigpu) does NOT yet consult + relative_speed when distributing conds across devices; it uses a uniform conds_per_device + round-robin via next_available_device(). Before re-enabling this node, wire its + relative_speed into the scheduler (e.g. via comfy.multigpu.load_balance_devices(), + which already implements the proportional split) so the input actually affects work + distribution. + """ + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="MultiGPU_Options", + display_name="MultiGPU Options", + category="advanced/multigpu", + description=cleandoc(cls.__doc__), + inputs=[ + io.Int.Input("device_index", default=0, min=0, max=64), + io.Float.Input("relative_speed", default=1.0, min=0.0, step=0.01), + io.Custom("GPU_OPTIONS").Input("gpu_options", optional=True), + ], + outputs=[ + io.Custom("GPU_OPTIONS").Output(), + ], + ) + + @classmethod + def execute(cls, device_index: int, relative_speed: float, gpu_options: comfy.multigpu.GPUOptionsGroup = None) -> io.NodeOutput: + if not gpu_options: + gpu_options = comfy.multigpu.GPUOptionsGroup() + else: + gpu_options = gpu_options.clone() + + opt = comfy.multigpu.GPUOptions(device_index=device_index, relative_speed=relative_speed) + gpu_options.add(opt) + + return io.NodeOutput(gpu_options) + + +class MultiGPUExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + MultiGPUCFGSplitNode, + SelectModelDeviceNode, + SelectCLIPDeviceNode, + SelectVAEDeviceNode, + # MultiGPUOptionsNode, + ] + + +async def comfy_entrypoint() -> MultiGPUExtension: + return MultiGPUExtension() diff --git a/comfy_extras/nodes_number_convert.py b/comfy_extras/nodes_number_convert.py index e38a33c15..d7e557e95 100644 --- a/comfy_extras/nodes_number_convert.py +++ b/comfy_extras/nodes_number_convert.py @@ -4,7 +4,6 @@ Provides a single node that converts INT, FLOAT, STRING, and BOOL inputs into FLOAT and INT outputs. """ -from __future__ import annotations import math @@ -21,7 +20,7 @@ class NumberConvertNode(io.ComfyNode): return io.Schema( node_id="ComfyNumberConvert", display_name="Convert Number", - category="utils", + category="utilities", search_aliases=[ "int to float", "float to int", "number convert", "int2float", "float2int", "cast", "parse number", diff --git a/comfy_extras/nodes_optimalsteps.py b/comfy_extras/nodes_optimalsteps.py index 5beeaa7db..19629790f 100644 --- a/comfy_extras/nodes_optimalsteps.py +++ b/comfy_extras/nodes_optimalsteps.py @@ -31,7 +31,7 @@ class OptimalStepsScheduler(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="OptimalStepsScheduler", - category="sampling/schedulers", + category="model/sampling/schedulers", inputs=[ io.Combo.Input("model_type", options=["FLUX", "Wan", "Chroma"]), io.Int.Input("steps", default=20, min=3, max=1000), diff --git a/comfy_extras/nodes_pag.py b/comfy_extras/nodes_pag.py index 79fea5f0c..c875e1e06 100644 --- a/comfy_extras/nodes_pag.py +++ b/comfy_extras/nodes_pag.py @@ -15,7 +15,7 @@ class PerturbedAttentionGuidance(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="PerturbedAttentionGuidance", - category="model_patches/unet", + category="model/patch/unet", inputs=[ io.Model.Input("model"), io.Float.Input("scale", default=3.0, min=0.0, max=100.0, step=0.01, round=0.01), diff --git a/comfy_extras/nodes_painter.py b/comfy_extras/nodes_painter.py index e104c8480..df7a0b76a 100644 --- a/comfy_extras/nodes_painter.py +++ b/comfy_extras/nodes_painter.py @@ -1,5 +1,3 @@ -from __future__ import annotations - import hashlib import os diff --git a/comfy_extras/nodes_pid.py b/comfy_extras/nodes_pid.py new file mode 100644 index 000000000..71855254e --- /dev/null +++ b/comfy_extras/nodes_pid.py @@ -0,0 +1,63 @@ +"""PiD (Pixel Diffusion Decoder) node""" + +import torch +from typing_extensions import override + +import node_helpers +import comfy.latent_formats +from comfy_api.latest import ComfyExtension, io + + +class PiDConditioning(io.ComfyNode): + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="PiDConditioning", + display_name="PiD Conditioning", + category="advanced/conditioning", + description=( + "Attaches a latent and a degrade_sigma scalar to a CONDITIONING for PiD decoding/upscaling" + ), + inputs=[ + io.Conditioning.Input("positive"), + io.Latent.Input("latent", tooltip="latent (from VAEEncode or a KSampler)."), + io.Combo.Input("latent_format", options=["flux", "sd3", "sdxl", "qwenimage"], default="flux", + tooltip="Flux1 (16-ch) and Flux2 (128-ch) latents are auto-detected from channel dim under 'flux'. For SD3 (16-ch), SDXL (4-ch), or QwenImage (16-ch), select manually."), + io.Float.Input( + "degrade_sigma", default=0.0, min=0.0, max=1.0, step=0.01, + tooltip="0 = clean latent. Increase to denoise corrupted latent outputs.", + ), + ], + outputs=[io.Conditioning.Output()], + ) + + @classmethod + def execute(cls, positive, latent, latent_format: str, degrade_sigma: float) -> io.NodeOutput: + samples = latent["samples"] + if latent_format == "flux": + fmt_cls = comfy.latent_formats.Flux2 if samples.shape[1] == 128 else comfy.latent_formats.Flux + elif latent_format == "sd3": + fmt_cls = comfy.latent_formats.SD3 + elif latent_format == "sdxl": + fmt_cls = comfy.latent_formats.SDXL + elif latent_format == "qwenimage": + fmt_cls = comfy.latent_formats.Wan21 + else: + raise ValueError(f"Unknown latent_format: {latent_format}") + lq_latent = fmt_cls().process_in(samples) + if lq_latent.ndim == 5: + lq_latent = lq_latent[:, :, 0] + sigma_t = torch.tensor([float(degrade_sigma)], dtype=torch.float32) + return io.NodeOutput(node_helpers.conditioning_set_values( + positive, {"lq_latent": lq_latent, "degrade_sigma": sigma_t}, + )) + + +class PiDExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [PiDConditioning] + + +async def comfy_entrypoint() -> PiDExtension: + return PiDExtension() diff --git a/comfy_extras/nodes_post_processing.py b/comfy_extras/nodes_post_processing.py index a25db277c..3e440433e 100644 --- a/comfy_extras/nodes_post_processing.py +++ b/comfy_extras/nodes_post_processing.py @@ -616,7 +616,7 @@ class BatchLatentsNode(io.ComfyNode): node_id="BatchLatentsNode", search_aliases=["combine latents", "stack latents", "merge latents"], display_name="Batch Latents", - category="latent", + category="model/latent", inputs=[ io.Autogrow.Input("latents", template=autogrow_template) ], diff --git a/comfy_extras/nodes_preview_any.py b/comfy_extras/nodes_preview_any.py index 17e25d514..1070a69d0 100644 --- a/comfy_extras/nodes_preview_any.py +++ b/comfy_extras/nodes_preview_any.py @@ -16,7 +16,7 @@ class PreviewAny(): FUNCTION = "main" OUTPUT_NODE = True - CATEGORY = "utils" + CATEGORY = "utilities" SEARCH_ALIASES = ["show output", "inspect", "debug", "print value", "show text"] def main(self, source=None): diff --git a/comfy_extras/nodes_primitive.py b/comfy_extras/nodes_primitive.py index 33373266b..c44b09098 100644 --- a/comfy_extras/nodes_primitive.py +++ b/comfy_extras/nodes_primitive.py @@ -11,7 +11,7 @@ class String(io.ComfyNode): node_id="PrimitiveString", search_aliases=["text", "string", "text box", "prompt"], display_name="Text String", - category="utils/primitive", + category="utilities/primitive", inputs=[ io.String.Input("value"), ], @@ -30,7 +30,7 @@ class StringMultiline(io.ComfyNode): node_id="PrimitiveStringMultiline", search_aliases=["text", "string", "text multiline", "string multiline", "text box", "prompt"], display_name="Text String (Multiline)", - category="utils/primitive", + category="utilities/primitive", essentials_category="Basics", inputs=[ io.String.Input("value", multiline=True), @@ -49,7 +49,7 @@ class Int(io.ComfyNode): return io.Schema( node_id="PrimitiveInt", display_name="Int", - category="utils/primitive", + category="utilities/primitive", inputs=[ io.Int.Input("value", min=-sys.maxsize, max=sys.maxsize, control_after_generate=io.ControlAfterGenerate.fixed), ], @@ -67,7 +67,7 @@ class Float(io.ComfyNode): return io.Schema( node_id="PrimitiveFloat", display_name="Float", - category="utils/primitive", + category="utilities/primitive", inputs=[ io.Float.Input("value", min=-sys.maxsize, max=sys.maxsize, step=0.1), ], @@ -85,7 +85,7 @@ class Boolean(io.ComfyNode): return io.Schema( node_id="PrimitiveBoolean", display_name="Boolean", - category="utils/primitive", + category="utilities/primitive", inputs=[ io.Boolean.Input("value"), ], diff --git a/comfy_extras/nodes_qwen.py b/comfy_extras/nodes_qwen.py index fde8fac9a..5b92814a4 100644 --- a/comfy_extras/nodes_qwen.py +++ b/comfy_extras/nodes_qwen.py @@ -112,7 +112,7 @@ class EmptyQwenImageLayeredLatentImage(io.ComfyNode): return io.Schema( node_id="EmptyQwenImageLayeredLatentImage", display_name="Empty Qwen Image Layered Latent", - category="latent/qwen", + category="model/latent/qwen", inputs=[ io.Int.Input("width", default=640, min=16, max=nodes.MAX_RESOLUTION, step=16), io.Int.Input("height", default=640, min=16, max=nodes.MAX_RESOLUTION, step=16), diff --git a/comfy_extras/nodes_rebatch.py b/comfy_extras/nodes_rebatch.py index 5f4e82aef..2185385f0 100644 --- a/comfy_extras/nodes_rebatch.py +++ b/comfy_extras/nodes_rebatch.py @@ -10,7 +10,7 @@ class LatentRebatch(io.ComfyNode): return io.Schema( node_id="RebatchLatents", display_name="Rebatch Latents", - category="latent/batch", + category="model/latent/batch", is_input_list=True, inputs=[ io.Latent.Input("latents"), diff --git a/comfy_extras/nodes_resolution.py b/comfy_extras/nodes_resolution.py index 520b4067e..083e47ae4 100644 --- a/comfy_extras/nodes_resolution.py +++ b/comfy_extras/nodes_resolution.py @@ -1,4 +1,3 @@ -from __future__ import annotations import math from enum import Enum from typing_extensions import override @@ -7,24 +6,24 @@ from comfy_api.latest import ComfyExtension, io class AspectRatio(str, Enum): SQUARE = "1:1 (Square)" + PHOTO_V = "2:3 (Portrait Photo)" PHOTO_H = "3:2 (Photo)" + STANDARD_V = "3:4 (Portrait Standard)" STANDARD_H = "4:3 (Standard)" + WIDESCREEN_V = "9:16 (Portrait Widescreen)" WIDESCREEN_H = "16:9 (Widescreen)" ULTRAWIDE_H = "21:9 (Ultrawide)" - PHOTO_V = "2:3 (Portrait Photo)" - STANDARD_V = "3:4 (Portrait Standard)" - WIDESCREEN_V = "9:16 (Portrait Widescreen)" ASPECT_RATIOS: dict[AspectRatio, tuple[int, int]] = { AspectRatio.SQUARE: (1, 1), + AspectRatio.PHOTO_V: (2, 3), AspectRatio.PHOTO_H: (3, 2), + AspectRatio.STANDARD_V: (3, 4), AspectRatio.STANDARD_H: (4, 3), + AspectRatio.WIDESCREEN_V: (9, 16), AspectRatio.WIDESCREEN_H: (16, 9), AspectRatio.ULTRAWIDE_H: (21, 9), - AspectRatio.PHOTO_V: (2, 3), - AspectRatio.STANDARD_V: (3, 4), - AspectRatio.WIDESCREEN_V: (9, 16), } @@ -36,7 +35,7 @@ class ResolutionSelector(io.ComfyNode): return io.Schema( node_id="ResolutionSelector", display_name="Resolution Selector", - category="utils", + category="utilities", description="Calculate width and height from aspect ratio and megapixel target. Useful for setting up Empty Latent Image dimensions.", inputs=[ io.Combo.Input( @@ -51,26 +50,35 @@ class ResolutionSelector(io.ComfyNode): min=0.1, max=16.0, step=0.1, - tooltip="Target total megapixels. 1.0 MP ≈ 1024×1024 for square.", + tooltip="Target total megapixels. 1.0 MP ≈ 1024x1024 for square.", + ), + io.Int.Input( + id="multiple", + default=8, + min=8, + max=128, + step=4, + tooltip="Nearest multiple of the result to set the selected resolution to.", + advanced=True, ), ], outputs=[ io.Int.Output( - "width", tooltip="Calculated width in pixels (multiple of 8)." + "width", tooltip="Calculated width in pixels multiplied by the selected multiple." ), io.Int.Output( - "height", tooltip="Calculated height in pixels (multiple of 8)." + "height", tooltip="Calculated height in pixels multiplied by the selected multiple." ), ], ) @classmethod - def execute(cls, aspect_ratio: str, megapixels: float) -> io.NodeOutput: + def execute(cls, aspect_ratio: str, megapixels: float, multiple: int) -> io.NodeOutput: w_ratio, h_ratio = ASPECT_RATIOS[aspect_ratio] total_pixels = megapixels * 1024 * 1024 scale = math.sqrt(total_pixels / (w_ratio * h_ratio)) - width = round(w_ratio * scale / 8) * 8 - height = round(h_ratio * scale / 8) * 8 + width = round(w_ratio * scale / multiple) * multiple + height = round(h_ratio * scale / multiple) * multiple return io.NodeOutput(width, height) diff --git a/comfy_extras/nodes_rope.py b/comfy_extras/nodes_rope.py index 918ddc02b..808eee29b 100644 --- a/comfy_extras/nodes_rope.py +++ b/comfy_extras/nodes_rope.py @@ -7,7 +7,7 @@ class ScaleROPE(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="ScaleROPE", - category="advanced/model_patches", + category="model/patch", description="Scale and shift the ROPE of the model.", is_experimental=True, inputs=[ diff --git a/comfy_extras/nodes_save_3d.py b/comfy_extras/nodes_save_3d.py index 8b5e1b8e0..403f268d4 100644 --- a/comfy_extras/nodes_save_3d.py +++ b/comfy_extras/nodes_save_3d.py @@ -19,7 +19,7 @@ from comfy.cli_args import args from comfy_api.latest import ComfyExtension, IO, Types -def pack_variable_mesh_batch(vertices, faces, colors=None, uvs=None, texture=None): +def pack_variable_mesh_batch(vertices, faces, colors=None, uvs=None, texture=None, unlit=False): # Pack lists of (Nᵢ, *) vertex/face/color/uv tensors into padded batched tensors, # stashing per-item lengths as runtime attrs so consumers can recover the real slice. # colors and uvs are 1:1 with vertices, so they're padded to max_vertices and read with vertex_counts. @@ -57,7 +57,7 @@ def pack_variable_mesh_batch(vertices, faces, colors=None, uvs=None, texture=Non return Types.MESH(packed_vertices, packed_faces, uvs=packed_uvs, vertex_colors=packed_colors, texture=texture, - vertex_counts=vertex_counts, face_counts=face_counts) + vertex_counts=vertex_counts, face_counts=face_counts, unlit=unlit) def get_mesh_batch_item(mesh, index): @@ -81,7 +81,7 @@ def get_mesh_batch_item(mesh, index): def save_glb(vertices, faces, filepath, metadata=None, uvs=None, vertex_colors=None, texture_image=None, - metallic_roughness_image=None): + metallic_roughness_image=None, unlit=False): """ Save PyTorch tensor vertices and faces as a GLB file without external dependencies. @@ -249,44 +249,55 @@ def save_glb(vertices, faces, filepath, metadata=None, textures = [] samplers = [] materials = [] - pbr = { - "metallicFactor": 0.0, - "roughnessFactor": 0.5, - "baseColorFactor": [0.22, 0.22, 0.22, 1.0], - } - - if texture_png_bytes is not None and "TEXCOORD_0" in primitive_attributes: - buffer_views.append({ - "buffer": 0, - "byteOffset": texture_byte_offset, - "byteLength": len(texture_buffer), + extensions_used = [] + if unlit and texture_png_bytes is None: + # Flat, light-independent shading (KHR_materials_unlit): COLOR_0 is shown as-is, matching how a + # gaussian splat renders (emissive). Without this the viewer lights the mesh and washes the colours. + materials.append({ + "pbrMetallicRoughness": {"baseColorFactor": [1.0, 1.0, 1.0, 1.0], "metallicFactor": 0.0, "roughnessFactor": 1.0}, + "extensions": {"KHR_materials_unlit": {}}, + "doubleSided": True, }) - images.append({"bufferView": len(buffer_views) - 1, "mimeType": "image/png"}) - samplers.append({"magFilter": 9729, "minFilter": 9729, "wrapS": 33071, "wrapT": 33071}) - textures.append({"source": len(images) - 1, "sampler": 0}) - pbr["baseColorTexture"] = {"index": len(textures) - 1, "texCoord": 0} - - if mr_png_bytes is not None and "TEXCOORD_0" in primitive_attributes: - buffer_views.append({ - "buffer": 0, - "byteOffset": mr_byte_offset, - "byteLength": len(mr_buffer), - }) - images.append({"bufferView": len(buffer_views) - 1, "mimeType": "image/png"}) - if not samplers: + extensions_used.append("KHR_materials_unlit") + primitive["material"] = 0 + else: + pbr = { + "metallicFactor": 0.0, + "roughnessFactor": 0.5, + "baseColorFactor": [0.22, 0.22, 0.22, 1.0], + } + if texture_png_bytes is not None and "TEXCOORD_0" in primitive_attributes: + buffer_views.append({ + "buffer": 0, + "byteOffset": texture_byte_offset, + "byteLength": len(texture_buffer), + }) + images.append({"bufferView": len(buffer_views) - 1, "mimeType": "image/png"}) samplers.append({"magFilter": 9729, "minFilter": 9729, "wrapS": 33071, "wrapT": 33071}) - textures.append({"source": len(images) - 1, "sampler": 0}) - pbr["metallicRoughnessTexture"] = {"index": len(textures) - 1, "texCoord": 0} - # When a metallicRoughness texture is present, the factors scale it; use 1.0 - # so the texture values pass through unchanged (glTF convention). - pbr["metallicFactor"] = 1.0 - pbr["roughnessFactor"] = 1.0 + textures.append({"source": len(images) - 1, "sampler": 0}) + pbr["baseColorTexture"] = {"index": len(textures) - 1, "texCoord": 0} - materials.append({ - "pbrMetallicRoughness": pbr, - "doubleSided": True, - }) - primitive["material"] = 0 + if mr_png_bytes is not None and "TEXCOORD_0" in primitive_attributes: + buffer_views.append({ + "buffer": 0, + "byteOffset": mr_byte_offset, + "byteLength": len(mr_buffer), + }) + images.append({"bufferView": len(buffer_views) - 1, "mimeType": "image/png"}) + if not samplers: + samplers.append({"magFilter": 9729, "minFilter": 9729, "wrapS": 33071, "wrapT": 33071}) + textures.append({"source": len(images) - 1, "sampler": 0}) + pbr["metallicRoughnessTexture"] = {"index": len(textures) - 1, "texCoord": 0} + # When a metallicRoughness texture is present, the factors scale it; use 1.0 + # so the texture values pass through unchanged (glTF convention). + pbr["metallicFactor"] = 1.0 + pbr["roughnessFactor"] = 1.0 + + materials.append({ + "pbrMetallicRoughness": pbr, + "doubleSided": True, + }) + primitive["material"] = 0 gltf = { "asset": {"version": "2.0", "generator": "ComfyUI"}, @@ -306,6 +317,8 @@ def save_glb(vertices, faces, filepath, metadata=None, gltf["textures"] = textures if materials: gltf["materials"] = materials + if extensions_used: + gltf["extensionsUsed"] = extensions_used if metadata: gltf["asset"]["extras"] = metadata @@ -359,6 +372,12 @@ class SaveGLB(IO.ComfyNode): IO.File3DFBX, IO.File3DSTL, IO.File3DUSDZ, + IO.File3DPLY, + IO.File3DSPLAT, + IO.File3DSPZ, + IO.File3DKSPLAT, + IO.File3DSplatAny, + IO.File3DPointCloudAny, IO.File3DAny, ], tooltip="Mesh or 3D file to save", @@ -424,6 +443,7 @@ class SaveGLB(IO.ComfyNode): vertex_colors=v_colors, texture_image=tex_img, metallic_roughness_image=mr_img, + unlit=getattr(mesh, "unlit", False), ) results.append({ "filename": f, diff --git a/comfy_extras/nodes_scail.py b/comfy_extras/nodes_scail.py new file mode 100644 index 000000000..a740442de --- /dev/null +++ b/comfy_extras/nodes_scail.py @@ -0,0 +1,321 @@ +"""SCAIL / SCAIL-2 nodes: the WanSCAILToVideo conditioning node and the SAM3 +preprocessing that turns video tracks into the bundle the SCAIL-2 model consumes.""" + +from typing_extensions import override + +import torch +import torch.nn.functional as F + +import nodes +import node_helpers +import comfy.model_management +import comfy.utils +from comfy_api.latest import ComfyExtension, io +from comfy.ldm.sam3.tracker import unpack_masks + +SAM3TrackData = io.Custom("SAM3_TRACK_DATA") + + +# Model was trained on these exact colors; deviating degrades multi-identity quality. +DEFAULT_PALETTE = [ + (0.0, 0.0, 1.0), # Blue + (1.0, 0.0, 0.0), # Red + (0.0, 1.0, 0.0), # Green + (1.0, 0.0, 1.0), # Magenta + (0.0, 1.0, 1.0), # Cyan + (1.0, 1.0, 0.0), # Yellow +] + + +def _unpack(track_data): + packed = track_data["packed_masks"] + if packed is None or packed.shape[1] == 0: + return None + return unpack_masks(packed) + + +def _first_frame_cx_area(masks_bool): + first = masks_bool[0].float() + H, W = first.shape[-2], first.shape[-1] + n_pixels = H * W + grid_x = torch.arange(W, device=first.device, dtype=first.dtype).view(1, W) + area = first.sum(dim=(-1, -2)).clamp_(min=1) + cx = (first * grid_x).sum(dim=(-1, -2)) / area + return (cx / W).tolist(), (area / n_pixels).tolist() + + +def _subset_track_data(track_data, obj_indices): + out = dict(track_data) + packed = track_data["packed_masks"] + if packed is None or not obj_indices: + out["packed_masks"] = None + if "scores" in out: + out["scores"] = [] + return out + out["packed_masks"] = packed[:, obj_indices].contiguous() + scores = track_data.get("scores") + if scores is not None: + out["scores"] = [scores[i] for i in obj_indices if i < len(scores)] + return out + + +def _render_colored_masks(track_data, background="black"): + packed = track_data["packed_masks"] + H, W = track_data["orig_size"] + device = comfy.model_management.intermediate_device() + dtype = comfy.model_management.intermediate_dtype() + bg_rgb = (1.0, 1.0, 1.0) if background.startswith("white") else (0.0, 0.0, 0.0) + if packed is None or packed.shape[1] == 0: + T = track_data.get("n_frames", 1) if packed is None else packed.shape[0] + out = torch.empty(T, H, W, 3, device=device, dtype=dtype) + out[..., 0], out[..., 1], out[..., 2] = bg_rgb[0], bg_rgb[1], bg_rgb[2] + return out + T, N_obj = packed.shape[0], packed.shape[1] + colors = torch.tensor( + [DEFAULT_PALETTE[i % len(DEFAULT_PALETTE)] for i in range(N_obj)], + device=device, dtype=dtype, + ) + masks_full = unpack_masks(packed.to(device)).float() + Hm, Wm = masks_full.shape[-2], masks_full.shape[-1] + masks_full = F.interpolate( + masks_full.view(T * N_obj, 1, Hm, Wm), size=(H, W), mode="nearest" + ).view(T, N_obj, H, W) > 0.5 + any_mask = masks_full.any(dim=1) + obj_idx_map = masks_full.to(torch.uint8).argmax(dim=1) + color_overlay = colors[obj_idx_map] + bg_tensor = torch.tensor(bg_rgb, device=device, dtype=color_overlay.dtype).view(1, 1, 1, 3) + return torch.where(any_mask.unsqueeze(-1), color_overlay, bg_tensor.expand_as(color_overlay)) + + +def _extract_mask_to_28ch(rgb_video): + """Colored RGB mask (T, H, W, 3) in [0, 1] -> SCAIL-2 28-channel binary latent + (1, T_lat, 28, H_lat, W_lat). 7 per-color binary channels (white/r/g/b/y/m/c) + threshold-extracted at 225/255, 8x spatial downsample, 4-frame temporal stacking.""" + T, H, W, _ = rgb_video.shape + _ON_THRESH = 225.0 / 255.0 + mask = rgb_video.movedim(-1, 1).float() + R = (mask[:, 0:1] > _ON_THRESH).float() + G = (mask[:, 1:2] > _ON_THRESH).float() + B = (mask[:, 2:3] > _ON_THRESH).float() + nR, nG, nB = 1 - R, 1 - G, 1 - B + binary_7ch = torch.cat([ + R * G * B, # white + R * nG * nB, # red + nR * G * nB, # green + nR * nG * B, # blue + R * G * nB, # yellow + R * nG * B, # magenta + nR * G * B, # cyan + ], dim=1) + H_lat, W_lat = H, W + for _ in range(3): + H_lat = (H_lat + 1) // 2 + W_lat = (W_lat + 1) // 2 + binary_7ch = torch.nn.functional.interpolate(binary_7ch, size=(H_lat, W_lat), mode='area') + T_latent = (T - 1) // 4 + 1 + padded = torch.cat([binary_7ch[:1].repeat(4, 1, 1, 1), binary_7ch[1:]], dim=0) + out = padded.view(T_latent, 28, H_lat, W_lat) + return out.unsqueeze(0) + + +class WanSCAILToVideo(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="WanSCAILToVideo", + category="model/conditioning/video_models", + inputs=[ + io.Conditioning.Input("positive"), + io.Conditioning.Input("negative"), + io.Vae.Input("vae"), + io.Int.Input("width", default=512, min=32, max=nodes.MAX_RESOLUTION, step=32), + io.Int.Input("height", default=896, min=32, max=nodes.MAX_RESOLUTION, step=32), + io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4), + io.Int.Input("batch_size", default=1, min=1, max=4096), + io.Image.Input("pose_video", optional=True, tooltip="Video used for pose conditioning. Will be downscaled to half the resolution of the main video."), + io.Image.Input("pose_video_mask", optional=True, tooltip="SCAIL-2 only. Colored per-identity SAM3 mask video at the same resolution as pose_video."), + io.Boolean.Input("replacement_mode", default=False, optional=True, tooltip="SCAIL-2 only. False = Animation Mode (pose_video_mask should have black background). True = Replacement Mode (pose_video_mask should have white background)."), + io.Float.Input("pose_strength", default=1.0, min=0.0, max=10.0, step=0.01, tooltip="Strength of the pose latent."), + io.Float.Input("pose_start", default=0.0, min=0.0, max=1.0, step=0.01, tooltip="Start step of the pose conditioning."), + io.Float.Input("pose_end", default=1.0, min=0.0, max=1.0, step=0.01, tooltip="End step of the pose conditioning."), + io.Image.Input("reference_image", optional=True, tooltip="Reference image, for multiple references composite all on single image."), + io.Image.Input("reference_image_mask", optional=True, tooltip="SCAIL-2 only. Colored reference mask at the same resolution as reference_image."), + io.ClipVisionOutput.Input("clip_vision_output", optional=True, tooltip="CLIP vision features for conditioning. Model is trained with stretch resize to aspect ratio."), + io.Int.Input("video_frame_offset", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1, tooltip="Cumulative output frame this chunk begins at. Wire from the previous chunk's video_frame_offset output."), + io.Int.Input("previous_frame_count", default=5, min=1, max=nodes.MAX_RESOLUTION, step=4, tooltip="Tail frames of previous_frames to anchor. SCAIL-2 trained at 5 (81-frame chunks, 76-frame step)."), + io.Image.Input("previous_frames", optional=True, tooltip="SCAIL-2 only. Full decoded output of the previous chunk. Only the last previous_frame_count are used as the extension anchor."), + ], + outputs=[ + io.Conditioning.Output(display_name="positive"), + io.Conditioning.Output(display_name="negative"), + io.Latent.Output(display_name="latent", tooltip="Empty latent of the generation size."), + io.Int.Output(display_name="video_frame_offset", tooltip="Adjusted offset + length. Wire into the next chunk."), + ], + is_experimental=True, + ) + + @classmethod + def execute(cls, positive, negative, vae, width, height, length, batch_size, pose_strength, pose_start, pose_end, + video_frame_offset, previous_frame_count, replacement_mode=False, reference_image=None, clip_vision_output=None, pose_video=None, + pose_video_mask=None, reference_image_mask=None, previous_frames=None) -> io.NodeOutput: + latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device()) + noise_mask = None + + ref_mask_flag = not replacement_mode + positive = node_helpers.conditioning_set_values(positive, {"ref_mask_flag": ref_mask_flag}) + negative = node_helpers.conditioning_set_values(negative, {"ref_mask_flag": ref_mask_flag}) + + prev_trimmed = None + if previous_frames is not None and previous_frames.shape[0] > 0: + prev_trimmed = previous_frames[-previous_frame_count:] + video_frame_offset -= prev_trimmed.shape[0] + video_frame_offset = max(0, video_frame_offset) + + ref_latent = None + if reference_image is not None: + reference_image = comfy.utils.common_upscale(reference_image[:1].movedim(-1, 1), width, height, "bicubic", "center").movedim(1, -1) + # Replacement Mode: composite ref on black bg using reference_image_mask as alpha matte + if replacement_mode and reference_image_mask is not None: + rm = comfy.utils.common_upscale(reference_image_mask[:1].movedim(-1, 1), width, height, "nearest-exact", "center").movedim(1, -1) + is_char = (rm[..., :3].max(dim=-1, keepdim=True).values > 0.1).to(reference_image.dtype) + reference_image = reference_image * is_char + ref_latent = vae.encode(reference_image[:, :, :, :3]) + + if ref_latent is not None: + positive = node_helpers.conditioning_set_values(positive, {"reference_latents": [ref_latent]}, append=True) + negative = node_helpers.conditioning_set_values(negative, {"reference_latents": [ref_latent]}, append=True) + + if clip_vision_output is not None: + positive = node_helpers.conditioning_set_values(positive, {"clip_vision_output": clip_vision_output}) + negative = node_helpers.conditioning_set_values(negative, {"clip_vision_output": clip_vision_output}) + + if pose_video is not None: + if pose_video.shape[0] <= video_frame_offset: + pose_video = None + else: + pose_video = pose_video[video_frame_offset:] + if pose_video_mask is not None: + if pose_video_mask.shape[0] <= video_frame_offset: + pose_video_mask = None + else: + pose_video_mask = pose_video_mask[video_frame_offset:] + + # Truncate pose+mask jointly to the shorter of the two, capped at length. + ts = [v.shape[0] for v in (pose_video, pose_video_mask) if v is not None] + if ts: + T_kept = ((min(min(ts), length) - 1) // 4) * 4 + 1 + if pose_video is not None: + pose_video = pose_video[:T_kept] + if pose_video_mask is not None: + pose_video_mask = pose_video_mask[:T_kept] + + if pose_video is not None: + pose_video = comfy.utils.common_upscale(pose_video[:length].movedim(-1, 1), width // 2, height // 2, "area", "center").movedim(1, -1) + pose_video_latent = vae.encode(pose_video[:, :, :, :3]) * pose_strength + positive = node_helpers.conditioning_set_values_with_timestep_range(positive, {"pose_video_latent": pose_video_latent}, pose_start, pose_end) + negative = node_helpers.conditioning_set_values_with_timestep_range(negative, {"pose_video_latent": pose_video_latent}, pose_start, pose_end) + + if pose_video_mask is not None: + mask_video_hw = comfy.utils.common_upscale(pose_video_mask[:length].movedim(-1, 1), width // 2, height // 2, "area", "center").movedim(1, -1) + driving_mask_28ch = _extract_mask_to_28ch(mask_video_hw) + positive = node_helpers.conditioning_set_values(positive, {"driving_mask_28ch": driving_mask_28ch}) + negative = node_helpers.conditioning_set_values(negative, {"driving_mask_28ch": driving_mask_28ch}) + + if reference_image_mask is not None: + ref_mask_hw = comfy.utils.common_upscale(reference_image_mask[:1].movedim(-1, 1), width, height, "bicubic", "center").movedim(1, -1) + ref_mask_1f = _extract_mask_to_28ch(ref_mask_hw) + zeros = torch.zeros((1, latent.shape[2], 28, ref_mask_1f.shape[-2], ref_mask_1f.shape[-1]), device=ref_mask_1f.device, dtype=ref_mask_1f.dtype) + ref_mask_28ch = torch.cat([ref_mask_1f, zeros], dim=1) + positive = node_helpers.conditioning_set_values(positive, {"ref_mask_28ch": ref_mask_28ch}) + negative = node_helpers.conditioning_set_values(negative, {"ref_mask_28ch": ref_mask_28ch}) + + if prev_trimmed is not None: + pf = comfy.utils.common_upscale(prev_trimmed.movedim(-1, 1), width, height, "bicubic", "center").movedim(1, -1) + prev_latent = vae.encode(pf[:, :, :, :3]) + prev_latent_frames = min(prev_latent.shape[2], latent.shape[2]) + latent[:, :, :prev_latent_frames] = prev_latent[:, :, :prev_latent_frames].to(latent.dtype) + noise_mask = torch.ones((1, 1, latent.shape[2], latent.shape[-2], latent.shape[-1]), device=latent.device, dtype=latent.dtype) + noise_mask[:, :, :prev_latent_frames] = 0.0 + + out_latent = {"samples": latent} + if noise_mask is not None: + out_latent["noise_mask"] = noise_mask + return io.NodeOutput(positive, negative, out_latent, video_frame_offset + length) + + +class SCAIL2ColoredMask(io.ComfyNode): + """Render SAM3 tracks for the driving pose video and (optionally) the reference + image into the two colored masks WanSCAILToVideo consumes. Shared `sort_by` + across both outputs guarantees identity K maps to the same color on both + sides, for multi-person workflow consistency. + reference_image_mask is always rendered black-bg (model convention) + pose_video_mask bg follows replacement_mode: black = Animation Mode, white = Replacement Mode + """ + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="SCAIL2ColoredMask", + display_name="Create SCAIL-2 Colored Mask", + category="conditioning/video_models/scail", + inputs=[ + SAM3TrackData.Input("driving_track_data", tooltip="SAM3 track of the driving pose video. Will be rendered into the pose_video_mask output."), + SAM3TrackData.Input("ref_track_data", optional=True, + tooltip="SAM3 track of the reference image."), + io.String.Input("object_indices", default="", + tooltip="Comma-separated list of person indices to include (e.g. '0,2,3'). Applied to both reference and pose video masks. Empty = all."), + io.Combo.Input("sort_by", options=["none", "left_to_right", "area"], default="left_to_right", + tooltip="Order in which palette colors are assigned to the tracked objects (applied to both reference and pose video so each identity keeps the same color). left_to_right = leftmost object (by first-frame centroid) gets the first color; area = biggest object (by first-frame mask area) gets the first color; none = keep SAM3's order."), + io.Boolean.Input("replacement_mode", default=False, + tooltip="False = mask_video has black bg (Animation Mode). True = white bg (Replacement Mode). Set the matching replacement_mode on WanSCAILToVideo. reference_image_mask is always black-bg regardless."), + ], + outputs=[ + io.Image.Output("pose_video_mask"), + io.Image.Output("reference_image_mask"), + ], + is_experimental=True, + ) + + @classmethod + def execute(cls, driving_track_data, object_indices, sort_by, replacement_mode, ref_track_data=None): + def _prep(td): + masks_bool = _unpack(td) + if sort_by != "none" and masks_bool is not None: + cx, area = _first_frame_cx_area(masks_bool) + if sort_by == "left_to_right": + order = sorted(range(len(cx)), key=lambda i: cx[i]) + else: # "area" + order = sorted(range(len(area)), key=lambda i: -area[i]) + td = _subset_track_data(td, order) + if object_indices.strip(): + indices = [int(i.strip()) for i in object_indices.split(",") if i.strip().isdigit()] + packed = td.get("packed_masks") + n_obj = packed.shape[1] if packed is not None else 0 + indices = [i for i in indices if 0 <= i < n_obj] + td = _subset_track_data(td, indices) + return td + + drv = _prep(driving_track_data) + mask_video = _render_colored_masks(drv, "white" if replacement_mode else "black") + + if ref_track_data is not None: + ref = _prep(ref_track_data) + reference_image_mask = _render_colored_masks(ref, "black") + else: + H, W = drv["orig_size"] + reference_image_mask = torch.zeros(1, H, W, 3, device=comfy.model_management.intermediate_device(), dtype=comfy.model_management.intermediate_dtype()) + + return io.NodeOutput(mask_video, reference_image_mask) + + +class SCAILExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + WanSCAILToVideo, + SCAIL2ColoredMask, + ] + + +async def comfy_entrypoint() -> SCAILExtension: + return SCAILExtension() diff --git a/comfy_extras/nodes_sd3.py b/comfy_extras/nodes_sd3.py index 6655c1ba7..38cbf117b 100644 --- a/comfy_extras/nodes_sd3.py +++ b/comfy_extras/nodes_sd3.py @@ -41,7 +41,7 @@ class EmptySD3LatentImage(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="EmptySD3LatentImage", - category="latent/sd3", + category="model/latent/sd3", inputs=[ io.Int.Input("width", default=1024, min=16, max=nodes.MAX_RESOLUTION, step=16), io.Int.Input("height", default=1024, min=16, max=nodes.MAX_RESOLUTION, step=16), @@ -113,7 +113,7 @@ class ControlNetApplySD3(io.ComfyNode): return io.Schema( node_id="ControlNetApplySD3", display_name="Apply Controlnet with VAE", - category="conditioning/controlnet", + category="model/conditioning/controlnet", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), diff --git a/comfy_extras/nodes_sdupscale.py b/comfy_extras/nodes_sdupscale.py index 5877719d3..ea283e971 100644 --- a/comfy_extras/nodes_sdupscale.py +++ b/comfy_extras/nodes_sdupscale.py @@ -9,7 +9,7 @@ class SD_4XUpscale_Conditioning(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SD_4XUpscale_Conditioning", - category="conditioning/upscale_diffusion", + category="model/conditioning/upscale_diffusion", inputs=[ io.Image.Input("images"), io.Conditioning.Input("positive"), diff --git a/comfy_extras/nodes_stable3d.py b/comfy_extras/nodes_stable3d.py index 829c837a1..8a6e5b726 100644 --- a/comfy_extras/nodes_stable3d.py +++ b/comfy_extras/nodes_stable3d.py @@ -27,7 +27,7 @@ class StableZero123_Conditioning(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="StableZero123_Conditioning", - category="conditioning/3d_models", + category="model/conditioning/3d_models", inputs=[ io.ClipVision.Input("clip_vision"), io.Image.Input("init_image"), @@ -65,7 +65,7 @@ class StableZero123_Conditioning_Batched(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="StableZero123_Conditioning_Batched", - category="conditioning/3d_models", + category="model/conditioning/3d_models", inputs=[ io.ClipVision.Input("clip_vision"), io.Image.Input("init_image"), @@ -112,7 +112,7 @@ class SV3D_Conditioning(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SV3D_Conditioning", - category="conditioning/3d_models", + category="model/conditioning/3d_models", inputs=[ io.ClipVision.Input("clip_vision"), io.Image.Input("init_image"), diff --git a/comfy_extras/nodes_stable_cascade.py b/comfy_extras/nodes_stable_cascade.py index 0dc6c9fcd..e55f248ae 100644 --- a/comfy_extras/nodes_stable_cascade.py +++ b/comfy_extras/nodes_stable_cascade.py @@ -29,7 +29,7 @@ class StableCascade_EmptyLatentImage(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="StableCascade_EmptyLatentImage", - category="latent/stable_cascade", + category="model/latent/stable_cascade", inputs=[ io.Int.Input("width", default=1024, min=256, max=nodes.MAX_RESOLUTION, step=8), io.Int.Input("height", default=1024, min=256, max=nodes.MAX_RESOLUTION, step=8), @@ -58,7 +58,7 @@ class StableCascade_StageC_VAEEncode(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="StableCascade_StageC_VAEEncode", - category="latent/stable_cascade", + category="model/latent/stable_cascade", inputs=[ io.Image.Input("image"), io.Vae.Input("vae"), @@ -93,7 +93,7 @@ class StableCascade_StageB_Conditioning(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="StableCascade_StageB_Conditioning", - category="conditioning/stable_cascade", + category="model/conditioning/stable_cascade", inputs=[ io.Conditioning.Input("conditioning"), io.Latent.Input("stage_c"), diff --git a/comfy_extras/nodes_tomesd.py b/comfy_extras/nodes_tomesd.py index 87bf29b8f..3667fac3a 100644 --- a/comfy_extras/nodes_tomesd.py +++ b/comfy_extras/nodes_tomesd.py @@ -151,7 +151,7 @@ class TomePatchModel(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="TomePatchModel", - category="model_patches/unet", + category="model/patch/unet", inputs=[ io.Model.Input("model"), io.Float.Input("ratio", default=0.3, min=0.0, max=1.0, step=0.01), diff --git a/comfy_extras/nodes_toolkit.py b/comfy_extras/nodes_toolkit.py index 71faf7226..9f709bbe3 100644 --- a/comfy_extras/nodes_toolkit.py +++ b/comfy_extras/nodes_toolkit.py @@ -1,4 +1,3 @@ -from __future__ import annotations from typing_extensions import override from comfy_api.latest import ComfyExtension, io @@ -14,7 +13,7 @@ class CreateList(io.ComfyNode): return io.Schema( node_id="CreateList", display_name="Create List", - category="logic", + category="utilities", is_input_list=True, search_aliases=["Image Iterator", "Text Iterator", "Iterator"], inputs=[io.Autogrow.Input("inputs", template=template_autogrow)], diff --git a/comfy_extras/nodes_train.py b/comfy_extras/nodes_train.py index e9871369b..bb68da6fa 100644 --- a/comfy_extras/nodes_train.py +++ b/comfy_extras/nodes_train.py @@ -15,6 +15,7 @@ import comfy.sampler_helpers import comfy.sd import comfy.utils import comfy.model_management +from comfy.conds import CONDRegular, CONDList from comfy.cli_args import args, PerformanceFeature import comfy_extras.nodes_custom_sampler import folder_paths @@ -120,6 +121,11 @@ def process_cond_list(d, prefix=""): process_cond_list(v, f"{prefix}.{k}") elif isinstance(v, torch.Tensor): d[k] = v.clone() + elif isinstance(v, CONDList): + v.cond = [t.detach() if isinstance(t, torch.Tensor) else t for t in v.cond] + elif isinstance(v, CONDRegular): + if isinstance(v.cond, torch.Tensor): + v.cond = v.cond.detach() elif isinstance(v, (list, tuple)): for index, item in enumerate(v): process_cond_list(item, f"{prefix}.{k}.{index}") @@ -951,7 +957,7 @@ class TrainLoraNode(io.ComfyNode): return io.Schema( node_id="TrainLoraNode", display_name="Train LoRA", - category="training", + category="model/training", is_experimental=True, is_input_list=True, # All inputs become lists inputs=[ @@ -1143,45 +1149,45 @@ class TrainLoraNode(io.ComfyNode): # Process conditioning positive = _process_conditioning(positive) - # Setup model and dtype - mp = model.clone() - use_grad_scaler = False - lora_dtype = node_helpers.string_to_torch_dtype(lora_dtype) - if training_dtype != "none": - dtype = node_helpers.string_to_torch_dtype(training_dtype) - mp.set_model_compute_dtype(dtype) - else: - # Detect model's native dtype for autocast - model_dtype = mp.model.get_dtype() - if model_dtype == torch.float16: - dtype = torch.float16 - # GradScaler only supports float16 gradients, not bfloat16. - # Only enable it when lora params will also be in float16. - if lora_dtype != torch.bfloat16: - use_grad_scaler = True - # Warn about fp16 accumulation instability during training - if PerformanceFeature.Fp16Accumulation in args.fast: - logging.warning( - "WARNING: FP16 model detected with fp16_accumulation enabled. " - "This combination can be numerically unstable during training and may cause NaN values. " - "Suggested fixes: 1) Set training_dtype to 'bf16', or 2) Disable fp16_accumulation (remove from --fast flags)." - ) - else: - # For fp8, bf16, or other dtypes, use bf16 autocast - dtype = torch.bfloat16 - - # Prepare latents and compute counts - latents_dtype = dtype if dtype not in (None,) else torch.bfloat16 - latents, num_images, multi_res = _prepare_latents_and_count( - latents, latents_dtype, bucket_mode - ) - - # Validate and expand conditioning - positive = _validate_and_expand_conditioning(positive, num_images, bucket_mode) - with torch.inference_mode(False): + # Setup model and dtype + mp = model.clone(force_deepcopy=True) + use_grad_scaler = False + lora_dtype = node_helpers.string_to_torch_dtype(lora_dtype) + if training_dtype != "none": + dtype = node_helpers.string_to_torch_dtype(training_dtype) + mp.set_model_compute_dtype(dtype) + else: + # Detect model's native dtype for autocast + model_dtype = mp.model.get_dtype() + if model_dtype == torch.float16: + dtype = torch.float16 + # GradScaler only supports float16 gradients, not bfloat16. + # Only enable it when lora params will also be in float16. + if lora_dtype != torch.bfloat16: + use_grad_scaler = True + # Warn about fp16 accumulation instability during training + if PerformanceFeature.Fp16Accumulation in args.fast: + logging.warning( + "WARNING: FP16 model detected with fp16_accumulation enabled. " + "This combination can be numerically unstable during training and may cause NaN values. " + "Suggested fixes: 1) Set training_dtype to 'bf16', or 2) Disable fp16_accumulation (remove from --fast flags)." + ) + else: + # For fp8, bf16, or other dtypes, use bf16 autocast + dtype = torch.bfloat16 + + # Prepare latents and compute counts + latents_dtype = dtype if dtype not in (None,) else torch.bfloat16 + latents, num_images, multi_res = _prepare_latents_and_count( + latents, latents_dtype, bucket_mode + ) + + # Validate and expand conditioning + positive = _validate_and_expand_conditioning(positive, num_images, bucket_mode) + # Setup models for training - mp.model.requires_grad_(False) + mp.model.requires_grad_(False).train() # Load existing LoRA weights if provided existing_weights, existing_steps = _load_existing_lora(existing_lora) @@ -1309,7 +1315,7 @@ class LoraModelLoader(io.ComfyNode): return io.Schema( node_id="LoraModelLoader", display_name="Load LoRA Model", - category="loaders", + category="model/loaders", is_experimental=True, inputs=[ io.Model.Input( @@ -1405,7 +1411,7 @@ class LossGraphNode(io.ComfyNode): node_id="LossGraphNode", search_aliases=["training chart", "training visualization", "plot loss"], display_name="Plot Loss Graph", - category="training", + category="model/training", is_experimental=True, is_output_node=True, inputs=[ diff --git a/comfy_extras/nodes_triposplat.py b/comfy_extras/nodes_triposplat.py new file mode 100644 index 000000000..1848ad31a --- /dev/null +++ b/comfy_extras/nodes_triposplat.py @@ -0,0 +1,270 @@ +# TripoSplat nodes: image -> 3D gaussian splat + +import logging + +import torch +import torch.nn.functional as F +from typing_extensions import override + +import comfy.model_management +import comfy.nested_tensor +import comfy.patcher_extension +import comfy.utils +from comfy_api.latest import ComfyExtension, IO, Types + + +_Q_TOKEN_LENGTH = 8192 +_LATENT_CHANNELS = 16 +_CAM_CHANNELS = 5 +_DINOV3_MEAN = [0.485, 0.456, 0.406] +_DINOV3_STD = [0.229, 0.224, 0.225] +_NUM_GAUSSIANS_MIN = 32768 +_NUM_GAUSSIANS_MAX = 1048576 + + +def _preprocess(image: torch.Tensor, mask: torch.Tensor, erode_radius: int, size: int) -> torch.Tensor: + # Match original preprocessing: + # resize min side to `size` -> erode alpha -> alpha bbox -> 1.2x square crop -> resize -> composite on black. + rgb = image[..., :3].clamp(0, 1).movedim(-1, 0) # (3, H, W) + alpha = mask.clamp(0, 1)[None] # (1, H, W) + rgba = torch.cat([rgb, alpha], 0)[None] # (1, 4, H, W) + + h, w = rgba.shape[-2:] + s = size / min(w, h) + rgba = comfy.utils.common_upscale(rgba, max(1, round(w * s)), max(1, round(h * s)), "lanczos", "disabled").clamp(0, 1) + + a = rgba[:, 3:4] + if erode_radius > 0: + # min filter over a (2r+1) window == morphological erosion of the alpha matte. + a = -F.max_pool2d(-a, 2 * erode_radius + 1, stride=1, padding=erode_radius) + rgba = torch.cat([rgba[:, :3], a], 1) + + ys, xs = torch.nonzero(a[0, 0] > 0, as_tuple=True) + if xs.numel() == 0: + raise ValueError("TripoSplatPreprocessImage: mask is empty (no foreground pixels).") + x0, x1 = int(xs.min()), int(xs.max()) + y0, y1 = int(ys.min()), int(ys.max()) + cx, cy = (x0 + x1) / 2, (y0 + y1) / 2 + half = max(x1 - x0, y1 - y0) / 2 * 1.2 + left, upper, right, lower = int(cx - half), int(cy - half), int(cx + half), int(cy + half) + + H, W = rgba.shape[-2:] + crop = rgba.new_zeros((1, 4, lower - upper, right - left)) # out-of-bounds stays 0, matching PIL.crop + sx0, sy0, sx1, sy1 = max(left, 0), max(upper, 0), min(right, W), min(lower, H) + if sx1 > sx0 and sy1 > sy0: + crop[:, :, sy0 - upper:sy1 - upper, sx0 - left:sx1 - left] = rgba[:, :, sy0:sy1, sx0:sx1] + + crop = comfy.utils.common_upscale(crop, size, size, "lanczos", "disabled").clamp(0, 1) + out = (crop[:, :3] * crop[:, 3:4])[0].movedim(0, -1) # composite over black == rgb * alpha + return out.unsqueeze(0) # (1, 1024, 1024, 3) + + +class TripoSplatPreprocessImage(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="TripoSplatPreprocessImage", + display_name="TripoSplat Preprocess Image", + category="3d/conditioning", + description="Crop center each image to a square canvas on a black background and add padding.", + inputs=[ + IO.Image.Input("image"), + IO.Mask.Input("mask"), + IO.Int.Input("erode_radius", default=1, min=0, max=16, + tooltip="Erode the alpha matte by this pixel radius before cropping (avoids border bleed)."), + IO.Int.Input("size", default=1024, min=256, max=4096, step=16, + tooltip="Square image size. The model is trained at 1024; other sizes run but are off-distribution."), + ], + outputs=[IO.Image.Output(display_name="image")], + ) + + @classmethod + def execute(cls, image, mask, erode_radius, size) -> IO.NodeOutput: + size = max(16, (int(size) // 16) * 16) # DINOv3 patch / Flux2 VAE stride is 16 + if mask.shape[0] != image.shape[0]: + mask = comfy.utils.repeat_to_batch_size(mask, image.shape[0]) + if tuple(mask.shape[1:]) != tuple(image.shape[1:3]): + mask = F.interpolate(mask[:, None].float(), size=tuple(image.shape[1:3]), mode="bilinear", align_corners=False)[:, 0] + prepared = torch.cat([_preprocess(image[i], mask[i], erode_radius, size) for i in range(image.shape[0])], dim=0) + return IO.NodeOutput(prepared) + + +class TripoSplatConditioning(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="TripoSplatConditioning", + display_name="TripoSplat Conditioning", + category="3d/conditioning", + description="Encode the image with DINOv3 and the Flux2 VAE into TripoSplat positive/negative " + "conditioning, and create the fixed size noise target (latent + camera) for the KSampler", + inputs=[ + IO.ClipVision.Input("clip_vision", tooltip="DINOv3 ViT-H/16+ image encoder"), + IO.Vae.Input("vae", tooltip="Flux2 VAE"), + IO.Image.Input("image"), + ], + outputs=[ + IO.Conditioning.Output(display_name="positive"), + IO.Conditioning.Output(display_name="negative"), + IO.Latent.Output(display_name="latent", tooltip="The fixed size noise target (latent +camera)."), + ], + ) + + @classmethod + def execute(cls, clip_vision, vae, image) -> IO.NodeOutput: + # feature1: DINOv3 token sequence (cls + registers + patches), ImageNet-normalized, with a final non-affine layer norm on top + comfy.model_management.load_model_gpu(clip_vision.patcher) + device = clip_vision.load_device + img = image.movedim(-1, 1).to(device) # (B,3,H,W) in [0,1] + mean = torch.tensor(_DINOV3_MEAN, device=device).view(1, 3, 1, 1) + std = torch.tensor(_DINOV3_STD, device=device).view(1, 3, 1, 1) + img = (img - mean) / std + seq = clip_vision.model(pixel_values=img.float())[0] + feature1 = F.layer_norm(seq.float(), seq.shape[-1:]).to(comfy.model_management.intermediate_device()) + + # Second conditioning: the Flux2 VAE latent of the image, carried as a standard reference_latents entry + ref = vae.encode(image).to(comfy.model_management.intermediate_device()) # (B, 128, H, W) + b = ref.shape[0] + + positive = [[feature1, {"reference_latents": [ref]}]] + negative = [[torch.zeros_like(feature1), {"reference_latents": [torch.zeros_like(ref)]}]] + + # Fixed noise target: the latent is a constant-shape (8192, 16) shape-code + a (1, 5) camera token + dev = comfy.model_management.intermediate_device() + latent_seq = torch.zeros([b, _Q_TOKEN_LENGTH, _LATENT_CHANNELS], device=dev) + camera = torch.zeros([b, 1, _CAM_CHANNELS], device=dev) + samples = comfy.nested_tensor.NestedTensor((latent_seq, camera)) + return IO.NodeOutput(positive, negative, {"samples": samples}) + + +class VAEDecodeTripoSplat(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="VAEDecodeTripoSplat", + display_name="TripoSplat Decode", + category="3d/latent", + description="Decode the sampled TripoSplat latent into a 3D gaussian splat. " + "Modify the number of gaussians to vary the density.", + inputs=[ + IO.Latent.Input("samples"), + IO.Vae.Input("vae", tooltip="TripoSplat VAE decoder"), + IO.Int.Input("num_gaussians", default=262144, min=_NUM_GAUSSIANS_MIN, max=_NUM_GAUSSIANS_MAX, step=32, + tooltip="Number of gaussians to produce (rounded to a multiple of 32). " + "262144 matches the octree's point density; higher oversamples the same points " + "(denser, but no new detail) and costs proportionally more VRAM/time."), + IO.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff, + tooltip="Seeds the octree point sampler (global RNG) for deterministic decodes."), + ], + outputs=[IO.Splat.Output(display_name="splat")], + ) + + @classmethod + def execute(cls, samples, vae, num_gaussians, seed) -> IO.NodeOutput: + s = samples["samples"] + latent = s.unbind()[0] if getattr(s, "is_nested", False) else s # take the latent stream, drop camera + + decoder = vae.first_stage_model + gpp = decoder.gaussians_per_point + n = max(_NUM_GAUSSIANS_MIN, min(_NUM_GAUSSIANS_MAX, int(num_gaussians))) + if n % gpp != 0: + n = round(n / gpp) * gpp + + dtype_size = comfy.model_management.dtype_size(vae.vae_dtype) + hidden = decoder.gs.model_channels + cond_tokens = latent.shape[1] + memory_required = (cond_tokens * 4 + (n // gpp) * 10) * hidden * dtype_size + comfy.model_management.load_models_gpu([vae.patcher], memory_required=memory_required) + latent = latent.to(device=vae.device, dtype=vae.vae_dtype) + generator = torch.Generator(device="cpu").manual_seed(seed) + parts = [g.render_tensors() for g in decoder.decode(latent, num_gaussians=n, generator=generator)] + positions, scales, rotations, opacities, sh = (torch.stack(t) for t in zip(*parts)) + return IO.NodeOutput(Types.SPLAT(positions, scales, rotations, opacities, sh)) + + +class TripoSplatSamplingPreview(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="TripoSplatSamplingPreview", + display_name="TripoSplat Sampling Preview", + category="3d/latent", + description="Patch the TripoSplat model for the standard Ksampler node to show a live decoded " + "gaussian splat preview at each step.", + inputs=[ + IO.Model.Input("model"), + IO.Vae.Input("vae", tooltip="TripoSplat VAE decoder"), + IO.Int.Input("octree_level", default=5, min=2, max=8, advanced=True, + tooltip="Octree depth for the preview decode (lower = cheaper/coarser)."), + IO.Int.Input("num_gaussians", default=16384, min=1024, max=262144, step=32, + tooltip="Number of gaussians to produce for the preview (rounded to a multiple of 32)."), + IO.Float.Input("yaw", default=90.0, min=-360.0, max=360.0, step=1.0, tooltip="Preview camera yaw in degrees.", advanced=True,), + IO.Float.Input("pitch", default=15.0, min=-89.0, max=89.0, step=1.0, tooltip="Preview camera pitch in degrees.", advanced=True,), + IO.Int.Input("point_size", default=3, min=1, max=16, + tooltip="Maximum splat radius in pixels. Each gaussian is sized from its scale and capped here; " + "lower = finer/pointier, higher = chunkier."), + ], + outputs=[IO.Model.Output()], + ) + + @classmethod + def execute(cls, model, vae, octree_level, num_gaussians, yaw, pitch, point_size) -> IO.NodeOutput: + from comfy.ldm.triposplat.preview import decode_x0_to_image + cfg = {"gaussians": num_gaussians, "level": octree_level, "yaw": yaw, "pitch": pitch, + "point_size": point_size} + + fsm = vae.first_stage_model + cond_tokens = model.model.diffusion_model.q_token_length + memory_required = (cond_tokens * 4 + (num_gaussians // fsm.gaussians_per_point) * 10) * fsm.gs.model_channels * comfy.model_management.dtype_size(vae.vae_dtype) + + # Live preview via WrappersMP.OUTER_SAMPLE + ProgressBar + # The wrapper augments the sampler's own callback to decode x0 -> gaussian splat -> preview image each step + def outer_sample_wrapper(executor, *args, **kwargs): + args = list(args) + cb_idx = 5 # outer_sample(noise, latent_image, sampler, sigmas, denoise_mask, callback, disable_pbar, seed) + orig_cb = args[cb_idx] if len(args) > cb_idx else kwargs.get("callback") + state = {"ok": True, "pbar": None, "loaded": False} + + def callback(step, x0, x, total_steps): + if orig_cb is not None: + orig_cb(step, x0, x, total_steps) + if not state["ok"]: + return + try: + if not state["loaded"]: + loaded_models = comfy.model_management.loaded_models(only_currently_used=True) + loaded_models.append(vae.patcher) + comfy.model_management.load_models_gpu(loaded_models, memory_required=memory_required) + state["loaded"] = True + img = decode_x0_to_image(vae, x0, cfg) + if state["pbar"] is None: + state["pbar"] = comfy.utils.ProgressBar(total_steps) + state["pbar"].update_absolute(step + 1, total_steps, ("JPEG", img, 512)) + except Exception as e: + logging.warning("TripoSplatSamplingPreview: preview failed, disabling ({})".format(e)) + state["ok"] = False + + if len(args) > cb_idx: + args[cb_idx] = callback + else: + kwargs["callback"] = callback + return executor(*args, **kwargs) + + m = model.clone() + m.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.OUTER_SAMPLE, "triposplat_sampling_preview", outer_sample_wrapper) + return IO.NodeOutput(m) + + +class TripoSplatExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + TripoSplatPreprocessImage, + TripoSplatConditioning, + VAEDecodeTripoSplat, + TripoSplatSamplingPreview, + ] + + +async def comfy_entrypoint() -> TripoSplatExtension: + return TripoSplatExtension() diff --git a/comfy_extras/nodes_upscale_model.py b/comfy_extras/nodes_upscale_model.py index d3ee3f1c1..1cf5a5d01 100644 --- a/comfy_extras/nodes_upscale_model.py +++ b/comfy_extras/nodes_upscale_model.py @@ -22,7 +22,7 @@ class UpscaleModelLoader(io.ComfyNode): return io.Schema( node_id="UpscaleModelLoader", display_name="Load Upscale Model", - category="loaders", + category="model/loaders", inputs=[ io.Combo.Input("model_name", options=folder_paths.get_filename_list("upscale_models")), ], diff --git a/comfy_extras/nodes_video.py b/comfy_extras/nodes_video.py index 78a2a28f8..6f6c416a6 100644 --- a/comfy_extras/nodes_video.py +++ b/comfy_extras/nodes_video.py @@ -1,5 +1,3 @@ -from __future__ import annotations - import os import av import torch @@ -21,7 +19,7 @@ class SaveWEBM(io.ComfyNode): category="video", is_experimental=True, inputs=[ - io.Image.Input("images"), + io.Image.Input("images", tooltip="RGBA images are saved with their alpha channel as transparency (vp9 codec only)."), io.String.Input("filename_prefix", default="ComfyUI"), io.Combo.Input("codec", options=["vp9", "av1"]), io.Float.Input("fps", default=24.0, min=0.01, max=1000.0, step=0.01), @@ -47,18 +45,25 @@ class SaveWEBM(io.ComfyNode): for x in cls.hidden.extra_pnginfo: container.metadata[x] = json.dumps(cls.hidden.extra_pnginfo[x]) + # Save transparency when the images carry an alpha channel (RGBA) and the codec supports it. + # vp9 -> yuva420p; other codecs have no usable alpha path, so the alpha is ignored. + save_alpha = images.shape[-1] == 4 and codec == "vp9" + codec_map = {"vp9": "libvpx-vp9", "av1": "libsvtav1"} stream = container.add_stream(codec_map[codec], rate=Fraction(round(fps * 1000), 1000)) stream.width = images.shape[-2] stream.height = images.shape[-3] - stream.pix_fmt = "yuv420p10le" if codec == "av1" else "yuv420p" + stream.pix_fmt = "yuva420p" if save_alpha else ("yuv420p10le" if codec == "av1" else "yuv420p") stream.bit_rate = 0 stream.options = {'crf': str(crf)} if codec == "av1": stream.options["preset"] = "6" for frame in images: - frame = av.VideoFrame.from_ndarray(torch.clamp(frame[..., :3] * 255, min=0, max=255).to(device=torch.device("cpu"), dtype=torch.uint8).numpy(), format="rgb24") + if save_alpha: + frame = av.VideoFrame.from_ndarray(torch.clamp(frame[..., :4] * 255, min=0, max=255).to(device=torch.device("cpu"), dtype=torch.uint8).numpy(), format="rgba") + else: + frame = av.VideoFrame.from_ndarray(torch.clamp(frame[..., :3] * 255, min=0, max=255).to(device=torch.device("cpu"), dtype=torch.uint8).numpy(), format="rgb24") for packet in stream.encode(frame): container.mux(packet) container.mux(stream.encode()) diff --git a/comfy_extras/nodes_video_model.py b/comfy_extras/nodes_video_model.py index 8f19895a1..0d6cae6a8 100644 --- a/comfy_extras/nodes_video_model.py +++ b/comfy_extras/nodes_video_model.py @@ -15,7 +15,7 @@ class ImageOnlyCheckpointLoader: RETURN_TYPES = ("MODEL", "CLIP_VISION", "VAE") FUNCTION = "load_checkpoint" - CATEGORY = "loaders" + CATEGORY = "model/loaders" def load_checkpoint(self, ckpt_name, output_vae=True, output_clip=True): ckpt_path = folder_paths.get_full_path_or_raise("checkpoints", ckpt_name) @@ -41,7 +41,7 @@ class SVD_img2vid_Conditioning: FUNCTION = "encode" - CATEGORY = "conditioning/video_models" + CATEGORY = "model/conditioning/video_models" def encode(self, clip_vision, init_image, vae, width, height, video_frames, motion_bucket_id, fps, augmentation_level): output = clip_vision.encode_image(init_image) @@ -65,7 +65,7 @@ class VideoLinearCFGGuidance: RETURN_TYPES = ("MODEL",) FUNCTION = "patch" - CATEGORY = "sampling/guiders" + CATEGORY = "model/sampling/guiders" def patch(self, model, min_cfg): def linear_cfg(args): @@ -89,7 +89,7 @@ class VideoTriangleCFGGuidance: RETURN_TYPES = ("MODEL",) FUNCTION = "patch" - CATEGORY = "sampling/guiders" + CATEGORY = "model/sampling/guiders" def patch(self, model, min_cfg): def linear_cfg(args): @@ -138,7 +138,7 @@ class ConditioningSetAreaPercentageVideo: RETURN_TYPES = ("CONDITIONING",) FUNCTION = "append" - CATEGORY = "conditioning" + CATEGORY = "model/conditioning" def append(self, conditioning, width, height, temporal, x, y, z, strength): c = node_helpers.conditioning_set_values(conditioning, {"area": ("percentage", temporal, height, width, z, y, x), diff --git a/comfy_extras/nodes_void.py b/comfy_extras/nodes_void.py index be724371a..b43154b8d 100644 --- a/comfy_extras/nodes_void.py +++ b/comfy_extras/nodes_void.py @@ -58,7 +58,7 @@ class OpticalFlowLoader(io.ComfyNode): return io.Schema( node_id="OpticalFlowLoader", display_name="Load Optical Flow Model", - category="loaders", + category="model/loaders", inputs=[ io.Combo.Input( "model_name", @@ -175,7 +175,7 @@ class VOIDInpaintConditioning(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="VOIDInpaintConditioning", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), @@ -288,7 +288,7 @@ class VOIDWarpedNoise(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="VOIDWarpedNoise", - category="latent/video", + category="model/latent/video", inputs=[ OpticalFlow.Input( "optical_flow", @@ -393,7 +393,7 @@ class VOIDWarpedNoiseSource(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="VOIDWarpedNoiseSource", - category="sampling/noise", + category="model/sampling/noise", inputs=[ io.Latent.Input("warped_noise", tooltip="Warped noise latent from VOIDWarpedNoise"), @@ -455,7 +455,7 @@ class VOIDSampler(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="VOIDSampler", - category="sampling/samplers", + category="model/sampling/samplers", inputs=[], outputs=[io.Sampler.Output()], ) diff --git a/comfy_extras/nodes_wan.py b/comfy_extras/nodes_wan.py index e50bfcd2c..d73be8e00 100644 --- a/comfy_extras/nodes_wan.py +++ b/comfy_extras/nodes_wan.py @@ -18,7 +18,7 @@ class WanImageToVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="WanImageToVideo", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), @@ -66,7 +66,7 @@ class WanFunControlToVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="WanFunControlToVideo", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), @@ -119,7 +119,7 @@ class Wan22FunControlToVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="Wan22FunControlToVideo", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), @@ -184,7 +184,7 @@ class WanFirstLastFrameToVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="WanFirstLastFrameToVideo", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), @@ -256,7 +256,7 @@ class WanFunInpaintToVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="WanFunInpaintToVideo", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), @@ -288,7 +288,7 @@ class WanVaceToVideo(io.ComfyNode): return io.Schema( node_id="WanVaceToVideo", search_aliases=["video conditioning", "video control"], - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), @@ -375,7 +375,7 @@ class TrimVideoLatent(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="TrimVideoLatent", - category="latent/video", + category="model/latent/video", inputs=[ io.Latent.Input("samples"), io.Int.Input("trim_amount", default=0, min=0, max=99999), @@ -398,7 +398,7 @@ class WanCameraImageToVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="WanCameraImageToVideo", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), @@ -452,7 +452,7 @@ class WanPhantomSubjectToVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="WanPhantomSubjectToVideo", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), @@ -707,7 +707,7 @@ class WanTrackToVideo(io.ComfyNode): return io.Schema( node_id="WanTrackToVideo", search_aliases=["motion tracking", "trajectory video", "point tracking", "keypoint animation"], - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), @@ -951,7 +951,7 @@ class WanSoundImageToVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="WanSoundImageToVideo", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), @@ -984,7 +984,7 @@ class WanSoundImageToVideoExtend(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="WanSoundImageToVideoExtend", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), @@ -1046,7 +1046,7 @@ class WanHuMoImageToVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="WanHuMoImageToVideo", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), @@ -1112,7 +1112,7 @@ class WanAnimateToVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="WanAnimateToVideo", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), @@ -1252,7 +1252,7 @@ class Wan22ImageToVideoLatent(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="Wan22ImageToVideoLatent", - category="conditioning/inpaint", + category="model/conditioning/inpaint", inputs=[ io.Vae.Input("vae"), io.Int.Input("width", default=1280, min=32, max=nodes.MAX_RESOLUTION, step=32), @@ -1302,7 +1302,7 @@ class WanInfiniteTalkToVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="WanInfiniteTalkToVideo", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.DynamicCombo.Input("mode", options=[ io.DynamicCombo.Option("single_speaker", []), @@ -1456,63 +1456,6 @@ class WanInfiniteTalkToVideo(io.ComfyNode): return io.NodeOutput(model_patched, positive, negative, out_latent, trim_image) -class WanSCAILToVideo(io.ComfyNode): - @classmethod - def define_schema(cls): - return io.Schema( - node_id="WanSCAILToVideo", - category="conditioning/video_models", - inputs=[ - io.Conditioning.Input("positive"), - io.Conditioning.Input("negative"), - io.Vae.Input("vae"), - io.Int.Input("width", default=512, min=32, max=nodes.MAX_RESOLUTION, step=32), - io.Int.Input("height", default=896, min=32, max=nodes.MAX_RESOLUTION, step=32), - io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4), - io.Int.Input("batch_size", default=1, min=1, max=4096), - io.ClipVisionOutput.Input("clip_vision_output", optional=True), - io.Image.Input("reference_image", optional=True), - io.Image.Input("pose_video", optional=True, tooltip="Video used for pose conditioning. Will be downscaled to half the resolution of the main video."), - io.Float.Input("pose_strength", default=1.0, min=0.0, max=10.0, step=0.01, tooltip="Strength of the pose latent."), - io.Float.Input("pose_start", default=0.0, min=0.0, max=1.0, step=0.01, tooltip="Start step to use pose conditioning."), - io.Float.Input("pose_end", default=1.0, min=0.0, max=1.0, step=0.01, tooltip="End step to use pose conditioning."), - ], - outputs=[ - io.Conditioning.Output(display_name="positive"), - io.Conditioning.Output(display_name="negative"), - io.Latent.Output(display_name="latent", tooltip="Empty latent of the generation size."), - ], - is_experimental=True, - ) - - @classmethod - def execute(cls, positive, negative, vae, width, height, length, batch_size, pose_strength, pose_start, pose_end, reference_image=None, clip_vision_output=None, pose_video=None) -> io.NodeOutput: - latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device()) - - ref_latent = None - if reference_image is not None: - reference_image = comfy.utils.common_upscale(reference_image[:1].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1) - ref_latent = vae.encode(reference_image[:, :, :, :3]) - - if ref_latent is not None: - positive = node_helpers.conditioning_set_values(positive, {"reference_latents": [ref_latent]}, append=True) - negative = node_helpers.conditioning_set_values(negative, {"reference_latents": [torch.zeros_like(ref_latent)]}, append=True) - - if clip_vision_output is not None: - positive = node_helpers.conditioning_set_values(positive, {"clip_vision_output": clip_vision_output}) - negative = node_helpers.conditioning_set_values(negative, {"clip_vision_output": clip_vision_output}) - - if pose_video is not None: - pose_video = comfy.utils.common_upscale(pose_video[:length].movedim(-1, 1), width // 2, height // 2, "area", "center").movedim(1, -1) - pose_video_latent = vae.encode(pose_video[:, :, :, :3]) * pose_strength - positive = node_helpers.conditioning_set_values_with_timestep_range(positive, {"pose_video_latent": pose_video_latent}, pose_start, pose_end) - negative = node_helpers.conditioning_set_values_with_timestep_range(negative, {"pose_video_latent": pose_video_latent}, pose_start, pose_end) - - out_latent = {} - out_latent["samples"] = latent - return io.NodeOutput(positive, negative, out_latent) - - class WanExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[io.ComfyNode]]: @@ -1533,7 +1476,6 @@ class WanExtension(ComfyExtension): WanAnimateToVideo, Wan22ImageToVideoLatent, WanInfiniteTalkToVideo, - WanSCAILToVideo, ] async def comfy_entrypoint() -> WanExtension: diff --git a/comfy_extras/nodes_wandancer.py b/comfy_extras/nodes_wandancer.py index fc005ed4c..a96885745 100644 --- a/comfy_extras/nodes_wandancer.py +++ b/comfy_extras/nodes_wandancer.py @@ -713,7 +713,7 @@ class WanDancerEncodeAudio(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="WanDancerEncodeAudio", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Audio.Input("audio"), io.Int.Input("video_frames", default=149, min=1, max=nodes.MAX_RESOLUTION, step=4), @@ -787,7 +787,7 @@ class WanDancerVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="WanDancerVideo", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), diff --git a/comfy_extras/nodes_wanmove.py b/comfy_extras/nodes_wanmove.py index 5acae03eb..2db064922 100644 --- a/comfy_extras/nodes_wanmove.py +++ b/comfy_extras/nodes_wanmove.py @@ -247,7 +247,7 @@ class WanMoveVisualizeTracks(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="WanMoveVisualizeTracks", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Image.Input("images"), io.Tracks.Input("tracks", optional=True), @@ -283,7 +283,7 @@ class WanMoveTracksFromCoords(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="WanMoveTracksFromCoords", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.String.Input("track_coords", force_input=True, default="[]", optional=True), io.Mask.Input("track_mask", optional=True), @@ -325,7 +325,7 @@ class GenerateTracks(io.ComfyNode): return io.Schema( node_id="GenerateTracks", search_aliases=["motion paths", "camera movement", "trajectory"], - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Int.Input("width", default=832, min=16, max=4096, step=16), io.Int.Input("height", default=480, min=16, max=4096, step=16), @@ -434,7 +434,7 @@ class WanMoveConcatTrack(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="WanMoveConcatTrack", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Tracks.Input("tracks_1"), io.Tracks.Input("tracks_2", optional=True), @@ -463,7 +463,7 @@ class WanMoveTrackToVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="WanMoveTrackToVideo", - category="conditioning/video_models", + category="model/conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), diff --git a/comfyui_version.py b/comfyui_version.py index 0bb0f780c..4e3c924e6 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.22.0" +__version__ = "0.24.0" diff --git a/cuda_malloc.py b/cuda_malloc.py index f7651981c..8c4422db8 100644 --- a/cuda_malloc.py +++ b/cuda_malloc.py @@ -2,6 +2,7 @@ import os import importlib.util from comfy.cli_args import args, PerformanceFeature import subprocess +import re #Can't use pytorch to get the GPU names because the cuda malloc has to be set before the first import. def get_gpu_names(): @@ -77,11 +78,24 @@ try: except: pass +def get_raw_cuda_version(version_str): + match = re.search(r'\+cu(\d+)', version_str) + if match: + try: + return int(match.group(1)) + except: + pass + return None + if not args.cuda_malloc: try: if int(version[0]) >= 2 and "+cu" in version: # enable by default for torch version 2.0 and up only on cuda torch if PerformanceFeature.AutoTune not in args.fast: # Autotune has issues with cuda malloc - args.cuda_malloc = cuda_malloc_supported() + cuda_version = get_raw_cuda_version(version) + if cuda_version is not None and cuda_version >= 130: + args.cuda_malloc = True + else: + args.cuda_malloc = cuda_malloc_supported() except: pass diff --git a/folder_paths.py b/folder_paths.py index 36d61fcd0..7304e1b73 100644 --- a/folder_paths.py +++ b/folder_paths.py @@ -1,5 +1,3 @@ -from __future__ import annotations - import os import time import mimetypes diff --git a/main.py b/main.py index 1e47cab84..7fcc8e97d 100644 --- a/main.py +++ b/main.py @@ -26,6 +26,7 @@ import utils.extra_config from utils.mime_types import init_mime_types import faulthandler import logging +import signal import sys from comfy_execution.progress import get_progress_state from comfy_execution.utils import get_executing_context @@ -37,7 +38,19 @@ if __name__ == "__main__": os.environ['HF_HUB_DISABLE_TELEMETRY'] = '1' os.environ['DO_NOT_TRACK'] = '1' -faulthandler.enable(file=sys.stderr, all_threads=False) +faulthandler.enable(file=sys.stderr, all_threads=args.debug_hang) +if __name__ == "__main__" and args.debug_hang: + dumping_traceback = False + + def dump_traceback_on_sigint(signum, frame): + global dumping_traceback + if dumping_traceback: + raise KeyboardInterrupt + dumping_traceback = True + faulthandler.dump_traceback(file=sys.stderr, all_threads=True) + raise KeyboardInterrupt + + signal.signal(signal.SIGINT, dump_traceback_on_sigint) import comfy_aimdo.control @@ -218,7 +231,7 @@ import comfy.model_patcher if args.enable_dynamic_vram or (enables_dynamic_vram() and comfy.model_management.is_nvidia() and not comfy.model_management.is_wsl()): if (not args.enable_dynamic_vram) and (comfy.model_management.torch_version_numeric < (2, 8)): logging.warning("Unsupported Pytorch detected. DynamicVRAM support requires Pytorch version 2.8 or later. Falling back to legacy ModelPatcher. VRAM estimates may be unreliable especially on Windows") - elif comfy_aimdo.control.init_device(comfy.model_management.get_torch_device().index): + elif comfy_aimdo.control.init_devices(d.index for d in comfy.model_management.get_all_torch_devices()): if args.verbose == 'DEBUG': comfy_aimdo.control.set_log_debug() elif args.verbose == 'CRITICAL': @@ -286,8 +299,8 @@ def prompt_worker(q, server_instance): cache_ram = 0 cache_ram_inactive = 0 if not args.cache_classic and not args.cache_none and args.cache_lru <= 0: - cache_ram = min(32.0, max(4.0, comfy.model_management.total_ram * 0.25 / 1024.0)) - cache_ram_inactive = min(96.0, max(12.0, comfy.model_management.total_ram * 0.75 / 1024.0)) + cache_ram = min(10.0, max(2.0, comfy.model_management.total_ram * 0.10 / 1024.0)) + cache_ram_inactive = min(96.0, comfy.model_management.total_ram / 1024.0) if len(args.cache_ram) > 0: cache_ram = args.cache_ram[0] if len(args.cache_ram) > 1: @@ -344,9 +357,9 @@ def prompt_worker(q, server_instance): # Log Time in a more readable way after 10 minutes if execution_time > 600: execution_time = time.strftime("%H:%M:%S", time.gmtime(execution_time)) - logging.info(f"Prompt executed in {execution_time}") + logging.info(f"Prompt executed in {execution_time}", extra={'color': 'green'}) else: - logging.info("Prompt executed in {:.2f} seconds".format(execution_time)) + logging.info("Prompt executed in {:.2f} seconds".format(execution_time), extra={'color': 'green'}) if not asset_seeder.is_disabled(): paths = _collect_output_absolute_paths(e.history_result) @@ -464,13 +477,6 @@ def start_comfyui(asyncio_loop=None): folder_paths.set_temp_directory(temp_dir) cleanup_temp() - if args.windows_standalone_build: - try: - import new_updater - new_updater.update_windows_updater() - except: - pass - if not asyncio_loop: asyncio_loop = asyncio.new_event_loop() asyncio.set_event_loop(asyncio_loop) diff --git a/new_updater.py b/new_updater.py deleted file mode 100644 index 9a203acdd..000000000 --- a/new_updater.py +++ /dev/null @@ -1,35 +0,0 @@ -import os -import shutil - -base_path = os.path.dirname(os.path.realpath(__file__)) - - -def update_windows_updater(): - top_path = os.path.dirname(base_path) - updater_path = os.path.join(base_path, ".ci/update_windows/update.py") - bat_path = os.path.join(base_path, ".ci/update_windows/update_comfyui.bat") - - dest_updater_path = os.path.join(top_path, "update/update.py") - dest_bat_path = os.path.join(top_path, "update/update_comfyui.bat") - dest_bat_deps_path = os.path.join(top_path, "update/update_comfyui_and_python_dependencies.bat") - - try: - with open(dest_bat_path, 'rb') as f: - contents = f.read() - except: - return - - if not contents.startswith(b"..\\python_embeded\\python.exe .\\update.py"): - return - - shutil.copy(updater_path, dest_updater_path) - try: - with open(dest_bat_deps_path, 'rb') as f: - contents = f.read() - contents = contents.replace(b'..\\python_embeded\\python.exe .\\update.py ..\\ComfyUI\\', b'call update_comfyui.bat nopause') - with open(dest_bat_deps_path, 'wb') as f: - f.write(contents) - except: - pass - shutil.copy(bat_path, dest_bat_path) - print("Updated the windows standalone package updater.") # noqa: T201 diff --git a/nodes.py b/nodes.py index f3fcb6656..1dc996588 100644 --- a/nodes.py +++ b/nodes.py @@ -1,4 +1,3 @@ -from __future__ import annotations import torch @@ -69,7 +68,7 @@ class CLIPTextEncode(ComfyNodeABC): OUTPUT_TOOLTIPS = ("A conditioning containing the embedded text used to guide the diffusion model.",) FUNCTION = "encode" - CATEGORY = "conditioning" + CATEGORY = "model/conditioning" DESCRIPTION = "Encodes a text prompt using a CLIP model into an embedding that can be used to guide the diffusion model towards generating specific images." SEARCH_ALIASES = ["text", "prompt", "text prompt", "positive prompt", "negative prompt", "encode text", "text encoder", "encode prompt"] @@ -88,7 +87,7 @@ class ConditioningCombine: RETURN_TYPES = ("CONDITIONING",) FUNCTION = "combine" - CATEGORY = "conditioning" + CATEGORY = "model/conditioning" SEARCH_ALIASES = ["combine", "merge conditioning", "combine prompts", "merge prompts", "mix prompts", "add prompt"] def combine(self, conditioning_1, conditioning_2): @@ -105,7 +104,7 @@ class ConditioningAverage : RETURN_TYPES = ("CONDITIONING",) FUNCTION = "addWeighted" - CATEGORY = "conditioning" + CATEGORY = "model/conditioning" def addWeighted(self, conditioning_to, conditioning_from, conditioning_to_strength): out = [] @@ -144,7 +143,7 @@ class ConditioningConcat: RETURN_TYPES = ("CONDITIONING",) FUNCTION = "concat" - CATEGORY = "conditioning" + CATEGORY = "model/conditioning" def concat(self, conditioning_to, conditioning_from): out = [] @@ -177,7 +176,7 @@ class ConditioningSetArea: RETURN_TYPES = ("CONDITIONING",) FUNCTION = "append" - CATEGORY = "conditioning" + CATEGORY = "model/conditioning" def append(self, conditioning, width, height, x, y, strength): c = node_helpers.conditioning_set_values(conditioning, {"area": (height // 8, width // 8, y // 8, x // 8), @@ -198,7 +197,7 @@ class ConditioningSetAreaPercentage: RETURN_TYPES = ("CONDITIONING",) FUNCTION = "append" - CATEGORY = "conditioning" + CATEGORY = "model/conditioning" def append(self, conditioning, width, height, x, y, strength): c = node_helpers.conditioning_set_values(conditioning, {"area": ("percentage", height, width, y, x), @@ -215,7 +214,7 @@ class ConditioningSetAreaStrength: RETURN_TYPES = ("CONDITIONING",) FUNCTION = "append" - CATEGORY = "conditioning" + CATEGORY = "model/conditioning" def append(self, conditioning, strength): c = node_helpers.conditioning_set_values(conditioning, {"strength": strength}) @@ -235,7 +234,7 @@ class ConditioningSetMask: RETURN_TYPES = ("CONDITIONING",) FUNCTION = "append" - CATEGORY = "conditioning" + CATEGORY = "model/conditioning" def append(self, conditioning, mask, set_cond_area, strength): set_area_to_bounds = False @@ -304,7 +303,7 @@ class VAEDecode: OUTPUT_TOOLTIPS = ("The decoded image.",) FUNCTION = "decode" - CATEGORY = "latent" + CATEGORY = "model/latent" DESCRIPTION = "Decodes latent images back into pixel space images." SEARCH_ALIASES = ["decode", "decode latent", "latent to image", "render latent"] @@ -358,7 +357,7 @@ class VAEEncode: RETURN_TYPES = ("LATENT",) FUNCTION = "encode" - CATEGORY = "latent" + CATEGORY = "model/latent" SEARCH_ALIASES = ["encode", "encode image", "image to latent"] def encode(self, vae, pixels): @@ -390,7 +389,7 @@ class VAEEncodeForInpaint: RETURN_TYPES = ("LATENT",) FUNCTION = "encode" - CATEGORY = "latent/inpaint" + CATEGORY = "model/latent/inpaint" def encode(self, vae, pixels, mask, grow_mask_by=6): downscale_ratio = vae.spacial_compression_encode() @@ -439,7 +438,7 @@ class InpaintModelConditioning: RETURN_NAMES = ("positive", "negative", "latent") FUNCTION = "encode" - CATEGORY = "conditioning/inpaint" + CATEGORY = "model/conditioning/inpaint" def encode(self, positive, negative, pixels, vae, mask, noise_mask=True): x = (pixels.shape[1] // 8) * 8 @@ -599,7 +598,7 @@ class CheckpointLoaderSimple: "The VAE model used for encoding and decoding images to and from latent space.") FUNCTION = "load_checkpoint" - CATEGORY = "loaders" + CATEGORY = "model/loaders" DESCRIPTION = "Loads a diffusion model checkpoint, diffusion models are used to denoise latents." SEARCH_ALIASES = ["load model", "checkpoint", "model loader", "load checkpoint", "ckpt", "model"] @@ -645,7 +644,7 @@ class unCLIPCheckpointLoader: RETURN_TYPES = ("MODEL", "CLIP", "VAE", "CLIP_VISION") FUNCTION = "load_checkpoint" - CATEGORY = "loaders" + CATEGORY = "model/loaders" def load_checkpoint(self, ckpt_name, output_vae=True, output_clip=True): ckpt_path = folder_paths.get_full_path_or_raise("checkpoints", ckpt_name) @@ -661,7 +660,7 @@ class CLIPSetLastLayer: RETURN_TYPES = ("CLIP",) FUNCTION = "set_last_layer" - CATEGORY = "conditioning" + CATEGORY = "model/conditioning" def set_last_layer(self, clip, stop_at_clip_layer): clip = clip.clone() @@ -690,7 +689,7 @@ class LoraLoader: OUTPUT_TOOLTIPS = ("The modified diffusion model.", "The modified CLIP model.") FUNCTION = "load_lora" - CATEGORY = "loaders" + CATEGORY = "model/loaders" DESCRIPTION = "This LoRA loader is used to modify both diffusion and CLIP models, altering the way in which latents are denoised such as applying styles. Multiple LoRA nodes can be linked together." SEARCH_ALIASES = ["lora", "load lora", "apply lora", "lora loader", "lora model"] @@ -790,11 +789,12 @@ class VAELoader: RETURN_TYPES = ("VAE",) FUNCTION = "load_vae" - CATEGORY = "loaders" + CATEGORY = "model/loaders" #TODO: scale factor? def load_vae(self, vae_name): metadata = None + vae_path = None if vae_name == "pixel_space": sd = {} sd["pixel_space_vae"] = torch.tensor(1.0) @@ -813,6 +813,14 @@ class VAELoader: metadata["tae_latent_channels"] = 128 vae = comfy.sd.VAE(sd=sd, metadata=metadata) vae.throw_exception_if_invalid() + # Register a reload factory on the patcher so multigpu deepclones + # (Select VAE Device, future MultiGPU VAE work-units) can produce + # per-device clones from the same loader context. Only set when we + # actually have a single backing file -- pixel_space and the + # image TAESDs (composed from separate encoder/decoder files via + # load_taesd) are not addressable by a single vae_path. + if vae_path is not None: + vae.patcher.cached_patcher_init = (comfy.sd.load_vae_patcher, (vae_path, metadata, None)) return (vae,) class ControlNetLoader: @@ -823,7 +831,7 @@ class ControlNetLoader: RETURN_TYPES = ("CONTROL_NET",) FUNCTION = "load_controlnet" - CATEGORY = "loaders" + CATEGORY = "model/loaders" SEARCH_ALIASES = ["controlnet", "control net", "cn", "load controlnet", "controlnet loader"] def load_controlnet(self, control_net_name): @@ -842,7 +850,7 @@ class DiffControlNetLoader: RETURN_TYPES = ("CONTROL_NET",) FUNCTION = "load_controlnet" - CATEGORY = "loaders" + CATEGORY = "model/loaders" def load_controlnet(self, model, control_net_name): controlnet_path = folder_paths.get_full_path_or_raise("controlnet", control_net_name) @@ -862,7 +870,7 @@ class ControlNetApply: FUNCTION = "apply_controlnet" DEPRECATED = True - CATEGORY = "conditioning/controlnet" + CATEGORY = "model/conditioning/controlnet" def apply_controlnet(self, conditioning, control_net, image, strength): if strength == 0: @@ -900,7 +908,7 @@ class ControlNetApplyAdvanced: RETURN_NAMES = ("positive", "negative") FUNCTION = "apply_controlnet" - CATEGORY = "conditioning/controlnet" + CATEGORY = "model/conditioning/controlnet" SEARCH_ALIASES = ["controlnet", "apply controlnet", "use controlnet", "control net"] def apply_controlnet(self, positive, negative, control_net, image, strength, start_percent, end_percent, vae=None, extra_concat=[]): @@ -961,7 +969,7 @@ class CLIPLoader: @classmethod def INPUT_TYPES(s): return {"required": { "clip_name": (folder_paths.get_filename_list("text_encoders"), ), - "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream", "chroma", "ace", "omnigen2", "qwen_image", "hunyuan_image", "flux2", "ovis", "longcat_image", "cogvideox"], ), + "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream", "chroma", "ace", "omnigen2", "qwen_image", "hunyuan_image", "flux2", "ovis", "longcat_image", "cogvideox", "lens", "pixeldit", "ideogram4"], ), }, "optional": { "device": (["default", "cpu"], {"advanced": True}), @@ -971,7 +979,7 @@ class CLIPLoader: CATEGORY = "advanced/loaders" - DESCRIPTION = "[Recipes]\n\nstable_diffusion: clip-l\nstable_cascade: clip-g\nsd3: t5 xxl/ clip-g / clip-l\nstable_audio: t5 base\nmochi: t5 xxl\ncogvideox: t5 xxl (226-token padding)\ncosmos: old t5 xxl\nlumina2: gemma 2 2B\nwan: umt5 xxl\n hidream: llama-3.1 (Recommend) or t5\nomnigen2: qwen vl 2.5 3B" + DESCRIPTION = "[Recipes]\n\nstable_diffusion: clip-l\nstable_cascade: clip-g\nsd3: t5 xxl/ clip-g / clip-l\nstable_audio: t5 base\nmochi: t5 xxl\ncogvideox: t5 xxl (226-token padding)\ncosmos: old t5 xxl\nlumina2: gemma 2 2B\nwan: umt5 xxl\n hidream: llama-3.1 (Recommend) or t5\nomnigen2: qwen vl 2.5 3B\nlens: gpt-oss-20b\n pixeldit: gemma 2 2B elm" def load_clip(self, clip_name, type="stable_diffusion", device="default"): clip_type = getattr(comfy.sd.CLIPType, type.upper(), comfy.sd.CLIPType.STABLE_DIFFUSION) @@ -1022,7 +1030,7 @@ class CLIPVisionLoader: RETURN_TYPES = ("CLIP_VISION",) FUNCTION = "load_clip" - CATEGORY = "loaders" + CATEGORY = "model/loaders" def load_clip(self, clip_name): clip_path = folder_paths.get_full_path_or_raise("clip_vision", clip_name) @@ -1041,7 +1049,7 @@ class CLIPVisionEncode: RETURN_TYPES = ("CLIP_VISION_OUTPUT",) FUNCTION = "encode" - CATEGORY = "conditioning" + CATEGORY = "model/conditioning" def encode(self, clip_vision, image, crop): crop_image = True @@ -1058,7 +1066,7 @@ class StyleModelLoader: RETURN_TYPES = ("STYLE_MODEL",) FUNCTION = "load_style_model" - CATEGORY = "loaders" + CATEGORY = "model/loaders" def load_style_model(self, style_model_name): style_model_path = folder_paths.get_full_path_or_raise("style_models", style_model_name) @@ -1080,7 +1088,7 @@ class StyleModelApply: RETURN_TYPES = ("CONDITIONING",) FUNCTION = "apply_stylemodel" - CATEGORY = "conditioning/style_model" + CATEGORY = "model/conditioning/style_model" def apply_stylemodel(self, conditioning, style_model, clip_vision_output, strength, strength_type): cond = style_model.get_cond(clip_vision_output).flatten(start_dim=0, end_dim=1).unsqueeze(dim=0) @@ -1140,7 +1148,7 @@ class unCLIPConditioning: RETURN_TYPES = ("CONDITIONING",) FUNCTION = "apply_adm" - CATEGORY = "conditioning" + CATEGORY = "model/conditioning" def apply_adm(self, conditioning, clip_vision_output, strength, noise_augmentation): if strength == 0: @@ -1157,7 +1165,7 @@ class GLIGENLoader: RETURN_TYPES = ("GLIGEN",) FUNCTION = "load_gligen" - CATEGORY = "loaders" + CATEGORY = "model/loaders" def load_gligen(self, gligen_name): gligen_path = folder_paths.get_full_path_or_raise("gligen", gligen_name) @@ -1179,7 +1187,7 @@ class GLIGENTextBoxApply: RETURN_TYPES = ("CONDITIONING",) FUNCTION = "append" - CATEGORY = "conditioning/gligen" + CATEGORY = "model/conditioning/gligen" def append(self, conditioning_to, clip, gligen_textbox_model, text, width, height, x, y): c = [] @@ -1209,7 +1217,7 @@ class EmptyLatentImage: OUTPUT_TOOLTIPS = ("The empty latent image batch.",) FUNCTION = "generate" - CATEGORY = "latent" + CATEGORY = "model/latent" DESCRIPTION = "Create a new batch of empty latent images to be denoised via sampling." SEARCH_ALIASES = ["empty", "empty latent", "new latent", "create latent", "blank latent", "blank"] @@ -1230,7 +1238,7 @@ class LatentFromBatch: RETURN_TYPES = ("LATENT",) FUNCTION = "frombatch" - CATEGORY = "latent/batch" + CATEGORY = "model/latent/batch" def frombatch(self, samples, batch_index, length): s = samples.copy() @@ -1265,7 +1273,7 @@ class RepeatLatentBatch: RETURN_TYPES = ("LATENT",) FUNCTION = "repeat" - CATEGORY = "latent/batch" + CATEGORY = "model/latent/batch" def repeat(self, samples, amount): s = samples.copy() @@ -1297,7 +1305,7 @@ class LatentUpscale: RETURN_TYPES = ("LATENT",) FUNCTION = "upscale" - CATEGORY = "latent" + CATEGORY = "model/latent" def upscale(self, samples, upscale_method, width, height, crop): if width == 0 and height == 0: @@ -1330,7 +1338,7 @@ class LatentUpscaleBy: RETURN_TYPES = ("LATENT",) FUNCTION = "upscale" - CATEGORY = "latent" + CATEGORY = "model/latent" def upscale(self, samples, upscale_method, scale_by): s = samples.copy() @@ -1348,7 +1356,7 @@ class LatentRotate: RETURN_TYPES = ("LATENT",) FUNCTION = "rotate" - CATEGORY = "latent/transform" + CATEGORY = "model/latent/transform" def rotate(self, samples, rotation): s = samples.copy() @@ -1374,7 +1382,7 @@ class LatentFlip: RETURN_TYPES = ("LATENT",) FUNCTION = "flip" - CATEGORY = "latent/transform" + CATEGORY = "model/latent/transform" def flip(self, samples, flip_method): s = samples.copy() @@ -1399,7 +1407,7 @@ class LatentComposite: RETURN_TYPES = ("LATENT",) FUNCTION = "composite" - CATEGORY = "latent" + CATEGORY = "model/latent" def composite(self, samples_to, samples_from, x, y, composite_method="normal", feather=0): x = x // 8 @@ -1486,7 +1494,7 @@ class LatentCrop: RETURN_TYPES = ("LATENT",) FUNCTION = "crop" - CATEGORY = "latent/transform" + CATEGORY = "model/latent/transform" def crop(self, samples, width, height, x, y): s = samples.copy() @@ -1516,7 +1524,7 @@ class SetLatentNoiseMask: RETURN_TYPES = ("LATENT",) FUNCTION = "set_mask" - CATEGORY = "latent/inpaint" + CATEGORY = "model/latent/inpaint" def set_mask(self, samples, mask): s = samples.copy() @@ -1570,7 +1578,7 @@ class KSampler: OUTPUT_TOOLTIPS = ("The denoised latent.",) FUNCTION = "sample" - CATEGORY = "sampling" + CATEGORY = "model/sampling" DESCRIPTION = "Uses the provided model, positive and negative conditioning to denoise the latent image." SEARCH_ALIASES = ["sampler", "sample", "generate", "denoise", "diffuse", "txt2img", "img2img"] @@ -1600,7 +1608,7 @@ class KSamplerAdvanced: RETURN_TYPES = ("LATENT",) FUNCTION = "sample" - CATEGORY = "sampling" + CATEGORY = "model/sampling" def sample(self, model, add_noise, noise_seed, steps, cfg, sampler_name, scheduler, positive, negative, latent_image, start_at_step, end_at_step, return_with_leftover_noise, denoise=1.0): force_full_denoise = True @@ -2354,6 +2362,7 @@ async def init_builtin_extra_nodes(): "nodes_model_downscale.py", "nodes_images.py", "nodes_video_model.py", + "nodes_ideogram4.py", "nodes_train.py", "nodes_dataset.py", "nodes_sag.py", @@ -2389,11 +2398,13 @@ async def init_builtin_extra_nodes(): "nodes_lt_audio.py", "nodes_lt.py", "nodes_hooks.py", + "nodes_multigpu.py", "nodes_load_3d.py", "nodes_cosmos.py", "nodes_video.py", "nodes_lumina2.py", "nodes_wan.py", + "nodes_bernini.py", "nodes_lotus.py", "nodes_hunyuan3d.py", "nodes_primitive.py", @@ -2411,6 +2422,7 @@ async def init_builtin_extra_nodes(): "nodes_context_windows.py", "nodes_qwen.py", "nodes_chroma_radiance.py", + "nodes_pid.py", "nodes_model_patch.py", "nodes_easycache.py", "nodes_audio_encoder.py", @@ -2441,12 +2453,16 @@ async def init_builtin_extra_nodes(): "nodes_rtdetr.py", "nodes_frame_interpolation.py", "nodes_sam3.py", + "nodes_scail.py", "nodes_void.py", "nodes_wandancer.py", "nodes_hidream_o1.py", "nodes_save_3d.py", "nodes_moge.py", "nodes_mediapipe.py", + "nodes_gaussian_splat.py", + "nodes_triposplat.py", + "nodes_depth_anything_3.py", ] import_failed = [] diff --git a/openapi.yaml b/openapi.yaml index 885231acc..c27ed7adf 100644 --- a/openapi.yaml +++ b/openapi.yaml @@ -1,8726 +1,4879 @@ -openapi: 3.1.0 -info: - title: ComfyUI API - description: | - API for ComfyUI - A powerful and modular stable diffusion GUI and backend. - - This API allows you to interact with ComfyUI programmatically, including: - - Submitting and managing workflow executions - - Querying node/object information - - Uploading and viewing files - - Managing user settings and data - - Asset management (feature-gated) - - ## Dual-path routing - Every route registered via `self.routes` in the ComfyUI server is available at - both its bare path (e.g. `/prompt`) and an `/api`-prefixed path (e.g. `/api/prompt`). - This spec uses the `/api`-prefixed versions as canonical. - - ## Multi-user mode - When ComfyUI is started with `--multi-user`, the `Comfy-User` header identifies - the active user for settings, userdata, and history isolation. This is **not** a - security mechanism — it is an organisational convenience with no authentication - or authorisation behind it. - version: 1.0.0 - license: - name: GNU General Public License v3.0 - url: https://github.com/comfyanonymous/ComfyUI/blob/master/LICENSE - -servers: - - url: / - description: Default ComfyUI server (typically http://127.0.0.1:8188) - -tags: - - name: prompt - description: Workflow submission and prompt info - - name: queue - description: Queue inspection and management - - name: history - description: Execution history - - name: upload - description: File upload endpoints - - name: view - description: File viewing / download - - name: system - description: System stats and feature flags - - name: node - description: Node / object_info definitions - - name: model - description: Model folder and file listing - - name: user - description: User management (multi-user mode) - - name: userdata - description: Per-user file storage - - name: settings - description: Per-user settings - - name: extensions - description: Frontend extension JS files - - name: subgraph - description: Global subgraph blueprints - - name: internal - description: Internal / debug endpoints - - name: assets - description: Asset management (feature-gated behind enable-assets) - - - name: auth - description: Authentication and session management (cloud-only) - - name: billing - description: Billing, subscriptions, and payment management (cloud-only) - - name: workspace - description: Workspace and team management (cloud-only) - - name: hub - description: "ComfyUI Hub: profiles, shared workflows, and labels (cloud-only)" - - name: workflows - description: Cloud workflow management and versioning (cloud-only) - - name: task - description: Background task management (cloud-only) - - name: runtime-only - description: Operations served exclusively by the cloud runtime with no local equivalent - -paths: - # --------------------------------------------------------------------------- - # WebSocket - # --------------------------------------------------------------------------- - /ws: - get: - operationId: connectWebSocket - tags: [system] - summary: WebSocket connection for real-time updates - description: | - Upgrades to a WebSocket connection that streams execution progress, - node status, and output messages. The server sends an initial `status` - message with the session ID (SID) on connect. - - ## Message types (server → client) - The server sends JSON messages with a `type` field. See the - `x-websocket-messages` list below for the schema of each message type. - parameters: - - name: clientId - in: query - required: false - schema: - type: string - description: Client identifier. If omitted the server assigns one. - responses: - "101": - description: WebSocket upgrade successful - x-websocket-messages: - - type: status - schema: - $ref: "#/components/schemas/StatusWsMessage" - - type: progress - schema: - $ref: "#/components/schemas/ProgressWsMessage" - - type: progress_text - schema: - $ref: "#/components/schemas/ProgressTextWsMessage" - - type: progress_state - schema: - $ref: "#/components/schemas/ProgressStateWsMessage" - - type: executing - schema: - $ref: "#/components/schemas/ExecutingWsMessage" - - type: executed - schema: - $ref: "#/components/schemas/ExecutedWsMessage" - - type: execution_start - schema: - $ref: "#/components/schemas/ExecutionStartWsMessage" - - type: execution_success - schema: - $ref: "#/components/schemas/ExecutionSuccessWsMessage" - - type: execution_cached - schema: - $ref: "#/components/schemas/ExecutionCachedWsMessage" - - type: execution_interrupted - schema: - $ref: "#/components/schemas/ExecutionInterruptedWsMessage" - - type: execution_error - schema: - $ref: "#/components/schemas/ExecutionErrorWsMessage" - - type: logs - schema: - $ref: "#/components/schemas/LogsWsMessage" - - type: notification - schema: - $ref: "#/components/schemas/NotificationWsMessage" - - type: feature_flags - schema: - $ref: "#/components/schemas/FeatureFlagsWsMessage" - - type: asset_download - schema: - $ref: "#/components/schemas/AssetDownloadWsMessage" - - type: asset_export - schema: - $ref: "#/components/schemas/AssetExportWsMessage" - - # --------------------------------------------------------------------------- - # Prompt - # --------------------------------------------------------------------------- - /api/prompt: - get: - operationId: getPromptInfo - tags: [prompt] - summary: Get queue status - description: Returns how many items remain in the execution queue. - responses: - "200": - description: Queue info - content: - application/json: - schema: - $ref: "#/components/schemas/PromptInfo" - post: - operationId: executePrompt - tags: [prompt] - summary: Submit a workflow for execution - description: Submits a workflow for execution. The server validates the graph, assigns a `prompt_id`, and enqueues it. Clients listen on `/ws` for execution progress and output messages. - requestBody: - required: true - content: - application/json: - schema: - $ref: "#/components/schemas/PromptRequest" - responses: - "200": - description: Prompt accepted - content: - application/json: - schema: - $ref: "#/components/schemas/PromptResponse" - "400": - description: Validation or node errors - content: - application/json: - schema: - $ref: "#/components/schemas/PromptErrorResponse" - - # --------------------------------------------------------------------------- - # Queue - # --------------------------------------------------------------------------- - /api/queue: - get: - operationId: getQueue - tags: [queue] - summary: Get running and pending queue items - description: Returns the server's current execution queue, split into the currently-running prompt and the list of pending prompts. - responses: - "200": - description: Queue contents - content: - application/json: - schema: - $ref: "#/components/schemas/QueueInfo" - post: - operationId: manageQueue - tags: [queue] - summary: Clear or delete items from the queue - description: Mutates the execution queue. Supports clearing all queued prompts or deleting individual prompts by ID. - requestBody: - required: true - content: - application/json: - schema: - $ref: "#/components/schemas/QueueManageRequest" - responses: - "200": - description: Queue updated - - /api/interrupt: - post: - operationId: interruptExecution - tags: [queue] - summary: Interrupt current execution - description: Interrupts the prompt that is currently executing. The next queued prompt (if any) will start immediately after. - requestBody: - required: false - content: - application/json: - schema: - type: object - properties: - prompt_id: - type: string - format: uuid - description: "If provided, only interrupts this specific running prompt. Otherwise interrupts all." - responses: - "200": - description: Interrupt signal sent - - /api/free: - post: - operationId: freeMemory - tags: [queue] - summary: Free GPU memory and/or unload models - description: Frees GPU memory by unloading models and/or freeing the resident model cache, controlled by the request flags. - requestBody: - required: false - content: - application/json: - schema: - type: object - properties: - unload_models: - type: boolean - description: Unload all models from VRAM/RAM - free_memory: - type: boolean - description: Run garbage collection and free cached memory - responses: - "200": - description: Memory freed - - # --------------------------------------------------------------------------- - # Jobs - # --------------------------------------------------------------------------- - /api/jobs: - get: - operationId: listJobs - tags: [queue] - summary: List jobs with filtering and pagination - description: Returns a paginated list of completed prompt executions, newest first. - parameters: - - name: status - in: query - schema: - type: string - description: Filter by job status - - name: workflow_id - in: query - schema: - type: string - description: Filter by workflow ID - - name: sort_by - in: query - schema: - type: string - description: Field to sort by - - name: sort_order - in: query - schema: - type: string - enum: [asc, desc] - description: Sort direction - - name: limit - in: query - schema: - type: integer - description: Maximum number of results (default is unlimited/None) - - name: offset - in: query - schema: - type: integer - default: 0 - description: Pagination offset - responses: - "200": - description: Jobs list - content: - application/json: - schema: - type: object - properties: - jobs: - type: array - items: - $ref: "#/components/schemas/JobEntry" - pagination: - $ref: "#/components/schemas/PaginationInfo" - - /api/jobs/{job_id}: - get: - operationId: getJob - tags: [queue] - summary: Get a single job by ID - description: Returns the full record for a single completed prompt execution, including its outputs, status, and metadata. - parameters: - - name: job_id - in: path - description: The job (prompt) ID to fetch. - required: true - schema: - type: string - format: uuid - responses: - "200": - description: Job detail - content: - application/json: - schema: - $ref: "#/components/schemas/JobDetailResponse" - "404": - description: Job not found - - # --------------------------------------------------------------------------- - # History - # --------------------------------------------------------------------------- - /api/history: - get: - operationId: getHistory - tags: [history] - summary: Get execution history - deprecated: true - description: | - **Deprecated.** Superseded by `GET /api/jobs`, which returns the same - execution records in a paginated, filterable format. Planned for removal - no earlier than a future major release; sunset timeline TBD. - - Returns a dictionary keyed by prompt_id. Each value is a HistoryEntry - containing prompt metadata, outputs, status, and node meta. - parameters: - - $ref: "#/components/parameters/ComfyUserHeader" - - name: max_items - in: query - schema: - type: integer - description: Maximum number of history entries to return - - name: offset - in: query - schema: - type: integer - description: Pagination offset (number of entries to skip) - responses: - "200": - description: History dictionary keyed by prompt_id - content: - application/json: - schema: - type: object - additionalProperties: - $ref: "#/components/schemas/HistoryEntry" - post: - operationId: manageHistory - tags: [history] - summary: Clear or delete history entries - deprecated: true - description: | - **Deprecated.** Superseded by the forthcoming job-management endpoints - under `/api/jobs`. Planned for removal no earlier than a future major - release; sunset timeline TBD. - parameters: - - $ref: "#/components/parameters/ComfyUserHeader" - requestBody: - required: true - content: - application/json: - schema: - $ref: "#/components/schemas/HistoryManageRequest" - responses: - "200": - description: History updated - - /api/history/{prompt_id}: - get: - operationId: getHistoryByPromptId - tags: [history] - summary: Get history for a specific prompt - deprecated: true - description: | - **Deprecated.** Superseded by `GET /api/jobs/{job_id}`, which returns - the same execution record. Planned for removal no earlier than a future - major release; sunset timeline TBD. - parameters: - - $ref: "#/components/parameters/ComfyUserHeader" - - name: prompt_id - in: path - description: The prompt ID to fetch history for. - required: true - schema: - type: string - format: uuid - responses: - "200": - description: Single-entry history dictionary. Returns an empty object `{}` if the prompt_id is not found. - content: - application/json: - schema: - type: object - additionalProperties: - $ref: "#/components/schemas/HistoryEntry" - - # --------------------------------------------------------------------------- - # Upload - # --------------------------------------------------------------------------- - /api/upload/image: - post: - operationId: uploadImage - tags: [upload] - summary: Upload an image file - description: Uploads an image file into one of the input/output/temp directories so it can be referenced by workflow nodes. - requestBody: - required: true - content: - multipart/form-data: - schema: - type: object - required: - - image - properties: - image: - type: string - format: binary - description: Image file to upload - type: - type: string - enum: [input, temp, output] - default: input - description: Target directory type - overwrite: - type: string - description: 'Set to "true" to overwrite existing files' - subfolder: - type: string - description: Subfolder within the target directory - responses: - "200": - description: Upload result - content: - application/json: - schema: - $ref: "#/components/schemas/UploadResult" - "400": - description: No file provided or invalid request - - /api/upload/mask: - post: - operationId: uploadMask - tags: [upload] - deprecated: true - summary: Upload a mask image (deprecated) - description: | - Deprecated. Clients should composite the mask onto the source image - client-side and upload the resulting image via POST /api/upload/image - instead. This endpoint will continue to function for older clients, - but will not receive new features. - - Uploads a mask image associated with a previously-uploaded reference image. - requestBody: - required: true - content: - multipart/form-data: - schema: - type: object - required: - - image - - original_ref - properties: - image: - type: string - format: binary - description: Mask image (alpha channel is used) - original_ref: - type: object - description: Reference to the original image file - required: - - filename - properties: - filename: - type: string - description: Filename of the original image - additionalProperties: true - type: - type: string - enum: [input, temp, output] - default: input - description: Target directory type - overwrite: - type: string - description: 'Set to "true" to overwrite existing files' - subfolder: - type: string - description: Subfolder within the target directory - responses: - "200": - description: Upload result - content: - application/json: - schema: - $ref: "#/components/schemas/UploadResult" - "400": - description: No file provided or invalid request - - # --------------------------------------------------------------------------- - # View - # --------------------------------------------------------------------------- - /api/view: - get: - operationId: viewFile - tags: [view] - summary: View or download a file - description: Serves a file (image, audio, or video) from the input/output/temp directory identified by the query parameters. - parameters: - - name: filename - in: query - required: true - schema: - type: string - description: Name of the file to view - - name: type - in: query - schema: - type: string - enum: [input, output, temp] - default: output - description: Directory type - - name: subfolder - in: query - schema: - type: string - description: Subfolder within the directory - - name: preview - in: query - schema: - type: string - description: Preview format hint (e.g. "webp;90") - - name: channel - in: query - schema: - type: string - enum: [rgba, rgb, a] - description: Channel extraction mode - responses: - "200": - description: File content - content: - image/*: - schema: - type: string - format: binary - video/*: - schema: - type: string - format: binary - audio/*: - schema: - type: string - format: binary - application/octet-stream: - schema: - type: string - format: binary - "404": - description: File not found - - /api/view_metadata/{folder_name}: - get: - operationId: viewMetadata - tags: [view] - summary: Get metadata for a file (e.g. safetensors header) - description: Returns embedded metadata parsed from a file in the given folder — for example, the header of a safetensors model. - parameters: - - name: folder_name - in: path - required: true - schema: - type: string - description: Folder type (output, input, temp, etc.) - - name: filename - in: query - required: true - schema: - type: string - description: Filename to read metadata from - responses: - "200": - description: File metadata - content: - application/json: - schema: - type: object - additionalProperties: true - "404": - description: File or metadata not found - - # --------------------------------------------------------------------------- - # System - # --------------------------------------------------------------------------- - /api/system_stats: - get: - operationId: getSystemStats - tags: [system] - summary: Get system statistics - description: Returns hardware, Python, VRAM, and runtime statistics for the running ComfyUI process. - responses: - "200": - description: System stats - content: - application/json: - schema: - $ref: "#/components/schemas/SystemStatsResponse" - - /api/features: - get: - operationId: getFeatures - tags: [system] - summary: Get enabled feature flags - description: Returns a dictionary of feature flag names to their enabled state. Cloud deployments may include additional typed fields alongside the boolean flags. - responses: - "200": - description: Feature flags - content: - application/json: - schema: - type: object - additionalProperties: - type: boolean - properties: - max_upload_size: - type: integer - format: int64 - minimum: 0 - description: "Maximum file upload size in bytes." - free_tier_credits: - type: integer - format: int32 - minimum: 0 - nullable: true - x-runtime: [cloud] - description: "[cloud-only] Credits available to free-tier users. Local ComfyUI returns null." - posthog_api_host: - type: string - format: uri - nullable: true - x-runtime: [cloud] - description: "[cloud-only] PostHog analytics proxy URL for frontend telemetry. Local ComfyUI returns null." - max_concurrent_jobs: - type: integer - format: int32 - minimum: 0 - nullable: true - x-runtime: [cloud] - description: "[cloud-only] Maximum concurrent jobs the authenticated user can run. Local ComfyUI returns null." - workflow_templates_version: - type: string - nullable: true - x-runtime: [cloud] - description: "[cloud-only] Version identifier for the workflow templates bundle. Local ComfyUI returns null." - workflow_templates_source: - type: string - nullable: true - enum: [dynamic_config_override, workflow_templates_version_json] - x-runtime: [cloud] - description: "[cloud-only] How the templates version was resolved. Local ComfyUI returns null." - - # --------------------------------------------------------------------------- - # Node / Object Info - # --------------------------------------------------------------------------- - /api/object_info: - get: - operationId: getObjectInfo - tags: [node] - summary: Get all node definitions - description: | - Returns a dictionary of every registered node class, keyed by class name. - Each value is a NodeInfo object describing inputs, outputs, category, etc. - responses: - "200": - description: All node definitions - content: - application/json: - schema: - type: object - additionalProperties: - $ref: "#/components/schemas/NodeInfo" - - /api/object_info/{node_class}: - get: - operationId: getObjectInfoByClass - tags: [node] - summary: Get a single node definition - description: Returns the `NodeInfo` definition for a single registered node class. - parameters: - - name: node_class - in: path - required: true - schema: - type: string - description: Node class name (e.g. "KSampler") - responses: - "200": - description: Single node definition - content: - application/json: - schema: - type: object - additionalProperties: - $ref: "#/components/schemas/NodeInfo" - "404": - description: Node class not found - - /api/embeddings: - get: - operationId: getEmbeddings - tags: [node] - summary: List available embedding names - description: Returns the list of text-encoder embeddings available on disk. - responses: - "200": - description: Embedding names - content: - application/json: - schema: - type: array - items: - type: string - - # --------------------------------------------------------------------------- - # Models - # --------------------------------------------------------------------------- - /api/models: - get: - operationId: getModelTypes - tags: [model] - summary: List model folder type names - description: Returns an array of model type names (e.g. checkpoints, loras, vae). - responses: - "200": - description: Model type names - content: - application/json: - schema: - type: array - items: - type: string - - /api/models/{folder}: - get: - operationId: getModelsByFolder - tags: [model] - summary: List model filenames in a folder - description: Returns the names of model files in the given folder. This endpoint predates `/api/experiment/models/{folder}` and returns names only — prefer the experiment endpoint for new integrations. - parameters: - - name: folder - in: path - required: true - schema: - type: string - description: Model folder type name - responses: - "200": - description: Model filenames - content: - application/json: - schema: - type: array - items: - type: string - "404": - description: Unknown folder type - - /api/experiment/models: - get: - operationId: getExperimentModels - tags: [model] - summary: List model folders with paths - description: Returns an array of model folder objects with name and folder paths. - responses: - "200": - description: Model folders - content: - application/json: - schema: - type: array - items: - $ref: "#/components/schemas/ModelFolder" - - /api/experiment/models/{folder}: - get: - operationId: getExperimentModelsByFolder - tags: [model] - summary: List model files with metadata - description: Returns the model files in the given folder with richer metadata (path index, mtime, size) than the legacy `/api/models/{folder}` endpoint. - parameters: - - name: folder - in: path - required: true - schema: - type: string - description: Model folder type name - responses: - "200": - description: Model files with metadata - content: - application/json: - schema: - type: array - items: - $ref: "#/components/schemas/ModelFile" - "404": - description: Unknown folder type - - /api/experiment/models/preview/{folder}/{path_index}/{filename}: - get: - operationId: getModelPreview - tags: [model] - summary: Get model preview image - description: Returns the preview image associated with a model file, if one exists alongside the model on disk. - parameters: - - name: folder - in: path - required: true - schema: - type: string - description: Model folder type name - - name: path_index - in: path - required: true - schema: - type: integer - description: Path index within the folder - - name: filename - in: path - required: true - schema: - type: string - description: Model filename - responses: - "200": - description: Preview image (WebP) - content: - image/webp: - schema: - type: string - format: binary - "404": - description: Preview not found - - # --------------------------------------------------------------------------- - # Users - # --------------------------------------------------------------------------- - /api/users: - get: - operationId: getUsers - tags: [user] - summary: Get user storage info - description: | - Returns user storage configuration. In single-user mode returns - `{"storage": "server", "migrated": true/false}`. In multi-user mode - returns `{"storage": "server", "users": {"user_id": "user_dir", ...}}`. - parameters: - - $ref: "#/components/parameters/ComfyUserHeader" - responses: - "200": - description: User info - content: - application/json: - schema: - type: object - properties: - storage: - type: string - description: Storage backend type (always "server") - migrated: - type: boolean - description: Whether migration from browser storage is complete (single-user) - users: - type: object - additionalProperties: - type: string - description: Map of user_id to directory name (multi-user) - post: - operationId: createUser - tags: [user] - summary: Create a new user (multi-user mode) - description: Creates a new user entry. Only meaningful when ComfyUI is running in multi-user mode. - parameters: - - $ref: "#/components/parameters/ComfyUserHeader" - requestBody: - required: true - content: - application/json: - schema: - type: object - required: - - username - properties: - username: - type: string - description: Username for the new user - responses: - "200": - description: Created user ID - content: - application/json: - schema: - type: string - description: The generated user_id - "400": - description: Username already exists or invalid - - # --------------------------------------------------------------------------- - # Userdata - # --------------------------------------------------------------------------- - /api/userdata: - get: - operationId: listUserdata - tags: [userdata] - summary: List files in a userdata directory - description: Lists files in the authenticated user's data directory. Returns either filename strings or full objects depending on the `full_info` query parameter. - parameters: - - $ref: "#/components/parameters/ComfyUserHeader" - - name: dir - in: query - required: true - schema: - type: string - description: Directory path relative to the user's data folder - - name: recurse - in: query - schema: - type: boolean - description: Recurse into subdirectories - - name: full_info - in: query - schema: - type: boolean - description: Return full file info objects instead of just names - - name: split - in: query - schema: - type: boolean - description: Split paths into directory components - responses: - "200": - description: File listing - content: - application/json: - schema: - $ref: "#/components/schemas/ListUserdataResponse" - "404": - description: Directory not found - - /api/v2/userdata: - get: - operationId: listUserdataV2 - tags: [userdata] - summary: List files in userdata (v2 format) - description: Lists files in the authenticated user's data directory using the v2 response shape, which always returns full objects. - parameters: - - $ref: "#/components/parameters/ComfyUserHeader" - - name: path - in: query - schema: - type: string - description: Directory path relative to user data root - responses: - "200": - description: File listing with metadata - content: - application/json: - schema: - type: array - items: - type: object - properties: - name: - type: string - path: - type: string - type: - type: string - enum: [file, directory] - size: - type: integer - modified: - type: number - description: Unix timestamp - - /api/userdata/{file}: - get: - operationId: getUserdataFile - tags: [userdata] - summary: Read a userdata file - description: Reads the contents of a file from the authenticated user's data directory. - parameters: - - $ref: "#/components/parameters/ComfyUserHeader" - - name: file - in: path - required: true - schema: - type: string - description: File path relative to user data directory - responses: - "200": - description: File content - content: - application/octet-stream: - schema: - type: string - format: binary - "404": - description: File not found - post: - operationId: writeUserdataFile - tags: [userdata] - summary: Write or create a userdata file - description: Writes (creates or replaces) a file in the authenticated user's data directory. - parameters: - - $ref: "#/components/parameters/ComfyUserHeader" - - name: file - in: path - required: true - schema: - type: string - description: File path relative to user data directory - - name: overwrite - in: query - schema: - type: boolean - description: Allow overwriting existing files - - name: full_info - in: query - schema: - type: boolean - description: Return full file info in response - requestBody: - required: true - content: - application/octet-stream: - schema: - type: string - format: binary - application/json: - schema: {} - responses: - "200": - description: File written - content: - application/json: - schema: - $ref: "#/components/schemas/UserDataResponse" - "409": - description: File exists and overwrite not set - delete: - operationId: deleteUserdataFile - tags: [userdata] - summary: Delete a userdata file - description: Deletes a file from the authenticated user's data directory. - parameters: - - $ref: "#/components/parameters/ComfyUserHeader" - - name: file - in: path - required: true - schema: - type: string - description: File path relative to user data directory - responses: - "204": - description: File deleted - "404": - description: File not found - - /api/userdata/{file}/move/{dest}: - post: - operationId: moveUserdataFile - tags: [userdata] - summary: Move or rename a userdata file - description: Renames or moves a file within the authenticated user's data directory. - parameters: - - $ref: "#/components/parameters/ComfyUserHeader" - - name: file - in: path - required: true - schema: - type: string - description: Source file path - - name: dest - in: path - required: true - schema: - type: string - description: Destination file path - - name: overwrite - in: query - schema: - type: boolean - description: Allow overwriting at destination - - name: full_info - in: query - schema: - type: boolean - description: Return full file info in response - responses: - "200": - description: File moved - content: - application/json: - schema: - $ref: "#/components/schemas/UserDataResponse" - "404": - description: Source file not found - "409": - description: Destination exists and overwrite not set - - # --------------------------------------------------------------------------- - # Settings - # --------------------------------------------------------------------------- - /api/settings: - get: - operationId: getSettings - tags: [settings] - summary: Get all user settings - description: Returns all settings for the authenticated user. - parameters: - - $ref: "#/components/parameters/ComfyUserHeader" - responses: - "200": - description: Settings object - content: - application/json: - schema: - type: object - additionalProperties: true - post: - operationId: updateSettings - tags: [settings] - summary: Update user settings (partial merge) - description: Replaces the authenticated user's settings with the provided object. - parameters: - - $ref: "#/components/parameters/ComfyUserHeader" - requestBody: - required: true - content: - application/json: - schema: - type: object - additionalProperties: true - description: Partial settings to merge - responses: - "200": - description: Settings updated - - /api/settings/{id}: - get: - operationId: getSetting - tags: [settings] - summary: Get a single setting by key - description: Returns the value of a single setting, identified by key. - parameters: - - $ref: "#/components/parameters/ComfyUserHeader" - - name: id - in: path - required: true - schema: - type: string - description: Setting key - responses: - "200": - description: Setting value (null if the setting does not exist) - content: - application/json: - schema: - nullable: true - description: The setting value (any JSON type), or null if not set - post: - operationId: updateSetting - tags: [settings] - summary: Set a single setting value - description: Sets the value of a single setting, identified by key. - parameters: - - $ref: "#/components/parameters/ComfyUserHeader" - - name: id - in: path - required: true - schema: - type: string - description: Setting key - requestBody: - required: true - content: - application/json: - schema: - description: The setting value (any JSON type) - responses: - "200": - description: Setting updated - - # --------------------------------------------------------------------------- - # Extensions / Templates / i18n - # --------------------------------------------------------------------------- - /api/extensions: - get: - operationId: getExtensions - tags: [extensions] - summary: List frontend extension JS file paths - description: Returns the list of frontend extension JS URLs registered by custom nodes, to be loaded by the frontend on startup. - responses: - "200": - description: Array of JS file paths - content: - application/json: - schema: - type: array - items: - type: string - description: Relative path to extension JS file - - /api/workflow_templates: - get: - operationId: getWorkflowTemplates - tags: [extensions] - summary: Get workflow template mappings - description: Returns a map of custom node names to their provided workflow template names. - responses: - "200": - description: Template mappings - content: - application/json: - schema: - type: object - additionalProperties: - type: array - items: - type: string - description: Map of node pack name to array of template names - - /api/i18n: - get: - operationId: getI18n - tags: [extensions] - summary: Get internationalisation translation strings - description: Returns the URLs of translation files contributed by custom nodes, keyed by locale. - responses: - "200": - description: Translation map - content: - application/json: - schema: - type: object - additionalProperties: true - description: Nested map of locale to translation key-value pairs - - # --------------------------------------------------------------------------- - # Subgraphs - # --------------------------------------------------------------------------- - /api/global_subgraphs: - get: - operationId: getGlobalSubgraphs - tags: [subgraph] - summary: List global subgraph blueprints - description: Returns a dictionary of subgraph IDs to their metadata. - responses: - "200": - description: Subgraph metadata dictionary - content: - application/json: - schema: - type: object - additionalProperties: - $ref: "#/components/schemas/GlobalSubgraphInfo" - - /api/global_subgraphs/{id}: - get: - operationId: getGlobalSubgraph - tags: [subgraph] - summary: Get a global subgraph with full data - description: Returns the blueprint for a globally-registered subgraph, used by the frontend to materialize the subgraph node. - parameters: - - name: id - in: path - required: true - schema: - type: string - description: Subgraph identifier - responses: - "200": - description: Full subgraph data - content: - application/json: - schema: - $ref: "#/components/schemas/GlobalSubgraphData" - "404": - description: Subgraph not found - - # --------------------------------------------------------------------------- - # Node Replacements - # --------------------------------------------------------------------------- - /api/node_replacements: - get: - operationId: getNodeReplacements - tags: [node] - summary: Get node replacement mappings - description: | - Returns a dictionary mapping deprecated or replaced node class names - to their replacement node information. - responses: - "200": - description: Replacement mappings - content: - application/json: - schema: - type: object - additionalProperties: true - - # --------------------------------------------------------------------------- - # Internal (x-internal: true) - # --------------------------------------------------------------------------- - /internal/logs: - get: - operationId: getInternalLogs - tags: [internal] - summary: Get server logs as text - description: Returns structured ComfyUI log entries from the in-memory log buffer. - x-internal: true - responses: - "200": - description: Log text - content: - text/plain: - schema: - type: string - - /internal/logs/raw: - get: - operationId: getInternalLogsRaw - tags: [internal] - summary: Get raw structured log entries - description: Returns the raw ComfyUI log buffer as text, together with metadata about the current size limit. - x-internal: true - responses: - "200": - description: Structured log data - content: - application/json: - schema: - type: object - properties: - entries: - type: array - items: - type: object - properties: - t: - type: number - description: Timestamp - m: - type: string - description: Message - size: - type: object - properties: - cols: - type: integer - rows: - type: integer - - /internal/logs/subscribe: - patch: - operationId: subscribeToLogs - tags: [internal] - summary: Subscribe or unsubscribe a WebSocket client to log streaming - description: Subscribes or unsubscribes the current client from live log streaming over the WebSocket. - x-internal: true - requestBody: - required: true - content: - application/json: - schema: - type: object - required: - - clientId - - enabled - properties: - clientId: - type: string - description: WebSocket client ID - enabled: - type: boolean - description: Enable or disable log streaming for this client - responses: - "200": - description: Subscription updated - - /internal/folder_paths: - get: - operationId: getInternalFolderPaths - tags: [internal] - summary: Get configured folder paths - description: Returns the filesystem paths ComfyUI is configured to load models and other assets from, keyed by folder type. - x-internal: true - responses: - "200": - description: Dictionary of folder type to paths - content: - application/json: - schema: - type: object - additionalProperties: - type: array - items: - type: array - items: - type: string - description: Map of folder type name to list of [path, ...] entries - - /internal/files/{directory_type}: - get: - operationId: getInternalFiles - tags: [internal] - summary: List files in a directory type - description: Lists the files present in one of ComfyUI's known directories (input, output, or temp). - x-internal: true - parameters: - - name: directory_type - in: path - required: true - schema: - type: string - description: Directory type (e.g. output, input, temp) - responses: - "200": - description: Array of filenames - content: - application/json: - schema: - type: array - items: - type: string - - # --------------------------------------------------------------------------- - # Assets (x-feature-gate: enable-assets) - # --------------------------------------------------------------------------- - /api/assets/hash/{hash}: - head: - operationId: checkAssetByHash - tags: [assets] - summary: Check if an asset with the given hash exists - description: Returns 204 if an asset with the given content hash already exists, 404 otherwise. Used by clients to deduplicate uploads before transferring bytes. - x-feature-gate: enable-assets - parameters: - - name: hash - in: path - required: true - schema: - type: string - description: "Blake3 hash of the asset (e.g. blake3:abc123...)" - responses: - "200": - description: Asset exists - "404": - description: No asset with this hash - - /api/assets: - get: - operationId: listAssets - tags: [assets] - summary: List assets with filtering and pagination - description: Returns a paginated list of assets, optionally filtered by tags, name, or other query parameters. - x-feature-gate: enable-assets - parameters: - - name: limit - in: query - schema: - type: integer - default: 50 - - name: offset - in: query - schema: - type: integer - default: 0 - - name: include_tags - in: query - schema: - type: array - items: - type: string - style: form - explode: true - description: Tags that assets must have (AND logic) - - name: exclude_tags - in: query - schema: - type: array - items: - type: string - style: form - explode: true - description: Tags that assets must not have - - name: name_contains - in: query - schema: - type: string - description: Filter assets whose name contains this substring - - name: metadata_filter - in: query - schema: - type: string - description: JSON-encoded metadata key/value filter - - name: sort - in: query - schema: - type: string - description: Field to sort by - - name: order - in: query - schema: - type: string - enum: [asc, desc] - description: Sort direction - - name: include_public - in: query - schema: - type: boolean - x-runtime: [cloud] - description: "[cloud-only] Include workspace-public assets in addition to the caller's own." - - name: asset_hash - in: query - schema: - type: string - x-runtime: [cloud] - description: "[cloud-only] Filter by exact content hash." - responses: - "200": - description: Asset list - content: - application/json: - schema: - $ref: "#/components/schemas/ListAssetsResponse" - post: - operationId: createAsset - tags: [assets] - summary: Upload a new asset - description: Uploads a new asset (binary content plus metadata) and registers it in the asset database. - x-feature-gate: enable-assets - requestBody: - required: true - content: - multipart/form-data: - schema: - type: object - required: - - file - properties: - file: - type: string - format: binary - description: Asset file to upload - name: - type: string - description: Display name for the asset - tags: - type: string - description: Comma-separated tags - user_metadata: - type: string - description: JSON-encoded user metadata - hash: - type: string - description: "Blake3 hash of the file content (e.g. blake3:abc123...)" - mime_type: - type: string - description: MIME type of the file (overrides auto-detected type) - preview_id: - type: string - format: uuid - description: ID of an existing asset to use as the preview image - id: - type: string - format: uuid - nullable: true - x-runtime: [cloud] - description: "[cloud-only] Client-supplied asset ID for idempotent creation. If an asset with this ID already exists, the existing asset is returned." - application/json: - schema: - type: object - x-runtime: [cloud] - description: "[cloud-only] URL-based asset upload. Caller supplies a URL instead of a file body; the server fetches the content." - required: - - url - properties: - url: - type: string - format: uri - description: "[cloud-only] URL of the file to import as an asset" - name: - type: string - description: Display name for the asset - tags: - type: string - description: Comma-separated tags - user_metadata: - type: string - description: JSON-encoded user metadata - hash: - type: string - description: "Blake3 hash of the file content (e.g. blake3:abc123...)" - mime_type: - type: string - description: MIME type of the file (overrides auto-detected type) - preview_id: - type: string - format: uuid - description: ID of an existing asset to use as the preview image - id: - type: string - format: uuid - nullable: true - x-runtime: [cloud] - description: "[cloud-only] Client-supplied asset ID for idempotent creation. If an asset with this ID already exists, the existing asset is returned." - responses: - "201": - description: Asset created - content: - application/json: - schema: - $ref: "#/components/schemas/AssetCreated" - - /api/assets/from-hash: - post: - operationId: createAssetFromHash - tags: [assets] - summary: Create an asset reference from an existing hash - description: Registers a new asset that references existing content by hash, without re-uploading the bytes. - x-feature-gate: enable-assets - requestBody: - required: true - content: - application/json: - schema: - type: object - required: - - hash - - name - properties: - hash: - type: string - description: Blake3 hash of existing content - name: - type: string - description: Display name - tags: - type: array - items: - type: string - user_metadata: - type: object - additionalProperties: true - mime_type: - type: string - nullable: true - x-runtime: [cloud] - description: "[cloud-only] MIME type of the content, so the type is preserved without re-inspecting content. Ignored by local ComfyUI." - responses: - "201": - description: Asset created from hash - content: - application/json: - schema: - $ref: "#/components/schemas/AssetCreated" - - /api/assets/{id}: - get: - operationId: getAsset - tags: [assets] - summary: Get asset metadata - description: Returns the metadata for a single asset. - x-feature-gate: enable-assets - parameters: - - name: id - in: path - description: The asset ID. - required: true - schema: - type: string - format: uuid - responses: - "200": - description: Asset metadata - content: - application/json: - schema: - $ref: "#/components/schemas/Asset" - "404": - description: Asset not found - put: - operationId: updateAsset - tags: [assets] - summary: Update asset metadata - description: Updates the mutable metadata of an asset (name, tags, etc.). Binary content is immutable. - x-feature-gate: enable-assets - parameters: - - name: id - in: path - description: The asset ID. - required: true - schema: - type: string - format: uuid - requestBody: - required: true - content: - application/json: - schema: - type: object - properties: - name: - type: string - description: New display name for the asset - user_metadata: - type: object - additionalProperties: true - description: Custom user metadata to set - preview_id: - type: string - format: uuid - description: ID of the asset to use as the preview - mime_type: - type: string - nullable: true - x-runtime: [cloud] - description: "[cloud-only] MIME type override when auto-detection was wrong. Ignored by local ComfyUI." - responses: - "200": - description: Asset updated - content: - application/json: - schema: - $ref: "#/components/schemas/AssetUpdated" - delete: - operationId: deleteAsset - tags: [assets] - summary: Delete an asset - description: Removes an asset entry. Depending on the server configuration, the underlying content may also be deleted. - x-feature-gate: enable-assets - parameters: - - name: id - in: path - description: The asset ID. - required: true - schema: - type: string - format: uuid - - name: delete_content - in: query - schema: - type: boolean - description: Also delete the underlying content file - responses: - "204": - description: Asset deleted - - /api/assets/{id}/content: - get: - operationId: getAssetContent - tags: [assets] - summary: Download asset file content - description: Returns the binary content of an asset. Supports range requests. - x-feature-gate: enable-assets - parameters: - - name: id - in: path - description: The asset ID. - required: true - schema: - type: string - format: uuid - responses: - "200": - description: Asset file content - content: - application/octet-stream: - schema: - type: string - format: binary - "404": - description: Asset not found - - /api/assets/{id}/tags: - post: - operationId: addAssetTags - tags: [assets] - summary: Add tags to an asset - description: Adds one or more tags to an asset. - x-feature-gate: enable-assets - parameters: - - name: id - in: path - description: The asset ID. - required: true - schema: - type: string - format: uuid - requestBody: - required: true - content: - application/json: - schema: - type: object - required: - - tags - properties: - tags: - type: array - items: - type: string - responses: - "200": - description: Tags added - content: - application/json: - schema: - $ref: "#/components/schemas/TagsModificationResponse" - delete: - operationId: removeAssetTags - tags: [assets] - summary: Remove tags from an asset - description: Removes one or more tags from an asset. - x-feature-gate: enable-assets - parameters: - - name: id - in: path - description: The asset ID. - required: true - schema: - type: string - format: uuid - requestBody: - required: true - content: - application/json: - schema: - type: object - required: - - tags - properties: - tags: - type: array - items: - type: string - responses: - "200": - description: Tags removed - content: - application/json: - schema: - $ref: "#/components/schemas/TagsModificationResponse" - - /api/tags: - get: - operationId: listTags - tags: [assets] - summary: List all known tags with counts - description: Returns the list of all tags known to the asset database, with counts. - x-feature-gate: enable-assets - parameters: - - name: limit - in: query - schema: - type: integer - - name: offset - in: query - schema: - type: integer - - name: search - in: query - schema: - type: string - description: Search term for tag name - responses: - "200": - description: Tag list - content: - application/json: - schema: - $ref: "#/components/schemas/ListTagsResponse" - - /api/assets/tags/refine: - get: - operationId: refineAssetTags - tags: [assets] - summary: Get tag counts for assets matching current filters - description: Returns suggested additional tags that would refine a filtered asset query, together with the count of assets each tag would select. - x-feature-gate: enable-assets - parameters: - - name: include_tags - in: query - schema: - type: array - items: - type: string - style: form - explode: true - description: Tags that assets must have (AND logic) - - name: exclude_tags - in: query - schema: - type: array - items: - type: string - style: form - explode: true - description: Tags that assets must not have - - name: name_contains - in: query - schema: - type: string - description: Filter assets whose name contains this substring - - name: metadata_filter - in: query - schema: - type: string - description: JSON-encoded metadata key/value filter - - name: limit - in: query - schema: - type: integer - - name: offset - in: query - schema: - type: integer - - name: sort - in: query - schema: - type: string - description: Field to sort by - - name: order - in: query - schema: - type: string - enum: [asc, desc] - description: Sort direction - responses: - "200": - description: Tag histogram - content: - application/json: - schema: - $ref: "#/components/schemas/AssetTagHistogramResponse" - - /api/assets/seed: - post: - operationId: seedAssets - tags: [assets] - summary: Trigger asset scan/seed from filesystem - description: Starts a background job that scans the configured directories and registers any assets not yet present in the asset database. - x-feature-gate: enable-assets - requestBody: - required: false - content: - application/json: - schema: - type: object - properties: - roots: - type: array - items: - type: string - description: Root folder paths to scan (if omitted, scans all) - responses: - "200": - description: Seed started - content: - application/json: - schema: - type: object - properties: - status: - type: string - - /api/assets/seed/status: - get: - operationId: getAssetSeedStatus - tags: [assets] - summary: Get asset scan progress - description: Returns the progress and status of the most recently-started asset seed job. - x-feature-gate: enable-assets - responses: - "200": - description: Scan progress - content: - application/json: - schema: - type: object - additionalProperties: true - description: Scan progress details (files scanned, total, status, etc.) - - /api/assets/seed/cancel: - post: - operationId: cancelAssetSeed - tags: [assets] - summary: Cancel an in-progress asset scan - description: Requests cancellation of the currently-running asset seed job. - x-feature-gate: enable-assets - responses: - "200": - description: Scan cancelled - content: - application/json: - schema: - type: object - properties: - status: - type: string - - /api/assets/prune: - post: - operationId: pruneAssets - tags: [assets] - summary: Mark assets whose backing files no longer exist on disk - description: Starts a background job that removes asset entries whose underlying content no longer exists on disk. - x-feature-gate: enable-assets - responses: - "200": - description: Prune result - content: - application/json: - schema: - type: object - properties: - status: - type: string - marked: - type: integer - description: Number of assets marked as missing - - # =========================================================================== - # Cloud-runtime FE-facing operations - # - # These operations are served by the cloud runtime. The local runtime returns - # 404 for all of these paths. Each operation is tagged x-runtime: [cloud]. - # =========================================================================== - - # --------------------------------------------------------------------------- - # Jobs / prompts (cloud) - # --------------------------------------------------------------------------- - /api/jobs/{job_id}/cancel: - post: - operationId: cancelJob - tags: [queue] - summary: Cancel a running or pending job - description: "[cloud-only] Requests cancellation of a job. If the job is currently executing, execution is interrupted. If it is pending in the queue, it is removed." - x-runtime: [cloud] - parameters: - - name: job_id - in: path - required: true - schema: - type: string - format: uuid - description: The job ID to cancel. - responses: - "200": - description: Cancellation accepted - content: - application/json: - schema: - $ref: "#/components/schemas/CloudJobStatus" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/job/{job_id}/status: - get: - operationId: getCloudJobStatus - tags: [queue] - summary: Get status of a cloud job - deprecated: true - description: | - **Deprecated.** This endpoint is superseded by `GET /api/jobs/{job_id}`. - Clients should migrate; the endpoint is retained for backward - compatibility but will be removed in a future release. - x-runtime: [cloud] - parameters: - - name: job_id - in: path - required: true - schema: - type: string - format: uuid - description: The job ID to check status for. - responses: - "200": - description: Job status - content: - application/json: - schema: - $ref: "#/components/schemas/CloudJobStatus" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/prompt/{prompt_id}: - get: - operationId: getCloudPrompt - tags: [prompt] - summary: Get a cloud prompt by ID - description: "[cloud-only] Returns the full prompt record for a cloud-executed prompt, including the submitted workflow graph and execution metadata." - x-runtime: [cloud] - parameters: - - name: prompt_id - in: path - required: true - schema: - type: string - format: uuid - description: The prompt ID to fetch. - responses: - "200": - description: Cloud prompt detail - content: - application/json: - schema: - $ref: "#/components/schemas/CloudPrompt" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/history_v2: - get: - operationId: getHistoryV2 - tags: [history] - summary: Get paginated execution history (v2) - deprecated: true - description: | - **Deprecated.** This endpoint is superseded by `GET /api/jobs`. - Clients should migrate; the endpoint is retained for backward - compatibility but will be removed in a future release. - x-runtime: [cloud] - parameters: - - name: limit - in: query - schema: - type: integer - default: 20 - description: Maximum number of results - - name: offset - in: query - schema: - type: integer - default: 0 - description: Pagination offset - - name: status - in: query - schema: - type: string - description: Filter by execution status - responses: - "200": - description: History list - content: - application/json: - schema: - $ref: "#/components/schemas/HistoryV2Response" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/history_v2/{prompt_id}: - get: - operationId: getHistoryV2ByPromptId - tags: [history] - summary: Get v2 history for a specific prompt - deprecated: true - description: | - **Deprecated.** This endpoint is superseded by `GET /api/jobs/{prompt_id}`. - Clients should migrate; the endpoint is retained for backward - compatibility but will be removed in a future release. - x-runtime: [cloud] - parameters: - - name: prompt_id - in: path - required: true - schema: - type: string - format: uuid - description: The prompt ID to fetch history for. - responses: - "200": - description: History entry - content: - application/json: - schema: - $ref: "#/components/schemas/HistoryV2Entry" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/logs: - get: - operationId: getCloudLogs - tags: [system] - summary: Get cloud execution logs - deprecated: true - description: | - **Deprecated.** This endpoint returns a static placeholder response and - provides no real log data. It is retained only to avoid breaking clients - that still call it. Clients should remove their dependency; the endpoint - will be removed in a future release. - x-runtime: [cloud] - parameters: - - name: job_id - in: query - schema: - type: string - description: Filter logs by job ID - - name: limit - in: query - schema: - type: integer - default: 100 - description: Maximum number of log entries - - name: offset - in: query - schema: - type: integer - default: 0 - description: Pagination offset - responses: - "200": - description: Log entries - content: - application/json: - schema: - $ref: "#/components/schemas/CloudLogsResponse" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - # --------------------------------------------------------------------------- - # Assets extensions (cloud) - # --------------------------------------------------------------------------- - /api/assets/download: - post: - operationId: downloadAssets - tags: [assets] - summary: Download assets to cloud runtime - description: "[cloud-only] Initiates a download of one or more assets to the cloud runtime environment. Returns a task ID for tracking download progress via WebSocket." - x-runtime: [cloud] - requestBody: - required: true - content: - application/json: - schema: - type: object - required: - - assets - properties: - assets: - type: array - items: - $ref: "#/components/schemas/AssetDownloadRequest" - description: Assets to download - responses: - "200": - description: Download initiated - content: - application/json: - schema: - type: object - properties: - task_id: - type: string - description: Task ID for tracking progress via WebSocket - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/assets/export: - post: - operationId: exportAssets - tags: [assets] - summary: Export assets as a downloadable archive - description: "[cloud-only] Initiates a bulk export of assets. Returns a task ID for tracking progress via WebSocket. When complete, the export can be downloaded via the exports endpoint." - x-runtime: [cloud] - requestBody: - required: true - content: - application/json: - schema: - type: object - required: - - asset_ids - properties: - asset_ids: - type: array - items: - type: string - format: uuid - description: IDs of assets to export - export_name: - type: string - description: Name for the export archive - responses: - "200": - description: Export initiated - content: - application/json: - schema: - type: object - properties: - task_id: - type: string - export_name: - type: string - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/assets/exports/{exportName}: - get: - operationId: getAssetExport - tags: [assets] - summary: Download a completed asset export - description: "[cloud-only] Returns the archive file for a completed asset export." - x-runtime: [cloud] - parameters: - - name: exportName - in: path - required: true - schema: - type: string - description: Name of the export to download - responses: - "200": - description: Export archive file - content: - application/zip: - schema: - type: string - format: binary - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/assets/from-workflow: - post: - operationId: createAssetsFromWorkflow - tags: [assets] - summary: Create asset records from a workflow execution - description: "[cloud-only] Registers output files from a workflow execution as assets in the asset database." - x-runtime: [cloud] - requestBody: - required: true - content: - application/json: - schema: - type: object - required: - - prompt_id - properties: - prompt_id: - type: string - format: uuid - description: Prompt ID whose outputs should be registered as assets - tags: - type: array - items: - type: string - description: Tags to apply to the created assets - responses: - "201": - description: Assets created - content: - application/json: - schema: - type: object - properties: - assets: - type: array - items: - $ref: "#/components/schemas/Asset" - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/assets/import: - post: - operationId: importPublishedAssets - tags: [assets] - summary: "[cloud-only] Import published assets into the caller's library" - description: | - [cloud-only] Imports the specified published assets into the caller's asset library. New DB records reference the same storage objects; no file copying occurs. Assets the caller already owns (by hash) are deduplicated. The `id` field on each returned `AssetInfo` is the caller's newly-created private asset ID, not the published asset ID supplied in the request. - x-runtime: [cloud] - requestBody: - required: true - content: - application/json: - schema: - $ref: "#/components/schemas/ImportPublishedAssetsRequest" - responses: - "200": - description: Successfully imported assets - content: - application/json: - schema: - $ref: "#/components/schemas/ImportPublishedAssetsResponse" - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/assets/remote-metadata: - get: - operationId: getAssetRemoteMetadata - tags: [assets] - summary: Fetch metadata for a remote asset URL - description: "[cloud-only] Fetches and returns metadata (content type, size, filename) for a remote URL without downloading the full content." - x-runtime: [cloud] - parameters: - - name: url - in: query - required: true - schema: - type: string - format: uri - description: URL to inspect - responses: - "200": - description: Remote metadata - content: - application/json: - schema: - $ref: "#/components/schemas/RemoteAssetMetadata" - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - # --------------------------------------------------------------------------- - # Custom nodes / hub (cloud) - # --------------------------------------------------------------------------- - /api/experiment/nodes: - get: - operationId: getNodeInfoSchema - tags: [runtime-only] - summary: Get pre-rendered node info schema - description: "[cloud-only] Returns the static ComfyUI object_info schema, identical for every caller, rendered once at startup with empty model/user-file context. Served by a raw HTTP handler that writes pre-rendered bytes with ETag + Cache-Control validators for RFC 7232 conditional GETs." - x-runtime: [cloud] - parameters: - - name: If-None-Match - in: header - required: false - schema: - type: string - description: Entity tag previously returned by this endpoint. When present and matching, the server returns 304 Not Modified. - responses: - "200": - description: Node info schema - headers: - ETag: - schema: - type: string - description: Entity tag for conditional request validation - Cache-Control: - schema: - type: string - description: Cache directives for the response - content: - application/json: - schema: - type: object - additionalProperties: - $ref: "#/components/schemas/NodeInfo" - "304": - description: Not Modified — returned when the client sends a matching If-None-Match header - post: - operationId: installCloudNode - tags: [node] - summary: Install a custom node package - description: "[cloud-only] Installs a custom node package in the cloud runtime by ID or repository URL." - x-runtime: [cloud] - requestBody: - required: true - content: - application/json: - schema: - type: object - required: - - id - properties: - id: - type: string - description: Node package ID or repository URL - version: - type: string - description: Specific version to install - responses: - "200": - description: Node installed - content: - application/json: - schema: - $ref: "#/components/schemas/CloudNode" - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/experiment/nodes/{id}: - get: - operationId: getNodeByID - tags: [runtime-only] - summary: Get a single node definition by ID - description: "[cloud-only] Returns one node's definition from the pre-indexed object_info schema. Served by a raw HTTP handler that writes pre-rendered bytes with ETag + Cache-Control validators for RFC 7232 conditional GETs." - x-runtime: [cloud] - parameters: - - name: id - in: path - required: true - schema: - type: string - description: Node class identifier - - name: If-None-Match - in: header - required: false - schema: - type: string - description: Entity tag previously returned by this endpoint. When present and matching, the server returns 304 Not Modified. - responses: - "200": - description: Single node definition - headers: - ETag: - schema: - type: string - description: Entity tag for conditional request validation - Cache-Control: - schema: - type: string - description: Cache directives for the response - content: - application/json: - schema: - $ref: "#/components/schemas/NodeInfo" - "304": - description: Not Modified — returned when the client sends a matching If-None-Match header - "404": - description: Node not found - delete: - operationId: uninstallCloudNode - tags: [node] - summary: Uninstall a custom node package - description: "[cloud-only] Removes a custom node package from the cloud runtime." - x-runtime: [cloud] - parameters: - - name: id - in: path - required: true - schema: - type: string - description: Custom node package ID - responses: - "204": - description: Node uninstalled - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/hub/assets/upload-url: - post: - operationId: getHubAssetUploadUrl - tags: [hub] - summary: Get a pre-signed upload URL for a hub asset - description: "[cloud-only] Returns a pre-signed URL that can be used to upload an asset file directly to storage." - x-runtime: [cloud] - requestBody: - required: true - content: - application/json: - schema: - type: object - required: - - filename - - content_type - properties: - filename: - type: string - description: Name of the file to upload - content_type: - type: string - description: MIME type of the file - size: - type: integer - format: int64 - description: File size in bytes - responses: - "200": - description: Upload URL - content: - application/json: - schema: - type: object - properties: - upload_url: - type: string - format: uri - description: Pre-signed upload URL - asset_url: - type: string - format: uri - description: Public URL after upload completes - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/hub/labels: - get: - operationId: listHubLabels - tags: [hub] - summary: List available hub labels - description: "[cloud-only] Returns the list of labels/categories available for tagging hub content." - x-runtime: [cloud] - responses: - "200": - description: Label list - content: - application/json: - schema: - type: array - items: - $ref: "#/components/schemas/HubLabel" - - /api/hub/profiles: - get: - operationId: listHubProfiles - tags: [hub] - summary: List hub user profiles - description: "[cloud-only] Returns a paginated list of public hub user profiles." - x-runtime: [cloud] - parameters: - - name: limit - in: query - schema: - type: integer - description: Maximum number of results - - name: offset - in: query - schema: - type: integer - description: Pagination offset - - name: search - in: query - schema: - type: string - description: Search by username or display name - responses: - "200": - description: Profile list - content: - application/json: - schema: - type: object - properties: - profiles: - type: array - items: - $ref: "#/components/schemas/HubProfile" - total: - type: integer - has_more: - type: boolean - post: - operationId: createHubProfile - tags: [hub] - summary: Create a Hub profile - description: "[cloud-only] Creates a hub profile for the specified workspace. Username is immutable after creation." - x-runtime: [cloud] - requestBody: - required: true - content: - application/json: - schema: - $ref: "#/components/schemas/CreateHubProfileRequest" - responses: - "201": - description: Hub profile created - content: - application/json: - schema: - $ref: "#/components/schemas/HubProfile" - "400": - description: Bad request (e.g. invalid username) - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "409": - description: Username already taken or profile already exists - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/hub/profiles/{username}: - get: - operationId: getHubProfile - tags: [hub] - summary: Get a hub profile by username - description: "[cloud-only] Returns the public hub profile for the given username." - x-runtime: [cloud] - parameters: - - name: username - in: path - required: true - schema: - type: string - description: Hub username - responses: - "200": - description: Profile - content: - application/json: - schema: - $ref: "#/components/schemas/HubProfile" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/hub/profiles/check: - get: - operationId: checkHubProfileUsername - tags: [hub] - summary: Check if a hub username is available - description: "[cloud-only] Returns whether the given username is available for registration." - x-runtime: [cloud] - parameters: - - name: username - in: query - required: true - schema: - type: string - description: Username to check - responses: - "200": - description: Availability result - content: - application/json: - schema: - type: object - properties: - available: - type: boolean - username: - type: string - - /api/hub/profiles/me: - get: - operationId: getMyHubProfile - tags: [hub] - summary: Get the authenticated user's hub profile - description: "[cloud-only] Returns the hub profile of the currently authenticated user." - x-runtime: [cloud] - responses: - "200": - description: Profile - content: - application/json: - schema: - $ref: "#/components/schemas/HubProfile" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - put: - operationId: updateMyHubProfile - tags: [hub] - summary: Update the authenticated user's hub profile - description: "[cloud-only] Updates the hub profile of the currently authenticated user." - x-runtime: [cloud] - requestBody: - required: true - content: - application/json: - schema: - type: object - properties: - username: - type: string - display_name: - type: string - bio: - type: string - avatar_url: - type: string - format: uri - links: - type: array - items: - type: string - format: uri - responses: - "200": - description: Updated profile - content: - application/json: - schema: - $ref: "#/components/schemas/HubProfile" - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "409": - description: Conflict - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/hub/workflows: - get: - operationId: listHubWorkflows - tags: [hub] - summary: List published hub workflows - description: "[cloud-only] Returns a paginated list of publicly shared workflows on the hub." - x-runtime: [cloud] - parameters: - - name: limit - in: query - schema: - type: integer - description: Maximum number of results - - name: offset - in: query - schema: - type: integer - description: Pagination offset - - name: sort - in: query - schema: - type: string - description: Sort field (e.g. created_at, likes) - - name: order - in: query - schema: - type: string - enum: [asc, desc] - description: Sort direction - - name: search - in: query - schema: - type: string - description: Search by title or description - - name: labels - in: query - schema: - type: string - description: Filter by label IDs (comma-separated) - responses: - "200": - description: Hub workflow list - content: - application/json: - schema: - $ref: "#/components/schemas/HubWorkflowList" - post: - operationId: publishHubWorkflow - tags: [hub] - summary: Publish a workflow to the hub - description: "[cloud-only] Publishes a workflow to the hub with metadata, thumbnail, and sample images." - x-runtime: [cloud] - requestBody: - required: true - content: - application/json: - schema: - $ref: "#/components/schemas/PublishHubWorkflowRequest" - responses: - "200": - description: Workflow published to hub - content: - application/json: - schema: - $ref: "#/components/schemas/HubWorkflowDetail" - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Workflow or profile not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/hub/workflows/{share_id}: - get: - operationId: getHubWorkflow - tags: [hub] - summary: Get a published hub workflow by share ID - description: "[cloud-only] Returns the full details of a published workflow on the hub." - x-runtime: [cloud] - parameters: - - name: share_id - in: path - required: true - schema: - type: string - description: Workflow share ID - responses: - "200": - description: Hub workflow - content: - application/json: - schema: - $ref: "#/components/schemas/HubWorkflow" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - delete: - operationId: deleteHubWorkflow - tags: [hub] - summary: Unpublish a workflow from the hub - description: "[cloud-only] Removes a workflow from the hub listing." - x-runtime: [cloud] - parameters: - - name: share_id - in: path - required: true - schema: - type: string - description: Workflow share ID - responses: - "204": - description: Successfully unpublished - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Workflow not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/hub/workflows/index: - get: - operationId: getHubWorkflowIndex - tags: [hub] - summary: Get the hub workflow index - description: "[cloud-only] Returns the lightweight index of all hub workflows for client-side search and navigation." - x-runtime: [cloud] - responses: - "200": - description: Workflow index - content: - application/json: - schema: - type: array - items: - $ref: "#/components/schemas/HubWorkflowIndexEntry" - - # --------------------------------------------------------------------------- - # Workflows (cloud) - # --------------------------------------------------------------------------- - /api/workflows: - get: - operationId: listCloudWorkflows - tags: [workflows] - summary: List cloud workflows - description: "[cloud-only] Returns a paginated list of the authenticated user's cloud workflows." - x-runtime: [cloud] - parameters: - - name: limit - in: query - schema: - type: integer - description: Maximum number of results - - name: offset - in: query - schema: - type: integer - description: Pagination offset - - name: sort - in: query - schema: - type: string - description: Sort field - - name: order - in: query - schema: - type: string - enum: [asc, desc] - description: Sort direction - - name: search - in: query - schema: - type: string - description: Search by workflow name - responses: - "200": - description: Workflow list - content: - application/json: - schema: - $ref: "#/components/schemas/CloudWorkflowList" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - post: - operationId: createCloudWorkflow - tags: [workflows] - summary: Create a new cloud workflow - description: "[cloud-only] Creates a new cloud workflow with the provided name and optional initial content." - x-runtime: [cloud] - requestBody: - required: true - content: - application/json: - schema: - type: object - required: - - name - properties: - name: - type: string - description: Workflow name - description: - type: string - description: Workflow description - content: - type: object - additionalProperties: true - description: Initial workflow graph JSON - responses: - "201": - description: Workflow created - content: - application/json: - schema: - $ref: "#/components/schemas/CloudWorkflow" - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/workflows/{workflow_id}: - get: - operationId: getCloudWorkflow - tags: [workflows] - summary: Get a cloud workflow by ID - description: "[cloud-only] Returns the metadata for a cloud workflow." - x-runtime: [cloud] - parameters: - - name: workflow_id - in: path - required: true - schema: - type: string - format: uuid - description: The workflow ID. - responses: - "200": - description: Workflow detail - content: - application/json: - schema: - $ref: "#/components/schemas/CloudWorkflow" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - patch: - operationId: updateCloudWorkflow - tags: [workflows] - summary: Update a cloud workflow - description: "[cloud-only] Updates the metadata (name, description) of an existing cloud workflow." - x-runtime: [cloud] - parameters: - - name: workflow_id - in: path - required: true - schema: - type: string - format: uuid - description: The workflow ID. - requestBody: - required: true - content: - application/json: - schema: - type: object - properties: - name: - type: string - description: - type: string - responses: - "200": - description: Workflow updated - content: - application/json: - schema: - $ref: "#/components/schemas/CloudWorkflow" - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - delete: - operationId: deleteCloudWorkflow - tags: [workflows] - summary: Delete a cloud workflow - description: "[cloud-only] Deletes a cloud workflow and all its versions." - x-runtime: [cloud] - parameters: - - name: workflow_id - in: path - required: true - schema: - type: string - format: uuid - description: The workflow ID. - responses: - "204": - description: Workflow deleted - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/workflows/{workflow_id}/content: - get: - operationId: getCloudWorkflowContent - tags: [workflows] - summary: Get the content of a cloud workflow - description: "[cloud-only] Returns the full workflow graph JSON for the latest version of a cloud workflow." - x-runtime: [cloud] - parameters: - - name: workflow_id - in: path - required: true - schema: - type: string - format: uuid - description: The workflow ID. - - name: version_id - in: query - schema: - type: string - description: Specific version ID to fetch - responses: - "200": - description: Workflow content - content: - application/json: - schema: - type: object - additionalProperties: true - description: The full workflow graph JSON - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - put: - operationId: updateCloudWorkflowContent - tags: [workflows] - summary: Update the content of a cloud workflow - description: "[cloud-only] Saves new workflow graph JSON as a new version of the cloud workflow." - x-runtime: [cloud] - parameters: - - name: workflow_id - in: path - required: true - schema: - type: string - format: uuid - description: The workflow ID. - requestBody: - required: true - content: - application/json: - schema: - type: object - additionalProperties: true - description: The workflow graph JSON to save - responses: - "200": - description: Content updated - content: - application/json: - schema: - $ref: "#/components/schemas/CloudWorkflowVersion" - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/workflows/{workflow_id}/fork: - post: - operationId: forkCloudWorkflow - tags: [workflows] - summary: Fork a cloud workflow - description: "[cloud-only] Creates a copy of a cloud workflow under the authenticated user's account." - x-runtime: [cloud] - parameters: - - name: workflow_id - in: path - required: true - schema: - type: string - format: uuid - description: The workflow ID to fork. - requestBody: - required: false - content: - application/json: - schema: - type: object - properties: - name: - type: string - description: Name for the forked workflow (defaults to original name) - responses: - "201": - description: Forked workflow - content: - application/json: - schema: - $ref: "#/components/schemas/CloudWorkflow" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/workflows/{workflow_id}/versions: - get: - operationId: listCloudWorkflowVersions - tags: [workflows] - summary: List versions of a cloud workflow - description: "[cloud-only] Returns the version history of a cloud workflow." - x-runtime: [cloud] - parameters: - - name: workflow_id - in: path - required: true - schema: - type: string - format: uuid - description: The workflow ID. - - name: limit - in: query - schema: - type: integer - description: Maximum number of results - - name: offset - in: query - schema: - type: integer - description: Pagination offset - responses: - "200": - description: Version list - content: - application/json: - schema: - type: object - properties: - versions: - type: array - items: - $ref: "#/components/schemas/CloudWorkflowVersion" - total: - type: integer - has_more: - type: boolean - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - post: - operationId: createCloudWorkflowVersion - tags: [workflows] - summary: Create a new cloud workflow version - description: "[cloud-only] Creates a new workflow version with updated workflow JSON. Uses optimistic concurrency via base_version." - x-runtime: [cloud] - parameters: - - name: workflow_id - in: path - required: true - schema: - type: string - format: uuid - description: The workflow ID. - requestBody: - required: true - content: - application/json: - schema: - $ref: "#/components/schemas/CreateWorkflowVersionRequest" - responses: - "201": - description: Version created - content: - application/json: - schema: - $ref: "#/components/schemas/WorkflowVersionResponse" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "403": - description: Forbidden — not the workflow owner - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "409": - description: Version conflict — base_version does not match latest - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/workflows/published/{share_id}: - get: - operationId: getPublishedWorkflow - tags: [workflows] - summary: Get a published workflow by share ID - description: "[cloud-only] Returns a publicly published cloud workflow by its share identifier." - x-runtime: [cloud] - parameters: - - name: share_id - in: path - required: true - schema: - type: string - description: The workflow share ID. - responses: - "200": - description: Published workflow - content: - application/json: - schema: - $ref: "#/components/schemas/CloudWorkflow" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - # --------------------------------------------------------------------------- - # Auth / session (cloud) - # --------------------------------------------------------------------------- - /api/auth/session: - get: - operationId: getAuthSession - tags: [auth] - summary: Get the current authentication session - description: "[cloud-only] Returns the current session state for the authenticated user, including user identity and active workspace." - x-runtime: [cloud] - responses: - "200": - description: Session info - content: - application/json: - schema: - $ref: "#/components/schemas/AuthSession" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - post: - operationId: createAuthSession - tags: [auth] - summary: Create a session cookie - description: "[cloud-only] Creates a session cookie from the bearer token in the Authorization header. Returns a Set-Cookie header with a secure HttpOnly session cookie. Cookie authentication is not allowed for this endpoint." - x-runtime: [cloud] - responses: - "200": - description: Session created - content: - application/json: - schema: - $ref: "#/components/schemas/CreateSessionResponse" - "400": - description: Bad request — invalid or expired ID token - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - delete: - operationId: deleteAuthSession - tags: [auth] - summary: Delete session cookie (logout) - description: "[cloud-only] Clears the session cookie and optionally revokes the session on the server." - x-runtime: [cloud] - responses: - "200": - description: Session deleted - content: - application/json: - schema: - $ref: "#/components/schemas/DeleteSessionResponse" - - /api/auth/token: - post: - operationId: createAuthToken - tags: [auth] - summary: Exchange credentials for an access token - description: "[cloud-only] Exchanges authentication credentials (e.g. an authorization code) for an access token." - x-runtime: [cloud] - requestBody: - required: true - content: - application/json: - schema: - type: object - required: - - grant_type - properties: - grant_type: - type: string - enum: [authorization_code, refresh_token] - description: OAuth2 grant type - code: - type: string - description: Authorization code (for authorization_code grant) - refresh_token: - type: string - description: Refresh token (for refresh_token grant) - redirect_uri: - type: string - format: uri - description: Redirect URI used in the authorization request - responses: - "200": - description: Token response - content: - application/json: - schema: - $ref: "#/components/schemas/AuthTokenResponse" - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /.well-known/jwks.json: - get: - operationId: getJwks - tags: [auth] - summary: Get JSON Web Key Set - description: "[cloud-only] Returns the JSON Web Key Set (JWKS) used to verify JWTs issued by the cloud authentication service." - x-runtime: [cloud] - responses: - "200": - description: JWKS - content: - application/json: - schema: - $ref: "#/components/schemas/JwksResponse" - - # --------------------------------------------------------------------------- - # OAuth 2.1 / RFC 7591 Dynamic Client Registration (cloud) - # --------------------------------------------------------------------------- - /.well-known/oauth-authorization-server: - get: - operationId: getOAuthAuthorizationServer - tags: [auth] - summary: "[cloud-only] OAuth 2.1 authorization-server metadata (RFC 8414)" - description: "[cloud-only] Public metadata document for OAuth 2.1 clients. Cached 5 minutes." - x-runtime: [cloud] - security: [] - responses: - "200": - description: Authorization-server metadata - content: - application/json: - schema: - $ref: "#/components/schemas/OAuthAuthorizationServerMetadata" - "404": - description: OAuth disabled - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /.well-known/oauth-protected-resource: - get: - operationId: getOAuthProtectedResource - tags: [auth] - summary: "[cloud-only] OAuth 2.1 protected-resource metadata (RFC 9728)" - description: "[cloud-only] Public metadata describing the currently advertised protected resource. Cached 5 minutes." - x-runtime: [cloud] - security: [] - responses: - "200": - description: Protected-resource metadata - content: - application/json: - schema: - $ref: "#/components/schemas/OAuthProtectedResourceMetadata" - "404": - description: OAuth disabled or no active resource configured - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /oauth/authorize: - get: - operationId: getOAuthAuthorize - tags: [auth] - summary: "[cloud-only] Begin or resume an OAuth 2.1 authorization request" - description: | - [cloud-only] Two modes: - - **Initial entry** (OAuth params present): validates client/redirect/resource/scopes, persists a server-side authorization-request row, and either redirects (no session / unverified email) to the configured frontend login URL carrying only the opaque `oauth_request_id`, or returns the JSON consent challenge for the frontend to render. - - **Resume** (`oauth_request_id` present): loads the server-side row, fails closed if expired/consumed/unknown, returns the JSON consent challenge. Browser-replayed OAuth params are intentionally ignored. - - The frontend renders the consent UI from the JSON payload and POSTs the user's decision back to this endpoint. - x-runtime: [cloud] - security: [] - parameters: - - { name: response_type, in: query, required: false, schema: { type: string } } - - { name: client_id, in: query, required: false, schema: { type: string } } - - { name: redirect_uri, in: query, required: false, schema: { type: string } } - - { name: scope, in: query, required: false, schema: { type: string } } - - name: state - in: query - required: false - schema: { type: string } - description: | - RFC 6749 §10.12 marks `state` as RECOMMENDED. Cloud hardening makes it REQUIRED on the initial-entry path (omitted only on the resume path where `oauth_request_id` is supplied instead). This parameter is `required: false` at the spec level only because the operation is dual-mode (initial entry vs. resume); the runtime rejects empty `state` on the initial-entry path with a stable `invalid_request` 400. - - { name: code_challenge, in: query, required: false, schema: { type: string } } - - { name: code_challenge_method, in: query, required: false, schema: { type: string } } - - { name: resource, in: query, required: false, schema: { type: string } } - - { name: oauth_request_id, in: query, required: false, schema: { type: string } } - responses: - "200": - description: Consent challenge payload (session present, email verified). Frontend renders the consent UI from this payload and POSTs back to /oauth/authorize. - content: - application/json: - schema: - $ref: "#/components/schemas/OAuthConsentChallenge" - "302": - description: Redirect to login (no session / unverified email) or to registered redirect_uri (pre-validated client error) - headers: - Location: - schema: - type: string - "400": - description: Invalid authorize request (pre-redirect failure — unknown client, redirect mismatch, malformed params) - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: OAuth disabled - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - post: - operationId: postOAuthAuthorize - tags: [auth] - summary: "[cloud-only] Submit OAuth consent decision" - description: | - [cloud-only] JSON-only consent submission. The handler verifies the per-row CSRF token, atomically marks the authorization request consumed (single-use covers both allow and deny paths), then returns the redirect URL the browser must navigate to. The URL contains either `code` + original `state` for allow, or the RFC 6749 §5.2 error and `state` for deny. - - Workspace membership is re-checked at submission time. Consent is persisted keyed by `(user_id, client_id, resource_id, workspace_id)`; broadening the previously approved scope set requires a fresh consent flow. - x-runtime: [cloud] - security: [] - requestBody: - required: true - content: - application/json: - schema: - type: object - required: [oauth_request_id, csrf_token, decision, workspace_id] - properties: - oauth_request_id: { type: string, format: uuid } - csrf_token: { type: string } - decision: { type: string, enum: [allow, deny] } - workspace_id: { type: string } - responses: - "200": - description: Redirect URL for the frontend to navigate to (allow → with code+state; deny → with error+state) - content: - application/json: - schema: - $ref: "#/components/schemas/OAuthAuthorizeRedirectResponse" - "400": - description: Bad request (CSRF mismatch, expired/consumed request, inaccessible workspace) - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "403": - description: Scope broadening on consent re-grant — fresh consent flow required - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: OAuth disabled - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /oauth/token: - post: - operationId: postOAuthToken - tags: [auth] - summary: "[cloud-only] Exchange authorization code or refresh token for a resource-bound access token" - description: | - [cloud-only] OAuth 2.1 token endpoint (RFC 6749 §3.2). Public clients only — `client_secret` is rejected. - - Two grant types are supported: - - `authorization_code` — exchanges the code minted by `/oauth/authorize` (with PKCE verifier) for an access token + first refresh token. Single-use; reuse fails closed. - - `refresh_token` — rotates the refresh token. Old token immediately invalid; presenting an already-rotated token revokes the entire token family and emits a security metric. - - Both grant types re-validate canonical user state, current workspace membership, and the resource's active flag at every mint. A code or refresh token bound to a deactivated resource fails closed. - - Errors follow RFC 6749 §5.2. Logs never contain raw codes, refresh tokens, or minted tokens. - - Per RFC 6749 §5.1, every 200 and 400 response carries `Cache-Control: no-store` and `Pragma: no-cache` so intermediaries cannot cache token-bearing or state-change-reason responses. - x-runtime: [cloud] - security: [] - requestBody: - required: true - content: - application/x-www-form-urlencoded: - schema: - type: object - required: [grant_type, client_id] - properties: - grant_type: { type: string, enum: [authorization_code, refresh_token] } - client_id: { type: string } - code: { type: string } - redirect_uri: { type: string } - code_verifier: { type: string } - refresh_token: { type: string } - scope: { type: string } - client_secret: { type: string } - responses: - "200": - description: New token pair - headers: - Cache-Control: - schema: - type: string - description: 'Always "no-store" per RFC 6749 §5.1' - Pragma: - schema: - type: string - description: 'Always "no-cache" per RFC 6749 §5.1' - content: - application/json: - schema: - $ref: "#/components/schemas/OAuthTokenResponse" - "400": - description: RFC 6749 §5.2 error - headers: - Cache-Control: - schema: - type: string - description: 'Always "no-store" per RFC 6749 §5.1' - Pragma: - schema: - type: string - description: 'Always "no-cache" per RFC 6749 §5.1' - content: - application/json: - schema: - $ref: "#/components/schemas/OAuthTokenError" - "404": - description: OAuth disabled - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /oauth/register: - post: - operationId: postOAuthRegister - tags: [auth] - summary: "[cloud-only] Dynamic Client Registration (RFC 7591)" - description: | - [cloud-only] Public, unauthenticated, insert-only RFC 7591 §3.1 client registration. Used by MCP-spec-compliant clients to self-register a public OAuth client without operator involvement. - - Policy: - - - Public clients only — `token_endpoint_auth_method` is forced to `none`. Confidential-client registration is out of scope this phase. - - Server-owned `resource_grants`. Caller-supplied `scope` or `resource_grants` is rejected as `invalid_client_metadata` (would be a privilege-escalation surface). Dynamic clients receive the same scopes the active resource publishes. - - Application-type-aware redirect URI policy. `application_type=native` accepts loopback (`127.0.0.1`, `::1`, `localhost`) and reverse-DNS-shaped custom schemes; `application_type=web` accepts HTTPS to hosts in an operator-controlled allowlist only. `application_type` is REQUIRED on the request — missing or empty rejects with `invalid_client_metadata`. - - Anti-impersonation: reserved client names are rejected from third parties via NFKC-folded compare. - - Generated `client_id` carries a stable prefix to distinguish dynamic from seeded clients in audit logs. - - Cache-Control: `no-store` on every 201 and 400 response (the response carries fresh credentials and rejection reasons). - x-runtime: [cloud] - security: [] - requestBody: - required: true - content: - application/json: - schema: - $ref: "#/components/schemas/OAuthRegisterRequest" - responses: - "201": - description: Registered. Body echoes the metadata RFC 7591 §3.2.1 requires. - headers: - Cache-Control: - schema: - type: string - description: 'Always "no-store"' - Pragma: - schema: - type: string - description: 'Always "no-cache"' - content: - application/json: - schema: - $ref: "#/components/schemas/OAuthRegisterResponse" - "400": - description: RFC 7591 §3.2.2 invalid client metadata - headers: - Cache-Control: - schema: - type: string - description: 'Always "no-store"' - Pragma: - schema: - type: string - description: 'Always "no-cache"' - content: - application/json: - schema: - $ref: "#/components/schemas/OAuthRegisterError" - "404": - description: OAuth disabled - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "503": - description: No active resource is configured — DCR cannot mint a usable client until an active resource row is seeded. - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - # --------------------------------------------------------------------------- - # Billing (cloud) - # --------------------------------------------------------------------------- - /api/billing/balance: - get: - operationId: getBillingBalance - tags: [billing] - summary: Get current credit balance - description: "[cloud-only] Returns the authenticated user's current credit balance and usage summary." - x-runtime: [cloud] - responses: - "200": - description: Balance info - content: - application/json: - schema: - $ref: "#/components/schemas/BillingBalance" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/billing/events: - get: - operationId: listBillingEvents - tags: [billing] - summary: List billing events - description: "[cloud-only] Returns a paginated list of billing events (charges, credits, refunds) for the authenticated user." - x-runtime: [cloud] - parameters: - - name: limit - in: query - schema: - type: integer - description: Maximum number of results - - name: offset - in: query - schema: - type: integer - description: Pagination offset - - name: type - in: query - schema: - type: string - description: Filter by event type - responses: - "200": - description: Billing events - content: - application/json: - schema: - $ref: "#/components/schemas/BillingEventList" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/billing/ops/{id}: - get: - operationId: getBillingOp - tags: [billing] - summary: Get a billing operation by ID - description: "[cloud-only] Returns details of a specific billing operation." - x-runtime: [cloud] - parameters: - - name: id - in: path - required: true - schema: - type: string - description: The billing operation ID. - responses: - "200": - description: Billing operation - content: - application/json: - schema: - $ref: "#/components/schemas/BillingOp" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/billing/payment-portal: - post: - operationId: createPaymentPortalSession - tags: [billing] - summary: Create a payment portal session - description: "[cloud-only] Creates a Stripe customer portal session for managing payment methods and invoices. Returns a URL to redirect the user to." - x-runtime: [cloud] - responses: - "200": - description: Portal session - content: - application/json: - schema: - type: object - properties: - url: - type: string - format: uri - description: Stripe portal URL - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/billing/plans: - get: - operationId: listBillingPlans - tags: [billing] - summary: List available billing plans - description: "[cloud-only] Returns the list of available subscription plans and their pricing." - x-runtime: [cloud] - responses: - "200": - description: Plan list - content: - application/json: - schema: - type: array - items: - $ref: "#/components/schemas/BillingPlan" - - /api/billing/preview-subscribe: - post: - operationId: previewSubscription - tags: [billing] - summary: Preview a subscription change - description: "[cloud-only] Returns a preview of what a subscription change would cost, including prorations." - x-runtime: [cloud] - requestBody: - required: true - content: - application/json: - schema: - type: object - required: - - plan_id - properties: - plan_id: - type: string - description: ID of the plan to preview - responses: - "200": - description: Subscription preview - content: - application/json: - schema: - $ref: "#/components/schemas/SubscriptionPreview" - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/billing/status: - get: - operationId: getBillingStatus - tags: [billing] - summary: Get billing status - description: "[cloud-only] Returns the authenticated user's current billing and subscription status." - x-runtime: [cloud] - responses: - "200": - description: Billing status - content: - application/json: - schema: - $ref: "#/components/schemas/BillingStatus" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/billing/subscribe: - post: - operationId: createSubscription - tags: [billing] - summary: Subscribe to a billing plan - description: "[cloud-only] Creates a new subscription to the specified billing plan." - x-runtime: [cloud] - requestBody: - required: true - content: - application/json: - schema: - type: object - required: - - plan_id - properties: - plan_id: - type: string - description: ID of the plan to subscribe to - payment_method_id: - type: string - description: Stripe payment method ID - responses: - "200": - description: Subscription created - content: - application/json: - schema: - $ref: "#/components/schemas/BillingSubscription" - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/billing/subscription/cancel: - post: - operationId: cancelSubscription - tags: [billing] - summary: Cancel the active subscription - description: "[cloud-only] Cancels the authenticated user's active subscription. The subscription remains active until the end of the current billing period." - x-runtime: [cloud] - responses: - "200": - description: Subscription cancelled - content: - application/json: - schema: - $ref: "#/components/schemas/BillingSubscription" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/billing/subscription/resubscribe: - post: - operationId: resubscribe - tags: [billing] - summary: Resubscribe after cancellation - description: "[cloud-only] Reactivates a subscription that was previously cancelled but has not yet expired." - x-runtime: [cloud] - responses: - "200": - description: Subscription reactivated - content: - application/json: - schema: - $ref: "#/components/schemas/BillingSubscription" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/billing/topup: - post: - operationId: topUpCredits - tags: [billing] - summary: Purchase additional credits - description: "[cloud-only] Purchases a one-time credit top-up using the user's payment method on file." - x-runtime: [cloud] - requestBody: - required: true - content: - application/json: - schema: - type: object - required: - - amount - properties: - amount: - type: integer - description: Number of credits to purchase - responses: - "200": - description: Top-up successful - content: - application/json: - schema: - $ref: "#/components/schemas/BillingBalance" - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - # --------------------------------------------------------------------------- - # Workspace (cloud) - # --------------------------------------------------------------------------- - /api/workspace/api-keys: - get: - operationId: listWorkspaceApiKeys - tags: [workspace] - summary: List workspace API keys - description: "[cloud-only] Returns the list of API keys for the current workspace." - x-runtime: [cloud] - responses: - "200": - description: API key list - content: - application/json: - schema: - type: array - items: - $ref: "#/components/schemas/WorkspaceApiKey" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "403": - description: Forbidden - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - post: - operationId: createWorkspaceApiKey - tags: [workspace] - summary: Create a workspace API key - description: "[cloud-only] Creates a new API key for the current workspace." - x-runtime: [cloud] - requestBody: - required: true - content: - application/json: - schema: - type: object - required: - - name - properties: - name: - type: string - description: Display name for the API key - description: - type: string - description: User-provided description of the key's purpose - maxLength: 5000 - responses: - "201": - description: API key created - content: - application/json: - schema: - $ref: "#/components/schemas/WorkspaceApiKeyCreated" - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "403": - description: Forbidden - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/workspace/api-keys/{id}: - delete: - operationId: deleteWorkspaceApiKey - tags: [workspace] - summary: Delete a workspace API key - description: "[cloud-only] Revokes and deletes a workspace API key." - x-runtime: [cloud] - parameters: - - name: id - in: path - required: true - schema: - type: string - description: The API key ID. - responses: - "204": - description: API key deleted - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "403": - description: Forbidden - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/workspace/invites: - get: - operationId: listWorkspaceInvites - tags: [workspace] - summary: List pending workspace invites - description: "[cloud-only] Returns the list of pending invitations for the current workspace." - x-runtime: [cloud] - responses: - "200": - description: Invite list - content: - application/json: - schema: - type: array - items: - $ref: "#/components/schemas/WorkspaceInvite" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "403": - description: Forbidden - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - post: - operationId: createWorkspaceInvite - tags: [workspace] - summary: Invite a user to the workspace - description: "[cloud-only] Creates an invitation for a user to join the current workspace." - x-runtime: [cloud] - requestBody: - required: true - content: - application/json: - schema: - type: object - required: - - email - properties: - email: - type: string - format: email - description: Email address to invite - role: - type: string - enum: [admin, member] - description: Role to assign - responses: - "201": - description: Invite created - content: - application/json: - schema: - $ref: "#/components/schemas/WorkspaceInvite" - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "403": - description: Forbidden - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "409": - description: Conflict - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/workspace/invites/{inviteId}: - delete: - operationId: deleteWorkspaceInvite - tags: [workspace] - summary: Cancel a workspace invite - description: "[cloud-only] Cancels a pending workspace invitation." - x-runtime: [cloud] - parameters: - - name: inviteId - in: path - required: true - schema: - type: string - description: The invite ID. - responses: - "204": - description: Invite cancelled - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "403": - description: Forbidden - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/workspace/leave: - post: - operationId: leaveWorkspace - tags: [workspace] - summary: Leave the current workspace - description: "[cloud-only] Removes the authenticated user from the current workspace." - x-runtime: [cloud] - responses: - "204": - description: Left workspace - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "403": - description: Forbidden - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/workspace/members: - get: - operationId: listWorkspaceMembers - tags: [workspace] - summary: List workspace members - description: "[cloud-only] Returns the list of members in the current workspace." - x-runtime: [cloud] - responses: - "200": - description: Member list - content: - application/json: - schema: - type: array - items: - $ref: "#/components/schemas/WorkspaceMember" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "403": - description: Forbidden - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/workspace/members/{user_id}/api-keys: - get: - operationId: listMemberApiKeys - tags: [workspace] - summary: List API keys for a workspace member - description: "[cloud-only] Returns the API keys belonging to a specific workspace member. Requires admin role." - x-runtime: [cloud] - parameters: - - name: user_id - in: path - required: true - schema: - type: string - description: The member's user ID. - responses: - "200": - description: API key list - content: - application/json: - schema: - type: array - items: - $ref: "#/components/schemas/WorkspaceApiKey" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "403": - description: Forbidden - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - delete: - operationId: bulkRevokeMemberApiKeys - tags: [workspace] - summary: Bulk revoke a member's API keys - description: "[cloud-only] Revokes all active API keys for a specific workspace member. Only workspace owners can perform this action." - x-runtime: [cloud] - parameters: - - name: user_id - in: path - required: true - schema: - type: string - minLength: 1 - description: The member's user ID. - responses: - "200": - description: Keys revoked - content: - application/json: - schema: - $ref: "#/components/schemas/BulkRevokeAPIKeysResponse" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "403": - description: Forbidden — must be workspace owner - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/workspace/members/{userId}: - patch: - operationId: updateWorkspaceMember - tags: [workspace] - summary: Update a workspace member's role - description: "[cloud-only] Updates the role of a workspace member. Requires admin role." - x-runtime: [cloud] - parameters: - - name: userId - in: path - required: true - schema: - type: string - description: The member's user ID. - requestBody: - required: true - content: - application/json: - schema: - type: object - required: - - role - properties: - role: - type: string - enum: [admin, member] - description: New role to assign - responses: - "200": - description: Member updated - content: - application/json: - schema: - $ref: "#/components/schemas/WorkspaceMember" - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "403": - description: Forbidden - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - delete: - operationId: removeWorkspaceMember - tags: [workspace] - summary: Remove a member from the workspace - description: "[cloud-only] Removes a member from the current workspace. Requires admin role." - x-runtime: [cloud] - parameters: - - name: userId - in: path - required: true - schema: - type: string - description: The member's user ID. - responses: - "204": - description: Member removed - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "403": - description: Forbidden - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/workspaces: - get: - operationId: listWorkspaces - tags: [workspace] - summary: List workspaces the user belongs to - description: "[cloud-only] Returns the list of workspaces the authenticated user is a member of." - x-runtime: [cloud] - responses: - "200": - description: Workspace list - content: - application/json: - schema: - type: array - items: - $ref: "#/components/schemas/Workspace" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - post: - operationId: createWorkspace - tags: [workspace] - summary: Create a new workspace - description: "[cloud-only] Creates a new workspace. The authenticated user becomes the owner." - x-runtime: [cloud] - requestBody: - required: true - content: - application/json: - schema: - type: object - required: - - name - properties: - name: - type: string - description: Workspace name - responses: - "201": - description: Workspace created - content: - application/json: - schema: - $ref: "#/components/schemas/Workspace" - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/workspaces/{id}: - get: - operationId: getWorkspace - tags: [workspace] - summary: Get a workspace by ID - description: "[cloud-only] Returns details of a workspace the user is a member of." - x-runtime: [cloud] - parameters: - - name: id - in: path - required: true - schema: - type: string - description: The workspace ID. - responses: - "200": - description: Workspace detail - content: - application/json: - schema: - $ref: "#/components/schemas/Workspace" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "403": - description: Forbidden - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - patch: - operationId: updateWorkspace - tags: [workspace] - summary: Update workspace settings - description: "[cloud-only] Updates the name or settings of a workspace. Requires admin role." - x-runtime: [cloud] - parameters: - - name: id - in: path - required: true - schema: - type: string - description: The workspace ID. - requestBody: - required: true - content: - application/json: - schema: - type: object - properties: - name: - type: string - description: New workspace name - responses: - "200": - description: Workspace updated - content: - application/json: - schema: - $ref: "#/components/schemas/Workspace" - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "403": - description: Forbidden - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - delete: - operationId: deleteWorkspace - tags: [workspace] - summary: Delete a workspace - description: "[cloud-only] Soft-deletes a workspace. Requires owner role. Personal workspaces cannot be deleted." - x-runtime: [cloud] - parameters: - - name: id - in: path - required: true - schema: - type: string - description: The workspace ID. - responses: - "204": - description: Workspace deleted - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "403": - description: Forbidden — must be workspace owner - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - # --------------------------------------------------------------------------- - # User / settings / misc (cloud) - # --------------------------------------------------------------------------- - /api/feedback: - post: - operationId: submitFeedback - tags: [user] - summary: Submit user feedback - description: "[cloud-only] Submits feedback from the user about their experience with the cloud runtime." - x-runtime: [cloud] - requestBody: - required: true - content: - application/json: - schema: - type: object - required: - - message - properties: - message: - type: string - description: Feedback message - rating: - type: integer - minimum: 1 - maximum: 5 - description: Optional satisfaction rating - context: - type: object - additionalProperties: true - description: Additional context metadata - responses: - "200": - description: Feedback submitted - content: - application/json: - schema: - type: object - properties: - id: - type: string - status: - type: string - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/files/mask-layers: - get: - operationId: getMaskLayers - tags: [assets] - summary: Get related mask layer filenames - description: "[cloud-only] Given a mask file (any of the 4 layers), returns all related mask layer filenames. Used by the mask editor to load the paint, mask, and painted layers when reopening a previously edited mask." - x-runtime: [cloud] - parameters: - - name: filename - in: query - required: true - schema: - type: string - description: Hash filename of any mask layer file - responses: - "200": - description: Related mask layers - content: - application/json: - schema: - type: object - properties: - mask: - type: string - description: Filename of the mask layer - nullable: true - paint: - type: string - description: Filename of the paint strokes layer - nullable: true - painted: - type: string - description: Filename of the painted image layer - nullable: true - painted_masked: - type: string - description: Filename of the final composite layer - nullable: true - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: File not found or not a mask file - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/internal/cloud_analytics: - post: - operationId: postCloudAnalytics - tags: [internal] - summary: Post client analytics events - description: "[cloud-only] Receives analytics events from the frontend for processing by the cloud analytics pipeline." - x-runtime: [cloud] - requestBody: - required: true - content: - application/json: - schema: - type: object - required: - - events - properties: - events: - type: array - items: - type: object - required: - - event_name - properties: - event_name: - type: string - timestamp: - type: string - format: date-time - properties: - type: object - additionalProperties: true - responses: - "200": - description: Events accepted - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/invites/{token}/accept: - post: - operationId: acceptInvite - tags: [workspace] - summary: Accept a workspace invitation - description: "[cloud-only] Accepts a workspace invitation using the invite token. The authenticated user is added to the workspace." - x-runtime: [cloud] - parameters: - - name: token - in: path - required: true - schema: - type: string - description: The invitation token. - responses: - "200": - description: Invite accepted - content: - application/json: - schema: - $ref: "#/components/schemas/Workspace" - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/secrets: - get: - operationId: listSecrets - tags: [settings] - summary: List user secrets - description: "[cloud-only] Returns the list of secrets (API keys for third-party services) stored for the authenticated user. Secret values are redacted." - x-runtime: [cloud] - responses: - "200": - description: Secret list - content: - application/json: - schema: - type: array - items: - $ref: "#/components/schemas/SecretMeta" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - post: - operationId: createSecret - tags: [settings] - summary: Create or update a secret - description: "[cloud-only] Stores a new secret or updates an existing one. Secrets are encrypted at rest." - x-runtime: [cloud] - requestBody: - required: true - content: - application/json: - schema: - type: object - required: - - name - - value - properties: - name: - type: string - description: Secret name (unique per user) - value: - type: string - description: Secret value - responses: - "201": - description: Secret created - content: - application/json: - schema: - $ref: "#/components/schemas/SecretMeta" - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/secrets/{id}: - get: - operationId: getSecret - tags: [settings] - summary: Get secret metadata - description: "[cloud-only] Returns metadata for a specific secret. Does not return the plaintext secret value." - x-runtime: [cloud] - parameters: - - name: id - in: path - required: true - schema: - type: string - format: uuid - description: The secret ID. - responses: - "200": - description: Secret metadata - content: - application/json: - schema: - $ref: "#/components/schemas/SecretMeta" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - patch: - operationId: updateSecret - tags: [settings] - summary: Update a secret - description: "[cloud-only] Updates an existing secret's name and/or value. Both fields are optional; only provided fields are updated." - x-runtime: [cloud] - parameters: - - name: id - in: path - required: true - schema: - type: string - format: uuid - description: The secret ID. - requestBody: - required: true - content: - application/json: - schema: - $ref: "#/components/schemas/UpdateSecretRequest" - responses: - "200": - description: Secret updated - content: - application/json: - schema: - $ref: "#/components/schemas/SecretMeta" - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "409": - description: Conflict — a secret with this name already exists - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - delete: - operationId: deleteSecret - tags: [settings] - summary: Delete a secret - description: "[cloud-only] Permanently deletes a stored secret." - x-runtime: [cloud] - parameters: - - name: id - in: path - required: true - schema: - type: string - description: The secret ID. - responses: - "204": - description: Secret deleted - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/user: - get: - operationId: getCloudUser - tags: [user] - summary: Get the authenticated cloud user - description: "[cloud-only] Returns the profile and account information for the currently authenticated user." - x-runtime: [cloud] - responses: - "200": - description: User profile - content: - application/json: - schema: - $ref: "#/components/schemas/CloudUser" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - put: - operationId: updateCloudUser - tags: [user] - summary: Update the authenticated cloud user profile - description: "[cloud-only] Updates the profile information for the currently authenticated user." - x-runtime: [cloud] - requestBody: - required: true - content: - application/json: - schema: - type: object - properties: - display_name: - type: string - avatar_url: - type: string - format: uri - responses: - "200": - description: Updated profile - content: - application/json: - schema: - $ref: "#/components/schemas/CloudUser" - "400": - description: Bad request - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/userdata/{file}/publish: - get: - operationId: getUserdataFilePublish - tags: [userdata] - summary: Get publish info for a userdata file - description: "[cloud-only] Returns the publish status and share info for a userdata workflow file." - x-runtime: [cloud] - parameters: - - name: file - in: path - required: true - schema: - type: string - description: File path relative to user data directory - responses: - "200": - description: Publish info (publish_time is null if never published) - content: - application/json: - schema: - $ref: "#/components/schemas/WorkflowPublishInfo" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Workflow not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - post: - operationId: publishUserdataFile - tags: [userdata] - summary: Publish a userdata file to the cloud - description: "[cloud-only] Makes a userdata file available via a public URL for sharing or embedding." - x-runtime: [cloud] - parameters: - - name: file - in: path - required: true - schema: - type: string - description: File path relative to user data directory - responses: - "200": - description: Published file URL - content: - application/json: - schema: - type: object - properties: - url: - type: string - format: uri - description: Public URL of the published file - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/vhs/queryvideo: - get: - operationId: queryVhsVideo - tags: [view] - summary: Query VHS video metadata - description: "[cloud-only] Returns metadata about a video file processed by the VHS (Video Helper Suite) integration." - x-runtime: [cloud] - parameters: - - name: filename - in: query - required: true - schema: - type: string - description: Video filename - - name: type - in: query - schema: - type: string - enum: [input, output, temp] - description: Directory type - - name: subfolder - in: query - schema: - type: string - description: Subfolder within the directory - responses: - "200": - description: Video metadata - content: - application/json: - schema: - type: object - additionalProperties: true - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/vhs/viewaudio: - get: - operationId: viewVhsAudio - tags: [view] - summary: View or download VHS audio - description: "[cloud-only] Returns audio content from a VHS-processed file." - x-runtime: [cloud] - parameters: - - name: filename - in: query - required: true - schema: - type: string - description: Audio filename - - name: type - in: query - schema: - type: string - enum: [input, output, temp] - description: Directory type - - name: subfolder - in: query - schema: - type: string - description: Subfolder within the directory - responses: - "200": - description: Audio content - content: - audio/*: - schema: - type: string - format: binary - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/vhs/viewvideo: - get: - operationId: viewVhsVideo - tags: [view] - summary: View or download VHS video - description: "[cloud-only] Returns video content from a VHS-processed file." - x-runtime: [cloud] - parameters: - - name: filename - in: query - required: true - schema: - type: string - description: Video filename - - name: type - in: query - schema: - type: string - enum: [input, output, temp] - description: Directory type - - name: subfolder - in: query - schema: - type: string - description: Subfolder within the directory - responses: - "200": - description: Video content - content: - video/*: - schema: - type: string - format: binary - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/viewvideo: - get: - operationId: viewVideo - tags: [view] - summary: View or download a video file - deprecated: true - description: | - **Deprecated.** This endpoint is an alias of `GET /api/view` added for - legacy history-queue video playback. Callers should use `/api/view` - directly; the endpoint is retained for backward compatibility but will - be removed in a future release. - x-runtime: [cloud] - parameters: - - name: filename - in: query - required: true - schema: - type: string - description: Video filename - - name: type - in: query - schema: - type: string - enum: [input, output, temp] - description: Directory type - - name: subfolder - in: query - schema: - type: string - description: Subfolder within the directory - responses: - "200": - description: Video content - content: - video/*: - schema: - type: string - format: binary - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/tasks: - get: - operationId: listTasks - tags: [task] - summary: List background tasks - description: "[cloud-only] Retrieve a paginated list of background tasks for the authenticated user. Supports filtering by task type, status, and creation time." - x-runtime: [cloud] - parameters: - - name: task_name - in: query - schema: - type: string - description: Filter by task type name (exact match). - - name: idempotency_key - in: query - schema: - type: string - description: Filter by idempotency key (exact match). - - name: status - in: query - schema: - type: string - description: Filter by one or more statuses (comma-separated). - - name: created_after - in: query - schema: - type: string - format: date-time - description: Filter tasks created after this timestamp. - - name: created_before - in: query - schema: - type: string - format: date-time - description: Filter tasks created before this timestamp. - - name: sort_order - in: query - schema: - type: string - enum: [asc, desc] - default: desc - description: Sort direction by create_time. - - name: offset - in: query - schema: - type: integer - minimum: 0 - default: 0 - description: Pagination offset (0-based). - - name: limit - in: query - schema: - type: integer - minimum: 1 - maximum: 100 - default: 20 - description: Maximum items per page (1-100). - responses: - "200": - description: Tasks retrieved - content: - application/json: - schema: - $ref: "#/components/schemas/TasksListResponse" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "422": - description: Validation error - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - - /api/tasks/{task_id}: - get: - operationId: getTask - tags: [task] - summary: Get task details - description: "[cloud-only] Retrieve full details for a specific background task." - x-runtime: [cloud] - parameters: - - name: task_id - in: path - required: true - schema: - type: string - format: uuid - description: Task identifier (UUID). - responses: - "200": - description: Task details - content: - application/json: - schema: - $ref: "#/components/schemas/TaskResponse" - "401": - description: Unauthorized - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - "404": - description: Task not found - content: - application/json: - schema: - $ref: "#/components/schemas/CloudError" - components: - parameters: - ComfyUserHeader: - name: Comfy-User - in: header - required: false - schema: - type: string - description: | - Identifies the active user in multi-user mode. Used for settings, - userdata, and history isolation. This is not a security mechanism — - it is an organisational convenience with no authentication behind it. - - schemas: - # ------------------------------------------------------------------- - # Prompt - # ------------------------------------------------------------------- - PromptRequest: - type: object - description: A workflow submission. Wraps the prompt graph plus optional client identifier and extra per-request data. - required: - - prompt - properties: - prompt: - type: object - description: | - The workflow graph to execute. Keys are node IDs (strings); - values are objects with class_type and inputs. - additionalProperties: true - number: - type: number - description: Priority number for the queue (lower numbers have higher priority) - front: - type: boolean - description: If true, adds the prompt to the front of the queue - extra_data: - type: object - description: Extra data associated with the prompt (e.g. extra_pnginfo) - additionalProperties: true - client_id: - type: string - description: WebSocket client ID to receive progress updates - prompt_id: - type: string - format: uuid - description: "Client-supplied prompt ID. Server generates a UUID if omitted." - partial_execution_targets: - type: array - items: - type: string - description: List of node IDs to execute (partial graph execution) - workflow_id: - type: string - format: uuid - nullable: true - x-runtime: [cloud] - description: "[cloud-only] Cloud workflow entity ID for tracking and gallery association. Ignored by local ComfyUI." - workflow_version_id: - type: string - format: uuid - nullable: true - x-runtime: [cloud] - description: "[cloud-only] Cloud workflow version ID for pinning execution to a specific version. Ignored by local ComfyUI." - - PromptResponse: - type: object - description: Server acknowledgement of a workflow submission. Includes the assigned `prompt_id` and current queue position. - properties: - prompt_id: - type: string - format: uuid - description: Unique identifier for the prompt execution - number: - type: number - description: Priority number in the queue - node_errors: - type: object - description: Validation errors keyed by node ID - additionalProperties: - $ref: "#/components/schemas/NodeError" - error: - description: Top-level prompt error (string message or structured error) - oneOf: - - type: string - - $ref: "#/components/schemas/PromptError" - - PromptErrorResponse: - type: object - description: Error response when prompt validation fails - additionalProperties: true - - PromptError: - type: object - description: Structured prompt validation error - properties: - type: - type: string - message: - type: string - details: - type: string - - Error: - type: object - description: Detailed node-level error - properties: - type: - type: string - message: - type: string - details: - type: string - extra_info: - type: object - properties: - input_name: - type: string - additionalProperties: true - - NodeError: - type: object - description: Error details for a single node - properties: - errors: - type: array - items: - $ref: "#/components/schemas/Error" - class_type: - type: string - description: The node's class type - dependent_outputs: - type: array - items: {} - - PromptInfo: - type: object - description: Summary of a queued or recently-executed prompt, as returned by the queue and history endpoints. - properties: - exec_info: - type: object - properties: - queue_remaining: - type: integer - description: Number of items remaining in the queue - - # ------------------------------------------------------------------- - # Queue - # ------------------------------------------------------------------- - QueueInfo: - type: object - description: Queue information with pending and running items - properties: - queue_running: - type: array - description: Currently running queue items - items: - type: array - description: | - Queue item tuple: [number, prompt_id, prompt, extra_data, outputs_to_execute, sensitive] - items: {} - prefixItems: - - type: number - description: Priority number - - type: string - format: uuid - description: prompt_id - - type: object - description: prompt graph - additionalProperties: true - - type: object - description: extra_data - additionalProperties: true - - type: array - description: outputs_to_execute (list of output node IDs) - items: - type: string - - type: object - description: sensitive data (may be omitted) - additionalProperties: true - queue_pending: - type: array - description: Pending queue items (oldest first) - items: - type: array - description: | - Queue item tuple: [number, prompt_id, prompt, extra_data, outputs_to_execute, sensitive] - items: {} - prefixItems: - - type: number - description: Priority number - - type: string - format: uuid - description: prompt_id - - type: object - description: prompt graph - additionalProperties: true - - type: object - description: extra_data - additionalProperties: true - - type: array - description: outputs_to_execute (list of output node IDs) - items: - type: string - - type: object - description: sensitive data (may be omitted) - additionalProperties: true - - QueueManageRequest: - type: object - description: Request to clear or delete from queue - properties: - clear: - type: boolean - description: If true, clear all pending items - delete: - type: array - items: - type: string - description: Array of prompt IDs to delete from queue - - # ------------------------------------------------------------------- - # History - # ------------------------------------------------------------------- - HistoryEntry: - type: object - description: A single execution history entry - properties: - prompt: - type: array - description: | - Prompt tuple: [number, prompt_id, prompt_graph, extra_data, output_node_ids] - items: {} - outputs: - type: object - description: Output data from execution keyed by node ID - additionalProperties: true - status: - type: object - description: Execution status (status_str, completed, messages, etc.) - additionalProperties: true - meta: - type: object - description: Metadata about the execution and nodes - additionalProperties: true - - HistoryManageRequest: - type: object - description: Request to clear or delete history entries - properties: - clear: - type: boolean - description: If true, clear all history - delete: - type: array - items: - type: string - description: Array of prompt IDs to delete from history - - # ------------------------------------------------------------------- - # Jobs - # ------------------------------------------------------------------- - JobEntry: - type: object - description: Lightweight job data for list views - required: - - id - - status - properties: - id: - type: string - format: uuid - description: Unique job identifier (same as prompt_id) - status: - type: string - description: Current job status - create_time: - type: number - description: Job creation timestamp - execution_start_time: - type: number - description: Workflow execution start timestamp - execution_end_time: - type: number - description: Workflow execution end timestamp - preview_output: - type: object - additionalProperties: true - description: Primary preview output - outputs_count: - type: integer - description: Total number of output files - - JobDetailResponse: - type: object - description: Full job details including workflow and outputs - required: - - id - - status - properties: - id: - type: string - format: uuid - status: - type: string - workflow: - type: object - additionalProperties: true - description: Full ComfyUI workflow - outputs: - type: object - additionalProperties: true - description: Full outputs object from execution - execution_error: - $ref: "#/components/schemas/ExecutionError" - create_time: - type: number - update_time: - type: number - execution_start_time: - type: number - execution_end_time: - type: number - preview_output: - type: object - additionalProperties: true - outputs_count: - type: integer - execution_status: - type: object - additionalProperties: true - execution_meta: - type: object - additionalProperties: true - - ExecutionError: - type: object - description: Detailed execution error from ComfyUI - properties: - node_id: - type: string - description: ID of the node that failed - node_type: - type: string - description: Type name of the node - exception_message: - type: string - description: Human-readable error message - exception_type: - type: string - description: Python exception type - traceback: - type: array - items: - type: string - description: Traceback lines - current_inputs: - type: object - additionalProperties: true - current_outputs: - type: object - additionalProperties: true - - PaginationInfo: - type: object - description: Pagination metadata returned alongside list responses. - properties: - offset: - type: integer - limit: - type: integer - total: - type: integer - has_more: - type: boolean - - # ------------------------------------------------------------------- - # Upload / View - # ------------------------------------------------------------------- - UploadResult: - type: object - description: Response body returned by the image/mask upload endpoints, describing where the uploaded file now lives. - properties: - name: - type: string - description: Saved filename (may be renamed to avoid collisions) - subfolder: - type: string - description: Subfolder the file was saved to - type: - type: string - description: Directory type (input, temp) - - # ------------------------------------------------------------------- - # System - # ------------------------------------------------------------------- - DeviceStats: - type: object - description: GPU/compute device statistics - required: - - name - - type - - index - properties: - name: - type: string - description: Device name - type: - type: string - description: Device type (cuda, mps, cpu, etc.) - index: - type: number - nullable: true - description: | - Device index within its type (e.g. CUDA ordinal for `cuda:0`, - `cuda:1`). `null` for devices with no index, including the CPU - device returned in `--cpu` mode (PyTorch's `torch.device('cpu').index` - is `None`). - vram_total: - type: number - description: Total VRAM in bytes - vram_free: - type: number - description: Free VRAM in bytes - torch_vram_total: - type: number - description: Total PyTorch-managed VRAM in bytes - torch_vram_free: - type: number - description: Free PyTorch-managed VRAM in bytes - - SystemStatsResponse: - type: object - description: Hardware, VRAM, Python, and ComfyUI version information for the running process. - required: - - system - - devices - properties: - system: - type: object - required: - - os - - python_version - - embedded_python - - comfyui_version - - pytorch_version - - argv - - ram_total - - ram_free - properties: - os: - type: string - description: Operating system - python_version: - type: string - description: Python version - embedded_python: - type: boolean - description: Whether using embedded Python - comfyui_version: - type: string - description: ComfyUI version string - pytorch_version: - type: string - description: PyTorch version - required_frontend_version: - type: string - description: Required frontend version - argv: - type: array - items: - type: string - description: Command line arguments - ram_total: - type: number - description: Total RAM in bytes - ram_free: - type: number - description: Free RAM in bytes - installed_templates_version: - type: string - nullable: true - description: Version of the currently installed workflow templates - required_templates_version: - type: string - nullable: true - description: Minimum required workflow templates version for this ComfyUI build - comfy_package_versions: - type: array - description: Installed and required versions for every comfy* package pinned in requirements.txt - items: - type: object - required: - - name - - installed - - required - properties: - name: - type: string - installed: - type: string - nullable: true - required: - type: string - nullable: true - devices: - type: array - items: - $ref: "#/components/schemas/DeviceStats" - - # ------------------------------------------------------------------- - # Node / Object Info - # ------------------------------------------------------------------- - NodeInfo: - type: object - description: 'Definition of a registered node class: its inputs, outputs, category, and display metadata.' - properties: - input: - type: object - description: Input specifications (required and optional groups) - additionalProperties: true - input_order: - type: object - description: Ordered input names per group - additionalProperties: - type: array - items: - type: string - output: - type: array - items: - type: string - description: Output type names - output_is_list: - type: array - items: - type: boolean - description: Whether each output is a list - output_name: - type: array - items: - type: string - description: Display names of outputs - name: - type: string - description: Internal class name - display_name: - type: string - description: Human-readable display name - description: - type: string - description: Node description - python_module: - type: string - description: Python module implementing the node - category: - type: string - description: Node category path - output_node: - type: boolean - description: Whether this is an output node - output_tooltips: - type: array - items: - type: string - description: Tooltips for each output - deprecated: - type: boolean - description: Whether the node is deprecated - experimental: - type: boolean - description: Whether the node is experimental - api_node: - type: boolean - description: Whether this is an API node - is_input_list: - type: boolean - description: Whether the node accepts list inputs - dev_only: - type: boolean - description: Whether the node is developer-only (hidden in production UI) - has_intermediate_output: - type: boolean - description: Whether the node emits intermediate output during execution - search_aliases: - type: array - items: - type: string - description: Alternative search terms for finding this node - essentials_category: - type: string - nullable: true - description: | - Category override used by the essentials pack. The - `essentials_category` key may be present with a string value, - present and `null`, or absent entirely: - - - V1 nodes: `essentials_category` is **omitted** when the node - class doesn't define an `ESSENTIALS_CATEGORY` attribute, and - **`null`** if the attribute is explicitly set to `None`. - - V3 nodes (`comfy_api.latest.io`): `essentials_category` is - **always present**, and **`null`** for nodes whose `Schema` - doesn't populate it. - - # ------------------------------------------------------------------- - # Models - # ------------------------------------------------------------------- - ModelFolder: - type: object - description: A configured model folder and the list of disk paths it resolves to. - required: - - name - - folders - properties: - name: - type: string - description: Model folder type name (e.g. "checkpoints") - folders: - type: array - items: - type: string - description: Filesystem paths for this model type - - ModelFile: - type: object - description: A single model file in a folder, with filesystem metadata. - required: - - name - - pathIndex - properties: - name: - type: string - description: Model filename - pathIndex: - type: integer - description: Index into the folder's paths array - modified: - type: number - description: File modification timestamp - created: - type: number - description: File creation timestamp - size: - type: integer - format: int64 - description: File size in bytes - - # ------------------------------------------------------------------- - # Subgraphs - # ------------------------------------------------------------------- - GlobalSubgraphInfo: - type: object - description: Metadata for a global subgraph blueprint (without full data) - required: - - source - - name - - info - properties: - source: - type: string - description: Source type ("templates" or "custom_node") - name: - type: string - description: Display name of the subgraph blueprint - info: - type: object - description: Additional information about the subgraph - required: - - node_pack - properties: - node_pack: - type: string - description: The node pack/module providing this subgraph - data: - type: string - description: The full subgraph JSON data (may be empty in list view) - - GlobalSubgraphData: - type: object - description: Full data for a global subgraph blueprint - required: - - source - - name - - info - - data - properties: - source: - type: string - description: Source type ("templates" or "custom_node") - name: - type: string - description: Display name of the subgraph blueprint - info: - type: object - description: Additional information about the subgraph - required: - - node_pack - properties: - node_pack: - type: string - description: The node pack/module providing this subgraph - data: - type: string - description: The full subgraph JSON data as a string - - # ------------------------------------------------------------------- - # Userdata - # ------------------------------------------------------------------- - UserDataResponse: - description: | - Response body for the POST endpoints `/api/userdata/{file}` and - `/api/userdata/{file}/move/{dest}`. Returns a single item whose - shape depends on the `full_info` query parameter. - x-variant-selector: - full_info=true: file-info object (`GetUserDataResponseFullFile`) - default: relative path string - oneOf: - - $ref: "#/components/schemas/GetUserDataResponseFullFile" - - type: string - description: Relative path of the written or moved file. Returned when `full_info` is absent or false. - - ListUserdataResponse: - description: | - Response body for `GET /api/userdata`. The array item shape is - determined by the `full_info` and `split` query parameters. - x-variant-selector: - full_info=true: array of file-info objects (`GetUserDataResponseFullFile`) - split=true: array of `[relative_path, ...path_components]` arrays - default: array of relative path strings - oneOf: - - type: array - items: - $ref: "#/components/schemas/GetUserDataResponseFullFile" - description: Returned when `full_info=true`. - - type: array - items: - type: array - items: - type: string - minItems: 2 - description: | - Returned when `split=true` and `full_info=false`. Each inner - array is `[relative_path, ...path_components]`. - - type: array - items: - type: string - description: Default shape — array of file paths relative to the user data root. - - GetUserDataResponseFullFile: - type: object - description: A single entry in a full-info user data listing. - properties: - path: - type: string - description: File name or path relative to the user directory - created: - type: number - description: Unix timestamp of file creation - size: - type: integer - description: File size in bytes - modified: - type: integer - format: int64 - description: Unix timestamp of last modification in milliseconds - - # ------------------------------------------------------------------- - # Assets - # ------------------------------------------------------------------- - Asset: - type: object - description: A registered asset — an input/output file tracked in the asset database with content hash and metadata. - required: - - id - - name - - size - - created_at - - updated_at - properties: - id: - type: string - format: uuid - description: Unique identifier for the asset - name: - type: string - description: Name of the asset file - hash: - type: string - nullable: true - description: Blake3 content hash of the asset (preferred over asset_hash) - pattern: "^blake3:[a-f0-9]{64}$" - asset_hash: - type: string - nullable: true - deprecated: true - description: "Deprecated: use `hash` instead. Blake3 hash of the asset content." - pattern: "^blake3:[a-f0-9]{64}$" - size: - type: integer - format: int64 - description: Size of the asset in bytes - mime_type: - type: string - description: MIME type of the asset - tags: - type: array - items: - type: string - description: Tags associated with the asset - user_metadata: - type: object - description: Custom user metadata - additionalProperties: true - metadata: - type: object - description: System-managed metadata (read-only) - additionalProperties: true - readOnly: true - preview_url: - type: string - format: uri - description: URL for asset preview/thumbnail - preview_id: - type: string - format: uuid - description: ID of the preview asset if available - prompt_id: - type: string - format: uuid - nullable: true - deprecated: true - description: "Deprecated: use job_id instead. ID of the prompt that created this asset." - job_id: - type: string - format: uuid - nullable: true - description: ID of the job that created this asset - created_at: - type: string - format: date-time - updated_at: - type: string - format: date-time - last_access_time: - type: string - format: date-time - is_immutable: - type: boolean - description: Whether this asset is immutable - - AssetCreated: - description: Response body returned after successfully registering a new asset. - allOf: - - $ref: "#/components/schemas/Asset" - - type: object - required: - - created_new - properties: - created_new: - type: boolean - description: Whether this was a new creation (true) or returned existing (false) - - AssetUpdated: - type: object - description: Response body returned after updating an asset's metadata. - required: - - id - - updated_at - properties: - id: - type: string - format: uuid - name: - type: string - hash: - type: string - nullable: true - description: Blake3 content hash of the asset (preferred over asset_hash) - pattern: "^blake3:[a-f0-9]{64}$" - asset_hash: - type: string - nullable: true - deprecated: true - description: "Deprecated: use `hash` instead. Blake3 hash of the asset content." - pattern: "^blake3:[a-f0-9]{64}$" - tags: - type: array - items: - type: string - mime_type: - type: string - user_metadata: - type: object - additionalProperties: true - prompt_id: - type: string - format: uuid - nullable: true - deprecated: true - description: "Deprecated: use job_id instead. ID of the prompt that created this asset." - job_id: - type: string - format: uuid - nullable: true - description: ID of the job that created this asset - updated_at: - type: string - format: date-time - - ListAssetsResponse: - type: object - description: Paginated list of assets. - required: - - assets - - total - - has_more - properties: - assets: - type: array - items: - $ref: "#/components/schemas/Asset" - total: - type: integer - has_more: - type: boolean - - TagInfo: - type: object - description: A tag known to the asset database, with the number of assets bearing it. - required: - - name - - count - properties: - name: - type: string - count: - type: integer - - ListTagsResponse: - type: object - description: Flat list of all tags, with counts. - required: - - tags - - total - - has_more - properties: - tags: - type: array - items: - $ref: "#/components/schemas/TagInfo" - total: - type: integer - has_more: - type: boolean - - AssetTagHistogramResponse: - type: object - description: Tags that would refine a filtered asset query, with the count of assets each tag would additionally select. - required: - - tag_counts - properties: - tag_counts: - type: object - additionalProperties: - type: integer - description: Map of tag names to occurrence counts - - TagsModificationResponse: - type: object - description: Response body returned after adding or removing tags on an asset. - required: - - total_tags - properties: - added: - type: array - items: - type: string - description: Tags successfully added - removed: - type: array - items: - type: string - description: Tags successfully removed - already_present: - type: array - items: - type: string - description: Tags already present (for add) - not_present: - type: array - items: - type: string - description: Tags not present (for remove) - total_tags: - type: array - items: - type: string - description: All tags on the asset after the operation - - # ------------------------------------------------------------------- - # Result / Output types - # ------------------------------------------------------------------- - ResultItem: - type: object - description: A single output file reference - properties: - filename: - type: string - subfolder: - type: string - type: - type: string - enum: [input, output, temp] - display_name: - type: string - - NodeOutputs: - type: object - description: | - Outputs from a single node execution. Known keys are listed below, - but custom nodes may add arbitrary keys (additionalProperties). - properties: - images: - type: array - items: - $ref: "#/components/schemas/ResultItem" - audio: - type: array - items: - $ref: "#/components/schemas/ResultItem" - video: - type: array - items: - $ref: "#/components/schemas/ResultItem" - animated: - type: array - items: - type: boolean - text: - oneOf: - - type: string - - type: array - items: - type: string - additionalProperties: true - - TerminalSize: - type: object - description: Terminal dimensions - properties: - cols: - type: number - row: - type: number - - LogEntry: - type: object - description: A single log entry - properties: - t: - type: string - description: Timestamp - m: - type: string - description: Log message - - StatusWsMessageStatus: - type: object - description: Inner payload of a `status` WebSocket message, describing the execution queue state. - properties: - exec_info: - type: object - required: - - queue_remaining - properties: - queue_remaining: - type: integer - - StatusWsMessage: - type: object - description: Initial status message sent on connect + queue status updates - properties: - status: - $ref: "#/components/schemas/StatusWsMessageStatus" - sid: - type: string - description: Session ID assigned by the server - - ProgressWsMessage: - type: object - description: Node execution progress (step N of M) - required: - - value - - max - - prompt_id - - node - properties: - value: - type: integer - description: Current step - max: - type: integer - description: Total steps - prompt_id: - type: string - node: - type: string - description: Node ID currently executing - - ProgressTextWsMessage: - type: object - description: Text-based progress update from a node - properties: - nodeId: - type: string - text: - type: string - prompt_id: - type: string - - NodeProgressState: - type: object - description: Progress state for a single node - properties: - value: - type: number - max: - type: number - state: - type: string - enum: [pending, running, finished, error] - node_id: - type: string - prompt_id: - type: string - display_node_id: - type: string - parent_node_id: - type: string - real_node_id: - type: string - - ProgressStateWsMessage: - type: object - description: Bulk progress state for all nodes in a prompt - required: - - prompt_id - - nodes - properties: - prompt_id: - type: string - nodes: - type: object - description: Map of node ID to progress state - additionalProperties: - $ref: "#/components/schemas/NodeProgressState" - - ExecutingWsMessage: - type: object - description: Fired when a node begins execution - required: - - node - - display_node - - prompt_id - properties: - node: - type: string - description: Node ID - display_node: - type: string - description: Display node ID (may differ for subgraphs) - prompt_id: - type: string - - ExecutedWsMessage: - type: object - description: Fired when a node completes execution with output - required: - - node - - display_node - - prompt_id - - output - properties: - node: - type: string - display_node: - type: string - prompt_id: - type: string - output: - $ref: "#/components/schemas/NodeOutputs" - merge: - type: boolean - description: Whether to merge with existing output - - ExecutionWsMessageBase: - type: object - description: Base fields for execution lifecycle messages - required: - - prompt_id - - timestamp - properties: - prompt_id: - type: string - timestamp: - type: integer - description: Unix timestamp in milliseconds - - ExecutionStartWsMessage: - allOf: - - $ref: "#/components/schemas/ExecutionWsMessageBase" - description: Fired when prompt execution begins - - ExecutionSuccessWsMessage: - allOf: - - $ref: "#/components/schemas/ExecutionWsMessageBase" - description: Fired when prompt execution completes successfully - - ExecutionCachedWsMessage: - allOf: - - $ref: "#/components/schemas/ExecutionWsMessageBase" - - type: object - properties: - nodes: - type: array - items: - type: string - description: List of node IDs that were cached - description: Fired when nodes are served from cache - - ExecutionInterruptedWsMessage: - allOf: - - $ref: "#/components/schemas/ExecutionWsMessageBase" - - type: object - properties: - node_id: - type: string - node_type: - type: string - executed: - type: array - items: - type: string - description: Node IDs that completed before interruption - description: Fired when execution is interrupted by user - - ExecutionErrorWsMessage: - allOf: - - $ref: "#/components/schemas/ExecutionWsMessageBase" - - type: object - properties: - node_id: - type: string - node_type: - type: string - executed: - type: array - items: - type: string - exception_message: - type: string - exception_type: - type: string - traceback: - type: array - items: - type: string - current_inputs: {} - current_outputs: {} - description: Fired when a node throws an exception during execution - - LogsWsMessage: - type: object - description: Streaming log entries from the server - properties: - size: - $ref: "#/components/schemas/TerminalSize" - entries: - type: array - items: - $ref: "#/components/schemas/LogEntry" - - NotificationWsMessage: - type: object - description: Server notification (e.g. model download complete) - properties: - value: - type: string - id: - type: string - - FeatureFlagsWsMessage: - type: object - description: Feature flags sent on connect - additionalProperties: true - - AssetDownloadWsMessage: - type: object - description: Asset download progress - required: - - task_id - - asset_name - - bytes_total - - bytes_downloaded - - progress - - status - properties: - task_id: - type: string - asset_name: - type: string - bytes_total: - type: number - bytes_downloaded: - type: number - progress: - type: number - description: 0.0 to 1.0 - status: - type: string - enum: [created, running, completed, failed] - asset_id: - type: string - error: - type: string - - AssetExportWsMessage: - type: object - description: Bulk asset export progress - required: - - task_id - - assets_total - - assets_attempted - - assets_failed - - bytes_total - - bytes_processed - - progress - - status - properties: - task_id: - type: string - export_name: - type: string - assets_total: - type: number - assets_attempted: - type: number - assets_failed: - type: number - bytes_total: - type: number - bytes_processed: - type: number - progress: - type: number - description: 0.0 to 1.0 - status: - type: string - enum: [created, running, completed, failed] - error: - type: string - - # ------------------------------------------------------------------- - # Cloud-runtime schemas - # - # These schemas are exclusively referenced by cloud-runtime operations. - # Tagged x-runtime: [cloud]. - # ------------------------------------------------------------------- - CloudError: - type: object - x-runtime: [cloud] - description: "[cloud-only] Standard error response from cloud endpoints." - required: - - error - properties: - error: - type: string - description: Error message - code: - type: string - description: Machine-readable error code - details: - type: object - additionalProperties: true - description: Additional error context - - CloudJobStatus: - type: object - x-runtime: [cloud] - description: "[cloud-only] Status of a cloud job." - required: - - id - - status - properties: - id: - type: string - format: uuid - status: - type: string - enum: [pending, running, completed, failed, cancelled] - progress: - type: number - minimum: 0 - maximum: 1 - description: "Execution progress (0.0 to 1.0)" - started_at: - type: string - format: date-time - nullable: true - completed_at: - type: string - format: date-time - nullable: true - - CloudPrompt: - type: object - x-runtime: [cloud] - description: "[cloud-only] A cloud-executed prompt record." - required: - - id - - status - properties: - id: - type: string - format: uuid - status: - type: string - workflow: - type: object - additionalProperties: true - outputs: - type: object - additionalProperties: true - created_at: - type: string - format: date-time - completed_at: - type: string - format: date-time - nullable: true - - HistoryV2Response: - type: object - x-runtime: [cloud] - description: "[cloud-only] Paginated execution history in v2 format." - required: - - items - - total - - has_more - properties: - items: - type: array - items: - $ref: "#/components/schemas/HistoryV2Entry" - total: - type: integer - has_more: - type: boolean - - HistoryV2Entry: - type: object - x-runtime: [cloud] - description: "[cloud-only] A single execution history entry in v2 format." - required: - - id - - status - properties: - id: - type: string - format: uuid - status: - type: string - workflow: - type: object - additionalProperties: true - outputs: - type: object - additionalProperties: true - created_at: - type: string - format: date-time - started_at: - type: string - format: date-time - nullable: true - completed_at: - type: string - format: date-time - nullable: true - preview_output: - type: object - additionalProperties: true - - CloudLogsResponse: - type: object - x-runtime: [cloud] - description: "[cloud-only] Paginated cloud execution logs." - required: - - entries - properties: - entries: - type: array - items: - type: object + schemas: + Asset: + description: Represents a user-owned asset (image, video, or other generated output). properties: - timestamp: - type: string - format: date-time - level: - type: string - enum: [debug, info, warn, error] - message: - type: string - job_id: - type: string - format: uuid - total: - type: integer - has_more: - type: boolean - - AssetDownloadRequest: - type: object - x-runtime: [cloud] - description: "[cloud-only] A single asset to download to the cloud runtime." - required: - - asset_id - properties: - asset_id: - type: string - format: uuid - description: ID of the asset to download - target_path: - type: string - description: Target path on the runtime filesystem - - ImportPublishedAssetsRequest: - type: object - x-runtime: [cloud] - description: "[cloud-only] Request body for importing published assets into the caller's library." - required: - - published_asset_ids - properties: - published_asset_ids: - type: array - description: IDs of published assets (inputs and models) to import. - items: - type: string - share_id: - type: string - nullable: true - description: | - Optional. Share ID of the published workflow these assets belong to. When provided (non-null, non-empty): all `published_asset_ids` must belong to this share's workflow version; returns 400 if the share is not found or any asset does not belong to it. When omitted, null, or empty string: no share-scoped validation is performed and the assets are validated only against global rules (preserved for clients that have not yet adopted `share_id`). - - ImportPublishedAssetsResponse: - type: object - x-runtime: [cloud] - description: "[cloud-only] Response after importing published assets. Each returned `AssetInfo.id` is the caller's newly-created private asset ID, not the published asset ID supplied in the request." - required: - - assets - properties: - assets: - type: array - items: - $ref: "#/components/schemas/AssetInfo" - - RemoteAssetMetadata: - type: object - x-runtime: [cloud] - description: "[cloud-only] Metadata fetched from a remote asset URL." - properties: - content_type: - type: string - description: MIME type of the remote file - content_length: - type: integer - format: int64 - description: Size in bytes - filename: - type: string - description: Suggested filename from Content-Disposition or URL - - CloudNode: - type: object - x-runtime: [cloud] - description: "[cloud-only] An installed custom node package in the cloud runtime." - required: - - id - - name - properties: - id: - type: string - name: - type: string - version: - type: string - description: - type: string - author: - type: string - repository: - type: string - format: uri - installed_at: - type: string - format: date-time - enabled: - type: boolean - - HubLabel: - type: object - x-runtime: [cloud] - description: "[cloud-only] A label/category used for tagging hub content." - required: - - id - - name - properties: - id: - type: string - name: - type: string - description: - type: string - color: - type: string - description: Hex color code for the label - - HubProfile: - type: object - x-runtime: [cloud] - description: "[cloud-only] A public user profile on the ComfyUI Hub." - required: - - username - properties: - username: - type: string - display_name: - type: string - bio: - type: string - avatar_url: - type: string - format: uri - links: - type: array - items: - type: string - format: uri - workflow_count: - type: integer - created_at: - type: string - format: date-time - - HubWorkflow: - type: object - x-runtime: [cloud] - description: "[cloud-only] A published workflow on the ComfyUI Hub." - required: - - share_id - - name - properties: - share_id: - type: string - name: - type: string - description: - type: string - author: - $ref: "#/components/schemas/HubProfile" - labels: - type: array - items: - $ref: "#/components/schemas/HubLabel" - thumbnail_url: - type: string - format: uri - content: - type: object - additionalProperties: true - description: Workflow graph JSON - likes: - type: integer - views: - type: integer - forks: - type: integer - created_at: - type: string - format: date-time - updated_at: - type: string - format: date-time - - HubWorkflowList: - type: object - x-runtime: [cloud] - description: "[cloud-only] Paginated list of hub workflows." - required: - - workflows - - total - - has_more - properties: - workflows: - type: array - items: - $ref: "#/components/schemas/HubWorkflow" - total: - type: integer - has_more: - type: boolean - - HubWorkflowIndexEntry: - type: object - x-runtime: [cloud] - description: "[cloud-only] Lightweight entry in the hub workflow index for client-side search." - required: - - share_id - - name - properties: - share_id: - type: string - name: - type: string - author_username: - type: string - labels: - type: array - items: - type: string - likes: - type: integer - updated_at: - type: string - format: date-time - - CloudWorkflow: - type: object - x-runtime: [cloud] - description: "[cloud-only] A cloud-managed workflow with version history." - required: - - id - - name - properties: - id: - type: string - format: uuid - name: - type: string - description: - type: string - share_id: - type: string - nullable: true - description: Public share identifier if published - latest_version_id: - type: string - format: uuid - nullable: true - thumbnail_url: - type: string - format: uri - nullable: true - created_at: - type: string - format: date-time - updated_at: - type: string - format: date-time - - CloudWorkflowList: - type: object - x-runtime: [cloud] - description: "[cloud-only] Paginated list of cloud workflows." - required: - - workflows - - total - - has_more - properties: - workflows: - type: array - items: - $ref: "#/components/schemas/CloudWorkflow" - total: - type: integer - has_more: - type: boolean - - CloudWorkflowVersion: - type: object - x-runtime: [cloud] - description: "[cloud-only] A version of a cloud workflow." - required: - - id - - workflow_id - properties: - id: - type: string - format: uuid - workflow_id: - type: string - format: uuid - version_number: - type: integer - created_at: - type: string - format: date-time - - AuthSession: - type: object - x-runtime: [cloud] - description: "[cloud-only] Current authentication session state." - required: - - user - properties: - user: - $ref: "#/components/schemas/CloudUser" - workspace: - $ref: "#/components/schemas/Workspace" - expires_at: - type: string - format: date-time - - AuthTokenResponse: - type: object - x-runtime: [cloud] - description: "[cloud-only] OAuth2 token response." - required: - - access_token - - token_type - properties: - access_token: - type: string - token_type: - type: string - description: Always "Bearer" - expires_in: - type: integer - description: Token lifetime in seconds - refresh_token: - type: string - nullable: true - scope: - type: string - - JwksResponse: - type: object - x-runtime: [cloud] - description: "[cloud-only] JSON Web Key Set for JWT verification." - required: - - keys - properties: - keys: - type: array - items: - type: object + created_at: + description: Timestamp when the asset was created + format: date-time + type: string + display_name: + description: Display name of the asset. Mirrors name for backwards compatibility. + nullable: true + type: string + file_path: + description: Relative path in global-namespace-root form (e.g. "models/checkpoints/flux.safetensors") + nullable: true + type: string + hash: + description: Blake3 hash of the asset content. + pattern: ^blake3:[a-f0-9]{64}$ + type: string + id: + description: Unique identifier for the asset + format: uuid + type: string + is_immutable: + description: Whether this asset is immutable (cannot be modified or deleted) + type: boolean + job_id: + description: ID of the job that created this asset, if available + format: uuid + nullable: true + type: string + last_access_time: + description: Timestamp when the asset was last accessed + format: date-time + type: string + metadata: + additionalProperties: true + description: System-managed metadata from download sources (HuggingFace, CivitAI, etc.) - read-only, not user-modifiable + readOnly: true + type: object + mime_type: + description: MIME type of the asset + type: string + name: + description: Name of the asset file + type: string + preview_id: + description: ID of the preview asset if available + format: uuid + nullable: true + type: string + preview_url: + description: URL for asset preview/thumbnail + format: uri + type: string + size: + description: Size of the asset in bytes + format: int64 + type: integer + tags: + description: Tags associated with the asset + items: + type: string + type: array + updated_at: + description: Timestamp when the asset was last updated + format: date-time + type: string + user_metadata: + additionalProperties: true + description: Custom user metadata for the asset + type: object required: - - kty - - kid - - use + - id + - name + - created_at + - updated_at + type: object + AssetCreated: + allOf: + - $ref: '#/components/schemas/Asset' + - properties: + created_new: + description: Whether this was a new asset creation (true) or returned existing (false) + type: boolean + required: + - created_new + type: object + description: Response returned when a new asset is successfully created. + AssetInfo: + description: Lightweight asset reference used in workflow publishing payloads. properties: - kty: - type: string - description: Key type (e.g. RSA) - kid: - type: string - description: Key ID - use: - type: string - description: Key use (e.g. sig) - alg: - type: string - description: Algorithm (e.g. RS256) - n: - type: string - description: RSA modulus (base64url) - e: - type: string - description: RSA exponent (base64url) - additionalProperties: true - - OAuthAuthorizationServerMetadata: - type: object - x-runtime: [cloud] - description: "[cloud-only] OAuth 2.1 authorization-server metadata (RFC 8414)." - required: - - issuer - - authorization_endpoint - - token_endpoint - - jwks_uri - - response_types_supported - - grant_types_supported - - code_challenge_methods_supported - - token_endpoint_auth_methods_supported - properties: - issuer: - type: string - format: uri - authorization_endpoint: - type: string - format: uri - token_endpoint: - type: string - format: uri - jwks_uri: - type: string - format: uri - registration_endpoint: - type: string - format: uri - description: "[cloud-only] RFC 7591 §3.1 Dynamic Client Registration endpoint. Advertised so MCP-spec-compliant clients can auto-discover and self-register without operator involvement. Present only when DCR is enabled." - response_types_supported: - type: array - items: - type: string - grant_types_supported: - type: array - items: - type: string - code_challenge_methods_supported: - type: array - items: - type: string - token_endpoint_auth_methods_supported: - type: array - items: - type: string - scopes_supported: - type: array - items: - type: string - - OAuthProtectedResourceMetadata: - type: object - x-runtime: [cloud] - description: "[cloud-only] OAuth 2.1 protected-resource metadata (RFC 9728)." - required: - - resource - - authorization_servers - - scopes_supported - properties: - resource: - type: string - format: uri - authorization_servers: - type: array - items: - type: string - format: uri - scopes_supported: - type: array - items: - type: string - bearer_methods_supported: - type: array - items: - type: string - - OAuthConsentChallenge: - type: object - x-runtime: [cloud] - description: "[cloud-only] Server-side state describing the OAuth consent decision the user is being asked to make. Returned by GET /oauth/authorize when a valid session exists; the frontend renders the consent UI from this payload and POSTs the decision back. Browser never sees the original OAuth params on resume." - required: - - oauth_request_id - - csrf_token - - client_display_name - - resource_display_name - - scopes - - workspaces - properties: - oauth_request_id: - type: string - format: uuid - description: Opaque server-side identifier for the authorization-request row. Carried back unchanged in the consent submission. - csrf_token: - type: string - description: Per-row CSRF token bound to this authorization request (not to the session). Must be echoed back on POST. - client_display_name: - type: string - description: Human-readable name of the OAuth client requesting authorization. - resource_display_name: - type: string - description: Human-readable name of the protected resource. - scopes: - type: array - description: Scopes the client is requesting for this resource. The frontend should present these for the user to approve. - items: - type: string - workspaces: - type: array - description: Workspaces the user can select from. Membership is re-checked on POST. - items: - $ref: "#/components/schemas/OAuthConsentChallengeWorkspace" - - OAuthConsentChallengeWorkspace: - type: object - x-runtime: [cloud] - description: "[cloud-only] One workspace option presented in the OAuth consent challenge." - required: [id, name, type, role] - properties: - id: { type: string } - name: { type: string } - type: { type: string, enum: [personal, team] } - role: { type: string, enum: [owner, member] } - - OAuthAuthorizeRedirectResponse: - type: object - x-runtime: [cloud] - description: "[cloud-only] Redirect target produced after a JSON consent submission. The frontend must navigate the browser to this URL so custom-scheme client callbacks work without relying on fetch-visible 302 headers." - required: - - redirect_url - properties: - redirect_url: - type: string - format: uri - description: OAuth client redirect URI with either code+state for allow, or error+state for deny. - - OAuthTokenResponse: - type: object - x-runtime: [cloud] - description: "[cloud-only] RFC 6749 §5.1 successful token response." - required: [access_token, token_type, expires_in, refresh_token, scope] - properties: - access_token: - type: string - description: Resource-bound access token (audience matches the protected resource). - token_type: - type: string - enum: [Bearer] - expires_in: - type: integer - description: Access token lifetime in seconds. - refresh_token: - type: string - description: Opaque refresh token. Rotates on every successful refresh; presenting an already-rotated token revokes the entire family. - scope: - type: string - description: Space-delimited scopes granted with this token. - - OAuthTokenError: - type: object - x-runtime: [cloud] - description: "[cloud-only] RFC 6749 §5.2 error response." - required: [error] - properties: - error: - type: string - description: 'RFC 6749 §5.2 error code: invalid_request, invalid_client, invalid_grant, unauthorized_client, unsupported_grant_type, invalid_scope.' - error_description: - type: string - description: Human-readable, no leak of internal storage state. - - OAuthRegisterRequest: - type: object - x-runtime: [cloud] - additionalProperties: false - description: "[cloud-only] RFC 7591 §2 client metadata document. Only the fields the server honors are listed; presence of `scope` or `resource_grants` in the request is rejected (`invalid_client_metadata`) because those are server-owned for dynamic clients." - required: - - redirect_uris - - application_type - properties: - redirect_uris: - type: array - items: - type: string - minItems: 1 - maxItems: 5 - description: 1–5 redirect URIs. Validated against `application_type` policy. - client_name: - type: string - maxLength: 100 - description: Human-readable name shown in the consent UI. Reserved-name list rejects impersonation of major clients. - application_type: - type: string - enum: [native, web] - description: | - RFC 7591 §2 application_type. **REQUIRED** — clients MUST declare intent; the server does not default this field. `native` for desktop / CLI / MCP-spec-strict clients (loopback redirects); `web` for hosted clients (HTTPS only, host must be allowlisted). A missing or explicitly empty `application_type` rejects with `invalid_client_metadata`. - token_endpoint_auth_method: - type: string - enum: [none] - description: 'Public clients only this phase — must be `none` if present. The server forces `none` regardless.' - grant_types: - type: array - items: - type: string - enum: [authorization_code, refresh_token] - description: Optional. Defaults to `["authorization_code","refresh_token"]`. - response_types: - type: array - items: - type: string - enum: [code] - description: Optional. Defaults to `["code"]`. - scope: - type: string - nullable: true - description: "**REJECTED IF PRESENT.** Dynamic clients do not pick scopes — the server assigns scopes from the active resource's published list. Sending `scope` in the registration body is treated as a privilege-escalation attempt and returns `invalid_client_metadata`." - resource_grants: - type: object - nullable: true - additionalProperties: - type: array + id: + description: Asset identifier. + type: string + in_library: + description: Whether the caller already owns this asset. + type: boolean + model: + description: Whether this asset is a model. + type: boolean + name: + type: string + preview_url: + description: Signed URL for previewing the asset. + type: string + public: + description: Whether this is a public (platform-provided) asset. + type: boolean + storage_url: + type: string + required: + - id + - name + - preview_url + - storage_url + - model + - public + - in_library + type: object + AssetTagHistogramResponse: + description: Histogram of tag counts used for refining asset search results. + properties: + tag_counts: + additionalProperties: + type: integer + description: Map of tag names to their occurrence counts on matching assets + example: + checkpoint: 32 + lora: 193 + vae: 6 + type: object + required: + - tag_counts + type: object + AssetUpdated: + description: Response returned when an existing asset is successfully updated. + properties: + display_name: + description: Display name of the asset. Mirrors name for backwards compatibility. + nullable: true + type: string + file_path: + description: Relative path in global-namespace-root form (e.g. "models/checkpoints/flux.safetensors") + nullable: true + type: string + hash: + description: Blake3 hash of the asset content. + pattern: ^blake3:[a-f0-9]{64}$ + type: string + id: + description: Asset ID + format: uuid + type: string + job_id: + description: ID of the job that created this asset, if available + format: uuid + nullable: true + type: string + mime_type: + description: Updated MIME type of the asset + type: string + name: + description: Updated name of the asset + type: string + tags: + description: Tags associated with the asset + items: + type: string + type: array + updated_at: + description: Timestamp of the update + format: date-time + type: string + user_metadata: + additionalProperties: true + description: Updated custom metadata + type: object + required: + - id + - updated_at + type: object + CreateWorkflowRequest: + description: Request body for creating a new saved workflow. + properties: + default_view: + description: Default view mode + enum: + - workflow + - app + type: string + description: + description: Description of the workflow + type: string + forked_from_workflow_id: + description: ID of the source workflow if forked + type: string + forked_from_workflow_version_id: + description: ID of the source workflow version if forked + type: string + name: + description: Display name for the workflow + type: string + workflow_json: + additionalProperties: true + description: The ComfyUI workflow JSON + type: object + required: + - workflow_json + type: object + CreateWorkflowVersionRequest: + description: Request body for creating a new version of a saved workflow. + properties: + base_version: + description: The version number this change is based on (for optimistic concurrency) + type: integer + workflow_json: + additionalProperties: true + description: The updated ComfyUI workflow JSON + type: object + required: + - base_version + - workflow_json + type: object + ErrorResponse: + description: Standard error response with a machine-readable code and human-readable message. + properties: + code: + type: string + details: + additionalProperties: true + description: Optional open object carrying structured, machine-readable context about the error (e.g. offending field names, validation specifics). Absent for most errors; consumers must not assume any particular shape. + type: object + message: + type: string + required: + - code + - message + type: object + ExecutionError: + description: Detailed execution error information from ComfyUI + properties: + current_inputs: + additionalProperties: true + description: Input values at time of failure (empty object if not available) + type: object + current_outputs: + additionalProperties: true + description: Output values at time of failure (empty object if not available) + type: object + exception_message: + description: Human-readable error message + type: string + exception_type: + description: Python exception type (e.g., "RuntimeError") + type: string + node_id: + description: ID of the node that failed + type: string + node_type: + description: Type name of the node (e.g., "KSampler") + type: string + traceback: + description: Array of traceback lines (empty array if not available) + items: + type: string + type: array + required: + - node_id + - node_type + - exception_message + - exception_type + - traceback + - current_inputs + - current_outputs + type: object + FeedbackRequest: + description: Request to submit user feedback + properties: + content: + description: The feedback content or message + type: string + metadata: + additionalProperties: true + description: Additional metadata about the feedback + type: object + rating: + description: User's rating of ComfyUI Cloud experience (1-5 stars) + maximum: 5 + minimum: 1 + type: integer + type: + description: Type of feedback being submitted + enum: + - missing_nodes + - general + - missing_models + type: string + required: + - type + type: object + FeedbackResponse: + description: Response after submitting feedback + type: object + ForkWorkflowRequest: + description: Request body for forking an existing workflow into the user's account. + properties: + name: + description: Name for the forked workflow + type: string + source_version: + description: Version number to fork from + type: integer + required: + - source_version + type: object + GetUserDataResponseFull: + description: List of user data file entries (each with path, size, and modification time) returned when full_info=true. items: - type: string - description: "**REJECTED IF PRESENT.** Same reason as `scope`. The set of resources and scopes a dynamic client may request is server-policy, not request-driven." - client_uri: - type: string - nullable: true - description: "**REJECTED IF PRESENT.** Unsupported RFC 7591 metadata for this public-client phase." - logo_uri: - type: string - nullable: true - description: "**REJECTED IF PRESENT.** Unsupported RFC 7591 metadata for this public-client phase." - tos_uri: - type: string - nullable: true - description: "**REJECTED IF PRESENT.** Unsupported RFC 7591 metadata for this public-client phase." - policy_uri: - type: string - nullable: true - description: "**REJECTED IF PRESENT.** Unsupported RFC 7591 metadata for this public-client phase." - software_id: - type: string - nullable: true - description: "**REJECTED IF PRESENT.** Unsupported RFC 7591 metadata for this public-client phase." - software_version: - type: string - nullable: true - description: "**REJECTED IF PRESENT.** Unsupported RFC 7591 metadata for this public-client phase." - contacts: - type: array - nullable: true - items: - type: string - description: "**REJECTED IF PRESENT.** Unsupported RFC 7591 metadata for this public-client phase." - jwks: - type: object - nullable: true - additionalProperties: true - description: "**REJECTED IF PRESENT.** Unsupported RFC 7591 metadata for this public-client phase." - jwks_uri: - type: string - nullable: true - description: "**REJECTED IF PRESENT.** Unsupported RFC 7591 metadata for this public-client phase." + $ref: '#/components/schemas/GetUserDataResponseFullFile' + type: array + GetUserDataResponseFullFile: + description: Individual file entry within a full user data response. + properties: + modified: + description: UNIX timestamp of the last modification in milliseconds. + format: int64 + type: integer + path: + description: File name or path relative to the user directory. + type: string + size: + description: File size in bytes. + type: integer + type: object + GlobalSubgraphData: + description: Full data for a global subgraph blueprint + properties: + data: + description: The full subgraph JSON data as a string + type: string + info: + description: Additional information about the subgraph + properties: + node_pack: + description: The node pack/module that provides this subgraph + type: string + required: + - node_pack + type: object + name: + description: Display name of the subgraph blueprint + type: string + source: + description: Source type of the subgraph - "templates" for workflow templates or "custom_node" for custom node subgraphs + type: string + required: + - source + - name + - info + - data + type: object + GlobalSubgraphInfo: + description: Metadata for a global subgraph blueprint (without full data) + properties: + data: + description: The full subgraph JSON data (may be empty in list view) + type: string + info: + description: Additional information about the subgraph + properties: + node_pack: + description: The node pack/module that provides this subgraph + type: string + required: + - node_pack + type: object + name: + description: Display name of the subgraph blueprint + type: string + source: + description: Source type of the subgraph - "templates" for workflow templates or "custom_node" for custom node subgraphs + type: string + required: + - source + - name + - info + type: object + HistoryDetailEntry: + description: History entry with full prompt data + properties: + meta: + additionalProperties: true + description: Metadata about the execution and nodes + type: object + outputs: + additionalProperties: true + description: Output data from execution (generated images, files, etc.) + type: object + prompt: + description: Full prompt execution data + properties: + extra_data: + additionalProperties: true + description: Additional execution data + type: object + outputs_to_execute: + description: Output nodes to execute + items: + type: string + type: array + priority: + description: Execution priority + format: double + type: number + prompt: + additionalProperties: true + description: The workflow nodes + type: object + prompt_id: + description: The prompt ID + type: string + type: object + status: + additionalProperties: true + description: Execution status and timeline information + type: object + type: object + HistoryDetailResponse: + additionalProperties: + $ref: '#/components/schemas/HistoryDetailEntry' + description: | + Detailed execution history response for a specific prompt. + Returns a dictionary with prompt_id as key and full history data as value. + type: object + HistoryEntry: + description: History entry with prompt_id and execution data + properties: + create_time: + description: Job creation timestamp (Unix timestamp in milliseconds) + format: int64 + type: integer + meta: + additionalProperties: true + description: Metadata about the execution and nodes + type: object + outputs: + additionalProperties: true + description: Output data from execution (generated images, files, etc.) + type: object + prompt: + description: Filtered prompt execution data (lightweight format) + properties: + extra_data: + additionalProperties: true + description: Additional execution data (workflow removed from extra_pnginfo) + type: object + priority: + description: Execution priority + format: double + type: number + prompt_id: + description: The prompt ID + type: string + type: object + prompt_id: + description: Unique identifier for this prompt execution + type: string + status: + additionalProperties: true + description: Execution status and timeline information + type: object + workflow_id: + description: UUID identifying the workflow graph definition + type: string + required: + - prompt_id + type: object + HistoryManageRequest: + additionalProperties: false + description: Request to manage history operations + properties: + clear: + description: If true, clear all history for the authenticated user + type: boolean + delete: + description: Array of job IDs to delete from history + items: + type: string + type: array + type: object + HistoryResponse: + description: | + Execution history response with history array. + Returns an object with a "history" key containing an array of history entries. + Each entry includes prompt_id as a property along with execution data. + properties: + history: + description: Array of history entries ordered by creation time (newest first) + items: + $ref: '#/components/schemas/HistoryEntry' + type: array + required: + - history + type: object + JobCancelResponse: + description: Response for POST /api/jobs/{job_id}/cancel. Returned on both fresh cancels and idempotent no-ops. + properties: + cancelled: + description: | + True when a cancel event was successfully dispatched by this call. + False when the job was already in a terminal or cancelling state, + in which case the call is a no-op (still 200 — idempotent). + type: boolean + required: + - cancelled + type: object + JobDetailResponse: + description: Full job details including workflow and outputs + properties: + create_time: + description: Job creation timestamp (Unix timestamp in milliseconds) + format: int64 + type: integer + execution_error: + allOf: + - $ref: '#/components/schemas/ExecutionError' + description: Detailed execution error from ComfyUI (only for failed jobs with structured error data) + execution_meta: + additionalProperties: true + description: Node-level execution metadata (only for terminal states) + type: object + execution_status: + additionalProperties: true + description: ComfyUI execution status and timeline (only for terminal states) + type: object + id: + description: Unique job identifier + format: uuid + type: string + outputs: + additionalProperties: true + description: Full outputs object from ComfyUI (only for terminal states) + type: object + outputs_count: + description: Total number of output files (omitted for non-terminal states) + type: integer + preview_output: + additionalProperties: true + description: Primary preview output (only for terminal states) + type: object + status: + description: User-friendly job status + enum: + - pending + - in_progress + - completed + - failed + - cancelled + type: string + update_time: + description: Last update timestamp (Unix timestamp in milliseconds) + format: int64 + type: integer + workflow: + additionalProperties: true + description: | + Full ComfyUI workflow (10-100KB, omitted if not available). - OAuthRegisterResponse: - type: object - x-runtime: [cloud] - description: "[cloud-only] RFC 7591 §3.2.1 successful registration response." - required: - - client_id - - client_id_issued_at - - redirect_uris - - grant_types - - response_types - - token_endpoint_auth_method - - application_type - properties: - client_id: - type: string - description: Server-generated client_id. - client_id_issued_at: - type: integer - format: int64 - description: Unix timestamp (seconds) when the client was registered. - client_name: - type: string - redirect_uris: - type: array - items: - type: string - grant_types: - type: array - items: - type: string - response_types: - type: array - items: - type: string - token_endpoint_auth_method: - type: string - enum: [none] - application_type: - type: string - enum: [native, web] + Sensitive credentials are redacted before the response is returned: + `extra_data.api_key_comfy_org`, when present, is replaced with the + literal string `"[REDACTED]"`. The field is preserved (not removed) + so existence checks still pass, but the value is not usable. + type: object + workflow_id: + description: UUID identifying the workflow graph definition + type: string + required: + - id + - status + - create_time + - update_time + type: object + JobEntry: + description: Lightweight job data for list views (workflow and full outputs excluded) + properties: + create_time: + description: Job creation timestamp (Unix timestamp in milliseconds) + format: int64 + type: integer + execution_end_time: + description: Workflow execution completion timestamp (Unix milliseconds, only present for terminal states) + format: int64 + type: integer + execution_error: + allOf: + - $ref: '#/components/schemas/ExecutionError' + description: Detailed execution error from ComfyUI (only for failed jobs with structured error data) + execution_start_time: + description: Workflow execution start timestamp (Unix milliseconds, only present for terminal states) + format: int64 + type: integer + id: + description: Unique job identifier + format: uuid + type: string + outputs_count: + description: Total number of output files (omitted for non-terminal states) + type: integer + preview_output: + additionalProperties: true + description: Primary preview output (only present for terminal states) + type: object + status: + description: User-friendly job status + enum: + - pending + - in_progress + - completed + - failed + - cancelled + type: string + workflow_id: + description: UUID identifying the workflow graph definition + type: string + required: + - id + - status + - create_time + type: object + JobStatusResponse: + description: Job status information + properties: + assigned_inference: + description: The inference instance assigned to this job (if any) + nullable: true + type: string + created_at: + description: When the job was created + format: date-time + type: string + error_message: + description: Error message if the job failed + nullable: true + type: string + id: + description: The job ID + format: uuid + type: string + last_state_update: + description: When the job status was last changed + format: date-time + type: string + status: + description: Current job status + enum: + - waiting_to_dispatch + - pending + - in_progress + - completed + - error + - cancelled + type: string + updated_at: + description: When the job was last updated + format: date-time + type: string + required: + - id + - status + - created_at + - updated_at + type: object + JobsListResponse: + description: Paginated list of jobs for the authenticated user. + properties: + jobs: + description: Array of jobs ordered by specified sort field + items: + $ref: '#/components/schemas/JobEntry' + type: array + pagination: + $ref: '#/components/schemas/PaginationInfo' + required: + - jobs + - pagination + type: object + ListAssetsResponse: + description: Paginated list of assets belonging to the authenticated user. + properties: + assets: + description: List of assets matching the query + items: + $ref: '#/components/schemas/Asset' + type: array + has_more: + description: Whether more assets are available beyond this page + type: boolean + next_cursor: + description: | + Opaque cursor to pass as the `after` query parameter to fetch the + next page. Omitted from the response when there are no more results. + type: string + total: + description: Total number of assets matching the filters + type: integer + required: + - assets + - total + - has_more + type: object + ListTagsResponse: + description: Paginated list of available asset tags. + properties: + has_more: + description: Whether more tags are available + type: boolean + tags: + description: List of tags + items: + $ref: '#/components/schemas/TagInfo' + type: array + total: + description: Total number of tags + type: integer + required: + - tags + - total + - has_more + type: object + ModelFile: + description: Represents a model file with metadata + properties: + name: + description: The filename of the model + example: model.safetensors + type: string + pathIndex: + description: Index of the path where this model is located + example: 0 + type: integer + required: + - name + - pathIndex + type: object + ModelFolder: + description: Represents a folder containing models + properties: + folders: + description: List of paths where models of this type are stored + example: + - checkpoints + items: + type: string + type: array + name: + description: The name of the model folder + example: checkpoints + type: string + required: + - name + - folders + type: object + NodeInfo: + description: Metadata describing a single ComfyUI node type and its inputs/outputs. + properties: + api_node: + description: Whether this is an API node + type: boolean + category: + description: Category of the node + type: string + deprecated: + description: Whether the node is deprecated + type: boolean + description: + description: Description of the node + type: string + display_name: + description: Display name of the node + type: string + experimental: + description: Whether the node is experimental + type: boolean + input: + additionalProperties: true + description: Input specifications for the node + type: object + input_order: + additionalProperties: + items: + type: string + type: array + description: Order of inputs for display + type: object + name: + description: Internal name of the node + type: string + output: + description: Output types of the node + items: + type: string + type: array + output_is_list: + description: Whether each output is a list + items: + type: boolean + type: array + output_name: + description: Names of the outputs + items: + type: string + type: array + output_node: + description: Whether this is an output node + type: boolean + output_tooltips: + description: Tooltips for outputs + items: + type: string + type: array + python_module: + description: Python module implementing the node + type: string + type: object + PaginationInfo: + description: | + Pagination metadata included in list responses. Supports both legacy + offset/limit pagination and cursor-based pagination. When cursor-based + pagination is used, `next_cursor` is the primary pagination token and + `offset`/`total` may be zero. + properties: + has_more: + description: Whether more items are available beyond this page + type: boolean + limit: + description: Items per page + minimum: 1 + type: integer + next_cursor: + description: | + Opaque cursor for the next page. Pass this value as the `after` + query parameter on the next request. Empty or absent when there + are no more results. + type: string + offset: + deprecated: true + description: 'Current offset (0-based). Deprecated: use cursor-based pagination.' + minimum: 0 + type: integer + total: + description: Total number of items matching filters (may be 0 when using cursor pagination) + minimum: 0 + type: integer + required: + - offset + - limit + - total + - has_more + type: object + PromptErrorResponse: + additionalProperties: true + description: Error response for ComfyUI prompt execution. + type: object + PromptInfo: + description: Metadata about the currently running and queued prompts. + properties: + exec_info: + properties: + queue_remaining: + description: Number of items remaining in the queue + type: integer + type: object + type: object + PromptRequest: + description: Request body for submitting a ComfyUI workflow prompt for execution. + properties: + extra_data: + additionalProperties: true + description: Extra data to be associated with the prompt + type: object + front: + description: If true, adds the prompt to the front of the queue + type: boolean + number: + description: Priority number for the queue (lower numbers have higher priority) + type: number + partial_execution_targets: + description: List of node names to execute + items: + type: string + type: array + prompt: + additionalProperties: true + description: The workflow graph to execute + type: object + workflow_id: + description: UUID identifying the cloud workflow entity to associate with this job + type: string + workflow_version_id: + description: UUID identifying the workflow version to associate with this job + type: string + required: + - prompt + type: object + PromptResponse: + description: Response returned after successfully queuing a workflow prompt. + properties: + node_errors: + additionalProperties: true + description: Any errors in the nodes of the prompt + type: object + number: + description: Priority number in the queue + type: number + prompt_id: + description: Unique identifier for the prompt execution + format: uuid + type: string + type: object + PublishWorkflowAssetsRequest: + description: Request body for publishing workflow assets to the Hub. + properties: + asset_ids: + description: IDs of assets (inputs and models) to snapshot. + items: + type: string + type: array + required: + - asset_ids + type: object + PublishedWorkflowDetail: + description: Full detail of a publicly published workflow on the Hub. + properties: + assets: + description: Published assets with their library status for the caller. + items: + $ref: '#/components/schemas/AssetInfo' + type: array + listed: + type: boolean + name: + description: Human-readable workflow name. + type: string + publish_time: + format: date-time + nullable: true + type: string + share_id: + type: string + workflow_id: + type: string + workflow_json: + additionalProperties: true + description: The workflow JSON content at publish time. + type: object + required: + - share_id + - workflow_id + - name + - listed + - workflow_json + - assets + type: object + QueueInfo: + description: Queue information with pending and running jobs + properties: + queue_pending: + description: Array of pending job items (ordered by creation time, oldest first) + items: + description: | + Queue item tuple format: [job_number, prompt_id, workflow_json, output_node_ids, metadata] + - [0] job_number (integer): Position in queue (1-based) + - [1] prompt_id (string): Job UUID + - [2] workflow_json (object): Full ComfyUI workflow + - [3] output_node_ids (array): Node IDs to return results from + - [4] metadata (object): Contains {create_time: } + items: {} + maxItems: 5 + minItems: 5 + type: array + type: array + queue_running: + description: Array of currently running job items + items: + description: | + Queue item tuple format: [job_number, prompt_id, workflow_json, output_node_ids, metadata] + - [0] job_number (integer): Position in queue (1-based) + - [1] prompt_id (string): Job UUID + - [2] workflow_json (object): Full ComfyUI workflow + - [3] output_node_ids (array): Node IDs to return results from + - [4] metadata (object): Contains {create_time: } + items: {} + maxItems: 5 + minItems: 5 + type: array + type: array + type: object + QueueManageRequest: + additionalProperties: false + description: Request to manage queue operations + properties: + clear: + description: If true, clear all pending jobs from the queue + type: boolean + delete: + description: Array of PENDING job IDs to cancel + items: + type: string + type: array + type: object + QueueManageResponse: + description: Response after a queue management action (delete or clear). + properties: + cleared: + description: Whether the queue was cleared + type: boolean + deleted: + description: Array of job IDs that were successfully cancelled + items: + type: string + type: array + type: object + SystemStatsResponse: + description: System statistics response + properties: + devices: + items: + properties: + name: + description: Device name + type: string + type: + description: Device type + type: string + vram_free: + description: Free VRAM in bytes + type: number + vram_total: + description: Total VRAM in bytes + type: number + required: + - name + - type + type: object + type: array + system: + properties: + argv: + description: Command line arguments + items: + type: string + type: array + cloud_version: + description: Cloud ingest service version (commit hash) + type: string + comfyui_frontend_version: + description: ComfyUI frontend version (commit hash or tag) + type: string + comfyui_version: + description: ComfyUI version + type: string + embedded_python: + description: Whether using embedded Python + type: boolean + os: + description: Operating system + type: string + python_version: + description: Python version + type: string + pytorch_version: + description: PyTorch version + type: string + ram_free: + description: Free RAM in bytes + type: number + ram_total: + description: Total RAM in bytes + type: number + workflow_templates_version: + description: Workflow templates version + type: string + required: + - os + - python_version + - embedded_python + - comfyui_version + - pytorch_version + - argv + - ram_total + - ram_free + type: object + required: + - system + - devices + type: object + TagInfo: + description: Metadata for a single tag that can be applied to assets. + properties: + count: + description: Number of assets using this tag + type: integer + name: + description: Tag name + type: string + required: + - name + - count + type: object + TagsModificationResponse: + description: Response after adding, updating, or removing tags on an asset. + properties: + added: + description: Tags that were successfully added (for add operation) + items: + type: string + type: array + already_present: + description: Tags that were already present (for add operation) + items: + type: string + type: array + not_present: + description: Tags that were not present (for remove operation) + items: + type: string + type: array + removed: + description: Tags that were successfully removed (for remove operation) + items: + type: string + type: array + total_tags: + description: All tags on the asset after the operation + items: + type: string + type: array + required: + - total_tags + type: object + TaskEntry: + description: Task data for list views + properties: + completed_at: + description: When task completed or failed (null if not finished) + format: date-time + type: string + create_time: + description: Task creation timestamp + format: date-time + type: string + id: + description: Unique task identifier + format: uuid + type: string + started_at: + description: When task execution started (null if not started) + format: date-time + type: string + status: + description: Current task status + enum: + - created + - running + - completed + - failed + type: string + task_name: + description: Task type name (e.g., model_upload) + type: string + required: + - id + - task_name + - status + - create_time + type: object + TaskResponse: + description: Full task details including payload and result + properties: + completed_at: + description: When task completed or failed (null if not finished) + format: date-time + type: string + create_time: + description: Task creation timestamp + format: date-time + type: string + error_message: + description: Error message on failure (null if not failed) + type: string + id: + description: Unique task identifier + format: uuid + type: string + idempotency_key: + description: Caller-provided key for idempotent task creation + type: string + payload: + additionalProperties: true + description: Task input data + type: object + result: + additionalProperties: true + description: Task output data (null if not completed) + type: object + started_at: + description: When task execution started (null if not started) + format: date-time + type: string + status: + description: Current task status + enum: + - created + - running + - completed + - failed + type: string + task_name: + description: Task type name (e.g., model_upload) + type: string + update_time: + description: Task last update timestamp + format: date-time + type: string + required: + - id + - idempotency_key + - task_name + - payload + - status + - create_time + - update_time + type: object + TasksListResponse: + description: Paginated list of background tasks for the authenticated user. + properties: + pagination: + $ref: '#/components/schemas/PaginationInfo' + tasks: + description: Array of tasks ordered by create_time + items: + $ref: '#/components/schemas/TaskEntry' + type: array + required: + - tasks + - pagination + type: object + UpdateWorkflowRequest: + description: Request body for updating an existing saved workflow. + properties: + default_view: + description: New default view mode + enum: + - workflow + - app + type: string + description: + description: New description + type: string + name: + description: New display name + type: string + type: object + UserDataResponseFull: + description: User data listing entry with file metadata (path, size, modification time). + properties: + modified: + description: UNIX timestamp of the last modification in milliseconds. + format: int64 + type: integer + path: + type: string + size: + type: integer + type: object + UserResponse: + description: User information response + properties: + id: + description: Firebase UID of the authenticated user + type: string + status: + description: User status (always "active" for authenticated users) + type: string + required: + - id + - status + type: object + WorkflowForkedFrom: + description: Reference to the parent workflow from which this workflow was forked. + properties: + workflow_id: + type: string + workflow_version_id: + type: string + type: object + WorkflowListResponse: + description: Paginated list of saved workflows. + properties: + data: + items: + $ref: '#/components/schemas/WorkflowResponse' + type: array + pagination: + $ref: '#/components/schemas/PaginationInfo' + required: + - data + - pagination + type: object + WorkflowPublishInfo: + description: Publishing metadata for a workflow shared to the Hub. + properties: + assets: + description: Published assets (inputs and models). + items: + $ref: '#/components/schemas/AssetInfo' + type: array + listed: + type: boolean + publish_time: + format: date-time + nullable: true + type: string + share_id: + type: string + workflow_id: + type: string + required: + - workflow_id + - share_id + - listed + - assets + type: object + WorkflowResponse: + description: Full workflow entity including metadata and version history. + properties: + created_at: + format: date-time + type: string + created_by: + type: string + default_view: + enum: + - workflow + - app + type: string + description: + type: string + forked_from: + $ref: '#/components/schemas/WorkflowForkedFrom' + id: + type: string + latest_version: + type: integer + name: + type: string + updated_at: + format: date-time + type: string + required: + - id + - latest_version + - created_by + - created_at + - updated_at + type: object + WorkflowVersionContentResponse: + description: Full workflow version including the serialized workflow JSON. + properties: + created_at: + format: date-time + type: string + created_by: + type: string + dependency_asset_ids: + items: + type: string + type: array + id: + type: string + version: + type: integer + workflow_json: + additionalProperties: true + type: object + required: + - id + - version + - workflow_json + - created_by + - created_at + type: object + WorkflowVersionResponse: + description: Metadata for a single workflow version. + properties: + created_at: + format: date-time + type: string + created_by: + type: string + id: + type: string + latest_version: + type: integer + version: + type: integer + required: + - id + - version + - latest_version + - created_by + - created_at + type: object + securitySchemes: + ApiKeyAuth: + description: | + API key authentication. Keys are prefixed with 'comfyui-' and can be + generated from user account settings. Example: 'comfyui-abc123...' + in: header + name: X-API-Key + type: apiKey + BearerAuth: + bearerFormat: JWT + description: | + Firebase JWT token authentication. Obtain a token by authenticating + with Firebase and pass it in the Authorization header. + scheme: bearer + type: http + CookieAuth: + description: | + Session cookie authentication. Set automatically after successful + login via the /api/auth/session endpoint. + in: cookie + name: session + type: apiKey +info: + description: | + API for ComfyUI - A powerful and modular UI for Stable Diffusion. - OAuthRegisterError: - type: object - x-runtime: [cloud] - description: "[cloud-only] RFC 7591 §3.2.2 error response." - required: - - error - properties: - error: - type: string - enum: [invalid_redirect_uri, invalid_client_metadata] - error_description: - type: string - nullable: true + This API allows you to interact with ComfyUI programmatically, including: + - Retrieving prompt information + - Retrieving node information + license: + name: GNU General Public License v3.0 + url: https://github.com/Comfy-Org/ComfyUI/blob/master/LICENSE + title: ComfyUI API + version: 1.0.0 +openapi: 3.0.3 +paths: + /api/assets: + get: + description: | + Retrieves a paginated list of assets belonging to the authenticated user. + Supports filtering by tags, name, metadata, and sorting options. + operationId: listAssets + parameters: + - description: Filter assets that have ALL of these tags + explode: false + in: query + name: include_tags + schema: + items: + type: string + type: array + style: form + - description: Exclude assets that have ANY of these tags + explode: false + in: query + name: exclude_tags + schema: + items: + type: string + type: array + style: form + - description: Filter assets where name contains this substring (case-insensitive) + in: query + name: name_contains + schema: + type: string + - description: JSON object for filtering by metadata fields + in: query + name: metadata_filter + schema: + type: string + - description: Maximum number of assets to return (1-500) + in: query + name: limit + schema: + default: 20 + maximum: 500 + minimum: 1 + type: integer + - description: Number of assets to skip for pagination + in: query + name: offset + schema: + default: 0 + minimum: 0 + type: integer + - description: Field to sort by + in: query + name: sort + schema: + default: created_at + enum: + - name + - created_at + - updated_at + - size + - last_access_time + type: string + - description: Sort order + in: query + name: order + schema: + default: desc + enum: + - asc + - desc + type: string + - description: Whether to include public/shared assets in results + in: query + name: include_public + schema: + default: true + type: boolean + - description: Filter assets by exact content hash. + in: query + name: hash + schema: + type: string + - description: | + Opaque cursor for keyset pagination. Pass the `next_cursor` value + from the previous response to fetch the next page. When provided, + `offset` is ignored. Cursor pagination is only supported with + `sort` values `created_at`, `updated_at`, `name`, or `size`; + requests combining `after` with other sort fields return 400. + The cursor must have been minted under the same `sort` value used + in the follow-up request. + in: query + name: after + schema: + type: string + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/ListAssetsResponse' + description: Success - Assets returned + "400": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Invalid request parameters + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: List user assets + tags: + - file + post: + description: | + Creates a new asset from a direct file upload (multipart/form-data) with associated metadata. - BillingBalance: - type: object - x-runtime: [cloud] - description: "[cloud-only] Current credit balance and usage summary." - required: - - credits_remaining - properties: - credits_remaining: - type: integer - description: Available credits - credits_used: - type: integer - description: Credits used in current billing period - credits_total: - type: integer - description: Total credits allocated in current period + If an asset with the same hash already exists, returns the existing asset. + operationId: createAsset + requestBody: + content: + multipart/form-data: + schema: + properties: + file: + description: The asset file to upload + format: binary + type: string + hash: + description: Content hash of the file. + pattern: ^(blake3|sha256):[a-f0-9]{64}$ + type: string + id: + description: Optional asset ID for idempotent creation. If provided and asset exists, returns existing asset. + format: uuid + type: string + mime_type: + description: MIME type of the asset (e.g., "image/png", "video/mp4") + type: string + name: + description: Display name for the asset + type: string + preview_id: + description: Optional preview asset ID. If not provided, images will use their own ID as preview. + format: uuid + type: string + tags: + description: JSON-encoded array of freeform tag strings, e.g. '["models","checkpoint"]'. Common types include "models", "input", "output", and "temp", but any tag can be used in any order. + type: string + user_metadata: + description: Custom JSON metadata as a string + type: string + required: + - file + type: object + required: true + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/AssetCreated' + description: | + Asset already existed for this user (deduplicated by content hash); the + existing asset is returned with created_new=false. + "201": + content: + application/json: + schema: + $ref: '#/components/schemas/AssetCreated' + description: Asset created successfully (created_new=true) + "400": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Invalid request (bad file, invalid content type, etc.) + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "413": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: File too large + "415": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unsupported media type + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Create a new asset + tags: + - file + /api/assets/{id}: + delete: + description: Deletes the asset record. + operationId: deleteAsset + parameters: + - description: Asset ID + in: path + name: id + required: true + schema: + format: uuid + type: string + responses: + "204": + description: Asset record deleted successfully + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Asset not found + "409": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: 'Asset cannot be deleted because it is referenced by another resource, e.g. a workflow version (error code: ASSET_IN_USE)' + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Delete asset + tags: + - file + get: + description: Retrieves detailed information about a specific asset + operationId: getAssetById + parameters: + - description: Asset ID + in: path + name: id + required: true + schema: + format: uuid + type: string + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/Asset' + description: Asset details retrieved successfully + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Asset not found + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Get asset details + tags: + - file + put: + description: | + Updates an asset's metadata. At least one field must be provided. + Only name, mime_type, preview_id, and user_metadata can be updated. + For tag management, use POST (add) and DELETE (remove) /api/assets/{id}/tags. + operationId: updateAsset + parameters: + - description: Asset ID + in: path + name: id + required: true + schema: + format: uuid + type: string + requestBody: + content: + application/json: + schema: + minProperties: 1 + properties: + mime_type: + description: Updated MIME type of the asset + type: string + name: + description: New display name for the asset + type: string + preview_id: + description: Updated preview asset ID + format: uuid + type: string + user_metadata: + additionalProperties: true + description: Updated custom metadata + type: object + type: object + required: true + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/AssetUpdated' + description: Asset updated successfully + "400": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Invalid request (no fields provided) + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Asset not found + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Update asset metadata + tags: + - file + /api/assets/{id}/tags: + delete: + description: Removes one or more tags from an existing asset + operationId: removeAssetTags + parameters: + - description: Asset ID + in: path + name: id + required: true + schema: + format: uuid + type: string + requestBody: + content: + application/json: + schema: + properties: + tags: + description: Tags to remove from the asset + items: + type: string + minItems: 1 + type: array + required: + - tags + type: object + required: true + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/TagsModificationResponse' + description: Tags removed successfully + "400": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Invalid request + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Asset not found + "422": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Validation error (e.g., reserved tag) + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Remove tags from asset + tags: + - file + post: + description: Adds one or more tags to an existing asset + operationId: addAssetTags + parameters: + - description: Asset ID + in: path + name: id + required: true + schema: + format: uuid + type: string + requestBody: + content: + application/json: + schema: + properties: + tags: + description: Tags to add to the asset + items: + type: string + minItems: 1 + type: array + required: + - tags + type: object + required: true + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/TagsModificationResponse' + description: Tags added successfully + "400": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Invalid request + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Asset not found + "422": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Validation error (e.g., reserved tag) + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Add tags to asset + tags: + - file + /api/assets/from-hash: + post: + description: | + Creates a new asset reference using an existing asset's hash. + This avoids re-uploading the file content when the asset already exists in storage. + The user can provide their own metadata and tags for the reference. + operationId: createAssetFromHash + requestBody: + content: + application/json: + schema: + properties: + hash: + description: 'Blake3 content hash of the existing asset (blake3: prefix)' + pattern: ^blake3:[a-f0-9]{64}$ + type: string + mime_type: + description: MIME type of the asset (e.g., "image/png", "video/mp4") + type: string + name: + description: Display name for the asset reference (optional) + type: string + tags: + description: Freeform tags for the asset. Common types include "models", "input", "output", and "temp", but any tag can be used in any order. + items: + type: string + minItems: 1 + type: array + user_metadata: + additionalProperties: true + description: Custom metadata for this asset reference + type: object + required: + - hash + - tags + type: object + required: true + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/AssetCreated' + description: | + Asset reference already existed for this user (deduplicated by content + hash); the existing asset is returned with created_new=false. + "201": + content: + application/json: + schema: + $ref: '#/components/schemas/AssetCreated' + description: Asset reference created successfully (created_new=true) + "400": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Invalid request (bad hash format, invalid tags, etc.) + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Source asset with given hash not found + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Create asset reference from existing hash + tags: + - file + /api/assets/hash/{hash}: + head: + description: | + Checks if an asset exists in the system by its blake3 hash. + Returns 200 if the asset exists, 404 if it doesn't. + operationId: checkAssetByHash + parameters: + - description: Blake3 hash of the asset in format 'blake3:hex_digest' + in: path + name: hash + required: true + schema: + example: blake3:a1b2c3d4e5f67890123456789012345678901234567890123456789012345678 + pattern: ^blake3:[a-f0-9]{64}$ + type: string + responses: + "200": + description: Asset exists + "400": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Invalid hash format + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "404": + description: Asset not found + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Check if asset exists by hash + tags: + - file + /api/assets/prune: + post: + description: Starts a background job that removes asset entries whose underlying content no longer exists on disk. + operationId: pruneAssets + responses: + "200": + content: + application/json: + schema: + properties: + marked: + description: Number of assets marked as missing + type: integer + status: + type: string + type: object + description: Prune result + summary: Mark assets whose backing files no longer exist on disk + /api/assets/seed: + post: + description: Starts a background job that scans configured directories and registers assets not yet in the asset database. + operationId: seedAssets + requestBody: + content: + application/json: + schema: + properties: + roots: + description: Root folder paths to scan (if omitted, scans all) + items: + type: string + type: array + type: object + responses: + "200": + content: + application/json: + schema: + properties: + status: + type: string + type: object + description: Seed started + summary: Trigger asset scan/seed from filesystem + /api/assets/seed/cancel: + post: + description: Requests cancellation of the currently-running asset seed job. + operationId: cancelAssetSeed + responses: + "200": + content: + application/json: + schema: + properties: + status: + type: string + type: object + description: Scan cancelled + summary: Cancel an in-progress asset scan + /api/assets/seed/status: + get: + description: Returns progress/status of the most recent asset seed job. + operationId: getAssetSeedStatus + responses: + "200": + content: + application/json: + schema: + additionalProperties: true + description: Scan progress details (files scanned, total, status, etc.) + type: object + description: Scan progress + summary: Get asset scan progress + /api/assets/tags/refine: + get: + description: | + Returns a histogram of tags appearing on assets matching the given filters. + Useful for refining asset searches by showing available tags and their counts. + Only returns tags with non-zero counts (tags that exist on matching assets). + operationId: getAssetTagHistogram + parameters: + - description: Filter assets that have ALL of these tags + explode: false + in: query + name: include_tags + schema: + items: + type: string + type: array + style: form + - description: Exclude assets that have ANY of these tags + explode: false + in: query + name: exclude_tags + schema: + items: + type: string + type: array + style: form + - description: Filter assets where name contains this substring (case-insensitive) + in: query + name: name_contains + schema: + type: string + - description: JSON object for filtering by metadata fields + in: query + name: metadata_filter + schema: + type: string + - description: Maximum number of tags to return (1-1000, default 100) + in: query + name: limit + schema: + default: 100 + maximum: 1000 + minimum: 1 + type: integer + - description: Whether to include public/shared assets in results + in: query + name: include_public + schema: + default: true + type: boolean + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/AssetTagHistogramResponse' + description: Success - Tag histogram returned + "400": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Invalid request parameters + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Get tag histogram for filtered assets + tags: + - file + /api/embeddings: + get: + description: Returns the list of text-encoder embeddings available on disk. + operationId: getEmbeddings + responses: + "200": + content: + application/json: + schema: + items: + type: string + type: array + description: Embedding names + summary: List available embedding names + /api/experiment/models: + get: + description: | + Returns a list of model folders available in the system. + This is an experimental endpoint that replaces the legacy /models endpoint. + operationId: getModelFolders + responses: + "200": + content: + application/json: + schema: + items: + $ref: '#/components/schemas/ModelFolder' + type: array + description: Success - List of model folders + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + security: [] + summary: Get available model folders + tags: + - file + /api/experiment/models/{folder}: + get: + description: | + Returns a list of models available in the specified folder. + This is an experimental endpoint that provides enhanced model information. + operationId: getModelsInFolder + parameters: + - description: The folder name to list models from + in: path + name: folder + required: true + schema: + example: checkpoints + type: string + responses: + "200": + content: + application/json: + schema: + items: + $ref: '#/components/schemas/ModelFile' + type: array + description: Success - List of models in the folder + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Folder not found or no models in folder + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + security: [] + summary: Get models in a specific folder + tags: + - file + /api/extensions: + get: + description: | + Returns the list of custom node web extension JS files available for + loading by the ComfyUI frontend. Paths are relative to the web root + (e.g. `/extensions/VHS.core.js`). + operationId: getExtensions + responses: + "200": + content: + application/json: + schema: + description: URL paths (relative to web root) of available extension JS files + items: + type: string + type: array + description: JSON array of extension file paths + security: [] + summary: List custom node JS extensions + tags: + - node + /api/features: + get: + description: Returns the server's feature capabilities + operationId: getFeatures + responses: + "200": + content: + application/json: + schema: + additionalProperties: true + properties: + max_upload_size: + description: Maximum upload size in bytes + type: integer + supports_preview_metadata: + description: Whether the server supports preview metadata + type: boolean + type: object + description: Success + headers: + Cache-Control: + description: Short-lived private cache to deduplicate rapid-fire calls from the frontend + schema: + type: string + Vary: + description: Cache key includes auth headers so anonymous and authenticated responses are stored separately + schema: + type: string + security: + - ApiKeyAuth: [] + - BearerAuth: [] + - CookieAuth: [] + - {} + summary: Get server feature flags + tags: + - node + /api/feedback: + post: + description: Submit feedback about the ComfyUI service + operationId: submitFeedback + requestBody: + content: + application/json: + schema: + $ref: '#/components/schemas/FeedbackRequest' + required: true + responses: + "201": + content: + application/json: + schema: + $ref: '#/components/schemas/FeedbackResponse' + description: Feedback submitted successfully + "400": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Invalid request + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Submit user feedback + tags: + - feedback + /api/files/mask-layers: + get: + description: | + Given a mask file (any of the 4 layers), returns all related mask layer files. + This is used by the mask editor to load the paint, mask, and painted layers + when reopening a previously edited mask. + operationId: getMaskLayers + parameters: + - description: Hash filename of any mask layer file + in: query + name: filename + required: true + schema: + example: abc123def456.png + type: string + responses: + "200": + content: + application/json: + schema: + properties: + mask: + description: Filename of the mask layer + nullable: true + type: string + paint: + description: Filename of the paint strokes layer + nullable: true + type: string + painted: + description: Filename of the painted image layer + nullable: true + type: string + painted_masked: + description: Filename of the final composite layer + nullable: true + type: string + type: object + description: Success - Related mask layers returned + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: File not found or not a mask file + summary: Get related mask layer files + tags: + - file + /api/free: + post: + description: Frees GPU memory by unloading models and/or freeing the resident model cache. + operationId: freeMemory + requestBody: + content: + application/json: + schema: + properties: + free_memory: + description: Run garbage collection and free cached memory + type: boolean + unload_models: + description: Unload all models from VRAM/RAM + type: boolean + type: object + responses: + "200": + description: Memory freed + summary: Free GPU memory and/or unload models + /api/global_subgraphs: + get: + description: | + Returns a list of globally available subgraph blueprints. + These are pre-built workflow components that can be used as nodes. + The data field contains a promise that resolves to the full subgraph JSON. + operationId: getGlobalSubgraphs + responses: + "200": + content: + application/json: + schema: + additionalProperties: + $ref: '#/components/schemas/GlobalSubgraphInfo' + type: object + description: Success - Map of subgraph IDs to their metadata + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + security: [] + summary: Get available subgraph blueprints + tags: + - workflow + /api/global_subgraphs/{id}: + get: + description: Returns the full data for a specific subgraph blueprint by ID + operationId: getGlobalSubgraph + parameters: + - description: The unique identifier of the subgraph blueprint + in: path + name: id + required: true + schema: + type: string + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/GlobalSubgraphData' + description: Success - Full subgraph data + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Subgraph not found + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + security: [] + summary: Get a specific subgraph blueprint + tags: + - workflow + /api/history: + post: + deprecated: true + description: | + **Deprecated.** Superseded by the job-management endpoints under + `/api/jobs`. Planned for removal no earlier than a future major + release; sunset timeline TBD. - BillingEvent: - type: object - x-runtime: [cloud] - description: "[cloud-only] A billing event (charge, credit, refund)." - required: - - id - - type - - amount - - created_at - properties: - id: - type: string - type: - type: string - enum: [charge, credit, refund, topup, subscription] - amount: - type: integer - description: Amount in credits - description: - type: string - job_id: - type: string - format: uuid - nullable: true - created_at: - type: string - format: date-time + Clear all history for the authenticated user or delete specific job IDs. + Supports clearing all history or deleting specific job IDs. + operationId: manageHistory + requestBody: + content: + application/json: + schema: + $ref: '#/components/schemas/HistoryManageRequest' + required: true + responses: + "200": + description: Success - History management operation completed + "400": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Invalid request parameters + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized - Authentication required + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Manage execution history + tags: + - workflow + /api/history_v2: + get: + deprecated: true + description: | + **Deprecated.** Superseded by `GET /api/jobs`, which returns the same + execution records in a paginated, filterable format. Planned for removal + no earlier than a future major release; sunset timeline TBD. - BillingEventList: - type: object - x-runtime: [cloud] - description: "[cloud-only] Paginated list of billing events." - required: - - events - - total - - has_more - properties: - events: - type: array - items: - $ref: "#/components/schemas/BillingEvent" - total: - type: integer - has_more: - type: boolean + Retrieve execution history for the authenticated user with pagination support. + Returns a lightweight history format with filtered prompt data (workflow removed from extra_pnginfo). + operationId: getHistory + parameters: + - description: Maximum number of items to return + in: query + name: max_items + schema: + type: integer + - description: Starting position (default 0) + in: query + name: offset + schema: + default: 0 + type: integer + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/HistoryResponse' + description: Success - Execution history retrieved + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized - Authentication required + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Get execution history (v2) + tags: + - workflow + /api/history_v2/{prompt_id}: + get: + deprecated: true + description: | + **Deprecated.** Superseded by `GET /api/jobs/{job_id}`, which returns + the same execution record. Planned for removal no earlier than a future + major release; sunset timeline TBD. - BillingOp: - type: object - x-runtime: [cloud] - description: "[cloud-only] A billing operation record." - required: - - id - - status - properties: - id: - type: string - status: - type: string - enum: [pending, completed, failed] - type: - type: string - amount: - type: integer - created_at: - type: string - format: date-time - completed_at: - type: string - format: date-time - nullable: true + Retrieve detailed execution history for a specific prompt ID. + Returns full history data including complete prompt information. + operationId: getHistoryForPrompt + parameters: + - description: The prompt ID to retrieve history for + in: path + name: prompt_id + required: true + schema: + type: string + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/HistoryDetailResponse' + description: Success - History for prompt retrieved + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized - Authentication required + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Prompt not found + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Get history for specific prompt + tags: + - workflow + /api/i18n: + get: + description: Returns translation file URLs contributed by custom nodes, keyed by locale. + operationId: getI18n + responses: + "200": + content: + application/json: + schema: + additionalProperties: true + description: Nested map of locale to translation key-value pairs + type: object + description: Translation map + summary: Get internationalisation translation strings + /api/interrupt: + post: + description: | + Cancel all currently RUNNING jobs for the authenticated user. + This will interrupt any job that is currently in 'in_progress' status. + Note: This endpoint only affects running jobs. To cancel pending jobs, use /api/queue. + operationId: interruptJob + responses: + "200": + description: Success - Job interrupted or no running job found + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized - Authentication required + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Interrupt currently running jobs + tags: + - queue + /api/job/{job_id}/status: + get: + deprecated: true + description: | + **Deprecated.** Superseded by `GET /api/jobs/{job_id}` (plural path). + Clients should migrate; the endpoint is retained for backward + compatibility but will be removed in a future release. + operationId: getJobStatus + parameters: + - description: The unique ID of the job + in: path + name: job_id + required: true + schema: + format: uuid + type: string + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/JobStatusResponse' + description: Success - Job status returned + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "403": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Forbidden - job belongs to another user + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Job not found + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Get job status (deprecated) + tags: + - job + /api/jobs: + get: + description: | + Retrieve a paginated list of jobs for the authenticated user. + Returns lightweight job data optimized for list views. + Workflow and full outputs are excluded to reduce payload size. + operationId: listJobs + parameters: + - description: Filter by one or more statuses (comma-separated). If not provided, returns all jobs. + example: pending,in_progress + in: query + name: status + schema: + type: string + - description: Filter by workflow ID (exact match) + example: 550e8400-e29b-41d4-a716-446655440000 + in: query + name: workflow_id + schema: + type: string + - description: Filter by output media type (only applies to completed jobs with outputs) + example: image + in: query + name: output_type + schema: + enum: + - image + - video + - audio + - 3d + type: string + - description: Field to sort by (create_time = when job was submitted, execution_time = how long workflow took to run) + example: execution_time + in: query + name: sort_by + schema: + default: create_time + enum: + - create_time + - execution_time + type: string + - description: Sort direction (asc = ascending, desc = descending) + in: query + name: sort_order + schema: + default: desc + enum: + - asc + - desc + type: string + - description: | + Opaque cursor for keyset pagination. Pass the `next_cursor` value + from a previous response to fetch the next page. + Cursor pagination is supported only when `sort_by=create_time` + (default). If `sort_by=execution_time`, `after` is ignored and + offset/limit pagination is used. + Cursors are opaque base64url payloads — clients should treat them + as strings and not parse the contents. + example: eyJzIjoiY3JlYXRlX3RpbWUiLCJ2IjoiMTcxNjIwMDAwMDAwMDAwMCIsImlkIjoiYTFiMmMzZDQtZTVmNi03YTg5LWIwYzEtZDJlM2Y0YTViNmM3In0 + in: query + name: after + schema: + type: string + - deprecated: true + description: 'Pagination offset (0-based). Deprecated: prefer cursor-based pagination via `after`.' + in: query + name: offset + schema: + default: 0 + minimum: 0 + type: integer + - description: Maximum items per page (1-1000) + in: query + name: limit + schema: + default: 100 + maximum: 1000 + minimum: 1 + type: integer + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/JobsListResponse' + description: Success - Jobs retrieved + "400": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Bad request (e.g. malformed pagination cursor). + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized - Authentication required + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: List jobs with pagination and filtering + tags: + - workflow + /api/jobs/{job_id}: + get: + description: | + Retrieve complete details for a specific job including workflow and outputs. + Used for detail views, workflow re-execution, and debugging. + operationId: getJobDetail + parameters: + - description: Job identifier (UUID) + in: path + name: job_id + required: true + schema: + format: uuid + type: string + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/JobDetailResponse' + description: Success - Job details retrieved + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized - Authentication required + "403": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Forbidden - Job does not belong to user + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Job not found + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Get full job details + tags: + - workflow + /api/jobs/{job_id}/cancel: + post: + description: | + Cancel a specific job for the authenticated user. - BillingPlan: - type: object - x-runtime: [cloud] - description: "[cloud-only] A subscription plan with pricing details." - required: - - id - - name - properties: - id: - type: string - name: - type: string - description: - type: string - credits_per_month: - type: integer - price_cents: - type: integer - description: Monthly price in cents (USD) - currency: - type: string - default: usd - features: - type: array - items: - type: string - description: List of plan features + Idempotent: a job that is already in a terminal state (completed, failed, + cancelled) or already cancelling is treated as a successful no-op and + returns 200. Only truly missing or cross-user jobs return 404. + operationId: cancelJob + parameters: + - description: Job identifier (UUID) + in: path + name: job_id + required: true + schema: + format: uuid + type: string + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/JobCancelResponse' + description: Success - Cancel request accepted (or job was already terminal) + "400": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Bad Request - job_id is not a valid UUID (emitted by request validation before the handler runs) + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized - Authentication required + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Job not found for this user + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error - cancellation failed + summary: Cancel a job + tags: + - workflow + /api/node_replacements: + get: + description: | + Returns mappings of unsupported node class names to their cloud-installed replacements. + Used by the frontend to offer "Quick Fix" when a workflow contains missing nodes. + operationId: getNodeReplacements + responses: + "200": + content: + application/json: + schema: + additionalProperties: true + type: object + description: Success - Node replacement mappings + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + security: [] + summary: Get node replacement mappings + tags: + - node + /api/object_info: + get: + description: Returns information about all available nodes + operationId: getNodeInfo + responses: + "200": + content: + application/json: + schema: + additionalProperties: + $ref: '#/components/schemas/NodeInfo' + type: object + description: Success + summary: Get all node information + tags: + - node + /api/prompt: + get: + description: Returns information about the current prompt in the execution queue + operationId: getPromptInfo + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/PromptInfo' + description: Success + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Get information about current prompt execution + tags: + - workflow + post: + description: | + Submit a workflow to be executed by the backend. + The workflow is a JSON object describing the nodes and their connections. + operationId: executePrompt + requestBody: + content: + application/json: + schema: + $ref: '#/components/schemas/PromptRequest' + required: true + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/PromptResponse' + description: Success - Prompt accepted + "400": + content: + application/json: + schema: + $ref: '#/components/schemas/PromptErrorResponse' + description: Invalid prompt + "402": + content: + application/json: + schema: + $ref: '#/components/schemas/PromptErrorResponse' + description: Payment required - Insufficient credits + "429": + content: + application/json: + schema: + $ref: '#/components/schemas/PromptErrorResponse' + description: Payment required - User has not paid + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/PromptErrorResponse' + description: Internal server error + "503": + content: + application/json: + schema: + $ref: '#/components/schemas/PromptErrorResponse' + description: Service unavailable + summary: Submit a workflow for execution + tags: + - workflow + /api/queue: + get: + description: Returns information about running and pending items in the queue + operationId: getQueueInfo + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/QueueInfo' + description: Success + "400": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Invalid request parameters + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Invalid request parameters + summary: Get queue information + tags: + - queue + post: + description: | + Cancel specific PENDING jobs by ID or clear all pending jobs in the queue. + Note: This endpoint only affects pending jobs. To cancel running jobs, use /api/interrupt. + operationId: manageQueue + requestBody: + content: + application/json: + schema: + $ref: '#/components/schemas/QueueManageRequest' + required: true + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/QueueManageResponse' + description: Success + "400": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Invalid request parameters + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Manage queue operations + tags: + - queue + /api/settings: + get: + description: Returns all settings for the authenticated user + operationId: getAllSettings + responses: + "200": + content: + application/json: + schema: + additionalProperties: true + description: User settings as key-value pairs + type: object + description: Success + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + summary: Get all user settings + tags: + - settings + post: + description: Update multiple settings (merge with existing) + operationId: updateMultipleSettings + requestBody: + content: + application/json: + schema: + additionalProperties: true + description: Settings to update as key-value pairs + type: object + text/plain: + schema: + description: JSON string of settings to update + type: string + required: true + responses: + "200": + content: + application/json: + schema: + additionalProperties: true + description: Updated user settings + type: object + description: Success + "400": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Invalid request + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + summary: Update multiple settings + tags: + - settings + /api/settings/{id}: + get: + description: Returns a specific setting value by its id + operationId: getSettingById + parameters: + - description: Setting id to retrieve + in: path + name: id + required: true + schema: + type: string + responses: + "200": + content: + application/json: + schema: + description: Setting value response + properties: + value: + description: The setting value + type: object + description: Success + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Setting not found + summary: Get a specific setting by id + tags: + - settings + post: + description: Update a specific setting by its id + operationId: updateSettingById + parameters: + - description: Setting id to update + in: path + name: id + required: true + schema: + type: string + requestBody: + content: + application/json: + schema: + description: New value for the setting + text/plain: + schema: + description: JSON string of the new setting value + type: string + required: true + responses: + "200": + content: + application/json: + schema: + description: Updated setting value response + properties: + value: + description: The updated setting value + type: object + description: Success + "400": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Invalid request + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + summary: Update a specific setting by id + tags: + - settings + /api/system_stats: + get: + description: Returns system statistics including ComfyUI version, device info, and system resources + operationId: getSystemStats + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/SystemStatsResponse' + description: Success + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + security: [] + summary: Get system statistics + tags: + - system + /api/tags: + get: + description: | + Retrieves a list of all tags used across assets. + Includes usage counts and filtering options. + operationId: listTags + parameters: + - description: Filter tags by prefix + in: query + name: prefix + schema: + type: string + - description: Maximum number of tags to return (1-1000) + in: query + name: limit + schema: + default: 100 + maximum: 1000 + minimum: 1 + type: integer + - description: Number of tags to skip for pagination + in: query + name: offset + schema: + default: 0 + minimum: 0 + type: integer + - description: Sort order for tags + in: query + name: order + schema: + default: count_desc + enum: + - count_desc + - name_asc + type: string + - description: Include tags with zero usage count + in: query + name: include_zero + schema: + default: false + type: boolean + - description: Whether to include public/shared assets when counting tags + in: query + name: include_public + schema: + default: true + type: boolean + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/ListTagsResponse' + description: Tags retrieved successfully + "400": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Invalid request parameters + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: List all tags + tags: + - file + /api/tasks: + get: + description: | + Retrieve a paginated list of background tasks for the authenticated user. + Supports filtering by task type, status, and creation time. + operationId: listTasks + parameters: + - description: Filter by task type name (exact match) + example: model_upload + in: query + name: task_name + schema: + type: string + - description: Filter by idempotency key (exact match). For best performance, specify task_name as well. + example: upload-model-abc123 + in: query + name: idempotency_key + schema: + type: string + - description: Filter by one or more statuses (comma-separated) + example: created,running + in: query + name: status + schema: + type: string + - description: Filter tasks created after this timestamp (RFC3339 format) + example: "2024-01-01T00:00:00Z" + in: query + name: created_after + schema: + format: date-time + type: string + - description: Filter tasks created before this timestamp (RFC3339 format) + example: "2024-12-31T23:59:59Z" + in: query + name: created_before + schema: + format: date-time + type: string + - description: Sort direction (asc = ascending, desc = descending by create_time) + in: query + name: sort_order + schema: + default: desc + enum: + - asc + - desc + type: string + - description: Pagination offset (0-based) + in: query + name: offset + schema: + default: 0 + minimum: 0 + type: integer + - description: Maximum items per page (1-100) + in: query + name: limit + schema: + default: 20 + maximum: 100 + minimum: 1 + type: integer + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/TasksListResponse' + description: Success - Tasks retrieved + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized - Authentication required + "422": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Validation error - Invalid filter values + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: List background tasks + tags: + - task + /api/tasks/{task_id}: + get: + description: | + Retrieve full details for a specific background task. + operationId: getTask + parameters: + - description: Task identifier (UUID) + in: path + name: task_id + required: true + schema: + format: uuid + type: string + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/TaskResponse' + description: Success - Task details retrieved + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized - Authentication required + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Task not found (also returned for ownership failures to avoid leaking task existence) + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Get task details + tags: + - task + /api/upload/image: + post: + description: | + Upload an image file to cloud storage. - BillingStatus: - type: object - x-runtime: [cloud] - description: "[cloud-only] Overall billing and subscription status." - properties: - subscription: - $ref: "#/components/schemas/BillingSubscription" - balance: - $ref: "#/components/schemas/BillingBalance" - has_payment_method: - type: boolean + Image limits: + - Maximum file size: 50 MB + - Maximum width/height per edge: 16384 px + - Maximum total pixel count: 64 megapixels (67108864 pixels) - BillingSubscription: - type: object - x-runtime: [cloud] - description: "[cloud-only] Active subscription details." - required: - - id - - status - - plan_id - properties: - id: - type: string - status: - type: string - enum: [active, cancelled, past_due, trialing] - plan_id: - type: string - plan_name: - type: string - current_period_start: - type: string - format: date-time - current_period_end: - type: string - format: date-time - cancel_at_period_end: - type: boolean + Uploads that exceed any of these limits are rejected with HTTP 400. + operationId: uploadImage + requestBody: + content: + multipart/form-data: + schema: + properties: + image: + description: The image file to upload + format: binary + type: string + overwrite: + description: Whether to overwrite existing file (true/false) + type: string + subfolder: + description: Optional subfolder path + type: string + type: + description: Upload type (defaults to "output") + type: string + required: + - image + type: object + required: true + responses: + "200": + content: + application/json: + schema: + properties: + name: + description: Filename of the uploaded image + type: string + subfolder: + description: Subfolder path where image was saved + type: string + type: + description: Type of upload (e.g., "output") + type: string + type: object + description: Image uploaded successfully + "400": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Bad request + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Upload an image file + tags: + - file + /api/upload/mask: + post: + description: | + Upload a mask image to be applied to an existing image. - SubscriptionPreview: - type: object - x-runtime: [cloud] - description: "[cloud-only] Preview of a subscription change including prorations." - properties: - plan_id: - type: string - plan_name: - type: string - amount_due: - type: integer - description: Amount due in cents - proration_amount: - type: integer - description: Proration adjustment in cents - currency: - type: string - next_billing_date: - type: string - format: date-time + Image limits apply to both the uploaded mask and the referenced + original image: + - Maximum file size: 50 MB + - Maximum width/height per edge: 16384 px + - Maximum total pixel count: 64 megapixels (67108864 pixels) - Workspace: - type: object - x-runtime: [cloud] - description: "[cloud-only] A cloud workspace for team collaboration." - required: - - id - - name - properties: - id: - type: string - name: - type: string - owner_id: - type: string - member_count: - type: integer - created_at: - type: string - format: date-time - updated_at: - type: string - format: date-time - - WorkspaceMember: - type: object - x-runtime: [cloud] - description: "[cloud-only] A member of a cloud workspace." - required: - - user_id - - role - properties: - user_id: - type: string - email: - type: string - format: email - display_name: - type: string - avatar_url: - type: string - format: uri - role: - type: string - enum: [owner, admin, member] - joined_at: - type: string - format: date-time - - WorkspaceInvite: - type: object - x-runtime: [cloud] - description: "[cloud-only] A pending workspace invitation." - required: - - id - - email - - role - properties: - id: - type: string - email: - type: string - format: email - role: - type: string - enum: [admin, member] - invited_by: - type: string - created_at: - type: string - format: date-time - expires_at: - type: string - format: date-time - - WorkspaceApiKey: - type: object - x-runtime: [cloud] - description: "[cloud-only] A workspace API key (secret value redacted)." - required: - - id - - name - - description - properties: - id: - type: string - name: - type: string - description: - type: string - maxLength: 5000 - description: User-provided description of the key's purpose. Always present in responses; empty string when no description was supplied on create. - prefix: - type: string - description: First few characters of the key for identification - created_at: - type: string - format: date-time - last_used_at: - type: string - format: date-time - nullable: true - created_by: - type: string - - WorkspaceApiKeyCreated: - type: object - x-runtime: [cloud] - description: "[cloud-only] A newly created workspace API key, including the full secret value (shown only once)." - required: - - id - - name - - description - - key - properties: - id: - type: string - name: - type: string - description: - type: string - maxLength: 5000 - description: User-provided description of the key's purpose. Always present in responses; empty string when no description was supplied on create. - key: - type: string - description: Full API key value (only returned on creation) - prefix: - type: string - created_at: - type: string - format: date-time - - CloudUser: - type: object - x-runtime: [cloud] - description: "[cloud-only] A cloud-authenticated user profile." - required: - - id - - email - properties: - id: - type: string - email: - type: string - format: email - display_name: - type: string - avatar_url: - type: string - format: uri - created_at: - type: string - format: date-time - - SecretMeta: - type: object - x-runtime: [cloud] - description: "[cloud-only] Metadata for a stored secret (value is never returned)." - required: - - id - - name - properties: - id: - type: string - name: - type: string - provider: - type: string - description: "[cloud-only] Provider identifier (e.g., huggingface, civitai)." - x-runtime: [cloud] - last_used_at: - type: string - format: date-time - description: "[cloud-only] When the secret was last used for decryption." - x-runtime: [cloud] - created_at: - type: string - format: date-time - updated_at: - type: string - format: date-time - - UpdateSecretRequest: - type: object - x-runtime: [cloud] - description: "[cloud-only] Request body for updating an existing user secret." - properties: - name: - type: string - description: New name for the secret - secret_value: - type: string - description: New secret value (API key, token, etc.) - - CreateSessionResponse: - type: object - x-runtime: [cloud] - description: "[cloud-only] Response after creating a session cookie." - required: - - success - properties: - success: - type: boolean - expiresIn: - type: integer - description: Session expiration time in seconds. - - DeleteSessionResponse: - type: object - x-runtime: [cloud] - description: "[cloud-only] Response after deleting a session cookie." - required: - - success - properties: - success: - type: boolean - - CreateHubProfileRequest: - type: object - x-runtime: [cloud] - description: "[cloud-only] Request body for creating a new Hub profile." - required: - - workspace_id - - username - properties: - workspace_id: - type: string - username: - type: string - description: Unique URL-safe slug. Immutable after creation. - display_name: - type: string - description: - type: string - avatar_token: - type: string - website_urls: - type: array - items: - type: string - - PublishHubWorkflowRequest: - type: object - x-runtime: [cloud] - description: "[cloud-only] Request body for publishing or updating a workflow on the Hub." - required: - - username - - name - - workflow_filename - - asset_ids - properties: - username: - type: string - name: - type: string - workflow_filename: - type: string - asset_ids: - type: array - items: - type: string - description: - type: string - tags: - type: array - items: - type: string - models: - type: array - items: - type: string - custom_nodes: - type: array - items: - type: string - tutorial_url: - type: string - metadata: - type: object - additionalProperties: true - thumbnail_type: - type: string - enum: [image, video, image_comparison] - thumbnail_token_or_url: - type: string - thumbnail_comparison_token_or_url: - type: string - sample_image_tokens_or_urls: - type: array - items: - type: string - - HubWorkflowDetail: - type: object - x-runtime: [cloud] - description: "[cloud-only] Full Hub workflow detail including versions, assets, and statistics." - required: - - share_id - - workflow_id - - name - - workflow_json - - assets - - profile - - status - properties: - share_id: - type: string - workflow_id: - type: string - name: - type: string - status: - type: string - enum: [pending, approved, rejected, deprecated] - description: - type: string - thumbnail_type: - type: string - enum: [image, video, image_comparison] - thumbnail_url: - type: string - thumbnail_comparison_url: - type: string - tutorial_url: - type: string - metadata: - type: object - additionalProperties: true - sample_image_urls: - type: array - items: - type: string - publish_time: - type: string - format: date-time - nullable: true - workflow_json: - type: object - additionalProperties: true - assets: - type: array - items: - $ref: "#/components/schemas/AssetInfo" - profile: - $ref: "#/components/schemas/HubProfile" - - AssetInfo: - type: object - x-runtime: [cloud] - description: "[cloud-only] Lightweight asset reference used in workflow publishing payloads." - required: - - id - - filename - properties: - id: - type: string - filename: - type: string - mime_type: - type: string - size_bytes: - type: integer - format: int64 - - BulkRevokeAPIKeysResponse: - type: object - x-runtime: [cloud] - description: "[cloud-only] Response after bulk-revoking API keys for a workspace member." - required: - - revoked_count - properties: - revoked_count: - type: integer - minimum: 0 - - CreateWorkflowVersionRequest: - type: object - x-runtime: [cloud] - description: "[cloud-only] Request body for creating a new version of a saved workflow." - required: - - base_version - - workflow_json - properties: - base_version: - type: integer - description: Version number this change is based on (for optimistic concurrency). - workflow_json: - type: object - additionalProperties: true - - WorkflowVersionResponse: - type: object - x-runtime: [cloud] - description: "[cloud-only] Metadata for a single workflow version." - required: - - id - - version - - latest_version - - created_by - - created_at - properties: - id: - type: string - version: - type: integer - latest_version: - type: integer - created_by: - type: string - created_at: - type: string - format: date-time - - WorkflowPublishInfo: - type: object - x-runtime: [cloud] - description: "[cloud-only] Publishing metadata for a workflow shared to the Hub." - required: - - workflow_id - - share_id - - listed - - assets - properties: - workflow_id: - type: string - share_id: - type: string - publish_time: - type: string - format: date-time - nullable: true - listed: - type: boolean - assets: - type: array - items: - $ref: "#/components/schemas/AssetInfo" - - TaskEntry: - type: object - x-runtime: [cloud] - description: "[cloud-only] Task data for list views." - required: - - id - - task_name - - status - - create_time - properties: - id: - type: string - format: uuid - task_name: - type: string - status: - type: string - enum: [created, running, completed, failed] - create_time: - type: string - format: date-time - started_at: - type: string - format: date-time - completed_at: - type: string - format: date-time - - TaskResponse: - type: object - x-runtime: [cloud] - description: "[cloud-only] Full task details including payload and result." - required: - - id - - idempotency_key - - task_name - - payload - - status - - create_time - - update_time - properties: - id: - type: string - format: uuid - idempotency_key: - type: string - task_name: - type: string - payload: - type: object - additionalProperties: true - status: - type: string - enum: [created, running, completed, failed] - result: - type: object - additionalProperties: true - create_time: - type: string - format: date-time - update_time: - type: string - format: date-time - started_at: - type: string - format: date-time - completed_at: - type: string - format: date-time - error: - type: string - - TasksListResponse: - type: object - x-runtime: [cloud] - description: "[cloud-only] Paginated list of background tasks for the authenticated user." - required: - - tasks - - pagination - properties: - tasks: - type: array - items: - $ref: "#/components/schemas/TaskEntry" - pagination: - $ref: "#/components/schemas/PaginationInfo" \ No newline at end of file + Uploads that exceed any of these limits are rejected with HTTP 400. + operationId: uploadMask + requestBody: + content: + multipart/form-data: + schema: + properties: + image: + description: The mask image file to upload + format: binary + type: string + original_ref: + description: JSON string containing reference to the original image + type: string + required: + - image + - original_ref + type: object + required: true + responses: + "200": + content: + application/json: + schema: + properties: + name: + description: Filename of the uploaded mask + type: string + subfolder: + description: Subfolder path where mask was saved + type: string + type: + description: Type of upload (e.g., "output") + type: string + type: object + description: Mask uploaded successfully + "400": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Bad request + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Upload a mask image + tags: + - file + /api/user: + get: + description: Returns information about the currently authenticated user + operationId: getUser + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/UserResponse' + description: Success + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + summary: Get current user information + tags: + - user + /api/userdata: + get: + description: Returns a list of user data files in the specified directory, optionally recursively and with full metadata. + operationId: getUserdata + parameters: + - description: The directory to list files from. + in: query + name: dir + schema: + type: string + - description: Whether to list files recursively. + in: query + name: recurse + schema: + default: false + type: boolean + - description: Whether to split file information by type. + in: query + name: split + schema: + default: false + type: boolean + - description: Whether to return full file metadata. + in: query + name: full_info + schema: + default: false + type: boolean + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/GetUserDataResponseFull' + description: A list of user data files. + "400": + content: + text/plain: + schema: + type: string + description: Bad request (e.g., invalid filename). + "401": + content: + text/plain: + schema: + type: string + description: Unauthorized. + "404": + content: + text/plain: + schema: + type: string + description: File not found or invalid path. + "500": + content: + text/plain: + schema: + type: string + description: General error + summary: List user data files + tags: + - user + /api/userdata/{file}: + delete: + description: | + Delete a user data file from the database. The file parameter should be + the relative path within the user's data directory. + operationId: deleteUserdataFile + parameters: + - description: The file path to delete (URL encoded if necessary). + in: path + name: file + required: true + schema: + type: string + responses: + "204": + description: File deleted successfully (No Content). + "401": + content: + text/plain: + schema: + type: string + description: Unauthorized. + "404": + content: + text/plain: + schema: + type: string + description: File not found. + "500": + content: + text/plain: + schema: + type: string + description: Internal server error. + summary: Delete a user data file + tags: + - user + get: + description: Returns the requested user data file if it exists. + operationId: getUserdataFile + parameters: + - description: The filename of the user data to retrieve. + in: path + name: file + required: true + schema: + type: string + responses: + "200": + content: + application/octet-stream: + schema: + format: binary + type: string + description: Successfully retrieved the file. + "400": + content: + text/plain: + schema: + type: string + description: Bad request (e.g., invalid filename). + "401": + content: + text/plain: + schema: + type: string + description: Unauthorized. + "404": + content: + text/plain: + schema: + type: string + description: File not found or invalid path. + "500": + content: + text/plain: + schema: + type: string + description: General error + summary: Get user data file + tags: + - user + post: + description: | + Upload a file to a user's data directory. Optional query parameters allow + control over overwrite behavior and response detail. + operationId: postUserdataFile + parameters: + - description: The target file path (URL encoded if necessary). + in: path + name: file + required: true + schema: + type: string + - description: If "false", prevents overwriting existing files. Defaults to "true". + in: query + name: overwrite + schema: + default: "true" + enum: + - "true" + - "false" + type: string + - description: If "true", returns detailed file info; if "false", returns only the relative path. + in: query + name: full_info + schema: + default: "false" + enum: + - "true" + - "false" + type: string + requestBody: + content: + application/octet-stream: + schema: + format: binary + type: string + required: true + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/UserDataResponseFull' + description: File uploaded successfully. + "400": + content: + text/plain: + schema: + type: string + description: Missing or invalid 'file' parameter. + "401": + content: + text/plain: + schema: + type: string + description: Unauthorized. + "403": + content: + text/plain: + schema: + type: string + description: The requested path is not allowed. + "409": + content: + text/plain: + schema: + type: string + description: File already exists and overwrite is set to false. + "500": + content: + text/plain: + schema: + type: string + description: General error + summary: Upload or update a user data file + tags: + - user + /api/userdata/{file}/move/{dest}: + post: + description: | + Move or rename a file within a user's data directory, with options for + controlling overwrite behavior and response format. + operationId: moveUserdataFile + parameters: + - description: The source file path (URL encoded if necessary). + in: path + name: file + required: true + schema: + type: string + - description: The destination file path (URL encoded if necessary). + in: path + name: dest + required: true + schema: + type: string + - description: If "false", prevents overwriting existing files. Defaults to "true". + in: query + name: overwrite + schema: + default: "true" + enum: + - "true" + - "false" + type: string + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/UserDataResponseFull' + description: File moved successfully. + "400": + content: + text/plain: + schema: + type: string + description: Missing or invalid parameters. + "401": + content: + text/plain: + schema: + type: string + description: Unauthorized. + "404": + content: + text/plain: + schema: + type: string + description: Source file not found. + "409": + content: + text/plain: + schema: + type: string + description: Destination file already exists and overwrite is set to false. + "500": + content: + text/plain: + schema: + type: string + description: General error + summary: Move or rename a user data file + tags: + - user + /api/userdata/{file}/publish: + get: + description: Returns the publish status and share info for a workflow identified by its userdata path. + operationId: getUserdataFilePublish + parameters: + - description: The workflow file path within the user's data directory (URL encoded if necessary). + in: path + name: file + required: true + schema: + type: string + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/WorkflowPublishInfo' + description: Publish info (publish_time is null if never published) + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Workflow not found + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Get publish info for a workflow file + tags: + - workflows + post: + description: Creates a new published_workflow record from the latest version and snapshots the provided assets. + operationId: postUserdataFilePublish + parameters: + - description: The workflow file path within the user's data directory (URL encoded if necessary). + in: path + name: file + required: true + schema: + type: string + requestBody: + content: + application/json: + schema: + $ref: '#/components/schemas/PublishWorkflowAssetsRequest' + required: true + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/WorkflowPublishInfo' + description: Workflow published + "400": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Bad request + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Workflow not found + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Publish a workflow file + tags: + - workflows + /api/users: + get: + description: | + ComfyUI legacy users endpoint. Returns information about how user + data is stored. In cloud this is always server-managed, so callers + receive a constant response indicating server-side storage. + operationId: getUsersInfo + responses: + "200": + content: + application/json: + schema: + properties: + migrated: + description: Whether user data has been migrated (always true in cloud) + type: boolean + storage: + description: Where user data is stored (always "server" in cloud) + type: string + required: + - storage + - migrated + type: object + description: Userdata storage information + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + summary: ComfyUI userdata storage info + tags: + - user + /api/vhs/queryvideo: + get: + description: | + VHS custom node endpoint that returns metadata about a video file + (frame count, fps, resolution, duration). Currently returns default + placeholder values; real ffprobe integration is a follow-up. + operationId: getVhsQueryVideo + parameters: + - description: Name of the video file to query + in: query + name: filename + required: true + schema: + type: string + responses: + "200": + content: + application/json: + schema: + properties: + source: + description: Source video metadata + properties: + duration: + description: Duration in seconds + type: number + fps: + description: Frames per second + type: number + frames: + description: Total frame count + type: integer + size: + description: '[width, height] in pixels' + items: + type: integer + maxItems: 2 + minItems: 2 + type: array + required: + - size + - fps + - frames + - duration + type: object + required: + - source + type: object + description: Video metadata + "400": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: | + Missing required query parameter. Produced by the oapi-codegen + wrapper via echo.NewHTTPError; the custom Echo HTTPErrorHandler + normalizes it to the standard ErrorResponse {code, message} shape + (BE-1178). + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + security: + - ApiKeyAuth: [] + - BearerAuth: [] + - CookieAuth: [] + summary: Query VHS video metadata + tags: + - file + /api/view: + get: + description: | + Retrieve and view a file from the ComfyUI file system. + This endpoint is typically used to view generated images or other output files. + Cookie auth is allowed on this endpoint because it's used by img/video tags in browsers. + operationId: viewFile + parameters: + - description: Name of the file to view + in: query + name: filename + required: true + schema: + example: ComfyUI_00004_.png + type: string + - description: Subfolder path where the file is located + in: query + name: subfolder + schema: + example: tests/foo/bar + type: string + - description: Type of file (e.g., output, input, temp) + in: query + name: type + schema: + example: output + type: string + - description: Full path to the file (used for temp files) + in: query + name: fullpath + schema: + type: string + - description: Format of the file + in: query + name: format + schema: + type: string + - description: Frame rate for video files + in: query + name: frame_rate + schema: + type: integer + - description: Workflow identifier + in: query + name: workflow + schema: + type: string + - description: Timestamp parameter + in: query + name: timestamp + schema: + example: 1234567890 + type: integer + - description: | + Image channel to extract from PNG images. + - 'rgb': Return only RGB channels (alpha set to fully opaque) + - 'a' or 'alpha': Return alpha channel as grayscale image + - If not specified, return original image unchanged via redirect + in: query + name: channel + schema: + example: rgb + type: string + - description: | + Maximum dimension (width or height) to resize the image to, preserving aspect ratio. + The image is fit within a res x res box. Returns a JPEG thumbnail. + Only applies to raster image files (PNG, JPEG, WebP, GIF). + in: query + name: res + schema: + example: 256 + maximum: 1024 + minimum: 64 + type: integer + responses: + "200": + content: + image/jpeg: + schema: + description: Resized JPEG thumbnail (returned when res parameter is used) + format: binary + type: string + image/png: + schema: + description: Processed PNG image with extracted channel + format: binary + type: string + description: Success - File content returned (used when channel or res parameter is present) + "302": + description: Redirect to GCS signed URL + headers: + Cache-Control: + description: Cache directive for the redirect response + schema: + type: string + Location: + description: Signed URL to access the file in GCS + schema: + type: string + Vary: + description: Headers that affect response caching + schema: + type: string + "400": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Invalid request parameters + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: File not found or unauthorized + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + security: + - ApiKeyAuth: [] + - BearerAuth: [] + - CookieAuth: [] + summary: View a file + tags: + - file + /api/workflow_templates: + get: + description: Returns available workflow templates + operationId: getWorkflowTemplates + responses: + "200": + content: + application/json: + schema: + description: Empty object for workflow templates + type: object + description: Success + security: [] + summary: Get available workflow templates + tags: + - workflow + /api/workflows: + get: + description: Returns a paginated list of workflows for the authenticated user in the current workspace. + operationId: listWorkflows + parameters: + - in: query + name: limit + schema: + default: 20 + maximum: 100 + type: integer + - in: query + name: offset + schema: + default: 0 + type: integer + - description: Search workflows by name (case-insensitive substring match) + in: query + name: name + schema: + type: string + - description: Filter by default view type + in: query + name: default_view + schema: + enum: + - workflow + - app + type: string + - description: Sort field + in: query + name: sort + schema: + default: create_time + enum: + - create_time + - update_time + - name + type: string + - description: Sort order + in: query + name: order + schema: + default: desc + enum: + - asc + - desc + type: string + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/WorkflowListResponse' + description: Success + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: List workflows + tags: + - workflows + post: + description: Creates a new workflow with its first version. + operationId: createWorkflow + requestBody: + content: + application/json: + schema: + $ref: '#/components/schemas/CreateWorkflowRequest' + required: true + responses: + "201": + content: + application/json: + schema: + $ref: '#/components/schemas/WorkflowResponse' + description: Workflow created successfully + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "422": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Validation error + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Create a new workflow + tags: + - workflows + /api/workflows/{workflow_id}: + delete: + description: Soft-deletes a workflow. + operationId: deleteWorkflow + parameters: + - description: The UUID of the workflow to delete. + in: path + name: workflow_id + required: true + schema: + type: string + responses: + "204": + description: Workflow deleted successfully + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Workflow not found + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Delete workflow + tags: + - workflows + get: + description: Retrieves workflow metadata by ID. + operationId: getWorkflow + parameters: + - description: The UUID of the workflow. + in: path + name: workflow_id + required: true + schema: + type: string + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/WorkflowResponse' + description: Success + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "403": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Forbidden + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Workflow not found + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Get workflow + tags: + - workflows + patch: + description: Updates mutable workflow metadata (name, description, default_view). + operationId: updateWorkflow + parameters: + - description: The UUID of the workflow to update. + in: path + name: workflow_id + required: true + schema: + type: string + requestBody: + content: + application/json: + schema: + $ref: '#/components/schemas/UpdateWorkflowRequest' + required: true + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/WorkflowResponse' + description: Success + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Workflow not found + "422": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Validation error + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Update workflow metadata + tags: + - workflows + /api/workflows/{workflow_id}/content: + get: + description: Retrieves the latest version of a workflow and its JSON content. + operationId: getWorkflowContent + parameters: + - description: The UUID of the workflow whose content should be retrieved. + in: path + name: workflow_id + required: true + schema: + type: string + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/WorkflowVersionContentResponse' + description: Success + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "403": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Forbidden + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Workflow not found + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Get workflow content + tags: + - workflows + /api/workflows/{workflow_id}/fork: + post: + description: Creates a new workflow by forking from an existing version. + operationId: forkWorkflow + parameters: + - description: The UUID of the source workflow to fork from. + in: path + name: workflow_id + required: true + schema: + type: string + requestBody: + content: + application/json: + schema: + $ref: '#/components/schemas/ForkWorkflowRequest' + required: true + responses: + "201": + content: + application/json: + schema: + $ref: '#/components/schemas/WorkflowResponse' + description: Workflow forked successfully + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "403": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Forbidden + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Source workflow or version not found + "422": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Validation error + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Fork a workflow + tags: + - workflows + /api/workflows/{workflow_id}/versions: + post: + description: Creates a new workflow version with updated workflow JSON. Uses optimistic concurrency via base_version. + operationId: createWorkflowVersion + parameters: + - description: The UUID of the workflow to create a new version for. + in: path + name: workflow_id + required: true + schema: + type: string + requestBody: + content: + application/json: + schema: + $ref: '#/components/schemas/CreateWorkflowVersionRequest' + required: true + responses: + "201": + content: + application/json: + schema: + $ref: '#/components/schemas/WorkflowVersionResponse' + description: Version created successfully + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "403": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Forbidden - not the workflow owner + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Workflow not found + "409": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Version conflict - base_version does not match latest + "422": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Validation error + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Create a new version + tags: + - workflows + /api/workflows/published/{share_id}: + get: + description: | + Returns the published workflow details including the status of each + published asset relative to the caller's library. Authentication is required. + operationId: getPublishedWorkflow + parameters: + - description: The share ID of the published workflow. + in: path + name: share_id + required: true + schema: + type: string + responses: + "200": + content: + application/json: + schema: + $ref: '#/components/schemas/PublishedWorkflowDetail' + description: Published workflow details with asset statuses + "401": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Unauthorized + "404": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Share not found + "413": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Workflow JSON too large + "500": + content: + application/json: + schema: + $ref: '#/components/schemas/ErrorResponse' + description: Internal server error + summary: Get a published workflow by share ID + tags: + - workflows + /health: + get: + description: | + Returns `200 OK` if the database is reachable and dynamic config has + loaded, otherwise `503 Service Unavailable`. Used by the GKE ingress + for health checks. Response body is plain text for probe simplicity. + operationId: getHealth + responses: + "200": + content: + text/plain: + schema: + example: OK + type: string + description: Service is healthy + "503": + content: + text/plain: + schema: + example: Service Unavailable + type: string + description: Service is unhealthy + security: [] + summary: Health probe for Kubernetes readiness/liveness + tags: + - system + /internal/folder_paths: + get: + description: Returns the filesystem paths ComfyUI loads models and assets from, keyed by folder type. + operationId: getInternalFolderPaths + responses: + "200": + content: + application/json: + schema: + additionalProperties: + items: + items: + type: string + type: array + type: array + description: Map of folder type name to list of path entries + type: object + description: Dictionary of folder type to paths + summary: Get configured folder paths + /internal/logs: + get: + description: Returns ComfyUI log entries from the in-memory log buffer. + operationId: getInternalLogs + responses: + "200": + content: + text/plain: + schema: + type: string + description: Log text + summary: Get server logs as text + /internal/logs/raw: + get: + description: Returns the raw ComfyUI log buffer plus size metadata. + operationId: getInternalLogsRaw + responses: + "200": + content: + application/json: + schema: + properties: + entries: + items: + properties: + m: + description: Message + type: string + t: + description: Timestamp + type: number + type: object + type: array + size: + properties: + cols: + type: integer + rows: + type: integer + type: object + type: object + description: Structured log data + summary: Get raw structured log entries + /internal/logs/subscribe: + patch: + description: Subscribes or unsubscribes the current client from live log streaming over the WebSocket. + operationId: subscribeToLogs + requestBody: + content: + application/json: + schema: + properties: + clientId: + description: WebSocket client ID + type: string + enabled: + description: Enable or disable log streaming for this client + type: boolean + required: + - clientId + - enabled + type: object + required: true + responses: + "200": + description: Subscription updated + summary: Subscribe or unsubscribe a WebSocket client to log streaming +security: + - ApiKeyAuth: [] + - BearerAuth: [] +servers: + - description: Default ComfyUI server + url: / +tags: + - description: Workflow execution and management + name: workflow + - description: Node information + name: node + - description: File operations + name: file + - description: User settings management + name: settings + - description: User feedback management + name: feedback + - description: System operations and monitoring + name: system + - description: User information and management + name: user + - description: Background task management + name: task + - description: Workflow storage and version management + name: workflows + - description: Job queue state and control + name: queue + - description: Job lifecycle queries + name: job diff --git a/pyproject.toml b/pyproject.toml index 1e449b4a3..4107b4911 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.22.0" +version = "0.24.0" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.10" diff --git a/requirements.txt b/requirements.txt index d2986eda8..a49d968af 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,6 +1,6 @@ -comfyui-frontend-package==1.43.18 -comfyui-workflow-templates==0.9.79 -comfyui-embedded-docs==0.5.0 +comfyui-frontend-package==1.45.15 +comfyui-workflow-templates==0.9.98 +comfyui-embedded-docs==0.5.3 torch torchsde torchvision @@ -21,9 +21,9 @@ psutil alembic SQLAlchemy>=2.0.0 filelock -av>=14.2.0 -comfy-kitchen>=0.2.8 -comfy-aimdo==0.4.3 +av>=16.0.0 +comfy-kitchen==0.2.10 +comfy-aimdo==0.4.9 requests simpleeval>=1.0.0 blake3 diff --git a/server.py b/server.py index 44470b904..a85c1e591 100644 --- a/server.py +++ b/server.py @@ -646,18 +646,37 @@ class PromptServer(): @routes.get("/system_stats") async def system_stats(request): - device = comfy.model_management.get_torch_device() - device_name = comfy.model_management.get_torch_device_name(device) + primary_device = comfy.model_management.get_torch_device() cpu_device = comfy.model_management.torch.device("cpu") ram_total = comfy.model_management.get_total_memory(cpu_device) ram_free = comfy.model_management.get_free_memory(cpu_device) - vram_total, torch_vram_total = comfy.model_management.get_total_memory(device, torch_total_too=True) - vram_free, torch_vram_free = comfy.model_management.get_free_memory(device, torch_free_too=True) required_frontend_version = FrontendManager.get_required_frontend_version() installed_templates_version = FrontendManager.get_installed_templates_version() required_templates_version = FrontendManager.get_required_templates_version() comfy_package_versions = FrontendManager.get_comfy_package_versions() + # Report every torch device visible to multigpu, with the primary + # device first so existing clients that read devices[0] keep working. + torch_devices = comfy.model_management.get_all_torch_devices() + if primary_device in torch_devices: + torch_devices = [primary_device] + [d for d in torch_devices if d != primary_device] + else: + torch_devices = [primary_device] + list(torch_devices) + + device_entries = [] + for d in torch_devices: + vram_total, torch_vram_total = comfy.model_management.get_total_memory(d, torch_total_too=True) + vram_free, torch_vram_free = comfy.model_management.get_free_memory(d, torch_free_too=True) + device_entries.append({ + "name": comfy.model_management.get_torch_device_name(d), + "type": d.type, + "index": d.index, + "vram_total": vram_total, + "vram_free": vram_free, + "torch_vram_total": torch_vram_total, + "torch_vram_free": torch_vram_free, + }) + system_stats = { "system": { "os": sys.platform, @@ -673,17 +692,7 @@ class PromptServer(): "embedded_python": os.path.split(os.path.split(sys.executable)[0])[1] == "python_embeded", "argv": sys.argv }, - "devices": [ - { - "name": device_name, - "type": device.type, - "index": device.index, - "vram_total": vram_total, - "vram_free": vram_free, - "torch_vram_total": torch_vram_total, - "torch_vram_free": torch_vram_free, - } - ] + "devices": device_entries } return web.json_response(system_stats) @@ -1244,6 +1253,15 @@ class PromptServer(): if verbose: logging.info("Starting server\n") + if args.debug_hang: + logging.info( + f"{'-' * 80}\n" + "ComfyUI has been started in debug-hang mode. Run your workflow as normal up to\n" + "the point of the hang or freeze, then use ctrl-C in the cmd or controlling\n" + "terminal to dump the python backtraces for debugging. Please attach the extra\n" + "debug info to your bug report.\n" + f"{'-' * 80}" + ) for addr in addresses: address = addr[0] port = addr[1] diff --git a/tests-unit/assets_test/conftest.py b/tests-unit/assets_test/conftest.py index 6c5c56113..4aa20372f 100644 --- a/tests-unit/assets_test/conftest.py +++ b/tests-unit/assets_test/conftest.py @@ -6,6 +6,7 @@ import subprocess import sys import tempfile import time +import uuid from pathlib import Path from typing import Callable, Iterator, Optional @@ -188,9 +189,17 @@ def _post_multipart_asset( @pytest.fixture def make_asset_bytes() -> Callable[[str, int], bytes]: + # Salt content per test so it never collides with assets left over from + # earlier tests. Delete is now always a soft delete (content is preserved), + # so the suite can no longer rely on hard-deleting content for isolation. + # Deterministic within a test: the same (name, size) yields the same bytes. + salt = uuid.uuid4().bytes + def _make(name: str, size: int = 8192) -> bytes: seed = sum(ord(c) for c in name) % 251 - return bytes((i * 31 + seed) % 256 for i in range(size)) + body = bytearray((i * 31 + seed) % 256 for i in range(size)) + body[: len(salt)] = salt[:size] + return bytes(body) return _make @@ -212,7 +221,7 @@ def asset_factory(http: requests.Session, api_base: str): for aid in created: with contextlib.suppress(Exception): - http.delete(f"{api_base}/api/assets/{aid}?delete_content=true", timeout=30) + http.delete(f"{api_base}/api/assets/{aid}", timeout=30) @pytest.fixture @@ -227,7 +236,11 @@ def seeded_asset(request: pytest.FixtureRequest, http: requests.Session, api_bas if tags is None: tags = ["models", "checkpoints", "unit-tests", "alpha"] meta = {"purpose": "test", "epoch": 1, "flags": ["x", "y"], "nullable": None} - files = {"file": (name, b"A" * 4096, "application/octet-stream")} + # Unique content per test so the seed always creates a fresh asset (201). + # Delete is now always a soft delete, so content from a prior test survives + # and would otherwise dedup this upload into an existing asset (200). + content = uuid.uuid4().bytes + b"A" * (4096 - 16) + files = {"file": (name, content, "application/octet-stream")} form_data = { "tags": json.dumps(tags), "name": name, @@ -236,6 +249,8 @@ def seeded_asset(request: pytest.FixtureRequest, http: requests.Session, api_bas r = http.post(api_base + "/api/assets", files=files, data=form_data, timeout=120) body = r.json() assert r.status_code == 201, body + from helpers import assert_hash_fields_consistent + assert_hash_fields_consistent(body) return body @@ -258,4 +273,4 @@ def autoclean_unit_test_assets(http: requests.Session, api_base: str): break for aid in ids: with contextlib.suppress(Exception): - http.delete(f"{api_base}/api/assets/{aid}?delete_content=true", timeout=30) + http.delete(f"{api_base}/api/assets/{aid}", timeout=30) diff --git a/tests-unit/assets_test/helpers.py b/tests-unit/assets_test/helpers.py index 770e011f4..ae3de6dc3 100644 --- a/tests-unit/assets_test/helpers.py +++ b/tests-unit/assets_test/helpers.py @@ -26,3 +26,26 @@ def trigger_sync_seed_assets(session: requests.Session, base_url: str) -> None: def get_asset_filename(asset_hash: str, extension: str) -> str: return asset_hash.removeprefix("blake3:") + extension + + +def assert_hash_fields_consistent(body: dict, expected_hash: str | None = None) -> None: + """Assert hash and asset_hash invariants on an Asset response. + + Both must be present or both absent (so a regression that drops only one + is caught). When present, they must equal each other and, if expected_hash + is provided, must equal that value. + """ + hash_present = "hash" in body + asset_hash_present = "asset_hash" in body + assert hash_present == asset_hash_present, ( + f"hash and asset_hash must both be present or both absent: " + f"hash present={hash_present}, asset_hash present={asset_hash_present}" + ) + if hash_present: + h = body["hash"] + ah = body["asset_hash"] + assert h == ah, f"hash and asset_hash must match: hash={h!r}, asset_hash={ah!r}" + if expected_hash is not None: + assert h == expected_hash, ( + f"hash must equal expected: got {h!r}, expected {expected_hash!r}" + ) diff --git a/tests-unit/assets_test/queries/test_asset_reference_keyset.py b/tests-unit/assets_test/queries/test_asset_reference_keyset.py new file mode 100644 index 000000000..d143d60f9 --- /dev/null +++ b/tests-unit/assets_test/queries/test_asset_reference_keyset.py @@ -0,0 +1,112 @@ +"""Keyset-pagination tiebreaker tests for list_references_page. + +When multiple rows share the same primary sort value (e.g. four assets +created in the same microsecond), the secondary `ORDER BY id` is what keeps +keyset pagination from losing or repeating rows. This file exercises that +branch directly against an in-memory SQLite session — engineering identical +timestamps via HTTP is unreliable enough that we work at the query layer. +""" +import uuid +from datetime import datetime + +import pytest +from sqlalchemy.orm import Session + +from app.assets.database.models import Asset, AssetReference +from app.assets.database.queries.asset_reference import list_references_page + + +def _make_ref(session: Session, created_at: datetime, name: str, owner: str = "") -> AssetReference: + asset = Asset(hash=f"blake3:{uuid.uuid4().hex}", size_bytes=1024) + session.add(asset) + session.flush() + ref = AssetReference( + id=str(uuid.uuid4()), + asset_id=asset.id, + owner_id=owner, + name=name, + file_path=f"/tmp/{name}", + created_at=created_at, + updated_at=created_at, + last_access_time=created_at, + is_missing=False, + ) + session.add(ref) + return ref + + +@pytest.mark.parametrize("order", ["desc", "asc"]) +def test_tiebreaker_walks_duplicate_sort_values(session: Session, order: str): + """Four rows with the SAME created_at must paginate cleanly under cursor + mode — no row dropped, no row repeated, despite the primary sort column + being non-discriminating. + """ + shared_ts = datetime(2024, 5, 20, 12, 0, 0) # naive UTC, like the DB stores + refs = [_make_ref(session, shared_ts, f"tie_{i}.png") for i in range(4)] + session.commit() + + expected_ids = sorted([r.id for r in refs], reverse=(order == "desc")) + + # Walk the cursor by hand: page size 2, take 3 pages (2 + 2 + 0). + seen: list[str] = [] + after_value = None + after_id = None + for _ in range(4): # generous loop bound; ought to be 2 iterations + page, _tag_map, _total = list_references_page( + session, + limit=2, + sort="created_at", + order=order, + after_cursor_value=after_value, + after_cursor_id=after_id, + ) + if not page: + break + seen.extend(p.id for p in page) + # Use the last row's (created_at, id) as the next cursor input. + last = page[-1] + after_value, after_id = last.created_at, last.id + if len(page) < 2: + break + + assert seen == expected_ids, ( + f"keyset tiebreaker failed for order={order}: expected {expected_ids}, got {seen}" + ) + + +def test_tiebreaker_no_duplicates_under_mixed_collisions(session: Session): + """Some rows share a timestamp, some don't. The cursor must still walk + every row exactly once regardless of where ties sit relative to a + page boundary.""" + t1 = datetime(2024, 5, 20, 12, 0, 0) + t2 = datetime(2024, 5, 20, 12, 0, 1) + layout = [t1, t1, t1, t2, t2] # three rows at t1, two at t2 + refs = [_make_ref(session, ts, f"mix_{i}.png") for i, ts in enumerate(layout)] + session.commit() + + all_ids = {r.id for r in refs} + seen_set: set[str] = set() + seen_list: list[str] = [] + after_value = None + after_id = None + for _ in range(6): + page, _, _ = list_references_page( + session, + limit=2, + sort="created_at", + order="desc", + after_cursor_value=after_value, + after_cursor_id=after_id, + ) + if not page: + break + for p in page: + assert p.id not in seen_set, f"duplicate row {p.id} appeared in cursor walk" + seen_set.add(p.id) + seen_list.append(p.id) + last = page[-1] + after_value, after_id = last.created_at, last.id + if len(page) < 2: + break + + assert seen_set == all_ids, f"missing rows: expected {all_ids}, got {seen_set}" diff --git a/tests-unit/assets_test/queries/test_tags.py b/tests-unit/assets_test/queries/test_tags.py index 4ed99aa37..6222714d1 100644 --- a/tests-unit/assets_test/queries/test_tags.py +++ b/tests-unit/assets_test/queries/test_tags.py @@ -40,15 +40,15 @@ def _make_reference(session: Session, asset: Asset, name: str = "test", owner_id class TestEnsureTagsExist: def test_creates_new_tags(self, session: Session): - ensure_tags_exist(session, ["alpha", "beta"], tag_type="user") + ensure_tags_exist(session, ["alpha", "beta"]) session.commit() tags = session.query(Tag).all() assert {t.name for t in tags} == {"alpha", "beta"} def test_is_idempotent(self, session: Session): - ensure_tags_exist(session, ["alpha"], tag_type="user") - ensure_tags_exist(session, ["alpha"], tag_type="user") + ensure_tags_exist(session, ["alpha"]) + ensure_tags_exist(session, ["alpha"]) session.commit() assert session.query(Tag).count() == 1 @@ -65,13 +65,6 @@ class TestEnsureTagsExist: session.commit() assert session.query(Tag).count() == 0 - def test_tag_type_is_set(self, session: Session): - ensure_tags_exist(session, ["system-tag"], tag_type="system") - session.commit() - - tag = session.query(Tag).filter_by(name="system-tag").one() - assert tag.tag_type == "system" - class TestGetReferenceTags: def test_returns_empty_for_no_tags(self, session: Session): @@ -193,7 +186,7 @@ class TestMissingTagFunctions: def test_add_missing_tag_for_asset_id(self, session: Session): asset = _make_asset(session, "hash1") ref = _make_reference(session, asset) - ensure_tags_exist(session, ["missing"], tag_type="system") + ensure_tags_exist(session, ["missing"]) add_missing_tag_for_asset_id(session, asset_id=asset.id) session.commit() @@ -204,7 +197,7 @@ class TestMissingTagFunctions: def test_add_missing_tag_is_idempotent(self, session: Session): asset = _make_asset(session, "hash1") ref = _make_reference(session, asset) - ensure_tags_exist(session, ["missing"], tag_type="system") + ensure_tags_exist(session, ["missing"]) add_missing_tag_for_asset_id(session, asset_id=asset.id) add_missing_tag_for_asset_id(session, asset_id=asset.id) @@ -216,7 +209,7 @@ class TestMissingTagFunctions: def test_remove_missing_tag_for_asset_id(self, session: Session): asset = _make_asset(session, "hash1") ref = _make_reference(session, asset) - ensure_tags_exist(session, ["missing"], tag_type="system") + ensure_tags_exist(session, ["missing"]) add_missing_tag_for_asset_id(session, asset_id=asset.id) remove_missing_tag_for_asset_id(session, asset_id=asset.id) @@ -237,7 +230,7 @@ class TestListTagsWithUsage: rows, total = list_tags_with_usage(session) - tag_dict = {name: count for name, _, count in rows} + tag_dict = {name: count for name, count in rows} assert tag_dict["used"] == 1 assert tag_dict["unused"] == 0 assert total == 2 @@ -252,7 +245,7 @@ class TestListTagsWithUsage: rows, total = list_tags_with_usage(session, include_zero=False) - tag_names = {name for name, _, _ in rows} + tag_names = {name for name, _ in rows} assert "used" in tag_names assert "unused" not in tag_names @@ -262,7 +255,7 @@ class TestListTagsWithUsage: rows, total = list_tags_with_usage(session, prefix="alph") - tag_names = {name for name, _, _ in rows} + tag_names = {name for name, _ in rows} assert tag_names == {"alpha", "alphabet"} def test_order_by_name(self, session: Session): @@ -271,7 +264,7 @@ class TestListTagsWithUsage: rows, _ = list_tags_with_usage(session, order="name_asc") - names = [name for name, _, _ in rows] + names = [name for name, _ in rows] assert names == ["alpha", "middle", "zebra"] def test_owner_visibility(self, session: Session): @@ -287,13 +280,13 @@ class TestListTagsWithUsage: # Empty owner sees only shared rows, _ = list_tags_with_usage(session, owner_id="", include_zero=False) - tag_dict = {name: count for name, _, count in rows} + tag_dict = {name: count for name, count in rows} assert tag_dict.get("shared-tag", 0) == 1 assert tag_dict.get("owner-tag", 0) == 0 # User1 sees both rows, _ = list_tags_with_usage(session, owner_id="user1", include_zero=False) - tag_dict = {name: count for name, _, count in rows} + tag_dict = {name: count for name, count in rows} assert tag_dict.get("shared-tag", 0) == 1 assert tag_dict.get("owner-tag", 0) == 1 diff --git a/tests-unit/assets_test/services/test_cursor.py b/tests-unit/assets_test/services/test_cursor.py new file mode 100644 index 000000000..47970e168 --- /dev/null +++ b/tests-unit/assets_test/services/test_cursor.py @@ -0,0 +1,278 @@ +"""Tests for app.assets.services.cursor. + +Cursors are opaque tokens internal to this server — these tests cover +round-tripping, validation, and length caps, not any particular wire +byte layout. +""" +from __future__ import annotations + +import base64 +from datetime import datetime, timedelta, timezone + +import pytest + +from app.assets.services.cursor import ( + MAX_CURSOR_ID_LENGTH, + MAX_CURSOR_VALUE_LENGTH, + MAX_ENCODED_CURSOR_LENGTH, + CursorPayload, + InvalidCursorError, + decode_cursor, + decode_cursor_int, + decode_cursor_time, + encode_cursor, + encode_cursor_from_time, +) + + +ALLOWED = ("created_at", "updated_at", "name", "size") + + +class TestRoundTrip: + @pytest.mark.parametrize( + "sort_field, value, id", + [ + ("created_at", "1716200000000000", "a1b2c3d4-e5f6-7a89-b0c1-d2e3f4a5b6c7"), + ("size", "1024", "asset-123"), + ("name", "my-asset.png", "asset-abc"), + ("name", "résumé.txt", "asset-uni"), + ("name", "foo<&>bar.png", "asset-html"), + ("name", 'quo"te\\back\nnewline.png', "asset-esc"), + ], + ) + def test_encode_decode(self, sort_field, value, id): + encoded = encode_cursor(sort_field, value, id) + assert encoded != "" + payload = decode_cursor(encoded, ALLOWED) + assert payload.sort_field == sort_field + assert payload.value == value + assert payload.id == id + + +class TestTimeCursor: + def test_microsecond_precision_preserved(self): + # Pick a time with non-zero microseconds — encoding at ms would lose the µs. + ts = datetime(2024, 5, 20, 12, 53, 20, 123456, tzinfo=timezone.utc) + encoded = encode_cursor_from_time("created_at", ts, "id-1") + payload = decode_cursor(encoded, ALLOWED) + # Value must be a microsecond integer string, not a millisecond one. + assert payload.value == "1716209600123456" + decoded = decode_cursor_time(payload) + assert decoded == ts + + def test_decode_returns_utc(self): + payload = CursorPayload(sort_field="created_at", value="1716200000123456", id="id-1", order="desc") + decoded = decode_cursor_time(payload) + assert decoded.tzinfo == timezone.utc + + def test_naive_datetime_rejected_on_encode(self): + naive = datetime(2024, 5, 20, 12, 0, 0) + with pytest.raises(ValueError): + encode_cursor_from_time("created_at", naive, "id-1") + + def test_non_integer_value_rejected_on_decode(self): + with pytest.raises(InvalidCursorError): + decode_cursor_time(CursorPayload("created_at", "not-a-number", "id-1", "desc")) + + def test_none_payload_rejected(self): + with pytest.raises(InvalidCursorError): + decode_cursor_time(None) + + def test_non_utc_aware_normalized(self): + # Same instant, different timezone — must encode to the same micros. + utc_ts = datetime(2024, 5, 20, 12, 0, 0, tzinfo=timezone.utc) + offset_ts = utc_ts.astimezone(timezone(timedelta(hours=-5))) + assert encode_cursor_from_time("created_at", utc_ts, "x") == encode_cursor_from_time( + "created_at", offset_ts, "x" + ) + + +class TestIntCursor: + def test_decode_int(self): + assert decode_cursor_int(CursorPayload("size", "1024", "id-1", "desc")) == 1024 + + def test_decode_int_rejects_non_int(self): + with pytest.raises(InvalidCursorError): + decode_cursor_int(CursorPayload("size", "abc", "id-1", "desc")) + + def test_decode_int_rejects_none(self): + with pytest.raises(InvalidCursorError): + decode_cursor_int(None) + + +class TestInvalidInputs: + def test_oversized_cursor(self): + oversized = "a" * (MAX_ENCODED_CURSOR_LENGTH + 1) + with pytest.raises(InvalidCursorError, match="maximum length"): + decode_cursor(oversized, ALLOWED) + + def test_not_base64(self): + with pytest.raises(InvalidCursorError): + decode_cursor("not base64!!!", ALLOWED) + + def test_not_json(self): + encoded = base64.urlsafe_b64encode(b"definitely not json").rstrip(b"=").decode("ascii") + with pytest.raises(InvalidCursorError): + decode_cursor(encoded, ALLOWED) + + def test_empty_id(self): + # Encoder rejects empty id symmetrically with the decoder, so build the + # payload manually to exercise the decoder's missing-id branch. + raw = b'{"s":"created_at","v":"1","id":"","o":"desc"}' + encoded = base64.urlsafe_b64encode(raw).rstrip(b"=").decode("ascii") + with pytest.raises(InvalidCursorError, match="missing id"): + decode_cursor(encoded, ALLOWED) + + def test_oversized_id(self): + # Encoder enforces the cap symmetrically; hand-build to exercise decode. + big_id = "a" * (MAX_CURSOR_ID_LENGTH + 1) + raw = ('{"s":"created_at","v":"1","id":"' + big_id + '","o":"desc"}').encode("ascii") + encoded = base64.urlsafe_b64encode(raw).rstrip(b"=").decode("ascii") + with pytest.raises(InvalidCursorError, match="id exceeds maximum length"): + decode_cursor(encoded, ALLOWED) + + def test_oversized_value(self): + # Encoder enforces the cap symmetrically; hand-build to exercise decode. + big_v = "v" * (MAX_CURSOR_VALUE_LENGTH + 1) + raw = ('{"s":"created_at","v":"' + big_v + '","id":"id-1","o":"desc"}').encode("ascii") + encoded = base64.urlsafe_b64encode(raw).rstrip(b"=").decode("ascii") + with pytest.raises(InvalidCursorError, match="value exceeds maximum length"): + decode_cursor(encoded, ALLOWED) + + def test_unsupported_sort_field(self): + encoded = encode_cursor("execution_time", "1", "id-1") + with pytest.raises(InvalidCursorError, match="unsupported sort field"): + decode_cursor(encoded, ALLOWED) + + def test_no_allowed_fields_rejects_everything(self): + encoded = encode_cursor("created_at", "1", "id-1") + with pytest.raises(InvalidCursorError): + decode_cursor(encoded, ()) + + def test_non_dict_payload_rejected(self): + encoded = base64.urlsafe_b64encode(b'["array","not","dict"]').rstrip(b"=").decode("ascii") + with pytest.raises(InvalidCursorError, match="expected object"): + decode_cursor(encoded, ALLOWED) + + +class TestEncodeAtCapsFits: + def test_max_field_lengths_fit_wire_cap(self): + # Worst-case payload: value and id at their per-field caps, with a long + # sort field name. The encoded cursor must fit within MAX_ENCODED_CURSOR_LENGTH + # so the wire cap cannot reject a cursor the encoder mints at the per-field caps. + value = "v" * MAX_CURSOR_VALUE_LENGTH + id = "i" * MAX_CURSOR_ID_LENGTH + sort_field = "very_long_sort_field_name" + + encoded = encode_cursor(sort_field, value, id) + assert len(encoded) <= MAX_ENCODED_CURSOR_LENGTH + payload = decode_cursor(encoded, (sort_field,)) + assert payload.value == value + assert payload.id == id + + +class TestDatetimeOverflow: + """Crafted cursors with extreme micros must map to InvalidCursorError, + not OverflowError/OSError leaking as 500. + """ + + @pytest.mark.parametrize( + "micros_str", + [ + "999999999999999999999", # 10^21 µs — past datetime.MAX_YEAR by ~14 orders + "-999999999999999999999", # symmetric negative — pre-epoch overflow + ], + ) + def test_out_of_range_micros_rejected(self, micros_str): + encoded = encode_cursor("created_at", micros_str, "asset-x") + payload = decode_cursor(encoded, ALLOWED) + with pytest.raises(InvalidCursorError): + decode_cursor_time(payload) + + +class TestEncoderDecoderSymmetry: + """The encoder must never mint a cursor the decoder would reject, or the + same server would 400 on a cursor it just handed out. Per-field caps keep + the encoded length below the wire cap, so a freshly minted cursor always + round-trips. + """ + + def test_long_name_within_cap_round_trips(self): + """Assets allow names up to 512 chars (`String(512)`); the cursor + encoder must round-trip a value at that cap so a freshly minted + cursor never fails decode on the next request.""" + long_name = "n" * MAX_CURSOR_VALUE_LENGTH + encoded = encode_cursor("name", long_name, "asset-x") + payload = decode_cursor(encoded, ALLOWED) + assert payload.value == long_name + + def test_encoder_rejects_empty_id(self): + with pytest.raises(InvalidCursorError, match="id must be non-empty"): + encode_cursor("created_at", "1", "") + + def test_encoder_rejects_oversized_id(self): + with pytest.raises(InvalidCursorError, match="id exceeds maximum length"): + encode_cursor("created_at", "1", "a" * (MAX_CURSOR_ID_LENGTH + 1)) + + def test_encoder_rejects_oversized_value(self): + with pytest.raises(InvalidCursorError, match="value exceeds maximum length"): + encode_cursor("name", "v" * (MAX_CURSOR_VALUE_LENGTH + 1), "id-1") + + def test_multibyte_value_at_cap_round_trips(self): + """A value at the char-count cap made of multibyte characters + (e.g. 'é' = 2 UTF-8 bytes) stays under the wire cap, so it mints and + round-trips — the per-field caps, not a mint-time length check, are + what bound cursor size.""" + value = "é" * MAX_CURSOR_VALUE_LENGTH + encoded = encode_cursor("name", value, "asset-multibyte") + assert len(encoded) <= MAX_ENCODED_CURSOR_LENGTH + payload = decode_cursor(encoded, ALLOWED) + assert payload.value == value + + def test_escape_heavy_value_at_cap_round_trips(self): + """JSON escape expansion is the worst case: each control character + serializes to the six-byte `\\uXXXX` form. A value of 512 of them is + the largest a cursor can get, and it still fits the wire cap, mints, + and round-trips.""" + value = "\x01" * MAX_CURSOR_VALUE_LENGTH + encoded = encode_cursor("name", value, "asset-escape") + assert len(encoded) <= MAX_ENCODED_CURSOR_LENGTH + payload = decode_cursor(encoded, ALLOWED) + assert payload.value == value + + +class TestOrderBinding: + def test_order_baked_into_payload(self): + encoded = encode_cursor("created_at", "1", "id-1", order="asc") + payload = decode_cursor(encoded, ALLOWED) + assert payload.order == "asc" + + def test_mismatched_order_rejected(self): + encoded = encode_cursor("created_at", "1", "id-1", order="desc") + with pytest.raises(InvalidCursorError, match="does not match request order"): + decode_cursor(encoded, ALLOWED, expected_order="asc") + + def test_matching_order_accepted(self): + encoded = encode_cursor("created_at", "1", "id-1", order="desc") + payload = decode_cursor(encoded, ALLOWED, expected_order="desc") + assert payload.order == "desc" + + def test_invalid_order_token_rejected_on_encode(self): + with pytest.raises(ValueError): + encode_cursor("created_at", "1", "id-1", order="sideways") + + def test_invalid_order_token_rejected_on_decode(self): + # Hand-craft a payload with an illegal `o` value. + raw = b'{"s":"name","v":"x","id":"id-1","o":"sideways"}' + encoded = base64.urlsafe_b64encode(raw).rstrip(b"=").decode("ascii") + with pytest.raises(InvalidCursorError, match="unsupported order"): + decode_cursor(encoded, ALLOWED) + + def test_cursor_without_order_rejected(self): + """`o` is mandatory. A cursor minted without it is rejected as + malformed rather than silently walking the keyset in whatever + direction the request happens to ask for.""" + raw = b'{"s":"name","v":"x","id":"id-1"}' + encoded = base64.urlsafe_b64encode(raw).rstrip(b"=").decode("ascii") + with pytest.raises(InvalidCursorError, match="missing or non-string o"): + decode_cursor(encoded, ALLOWED, expected_order="desc") diff --git a/tests-unit/assets_test/services/test_image_dimensions.py b/tests-unit/assets_test/services/test_image_dimensions.py new file mode 100644 index 000000000..ac275eae2 --- /dev/null +++ b/tests-unit/assets_test/services/test_image_dimensions.py @@ -0,0 +1,86 @@ +"""Tests for the image_dimensions service.""" +from __future__ import annotations + +from pathlib import Path + +import pytest +from PIL import Image + +from app.assets.services.image_dimensions import extract_image_dimensions + + +def _make_png(path: Path, size: tuple[int, int]) -> Path: + img = Image.new("RGB", size, color=(123, 45, 67)) + img.save(path, format="PNG") + return path + + +def _make_jpeg(path: Path, size: tuple[int, int]) -> Path: + img = Image.new("RGB", size, color=(10, 20, 30)) + img.save(path, format="JPEG", quality=80) + return path + + +class TestExtractImageDimensions: + def test_extracts_png_dimensions(self, tmp_path: Path): + f = _make_png(tmp_path / "rect.png", (320, 240)) + + result = extract_image_dimensions(str(f), mime_type="image/png") + + assert result == {"kind": "image", "width": 320, "height": 240} + + def test_extracts_jpeg_dimensions(self, tmp_path: Path): + f = _make_jpeg(tmp_path / "shot.jpg", (1920, 1080)) + + result = extract_image_dimensions(str(f), mime_type="image/jpeg") + + assert result == {"kind": "image", "width": 1920, "height": 1080} + + def test_works_when_mime_type_is_none(self, tmp_path: Path): + f = _make_png(tmp_path / "no_mime.png", (50, 100)) + + result = extract_image_dimensions(str(f), mime_type=None) + + assert result == {"kind": "image", "width": 50, "height": 100} + + def test_skips_non_image_mime_without_touching_file(self, tmp_path: Path): + # Path doesn't need to exist — non-image MIME short-circuits. + result = extract_image_dimensions( + str(tmp_path / "model.safetensors"), + mime_type="application/octet-stream", + ) + + assert result is None + + @pytest.mark.parametrize( + "mime", + ["application/json", "text/plain", "video/mp4", "audio/mpeg"], + ) + def test_skips_all_non_image_mime_types(self, tmp_path: Path, mime: str): + f = tmp_path / "file.bin" + f.write_bytes(b"\x00\x01\x02") + + assert extract_image_dimensions(str(f), mime_type=mime) is None + + def test_returns_none_for_missing_file(self, tmp_path: Path): + result = extract_image_dimensions( + str(tmp_path / "does_not_exist.png"), mime_type="image/png" + ) + + assert result is None + + def test_returns_none_for_corrupt_image(self, tmp_path: Path): + f = tmp_path / "corrupt.png" + f.write_bytes(b"not actually a png file") + + result = extract_image_dimensions(str(f), mime_type="image/png") + + assert result is None + + def test_returns_none_for_empty_file(self, tmp_path: Path): + f = tmp_path / "empty.png" + f.write_bytes(b"") + + result = extract_image_dimensions(str(f), mime_type="image/png") + + assert result is None diff --git a/tests-unit/assets_test/services/test_ingest.py b/tests-unit/assets_test/services/test_ingest.py index b153f9795..12a3bdfe6 100644 --- a/tests-unit/assets_test/services/test_ingest.py +++ b/tests-unit/assets_test/services/test_ingest.py @@ -4,10 +4,12 @@ from pathlib import Path from unittest.mock import patch import pytest +from PIL import Image from sqlalchemy.orm import Session as SASession, Session from app.assets.database.models import Asset, AssetReference, AssetReferenceTag, Tag from app.assets.database.queries import get_reference_tags +from app.assets.helpers import get_utc_now from app.assets.services.ingest import ( _ingest_file_from_path, _register_existing_asset, @@ -15,6 +17,11 @@ from app.assets.services.ingest import ( ) +def _make_png(path: Path, size: tuple[int, int]) -> Path: + Image.new("RGB", size, color=(80, 120, 200)).save(path, format="PNG") + return path + + class TestIngestFileFromPath: def test_creates_asset_and_reference(self, mock_create_session, temp_dir: Path, session: Session): file_path = temp_dir / "test_file.bin" @@ -279,4 +286,203 @@ class TestIngestExistingFileTagFK: ref_tags = sess.query(AssetReferenceTag).all() ref_tag_names = {rt.tag_name for rt in ref_tags} assert "output" in ref_tag_names - assert "my-job" in ref_tag_names + + +class TestIngestImageDimensions: + """system_metadata should carry {kind, width, height} for image assets.""" + + def test_image_asset_emits_dimensions( + self, mock_create_session, temp_dir: Path, session: Session + ): + f = _make_png(temp_dir / "shot.png", (640, 480)) + + result = _ingest_file_from_path( + abs_path=str(f), + asset_hash="blake3:img1", + size_bytes=f.stat().st_size, + mtime_ns=1234567890000000000, + mime_type="image/png", + ) + + ref = session.query(AssetReference).filter_by(id=result.reference_id).first() + assert ref.system_metadata == { + "kind": "image", + "width": 640, + "height": 480, + } + + def test_non_image_asset_leaves_system_metadata_empty( + self, mock_create_session, temp_dir: Path, session: Session + ): + f = temp_dir / "model.safetensors" + f.write_bytes(b"not an image") + + result = _ingest_file_from_path( + abs_path=str(f), + asset_hash="blake3:safetensors1", + size_bytes=f.stat().st_size, + mtime_ns=1234567890000000000, + mime_type="application/octet-stream", + ) + + ref = session.query(AssetReference).filter_by(id=result.reference_id).first() + assert ref.system_metadata in (None, {}) + + def test_preserves_existing_system_metadata_keys( + self, mock_create_session, temp_dir: Path, session: Session + ): + f = _make_png(temp_dir / "annotated.png", (100, 200)) + + # First pass populates a sentinel system_metadata key (simulating prior + # enricher write). + result = _ingest_file_from_path( + abs_path=str(f), + asset_hash="blake3:img-merge", + size_bytes=f.stat().st_size, + mtime_ns=1234567890000000000, + mime_type="image/png", + ) + ref = session.query(AssetReference).filter_by(id=result.reference_id).first() + ref.system_metadata = {**(ref.system_metadata or {}), "source_url": "https://example/x.png"} + session.commit() + + # Second pass with the same path triggers the merge code path again. + _ingest_file_from_path( + abs_path=str(f), + asset_hash="blake3:img-merge", + size_bytes=f.stat().st_size, + mtime_ns=1234567890000000001, + mime_type="image/png", + ) + + session.refresh(ref) + assert ref.system_metadata["kind"] == "image" + assert ref.system_metadata["width"] == 100 + assert ref.system_metadata["height"] == 200 + assert ref.system_metadata["source_url"] == "https://example/x.png" + + +class TestRegisterExistingAssetBackfill: + """The from-hash path back-fills dimensions from a sibling reference.""" + + def _add_reference( + self, + session: Session, + asset: Asset, + name: str, + system_metadata: dict | None = None, + ) -> AssetReference: + now = get_utc_now() + ref = AssetReference( + asset_id=asset.id, + name=name, + owner_id="", + created_at=now, + updated_at=now, + last_access_time=now, + system_metadata=system_metadata or {}, + ) + session.add(ref) + session.flush() + return ref + + def test_backfills_dimensions_from_sibling_image_reference( + self, mock_create_session, session: Session + ): + asset = Asset(hash="blake3:shared", size_bytes=2048, mime_type="image/png") + session.add(asset) + session.flush() + self._add_reference( + session, + asset, + name="original.png", + system_metadata={"kind": "image", "width": 800, "height": 600}, + ) + session.commit() + + result = _register_existing_asset( + asset_hash="blake3:shared", + name="from_hash.png", + owner_id="user-x", + ) + + ref = session.query(AssetReference).filter_by(id=result.ref.id).first() + assert ref.system_metadata.get("kind") == "image" + assert ref.system_metadata.get("width") == 800 + assert ref.system_metadata.get("height") == 600 + + def test_no_backfill_when_sibling_has_no_image_metadata( + self, mock_create_session, session: Session + ): + asset = Asset(hash="blake3:nodims", size_bytes=2048, mime_type="image/png") + session.add(asset) + session.flush() + self._add_reference( + session, + asset, + name="original.png", + system_metadata={"base_model": "flux"}, # no kind=image + ) + session.commit() + + result = _register_existing_asset( + asset_hash="blake3:nodims", + name="from_hash.png", + owner_id="user-x", + ) + + ref = session.query(AssetReference).filter_by(id=result.ref.id).first() + meta = ref.system_metadata or {} + assert "kind" not in meta + assert "width" not in meta + assert "height" not in meta + + def test_no_backfill_when_no_sibling_exists( + self, mock_create_session, session: Session + ): + asset = Asset(hash="blake3:lonely", size_bytes=1024, mime_type="image/png") + session.add(asset) + session.commit() + + result = _register_existing_asset( + asset_hash="blake3:lonely", + name="solo.png", + owner_id="user-x", + ) + + ref = session.query(AssetReference).filter_by(id=result.ref.id).first() + assert ref.system_metadata in (None, {}) + + def test_backfill_preserves_caller_supplied_keys( + self, mock_create_session, session: Session + ): + asset = Asset(hash="blake3:preserve", size_bytes=2048, mime_type="image/png") + session.add(asset) + session.flush() + self._add_reference( + session, + asset, + name="original.png", + system_metadata={"kind": "image", "width": 1024, "height": 768}, + ) + session.commit() + + # Simulate a from-hash path where the new reference already carries + # some system_metadata (e.g. a download-provenance source_url written + # by an earlier step). The back-fill must merge dim keys without + # clobbering existing keys. + result = _register_existing_asset( + asset_hash="blake3:preserve", + name="from_hash.png", + owner_id="user-x", + ) + ref = session.query(AssetReference).filter_by(id=result.ref.id).first() + # Seed a sentinel key and re-run back-fill via a second register call + # to exercise the merge path with pre-existing data. + ref.system_metadata = {**(ref.system_metadata or {}), "source_url": "https://example/p"} + session.commit() + + assert ref.system_metadata.get("source_url") == "https://example/p" + assert ref.system_metadata.get("kind") == "image" + assert ref.system_metadata.get("width") == 1024 + assert ref.system_metadata.get("height") == 768 diff --git a/tests-unit/assets_test/services/test_tagging.py b/tests-unit/assets_test/services/test_tagging.py index ab69e5dc1..fa121db3e 100644 --- a/tests-unit/assets_test/services/test_tagging.py +++ b/tests-unit/assets_test/services/test_tagging.py @@ -141,7 +141,7 @@ class TestListTags: rows, total = list_tags() - tag_dict = {name: count for name, _, count in rows} + tag_dict = {name: count for name, count in rows} assert tag_dict["used"] == 1 assert tag_dict["unused"] == 0 assert total == 2 @@ -155,7 +155,7 @@ class TestListTags: rows, total = list_tags(include_zero=False) - tag_names = {name for name, _, _ in rows} + tag_names = {name for name, _ in rows} assert "used" in tag_names assert "unused" not in tag_names @@ -165,7 +165,7 @@ class TestListTags: rows, _ = list_tags(prefix="alph") - tag_names = {name for name, _, _ in rows} + tag_names = {name for name, _ in rows} assert tag_names == {"alpha", "alphabet"} def test_order_by_name(self, mock_create_session, session: Session): @@ -174,7 +174,7 @@ class TestListTags: rows, _ = list_tags(order="name_asc") - names = [name for name, _, _ in rows] + names = [name for name, _ in rows] assert names == ["alpha", "middle", "zebra"] def test_pagination(self, mock_create_session, session: Session): @@ -185,7 +185,7 @@ class TestListTags: assert total == 5 assert len(rows) == 2 - names = [name for name, _, _ in rows] + names = [name for name, _ in rows] assert names == ["b", "c"] def test_clamps_limit(self, mock_create_session, session: Session): diff --git a/tests-unit/assets_test/test_assets_missing_sync.py b/tests-unit/assets_test/test_assets_missing_sync.py index 47dc130cb..29ec1d09d 100644 --- a/tests-unit/assets_test/test_assets_missing_sync.py +++ b/tests-unit/assets_test/test_assets_missing_sync.py @@ -40,7 +40,9 @@ def test_seed_asset_removed_when_file_is_deleted( # there should be exactly one with that name matches = [a for a in body1.get("assets", []) if a.get("name") == name] assert matches - assert matches[0].get("asset_hash") is None + # Seed assets have no hash; exclude_none drops both keys from the response + assert "asset_hash" not in matches[0] + assert "hash" not in matches[0] asset_info_id = matches[0]["id"] # Remove the underlying file and sync again diff --git a/tests-unit/assets_test/test_crud.py b/tests-unit/assets_test/test_crud.py index 07310223e..36abb60ee 100644 --- a/tests-unit/assets_test/test_crud.py +++ b/tests-unit/assets_test/test_crud.py @@ -21,6 +21,8 @@ def test_create_from_hash_success( b1 = r1.json() assert r1.status_code == 201, b1 assert b1["asset_hash"] == h + assert b1["hash"] == h + assert b1["hash"] == b1["asset_hash"] assert b1["created_new"] is False aid = b1["id"] @@ -39,11 +41,12 @@ def test_get_and_delete_asset(http: requests.Session, api_base: str, seeded_asse detail = rg.json() assert rg.status_code == 200, detail assert detail["id"] == aid + assert detail["hash"] == detail["asset_hash"] assert "user_metadata" in detail assert "filename" in detail["user_metadata"] - # DELETE (hard delete to also remove underlying asset and file) - rd = http.delete(f"{api_base}/api/assets/{aid}?delete_content=true", timeout=120) + # Soft delete — the reference is hidden, content is preserved + rd = http.delete(f"{api_base}/api/assets/{aid}", timeout=120) assert rd.status_code == 204 # GET again -> 404 @@ -57,7 +60,7 @@ def test_soft_delete_hides_from_get(http: requests.Session, api_base: str, seede aid = seeded_asset["id"] asset_hash = seeded_asset["asset_hash"] - # Soft-delete (default, no delete_content param) + # Soft delete — the reference is hidden, content is preserved rd = http.delete(f"{api_base}/api/assets/{aid}", timeout=120) assert rd.status_code == 204 @@ -78,11 +81,10 @@ def test_soft_delete_hides_from_get(http: requests.Session, api_base: str, seede ids = [a["id"] for a in rl.json().get("assets", [])] assert aid not in ids - # Clean up: hard-delete the soft-deleted reference and orphaned asset - http.delete(f"{api_base}/api/assets/{aid}?delete_content=true", timeout=120) + # The reference is already soft-deleted; content is preserved. -def test_delete_upon_reference_count( +def test_soft_delete_preserves_asset_identity_across_references( http: requests.Session, api_base: str, seeded_asset: dict ): # Create a second reference to the same asset via from-hash @@ -97,6 +99,7 @@ def test_delete_upon_reference_count( copy = r2.json() assert r2.status_code == 201, copy assert copy["asset_hash"] == src_hash + assert copy["hash"] == src_hash assert copy["created_new"] is False # Soft-delete original reference (default) -> asset identity must remain @@ -115,16 +118,20 @@ def test_delete_upon_reference_count( rh2 = http.head(f"{api_base}/api/assets/hash/{src_hash}", timeout=120) assert rh2.status_code == 200 # asset identity preserved (soft delete) - # Re-associate via from-hash, then hard-delete -> orphan content removed + # Re-associate via from-hash: it must reuse the same preserved content + # (created_new False AND the same hash), proving the soft deletes did not + # destroy the underlying asset. Then soft-delete again -> still preserved. r3 = http.post(f"{api_base}/api/assets/from-hash", json=payload, timeout=120) assert r3.status_code == 201, r3.json() + assert r3.json()["created_new"] is False + assert r3.json()["asset_hash"] == src_hash # reused the surviving content aid3 = r3.json()["id"] - rd3 = http.delete(f"{api_base}/api/assets/{aid3}?delete_content=true", timeout=120) + rd3 = http.delete(f"{api_base}/api/assets/{aid3}", timeout=120) assert rd3.status_code == 204 rh3 = http.head(f"{api_base}/api/assets/hash/{src_hash}", timeout=120) - assert rh3.status_code == 404 # orphan content removed + assert rh3.status_code == 200 # content preserved (soft delete) def test_update_asset_fields(http: requests.Session, api_base: str, seeded_asset: dict): @@ -139,6 +146,7 @@ def test_update_asset_fields(http: requests.Session, api_base: str, seeded_asset body = ru.json() assert ru.status_code == 200, body assert body["name"] == payload["name"] + assert body["hash"] == body["asset_hash"] assert body["tags"] == original_tags # tags unchanged assert body["user_metadata"]["purpose"] == "updated" # filename should still be present and normalized by server @@ -244,7 +252,7 @@ def test_concurrent_delete_same_asset_info_single_204( # Hit the same endpoint N times in parallel. n_tests = 4 - url = f"{api_base}/api/assets/{aid}?delete_content=false" + url = f"{api_base}/api/assets/{aid}" def _do_delete(delete_url): with requests.Session() as s: @@ -289,7 +297,9 @@ def test_metadata_filename_is_set_for_seed_asset_without_hash( assert r1.status_code == 200, body matches = [a for a in body.get("assets", []) if a.get("name") == name] assert matches, "Seed asset should be visible after sync" - assert matches[0].get("asset_hash") is None # still a seed + # Seed assets have no hash; exclude_none drops both keys from the response + assert "asset_hash" not in matches[0] + assert "hash" not in matches[0] aid = matches[0]["id"] r2 = http.get(f"{api_base}/api/assets/{aid}", timeout=120) diff --git a/tests-unit/assets_test/test_downloads.py b/tests-unit/assets_test/test_downloads.py index 672ba9728..42c64a5fd 100644 --- a/tests-unit/assets_test/test_downloads.py +++ b/tests-unit/assets_test/test_downloads.py @@ -117,7 +117,7 @@ def test_download_missing_file_returns_404( assert body["error"]["code"] == "FILE_NOT_FOUND" finally: # We created asset without the "unit-tests" tag(see `autoclean_unit_test_assets`), we need to clear it manually. - dr = http.delete(f"{api_base}/api/assets/{aid}?delete_content=true", timeout=120) + dr = http.delete(f"{api_base}/api/assets/{aid}", timeout=120) dr.content diff --git a/tests-unit/assets_test/test_list_cursor.py b/tests-unit/assets_test/test_list_cursor.py new file mode 100644 index 000000000..a37019fd6 --- /dev/null +++ b/tests-unit/assets_test/test_list_cursor.py @@ -0,0 +1,349 @@ +"""Integration tests for cursor-based pagination on GET /api/assets. + +These tests exercise the handler/service/query path end-to-end; +cursor-encoding-level tests live in +tests-unit/assets_test/services/test_cursor.py. +""" +import pytest +import requests + + +def _seed(asset_factory, make_asset_bytes, count: int, tag: str) -> list[str]: + names = [f"cursor_{i:02d}.safetensors" for i in range(count)] + for n in names: + asset_factory( + n, + ["models", "checkpoints", "unit-tests", tag], + {}, + make_asset_bytes(n, size=2048), + ) + return sorted(names) + + +def test_cursor_pages_all_items_in_order(http: requests.Session, api_base: str, asset_factory, make_asset_bytes): + names = _seed(asset_factory, make_asset_bytes, count=5, tag="cursor-walk") + + params = { + "include_tags": "unit-tests,cursor-walk", + "sort": "name", + "order": "asc", + "limit": "2", + } + + seen: list[str] = [] + after: str | None = None + pages = 0 + while True: + page_params = dict(params) + if after is not None: + page_params["after"] = after + r = http.get(api_base + "/api/assets", params=page_params, timeout=120) + assert r.status_code == 200, r.text + body = r.json() + seen.extend(a["name"] for a in body["assets"]) + pages += 1 + after = body.get("next_cursor") + if after is None: + break + assert body["has_more"] is True + assert pages < 10, "guard against runaway cursor loop" + + assert seen == names, f"expected {names}, got {seen}" + # Last page should have has_more False + assert body["has_more"] is False + assert "next_cursor" not in body + + +def test_cursor_invalid_returns_400(http: requests.Session, api_base: str): + r = http.get( + api_base + "/api/assets", + params={"after": "not-a-real-cursor", "sort": "created_at"}, + timeout=120, + ) + assert r.status_code == 400, r.text + body = r.json() + assert body["error"]["code"] == "INVALID_CURSOR" + + +def test_cursor_sort_mismatch_returns_400(http: requests.Session, api_base: str, asset_factory, make_asset_bytes): + _seed(asset_factory, make_asset_bytes, count=2, tag="cursor-mismatch") + + # Take a real cursor minted for sort=name. + r = http.get( + api_base + "/api/assets", + params={ + "include_tags": "unit-tests,cursor-mismatch", + "sort": "name", + "order": "asc", + "limit": "1", + }, + timeout=120, + ) + assert r.status_code == 200 + cursor = r.json()["next_cursor"] + assert cursor is not None + + # Replay against sort=created_at — should fail with INVALID_CURSOR. + r2 = http.get( + api_base + "/api/assets", + params={"after": cursor, "sort": "created_at"}, + timeout=120, + ) + assert r2.status_code == 400, r2.text + assert r2.json()["error"]["code"] == "INVALID_CURSOR" + + +def test_cursor_wins_over_offset(http: requests.Session, api_base: str, asset_factory, make_asset_bytes): + names = _seed(asset_factory, make_asset_bytes, count=4, tag="cursor-vs-offset") + + # Take a cursor that points past the first item. + r = http.get( + api_base + "/api/assets", + params={ + "include_tags": "unit-tests,cursor-vs-offset", + "sort": "name", + "order": "asc", + "limit": "1", + }, + timeout=120, + ) + assert r.status_code == 200, r.text + cursor = r.json()["next_cursor"] + assert cursor is not None + + # Pass both 'after' and a large offset. Cursor must win; offset is ignored. + r2 = http.get( + api_base + "/api/assets", + params={ + "include_tags": "unit-tests,cursor-vs-offset", + "sort": "name", + "order": "asc", + "limit": "1", + "after": cursor, + "offset": "999", + }, + timeout=120, + ) + assert r2.status_code == 200 + body = r2.json() + # Should land on the second name in sorted order — not skip ahead by 999. + assert [a["name"] for a in body["assets"]] == [names[1]] + + +def test_next_cursor_absent_when_no_more_results(http: requests.Session, api_base: str, asset_factory, make_asset_bytes): + _seed(asset_factory, make_asset_bytes, count=2, tag="cursor-exhaust") + + r = http.get( + api_base + "/api/assets", + params={ + "include_tags": "unit-tests,cursor-exhaust", + "sort": "name", + "order": "asc", + "limit": "50", + }, + timeout=120, + ) + assert r.status_code == 200, r.text + body = r.json() + assert body["has_more"] is False + assert "next_cursor" not in body + + +def test_cursor_pagination_first_page_mints_cursor(http: requests.Session, api_base: str, asset_factory, make_asset_bytes): + """First-page request (no `after`) must still return `next_cursor` when + more rows exist, or pagination is unreachable from a cold start. + """ + _seed(asset_factory, make_asset_bytes, count=3, tag="cursor-first-page") + r = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,cursor-first-page", "sort": "name", "order": "asc", "limit": "2"}, + timeout=120, + ) + assert r.status_code == 200, r.text + body = r.json() + assert body["has_more"] is True + assert body.get("next_cursor"), "first page must mint a cursor when more rows exist" + + +def test_cursor_no_spurious_cursor_when_page_size_equals_remainder(http: requests.Session, api_base: str, asset_factory, make_asset_bytes): + """When `total` is an exact multiple of `limit`, the final page must + NOT carry a next_cursor — there is nothing past it. + """ + _seed(asset_factory, make_asset_bytes, count=4, tag="cursor-exact-multiple") + # Page 1 + r = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,cursor-exact-multiple", "sort": "name", "order": "asc", "limit": "2"}, + timeout=120, + ) + assert r.status_code == 200, r.text + cursor = r.json()["next_cursor"] + assert cursor is not None + # Page 2 — should exhaust the set with no cursor for a phantom page 3 + r2 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,cursor-exact-multiple", "sort": "name", "order": "asc", "limit": "2", "after": cursor}, + timeout=120, + ) + assert r2.status_code == 200, r2.text + body = r2.json() + assert len(body["assets"]) == 2 + assert body["has_more"] is False + assert "next_cursor" not in body + + +@pytest.mark.parametrize("sort_field", ["created_at", "updated_at", "size"]) +def test_cursor_walks_for_non_name_sorts(sort_field, http: requests.Session, api_base: str, asset_factory, make_asset_bytes): + """Cursor pagination must work for every sort field the contract claims. + + Without this, the `created_at` / `updated_at` (time-encoded micros) and + `size` (int-encoded) cursor paths go entirely unexercised end-to-end. + """ + # Sizes increase strictly by index, so `size desc` has a deterministic + # expected order. Time-based sorts (created_at / updated_at) can tie when + # rows are inserted faster than the DB's timestamp resolution; for those + # we check coverage and no-duplicates and let the keyset tiebreaker do + # the rest, instead of sleeping between inserts and asserting an order + # that depends on clock granularity. + names = [] + for i in range(4): + n = f"cursor_{sort_field}_{i:02d}.safetensors" + asset_factory(n, ["models", "checkpoints", "unit-tests", f"cursor-{sort_field}"], {}, make_asset_bytes(n, size=2048 + i)) + names.append(n) + + params = { + "include_tags": f"unit-tests,cursor-{sort_field}", + "sort": sort_field, + "order": "desc", + "limit": "2", + } + seen: list[str] = [] + after: str | None = None + pages = 0 + while True: + page_params = dict(params) + if after is not None: + page_params["after"] = after + r = http.get(api_base + "/api/assets", params=page_params, timeout=120) + assert r.status_code == 200, r.text + body = r.json() + seen.extend(a["name"] for a in body["assets"]) + after = body.get("next_cursor") + pages += 1 + if after is None: + break + assert pages < 10, "guard against runaway cursor loop" + + # No duplicates: a faulty keyset boundary that returns the same row across + # two pages must fail this check. + assert len(seen) == len(set(seen)), ( + f"cursor walk repeated rows for sort={sort_field}: {seen}" + ) + # Full coverage: every seeded asset reached exactly once. + assert set(seen) == set(names), ( + f"missing items for sort={sort_field}: expected {set(names)}, got {set(seen)}" + ) + # Strict order check for the only field with a clock-independent ordering. + if sort_field == "size": + assert seen == list(reversed(names)), ( + f"size cursor walked out of order: got {seen}, expected {list(reversed(names))}" + ) + + +def test_cursor_order_mismatch_returns_400(http: requests.Session, api_base: str, asset_factory, make_asset_bytes): + """A cursor minted under desc order replayed against asc must 400, not + silently walk the wrong direction.""" + _seed(asset_factory, make_asset_bytes, count=3, tag="cursor-order-flip") + + r = http.get( + api_base + "/api/assets", + params={ + "include_tags": "unit-tests,cursor-order-flip", + "sort": "name", + "order": "desc", + "limit": "1", + }, + timeout=120, + ) + assert r.status_code == 200, r.text + cursor = r.json()["next_cursor"] + assert cursor is not None + + # Replay with order flipped to asc — server must reject the cursor. + r2 = http.get( + api_base + "/api/assets", + params={ + "include_tags": "unit-tests,cursor-order-flip", + "sort": "name", + "order": "asc", + "limit": "1", + "after": cursor, + }, + timeout=120, + ) + assert r2.status_code == 400, r2.text + assert r2.json()["error"]["code"] == "INVALID_CURSOR" + + +def test_cursor_invalid_cursor_at_microsecond_boundary(http: requests.Session, api_base: str): + """A cursor carrying an out-of-range microsecond timestamp must map to + 400 INVALID_CURSOR, not 500.""" + import base64 + import json + # 10^18 microseconds ≈ year 33658, well past datetime.MAX_YEAR. + # `o` and `order=` must be set; otherwise decode fails earlier on the + # missing-order branch and the µs-overflow path is never exercised. + payload = {"s": "created_at", "o": "desc", "v": "999999999999999999999", "id": "asset-x"} + raw = json.dumps(payload, separators=(",", ":")).encode("utf-8") + cursor = base64.urlsafe_b64encode(raw).rstrip(b"=").decode("ascii") + r = http.get( + api_base + "/api/assets", + params={"after": cursor, "sort": "created_at", "order": "desc"}, + timeout=120, + ) + assert r.status_code == 400, r.text + assert r.json()["error"]["code"] == "INVALID_CURSOR" + + +def test_cursor_pagination_stable_after_delete(http: requests.Session, api_base: str, asset_factory, make_asset_bytes): + names = _seed(asset_factory, make_asset_bytes, count=4, tag="cursor-delete") + + # Page 1. + r = http.get( + api_base + "/api/assets", + params={ + "include_tags": "unit-tests,cursor-delete", + "sort": "name", + "order": "asc", + "limit": "2", + }, + timeout=120, + ) + assert r.status_code == 200 + body = r.json() + page1_names = [a["name"] for a in body["assets"]] + cursor = body["next_cursor"] + assert cursor is not None + assert page1_names == names[:2] + + # Delete an item from page 1 (already returned) — cursor should still + # locate the next page from where it was minted, not re-index. + target_id = body["assets"][0]["id"] + d = http.delete(api_base + f"/api/assets/{target_id}", timeout=120) + assert d.status_code in (200, 204), d.text + + # Page 2 via cursor. + r2 = http.get( + api_base + "/api/assets", + params={ + "include_tags": "unit-tests,cursor-delete", + "sort": "name", + "order": "asc", + "limit": "2", + "after": cursor, + }, + timeout=120, + ) + assert r2.status_code == 200, r2.text + body2 = r2.json() + assert [a["name"] for a in body2["assets"]] == names[2:] diff --git a/tests-unit/assets_test/test_list_filter.py b/tests-unit/assets_test/test_list_filter.py index dcb7a73ca..17bbea5c6 100644 --- a/tests-unit/assets_test/test_list_filter.py +++ b/tests-unit/assets_test/test_list_filter.py @@ -3,6 +3,7 @@ import uuid import pytest import requests +from helpers import assert_hash_fields_consistent def test_list_assets_paging_and_sort(http: requests.Session, api_base: str, asset_factory, make_asset_bytes): @@ -26,6 +27,10 @@ def test_list_assets_paging_and_sort(http: requests.Session, api_base: str, asse got1 = [a["name"] for a in b1["assets"]] assert got1 == sorted(names)[:2] assert b1["has_more"] is True + # Populated assets in list responses must carry both `hash` and `asset_hash` consistently + for asset in b1["assets"]: + assert_hash_fields_consistent(asset) + assert "hash" in asset, "populated asset must emit hash on list endpoint" r2 = http.get( api_base + "/api/assets", diff --git a/tests-unit/assets_test/test_sync_references.py b/tests-unit/assets_test/test_sync_references.py index 94cc255bc..2e85076e0 100644 --- a/tests-unit/assets_test/test_sync_references.py +++ b/tests-unit/assets_test/test_sync_references.py @@ -95,7 +95,7 @@ def _make_asset( def _ensure_missing_tag(session: Session): """Ensure the 'missing' tag exists.""" if not session.get(Tag, "missing"): - session.add(Tag(name="missing", tag_type="system")) + session.add(Tag(name="missing")) session.flush() diff --git a/tests-unit/assets_test/test_tags_api.py b/tests-unit/assets_test/test_tags_api.py index 595bf29c6..9729b7d03 100644 --- a/tests-unit/assets_test/test_tags_api.py +++ b/tests-unit/assets_test/test_tags_api.py @@ -69,8 +69,8 @@ def test_tags_empty_usage(http: requests.Session, api_base: str, asset_factory, used_names = [t["name"] for t in body2["tags"]] assert custom_tag in used_names - # Hard-delete the asset so the tag usage drops to zero - rd = http.delete(f"{api_base}/api/assets/{_asset['id']}?delete_content=true", timeout=120) + # Delete the asset reference so the tag usage drops to zero + rd = http.delete(f"{api_base}/api/assets/{_asset['id']}", timeout=120) assert rd.status_code == 204 # Now the custom tag must not be returned when include_zero=false diff --git a/tests-unit/assets_test/test_uploads.py b/tests-unit/assets_test/test_uploads.py index 0f2b124a3..427a417cc 100644 --- a/tests-unit/assets_test/test_uploads.py +++ b/tests-unit/assets_test/test_uploads.py @@ -5,6 +5,20 @@ from concurrent.futures import ThreadPoolExecutor import requests import pytest +from app.assets.api.schemas_out import Asset, AssetCreated + + +def test_asset_created_inherits_hash_field(): + """AssetCreated must inherit `hash` from Asset so POST /api/assets responses emit it. + + Schema-level guard: integration tests cover the wire shape, but inheritance + drift (e.g. AssetCreated ever being redefined to no longer extend Asset) + would silently drop `hash` from a major endpoint without this check. + """ + assert "hash" in Asset.model_fields + assert "hash" in AssetCreated.model_fields + assert AssetCreated.model_fields["hash"].annotation == Asset.model_fields["hash"].annotation + def test_upload_ok_duplicate_reference(http: requests.Session, api_base: str, make_asset_bytes): name = "dup_a.safetensors" @@ -17,6 +31,7 @@ def test_upload_ok_duplicate_reference(http: requests.Session, api_base: str, ma a1 = r1.json() assert r1.status_code == 201, a1 assert a1["created_new"] is True + assert a1["hash"] == a1["asset_hash"] # Second upload with the same data and name creates a new AssetReference (duplicates allowed) # Returns 200 because Asset already exists, but a new AssetReference is created @@ -26,6 +41,7 @@ def test_upload_ok_duplicate_reference(http: requests.Session, api_base: str, ma a2 = r2.json() assert r2.status_code in (200, 201), a2 assert a2["asset_hash"] == a1["asset_hash"] + assert a2["hash"] == a1["hash"] assert a2["id"] != a1["id"] # new reference with same content # Third upload with the same data but different name also creates new AssetReference @@ -50,6 +66,7 @@ def test_upload_fastpath_from_existing_hash_no_file(http: requests.Session, api_ b1 = r1.json() assert r1.status_code == 201, b1 h = b1["asset_hash"] + assert b1["hash"] == h # Now POST /api/assets with only hash and no file files = [ @@ -63,6 +80,7 @@ def test_upload_fastpath_from_existing_hash_no_file(http: requests.Session, api_ assert r2.status_code == 200, b2 # fast path returns 200 with created_new == False assert b2["created_new"] is False assert b2["asset_hash"] == h + assert b2["hash"] == h def test_upload_fastpath_with_known_hash_and_file( @@ -75,6 +93,7 @@ def test_upload_fastpath_with_known_hash_and_file( b1 = r1.json() assert r1.status_code == 201, b1 h = b1["asset_hash"] + assert b1["hash"] == h # Send both file and hash of existing content -> server must drain file and create from hash (200) files = {"file": ("ignored.bin", b"ignored" * 10, "application/octet-stream")} @@ -84,6 +103,7 @@ def test_upload_fastpath_with_known_hash_and_file( assert r2.status_code == 200, b2 assert b2["created_new"] is False assert b2["asset_hash"] == h + assert b2["hash"] == h def test_upload_multiple_tags_fields_are_merged(http: requests.Session, api_base: str): @@ -142,6 +162,8 @@ def test_concurrent_upload_identical_bytes_different_names( assert r1.status_code in (200, 201), b1 assert r2.status_code in (200, 201), b2 assert b1["asset_hash"] == b2["asset_hash"] + assert b1["hash"] == b2["hash"] + assert b1["hash"] == b1["asset_hash"] assert b1["id"] != b2["id"] created_flags = sorted([bool(b1.get("created_new")), bool(b2.get("created_new"))]) diff --git a/tests-unit/comfy_extras_test/nodes_math_test.py b/tests-unit/comfy_extras_test/nodes_math_test.py index 714e37c32..030accc5e 100644 --- a/tests-unit/comfy_extras_test/nodes_math_test.py +++ b/tests-unit/comfy_extras_test/nodes_math_test.py @@ -197,3 +197,10 @@ class TestMathExpressionExecute: def test_pow_huge_exponent_raises(self): with pytest.raises(ValueError, match="Exponent .* exceeds maximum"): self._exec("pow(a, b)", a=10, b=10000000) + + def test_huge_int_result_raises_value_error(self): + # Exponent is within the allowed MAX_EXPONENT range, so the result is a + # finite Python int that is nonetheless too large to convert to float. + # This must raise a clean ValueError, not an uncaught OverflowError. + with pytest.raises(ValueError, match="too large to represent as a float"): + self._exec("2 ** 3999")