diff --git a/.ci/windows_amd_base_files/run_amd_gpu_disable_smart_memory.bat b/.ci/windows_amd_base_files/run_amd_gpu_enable_dynamic_vram.bat similarity index 66% rename from .ci/windows_amd_base_files/run_amd_gpu_disable_smart_memory.bat rename to .ci/windows_amd_base_files/run_amd_gpu_enable_dynamic_vram.bat index cece0aeb2..94ad31942 100755 --- a/.ci/windows_amd_base_files/run_amd_gpu_disable_smart_memory.bat +++ b/.ci/windows_amd_base_files/run_amd_gpu_enable_dynamic_vram.bat @@ -1,2 +1,2 @@ -.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --disable-smart-memory +.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --enable-dynamic-vram pause diff --git a/.github/workflows/openapi-lint.yml b/.github/workflows/openapi-lint.yml new file mode 100644 index 000000000..be949de2a --- /dev/null +++ b/.github/workflows/openapi-lint.yml @@ -0,0 +1,31 @@ +name: OpenAPI Lint + +on: + pull_request: + paths: + - 'openapi.yaml' + - '.spectral.yaml' + - '.github/workflows/openapi-lint.yml' + +permissions: + contents: read + +jobs: + spectral: + name: Run Spectral + runs-on: ubuntu-latest + + steps: + - name: Checkout repository + uses: actions/checkout@v4 + + - name: Set up Node.js + uses: actions/setup-node@v4 + with: + node-version: '20' + + - name: Install Spectral + run: npm install -g @stoplight/spectral-cli@6 + + - name: Lint openapi.yaml + run: spectral lint openapi.yaml --ruleset .spectral.yaml --fail-severity=error diff --git a/.github/workflows/stable-release.yml b/.github/workflows/stable-release.yml index f501b7b31..bc64ed74d 100644 --- a/.github/workflows/stable-release.yml +++ b/.github/workflows/stable-release.yml @@ -145,6 +145,8 @@ jobs: cp -r ComfyUI/.ci/windows_${{ inputs.rel_name }}_base_files/* ./ cp ../update_comfyui_and_python_dependencies.bat ./update/ + echo 'local-portable' > ComfyUI/.comfy_environment + cd .. "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=9 -mfb=128 -md=768m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable diff --git a/.gitignore b/.gitignore index 0ab4ba75e..fc426eda4 100644 --- a/.gitignore +++ b/.gitignore @@ -23,3 +23,4 @@ web_custom_versions/ .DS_Store filtered-openapi.yaml uv.lock +.comfy_environment diff --git a/.spectral.yaml b/.spectral.yaml new file mode 100644 index 000000000..a4b137628 --- /dev/null +++ b/.spectral.yaml @@ -0,0 +1,100 @@ +extends: + - spectral:oas + +# Severity levels: error, warn, info, hint, off +# Rules from the built-in "spectral:oas" ruleset are active by default. +# Below we tune severity and add custom rules for our conventions. +# +# This ruleset mirrors Comfy-Org/cloud/.spectral.yaml so specs across the +# organization are linted against a single consistent standard. + +rules: + # ----------------------------------------------------------------------- + # Built-in rule severity overrides + # ----------------------------------------------------------------------- + operation-operationId: error + operation-description: warn + operation-tag-defined: error + info-contact: off + info-description: warn + no-eval-in-markdown: error + no-$ref-siblings: error + + # ----------------------------------------------------------------------- + # Custom rules: naming conventions + # ----------------------------------------------------------------------- + + # Property names should be snake_case + property-name-snake-case: + description: Property names must be snake_case + severity: warn + given: "$.components.schemas.*.properties[*]~" + then: + function: pattern + functionOptions: + match: "^[a-z][a-z0-9]*(_[a-z0-9]+)*$" + + # Operation IDs should be camelCase + operation-id-camel-case: + description: Operation IDs must be camelCase + severity: warn + given: "$.paths.*.*.operationId" + then: + function: pattern + functionOptions: + match: "^[a-z][a-zA-Z0-9]*$" + + # ----------------------------------------------------------------------- + # Custom rules: response conventions + # ----------------------------------------------------------------------- + + # Error responses (4xx, 5xx) should use a consistent shape + error-response-schema: + description: Error responses should reference a standard error schema + severity: hint + given: "$.paths.*.*.responses[?(@property >= '400' && @property < '600')].content['application/json'].schema" + then: + field: "$ref" + function: truthy + + # All 2xx responses with JSON body should have a schema + response-schema-defined: + description: Success responses with JSON content should define a schema + severity: warn + given: "$.paths.*.*.responses[?(@property >= '200' && @property < '300')].content['application/json']" + then: + field: schema + function: truthy + + # ----------------------------------------------------------------------- + # Custom rules: best practices + # ----------------------------------------------------------------------- + + # Path parameters must have a description + path-param-description: + description: Path parameters should have a description + severity: warn + given: + - "$.paths.*.parameters[?(@.in == 'path')]" + - "$.paths.*.*.parameters[?(@.in == 'path')]" + then: + field: description + function: truthy + + # Schemas should have a description + schema-description: + description: Component schemas should have a description + severity: hint + given: "$.components.schemas.*" + then: + field: description + function: truthy + +overrides: + # /ws uses HTTP 101 (Switching Protocols) — a legitimate response for a + # WebSocket upgrade, but not a 2xx, so operation-success-response fires + # as a false positive. OpenAPI 3.x has no native WebSocket support. + - files: + - "openapi.yaml#/paths/~1ws" + rules: + operation-success-response: off diff --git a/CODEOWNERS b/CODEOWNERS index 4d5448636..946dbf946 100644 --- a/CODEOWNERS +++ b/CODEOWNERS @@ -1,2 +1,2 @@ # Admins -* @comfyanonymous @kosinkadink @guill +* @comfyanonymous @kosinkadink @guill @alexisrolland @rattus128 @kijai diff --git a/README.md b/README.md index f05311421..0eecd8a4b 100644 --- a/README.md +++ b/README.md @@ -1,7 +1,7 @@
# ComfyUI -**The most powerful and modular visual AI engine and application.** +**The most powerful and modular AI engine for content creation.** [![Website][website-shield]][website-url] @@ -31,10 +31,16 @@ [github-downloads-latest-shield]: https://img.shields.io/github/downloads/comfyanonymous/ComfyUI/latest/total?style=flat&label=downloads%40latest [github-downloads-link]: https://github.com/comfyanonymous/ComfyUI/releases -![ComfyUI Screenshot](https://github.com/user-attachments/assets/7ccaf2c1-9b72-41ae-9a89-5688c94b7abe) +ComfyUI Screenshot +
-ComfyUI lets you design and execute advanced stable diffusion pipelines using a graph/nodes/flowchart based interface. Available on Windows, Linux, and macOS. +ComfyUI is the AI creation engine for visual professionals who demand control over every model, every parameter, and every output. Its powerful and modular node graph interface empowers creatives to generate images, videos, 3D models, audio, and more... +- ComfyUI natively supports the latest open-source state of the art models. +- API nodes provide access to the best closed source models such as Nano Banana, Seedance, Hunyuan3D, etc. +- It is available on Windows, Linux, and macOS, locally with our [desktop application](https://www.comfy.org/download), our [portable install](#installing) or on our [cloud](https://www.comfy.org/cloud). +- The most sophisticated workflows can be exposed through a simple UI thanks to App Mode. +- It integrates seamlessly into production pipelines with our API endpoints. ## Get Started @@ -77,6 +83,7 @@ See what ComfyUI can do with the [newer template workflows](https://comfy.org/wo - [Hunyuan Image 2.1](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_image/) - [Flux 2](https://comfyanonymous.github.io/ComfyUI_examples/flux2/) - [Z Image](https://comfyanonymous.github.io/ComfyUI_examples/z_image/) + - Ernie Image - Image Editing Models - [Omnigen 2](https://comfyanonymous.github.io/ComfyUI_examples/omnigen/) - [Flux Kontext](https://comfyanonymous.github.io/ComfyUI_examples/flux/#flux-kontext-image-editing-model) @@ -126,7 +133,7 @@ Workflow examples can be found on the [Examples page](https://comfyanonymous.git ComfyUI follows a weekly release cycle targeting Monday but this regularly changes because of model releases or large changes to the codebase. There are three interconnected repositories: 1. **[ComfyUI Core](https://github.com/comfyanonymous/ComfyUI)** - - Releases a new stable version (e.g., v0.7.0) roughly every week. + - Releases a new major stable version (e.g., v0.7.0) roughly every 2 weeks. - Starting from v0.4.0 patch versions will be used for fixes backported onto the current stable release. - Minor versions will be used for releases off the master branch. - Patch versions may still be used for releases on the master branch in cases where a backport would not make sense. @@ -193,13 +200,15 @@ If you have trouble extracting it, right click the file -> properties -> unblock The portable above currently comes with python 3.13 and pytorch cuda 13.0. Update your Nvidia drivers if it doesn't start. -#### Alternative Downloads: +#### All Official Portable Downloads: [Portable for AMD GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_amd.7z) -[Experimental portable for Intel GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_intel.7z) +[Portable for Intel GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_intel.7z) -[Portable with pytorch cuda 12.6 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu126.7z) (Supports Nvidia 10 series and older GPUs). +[Portable for Nvidia GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia.7z) (supports 20 series and above). + +[Portable for Nvidia GPUs with pytorch cuda 12.6 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu126.7z) (Supports Nvidia 10 series and older GPUs). #### How do I share models between another UI and ComfyUI? @@ -420,6 +429,8 @@ Use `--tls-keyfile key.pem --tls-certfile cert.pem` to enable TLS/SSL, the app w See also: [https://www.comfy.org/](https://www.comfy.org/) +> _psst — we're hiring!_ Help build ComfyUI: [comfy.org/careers](https://www.comfy.org/careers) + ## 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. diff --git a/SECURITY.md b/SECURITY.md new file mode 100644 index 000000000..299b0067b --- /dev/null +++ b/SECURITY.md @@ -0,0 +1,44 @@ +# Security Policy + +## Scope + +ComfyUI is designed to run locally. By default, the server binds to `127.0.0.1`, meaning only the user's own machine can reach it. Our threat model assumes: + +- The user installed ComfyUI through a supported channel: the desktop application, the portable build, or a manual install following the README. +- The user has not installed untrusted custom nodes. Custom nodes are arbitrary Python code and are trusted as much as any other software the user chooses to install. +- Anyone with access to the ComfyUI URL is trusted (a direct consequence of the localhost-only default). +- PyTorch and other dependencies are at the versions we ship or recommend in the README. + +A report is in scope only if it affects a user operating within this threat model. + +## What We Consider a Vulnerability + +We want to hear about issues where a **reasonable user** — someone who does not install random untrusted nodes and who reads UI prompts and warnings before clicking through them — can be harmed by ComfyUI itself. + +The clearest example: a workflow file that such a user might plausibly load and run, using only built-in nodes, that results in **untrusted code execution, arbitrary file read/write outside expected directories, or credential/data exfiltration**. + +When submitting a report, please include a clear description of *why this is a problem for a typical local ComfyUI user*. Reports without this context are difficult to act on. + +## What We Do Not Consider a Security Vulnerability + +Please report the following through our regular [GitHub issues](https://github.com/comfyanonymous/ComfyUI/issues) instead. Filing them as security reports will likely cause them to be deprioritized or closed. + +- **Issues requiring `--listen` or any non-default network exposure.** ComfyUI binds to localhost by default. If a remote attacker needs to reach the server for the attack to work, the user has chosen to expose it and is responsible for securing that deployment (firewall, reverse proxy, authentication, etc.). These are bugs, not vulnerabilities. +- **`torch.load` and related deserialization issues in old PyTorch versions.** These are upstream PyTorch issues. Our distributions ship with — and our documentation recommends — recent PyTorch versions where these are addressed. +- **Vulnerabilities that depend on outdated library versions** that we neither ship nor recommend (e.g., requiring PyTorch 2.6 or older). +- **Issues that require a specific custom node to be installed.** Custom nodes are third-party code. Report these to the maintainer of that node. +- **Crashes, hangs, or resource exhaustion from a loaded workflow.** Annoying, but not a security issue in our model. File a regular bug. +- **Social-engineering scenarios** where the user is expected to ignore an explicit UI warning or prompt. + +## Reporting + +If you believe you have found an issue that falls within the scope above, please report it privately via GitHub's [Report a vulnerability](https://github.com/comfyanonymous/ComfyUI/security/advisories/new) feature rather than opening a public issue. + +Please include: + +1. A description of the vulnerability and the affected component. +2. Reproduction steps, ideally with a minimal workflow file or proof-of-concept. +3. The ComfyUI version, install method (desktop / portable / manual), and OS. +4. An explanation of how this affects a typical local user as described in the threat model. + +We will acknowledge valid reports and coordinate a fix and disclosure timeline with you. diff --git a/app/frontend_management.py b/app/frontend_management.py index f753ef0de..d0596b276 100644 --- a/app/frontend_management.py +++ b/app/frontend_management.py @@ -27,7 +27,7 @@ def frontend_install_warning_message(): return f""" {get_missing_requirements_message()} -This error is happening because the ComfyUI frontend is no longer shipped as part of the main repo but as a pip package instead. +The ComfyUI frontend is shipped in a pip package so it needs to be updated separately from the ComfyUI code. """.strip() def parse_version(version: str) -> tuple[int, int, int]: @@ -38,40 +38,54 @@ def is_valid_version(version: str) -> bool: pattern = r"^(\d+)\.(\d+)\.(\d+)$" return bool(re.match(pattern, version)) -def get_installed_frontend_version(): - """Get the currently installed frontend package version.""" - frontend_version_str = version("comfyui-frontend-package") - return frontend_version_str - - def get_required_frontend_version(): return get_required_packages_versions().get("comfyui-frontend-package", None) -def check_frontend_version(): - """Check if the frontend version is up to date.""" +COMFY_PACKAGE_VERSIONS = [] +def get_comfy_package_versions(): + """List installed/required versions for every comfy* package in requirements.txt.""" + if COMFY_PACKAGE_VERSIONS: + return COMFY_PACKAGE_VERSIONS.copy() + out = COMFY_PACKAGE_VERSIONS + for name, required in (get_required_packages_versions() or {}).items(): + if not name.startswith("comfy"): + continue + try: + installed = version(name) + except Exception: + installed = None + out.append({"name": name, "installed": installed, "required": required}) + return out.copy() - try: - frontend_version_str = get_installed_frontend_version() - frontend_version = parse_version(frontend_version_str) - required_frontend_str = get_required_frontend_version() - required_frontend = parse_version(required_frontend_str) - if frontend_version < required_frontend: + +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 + for pkg in get_comfy_package_versions(): + installed_str = pkg["installed"] + required_str = pkg["required"] + if not installed_str or not required_str: + continue + try: + outdated = parse_pep440(installed_str) < parse_pep440(required_str) + except InvalidVersion as e: + logging.error(f"Failed to check {pkg['name']} version: {e}") + continue + if outdated: app.logger.log_startup_warning( f""" ________________________________________________________________________ WARNING WARNING WARNING WARNING WARNING -Installed frontend version {".".join(map(str, frontend_version))} is lower than the recommended version {".".join(map(str, required_frontend))}. +Installed {pkg["name"]} version {installed_str} is lower than the recommended version {required_str}. -{frontend_install_warning_message()} +{get_missing_requirements_message()} ________________________________________________________________________ """.strip() ) else: - logging.info("ComfyUI frontend version: {}".format(frontend_version_str)) - except Exception as e: - logging.error(f"Failed to check frontend version: {e}") + logging.info("{} version: {}".format(pkg["name"], installed_str)) REQUEST_TIMEOUT = 10 # seconds @@ -201,6 +215,11 @@ class FrontendManager: def get_required_templates_version(cls) -> str: return get_required_packages_versions().get("comfyui-workflow-templates", None) + @classmethod + def get_comfy_package_versions(cls): + """List installed/required versions for every comfy* package in requirements.txt.""" + return get_comfy_package_versions() + @classmethod def default_frontend_path(cls) -> str: try: @@ -341,7 +360,7 @@ comfyui-workflow-templates is not installed. main error source might be request timeout or invalid URL. """ if version_string == DEFAULT_VERSION_STRING: - check_frontend_version() + check_comfy_packages_versions() return cls.default_frontend_path() repo_owner, repo_name, version = cls.parse_version_string(version_string) @@ -403,7 +422,7 @@ comfyui-workflow-templates is not installed. except Exception as e: logging.error("Failed to initialize frontend: %s", e) logging.info("Falling back to the default frontend.") - check_frontend_version() + check_comfy_packages_versions() return cls.default_frontend_path() @classmethod def template_asset_handler(cls): diff --git a/app/node_replace_manager.py b/app/node_replace_manager.py index d9aab5b22..72e8ac2b1 100644 --- a/app/node_replace_manager.py +++ b/app/node_replace_manager.py @@ -1,5 +1,7 @@ from __future__ import annotations +import logging + from aiohttp import web from typing import TYPE_CHECKING, TypedDict @@ -31,8 +33,22 @@ class NodeReplaceManager: self._replacements: dict[str, list[NodeReplace]] = {} def register(self, node_replace: NodeReplace): - """Register a node replacement mapping.""" - self._replacements.setdefault(node_replace.old_node_id, []).append(node_replace) + """Register a node replacement mapping. + + Idempotent: if a replacement with the same (old_node_id, new_node_id) + is already registered, the duplicate is ignored. This prevents stale + entries from accumulating when custom nodes are reloaded in the same + process (e.g. via ComfyUI-Manager). + """ + existing = self._replacements.setdefault(node_replace.old_node_id, []) + for entry in existing: + if entry.new_node_id == node_replace.new_node_id: + logging.debug( + "Node replacement %s -> %s already registered, ignoring duplicate.", + node_replace.old_node_id, node_replace.new_node_id, + ) + return + existing.append(node_replace) def get_replacement(self, old_node_id: str) -> list[NodeReplace] | None: """Get replacements for an old node ID.""" diff --git a/app/user_manager.py b/app/user_manager.py index e18afb71b..0517b3344 100644 --- a/app/user_manager.py +++ b/app/user_manager.py @@ -28,8 +28,8 @@ def get_file_info(path: str, relative_to: str) -> FileInfo: return { "path": os.path.relpath(path, relative_to).replace(os.sep, '/'), "size": os.path.getsize(path), - "modified": os.path.getmtime(path), - "created": os.path.getctime(path) + "modified": int(os.path.getmtime(path) * 1000), + "created": int(os.path.getctime(path) * 1000), } diff --git a/blueprints/Brightness and Contrast.json b/blueprints/Brightness and Contrast.json index 90bfe999d..78fc52f29 100644 --- a/blueprints/Brightness and Contrast.json +++ b/blueprints/Brightness and Contrast.json @@ -431,9 +431,10 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image Tools/Color adjust" + "category": "Image Tools/Color adjust", + "description": "Adjusts image brightness and contrast using a real-time GPU fragment shader." } ] }, "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 ff9717308..14deb64cc 100644 --- a/blueprints/Canny to Image (Z-Image-Turbo).json +++ b/blueprints/Canny to Image (Z-Image-Turbo).json @@ -162,7 +162,7 @@ }, "revision": 0, "config": {}, - "name": "local-Canny to Image (Z-Image-Turbo)", + "name": "Canny to Image (Z-Image-Turbo)", "inputNode": { "id": -10, "bounding": [ @@ -1553,7 +1553,8 @@ "VHS_MetadataImage": true, "VHS_KeepIntermediate": true }, - "category": "Image generation and editing/Canny to image" + "category": "Image generation and editing/Canny to image", + "description": "Generates an image from a Canny edge map using Z-Image-Turbo, with text conditioning." } ] }, @@ -1574,4 +1575,4 @@ } }, "version": 0.4 -} +} \ No newline at end of file diff --git a/blueprints/Canny to Video (LTX 2.0).json b/blueprints/Canny to Video (LTX 2.0).json index fae8321b9..a9682c8a4 100644 --- a/blueprints/Canny to Video (LTX 2.0).json +++ b/blueprints/Canny to Video (LTX 2.0).json @@ -192,7 +192,7 @@ }, "revision": 0, "config": {}, - "name": "local-Canny to Video (LTX 2.0)", + "name": "Canny to Video (LTX 2.0)", "inputNode": { "id": -10, "bounding": [ @@ -3600,7 +3600,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Video generation and editing/Canny to video" + "category": "Video generation and editing/Canny to video", + "description": "Generates video from Canny edge maps using LTX-2, with optional synchronized audio." } ] }, @@ -3616,4 +3617,4 @@ } }, "version": 0.4 -} +} \ No newline at end of file diff --git a/blueprints/Chromatic Aberration.json b/blueprints/Chromatic Aberration.json index ae8037b1b..893fb1190 100644 --- a/blueprints/Chromatic Aberration.json +++ b/blueprints/Chromatic Aberration.json @@ -377,8 +377,9 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image Tools/Color adjust" + "category": "Image Tools/Color adjust", + "description": "Adds lens-style chromatic aberration (color fringing) using a real-time GPU fragment shader." } ] } -} +} \ No newline at end of file diff --git a/blueprints/Color Adjustment.json b/blueprints/Color Adjustment.json index 622bf28af..5abbf8baa 100644 --- a/blueprints/Color Adjustment.json +++ b/blueprints/Color Adjustment.json @@ -596,7 +596,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image Tools/Color adjust" + "category": "Image Tools/Color adjust", + "description": "Adjusts saturation, temperature, tint, and vibrance using a real-time GPU fragment shader." } ] } diff --git a/blueprints/Color Balance.json b/blueprints/Color Balance.json index 21d6319ed..d921eab37 100644 --- a/blueprints/Color Balance.json +++ b/blueprints/Color Balance.json @@ -1129,7 +1129,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image Tools/Color adjust" + "category": "Image Tools/Color adjust", + "description": "Balances colors across shadows, midtones, and highlights using a real-time GPU fragment shader." } ] } diff --git a/blueprints/Color Curves.json b/blueprints/Color Curves.json index 1461cf396..b9bfb7029 100644 --- a/blueprints/Color Curves.json +++ b/blueprints/Color Curves.json @@ -608,7 +608,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image Tools/Color adjust" + "category": "Image Tools/Color adjust", + "description": "Fine-tunes tone and color with per-channel curve adjustments using a real-time GPU fragment shader." } ] } diff --git a/blueprints/ControlNet (Z-Image-Turbo).json b/blueprints/ControlNet (Z-Image-Turbo).json new file mode 100644 index 000000000..fbec95a97 --- /dev/null +++ b/blueprints/ControlNet (Z-Image-Turbo).json @@ -0,0 +1,1412 @@ +{ + "revision": 0, + "last_node_id": 85, + "last_link_id": 0, + "nodes": [ + { + "id": 85, + "type": "d2e76ecf-6e84-4b8c-8913-48efc09ec1c4", + "pos": [ + 440, + 1220 + ], + "size": [ + 480, + 0 + ], + "flags": {}, + "order": 6, + "mode": 0, + "inputs": [ + { + "label": "control_image", + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": null + }, + { + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "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 + }, + { + "label": "patch_model", + "name": "name", + "type": "COMBO", + "widget": { + "name": "name" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [] + } + ], + "title": "ControlNet (Z-Image-Turbo)", + "properties": { + "proxyWidgets": [ + [ + "83", + "text" + ], + [ + "79", + "seed" + ], + [ + "74", + "unet_name" + ], + [ + "73", + "clip_name" + ], + [ + "75", + "vae_name" + ], + [ + "76", + "name" + ], + [ + "79", + "control_after_generate" + ] + ], + "cnr_id": "comfy-core", + "ver": "0.18.1", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [] + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "d2e76ecf-6e84-4b8c-8913-48efc09ec1c4", + "version": 1, + "state": { + "lastGroupId": 9, + "lastNodeId": 85, + "lastLinkId": 87, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "ControlNet (Z-Image-Turbo)", + "inputNode": { + "id": -10, + "bounding": [ + -500, + 620, + 120, + 180 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + 1390, + 1100, + 120, + 60 + ] + }, + "inputs": [ + { + "id": "fbbb968e-d3cf-40e4-b3ce-7abb074e5bd8", + "name": "image", + "type": "IMAGE", + "linkIds": [ + 65, + 80 + ], + "localized_name": "image", + "label": "control_image", + "pos": [ + -400, + 640 + ] + }, + { + "id": "c1b19877-5417-4580-aea1-44439c70c1dd", + "name": "text", + "type": "STRING", + "linkIds": [ + 81 + ], + "pos": [ + -400, + 660 + ] + }, + { + "id": "b5671515-bc7a-4be5-b1e7-d4f0f68907d6", + "name": "seed", + "type": "INT", + "linkIds": [ + 83 + ], + "pos": [ + -400, + 680 + ] + }, + { + "id": "2838be23-8034-4f16-87a5-d29d790e8391", + "name": "unet_name", + "type": "COMBO", + "linkIds": [ + 84 + ], + "pos": [ + -400, + 700 + ] + }, + { + "id": "8a6643b5-8f78-41ff-bbc6-e87b95459706", + "name": "clip_name", + "type": "COMBO", + "linkIds": [ + 85 + ], + "pos": [ + -400, + 720 + ] + }, + { + "id": "b103dc94-8ca7-456b-a809-414d7e341a1b", + "name": "vae_name", + "type": "COMBO", + "linkIds": [ + 86 + ], + "pos": [ + -400, + 740 + ] + }, + { + "id": "4a7d65af-f0fd-4a5c-832a-bdc0d15b1f30", + "name": "name", + "type": "COMBO", + "linkIds": [ + 87 + ], + "label": "patch_model", + "pos": [ + -400, + 760 + ] + } + ], + "outputs": [ + { + "id": "ccb7fa39-4a3d-4eb2-8fd2-91d08fad9570", + "name": "IMAGE", + "type": "IMAGE", + "linkIds": [ + 45 + ], + "localized_name": "IMAGE", + "pos": [ + 1410, + 1120 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 73, + "type": "CLIPLoader", + "pos": [ + 20, + 500 + ], + "size": [ + 270, + 150 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "clip_name", + "name": "clip_name", + "type": "COMBO", + "widget": { + "name": "clip_name" + }, + "link": 85 + }, + { + "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": [ + 44 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.73", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "CLIPLoader", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "models": [ + { + "name": "qwen_3_4b.safetensors", + "url": "https://huggingface.co/Comfy-Org/z_image_turbo/resolve/main/split_files/text_encoders/qwen_3_4b.safetensors", + "directory": "text_encoders" + } + ] + }, + "widgets_values": [ + "qwen_3_4b.safetensors", + "lumina2", + "default" + ] + }, + { + "id": 74, + "type": "UNETLoader", + "pos": [ + 20, + 320 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "unet_name", + "name": "unet_name", + "type": "COMBO", + "widget": { + "name": "unet_name" + }, + "link": 84 + }, + { + "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": { + "cnr_id": "comfy-core", + "ver": "0.3.73", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "UNETLoader", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "models": [ + { + "name": "z_image_turbo_bf16.safetensors", + "url": "https://huggingface.co/Comfy-Org/z_image_turbo/resolve/main/split_files/diffusion_models/z_image_turbo_bf16.safetensors", + "directory": "diffusion_models" + } + ] + }, + "widgets_values": [ + "z_image_turbo_bf16.safetensors", + "default" + ] + }, + { + "id": 75, + "type": "VAELoader", + "pos": [ + 20, + 760 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "vae_name", + "name": "vae_name", + "type": "COMBO", + "widget": { + "name": "vae_name" + }, + "link": 86 + } + ], + "outputs": [ + { + "localized_name": "VAE", + "name": "VAE", + "type": "VAE", + "links": [ + 39, + 70 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.73", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "VAELoader", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "models": [ + { + "name": "ae.safetensors", + "url": "https://huggingface.co/Comfy-Org/z_image_turbo/resolve/main/split_files/vae/ae.safetensors", + "directory": "vae" + } + ] + }, + "widgets_values": [ + "ae.safetensors" + ] + }, + { + "id": 76, + "type": "ModelPatchLoader", + "pos": [ + 20, + 940 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "name", + "name": "name", + "type": "COMBO", + "widget": { + "name": "name" + }, + "link": 87 + } + ], + "outputs": [ + { + "localized_name": "MODEL_PATCH", + "name": "MODEL_PATCH", + "type": "MODEL_PATCH", + "links": [ + 74 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.51", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "ModelPatchLoader", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "models": [ + { + "name": "Z-Image-Turbo-Fun-Controlnet-Union.safetensors", + "url": "https://huggingface.co/alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union/resolve/main/Z-Image-Turbo-Fun-Controlnet-Union.safetensors", + "directory": "model_patches" + } + ] + }, + "widgets_values": [ + "Z-Image-Turbo-Fun-Controlnet-Union.safetensors" + ] + }, + { + "id": 77, + "type": "VAEDecode", + "pos": [ + 940, + 1100 + ], + "size": [ + 230, + 100 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "localized_name": "samples", + "name": "samples", + "type": "LATENT", + "link": 38 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 39 + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "slot_index": 0, + "links": [ + 45 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.64", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "VAEDecode", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 78, + "type": "ModelSamplingAuraFlow", + "pos": [ + 910, + 270 + ], + "size": [ + 290, + 110 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 69 + }, + { + "localized_name": "shift", + "name": "shift", + "type": "FLOAT", + "widget": { + "name": "shift" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "slot_index": 0, + "links": [ + 40 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.64", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "ModelSamplingAuraFlow", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 3 + ] + }, + { + "id": 79, + "type": "KSampler", + "pos": [ + 910, + 430 + ], + "size": [ + 300, + 570 + ], + "flags": {}, + "order": 6, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 40 + }, + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "link": 41 + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "link": 42 + }, + { + "localized_name": "latent_image", + "name": "latent_image", + "type": "LATENT", + "link": 78 + }, + { + "localized_name": "seed", + "name": "seed", + "type": "INT", + "widget": { + "name": "seed" + }, + "link": 83 + }, + { + "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": [ + 38 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.64", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "KSampler", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 729703840979498, + "randomize", + 8, + 1, + "res_multistep", + "simple", + 1 + ] + }, + { + "id": 80, + "type": "ConditioningZeroOut", + "pos": [ + 610, + 830 + ], + "size": [ + 230, + 80 + ], + "flags": { + "collapsed": true + }, + "order": 7, + "mode": 0, + "inputs": [ + { + "localized_name": "conditioning", + "name": "conditioning", + "type": "CONDITIONING", + "link": 36 + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "links": [ + 42 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.73", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "ConditioningZeroOut", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 81, + "type": "QwenImageDiffsynthControlnet", + "pos": [ + 490, + 970 + ], + "size": [ + 290, + 200 + ], + "flags": {}, + "order": 8, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 79 + }, + { + "localized_name": "model_patch", + "name": "model_patch", + "type": "MODEL_PATCH", + "link": 74 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 70 + }, + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 65 + }, + { + "localized_name": "mask", + "name": "mask", + "shape": 7, + "type": "MASK", + "link": null + }, + { + "localized_name": "strength", + "name": "strength", + "type": "FLOAT", + "widget": { + "name": "strength" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "links": [ + 69 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.76", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "QwenImageDiffsynthControlnet", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 1 + ] + }, + { + "id": 82, + "type": "EmptySD3LatentImage", + "pos": [ + 40, + 1200 + ], + "size": [ + 260, + 170 + ], + "flags": {}, + "order": 9, + "mode": 0, + "inputs": [ + { + "localized_name": "width", + "name": "width", + "type": "INT", + "widget": { + "name": "width" + }, + "link": 76 + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "widget": { + "name": "height" + }, + "link": 77 + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "widget": { + "name": "batch_size" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "LATENT", + "name": "LATENT", + "type": "LATENT", + "slot_index": 0, + "links": [ + 78 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.64", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "EmptySD3LatentImage", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 1024, + 1024, + 1 + ] + }, + { + "id": 83, + "type": "CLIPTextEncode", + "pos": [ + 430, + 310 + ], + "size": [ + 400, + 440 + ], + "flags": {}, + "order": 10, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 44 + }, + { + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": 81 + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "links": [ + 36, + 41 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.73", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "CLIPTextEncode", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "" + ], + "color": "#232", + "bgcolor": "#353" + }, + { + "id": 84, + "type": "GetImageSize", + "pos": [ + 50, + 1410 + ], + "size": [ + 230, + 120 + ], + "flags": { + "collapsed": true + }, + "order": 11, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 80 + } + ], + "outputs": [ + { + "localized_name": "width", + "name": "width", + "type": "INT", + "links": [ + 76 + ] + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "links": [ + 77 + ] + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "links": null + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.76", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "GetImageSize", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + } + ], + "groups": [ + { + "id": 3, + "title": "Prompt", + "bounding": [ + 410, + 230, + 440, + 630 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 4, + "title": "Model", + "bounding": [ + -50, + 230, + 430, + 840 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 8, + "title": "Apple ControlNet", + "bounding": [ + 410, + 890, + 440, + 330 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 9, + "title": "Image Size", + "bounding": [ + -50, + 1100, + 430, + 350 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + } + ], + "links": [ + { + "id": 38, + "origin_id": 79, + "origin_slot": 0, + "target_id": 77, + "target_slot": 0, + "type": "LATENT" + }, + { + "id": 39, + "origin_id": 75, + "origin_slot": 0, + "target_id": 77, + "target_slot": 1, + "type": "VAE" + }, + { + "id": 69, + "origin_id": 81, + "origin_slot": 0, + "target_id": 78, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 40, + "origin_id": 78, + "origin_slot": 0, + "target_id": 79, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 41, + "origin_id": 83, + "origin_slot": 0, + "target_id": 79, + "target_slot": 1, + "type": "CONDITIONING" + }, + { + "id": 42, + "origin_id": 80, + "origin_slot": 0, + "target_id": 79, + "target_slot": 2, + "type": "CONDITIONING" + }, + { + "id": 78, + "origin_id": 82, + "origin_slot": 0, + "target_id": 79, + "target_slot": 3, + "type": "LATENT" + }, + { + "id": 36, + "origin_id": 83, + "origin_slot": 0, + "target_id": 80, + "target_slot": 0, + "type": "CONDITIONING" + }, + { + "id": 79, + "origin_id": 74, + "origin_slot": 0, + "target_id": 81, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 74, + "origin_id": 76, + "origin_slot": 0, + "target_id": 81, + "target_slot": 1, + "type": "MODEL_PATCH" + }, + { + "id": 70, + "origin_id": 75, + "origin_slot": 0, + "target_id": 81, + "target_slot": 2, + "type": "VAE" + }, + { + "id": 76, + "origin_id": 84, + "origin_slot": 0, + "target_id": 82, + "target_slot": 0, + "type": "INT" + }, + { + "id": 77, + "origin_id": 84, + "origin_slot": 1, + "target_id": 82, + "target_slot": 1, + "type": "INT" + }, + { + "id": 44, + "origin_id": 73, + "origin_slot": 0, + "target_id": 83, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 65, + "origin_id": -10, + "origin_slot": 0, + "target_id": 81, + "target_slot": 3, + "type": "IMAGE" + }, + { + "id": 80, + "origin_id": -10, + "origin_slot": 0, + "target_id": 84, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 45, + "origin_id": 77, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 81, + "origin_id": -10, + "origin_slot": 1, + "target_id": 83, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 83, + "origin_id": -10, + "origin_slot": 2, + "target_id": 79, + "target_slot": 4, + "type": "INT" + }, + { + "id": 84, + "origin_id": -10, + "origin_slot": 3, + "target_id": 74, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 85, + "origin_id": -10, + "origin_slot": 4, + "target_id": 73, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 86, + "origin_id": -10, + "origin_slot": 5, + "target_id": 75, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 87, + "origin_id": -10, + "origin_slot": 6, + "target_id": 76, + "target_slot": 0, + "type": "COMBO" + } + ], + "extra": { + "workflowRendererVersion": "LG" + }, + "category": "Image generation and editing/ControlNet", + "description": "Generates images from a text prompt and ControlNet conditioning (e.g. depth, canny) using Z-Image-Turbo." + } + ] + }, + "extra": { + "ue_links": [] + } +} \ No newline at end of file diff --git a/blueprints/Crop Images 2x2.json b/blueprints/Crop Images 2x2.json index 2aa42cfc3..99b89b608 100644 --- a/blueprints/Crop Images 2x2.json +++ b/blueprints/Crop Images 2x2.json @@ -1609,7 +1609,8 @@ } ], "extra": {}, - "category": "Image Tools/Crop" + "category": "Image Tools/Crop", + "description": "Splits an image into a 2×2 grid of four equal tiles." } ] }, diff --git a/blueprints/Crop Images 3x3.json b/blueprints/Crop Images 3x3.json index 3a3615ac8..6ac636da4 100644 --- a/blueprints/Crop Images 3x3.json +++ b/blueprints/Crop Images 3x3.json @@ -2946,7 +2946,8 @@ } ], "extra": {}, - "category": "Image Tools/Crop" + "category": "Image Tools/Crop", + "description": "Splits an image into a 3×3 grid of nine equal tiles." } ] }, diff --git a/blueprints/Depth to Image (Z-Image-Turbo).json b/blueprints/Depth to Image (Z-Image-Turbo).json index 4f69a8149..fe9ef0f72 100644 --- a/blueprints/Depth to Image (Z-Image-Turbo).json +++ b/blueprints/Depth to Image (Z-Image-Turbo).json @@ -1579,7 +1579,8 @@ "VHS_MetadataImage": true, "VHS_KeepIntermediate": true }, - "category": "Image generation and editing/Depth to image" + "category": "Image generation and editing/Depth to image", + "description": "Generates an image from a depth map using Z-Image-Turbo with text conditioning." }, { "id": "458bdf3c-4b58-421c-af50-c9c663a4d74c", @@ -2461,7 +2462,8 @@ ] }, "workflowRendererVersion": "LG" - } + }, + "description": "Estimates a monocular depth map from an input image using the Lotus depth estimation model." } ] }, diff --git a/blueprints/Depth to Video (ltx 2.0).json b/blueprints/Depth to Video (ltx 2.0).json index f15212520..bd51e4476 100644 --- a/blueprints/Depth to Video (ltx 2.0).json +++ b/blueprints/Depth to Video (ltx 2.0).json @@ -4233,7 +4233,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Video generation and editing/Depth to video" + "category": "Video generation and editing/Depth to video", + "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." }, { "id": "38b60539-50a7-42f9-a5fe-bdeca26272e2", @@ -5192,7 +5193,8 @@ ], "extra": { "workflowRendererVersion": "LG" - } + }, + "description": "Estimates a monocular depth map from an input image using the Lotus depth estimation model." } ] }, diff --git a/blueprints/Edge-Preserving Blur.json b/blueprints/Edge-Preserving Blur.json index 18012beb1..fbda9f126 100644 --- a/blueprints/Edge-Preserving Blur.json +++ b/blueprints/Edge-Preserving Blur.json @@ -450,9 +450,10 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image Tools/Blur" + "category": "Image Tools/Blur", + "description": "Applies bilateral (edge-preserving) blur to soften images while retaining detail." } ] }, "extra": {} -} +} \ No newline at end of file diff --git a/blueprints/Film Grain.json b/blueprints/Film Grain.json index a680b3ece..3226ea9aa 100644 --- a/blueprints/Film Grain.json +++ b/blueprints/Film Grain.json @@ -580,8 +580,9 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image Tools/Color adjust" + "category": "Image Tools/Color adjust", + "description": "Adds procedural film grain texture for a cinematic look via GPU fragment shader." } ] } -} +} \ No newline at end of file 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 8ec9ed61a..f509aefe0 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 +3350,8 @@ } ], "extra": {}, - "category": "Video generation and editing/First-Last-Frame to Video" + "category": "Video generation and editing/First-Last-Frame to Video", + "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 new file mode 100644 index 000000000..84dfafbcd --- /dev/null +++ b/blueprints/First-Last-Frame to Video.json @@ -0,0 +1,3361 @@ +{ + "revision": 0, + "last_node_id": 227, + "last_link_id": 0, + "nodes": [ + { + "id": 227, + "type": "283e4561-61a2-4538-b960-265736eb041f", + "pos": [ + 620, + 3140 + ], + "size": [ + 540, + 0 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "label": "first_frame", + "localized_name": "input", + "name": "input", + "type": "IMAGE,MASK", + "link": null + }, + { + "label": "last_frame", + "localized_name": "input_1", + "name": "input_1", + "type": "IMAGE,MASK", + "link": null + }, + { + "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": "duration", + "name": "value_2", + "type": "INT", + "widget": { + "name": "value_2" + }, + "link": null + }, + { + "label": "fps", + "name": "value_3", + "type": "INT", + "widget": { + "name": "value_3" + }, + "link": null + }, + { + "name": "noise_seed", + "type": "INT", + "widget": { + "name": "noise_seed" + }, + "link": null + }, + { + "label": "ckpt_name", + "name": "ckpt_name_1", + "type": "COMBO", + "widget": { + "name": "ckpt_name_1" + }, + "link": null + }, + { + "name": "text_encoder", + "type": "COMBO", + "widget": { + "name": "text_encoder" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "VIDEO", + "name": "VIDEO", + "type": "VIDEO", + "links": [] + } + ], + "title": "First-Last-Frame to Video", + "properties": { + "proxyWidgets": [ + [ + "222", + "text" + ], + [ + "215", + "value" + ], + [ + "216", + "value" + ], + [ + "198", + "value" + ], + [ + "205", + "value" + ], + [ + "196", + "noise_seed" + ], + [ + "224", + "ckpt_name" + ], + [ + "225", + "text_encoder" + ] + ], + "cnr_id": "comfy-core", + "ver": "0.18.1", + "ue_properties": { + "widget_ue_connectable": {}, + "input_ue_unconnectable": {}, + "version": "7.7" + } + }, + "widgets_values": [] + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "283e4561-61a2-4538-b960-265736eb041f", + "version": 1, + "state": { + "lastGroupId": 22, + "lastNodeId": 227, + "lastLinkId": 276, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "First-Last-Frame to Video", + "inputNode": { + "id": -10, + "bounding": [ + 270, + 3100, + 120, + 240 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + 3620, + 3120, + 120, + 60 + ] + }, + "inputs": [ + { + "id": "6fe179c4-d96f-4383-b202-844f6de4922e", + "name": "input", + "type": "IMAGE,MASK", + "linkIds": [ + 251 + ], + "localized_name": "input", + "label": "first_frame", + "pos": [ + 370, + 3120 + ] + }, + { + "id": "e80df1ae-5f39-4f86-91bd-0467635e2f2d", + "name": "input_1", + "type": "IMAGE,MASK", + "linkIds": [ + 253 + ], + "localized_name": "input_1", + "label": "last_frame", + "pos": [ + 370, + 3140 + ] + }, + { + "id": "433148fa-bf73-4ab1-81d9-09e2e38ed861", + "name": "text", + "type": "STRING", + "linkIds": [ + 265 + ], + "pos": [ + 370, + 3160 + ] + }, + { + "id": "36915bc8-a6ed-4d48-8619-e0e8723228e9", + "name": "value", + "type": "INT", + "linkIds": [ + 266 + ], + "label": "width", + "pos": [ + 370, + 3180 + ] + }, + { + "id": "425a36b8-91ab-41b7-81e9-496eba064ec8", + "name": "value_1", + "type": "INT", + "linkIds": [ + 267 + ], + "label": "height", + "pos": [ + 370, + 3200 + ] + }, + { + "id": "0c9e003b-bd07-4b7d-aa6d-789e138ed161", + "name": "value_2", + "type": "INT", + "linkIds": [ + 268 + ], + "label": "duration", + "pos": [ + 370, + 3220 + ] + }, + { + "id": "581b52ff-21c5-4774-ac2a-8f69a7e09e2e", + "name": "value_3", + "type": "INT", + "linkIds": [ + 269 + ], + "label": "fps", + "pos": [ + 370, + 3240 + ] + }, + { + "id": "d03cc171-45da-4658-99aa-77252bbcf522", + "name": "noise_seed", + "type": "INT", + "linkIds": [ + 270 + ], + "pos": [ + 370, + 3260 + ] + }, + { + "id": "e68e61c8-905e-43ac-8c76-65ac52270a08", + "name": "ckpt_name_1", + "type": "COMBO", + "linkIds": [ + 272, + 275, + 276 + ], + "label": "ckpt_name", + "pos": [ + 370, + 3280 + ] + }, + { + "id": "5d065f3b-891b-499f-950b-c2df0be24536", + "name": "text_encoder", + "type": "COMBO", + "linkIds": [ + 273 + ], + "pos": [ + 370, + 3300 + ] + } + ], + "outputs": [ + { + "id": "0c8c2dc0-c67c-4bc2-9e57-6aa00db2e3a9", + "name": "VIDEO", + "type": "VIDEO", + "linkIds": [ + 252 + ], + "localized_name": "VIDEO", + "pos": [ + 3640, + 3140 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 195, + "type": "LTXVPreprocess", + "pos": [ + 1480, + 3780 + ], + "size": [ + 230, + 110 + ], + "flags": { + "collapsed": false + }, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 203 + }, + { + "localized_name": "img_compression", + "name": "img_compression", + "type": "INT", + "widget": { + "name": "img_compression" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "output_image", + "name": "output_image", + "type": "IMAGE", + "links": [ + 229 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.7.0", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "LTXVPreprocess", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 25 + ] + }, + { + "id": 196, + "type": "RandomNoise", + "pos": [ + 1990, + 2320 + ], + "size": [ + 280, + 110 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "noise_seed", + "name": "noise_seed", + "type": "INT", + "widget": { + "name": "noise_seed" + }, + "link": 270 + } + ], + "outputs": [ + { + "localized_name": "NOISE", + "name": "NOISE", + "type": "NOISE", + "links": [ + 246 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.14.1", + "ue_properties": { + "widget_ue_connectable": { + "noise_seed": true + }, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "RandomNoise", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 315253765879496, + "randomize" + ] + }, + { + "id": 197, + "type": "LTXVEmptyLatentAudio", + "pos": [ + 2090, + 3820 + ], + "size": [ + 280, + 170 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "localized_name": "audio_vae", + "name": "audio_vae", + "type": "VAE", + "link": 205 + }, + { + "localized_name": "frames_number", + "name": "frames_number", + "type": "INT", + "widget": { + "name": "frames_number" + }, + "link": 262 + }, + { + "localized_name": "frame_rate", + "name": "frame_rate", + "type": "INT", + "widget": { + "name": "frame_rate" + }, + "link": 207 + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "widget": { + "name": "batch_size" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "Latent", + "name": "Latent", + "type": "LATENT", + "links": [ + 245 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.68", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.5.2", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "LTXVEmptyLatentAudio", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 97, + 25, + 1 + ] + }, + { + "id": 198, + "type": "PrimitiveInt", + "pos": [ + 760, + 3650 + ], + "size": [ + 230, + 110 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "INT", + "widget": { + "name": "value" + }, + "link": 268 + } + ], + "outputs": [ + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 260 + ] + } + ], + "title": "Duration", + "properties": { + "cnr_id": "comfy-core", + "ver": "0.7.0", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.5.2", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "PrimitiveInt", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 5, + "fixed" + ] + }, + { + "id": 199, + "type": "LTXVPreprocess", + "pos": [ + 1480, + 3340 + ], + "size": [ + 230, + 110 + ], + "flags": { + "collapsed": false + }, + "order": 6, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 210 + }, + { + "localized_name": "img_compression", + "name": "img_compression", + "type": "INT", + "widget": { + "name": "img_compression" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "output_image", + "name": "output_image", + "type": "IMAGE", + "links": [ + 240 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.7.0", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "LTXVPreprocess", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 25 + ] + }, + { + "id": 200, + "type": "LTXVCropGuides", + "pos": [ + 2820, + 2450 + ], + "size": [ + 280, + 120 + ], + "flags": {}, + "order": 7, + "mode": 0, + "inputs": [ + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "link": 213 + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "link": 214 + }, + { + "localized_name": "latent", + "name": "latent", + "type": "LATENT", + "link": 215 + } + ], + "outputs": [ + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "links": [] + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "links": [] + }, + { + "localized_name": "latent", + "name": "latent", + "type": "LATENT", + "links": [ + 211 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.8.2", + "ue_properties": { + "widget_ue_connectable": {}, + "input_ue_unconnectable": {}, + "version": "7.5.2" + }, + "Node name for S&R": "LTXVCropGuides", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 201, + "type": "EmptyLTXVLatentVideo", + "pos": [ + 2090, + 3580 + ], + "size": [ + 280, + 200 + ], + "flags": {}, + "order": 8, + "mode": 0, + "inputs": [ + { + "localized_name": "width", + "name": "width", + "type": "INT", + "widget": { + "name": "width" + }, + "link": 218 + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "widget": { + "name": "height" + }, + "link": 219 + }, + { + "localized_name": "length", + "name": "length", + "type": "INT", + "widget": { + "name": "length" + }, + "link": 263 + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "widget": { + "name": "batch_size" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "LATENT", + "name": "LATENT", + "type": "LATENT", + "links": [ + 239 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.60", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.5.2", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "EmptyLTXVLatentVideo", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 768, + 512, + 97, + 1 + ] + }, + { + "id": 202, + "type": "LTXVConditioning", + "pos": [ + 2090, + 3400 + ], + "size": [ + 280, + 130 + ], + "flags": {}, + "order": 9, + "mode": 0, + "inputs": [ + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "link": 221 + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "link": 222 + }, + { + "localized_name": "frame_rate", + "name": "frame_rate", + "type": "FLOAT", + "widget": { + "name": "frame_rate" + }, + "link": 223 + } + ], + "outputs": [ + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "links": [ + 236 + ] + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "links": [ + 237 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.56", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.5.2", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "LTXVConditioning", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 25 + ] + }, + { + "id": 203, + "type": "GetImageSize", + "pos": [ + 1480, + 3500 + ], + "size": [ + 230, + 130 + ], + "flags": {}, + "order": 10, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 224 + } + ], + "outputs": [ + { + "localized_name": "width", + "name": "width", + "type": "INT", + "links": [ + 218 + ] + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "links": [ + 219 + ] + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "links": [] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.14.1", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "GetImageSize", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 204, + "type": "LTXVAddGuide", + "pos": [ + 2750, + 3700 + ], + "size": [ + 280, + 240 + ], + "flags": {}, + "order": 11, + "mode": 0, + "inputs": [ + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "link": 225 + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "link": 226 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 227 + }, + { + "localized_name": "latent", + "name": "latent", + "type": "LATENT", + "link": 228 + }, + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 229 + }, + { + "localized_name": "frame_idx", + "name": "frame_idx", + "type": "INT", + "widget": { + "name": "frame_idx" + }, + "link": null + }, + { + "localized_name": "strength", + "name": "strength", + "type": "FLOAT", + "widget": { + "name": "strength" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "links": [ + 213, + 242 + ] + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "links": [ + 214, + 243 + ] + }, + { + "localized_name": "latent", + "name": "latent", + "type": "LATENT", + "links": [ + 244 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.12.3", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "LTXVAddGuide", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + -1, + 0.7 + ] + }, + { + "id": 205, + "type": "PrimitiveInt", + "pos": [ + 760, + 3800 + ], + "size": [ + 230, + 110 + ], + "flags": {}, + "order": 12, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "INT", + "widget": { + "name": "value" + }, + "link": 269 + } + ], + "outputs": [ + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 207, + 235, + 261 + ] + } + ], + "title": "Frame Rate(int)", + "properties": { + "cnr_id": "comfy-core", + "ver": "0.7.0", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.5.2", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "PrimitiveInt", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 25, + "fixed" + ] + }, + { + "id": 206, + "type": "LTXVAddGuide", + "pos": [ + 2750, + 3430 + ], + "size": [ + 280, + 240 + ], + "flags": {}, + "order": 13, + "mode": 0, + "inputs": [ + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "link": 236 + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "link": 237 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 238 + }, + { + "localized_name": "latent", + "name": "latent", + "type": "LATENT", + "link": 239 + }, + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 240 + }, + { + "localized_name": "frame_idx", + "name": "frame_idx", + "type": "INT", + "widget": { + "name": "frame_idx" + }, + "link": null + }, + { + "localized_name": "strength", + "name": "strength", + "type": "FLOAT", + "widget": { + "name": "strength" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "links": [ + 225 + ] + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "links": [ + 226 + ] + }, + { + "localized_name": "latent", + "name": "latent", + "type": "LATENT", + "links": [ + 228 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.12.3", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "LTXVAddGuide", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 0, + 0.7 + ] + }, + { + "id": 207, + "type": "CFGGuider", + "pos": [ + 1990, + 2500 + ], + "size": [ + 280, + 160 + ], + "flags": {}, + "order": 14, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 241 + }, + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "link": 242 + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "link": 243 + }, + { + "localized_name": "cfg", + "name": "cfg", + "type": "FLOAT", + "widget": { + "name": "cfg" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "GUIDER", + "name": "GUIDER", + "type": "GUIDER", + "links": [ + 247 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.14.1", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "CFGGuider", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 1 + ] + }, + { + "id": 208, + "type": "SamplerEulerAncestral", + "pos": [ + 1990, + 2720 + ], + "size": [ + 280, + 120 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "eta", + "name": "eta", + "type": "FLOAT", + "widget": { + "name": "eta" + }, + "link": null + }, + { + "localized_name": "s_noise", + "name": "s_noise", + "type": "FLOAT", + "widget": { + "name": "s_noise" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "SAMPLER", + "name": "SAMPLER", + "type": "SAMPLER", + "links": [ + 248 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.14.1", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "SamplerEulerAncestral", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 0, + 1 + ] + }, + { + "id": 209, + "type": "ManualSigmas", + "pos": [ + 1990, + 2910 + ], + "size": [ + 280, + 110 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "sigmas", + "name": "sigmas", + "type": "STRING", + "widget": { + "name": "sigmas" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "SIGMAS", + "name": "SIGMAS", + "type": "SIGMAS", + "links": [ + 249 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.14.1", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "ManualSigmas", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "1., 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875, 0.0" + ] + }, + { + "id": 210, + "type": "LTXVConcatAVLatent", + "pos": [ + 1990, + 3090 + ], + "size": [ + 280, + 100 + ], + "flags": {}, + "order": 15, + "mode": 0, + "inputs": [ + { + "localized_name": "video_latent", + "name": "video_latent", + "type": "LATENT", + "link": 244 + }, + { + "localized_name": "audio_latent", + "name": "audio_latent", + "type": "LATENT", + "link": 245 + } + ], + "outputs": [ + { + "localized_name": "latent", + "name": "latent", + "type": "LATENT", + "links": [ + 250 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.7.0", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.5.2", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "LTXVConcatAVLatent", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 211, + "type": "SamplerCustomAdvanced", + "pos": [ + 2460, + 2330 + ], + "size": [ + 230, + 170 + ], + "flags": {}, + "order": 16, + "mode": 0, + "inputs": [ + { + "localized_name": "noise", + "name": "noise", + "type": "NOISE", + "link": 246 + }, + { + "localized_name": "guider", + "name": "guider", + "type": "GUIDER", + "link": 247 + }, + { + "localized_name": "sampler", + "name": "sampler", + "type": "SAMPLER", + "link": 248 + }, + { + "localized_name": "sigmas", + "name": "sigmas", + "type": "SIGMAS", + "link": 249 + }, + { + "localized_name": "latent_image", + "name": "latent_image", + "type": "LATENT", + "link": 250 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "LATENT", + "links": [] + }, + { + "localized_name": "denoised_output", + "name": "denoised_output", + "type": "LATENT", + "links": [ + 204 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.14.1", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "SamplerCustomAdvanced", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 212, + "type": "ComfyMathExpression", + "pos": [ + 760, + 3970 + ], + "size": [ + 230, + 170 + ], + "flags": { + "collapsed": true + }, + "order": 17, + "mode": 0, + "inputs": [ + { + "label": "a", + "localized_name": "values.a", + "name": "values.a", + "type": "FLOAT,INT", + "link": 235 + }, + { + "label": "b", + "localized_name": "values.b", + "name": "values.b", + "shape": 7, + "type": "FLOAT,INT", + "link": null + }, + { + "localized_name": "expression", + "name": "expression", + "type": "STRING", + "widget": { + "name": "expression" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "FLOAT", + "name": "FLOAT", + "type": "FLOAT", + "links": [ + 223, + 234 + ] + }, + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.17.0", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "ComfyMathExpression", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "a" + ] + }, + { + "id": 213, + "type": "ResizeImageMaskNode", + "pos": [ + 1130, + 3340 + ], + "size": [ + 280, + 160 + ], + "flags": {}, + "order": 18, + "mode": 0, + "inputs": [ + { + "localized_name": "input", + "name": "input", + "type": "IMAGE,MASK", + "link": 251 + }, + { + "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": 208 + }, + { + "localized_name": "height", + "name": "resize_type.height", + "type": "INT", + "widget": { + "name": "resize_type.height" + }, + "link": 209 + }, + { + "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": [ + 210, + 224 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.14.1", + "ue_properties": { + "widget_ue_connectable": { + "resize_type.width": true, + "resize_type.height": true + }, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "ResizeImageMaskNode", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "scale dimensions", + 640, + 360, + "center", + "nearest-exact" + ] + }, + { + "id": 214, + "type": "ResizeImageMaskNode", + "pos": [ + 1130, + 3780 + ], + "size": [ + 280, + 160 + ], + "flags": {}, + "order": 19, + "mode": 0, + "inputs": [ + { + "localized_name": "input", + "name": "input", + "type": "IMAGE,MASK", + "link": 253 + }, + { + "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": 201 + }, + { + "localized_name": "height", + "name": "resize_type.height", + "type": "INT", + "widget": { + "name": "resize_type.height" + }, + "link": 202 + }, + { + "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": [ + 203 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.14.1", + "ue_properties": { + "widget_ue_connectable": { + "resize_type.width": true, + "resize_type.height": true + }, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "ResizeImageMaskNode", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "scale dimensions", + 640, + 360, + "center", + "nearest-exact" + ] + }, + { + "id": 215, + "type": "PrimitiveInt", + "pos": [ + 760, + 3340 + ], + "size": [ + 230, + 110 + ], + "flags": {}, + "order": 20, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "INT", + "widget": { + "name": "value" + }, + "link": 266 + } + ], + "outputs": [ + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 201, + 208 + ] + } + ], + "title": "Width", + "properties": { + "cnr_id": "comfy-core", + "ver": "0.7.0", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.5.2", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "PrimitiveInt", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 1280, + "fixed" + ] + }, + { + "id": 216, + "type": "PrimitiveInt", + "pos": [ + 760, + 3490 + ], + "size": [ + 230, + 110 + ], + "flags": {}, + "order": 21, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "INT", + "widget": { + "name": "value" + }, + "link": 267 + } + ], + "outputs": [ + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 202, + 209 + ] + } + ], + "title": "height", + "properties": { + "cnr_id": "comfy-core", + "ver": "0.7.0", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.5.2", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "PrimitiveInt", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 720, + "fixed" + ] + }, + { + "id": 217, + "type": "CLIPTextEncode", + "pos": [ + 1320, + 2870 + ], + "size": [ + 590, + 200 + ], + "flags": { + "collapsed": false + }, + "order": 22, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 230 + }, + { + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "links": [ + 222 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.56", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.5.2", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "CLIPTextEncode", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "blurry, out of focus, overexposed, underexposed, low contrast, washed out colors, excessive noise, grainy texture, poor lighting, flickering, motion blur, distorted proportions, unnatural skin tones, deformed facial features, asymmetrical face, missing facial features, extra limbs, disfigured hands, wrong hand count, artifacts around text, unreadable text on shirt or hat, incorrect lettering on cap (“PNTR”), incorrect t-shirt slogan (“JUST DO IT”), missing microphone, misplaced microphone, inconsistent perspective, camera shake, incorrect depth of field, background too sharp, background clutter, distracting reflections, harsh shadows, inconsistent lighting direction, color banding, cartoonish rendering, 3D CGI look, unrealistic materials, uncanny valley effect, incorrect ethnicity, wrong gender, exaggerated expressions, smiling, laughing, exaggerated sadness, wrong gaze direction, eyes looking at camera, mismatched lip sync, silent or muted audio, distorted voice, robotic voice, echo, background noise, off-sync audio, missing sniff sounds, incorrect dialogue, added dialogue, repetitive speech, jittery movement, awkward pauses, incorrect timing, unnatural transitions, inconsistent framing, tilted camera, missing door or shelves, missing shallow depth of field, flat lighting, inconsistent tone, cinematic oversaturation, stylized filters, or AI artifacts." + ], + "color": "#323", + "bgcolor": "#535" + }, + { + "id": 218, + "type": "CreateVideo", + "pos": [ + 3280, + 2320 + ], + "size": [ + 280, + 130 + ], + "flags": {}, + "order": 23, + "mode": 0, + "inputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "link": 232 + }, + { + "localized_name": "audio", + "name": "audio", + "shape": 7, + "type": "AUDIO", + "link": 233 + }, + { + "localized_name": "fps", + "name": "fps", + "type": "FLOAT", + "widget": { + "name": "fps" + }, + "link": 234 + } + ], + "outputs": [ + { + "localized_name": "VIDEO", + "name": "VIDEO", + "type": "VIDEO", + "links": [ + 252 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.14.1", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "CreateVideo", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 24 + ] + }, + { + "id": 219, + "type": "VAEDecodeTiled", + "pos": [ + 2820, + 2630 + ], + "size": [ + 280, + 200 + ], + "flags": {}, + "order": 24, + "mode": 0, + "inputs": [ + { + "localized_name": "samples", + "name": "samples", + "type": "LATENT", + "link": 211 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 212 + }, + { + "localized_name": "tile_size", + "name": "tile_size", + "type": "INT", + "widget": { + "name": "tile_size" + }, + "link": null + }, + { + "localized_name": "overlap", + "name": "overlap", + "type": "INT", + "widget": { + "name": "overlap" + }, + "link": null + }, + { + "localized_name": "temporal_size", + "name": "temporal_size", + "type": "INT", + "widget": { + "name": "temporal_size" + }, + "link": null + }, + { + "localized_name": "temporal_overlap", + "name": "temporal_overlap", + "type": "INT", + "widget": { + "name": "temporal_overlap" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 232 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.7.0", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.5.2", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "VAEDecodeTiled", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 768, + 64, + 4096, + 64 + ] + }, + { + "id": 220, + "type": "LTXVAudioVAEDecode", + "pos": [ + 2820, + 2920 + ], + "size": [ + 280, + 100 + ], + "flags": {}, + "order": 25, + "mode": 0, + "inputs": [ + { + "localized_name": "samples", + "name": "samples", + "type": "LATENT", + "link": 216 + }, + { + "label": "Audio VAE", + "localized_name": "audio_vae", + "name": "audio_vae", + "type": "VAE", + "link": 217 + } + ], + "outputs": [ + { + "localized_name": "Audio", + "name": "Audio", + "type": "AUDIO", + "links": [ + 233 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.7.0", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.5.2", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "LTXVAudioVAEDecode", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 221, + "type": "LTXVSeparateAVLatent", + "pos": [ + 2460, + 2580 + ], + "size": [ + 250, + 100 + ], + "flags": {}, + "order": 26, + "mode": 0, + "inputs": [ + { + "localized_name": "av_latent", + "name": "av_latent", + "type": "LATENT", + "link": 204 + } + ], + "outputs": [ + { + "localized_name": "video_latent", + "name": "video_latent", + "type": "LATENT", + "links": [ + 215 + ] + }, + { + "localized_name": "audio_latent", + "name": "audio_latent", + "type": "LATENT", + "links": [ + 216 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.5.1", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.5.2", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "LTXVSeparateAVLatent", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 222, + "type": "CLIPTextEncode", + "pos": [ + 1310, + 2380 + ], + "size": [ + 620, + 420 + ], + "flags": {}, + "order": 27, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 231 + }, + { + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": 265 + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "links": [ + 221 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.56", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.5.2", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "CLIPTextEncode", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "" + ], + "color": "#232", + "bgcolor": "#353" + }, + { + "id": 223, + "type": "CheckpointLoaderSimple", + "pos": [ + 770, + 2380 + ], + "size": [ + 420, + 160 + ], + "flags": {}, + "order": 28, + "mode": 0, + "inputs": [ + { + "localized_name": "ckpt_name", + "name": "ckpt_name", + "type": "COMBO", + "widget": { + "name": "ckpt_name" + }, + "link": 276 + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "links": [ + 241 + ] + }, + { + "localized_name": "CLIP", + "name": "CLIP", + "type": "CLIP", + "links": [] + }, + { + "localized_name": "VAE", + "name": "VAE", + "type": "VAE", + "links": [ + 212, + 227, + 238 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.10.0", + "ue_properties": { + "widget_ue_connectable": {}, + "input_ue_unconnectable": {}, + "version": "7.5.2" + }, + "Node name for S&R": "CheckpointLoaderSimple", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "models": [ + { + "name": "ltx-2.3-22b-distilled-fp8.safetensors", + "url": "https://huggingface.co/Lightricks/LTX-2.3-fp8/resolve/main/ltx-2.3-22b-distilled-fp8.safetensors", + "directory": "checkpoints" + } + ] + }, + "widgets_values": [ + "ltx-2.3-22b-distilled-fp8.safetensors" + ] + }, + { + "id": 224, + "type": "LTXVAudioVAELoader", + "pos": [ + 770, + 2660 + ], + "size": [ + 420, + 110 + ], + "flags": {}, + "order": 29, + "mode": 0, + "inputs": [ + { + "localized_name": "ckpt_name", + "name": "ckpt_name", + "type": "COMBO", + "widget": { + "name": "ckpt_name" + }, + "link": 272 + } + ], + "outputs": [ + { + "localized_name": "Audio VAE", + "name": "Audio VAE", + "type": "VAE", + "links": [ + 205, + 217 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.10.0", + "ue_properties": { + "widget_ue_connectable": {}, + "input_ue_unconnectable": {}, + "version": "7.5.2" + }, + "Node name for S&R": "LTXVAudioVAELoader", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "models": [ + { + "name": "ltx-2.3-22b-distilled-fp8.safetensors", + "url": "https://huggingface.co/Lightricks/LTX-2.3-fp8/resolve/main/ltx-2.3-22b-distilled-fp8.safetensors", + "directory": "checkpoints" + } + ] + }, + "widgets_values": [ + "ltx-2.3-22b-distilled-fp8.safetensors" + ] + }, + { + "id": 225, + "type": "LTXAVTextEncoderLoader", + "pos": [ + 770, + 2890 + ], + "size": [ + 410, + 160 + ], + "flags": {}, + "order": 30, + "mode": 0, + "inputs": [ + { + "localized_name": "text_encoder", + "name": "text_encoder", + "type": "COMBO", + "widget": { + "name": "text_encoder" + }, + "link": 273 + }, + { + "localized_name": "ckpt_name", + "name": "ckpt_name", + "type": "COMBO", + "widget": { + "name": "ckpt_name" + }, + "link": 275 + }, + { + "localized_name": "device", + "name": "device", + "type": "COMBO", + "widget": { + "name": "device" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "CLIP", + "name": "CLIP", + "type": "CLIP", + "links": [ + 230, + 231 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.10.0", + "ue_properties": { + "widget_ue_connectable": {}, + "input_ue_unconnectable": {}, + "version": "7.5.2" + }, + "Node name for S&R": "LTXAVTextEncoderLoader", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "models": [ + { + "name": "gemma_3_12B_it_fp4_mixed.safetensors", + "url": "https://huggingface.co/Comfy-Org/ltx-2/resolve/main/split_files/text_encoders/gemma_3_12B_it_fp4_mixed.safetensors", + "directory": "text_encoders" + }, + { + "name": "ltx-2.3-22b-distilled-fp8.safetensors", + "url": "https://huggingface.co/Lightricks/LTX-2.3-fp8/resolve/main/ltx-2.3-22b-distilled-fp8.safetensors", + "directory": "checkpoints" + } + ] + }, + "widgets_values": [ + "gemma_3_12B_it_fp4_mixed.safetensors", + "ltx-2.3-22b-distilled-fp8.safetensors", + "default" + ] + }, + { + "id": 226, + "type": "ComfyMathExpression", + "pos": [ + 760, + 4020 + ], + "size": [ + 400, + 200 + ], + "flags": { + "collapsed": true + }, + "order": 31, + "mode": 0, + "inputs": [ + { + "label": "a", + "localized_name": "values.a", + "name": "values.a", + "type": "FLOAT,INT", + "link": 260 + }, + { + "label": "b", + "localized_name": "values.b", + "name": "values.b", + "shape": 7, + "type": "FLOAT,INT", + "link": 261 + }, + { + "label": "c", + "localized_name": "values.c", + "name": "values.c", + "shape": 7, + "type": "FLOAT,INT", + "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": [ + 262, + 263 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.18.1", + "ue_properties": { + "widget_ue_connectable": {}, + "input_ue_unconnectable": {}, + "version": "7.7" + }, + "Node name for S&R": "ComfyMathExpression" + }, + "widgets_values": [ + "a * b + 1" + ] + } + ], + "groups": [ + { + "id": 1, + "title": "Conditioning", + "bounding": [ + 1850, + 3250, + 1370, + 800 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 2, + "title": "Settings", + "bounding": [ + 730, + 3250, + 290, + 800 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 3, + "title": "FIrst Frame", + "bounding": [ + 1050, + 3250, + 770, + 400 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 4, + "title": "Last Frame", + "bounding": [ + 1050, + 3680, + 770, + 370 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 5, + "title": "Model", + "bounding": [ + 730, + 2240, + 500, + 980 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 6, + "title": "Prompt", + "bounding": [ + 1260, + 2240, + 680, + 980 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 7, + "title": "Sampling", + "bounding": [ + 1970, + 2240, + 770, + 980 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 8, + "title": "Decoding", + "bounding": [ + 2770, + 2240, + 450, + 980 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + } + ], + "links": [ + { + "id": 203, + "origin_id": 214, + "origin_slot": 0, + "target_id": 195, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 205, + "origin_id": 224, + "origin_slot": 0, + "target_id": 197, + "target_slot": 0, + "type": "VAE" + }, + { + "id": 207, + "origin_id": 205, + "origin_slot": 0, + "target_id": 197, + "target_slot": 2, + "type": "INT" + }, + { + "id": 210, + "origin_id": 213, + "origin_slot": 0, + "target_id": 199, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 213, + "origin_id": 204, + "origin_slot": 0, + "target_id": 200, + "target_slot": 0, + "type": "CONDITIONING" + }, + { + "id": 214, + "origin_id": 204, + "origin_slot": 1, + "target_id": 200, + "target_slot": 1, + "type": "CONDITIONING" + }, + { + "id": 215, + "origin_id": 221, + "origin_slot": 0, + "target_id": 200, + "target_slot": 2, + "type": "LATENT" + }, + { + "id": 218, + "origin_id": 203, + "origin_slot": 0, + "target_id": 201, + "target_slot": 0, + "type": "INT" + }, + { + "id": 219, + "origin_id": 203, + "origin_slot": 1, + "target_id": 201, + "target_slot": 1, + "type": "INT" + }, + { + "id": 221, + "origin_id": 222, + "origin_slot": 0, + "target_id": 202, + "target_slot": 0, + "type": "CONDITIONING" + }, + { + "id": 222, + "origin_id": 217, + "origin_slot": 0, + "target_id": 202, + "target_slot": 1, + "type": "CONDITIONING" + }, + { + "id": 223, + "origin_id": 212, + "origin_slot": 0, + "target_id": 202, + "target_slot": 2, + "type": "FLOAT" + }, + { + "id": 224, + "origin_id": 213, + "origin_slot": 0, + "target_id": 203, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 225, + "origin_id": 206, + "origin_slot": 0, + "target_id": 204, + "target_slot": 0, + "type": "CONDITIONING" + }, + { + "id": 226, + "origin_id": 206, + "origin_slot": 1, + "target_id": 204, + "target_slot": 1, + "type": "CONDITIONING" + }, + { + "id": 227, + "origin_id": 223, + "origin_slot": 2, + "target_id": 204, + "target_slot": 2, + "type": "VAE" + }, + { + "id": 228, + "origin_id": 206, + "origin_slot": 2, + "target_id": 204, + "target_slot": 3, + "type": "LATENT" + }, + { + "id": 229, + "origin_id": 195, + "origin_slot": 0, + "target_id": 204, + "target_slot": 4, + "type": "IMAGE" + }, + { + "id": 236, + "origin_id": 202, + "origin_slot": 0, + "target_id": 206, + "target_slot": 0, + "type": "CONDITIONING" + }, + { + "id": 237, + "origin_id": 202, + "origin_slot": 1, + "target_id": 206, + "target_slot": 1, + "type": "CONDITIONING" + }, + { + "id": 238, + "origin_id": 223, + "origin_slot": 2, + "target_id": 206, + "target_slot": 2, + "type": "VAE" + }, + { + "id": 239, + "origin_id": 201, + "origin_slot": 0, + "target_id": 206, + "target_slot": 3, + "type": "LATENT" + }, + { + "id": 240, + "origin_id": 199, + "origin_slot": 0, + "target_id": 206, + "target_slot": 4, + "type": "IMAGE" + }, + { + "id": 241, + "origin_id": 223, + "origin_slot": 0, + "target_id": 207, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 242, + "origin_id": 204, + "origin_slot": 0, + "target_id": 207, + "target_slot": 1, + "type": "CONDITIONING" + }, + { + "id": 243, + "origin_id": 204, + "origin_slot": 1, + "target_id": 207, + "target_slot": 2, + "type": "CONDITIONING" + }, + { + "id": 244, + "origin_id": 204, + "origin_slot": 2, + "target_id": 210, + "target_slot": 0, + "type": "LATENT" + }, + { + "id": 245, + "origin_id": 197, + "origin_slot": 0, + "target_id": 210, + "target_slot": 1, + "type": "LATENT" + }, + { + "id": 246, + "origin_id": 196, + "origin_slot": 0, + "target_id": 211, + "target_slot": 0, + "type": "NOISE" + }, + { + "id": 247, + "origin_id": 207, + "origin_slot": 0, + "target_id": 211, + "target_slot": 1, + "type": "GUIDER" + }, + { + "id": 248, + "origin_id": 208, + "origin_slot": 0, + "target_id": 211, + "target_slot": 2, + "type": "SAMPLER" + }, + { + "id": 249, + "origin_id": 209, + "origin_slot": 0, + "target_id": 211, + "target_slot": 3, + "type": "SIGMAS" + }, + { + "id": 250, + "origin_id": 210, + "origin_slot": 0, + "target_id": 211, + "target_slot": 4, + "type": "LATENT" + }, + { + "id": 235, + "origin_id": 205, + "origin_slot": 0, + "target_id": 212, + "target_slot": 0, + "type": "INT" + }, + { + "id": 208, + "origin_id": 215, + "origin_slot": 0, + "target_id": 213, + "target_slot": 2, + "type": "INT" + }, + { + "id": 209, + "origin_id": 216, + "origin_slot": 0, + "target_id": 213, + "target_slot": 3, + "type": "INT" + }, + { + "id": 201, + "origin_id": 215, + "origin_slot": 0, + "target_id": 214, + "target_slot": 2, + "type": "INT" + }, + { + "id": 202, + "origin_id": 216, + "origin_slot": 0, + "target_id": 214, + "target_slot": 3, + "type": "INT" + }, + { + "id": 230, + "origin_id": 225, + "origin_slot": 0, + "target_id": 217, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 232, + "origin_id": 219, + "origin_slot": 0, + "target_id": 218, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 233, + "origin_id": 220, + "origin_slot": 0, + "target_id": 218, + "target_slot": 1, + "type": "AUDIO" + }, + { + "id": 234, + "origin_id": 212, + "origin_slot": 0, + "target_id": 218, + "target_slot": 2, + "type": "FLOAT" + }, + { + "id": 211, + "origin_id": 200, + "origin_slot": 2, + "target_id": 219, + "target_slot": 0, + "type": "LATENT" + }, + { + "id": 212, + "origin_id": 223, + "origin_slot": 2, + "target_id": 219, + "target_slot": 1, + "type": "VAE" + }, + { + "id": 216, + "origin_id": 221, + "origin_slot": 1, + "target_id": 220, + "target_slot": 0, + "type": "LATENT" + }, + { + "id": 217, + "origin_id": 224, + "origin_slot": 0, + "target_id": 220, + "target_slot": 1, + "type": "VAE" + }, + { + "id": 204, + "origin_id": 211, + "origin_slot": 1, + "target_id": 221, + "target_slot": 0, + "type": "LATENT" + }, + { + "id": 231, + "origin_id": 225, + "origin_slot": 0, + "target_id": 222, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 251, + "origin_id": -10, + "origin_slot": 0, + "target_id": 213, + "target_slot": 0, + "type": "IMAGE,MASK" + }, + { + "id": 253, + "origin_id": -10, + "origin_slot": 1, + "target_id": 214, + "target_slot": 0, + "type": "IMAGE,MASK" + }, + { + "id": 252, + "origin_id": 218, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "VIDEO" + }, + { + "id": 260, + "origin_id": 198, + "origin_slot": 0, + "target_id": 226, + "target_slot": 0, + "type": "INT" + }, + { + "id": 261, + "origin_id": 205, + "origin_slot": 0, + "target_id": 226, + "target_slot": 1, + "type": "INT" + }, + { + "id": 262, + "origin_id": 226, + "origin_slot": 1, + "target_id": 197, + "target_slot": 1, + "type": "INT" + }, + { + "id": 263, + "origin_id": 226, + "origin_slot": 1, + "target_id": 201, + "target_slot": 2, + "type": "INT" + }, + { + "id": 265, + "origin_id": -10, + "origin_slot": 2, + "target_id": 222, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 266, + "origin_id": -10, + "origin_slot": 3, + "target_id": 215, + "target_slot": 0, + "type": "INT" + }, + { + "id": 267, + "origin_id": -10, + "origin_slot": 4, + "target_id": 216, + "target_slot": 0, + "type": "INT" + }, + { + "id": 268, + "origin_id": -10, + "origin_slot": 5, + "target_id": 198, + "target_slot": 0, + "type": "INT" + }, + { + "id": 269, + "origin_id": -10, + "origin_slot": 6, + "target_id": 205, + "target_slot": 0, + "type": "INT" + }, + { + "id": 270, + "origin_id": -10, + "origin_slot": 7, + "target_id": 196, + "target_slot": 0, + "type": "INT" + }, + { + "id": 272, + "origin_id": -10, + "origin_slot": 8, + "target_id": 224, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 273, + "origin_id": -10, + "origin_slot": 9, + "target_id": 225, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 275, + "origin_id": -10, + "origin_slot": 8, + "target_id": 225, + "target_slot": 1, + "type": "COMBO" + }, + { + "id": 276, + "origin_id": -10, + "origin_slot": 8, + "target_id": 223, + "target_slot": 0, + "type": "COMBO" + } + ], + "extra": {}, + "category": "Video generation and editing/First-Last-Frame to Video", + "description": "Generates a video that interpolates between the first and last keyframes using LTX-2.3, including optional audio." + } + ] + }, + "extra": { + "ue_links": [] + } +} \ No newline at end of file diff --git a/blueprints/Frame Interpolation.json b/blueprints/Frame Interpolation.json new file mode 100644 index 000000000..8e183de7e --- /dev/null +++ b/blueprints/Frame Interpolation.json @@ -0,0 +1,858 @@ +{ + "revision": 0, + "last_node_id": 16, + "last_link_id": 0, + "nodes": [ + { + "id": 16, + "type": "022693be-2baa-4009-870a-28921508a7ef", + "pos": [ + -2990, + -3240 + ], + "size": [ + 410, + 200 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "video", + "name": "video", + "type": "VIDEO", + "link": null + }, + { + "label": "multiplier", + "name": "value", + "type": "INT", + "widget": { + "name": "value" + }, + "link": null + }, + { + "label": "enable_fps_multiplier", + "name": "value_1", + "type": "BOOLEAN", + "widget": { + "name": "value_1" + }, + "link": null + }, + { + "name": "model_name", + "type": "COMBO", + "widget": { + "name": "model_name" + }, + "link": null + } + ], + "outputs": [ + { + "label": "VIDEO", + "name": "VIDEO_1", + "type": "VIDEO", + "links": [] + }, + { + "name": "IMAGE", + "type": "IMAGE", + "links": null + } + ], + "properties": { + "proxyWidgets": [ + [ + "9", + "value" + ], + [ + "13", + "value" + ], + [ + "1", + "model_name" + ] + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.19.3" + }, + "widgets_values": [], + "title": "Frame Interpolation" + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "022693be-2baa-4009-870a-28921508a7ef", + "version": 1, + "state": { + "lastGroupId": 0, + "lastNodeId": 17, + "lastLinkId": 28, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Frame Interpolation", + "inputNode": { + "id": -10, + "bounding": [ + -2810, + -3070, + 159.7421875, + 120 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + -1270, + -3075, + 120, + 80 + ] + }, + "inputs": [ + { + "id": "05e31c51-dcb6-4a1e-9651-1b9ad4f7a287", + "name": "video", + "type": "VIDEO", + "linkIds": [ + 2 + ], + "localized_name": "video", + "pos": [ + -2670.2578125, + -3050 + ] + }, + { + "id": "feecb409-7d1c-4a99-9c63-50c5fecdd3c9", + "name": "value", + "type": "INT", + "linkIds": [ + 22 + ], + "label": "multiplier", + "pos": [ + -2670.2578125, + -3030 + ] + }, + { + "id": "0b8a861b-b581-4068-9e8c-f8d15daf1ca6", + "name": "value_1", + "type": "BOOLEAN", + "linkIds": [ + 23 + ], + "label": "enable_fps_multiplier", + "pos": [ + -2670.2578125, + -3010 + ] + }, + { + "id": "a22b101e-8773-4e17-a297-7ee3aae09162", + "name": "model_name", + "type": "COMBO", + "linkIds": [ + 24 + ], + "pos": [ + -2670.2578125, + -2990 + ] + } + ], + "outputs": [ + { + "id": "ef2ada05-d5aa-492a-9394-6c3e71e39ebb", + "name": "VIDEO_1", + "type": "VIDEO", + "linkIds": [ + 26 + ], + "label": "VIDEO", + "pos": [ + -1250, + -3055 + ] + }, + { + "id": "5aacc622-2a07-4983-b31c-e04461f7f953", + "name": "IMAGE", + "type": "IMAGE", + "linkIds": [ + 28 + ], + "pos": [ + -1250, + -3035 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 1, + "type": "FrameInterpolationModelLoader", + "pos": [ + -2510, + -3370 + ], + "size": [ + 370, + 90 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "model_name", + "name": "model_name", + "type": "COMBO", + "widget": { + "name": "model_name" + }, + "link": 24 + } + ], + "outputs": [ + { + "localized_name": "INTERP_MODEL", + "name": "INTERP_MODEL", + "type": "INTERP_MODEL", + "links": [ + 1 + ] + } + ], + "properties": { + "Node name for S&R": "FrameInterpolationModelLoader", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.19.3", + "models": [ + { + "name": "film_net_fp16.safetensors", + "url": "https://huggingface.co/Comfy-Org/frame_interpolation/resolve/main/frame_interpolation/film_net_fp16.safetensors", + "directory": "frame_interpolation" + } + ] + }, + "widgets_values": [ + "film_net_fp16.safetensors" + ] + }, + { + "id": 2, + "type": "FrameInterpolate", + "pos": [ + -2040, + -3370 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "interp_model", + "name": "interp_model", + "type": "INTERP_MODEL", + "link": 1 + }, + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "link": 3 + }, + { + "localized_name": "multiplier", + "name": "multiplier", + "type": "INT", + "widget": { + "name": "multiplier" + }, + "link": 8 + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 4, + 28 + ] + } + ], + "properties": { + "Node name for S&R": "FrameInterpolate", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.19.3" + }, + "widgets_values": [ + 2 + ] + }, + { + "id": 5, + "type": "CreateVideo", + "pos": [ + -1600, + -3370 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "link": 4 + }, + { + "localized_name": "audio", + "name": "audio", + "shape": 7, + "type": "AUDIO", + "link": 5 + }, + { + "localized_name": "fps", + "name": "fps", + "type": "FLOAT", + "widget": { + "name": "fps" + }, + "link": 12 + } + ], + "outputs": [ + { + "localized_name": "VIDEO", + "name": "VIDEO", + "type": "VIDEO", + "links": [ + 26 + ] + } + ], + "properties": { + "Node name for S&R": "CreateVideo", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.19.3" + }, + "widgets_values": [ + 30 + ] + }, + { + "id": 9, + "type": "PrimitiveInt", + "pos": [ + -2500, + -2970 + ], + "size": [ + 270, + 90 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "INT", + "widget": { + "name": "value" + }, + "link": 22 + } + ], + "outputs": [ + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 8, + 19 + ] + } + ], + "title": "Int (Multiplier)", + "properties": { + "Node name for S&R": "PrimitiveInt", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.19.3" + }, + "widgets_values": [ + 2, + "fixed" + ] + }, + { + "id": 10, + "type": "ComfySwitchNode", + "pos": [ + -1610, + -3120 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 11 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 13 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 15 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 12 + ] + } + ], + "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": "0.19.3" + }, + "widgets_values": [ + true + ] + }, + { + "id": 13, + "type": "PrimitiveBoolean", + "pos": [ + -2500, + -2770 + ], + "size": [ + 310, + 90 + ], + "flags": {}, + "order": 7, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "BOOLEAN", + "widget": { + "name": "value" + }, + "link": 23 + } + ], + "outputs": [ + { + "localized_name": "BOOLEAN", + "name": "BOOLEAN", + "type": "BOOLEAN", + "links": [ + 15 + ] + } + ], + "title": "Boolean (Apply multiplier to FPS?)", + "properties": { + "Node name for S&R": "PrimitiveBoolean", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "cnr_id": "comfy-core", + "ver": "0.19.3" + }, + "widgets_values": [ + true + ] + }, + { + "id": 3, + "type": "GetVideoComponents", + "pos": [ + -2500, + -3170 + ], + "size": [ + 230, + 100 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "video", + "name": "video", + "type": "VIDEO", + "link": 2 + } + ], + "outputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "links": [ + 3 + ] + }, + { + "localized_name": "audio", + "name": "audio", + "type": "AUDIO", + "links": [ + 5 + ] + }, + { + "localized_name": "fps", + "name": "fps", + "type": "FLOAT", + "links": [ + 11, + 18 + ] + } + ], + "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": "0.19.3" + } + }, + { + "id": 11, + "type": "ComfyMathExpression", + "pos": [ + -2090, + -3070 + ], + "size": [ + 400, + 210 + ], + "flags": { + "collapsed": false + }, + "order": 6, + "mode": 0, + "inputs": [ + { + "label": "a", + "localized_name": "values.a", + "name": "values.a", + "type": "FLOAT,INT", + "link": 18 + }, + { + "label": "b", + "localized_name": "values.b", + "name": "values.b", + "shape": 7, + "type": "FLOAT,INT", + "link": 19 + }, + { + "label": "c", + "localized_name": "values.c", + "name": "values.c", + "shape": 7, + "type": "FLOAT,INT", + "link": null + }, + { + "localized_name": "expression", + "name": "expression", + "type": "STRING", + "widget": { + "name": "expression" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "FLOAT", + "name": "FLOAT", + "type": "FLOAT", + "links": [ + 13 + ] + }, + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": null + } + ], + "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": "0.19.3" + }, + "widgets_values": [ + "min(abs(b), 16) * a" + ] + } + ], + "groups": [], + "links": [ + { + "id": 1, + "origin_id": 1, + "origin_slot": 0, + "target_id": 2, + "target_slot": 0, + "type": "INTERP_MODEL" + }, + { + "id": 3, + "origin_id": 3, + "origin_slot": 0, + "target_id": 2, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 8, + "origin_id": 9, + "origin_slot": 0, + "target_id": 2, + "target_slot": 2, + "type": "INT" + }, + { + "id": 4, + "origin_id": 2, + "origin_slot": 0, + "target_id": 5, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 5, + "origin_id": 3, + "origin_slot": 1, + "target_id": 5, + "target_slot": 1, + "type": "AUDIO" + }, + { + "id": 12, + "origin_id": 10, + "origin_slot": 0, + "target_id": 5, + "target_slot": 2, + "type": "FLOAT" + }, + { + "id": 11, + "origin_id": 3, + "origin_slot": 2, + "target_id": 10, + "target_slot": 0, + "type": "FLOAT" + }, + { + "id": 13, + "origin_id": 11, + "origin_slot": 0, + "target_id": 10, + "target_slot": 1, + "type": "FLOAT" + }, + { + "id": 15, + "origin_id": 13, + "origin_slot": 0, + "target_id": 10, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 18, + "origin_id": 3, + "origin_slot": 2, + "target_id": 11, + "target_slot": 0, + "type": "FLOAT" + }, + { + "id": 19, + "origin_id": 9, + "origin_slot": 0, + "target_id": 11, + "target_slot": 1, + "type": "INT" + }, + { + "id": 2, + "origin_id": -10, + "origin_slot": 0, + "target_id": 3, + "target_slot": 0, + "type": "VIDEO" + }, + { + "id": 22, + "origin_id": -10, + "origin_slot": 1, + "target_id": 9, + "target_slot": 0, + "type": "INT" + }, + { + "id": 23, + "origin_id": -10, + "origin_slot": 2, + "target_id": 13, + "target_slot": 0, + "type": "BOOLEAN" + }, + { + "id": 24, + "origin_id": -10, + "origin_slot": 3, + "target_id": 1, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 26, + "origin_id": 5, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "VIDEO" + }, + { + "id": 28, + "origin_id": 2, + "origin_slot": 0, + "target_id": -20, + "target_slot": 1, + "type": "IMAGE" + } + ], + "extra": {}, + "category": "Video Tools", + "description": "Increases video frame rate by synthesizing intermediate frames with a frame interpolation model." + } + ] + }, + "extra": {} +} \ No newline at end of file diff --git a/blueprints/Get Any Video Frame.json b/blueprints/Get Any Video Frame.json new file mode 100644 index 000000000..9ff0f8e6e --- /dev/null +++ b/blueprints/Get Any Video Frame.json @@ -0,0 +1,485 @@ +{ + "revision": 0, + "last_node_id": 98, + "last_link_id": 0, + "nodes": [ + { + "id": 98, + "type": "dca6e78d-fb06-421e-97f7-6ce17a665260", + "pos": [ + -410, + -2230 + ], + "size": [ + 270, + 104 + ], + "flags": {}, + "order": 7, + "mode": 0, + "inputs": [ + { + "name": "video", + "type": "VIDEO", + "link": null + }, + { + "label": "frame_index", + "name": "value", + "type": "INT", + "widget": { + "name": "value" + }, + "link": null + } + ], + "outputs": [ + { + "name": "IMAGE", + "type": "IMAGE", + "links": [] + } + ], + "title": "Get Any Video Frame", + "properties": { + "proxyWidgets": [ + [ + "100", + "value" + ] + ] + }, + "widgets_values": [] + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "dca6e78d-fb06-421e-97f7-6ce17a665260", + "version": 1, + "state": { + "lastGroupId": 1, + "lastNodeId": 136, + "lastLinkId": 302, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Get Any Video Frame", + "inputNode": { + "id": -10, + "bounding": [ + 380, + -57, + 120, + 80 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + 1460, + -57, + 120, + 60 + ] + }, + "inputs": [ + { + "id": "2ceec378-8dcf-4340-8570-155967f59a93", + "name": "video", + "type": "VIDEO", + "linkIds": [ + 4 + ], + "pos": [ + 480, + -37 + ] + }, + { + "id": "819955f6-c686-4896-8032-ff2d0059109a", + "name": "value", + "type": "INT", + "linkIds": [ + 283 + ], + "label": "frame_index", + "pos": [ + 480, + -17 + ] + } + ], + "outputs": [ + { + "id": "1ab0684d-6a44-45b6-8aa4-a0b971a1d41e", + "name": "IMAGE", + "type": "IMAGE", + "linkIds": [ + 5 + ], + "pos": [ + 1480, + -37 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 1, + "type": "GetVideoComponents", + "pos": [ + 560, + -150 + ], + "size": [ + 230, + 120 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "video", + "name": "video", + "type": "VIDEO", + "link": 4 + } + ], + "outputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "links": [ + 1, + 2 + ] + }, + { + "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" + } + }, + { + "id": 2, + "type": "GetImageSize", + "pos": [ + 560, + 50 + ], + "size": [ + 230, + 120 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 1 + } + ], + "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": [ + 285 + ] + } + ], + "properties": { + "Node name for S&R": "GetImageSize" + } + }, + { + "id": 3, + "type": "ImageFromBatch", + "pos": [ + 1130, + -150 + ], + "size": [ + 270, + 140 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 2 + }, + { + "localized_name": "batch_index", + "name": "batch_index", + "type": "INT", + "widget": { + "name": "batch_index" + }, + "link": 286 + }, + { + "localized_name": "length", + "name": "length", + "type": "INT", + "widget": { + "name": "length" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 5 + ] + } + ], + "properties": { + "Node name for S&R": "ImageFromBatch" + }, + "widgets_values": [ + 0, + 1 + ] + }, + { + "id": 99, + "type": "ComfyMathExpression", + "pos": [ + 910, + 100 + ], + "size": [ + 400, + 200 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "label": "a", + "localized_name": "values.a", + "name": "values.a", + "type": "FLOAT,INT", + "link": 284 + }, + { + "label": "b", + "localized_name": "values.b", + "name": "values.b", + "shape": 7, + "type": "FLOAT,INT", + "link": 285 + }, + { + "label": "c", + "localized_name": "values.c", + "name": "values.c", + "shape": 7, + "type": "FLOAT,INT", + "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": [ + 286 + ] + } + ], + "properties": { + "Node name for S&R": "ComfyMathExpression" + }, + "widgets_values": [ + "min(max(int(a if a >= 0 else b + a), 0), b - 1)" + ] + }, + { + "id": 100, + "type": "PrimitiveInt", + "pos": [ + 560, + 250 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "INT", + "widget": { + "name": "value" + }, + "link": 283 + } + ], + "outputs": [ + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 284 + ] + } + ], + "properties": { + "Node name for S&R": "PrimitiveInt" + }, + "widgets_values": [ + 0, + "fixed" + ] + } + ], + "groups": [], + "links": [ + { + "id": 1, + "origin_id": 1, + "origin_slot": 0, + "target_id": 2, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 2, + "origin_id": 1, + "origin_slot": 0, + "target_id": 3, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 4, + "origin_id": -10, + "origin_slot": 0, + "target_id": 1, + "target_slot": 0, + "type": "VIDEO" + }, + { + "id": 5, + "origin_id": 3, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 283, + "origin_id": -10, + "origin_slot": 1, + "target_id": 100, + "target_slot": 0, + "type": "INT" + }, + { + "id": 284, + "origin_id": 100, + "origin_slot": 0, + "target_id": 99, + "target_slot": 0, + "type": "INT" + }, + { + "id": 285, + "origin_id": 2, + "origin_slot": 2, + "target_id": 99, + "target_slot": 1, + "type": "INT" + }, + { + "id": 286, + "origin_id": 99, + "origin_slot": 1, + "target_id": 3, + "target_slot": 1, + "type": "INT" + } + ], + "extra": {}, + "category": "Video Tools", + "description": "Extracts one image frame from a video at a chosen index, with optional trim and FPS control." + } + ] + }, + "extra": { + "ds": { + "scale": 1.197015527856339, + "offset": [ + -168.76833554248222, + 540.6638955283997 + ] + }, + "frontendVersion": "1.42.8" + } +} \ No newline at end of file diff --git a/blueprints/Glow.json b/blueprints/Glow.json index 1dafb2d35..2bbfdee51 100644 --- a/blueprints/Glow.json +++ b/blueprints/Glow.json @@ -575,8 +575,9 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image Tools/Color adjust" + "category": "Image Tools/Color adjust", + "description": "Adds a glow/bloom effect around bright image areas via GPU fragment shader." } ] } -} +} \ No newline at end of file diff --git a/blueprints/Hue and Saturation.json b/blueprints/Hue and Saturation.json index 1a2df8937..cddf0154a 100644 --- a/blueprints/Hue and Saturation.json +++ b/blueprints/Hue and Saturation.json @@ -752,8 +752,9 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image Tools/Color adjust" + "category": "Image Tools/Color adjust", + "description": "Adjusts hue, saturation, and lightness of an image using a real-time GPU fragment shader." } ] } -} +} \ No newline at end of file diff --git a/blueprints/Image Blur.json b/blueprints/Image Blur.json index 3c7a784b0..0ca8d9931 100644 --- a/blueprints/Image Blur.json +++ b/blueprints/Image Blur.json @@ -374,7 +374,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image Tools/Blur" + "category": "Image Tools/Blur", + "description": "Applies Gaussian, Box, or Radial blur to soften images and create stylized depth or motion effects." } ] } diff --git a/blueprints/Image Captioning (gemini).json b/blueprints/Image Captioning (gemini).json index 98cfb8999..2fc5d6746 100644 --- a/blueprints/Image Captioning (gemini).json +++ b/blueprints/Image Captioning (gemini).json @@ -310,7 +310,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Text generation/Image Captioning" + "category": "Text generation/Image Captioning", + "description": "Generates descriptive captions for images using Google's Gemini multimodal LLM." } ] } diff --git a/blueprints/Image Channels.json b/blueprints/Image Channels.json index 9c7b675b2..b6fdff5be 100644 --- a/blueprints/Image Channels.json +++ b/blueprints/Image Channels.json @@ -315,8 +315,9 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image Tools/Color adjust" + "category": "Image Tools/Color adjust", + "description": "Manipulates individual RGBA channels for masking, compositing, and channel effects." } ] } -} +} \ No newline at end of file diff --git a/blueprints/Image Edit (FireRed Image Edit 1.1).json b/blueprints/Image Edit (FireRed Image Edit 1.1).json index c34246ce6..b82c7d18b 100644 --- a/blueprints/Image Edit (FireRed Image Edit 1.1).json +++ b/blueprints/Image Edit (FireRed Image Edit 1.1).json @@ -1,18 +1,18 @@ { "revision": 0, - "last_node_id": 172, + "last_node_id": 213, "last_link_id": 0, "nodes": [ { - "id": 172, - "type": "edf73971-14ee-4d39-b58e-46ce2a89d3d0", + "id": 213, + "type": "e35fbbeb-d7b1-46d1-a74e-959517d0fb1a", "pos": [ - 30, - 200 + -700, + -470 ], "size": [ 500, - 570 + 0 ], "flags": {}, "order": 2, @@ -105,44 +105,44 @@ "properties": { "proxyWidgets": [ [ - "118", + "208", "prompt" ], [ - "153", + "207", "value" ], [ - "130", + "210", "seed" ], [ - "128", + "205", "unet_name" ], [ - "115", + "203", "clip_name" ], [ - "116", + "202", "vae_name" ], [ - "151", + "204", "lora_name" ], [ - "130", + "210", "control_after_generate" ] ], + "cnr_id": "comfy-core", + "ver": "0.15.1", "ue_properties": { "widget_ue_connectable": {}, "input_ue_unconnectable": {} }, - "cnr_id": "comfy-core", - "ver": "0.15.1", "enableTabs": false, "tabWidth": 65, "tabXOffset": 10, @@ -160,12 +160,12 @@ "definitions": { "subgraphs": [ { - "id": "edf73971-14ee-4d39-b58e-46ce2a89d3d0", + "id": "e35fbbeb-d7b1-46d1-a74e-959517d0fb1a", "version": 1, "state": { "lastGroupId": 8, - "lastNodeId": 174, - "lastLinkId": 376, + "lastNodeId": 213, + "lastLinkId": 378, "lastRerouteId": 0 }, "revision": 0, @@ -183,8 +183,8 @@ "outputNode": { "id": -20, "bounding": [ - 1147.5, - -1215, + 1860, + -1340, 120, 60 ] @@ -327,26 +327,26 @@ ], "localized_name": "IMAGE", "pos": [ - 1167.5, - -1195 + 1880, + -1320 ] } ], "widgets": [], "nodes": [ { - "id": 120, + "id": 193, "type": "ModelSamplingAuraFlow", "pos": [ - 1060, - -1760 + 1010, + -1680 ], "size": [ 290, 110 ], "flags": {}, - "order": 8, + "order": 4, "mode": 0, "inputs": [ { @@ -376,13 +376,13 @@ } ], "properties": { + "Node name for S&R": "ModelSamplingAuraFlow", + "cnr_id": "comfy-core", + "ver": "0.5.1", "ue_properties": { "widget_ue_connectable": {}, "input_ue_unconnectable": {} }, - "cnr_id": "comfy-core", - "ver": "0.5.1", - "Node name for S&R": "ModelSamplingAuraFlow", "enableTabs": false, "tabWidth": 65, "tabXOffset": 10, @@ -396,7 +396,7 @@ ] }, { - "id": 154, + "id": 194, "type": "ComfySwitchNode", "pos": [ 680, @@ -407,7 +407,7 @@ 140 ], "flags": {}, - "order": 16, + "order": 5, "mode": 0, "inputs": [ { @@ -444,13 +444,13 @@ ], "title": "Switch (Model)", "properties": { + "Node name for S&R": "ComfySwitchNode", + "cnr_id": "comfy-core", + "ver": "0.15.1", "ue_properties": { "widget_ue_connectable": {}, "input_ue_unconnectable": {} }, - "cnr_id": "comfy-core", - "ver": "0.15.1", - "Node name for S&R": "ComfySwitchNode", "enableTabs": false, "tabWidth": 65, "tabXOffset": 10, @@ -464,7 +464,7 @@ ] }, { - "id": 155, + "id": 195, "type": "PrimitiveInt", "pos": [ 190, @@ -500,13 +500,13 @@ ], "title": "Int (Steps)", "properties": { + "Node name for S&R": "PrimitiveInt", + "cnr_id": "comfy-core", + "ver": "0.15.1", "ue_properties": { "widget_ue_connectable": {}, "input_ue_unconnectable": {} }, - "cnr_id": "comfy-core", - "ver": "0.15.1", - "Node name for S&R": "PrimitiveInt", "enableTabs": false, "tabWidth": 65, "tabXOffset": 10, @@ -521,18 +521,18 @@ ] }, { - "id": 123, + "id": 196, "type": "CFGNorm", "pos": [ - 1060, - -1590 + 1010, + -1510 ], "size": [ 290, 110 ], "flags": {}, - "order": 9, + "order": 6, "mode": 0, "inputs": [ { @@ -562,13 +562,13 @@ } ], "properties": { + "Node name for S&R": "CFGNorm", + "cnr_id": "comfy-core", + "ver": "0.5.1", "ue_properties": { "widget_ue_connectable": {}, "input_ue_unconnectable": {} }, - "cnr_id": "comfy-core", - "ver": "0.5.1", - "Node name for S&R": "CFGNorm", "enableTabs": false, "tabWidth": 65, "tabXOffset": 10, @@ -582,7 +582,7 @@ ] }, { - "id": 164, + "id": 197, "type": "ComfySwitchNode", "pos": [ 680, @@ -593,7 +593,7 @@ 130 ], "flags": {}, - "order": 18, + "order": 7, "mode": 0, "inputs": [ { @@ -630,13 +630,13 @@ ], "title": "Switch (CFG)", "properties": { + "Node name for S&R": "ComfySwitchNode", + "cnr_id": "comfy-core", + "ver": "0.15.1", "ue_properties": { "widget_ue_connectable": {}, "input_ue_unconnectable": {} }, - "cnr_id": "comfy-core", - "ver": "0.15.1", - "Node name for S&R": "ComfySwitchNode", "enableTabs": false, "tabWidth": 65, "tabXOffset": 10, @@ -650,7 +650,7 @@ ] }, { - "id": 156, + "id": 198, "type": "PrimitiveInt", "pos": [ 190, @@ -686,13 +686,13 @@ ], "title": "Float (Steps)", "properties": { + "Node name for S&R": "PrimitiveInt", + "cnr_id": "comfy-core", + "ver": "0.15.1", "ue_properties": { "widget_ue_connectable": {}, "input_ue_unconnectable": {} }, - "cnr_id": "comfy-core", - "ver": "0.15.1", - "Node name for S&R": "PrimitiveInt", "enableTabs": false, "tabWidth": 65, "tabXOffset": 10, @@ -707,7 +707,7 @@ ] }, { - "id": 162, + "id": 199, "type": "PrimitiveFloat", "pos": [ 190, @@ -743,13 +743,13 @@ ], "title": "Float (CFG)", "properties": { + "Node name for S&R": "PrimitiveFloat", + "cnr_id": "comfy-core", + "ver": "0.15.1", "ue_properties": { "widget_ue_connectable": {}, "input_ue_unconnectable": {} }, - "cnr_id": "comfy-core", - "ver": "0.15.1", - "Node name for S&R": "PrimitiveFloat", "enableTabs": false, "tabWidth": 65, "tabXOffset": 10, @@ -763,7 +763,7 @@ ] }, { - "id": 163, + "id": 200, "type": "PrimitiveFloat", "pos": [ 190, @@ -799,13 +799,13 @@ ], "title": "Float (CFG)", "properties": { + "Node name for S&R": "PrimitiveFloat", + "cnr_id": "comfy-core", + "ver": "0.15.1", "ue_properties": { "widget_ue_connectable": {}, "input_ue_unconnectable": {} }, - "cnr_id": "comfy-core", - "ver": "0.15.1", - "Node name for S&R": "PrimitiveFloat", "enableTabs": false, "tabWidth": 65, "tabXOffset": 10, @@ -819,7 +819,7 @@ ] }, { - "id": 157, + "id": 201, "type": "ComfySwitchNode", "pos": [ 680, @@ -830,7 +830,7 @@ 130 ], "flags": {}, - "order": 17, + "order": 8, "mode": 0, "inputs": [ { @@ -867,13 +867,13 @@ ], "title": "Switch (Steps)", "properties": { + "Node name for S&R": "ComfySwitchNode", + "cnr_id": "comfy-core", + "ver": "0.15.1", "ue_properties": { "widget_ue_connectable": {}, "input_ue_unconnectable": {} }, - "cnr_id": "comfy-core", - "ver": "0.15.1", - "Node name for S&R": "ComfySwitchNode", "enableTabs": false, "tabWidth": 65, "tabXOffset": 10, @@ -887,11 +887,11 @@ ] }, { - "id": 116, + "id": 202, "type": "VAELoader", "pos": [ - -950, - -1040 + -960, + -1100 ], "size": [ 400, @@ -900,7 +900,7 @@ "flags": { "collapsed": false }, - "order": 5, + "order": 9, "mode": 0, "inputs": [ { @@ -928,45 +928,45 @@ } ], "properties": { - "ue_properties": { - "widget_ue_connectable": {}, - "input_ue_unconnectable": {} - }, + "Node name for S&R": "VAELoader", "cnr_id": "comfy-core", "ver": "0.5.1", - "Node name for S&R": "VAELoader", - "enableTabs": false, - "tabWidth": 65, - "tabXOffset": 10, - "hasSecondTab": false, - "secondTabText": "Send Back", - "secondTabOffset": 80, - "secondTabWidth": 65, "models": [ { "name": "qwen_image_vae.safetensors", "url": "https://huggingface.co/FireRedTeam/FireRed-Image-Edit-1.0-ComfyUI/resolve/main/qwen_image_vae.safetensors", "directory": "vae" } - ] + ], + "ue_properties": { + "widget_ue_connectable": {}, + "input_ue_unconnectable": {} + }, + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 }, "widgets_values": [ "qwen_image_vae.safetensors" ] }, { - "id": 115, + "id": 203, "type": "CLIPLoader", "pos": [ -960, - -1370 + -1400 ], "size": [ 400, 150 ], "flags": {}, - "order": 4, + "order": 10, "mode": 0, "inputs": [ { @@ -1010,27 +1010,27 @@ } ], "properties": { - "ue_properties": { - "widget_ue_connectable": {}, - "input_ue_unconnectable": {} - }, + "Node name for S&R": "CLIPLoader", "cnr_id": "comfy-core", "ver": "0.5.1", - "Node name for S&R": "CLIPLoader", - "enableTabs": false, - "tabWidth": 65, - "tabXOffset": 10, - "hasSecondTab": false, - "secondTabText": "Send Back", - "secondTabOffset": 80, - "secondTabWidth": 65, "models": [ { "name": "qwen_2.5_vl_7b_fp8_scaled.safetensors", "url": "https://huggingface.co/Comfy-Org/HunyuanVideo_1.5_repackaged/resolve/main/split_files/text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors", "directory": "text_encoders" } - ] + ], + "ue_properties": { + "widget_ue_connectable": {}, + "input_ue_unconnectable": {} + }, + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 }, "widgets_values": [ "qwen_2.5_vl_7b_fp8_scaled.safetensors", @@ -1039,7 +1039,7 @@ ] }, { - "id": 151, + "id": 204, "type": "LoraLoaderModelOnly", "pos": [ 100, @@ -1050,7 +1050,7 @@ 140 ], "flags": {}, - "order": 14, + "order": 11, "mode": 0, "inputs": [ { @@ -1089,27 +1089,27 @@ } ], "properties": { - "ue_properties": { - "widget_ue_connectable": {}, - "input_ue_unconnectable": {} - }, + "Node name for S&R": "LoraLoaderModelOnly", "cnr_id": "comfy-core", "ver": "0.15.1", - "Node name for S&R": "LoraLoaderModelOnly", - "enableTabs": false, - "tabWidth": 65, - "tabXOffset": 10, - "hasSecondTab": false, - "secondTabText": "Send Back", - "secondTabOffset": 80, - "secondTabWidth": 65, "models": [ { "name": "FireRed-Image-Edit-1.0-Lightning-8steps-v1.0.safetensors", "url": "https://huggingface.co/FireRedTeam/FireRed-Image-Edit-1.0-ComfyUI/resolve/main/FireRed-Image-Edit-1.0-Lightning-8steps-v1.0.safetensors", "directory": "loras" } - ] + ], + "ue_properties": { + "widget_ue_connectable": {}, + "input_ue_unconnectable": {} + }, + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 }, "widgets_values": [ "FireRed-Image-Edit-1.0-Lightning-8steps-v1.0.safetensors", @@ -1117,7 +1117,7 @@ ] }, { - "id": 128, + "id": 205, "type": "UNETLoader", "pos": [ -960, @@ -1163,27 +1163,27 @@ } ], "properties": { - "ue_properties": { - "widget_ue_connectable": {}, - "input_ue_unconnectable": {} - }, + "Node name for S&R": "UNETLoader", "cnr_id": "comfy-core", "ver": "0.5.1", - "Node name for S&R": "UNETLoader", - "enableTabs": false, - "tabWidth": 65, - "tabXOffset": 10, - "hasSecondTab": false, - "secondTabText": "Send Back", - "secondTabOffset": 80, - "secondTabWidth": 65, "models": [ { "name": "FireRed-Image-Edit-1.1-transformer.safetensors", "url": "https://huggingface.co/FireRedTeam/FireRed-Image-Edit-1.1-ComfyUI/resolve/main/FireRed-Image-Edit-1.1-transformer.safetensors", "directory": "diffusion_models" } - ] + ], + "ue_properties": { + "widget_ue_connectable": {}, + "input_ue_unconnectable": {} + }, + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 }, "widgets_values": [ "FireRed-Image-Edit-1.1-transformer.safetensors", @@ -1191,7 +1191,7 @@ ] }, { - "id": 125, + "id": 206, "type": "VAEEncode", "pos": [ -390, @@ -1202,7 +1202,7 @@ 100 ], "flags": {}, - "order": 10, + "order": 13, "mode": 0, "inputs": [ { @@ -1229,13 +1229,13 @@ } ], "properties": { + "Node name for S&R": "VAEEncode", + "cnr_id": "comfy-core", + "ver": "0.5.1", "ue_properties": { "widget_ue_connectable": {}, "input_ue_unconnectable": {} }, - "cnr_id": "comfy-core", - "ver": "0.5.1", - "Node name for S&R": "VAEEncode", "enableTabs": false, "tabWidth": 65, "tabXOffset": 10, @@ -1246,7 +1246,7 @@ } }, { - "id": 153, + "id": 207, "type": "PrimitiveBoolean", "pos": [ 160, @@ -1257,7 +1257,7 @@ 100 ], "flags": {}, - "order": 15, + "order": 14, "mode": 0, "inputs": [ { @@ -1284,13 +1284,13 @@ ], "title": "Enable Lightning LoRA?", "properties": { + "Node name for S&R": "PrimitiveBoolean", + "cnr_id": "comfy-core", + "ver": "0.15.1", "ue_properties": { "widget_ue_connectable": {}, "input_ue_unconnectable": {} }, - "cnr_id": "comfy-core", - "ver": "0.15.1", - "Node name for S&R": "PrimitiveBoolean", "enableTabs": false, "tabWidth": 65, "tabXOffset": 10, @@ -1304,7 +1304,7 @@ ] }, { - "id": 118, + "id": 208, "type": "TextEncodeQwenImageEditPlus", "pos": [ -480, @@ -1315,7 +1315,7 @@ 370 ], "flags": {}, - "order": 7, + "order": 15, "mode": 0, "inputs": [ { @@ -1374,13 +1374,13 @@ ], "title": "TextEncodeQwenImageEditPlus (Positive)", "properties": { + "Node name for S&R": "TextEncodeQwenImageEditPlus", + "cnr_id": "comfy-core", + "ver": "0.5.1", "ue_properties": { "widget_ue_connectable": {}, "input_ue_unconnectable": {} }, - "cnr_id": "comfy-core", - "ver": "0.5.1", - "Node name for S&R": "TextEncodeQwenImageEditPlus", "enableTabs": false, "tabWidth": 65, "tabXOffset": 10, @@ -1396,7 +1396,7 @@ "bgcolor": "#353" }, { - "id": 117, + "id": 209, "type": "TextEncodeQwenImageEditPlus", "pos": [ -470, @@ -1407,7 +1407,7 @@ 290 ], "flags": {}, - "order": 6, + "order": 16, "mode": 0, "inputs": [ { @@ -1465,13 +1465,13 @@ } ], "properties": { + "Node name for S&R": "TextEncodeQwenImageEditPlus", + "cnr_id": "comfy-core", + "ver": "0.5.1", "ue_properties": { "widget_ue_connectable": {}, "input_ue_unconnectable": {} }, - "cnr_id": "comfy-core", - "ver": "0.5.1", - "Node name for S&R": "TextEncodeQwenImageEditPlus", "enableTabs": false, "tabWidth": 65, "tabXOffset": 10, @@ -1487,18 +1487,18 @@ "bgcolor": "#535" }, { - "id": 130, + "id": 210, "type": "KSampler", "pos": [ - 1060, - -1420 + 1010, + -1340 ], "size": [ 270, 480 ], "flags": {}, - "order": 13, + "order": 17, "mode": 0, "inputs": [ { @@ -1591,13 +1591,13 @@ } ], "properties": { + "Node name for S&R": "KSampler", + "cnr_id": "comfy-core", + "ver": "0.5.1", "ue_properties": { "widget_ue_connectable": {}, "input_ue_unconnectable": {} }, - "cnr_id": "comfy-core", - "ver": "0.5.1", - "Node name for S&R": "KSampler", "enableTabs": false, "tabWidth": 65, "tabXOffset": 10, @@ -1617,11 +1617,11 @@ ] }, { - "id": 126, + "id": 211, "type": "VAEDecode", "pos": [ - 1360, - -1420 + 1440, + -1340 ], "size": [ 230, @@ -1630,7 +1630,7 @@ "flags": { "collapsed": false }, - "order": 11, + "order": 18, "mode": 0, "inputs": [ { @@ -1658,13 +1658,13 @@ } ], "properties": { + "Node name for S&R": "VAEDecode", + "cnr_id": "comfy-core", + "ver": "0.5.1", "ue_properties": { "widget_ue_connectable": {}, "input_ue_unconnectable": {} }, - "cnr_id": "comfy-core", - "ver": "0.5.1", - "Node name for S&R": "VAEDecode", "enableTabs": false, "tabWidth": 65, "tabXOffset": 10, @@ -1675,7 +1675,7 @@ } }, { - "id": 174, + "id": 212, "type": "ResizeImageMaskNode", "pos": [ -900, @@ -1736,18 +1736,18 @@ } ], "properties": { + "Node name for S&R": "ResizeImageMaskNode", + "cnr_id": "comfy-core", + "ver": "0.18.1", "ue_properties": { "widget_ue_connectable": {}, "input_ue_unconnectable": {} - }, - "cnr_id": "comfy-core", - "ver": "0.18.1", - "Node name for S&R": "ResizeImageMaskNode" + } }, "widgets_values": [ "scale total pixels", 1, - "area" + "lanczos" ] } ], @@ -1808,207 +1808,207 @@ "links": [ { "id": 326, - "origin_id": 154, + "origin_id": 194, "origin_slot": 0, - "target_id": 120, + "target_id": 193, "target_slot": 0, "type": "MODEL" }, { "id": 324, - "origin_id": 128, + "origin_id": 205, "origin_slot": 0, - "target_id": 154, + "target_id": 194, "target_slot": 0, "type": "MODEL" }, { "id": 325, - "origin_id": 151, + "origin_id": 204, "origin_slot": 0, - "target_id": 154, + "target_id": 194, "target_slot": 1, "type": "MODEL" }, { "id": 323, - "origin_id": 153, + "origin_id": 207, "origin_slot": 0, - "target_id": 154, + "target_id": 194, "target_slot": 2, "type": "BOOLEAN" }, { "id": 294, - "origin_id": 120, + "origin_id": 193, "origin_slot": 0, - "target_id": 123, + "target_id": 196, "target_slot": 0, "type": "MODEL" }, { "id": 333, - "origin_id": 162, + "origin_id": 199, "origin_slot": 0, - "target_id": 164, + "target_id": 197, "target_slot": 0, "type": "FLOAT" }, { "id": 334, - "origin_id": 163, + "origin_id": 200, "origin_slot": 0, - "target_id": 164, + "target_id": 197, "target_slot": 1, "type": "FLOAT" }, { "id": 336, - "origin_id": 153, + "origin_id": 207, "origin_slot": 0, - "target_id": 164, + "target_id": 197, "target_slot": 2, "type": "BOOLEAN" }, { "id": 329, - "origin_id": 155, + "origin_id": 195, "origin_slot": 0, - "target_id": 157, + "target_id": 201, "target_slot": 0, "type": "INT" }, { "id": 337, - "origin_id": 156, + "origin_id": 198, "origin_slot": 0, - "target_id": 157, + "target_id": 201, "target_slot": 1, "type": "INT" }, { "id": 330, - "origin_id": 153, + "origin_id": 207, "origin_slot": 0, - "target_id": 157, + "target_id": 201, "target_slot": 2, "type": "BOOLEAN" }, { "id": 297, - "origin_id": 115, + "origin_id": 203, "origin_slot": 0, - "target_id": 117, + "target_id": 209, "target_slot": 0, "type": "CLIP" }, { "id": 299, - "origin_id": 116, + "origin_id": 202, "origin_slot": 0, - "target_id": 117, + "target_id": 209, "target_slot": 1, "type": "VAE" }, { "id": 316, - "origin_id": 128, + "origin_id": 205, "origin_slot": 0, - "target_id": 151, + "target_id": 204, "target_slot": 0, "type": "MODEL" }, { "id": 296, - "origin_id": 115, + "origin_id": 203, "origin_slot": 0, - "target_id": 118, + "target_id": 208, "target_slot": 0, "type": "CLIP" }, { "id": 298, - "origin_id": 116, + "origin_id": 202, "origin_slot": 0, - "target_id": 118, + "target_id": 208, "target_slot": 1, "type": "VAE" }, { "id": 300, - "origin_id": 116, + "origin_id": 202, "origin_slot": 0, - "target_id": 125, + "target_id": 206, "target_slot": 1, "type": "VAE" }, { "id": 295, - "origin_id": 123, + "origin_id": 196, "origin_slot": 0, - "target_id": 130, + "target_id": 210, "target_slot": 0, "type": "MODEL" }, { "id": 312, - "origin_id": 118, + "origin_id": 208, "origin_slot": 0, - "target_id": 130, + "target_id": 210, "target_slot": 1, "type": "CONDITIONING" }, { "id": 313, - "origin_id": 117, + "origin_id": 209, "origin_slot": 0, - "target_id": 130, + "target_id": 210, "target_slot": 2, "type": "CONDITIONING" }, { "id": 303, - "origin_id": 125, + "origin_id": 206, "origin_slot": 0, - "target_id": 130, + "target_id": 210, "target_slot": 3, "type": "LATENT" }, { "id": 345, - "origin_id": 157, + "origin_id": 201, "origin_slot": 0, - "target_id": 130, + "target_id": 210, "target_slot": 5, "type": "INT" }, { "id": 335, - "origin_id": 164, + "origin_id": 197, "origin_slot": 0, - "target_id": 130, + "target_id": 210, "target_slot": 6, "type": "FLOAT" }, { "id": 273, - "origin_id": 130, + "origin_id": 210, "origin_slot": 0, - "target_id": 126, + "target_id": 211, "target_slot": 0, "type": "LATENT" }, { "id": 314, - "origin_id": 116, + "origin_id": 202, "origin_slot": 0, - "target_id": 126, + "target_id": 211, "target_slot": 1, "type": "VAE" }, { "id": 292, - "origin_id": 126, + "origin_id": 211, "origin_slot": 0, "target_id": -20, "target_slot": 0, @@ -2018,7 +2018,7 @@ "id": 355, "origin_id": -10, "origin_slot": 1, - "target_id": 118, + "target_id": 208, "target_slot": 3, "type": "IMAGE" }, @@ -2026,7 +2026,7 @@ "id": 356, "origin_id": -10, "origin_slot": 1, - "target_id": 117, + "target_id": 209, "target_slot": 3, "type": "IMAGE" }, @@ -2034,7 +2034,7 @@ "id": 357, "origin_id": -10, "origin_slot": 2, - "target_id": 118, + "target_id": 208, "target_slot": 4, "type": "IMAGE" }, @@ -2042,7 +2042,7 @@ "id": 358, "origin_id": -10, "origin_slot": 2, - "target_id": 117, + "target_id": 209, "target_slot": 4, "type": "IMAGE" }, @@ -2050,7 +2050,7 @@ "id": 359, "origin_id": -10, "origin_slot": 3, - "target_id": 118, + "target_id": 208, "target_slot": 5, "type": "STRING" }, @@ -2058,31 +2058,31 @@ "id": 364, "origin_id": -10, "origin_slot": 4, - "target_id": 153, + "target_id": 207, "target_slot": 0, "type": "BOOLEAN" }, { "id": 368, - "origin_id": 174, + "origin_id": 212, "origin_slot": 0, - "target_id": 125, + "target_id": 206, "target_slot": 0, "type": "IMAGE" }, { "id": 369, - "origin_id": 174, + "origin_id": 212, "origin_slot": 0, - "target_id": 118, + "target_id": 208, "target_slot": 2, "type": "IMAGE" }, { "id": 370, - "origin_id": 174, + "origin_id": 212, "origin_slot": 0, - "target_id": 117, + "target_id": 209, "target_slot": 2, "type": "IMAGE" }, @@ -2090,7 +2090,7 @@ "id": 371, "origin_id": -10, "origin_slot": 0, - "target_id": 174, + "target_id": 212, "target_slot": 0, "type": "IMAGE" }, @@ -2098,7 +2098,7 @@ "id": 372, "origin_id": -10, "origin_slot": 5, - "target_id": 130, + "target_id": 210, "target_slot": 4, "type": "INT" }, @@ -2106,7 +2106,7 @@ "id": 373, "origin_id": -10, "origin_slot": 6, - "target_id": 128, + "target_id": 205, "target_slot": 0, "type": "COMBO" }, @@ -2114,7 +2114,7 @@ "id": 374, "origin_id": -10, "origin_slot": 7, - "target_id": 115, + "target_id": 203, "target_slot": 0, "type": "COMBO" }, @@ -2122,7 +2122,7 @@ "id": 375, "origin_id": -10, "origin_slot": 8, - "target_id": 116, + "target_id": 202, "target_slot": 0, "type": "COMBO" }, @@ -2130,7 +2130,7 @@ "id": 376, "origin_id": -10, "origin_slot": 9, - "target_id": 151, + "target_id": 204, "target_slot": 1, "type": "COMBO" } @@ -2138,7 +2138,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image generation and editing/Edit image" + "category": "Image generation and editing/Edit image", + "description": "Edits images via text instructions using FireRed Image Edit 1.1, a diffusion-based instruction-following editing model." } ] }, diff --git a/blueprints/Image Edit (Flux.2 Dev).json b/blueprints/Image Edit (Flux.2 Dev).json new file mode 100644 index 000000000..92827bf17 --- /dev/null +++ b/blueprints/Image Edit (Flux.2 Dev).json @@ -0,0 +1,2050 @@ +{ + "revision": 0, + "last_node_id": 139, + "last_link_id": 0, + "nodes": [ + { + "id": 139, + "type": "41b0c117-7470-454c-914e-b8742dc06d62", + "pos": [ + -650, + 570 + ], + "size": [ + 400, + 0 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "label": "image", + "localized_name": "pixels", + "name": "pixels", + "type": "IMAGE", + "link": null + }, + { + "label": "prompt", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "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 + }, + { + "label": "enable_turbo_mode", + "name": "value", + "type": "BOOLEAN", + "widget": { + "name": "value" + }, + "link": null + }, + { + "label": "turbo_lora", + "name": "lora_name", + "type": "COMBO", + "widget": { + "name": "lora_name" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [] + } + ], + "properties": { + "proxyWidgets": [ + [ + "123", + "text" + ], + [ + "129", + "unet_name" + ], + [ + "124", + "clip_name" + ], + [ + "121", + "vae_name" + ], + [ + "138", + "value" + ], + [ + "128", + "lora_name" + ], + [ + "125", + "noise_seed" + ], + [ + "125", + "control_after_generate" + ] + ], + "cnr_id": "comfy-core", + "ver": "0.7.0", + "ue_properties": { + "widget_ue_connectable": { + "text": true, + "value": true, + "lora_name": true + }, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [], + "title": "Image Edit (Flux.2 Dev)" + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "41b0c117-7470-454c-914e-b8742dc06d62", + "version": 1, + "state": { + "lastGroupId": 8, + "lastNodeId": 139, + "lastLinkId": 194, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Image Edit (Flux.2 Dev)", + "inputNode": { + "id": -10, + "bounding": [ + -1520, + 400, + 151.744140625, + 180 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + 1240, + 420, + 120, + 60 + ] + }, + "inputs": [ + { + "id": "fc74acd5-30a9-410b-abb5-4a4171ba3d25", + "name": "pixels", + "type": "IMAGE", + "linkIds": [ + 126, + 169 + ], + "localized_name": "pixels", + "label": "image", + "pos": [ + -1388.255859375, + 420 + ] + }, + { + "id": "3e69affa-397b-4d52-82d7-68dfcef9e761", + "name": "text", + "type": "STRING", + "linkIds": [ + 168 + ], + "label": "prompt", + "pos": [ + -1388.255859375, + 440 + ] + }, + { + "id": "2f016a8a-fb3e-4cb9-97f2-a991defe4fa2", + "name": "unet_name", + "type": "COMBO", + "linkIds": [ + 177 + ], + "pos": [ + -1388.255859375, + 460 + ] + }, + { + "id": "799b9dc7-0c90-4b19-9a13-e01d896bea1f", + "name": "clip_name", + "type": "COMBO", + "linkIds": [ + 178 + ], + "pos": [ + -1388.255859375, + 480 + ] + }, + { + "id": "e58a83c9-1b93-4378-9598-f24068820313", + "name": "vae_name", + "type": "COMBO", + "linkIds": [ + 179 + ], + "pos": [ + -1388.255859375, + 500 + ] + }, + { + "id": "8335a4a9-0ce4-4e67-a641-1c9d7a762977", + "name": "value", + "type": "BOOLEAN", + "linkIds": [ + 191 + ], + "label": "enable_turbo_mode", + "pos": [ + -1388.255859375, + 520 + ] + }, + { + "id": "890b22b4-44a7-4707-912a-ca8b4ee7b7c9", + "name": "lora_name", + "type": "COMBO", + "linkIds": [ + 192 + ], + "label": "turbo_lora", + "pos": [ + -1388.255859375, + 540 + ] + } + ], + "outputs": [ + { + "id": "3eaa05d6-4960-4a7c-bf2a-8b585fbb7c9c", + "name": "IMAGE", + "type": "IMAGE", + "linkIds": [ + 9 + ], + "localized_name": "IMAGE", + "pos": [ + 1260, + 440 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 118, + "type": "Flux2Scheduler", + "pos": [ + 540, + 430 + ], + "size": [ + 230, + 170 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "localized_name": "steps", + "name": "steps", + "type": "INT", + "widget": { + "name": "steps" + }, + "link": 188 + }, + { + "localized_name": "width", + "name": "width", + "type": "INT", + "widget": { + "name": "width" + }, + "link": 170 + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "widget": { + "name": "height" + }, + "link": 172 + } + ], + "outputs": [ + { + "localized_name": "SIGMAS", + "name": "SIGMAS", + "type": "SIGMAS", + "links": [ + 132 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "Flux2Scheduler", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 20, + 1248, + 832 + ] + }, + { + "id": 119, + "type": "BasicGuider", + "pos": [ + 530, + 120 + ], + "size": [ + 230, + 100 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 185 + }, + { + "localized_name": "conditioning", + "name": "conditioning", + "type": "CONDITIONING", + "link": 166 + } + ], + "outputs": [ + { + "localized_name": "GUIDER", + "name": "GUIDER", + "type": "GUIDER", + "slot_index": 0, + "links": [ + 30 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "BasicGuider", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 120, + "type": "KSamplerSelect", + "pos": [ + 530, + 270 + ], + "size": [ + 230, + 110 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "sampler_name", + "name": "sampler_name", + "type": "COMBO", + "widget": { + "name": "sampler_name" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "SAMPLER", + "name": "SAMPLER", + "type": "SAMPLER", + "links": [ + 19 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "KSamplerSelect", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "euler" + ] + }, + { + "id": 121, + "type": "VAELoader", + "pos": [ + -970, + 390 + ], + "size": [ + 300, + 110 + ], + "flags": {}, + "order": 6, + "mode": 0, + "inputs": [ + { + "localized_name": "vae_name", + "name": "vae_name", + "type": "COMBO", + "widget": { + "name": "vae_name" + }, + "link": 179 + } + ], + "outputs": [ + { + "localized_name": "VAE", + "name": "VAE", + "type": "VAE", + "slot_index": 0, + "links": [ + 127, + 159 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "VAELoader", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "models": [ + { + "name": "full_encoder_small_decoder.safetensors", + "url": "https://huggingface.co/black-forest-labs/FLUX.2-small-decoder/resolve/main/full_encoder_small_decoder.safetensors", + "directory": "vae" + } + ] + }, + "widgets_values": [ + "full_encoder_small_decoder.safetensors" + ] + }, + { + "id": 122, + "type": "SamplerCustomAdvanced", + "pos": [ + 790, + -50 + ], + "size": [ + 280, + 170 + ], + "flags": {}, + "order": 7, + "mode": 0, + "inputs": [ + { + "localized_name": "noise", + "name": "noise", + "type": "NOISE", + "link": 37 + }, + { + "localized_name": "guider", + "name": "guider", + "type": "GUIDER", + "link": 30 + }, + { + "localized_name": "sampler", + "name": "sampler", + "type": "SAMPLER", + "link": 19 + }, + { + "localized_name": "sigmas", + "name": "sigmas", + "type": "SIGMAS", + "link": 132 + }, + { + "localized_name": "latent_image", + "name": "latent_image", + "type": "LATENT", + "link": 161 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "LATENT", + "slot_index": 0, + "links": [ + 24 + ] + }, + { + "localized_name": "denoised_output", + "name": "denoised_output", + "type": "LATENT", + "links": null + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "SamplerCustomAdvanced", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 123, + "type": "CLIPTextEncode", + "pos": [ + -630, + -50 + ], + "size": [ + 430, + 360 + ], + "flags": {}, + "order": 8, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 117 + }, + { + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": 168 + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "slot_index": 0, + "links": [ + 41 + ] + } + ], + "title": "CLIP Text Encode (Positive Prompt)", + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "CLIPTextEncode", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "" + ], + "color": "#232", + "bgcolor": "#353" + }, + { + "id": 124, + "type": "CLIPLoader", + "pos": [ + -970, + 160 + ], + "size": [ + 300, + 150 + ], + "flags": {}, + "order": 9, + "mode": 0, + "inputs": [ + { + "localized_name": "clip_name", + "name": "clip_name", + "type": "COMBO", + "widget": { + "name": "clip_name" + }, + "link": 178 + }, + { + "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": [ + 117 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "CLIPLoader", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "models": [ + { + "name": "mistral_3_small_flux2_bf16.safetensors", + "url": "https://huggingface.co/Comfy-Org/flux2-dev/resolve/main/split_files/text_encoders/mistral_3_small_flux2_bf16.safetensors", + "directory": "text_encoders" + } + ] + }, + "widgets_values": [ + "mistral_3_small_flux2_bf16.safetensors", + "flux2", + "default" + ] + }, + { + "id": 125, + "type": "RandomNoise", + "pos": [ + 530, + -50 + ], + "size": [ + 230, + 110 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "noise_seed", + "name": "noise_seed", + "type": "INT", + "widget": { + "name": "noise_seed" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "NOISE", + "name": "NOISE", + "type": "NOISE", + "links": [ + 37 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "RandomNoise", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 342971778941390, + "randomize" + ] + }, + { + "id": 126, + "type": "VAEDecode", + "pos": [ + 830, + 410 + ], + "size": [ + 230, + 100 + ], + "flags": {}, + "order": 10, + "mode": 0, + "inputs": [ + { + "localized_name": "samples", + "name": "samples", + "type": "LATENT", + "link": 24 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 159 + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "slot_index": 0, + "links": [ + 9 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "VAEDecode", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 127, + "type": "FluxGuidance", + "pos": [ + -520, + 390 + ], + "size": [ + 320, + 110 + ], + "flags": {}, + "order": 11, + "mode": 0, + "inputs": [ + { + "localized_name": "conditioning", + "name": "conditioning", + "type": "CONDITIONING", + "link": 41 + }, + { + "localized_name": "guidance", + "name": "guidance", + "type": "FLOAT", + "widget": { + "name": "guidance" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "slot_index": 0, + "links": [ + 144 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "FluxGuidance", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 4 + ], + "color": "#233", + "bgcolor": "#355" + }, + { + "id": 128, + "type": "LoraLoaderModelOnly", + "pos": [ + -150, + 200 + ], + "size": [ + 300, + 140 + ], + "flags": {}, + "order": 12, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 181 + }, + { + "localized_name": "lora_name", + "name": "lora_name", + "type": "COMBO", + "widget": { + "name": "lora_name" + }, + "link": 192 + }, + { + "localized_name": "strength_model", + "name": "strength_model", + "type": "FLOAT", + "widget": { + "name": "strength_model" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "links": [ + 183 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.7.0", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "LoraLoaderModelOnly", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "models": [ + { + "name": "Flux_2-Turbo-LoRA_comfyui.safetensors", + "url": "https://huggingface.co/ByteZSzn/Flux.2-Turbo-ComfyUI/resolve/main/Flux_2-Turbo-LoRA_comfyui.safetensors", + "directory": "loras" + } + ] + }, + "widgets_values": [ + "Flux_2-Turbo-LoRA_comfyui.safetensors", + 1 + ] + }, + { + "id": 129, + "type": "UNETLoader", + "pos": [ + -970, + -40 + ], + "size": [ + 300, + 110 + ], + "flags": {}, + "order": 13, + "mode": 0, + "inputs": [ + { + "localized_name": "unet_name", + "name": "unet_name", + "type": "COMBO", + "widget": { + "name": "unet_name" + }, + "link": 177 + }, + { + "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": [ + 181, + 184 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "UNETLoader", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "models": [ + { + "name": "flux2_dev_fp8mixed.safetensors", + "url": "https://huggingface.co/Comfy-Org/flux2-dev/resolve/main/split_files/diffusion_models/flux2_dev_fp8mixed.safetensors", + "directory": "diffusion_models" + } + ] + }, + "widgets_values": [ + "flux2_dev_fp8mixed.safetensors", + "default" + ] + }, + { + "id": 130, + "type": "ComfySwitchNode", + "pos": [ + 220, + 10 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 14, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 184 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 183 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 190 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 185 + ] + } + ], + "title": "Switch(model)", + "properties": { + "cnr_id": "comfy-core", + "ver": "0.18.1", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "ComfySwitchNode", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + false + ] + }, + { + "id": 131, + "type": "PrimitiveInt", + "pos": [ + -150, + 430 + ], + "size": [ + 300, + 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": [ + 186 + ] + } + ], + "title": "Steps", + "properties": { + "cnr_id": "comfy-core", + "ver": "0.18.1", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "PrimitiveInt", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 8, + "fixed" + ] + }, + { + "id": 132, + "type": "PrimitiveInt", + "pos": [ + -150, + -50 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "INT", + "widget": { + "name": "value" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 187 + ] + } + ], + "title": "Steps", + "properties": { + "cnr_id": "comfy-core", + "ver": "0.18.1", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "PrimitiveInt", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 20, + "fixed" + ] + }, + { + "id": 133, + "type": "ComfySwitchNode", + "pos": [ + 220, + 280 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 15, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 187 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 186 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 189 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 188 + ] + } + ], + "title": "Switch(steps)", + "properties": { + "cnr_id": "comfy-core", + "ver": "0.18.1", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "ComfySwitchNode", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + false + ] + }, + { + "id": 134, + "type": "EmptyFlux2LatentImage", + "pos": [ + 530, + 790 + ], + "size": [ + 270, + 170 + ], + "flags": {}, + "order": 16, + "mode": 0, + "inputs": [ + { + "localized_name": "width", + "name": "width", + "type": "INT", + "widget": { + "name": "width" + }, + "link": 171 + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "widget": { + "name": "height" + }, + "link": 173 + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "widget": { + "name": "batch_size" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "LATENT", + "name": "LATENT", + "type": "LATENT", + "links": [ + 161 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "EmptyFlux2LatentImage", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 1248, + 832, + 1 + ] + }, + { + "id": 135, + "type": "GetImageSize", + "pos": [ + -100, + 810 + ], + "size": [ + 230, + 120 + ], + "flags": {}, + "order": 17, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 169 + } + ], + "outputs": [ + { + "localized_name": "width", + "name": "width", + "type": "INT", + "links": [ + 170, + 171 + ] + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "links": [ + 172, + 173 + ] + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "links": null + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.7.0", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "GetImageSize", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 136, + "type": "VAEEncode", + "pos": [ + -910, + 600 + ], + "size": [ + 230, + 100 + ], + "flags": { + "collapsed": true + }, + "order": 18, + "mode": 0, + "inputs": [ + { + "localized_name": "pixels", + "name": "pixels", + "type": "IMAGE", + "link": 126 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 127 + } + ], + "outputs": [ + { + "localized_name": "LATENT", + "name": "LATENT", + "type": "LATENT", + "links": [ + 125 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "VAEEncode", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 137, + "type": "ReferenceLatent", + "pos": [ + -470, + 580 + ], + "size": [ + 230, + 100 + ], + "flags": {}, + "order": 19, + "mode": 0, + "inputs": [ + { + "localized_name": "conditioning", + "name": "conditioning", + "type": "CONDITIONING", + "link": 144 + }, + { + "localized_name": "latent", + "name": "latent", + "shape": 7, + "type": "LATENT", + "link": 125 + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "links": [ + 166 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "ReferenceLatent", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 138, + "type": "PrimitiveBoolean", + "pos": [ + -130, + 640 + ], + "size": [ + 270, + 100 + ], + "flags": {}, + "order": 20, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "BOOLEAN", + "widget": { + "name": "value" + }, + "link": 191 + } + ], + "outputs": [ + { + "localized_name": "BOOLEAN", + "name": "BOOLEAN", + "type": "BOOLEAN", + "links": [ + 189, + 190 + ] + } + ], + "title": "Enable 8 steps lora", + "properties": { + "cnr_id": "comfy-core", + "ver": "0.18.1", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "PrimitiveBoolean", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + false + ] + } + ], + "groups": [ + { + "id": 1, + "title": "Models", + "bounding": [ + -980, + -120, + 320, + 640 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 2, + "title": "Custom sampler", + "bounding": [ + 520, + -120, + 590, + 740 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 3, + "title": "Image size", + "bounding": [ + 510, + 690, + 590, + 290 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 4, + "title": "Prompt", + "bounding": [ + -640, + -120, + 450, + 640 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 7, + "title": "Original", + "bounding": [ + -160, + -120, + 340, + 230 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 8, + "title": "8 Steps LoRA", + "bounding": [ + -160, + 130, + 340, + 430 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + } + ], + "links": [ + { + "id": 41, + "origin_id": 123, + "origin_slot": 0, + "target_id": 127, + "target_slot": 0, + "type": "CONDITIONING" + }, + { + "id": 144, + "origin_id": 127, + "origin_slot": 0, + "target_id": 137, + "target_slot": 0, + "type": "CONDITIONING" + }, + { + "id": 125, + "origin_id": 136, + "origin_slot": 0, + "target_id": 137, + "target_slot": 1, + "type": "LATENT" + }, + { + "id": 37, + "origin_id": 125, + "origin_slot": 0, + "target_id": 122, + "target_slot": 0, + "type": "NOISE" + }, + { + "id": 30, + "origin_id": 119, + "origin_slot": 0, + "target_id": 122, + "target_slot": 1, + "type": "GUIDER" + }, + { + "id": 19, + "origin_id": 120, + "origin_slot": 0, + "target_id": 122, + "target_slot": 2, + "type": "SAMPLER" + }, + { + "id": 132, + "origin_id": 118, + "origin_slot": 0, + "target_id": 122, + "target_slot": 3, + "type": "SIGMAS" + }, + { + "id": 161, + "origin_id": 134, + "origin_slot": 0, + "target_id": 122, + "target_slot": 4, + "type": "LATENT" + }, + { + "id": 24, + "origin_id": 122, + "origin_slot": 0, + "target_id": 126, + "target_slot": 0, + "type": "LATENT" + }, + { + "id": 159, + "origin_id": 121, + "origin_slot": 0, + "target_id": 126, + "target_slot": 1, + "type": "VAE" + }, + { + "id": 117, + "origin_id": 124, + "origin_slot": 0, + "target_id": 123, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 127, + "origin_id": 121, + "origin_slot": 0, + "target_id": 136, + "target_slot": 1, + "type": "VAE" + }, + { + "id": 126, + "origin_id": -10, + "origin_slot": 0, + "target_id": 136, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 9, + "origin_id": 126, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 166, + "origin_id": 137, + "origin_slot": 0, + "target_id": 119, + "target_slot": 1, + "type": "CONDITIONING" + }, + { + "id": 168, + "origin_id": -10, + "origin_slot": 1, + "target_id": 123, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 169, + "origin_id": -10, + "origin_slot": 0, + "target_id": 135, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 170, + "origin_id": 135, + "origin_slot": 0, + "target_id": 118, + "target_slot": 1, + "type": "INT" + }, + { + "id": 171, + "origin_id": 135, + "origin_slot": 0, + "target_id": 134, + "target_slot": 0, + "type": "INT" + }, + { + "id": 172, + "origin_id": 135, + "origin_slot": 1, + "target_id": 118, + "target_slot": 2, + "type": "INT" + }, + { + "id": 173, + "origin_id": 135, + "origin_slot": 1, + "target_id": 134, + "target_slot": 1, + "type": "INT" + }, + { + "id": 177, + "origin_id": -10, + "origin_slot": 2, + "target_id": 129, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 178, + "origin_id": -10, + "origin_slot": 3, + "target_id": 124, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 179, + "origin_id": -10, + "origin_slot": 4, + "target_id": 121, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 181, + "origin_id": 129, + "origin_slot": 0, + "target_id": 128, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 183, + "origin_id": 128, + "origin_slot": 0, + "target_id": 130, + "target_slot": 1, + "type": "MODEL" + }, + { + "id": 184, + "origin_id": 129, + "origin_slot": 0, + "target_id": 130, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 185, + "origin_id": 130, + "origin_slot": 0, + "target_id": 119, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 186, + "origin_id": 131, + "origin_slot": 0, + "target_id": 133, + "target_slot": 1, + "type": "INT" + }, + { + "id": 187, + "origin_id": 132, + "origin_slot": 0, + "target_id": 133, + "target_slot": 0, + "type": "INT" + }, + { + "id": 188, + "origin_id": 133, + "origin_slot": 0, + "target_id": 118, + "target_slot": 0, + "type": "INT" + }, + { + "id": 189, + "origin_id": 138, + "origin_slot": 0, + "target_id": 133, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 190, + "origin_id": 138, + "origin_slot": 0, + "target_id": 130, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 191, + "origin_id": -10, + "origin_slot": 5, + "target_id": 138, + "target_slot": 0, + "type": "BOOLEAN" + }, + { + "id": 192, + "origin_id": -10, + "origin_slot": 6, + "target_id": 128, + "target_slot": 1, + "type": "COMBO" + } + ], + "extra": { + "workflowRendererVersion": "LG" + }, + "category": "Image generation and editing/Edit image", + "description": "Edits an image from text instructions using Flux.2 [dev], with guidance, schedulers, and optional Turbo LoRAs." + } + ] + }, + "extra": { + "ue_links": [] + } +} \ No newline at end of file diff --git a/blueprints/Image Edit (Flux.2 Klein 4B).json b/blueprints/Image Edit (Flux.2 Klein 4B).json index 6f2f7dc01..7f6fa7a4b 100644 --- a/blueprints/Image Edit (Flux.2 Klein 4B).json +++ b/blueprints/Image Edit (Flux.2 Klein 4B).json @@ -1472,7 +1472,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image generation and editing/Edit image" + "category": "Image generation and editing/Edit image", + "description": "Edits an input image via text instructions using FLUX.2 [klein] 4B." }, { "id": "6007e698-2ebd-4917-84d8-299b35d7b7ab", @@ -1821,7 +1822,8 @@ ], "extra": { "workflowRendererVersion": "LG" - } + }, + "description": "Applies reference image conditioning for style/identity transfer (Flux.2 Klein 4B)." } ] }, @@ -1837,4 +1839,4 @@ } }, "version": 0.4 -} \ No newline at end of file +} diff --git a/blueprints/Image Edit (LongCat Image Edit).json b/blueprints/Image Edit (LongCat Image Edit).json index 5b4eb18f0..de1c155a2 100644 --- a/blueprints/Image Edit (LongCat Image Edit).json +++ b/blueprints/Image Edit (LongCat Image Edit).json @@ -1417,7 +1417,8 @@ } ], "extra": {}, - "category": "Image generation and editing/Edit image" + "category": "Image generation and editing/Edit image", + "description": "Edits images via text instructions using LongCat Image Edit, an instruction-following image editing diffusion model." } ] }, diff --git a/blueprints/Image Edit (Qwen 2509).json b/blueprints/Image Edit (Qwen 2509).json new file mode 100644 index 000000000..f7be322a0 --- /dev/null +++ b/blueprints/Image Edit (Qwen 2509).json @@ -0,0 +1,1947 @@ +{ + "revision": 0, + "last_node_id": 433, + "last_link_id": 0, + "nodes": [ + { + "id": 433, + "type": "eba40a3a-f6c5-48ac-b58e-55525d06b373", + "pos": [ + 90, + -160 + ], + "size": [ + 390, + 610 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "label": "image", + "name": "image", + "type": "IMAGE", + "link": null + }, + { + "label": "image2 (optional)", + "name": "image2", + "type": "IMAGE", + "link": null + }, + { + "label": "image3 (optional)", + "name": "image3", + "type": "IMAGE", + "link": null + }, + { + "name": "prompt", + "type": "STRING", + "widget": { + "name": "prompt" + }, + "link": null + }, + { + "name": "seed", + "type": "INT", + "widget": { + "name": "seed" + }, + "link": null + }, + { + "label": "enable_turbo_mode", + "name": "value", + "type": "BOOLEAN", + "widget": { + "name": "value" + }, + "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": [ + [ + "111", + "prompt" + ], + [ + "3", + "seed" + ], + [ + "443", + "value" + ], + [ + "37", + "unet_name" + ], + [ + "38", + "clip_name" + ], + [ + "39", + "vae_name" + ], + [ + "3", + "control_after_generate" + ] + ], + "cnr_id": "comfy-core", + "ver": "0.3.62" + }, + "widgets_values": [], + "title": "Image Edit (Qwen 2509)" + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "eba40a3a-f6c5-48ac-b58e-55525d06b373", + "version": 1, + "state": { + "lastGroupId": 51, + "lastNodeId": 468, + "lastLinkId": 731, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Image Edit (Qwen 2509)", + "inputNode": { + "id": -10, + "bounding": [ + -1160, + 280, + 151.744140625, + 220 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + 2030, + -20, + 120, + 60 + ] + }, + "inputs": [ + { + "id": "d5089bd3-63bc-4a24-b478-6565ed2364e3", + "name": "image", + "type": "IMAGE", + "linkIds": [ + 248 + ], + "label": "image", + "pos": [ + -1028.255859375, + 300 + ] + }, + { + "id": "9e80fff0-ed0a-439f-a16e-a4a6cc1eb601", + "name": "image2", + "type": "IMAGE", + "linkIds": [ + 235, + 236 + ], + "label": "image2 (optional)", + "pos": [ + -1028.255859375, + 320 + ] + }, + { + "id": "49d98fd6-01b5-440b-8603-579252fd7fef", + "name": "image3", + "type": "IMAGE", + "linkIds": [ + 237, + 238 + ], + "label": "image3 (optional)", + "pos": [ + -1028.255859375, + 340 + ] + }, + { + "id": "5de32f24-a7b5-4423-b772-72824005f585", + "name": "prompt", + "type": "STRING", + "linkIds": [ + 244 + ], + "pos": [ + -1028.255859375, + 360 + ] + }, + { + "id": "85fb3d74-7881-4c71-bc8c-624be5eedc3d", + "name": "seed", + "type": "INT", + "linkIds": [ + 718 + ], + "pos": [ + -1028.255859375, + 380 + ] + }, + { + "id": "b0c828de-d7eb-42a3-8dfb-4f53360d4fc9", + "name": "value", + "type": "BOOLEAN", + "linkIds": [ + 719 + ], + "label": "enable_turbo_mode", + "pos": [ + -1028.255859375, + 400 + ] + }, + { + "id": "072baa05-5551-4a98-bd66-015a36833ac2", + "name": "unet_name", + "type": "COMBO", + "linkIds": [ + 720 + ], + "pos": [ + -1028.255859375, + 420 + ] + }, + { + "id": "d2891d11-b336-4750-9742-b93717c9ae39", + "name": "clip_name", + "type": "COMBO", + "linkIds": [ + 721 + ], + "pos": [ + -1028.255859375, + 440 + ] + }, + { + "id": "4218135f-5128-4b7e-8572-92cc55615793", + "name": "vae_name", + "type": "COMBO", + "linkIds": [ + 722 + ], + "pos": [ + -1028.255859375, + 460 + ] + } + ], + "outputs": [ + { + "id": "c4ebfc18-de83-4361-8e42-767c3c8c25c0", + "name": "IMAGE", + "type": "IMAGE", + "linkIds": [ + 110 + ], + "localized_name": "IMAGE", + "pos": [ + 2050, + 0 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 75, + "type": "CFGNorm", + "pos": [ + 1080, + 30 + ], + "size": [ + 290, + 110 + ], + "flags": {}, + "order": 11, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 141 + }, + { + "localized_name": "strength", + "name": "strength", + "type": "FLOAT", + "widget": { + "name": "strength" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "patched_model", + "name": "patched_model", + "type": "MODEL", + "links": [ + 186 + ] + } + ], + "properties": { + "Node name for S&R": "CFGNorm", + "cnr_id": "comfy-core", + "ver": "0.3.50", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "ue_properties": { + "widget_ue_connectable": { + "strength": true + } + } + }, + "widgets_values": [ + 1 + ] + }, + { + "id": 39, + "type": "VAELoader", + "pos": [ + -730, + 410 + ], + "size": [ + 330, + 110 + ], + "flags": {}, + "order": 9, + "mode": 0, + "inputs": [ + { + "localized_name": "vae_name", + "name": "vae_name", + "type": "COMBO", + "widget": { + "name": "vae_name" + }, + "link": 722 + } + ], + "outputs": [ + { + "localized_name": "VAE", + "name": "VAE", + "type": "VAE", + "slot_index": 0, + "links": [ + 76, + 168, + 206, + 207 + ] + } + ], + "properties": { + "Node name for S&R": "VAELoader", + "cnr_id": "comfy-core", + "ver": "0.3.48", + "models": [ + { + "name": "qwen_image_vae.safetensors", + "url": "https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/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, + "widget_ue_connectable": {} + }, + "widgets_values": [ + "qwen_image_vae.safetensors" + ] + }, + { + "id": 38, + "type": "CLIPLoader", + "pos": [ + -730, + 150 + ], + "size": [ + 330, + 150 + ], + "flags": {}, + "order": 8, + "mode": 0, + "inputs": [ + { + "localized_name": "clip_name", + "name": "clip_name", + "type": "COMBO", + "widget": { + "name": "clip_name" + }, + "link": 721 + }, + { + "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": [ + 204, + 205 + ] + } + ], + "properties": { + "Node name for S&R": "CLIPLoader", + "cnr_id": "comfy-core", + "ver": "0.3.48", + "models": [ + { + "name": "qwen_2.5_vl_7b_fp8_scaled.safetensors", + "url": "https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors", + "directory": "text_encoders" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "widget_ue_connectable": {} + }, + "widgets_values": [ + "qwen_2.5_vl_7b_fp8_scaled.safetensors", + "qwen_image", + "default" + ] + }, + { + "id": 37, + "type": "UNETLoader", + "pos": [ + -730, + -60 + ], + "size": [ + 330, + 110 + ], + "flags": {}, + "order": 7, + "mode": 0, + "inputs": [ + { + "localized_name": "unet_name", + "name": "unet_name", + "type": "COMBO", + "widget": { + "name": "unet_name" + }, + "link": 720 + }, + { + "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": [ + 184, + 710 + ] + } + ], + "properties": { + "Node name for S&R": "UNETLoader", + "cnr_id": "comfy-core", + "ver": "0.3.48", + "models": [ + { + "name": "qwen_image_edit_2509_fp8_e4m3fn.safetensors", + "url": "https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/resolve/main/split_files/diffusion_models/qwen_image_edit_2509_fp8_e4m3fn.safetensors", + "directory": "diffusion_models" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "widget_ue_connectable": {} + }, + "widgets_values": [ + "qwen_image_edit_2509_fp8_e4m3fn.safetensors", + "default" + ] + }, + { + "id": 110, + "type": "TextEncodeQwenImageEditPlus", + "pos": [ + -240, + 320 + ], + "size": [ + 400, + 240 + ], + "flags": {}, + "order": 14, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 204 + }, + { + "localized_name": "vae", + "name": "vae", + "shape": 7, + "type": "VAE", + "link": 206 + }, + { + "localized_name": "image1", + "name": "image1", + "shape": 7, + "type": "IMAGE", + "link": 251 + }, + { + "localized_name": "image2", + "name": "image2", + "shape": 7, + "type": "IMAGE", + "link": 236 + }, + { + "localized_name": "image3", + "name": "image3", + "shape": 7, + "type": "IMAGE", + "link": 238 + }, + { + "localized_name": "prompt", + "name": "prompt", + "type": "STRING", + "widget": { + "name": "prompt" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "links": [ + 210 + ] + } + ], + "properties": { + "Node name for S&R": "TextEncodeQwenImageEditPlus", + "cnr_id": "comfy-core", + "ver": "0.3.59" + }, + "widgets_values": [ + "" + ], + "color": "#223", + "bgcolor": "#335" + }, + { + "id": 66, + "type": "ModelSamplingAuraFlow", + "pos": [ + 1070, + -120 + ], + "size": [ + 290, + 110 + ], + "flags": {}, + "order": 10, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 708 + }, + { + "localized_name": "shift", + "name": "shift", + "type": "FLOAT", + "widget": { + "name": "shift" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "links": [ + 141 + ] + } + ], + "properties": { + "Node name for S&R": "ModelSamplingAuraFlow", + "cnr_id": "comfy-core", + "ver": "0.3.48", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "widget_ue_connectable": {} + }, + "widgets_values": [ + 3 + ] + }, + { + "id": 111, + "type": "TextEncodeQwenImageEditPlus", + "pos": [ + -250, + -70 + ], + "size": [ + 410, + 330 + ], + "flags": {}, + "order": 15, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 205 + }, + { + "localized_name": "vae", + "name": "vae", + "shape": 7, + "type": "VAE", + "link": 207 + }, + { + "localized_name": "image1", + "name": "image1", + "shape": 7, + "type": "IMAGE", + "link": 250 + }, + { + "localized_name": "image2", + "name": "image2", + "shape": 7, + "type": "IMAGE", + "link": 235 + }, + { + "localized_name": "image3", + "name": "image3", + "shape": 7, + "type": "IMAGE", + "link": 237 + }, + { + "localized_name": "prompt", + "name": "prompt", + "type": "STRING", + "widget": { + "name": "prompt" + }, + "link": 244 + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "links": [ + 211 + ] + } + ], + "properties": { + "Node name for S&R": "TextEncodeQwenImageEditPlus", + "cnr_id": "comfy-core", + "ver": "0.3.59" + }, + "widgets_values": [ + "" + ], + "color": "#232", + "bgcolor": "#353" + }, + { + "id": 88, + "type": "VAEEncode", + "pos": [ + -70, + 640 + ], + "size": [ + 230, + 100 + ], + "flags": {}, + "order": 12, + "mode": 0, + "inputs": [ + { + "localized_name": "pixels", + "name": "pixels", + "type": "IMAGE", + "link": 249 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 168 + } + ], + "outputs": [ + { + "localized_name": "LATENT", + "name": "LATENT", + "type": "LATENT", + "links": [ + 246 + ] + } + ], + "properties": { + "Node name for S&R": "VAEEncode", + "cnr_id": "comfy-core", + "ver": "0.3.50", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "ue_properties": { + "widget_ue_connectable": {} + } + } + }, + { + "id": 8, + "type": "VAEDecode", + "pos": [ + 1590, + -60 + ], + "size": [ + 230, + 100 + ], + "flags": { + "collapsed": false + }, + "order": 6, + "mode": 0, + "inputs": [ + { + "localized_name": "samples", + "name": "samples", + "type": "LATENT", + "link": 128 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 76 + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "slot_index": 0, + "links": [ + 110 + ] + } + ], + "properties": { + "Node name for S&R": "VAEDecode", + "cnr_id": "comfy-core", + "ver": "0.3.48", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "widget_ue_connectable": {} + } + }, + { + "id": 89, + "type": "LoraLoaderModelOnly", + "pos": [ + 320, + 300 + ], + "size": [ + 300, + 140 + ], + "flags": {}, + "order": 13, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 184 + }, + { + "localized_name": "lora_name", + "name": "lora_name", + "type": "COMBO", + "widget": { + "name": "lora_name" + }, + "link": null + }, + { + "localized_name": "strength_model", + "name": "strength_model", + "type": "FLOAT", + "widget": { + "name": "strength_model" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "links": [ + 709 + ] + } + ], + "properties": { + "Node name for S&R": "LoraLoaderModelOnly", + "cnr_id": "comfy-core", + "ver": "0.3.50", + "models": [ + { + "name": "Qwen-Image-Edit-2509-Lightning-4steps-V1.0-bf16.safetensors", + "url": "https://huggingface.co/lightx2v/Qwen-Image-Lightning/resolve/main/Qwen-Image-Edit-2509/Qwen-Image-Edit-2509-Lightning-4steps-V1.0-bf16.safetensors", + "directory": "loras" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "ue_properties": { + "widget_ue_connectable": { + "lora_name": true, + "strength_model": true + } + } + }, + "widgets_values": [ + "Qwen-Image-Edit-2509-Lightning-4steps-V1.0-bf16.safetensors", + 1 + ] + }, + { + "id": 117, + "type": "FluxKontextImageScale", + "pos": [ + -680, + 630 + ], + "size": [ + 230, + 80 + ], + "flags": {}, + "order": 16, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 248 + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 249, + 250, + 251 + ] + } + ], + "properties": { + "Node name for S&R": "FluxKontextImageScale" + } + }, + { + "id": 3, + "type": "KSampler", + "pos": [ + 1070, + 210 + ], + "size": [ + 300, + 590 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 186 + }, + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "link": 211 + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "link": 210 + }, + { + "localized_name": "latent_image", + "name": "latent_image", + "type": "LATENT", + "link": 246 + }, + { + "localized_name": "seed", + "name": "seed", + "type": "INT", + "widget": { + "name": "seed" + }, + "link": 718 + }, + { + "localized_name": "steps", + "name": "steps", + "type": "INT", + "widget": { + "name": "steps" + }, + "link": 707 + }, + { + "localized_name": "cfg", + "name": "cfg", + "type": "FLOAT", + "widget": { + "name": "cfg" + }, + "link": 706 + }, + { + "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": [ + 128 + ] + } + ], + "properties": { + "Node name for S&R": "KSampler", + "cnr_id": "comfy-core", + "ver": "0.3.48", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "widget_ue_connectable": {} + }, + "widgets_values": [ + 973414316252139, + "randomize", + 4, + 1, + "euler", + "simple", + 1 + ] + }, + { + "id": 436, + "type": "PrimitiveInt", + "pos": [ + 320, + 500 + ], + "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": [ + 713 + ] + } + ], + "title": "Steps", + "properties": { + "Node name for S&R": "PrimitiveInt" + }, + "widgets_values": [ + 4, + "fixed" + ] + }, + { + "id": 437, + "type": "PrimitiveFloat", + "pos": [ + 320, + 670 + ], + "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": [ + 714 + ] + } + ], + "title": "CFG", + "properties": { + "Node name for S&R": "PrimitiveFloat" + }, + "widgets_values": [ + 1 + ] + }, + { + "id": 438, + "type": "PrimitiveInt", + "pos": [ + 320, + -100 + ], + "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": [ + 711 + ] + } + ], + "title": "Steps", + "properties": { + "Node name for S&R": "PrimitiveInt" + }, + "widgets_values": [ + 20, + "fixed" + ] + }, + { + "id": 439, + "type": "PrimitiveFloat", + "pos": [ + 320, + 70 + ], + "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": [ + 712 + ] + } + ], + "title": "CFG", + "properties": { + "Node name for S&R": "PrimitiveFloat" + }, + "widgets_values": [ + 4 + ] + }, + { + "id": 440, + "type": "ComfySwitchNode", + "pos": [ + 750, + -80 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 17, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 710 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 709 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 715 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 708 + ] + } + ], + "title": "Switch (Model)", + "properties": { + "Node name for S&R": "ComfySwitchNode" + }, + "widgets_values": [ + false + ] + }, + { + "id": 441, + "type": "ComfySwitchNode", + "pos": [ + 730, + 340 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 18, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 711 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 713 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 716 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 707 + ] + } + ], + "title": "Switch (Steps)", + "properties": { + "Node name for S&R": "ComfySwitchNode" + }, + "widgets_values": [ + false + ] + }, + { + "id": 442, + "type": "ComfySwitchNode", + "pos": [ + 730, + 520 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 19, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 712 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 714 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 717 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 706 + ] + } + ], + "title": "Switch (CFG)", + "properties": { + "Node name for S&R": "ComfySwitchNode" + }, + "widgets_values": [ + false + ] + }, + { + "id": 443, + "type": "PrimitiveBoolean", + "pos": [ + 330, + 850 + ], + "size": [ + 270, + 100 + ], + "flags": {}, + "order": 20, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "BOOLEAN", + "widget": { + "name": "value" + }, + "link": 719 + } + ], + "outputs": [ + { + "localized_name": "BOOLEAN", + "name": "BOOLEAN", + "type": "BOOLEAN", + "links": [ + 715, + 716, + 717 + ] + } + ], + "title": "Enable Lightning LoRA", + "properties": { + "Node name for S&R": "PrimitiveBoolean" + }, + "widgets_values": [ + true + ] + }, + { + "id": 444, + "type": "MarkdownNote", + "pos": [ + 240, + -500 + ], + "size": [ + 450, + 310 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [], + "outputs": [], + "title": "Note: KSampler settings", + "properties": {}, + "widgets_values": [ + "You can test and find the best setting by yourself. The following table is for reference.\n| Parameters | Qwen Team | Comfy Original | with 4steps LoRA |\n|--------|---------|------------|---------------------------|\n| Steps | 50 | 20 | 4 |\n| CFG | 4.0 | 2.5 | 1.0 |" + ], + "color": "#432", + "bgcolor": "#000" + } + ], + "groups": [ + { + "id": 1, + "title": "Step1 - Load models", + "bounding": [ + -770, + -170, + 410, + 750 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 3, + "title": "Step 4 - Prompt", + "bounding": [ + -330, + -170, + 570, + 750 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 50, + "title": "Lightning LoRA", + "bounding": [ + 270, + 220, + 390, + 570 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 51, + "title": "Original Settings", + "bounding": [ + 270, + -170, + 390, + 360 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + } + ], + "links": [ + { + "id": 141, + "origin_id": 66, + "origin_slot": 0, + "target_id": 75, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 128, + "origin_id": 3, + "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": 184, + "origin_id": 37, + "origin_slot": 0, + "target_id": 89, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 186, + "origin_id": 75, + "origin_slot": 0, + "target_id": 3, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 211, + "origin_id": 111, + "origin_slot": 0, + "target_id": 3, + "target_slot": 1, + "type": "CONDITIONING" + }, + { + "id": 210, + "origin_id": 110, + "origin_slot": 0, + "target_id": 3, + "target_slot": 2, + "type": "CONDITIONING" + }, + { + "id": 168, + "origin_id": 39, + "origin_slot": 0, + "target_id": 88, + "target_slot": 1, + "type": "VAE" + }, + { + "id": 204, + "origin_id": 38, + "origin_slot": 0, + "target_id": 110, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 206, + "origin_id": 39, + "origin_slot": 0, + "target_id": 110, + "target_slot": 1, + "type": "VAE" + }, + { + "id": 205, + "origin_id": 38, + "origin_slot": 0, + "target_id": 111, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 207, + "origin_id": 39, + "origin_slot": 0, + "target_id": 111, + "target_slot": 1, + "type": "VAE" + }, + { + "id": 110, + "origin_id": 8, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 235, + "origin_id": -10, + "origin_slot": 1, + "target_id": 111, + "target_slot": 3, + "type": "IMAGE" + }, + { + "id": 236, + "origin_id": -10, + "origin_slot": 1, + "target_id": 110, + "target_slot": 3, + "type": "IMAGE" + }, + { + "id": 237, + "origin_id": -10, + "origin_slot": 2, + "target_id": 111, + "target_slot": 4, + "type": "IMAGE" + }, + { + "id": 238, + "origin_id": -10, + "origin_slot": 2, + "target_id": 110, + "target_slot": 4, + "type": "IMAGE" + }, + { + "id": 244, + "origin_id": -10, + "origin_slot": 3, + "target_id": 111, + "target_slot": 5, + "type": "STRING" + }, + { + "id": 246, + "origin_id": 88, + "origin_slot": 0, + "target_id": 3, + "target_slot": 3, + "type": "LATENT" + }, + { + "id": 248, + "origin_id": -10, + "origin_slot": 0, + "target_id": 117, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 249, + "origin_id": 117, + "origin_slot": 0, + "target_id": 88, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 250, + "origin_id": 117, + "origin_slot": 0, + "target_id": 111, + "target_slot": 2, + "type": "IMAGE" + }, + { + "id": 251, + "origin_id": 117, + "origin_slot": 0, + "target_id": 110, + "target_slot": 2, + "type": "IMAGE" + }, + { + "id": 706, + "origin_id": 442, + "origin_slot": 0, + "target_id": 3, + "target_slot": 6, + "type": "FLOAT" + }, + { + "id": 707, + "origin_id": 441, + "origin_slot": 0, + "target_id": 3, + "target_slot": 5, + "type": "INT" + }, + { + "id": 708, + "origin_id": 440, + "origin_slot": 0, + "target_id": 66, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 709, + "origin_id": 89, + "origin_slot": 0, + "target_id": 440, + "target_slot": 1, + "type": "MODEL" + }, + { + "id": 710, + "origin_id": 37, + "origin_slot": 0, + "target_id": 440, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 711, + "origin_id": 438, + "origin_slot": 0, + "target_id": 441, + "target_slot": 0, + "type": "INT" + }, + { + "id": 712, + "origin_id": 439, + "origin_slot": 0, + "target_id": 442, + "target_slot": 0, + "type": "FLOAT" + }, + { + "id": 713, + "origin_id": 436, + "origin_slot": 0, + "target_id": 441, + "target_slot": 1, + "type": "INT" + }, + { + "id": 714, + "origin_id": 437, + "origin_slot": 0, + "target_id": 442, + "target_slot": 1, + "type": "FLOAT" + }, + { + "id": 715, + "origin_id": 443, + "origin_slot": 0, + "target_id": 440, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 716, + "origin_id": 443, + "origin_slot": 0, + "target_id": 441, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 717, + "origin_id": 443, + "origin_slot": 0, + "target_id": 442, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 718, + "origin_id": -10, + "origin_slot": 4, + "target_id": 3, + "target_slot": 4, + "type": "INT" + }, + { + "id": 719, + "origin_id": -10, + "origin_slot": 5, + "target_id": 443, + "target_slot": 0, + "type": "BOOLEAN" + }, + { + "id": 720, + "origin_id": -10, + "origin_slot": 6, + "target_id": 37, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 721, + "origin_id": -10, + "origin_slot": 7, + "target_id": 38, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 722, + "origin_id": -10, + "origin_slot": 8, + "target_id": 39, + "target_slot": 0, + "type": "COMBO" + } + ], + "extra": { + "workflowRendererVersion": "LG" + }, + "category": "Image generation and editing/Edit image", + "description": "Edits images from text instructions using Qwen-Image-Edit-2509 with optional Lightning LoRA for few-step sampling." + } + ] + }, + "extra": {} +} diff --git a/blueprints/Image Edit (Qwen 2511).json b/blueprints/Image Edit (Qwen 2511).json index 582171fa0..1aa7e5765 100644 --- a/blueprints/Image Edit (Qwen 2511).json +++ b/blueprints/Image Edit (Qwen 2511).json @@ -132,7 +132,7 @@ }, "revision": 0, "config": {}, - "name": "local-Image Edit (Qwen 2511)", + "name": "Image Edit (Qwen 2511)", "inputNode": { "id": -10, "bounding": [ @@ -1468,7 +1468,8 @@ "VHS_MetadataImage": true, "VHS_KeepIntermediate": true }, - "category": "Image generation and editing/Edit image" + "category": "Image generation and editing/Edit image", + "description": "Edits images via text instructions using Qwen-Image-Edit-2511 with improved character consistency and integrated LoRA." } ] }, @@ -1489,4 +1490,4 @@ } }, "version": 0.4 -} +} \ No newline at end of file diff --git a/blueprints/Image Inpainting (Flux.1 Fill Dev).json b/blueprints/Image Inpainting (Flux.1 Fill Dev).json index d40d63594..c1326ed3d 100644 --- a/blueprints/Image Inpainting (Flux.1 Fill Dev).json +++ b/blueprints/Image Inpainting (Flux.1 Fill Dev).json @@ -1188,7 +1188,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image generation and editing/Inpaint image" + "category": "Image generation and editing/Inpaint image", + "description": "Inpaints masked image regions using Flux.1 fill [dev], Black Forest Labs' inpainting/outpainting model." } ] }, @@ -1202,4 +1203,4 @@ }, "ue_links": [] } -} \ No newline at end of file +} diff --git a/blueprints/Image Inpainting (Qwen-image).json b/blueprints/Image Inpainting (Qwen-image).json index 95b2909fa..a06d57e19 100644 --- a/blueprints/Image Inpainting (Qwen-image).json +++ b/blueprints/Image Inpainting (Qwen-image).json @@ -1548,7 +1548,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image generation and editing/Inpaint image" + "category": "Image generation and editing/Inpaint image", + "description": "Inpaints masked regions using Qwen-Image, extending its multilingual text rendering to inpainting tasks." }, { "id": "56a1f603-fbd2-40ed-94ef-c9ecbd96aca8", @@ -1907,7 +1908,8 @@ ], "extra": { "workflowRendererVersion": "LG" - } + }, + "description": "Expands and softens mask edges to reduce visible seams after image processing." } ] }, diff --git a/blueprints/Image Levels.json b/blueprints/Image Levels.json index ef256a1aa..1a1b18932 100644 --- a/blueprints/Image Levels.json +++ b/blueprints/Image Levels.json @@ -742,9 +742,10 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image Tools/Color adjust" + "category": "Image Tools/Color adjust", + "description": "Adjusts black point, white point, and gamma for tonal range control via GPU shader." } ] }, "extra": {} -} +} \ No newline at end of file diff --git a/blueprints/Image Outpainting (Qwen-Image).json b/blueprints/Image Outpainting (Qwen-Image).json index 218fdc775..6c07227c0 100644 --- a/blueprints/Image Outpainting (Qwen-Image).json +++ b/blueprints/Image Outpainting (Qwen-Image).json @@ -1919,7 +1919,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image generation and editing/Outpaint image" + "category": "Image generation and editing/Outpaint image", + "description": "Outpaints beyond image boundaries using Qwen-Image's outpainting capabilities." }, { "id": "f93c215e-c393-460e-9534-ed2c3d8a652e", @@ -2278,7 +2279,8 @@ ], "extra": { "workflowRendererVersion": "LG" - } + }, + "description": "Expands and softens mask edges to reduce visible seams after image processing." }, { "id": "2a4b2cc0-db37-4302-a067-da392f38f06b", @@ -2733,7 +2735,8 @@ ], "extra": { "workflowRendererVersion": "LG" - } + }, + "description": "Scales both image and mask together while preserving alignment for editing workflows." } ] }, diff --git a/blueprints/Image Segmentation (SAM3).json b/blueprints/Image Segmentation (SAM3).json new file mode 100644 index 000000000..b405bf623 --- /dev/null +++ b/blueprints/Image Segmentation (SAM3).json @@ -0,0 +1,714 @@ +{ + "revision": 0, + "last_node_id": 99, + "last_link_id": 0, + "nodes": [ + { + "id": 99, + "type": "6e7ab3ea-96aa-470f-9b94-3d9d0e01f481", + "pos": [ + -1630, + -3270 + ], + "size": [ + 290, + 370 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "label": "image", + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": null + }, + { + "label": "object", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": null + }, + { + "name": "bboxes", + "type": "BOUNDING_BOX", + "link": null + }, + { + "name": "positive_coords", + "type": "STRING", + "link": null + }, + { + "name": "negative_coords", + "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": null + } + ], + "outputs": [ + { + "localized_name": "masks", + "name": "masks", + "type": "MASK", + "links": [] + }, + { + "localized_name": "bboxes", + "name": "bboxes", + "type": "BOUNDING_BOX", + "links": [] + } + ], + "properties": { + "proxyWidgets": [ + [ + "78", + "text" + ], + [ + "75", + "threshold" + ], + [ + "75", + "refine_iterations" + ], + [ + "75", + "individual_masks" + ], + [ + "77", + "ckpt_name" + ] + ], + "ue_properties": { + "widget_ue_connectable": { + "text": true + }, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "cnr_id": "comfy-core", + "ver": "0.19.3", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [], + "title": "Image Segmentation (SAM3)" + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "6e7ab3ea-96aa-470f-9b94-3d9d0e01f481", + "version": 1, + "state": { + "lastGroupId": 0, + "lastNodeId": 113, + "lastLinkId": 283, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Image Segmentation (SAM3)", + "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": { + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "cnr_id": "comfy-core", + "ver": "0.19.3", + "Node name for S&R": "SAM3_Detect", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "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": { + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "cnr_id": "comfy-core", + "ver": "0.19.3", + "Node name for S&R": "CheckpointLoaderSimple", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "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" + } + ] + }, + "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": { + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "cnr_id": "comfy-core", + "ver": "0.19.3", + "Node name for S&R": "CLIPTextEncode", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "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": {}, + "category": "Image Tools/Image Segmentation", + "description": "Segments images into masks using Meta SAM3 from text prompts, points, or boxes." + } + ] + }, + "extra": { + "ue_links": [] + } +} diff --git a/blueprints/Image Upscale(Z-image-Turbo).json b/blueprints/Image Upscale(Z-image-Turbo).json index 0d2b6e240..bd803a0b1 100644 --- a/blueprints/Image Upscale(Z-image-Turbo).json +++ b/blueprints/Image Upscale(Z-image-Turbo).json @@ -141,7 +141,7 @@ }, "revision": 0, "config": {}, - "name": "local-Image Upscale(Z-image-Turbo)", + "name": "Image Upscale (Z-image-Turbo)", "inputNode": { "id": -10, "bounding": [ @@ -1302,7 +1302,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image generation and editing/Enhance" + "category": "Image generation and editing/Enhance", + "description": "Upscales images to higher resolution using Z-Image-Turbo." } ] }, diff --git a/blueprints/Image to Depth Map (Lotus).json b/blueprints/Image to Depth Map (Lotus).json index 089f2cd42..12f10ba5b 100644 --- a/blueprints/Image to Depth Map (Lotus).json +++ b/blueprints/Image to Depth Map (Lotus).json @@ -99,7 +99,7 @@ }, "revision": 0, "config": {}, - "name": "local-Image to Depth Map (Lotus)", + "name": "Image to Depth Map (Lotus)", "inputNode": { "id": -10, "bounding": [ @@ -948,7 +948,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image generation and editing/Depth to image" + "category": "Image generation and editing/Depth to image", + "description": "Estimates a monocular depth map from an input image using the Lotus depth estimation model." } ] }, @@ -964,4 +965,4 @@ "workflowRendererVersion": "LG" }, "version": 0.4 -} +} \ No newline at end of file diff --git a/blueprints/Image to Layers(Qwen-Image-Layered).json b/blueprints/Image to Layers(Qwen-Image-Layered).json index 8a525e7a5..7b44f0563 100644 --- a/blueprints/Image to Layers(Qwen-Image-Layered).json +++ b/blueprints/Image to Layers(Qwen-Image-Layered).json @@ -1586,7 +1586,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image generation and editing/Image to layers" + "category": "Image generation and editing/Image to layers", + "description": "Decomposes an image into variable-resolution RGBA layers for independent editing using Qwen-Image-Layered." } ] }, diff --git a/blueprints/Image to Model (Hunyuan3d 2.1).json b/blueprints/Image to Model (Hunyuan3d 2.1).json index 4705603a8..ee5552656 100644 --- a/blueprints/Image to Model (Hunyuan3d 2.1).json +++ b/blueprints/Image to Model (Hunyuan3d 2.1).json @@ -72,7 +72,7 @@ }, "revision": 0, "config": {}, - "name": "local-Image to Model (Hunyuan3d 2.1)", + "name": "Image to 3D Model (Hunyuan3d 2.1)", "inputNode": { "id": -10, "bounding": [ @@ -765,7 +765,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "3D/Image to 3D Model" + "category": "3D/Image to 3D Model", + "description": "Generates 3D mesh models from a single input image using Hunyuan3D 2.0/2.1." } ] }, diff --git a/blueprints/Image to Video (LTX-2.3).json b/blueprints/Image to Video (LTX-2.3).json index 86a601130..3db524ea0 100644 --- a/blueprints/Image to Video (LTX-2.3).json +++ b/blueprints/Image to Video (LTX-2.3).json @@ -4223,7 +4223,8 @@ "extra": { "workflowRendererVersion": "Vue-corrected" }, - "category": "Video generation and editing/Image to video" + "category": "Video generation and editing/Image to video", + "description": "Generates video from a single input image using LTX-2.3." } ] }, diff --git a/blueprints/Image to Video (Wan 2.2).json b/blueprints/Image to Video (Wan 2.2).json index a8dafd3c9..a24adcfb6 100644 --- a/blueprints/Image to Video (Wan 2.2).json +++ b/blueprints/Image to Video (Wan 2.2).json @@ -206,7 +206,7 @@ }, "revision": 0, "config": {}, - "name": "local-Image to Video (Wan 2.2)", + "name": "Image to Video (Wan 2.2)", "inputNode": { "id": -10, "bounding": [ @@ -2027,7 +2027,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Video generation and editing/Image to video" + "category": "Video generation and editing/Image to video", + "description": "Image-to-video with Wan 2.2 using a start image plus text prompt to extend motion from the still frame." } ] }, diff --git a/blueprints/Pose to Image (Z-Image-Turbo).json b/blueprints/Pose to Image (Z-Image-Turbo).json index a55410ba4..5c2749efe 100644 --- a/blueprints/Pose to Image (Z-Image-Turbo).json +++ b/blueprints/Pose to Image (Z-Image-Turbo).json @@ -134,7 +134,7 @@ }, "revision": 0, "config": {}, - "name": "local-Pose to Image (Z-Image-Turbo)", + "name": "Pose to Image (Z-Image-Turbo)", "inputNode": { "id": -10, "bounding": [ @@ -1298,7 +1298,8 @@ "VHS_MetadataImage": true, "VHS_KeepIntermediate": true }, - "category": "Image generation and editing/Pose to image" + "category": "Image generation and editing/Pose to image", + "description": "Generates an image from pose keypoints using Z-Image-Turbo with text conditioning." } ] }, @@ -1319,4 +1320,4 @@ } }, "version": 0.4 -} +} \ No newline at end of file diff --git a/blueprints/Pose to Video (LTX 2.0).json b/blueprints/Pose to Video (LTX 2.0).json index 580900bc0..1ce49351a 100644 --- a/blueprints/Pose to Video (LTX 2.0).json +++ b/blueprints/Pose to Video (LTX 2.0).json @@ -3870,7 +3870,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Video generation and editing/Pose to video" + "category": "Video generation and editing/Pose to video", + "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 5e57548ff..e260b1203 100644 --- a/blueprints/Prompt Enhance.json +++ b/blueprints/Prompt Enhance.json @@ -270,9 +270,10 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Text generation/Prompt enhance" + "category": "Text generation/Prompt enhance", + "description": "Expands short text prompts into detailed descriptions using a text generation model for better generation quality." } ] }, "extra": {} -} +} \ No newline at end of file diff --git a/blueprints/Remove Background (BiRefNet).json b/blueprints/Remove Background (BiRefNet).json new file mode 100644 index 000000000..732a4adc4 --- /dev/null +++ b/blueprints/Remove Background (BiRefNet).json @@ -0,0 +1,397 @@ +{ + "revision": 0, + "last_node_id": 19, + "last_link_id": 0, + "nodes": [ + { + "id": 19, + "type": "5b40ca21-ba1a-41d5-b403-4d2d7acdc195", + "pos": [ + -6411.330578108367, + 1940.2638932730042 + ], + "size": [ + 349.609375, + 145.9375 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": null + }, + { + "name": "bg_removal_name", + "type": "COMBO", + "widget": { + "name": "bg_removal_name" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [] + }, + { + "name": "mask", + "type": "MASK", + "links": [] + } + ], + "properties": { + "proxyWidgets": [ + [ + "14", + "bg_removal_name" + ] + ] + }, + "widgets_values": [], + "title": "Remove Background (BiRefNet)" + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "5b40ca21-ba1a-41d5-b403-4d2d7acdc195", + "version": 1, + "state": { + "lastGroupId": 0, + "lastNodeId": 21, + "lastLinkId": 16, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Remove Background (BiRefNet)", + "description": "Removes or replaces image backgrounds using BiRefNet segmentation and alpha compositing.", + "inputNode": { + "id": -10, + "bounding": [ + -6728.534070722246, + 1475.2619799128663, + 150.9140625, + 88 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + -6169.049695722246, + 1475.2619799128663, + 128, + 88 + ] + }, + "inputs": [ + { + "id": "7bc321cd-df31-4c39-aaf7-7f0d01326189", + "name": "image", + "type": "IMAGE", + "linkIds": [ + 5, + 7 + ], + "localized_name": "image", + "pos": [ + -6601.620008222246, + 1499.2619799128663 + ] + }, + { + "id": "e89d2cd8-daa3-4e29-8a69-851db85072cb", + "name": "bg_removal_name", + "type": "COMBO", + "linkIds": [ + 12 + ], + "pos": [ + -6601.620008222246, + 1519.2619799128663 + ] + } + ], + "outputs": [ + { + "id": "16e7863c-4c38-46c2-aa74-e82991fbfe8d", + "name": "IMAGE", + "type": "IMAGE", + "linkIds": [ + 8 + ], + "localized_name": "IMAGE", + "pos": [ + -6145.049695722246, + 1499.2619799128663 + ] + }, + { + "id": "f7240c19-5b80-406e-a8e2-9b12440ee2d6", + "name": "mask", + "type": "MASK", + "linkIds": [ + 11 + ], + "pos": [ + -6145.049695722246, + 1519.2619799128663 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 13, + "type": "RemoveBackground", + "pos": [ + -6536.764823982709, + 1444.9963409012412 + ], + "size": [ + 302.25, + 72 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 5 + }, + { + "localized_name": "bg_removal_model", + "name": "bg_removal_model", + "type": "BACKGROUND_REMOVAL", + "link": 3 + } + ], + "outputs": [ + { + "localized_name": "mask", + "name": "mask", + "type": "MASK", + "links": [ + 4, + 11 + ] + } + ], + "properties": { + "Node name for S&R": "RemoveBackground" + } + }, + { + "id": 14, + "type": "LoadBackgroundRemovalModel", + "pos": [ + -6540.534070722246, + 1302.223464635445 + ], + "size": [ + 311.484375, + 85.515625 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "bg_removal_name", + "name": "bg_removal_name", + "type": "COMBO", + "widget": { + "name": "bg_removal_name" + }, + "link": 12 + } + ], + "outputs": [ + { + "localized_name": "bg_model", + "name": "bg_model", + "type": "BACKGROUND_REMOVAL", + "links": [ + 3 + ] + } + ], + "properties": { + "Node name for S&R": "LoadBackgroundRemovalModel", + "models": [ + { + "name": "birefnet.safetensors", + "url": "https://huggingface.co/Comfy-Org/BiRefNet/resolve/main/background_removal/birefnet.safetensors", + "directory": "background_removal" + } + ] + }, + "widgets_values": [ + "birefnet.safetensors" + ] + }, + { + "id": 15, + "type": "InvertMask", + "pos": [ + -6532.446160529669, + 1571.1111286839914 + ], + "size": [ + 285.984375, + 48 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "mask", + "name": "mask", + "type": "MASK", + "link": 4 + } + ], + "outputs": [ + { + "localized_name": "MASK", + "name": "MASK", + "type": "MASK", + "links": [ + 6 + ] + } + ], + "properties": { + "Node name for S&R": "InvertMask" + } + }, + { + "id": 16, + "type": "JoinImageWithAlpha", + "pos": [ + -6527.4370171636665, + 1674.3004951902876 + ], + "size": [ + 284.96875, + 72 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 7 + }, + { + "localized_name": "alpha", + "name": "alpha", + "type": "MASK", + "link": 6 + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [ + 8 + ] + } + ], + "properties": { + "Node name for S&R": "JoinImageWithAlpha" + } + } + ], + "groups": [], + "links": [ + { + "id": 3, + "origin_id": 14, + "origin_slot": 0, + "target_id": 13, + "target_slot": 1, + "type": "BACKGROUND_REMOVAL" + }, + { + "id": 4, + "origin_id": 13, + "origin_slot": 0, + "target_id": 15, + "target_slot": 0, + "type": "MASK" + }, + { + "id": 6, + "origin_id": 15, + "origin_slot": 0, + "target_id": 16, + "target_slot": 1, + "type": "MASK" + }, + { + "id": 5, + "origin_id": -10, + "origin_slot": 0, + "target_id": 13, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 7, + "origin_id": -10, + "origin_slot": 0, + "target_id": 16, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 8, + "origin_id": 16, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 11, + "origin_id": 13, + "origin_slot": 0, + "target_id": -20, + "target_slot": 1, + "type": "MASK" + }, + { + "id": 12, + "origin_id": -10, + "origin_slot": 1, + "target_id": 14, + "target_slot": 0, + "type": "COMBO" + } + ], + "extra": {}, + "category": "Image generation and editing/Background Removal" + } + ] + }, + "extra": {} +} \ No newline at end of file diff --git a/blueprints/Sharpen.json b/blueprints/Sharpen.json index f332400fd..3c4099c6b 100644 --- a/blueprints/Sharpen.json +++ b/blueprints/Sharpen.json @@ -302,8 +302,9 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image Tools/Sharpen" + "category": "Image Tools/Sharpen", + "description": "Sharpens image details using a GPU fragment shader for enhanced clarity." } ] } -} +} \ No newline at end of file diff --git a/blueprints/Text to Audio (ACE-Step 1.5).json b/blueprints/Text to Audio (ACE-Step 1.5).json index 206cf16be..5b8b8626f 100644 --- a/blueprints/Text to Audio (ACE-Step 1.5).json +++ b/blueprints/Text to Audio (ACE-Step 1.5).json @@ -222,7 +222,7 @@ }, "revision": 0, "config": {}, - "name": "local-Text to Audio (ACE-Step 1.5)", + "name": "Text to Audio (ACE-Step 1.5)", "inputNode": { "id": -10, "bounding": [ @@ -1502,7 +1502,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Audio/Music generation" + "category": "Audio/Music generation", + "description": "Generates audio/music from text prompts using ACE-Step 1.5, a diffusion-based audio generation model." } ] }, @@ -1518,4 +1519,4 @@ } }, "version": 0.4 -} +} \ No newline at end of file diff --git a/blueprints/Text to Image (Ernie Image Turbo).json b/blueprints/Text to Image (Ernie Image Turbo).json new file mode 100644 index 000000000..4ecdd1883 --- /dev/null +++ b/blueprints/Text to Image (Ernie Image Turbo).json @@ -0,0 +1,2112 @@ +{ + "revision": 0, + "last_node_id": 88, + "last_link_id": 0, + "nodes": [ + { + "id": 88, + "type": "2a4f0815-c4d2-4e8b-9bdf-991a8403889d", + "pos": [ + -120, + 240 + ], + "size": [ + 400, + 540 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "label": "prompt", + "name": "value", + "type": "STRING", + "widget": { + "name": "value" + }, + "link": null + }, + { + "label": "prompt_enhancement", + "name": "value_1", + "type": "BOOLEAN", + "widget": { + "name": "value_1" + }, + "link": null + }, + { + "name": "width", + "type": "INT", + "widget": { + "name": "width" + }, + "link": null + }, + { + "name": "height", + "type": "INT", + "widget": { + "name": "height" + }, + "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 + }, + { + "label": "prompt_enhancer", + "name": "clip_name_1", + "type": "COMBO", + "widget": { + "name": "clip_name_1" + }, + "link": null + }, + { + "name": "vae_name", + "type": "COMBO", + "widget": { + "name": "vae_name" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [] + } + ], + "properties": { + "proxyWidgets": [ + [ + "94", + "value" + ], + [ + "96", + "value" + ], + [ + "71", + "width" + ], + [ + "71", + "height" + ], + [ + "70", + "seed" + ], + [ + "66", + "unet_name" + ], + [ + "62", + "clip_name" + ], + [ + "98", + "clip_name" + ], + [ + "63", + "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, + "ue_properties": { + "widget_ue_connectable": { + "value": true, + "value_1": true + }, + "version": "7.7", + "input_ue_unconnectable": {} + } + }, + "widgets_values": [], + "title": "Text to Image (Ernie Image Turbo)" + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "2a4f0815-c4d2-4e8b-9bdf-991a8403889d", + "version": 1, + "state": { + "lastGroupId": 7, + "lastNodeId": 103, + "lastLinkId": 134, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Text to Image (Ernie Image Turbo)", + "inputNode": { + "id": -10, + "bounding": [ + -1350, + 370, + 163.50390625, + 220 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + 1110, + 260, + 120, + 60 + ] + }, + "inputs": [ + { + "id": "74a4609c-67df-4ae9-ab96-9ff4e3a1c3b1", + "name": "value", + "type": "STRING", + "linkIds": [ + 128 + ], + "label": "prompt", + "pos": [ + -1206.49609375, + 390 + ] + }, + { + "id": "996f1854-7ae3-450e-821c-a9b5b7c310f9", + "name": "value_1", + "type": "BOOLEAN", + "linkIds": [ + 127 + ], + "label": "prompt_enhancement", + "pos": [ + -1206.49609375, + 410 + ] + }, + { + "id": "71e9c6e8-4285-4543-b1d3-81520088f6a4", + "name": "width", + "type": "INT", + "linkIds": [ + 104, + 129 + ], + "pos": [ + -1206.49609375, + 430 + ] + }, + { + "id": "bdb6cd97-67d9-440c-8c4c-9b7a7540edd0", + "name": "height", + "type": "INT", + "linkIds": [ + 105, + 130 + ], + "pos": [ + -1206.49609375, + 450 + ] + }, + { + "id": "18abb56c-30bf-4de5-83c1-c12376e8d14e", + "name": "seed", + "type": "INT", + "linkIds": [ + 108 + ], + "pos": [ + -1206.49609375, + 470 + ] + }, + { + "id": "e5cd06f9-64ed-4778-97ba-b165f7a79c4e", + "name": "unet_name", + "type": "COMBO", + "linkIds": [ + 109 + ], + "pos": [ + -1206.49609375, + 490 + ] + }, + { + "id": "06480e4c-4043-489b-ae68-1cf2b4246260", + "name": "clip_name", + "type": "COMBO", + "linkIds": [ + 110 + ], + "pos": [ + -1206.49609375, + 510 + ] + }, + { + "id": "8d65d01b-16b2-420d-8b7b-42077c2e4976", + "name": "clip_name_1", + "type": "COMBO", + "linkIds": [ + 132 + ], + "label": "prompt_enhancer", + "pos": [ + -1206.49609375, + 530 + ] + }, + { + "id": "697f2fdb-0fd9-4008-a895-0f9ce9e8fd88", + "name": "vae_name", + "type": "COMBO", + "linkIds": [ + 133 + ], + "pos": [ + -1206.49609375, + 550 + ] + } + ], + "outputs": [ + { + "id": "21d5fbe0-9f91-4d93-8ea8-5bbf2cd5b698", + "name": "IMAGE", + "type": "IMAGE", + "linkIds": [ + 84 + ], + "localized_name": "IMAGE", + "pos": [ + 1130, + 280 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 71, + "type": "EmptyFlux2LatentImage", + "pos": [ + -470, + 1050 + ], + "size": [ + 270, + 170 + ], + "flags": {}, + "order": 6, + "mode": 0, + "inputs": [ + { + "localized_name": "width", + "name": "width", + "type": "INT", + "widget": { + "name": "width" + }, + "link": 104 + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "widget": { + "name": "height" + }, + "link": 105 + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "widget": { + "name": "batch_size" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "LATENT", + "name": "LATENT", + "type": "LATENT", + "links": [ + 80 + ] + } + ], + "properties": { + "Node name for S&R": "EmptyFlux2LatentImage", + "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": [ + 1024, + 1024, + 1 + ] + }, + { + "id": 66, + "type": "UNETLoader", + "pos": [ + -470, + 320 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "unet_name", + "name": "unet_name", + "type": "COMBO", + "widget": { + "name": "unet_name" + }, + "link": 109 + }, + { + "localized_name": "weight_dtype", + "name": "weight_dtype", + "type": "COMBO", + "widget": { + "name": "weight_dtype" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "links": [ + 85 + ] + } + ], + "properties": { + "Node name for S&R": "UNETLoader", + "cnr_id": "comfy-core", + "ver": "0.3.73", + "models": [ + { + "name": "ernie-image-turbo.safetensors", + "url": "https://huggingface.co/Comfy-Org/ERNIE-Image/resolve/main/diffusion_models/ernie-image-turbo.safetensors", + "directory": "diffusion_models" + } + ], + "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": [ + "ernie-image-turbo.safetensors", + "default" + ] + }, + { + "id": 65, + "type": "VAEDecode", + "pos": [ + 710, + 280 + ], + "size": [ + 230, + 100 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "samples", + "name": "samples", + "type": "LATENT", + "link": 73 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 74 + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "slot_index": 0, + "links": [ + 84 + ] + } + ], + "properties": { + "Node name for S&R": "VAEDecode", + "cnr_id": "comfy-core", + "ver": "0.3.64", + "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": 70, + "type": "KSampler", + "pos": [ + 350, + 280 + ], + "size": [ + 320, + 350 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 85 + }, + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "link": 76 + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "link": 113 + }, + { + "localized_name": "latent_image", + "name": "latent_image", + "type": "LATENT", + "link": 80 + }, + { + "localized_name": "seed", + "name": "seed", + "type": "INT", + "widget": { + "name": "seed" + }, + "link": 108 + }, + { + "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": [ + 73 + ] + } + ], + "properties": { + "Node name for S&R": "KSampler", + "cnr_id": "comfy-core", + "ver": "0.3.64", + "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": [ + 423299999918804, + "randomize", + 8, + 1, + "euler", + "simple", + 1 + ] + }, + { + "id": 67, + "type": "CLIPTextEncode", + "pos": [ + -140, + 320 + ], + "size": [ + 410, + 370 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 79 + }, + { + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": 131 + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "links": [ + 76, + 112 + ] + } + ], + "properties": { + "Node name for S&R": "CLIPTextEncode", + "cnr_id": "comfy-core", + "ver": "0.3.73", + "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": [ + "" + ], + "color": "#232", + "bgcolor": "#353" + }, + { + "id": 62, + "type": "CLIPLoader", + "pos": [ + -470, + 530 + ], + "size": [ + 270, + 150 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "clip_name", + "name": "clip_name", + "type": "COMBO", + "widget": { + "name": "clip_name" + }, + "link": 110 + }, + { + "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": [ + 79 + ] + } + ], + "properties": { + "Node name for S&R": "CLIPLoader", + "cnr_id": "comfy-core", + "ver": "0.3.73", + "models": [ + { + "name": "ministral-3-3b.safetensors", + "url": "https://huggingface.co/Comfy-Org/ERNIE-Image/resolve/main/text_encoders/ministral-3-3b.safetensors", + "directory": "text_encoders" + } + ], + "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": [ + "ministral-3-3b.safetensors", + "flux2", + "default" + ] + }, + { + "id": 63, + "type": "VAELoader", + "pos": [ + -470, + 780 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "vae_name", + "name": "vae_name", + "type": "COMBO", + "widget": { + "name": "vae_name" + }, + "link": 133 + } + ], + "outputs": [ + { + "localized_name": "VAE", + "name": "VAE", + "type": "VAE", + "links": [ + 74 + ] + } + ], + "properties": { + "Node name for S&R": "VAELoader", + "cnr_id": "comfy-core", + "ver": "0.3.73", + "models": [ + { + "name": "flux2-vae.safetensors", + "url": "https://huggingface.co/Comfy-Org/ERNIE-Image/resolve/main/vae/flux2-vae.safetensors", + "directory": "vae" + } + ], + "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": [ + "flux2-vae.safetensors" + ] + }, + { + "id": 91, + "type": "ConditioningZeroOut", + "pos": [ + 30, + 760 + ], + "size": [ + 230, + 80 + ], + "flags": {}, + "order": 7, + "mode": 0, + "inputs": [ + { + "localized_name": "conditioning", + "name": "conditioning", + "type": "CONDITIONING", + "link": 112 + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "links": [ + 113 + ] + } + ], + "properties": { + "Node name for S&R": "ConditioningZeroOut", + "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": 93, + "type": "StringReplace", + "pos": [ + -500, + -650 + ], + "size": [ + 430, + 450 + ], + "flags": {}, + "order": 8, + "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": 115 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 121 + ] + } + ], + "properties": { + "Node name for S&R": "StringReplace", + "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": [ + "[SYSTEM_PROMPT]你是一个专业的文生图 Prompt 增强助手。你将收到用户的简短图片描述及目标生成分辨率，请据此扩写为一段内容丰富、细节充分的视觉描述，以帮助文生图模型生成高质量的图片。仅输出增强后的描述，不要包含任何解释或前缀。[/SYSTEM_PROMPT][INST]{\"prompt\": \"{prompt}\", \"width\": {width}, \"height\": {height}}[/INST]", + "{prompt}", + "" + ] + }, + { + "id": 94, + "type": "PrimitiveStringMultiline", + "pos": [ + -950, + -660 + ], + "size": [ + 400, + 200 + ], + "flags": {}, + "order": 9, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "STRING", + "widget": { + "name": "value" + }, + "link": 128 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 115, + 118 + ] + } + ], + "title": "String (Multiline - Prompt)", + "properties": { + "Node name for S&R": "PrimitiveStringMultiline", + "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": [ + "" + ] + }, + { + "id": 95, + "type": "TextGenerate", + "pos": [ + 530, + -660 + ], + "size": [ + 400, + 380 + ], + "flags": {}, + "order": 10, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 116 + }, + { + "localized_name": "image", + "name": "image", + "shape": 7, + "type": "IMAGE", + "link": null + }, + { + "localized_name": "prompt", + "name": "prompt", + "type": "STRING", + "widget": { + "name": "prompt" + }, + "link": 117 + }, + { + "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": "sampling_mode.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": [ + 119 + ] + } + ], + "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, + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + } + }, + "widgets_values": [ + "", + 2048, + "on", + 0.6, + 64, + 0.8, + 0.05, + 1.05, + 0, + 0, + false, + true + ] + }, + { + "id": 96, + "type": "PrimitiveBoolean", + "pos": [ + -490, + 60 + ], + "size": [ + 270, + 100 + ], + "flags": {}, + "order": 11, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "BOOLEAN", + "widget": { + "name": "value" + }, + "link": 127 + } + ], + "outputs": [ + { + "localized_name": "BOOLEAN", + "name": "BOOLEAN", + "type": "BOOLEAN", + "links": [ + 120 + ] + } + ], + "title": "Enable prompt enhancement?", + "properties": { + "Node name for S&R": "PrimitiveBoolean", + "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": [ + true + ] + }, + { + "id": 97, + "type": "ComfySwitchNode", + "pos": [ + 550, + -10 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 12, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 118 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 119 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 120 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 131, + 134 + ] + } + ], + "properties": { + "Node name for S&R": "ComfySwitchNode", + "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": [ + false + ] + }, + { + "id": 98, + "type": "CLIPLoader", + "pos": [ + -490, + -150 + ], + "size": [ + 510, + 150 + ], + "flags": {}, + "order": 13, + "mode": 0, + "inputs": [ + { + "localized_name": "clip_name", + "name": "clip_name", + "type": "COMBO", + "widget": { + "name": "clip_name" + }, + "link": 132 + }, + { + "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": [ + 116 + ] + } + ], + "title": "Load CLIP (PE)", + "properties": { + "Node name for S&R": "CLIPLoader", + "cnr_id": "comfy-core", + "ver": "0.19.0", + "models": [ + { + "name": "ernie-image-prompt-enhancer.safetensors", + "url": "https://huggingface.co/Comfy-Org/ERNIE-Image/resolve/main/text_encoders/ernie-image-prompt-enhancer.safetensors", + "directory": "text_encoders" + } + ], + "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": [ + "ernie-image-prompt-enhancer.safetensors", + "flux2", + "default" + ] + }, + { + "id": 99, + "type": "PreviewAny", + "pos": [ + -950, + -410 + ], + "size": [ + 400, + 180 + ], + "flags": {}, + "order": 14, + "mode": 0, + "inputs": [ + { + "localized_name": "source", + "name": "source", + "type": "*", + "link": 129 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 122 + ] + } + ], + "title": "Preview as Text (Int to String)", + "properties": { + "Node name for S&R": "PreviewAny", + "cnr_id": "comfy-core", + "ver": "0.19.0", + "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": [ + null, + null, + null + ] + }, + { + "id": 100, + "type": "PreviewAny", + "pos": [ + -950, + -190 + ], + "size": [ + 400, + 180 + ], + "flags": {}, + "order": 15, + "mode": 0, + "inputs": [ + { + "localized_name": "source", + "name": "source", + "type": "*", + "link": 130 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 124 + ] + } + ], + "title": "Preview as Text (Int to String)", + "properties": { + "Node name for S&R": "PreviewAny", + "cnr_id": "comfy-core", + "ver": "0.19.0", + "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": [ + null, + null, + null + ] + }, + { + "id": 101, + "type": "StringReplace", + "pos": [ + -30, + -650 + ], + "size": [ + 230, + 450 + ], + "flags": {}, + "order": 16, + "mode": 0, + "inputs": [ + { + "localized_name": "string", + "name": "string", + "type": "STRING", + "widget": { + "name": "string" + }, + "link": 121 + }, + { + "localized_name": "find", + "name": "find", + "type": "STRING", + "widget": { + "name": "find" + }, + "link": null + }, + { + "localized_name": "replace", + "name": "replace", + "type": "STRING", + "widget": { + "name": "replace" + }, + "link": 122 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 123 + ] + } + ], + "properties": { + "Node name for S&R": "StringReplace", + "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": [ + "", + "{width}", + "" + ] + }, + { + "id": 102, + "type": "StringReplace", + "pos": [ + 220, + -650 + ], + "size": [ + 250, + 450 + ], + "flags": {}, + "order": 17, + "mode": 0, + "inputs": [ + { + "localized_name": "string", + "name": "string", + "type": "STRING", + "widget": { + "name": "string" + }, + "link": 123 + }, + { + "localized_name": "find", + "name": "find", + "type": "STRING", + "widget": { + "name": "find" + }, + "link": null + }, + { + "localized_name": "replace", + "name": "replace", + "type": "STRING", + "widget": { + "name": "replace" + }, + "link": 124 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 117 + ] + } + ], + "properties": { + "Node name for S&R": "StringReplace", + "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": [ + "", + "{height}", + "" + ] + }, + { + "id": 103, + "type": "PreviewAny", + "pos": [ + 970, + -660 + ], + "size": [ + 570, + 790 + ], + "flags": {}, + "order": 18, + "mode": 0, + "inputs": [ + { + "localized_name": "source", + "name": "source", + "type": "*", + "link": 134 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [] + } + ], + "title": "Preview as Text (Int to String)", + "properties": { + "Node name for S&R": "PreviewAny", + "cnr_id": "comfy-core", + "ver": "0.19.0", + "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": [ + null, + null, + null + ] + } + ], + "groups": [ + { + "id": 6, + "title": "Text to Image", + "bounding": [ + -510, + 200, + 1450, + 1060 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 2, + "title": "Image Size", + "bounding": [ + -490, + 950, + 300, + 290 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 3, + "title": "Prompt", + "bounding": [ + -160, + 250, + 470, + 670 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 4, + "title": "Model", + "bounding": [ + -490, + 250, + 300, + 670 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 7, + "title": "Prompt Enhancement", + "bounding": [ + -510, + -720, + 1450, + 890 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + } + ], + "links": [ + { + "id": 73, + "origin_id": 70, + "origin_slot": 0, + "target_id": 65, + "target_slot": 0, + "type": "LATENT" + }, + { + "id": 74, + "origin_id": 63, + "origin_slot": 0, + "target_id": 65, + "target_slot": 1, + "type": "VAE" + }, + { + "id": 85, + "origin_id": 66, + "origin_slot": 0, + "target_id": 70, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 76, + "origin_id": 67, + "origin_slot": 0, + "target_id": 70, + "target_slot": 1, + "type": "CONDITIONING" + }, + { + "id": 80, + "origin_id": 71, + "origin_slot": 0, + "target_id": 70, + "target_slot": 3, + "type": "LATENT" + }, + { + "id": 79, + "origin_id": 62, + "origin_slot": 0, + "target_id": 67, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 84, + "origin_id": 65, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 104, + "origin_id": -10, + "origin_slot": 2, + "target_id": 71, + "target_slot": 0, + "type": "INT" + }, + { + "id": 105, + "origin_id": -10, + "origin_slot": 3, + "target_id": 71, + "target_slot": 1, + "type": "INT" + }, + { + "id": 108, + "origin_id": -10, + "origin_slot": 4, + "target_id": 70, + "target_slot": 4, + "type": "INT" + }, + { + "id": 109, + "origin_id": -10, + "origin_slot": 5, + "target_id": 66, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 110, + "origin_id": -10, + "origin_slot": 6, + "target_id": 62, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 112, + "origin_id": 67, + "origin_slot": 0, + "target_id": 91, + "target_slot": 0, + "type": "CONDITIONING" + }, + { + "id": 113, + "origin_id": 91, + "origin_slot": 0, + "target_id": 70, + "target_slot": 2, + "type": "CONDITIONING" + }, + { + "id": 115, + "origin_id": 94, + "origin_slot": 0, + "target_id": 93, + "target_slot": 2, + "type": "STRING" + }, + { + "id": 116, + "origin_id": 98, + "origin_slot": 0, + "target_id": 95, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 117, + "origin_id": 102, + "origin_slot": 0, + "target_id": 95, + "target_slot": 2, + "type": "STRING" + }, + { + "id": 118, + "origin_id": 94, + "origin_slot": 0, + "target_id": 97, + "target_slot": 0, + "type": "STRING" + }, + { + "id": 119, + "origin_id": 95, + "origin_slot": 0, + "target_id": 97, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 120, + "origin_id": 96, + "origin_slot": 0, + "target_id": 97, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 121, + "origin_id": 93, + "origin_slot": 0, + "target_id": 101, + "target_slot": 0, + "type": "STRING" + }, + { + "id": 122, + "origin_id": 99, + "origin_slot": 0, + "target_id": 101, + "target_slot": 2, + "type": "STRING" + }, + { + "id": 123, + "origin_id": 101, + "origin_slot": 0, + "target_id": 102, + "target_slot": 0, + "type": "STRING" + }, + { + "id": 124, + "origin_id": 100, + "origin_slot": 0, + "target_id": 102, + "target_slot": 2, + "type": "STRING" + }, + { + "id": 127, + "origin_id": -10, + "origin_slot": 1, + "target_id": 96, + "target_slot": 0, + "type": "BOOLEAN" + }, + { + "id": 128, + "origin_id": -10, + "origin_slot": 0, + "target_id": 94, + "target_slot": 0, + "type": "STRING" + }, + { + "id": 129, + "origin_id": -10, + "origin_slot": 2, + "target_id": 99, + "target_slot": 0, + "type": "*" + }, + { + "id": 130, + "origin_id": -10, + "origin_slot": 3, + "target_id": 100, + "target_slot": 0, + "type": "*" + }, + { + "id": 131, + "origin_id": 97, + "origin_slot": 0, + "target_id": 67, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 132, + "origin_id": -10, + "origin_slot": 7, + "target_id": 98, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 133, + "origin_id": -10, + "origin_slot": 8, + "target_id": 63, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 134, + "origin_id": 97, + "origin_slot": 0, + "target_id": 103, + "target_slot": 0, + "type": "STRING" + } + ], + "extra": {}, + "category": "Image generation and editing/Text to image", + "description": "Faster ERNIE Image Turbo variant (~8B DiT, distilled for fewer sampling steps): same strengths in Chinese/English on-image text and layout-heavy graphics as the base ERNIE Image lineup, with bundled encoders and VAE." + } + ] + }, + "extra": { + "ue_links": [] + } +} diff --git a/blueprints/Text to Image (Ernie Image).json b/blueprints/Text to Image (Ernie Image).json new file mode 100644 index 000000000..2bab20d69 --- /dev/null +++ b/blueprints/Text to Image (Ernie Image).json @@ -0,0 +1,2190 @@ +{ + "revision": 0, + "last_node_id": 88, + "last_link_id": 0, + "nodes": [ + { + "id": 88, + "type": "03921aea-a70e-44b4-bc77-f6bda10f2120", + "pos": [ + -120, + 240 + ], + "size": [ + 400, + 540 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "label": "prompt", + "name": "value", + "type": "STRING", + "widget": { + "name": "value" + }, + "link": null + }, + { + "label": "prompt_enhancement", + "name": "value_1", + "type": "BOOLEAN", + "widget": { + "name": "value_1" + }, + "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 + }, + { + "label": "prompt_enhancer", + "name": "clip_name_1", + "type": "COMBO", + "widget": { + "name": "clip_name_1" + }, + "link": null + }, + { + "name": "vae_name", + "type": "COMBO", + "widget": { + "name": "vae_name" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [] + } + ], + "properties": { + "proxyWidgets": [ + [ + "78", + "value" + ], + [ + "76", + "value" + ], + [ + "71", + "width" + ], + [ + "71", + "height" + ], + [ + "70", + "steps" + ], + [ + "70", + "cfg" + ], + [ + "70", + "seed" + ], + [ + "66", + "unet_name" + ], + [ + "62", + "clip_name" + ], + [ + "91", + "clip_name" + ], + [ + "63", + "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, + "ue_properties": { + "widget_ue_connectable": { + "value": true, + "value_1": true + }, + "version": "7.7", + "input_ue_unconnectable": {} + } + }, + "widgets_values": [], + "title": "Text to Image (Ernie Image)" + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "03921aea-a70e-44b4-bc77-f6bda10f2120", + "version": 1, + "state": { + "lastGroupId": 6, + "lastNodeId": 99, + "lastLinkId": 124, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Text to Image (Ernie Image)", + "inputNode": { + "id": -10, + "bounding": [ + -1350, + 370, + 163.50390625, + 260 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + 1110, + 260, + 120, + 60 + ] + }, + "inputs": [ + { + "id": "504de359-52a4-49aa-b6be-23c1cdb0cbde", + "name": "value", + "type": "STRING", + "linkIds": [ + 102 + ], + "label": "prompt", + "pos": [ + -1206.49609375, + 390 + ] + }, + { + "id": "29f699c6-9263-41f6-b37d-69b9fc3913dd", + "name": "value_1", + "type": "BOOLEAN", + "linkIds": [ + 103 + ], + "label": "prompt_enhancement", + "pos": [ + -1206.49609375, + 410 + ] + }, + { + "id": "968e6213-d1e9-4268-8f47-1d6b9a39a43e", + "name": "width", + "type": "INT", + "linkIds": [ + 104, + 113 + ], + "pos": [ + -1206.49609375, + 430 + ] + }, + { + "id": "181c49ef-740d-4385-aa11-79718951ccb9", + "name": "height", + "type": "INT", + "linkIds": [ + 105, + 114 + ], + "pos": [ + -1206.49609375, + 450 + ] + }, + { + "id": "1e85f808-66a1-41df-be52-334142b35419", + "name": "steps", + "type": "INT", + "linkIds": [ + 106 + ], + "pos": [ + -1206.49609375, + 470 + ] + }, + { + "id": "2806addf-a252-4aa3-a5b7-397ab36dccec", + "name": "cfg", + "type": "FLOAT", + "linkIds": [ + 107 + ], + "pos": [ + -1206.49609375, + 490 + ] + }, + { + "id": "5d036a66-5dc0-4d7c-b9a9-349e454738aa", + "name": "seed", + "type": "INT", + "linkIds": [ + 108 + ], + "pos": [ + -1206.49609375, + 510 + ] + }, + { + "id": "360f9a40-aac5-4e9c-bc98-9d55a4a58be2", + "name": "unet_name", + "type": "COMBO", + "linkIds": [ + 109 + ], + "pos": [ + -1206.49609375, + 530 + ] + }, + { + "id": "886301c7-6e88-4cec-96fa-8ae20e8340c5", + "name": "clip_name", + "type": "COMBO", + "linkIds": [ + 110 + ], + "pos": [ + -1206.49609375, + 550 + ] + }, + { + "id": "1d73a545-6d01-462f-bc61-966d4b918ff2", + "name": "clip_name_1", + "type": "COMBO", + "linkIds": [ + 120 + ], + "label": "prompt_enhancer", + "pos": [ + -1206.49609375, + 570 + ] + }, + { + "id": "8c61dc8c-e260-4b36-b73e-d36f90a0bbe3", + "name": "vae_name", + "type": "COMBO", + "linkIds": [ + 121 + ], + "pos": [ + -1206.49609375, + 590 + ] + } + ], + "outputs": [ + { + "id": "f4cb34c8-4090-4281-b428-7338a339d274", + "name": "IMAGE", + "type": "IMAGE", + "linkIds": [ + 84 + ], + "localized_name": "IMAGE", + "pos": [ + 1130, + 280 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 71, + "type": "EmptyFlux2LatentImage", + "pos": [ + -460, + 1040 + ], + "size": [ + 270, + 170 + ], + "flags": {}, + "order": 6, + "mode": 0, + "inputs": [ + { + "localized_name": "width", + "name": "width", + "type": "INT", + "widget": { + "name": "width" + }, + "link": 104 + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "widget": { + "name": "height" + }, + "link": 105 + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "widget": { + "name": "batch_size" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "LATENT", + "name": "LATENT", + "type": "LATENT", + "links": [ + 80 + ] + } + ], + "properties": { + "Node name for S&R": "EmptyFlux2LatentImage", + "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": [ + 1024, + 1024, + 1 + ] + }, + { + "id": 66, + "type": "UNETLoader", + "pos": [ + -470, + 320 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "unet_name", + "name": "unet_name", + "type": "COMBO", + "widget": { + "name": "unet_name" + }, + "link": 109 + }, + { + "localized_name": "weight_dtype", + "name": "weight_dtype", + "type": "COMBO", + "widget": { + "name": "weight_dtype" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "links": [ + 85 + ] + } + ], + "properties": { + "Node name for S&R": "UNETLoader", + "cnr_id": "comfy-core", + "ver": "0.3.73", + "models": [ + { + "name": "ernie-image.safetensors", + "url": "https://huggingface.co/Comfy-Org/ERNIE-Image/resolve/main/diffusion_models/ernie-image.safetensors", + "directory": "diffusion_models" + } + ], + "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": [ + "ernie-image.safetensors", + "default" + ] + }, + { + "id": 65, + "type": "VAEDecode", + "pos": [ + 710, + 280 + ], + "size": [ + 230, + 100 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "samples", + "name": "samples", + "type": "LATENT", + "link": 73 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 74 + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "slot_index": 0, + "links": [ + 84 + ] + } + ], + "properties": { + "Node name for S&R": "VAEDecode", + "cnr_id": "comfy-core", + "ver": "0.3.64", + "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": 70, + "type": "KSampler", + "pos": [ + 350, + 280 + ], + "size": [ + 320, + 350 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 85 + }, + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "link": 76 + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "link": 83 + }, + { + "localized_name": "latent_image", + "name": "latent_image", + "type": "LATENT", + "link": 80 + }, + { + "localized_name": "seed", + "name": "seed", + "type": "INT", + "widget": { + "name": "seed" + }, + "link": 108 + }, + { + "localized_name": "steps", + "name": "steps", + "type": "INT", + "widget": { + "name": "steps" + }, + "link": 106 + }, + { + "localized_name": "cfg", + "name": "cfg", + "type": "FLOAT", + "widget": { + "name": "cfg" + }, + "link": 107 + }, + { + "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": [ + 73 + ] + } + ], + "properties": { + "Node name for S&R": "KSampler", + "cnr_id": "comfy-core", + "ver": "0.3.64", + "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": [ + 182596410725960, + "randomize", + 20, + 4, + "euler", + "simple", + 1 + ] + }, + { + "id": 67, + "type": "CLIPTextEncode", + "pos": [ + -140, + 320 + ], + "size": [ + 410, + 370 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 79 + }, + { + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": 100 + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "links": [ + 76 + ] + } + ], + "properties": { + "Node name for S&R": "CLIPTextEncode", + "cnr_id": "comfy-core", + "ver": "0.3.73", + "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": [ + "" + ], + "color": "#232", + "bgcolor": "#353" + }, + { + "id": 72, + "type": "CLIPTextEncode", + "pos": [ + -130, + 770 + ], + "size": [ + 390, + 140 + ], + "flags": {}, + "order": 7, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 82 + }, + { + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "links": [ + 83 + ] + } + ], + "properties": { + "Node name for S&R": "CLIPTextEncode", + "cnr_id": "comfy-core", + "ver": "0.3.73", + "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": [ + "" + ], + "color": "#223", + "bgcolor": "#335" + }, + { + "id": 83, + "type": "StringReplace", + "pos": [ + -500, + -640 + ], + "size": [ + 430, + 450 + ], + "flags": {}, + "order": 12, + "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": 92 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 115 + ] + } + ], + "properties": { + "Node name for S&R": "StringReplace", + "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": [ + "[SYSTEM_PROMPT]你是一个专业的文生图 Prompt 增强助手。你将收到用户的简短图片描述及目标生成分辨率，请据此扩写为一段内容丰富、细节充分的视觉描述，以帮助文生图模型生成高质量的图片。仅输出增强后的描述，不要包含任何解释或前缀。[/SYSTEM_PROMPT][INST]{\"prompt\": \"{prompt}\", \"width\": {width}, \"height\": {height}}[/INST]", + "{prompt}", + "" + ] + }, + { + "id": 78, + "type": "PrimitiveStringMultiline", + "pos": [ + -950, + -650 + ], + "size": [ + 400, + 200 + ], + "flags": {}, + "order": 11, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "STRING", + "widget": { + "name": "value" + }, + "link": 102 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 87, + 92 + ] + } + ], + "title": "String (Multiline - Prompt)", + "properties": { + "Node name for S&R": "PrimitiveStringMultiline", + "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": [ + "" + ] + }, + { + "id": 74, + "type": "TextGenerate", + "pos": [ + 530, + -650 + ], + "size": [ + 400, + 380 + ], + "flags": {}, + "order": 8, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 112 + }, + { + "localized_name": "image", + "name": "image", + "shape": 7, + "type": "IMAGE", + "link": null + }, + { + "localized_name": "prompt", + "name": "prompt", + "type": "STRING", + "widget": { + "name": "prompt" + }, + "link": 119 + }, + { + "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": "sampling_mode.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": [ + 89 + ] + } + ], + "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, + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + } + }, + "widgets_values": [ + "", + 2048, + "on", + 0.6, + 64, + 0.8, + 0.05, + 1.05, + 0, + 0, + false, + true + ] + }, + { + "id": 76, + "type": "PrimitiveBoolean", + "pos": [ + -500, + 60 + ], + "size": [ + 270, + 100 + ], + "flags": {}, + "order": 10, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "BOOLEAN", + "widget": { + "name": "value" + }, + "link": 103 + } + ], + "outputs": [ + { + "localized_name": "BOOLEAN", + "name": "BOOLEAN", + "type": "BOOLEAN", + "links": [ + 88 + ] + } + ], + "title": "Enable prompt enhancement?", + "properties": { + "Node name for S&R": "PrimitiveBoolean", + "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": [ + true + ] + }, + { + "id": 75, + "type": "ComfySwitchNode", + "pos": [ + 530, + 20 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 9, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 87 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 89 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 88 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 100, + 124 + ] + } + ], + "properties": { + "Node name for S&R": "ComfySwitchNode", + "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": [ + false + ] + }, + { + "id": 62, + "type": "CLIPLoader", + "pos": [ + -460, + 520 + ], + "size": [ + 270, + 150 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "clip_name", + "name": "clip_name", + "type": "COMBO", + "widget": { + "name": "clip_name" + }, + "link": 110 + }, + { + "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": [ + 79, + 82 + ] + } + ], + "properties": { + "Node name for S&R": "CLIPLoader", + "cnr_id": "comfy-core", + "ver": "0.3.73", + "models": [ + { + "name": "ministral-3-3b.safetensors", + "url": "https://huggingface.co/Comfy-Org/ERNIE-Image/resolve/main/text_encoders/ministral-3-3b.safetensors", + "directory": "text_encoders" + } + ], + "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": [ + "ministral-3-3b.safetensors", + "flux2", + "default" + ] + }, + { + "id": 63, + "type": "VAELoader", + "pos": [ + -460, + 770 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "vae_name", + "name": "vae_name", + "type": "COMBO", + "widget": { + "name": "vae_name" + }, + "link": 121 + } + ], + "outputs": [ + { + "localized_name": "VAE", + "name": "VAE", + "type": "VAE", + "links": [ + 74 + ] + } + ], + "properties": { + "Node name for S&R": "VAELoader", + "cnr_id": "comfy-core", + "ver": "0.3.73", + "models": [ + { + "name": "flux2-vae.safetensors", + "url": "https://huggingface.co/Comfy-Org/ERNIE-Image/resolve/main/vae/flux2-vae.safetensors", + "directory": "vae" + } + ], + "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": [ + "flux2-vae.safetensors" + ] + }, + { + "id": 91, + "type": "CLIPLoader", + "pos": [ + -500, + -150 + ], + "size": [ + 510, + 150 + ], + "flags": {}, + "order": 13, + "mode": 0, + "inputs": [ + { + "localized_name": "clip_name", + "name": "clip_name", + "type": "COMBO", + "widget": { + "name": "clip_name" + }, + "link": 120 + }, + { + "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": [ + 112 + ] + } + ], + "title": "Load CLIP (PE)", + "properties": { + "Node name for S&R": "CLIPLoader", + "cnr_id": "comfy-core", + "ver": "0.19.0", + "models": [ + { + "name": "ernie-image-prompt-enhancer.safetensors", + "url": "https://huggingface.co/Comfy-Org/ERNIE-Image/resolve/main/text_encoders/ernie-image-prompt-enhancer.safetensors", + "directory": "text_encoders" + } + ], + "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": [ + "ernie-image-prompt-enhancer.safetensors", + "flux2", + "default" + ] + }, + { + "id": 92, + "type": "PreviewAny", + "pos": [ + -950, + -400 + ], + "size": [ + 400, + 180 + ], + "flags": {}, + "order": 14, + "mode": 0, + "inputs": [ + { + "localized_name": "source", + "name": "source", + "type": "*", + "link": 113 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 116 + ] + } + ], + "title": "Preview as Text (Int to String)", + "properties": { + "Node name for S&R": "PreviewAny", + "cnr_id": "comfy-core", + "ver": "0.19.0", + "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": [ + null, + null, + null + ] + }, + { + "id": 93, + "type": "PreviewAny", + "pos": [ + -950, + -180 + ], + "size": [ + 400, + 180 + ], + "flags": {}, + "order": 15, + "mode": 0, + "inputs": [ + { + "localized_name": "source", + "name": "source", + "type": "*", + "link": 114 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 118 + ] + } + ], + "title": "Preview as Text (Int to String)", + "properties": { + "Node name for S&R": "PreviewAny", + "cnr_id": "comfy-core", + "ver": "0.19.0", + "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": [ + null, + null, + null + ] + }, + { + "id": 94, + "type": "StringReplace", + "pos": [ + -30, + -640 + ], + "size": [ + 230, + 450 + ], + "flags": {}, + "order": 16, + "mode": 0, + "inputs": [ + { + "localized_name": "string", + "name": "string", + "type": "STRING", + "widget": { + "name": "string" + }, + "link": 115 + }, + { + "localized_name": "find", + "name": "find", + "type": "STRING", + "widget": { + "name": "find" + }, + "link": null + }, + { + "localized_name": "replace", + "name": "replace", + "type": "STRING", + "widget": { + "name": "replace" + }, + "link": 116 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 117 + ] + } + ], + "properties": { + "Node name for S&R": "StringReplace", + "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": [ + "", + "{width}", + "" + ] + }, + { + "id": 95, + "type": "StringReplace", + "pos": [ + 220, + -640 + ], + "size": [ + 250, + 450 + ], + "flags": {}, + "order": 17, + "mode": 0, + "inputs": [ + { + "localized_name": "string", + "name": "string", + "type": "STRING", + "widget": { + "name": "string" + }, + "link": 117 + }, + { + "localized_name": "find", + "name": "find", + "type": "STRING", + "widget": { + "name": "find" + }, + "link": null + }, + { + "localized_name": "replace", + "name": "replace", + "type": "STRING", + "widget": { + "name": "replace" + }, + "link": 118 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [ + 119 + ] + } + ], + "properties": { + "Node name for S&R": "StringReplace", + "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": [ + "", + "{height}", + "" + ] + }, + { + "id": 97, + "type": "PreviewAny", + "pos": [ + 970, + -650 + ], + "size": [ + 570, + 790 + ], + "flags": {}, + "order": 18, + "mode": 0, + "inputs": [ + { + "localized_name": "source", + "name": "source", + "type": "*", + "link": 124 + } + ], + "outputs": [ + { + "localized_name": "STRING", + "name": "STRING", + "type": "STRING", + "links": [] + } + ], + "title": "Preview as Text (Int to String)", + "properties": { + "Node name for S&R": "PreviewAny", + "cnr_id": "comfy-core", + "ver": "0.19.0", + "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": [ + null, + null, + null + ] + } + ], + "groups": [ + { + "id": 6, + "title": "Text to Image", + "bounding": [ + -510, + 200, + 1450, + 1060 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 2, + "title": "Image Size", + "bounding": [ + -480, + 940, + 310, + 290 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 3, + "title": "Prompt", + "bounding": [ + -160, + 250, + 470, + 670 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 4, + "title": "Model", + "bounding": [ + -490, + 250, + 320, + 670 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 5, + "title": "Prompt Enhancement", + "bounding": [ + -510, + -720, + 1450, + 890 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + } + ], + "links": [ + { + "id": 73, + "origin_id": 70, + "origin_slot": 0, + "target_id": 65, + "target_slot": 0, + "type": "LATENT" + }, + { + "id": 74, + "origin_id": 63, + "origin_slot": 0, + "target_id": 65, + "target_slot": 1, + "type": "VAE" + }, + { + "id": 85, + "origin_id": 66, + "origin_slot": 0, + "target_id": 70, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 76, + "origin_id": 67, + "origin_slot": 0, + "target_id": 70, + "target_slot": 1, + "type": "CONDITIONING" + }, + { + "id": 83, + "origin_id": 72, + "origin_slot": 0, + "target_id": 70, + "target_slot": 2, + "type": "CONDITIONING" + }, + { + "id": 80, + "origin_id": 71, + "origin_slot": 0, + "target_id": 70, + "target_slot": 3, + "type": "LATENT" + }, + { + "id": 79, + "origin_id": 62, + "origin_slot": 0, + "target_id": 67, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 100, + "origin_id": 75, + "origin_slot": 0, + "target_id": 67, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 82, + "origin_id": 62, + "origin_slot": 0, + "target_id": 72, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 92, + "origin_id": 78, + "origin_slot": 0, + "target_id": 83, + "target_slot": 2, + "type": "STRING" + }, + { + "id": 87, + "origin_id": 78, + "origin_slot": 0, + "target_id": 75, + "target_slot": 0, + "type": "STRING" + }, + { + "id": 89, + "origin_id": 74, + "origin_slot": 0, + "target_id": 75, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 88, + "origin_id": 76, + "origin_slot": 0, + "target_id": 75, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 84, + "origin_id": 65, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 102, + "origin_id": -10, + "origin_slot": 0, + "target_id": 78, + "target_slot": 0, + "type": "STRING" + }, + { + "id": 103, + "origin_id": -10, + "origin_slot": 1, + "target_id": 76, + "target_slot": 0, + "type": "BOOLEAN" + }, + { + "id": 104, + "origin_id": -10, + "origin_slot": 2, + "target_id": 71, + "target_slot": 0, + "type": "INT" + }, + { + "id": 105, + "origin_id": -10, + "origin_slot": 3, + "target_id": 71, + "target_slot": 1, + "type": "INT" + }, + { + "id": 106, + "origin_id": -10, + "origin_slot": 4, + "target_id": 70, + "target_slot": 5, + "type": "INT" + }, + { + "id": 107, + "origin_id": -10, + "origin_slot": 5, + "target_id": 70, + "target_slot": 6, + "type": "FLOAT" + }, + { + "id": 108, + "origin_id": -10, + "origin_slot": 6, + "target_id": 70, + "target_slot": 4, + "type": "INT" + }, + { + "id": 109, + "origin_id": -10, + "origin_slot": 7, + "target_id": 66, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 110, + "origin_id": -10, + "origin_slot": 8, + "target_id": 62, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 112, + "origin_id": 91, + "origin_slot": 0, + "target_id": 74, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 113, + "origin_id": -10, + "origin_slot": 2, + "target_id": 92, + "target_slot": 0, + "type": "*" + }, + { + "id": 114, + "origin_id": -10, + "origin_slot": 3, + "target_id": 93, + "target_slot": 0, + "type": "*" + }, + { + "id": 115, + "origin_id": 83, + "origin_slot": 0, + "target_id": 94, + "target_slot": 0, + "type": "STRING" + }, + { + "id": 116, + "origin_id": 92, + "origin_slot": 0, + "target_id": 94, + "target_slot": 2, + "type": "STRING" + }, + { + "id": 117, + "origin_id": 94, + "origin_slot": 0, + "target_id": 95, + "target_slot": 0, + "type": "STRING" + }, + { + "id": 118, + "origin_id": 93, + "origin_slot": 0, + "target_id": 95, + "target_slot": 2, + "type": "STRING" + }, + { + "id": 119, + "origin_id": 95, + "origin_slot": 0, + "target_id": 74, + "target_slot": 2, + "type": "STRING" + }, + { + "id": 120, + "origin_id": -10, + "origin_slot": 9, + "target_id": 91, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 121, + "origin_id": -10, + "origin_slot": 10, + "target_id": 63, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 124, + "origin_id": 75, + "origin_slot": 0, + "target_id": 97, + "target_slot": 0, + "type": "STRING" + } + ], + "extra": {}, + "category": "Image generation and editing/Text to image", + "description": "Generates images from text prompts using Baidu’s open ERNIE Image (~8B DiT): bilingual in-image typography and layouts (posters, infographics, multi-panel compositions) alongside general scenes, with bundled encoders and VAE." + } + ] + }, + "extra": { + "ue_links": [] + } +} diff --git a/blueprints/Text to Image (Flux.1 Dev).json b/blueprints/Text to Image (Flux.1 Dev).json index 04c3cb95a..6d8446e81 100644 --- a/blueprints/Text to Image (Flux.1 Dev).json +++ b/blueprints/Text to Image (Flux.1 Dev).json @@ -1029,7 +1029,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image generation and editing/Text to image" + "category": "Image generation and editing/Text to image", + "description": "Generates images from prompts using FLUX.1 [dev]: a 12B rectified-flow MMDiT with dual CLIP plus T5-XXL text encoders and guidance-distilled sampling for sharp prompt following versus classic DDPM diffusion." } ] }, @@ -1043,4 +1044,4 @@ }, "ue_links": [] } -} \ No newline at end of file +} diff --git a/blueprints/Text to Image (Flux.1 Krea Dev).json b/blueprints/Text to Image (Flux.1 Krea Dev).json index fe4db1cfc..0d7fa03c4 100644 --- a/blueprints/Text to Image (Flux.1 Krea Dev).json +++ b/blueprints/Text to Image (Flux.1 Krea Dev).json @@ -1023,7 +1023,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image generation and editing/Text to image" + "category": "Image generation and editing/Text to image", + "description": "FLUX.1 Krea [dev] (Black Forest Labs × Krea): open-weight 12B rectified-flow text-to-image drop-in alongside FLUX.1 [dev], tuned away from overcooked saturation toward more natural diversity in people, realism, and style while keeping ecosystem compatibility." } ] }, @@ -1037,4 +1038,4 @@ }, "ue_links": [] } -} \ No newline at end of file +} diff --git a/blueprints/Text to Image (Flux.2 Dev).json b/blueprints/Text to Image (Flux.2 Dev).json new file mode 100644 index 000000000..d5ca3077d --- /dev/null +++ b/blueprints/Text to Image (Flux.2 Dev).json @@ -0,0 +1,1870 @@ +{ + "revision": 0, + "last_node_id": 123, + "last_link_id": 0, + "nodes": [ + { + "id": 123, + "type": "85066daf-feda-4c7b-bbc3-d4797e8ccf0f", + "pos": [ + -800, + 640 + ], + "size": [ + 400, + 0 + ], + "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": "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 + }, + { + "label": "turbo_lora", + "name": "lora_name", + "type": "COMBO", + "widget": { + "name": "lora_name" + }, + "link": null + }, + { + "label": "enable_turbo_mode", + "name": "value", + "type": "BOOLEAN", + "widget": { + "name": "value" + }, + "link": null + }, + { + "name": "noise_seed", + "type": "INT", + "widget": { + "name": "noise_seed" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [] + } + ], + "properties": { + "proxyWidgets": [ + [ + "115", + "text" + ], + [ + "113", + "width" + ], + [ + "113", + "height" + ], + [ + "122", + "unet_name" + ], + [ + "111", + "clip_name" + ], + [ + "108", + "vae_name" + ], + [ + "116", + "lora_name" + ], + [ + "121", + "value" + ], + [ + "114", + "noise_seed" + ], + [ + "114", + "control_after_generate" + ] + ], + "cnr_id": "comfy-core", + "ver": "0.15.1", + "ue_properties": { + "widget_ue_connectable": { + "value": true, + "lora_name": true + }, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [], + "title": "Text to Image (Flux.2 Dev)" + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "85066daf-feda-4c7b-bbc3-d4797e8ccf0f", + "version": 1, + "state": { + "lastGroupId": 6, + "lastNodeId": 123, + "lastLinkId": 232, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Text to Image (Flux.2 Dev)", + "inputNode": { + "id": -10, + "bounding": [ + -1500, + 250, + 151.744140625, + 220 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + 1560, + -20, + 120, + 60 + ] + }, + "inputs": [ + { + "id": "1f4f1091-3f97-41d8-8ed8-e8b02260cf3c", + "name": "text", + "type": "STRING", + "linkIds": [ + 206 + ], + "label": "prompt", + "pos": [ + -1368.255859375, + 270 + ] + }, + { + "id": "b9b59411-4f5f-4482-8f78-369e6d50e71c", + "name": "width", + "type": "INT", + "linkIds": [ + 222, + 231 + ], + "pos": [ + -1368.255859375, + 290 + ] + }, + { + "id": "c6de9a28-3bf6-40d0-be16-f75ec517a766", + "name": "height", + "type": "INT", + "linkIds": [ + 223, + 232 + ], + "pos": [ + -1368.255859375, + 310 + ] + }, + { + "id": "8f1b1c75-e47c-45f5-af57-74abcfe8967c", + "name": "unet_name", + "type": "COMBO", + "linkIds": [ + 225 + ], + "pos": [ + -1368.255859375, + 330 + ] + }, + { + "id": "6ac27631-1bf0-4161-9670-a662f6180b94", + "name": "clip_name", + "type": "COMBO", + "linkIds": [ + 226 + ], + "pos": [ + -1368.255859375, + 350 + ] + }, + { + "id": "932e6cbe-f716-4905-ae54-d2b3543497bd", + "name": "vae_name", + "type": "COMBO", + "linkIds": [ + 227 + ], + "pos": [ + -1368.255859375, + 370 + ] + }, + { + "id": "37400048-5e7b-427b-8b79-ea35841d5306", + "name": "lora_name", + "type": "COMBO", + "linkIds": [ + 228 + ], + "label": "turbo_lora", + "pos": [ + -1368.255859375, + 390 + ] + }, + { + "id": "333212d0-f027-476f-8b97-a921e20e340a", + "name": "value", + "type": "BOOLEAN", + "linkIds": [ + 229 + ], + "label": "enable_turbo_mode", + "pos": [ + -1368.255859375, + 410 + ] + }, + { + "id": "e7e73fad-ce6e-48d5-b719-e2abed685185", + "name": "noise_seed", + "type": "INT", + "linkIds": [ + 230 + ], + "pos": [ + -1368.255859375, + 430 + ] + } + ], + "outputs": [ + { + "id": "ed3c0a0f-a39f-453e-907f-8249c8e3335d", + "name": "IMAGE", + "type": "IMAGE", + "linkIds": [ + 9 + ], + "localized_name": "IMAGE", + "pos": [ + 1580, + 0 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 105, + "type": "BasicGuider", + "pos": [ + 570, + 170 + ], + "size": [ + 230, + 100 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 210 + }, + { + "localized_name": "conditioning", + "name": "conditioning", + "type": "CONDITIONING", + "link": 165 + } + ], + "outputs": [ + { + "localized_name": "GUIDER", + "name": "GUIDER", + "type": "GUIDER", + "slot_index": 0, + "links": [ + 30 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "BasicGuider", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 106, + "type": "FluxGuidance", + "pos": [ + -510, + 470 + ], + "size": [ + 320, + 110 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "localized_name": "conditioning", + "name": "conditioning", + "type": "CONDITIONING", + "link": 41 + }, + { + "localized_name": "guidance", + "name": "guidance", + "type": "FLOAT", + "widget": { + "name": "guidance" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "slot_index": 0, + "links": [ + 165 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "FluxGuidance", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 4 + ], + "color": "#233", + "bgcolor": "#355" + }, + { + "id": 107, + "type": "KSamplerSelect", + "pos": [ + 570, + 350 + ], + "size": [ + 230, + 110 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "sampler_name", + "name": "sampler_name", + "type": "COMBO", + "widget": { + "name": "sampler_name" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "SAMPLER", + "name": "SAMPLER", + "type": "SAMPLER", + "links": [ + 19 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "KSamplerSelect", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "euler" + ] + }, + { + "id": 108, + "type": "VAELoader", + "pos": [ + -1000, + 460 + ], + "size": [ + 300, + 110 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "localized_name": "vae_name", + "name": "vae_name", + "type": "COMBO", + "widget": { + "name": "vae_name" + }, + "link": 227 + } + ], + "outputs": [ + { + "localized_name": "VAE", + "name": "VAE", + "type": "VAE", + "slot_index": 0, + "links": [ + 159 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "VAELoader", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "models": [ + { + "name": "full_encoder_small_decoder.safetensors", + "url": "https://huggingface.co/black-forest-labs/FLUX.2-small-decoder/resolve/main/full_encoder_small_decoder.safetensors", + "directory": "vae" + } + ] + }, + "widgets_values": [ + "full_encoder_small_decoder.safetensors" + ] + }, + { + "id": 109, + "type": "SamplerCustomAdvanced", + "pos": [ + 860, + -20 + ], + "size": [ + 280, + 330 + ], + "flags": {}, + "order": 6, + "mode": 0, + "inputs": [ + { + "localized_name": "noise", + "name": "noise", + "type": "NOISE", + "link": 37 + }, + { + "localized_name": "guider", + "name": "guider", + "type": "GUIDER", + "link": 30 + }, + { + "localized_name": "sampler", + "name": "sampler", + "type": "SAMPLER", + "link": 19 + }, + { + "localized_name": "sigmas", + "name": "sigmas", + "type": "SIGMAS", + "link": 132 + }, + { + "localized_name": "latent_image", + "name": "latent_image", + "type": "LATENT", + "link": 161 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "LATENT", + "slot_index": 0, + "links": [ + 24 + ] + }, + { + "localized_name": "denoised_output", + "name": "denoised_output", + "type": "LATENT", + "links": null + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "SamplerCustomAdvanced", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 110, + "type": "VAEDecode", + "pos": [ + 1220, + -20 + ], + "size": [ + 230, + 100 + ], + "flags": {}, + "order": 7, + "mode": 0, + "inputs": [ + { + "localized_name": "samples", + "name": "samples", + "type": "LATENT", + "link": 24 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 159 + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "slot_index": 0, + "links": [ + 9 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "VAEDecode", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 111, + "type": "CLIPLoader", + "pos": [ + -1000, + 200 + ], + "size": [ + 300, + 150 + ], + "flags": {}, + "order": 8, + "mode": 0, + "inputs": [ + { + "localized_name": "clip_name", + "name": "clip_name", + "type": "COMBO", + "widget": { + "name": "clip_name" + }, + "link": 226 + }, + { + "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": [ + 117 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "CLIPLoader", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "models": [ + { + "name": "mistral_3_small_flux2_bf16.safetensors", + "url": "https://huggingface.co/Comfy-Org/flux2-dev/resolve/main/split_files/text_encoders/mistral_3_small_flux2_bf16.safetensors", + "directory": "text_encoders" + } + ] + }, + "widgets_values": [ + "mistral_3_small_flux2_bf16.safetensors", + "flux2", + "default" + ] + }, + { + "id": 112, + "type": "Flux2Scheduler", + "pos": [ + 570, + 550 + ], + "size": [ + 230, + 170 + ], + "flags": {}, + "order": 9, + "mode": 0, + "inputs": [ + { + "localized_name": "steps", + "name": "steps", + "type": "INT", + "widget": { + "name": "steps" + }, + "link": 213 + }, + { + "localized_name": "width", + "name": "width", + "type": "INT", + "widget": { + "name": "width" + }, + "link": 231 + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "widget": { + "name": "height" + }, + "link": 232 + } + ], + "outputs": [ + { + "localized_name": "SIGMAS", + "name": "SIGMAS", + "type": "SIGMAS", + "links": [ + 132 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "Flux2Scheduler", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 20, + 1024, + 1024 + ] + }, + { + "id": 113, + "type": "EmptyFlux2LatentImage", + "pos": [ + -980, + 660 + ], + "size": [ + 270, + 170 + ], + "flags": {}, + "order": 10, + "mode": 0, + "inputs": [ + { + "localized_name": "width", + "name": "width", + "type": "INT", + "widget": { + "name": "width" + }, + "link": 222 + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "widget": { + "name": "height" + }, + "link": 223 + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "widget": { + "name": "batch_size" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "LATENT", + "name": "LATENT", + "type": "LATENT", + "links": [ + 161 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "EmptyFlux2LatentImage", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 1024, + 1024, + 1 + ] + }, + { + "id": 114, + "type": "RandomNoise", + "pos": [ + 570, + -20 + ], + "size": [ + 230, + 110 + ], + "flags": {}, + "order": 11, + "mode": 0, + "inputs": [ + { + "localized_name": "noise_seed", + "name": "noise_seed", + "type": "INT", + "widget": { + "name": "noise_seed" + }, + "link": 230 + } + ], + "outputs": [ + { + "localized_name": "NOISE", + "name": "NOISE", + "type": "NOISE", + "links": [ + 37 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "RandomNoise", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 1027111520328378, + "randomize" + ] + }, + { + "id": 115, + "type": "CLIPTextEncode", + "pos": [ + -630, + -40 + ], + "size": [ + 440, + 450 + ], + "flags": {}, + "order": 12, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 117 + }, + { + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": 206 + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "slot_index": 0, + "links": [ + 41 + ] + } + ], + "title": "CLIP Text Encode (Positive Prompt)", + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "CLIPTextEncode", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "" + ], + "color": "#232", + "bgcolor": "#353" + }, + { + "id": 116, + "type": "LoraLoaderModelOnly", + "pos": [ + -150, + 220 + ], + "size": [ + 300, + 140 + ], + "flags": {}, + "order": 13, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 221 + }, + { + "localized_name": "lora_name", + "name": "lora_name", + "type": "COMBO", + "widget": { + "name": "lora_name" + }, + "link": 228 + }, + { + "localized_name": "strength_model", + "name": "strength_model", + "type": "FLOAT", + "widget": { + "name": "strength_model" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "links": [ + 209 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.7.0", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "LoraLoaderModelOnly", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "models": [ + { + "name": "Flux_2-Turbo-LoRA_comfyui.safetensors", + "url": "https://huggingface.co/ByteZSzn/Flux.2-Turbo-ComfyUI/resolve/main/Flux_2-Turbo-LoRA_comfyui.safetensors", + "directory": "loras" + } + ] + }, + "widgets_values": [ + "Flux_2-Turbo-LoRA_comfyui.safetensors", + 1 + ] + }, + { + "id": 117, + "type": "ComfySwitchNode", + "pos": [ + 220, + -30 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 14, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 208 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 209 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 215 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 210 + ] + } + ], + "title": "Switch(model)", + "properties": { + "cnr_id": "comfy-core", + "ver": "0.15.1", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "ComfySwitchNode", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + false + ] + }, + { + "id": 118, + "type": "PrimitiveInt", + "pos": [ + -140, + -30 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "INT", + "widget": { + "name": "value" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "INT", + "name": "INT", + "type": "INT", + "links": [ + 211 + ] + } + ], + "title": "Steps", + "properties": { + "cnr_id": "comfy-core", + "ver": "0.15.1", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "PrimitiveInt", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 20, + "fixed" + ] + }, + { + "id": 119, + "type": "PrimitiveInt", + "pos": [ + -150, + 460 + ], + "size": [ + 300, + 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": [ + 212 + ] + } + ], + "title": "Steps", + "properties": { + "cnr_id": "comfy-core", + "ver": "0.15.1", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "PrimitiveInt", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 8, + "fixed" + ] + }, + { + "id": 120, + "type": "ComfySwitchNode", + "pos": [ + 220, + 260 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 15, + "mode": 0, + "inputs": [ + { + "localized_name": "on_false", + "name": "on_false", + "type": "*", + "link": 211 + }, + { + "localized_name": "on_true", + "name": "on_true", + "type": "*", + "link": 212 + }, + { + "localized_name": "switch", + "name": "switch", + "type": "BOOLEAN", + "widget": { + "name": "switch" + }, + "link": 214 + } + ], + "outputs": [ + { + "localized_name": "output", + "name": "output", + "type": "*", + "links": [ + 213 + ] + } + ], + "title": "Switch(steps)", + "properties": { + "cnr_id": "comfy-core", + "ver": "0.15.1", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "ComfySwitchNode", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + false + ] + }, + { + "id": 121, + "type": "PrimitiveBoolean", + "pos": [ + -110, + 690 + ], + "size": [ + 270, + 100 + ], + "flags": {}, + "order": 16, + "mode": 0, + "inputs": [ + { + "localized_name": "value", + "name": "value", + "type": "BOOLEAN", + "widget": { + "name": "value" + }, + "link": 229 + } + ], + "outputs": [ + { + "localized_name": "BOOLEAN", + "name": "BOOLEAN", + "type": "BOOLEAN", + "links": [ + 214, + 215 + ] + } + ], + "title": "Enable Turbo LoRA", + "properties": { + "cnr_id": "comfy-core", + "ver": "0.15.1", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "PrimitiveBoolean", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + false + ] + }, + { + "id": 122, + "type": "UNETLoader", + "pos": [ + -1000, + -30 + ], + "size": [ + 300, + 110 + ], + "flags": {}, + "order": 17, + "mode": 0, + "inputs": [ + { + "localized_name": "unet_name", + "name": "unet_name", + "type": "COMBO", + "widget": { + "name": "unet_name" + }, + "link": 225 + }, + { + "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": [ + 208, + 221 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.71", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "UNETLoader", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "models": [ + { + "name": "flux2_dev_fp8mixed.safetensors", + "url": "https://huggingface.co/Comfy-Org/flux2-dev/resolve/main/split_files/diffusion_models/flux2_dev_fp8mixed.safetensors", + "directory": "diffusion_models" + } + ] + }, + "widgets_values": [ + "flux2_dev_fp8mixed.safetensors", + "default" + ] + } + ], + "groups": [ + { + "id": 1, + "title": "Step 1 - Upload models", + "bounding": [ + -1040, + -110, + 380, + 710 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 2, + "title": "Custom sampler", + "bounding": [ + 540, + -110, + 640, + 870 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 4, + "title": "Step2 - Prompt", + "bounding": [ + -640, + -110, + 460, + 710 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 5, + "title": "Original", + "bounding": [ + -160, + -110, + 320, + 230 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 6, + "title": "8 Steps LoRA", + "bounding": [ + -160, + 140, + 320, + 460 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + } + ], + "links": [ + { + "id": 165, + "origin_id": 106, + "origin_slot": 0, + "target_id": 105, + "target_slot": 1, + "type": "CONDITIONING" + }, + { + "id": 41, + "origin_id": 115, + "origin_slot": 0, + "target_id": 106, + "target_slot": 0, + "type": "CONDITIONING" + }, + { + "id": 37, + "origin_id": 114, + "origin_slot": 0, + "target_id": 109, + "target_slot": 0, + "type": "NOISE" + }, + { + "id": 30, + "origin_id": 105, + "origin_slot": 0, + "target_id": 109, + "target_slot": 1, + "type": "GUIDER" + }, + { + "id": 19, + "origin_id": 107, + "origin_slot": 0, + "target_id": 109, + "target_slot": 2, + "type": "SAMPLER" + }, + { + "id": 132, + "origin_id": 112, + "origin_slot": 0, + "target_id": 109, + "target_slot": 3, + "type": "SIGMAS" + }, + { + "id": 161, + "origin_id": 113, + "origin_slot": 0, + "target_id": 109, + "target_slot": 4, + "type": "LATENT" + }, + { + "id": 117, + "origin_id": 111, + "origin_slot": 0, + "target_id": 115, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 24, + "origin_id": 109, + "origin_slot": 0, + "target_id": 110, + "target_slot": 0, + "type": "LATENT" + }, + { + "id": 159, + "origin_id": 108, + "origin_slot": 0, + "target_id": 110, + "target_slot": 1, + "type": "VAE" + }, + { + "id": 9, + "origin_id": 110, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 206, + "origin_id": -10, + "origin_slot": 0, + "target_id": 115, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 208, + "origin_id": 122, + "origin_slot": 0, + "target_id": 117, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 209, + "origin_id": 116, + "origin_slot": 0, + "target_id": 117, + "target_slot": 1, + "type": "MODEL" + }, + { + "id": 210, + "origin_id": 117, + "origin_slot": 0, + "target_id": 105, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 211, + "origin_id": 118, + "origin_slot": 0, + "target_id": 120, + "target_slot": 0, + "type": "INT" + }, + { + "id": 212, + "origin_id": 119, + "origin_slot": 0, + "target_id": 120, + "target_slot": 1, + "type": "INT" + }, + { + "id": 213, + "origin_id": 120, + "origin_slot": 0, + "target_id": 112, + "target_slot": 0, + "type": "INT" + }, + { + "id": 214, + "origin_id": 121, + "origin_slot": 0, + "target_id": 120, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 215, + "origin_id": 121, + "origin_slot": 0, + "target_id": 117, + "target_slot": 2, + "type": "BOOLEAN" + }, + { + "id": 221, + "origin_id": 122, + "origin_slot": 0, + "target_id": 116, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 222, + "origin_id": -10, + "origin_slot": 1, + "target_id": 113, + "target_slot": 0, + "type": "INT" + }, + { + "id": 223, + "origin_id": -10, + "origin_slot": 2, + "target_id": 113, + "target_slot": 1, + "type": "INT" + }, + { + "id": 225, + "origin_id": -10, + "origin_slot": 3, + "target_id": 122, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 226, + "origin_id": -10, + "origin_slot": 4, + "target_id": 111, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 227, + "origin_id": -10, + "origin_slot": 5, + "target_id": 108, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 228, + "origin_id": -10, + "origin_slot": 6, + "target_id": 116, + "target_slot": 1, + "type": "COMBO" + }, + { + "id": 229, + "origin_id": -10, + "origin_slot": 7, + "target_id": 121, + "target_slot": 0, + "type": "BOOLEAN" + }, + { + "id": 230, + "origin_id": -10, + "origin_slot": 8, + "target_id": 114, + "target_slot": 0, + "type": "INT" + }, + { + "id": 231, + "origin_id": -10, + "origin_slot": 1, + "target_id": 112, + "target_slot": 1, + "type": "INT" + }, + { + "id": 232, + "origin_id": -10, + "origin_slot": 2, + "target_id": 112, + "target_slot": 2, + "type": "INT" + } + ], + "extra": { + "workflowRendererVersion": "LG" + }, + "category": "Image generation and editing/Text to image", + "description": "Generates images from prompts using FLUX.2 [dev]: a newer 32B rectified-flow stack with distilled guidance plus a stronger long-context multimodal encoder for complex scenes, sharper typography/UI text, anatomy, lighting, and high-resolution latent decoding." + } + ] + }, + "extra": { + "ue_links": [] + } +} diff --git a/blueprints/Text to Image (NetaYume Lumina).json b/blueprints/Text to Image (NetaYume Lumina).json index 394ad1608..9e11b7a86 100644 --- a/blueprints/Text to Image (NetaYume Lumina).json +++ b/blueprints/Text to Image (NetaYume Lumina).json @@ -1104,7 +1104,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image generation and editing/Text to image" + "category": "Image generation and editing/Text to image", + "description": "Generates images from text prompts using NetaYume Lumina, fine-tuned from Neta Lumina for anime-style and illustration generation." }, { "id": "a07fdf06-1bda-4dac-bdbd-63ee8ebca1c9", @@ -1458,11 +1459,12 @@ ], "extra": { "workflowRendererVersion": "LG" - } + }, + "description": "Encodes a negative text prompt via CLIP for classifier-free guidance in anime-style generation (NetaYume Lumina)." } ] }, "extra": { "ue_links": [] } -} \ No newline at end of file +} diff --git a/blueprints/Text to Image (Qwen-Image 2512).json b/blueprints/Text to Image (Qwen-Image 2512).json index f52ea2ef2..09612be8b 100644 --- a/blueprints/Text to Image (Qwen-Image 2512).json +++ b/blueprints/Text to Image (Qwen-Image 2512).json @@ -1941,7 +1941,8 @@ "extra": { "workflowRendererVersion": "Vue-corrected" }, - "category": "Image generation and editing/Text to image" + "category": "Image generation and editing/Text to image", + "description": "Generates images from text prompts using Qwen-Image-2512, with enhanced human realism and finer natural detail over the base version." } ] }, diff --git a/blueprints/Text to Image (Qwen-Image).json b/blueprints/Text to Image (Qwen-Image).json index 70b4b44b3..e78d5a962 100644 --- a/blueprints/Text to Image (Qwen-Image).json +++ b/blueprints/Text to Image (Qwen-Image).json @@ -1873,7 +1873,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image generation and editing/Text to image" + "category": "Image generation and editing/Text to image", + "description": "Generates images from text prompts using Qwen-Image, Alibaba's 20B MMDiT model with excellent multilingual text rendering." } ] }, diff --git a/blueprints/Text to Image (Z-Image-Base).json b/blueprints/Text to Image (Z-Image-Base).json new file mode 100644 index 000000000..169263712 --- /dev/null +++ b/blueprints/Text to Image (Z-Image-Base).json @@ -0,0 +1,1184 @@ +{ + "revision": 0, + "last_node_id": 126, + "last_link_id": 0, + "nodes": [ + { + "id": 126, + "type": "8a2bb267-5858-4aaf-bdcd-61002711af19", + "pos": [ + -2280, + 2850 + ], + "size": [ + 410, + 560 + ], + "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": [ + [ + "67", + "text" + ], + [ + "68", + "width" + ], + [ + "68", + "height" + ], + [ + "69", + "steps" + ], + [ + "69", + "cfg" + ], + [ + "69", + "seed" + ], + [ + "66", + "unet_name" + ], + [ + "62", + "clip_name" + ], + [ + "63", + "vae_name" + ], + [ + "69", + "control_after_generate" + ] + ], + "cnr_id": "comfy-core", + "ver": "0.13.0", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [], + "title": "Text to Image (Z-Image-Base)" + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "8a2bb267-5858-4aaf-bdcd-61002711af19", + "version": 1, + "state": { + "lastGroupId": 16, + "lastNodeId": 126, + "lastLinkId": 229, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Text to Image (Z-Image-Base)", + "description": "Generates images from text prompts using Z-Image base weights with Qwen3 text encoder and bundled VAE.", + "inputNode": { + "id": -10, + "bounding": [ + -220, + 40, + 120, + 220 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + 1840, + -150, + 120, + 60 + ] + }, + "inputs": [ + { + "id": "af36fee5-4f8b-4a8e-bfa8-cb8fe7006cc3", + "name": "text", + "type": "STRING", + "linkIds": [ + 108 + ], + "label": "prompt", + "pos": [ + -120, + 60 + ] + }, + { + "id": "357f0059-e8e6-41f6-a290-c53b0a60c0ed", + "name": "width", + "type": "INT", + "linkIds": [ + 114 + ], + "pos": [ + -120, + 80 + ] + }, + { + "id": "4a442743-a9c2-4aa5-9efd-05d43f3322d3", + "name": "height", + "type": "INT", + "linkIds": [ + 115 + ], + "pos": [ + -120, + 100 + ] + }, + { + "id": "a0fc336b-d349-418e-8415-318653f7b6b3", + "name": "steps", + "type": "INT", + "linkIds": [ + 116 + ], + "pos": [ + -120, + 120 + ] + }, + { + "id": "2f253ace-1e1a-415f-9b95-a10430bd5749", + "name": "cfg", + "type": "FLOAT", + "linkIds": [ + 117 + ], + "pos": [ + -120, + 140 + ] + }, + { + "id": "18a6ad37-23aa-4bf7-a0cd-1d6ca6e2a128", + "name": "seed", + "type": "INT", + "linkIds": [ + 118 + ], + "pos": [ + -120, + 160 + ] + }, + { + "id": "d1fc4937-8505-4ec6-9fc4-a33ef7b45eee", + "name": "unet_name", + "type": "COMBO", + "linkIds": [ + 119 + ], + "pos": [ + -120, + 180 + ] + }, + { + "id": "db45dd49-d990-4ceb-a849-f96341874cdd", + "name": "clip_name", + "type": "COMBO", + "linkIds": [ + 120 + ], + "pos": [ + -120, + 200 + ] + }, + { + "id": "37b8eac6-9b1b-452b-81f3-0ba9e34a576a", + "name": "vae_name", + "type": "COMBO", + "linkIds": [ + 121 + ], + "pos": [ + -120, + 220 + ] + } + ], + "outputs": [ + { + "id": "f2bea309-bfe7-4ccb-9ffe-9475bf1da2ae", + "name": "IMAGE", + "type": "IMAGE", + "linkIds": [ + 79 + ], + "localized_name": "IMAGE", + "pos": [ + 1860, + -130 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 67, + "type": "CLIPTextEncode", + "pos": [ + 600, + -90 + ], + "size": [ + 410, + 320 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 78 + }, + { + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": 108 + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "links": [ + 75 + ] + } + ], + "title": "CLIP Text Encode (Positive Prompt)", + "properties": { + "Node name for S&R": "CLIPTextEncode", + "cnr_id": "comfy-core", + "ver": "0.3.73", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "" + ], + "color": "#232", + "bgcolor": "#353" + }, + { + "id": 68, + "type": "EmptySD3LatentImage", + "pos": [ + 240, + 620 + ], + "size": [ + 260, + 170 + ], + "flags": {}, + "order": 6, + "mode": 0, + "inputs": [ + { + "localized_name": "width", + "name": "width", + "type": "INT", + "widget": { + "name": "width" + }, + "link": 114 + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "widget": { + "name": "height" + }, + "link": 115 + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "widget": { + "name": "batch_size" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "LATENT", + "name": "LATENT", + "type": "LATENT", + "slot_index": 0, + "links": [ + 77 + ] + } + ], + "properties": { + "Node name for S&R": "EmptySD3LatentImage", + "cnr_id": "comfy-core", + "ver": "0.3.64", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 1024, + 1024, + 1 + ] + }, + { + "id": 63, + "type": "VAELoader", + "pos": [ + 230, + 340 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "vae_name", + "name": "vae_name", + "type": "COMBO", + "widget": { + "name": "vae_name" + }, + "link": 121 + } + ], + "outputs": [ + { + "localized_name": "VAE", + "name": "VAE", + "type": "VAE", + "links": [ + 73 + ] + } + ], + "properties": { + "Node name for S&R": "VAELoader", + "cnr_id": "comfy-core", + "ver": "0.3.73", + "models": [ + { + "name": "ae.safetensors", + "url": "https://huggingface.co/Comfy-Org/z_image_turbo/resolve/main/split_files/vae/ae.safetensors", + "directory": "vae" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "ae.safetensors" + ] + }, + { + "id": 62, + "type": "CLIPLoader", + "pos": [ + 230, + 110 + ], + "size": [ + 270, + 150 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "clip_name", + "name": "clip_name", + "type": "COMBO", + "widget": { + "name": "clip_name" + }, + "link": 120 + }, + { + "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": [ + 78, + 82 + ] + } + ], + "properties": { + "Node name for S&R": "CLIPLoader", + "cnr_id": "comfy-core", + "ver": "0.3.73", + "models": [ + { + "name": "qwen_3_4b.safetensors", + "url": "https://huggingface.co/Comfy-Org/z_image_turbo/resolve/main/split_files/text_encoders/qwen_3_4b.safetensors", + "directory": "text_encoders" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "qwen_3_4b.safetensors", + "lumina2", + "default" + ] + }, + { + "id": 65, + "type": "VAEDecode", + "pos": [ + 1450, + -150 + ], + "size": [ + 230, + 100 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "samples", + "name": "samples", + "type": "LATENT", + "link": 72 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 73 + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "slot_index": 0, + "links": [ + 79 + ] + } + ], + "properties": { + "Node name for S&R": "VAEDecode", + "cnr_id": "comfy-core", + "ver": "0.3.64", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 70, + "type": "ModelSamplingAuraFlow", + "pos": [ + 1100, + -150 + ], + "size": [ + 310, + 110 + ], + "flags": {}, + "order": 8, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 109 + }, + { + "localized_name": "shift", + "name": "shift", + "type": "FLOAT", + "widget": { + "name": "shift" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "slot_index": 0, + "links": [ + 74 + ] + } + ], + "properties": { + "Node name for S&R": "ModelSamplingAuraFlow", + "cnr_id": "comfy-core", + "ver": "0.3.64", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 3 + ] + }, + { + "id": 66, + "type": "UNETLoader", + "pos": [ + 230, + -90 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "localized_name": "unet_name", + "name": "unet_name", + "type": "COMBO", + "widget": { + "name": "unet_name" + }, + "link": 119 + }, + { + "localized_name": "weight_dtype", + "name": "weight_dtype", + "type": "COMBO", + "widget": { + "name": "weight_dtype" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "links": [ + 109 + ] + } + ], + "properties": { + "Node name for S&R": "UNETLoader", + "cnr_id": "comfy-core", + "ver": "0.3.73", + "models": [ + { + "name": "z_image_bf16.safetensors", + "url": "https://huggingface.co/Comfy-Org/z_image/resolve/main/split_files/diffusion_models/z_image_bf16.safetensors", + "directory": "diffusion_models" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "z_image_bf16.safetensors", + "default" + ] + }, + { + "id": 71, + "type": "CLIPTextEncode", + "pos": [ + 600, + 310 + ], + "size": [ + 390, + 140 + ], + "flags": {}, + "order": 9, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 82 + }, + { + "label": "prompt", + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "links": [ + 83 + ] + } + ], + "title": "CLIP Text Encode (Negative Prompt)", + "properties": { + "Node name for S&R": "CLIPTextEncode", + "cnr_id": "comfy-core", + "ver": "0.3.73", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "" + ], + "color": "#323", + "bgcolor": "#535" + }, + { + "id": 69, + "type": "KSampler", + "pos": [ + 1100, + 10 + ], + "size": [ + 310, + 440 + ], + "flags": {}, + "order": 7, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 74 + }, + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "link": 75 + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "link": 83 + }, + { + "localized_name": "latent_image", + "name": "latent_image", + "type": "LATENT", + "link": 77 + }, + { + "localized_name": "seed", + "name": "seed", + "type": "INT", + "widget": { + "name": "seed" + }, + "link": 118 + }, + { + "localized_name": "steps", + "name": "steps", + "type": "INT", + "widget": { + "name": "steps" + }, + "link": 116 + }, + { + "localized_name": "cfg", + "name": "cfg", + "type": "FLOAT", + "widget": { + "name": "cfg" + }, + "link": 117 + }, + { + "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": [ + 72 + ] + } + ], + "properties": { + "Node name for S&R": "KSampler", + "cnr_id": "comfy-core", + "ver": "0.3.64", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 0, + "randomize", + 25, + 4, + "res_multistep", + "simple", + 1 + ] + }, + { + "id": 87, + "type": "MarkdownNote", + "pos": [ + 1110, + -360 + ], + "size": [ + 300, + 120 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [], + "outputs": [], + "properties": {}, + "widgets_values": [ + "- Steps: 30～50\n- cfg: 3～5" + ], + "color": "#222", + "bgcolor": "#000", + "title": "Original Settings" + } + ], + "groups": [ + { + "id": 2, + "title": "Step2 - Image size", + "bounding": [ + 200, + 530, + 330, + 287.9999544955691 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 3, + "title": "Step3 - Prompt", + "bounding": [ + 570, + -200, + 470, + 700 + ], + "color": "#3f789e", + "flags": {} + }, + { + "id": 4, + "title": "Step1 - Load models", + "bounding": [ + 200, + -200, + 330, + 700 + ], + "color": "#3f789e", + "flags": {} + } + ], + "links": [ + { + "id": 78, + "origin_id": 62, + "origin_slot": 0, + "target_id": 67, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 74, + "origin_id": 70, + "origin_slot": 0, + "target_id": 69, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 75, + "origin_id": 67, + "origin_slot": 0, + "target_id": 69, + "target_slot": 1, + "type": "CONDITIONING" + }, + { + "id": 83, + "origin_id": 71, + "origin_slot": 0, + "target_id": 69, + "target_slot": 2, + "type": "CONDITIONING" + }, + { + "id": 77, + "origin_id": 68, + "origin_slot": 0, + "target_id": 69, + "target_slot": 3, + "type": "LATENT" + }, + { + "id": 82, + "origin_id": 62, + "origin_slot": 0, + "target_id": 71, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 72, + "origin_id": 69, + "origin_slot": 0, + "target_id": 65, + "target_slot": 0, + "type": "LATENT" + }, + { + "id": 73, + "origin_id": 63, + "origin_slot": 0, + "target_id": 65, + "target_slot": 1, + "type": "VAE" + }, + { + "id": 79, + "origin_id": 65, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 108, + "origin_id": -10, + "origin_slot": 0, + "target_id": 67, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 109, + "origin_id": 66, + "origin_slot": 0, + "target_id": 70, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 114, + "origin_id": -10, + "origin_slot": 1, + "target_id": 68, + "target_slot": 0, + "type": "INT" + }, + { + "id": 115, + "origin_id": -10, + "origin_slot": 2, + "target_id": 68, + "target_slot": 1, + "type": "INT" + }, + { + "id": 116, + "origin_id": -10, + "origin_slot": 3, + "target_id": 69, + "target_slot": 5, + "type": "INT" + }, + { + "id": 117, + "origin_id": -10, + "origin_slot": 4, + "target_id": 69, + "target_slot": 6, + "type": "FLOAT" + }, + { + "id": 118, + "origin_id": -10, + "origin_slot": 5, + "target_id": 69, + "target_slot": 4, + "type": "INT" + }, + { + "id": 119, + "origin_id": -10, + "origin_slot": 6, + "target_id": 66, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 120, + "origin_id": -10, + "origin_slot": 7, + "target_id": 62, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 121, + "origin_id": -10, + "origin_slot": 8, + "target_id": 63, + "target_slot": 0, + "type": "COMBO" + } + ], + "extra": { + "workflowRendererVersion": "LG" + }, + "category": "Image generation and editing/Text to image" + } + ] + }, + "extra": {} +} \ No newline at end of file diff --git a/blueprints/Text to Image (Z-Image-Turbo).json b/blueprints/Text to Image (Z-Image-Turbo).json index 6aa80e327..2501486fa 100644 --- a/blueprints/Text to Image (Z-Image-Turbo).json +++ b/blueprints/Text to Image (Z-Image-Turbo).json @@ -1,22 +1,21 @@ { - "id": "1c3eaa76-5cfa-4dc7-8571-97a570324e01", "revision": 0, - "last_node_id": 34, - "last_link_id": 40, + "last_node_id": 57, + "last_link_id": 0, "nodes": [ { - "id": 5, - "type": "dfe9eb32-97c0-43a5-90d5-4fd37768d91b", + "id": 57, + "type": "f2fdebf6-dfaf-43b6-9eb2-7f70613cfdc1", "pos": [ - -2.5766491043910378e-05, - 1229.999928629805 + 130, + 200 ], "size": [ 400, 470 ], "flags": {}, - "order": 0, + "order": 1, "mode": 0, "inputs": [ { @@ -44,6 +43,22 @@ }, "link": null }, + { + "name": "seed", + "type": "INT", + "widget": { + "name": "seed" + }, + "link": null + }, + { + "name": "steps", + "type": "INT", + "widget": { + "name": "steps" + }, + "link": null + }, { "name": "unet_name", "type": "COMBO", @@ -80,15 +95,15 @@ "properties": { "proxyWidgets": [ [ - "-1", + "27", "text" ], [ - "-1", + "13", "width" ], [ - "-1", + "13", "height" ], [ @@ -97,19 +112,23 @@ ], [ "3", - "control_after_generate" + "steps" ], [ - "-1", + "28", "unet_name" ], [ - "-1", + "30", "clip_name" ], [ - "-1", + "29", "vae_name" + ], + [ + "3", + "control_after_generate" ] ], "cnr_id": "comfy-core", @@ -122,48 +141,40 @@ "secondTabOffset": 80, "secondTabWidth": 65 }, - "widgets_values": [ - "", - 1024, - 1024, - null, - null, - "z_image_turbo_bf16.safetensors", - "qwen_3_4b.safetensors", - "ae.safetensors" - ] + "widgets_values": [], + "title": "Text to Image (Z-Image-Turbo)" } ], "links": [], - "groups": [], + "version": 0.4, "definitions": { "subgraphs": [ { - "id": "dfe9eb32-97c0-43a5-90d5-4fd37768d91b", + "id": "f2fdebf6-dfaf-43b6-9eb2-7f70613cfdc1", "version": 1, "state": { "lastGroupId": 4, - "lastNodeId": 34, - "lastLinkId": 40, + "lastNodeId": 61, + "lastLinkId": 75, "lastRerouteId": 0 }, "revision": 0, "config": {}, - "name": "local-Text to Image (Z-Image-Turbo)", + "name": "Text to Image (Z-Image-Turbo)", "inputNode": { "id": -10, "bounding": [ - -80, - 425, + -560, + 480, 120, - 160 + 200 ] }, "outputNode": { "id": -20, "bounding": [ - 1490, - 415, + 1670, + 320, 120, 60 ] @@ -178,8 +189,8 @@ ], "label": "prompt", "pos": [ - 20, - 445 + -460, + 500 ] }, { @@ -190,8 +201,8 @@ 35 ], "pos": [ - 20, - 465 + -460, + 520 ] }, { @@ -202,44 +213,68 @@ 36 ], "pos": [ - 20, - 485 + -460, + 540 ] }, { - "id": "23087d15-8412-4fbd-b71e-9b6d7ef76de1", + "id": "f77677f7-6bf6-4c19-a71f-c4a553d5981e", + "name": "seed", + "type": "INT", + "linkIds": [ + 71 + ], + "pos": [ + -460, + 560 + ] + }, + { + "id": "ef9a9fb1-5983-4bc9-a60b-cf5aec48bff1", + "name": "steps", + "type": "INT", + "linkIds": [ + 72 + ], + "pos": [ + -460, + 580 + ] + }, + { + "id": "a20a1b30-785f-4a04-bb6d-3d61adab9764", "name": "unet_name", "type": "COMBO", "linkIds": [ - 38 + 73 ], "pos": [ - 20, - 505 + -460, + 600 ] }, { - "id": "0677f5c3-2a3f-43d4-98ac-a4c56d5efdc0", + "id": "4af8fc2b-4655-4086-8240-45f8cb38c6f6", "name": "clip_name", "type": "COMBO", "linkIds": [ - 39 + 74 ], "pos": [ - 20, - 525 + -460, + 620 ] }, { - "id": "c85c0445-2641-48b1-bbca-95057edf2fcf", + "id": "4d518693-2807-439c-9cb6-cffd23ccba2c", "name": "vae_name", "type": "COMBO", "linkIds": [ - 40 + 75 ], "pos": [ - 20, - 545 + -460, + 640 ] } ], @@ -253,8 +288,8 @@ ], "localized_name": "IMAGE", "pos": [ - 1510, - 435 + 1690, + 340 ] } ], @@ -264,15 +299,15 @@ "id": 30, "type": "CLIPLoader", "pos": [ - 109.99997264844609, - 329.99999029608756 + 30, + 420 ], "size": [ - 269.9869791666667, - 106 + 270, + 150 ], "flags": {}, - "order": 0, + "order": 7, "mode": 0, "inputs": [ { @@ -282,7 +317,7 @@ "widget": { "name": "clip_name" }, - "link": 39 + "link": 74 }, { "localized_name": "type", @@ -315,9 +350,9 @@ } ], "properties": { + "Node name for S&R": "CLIPLoader", "cnr_id": "comfy-core", "ver": "0.3.73", - "Node name for S&R": "CLIPLoader", "models": [ { "name": "qwen_3_4b.safetensors", @@ -343,15 +378,15 @@ "id": 29, "type": "VAELoader", "pos": [ - 109.99997264844609, - 479.9999847172637 + 30, + 650 ], "size": [ - 269.9869791666667, - 58 + 270, + 110 ], "flags": {}, - "order": 1, + "order": 6, "mode": 0, "inputs": [ { @@ -361,7 +396,7 @@ "widget": { "name": "vae_name" }, - "link": 40 + "link": 75 } ], "outputs": [ @@ -375,9 +410,9 @@ } ], "properties": { + "Node name for S&R": "VAELoader", "cnr_id": "comfy-core", "ver": "0.3.73", - "Node name for S&R": "VAELoader", "models": [ { "name": "ae.safetensors", @@ -401,12 +436,12 @@ "id": 33, "type": "ConditioningZeroOut", "pos": [ - 639.9999103333332, - 620.0000271257795 + 630, + 960 ], "size": [ - 204.134765625, - 26 + 230, + 80 ], "flags": {}, "order": 8, @@ -430,9 +465,9 @@ } ], "properties": { + "Node name for S&R": "ConditioningZeroOut", "cnr_id": "comfy-core", "ver": "0.3.73", - "Node name for S&R": "ConditioningZeroOut", "enableTabs": false, "tabWidth": 65, "tabXOffset": 10, @@ -440,22 +475,21 @@ "secondTabText": "Send Back", "secondTabOffset": 80, "secondTabWidth": 65 - }, - "widgets_values": [] + } }, { "id": 8, "type": "VAEDecode", "pos": [ - 1219.9999088104782, - 160.00009184959066 + 1320, + 230 ], "size": [ - 209.98697916666669, - 46 + 230, + 100 ], "flags": {}, - "order": 5, + "order": 1, "mode": 0, "inputs": [ { @@ -483,9 +517,9 @@ } ], "properties": { + "Node name for S&R": "VAEDecode", "cnr_id": "comfy-core", "ver": "0.3.64", - "Node name for S&R": "VAEDecode", "enableTabs": false, "tabWidth": 65, "tabXOffset": 10, @@ -493,22 +527,21 @@ "secondTabText": "Send Back", "secondTabOffset": 80, "secondTabWidth": 65 - }, - "widgets_values": [] + } }, { "id": 28, "type": "UNETLoader", "pos": [ - 109.99997264844609, - 200.0000502647102 + 30, + 230 ], "size": [ - 269.9869791666667, - 82 + 270, + 110 ], "flags": {}, - "order": 2, + "order": 5, "mode": 0, "inputs": [ { @@ -518,7 +551,7 @@ "widget": { "name": "unet_name" }, - "link": 38 + "link": 73 }, { "localized_name": "weight_dtype", @@ -541,9 +574,9 @@ } ], "properties": { + "Node name for S&R": "UNETLoader", "cnr_id": "comfy-core", "ver": "0.3.73", - "Node name for S&R": "UNETLoader", "models": [ { "name": "z_image_turbo_bf16.safetensors", @@ -568,15 +601,15 @@ "id": 27, "type": "CLIPTextEncode", "pos": [ - 429.99997828947767, - 200.0000502647102 + 400, + 230 ], "size": [ - 409.9869791666667, - 319.9869791666667 + 450, + 650 ], "flags": {}, - "order": 7, + "order": 4, "mode": 0, "inputs": [ { @@ -607,9 +640,9 @@ } ], "properties": { + "Node name for S&R": "CLIPTextEncode", "cnr_id": "comfy-core", "ver": "0.3.73", - "Node name for S&R": "CLIPTextEncode", "enableTabs": false, "tabWidth": 65, "tabXOffset": 10, @@ -626,15 +659,15 @@ "id": 13, "type": "EmptySD3LatentImage", "pos": [ - 109.99997264844609, - 629.9999791384399 + 40, + 890 ], "size": [ - 259.9869791666667, - 106 + 260, + 170 ], "flags": {}, - "order": 6, + "order": 3, "mode": 0, "inputs": [ { @@ -677,9 +710,9 @@ } ], "properties": { + "Node name for S&R": "EmptySD3LatentImage", "cnr_id": "comfy-core", "ver": "0.3.64", - "Node name for S&R": "EmptySD3LatentImage", "enableTabs": false, "tabWidth": 65, "tabXOffset": 10, @@ -694,19 +727,77 @@ 1 ] }, + { + "id": 11, + "type": "ModelSamplingAuraFlow", + "pos": [ + 950, + 230 + ], + "size": [ + 310, + 110 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 26 + }, + { + "localized_name": "shift", + "name": "shift", + "type": "FLOAT", + "widget": { + "name": "shift" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "slot_index": 0, + "links": [ + 13 + ] + } + ], + "properties": { + "Node name for S&R": "ModelSamplingAuraFlow", + "cnr_id": "comfy-core", + "ver": "0.3.64", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 3 + ] + }, { "id": 3, "type": "KSampler", "pos": [ - 879.9999615530063, - 269.9999774911694 + 950, + 400 ], "size": [ - 314.9869791666667, - 262 + 320, + 350 ], "flags": {}, - "order": 4, + "order": 0, "mode": 0, "inputs": [ { @@ -740,7 +831,7 @@ "widget": { "name": "seed" }, - "link": null + "link": 71 }, { "localized_name": "steps", @@ -749,7 +840,7 @@ "widget": { "name": "steps" }, - "link": null + "link": 72 }, { "localized_name": "cfg", @@ -800,9 +891,9 @@ } ], "properties": { + "Node name for S&R": "KSampler", "cnr_id": "comfy-core", "ver": "0.3.64", - "Node name for S&R": "KSampler", "enableTabs": false, "tabWidth": 65, "tabXOffset": 10, @@ -814,81 +905,23 @@ "widgets_values": [ 0, "randomize", - 4, + 8, 1, "res_multistep", "simple", 1 ] - }, - { - "id": 11, - "type": "ModelSamplingAuraFlow", - "pos": [ - 879.9999615530063, - 160.00009184959066 - ], - "size": [ - 309.9869791666667, - 58 - ], - "flags": {}, - "order": 3, - "mode": 0, - "inputs": [ - { - "localized_name": "model", - "name": "model", - "type": "MODEL", - "link": 26 - }, - { - "localized_name": "shift", - "name": "shift", - "type": "FLOAT", - "widget": { - "name": "shift" - }, - "link": null - } - ], - "outputs": [ - { - "localized_name": "MODEL", - "name": "MODEL", - "type": "MODEL", - "slot_index": 0, - "links": [ - 13 - ] - } - ], - "properties": { - "cnr_id": "comfy-core", - "ver": "0.3.64", - "Node name for S&R": "ModelSamplingAuraFlow", - "enableTabs": false, - "tabWidth": 65, - "tabXOffset": 10, - "hasSecondTab": false, - "secondTabText": "Send Back", - "secondTabOffset": 80, - "secondTabWidth": 65 - }, - "widgets_values": [ - 3 - ] } ], "groups": [ { "id": 2, - "title": "Image size", + "title": "Step2 - Image size", "bounding": [ - 100, - 560, - 290, - 200 + 10, + 820, + 320, + 280 ], "color": "#3f789e", "font_size": 24, @@ -896,12 +929,12 @@ }, { "id": 3, - "title": "Prompt", + "title": "Step3 - Prompt", "bounding": [ - 410, + 360, 130, - 450, - 540 + 530, + 970 ], "color": "#3f789e", "font_size": 24, @@ -909,12 +942,12 @@ }, { "id": 4, - "title": "Models", + "title": "Step1 - Load models", "bounding": [ - 100, + 0, 130, - 290, - 413.6 + 330, + 660 ], "color": "#3f789e", "font_size": 24, @@ -1027,25 +1060,41 @@ "type": "INT" }, { - "id": 38, + "id": 71, "origin_id": -10, "origin_slot": 3, + "target_id": 3, + "target_slot": 4, + "type": "INT" + }, + { + "id": 72, + "origin_id": -10, + "origin_slot": 4, + "target_id": 3, + "target_slot": 5, + "type": "INT" + }, + { + "id": 73, + "origin_id": -10, + "origin_slot": 5, "target_id": 28, "target_slot": 0, "type": "COMBO" }, { - "id": 39, + "id": 74, "origin_id": -10, - "origin_slot": 4, + "origin_slot": 6, "target_id": 30, "target_slot": 0, "type": "COMBO" }, { - "id": 40, + "id": 75, "origin_id": -10, - "origin_slot": 5, + "origin_slot": 7, "target_id": 29, "target_slot": 0, "type": "COMBO" @@ -1054,25 +1103,10 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image generation and editing/Text to image" + "category": "Image generation and editing/Text to image", + "description": "Generates images from text prompts using Z-Image-Turbo, Alibaba's distilled 6B DiT model." } ] }, - "config": {}, - "extra": { - "frontendVersion": "1.37.10", - "workflowRendererVersion": "LG", - "VHS_latentpreview": false, - "VHS_latentpreviewrate": 0, - "VHS_MetadataImage": true, - "VHS_KeepIntermediate": true, - "ds": { - "scale": 0.8401370345180755, - "offset": [ - 940.0587067393087, - -830.7121087564725 - ] - } - }, - "version": 0.4 -} + "extra": {} +} \ No newline at end of file diff --git a/blueprints/Text to Image.json b/blueprints/Text to Image.json new file mode 100644 index 000000000..ffe3682ff --- /dev/null +++ b/blueprints/Text to Image.json @@ -0,0 +1,1132 @@ +{ + "revision": 0, + "last_node_id": 71, + "last_link_id": 0, + "nodes": [ + { + "id": 71, + "type": "2d5985c9-deef-41ae-9c34-6353d3d7d1ef", + "pos": [ + 90, + 800 + ], + "size": [ + 400, + 80 + ], + "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": "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 + }, + { + "name": "steps", + "type": "INT", + "widget": { + "name": "steps" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "links": [] + } + ], + "title": "Text to Image", + "properties": { + "proxyWidgets": [ + [ + "67", + "text" + ], + [ + "68", + "width" + ], + [ + "68", + "height" + ], + [ + "66", + "unet_name" + ], + [ + "62", + "clip_name" + ], + [ + "63", + "vae_name" + ], + [ + "70", + "steps" + ], + [ + "70", + "control_after_generate" + ] + ], + "cnr_id": "comfy-core", + "ver": "0.3.73", + "ue_properties": { + "widget_ue_connectable": { + "text": true + }, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [] + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "2d5985c9-deef-41ae-9c34-6353d3d7d1ef", + "version": 1, + "state": { + "lastGroupId": 4, + "lastNodeId": 71, + "lastLinkId": 70, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Text to Image", + "inputNode": { + "id": -10, + "bounding": [ + -80, + 425, + 120, + 180 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + 1490, + 415, + 120, + 60 + ] + }, + "inputs": [ + { + "id": "fb178669-e742-4a53-8a69-7df59834dfd8", + "name": "text", + "type": "STRING", + "linkIds": [ + 34 + ], + "label": "prompt", + "pos": [ + 20, + 445 + ] + }, + { + "id": "dd780b3c-23e9-46ff-8469-156008f42e5a", + "name": "width", + "type": "INT", + "linkIds": [ + 35 + ], + "pos": [ + 20, + 465 + ] + }, + { + "id": "7b08d546-6bb0-4ef9-82e9-ffae5e1ee6bc", + "name": "height", + "type": "INT", + "linkIds": [ + 36 + ], + "pos": [ + 20, + 485 + ] + }, + { + "id": "8ed4eb73-a2bf-4766-8bf4-c5890b560596", + "name": "unet_name", + "type": "COMBO", + "linkIds": [ + 38 + ], + "pos": [ + 20, + 505 + ] + }, + { + "id": "f362d639-d412-4b5d-8490-1e9995dc5f82", + "name": "clip_name", + "type": "COMBO", + "linkIds": [ + 39 + ], + "pos": [ + 20, + 525 + ] + }, + { + "id": "ee25ac16-de63-4b74-bbbb-5b29fdc1efcf", + "name": "vae_name", + "type": "COMBO", + "linkIds": [ + 40 + ], + "pos": [ + 20, + 545 + ] + }, + { + "id": "51cbcd61-9218-4bcb-89ac-ecdfb1ef8892", + "name": "steps", + "type": "INT", + "linkIds": [ + 70 + ], + "pos": [ + 20, + 565 + ] + } + ], + "outputs": [ + { + "id": "1fa72a21-ce00-4952-814e-1f2ffbe87d1d", + "name": "IMAGE", + "type": "IMAGE", + "linkIds": [ + 16 + ], + "localized_name": "IMAGE", + "pos": [ + 1510, + 435 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 62, + "type": "CLIPLoader", + "pos": [ + 110, + 330 + ], + "size": [ + 270, + 110 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [ + { + "localized_name": "clip_name", + "name": "clip_name", + "type": "COMBO", + "widget": { + "name": "clip_name" + }, + "link": 39 + }, + { + "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": [ + 28 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.73", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "CLIPLoader", + "models": [ + { + "name": "qwen_3_4b.safetensors", + "url": "https://huggingface.co/Comfy-Org/z_image_turbo/resolve/main/split_files/text_encoders/qwen_3_4b.safetensors", + "directory": "text_encoders" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "qwen_3_4b.safetensors", + "lumina2", + "default" + ] + }, + { + "id": 63, + "type": "VAELoader", + "pos": [ + 110, + 480 + ], + "size": [ + 270, + 60 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "vae_name", + "name": "vae_name", + "type": "COMBO", + "widget": { + "name": "vae_name" + }, + "link": 40 + } + ], + "outputs": [ + { + "localized_name": "VAE", + "name": "VAE", + "type": "VAE", + "links": [ + 27 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.73", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "VAELoader", + "models": [ + { + "name": "ae.safetensors", + "url": "https://huggingface.co/Comfy-Org/z_image_turbo/resolve/main/split_files/vae/ae.safetensors", + "directory": "vae" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "ae.safetensors" + ] + }, + { + "id": 64, + "type": "ConditioningZeroOut", + "pos": [ + 640, + 620 + ], + "size": [ + 210, + 30 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "localized_name": "conditioning", + "name": "conditioning", + "type": "CONDITIONING", + "link": 32 + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "links": [ + 33 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.73", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "ConditioningZeroOut", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 65, + "type": "VAEDecode", + "pos": [ + 1220, + 160 + ], + "size": [ + 210, + 50 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "samples", + "name": "samples", + "type": "LATENT", + "link": 14 + }, + { + "localized_name": "vae", + "name": "vae", + "type": "VAE", + "link": 27 + } + ], + "outputs": [ + { + "localized_name": "IMAGE", + "name": "IMAGE", + "type": "IMAGE", + "slot_index": 0, + "links": [ + 16 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.64", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "VAEDecode", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 66, + "type": "UNETLoader", + "pos": [ + 110, + 200 + ], + "size": [ + 270, + 90 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "localized_name": "unet_name", + "name": "unet_name", + "type": "COMBO", + "widget": { + "name": "unet_name" + }, + "link": 38 + }, + { + "localized_name": "weight_dtype", + "name": "weight_dtype", + "type": "COMBO", + "widget": { + "name": "weight_dtype" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "links": [ + 26 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.73", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "UNETLoader", + "models": [ + { + "name": "z_image_turbo_bf16.safetensors", + "url": "https://huggingface.co/Comfy-Org/z_image_turbo/resolve/main/split_files/diffusion_models/z_image_turbo_bf16.safetensors", + "directory": "diffusion_models" + } + ], + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "z_image_turbo_bf16.safetensors", + "default" + ] + }, + { + "id": 67, + "type": "CLIPTextEncode", + "pos": [ + 430, + 200 + ], + "size": [ + 410, + 370 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 28 + }, + { + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": 34 + } + ], + "outputs": [ + { + "localized_name": "CONDITIONING", + "name": "CONDITIONING", + "type": "CONDITIONING", + "links": [ + 30, + 32 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.73", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "CLIPTextEncode", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + "" + ] + }, + { + "id": 68, + "type": "EmptySD3LatentImage", + "pos": [ + 110, + 630 + ], + "size": [ + 260, + 110 + ], + "flags": {}, + "order": 6, + "mode": 0, + "inputs": [ + { + "localized_name": "width", + "name": "width", + "type": "INT", + "widget": { + "name": "width" + }, + "link": 35 + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "widget": { + "name": "height" + }, + "link": 36 + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "widget": { + "name": "batch_size" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "LATENT", + "name": "LATENT", + "type": "LATENT", + "slot_index": 0, + "links": [ + 17 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.64", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "EmptySD3LatentImage", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 1024, + 1024, + 1 + ] + }, + { + "id": 69, + "type": "ModelSamplingAuraFlow", + "pos": [ + 880, + 160 + ], + "size": [ + 310, + 60 + ], + "flags": {}, + "order": 7, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 26 + }, + { + "localized_name": "shift", + "name": "shift", + "type": "FLOAT", + "widget": { + "name": "shift" + }, + "link": null + } + ], + "outputs": [ + { + "localized_name": "MODEL", + "name": "MODEL", + "type": "MODEL", + "slot_index": 0, + "links": [ + 13 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.64", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "ModelSamplingAuraFlow", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 3 + ] + }, + { + "id": 70, + "type": "KSampler", + "pos": [ + 880, + 270 + ], + "size": [ + 320, + 270 + ], + "flags": {}, + "order": 8, + "mode": 0, + "inputs": [ + { + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 13 + }, + { + "localized_name": "positive", + "name": "positive", + "type": "CONDITIONING", + "link": 30 + }, + { + "localized_name": "negative", + "name": "negative", + "type": "CONDITIONING", + "link": 33 + }, + { + "localized_name": "latent_image", + "name": "latent_image", + "type": "LATENT", + "link": 17 + }, + { + "localized_name": "seed", + "name": "seed", + "type": "INT", + "widget": { + "name": "seed" + }, + "link": null + }, + { + "localized_name": "steps", + "name": "steps", + "type": "INT", + "widget": { + "name": "steps" + }, + "link": 70 + }, + { + "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": [ + 14 + ] + } + ], + "properties": { + "cnr_id": "comfy-core", + "ver": "0.3.64", + "ue_properties": { + "widget_ue_connectable": {}, + "version": "7.7", + "input_ue_unconnectable": {} + }, + "Node name for S&R": "KSampler", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + }, + "widgets_values": [ + 0, + "randomize", + 8, + 1, + "res_multistep", + "simple", + 1 + ] + } + ], + "groups": [ + { + "id": 2, + "title": "Step2 - Image size", + "bounding": [ + 100, + 560, + 290, + 200 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 3, + "title": "Step3 - Prompt", + "bounding": [ + 410, + 130, + 450, + 540 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + }, + { + "id": 4, + "title": "Step1 - Load models", + "bounding": [ + 100, + 130, + 290, + 413.6 + ], + "color": "#3f789e", + "font_size": 24, + "flags": {} + } + ], + "links": [ + { + "id": 32, + "origin_id": 67, + "origin_slot": 0, + "target_id": 64, + "target_slot": 0, + "type": "CONDITIONING" + }, + { + "id": 26, + "origin_id": 66, + "origin_slot": 0, + "target_id": 69, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 14, + "origin_id": 70, + "origin_slot": 0, + "target_id": 65, + "target_slot": 0, + "type": "LATENT" + }, + { + "id": 27, + "origin_id": 63, + "origin_slot": 0, + "target_id": 65, + "target_slot": 1, + "type": "VAE" + }, + { + "id": 13, + "origin_id": 69, + "origin_slot": 0, + "target_id": 70, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 30, + "origin_id": 67, + "origin_slot": 0, + "target_id": 70, + "target_slot": 1, + "type": "CONDITIONING" + }, + { + "id": 33, + "origin_id": 64, + "origin_slot": 0, + "target_id": 70, + "target_slot": 2, + "type": "CONDITIONING" + }, + { + "id": 17, + "origin_id": 68, + "origin_slot": 0, + "target_id": 70, + "target_slot": 3, + "type": "LATENT" + }, + { + "id": 28, + "origin_id": 62, + "origin_slot": 0, + "target_id": 67, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 16, + "origin_id": 65, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 34, + "origin_id": -10, + "origin_slot": 0, + "target_id": 67, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 35, + "origin_id": -10, + "origin_slot": 1, + "target_id": 68, + "target_slot": 0, + "type": "INT" + }, + { + "id": 36, + "origin_id": -10, + "origin_slot": 2, + "target_id": 68, + "target_slot": 1, + "type": "INT" + }, + { + "id": 38, + "origin_id": -10, + "origin_slot": 3, + "target_id": 66, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 39, + "origin_id": -10, + "origin_slot": 4, + "target_id": 62, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 40, + "origin_id": -10, + "origin_slot": 5, + "target_id": 63, + "target_slot": 0, + "type": "COMBO" + }, + { + "id": 70, + "origin_id": -10, + "origin_slot": 6, + "target_id": 70, + "target_slot": 5, + "type": "INT" + } + ], + "extra": { + "workflowRendererVersion": "LG" + }, + "category": "Image generation and editing/Text to image", + "description": "Generates images from text prompts using Z-Image-Turbo defaults with Qwen3 text encoder and VAE." + } + ] + }, + "extra": {} +} diff --git a/blueprints/Text to Video (LTX-2.3).json b/blueprints/Text to Video (LTX-2.3).json index ff9bc6ccf..f44a216dd 100644 --- a/blueprints/Text to Video (LTX-2.3).json +++ b/blueprints/Text to Video (LTX-2.3).json @@ -4286,7 +4286,8 @@ "extra": { "workflowRendererVersion": "Vue-corrected" }, - "category": "Video generation and editing/Text to video" + "category": "Video generation and editing/Text to video", + "description": "Generates video from text prompts using LTX-2.3, Lightricks' video diffusion model." } ] }, diff --git a/blueprints/Text to Video (Wan 2.2).json b/blueprints/Text to Video (Wan 2.2).json index 0ce485b67..a264a490d 100644 --- a/blueprints/Text to Video (Wan 2.2).json +++ b/blueprints/Text to Video (Wan 2.2).json @@ -1572,7 +1572,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Video generation and editing/Text to video" + "category": "Video generation and editing/Text to video", + "description": "Generates video from text prompts using Wan2.2, Alibaba's diffusion video model." } ] }, @@ -1586,4 +1587,4 @@ "VHS_KeepIntermediate": true }, "version": 0.4 -} +} \ No newline at end of file diff --git a/blueprints/Unsharp Mask.json b/blueprints/Unsharp Mask.json index 137acaa43..79a4c954f 100644 --- a/blueprints/Unsharp Mask.json +++ b/blueprints/Unsharp Mask.json @@ -434,8 +434,9 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Image Tools/Sharpen" + "category": "Image Tools/Sharpen", + "description": "Enhances edge contrast via unsharp masking for a sharper image appearance." } ] } -} +} \ No newline at end of file diff --git a/blueprints/Video Captioning (Gemini).json b/blueprints/Video Captioning (Gemini).json index ea6dc8bee..7642b23c1 100644 --- a/blueprints/Video Captioning (Gemini).json +++ b/blueprints/Video Captioning (Gemini).json @@ -307,7 +307,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Text generation/Video Captioning" + "category": "Text generation/Video Captioning", + "description": "Generates descriptive captions for video input using Google's Gemini multimodal LLM." } ] } diff --git a/blueprints/Video Inpaint(Wan2.1 VACE).json b/blueprints/Video Inpaint(Wan2.1 VACE).json index f404e6773..a658be5f8 100644 --- a/blueprints/Video Inpaint(Wan2.1 VACE).json +++ b/blueprints/Video Inpaint(Wan2.1 VACE).json @@ -165,7 +165,7 @@ }, "revision": 0, "config": {}, - "name": "local-Video Inpaint(Wan2.1 VACE)", + "name": "Video Inpaint (Wan 2.1 VACE)", "inputNode": { "id": -10, "bounding": [ @@ -2368,7 +2368,8 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Video generation and editing/Inpaint video" + "category": "Video generation and editing/Inpaint video", + "description": "Inpaints masked regions in video frames using Wan 2.1 VACE." } ] }, diff --git a/blueprints/Video Segmentation (SAM3).json b/blueprints/Video Segmentation (SAM3).json new file mode 100644 index 000000000..4d9a13412 --- /dev/null +++ b/blueprints/Video Segmentation (SAM3).json @@ -0,0 +1,827 @@ +{ + "revision": 0, + "last_node_id": 130, + "last_link_id": 0, + "nodes": [ + { + "id": 130, + "type": "7937cf78-b52b-40a3-93b2-b4e2e5f98df1", + "pos": [ + -1210, + -2780 + ], + "size": [ + 300, + 370 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "name": "video", + "type": "VIDEO", + "link": null + }, + { + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": null + }, + { + "name": "bboxes", + "type": "BOUNDING_BOX", + "link": null + }, + { + "name": "positive_coords", + "type": "STRING", + "link": null + }, + { + "name": "negative_coords", + "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": null + } + ], + "outputs": [ + { + "localized_name": "masks", + "name": "masks", + "type": "MASK", + "links": [] + }, + { + "localized_name": "bboxes", + "name": "bboxes", + "type": "BOUNDING_BOX", + "links": [] + }, + { + "name": "audio", + "type": "AUDIO", + "links": null + }, + { + "name": "fps", + "type": "FLOAT", + "links": null + } + ], + "properties": { + "proxyWidgets": [ + [ + "125", + "text" + ], + [ + "126", + "threshold" + ], + [ + "126", + "refine_iterations" + ], + [ + "126", + "individual_masks" + ], + [ + "127", + "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 + }, + "widgets_values": [], + "title": "Video Segmentation (SAM3)" + } + ], + "links": [], + "version": 0.4, + "definitions": { + "subgraphs": [ + { + "id": "7937cf78-b52b-40a3-93b2-b4e2e5f98df1", + "version": 1, + "state": { + "lastGroupId": 0, + "lastNodeId": 130, + "lastLinkId": 299, + "lastRerouteId": 0 + }, + "revision": 0, + "config": {}, + "name": "Video Segmentation (SAM3)", + "inputNode": { + "id": -10, + "bounding": [ + -2260, + -3450, + 136.369140625, + 220 + ] + }, + "outputNode": { + "id": -20, + "bounding": [ + -1050, + -3510, + 120, + 120 + ] + }, + "inputs": [ + { + "id": "680ffd88-32fe-48be-88d6-91ea44d5eaee", + "name": "video", + "type": "VIDEO", + "linkIds": [ + 252 + ], + "pos": [ + -2143.630859375, + -3430 + ] + }, + { + "id": "ceaf249c-32d7-4624-8bf6-e590e347ed90", + "name": "text", + "type": "STRING", + "linkIds": [ + 254 + ], + "pos": [ + -2143.630859375, + -3410 + ] + }, + { + "id": "1ffbff36-da0c-4854-8cb4-88ad31e64f99", + "name": "bboxes", + "type": "BOUNDING_BOX", + "linkIds": [ + 255 + ], + "pos": [ + -2143.630859375, + -3390 + ] + }, + { + "id": "67b7f4c7-cec0-4e00-b154-23cc1abf880e", + "name": "positive_coords", + "type": "STRING", + "linkIds": [ + 256 + ], + "pos": [ + -2143.630859375, + -3370 + ] + }, + { + "id": "b090a498-2bde-46b9-9554-18501401d687", + "name": "negative_coords", + "type": "STRING", + "linkIds": [ + 257 + ], + "pos": [ + -2143.630859375, + -3350 + ] + }, + { + "id": "1a76dfcf-ce95-46af-bba5-c42160c683dd", + "name": "threshold", + "type": "FLOAT", + "linkIds": [ + 261 + ], + "pos": [ + -2143.630859375, + -3330 + ] + }, + { + "id": "999523fa-c476-4c53-80c3-0a2f554d18ab", + "name": "refine_iterations", + "type": "INT", + "linkIds": [ + 262 + ], + "pos": [ + -2143.630859375, + -3310 + ] + }, + { + "id": "d2371011-7fe5-4a39-b0c1-df2e0bbd6ece", + "name": "individual_masks", + "type": "BOOLEAN", + "linkIds": [ + 263 + ], + "pos": [ + -2143.630859375, + -3290 + ] + }, + { + "id": "675a8b37-17db-48d1-853c-2fe5d6a74582", + "name": "ckpt_name", + "type": "COMBO", + "linkIds": [ + 273 + ], + "pos": [ + -2143.630859375, + -3270 + ] + } + ], + "outputs": [ + { + "id": "ff50da09-1e59-4a58-9b7f-be1a00aa5913", + "name": "masks", + "type": "MASK", + "linkIds": [ + 231 + ], + "localized_name": "masks", + "pos": [ + -1030, + -3490 + ] + }, + { + "id": "8f622e40-8528-4078-b7d3-147e9f872194", + "name": "bboxes", + "type": "BOUNDING_BOX", + "linkIds": [ + 232 + ], + "localized_name": "bboxes", + "pos": [ + -1030, + -3470 + ] + }, + { + "id": "6c9924ec-f0fa-4509-83ea-8f97f5889bcc", + "name": "audio", + "type": "AUDIO", + "linkIds": [ + 259 + ], + "pos": [ + -1030, + -3450 + ] + }, + { + "id": "82c1cddc-ab11-44eb-9e2f-1a5c7ea5645b", + "name": "fps", + "type": "FLOAT", + "linkIds": [ + 260 + ], + "pos": [ + -1030, + -3430 + ] + } + ], + "widgets": [], + "nodes": [ + { + "id": 125, + "type": "CLIPTextEncode", + "pos": [ + -2010, + -3040 + ], + "size": [ + 400, + 200 + ], + "flags": {}, + "order": 1, + "mode": 0, + "inputs": [ + { + "localized_name": "clip", + "name": "clip", + "type": "CLIP", + "link": 240 + }, + { + "localized_name": "text", + "name": "text", + "type": "STRING", + "widget": { + "name": "text" + }, + "link": 254 + } + ], + "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 + }, + "widgets_values": [ + "" + ] + }, + { + "id": 126, + "type": "SAM3_Detect", + "pos": [ + -1520, + -3520 + ], + "size": [ + 270, + 290 + ], + "flags": {}, + "order": 2, + "mode": 0, + "inputs": [ + { + "label": "model", + "localized_name": "model", + "name": "model", + "type": "MODEL", + "link": 237 + }, + { + "label": "image", + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 253 + }, + { + "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": 255 + }, + { + "label": "positive_coords", + "localized_name": "positive_coords", + "name": "positive_coords", + "shape": 7, + "type": "STRING", + "link": 256 + }, + { + "label": "negative_coords", + "localized_name": "negative_coords", + "name": "negative_coords", + "shape": 7, + "type": "STRING", + "link": 257 + }, + { + "localized_name": "threshold", + "name": "threshold", + "type": "FLOAT", + "widget": { + "name": "threshold" + }, + "link": 261 + }, + { + "localized_name": "refine_iterations", + "name": "refine_iterations", + "type": "INT", + "widget": { + "name": "refine_iterations" + }, + "link": 262 + }, + { + "localized_name": "individual_masks", + "name": "individual_masks", + "type": "BOOLEAN", + "widget": { + "name": "individual_masks" + }, + "link": 263 + } + ], + "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 + }, + "widgets_values": [ + 0.5, + 2, + false + ] + }, + { + "id": 127, + "type": "CheckpointLoaderSimple", + "pos": [ + -1970, + -3310 + ], + "size": [ + 330, + 160 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "ckpt_name", + "name": "ckpt_name", + "type": "COMBO", + "widget": { + "name": "ckpt_name" + }, + "link": 273 + } + ], + "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", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65, + "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" + } + ] + }, + "widgets_values": [ + "sam3.1_multiplex_fp16.safetensors" + ] + }, + { + "id": 128, + "type": "GetVideoComponents", + "pos": [ + -1910, + -3540 + ], + "size": [ + 230, + 120 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "localized_name": "video", + "name": "video", + "type": "VIDEO", + "link": 252 + } + ], + "outputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "links": [ + 253 + ] + }, + { + "localized_name": "audio", + "name": "audio", + "type": "AUDIO", + "links": [ + 259 + ] + }, + { + "localized_name": "fps", + "name": "fps", + "type": "FLOAT", + "links": [ + 260 + ] + } + ], + "properties": { + "Node name for S&R": "GetVideoComponents", + "cnr_id": "comfy-core", + "ver": "0.19.3", + "enableTabs": false, + "tabWidth": 65, + "tabXOffset": 10, + "hasSecondTab": false, + "secondTabText": "Send Back", + "secondTabOffset": 80, + "secondTabWidth": 65 + } + }, + { + "id": 129, + "type": "Note", + "pos": [ + -1980, + -2790 + ], + "size": [ + 370, + 250 + ], + "flags": {}, + "order": 0, + "mode": 0, + "inputs": [], + "outputs": [], + "title": "Note: Prompt format", + "properties": {}, + "widgets_values": [ + "Max tokens for this model is only 32, to separately prompt multiple subjects you can separate prompts with comma, and set the max amount of objects detected for each prompt with :N\n\nFor example above test prompt finds 2 cakes, one apron, 4 window panels" + ], + "color": "#432", + "bgcolor": "#653" + } + ], + "groups": [], + "links": [ + { + "id": 237, + "origin_id": 127, + "origin_slot": 0, + "target_id": 126, + "target_slot": 0, + "type": "MODEL" + }, + { + "id": 200, + "origin_id": 125, + "origin_slot": 0, + "target_id": 126, + "target_slot": 2, + "type": "CONDITIONING" + }, + { + "id": 240, + "origin_id": 127, + "origin_slot": 1, + "target_id": 125, + "target_slot": 0, + "type": "CLIP" + }, + { + "id": 231, + "origin_id": 126, + "origin_slot": 0, + "target_id": -20, + "target_slot": 0, + "type": "MASK" + }, + { + "id": 232, + "origin_id": 126, + "origin_slot": 1, + "target_id": -20, + "target_slot": 1, + "type": "BOUNDING_BOX" + }, + { + "id": 252, + "origin_id": -10, + "origin_slot": 0, + "target_id": 128, + "target_slot": 0, + "type": "VIDEO" + }, + { + "id": 253, + "origin_id": 128, + "origin_slot": 0, + "target_id": 126, + "target_slot": 1, + "type": "IMAGE" + }, + { + "id": 254, + "origin_id": -10, + "origin_slot": 1, + "target_id": 125, + "target_slot": 1, + "type": "STRING" + }, + { + "id": 255, + "origin_id": -10, + "origin_slot": 2, + "target_id": 126, + "target_slot": 3, + "type": "BOUNDING_BOX" + }, + { + "id": 256, + "origin_id": -10, + "origin_slot": 3, + "target_id": 126, + "target_slot": 4, + "type": "STRING" + }, + { + "id": 257, + "origin_id": -10, + "origin_slot": 4, + "target_id": 126, + "target_slot": 5, + "type": "STRING" + }, + { + "id": 259, + "origin_id": 128, + "origin_slot": 1, + "target_id": -20, + "target_slot": 2, + "type": "AUDIO" + }, + { + "id": 260, + "origin_id": 128, + "origin_slot": 2, + "target_id": -20, + "target_slot": 3, + "type": "FLOAT" + }, + { + "id": 261, + "origin_id": -10, + "origin_slot": 5, + "target_id": 126, + "target_slot": 6, + "type": "FLOAT" + }, + { + "id": 262, + "origin_id": -10, + "origin_slot": 6, + "target_id": 126, + "target_slot": 7, + "type": "INT" + }, + { + "id": 263, + "origin_id": -10, + "origin_slot": 7, + "target_id": 126, + "target_slot": 8, + "type": "BOOLEAN" + }, + { + "id": 273, + "origin_id": -10, + "origin_slot": 8, + "target_id": 127, + "target_slot": 0, + "type": "COMBO" + } + ], + "extra": {}, + "category": "Video Tools", + "description": "Segments video into temporally consistent masks using Meta SAM3 from text or interactive prompts." + } + ] + }, + "extra": {} +} diff --git a/blueprints/Video Stitch.json b/blueprints/Video Stitch.json index 020896d78..2ac78b328 100644 --- a/blueprints/Video Stitch.json +++ b/blueprints/Video Stitch.json @@ -1,21 +1,21 @@ { "revision": 0, - "last_node_id": 84, + "last_node_id": 85, "last_link_id": 0, "nodes": [ { - "id": 84, - "type": "8e8aa94a-647e-436d-8440-8ee4691864de", + "id": 85, + "type": "637913e7-0206-46ba-8ded-70ae3a7c2e19", "pos": [ - -6100, - 2620 + -880, + -2260 ], "size": [ 290, 160 ], "flags": {}, - "order": 0, + "order": 2, "mode": 0, "inputs": [ { @@ -76,31 +76,26 @@ "properties": { "proxyWidgets": [ [ - "-1", + "79", "direction" ], [ - "-1", + "79", "match_image_size" ], [ - "-1", + "79", "spacing_width" ], [ - "-1", + "79", "spacing_color" ] ], "cnr_id": "comfy-core", "ver": "0.13.0" }, - "widgets_values": [ - "right", - true, - 0, - "white" - ], + "widgets_values": [], "title": "Video Stitch" } ], @@ -109,12 +104,12 @@ "definitions": { "subgraphs": [ { - "id": "8e8aa94a-647e-436d-8440-8ee4691864de", + "id": "637913e7-0206-46ba-8ded-70ae3a7c2e19", "version": 1, "state": { "lastGroupId": 1, - "lastNodeId": 84, - "lastLinkId": 262, + "lastNodeId": 97, + "lastLinkId": 282, "lastRerouteId": 0 }, "revision": 0, @@ -123,8 +118,8 @@ "inputNode": { "id": -10, "bounding": [ - -6580, - 2649, + -6810, + 2580, 143.55859375, 160 ] @@ -132,8 +127,8 @@ "outputNode": { "id": -20, "bounding": [ - -5720, - 2659, + -4770, + 2600, 120, 60 ] @@ -149,8 +144,8 @@ "localized_name": "video", "label": "Before Video", "pos": [ - -6456.44140625, - 2669 + -6686.44140625, + 2600 ] }, { @@ -163,8 +158,8 @@ "localized_name": "video_1", "label": "After Video", "pos": [ - -6456.44140625, - 2689 + -6686.44140625, + 2620 ] }, { @@ -175,8 +170,8 @@ 259 ], "pos": [ - -6456.44140625, - 2709 + -6686.44140625, + 2640 ] }, { @@ -187,8 +182,8 @@ 260 ], "pos": [ - -6456.44140625, - 2729 + -6686.44140625, + 2660 ] }, { @@ -199,8 +194,8 @@ 261 ], "pos": [ - -6456.44140625, - 2749 + -6686.44140625, + 2680 ] }, { @@ -211,8 +206,8 @@ 262 ], "pos": [ - -6456.44140625, - 2769 + -6686.44140625, + 2700 ] } ], @@ -226,8 +221,8 @@ ], "localized_name": "VIDEO", "pos": [ - -5700, - 2679 + -4750, + 2620 ] } ], @@ -238,11 +233,11 @@ "type": "GetVideoComponents", "pos": [ -6390, - 2560 + 2600 ], "size": [ - 193.530859375, - 66 + 230, + 120 ], "flags": {}, "order": 1, @@ -278,9 +273,9 @@ } ], "properties": { + "Node name for S&R": "GetVideoComponents", "cnr_id": "comfy-core", - "ver": "0.13.0", - "Node name for S&R": "GetVideoComponents" + "ver": "0.13.0" } }, { @@ -291,8 +286,8 @@ 2420 ], "size": [ - 193.530859375, - 66 + 230, + 120 ], "flags": {}, "order": 0, @@ -332,21 +327,254 @@ } ], "properties": { + "Node name for S&R": "GetVideoComponents", "cnr_id": "comfy-core", - "ver": "0.13.0", - "Node name for S&R": "GetVideoComponents" + "ver": "0.13.0" } }, + { + "id": 90, + "type": "GetImageSize", + "pos": [ + -6390, + 3030 + ], + "size": [ + 230, + 120 + ], + "flags": {}, + "order": 4, + "mode": 0, + "inputs": [ + { + "localized_name": "image", + "name": "image", + "type": "IMAGE", + "link": 266 + } + ], + "outputs": [ + { + "localized_name": "width", + "name": "width", + "type": "INT", + "links": [ + 274 + ] + }, + { + "localized_name": "height", + "name": "height", + "type": "INT", + "links": [ + 276 + ] + }, + { + "localized_name": "batch_size", + "name": "batch_size", + "type": "INT", + "links": null + } + ], + "properties": { + "Node name for S&R": "GetImageSize" + } + }, + { + "id": 80, + "type": "CreateVideo", + "pos": [ + -5190, + 2420 + ], + "size": [ + 270, + 130 + ], + "flags": {}, + "order": 3, + "mode": 0, + "inputs": [ + { + "localized_name": "images", + "name": "images", + "type": "IMAGE", + "link": 282 + }, + { + "localized_name": "audio", + "name": "audio", + "shape": 7, + "type": "AUDIO", + "link": 251 + }, + { + "localized_name": "fps", + "name": "fps", + "type": "FLOAT", + "widget": { + "name": "fps" + }, + "link": 252 + } + ], + "outputs": [ + { + "localized_name": "VIDEO", + "name": "VIDEO", + "type": "VIDEO", + "links": [ + 255 + ] + } + ], + "properties": { + "Node name for S&R": "CreateVideo", + "cnr_id": "comfy-core", + "ver": "0.13.0" + }, + "widgets_values": [ + 30 + ] + }, + { + "id": 95, + "type": "ComfyMathExpression", + "pos": [ + -6040, + 3020 + ], + "size": [ + 400, + 200 + ], + "flags": {}, + "order": 5, + "mode": 0, + "inputs": [ + { + "label": "a", + "localized_name": "values.a", + "name": "values.a", + "type": "FLOAT,INT", + "link": 274 + }, + { + "label": "b", + "localized_name": "values.b", + "name": "values.b", + "shape": 7, + "type": "FLOAT,INT", + "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": [ + 279 + ] + } + ], + "properties": { + "Node name for S&R": "ComfyMathExpression" + }, + "widgets_values": [ + "a & ~1" + ] + }, + { + "id": 96, + "type": "ComfyMathExpression", + "pos": [ + -6040, + 3290 + ], + "size": [ + 400, + 200 + ], + "flags": {}, + "order": 6, + "mode": 0, + "inputs": [ + { + "label": "a", + "localized_name": "values.a", + "name": "values.a", + "type": "FLOAT,INT", + "link": 276 + }, + { + "label": "b", + "localized_name": "values.b", + "name": "values.b", + "shape": 7, + "type": "FLOAT,INT", + "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": [ + 280 + ] + } + ], + "properties": { + "Node name for S&R": "ComfyMathExpression" + }, + "widgets_values": [ + "a & ~1" + ] + }, { "id": 79, "type": "ImageStitch", "pos": [ -6390, - 2700 + 2780 ], "size": [ 270, - 150 + 160 ], "flags": {}, "order": 2, @@ -408,14 +636,15 @@ "name": "IMAGE", "type": "IMAGE", "links": [ - 250 + 266, + 281 ] } ], "properties": { + "Node name for S&R": "ImageStitch", "cnr_id": "comfy-core", - "ver": "0.13.0", - "Node name for S&R": "ImageStitch" + "ver": "0.13.0" }, "widgets_values": [ "right", @@ -425,60 +654,91 @@ ] }, { - "id": 80, - "type": "CreateVideo", + "id": 97, + "type": "ResizeImageMaskNode", "pos": [ - -6040, - 2610 + -5560, + 2790 ], "size": [ 270, - 78 + 160 ], "flags": {}, - "order": 3, + "order": 7, "mode": 0, "inputs": [ { - "localized_name": "images", - "name": "images", - "type": "IMAGE", - "link": 250 + "localized_name": "input", + "name": "input", + "type": "IMAGE,MASK", + "link": 281 }, { - "localized_name": "audio", - "name": "audio", - "shape": 7, - "type": "AUDIO", - "link": 251 - }, - { - "localized_name": "fps", - "name": "fps", - "type": "FLOAT", + "localized_name": "resize_type", + "name": "resize_type", + "type": "COMFY_DYNAMICCOMBO_V3", "widget": { - "name": "fps" + "name": "resize_type" }, - "link": 252 + "link": null + }, + { + "localized_name": "width", + "name": "resize_type.width", + "type": "INT", + "widget": { + "name": "resize_type.width" + }, + "link": 279 + }, + { + "localized_name": "height", + "name": "resize_type.height", + "type": "INT", + "widget": { + "name": "resize_type.height" + }, + "link": 280 + }, + { + "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": "VIDEO", - "name": "VIDEO", - "type": "VIDEO", + "localized_name": "resized", + "name": "resized", + "type": "*", "links": [ - 255 + 282 ] } ], "properties": { - "cnr_id": "comfy-core", - "ver": "0.13.0", - "Node name for S&R": "CreateVideo" + "Node name for S&R": "ResizeImageMaskNode" }, "widgets_values": [ - 30 + "scale dimensions", + 512, + 512, + "center", + "area" ] } ], @@ -500,14 +760,6 @@ "target_slot": 1, "type": "IMAGE" }, - { - "id": 250, - "origin_id": 79, - "origin_slot": 0, - "target_id": 80, - "target_slot": 0, - "type": "IMAGE" - }, { "id": 251, "origin_id": 77, @@ -579,13 +831,71 @@ "target_id": 79, "target_slot": 5, "type": "COMBO" + }, + { + "id": 266, + "origin_id": 79, + "origin_slot": 0, + "target_id": 90, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 274, + "origin_id": 90, + "origin_slot": 0, + "target_id": 95, + "target_slot": 0, + "type": "INT" + }, + { + "id": 276, + "origin_id": 90, + "origin_slot": 1, + "target_id": 96, + "target_slot": 0, + "type": "INT" + }, + { + "id": 279, + "origin_id": 95, + "origin_slot": 1, + "target_id": 97, + "target_slot": 2, + "type": "INT" + }, + { + "id": 280, + "origin_id": 96, + "origin_slot": 1, + "target_id": 97, + "target_slot": 3, + "type": "INT" + }, + { + "id": 281, + "origin_id": 79, + "origin_slot": 0, + "target_id": 97, + "target_slot": 0, + "type": "IMAGE" + }, + { + "id": 282, + "origin_id": 97, + "origin_slot": 0, + "target_id": 80, + "target_slot": 0, + "type": "IMAGE" } ], "extra": { "workflowRendererVersion": "LG" }, - "category": "Video Tools/Stitch videos" + "category": "Video Tools/Stitch videos", + "description": "Stitches multiple video clips into a single sequential video file." } ] - } -} + }, + "extra": {} +} \ No newline at end of file diff --git a/blueprints/Video Upscale(GAN x4).json b/blueprints/Video Upscale(GAN x4).json index b61dc88d7..73476e36b 100644 --- a/blueprints/Video Upscale(GAN x4).json +++ b/blueprints/Video Upscale(GAN x4).json @@ -412,9 +412,10 @@ "extra": { "workflowRendererVersion": "LG" }, - "category": "Video generation and editing/Enhance video" + "category": "Video generation and editing/Enhance video", + "description": "Upscales video to 4× resolution using a GAN-based upscaling model." } ] }, "extra": {} -} +} \ No newline at end of file diff --git a/comfy/background_removal/birefnet.json b/comfy/background_removal/birefnet.json new file mode 100644 index 000000000..f0960af39 --- /dev/null +++ b/comfy/background_removal/birefnet.json @@ -0,0 +1,7 @@ +{ + "model_type": "birefnet", + "image_std": [1.0, 1.0, 1.0], + "image_mean": [0.0, 0.0, 0.0], + "image_size": 1024, + "resize_to_original": true +} diff --git a/comfy/background_removal/birefnet.py b/comfy/background_removal/birefnet.py new file mode 100644 index 000000000..df54b2b90 --- /dev/null +++ b/comfy/background_removal/birefnet.py @@ -0,0 +1,689 @@ +import torch +import comfy.ops +import numpy as np +import torch.nn as nn +from functools import partial +import torch.nn.functional as F +from torchvision.ops import deform_conv2d +from comfy.ldm.modules.attention import optimized_attention_for_device + +CXT = [3072, 1536, 768, 384][1:][::-1][-3:] + +class Attention(nn.Module): + def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, device=None, dtype=None, operations=None): + super().__init__() + + self.dim = dim + self.num_heads = num_heads + head_dim = dim // num_heads + self.scale = qk_scale or head_dim ** -0.5 + + self.q = operations.Linear(dim, dim, bias=qkv_bias, device=device, dtype=dtype) + self.kv = operations.Linear(dim, dim * 2, bias=qkv_bias, device=device, dtype=dtype) + self.proj = operations.Linear(dim, dim, device=device, dtype=dtype) + + def forward(self, x): + B, N, C = x.shape + optimized_attention = optimized_attention_for_device(x.device, mask=False, small_input=True) + q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3) + kv = self.kv(x).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) + k, v = kv[0], kv[1] + + x = optimized_attention( + q, k, v, heads=self.num_heads, skip_output_reshape=True, skip_reshape=True + ).transpose(1, 2).reshape(B, N, C) + x = self.proj(x) + + return x + +class Mlp(nn.Module): + 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, device=device, dtype=dtype) + self.act = nn.GELU() + self.fc2 = operations.Linear(hidden_features, out_features, device=device, dtype=dtype) + + def forward(self, x): + x = self.fc1(x) + x = self.act(x) + x = self.fc2(x) + return x + + +def window_partition(x, window_size): + B, H, W, C = x.shape + x = x.view(B, H // window_size, window_size, W // window_size, window_size, C) + windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C) + return windows + + +def window_reverse(windows, window_size, H, W): + B = int(windows.shape[0] / (H * W / window_size / window_size)) + x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1) + x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1) + return x + + +class WindowAttention(nn.Module): + def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, device=None, dtype=None, operations=None): + + super().__init__() + self.dim = dim + self.window_size = window_size # Wh, Ww + self.num_heads = num_heads + head_dim = dim // num_heads + self.scale = qk_scale or head_dim ** -0.5 + + self.relative_position_bias_table = nn.Parameter( + torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads, device=device, dtype=dtype)) + + coords_h = torch.arange(self.window_size[0]) + coords_w = torch.arange(self.window_size[1]) + coords = torch.stack(torch.meshgrid([coords_h, coords_w], indexing='ij')) # 2, Wh, Ww + coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww + relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww + relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2 + relative_coords[:, :, 0] += self.window_size[0] - 1 + relative_coords[:, :, 1] += self.window_size[1] - 1 + relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1 + relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww + self.register_buffer("relative_position_index", relative_position_index) + + self.qkv = operations.Linear(dim, dim * 3, bias=qkv_bias, device=device, dtype=dtype) + self.proj = operations.Linear(dim, dim, device=device, dtype=dtype) + self.softmax = nn.Softmax(dim=-1) + + def forward(self, x, mask=None): + B_, N, C = x.shape + qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) + q, k, v = qkv[0], qkv[1], qkv[2] + + q = q * self.scale + attn = (q @ k.transpose(-2, -1)) + + 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 + attn = attn + relative_position_bias.unsqueeze(0) + + if mask is not None: + nW = mask.shape[0] + attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0) + attn = attn.view(-1, self.num_heads, N, N) + attn = self.softmax(attn) + else: + attn = self.softmax(attn) + + x = (attn @ v).transpose(1, 2).reshape(B_, N, C) + x = self.proj(x) + return x + + +class SwinTransformerBlock(nn.Module): + def __init__(self, dim, num_heads, window_size=7, shift_size=0, + mlp_ratio=4., qkv_bias=True, qk_scale=None, + norm_layer=nn.LayerNorm, device=None, dtype=None, operations=None): + super().__init__() + self.dim = dim + self.num_heads = num_heads + self.window_size = window_size + self.shift_size = shift_size + self.mlp_ratio = mlp_ratio + + self.norm1 = norm_layer(dim, device=device, dtype=dtype) + self.attn = WindowAttention( + dim, window_size=(self.window_size, self.window_size), num_heads=num_heads, + qkv_bias=qkv_bias, qk_scale=qk_scale, device=device, dtype=dtype, operations=operations) + + self.norm2 = norm_layer(dim, device=device, dtype=dtype) + mlp_hidden_dim = int(dim * mlp_ratio) + self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, device=device, dtype=dtype, operations=operations) + + self.H = None + self.W = None + + def forward(self, x, mask_matrix): + B, L, C = x.shape + H, W = self.H, self.W + + shortcut = x + x = self.norm1(x) + x = x.view(B, H, W, C) + + pad_l = pad_t = 0 + pad_r = (self.window_size - W % self.window_size) % self.window_size + pad_b = (self.window_size - H % self.window_size) % self.window_size + x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b)) + _, Hp, Wp, _ = x.shape + + if self.shift_size > 0: + shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)) + attn_mask = mask_matrix + else: + shifted_x = x + attn_mask = None + + x_windows = window_partition(shifted_x, self.window_size) + x_windows = x_windows.view(-1, self.window_size * self.window_size, C) + + attn_windows = self.attn(x_windows, mask=attn_mask) + + attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C) + shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp) # B H' W' C + + if self.shift_size > 0: + x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2)) + else: + x = shifted_x + + if pad_r > 0 or pad_b > 0: + x = x[:, :H, :W, :].contiguous() + + x = x.view(B, H * W, C) + + x = shortcut + x + x = x + self.mlp(self.norm2(x)) + + return x + + +class PatchMerging(nn.Module): + def __init__(self, dim, device=None, dtype=None, operations=None): + super().__init__() + self.dim = dim + self.reduction = operations.Linear(4 * dim, 2 * dim, bias=False, device=device, dtype=dtype) + self.norm = operations.LayerNorm(4 * dim, device=device, dtype=dtype) + + def forward(self, x, H, W): + B, L, C = x.shape + x = x.view(B, H, W, C) + + # padding + pad_input = (H % 2 == 1) or (W % 2 == 1) + if pad_input: + x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2)) + + x0 = x[:, 0::2, 0::2, :] # B H/2 W/2 C + x1 = x[:, 1::2, 0::2, :] # B H/2 W/2 C + x2 = x[:, 0::2, 1::2, :] # B H/2 W/2 C + x3 = x[:, 1::2, 1::2, :] # B H/2 W/2 C + x = torch.cat([x0, x1, x2, x3], -1) # B H/2 W/2 4*C + x = x.view(B, -1, 4 * C) # B H/2*W/2 4*C + + x = self.norm(x) + x = self.reduction(x) + + return x + + +class BasicLayer(nn.Module): + def __init__(self, + dim, + depth, + num_heads, + window_size=7, + mlp_ratio=4., + qkv_bias=True, + qk_scale=None, + norm_layer=nn.LayerNorm, + downsample=None, + device=None, dtype=None, operations=None): + super().__init__() + self.window_size = window_size + self.shift_size = window_size // 2 + self.depth = depth + + # build blocks + self.blocks = nn.ModuleList([ + SwinTransformerBlock( + dim=dim, + num_heads=num_heads, + window_size=window_size, + shift_size=0 if (i % 2 == 0) else window_size // 2, + mlp_ratio=mlp_ratio, + qkv_bias=qkv_bias, + qk_scale=qk_scale, + norm_layer=norm_layer, + device=device, dtype=dtype, operations=operations) + for i in range(depth)]) + + # patch merging layer + if downsample is not None: + self.downsample = downsample(dim=dim, device=device, dtype=dtype, operations=operations) + else: + self.downsample = None + + def forward(self, x, H, W): + Hp = int(np.ceil(H / self.window_size)) * self.window_size + Wp = int(np.ceil(W / self.window_size)) * self.window_size + img_mask = torch.zeros((1, Hp, Wp, 1), device=x.device) # 1 Hp Wp 1 + h_slices = (slice(0, -self.window_size), + slice(-self.window_size, -self.shift_size), + slice(-self.shift_size, None)) + w_slices = (slice(0, -self.window_size), + slice(-self.window_size, -self.shift_size), + slice(-self.shift_size, None)) + cnt = 0 + for h in h_slices: + for w in w_slices: + img_mask[:, h, w, :] = cnt + cnt += 1 + + mask_windows = window_partition(img_mask, self.window_size) + mask_windows = mask_windows.view(-1, self.window_size * self.window_size) + attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2) + attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0)) + + for blk in self.blocks: + blk.H, blk.W = H, W + x = blk(x, attn_mask) + if self.downsample is not None: + x_down = self.downsample(x, H, W) + Wh, Ww = (H + 1) // 2, (W + 1) // 2 + return x, H, W, x_down, Wh, Ww + else: + return x, H, W, x, H, W + + +class PatchEmbed(nn.Module): + def __init__(self, patch_size=4, in_channels=3, embed_dim=96, norm_layer=None, device=None, dtype=None, operations=None): + super().__init__() + patch_size = (patch_size, patch_size) + self.patch_size = patch_size + + self.in_channels = in_channels + self.embed_dim = embed_dim + + self.proj = operations.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=patch_size, device=device, dtype=dtype) + if norm_layer is not None: + self.norm = norm_layer(embed_dim, device=device, dtype=dtype) + else: + self.norm = None + + def forward(self, x): + _, _, H, W = x.size() + if W % self.patch_size[1] != 0: + x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1])) + if H % self.patch_size[0] != 0: + x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0])) + + x = self.proj(x) # B C Wh Ww + if self.norm is not None: + Wh, Ww = x.size(2), x.size(3) + x = x.flatten(2).transpose(1, 2) + x = self.norm(x) + x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww) + + return x + + +class SwinTransformer(nn.Module): + def __init__(self, + pretrain_img_size=224, + patch_size=4, + in_channels=3, + embed_dim=96, + depths=[2, 2, 6, 2], + num_heads=[3, 6, 12, 24], + window_size=7, + mlp_ratio=4., + qkv_bias=True, + qk_scale=None, + patch_norm=True, + out_indices=(0, 1, 2, 3), + frozen_stages=-1, + device=None, dtype=None, operations=None): + super().__init__() + + norm_layer = partial(operations.LayerNorm, device=device, dtype=dtype) + self.pretrain_img_size = pretrain_img_size + self.num_layers = len(depths) + self.embed_dim = embed_dim + self.patch_norm = patch_norm + self.out_indices = out_indices + self.frozen_stages = frozen_stages + + self.patch_embed = PatchEmbed( + patch_size=patch_size, in_channels=in_channels, embed_dim=embed_dim, + device=device, dtype=dtype, operations=operations, + norm_layer=norm_layer if self.patch_norm else None) + + self.layers = nn.ModuleList() + for i_layer in range(self.num_layers): + layer = BasicLayer( + dim=int(embed_dim * 2 ** i_layer), + depth=depths[i_layer], + num_heads=num_heads[i_layer], + window_size=window_size, + mlp_ratio=mlp_ratio, + qkv_bias=qkv_bias, + qk_scale=qk_scale, + norm_layer=norm_layer, + downsample=PatchMerging if (i_layer < self.num_layers - 1) else None, + device=device, dtype=dtype, operations=operations) + self.layers.append(layer) + + num_features = [int(embed_dim * 2 ** i) for i in range(self.num_layers)] + self.num_features = num_features + + for i_layer in out_indices: + layer = norm_layer(num_features[i_layer]) + layer_name = f'norm{i_layer}' + self.add_module(layer_name, layer) + + + def forward(self, x): + x = self.patch_embed(x) + + Wh, Ww = x.size(2), x.size(3) + + outs = [] + x = x.flatten(2).transpose(1, 2) + for i in range(self.num_layers): + layer = self.layers[i] + x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww) + + if i in self.out_indices: + norm_layer = getattr(self, f'norm{i}') + x_out = norm_layer(x_out) + + out = x_out.view(-1, H, W, self.num_features[i]).permute(0, 3, 1, 2).contiguous() + outs.append(out) + + return tuple(outs) + +class DeformableConv2d(nn.Module): + def __init__(self, + in_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1, + bias=False, device=None, dtype=None, operations=None): + + super(DeformableConv2d, self).__init__() + + kernel_size = kernel_size if type(kernel_size) is tuple else (kernel_size, kernel_size) + self.stride = stride if type(stride) is tuple else (stride, stride) + self.padding = padding + + self.offset_conv = operations.Conv2d(in_channels, + 2 * kernel_size[0] * kernel_size[1], + kernel_size=kernel_size, + stride=stride, + padding=self.padding, + bias=True, device=device, dtype=dtype) + + self.modulator_conv = operations.Conv2d(in_channels, + 1 * kernel_size[0] * kernel_size[1], + kernel_size=kernel_size, + stride=stride, + padding=self.padding, + bias=True, device=device, dtype=dtype) + + self.regular_conv = operations.Conv2d(in_channels, + out_channels=out_channels, + kernel_size=kernel_size, + stride=stride, + padding=self.padding, + bias=bias, device=device, dtype=dtype) + + def forward(self, x): + offset = self.offset_conv(x) + modulator = 2. * torch.sigmoid(self.modulator_conv(x)) + weight, bias, offload_info = comfy.ops.cast_bias_weight(self.regular_conv, x, offloadable=True) + + x = deform_conv2d( + input=x, + offset=offset, + weight=weight, + bias=None, + padding=self.padding, + mask=modulator, + stride=self.stride, + ) + comfy.ops.uncast_bias_weight(self.regular_conv, weight, bias, offload_info) + return x + +class BasicDecBlk(nn.Module): + def __init__(self, in_channels=64, out_channels=64, inter_channels=64, device=None, dtype=None, operations=None): + super(BasicDecBlk, self).__init__() + inter_channels = 64 + self.conv_in = operations.Conv2d(in_channels, inter_channels, 3, 1, padding=1, device=device, dtype=dtype) + self.relu_in = nn.ReLU(inplace=True) + self.dec_att = ASPPDeformable(in_channels=inter_channels, device=device, dtype=dtype, operations=operations) + self.conv_out = operations.Conv2d(inter_channels, out_channels, 3, 1, padding=1, device=device, dtype=dtype) + self.bn_in = operations.BatchNorm2d(inter_channels, device=device, dtype=dtype) + self.bn_out = operations.BatchNorm2d(out_channels, device=device, dtype=dtype) + + def forward(self, x): + x = self.conv_in(x) + x = self.bn_in(x) + x = self.relu_in(x) + x = self.dec_att(x) + x = self.conv_out(x) + x = self.bn_out(x) + return x + + +class BasicLatBlk(nn.Module): + def __init__(self, in_channels=64, out_channels=64, device=None, dtype=None, operations=None): + super(BasicLatBlk, self).__init__() + self.conv = operations.Conv2d(in_channels, out_channels, 1, 1, 0, device=device, dtype=dtype) + + def forward(self, x): + x = self.conv(x) + return x + + +class _ASPPModuleDeformable(nn.Module): + def __init__(self, in_channels, planes, kernel_size, padding, device, dtype, operations): + super(_ASPPModuleDeformable, self).__init__() + self.atrous_conv = DeformableConv2d(in_channels, planes, kernel_size=kernel_size, + stride=1, padding=padding, bias=False, device=device, dtype=dtype, operations=operations) + self.bn = operations.BatchNorm2d(planes, device=device, dtype=dtype) + self.relu = nn.ReLU(inplace=True) + + def forward(self, x): + x = self.atrous_conv(x) + x = self.bn(x) + + return self.relu(x) + + +class ASPPDeformable(nn.Module): + def __init__(self, in_channels, out_channels=None, parallel_block_sizes=[1, 3, 7], device=None, dtype=None, operations=None): + super(ASPPDeformable, self).__init__() + self.down_scale = 1 + if out_channels is None: + out_channels = in_channels + self.in_channelster = 256 // self.down_scale + + self.aspp1 = _ASPPModuleDeformable(in_channels, self.in_channelster, 1, padding=0, device=device, dtype=dtype, operations=operations) + self.aspp_deforms = nn.ModuleList([ + _ASPPModuleDeformable(in_channels, self.in_channelster, conv_size, padding=int(conv_size//2), device=device, dtype=dtype, operations=operations) + for conv_size in parallel_block_sizes + ]) + + self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)), + operations.Conv2d(in_channels, self.in_channelster, 1, stride=1, bias=False, device=device, dtype=dtype), + operations.BatchNorm2d(self.in_channelster, device=device, dtype=dtype), + nn.ReLU(inplace=True)) + self.conv1 = operations.Conv2d(self.in_channelster * (2 + len(self.aspp_deforms)), out_channels, 1, bias=False, device=device, dtype=dtype) + self.bn1 = operations.BatchNorm2d(out_channels, device=device, dtype=dtype) + self.relu = nn.ReLU(inplace=True) + + def forward(self, x): + x1 = self.aspp1(x) + x_aspp_deforms = [aspp_deform(x) for aspp_deform in self.aspp_deforms] + x5 = self.global_avg_pool(x) + x5 = F.interpolate(x5, size=x1.size()[2:], mode='bilinear', align_corners=True) + x = torch.cat((x1, *x_aspp_deforms, x5), dim=1) + + x = self.conv1(x) + x = self.bn1(x) + x = self.relu(x) + + return x + +class BiRefNet(nn.Module): + def __init__(self, config=None, dtype=None, device=None, operations=None): + super(BiRefNet, self).__init__() + self.bb = SwinTransformer(embed_dim=192, depths=[2, 2, 18, 2], num_heads=[6, 12, 24, 48], window_size=12, device=device, dtype=dtype, operations=operations) + + channels = [1536, 768, 384, 192] + channels = [c * 2 for c in channels] + self.cxt = channels[1:][::-1][-3:] + self.squeeze_module = nn.Sequential(*[ + BasicDecBlk(channels[0]+sum(self.cxt), channels[0], device=device, dtype=dtype, operations=operations) + for _ in range(1) + ]) + + self.decoder = Decoder(channels, device=device, dtype=dtype, operations=operations) + + def forward_enc(self, x): + x1, x2, x3, x4 = self.bb(x) + B, C, H, W = x.shape + x1_, x2_, x3_, x4_ = self.bb(F.interpolate(x, size=(H//2, W//2), mode='bilinear', align_corners=True)) + x1 = torch.cat([x1, F.interpolate(x1_, size=x1.shape[2:], mode='bilinear', align_corners=True)], dim=1) + x2 = torch.cat([x2, F.interpolate(x2_, size=x2.shape[2:], mode='bilinear', align_corners=True)], dim=1) + x3 = torch.cat([x3, F.interpolate(x3_, size=x3.shape[2:], mode='bilinear', align_corners=True)], dim=1) + x4 = torch.cat([x4, F.interpolate(x4_, size=x4.shape[2:], mode='bilinear', align_corners=True)], dim=1) + x4 = torch.cat( + ( + *[ + F.interpolate(x1, size=x4.shape[2:], mode='bilinear', align_corners=True), + F.interpolate(x2, size=x4.shape[2:], mode='bilinear', align_corners=True), + F.interpolate(x3, size=x4.shape[2:], mode='bilinear', align_corners=True), + ][-len(CXT):], + x4 + ), + dim=1 + ) + return (x1, x2, x3, x4) + + def forward_ori(self, x): + (x1, x2, x3, x4) = self.forward_enc(x) + x4 = self.squeeze_module(x4) + features = [x, x1, x2, x3, x4] + scaled_preds = self.decoder(features) + return scaled_preds + + def forward(self, pixel_values, intermediate_output=None): + scaled_preds = self.forward_ori(pixel_values) + return scaled_preds + + +class Decoder(nn.Module): + def __init__(self, channels, device, dtype, operations): + super(Decoder, self).__init__() + # factory kwargs + fk = {"device":device, "dtype":dtype, "operations":operations} + DecoderBlock = partial(BasicDecBlk, **fk) + LateralBlock = partial(BasicLatBlk, **fk) + DBlock = partial(SimpleConvs, **fk) + + self.split = True + N_dec_ipt = 64 + ic = 64 + ipt_cha_opt = 1 + self.ipt_blk5 = DBlock(2**10*3 if self.split else 3, [N_dec_ipt, channels[0]//8][ipt_cha_opt], inter_channels=ic) + self.ipt_blk4 = DBlock(2**8*3 if self.split else 3, [N_dec_ipt, channels[0]//8][ipt_cha_opt], inter_channels=ic) + self.ipt_blk3 = DBlock(2**6*3 if self.split else 3, [N_dec_ipt, channels[1]//8][ipt_cha_opt], inter_channels=ic) + self.ipt_blk2 = DBlock(2**4*3 if self.split else 3, [N_dec_ipt, channels[2]//8][ipt_cha_opt], inter_channels=ic) + self.ipt_blk1 = DBlock(2**0*3 if self.split else 3, [N_dec_ipt, channels[3]//8][ipt_cha_opt], inter_channels=ic) + + self.decoder_block4 = DecoderBlock(channels[0]+([N_dec_ipt, channels[0]//8][ipt_cha_opt]), channels[1]) + self.decoder_block3 = DecoderBlock(channels[1]+([N_dec_ipt, channels[0]//8][ipt_cha_opt]), channels[2]) + self.decoder_block2 = DecoderBlock(channels[2]+([N_dec_ipt, channels[1]//8][ipt_cha_opt]), channels[3]) + self.decoder_block1 = DecoderBlock(channels[3]+([N_dec_ipt, channels[2]//8][ipt_cha_opt]), channels[3]//2) + + fk = {"device":device, "dtype":dtype} + + self.conv_out1 = nn.Sequential(operations.Conv2d(channels[3]//2+([N_dec_ipt, channels[3]//8][ipt_cha_opt]), 1, 1, 1, 0, **fk)) + + self.lateral_block4 = LateralBlock(channels[1], channels[1]) + self.lateral_block3 = LateralBlock(channels[2], channels[2]) + self.lateral_block2 = LateralBlock(channels[3], channels[3]) + + self.conv_ms_spvn_4 = operations.Conv2d(channels[1], 1, 1, 1, 0, **fk) + self.conv_ms_spvn_3 = operations.Conv2d(channels[2], 1, 1, 1, 0, **fk) + self.conv_ms_spvn_2 = operations.Conv2d(channels[3], 1, 1, 1, 0, **fk) + + _N = 16 + + self.gdt_convs_4 = nn.Sequential(operations.Conv2d(channels[0] // 2, _N, 3, 1, 1, **fk), operations.BatchNorm2d(_N, **fk), nn.ReLU(inplace=True)) + self.gdt_convs_3 = nn.Sequential(operations.Conv2d(channels[1] // 2, _N, 3, 1, 1, **fk), operations.BatchNorm2d(_N, **fk), nn.ReLU(inplace=True)) + self.gdt_convs_2 = nn.Sequential(operations.Conv2d(channels[2] // 2, _N, 3, 1, 1, **fk), operations.BatchNorm2d(_N, **fk), nn.ReLU(inplace=True)) + + [setattr(self, f"gdt_convs_pred_{i}", nn.Sequential(operations.Conv2d(_N, 1, 1, 1, 0, **fk))) for i in range(2, 5)] + [setattr(self, f"gdt_convs_attn_{i}", nn.Sequential(operations.Conv2d(_N, 1, 1, 1, 0, **fk))) for i in range(2, 5)] + + def get_patches_batch(self, x, p): + _size_h, _size_w = p.shape[2:] + patches_batch = [] + for idx in range(x.shape[0]): + columns_x = torch.split(x[idx], split_size_or_sections=_size_w, dim=-1) + patches_x = [] + for column_x in columns_x: + patches_x += [p.unsqueeze(0) for p in torch.split(column_x, split_size_or_sections=_size_h, dim=-2)] + patch_sample = torch.cat(patches_x, dim=1) + patches_batch.append(patch_sample) + return torch.cat(patches_batch, dim=0) + + def forward(self, features): + x, x1, x2, x3, x4 = features + + patches_batch = self.get_patches_batch(x, x4) if self.split else x + x4 = torch.cat((x4, self.ipt_blk5(F.interpolate(patches_batch, size=x4.shape[2:], mode='bilinear', align_corners=True))), 1) + p4 = self.decoder_block4(x4) + p4_gdt = self.gdt_convs_4(p4) + gdt_attn_4 = self.gdt_convs_attn_4(p4_gdt).sigmoid() + p4 = p4 * gdt_attn_4 + _p4 = F.interpolate(p4, size=x3.shape[2:], mode='bilinear', align_corners=True) + _p3 = _p4 + self.lateral_block4(x3) + + patches_batch = self.get_patches_batch(x, _p3) if self.split else x + _p3 = torch.cat((_p3, self.ipt_blk4(F.interpolate(patches_batch, size=x3.shape[2:], mode='bilinear', align_corners=True))), 1) + p3 = self.decoder_block3(_p3) + + p3_gdt = self.gdt_convs_3(p3) + gdt_attn_3 = self.gdt_convs_attn_3(p3_gdt).sigmoid() + p3 = p3 * gdt_attn_3 + _p3 = F.interpolate(p3, size=x2.shape[2:], mode='bilinear', align_corners=True) + _p2 = _p3 + self.lateral_block3(x2) + + patches_batch = self.get_patches_batch(x, _p2) if self.split else x + _p2 = torch.cat((_p2, self.ipt_blk3(F.interpolate(patches_batch, size=x2.shape[2:], mode='bilinear', align_corners=True))), 1) + p2 = self.decoder_block2(_p2) + + p2_gdt = self.gdt_convs_2(p2) + gdt_attn_2 = self.gdt_convs_attn_2(p2_gdt).sigmoid() + p2 = p2 * gdt_attn_2 + + _p2 = F.interpolate(p2, size=x1.shape[2:], mode='bilinear', align_corners=True) + _p1 = _p2 + self.lateral_block2(x1) + + patches_batch = self.get_patches_batch(x, _p1) if self.split else x + _p1 = torch.cat((_p1, self.ipt_blk2(F.interpolate(patches_batch, size=x1.shape[2:], mode='bilinear', align_corners=True))), 1) + _p1 = self.decoder_block1(_p1) + _p1 = F.interpolate(_p1, size=x.shape[2:], mode='bilinear', align_corners=True) + + patches_batch = self.get_patches_batch(x, _p1) if self.split else x + _p1 = torch.cat((_p1, self.ipt_blk1(F.interpolate(patches_batch, size=x.shape[2:], mode='bilinear', align_corners=True))), 1) + p1_out = self.conv_out1(_p1) + return p1_out + + +class SimpleConvs(nn.Module): + def __init__( + self, in_channels: int, out_channels: int, inter_channels=64, device=None, dtype=None, operations=None + ) -> None: + super().__init__() + self.conv1 = operations.Conv2d(in_channels, inter_channels, 3, 1, 1, device=device, dtype=dtype) + self.conv_out = operations.Conv2d(inter_channels, out_channels, 3, 1, 1, device=device, dtype=dtype) + + def forward(self, x): + return self.conv_out(self.conv1(x)) diff --git a/comfy/bg_removal_model.py b/comfy/bg_removal_model.py new file mode 100644 index 000000000..6dec65e63 --- /dev/null +++ b/comfy/bg_removal_model.py @@ -0,0 +1,85 @@ +from .utils import load_torch_file +import os +import json +import torch +import logging + +import comfy.ops +import comfy.model_patcher +import comfy.model_management +import comfy.clip_model +import comfy.background_removal.birefnet + +BG_REMOVAL_MODELS = { + "birefnet": comfy.background_removal.birefnet.BiRefNet +} + +class BackgroundRemovalModel(): + def __init__(self, json_config): + with open(json_config) as f: + config = json.load(f) + + self.image_size = config.get("image_size", 1024) + self.image_mean = config.get("image_mean", [0.0, 0.0, 0.0]) + self.image_std = config.get("image_std", [1.0, 1.0, 1.0]) + self.model_type = config.get("model_type", "birefnet") + self.config = config.copy() + model_class = BG_REMOVAL_MODELS.get(self.model_type) + + self.load_device = comfy.model_management.text_encoder_device() + offload_device = comfy.model_management.text_encoder_offload_device() + self.dtype = comfy.model_management.text_encoder_dtype(self.load_device) + self.model = model_class(config, self.dtype, offload_device, comfy.ops.manual_cast) + self.model.eval() + + self.patcher = comfy.model_patcher.CoreModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device) + + def load_sd(self, sd): + return self.model.load_state_dict(sd, strict=False, assign=self.patcher.is_dynamic()) + + def get_sd(self): + return self.model.state_dict() + + def encode_image(self, image): + comfy.model_management.load_model_gpu(self.patcher) + H, W = image.shape[1], image.shape[2] + pixel_values = comfy.clip_model.clip_preprocess(image.to(self.load_device), size=self.image_size, mean=self.image_mean, std=self.image_std, crop=False) + + if pixel_values.shape[0] > 1: + out = torch.cat([ + self.model(pixel_values=pixel_values[i:i+1]) + for i in range(pixel_values.shape[0]) + ], dim=0) + else: + out = self.model(pixel_values=pixel_values) + 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 + + +def load_background_removal_model(sd): + if "bb.layers.1.blocks.0.attn.relative_position_index" in sd: + json_config = os.path.join(os.path.join(os.path.dirname(os.path.realpath(__file__)), "background_removal"), "birefnet.json") + else: + return None + + bg_model = BackgroundRemovalModel(json_config) + m, u = bg_model.load_sd(sd) + if len(m) > 0: + logging.warning("missing background removal: {}".format(m)) + u = set(u) + keys = list(sd.keys()) + for k in keys: + if k not in u: + sd.pop(k) + return bg_model + +def load(ckpt_path): + sd = load_torch_file(ckpt_path) + return load_background_removal_model(sd) diff --git a/comfy/cli_args.py b/comfy/cli_args.py index dbaadf723..76faed3ad 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -90,8 +90,8 @@ parser.add_argument("--force-channels-last", action="store_true", help="Force ch parser.add_argument("--directml", type=int, nargs="?", metavar="DIRECTML_DEVICE", const=-1, help="Use torch-directml.") parser.add_argument("--oneapi-device-selector", type=str, default=None, metavar="SELECTOR_STRING", help="Sets the oneAPI device(s) this instance will use.") -parser.add_argument("--disable-ipex-optimize", action="store_true", help="Disables ipex.optimize default when loading models with Intel's Extension for Pytorch.") parser.add_argument("--supports-fp8-compute", action="store_true", help="ComfyUI will act like if the device supports fp8 compute.") +parser.add_argument("--enable-triton-backend", action="store_true", help="ComfyUI will enable the use of Triton backend in comfy-kitchen. Is disabled at launch by default.") class LatentPreviewMethod(enum.Enum): NoPreviews = "none" @@ -141,8 +141,7 @@ manager_group.add_argument("--enable-manager-legacy-ui", action="store_true", he vram_group = parser.add_mutually_exclusive_group() vram_group.add_argument("--gpu-only", action="store_true", help="Store and run everything (text encoders/CLIP models, etc... on the GPU).") vram_group.add_argument("--highvram", action="store_true", help="By default models will be unloaded to CPU memory after being used. This option keeps them in GPU memory.") -vram_group.add_argument("--normalvram", action="store_true", help="Used to force normal vram use if lowvram gets automatically enabled.") -vram_group.add_argument("--lowvram", action="store_true", help="Split the unet in parts to use less vram.") +vram_group.add_argument("--lowvram", action="store_true", help="Doesn't do anything if dynamic vram is enabled. If dynamic vram isn't being used this option makes the text encoders run on the CPU.") vram_group.add_argument("--novram", action="store_true", help="When lowvram isn't enough.") vram_group.add_argument("--cpu", action="store_true", help="To use the CPU for everything (slow).") @@ -238,6 +237,8 @@ database_default_path = os.path.abspath( ) parser.add_argument("--database-url", type=str, default=f"sqlite:///{database_default_path}", help="Specify the database URL, e.g. for an in-memory database you can use 'sqlite:///:memory:'.") parser.add_argument("--enable-assets", action="store_true", help="Enable the assets system (API routes, database synchronization, and background scanning).") +parser.add_argument("--feature-flag", type=str, action='append', default=[], metavar="KEY[=VALUE]", help="Set a server feature flag. Use KEY=VALUE to set an explicit value, or bare KEY to set it to true. Can be specified multiple times. Boolean values (true/false) and numbers are auto-converted. Examples: --feature-flag show_signin_button=true or --feature-flag show_signin_button") +parser.add_argument("--list-feature-flags", action="store_true", help="Print the registry of known CLI-settable feature flags as JSON and exit.") if comfy.options.args_parsing: args = parser.parse_args() diff --git a/comfy/context_windows.py b/comfy/context_windows.py index cb44ee6e8..db57537a2 100644 --- a/comfy/context_windows.py +++ b/comfy/context_windows.py @@ -63,7 +63,11 @@ class IndexListContextWindow(ContextWindowABC): dim = self.dim if dim == 0 and full.shape[dim] == 1: return full - idx = tuple([slice(None)] * dim + [self.index_list]) + indices = self.index_list + anchor_idx = getattr(self, 'causal_anchor_index', None) + if anchor_idx is not None and anchor_idx >= 0: + indices = [anchor_idx] + list(indices) + idx = tuple([slice(None)] * dim + [indices]) window = full[idx] if retain_index_list: idx = tuple([slice(None)] * dim + [retain_index_list]) @@ -113,7 +117,14 @@ def slice_cond(cond_value, window: IndexListContextWindow, x_in: torch.Tensor, d # skip leading latent positions that have no corresponding conditioning (e.g. reference frames) if temporal_offset > 0: - indices = [i - temporal_offset for i in window.index_list[temporal_offset:]] + anchor_idx = getattr(window, 'causal_anchor_index', None) + if anchor_idx is not None and anchor_idx >= 0: + # anchor occupies one of the no-cond positions, so skip one fewer from window.index_list + skip_count = temporal_offset - 1 + else: + skip_count = temporal_offset + + indices = [i - temporal_offset for i in window.index_list[skip_count:]] indices = [i for i in indices if 0 <= i] else: indices = list(window.index_list) @@ -150,7 +161,8 @@ class ContextFuseMethod: ContextResults = collections.namedtuple("ContextResults", ['window_idx', 'sub_conds_out', 'sub_conds', 'window']) class IndexListContextHandler(ContextHandlerABC): def __init__(self, context_schedule: ContextSchedule, fuse_method: ContextFuseMethod, context_length: int=1, context_overlap: int=0, context_stride: int=1, - closed_loop: bool=False, dim:int=0, freenoise: bool=False, cond_retain_index_list: list[int]=[], split_conds_to_windows: bool=False): + closed_loop: bool=False, dim:int=0, freenoise: bool=False, cond_retain_index_list: list[int]=[], split_conds_to_windows: bool=False, + causal_window_fix: bool=True): self.context_schedule = context_schedule self.fuse_method = fuse_method self.context_length = context_length @@ -162,6 +174,7 @@ class IndexListContextHandler(ContextHandlerABC): self.freenoise = freenoise self.cond_retain_index_list = [int(x.strip()) for x in cond_retain_index_list.split(",")] if cond_retain_index_list else [] self.split_conds_to_windows = split_conds_to_windows + self.causal_window_fix = causal_window_fix self.callbacks = {} @@ -318,6 +331,14 @@ class IndexListContextHandler(ContextHandlerABC): # allow processing to end between context window executions for faster Cancel comfy.model_management.throw_exception_if_processing_interrupted() + # causal_window_fix: prepend a pre-window frame that will be stripped post-forward + anchor_applied = False + if self.causal_window_fix: + anchor_idx = window.index_list[0] - 1 + if 0 <= anchor_idx < x_in.size(self.dim): + window.causal_anchor_index = anchor_idx + anchor_applied = True + for callback in comfy.patcher_extension.get_all_callbacks(IndexListCallbacks.EVALUATE_CONTEXT_WINDOWS, self.callbacks): callback(self, model, x_in, conds, timestep, model_options, window_idx, window, model_options, device, first_device) @@ -332,6 +353,12 @@ class IndexListContextHandler(ContextHandlerABC): if device is not None: for i in range(len(sub_conds_out)): sub_conds_out[i] = sub_conds_out[i].to(x_in.device) + + # strip causal_window_fix anchor if applied + if anchor_applied: + for i in range(len(sub_conds_out)): + sub_conds_out[i] = sub_conds_out[i].narrow(self.dim, 1, sub_conds_out[i].shape[self.dim] - 1) + results.append(ContextResults(window_idx, sub_conds_out, sub_conds, window)) return results diff --git a/comfy/deploy_environment.py b/comfy/deploy_environment.py new file mode 100644 index 000000000..8c99a3584 --- /dev/null +++ b/comfy/deploy_environment.py @@ -0,0 +1,34 @@ +import functools +import logging +import os + +logger = logging.getLogger(__name__) + +_DEFAULT_DEPLOY_ENV = "local-git" +_ENV_FILENAME = ".comfy_environment" + +# Resolve the ComfyUI install directory (the parent of this `comfy/` package). +# We deliberately avoid `folder_paths.base_path` here because that is overridden +# by the `--base-directory` CLI arg to a user-supplied path, whereas the +# `.comfy_environment` marker is written by launchers/installers next to the +# ComfyUI install itself. +_COMFY_INSTALL_DIR = os.path.dirname(os.path.dirname(os.path.realpath(__file__))) + + +@functools.cache +def get_deploy_environment() -> str: + env_file = os.path.join(_COMFY_INSTALL_DIR, _ENV_FILENAME) + try: + with open(env_file, encoding="utf-8") as f: + # Cap the read so a malformed or maliciously crafted file (e.g. + # a single huge line with no newline) can't blow up memory. + first_line = f.readline(128).strip() + value = "".join(c for c in first_line if 32 <= ord(c) < 127) + if value: + return value + except FileNotFoundError: + pass + except Exception as e: + logger.error("Failed to read %s: %s", env_file, e) + + return _DEFAULT_DEPLOY_ENV diff --git a/comfy/hooks.py b/comfy/hooks.py index 1a76c7ba4..5458fc3d8 100644 --- a/comfy/hooks.py +++ b/comfy/hooks.py @@ -93,7 +93,7 @@ class Hook: self.hook_scope = hook_scope '''Scope of where this hook should apply in terms of the conds used in sampling run.''' self.custom_should_register = default_should_register - '''Can be overriden with a compatible function to decide if this hook should be registered without the need to override .should_register''' + '''Can be overridden with a compatible function to decide if this hook should be registered without the need to override .should_register''' @property def strength(self): diff --git a/comfy/image_encoders/dino2.py b/comfy/image_encoders/dino2.py index 9b6dace9d..ee86f8309 100644 --- a/comfy/image_encoders/dino2.py +++ b/comfy/image_encoders/dino2.py @@ -106,6 +106,7 @@ class Dino2Encoder(torch.nn.Module): class Dino2PatchEmbeddings(torch.nn.Module): def __init__(self, dim, num_channels=3, patch_size=14, image_size=518, dtype=None, device=None, operations=None): super().__init__() + self.patch_size = patch_size self.projection = operations.Conv2d( in_channels=num_channels, out_channels=dim, @@ -125,17 +126,37 @@ class Dino2Embeddings(torch.nn.Module): super().__init__() patch_size = 14 image_size = 518 + self.patch_size = patch_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)) + 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)) + 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) + + class_pos = pos_embed[:, 0:1] + patch_pos = pos_embed[:, 1:] + N = patch_pos.shape[1] + M = int(N ** 0.5) + 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). + + 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) + patch_pos = patch_pos.permute(0, 2, 3, 1).flatten(1, 2) + return torch.cat((class_pos, patch_pos), dim=1).to(x.dtype) + def forward(self, pixel_values): x = self.patch_embeddings(pixel_values) - # TODO: mask_token? x = torch.cat((self.cls_token.to(device=x.device, dtype=x.dtype).expand(x.shape[0], -1, -1), x), dim=1) - x = x + comfy.model_management.cast_to_device(self.position_embeddings, x.device, x.dtype) + if x.shape[1] - 1 == self.position_embeddings.shape[1] - 1: + x = x + comfy.model_management.cast_to_device(self.position_embeddings, x.device, x.dtype) + else: + h, w = pixel_values.shape[-2:] + x = x + self.interpolate_pos_encoding(x, h, w) return x @@ -158,3 +179,21 @@ class Dinov2Model(torch.nn.Module): x = self.layernorm(x) pooled_output = x[:, 0, :] return x, i, pooled_output, None + + def get_intermediate_layers(self, pixel_values, indices, apply_norm=True): + x = self.embeddings(pixel_values) + optimized_attention = optimized_attention_for_device(x.device, False, small_input=True) + n_layers = len(self.encoder.layer) + resolved = [(i if i >= 0 else n_layers + i) for i in indices] + target = set(resolved) + max_idx = max(resolved) + n_skip = 1 # skip cls token + cache = {} + for i, layer in enumerate(self.encoder.layer): + x = layer(x, optimized_attention) + if i in target: + normed = self.layernorm(x) if apply_norm else x + cache[i] = (normed[:, n_skip:], normed[:, 0]) + if i >= max_idx: + break + return [cache[i] for i in resolved] diff --git a/comfy/k_diffusion/sampling.py b/comfy/k_diffusion/sampling.py index 6978eb717..11db46d94 100644 --- a/comfy/k_diffusion/sampling.py +++ b/comfy/k_diffusion/sampling.py @@ -242,6 +242,7 @@ def sample_euler_ancestral_RF(model, x, sigmas, extra_args=None, callback=None, extra_args = {} if extra_args is None else extra_args seed = extra_args.get("seed", None) noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler + s_noise = s_noise * getattr(model.inner_model.model_patcher.get_model_object('model_sampling'), "noise_scale", 1.0) s_in = x.new_ones([x.shape[0]]) for i in trange(len(sigmas) - 1, disable=disable): denoised = model(x, sigmas[i] * s_in, **extra_args) @@ -373,6 +374,7 @@ def sample_dpm_2_ancestral_RF(model, x, sigmas, extra_args=None, callback=None, extra_args = {} if extra_args is None else extra_args seed = extra_args.get("seed", None) noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler + s_noise = s_noise * getattr(model.inner_model.model_patcher.get_model_object('model_sampling'), "noise_scale", 1.0) s_in = x.new_ones([x.shape[0]]) for i in trange(len(sigmas) - 1, disable=disable): denoised = model(x, sigmas[i] * s_in, **extra_args) @@ -686,6 +688,7 @@ def sample_dpmpp_2s_ancestral_RF(model, x, sigmas, extra_args=None, callback=Non extra_args = {} if extra_args is None else extra_args seed = extra_args.get("seed", None) noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler + s_noise = s_noise * getattr(model.inner_model.model_patcher.get_model_object('model_sampling'), "noise_scale", 1.0) s_in = x.new_ones([x.shape[0]]) sigma_fn = lambda lbda: (lbda.exp() + 1) ** -1 lambda_fn = lambda sigma: ((1-sigma)/sigma).log() @@ -747,6 +750,7 @@ def sample_dpmpp_sde(model, x, sigmas, extra_args=None, callback=None, disable=N sigma_fn = partial(half_log_snr_to_sigma, model_sampling=model_sampling) lambda_fn = partial(sigma_to_half_log_snr, model_sampling=model_sampling) sigmas = offset_first_sigma_for_snr(sigmas, model_sampling) + s_noise = s_noise * getattr(model_sampling, "noise_scale", 1.0) for i in trange(len(sigmas) - 1, disable=disable): denoised = model(x, sigmas[i] * s_in, **extra_args) @@ -832,6 +836,7 @@ def sample_dpmpp_2m_sde(model, x, sigmas, extra_args=None, callback=None, disabl model_sampling = model.inner_model.model_patcher.get_model_object('model_sampling') lambda_fn = partial(sigma_to_half_log_snr, model_sampling=model_sampling) sigmas = offset_first_sigma_for_snr(sigmas, model_sampling) + s_noise = s_noise * getattr(model_sampling, "noise_scale", 1.0) old_denoised = None h, h_last = None, None @@ -889,6 +894,7 @@ def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disabl model_sampling = model.inner_model.model_patcher.get_model_object('model_sampling') lambda_fn = partial(sigma_to_half_log_snr, model_sampling=model_sampling) sigmas = offset_first_sigma_for_snr(sigmas, model_sampling) + s_noise = s_noise * getattr(model_sampling, "noise_scale", 1.0) denoised_1, denoised_2 = None, None h, h_1, h_2 = None, None, None @@ -1006,23 +1012,39 @@ def sample_ddpm(model, x, sigmas, extra_args=None, callback=None, disable=None, return generic_step_sampler(model, x, sigmas, extra_args, callback, disable, noise_sampler, DDPMSampler_step) @torch.no_grad() -def sample_lcm(model, x, sigmas, extra_args=None, callback=None, disable=None, noise_sampler=None): +def sample_lcm(model, x, sigmas, extra_args=None, callback=None, disable=None, noise_sampler=None, s_noise=1.0, s_noise_end=None, noise_clip_std=0.0): + + # s_noise / s_noise_end: per-step noise multiplier, linearly interpolated across steps + # noise_clip_std: clamp injected noise to +/- N stddevs (0 disables). + extra_args = {} if extra_args is None else extra_args seed = extra_args.get("seed", None) noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler s_in = x.new_ones([x.shape[0]]) - for i in trange(len(sigmas) - 1, disable=disable): + n_steps = max(1, len(sigmas) - 1) + model_sampling = model.inner_model.model_patcher.get_model_object('model_sampling') + + s_start = float(s_noise) + s_end = s_start if s_noise_end is None else float(s_noise_end) + for i in trange(n_steps, disable=disable): denoised = model(x, sigmas[i] * s_in, **extra_args) if callback is not None: callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised}) x = denoised if sigmas[i + 1] > 0: - x = model.inner_model.inner_model.model_sampling.noise_scaling(sigmas[i + 1], noise_sampler(sigmas[i], sigmas[i + 1]), x) + noise = noise_sampler(sigmas[i], sigmas[i + 1]) + if noise_clip_std > 0: + clip_val = noise_clip_std * noise.std() + noise = noise.clamp(min=-clip_val, max=clip_val) + t = (i / (n_steps - 1)) if n_steps > 1 else 0.0 + s_noise_i = s_start + (s_end - s_start) * t + if s_noise_i != 1.0: + noise = noise * s_noise_i + x = model_sampling.noise_scaling(sigmas[i + 1], noise, x) return x - @torch.no_grad() def sample_heunpp2(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1.): # From MIT licensed: https://github.com/Carzit/sd-webui-samplers-scheduler/ @@ -1249,6 +1271,7 @@ def sample_euler_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback=No model_sampling = model.inner_model.model_patcher.get_model_object("model_sampling") lambda_fn = partial(sigma_to_half_log_snr, model_sampling=model_sampling) + s_noise = s_noise * getattr(model_sampling, "noise_scale", 1.0) uncond_denoised = None @@ -1296,6 +1319,7 @@ def sample_dpmpp_2s_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback extra_args = {} if extra_args is None else extra_args seed = extra_args.get("seed", None) noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler + s_noise = s_noise * getattr(model.inner_model.model_patcher.get_model_object('model_sampling'), "noise_scale", 1.0) temp = [0] def post_cfg_function(args): @@ -1371,6 +1395,7 @@ def res_multistep(model, x, sigmas, extra_args=None, callback=None, disable=None extra_args = {} if extra_args is None else extra_args seed = extra_args.get("seed", None) noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler + s_noise = s_noise * getattr(model.inner_model.model_patcher.get_model_object('model_sampling'), "noise_scale", 1.0) s_in = x.new_ones([x.shape[0]]) sigma_fn = lambda t: t.neg().exp() t_fn = lambda sigma: sigma.log().neg() @@ -1504,6 +1529,7 @@ def sample_er_sde(model, x, sigmas, extra_args=None, callback=None, disable=None extra_args = {} if extra_args is None else extra_args seed = extra_args.get("seed", None) noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler + s_noise = s_noise * getattr(model.inner_model.model_patcher.get_model_object('model_sampling'), "noise_scale", 1.0) s_in = x.new_ones([x.shape[0]]) def default_er_sde_noise_scaler(x): @@ -1574,9 +1600,10 @@ def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=Non seed = extra_args.get("seed", None) noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler s_in = x.new_ones([x.shape[0]]) - inject_noise = eta > 0 and s_noise > 0 model_sampling = model.inner_model.model_patcher.get_model_object('model_sampling') + s_noise = s_noise * getattr(model_sampling, "noise_scale", 1.0) + inject_noise = eta > 0 and s_noise > 0 sigma_fn = partial(half_log_snr_to_sigma, model_sampling=model_sampling) lambda_fn = partial(sigma_to_half_log_snr, model_sampling=model_sampling) sigmas = offset_first_sigma_for_snr(sigmas, model_sampling) @@ -1645,9 +1672,10 @@ def sample_seeds_3(model, x, sigmas, extra_args=None, callback=None, disable=Non seed = extra_args.get("seed", None) noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler s_in = x.new_ones([x.shape[0]]) - inject_noise = eta > 0 and s_noise > 0 model_sampling = model.inner_model.model_patcher.get_model_object('model_sampling') + s_noise = s_noise * getattr(model_sampling, "noise_scale", 1.0) + inject_noise = eta > 0 and s_noise > 0 sigma_fn = partial(half_log_snr_to_sigma, model_sampling=model_sampling) lambda_fn = partial(sigma_to_half_log_snr, model_sampling=model_sampling) sigmas = offset_first_sigma_for_snr(sigmas, model_sampling) @@ -1713,6 +1741,7 @@ def sample_sa_solver(model, x, sigmas, extra_args=None, callback=None, disable=F s_in = x.new_ones([x.shape[0]]) model_sampling = model.inner_model.model_patcher.get_model_object("model_sampling") + s_noise = s_noise * getattr(model_sampling, "noise_scale", 1.0) sigmas = offset_first_sigma_for_snr(sigmas, model_sampling) lambdas = sigma_to_half_log_snr(sigmas, model_sampling=model_sampling) @@ -1810,3 +1839,119 @@ def sample_sa_solver(model, x, sigmas, extra_args=None, callback=None, disable=F def sample_sa_solver_pece(model, x, sigmas, extra_args=None, callback=None, disable=False, tau_func=None, s_noise=1.0, noise_sampler=None, predictor_order=3, corrector_order=4, simple_order_2=False): """Stochastic Adams Solver with PECE (Predict–Evaluate–Correct–Evaluate) mode (NeurIPS 2023).""" return sample_sa_solver(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, tau_func=tau_func, s_noise=s_noise, noise_sampler=noise_sampler, predictor_order=predictor_order, corrector_order=corrector_order, use_pece=True, simple_order_2=simple_order_2) + + +@torch.no_grad() +def sample_ar_video(model, x, sigmas, extra_args=None, callback=None, disable=None, + num_frame_per_block=1): + """ + Autoregressive video sampler: block-by-block denoising with KV cache + and flow-match re-noising for Causal Forcing / Self-Forcing models. + + Requires a Causal-WAN compatible model (diffusion_model must expose + init_kv_caches / init_crossattn_caches) and 5-D latents [B,C,T,H,W]. + + All AR-loop parameters are passed via the SamplerARVideo node, not read + from the checkpoint or transformer_options. + """ + extra_args = {} if extra_args is None else extra_args + model_options = extra_args.get("model_options", {}) + transformer_options = model_options.get("transformer_options", {}) + + if x.ndim != 5: + raise ValueError( + f"ar_video sampler requires 5-D video latents [B,C,T,H,W], got {x.ndim}-D tensor with shape {x.shape}. " + "This sampler is only compatible with autoregressive video models (e.g. Causal-WAN)." + ) + + inner_model = model.inner_model.inner_model + causal_model = inner_model.diffusion_model + + if not (hasattr(causal_model, "init_kv_caches") and hasattr(causal_model, "init_crossattn_caches")): + raise TypeError( + "ar_video sampler requires a Causal-WAN compatible model whose diffusion_model " + "exposes init_kv_caches() and init_crossattn_caches(). The loaded checkpoint " + "does not support this interface — choose a different sampler." + ) + + seed = extra_args.get("seed", 0) + + bs, c, lat_t, lat_h, lat_w = x.shape + frame_seq_len = -(-lat_h // 2) * -(-lat_w // 2) # ceiling division + num_blocks = -(-lat_t // num_frame_per_block) # ceiling division + device = x.device + model_dtype = inner_model.get_dtype() + + kv_caches = causal_model.init_kv_caches(bs, lat_t * frame_seq_len, device, model_dtype) + crossattn_caches = causal_model.init_crossattn_caches(bs, device, model_dtype) + + output = torch.zeros_like(x) + s_in = x.new_ones([x.shape[0]]) + current_start_frame = 0 + + # I2V: seed KV cache with the initial image latent before the denoising loop + initial_latent = transformer_options.get("ar_config", {}).get("initial_latent", None) + if initial_latent is not None: + initial_latent = inner_model.process_latent_in(initial_latent).to(device=device, dtype=model_dtype) + n_init = initial_latent.shape[2] + output[:, :, :n_init] = initial_latent + + ar_state = {"start_frame": 0, "kv_caches": kv_caches, "crossattn_caches": crossattn_caches} + transformer_options["ar_state"] = ar_state + zero_sigma = sigmas.new_zeros([1]) + _ = model(initial_latent, zero_sigma * s_in, **extra_args) + + current_start_frame = n_init + remaining = lat_t - n_init + num_blocks = -(-remaining // num_frame_per_block) + + num_sigma_steps = len(sigmas) - 1 + total_real_steps = num_blocks * num_sigma_steps + step_count = 0 + + try: + for block_idx in trange(num_blocks, disable=disable): + bf = min(num_frame_per_block, lat_t - current_start_frame) + fs, fe = current_start_frame, current_start_frame + bf + noisy_input = x[:, :, fs:fe] + + ar_state = { + "start_frame": current_start_frame, + "kv_caches": kv_caches, + "crossattn_caches": crossattn_caches, + } + transformer_options["ar_state"] = ar_state + + for i in range(num_sigma_steps): + denoised = model(noisy_input, sigmas[i] * s_in, **extra_args) + + if callback is not None: + scaled_i = step_count * num_sigma_steps // total_real_steps + callback({"x": noisy_input, "i": scaled_i, "sigma": sigmas[i], + "sigma_hat": sigmas[i], "denoised": denoised}) + + if sigmas[i + 1] == 0: + noisy_input = denoised + else: + sigma_next = sigmas[i + 1] + torch.manual_seed(seed + block_idx * 1000 + i) + fresh_noise = torch.randn_like(denoised) + noisy_input = (1.0 - sigma_next) * denoised + sigma_next * fresh_noise + + for cache in kv_caches: + cache["end"] -= bf * frame_seq_len + + step_count += 1 + + output[:, :, fs:fe] = noisy_input + + for cache in kv_caches: + cache["end"] -= bf * frame_seq_len + zero_sigma = sigmas.new_zeros([1]) + _ = model(noisy_input, zero_sigma * s_in, **extra_args) + + current_start_frame += bf + finally: + transformer_options.pop("ar_state", None) + + return output diff --git a/comfy/latent_formats.py b/comfy/latent_formats.py index 3dac5be18..6e37080bb 100644 --- a/comfy/latent_formats.py +++ b/comfy/latent_formats.py @@ -9,6 +9,7 @@ class LatentFormat: latent_rgb_factors_reshape = None taesd_decoder_name = None spacial_downscale_ratio = 8 + temporal_downscale_ratio = 1 def process_in(self, latent): return latent * self.scale_factor @@ -149,6 +150,7 @@ class SD3(LatentFormat): class StableAudio1(LatentFormat): latent_channels = 64 latent_dimensions = 1 + temporal_downscale_ratio = 2048 class Flux(SD3): latent_channels = 16 @@ -235,6 +237,7 @@ class Flux2(LatentFormat): class Mochi(LatentFormat): latent_channels = 12 latent_dimensions = 3 + temporal_downscale_ratio = 6 def __init__(self): self.scale_factor = 1.0 @@ -278,6 +281,7 @@ class LTXV(LatentFormat): latent_channels = 128 latent_dimensions = 3 spacial_downscale_ratio = 32 + temporal_downscale_ratio = 8 def __init__(self): self.latent_rgb_factors = [ @@ -421,6 +425,7 @@ class LTXAV(LTXV): class HunyuanVideo(LatentFormat): latent_channels = 16 latent_dimensions = 3 + temporal_downscale_ratio = 4 scale_factor = 0.476986 latent_rgb_factors = [ [-0.0395, -0.0331, 0.0445], @@ -447,6 +452,7 @@ class HunyuanVideo(LatentFormat): class Cosmos1CV8x8x8(LatentFormat): latent_channels = 16 latent_dimensions = 3 + temporal_downscale_ratio = 8 latent_rgb_factors = [ [ 0.1817, 0.2284, 0.2423], @@ -472,6 +478,7 @@ class Cosmos1CV8x8x8(LatentFormat): class Wan21(LatentFormat): latent_channels = 16 latent_dimensions = 3 + temporal_downscale_ratio = 4 latent_rgb_factors = [ [-0.1299, -0.1692, 0.2932], @@ -734,6 +741,7 @@ class HunyuanVideo15(LatentFormat): latent_channels = 32 latent_dimensions = 3 spacial_downscale_ratio = 16 + temporal_downscale_ratio = 4 scale_factor = 1.03682 taesd_decoder_name = "lighttaehy1_5" @@ -759,6 +767,7 @@ class ACEAudio(LatentFormat): class ACEAudio15(LatentFormat): latent_channels = 64 latent_dimensions = 1 + temporal_downscale_ratio = 1764 class ChromaRadiance(LatentFormat): latent_channels = 3 @@ -785,9 +794,35 @@ 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 CogVideoX(LatentFormat): + """Latent format for CogVideoX-2b (THUDM/CogVideoX-2b). + + scale_factor matches the vae/config.json scaling_factor for the 2b variant. + The 5b-class checkpoints (CogVideoX-5b, CogVideoX-1.5-5B, CogVideoX-Fun-V1.5-*) + use a different value; see CogVideoX1_5 below. + """ latent_channels = 16 latent_dimensions = 3 + temporal_downscale_ratio = 4 def __init__(self): self.scale_factor = 1.15258426 + + +class CogVideoX1_5(CogVideoX): + """Latent format for 5b-class CogVideoX checkpoints. + + Covers THUDM/CogVideoX-5b, THUDM/CogVideoX-1.5-5B, and the CogVideoX-Fun + V1.5-5b family (including VOID inpainting). All of these have + scaling_factor=0.7 in their vae/config.json. Auto-selected in + supported_models.CogVideoX_T2V based on transformer hidden dim. + """ + def __init__(self): + self.scale_factor = 0.7 diff --git a/comfy/ldm/hidream_o1/attention.py b/comfy/ldm/hidream_o1/attention.py new file mode 100644 index 000000000..1b68f1771 --- /dev/null +++ b/comfy/ldm/hidream_o1/attention.py @@ -0,0 +1,41 @@ +"""HiDream-O1 two-pass attention: tokens [0, ar_len) are causal, [ar_len, T) +attend full K/V. Splitting Q at the boundary avoids the (B, 1, T, T) additive +mask the general-purpose path would build (~500 MB at T~16K) and lets the +gen half hit the user's preferred backend via optimized_attention. +""" + +import torch + +import comfy.ops +from comfy.ldm.modules.attention import optimized_attention + + +def make_two_pass_attention(ar_len: int, transformer_options=None): + """Build a two-pass attention callable. AR pass uses SDPA-causal directly, gen pass routes through optimized_attention. + The AR pass goes through SDPA directand bypasses wrappers, it is only ~1% of T at typical edit sizes. + """ + + def two_pass_attention(q, k, v, heads, **kwargs): + B, H, T, D = q.shape + + if T < k.shape[2]: # KV-cache hot path: Q is shorter than K/V (cached AR prefix is in K/V only), all fresh Q positions are in the gen region, single full-attention call + out = optimized_attention(q, k, v, heads, mask=None, skip_reshape=True, skip_output_reshape=True, transformer_options=transformer_options) + elif ar_len >= T: + out = comfy.ops.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=True) + elif ar_len <= 0: + out = optimized_attention(q, k, v, heads, mask=None, skip_reshape=True, skip_output_reshape=True, transformer_options=transformer_options) + else: + out_ar = comfy.ops.scaled_dot_product_attention( + q[:, :, :ar_len], k[:, :, :ar_len], v[:, :, :ar_len], + attn_mask=None, dropout_p=0.0, is_causal=True, + ) + out_gen = optimized_attention( + q[:, :, ar_len:], k, v, heads, + mask=None, skip_reshape=True, skip_output_reshape=True, + transformer_options=transformer_options, + ) + out = torch.cat([out_ar, out_gen], dim=2) + + return out.transpose(1, 2).reshape(B, T, H * D) + + return two_pass_attention diff --git a/comfy/ldm/hidream_o1/conditioning.py b/comfy/ldm/hidream_o1/conditioning.py new file mode 100644 index 000000000..7496f0035 --- /dev/null +++ b/comfy/ldm/hidream_o1/conditioning.py @@ -0,0 +1,230 @@ +"""HiDream-O1 conditioning prep — ref-image dual path + extra_conds assembly. + +Each ref image goes through two paths: a 32x32 patchified stream concatenated +to the noised target, and a Qwen3-VL ViT path producing tokens that scatter +into input_ids at <|image_pad|> positions. +""" + +from typing import List + +import torch + +import comfy.utils +from comfy.text_encoders.qwen_vl import process_qwen2vl_images + +from .utils import (PATCH_SIZE, calculate_dimensions, cond_image_size, ref_max_size, resize_tensor) + +# Qwen3-VL ViT preprocessing constants (preprocessor_config.json). +VIT_PATCH = 16 +VIT_MERGE = 2 +VIT_IMAGE_MEAN = [0.5, 0.5, 0.5] +VIT_IMAGE_STD = [0.5, 0.5, 0.5] + + +def prepare_ref_images( + ref_images: List[torch.Tensor], + target_h: int, + target_w: int, + device: torch.device, + dtype: torch.dtype, +): + """Build the dual-path tensors for K reference images at (target_h, target_w). + + Returns None for K=0, else a dict with ref_patches, ref_pixel_values, + ref_image_grid_thw, per_ref_vit_tokens, per_ref_patch_grids. + """ + K = len(ref_images) + if K == 0: + return None + max_size = ref_max_size(max(target_h, target_w), K) + cis = cond_image_size(K) + + refs_t = [img[0].clamp(0, 1).permute(2, 0, 1).unsqueeze(0).contiguous().float() for img in ref_images] + refs_t = [resize_tensor(t, max_size, PATCH_SIZE) for t in refs_t] + + # 32-patch path. + ref_patches_per = [] + per_ref_patch_grids = [] + for t in refs_t: + t_norm = (t.squeeze(0) - 0.5) / 0.5 # (3, H, W) in [-1, 1] + h_p, w_p = t_norm.shape[-2] // PATCH_SIZE, t_norm.shape[-1] // PATCH_SIZE + per_ref_patch_grids.append((h_p, w_p)) + patches = ( + t_norm.reshape(3, h_p, PATCH_SIZE, w_p, PATCH_SIZE) + .permute(1, 3, 0, 2, 4) + .reshape(h_p * w_p, 3 * PATCH_SIZE * PATCH_SIZE) + ) + ref_patches_per.append(patches) + ref_patches = torch.cat(ref_patches_per, dim=0).unsqueeze(0).to(device=device, dtype=dtype) + + # ViT path. + refs_vlm_t = [] + for t in refs_t: + _, _, h, w = t.shape + cond_w, cond_h = calculate_dimensions(cis, w / h) + cond_w = max(cond_w, VIT_PATCH * VIT_MERGE) + cond_h = max(cond_h, VIT_PATCH * VIT_MERGE) + refs_vlm_t.append(comfy.utils.common_upscale(t, cond_w, cond_h, "lanczos", "disabled")) + + pv_list, grid_list, per_ref_vit_tokens = [], [], [] + for t_v in refs_vlm_t: + pv, grid_thw = process_qwen2vl_images( + t_v.permute(0, 2, 3, 1), + min_pixels=0, max_pixels=10**12, + patch_size=VIT_PATCH, merge_size=VIT_MERGE, + image_mean=VIT_IMAGE_MEAN, image_std=VIT_IMAGE_STD, + ) + grid_thw = grid_thw[0] + pv_list.append(pv.to(device=device, dtype=dtype)) + grid_list.append(grid_thw.to(device=device)) + # Post-merge token count = number of <|image_pad|> tokens this image expands to in input_ids. + gh, gw = int(grid_thw[1].item()), int(grid_thw[2].item()) + per_ref_vit_tokens.append((gh // VIT_MERGE) * (gw // VIT_MERGE)) + + return { + "ref_patches": ref_patches, + "ref_pixel_values": torch.cat(pv_list, dim=0), + "ref_image_grid_thw": torch.stack(grid_list, dim=0), + "per_ref_vit_tokens": per_ref_vit_tokens, + "per_ref_patch_grids": per_ref_patch_grids, + } + + +def build_ref_input_ids( + text_input_ids: torch.Tensor, + per_ref_vit_tokens: List[int], + image_token_id: int, + vision_start_id: int, + vision_end_id: int, +): + """Splice [vision_start, image_pad*N, vision_end] blocks into input_ids + after the [im_start, user, \\n] prefix (matches original chat template). + """ + ids = text_input_ids[0].tolist() + inserted = [] + for n_pad in per_ref_vit_tokens: + inserted.extend([vision_start_id] + [image_token_id] * n_pad + [vision_end_id]) + new_ids = ids[:3] + inserted + ids[3:] # 3 = len([im_start, user, \n]) + return torch.tensor([new_ids], dtype=text_input_ids.dtype, device=text_input_ids.device) + + +def build_extra_conds( + text_input_ids: torch.Tensor, + noise: torch.Tensor, + ref_images: List[torch.Tensor] = None, + target_patch_size: int = 32, +): + """Assemble all conditioning tensors for HiDreamO1Transformer.forward: + input_ids (with ref-vision tokens spliced in for the edit/IP path), + position_ids (MRoPE), token_types, vinput_mask, plus the ref + dual-path tensors when refs are provided. + """ + from .utils import get_rope_index_fix_point + from comfy.text_encoders.hidream_o1 import ( + IMAGE_TOKEN_ID, VISION_START_ID, VISION_END_ID, + ) + + if text_input_ids.dim() == 1: + text_input_ids = text_input_ids.unsqueeze(0) + text_input_ids = text_input_ids.long().to(noise.device) + B = noise.shape[0] + if text_input_ids.shape[0] == 1 and B > 1: + text_input_ids = text_input_ids.expand(B, -1) + + H, W = noise.shape[-2], noise.shape[-1] + h_p, w_p = H // target_patch_size, W // target_patch_size + image_len = h_p * w_p + image_grid_thw_tgt = torch.tensor( + [[1, h_p, w_p]], dtype=torch.long, device=text_input_ids.device, + ) + + out = {} + if ref_images: + ref = prepare_ref_images(ref_images, H, W, device=noise.device, dtype=noise.dtype) + text_input_ids = build_ref_input_ids( + text_input_ids, ref["per_ref_vit_tokens"], + IMAGE_TOKEN_ID, VISION_START_ID, VISION_END_ID, + ) + new_txt_len = text_input_ids.shape[1] + + # Each ref's patchified stream gets a [vision_start, image_pad*N-1] + # block in the position-id stream after the noised target. + ref_grid_lengths = [hp * wp for (hp, wp) in ref["per_ref_patch_grids"]] + tgt_vision = torch.full((1, image_len), IMAGE_TOKEN_ID, + dtype=text_input_ids.dtype, device=text_input_ids.device) + tgt_vision[:, 0] = VISION_START_ID + ref_vision_blocks = [] + for rl in ref_grid_lengths: + blk = torch.full((1, rl), IMAGE_TOKEN_ID, + dtype=text_input_ids.dtype, device=text_input_ids.device) + blk[:, 0] = VISION_START_ID + ref_vision_blocks.append(blk) + ref_vision_cat = torch.cat([tgt_vision] + ref_vision_blocks, dim=1) + input_ids_pad = torch.cat([text_input_ids, ref_vision_cat], dim=-1) + total_ref_patches_len = sum(ref_grid_lengths) + total_len = new_txt_len + image_len + total_ref_patches_len + + # K (ViT, post-merge) + 1 (target) + K (ref-patches) image grids. + K = len(ref_images) + igthw_cond = ref["ref_image_grid_thw"].clone() + igthw_cond[:, 1] //= 2 + igthw_cond[:, 2] //= 2 + image_grid_thw_ref = torch.tensor( + [[1, hp, wp] for (hp, wp) in ref["per_ref_patch_grids"]], + dtype=torch.long, device=text_input_ids.device, + ) + igthw_all = torch.cat([ + igthw_cond.to(text_input_ids.device), + image_grid_thw_tgt, + image_grid_thw_ref, + ], dim=0) + position_ids, _ = get_rope_index_fix_point( + spatial_merge_size=1, + image_token_id=IMAGE_TOKEN_ID, + vision_start_token_id=VISION_START_ID, + input_ids=input_ids_pad, image_grid_thw=igthw_all, + attention_mask=None, + skip_vision_start_token=[0] * K + [1] + [1] * K, + fix_point=4096, + ) + + # tms + target_image + ref_patches are all gen. + tms_pos = new_txt_len - 1 + ar_len = tms_pos + token_types = torch.zeros(B, total_len, dtype=torch.long, device=noise.device) + token_types[:, tms_pos:] = 1 + vinput_mask = torch.zeros(B, total_len, dtype=torch.bool, device=noise.device) + vinput_mask[:, new_txt_len:] = True + + # Leading batch dim sidesteps CONDRegular.process_cond's repeat_to_batch_size truncation + out["ref_pixel_values"] = ref["ref_pixel_values"].unsqueeze(0) + out["ref_image_grid_thw"] = ref["ref_image_grid_thw"].unsqueeze(0) + out["ref_patches"] = ref["ref_patches"] + else: + # T2I: text + noised target only, vision_start replaces the first image token + txt_len = text_input_ids.shape[1] + total_len = txt_len + image_len + vision_tokens = torch.full((B, image_len), IMAGE_TOKEN_ID, + dtype=text_input_ids.dtype, device=text_input_ids.device) + vision_tokens[:, 0] = VISION_START_ID + input_ids_pad = torch.cat([text_input_ids, vision_tokens], dim=-1) + position_ids, _ = get_rope_index_fix_point( + spatial_merge_size=1, + image_token_id=IMAGE_TOKEN_ID, + vision_start_token_id=VISION_START_ID, + input_ids=input_ids_pad, image_grid_thw=image_grid_thw_tgt, + attention_mask=None, + skip_vision_start_token=[1], + ) + ar_len = txt_len - 1 + token_types = torch.zeros(B, total_len, dtype=torch.long, device=noise.device) + token_types[:, ar_len:] = 1 + vinput_mask = torch.zeros(B, total_len, dtype=torch.bool, device=noise.device) + vinput_mask[:, txt_len:] = True + + out["input_ids"] = text_input_ids + out["position_ids"] = position_ids[:, 0].unsqueeze(0) # Collapse position_ids batch and add a leading dim so CONDRegular's batch-resize doesn't truncate the 3-axis MRoPE dim + out["token_types"] = token_types + out["vinput_mask"] = vinput_mask + out["ar_len"] = ar_len + return out diff --git a/comfy/ldm/hidream_o1/model.py b/comfy/ldm/hidream_o1/model.py new file mode 100644 index 000000000..a223e706f --- /dev/null +++ b/comfy/ldm/hidream_o1/model.py @@ -0,0 +1,306 @@ +"""HiDream-O1-Image transformer. + +Pixel-space DiT built on Qwen3-VL: the vision tower (Qwen35VisionModel) +encodes ref images, the Qwen3-VL-8B decoder (Llama2_ with interleaved MRoPE) +processes a unified text+image sequence, and 32x32 patch embed/unembed +shims map raw RGB in and out of LLM hidden space. The Qwen3-VL deepstack +mergers go unused — their weights are dropped at load. +""" + +from dataclasses import dataclass, field +from typing import List, Optional + +import einops +import torch +import torch.nn as nn + +import comfy.patcher_extension +from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder +from comfy.text_encoders.llama import Llama2_ +from comfy.text_encoders.qwen35 import Qwen35VisionModel + +from .attention import make_two_pass_attention + + +IMAGE_TOKEN_ID = 151655 # Qwen3-VL <|image_pad|> +TMS_TOKEN_ID = 151673 # HiDream-O1 <|tms_token|> +PATCH_SIZE = 32 + + +@dataclass +class HiDreamO1TextConfig: + """Qwen3-VL-8B text-decoder dims (matches public Qwen3-VL-8B-Instruct).""" + vocab_size: int = 151936 + hidden_size: int = 4096 + intermediate_size: int = 12288 + num_hidden_layers: int = 36 + num_attention_heads: int = 32 + num_key_value_heads: int = 8 + head_dim: int = 128 + max_position_embeddings: int = 128000 + rms_norm_eps: float = 1e-6 + rope_theta: float = 5000000.0 + rope_scale: Optional[float] = None + rope_dims: List[int] = field(default_factory=lambda: [24, 20, 20]) + interleaved_mrope: bool = True + transformer_type: str = "llama" + rms_norm_add: bool = False + mlp_activation: str = "silu" + qkv_bias: bool = False + q_norm: str = "gemma3" + k_norm: str = "gemma3" + final_norm: bool = True + lm_head: bool = False + stop_tokens: List[int] = field(default_factory=lambda: [151643, 151645]) + + +QWEN3VL_VISION_DEFAULTS = dict( + hidden_size=1152, + num_heads=16, + intermediate_size=4304, + depth=27, + patch_size=16, + temporal_patch_size=2, + in_channels=3, + spatial_merge_size=2, + num_position_embeddings=2304, + deepstack_visual_indexes=(8, 16, 24), + out_hidden_size=4096, # final merger projects directly into LLM hidden +) + + +class BottleneckPatchEmbed(nn.Module): + # 3072 -> 1024 -> 4096 (raw 32x32 RGB patch -> bottleneck -> LLM hidden). + def __init__(self, patch_size=32, in_chans=3, pca_dim=1024, embed_dim=4096, bias=True, device=None, dtype=None, ops=None): + super().__init__() + self.proj1 = ops.Linear(patch_size * patch_size * in_chans, pca_dim, bias=False, device=device, dtype=dtype) + self.proj2 = ops.Linear(pca_dim, embed_dim, bias=bias, device=device, dtype=dtype) + + def forward(self, x): + return self.proj2(self.proj1(x)) + + +class FinalLayer(nn.Module): + # 4096 -> 3072 (LLM hidden -> flat pixel patch). + def __init__(self, hidden_size, patch_size=32, out_channels=3, device=None, dtype=None, ops=None): + super().__init__() + self.linear = ops.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True, device=device, dtype=dtype) + + def forward(self, x): + return self.linear(x) + + +class HiDreamO1Transformer(nn.Module): + """HiDream-O1 unified pixel-level transformer.""" + + def __init__(self, image_model=None, dtype=None, device=None, operations=None, + text_config_overrides=None, vision_config_overrides=None, **kwargs): + super().__init__() + self.dtype = dtype + + text_cfg = HiDreamO1TextConfig(**(text_config_overrides or {})) + vision_cfg = dict(QWEN3VL_VISION_DEFAULTS) + if vision_config_overrides: + vision_cfg.update(vision_config_overrides) + vision_cfg["out_hidden_size"] = text_cfg.hidden_size + + self.text_config = text_cfg + self.vision_config = vision_cfg + self.hidden_size = text_cfg.hidden_size + self.patch_size = PATCH_SIZE + self.in_channels = 3 + self.tms_token_id = TMS_TOKEN_ID + + self.visual = Qwen35VisionModel(vision_cfg, device=device, dtype=dtype, ops=operations) + self.language_model = Llama2_(text_cfg, device=device, dtype=dtype, ops=operations) + self.t_embedder1 = TimestepEmbedder( + text_cfg.hidden_size, device=device, dtype=dtype, operations=operations, + ) + self.x_embedder = BottleneckPatchEmbed( + patch_size=self.patch_size, in_chans=self.in_channels, + pca_dim=text_cfg.hidden_size // 4, embed_dim=text_cfg.hidden_size, + bias=True, device=device, dtype=dtype, ops=operations, + ) + self.final_layer2 = FinalLayer( + text_cfg.hidden_size, patch_size=self.patch_size, + out_channels=self.in_channels, device=device, dtype=dtype, ops=operations, + ) + + self._visual_cache = None + self._kv_cache_entries = [] + + def clear_kv_cache(self): + self._kv_cache_entries = [] + self._visual_cache = None + + def forward(self, x, timesteps, context=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, transformer_options, **kwargs) + + def _forward(self, x, timesteps, context=None, transformer_options={}, input_ids=None, attention_mask=None, position_ids=None, + vinput_mask=None, ar_len=None, ref_pixel_values=None, ref_image_grid_thw=None, ref_patches=None, **kwargs): + """Returns flow-match velocity (x - x_pred) / sigma""" + + if input_ids is None or position_ids is None: + raise ValueError("HiDreamO1Transformer requires input_ids and position_ids in conditioning") + + B, _, H, W = x.shape + h_p, w_p = H // self.patch_size, W // self.patch_size + tgt_image_len = h_p * w_p + + z = einops.rearrange( + x, 'B C (H p1) (W p2) -> B (H W) (C p1 p2)', + p1=self.patch_size, p2=self.patch_size, + ) + vinputs = torch.cat([z, ref_patches.to(z.dtype)], dim=1) if ref_patches is not None else z + + inputs_embeds = self.language_model.embed_tokens(input_ids).to(x.dtype) + + if ref_pixel_values is not None and ref_image_grid_thw is not None: + # ViT output is constant across sampling steps within a generation + # identity-key by the input tensor so refs don't recompute every step. + cached = self._visual_cache + if cached is not None and cached[0] is ref_pixel_values: + image_embeds = cached[1] + else: + ref_pv = ref_pixel_values.to(inputs_embeds.device) + ref_grid = ref_image_grid_thw.to(inputs_embeds.device).long() + # extra_conds wraps with a leading batch dim; refs are model-level so [0] always recovers them. + if ref_pv.dim() == 3: + ref_pv = ref_pv[0] + if ref_grid.dim() == 3: + ref_grid = ref_grid[0] + image_embeds = self.visual(ref_pv, ref_grid).to(inputs_embeds.dtype) + self._visual_cache = (ref_pixel_values, image_embeds) + # image_pad positions identical across batch (input_ids shared cond/uncond). + image_idx = (input_ids[0] == IMAGE_TOKEN_ID).nonzero(as_tuple=True)[0] + if image_idx.shape[0] != image_embeds.shape[0]: + raise ValueError( + f"Image-token count {image_idx.shape[0]} != ViT output count " + f"{image_embeds.shape[0]}; check tokenizer/processor alignment." + ) + inputs_embeds[:, image_idx] = image_embeds.unsqueeze(0).expand(B, -1, -1) + + sigma = timesteps.float() / 1000.0 + t_pixeldit = 1.0 - sigma + t_emb = self.t_embedder1(t_pixeldit * 1000, inputs_embeds.dtype) + tms_mask_3d = (input_ids == self.tms_token_id).unsqueeze(-1).expand_as(inputs_embeds) + inputs_embeds = torch.where(tms_mask_3d, t_emb.unsqueeze(1).expand_as(inputs_embeds), inputs_embeds) + + vinputs_embedded = self.x_embedder(vinputs.to(inputs_embeds.dtype)) + inputs_embeds = torch.cat([inputs_embeds, vinputs_embedded], dim=1) + + # extra_conds stores position_ids as (1, 3, T); process_cond repeats dim 0 to B. Take row 0. + freqs_cis = self.language_model.compute_freqs_cis(position_ids[0].to(x.device), x.device) + freqs_cis = tuple(t.to(x.dtype) for t in freqs_cis) + + two_pass_attn = make_two_pass_attention(ar_len, transformer_options=transformer_options) + patches_replace = transformer_options.get("patches_replace", {}) + blocks_replace = patches_replace.get("dit", {}) + transformer_options["total_blocks"] = len(self.language_model.layers) + transformer_options["block_type"] = "double" + + # Cache prefix K/V across steps. Key includes input_ids (prompt), ref_id + # (refs scatter into inputs_embeds), and position_ids (RoPE baked into cached K). + can_cache = not blocks_replace and ar_len > 0 + cache_len = ar_len if can_cache else 0 + ref_id = id(ref_pixel_values) if ref_pixel_values is not None else None + pos_ids_key = position_ids[..., :cache_len] if can_cache else position_ids + cache_entries = self._kv_cache_entries + # Drop stale entries from a previous device (model was unloaded and reloaded). + if cache_entries and cache_entries[0]["input_ids"].device != input_ids.device: + cache_entries = [] + self._kv_cache_entries = [] + kv_cache = None + if can_cache: + for entry in cache_entries: + ck = entry["input_ids"] + ep = entry["position_ids"] + if (entry["cache_len"] == cache_len + and ck.shape == input_ids.shape and torch.equal(ck, input_ids) + and entry["ref_id"] == ref_id + and ep.shape == pos_ids_key.shape and torch.equal(ep, pos_ids_key)): + kv_cache = entry + break + + if kv_cache is not None: + # Hot path: project Q/K/V only for fresh positions; past_key_value prepends cached AR K/V. + hidden_states = inputs_embeds[:, cache_len:] + sliced_freqs = tuple(t[..., cache_len:, :] for t in freqs_cis) + for i, layer in enumerate(self.language_model.layers): + transformer_options["block_index"] = i + K_i, V_i = kv_cache["kv"][i] + hidden_states, _ = layer( + x=hidden_states, attention_mask=None, freqs_cis=sliced_freqs, optimized_attention=two_pass_attn, + past_key_value=(K_i, V_i, cache_len), + ) + else: + # Cold path: run full sequence; if cacheable, snapshot K/V at AR positions. + snapshots = [] if can_cache else None + past_kv_cold = () if can_cache else None + hidden_states = inputs_embeds + for i, layer in enumerate(self.language_model.layers): + transformer_options["block_index"] = i + if ("double_block", i) in blocks_replace: + def block_wrap(args, _layer=layer): + out = {} + out["x"], _ = _layer( + x=args["x"], attention_mask=args.get("attention_mask"), + freqs_cis=args["freqs_cis"], optimized_attention=args["optimized_attention"], + past_key_value=None, + ) + return out + out = blocks_replace[("double_block", i)]( + {"x": hidden_states, "attention_mask": None, + "freqs_cis": freqs_cis, "optimized_attention": two_pass_attn, + "transformer_options": transformer_options}, + {"original_block": block_wrap}, + ) + hidden_states = out["x"] + else: + hidden_states, present_kv = layer( + x=hidden_states, attention_mask=None, + freqs_cis=freqs_cis, optimized_attention=two_pass_attn, + past_key_value=past_kv_cold, + ) + if snapshots is not None: + K, V, _ = present_kv + snapshots.append((K[:, :, :cache_len].contiguous(), + V[:, :, :cache_len].contiguous())) + if snapshots is not None: + # Cap at 2 entries (cond + uncond). Multi-cond workflows LRU-evict. + new_entry = { + "input_ids": input_ids.clone(), + "cache_len": cache_len, + "kv": snapshots, + "ref_id": ref_id, + "position_ids": pos_ids_key.clone(), + } + self._kv_cache_entries = (cache_entries + [new_entry])[-2:] + + if self.language_model.norm is not None: + hidden_states = self.language_model.norm(hidden_states) + + # Slice target-image positions before the final projection so the Linear only runs on tgt_image_len tokens. + # In the hot path hidden_states starts at original position cache_len, so masks/indices shift by cache_len. + sliced_offset = cache_len if kv_cache is not None else 0 + if vinput_mask is not None: + vmask = vinput_mask.to(x.device).bool() + if sliced_offset > 0: + vmask = vmask[:, sliced_offset:] + target_hidden = hidden_states[vmask].view(B, -1, hidden_states.shape[-1])[:, :tgt_image_len] + else: + txt_seq_len = input_ids.shape[1] + start = txt_seq_len - sliced_offset + target_hidden = hidden_states[:, start:start + tgt_image_len] + x_pred_tgt = self.final_layer2(target_hidden) + + # fp32 final subtraction, bf16 here noticeably degrades samples. + x_pred_img = einops.rearrange( + x_pred_tgt, 'B (H W) (C p1 p2) -> B C (H p1) (W p2)', + H=h_p, W=w_p, p1=self.patch_size, p2=self.patch_size, + ) + return (x.float() - x_pred_img.float()) / sigma.view(B, 1, 1, 1).clamp_min(1e-3) diff --git a/comfy/ldm/hidream_o1/utils.py b/comfy/ldm/hidream_o1/utils.py new file mode 100644 index 000000000..5a1249c72 --- /dev/null +++ b/comfy/ldm/hidream_o1/utils.py @@ -0,0 +1,173 @@ +"""HiDream-O1 input-prep helpers: image/resolution math and unified-sequence +RoPE position-id assembly. The fix_point offset in get_rope_index_fix_point +lets the target image and patchified ref images share spatial RoPE positions +despite living at different sequence indices — same 2D image plane. +""" + +import math +from typing import Optional + +import torch + + +PATCH_SIZE = 32 +CONDITION_IMAGE_SIZE = 384 # ViT-side base size for ref images + + +def resize_tensor(img_t, image_size, patch_size=16): + """img_t: (1, 3, H, W) float [0, 1]. Fit to image_size**2 area, patch-aligned, center-cropped.""" + + while min(img_t.shape[-2], img_t.shape[-1]) >= 2 * image_size: # Pre-halves with 2x2 box averaging while the image is still very large + img_t = torch.nn.functional.avg_pool2d(img_t, kernel_size=2, stride=2) + + _, _, height, width = img_t.shape + m = patch_size + s_max = image_size * image_size + scale = math.sqrt(s_max / (width * height)) + + candidates = [ + (round(width * scale) // m * m, round(height * scale) // m * m), + (round(width * scale) // m * m, math.floor(height * scale) // m * m), + (math.floor(width * scale) // m * m, round(height * scale) // m * m), + (math.floor(width * scale) // m * m, math.floor(height * scale) // m * m), + ] + candidates = sorted(candidates, key=lambda x: x[0] * x[1], reverse=True) + new_size = candidates[-1] + for c in candidates: + if c[0] * c[1] <= s_max: + new_size = c + break + + new_w, new_h = new_size + s1 = width / new_w + s2 = height / new_h + if s1 < s2: + resize_w, resize_h = new_w, round(height / s1) + else: + resize_w, resize_h = round(width / s2), new_h + img_t = torch.nn.functional.interpolate(img_t, size=(resize_h, resize_w), mode="bicubic") + top = (resize_h - new_h) // 2 + left = (resize_w - new_w) // 2 + return img_t[..., top:top + new_h, left:left + new_w] + + +def calculate_dimensions(max_size, ratio): + """(W, H) for an aspect ratio fitting in max_size**2 area, 32-aligned.""" + width = math.sqrt(max_size * max_size * ratio) + height = width / ratio + width = int(width / 32) * 32 + height = int(height / 32) * 32 + return width, height + + +def ref_max_size(target_max_dim, k): + """K-dependent ref-image max dim before patchifying.""" + if k == 1: + return target_max_dim + if k == 2: + return target_max_dim * 48 // 64 + if k <= 4: + return target_max_dim // 2 + if k <= 8: + return target_max_dim * 24 // 64 + return target_max_dim // 4 + + +def cond_image_size(k): + """K-dependent ViT-side image size.""" + if k <= 4: + return CONDITION_IMAGE_SIZE + if k <= 8: + return CONDITION_IMAGE_SIZE * 48 // 64 + return CONDITION_IMAGE_SIZE // 2 + + +def get_rope_index_fix_point( + spatial_merge_size: int, + image_token_id: int, + vision_start_token_id: int, + input_ids: Optional[torch.LongTensor] = None, + image_grid_thw: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.Tensor] = None, + skip_vision_start_token=None, + fix_point: int = 4096, +): + mrope_position_deltas = [] + if input_ids is not None and image_grid_thw is not None: + total_input_ids = input_ids + if attention_mask is None: + attention_mask = torch.ones_like(total_input_ids) + position_ids = torch.ones( + 3, input_ids.shape[0], input_ids.shape[1], + dtype=input_ids.dtype, device=input_ids.device, + ) + attention_mask = attention_mask.to(total_input_ids.device) + for i, input_ids_b in enumerate(total_input_ids): + fp = fix_point + image_index = 0 + input_ids_b = input_ids_b[attention_mask[i] == 1] + vision_start_indices = torch.argwhere(input_ids_b == vision_start_token_id).squeeze(1) + vision_tokens = input_ids_b[vision_start_indices + 1] + image_nums = (vision_tokens == image_token_id).sum() + input_tokens = input_ids_b.tolist() + llm_pos_ids_list = [] + st = 0 + remain_images = image_nums + for _ in range(image_nums): + if image_token_id in input_tokens and remain_images > 0: + ed = input_tokens.index(image_token_id, st) + else: + ed = len(input_tokens) + 1 + t = image_grid_thw[image_index][0] + h = image_grid_thw[image_index][1] + w = image_grid_thw[image_index][2] + image_index += 1 + remain_images -= 1 + llm_grid_t = t.item() + llm_grid_h = h.item() // spatial_merge_size + llm_grid_w = w.item() // spatial_merge_size + text_len = ed - st + text_len -= skip_vision_start_token[image_index - 1] + text_len = max(0, text_len) + st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0 + llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx) + + t_index = torch.arange(llm_grid_t).view(-1, 1).expand(-1, llm_grid_h * llm_grid_w).flatten() + h_index = torch.arange(llm_grid_h).view(1, -1, 1).expand(llm_grid_t, -1, llm_grid_w).flatten() + w_index = torch.arange(llm_grid_w).view(1, 1, -1).expand(llm_grid_t, llm_grid_h, -1).flatten() + + if skip_vision_start_token[image_index - 1]: + if fp > 0: + fp = fp - st_idx + llm_pos_ids_list.append(torch.stack([t_index, h_index, w_index]) + fp + st_idx) + fp = 0 + else: + llm_pos_ids_list.append(torch.stack([t_index, h_index, w_index]) + text_len + st_idx) + st = ed + llm_grid_t * llm_grid_h * llm_grid_w + + if st < len(input_tokens): + st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0 + text_len = len(input_tokens) - st + llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx) + + llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1) + position_ids[..., i, attention_mask[i] == 1] = llm_positions.to(position_ids.device) + mrope_position_deltas.append(llm_positions.max() + 1 - len(total_input_ids[i])) + mrope_position_deltas = torch.tensor(mrope_position_deltas, device=input_ids.device).unsqueeze(1) + return position_ids, mrope_position_deltas + + if attention_mask is not None: + position_ids = attention_mask.long().cumsum(-1) - 1 + position_ids.masked_fill_(attention_mask == 0, 1) + position_ids = position_ids.unsqueeze(0).expand(3, -1, -1).to(attention_mask.device) + max_position_ids = position_ids.max(0, keepdim=False)[0].max(-1, keepdim=True)[0] + mrope_position_deltas = max_position_ids + 1 - attention_mask.shape[-1] + else: + position_ids = ( + torch.arange(input_ids.shape[1], device=input_ids.device) + .view(1, 1, -1).expand(3, input_ids.shape[0], -1) + ) + mrope_position_deltas = torch.zeros( + [input_ids.shape[0], 1], device=input_ids.device, dtype=input_ids.dtype, + ) + return position_ids, mrope_position_deltas diff --git a/comfy/ldm/lightricks/av_model.py b/comfy/ldm/lightricks/av_model.py index 6f2ba41ef..bc09fb77e 100644 --- a/comfy/ldm/lightricks/av_model.py +++ b/comfy/ldm/lightricks/av_model.py @@ -16,31 +16,31 @@ from comfy.ldm.lightricks.model import ( from comfy.ldm.lightricks.symmetric_patchifier import AudioPatchifier from comfy.ldm.lightricks.embeddings_connector import Embeddings1DConnector import comfy.ldm.common_dit +import comfy.model_prefetch class CompressedTimestep: """Store video timestep embeddings in compressed form using per-frame indexing.""" __slots__ = ('data', 'batch_size', 'num_frames', 'patches_per_frame', 'feature_dim') - def __init__(self, tensor: torch.Tensor, patches_per_frame: int): + def __init__(self, tensor: torch.Tensor, patches_per_frame: int, per_frame: bool = False): """ - tensor: [batch_size, num_tokens, feature_dim] tensor where num_tokens = num_frames * patches_per_frame - patches_per_frame: Number of spatial patches per frame (height * width in latent space), or None to disable compression + tensor: [batch, num_tokens, feature_dim] (per-token, default) or + [batch, num_frames, feature_dim] (per_frame=True, already compressed). + patches_per_frame: spatial patches per frame; pass None to disable compression. """ - self.batch_size, num_tokens, self.feature_dim = tensor.shape - - # Check if compression is valid (num_tokens must be divisible by patches_per_frame) - if patches_per_frame is not None and num_tokens % patches_per_frame == 0 and num_tokens >= patches_per_frame: + self.batch_size, n, self.feature_dim = tensor.shape + if per_frame: self.patches_per_frame = patches_per_frame - self.num_frames = num_tokens // patches_per_frame - - # Reshape to [batch, frames, patches_per_frame, feature_dim] and store one value per frame - # All patches in a frame are identical, so we only keep the first one - reshaped = tensor.view(self.batch_size, self.num_frames, patches_per_frame, self.feature_dim) - self.data = reshaped[:, :, 0, :].contiguous() # [batch, frames, feature_dim] + self.num_frames = n + self.data = tensor + elif patches_per_frame is not None and n >= patches_per_frame and n % patches_per_frame == 0: + self.patches_per_frame = patches_per_frame + self.num_frames = n // patches_per_frame + # All patches in a frame are identical — keep only the first. + self.data = tensor.view(self.batch_size, self.num_frames, patches_per_frame, self.feature_dim)[:, :, 0, :].contiguous() else: - # Not divisible or too small - store directly without compression self.patches_per_frame = 1 - self.num_frames = num_tokens + self.num_frames = n self.data = tensor def expand(self): @@ -715,32 +715,35 @@ class LTXAVModel(LTXVModel): def _prepare_timestep(self, timestep, batch_size, hidden_dtype, **kwargs): """Prepare timestep embeddings.""" - # TODO: some code reuse is needed here. grid_mask = kwargs.get("grid_mask", None) - if grid_mask is not None: - timestep = timestep[:, grid_mask] - - timestep_scaled = timestep * self.timestep_scale_multiplier - - v_timestep, v_embedded_timestep = self.adaln_single( - timestep_scaled.flatten(), - {"resolution": None, "aspect_ratio": None}, - batch_size=batch_size, - hidden_dtype=hidden_dtype, - ) - - # Calculate patches_per_frame from orig_shape: [batch, channels, frames, height, width] - # Video tokens are arranged as (frames * height * width), so patches_per_frame = height * width orig_shape = kwargs.get("orig_shape") has_spatial_mask = kwargs.get("has_spatial_mask", None) v_patches_per_frame = None if not has_spatial_mask and orig_shape is not None and len(orig_shape) == 5: - # orig_shape[3] = height, orig_shape[4] = width (in latent space) v_patches_per_frame = orig_shape[3] * orig_shape[4] - # Reshape to [batch_size, num_tokens, dim] and compress for storage - v_timestep = CompressedTimestep(v_timestep.view(batch_size, -1, v_timestep.shape[-1]), v_patches_per_frame) - v_embedded_timestep = CompressedTimestep(v_embedded_timestep.view(batch_size, -1, v_embedded_timestep.shape[-1]), v_patches_per_frame) + # Used by compute_prompt_timestep and the audio cross-attention paths. + timestep_scaled = (timestep[:, grid_mask] if grid_mask is not None else timestep) * self.timestep_scale_multiplier + + # When patches in a frame share a timestep (no spatial mask), project one row per frame instead of one per token + per_frame_path = v_patches_per_frame is not None and (timestep.numel() // batch_size) % v_patches_per_frame == 0 + if per_frame_path: + per_frame = timestep.reshape(batch_size, -1, v_patches_per_frame)[:, :, 0] + if grid_mask is not None: + # All-or-nothing per frame when has_spatial_mask=False. + per_frame = per_frame[:, grid_mask[::v_patches_per_frame]] + ts_input = per_frame * self.timestep_scale_multiplier + else: + ts_input = timestep_scaled + + v_timestep, v_embedded_timestep = self.adaln_single( + ts_input.flatten(), + {"resolution": None, "aspect_ratio": None}, + batch_size=batch_size, + hidden_dtype=hidden_dtype, + ) + v_timestep = CompressedTimestep(v_timestep.view(batch_size, -1, v_timestep.shape[-1]), v_patches_per_frame, per_frame=per_frame_path) + v_embedded_timestep = CompressedTimestep(v_embedded_timestep.view(batch_size, -1, v_embedded_timestep.shape[-1]), v_patches_per_frame, per_frame=per_frame_path) v_prompt_timestep = compute_prompt_timestep( self.prompt_adaln_single, timestep_scaled, batch_size, hidden_dtype @@ -907,9 +910,11 @@ class LTXAVModel(LTXVModel): """Process transformer blocks for LTXAV.""" patches_replace = transformer_options.get("patches_replace", {}) blocks_replace = patches_replace.get("dit", {}) + prefetch_queue = comfy.model_prefetch.make_prefetch_queue(list(self.transformer_blocks), vx.device, transformer_options) # Process transformer blocks for i, block in enumerate(self.transformer_blocks): + comfy.model_prefetch.prefetch_queue_pop(prefetch_queue, vx.device, block) if ("double_block", i) in blocks_replace: def block_wrap(args): @@ -982,6 +987,8 @@ class LTXAVModel(LTXVModel): a_prompt_timestep=a_prompt_timestep, ) + comfy.model_prefetch.prefetch_queue_pop(prefetch_queue, vx.device, None) + return [vx, ax] def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs): diff --git a/comfy/ldm/lightricks/model.py b/comfy/ldm/lightricks/model.py index bfbc08357..e0a4a0f9b 100644 --- a/comfy/ldm/lightricks/model.py +++ b/comfy/ldm/lightricks/model.py @@ -358,6 +358,61 @@ def apply_split_rotary_emb(input_tensor, cos, sin): return output.swapaxes(1, 2).reshape(B, T, -1) if needs_reshape else output +class GuideAttentionMask: + """Holds the two per-group masks for LTXV guide self-attention. + _attention_with_guide_mask splits queries into noisy and tracked-guide + groups, so the largest mask is (1, 1, tracked_count, T). + """ + __slots__ = ("guide_start", "tracked_count", "noisy_mask", "tracked_mask") + + def __init__(self, total_tokens, guide_start, tracked_count, tracked_weights): + device = tracked_weights.device + dtype = tracked_weights.dtype + finfo = torch.finfo(dtype) + + pos = tracked_weights > 0 + log_w = torch.full_like(tracked_weights, finfo.min) + log_w[pos] = torch.log(tracked_weights[pos].clamp(min=finfo.tiny)) + + self.guide_start = guide_start + self.tracked_count = tracked_count + + self.noisy_mask = torch.zeros((1, 1, 1, total_tokens), device=device, dtype=dtype) + self.noisy_mask[:, :, :, guide_start:guide_start + tracked_count] = log_w.view(1, 1, 1, -1) + + self.tracked_mask = torch.zeros((1, 1, tracked_count, total_tokens), device=device, dtype=dtype) + self.tracked_mask[:, :, :, :guide_start] = log_w.view(1, 1, -1, 1) + + +def _attention_with_guide_mask(q, k, v, heads, guide_mask, attn_precision, transformer_options): + """Apply the guide mask by partitioning Q into noisy and tracked-guide + groups, so each group needs only its own sub-mask. Avoids materializing + the (1,1,T,T) dense mask. + """ + guide_start = guide_mask.guide_start + tracked_end = guide_start + guide_mask.tracked_count + + out = torch.empty_like(q) + + if guide_start > 0: # In practice currently guides are always after noise, guard for safety if this changes. + out[:, :guide_start, :] = comfy.ldm.modules.attention.optimized_attention( + q[:, :guide_start, :], k, v, heads, mask=guide_mask.noisy_mask, + attn_precision=attn_precision, transformer_options=transformer_options, + low_precision_attention=False, # sageattn mask support is unreliable + ) + out[:, guide_start:tracked_end, :] = comfy.ldm.modules.attention.optimized_attention( + q[:, guide_start:tracked_end, :], k, v, heads, mask=guide_mask.tracked_mask, + attn_precision=attn_precision, transformer_options=transformer_options, + low_precision_attention=False, + ) + if tracked_end < q.shape[1]: # Every guide token is tracked, and nothing comes after them, guard for safety if this changes. + out[:, tracked_end:, :] = comfy.ldm.modules.attention.optimized_attention( + q[:, tracked_end:, :], k, v, heads, + attn_precision=attn_precision, transformer_options=transformer_options, + ) + return out + + class CrossAttention(nn.Module): def __init__( self, @@ -412,8 +467,10 @@ class CrossAttention(nn.Module): if mask is None: out = comfy.ldm.modules.attention.optimized_attention(q, k, v, self.heads, attn_precision=self.attn_precision, transformer_options=transformer_options) + elif isinstance(mask, GuideAttentionMask): + out = _attention_with_guide_mask(q, k, v, self.heads, mask, attn_precision=self.attn_precision, transformer_options=transformer_options) else: - out = comfy.ldm.modules.attention.optimized_attention_masked(q, k, v, self.heads, mask, attn_precision=self.attn_precision, transformer_options=transformer_options) + out = comfy.ldm.modules.attention.optimized_attention(q, k, v, self.heads, mask=mask, attn_precision=self.attn_precision, transformer_options=transformer_options) # Apply per-head gating if enabled if self.to_gate_logits is not None: @@ -1063,7 +1120,9 @@ class LTXVModel(LTXBaseModel): additional_args["resolved_guide_entries"] = resolved_entries keyframe_idxs = keyframe_idxs[..., kf_grid_mask, :] - pixel_coords[:, :, -keyframe_idxs.shape[2]:, :] = keyframe_idxs + + if keyframe_idxs.shape[2] > 0: # Guard for the case of no keyframes surviving + pixel_coords[:, :, -keyframe_idxs.shape[2]:, :] = keyframe_idxs # Total surviving guide tokens (all guides) additional_args["num_guide_tokens"] = keyframe_idxs.shape[2] @@ -1099,12 +1158,12 @@ class LTXVModel(LTXBaseModel): if not resolved_entries: return None - # Check if any attenuation is actually needed - needs_attenuation = any( - e["strength"] < 1.0 or e.get("pixel_mask") is not None + # strength != 1.0 means we want to either attenuate (< 1) or amplify (> 1) guide attention. + needs_mask = any( + e["strength"] != 1.0 or e.get("pixel_mask") is not None for e in resolved_entries ) - if not needs_attenuation: + if not needs_mask: return None # Build per-guide-token weights for all tracked guide tokens. @@ -1159,16 +1218,11 @@ class LTXVModel(LTXBaseModel): # Concatenate per-token weights for all tracked guides tracked_weights = torch.cat(all_weights, dim=1) # (1, total_tracked) - # Check if any weight is actually < 1.0 (otherwise no attenuation needed) - if (tracked_weights >= 1.0).all(): + # Skip when every weight is exactly 1.0 (additive bias would be 0). + if (tracked_weights == 1.0).all(): return None - # Build the mask: guide tokens are at the end of the sequence. - # Tracked guides come first (in order), untracked follow. - return self._build_self_attention_mask( - total_tokens, num_guide_tokens, total_tracked, - tracked_weights, guide_start, device, dtype, - ) + return GuideAttentionMask(total_tokens, guide_start, total_tracked, tracked_weights) @staticmethod def _downsample_mask_to_latent(mask, f_lat, h_lat, w_lat): @@ -1234,45 +1288,6 @@ class LTXVModel(LTXBaseModel): return rearrange(latent_mask, "b 1 f h w -> b (f h w)") - @staticmethod - def _build_self_attention_mask(total_tokens, num_guide_tokens, tracked_count, - tracked_weights, guide_start, device, dtype): - """Build a log-space additive self-attention bias mask. - - Attenuates attention between noisy tokens and tracked guide tokens. - Untracked guide tokens (at the end of the guide portion) keep full attention. - - Args: - total_tokens: Total sequence length. - num_guide_tokens: Total guide tokens (all guides) at end of sequence. - tracked_count: Number of tracked guide tokens (first in the guide portion). - tracked_weights: (1, tracked_count) tensor, values in [0, 1]. - guide_start: Index where guide tokens begin in the sequence. - device: Target device. - dtype: Target dtype. - - Returns: - (1, 1, total_tokens, total_tokens) additive bias mask. - 0.0 = full attention, negative = attenuated, finfo.min = effectively fully masked. - """ - finfo = torch.finfo(dtype) - mask = torch.zeros((1, 1, total_tokens, total_tokens), device=device, dtype=dtype) - tracked_end = guide_start + tracked_count - - # Convert weights to log-space bias - w = tracked_weights.to(device=device, dtype=dtype) # (1, tracked_count) - log_w = torch.full_like(w, finfo.min) - positive_mask = w > 0 - if positive_mask.any(): - log_w[positive_mask] = torch.log(w[positive_mask].clamp(min=finfo.tiny)) - - # noisy → tracked guides: each noisy row gets the same per-guide weight - mask[:, :, :guide_start, guide_start:tracked_end] = log_w.view(1, 1, 1, -1) - # tracked guides → noisy: each guide row broadcasts its weight across noisy cols - mask[:, :, guide_start:tracked_end, :guide_start] = log_w.view(1, 1, -1, 1) - - return mask - def _process_transformer_blocks(self, x, context, attention_mask, timestep, pe, transformer_options={}, self_attention_mask=None, **kwargs): """Process transformer blocks for LTXV.""" patches_replace = transformer_options.get("patches_replace", {}) diff --git a/comfy/ldm/modules/attention.py b/comfy/ldm/modules/attention.py index b193fe5e8..a68cb8439 100644 --- a/comfy/ldm/modules/attention.py +++ b/comfy/ldm/modules/attention.py @@ -14,6 +14,8 @@ from .sub_quadratic_attention import efficient_dot_product_attention from comfy import model_management +TORCH_HAS_GQA = model_management.torch_version_numeric >= (2, 5) + if model_management.xformers_enabled(): import xformers import xformers.ops @@ -150,7 +152,12 @@ def attention_basic(q, k, v, heads, mask=None, attn_precision=None, skip_reshape b, _, dim_head = q.shape dim_head //= heads - scale = dim_head ** -0.5 + if kwargs.get("enable_gqa", False) and q.shape[-3] != k.shape[-3]: + n_rep = q.shape[-3] // k.shape[-3] + k = k.repeat_interleave(n_rep, dim=-3) + v = v.repeat_interleave(n_rep, dim=-3) + + scale = kwargs.get("scale", dim_head ** -0.5) h = heads if skip_reshape: @@ -219,6 +226,10 @@ def attention_sub_quad(query, key, value, heads, mask=None, attn_precision=None, b, _, dim_head = query.shape dim_head //= heads + if "scale" in kwargs: + # Pre-scale query to match requested scale (cancels internal 1/sqrt(dim_head)) + query = query * (kwargs["scale"] * dim_head ** 0.5) + if skip_reshape: query = query.reshape(b * heads, -1, dim_head) value = value.reshape(b * heads, -1, dim_head) @@ -290,7 +301,7 @@ def attention_split(q, k, v, heads, mask=None, attn_precision=None, skip_reshape b, _, dim_head = q.shape dim_head //= heads - scale = dim_head ** -0.5 + scale = kwargs.get("scale", dim_head ** -0.5) if skip_reshape: q, k, v = map( @@ -500,8 +511,13 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha if mask.ndim == 3: mask = mask.unsqueeze(1) + # Pass through extra SDPA kwargs (scale, enable_gqa) if provided + # enable_gqa requires PyTorch 2.5+; older versions use manual KV expansion above + sdpa_keys = ("scale", "enable_gqa") if TORCH_HAS_GQA else ("scale",) + sdpa_extra = {k: v for k, v in kwargs.items() if k in sdpa_keys} + if SDP_BATCH_LIMIT >= b: - out = comfy.ops.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False) + out = comfy.ops.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False, **sdpa_extra) if not skip_output_reshape: out = ( out.transpose(1, 2).reshape(b, -1, heads * dim_head) @@ -519,7 +535,7 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha k[i : i + SDP_BATCH_LIMIT], v[i : i + SDP_BATCH_LIMIT], attn_mask=m, - dropout_p=0.0, is_causal=False + dropout_p=0.0, is_causal=False, **sdpa_extra ).transpose(1, 2).reshape(-1, q.shape[2], heads * dim_head) return out diff --git a/comfy/ldm/modules/diffusionmodules/util.py b/comfy/ldm/modules/diffusionmodules/util.py index 233011dc9..aed5c149c 100644 --- a/comfy/ldm/modules/diffusionmodules/util.py +++ b/comfy/ldm/modules/diffusionmodules/util.py @@ -140,7 +140,7 @@ def make_ddim_sampling_parameters(alphacums, ddim_timesteps, eta, verbose=True): alphas = alphacums[ddim_timesteps] alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist()) - # according the the formula provided in https://arxiv.org/abs/2010.02502 + # according to the formula provided in https://arxiv.org/abs/2010.02502 sigmas = eta * np.sqrt((1 - alphas_prev) / (1 - alphas) * (1 - alphas / alphas_prev)) if verbose: logging.info(f'Selected alphas for ddim sampler: a_t: {alphas}; a_(t-1): {alphas_prev}') diff --git a/comfy/ldm/moge/geometry.py b/comfy/ldm/moge/geometry.py new file mode 100644 index 000000000..7fdc97871 --- /dev/null +++ b/comfy/ldm/moge/geometry.py @@ -0,0 +1,189 @@ +"""Pure-torch + scipy geometry helpers for MoGe inference and mesh export.""" + +from __future__ import annotations + +from typing import Optional, Tuple + +import numpy as np +import torch +import torch.nn.functional as F + +from scipy.optimize import least_squares + +def normalized_view_plane_uv(width: int, height: int, aspect_ratio: Optional[float] = None, + dtype: Optional[torch.dtype] = None, device: Optional[torch.device] = None) -> torch.Tensor: + """Normalized view-plane UV coordinates with corners at +/-(W, H)/diagonal.""" + if aspect_ratio is None: + aspect_ratio = width / height + span_x = aspect_ratio / (1 + aspect_ratio ** 2) ** 0.5 + span_y = 1.0 / (1 + aspect_ratio ** 2) ** 0.5 + u = torch.linspace(-span_x * (width - 1) / width, span_x * (width - 1) / width, width, dtype=dtype, device=device) + v = torch.linspace(-span_y * (height - 1) / height, span_y * (height - 1) / height, height, dtype=dtype, device=device) + u, v = torch.meshgrid(u, v, indexing="xy") + return torch.stack([u, v], dim=-1) + + +def intrinsics_from_focal_center(fx: torch.Tensor, fy: torch.Tensor, cx: torch.Tensor, cy: torch.Tensor) -> torch.Tensor: + """Assemble (..., 3, 3) intrinsics from broadcastable fx, fy, cx, cy.""" + fx, fy, cx, cy = [torch.as_tensor(v) for v in (fx, fy, cx, cy)] + fx, fy, cx, cy = torch.broadcast_tensors(fx, fy, cx, cy) + zero = torch.zeros_like(fx) + one = torch.ones_like(fx) + return torch.stack([ + torch.stack([fx, zero, cx], dim=-1), + torch.stack([zero, fy, cy], dim=-1), + torch.stack([zero, zero, one], dim=-1), + ], dim=-2) + + +def depth_map_to_point_map(depth: torch.Tensor, intrinsics: torch.Tensor) -> torch.Tensor: + """Back-project a (..., H, W) depth map through K^-1 to (..., H, W, 3) camera-space points. + + Intrinsics use normalized image coords (x in [0, 1] left->right, y in [0, 1] top->bottom). + """ + H, W = depth.shape[-2:] + device, dtype = depth.device, depth.dtype + u = (torch.arange(W, dtype=dtype, device=device) + 0.5) / W + v = (torch.arange(H, dtype=dtype, device=device) + 0.5) / H + grid_v, grid_u = torch.meshgrid(v, u, indexing="ij") + pix = torch.stack([grid_u, grid_v, torch.ones_like(grid_u)], dim=-1) + K_inv = torch.linalg.inv(intrinsics) + rays = torch.einsum("...ij,hwj->...hwi", K_inv, pix) + return rays * depth.unsqueeze(-1) + + +def _solve_optimal_shift(uv: np.ndarray, xyz: np.ndarray, + focal: Optional[float] = None) -> Tuple[float, float]: + """LM-solve for z-shift; when focal is None, also recovers the optimal focal.""" + uv = uv.reshape(-1, 2) + xy = xyz[..., :2].reshape(-1, 2) + z = xyz[..., 2].reshape(-1) + + def fn(shift): + xy_proj = xy / (z + shift)[:, None] + f = focal if focal is not None else (xy_proj * uv).sum() / np.square(xy_proj).sum() + return (f * xy_proj - uv).ravel() + + sol = least_squares(fn, x0=0.0, ftol=1e-3, method="lm") + shift = float(np.asarray(sol["x"]).squeeze()) + if focal is None: + xy_proj = xy / (z + shift)[:, None] + focal = float((xy_proj * uv).sum() / np.square(xy_proj).sum()) + return shift, focal + + +def recover_focal_shift(points: torch.Tensor, mask: Optional[torch.Tensor] = None, + focal: Optional[torch.Tensor] = None, downsample_size: Tuple[int, int] = (64, 64) + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Recover the focal length and z-shift that turn points into a metric point map. + + Optical center is at the image center; returned focal is relative to half the image diagonal. + Returns (focal, shift) on the same device/dtype as points. + """ + shape = points.shape + H, W = shape[-3], shape[-2] + points_b = points.reshape(-1, H, W, 3) + mask_b = None if mask is None else mask.reshape(-1, H, W) + focal_b = None if focal is None else focal.reshape(-1) + + uv = normalized_view_plane_uv(W, H, dtype=points.dtype, device=points.device) + + points_lr = F.interpolate(points_b.permute(0, 3, 1, 2), downsample_size, mode="nearest").permute(0, 2, 3, 1) + uv_lr = F.interpolate(uv.unsqueeze(0).permute(0, 3, 1, 2), downsample_size, mode="nearest").squeeze(0).permute(1, 2, 0) + mask_lr = None + if mask_b is not None: + mask_lr = F.interpolate(mask_b.to(torch.float32).unsqueeze(1), downsample_size, mode="nearest").squeeze(1) > 0 + + uv_np = uv_lr.detach().cpu().numpy() + points_np = points_lr.detach().cpu().numpy() + mask_np = None if mask_lr is None else mask_lr.detach().cpu().numpy() + focal_np = None if focal_b is None else focal_b.detach().cpu().numpy() + + out_focal: list = [] + out_shift: list = [] + for i in range(points_b.shape[0]): + if mask_np is None: + xyz_i = points_np[i].reshape(-1, 3) + uv_i = uv_np.reshape(-1, 2) + else: + sel = mask_np[i] + if sel.sum() < 2: + out_focal.append(1.0) + out_shift.append(0.0) + continue + xyz_i = points_np[i][sel] + uv_i = uv_np[sel] + if focal_np is None: + shift_i, focal_i = _solve_optimal_shift(uv_i, xyz_i) + out_focal.append(focal_i) + else: + shift_i, _ = _solve_optimal_shift(uv_i, xyz_i, focal=float(focal_np[i])) + out_shift.append(shift_i) + + shift_t = torch.tensor(out_shift, device=points.device, dtype=points.dtype).reshape(shape[:-3]) + if focal is None: + focal_t = torch.tensor(out_focal, device=points.device, dtype=points.dtype).reshape(shape[:-3]) + else: + focal_t = focal.reshape(shape[:-3]) + return focal_t, shift_t + + +def depth_map_edge(depth: torch.Tensor, atol: Optional[float] = None, rtol: Optional[float] = None, kernel_size: int = 3) -> torch.Tensor: + """Per-pixel boolean: True where the local depth window's max-min span exceeds atol or rtol*depth.""" + shape = depth.shape + d = depth.reshape(-1, 1, *shape[-2:]) + pad = kernel_size // 2 + diff = F.max_pool2d(d, kernel_size, stride=1, padding=pad) + F.max_pool2d(-d, kernel_size, stride=1, padding=pad) + edge = torch.zeros_like(d, dtype=torch.bool) + if atol is not None: + edge |= diff > atol + if rtol is not None: + edge |= (diff / d.clamp_min(1e-6)).nan_to_num_() > rtol + return edge.reshape(*shape) + + +def triangulate_grid_mesh(points: torch.Tensor, mask: Optional[torch.Tensor] = None, decimation: int = 1, discontinuity_threshold: float = 0.04, + depth: Optional[torch.Tensor] = None) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """Triangulate a (H, W, 3) point map into (vertices, faces, uvs) on CPU. + + Vertices: pixels with finite coords (passing optional mask). Quads with four valid corners + become two triangles. depth overrides the scalar used for the rtol edge check; pass radial + depth for panoramas (the default points[..., 2] goes negative below the equator). + """ + points = points.detach().cpu() + finite = torch.isfinite(points).all(dim=-1) + if mask is None: + mask = finite + else: + mask = mask.detach().cpu().to(torch.bool) & finite + + if discontinuity_threshold > 0: + d = depth.detach().cpu() if depth is not None else points[..., 2] + # Replace inf with 0 so max-pool doesn't poison neighbourhoods (mask above already excludes those pixels). + d_finite = torch.where(finite, d, torch.zeros_like(d)) + edge = depth_map_edge(d_finite, rtol=discontinuity_threshold) + mask = mask & ~edge + + if decimation > 1: + points = points[::decimation, ::decimation].contiguous() + mask = mask[::decimation, ::decimation].contiguous() + H, W = points.shape[:2] + + flat_mask = mask.reshape(-1) + idx = torch.full((H * W,), -1, dtype=torch.long) + n_valid = int(flat_mask.sum().item()) + idx[flat_mask] = torch.arange(n_valid, dtype=torch.long) + idx = idx.reshape(H, W) + + vertices = points.reshape(-1, 3)[flat_mask].contiguous() + + yy, xx = torch.meshgrid(torch.arange(H), torch.arange(W), indexing="ij") + u = xx.float() / max(W - 1, 1) + v = yy.float() / max(H - 1, 1) + uvs = torch.stack([u, v], dim=-1).reshape(-1, 2)[flat_mask].contiguous() + + a, b, c, d = idx[:-1, :-1], idx[:-1, 1:], idx[1:, 1:], idx[1:, :-1] + quad_ok = (a >= 0) & (b >= 0) & (c >= 0) & (d >= 0) + a, b, c, d = a[quad_ok], b[quad_ok], c[quad_ok], d[quad_ok] + faces = torch.cat([torch.stack([a, b, c], dim=-1), torch.stack([a, c, d], dim=-1)], dim=0).contiguous() + return vertices, faces, uvs diff --git a/comfy/ldm/moge/model.py b/comfy/ldm/moge/model.py new file mode 100644 index 000000000..6876c4af2 --- /dev/null +++ b/comfy/ldm/moge/model.py @@ -0,0 +1,347 @@ +"""MoGe v1 / v2 inference modules and a state-dict-driven builder. + +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 + +import torch +import torch.nn as nn +import torch.nn.functional as F + +import comfy.ops +import comfy.model_management +import comfy.model_patcher + +from comfy.image_encoders.dino2 import Dinov2Model + +from .geometry import depth_map_to_point_map, intrinsics_from_focal_center, recover_focal_shift +from .modules import ConvStack, DINOv2Encoder, HeadV1, MLP, _view_plane_uv_grid + + +def _remap_points(points: torch.Tensor) -> torch.Tensor: + """Apply the exp remap: z -> exp(z), xy stays linear and gets scaled by the new z.""" + xy, z = points.split([2, 1], dim=-1) + z = torch.exp(z) + return torch.cat([xy * z, z], dim=-1) + + +def _detect_dinov2(sd: dict, prefix: str) -> Dict[str, Any]: + # All shipped MoGe checkpoints use plain DINOv2 + hidden = sd[prefix + "embeddings.cls_token"].shape[-1] + layer_prefix = prefix + "encoder.layer." + depth = 1 + max(int(k[len(layer_prefix):].split(".")[0]) for k in sd if k.startswith(layer_prefix)) + return { + "hidden_size": hidden, + "num_attention_heads": hidden // 64, + "num_hidden_layers": depth, + "layer_norm_eps": 1e-6, + "use_swiglu_ffn": False, + } + + +class MoGeModelV1(nn.Module): + """MoGe v1: DINOv2 backbone + HeadV1 (points, mask).""" + + image_mean: torch.Tensor + image_std: torch.Tensor + + intermediate_layers = 4 + num_tokens_range: Tuple[Number, Number] = (1200, 2500) + mask_threshold = 0.5 + + def __init__(self, backbone: Dict[str, Any], dim_upsample: List[int] = (256, 128, 128), + num_res_blocks: int = 1, dim_times_res_block_hidden: int = 1, + dtype=None, device=None, operations=comfy.ops.manual_cast): + super().__init__() + self.backbone = Dinov2Model(backbone, dtype, device, operations) + self.head = HeadV1(dim_in=backbone["hidden_size"], dim_upsample=list(dim_upsample), + num_res_blocks=num_res_blocks, dim_times_res_block_hidden=dim_times_res_block_hidden, + dtype=dtype, device=device, operations=operations) + self.register_buffer("image_mean", torch.tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1)) + self.register_buffer("image_std", torch.tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1)) + + def forward(self, image: torch.Tensor, num_tokens: int) -> Dict[str, torch.Tensor]: + H, W = image.shape[-2:] + resize = ((num_tokens * 14 ** 2) / (H * W)) ** 0.5 + rh, rw = int(H * resize), int(W * resize) + x = F.interpolate(image, (rh, rw), mode="bicubic", align_corners=False, antialias=True) + x = (x - self.image_mean) / self.image_std + x14 = F.interpolate(x, (rh // 14 * 14, rw // 14 * 14), mode="bilinear", align_corners=False, antialias=True) + + n_layers = len(self.backbone.encoder.layer) + indices = list(range(n_layers - self.intermediate_layers, n_layers)) + feats = self.backbone.get_intermediate_layers(x14, indices, apply_norm=True) + + points, mask = self.head(feats, x) + points = F.interpolate(points.float(), (H, W), mode="bilinear", align_corners=False) + points = _remap_points(points.permute(0, 2, 3, 1)) + + mask = F.interpolate(mask.float(), (H, W), mode="bilinear", align_corners=False).squeeze(1) + + return {"points": points, "mask": mask} + + @classmethod + def from_state_dict(cls, sd, dtype=None, device=None, operations=comfy.ops.manual_cast): + """Detect the v1 head config from sd, build a model, and load weights.""" + n_up = 1 + max(int(k.split(".")[2]) for k in sd if k.startswith("head.upsample_blocks.")) + dim_upsample = [sd[f"head.upsample_blocks.{i}.0.0.weight"].shape[1] for i in range(n_up)] + # Each upsample stage is Sequential[upsampler, *res_blocks]; count res blocks at level 0. + num_res_blocks = max({int(k.split(".")[3]) for k in sd if k.startswith("head.upsample_blocks.0.")}) + hidden_out = sd["head.upsample_blocks.0.1.layers.2.weight"].shape[0] + dim_times = max(hidden_out // dim_upsample[0], 1) + model = cls(backbone=_detect_dinov2(sd, prefix="backbone."), + dim_upsample=dim_upsample, num_res_blocks=num_res_blocks, dim_times_res_block_hidden=dim_times, + dtype=dtype, device=device, operations=operations) + model.load_state_dict(sd, strict=True) + return model + + +class MoGeModelV2(nn.Module): + """MoGe v2: DINOv2 encoder + neck + per-output heads (points/mask/normal/metric-scale).""" + + intermediate_layers = 4 + num_tokens_range: Tuple[Number, Number] = (1200, 3600) + + def __init__(self, + encoder: Dict[str, Any], + neck: Dict[str, Any], + points_head: Dict[str, Any], + mask_head: Dict[str, Any], + scale_head: Dict[str, Any], + normal_head: Optional[Dict[str, Any]] = None, + dtype=None, device=None, operations=comfy.ops.manual_cast): + super().__init__() + self.encoder = DINOv2Encoder(**encoder, dtype=dtype, device=device, operations=operations) + self.neck = ConvStack(**neck, dtype=dtype, device=device, operations=operations) + self.points_head = ConvStack(**points_head, dtype=dtype, device=device, operations=operations) + self.mask_head = ConvStack(**mask_head, dtype=dtype, device=device, operations=operations) + self.scale_head = MLP(**scale_head, dtype=dtype, device=device, operations=operations) + if normal_head is not None: + self.normal_head = ConvStack(**normal_head, dtype=dtype, device=device, operations=operations) + + def forward(self, image: torch.Tensor, num_tokens: int) -> Dict[str, torch.Tensor]: + B, _, H, W = image.shape + device, dtype = image.device, image.dtype + aspect_ratio = W / H + base_h = round((num_tokens / aspect_ratio) ** 0.5) + base_w = round((num_tokens * aspect_ratio) ** 0.5) + + feat_top, cls_token = self.encoder(image, base_h, base_w, return_class_token=True) + + # 5-level pyramid: feat at level 0 concatenated with UV, other levels UV-only. + levels = [_view_plane_uv_grid(B, base_h * (2 ** L), base_w * (2 ** L), aspect_ratio, dtype, device) + for L in range(5)] + levels[0] = torch.cat([feat_top, levels[0]], dim=1) + + feats = self.neck(levels) + + def _resize(v): + return F.interpolate(v, (H, W), mode="bilinear", align_corners=False) + + points = _remap_points(_resize(self.points_head(feats)[-1]).permute(0, 2, 3, 1)) + mask = _resize(self.mask_head(feats)[-1]).squeeze(1).sigmoid() + metric_scale = self.scale_head(cls_token).squeeze(1).exp() + + result = {"points": points, "mask": mask, "metric_scale": metric_scale} + if hasattr(self, "normal_head"): + normal = _resize(self.normal_head(feats)[-1]) + result["normal"] = F.normalize(normal.permute(0, 2, 3, 1), dim=-1) + return result + + @classmethod + def from_state_dict(cls, sd, dtype=None, device=None, operations=comfy.ops.manual_cast): + """Detect the v2 encoder/neck/heads config from sd, build a model, and load weights.""" + backbone = _detect_dinov2(sd, prefix="encoder.backbone.") + depth = backbone["num_hidden_layers"] + n = cls.intermediate_layers + encoder = { + "backbone": backbone, + "intermediate_layers": [(depth // n) * (i + 1) - 1 for i in range(n)], + "dim_out": sd["encoder.output_projections.0.weight"].shape[0], + } + # scale_head is an MLP: Sequential of [Linear, ReLU, ..., Linear]; Linear weight is (out, in). + scale_idxs = sorted({int(k.split(".")[1]) for k in sd if k.startswith("scale_head.")}) + scale_first = sd[f"scale_head.{scale_idxs[0]}.weight"] + cfg: Dict[str, Any] = { + "encoder": encoder, + "neck": cls._detect_convstack(sd, "neck."), + "points_head": cls._detect_convstack(sd, "points_head."), + "mask_head": cls._detect_convstack(sd, "mask_head."), + "scale_head": {"dims": [scale_first.shape[1]] + [sd[f"scale_head.{i}.weight"].shape[0] for i in scale_idxs]}, + } + if any(k.startswith("normal_head.") for k in sd): + cfg["normal_head"] = cls._detect_convstack(sd, "normal_head.") + model = cls(**cfg, dtype=dtype, device=device, operations=operations) + model.load_state_dict(sd, strict=True) + return model + + @staticmethod + def _detect_convstack(sd: dict, prefix: str) -> Dict[str, Any]: + """Reconstruct a ConvStack config from the keys under prefix""" + in_keys = [k for k in sd if k.startswith(f"{prefix}input_blocks.") and k.endswith(".weight")] + n = 1 + max(int(k[len(f"{prefix}input_blocks."):].split(".")[0]) for k in in_keys) + + in_shapes = [sd[f"{prefix}input_blocks.{i}.weight"].shape for i in range(n)] + has_out = lambda i: f"{prefix}output_blocks.{i}.weight" in sd + has_norm = f"{prefix}res_blocks.0.0.layers.0.weight" in sd + + def num_res_at(i): + rb_prefix = f"{prefix}res_blocks.{i}." + return len({int(k[len(rb_prefix):].split(".")[0]) for k in sd if k.startswith(rb_prefix)}) + + return { + "dim_in": [s[1] for s in in_shapes], + "dim_res_blocks": [s[0] for s in in_shapes], + "dim_out": [sd[f"{prefix}output_blocks.{i}.weight"].shape[0] if has_out(i) else None for i in range(n)], + "num_res_blocks": [num_res_at(i) for i in range(n)], + "resamplers": ["conv_transpose" if f"{prefix}resamplers.{i}.0.weight" in sd else "bilinear" + for i in range(n - 1)], + "res_block_in_norm": "layer_norm" if has_norm else "none", + "res_block_hidden_norm": "group_norm" if has_norm else "none", + } + + +# Translate the Meta-style DINOv2 keys MoGe ships to the naming ComfyUI DINOv2 port expects, +# and split each fused qkv tensor into Q/K/V. +_DINOV2_TOPLEVEL_RENAMES = { + "patch_embed.proj.weight": "embeddings.patch_embeddings.projection.weight", + "patch_embed.proj.bias": "embeddings.patch_embeddings.projection.bias", + "cls_token": "embeddings.cls_token", + "pos_embed": "embeddings.position_embeddings", + "register_tokens": "embeddings.register_tokens", + "mask_token": "embeddings.mask_token", + "norm.weight": "layernorm.weight", + "norm.bias": "layernorm.bias", +} +_DINOV2_BLOCK_RENAMES = [ + ("ls1.gamma", "layer_scale1.lambda1"), + ("ls2.gamma", "layer_scale2.lambda1"), + ("attn.proj.", "attention.output.dense."), + ("mlp.w12.", "mlp.weights_in."), + ("mlp.w3.", "mlp.weights_out."), +] + + +def _remap_state_dict(sd: dict) -> dict: + if "model" in sd and "model_config" in sd: + sd = sd["model"] + prefix = "encoder.backbone." if any(k.startswith("encoder.backbone.") for k in sd) else "backbone." + out: dict = {} + for k, v in sd.items(): + if not k.startswith(prefix): + out[k] = v + continue + rel = k[len(prefix):] + if rel in _DINOV2_TOPLEVEL_RENAMES: + out[prefix + _DINOV2_TOPLEVEL_RENAMES[rel]] = v + continue + if not rel.startswith("blocks."): + out[k] = v + continue + _, idx, sub = rel.split(".", 2) + if sub in ("attn.qkv.weight", "attn.qkv.bias"): + tail = sub.rsplit(".", 1)[1] + q, kw, vw = v.chunk(3, dim=0) + base = f"{prefix}encoder.layer.{idx}.attention.attention" + out[f"{base}.query.{tail}"] = q + out[f"{base}.key.{tail}"] = kw + out[f"{base}.value.{tail}"] = vw + continue + for old, new in _DINOV2_BLOCK_RENAMES: + sub = sub.replace(old, new) + out[f"{prefix}encoder.layer.{idx}.{sub}"] = v + return out + + +def build_from_state_dict(sd: dict, dtype=None, device=None, operations=comfy.ops.manual_cast) -> nn.Module: + """Dispatch to v1 or v2 based on the DINOv2 backbone prefix.""" + sd = _remap_state_dict(sd) + cls = MoGeModelV2 if any(k.startswith("encoder.backbone.") for k in sd) else MoGeModelV1 + return cls.from_state_dict(sd, dtype=dtype, device=device, operations=operations) + + +class MoGeModel: + """Loaded MoGe model + ComfyUI memory management.""" + + def __init__(self, state_dict: dict): + # text encoder dtype closest match + self.load_device = comfy.model_management.text_encoder_device() + offload_device = comfy.model_management.text_encoder_offload_device() + self.dtype = comfy.model_management.text_encoder_dtype(self.load_device) + + self.model = build_from_state_dict(state_dict, dtype=self.dtype, device=offload_device, operations=comfy.ops.manual_cast).eval() + self.patcher = comfy.model_patcher.CoreModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device) + self.version = "v2" if hasattr(self.model, "encoder") else "v1" + self.mask_threshold = float(getattr(self.model, "mask_threshold", 0.5)) + nt = getattr(self.model, "num_tokens_range", (1200, 2500 if self.version == "v1" else 3600)) + self.num_tokens_range = (int(nt[0]), int(nt[1])) + + def infer(self, image: torch.Tensor, num_tokens: Optional[int] = None, + resolution_level: int = 9, fov_x: Optional[Union[Number, torch.Tensor]] = None, + force_projection: bool = True, apply_mask: bool = True, + apply_metric_scale: bool = True + ) -> Dict[str, torch.Tensor]: + """Run a single MoGe forward + post-process pass. image is (B, 3, H, W) in [0, 1].""" + comfy.model_management.load_model_gpu(self.patcher) + image = image.to(device=self.load_device, dtype=self.dtype) + H, W = image.shape[-2:] + aspect_ratio = W / H + + if num_tokens is None: + lo, hi = self.num_tokens_range + num_tokens = int(lo + (resolution_level / 9) * (hi - lo)) + + out = self.model.forward(image, num_tokens=num_tokens) + points = out["points"].float() # recover_focal_shift goes through scipy on CPU; needs fp32. + mask_binary = out["mask"] > self.mask_threshold + normal = out.get("normal") + metric_scale = out.get("metric_scale") + + diag = (1 + aspect_ratio ** 2) ** 0.5 + + def focal_from_fov_deg(deg): + fov = torch.as_tensor(deg, device=points.device, dtype=points.dtype) + return aspect_ratio / diag / torch.tan(torch.deg2rad(fov / 2)) + + if fov_x is None: + focal, shift = recover_focal_shift(points, mask_binary) + # Fall back to 60 deg FoV when the least-squares solver flips the focal sign. + bad = ~torch.isfinite(focal) | (focal <= 0) + if bool(bad.any()): + focal = torch.where(bad, focal_from_fov_deg(60.0), focal) + _, shift = recover_focal_shift(points, mask_binary, focal=focal) + else: + focal = focal_from_fov_deg(fov_x).expand(points.shape[0]) + _, shift = recover_focal_shift(points, mask_binary, focal=focal) + + f_diag = focal / 2 * diag + half = torch.tensor(0.5, device=points.device, dtype=points.dtype) + intrinsics = intrinsics_from_focal_center(f_diag / aspect_ratio, f_diag, half, half) + points[..., 2] = points[..., 2] + shift[..., None, None] + # v2 only: filter mask by depth>0 to drop metric-scale negative-depth artifacts. + if self.version == "v2": + mask_binary = mask_binary & (points[..., 2] > 0) + depth = points[..., 2].clone() + + if force_projection: + points = depth_map_to_point_map(depth, intrinsics=intrinsics) + + if apply_metric_scale and metric_scale is not None: + points = points * metric_scale[:, None, None, None] + depth = depth * metric_scale[:, None, None] + + if apply_mask: + points = torch.where(mask_binary[..., None], points, torch.full_like(points, float("inf"))) + depth = torch.where(mask_binary, depth, torch.full_like(depth, float("inf"))) + if normal is not None: + normal = torch.where(mask_binary[..., None], normal, torch.zeros_like(normal)) + + result = {"points": points, "depth": depth, "intrinsics": intrinsics, "mask": mask_binary} + if normal is not None: + result["normal"] = normal + return result diff --git a/comfy/ldm/moge/modules.py b/comfy/ldm/moge/modules.py new file mode 100644 index 000000000..235a59212 --- /dev/null +++ b/comfy/ldm/moge/modules.py @@ -0,0 +1,204 @@ +"""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 + +import torch +import torch.nn as nn +import torch.nn.functional as F + +import comfy.ops +from comfy.image_encoders.dino2 import Dinov2Model + +from .geometry import normalized_view_plane_uv + + +def _conv2d(operations, c_in: int, c_out: int, k: int = 3, *, dtype=None, device=None): + return operations.Conv2d(c_in, c_out, kernel_size=k, padding=k // 2, padding_mode="replicate", dtype=dtype, device=device) + + +def _view_plane_uv_grid(batch: int, height: int, width: int, aspect_ratio: float, dtype, device) -> torch.Tensor: + """Batched normalized view-plane UV grid as a (B, 2, H, W) tensor.""" + uv = normalized_view_plane_uv(width, height, aspect_ratio=aspect_ratio, dtype=dtype, device=device) + return uv.permute(2, 0, 1).unsqueeze(0).expand(batch, -1, -1, -1) + + +def _concat_view_plane_uv(x: torch.Tensor, aspect_ratio: float) -> torch.Tensor: + """Append a 2-channel normalized view-plane UV grid to x along the channel dim.""" + uv = _view_plane_uv_grid(x.shape[0], x.shape[-2], x.shape[-1], aspect_ratio, x.dtype, x.device) + return torch.cat([x, uv], dim=1) + + +class ResidualConvBlock(nn.Module): + def __init__(self, channels: int, hidden_channels: Optional[int] = None, in_norm: str = "layer_norm", hidden_norm: str = "group_norm", + dtype=None, device=None, operations=comfy.ops.manual_cast): + super().__init__() + hidden_channels = hidden_channels if hidden_channels is not None else channels + + in_norm_layer = operations.GroupNorm(1, channels, dtype=dtype, device=device) if in_norm == "layer_norm" else nn.Identity() + hidden_norm_layer = (operations.GroupNorm(max(hidden_channels // 32, 1), hidden_channels, dtype=dtype, device=device) + if hidden_norm == "group_norm" else nn.Identity()) + + self.layers = nn.Sequential( + in_norm_layer, nn.ReLU(), _conv2d(operations, channels, hidden_channels, dtype=dtype, device=device), + hidden_norm_layer, nn.ReLU(), _conv2d(operations, hidden_channels, channels, dtype=dtype, device=device), + ) + + def forward(self, x): + return self.layers(x) + x + + +class Resampler(nn.Sequential): + """2x upsampler: ConvTranspose2d(2x2) or bilinear upsample, followed by a 3x3 conv.""" + + def __init__(self, in_channels: int, out_channels: int, type_: str, dtype=None, device=None, operations=comfy.ops.manual_cast): + if type_ == "conv_transpose": + up = operations.ConvTranspose2d(in_channels, out_channels, kernel_size=2, stride=2, dtype=dtype, device=device) + conv_in = out_channels + else: # "bilinear" + up = nn.Upsample(scale_factor=2, mode="bilinear", align_corners=False) + conv_in = in_channels + super().__init__(up, _conv2d(operations, conv_in, out_channels, dtype=dtype, device=device)) + + +class MLP(nn.Sequential): + def __init__(self, dims: Sequence[int], dtype=None, device=None, operations=comfy.ops.manual_cast): + layers = [] + for d_in, d_out in zip(dims[:-2], dims[1:-1]): + layers.append(operations.Linear(d_in, d_out, dtype=dtype, device=device)) + layers.append(nn.ReLU(inplace=True)) + layers.append(operations.Linear(dims[-2], dims[-1], dtype=dtype, device=device)) + super().__init__(*layers) + + +class ConvStack(nn.Module): + def __init__(self, dim_in: List[Optional[int]], dim_res_blocks: List[int], dim_out: List[Optional[int]], resamplers: List[str], + num_res_blocks: List[int], dim_times_res_block_hidden: int = 1, res_block_in_norm: str = "layer_norm", res_block_hidden_norm: str = "group_norm", + dtype=None, device=None, operations=comfy.ops.manual_cast): + super().__init__() + + self.input_blocks = nn.ModuleList([ + (_conv2d(operations, d_in, d_res, k=1, dtype=dtype, device=device) + if d_in is not None else nn.Identity()) + for d_in, d_res in zip(dim_in, dim_res_blocks) + ]) + + self.resamplers = nn.ModuleList([ + Resampler(prev, succ, type_=r, dtype=dtype, device=device, operations=operations) + for prev, succ, r in zip(dim_res_blocks[:-1], dim_res_blocks[1:], resamplers) + ]) + + self.res_blocks = nn.ModuleList([ + nn.Sequential(*[ + ResidualConvBlock(d_res, dim_times_res_block_hidden * d_res, in_norm=res_block_in_norm, hidden_norm=res_block_hidden_norm, dtype=dtype, device=device, operations=operations) + for _ in range(num_res_blocks[i]) + ]) + for i, d_res in enumerate(dim_res_blocks) + ]) + + self.output_blocks = nn.ModuleList([ + (_conv2d(operations, d_res, d_out, k=1, dtype=dtype, device=device) + if d_out is not None else nn.Identity()) + for d_out, d_res in zip(dim_out, dim_res_blocks) + ]) + + def forward(self, in_features: List[Optional[torch.Tensor]]): + out_features = [] + x = None + for i in range(len(self.res_blocks)): + feat = self.input_blocks[i](in_features[i]) if in_features[i] is not None else None + if i == 0: + x = feat + elif feat is not None: + x = x + feat + x = self.res_blocks[i](x) + out_features.append(self.output_blocks[i](x)) + if i < len(self.res_blocks) - 1: + x = self.resamplers[i](x) + return out_features + + +class DINOv2Encoder(nn.Module): + """Comfy DINOv2 backbone with per-layer 1x1 projection heads.""" + + def __init__(self, backbone: dict, intermediate_layers: List[int], dim_out: int, dtype=None, device=None, operations=comfy.ops.manual_cast): + super().__init__() + self.intermediate_layers = list(intermediate_layers) + dim_features = backbone["hidden_size"] + self.backbone = Dinov2Model(backbone, dtype, device, operations) + self.output_projections = nn.ModuleList([ + _conv2d(operations, dim_features, dim_out, k=1, dtype=dtype, device=device) + for _ in range(len(self.intermediate_layers)) + ]) + self.register_buffer("image_mean", torch.tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1)) + self.register_buffer("image_std", torch.tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1)) + + def forward(self, image: torch.Tensor, token_rows: int, token_cols: int, + return_class_token: bool = False) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]: + image_14 = F.interpolate(image, (token_rows * 14, token_cols * 14), mode="bilinear", align_corners=False, antialias=True) + image_14 = (image_14 - self.image_mean) / self.image_std + feats = self.backbone.get_intermediate_layers(image_14, self.intermediate_layers, apply_norm=True) + x = torch.stack([ + proj(feat.permute(0, 2, 1).unflatten(2, (token_rows, token_cols)).contiguous()) + for proj, (feat, _cls) in zip(self.output_projections, feats) + ], dim=1).sum(dim=1) + if return_class_token: + return x, feats[-1][1] + return x + + +class HeadV1(nn.Module): + """v1 head: 4 backbone-feature projections -> shared upsample stack -> per-target output convs (points, mask).""" + + NUM_FEATURES = 4 + DIM_PROJ = 512 + DIM_OUT = (3, 1) # 3 channels for points, 1 for mask + LAST_CONV_CHANNELS = 32 + + def __init__(self, dim_in: int, dim_upsample: List[int] = (256, 128, 128), num_res_blocks: int = 1, dim_times_res_block_hidden: int = 1, + dtype=None, device=None, operations=comfy.ops.manual_cast): + super().__init__() + self.projects = nn.ModuleList([ + _conv2d(operations, dim_in, self.DIM_PROJ, k=1, dtype=dtype, device=device) + for _ in range(self.NUM_FEATURES) + ]) + def upsampler(in_ch, out_ch): + return nn.Sequential( + operations.ConvTranspose2d(in_ch, out_ch, kernel_size=2, stride=2, dtype=dtype, device=device), + _conv2d(operations, out_ch, out_ch, dtype=dtype, device=device), + ) + + in_chs = [self.DIM_PROJ] + list(dim_upsample[:-1]) + self.upsample_blocks = nn.ModuleList([ + nn.Sequential( + upsampler(in_ch + 2, out_ch), + *(ResidualConvBlock(out_ch, dim_times_res_block_hidden * out_ch, dtype=dtype, device=device, operations=operations) + for _ in range(num_res_blocks)) + ) + for in_ch, out_ch in zip(in_chs, dim_upsample) + ]) + self.output_block = nn.ModuleList([ + nn.Sequential( + _conv2d(operations, dim_upsample[-1] + 2, self.LAST_CONV_CHANNELS, dtype=dtype, device=device), + nn.ReLU(inplace=True), + _conv2d(operations, self.LAST_CONV_CHANNELS, d_out, k=1, dtype=dtype, device=device), + ) + for d_out in self.DIM_OUT + ]) + + def forward(self, hidden_states, image: torch.Tensor): + img_h, img_w = image.shape[-2:] + patch_h, patch_w = img_h // 14, img_w // 14 + aspect = img_w / img_h + x = torch.stack([ + proj(feat.permute(0, 2, 1).unflatten(2, (patch_h, patch_w)).contiguous()) + for proj, (feat, _cls) in zip(self.projects, hidden_states) + ], dim=1).sum(dim=1) + + for block in self.upsample_blocks: + x = block(_concat_view_plane_uv(x, aspect)) + + x = F.interpolate(x, (img_h, img_w), mode="bilinear", align_corners=False) + x = _concat_view_plane_uv(x, aspect) + return [block(x) for block in self.output_block] diff --git a/comfy/ldm/moge/panorama.py b/comfy/ldm/moge/panorama.py new file mode 100644 index 000000000..de53ebe68 --- /dev/null +++ b/comfy/ldm/moge/panorama.py @@ -0,0 +1,313 @@ +"""Panorama (equirectangular) inference helpers for MoGe. + +Splits an equirect into 12 perspective views via an icosahedron camera rig, runs +the model per view, and stitches per-view distance maps back into a single +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 + +import numpy as np +import torch +import torch.nn.functional as F + +from scipy.ndimage import convolve, map_coordinates +from scipy.sparse import vstack, csr_array +from scipy.sparse.linalg import lsmr + + +def _icosahedron_directions() -> np.ndarray: + """12 icosahedron-vertex directions (non-normalised, matching upstream's vertex order).""" + A = (1.0 + np.sqrt(5.0)) / 2.0 + return np.array([ + [0, 1, A], [0, -1, A], [0, 1, -A], [0, -1, -A], + [1, A, 0], [-1, A, 0], [1, -A, 0], [-1, -A, 0], + [A, 0, 1], [A, 0, -1], [-A, 0, 1], [-A, 0, -1], + ], dtype=np.float32) + + +def _intrinsics_from_fov(fov_x_rad: float, fov_y_rad: float) -> np.ndarray: + """Normalised-image (unit-square) K matrix.""" + fx = 0.5 / np.tan(fov_x_rad / 2) + fy = 0.5 / np.tan(fov_y_rad / 2) + return np.array([[fx, 0, 0.5], [0, fy, 0.5], [0, 0, 1]], dtype=np.float32) + + +def _extrinsics_look_at(eye: np.ndarray, target: np.ndarray, up: np.ndarray) -> np.ndarray: + """OpenCV-convention world->camera extrinsics for an array of look-at targets (N, 4, 4).""" + eye = np.asarray(eye, dtype=np.float32) + target = np.asarray(target, dtype=np.float32) + up = np.asarray(up, dtype=np.float32) + if target.ndim == 1: + target = target[None] + + fwd = target - eye + fwd = fwd / np.linalg.norm(fwd, axis=-1, keepdims=True).clip(1e-12) + right = np.cross(fwd, up) + right_norm = np.linalg.norm(right, axis=-1, keepdims=True) + # Fall back to an arbitrary perpendicular if forward is parallel to up. + parallel = right_norm.squeeze(-1) < 1e-6 + if parallel.any(): + alt_up = np.array([1, 0, 0], dtype=np.float32) + right = np.where(parallel[:, None], np.cross(fwd, alt_up), right) + right_norm = np.linalg.norm(right, axis=-1, keepdims=True) + right = right / right_norm.clip(1e-12) + new_up = np.cross(fwd, right) + + R = np.stack([right, new_up, fwd], axis=-2) + t = -np.einsum("nij,j->ni", R, eye) + E = np.zeros((R.shape[0], 4, 4), dtype=np.float32) + E[:, :3, :3] = R + E[:, :3, 3] = t + E[:, 3, 3] = 1.0 + return E + + +def get_panorama_cameras() -> Tuple[np.ndarray, List[np.ndarray]]: + """Returns (extrinsics (12, 4, 4), [intrinsics] * 12) for icosahedron views at 90 deg FoV.""" + targets = _icosahedron_directions() + eye = np.zeros(3, dtype=np.float32) + up = np.array([0, 0, 1], dtype=np.float32) + extrinsics = _extrinsics_look_at(eye, targets, up) + K = _intrinsics_from_fov(np.deg2rad(90.0), np.deg2rad(90.0)) + return extrinsics, [K] * len(targets) + + +def spherical_uv_to_directions(uv: np.ndarray) -> np.ndarray: + """Equirect UV in [0, 1] -> 3D unit-direction (Z up).""" + theta = (1 - uv[..., 0]) * (2 * np.pi) + phi = uv[..., 1] * np.pi + return np.stack([ + np.sin(phi) * np.cos(theta), + np.sin(phi) * np.sin(theta), + np.cos(phi), + ], axis=-1).astype(np.float32) + + +def directions_to_spherical_uv(directions: np.ndarray) -> np.ndarray: + """3D direction -> equirect UV in [0, 1].""" + n = np.linalg.norm(directions, axis=-1, keepdims=True).clip(1e-12) + d = directions / n + u = 1 - np.arctan2(d[..., 1], d[..., 0]) / (2 * np.pi) % 1.0 + v = np.arccos(d[..., 2].clip(-1, 1)) / np.pi + return np.stack([u, v], axis=-1).astype(np.float32) + + +def _uv_grid(H: int, W: int) -> np.ndarray: + """Pixel-center UV grid in [0, 1]; (H, W, 2).""" + u = (np.arange(W, dtype=np.float32) + 0.5) / W + v = (np.arange(H, dtype=np.float32) + 0.5) / H + return np.stack(np.meshgrid(u, v, indexing="xy"), axis=-1) + + +def _unproject_cv(uv: np.ndarray, depth: np.ndarray, + extrinsics: np.ndarray, intrinsics: np.ndarray) -> np.ndarray: + """Back-project pixels into world coords (OpenCV convention).""" + pix = np.concatenate([uv, np.ones_like(uv[..., :1])], axis=-1) + K_inv = np.linalg.inv(intrinsics) + cam = pix @ K_inv.T * depth[..., None] + cam_h = np.concatenate([cam, np.ones_like(cam[..., :1])], axis=-1) + E_inv = np.linalg.inv(extrinsics) + return (cam_h @ E_inv.T)[..., :3] + + +def _project_cv(points: np.ndarray, extrinsics: np.ndarray, intrinsics: np.ndarray) -> Tuple[np.ndarray, np.ndarray]: + """World coords -> (uv, depth) in the camera (OpenCV convention).""" + pts_h = np.concatenate([points, np.ones_like(points[..., :1])], axis=-1) + cam = pts_h @ extrinsics.T + cam_xyz = cam[..., :3] + depth = cam_xyz[..., 2] + proj = cam_xyz @ intrinsics.T + uv = proj[..., :2] / proj[..., 2:3].clip(1e-12) + return uv.astype(np.float32), depth.astype(np.float32) + + +def _grid_sample_uv(img_bchw: torch.Tensor, uv: torch.Tensor, mode: str = "bilinear") -> torch.Tensor: + """Sample img_bchw at UV-in-[0,1] coords uv of shape (B, H, W, 2); replicate-border.""" + grid = uv * 2.0 - 1.0 + return F.grid_sample(img_bchw, grid, mode=mode, padding_mode="border", align_corners=False) + + +def split_panorama_image(image: torch.Tensor, extrinsics: np.ndarray, intrinsics: List[np.ndarray], resolution: int) -> torch.Tensor: + """(3, Hp, Wp) equirect on any device -> (N, 3, R, R) perspective crops on the same device.""" + device = image.device + N = len(extrinsics) + uv = _uv_grid(resolution, resolution) + sample_uvs = [] + for i in range(N): + world = _unproject_cv(uv, np.ones(uv.shape[:-1], dtype=np.float32), extrinsics[i], intrinsics[i]) + sample_uvs.append(directions_to_spherical_uv(world)) + sample_uvs = np.stack(sample_uvs, axis=0) + + img_bchw = image.unsqueeze(0).expand(N, -1, -1, -1).contiguous() + sample_uvs_t = torch.from_numpy(sample_uvs).to(device=device, dtype=image.dtype) + return _grid_sample_uv(img_bchw, sample_uvs_t, mode="bilinear") + + +def _poisson_equation(W: int, H: int, wrap_x: bool = False, wrap_y: bool = False): + """Sparse Laplacian operator over the H x W grid.""" + grid_index = np.arange(H * W).reshape(H, W) + grid_index = np.pad(grid_index, ((0, 0), (1, 1)), mode="wrap" if wrap_x else "edge") + grid_index = np.pad(grid_index, ((1, 1), (0, 0)), mode="wrap" if wrap_y else "edge") + + data = np.array([[-4, 1, 1, 1, 1]], dtype=np.float32).repeat(H * W, axis=0).reshape(-1) + indices = np.stack([ + grid_index[1:-1, 1:-1], + grid_index[:-2, 1:-1], grid_index[2:, 1:-1], + grid_index[1:-1, :-2], grid_index[1:-1, 2:], + ], axis=-1).reshape(-1) + indptr = np.arange(0, H * W * 5 + 1, 5) + return csr_array((data, indices, indptr), shape=(H * W, H * W)) + + +def _grad_equation(W: int, H: int, wrap_x: bool = False, wrap_y: bool = False): + """Sparse forward-difference operator over the H x W grid.""" + grid_index = np.arange(W * H).reshape(H, W) + if wrap_x: + grid_index = np.pad(grid_index, ((0, 0), (0, 1)), mode="wrap") + if wrap_y: + grid_index = np.pad(grid_index, ((0, 1), (0, 0)), mode="wrap") + + data = np.concatenate([ + np.concatenate([ + np.ones((grid_index.shape[0], grid_index.shape[1] - 1), dtype=np.float32).reshape(-1, 1), + -np.ones((grid_index.shape[0], grid_index.shape[1] - 1), dtype=np.float32).reshape(-1, 1), + ], axis=1).reshape(-1), + np.concatenate([ + np.ones((grid_index.shape[0] - 1, grid_index.shape[1]), dtype=np.float32).reshape(-1, 1), + -np.ones((grid_index.shape[0] - 1, grid_index.shape[1]), dtype=np.float32).reshape(-1, 1), + ], axis=1).reshape(-1), + ]) + indices = np.concatenate([ + np.concatenate([grid_index[:, :-1].reshape(-1, 1), grid_index[:, 1:].reshape(-1, 1)], axis=1).reshape(-1), + np.concatenate([grid_index[:-1, :].reshape(-1, 1), grid_index[1:, :].reshape(-1, 1)], axis=1).reshape(-1), + ]) + nx = grid_index.shape[0] * (grid_index.shape[1] - 1) + ny = (grid_index.shape[0] - 1) * grid_index.shape[1] + indptr = np.arange(0, nx * 2 + ny * 2 + 1, 2) + return csr_array((data, indices, indptr), shape=(nx + ny, H * W)) + + +def _scipy_remap_bilinear(img: np.ndarray, sample_pixels: np.ndarray, mode: str = "bilinear") -> np.ndarray: + """Bilinear/nearest sampling at fractional pixel coords; out-of-range clamps to nearest border.""" + H, W = img.shape[:2] + yy = np.clip(sample_pixels[..., 1], 0, H - 1) + xx = np.clip(sample_pixels[..., 0], 0, W - 1) + order = 1 if mode == "bilinear" else 0 + if img.ndim == 2: + return map_coordinates(img, [yy, xx], order=order, mode="nearest").astype(img.dtype) + out = np.stack([ + map_coordinates(img[..., c], [yy, xx], order=order, mode="nearest") + for c in range(img.shape[-1]) + ], axis=-1) + return out.astype(img.dtype) + + +def merge_panorama_depth(width: int, height: int, + distance_maps: List[np.ndarray], pred_masks: List[np.ndarray], + extrinsics: List[np.ndarray], intrinsics: List[np.ndarray], + on_view: Optional[Callable[[], None]] = None, + on_solve_start: Optional[Callable[[int, int], None]] = None, + on_solve_end: Optional[Callable[[int, int], None]] = None, + ) -> Tuple[np.ndarray, np.ndarray]: + """Stitch per-view distance maps into a single equirect distance map. + + Recursive multi-scale solve: solves at half resolution first and uses that as the lsmr init + for the full-resolution solve. Optional callbacks fire per view processed and around each + lsmr solve so callers can drive a progress bar. + """ + + if max(width, height) > 256: + coarse_depth, _ = merge_panorama_depth(width // 2, height // 2, + distance_maps, pred_masks, extrinsics, intrinsics, + on_view=on_view, + on_solve_start=on_solve_start, + on_solve_end=on_solve_end) + t = torch.from_numpy(coarse_depth).unsqueeze(0).unsqueeze(0) + t = F.interpolate(t, size=(height, width), mode="bilinear", align_corners=False) + depth_init = t.squeeze().numpy().astype(np.float32) + else: + depth_init = None + + spherical_directions = spherical_uv_to_directions(_uv_grid(height, width)) + + pano_log_grad_maps, pano_grad_masks = [], [] + pano_log_lap_maps, pano_lap_masks = [], [] + pano_pred_masks: List[np.ndarray] = [] + + for i in range(len(distance_maps)): + proj_uv, proj_depth = _project_cv(spherical_directions, extrinsics[i], intrinsics[i]) + proj_valid = (proj_depth > 0) & (proj_uv > 0).all(axis=-1) & (proj_uv < 1).all(axis=-1) + + Hd, Wd = distance_maps[i].shape[:2] + proj_pixels = np.clip(proj_uv, 0, 1) * np.array([Wd - 1, Hd - 1], dtype=np.float32) + + log_dist = np.log(np.clip(distance_maps[i], 1e-6, None)) + sampled = _scipy_remap_bilinear(log_dist, proj_pixels, mode="bilinear") + pano_log = np.where(proj_valid, sampled, 0.0).astype(np.float32) + + sampled_mask = _scipy_remap_bilinear(pred_masks[i].astype(np.uint8), proj_pixels, mode="nearest") + pano_pred = proj_valid & (sampled_mask > 0) + + # Equirect wraps horizontally but not vertically: wrap pad along x, edge pad along y. + padded = np.pad(pano_log, ((0, 0), (0, 1)), mode="wrap") + gx, gy = padded[:, :-1] - padded[:, 1:], padded[:-1, :] - padded[1:, :] + padded_m = np.pad(pano_pred, ((0, 0), (0, 1)), mode="wrap") + mx, my = padded_m[:, :-1] & padded_m[:, 1:], padded_m[:-1, :] & padded_m[1:, :] + pano_log_grad_maps.append((gx, gy)) + pano_grad_masks.append((mx, my)) + + padded = np.pad(pano_log, ((1, 1), (0, 0)), mode="edge") + padded = np.pad(padded, ((0, 0), (1, 1)), mode="wrap") + lap_kernel = np.array([[0, 1, 0], [1, -4, 1], [0, 1, 0]], dtype=np.float32) + lap = convolve(padded, lap_kernel)[1:-1, 1:-1] + padded_m = np.pad(pano_pred, ((1, 1), (0, 0)), mode="edge") + padded_m = np.pad(padded_m, ((0, 0), (1, 1)), mode="wrap") + m_kernel = np.array([[0, 1, 0], [1, 1, 1], [0, 1, 0]], dtype=np.uint8) + lap_mask = convolve(padded_m.astype(np.uint8), m_kernel)[1:-1, 1:-1] == 5 + pano_log_lap_maps.append(lap) + pano_lap_masks.append(lap_mask) + pano_pred_masks.append(pano_pred) + + if on_view is not None: + on_view() + + gx = np.stack([m[0] for m in pano_log_grad_maps], axis=0) + gy = np.stack([m[1] for m in pano_log_grad_maps], axis=0) + mx = np.stack([m[0] for m in pano_grad_masks], axis=0) + my = np.stack([m[1] for m in pano_grad_masks], axis=0) + gx_avg = (gx * mx).sum(axis=0) / mx.sum(axis=0).clip(1e-3) + gy_avg = (gy * my).sum(axis=0) / my.sum(axis=0).clip(1e-3) + + laps = np.stack(pano_log_lap_maps, axis=0) + lap_masks = np.stack(pano_lap_masks, axis=0) + lap_avg = (laps * lap_masks).sum(axis=0) / lap_masks.sum(axis=0).clip(1e-3) + + grad_x_mask = mx.any(axis=0).reshape(-1) + grad_y_mask = my.any(axis=0).reshape(-1) + grad_mask = np.concatenate([grad_x_mask, grad_y_mask]) + lap_mask_flat = lap_masks.any(axis=0).reshape(-1) + + A = vstack([ + _grad_equation(width, height, wrap_x=True, wrap_y=False)[grad_mask], + _poisson_equation(width, height, wrap_x=True, wrap_y=False)[lap_mask_flat], + ]) + b = np.concatenate([ + gx_avg.reshape(-1)[grad_x_mask], + gy_avg.reshape(-1)[grad_y_mask], + lap_avg.reshape(-1)[lap_mask_flat], + ]) + x0 = np.log(np.clip(depth_init, 1e-6, None)).reshape(-1) if depth_init is not None else None + + if on_solve_start is not None: + on_solve_start(width, height) + x, *_ = lsmr(A, b, atol=1e-5, btol=1e-5, x0=x0, show=False) + if on_solve_end is not None: + on_solve_end(width, height) + + pano_depth = np.exp(x).reshape(height, width).astype(np.float32) + pano_mask = np.any(pano_pred_masks, axis=0) + return pano_depth, pano_mask diff --git a/comfy/ldm/sam3/detector.py b/comfy/ldm/sam3/detector.py index 12d3a01ab..23a972ac7 100644 --- a/comfy/ldm/sam3/detector.py +++ b/comfy/ldm/sam3/detector.py @@ -561,7 +561,8 @@ class SAM3Model(nn.Module): return high_res_masks def forward_video(self, images, initial_masks, pbar=None, text_prompts=None, - new_det_thresh=0.5, max_objects=0, detect_interval=1): + new_det_thresh=0.5, max_objects=0, detect_interval=1, + target_device=None, target_dtype=None): """Track video with optional per-frame text-prompted detection.""" bb = self.detector.backbone["vision_backbone"] @@ -589,8 +590,10 @@ class SAM3Model(nn.Module): return self.tracker.track_video_with_detection( backbone_fn, images, initial_masks, detect_fn, new_det_thresh=new_det_thresh, max_objects=max_objects, - detect_interval=detect_interval, backbone_obj=bb, pbar=pbar) + detect_interval=detect_interval, backbone_obj=bb, pbar=pbar, + target_device=target_device, target_dtype=target_dtype) # SAM3 (non-multiplex) — no detection support, requires initial masks if initial_masks is None: raise ValueError("SAM3 (non-multiplex) requires initial_mask for video tracking") - return self.tracker.track_video(backbone_fn, images, initial_masks, pbar=pbar, backbone_obj=bb) + return self.tracker.track_video(backbone_fn, images, initial_masks, pbar=pbar, backbone_obj=bb, + target_device=target_device, target_dtype=target_dtype) diff --git a/comfy/ldm/sam3/tracker.py b/comfy/ldm/sam3/tracker.py index 8f7481003..8456e90a6 100644 --- a/comfy/ldm/sam3/tracker.py +++ b/comfy/ldm/sam3/tracker.py @@ -200,8 +200,13 @@ def pack_masks(masks): def unpack_masks(packed): """Unpack bit-packed [*, H, W//8] uint8 to bool [*, H, W*8].""" - shifts = torch.arange(8, device=packed.device) - return ((packed.unsqueeze(-1) >> shifts) & 1).view(*packed.shape[:-1], -1).bool() + bits = torch.tensor([1, 2, 4, 8, 16, 32, 64, 128], dtype=torch.uint8, device=packed.device) + return (packed.unsqueeze(-1) & bits).bool().view(*packed.shape[:-1], -1) + + +def _prep_frame(images, idx, device, dt, size): + """Slice CPU full-res frames, transfer to GPU in target dtype, and resize to (size, size).""" + return comfy.utils.common_upscale(images[idx].to(device=device, dtype=dt), size, size, "bicubic", crop="disabled") def _compute_backbone(backbone_fn, frame, frame_idx=None): @@ -1078,16 +1083,19 @@ class SAM3Tracker(nn.Module): # SAM3: drop last FPN level return vision_feats[:-1], vision_pos[:-1], feat_sizes[:-1] - def _track_single_object(self, backbone_fn, images, initial_mask, pbar=None): + def _track_single_object(self, backbone_fn, images, initial_mask, pbar=None, + target_device=None, target_dtype=None): """Track one object, computing backbone per frame to save VRAM.""" N = images.shape[0] - device, dt = images.device, images.dtype + device = target_device if target_device is not None else images.device + dt = target_dtype if target_dtype is not None else images.dtype + size = self.image_size output_dict = {"cond_frame_outputs": {}, "non_cond_frame_outputs": {}} all_masks = [] for frame_idx in tqdm(range(N), desc="tracking"): vision_feats, vision_pos, feat_sizes = self._compute_backbone_frame( - backbone_fn, images[frame_idx:frame_idx + 1], frame_idx=frame_idx) + backbone_fn, _prep_frame(images, slice(frame_idx, frame_idx + 1), device, dt, size), frame_idx=frame_idx) mask_input = None if frame_idx == 0: mask_input = F.interpolate(initial_mask.to(device=device, dtype=dt), @@ -1114,12 +1122,13 @@ class SAM3Tracker(nn.Module): return torch.cat(all_masks, dim=0) # [N, 1, H, W] - def track_video(self, backbone_fn, images, initial_masks, pbar=None, **kwargs): + def track_video(self, backbone_fn, images, initial_masks, pbar=None, + target_device=None, target_dtype=None, **kwargs): """Track one or more objects across video frames. Args: backbone_fn: callable that returns (sam2_features, sam2_positions, trunk_out) for a frame - images: [N, 3, 1008, 1008] video frames + images: [N, 3, H, W] CPU full-res video frames (resized per-frame to self.image_size) initial_masks: [N_obj, 1, H, W] binary masks for first frame (one per object) pbar: optional progress bar @@ -1130,7 +1139,8 @@ class SAM3Tracker(nn.Module): per_object = [] for obj_idx in range(N_obj): obj_masks = self._track_single_object( - backbone_fn, images, initial_masks[obj_idx:obj_idx + 1], pbar=pbar) + backbone_fn, images, initial_masks[obj_idx:obj_idx + 1], pbar=pbar, + target_device=target_device, target_dtype=target_dtype) per_object.append(obj_masks) return torch.cat(per_object, dim=1) # [N, N_obj, H, W] @@ -1632,11 +1642,18 @@ class SAM31Tracker(nn.Module): return det_scores[new_dets].tolist() if det_scores is not None else [0.0] * new_dets.sum().item() return [] + INTERNAL_MAX_OBJECTS = 64 # Hard ceiling on accumulated tracks; max_objects=0 or any value above this is clamped here. + def track_video_with_detection(self, backbone_fn, images, initial_masks, detect_fn=None, new_det_thresh=0.5, max_objects=0, detect_interval=1, - backbone_obj=None, pbar=None): + backbone_obj=None, pbar=None, target_device=None, target_dtype=None): """Track with optional per-frame detection. Returns [N, max_N_obj, H, W] mask logits.""" - N, device, dt = images.shape[0], images.device, images.dtype + if max_objects <= 0 or max_objects > self.INTERNAL_MAX_OBJECTS: + max_objects = self.INTERNAL_MAX_OBJECTS + N = images.shape[0] + device = target_device if target_device is not None else images.device + dt = target_dtype if target_dtype is not None else images.dtype + size = self.image_size output_dict = {"cond_frame_outputs": {}, "non_cond_frame_outputs": {}} all_masks = [] idev = comfy.model_management.intermediate_device() @@ -1656,7 +1673,7 @@ class SAM31Tracker(nn.Module): prefetch = True except RuntimeError: pass - cur_bb = self._compute_backbone_frame(backbone_fn, images[0:1], frame_idx=0) + cur_bb = self._compute_backbone_frame(backbone_fn, _prep_frame(images, slice(0, 1), device, dt, size), frame_idx=0) for frame_idx in tqdm(range(N), desc="tracking"): vision_feats, vision_pos, feat_sizes, high_res_prop, trunk_out = cur_bb @@ -1666,7 +1683,7 @@ class SAM31Tracker(nn.Module): backbone_stream.wait_stream(torch.cuda.current_stream(device)) with torch.cuda.stream(backbone_stream): next_bb = self._compute_backbone_frame( - backbone_fn, images[frame_idx + 1:frame_idx + 2], frame_idx=frame_idx + 1) + backbone_fn, _prep_frame(images, slice(frame_idx + 1, frame_idx + 2), device, dt, size), frame_idx=frame_idx + 1) # Per-frame detection with NMS (skip if no detect_fn, or interval/max not met) det_masks = torch.empty(0, device=device) @@ -1687,7 +1704,7 @@ class SAM31Tracker(nn.Module): current_out = self._condition_with_masks( initial_masks.to(device=device, dtype=dt), frame_idx, vision_feats, vision_pos, feat_sizes, high_res_prop, output_dict, N, mux_state, backbone_obj, - images[frame_idx:frame_idx + 1], trunk_out) + _prep_frame(images, slice(frame_idx, frame_idx + 1), device, dt, size), trunk_out) last_occluded = torch.full((mux_state.total_valid_entries,), -1, device=device, dtype=torch.long) obj_scores = [1.0] * mux_state.total_valid_entries if keep_alive is not None: @@ -1702,7 +1719,7 @@ class SAM31Tracker(nn.Module): current_out = self._condition_with_masks( det_masks, frame_idx, vision_feats, vision_pos, feat_sizes, high_res_prop, output_dict, N, mux_state, backbone_obj, - images[frame_idx:frame_idx + 1], trunk_out, threshold=0.0) + _prep_frame(images, slice(frame_idx, frame_idx + 1), device, dt, size), trunk_out, threshold=0.0) last_occluded = torch.full((mux_state.total_valid_entries,), -1, device=device, dtype=torch.long) obj_scores = det_scores[:mux_state.total_valid_entries].tolist() if keep_alive is not None: @@ -1718,7 +1735,7 @@ class SAM31Tracker(nn.Module): torch.cuda.current_stream(device).wait_stream(backbone_stream) cur_bb = next_bb else: - cur_bb = self._compute_backbone_frame(backbone_fn, images[frame_idx + 1:frame_idx + 2], frame_idx=frame_idx + 1) + cur_bb = self._compute_backbone_frame(backbone_fn, _prep_frame(images, slice(frame_idx + 1, frame_idx + 2), device, dt, size), frame_idx=frame_idx + 1) continue else: N_obj = mux_state.total_valid_entries @@ -1768,7 +1785,7 @@ class SAM31Tracker(nn.Module): torch.cuda.current_stream(device).wait_stream(backbone_stream) cur_bb = next_bb else: - cur_bb = self._compute_backbone_frame(backbone_fn, images[frame_idx + 1:frame_idx + 2], frame_idx=frame_idx + 1) + cur_bb = self._compute_backbone_frame(backbone_fn, _prep_frame(images, slice(frame_idx + 1, frame_idx + 2), device, dt, size), frame_idx=frame_idx + 1) if not all_masks or all(m is None for m in all_masks): return {"packed_masks": None, "n_frames": N, "scores": []} diff --git a/comfy/ldm/wan/ar_model.py b/comfy/ldm/wan/ar_model.py new file mode 100644 index 000000000..d72f53602 --- /dev/null +++ b/comfy/ldm/wan/ar_model.py @@ -0,0 +1,276 @@ +""" +CausalWanModel: Wan 2.1 backbone with KV-cached causal self-attention for +autoregressive (frame-by-frame) video generation via Causal Forcing. + +Weight-compatible with the standard WanModel -- same layer names, same shapes. +The difference is purely in the forward pass: this model processes one temporal +block at a time and maintains a KV cache across blocks. + +Reference: https://github.com/thu-ml/Causal-Forcing +""" + +import torch +import torch.nn as nn + +from comfy.ldm.modules.attention import optimized_attention +from comfy.ldm.flux.math import apply_rope1 +from comfy.ldm.wan.model import ( + sinusoidal_embedding_1d, + repeat_e, + WanModel, + WanAttentionBlock, +) +import comfy.ldm.common_dit +import comfy.model_management + + +class CausalWanSelfAttention(nn.Module): + """Self-attention with KV cache support for autoregressive inference.""" + + def __init__(self, dim, num_heads, window_size=(-1, -1), qk_norm=True, + eps=1e-6, operation_settings={}): + assert dim % num_heads == 0 + super().__init__() + self.dim = dim + self.num_heads = num_heads + self.head_dim = dim // num_heads + self.qk_norm = qk_norm + self.eps = eps + + ops = operation_settings.get("operations") + device = operation_settings.get("device") + dtype = operation_settings.get("dtype") + + self.q = ops.Linear(dim, dim, device=device, dtype=dtype) + self.k = ops.Linear(dim, dim, device=device, dtype=dtype) + self.v = ops.Linear(dim, dim, device=device, dtype=dtype) + self.o = ops.Linear(dim, dim, device=device, dtype=dtype) + self.norm_q = ops.RMSNorm(dim, eps=eps, elementwise_affine=True, device=device, dtype=dtype) if qk_norm else nn.Identity() + self.norm_k = ops.RMSNorm(dim, eps=eps, elementwise_affine=True, device=device, dtype=dtype) if qk_norm else nn.Identity() + + def forward(self, x, freqs, kv_cache=None, transformer_options={}): + b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim + + q = apply_rope1(self.norm_q(self.q(x)).view(b, s, n, d), freqs) + k = apply_rope1(self.norm_k(self.k(x)).view(b, s, n, d), freqs) + v = self.v(x).view(b, s, n, d) + + if kv_cache is None: + x = optimized_attention( + q.view(b, s, n * d), + k.view(b, s, n * d), + v.view(b, s, n * d), + heads=self.num_heads, + transformer_options=transformer_options, + ) + else: + end = kv_cache["end"] + new_end = end + s + + # Roped K and plain V go into cache + kv_cache["k"][:, end:new_end] = k + kv_cache["v"][:, end:new_end] = v + kv_cache["end"] = new_end + + x = optimized_attention( + q.view(b, s, n * d), + kv_cache["k"][:, :new_end].view(b, new_end, n * d), + kv_cache["v"][:, :new_end].view(b, new_end, n * d), + heads=self.num_heads, + transformer_options=transformer_options, + ) + + x = self.o(x) + return x + + +class CausalWanAttentionBlock(WanAttentionBlock): + """Transformer block with KV-cached self-attention and cross-attention caching.""" + + def __init__(self, cross_attn_type, dim, ffn_dim, num_heads, + window_size=(-1, -1), qk_norm=True, cross_attn_norm=False, + eps=1e-6, operation_settings={}): + super().__init__(cross_attn_type, dim, ffn_dim, num_heads, + window_size, qk_norm, cross_attn_norm, eps, + operation_settings=operation_settings) + self.self_attn = CausalWanSelfAttention( + dim, num_heads, window_size, qk_norm, eps, + operation_settings=operation_settings) + + def forward(self, x, e, freqs, context, context_img_len=257, + kv_cache=None, crossattn_cache=None, transformer_options={}): + if e.ndim < 4: + e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device) + e).chunk(6, dim=1) + else: + e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device).unsqueeze(0) + e).unbind(2) + + # Self-attention with optional KV cache + x = x.contiguous() + y = self.self_attn( + torch.addcmul(repeat_e(e[0], x), self.norm1(x), 1 + repeat_e(e[1], x)), + freqs, kv_cache=kv_cache, transformer_options=transformer_options) + x = torch.addcmul(x, y, repeat_e(e[2], x)) + del y + + # Cross-attention with optional caching + if crossattn_cache is not None and crossattn_cache.get("is_init"): + q = self.cross_attn.norm_q(self.cross_attn.q(self.norm3(x))) + x_ca = optimized_attention( + q, crossattn_cache["k"], crossattn_cache["v"], + heads=self.num_heads, transformer_options=transformer_options) + x = x + self.cross_attn.o(x_ca) + else: + x = x + self.cross_attn(self.norm3(x), context, context_img_len=context_img_len, transformer_options=transformer_options) + if crossattn_cache is not None: + crossattn_cache["k"] = self.cross_attn.norm_k(self.cross_attn.k(context)) + crossattn_cache["v"] = self.cross_attn.v(context) + crossattn_cache["is_init"] = True + + # FFN + y = self.ffn(torch.addcmul(repeat_e(e[3], x), self.norm2(x), 1 + repeat_e(e[4], x))) + x = torch.addcmul(x, y, repeat_e(e[5], x)) + return x + + +class CausalWanModel(WanModel): + """ + Wan 2.1 diffusion backbone with causal KV-cache support. + + Same weight structure as WanModel -- loads identical state dicts. + Adds forward_block() for frame-by-frame autoregressive inference. + """ + + def __init__(self, + model_type='t2v', + patch_size=(1, 2, 2), + text_len=512, + in_dim=16, + dim=2048, + ffn_dim=8192, + freq_dim=256, + text_dim=4096, + out_dim=16, + num_heads=16, + num_layers=32, + window_size=(-1, -1), + qk_norm=True, + cross_attn_norm=True, + eps=1e-6, + image_model=None, + device=None, + dtype=None, + operations=None): + super().__init__( + model_type=model_type, patch_size=patch_size, text_len=text_len, + in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, + text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, + num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, + cross_attn_norm=cross_attn_norm, eps=eps, image_model=image_model, + wan_attn_block_class=CausalWanAttentionBlock, + device=device, dtype=dtype, operations=operations) + + def forward_block(self, x, timestep, context, start_frame, + kv_caches, crossattn_caches, clip_fea=None): + """ + Forward one temporal block for autoregressive inference. + + Args: + x: [B, C, block_frames, H, W] input latent for the current block + timestep: [B, block_frames] per-frame timesteps + context: [B, L, text_dim] raw text embeddings (pre-text_embedding) + start_frame: temporal frame index for RoPE offset + kv_caches: list of per-layer KV cache dicts + crossattn_caches: list of per-layer cross-attention cache dicts + clip_fea: optional CLIP features for I2V + + Returns: + flow_pred: [B, C_out, block_frames, H, W] flow prediction + """ + x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size) + bs, c, t, h, w = x.shape + + x = self.patch_embedding(x.float()).to(x.dtype) + grid_sizes = x.shape[2:] + x = x.flatten(2).transpose(1, 2) + + # Per-frame time embedding + e = self.time_embedding( + sinusoidal_embedding_1d(self.freq_dim, timestep.flatten()).to(dtype=x.dtype)) + e = e.reshape(timestep.shape[0], -1, e.shape[-1]) + e0 = self.time_projection(e).unflatten(2, (6, self.dim)) + + # Text embedding (reuses crossattn_cache after first block) + context = self.text_embedding(context) + + context_img_len = None + if clip_fea is not None and self.img_emb is not None: + context_clip = self.img_emb(clip_fea) + context = torch.concat([context_clip, context], dim=1) + context_img_len = clip_fea.shape[-2] + + # RoPE for current block's temporal position + freqs = self.rope_encode(t, h, w, t_start=start_frame, device=x.device, dtype=x.dtype) + + # Transformer blocks + for i, block in enumerate(self.blocks): + x = block(x, e=e0, freqs=freqs, context=context, + context_img_len=context_img_len, + kv_cache=kv_caches[i], + crossattn_cache=crossattn_caches[i]) + + # Head + x = self.head(x, e) + + # Unpatchify + x = self.unpatchify(x, grid_sizes) + return x[:, :, :t, :h, :w] + + def init_kv_caches(self, batch_size, max_seq_len, device, dtype): + """Create fresh KV caches for all layers.""" + caches = [] + for _ in range(self.num_layers): + caches.append({ + "k": torch.zeros(batch_size, max_seq_len, self.num_heads, self.head_dim, device=device, dtype=dtype), + "v": torch.zeros(batch_size, max_seq_len, self.num_heads, self.head_dim, device=device, dtype=dtype), + "end": 0, + }) + return caches + + def init_crossattn_caches(self, batch_size, device, dtype): + """Create fresh cross-attention caches for all layers.""" + caches = [] + for _ in range(self.num_layers): + caches.append({"is_init": False}) + return caches + + def reset_kv_caches(self, kv_caches): + """Reset KV caches to empty (reuse allocated memory).""" + for cache in kv_caches: + cache["end"] = 0 + + def reset_crossattn_caches(self, crossattn_caches): + """Reset cross-attention caches.""" + for cache in crossattn_caches: + cache["is_init"] = False + + @property + def head_dim(self): + return self.dim // self.num_heads + + def forward(self, x, timestep, context, clip_fea=None, time_dim_concat=None, transformer_options={}, **kwargs): + ar_state = transformer_options.get("ar_state") + if ar_state is not None: + bs = x.shape[0] + block_frames = x.shape[2] + t_per_frame = timestep.unsqueeze(1).expand(bs, block_frames) + return self.forward_block( + x=x, timestep=t_per_frame, context=context, + start_frame=ar_state["start_frame"], + kv_caches=ar_state["kv_caches"], + crossattn_caches=ar_state["crossattn_caches"], + clip_fea=clip_fea, + ) + + return super().forward(x, timestep, context, clip_fea=clip_fea, + time_dim_concat=time_dim_concat, + transformer_options=transformer_options, **kwargs) diff --git a/comfy/ldm/wan/model.py b/comfy/ldm/wan/model.py index b2287dba9..70dfe7b16 100644 --- a/comfy/ldm/wan/model.py +++ b/comfy/ldm/wan/model.py @@ -1135,7 +1135,7 @@ class AudioInjector_WAN(nn.Module): self.injector_adain_output_layers = nn.ModuleList( [operations.Linear(dim, dim, dtype=dtype, device=device) for _ in range(audio_injector_id)]) - def forward(self, x, block_id, audio_emb, audio_emb_global, seq_len): + def forward(self, x, block_id, audio_emb, audio_emb_global, seq_len, scale=1.0): audio_attn_id = self.injected_block_id.get(block_id, None) if audio_attn_id is None: return x @@ -1148,12 +1148,15 @@ class AudioInjector_WAN(nn.Module): attn_hidden_states = adain_hidden_states else: attn_hidden_states = self.injector_pre_norm_feat[audio_attn_id](input_hidden_states) - audio_emb = rearrange(audio_emb, "b t n c -> (b t) n c", t=num_frames) - attn_audio_emb = audio_emb + + if audio_emb.dim() == 3: # WanDancer case + attn_audio_emb = rearrange(audio_emb, "b t c -> (b t) 1 c", t=num_frames) + else: # S2V case + attn_audio_emb = rearrange(audio_emb, "b t n c -> (b t) n c", t=num_frames) + residual_out = self.injector[audio_attn_id](x=attn_hidden_states, context=attn_audio_emb) - residual_out = rearrange( - residual_out, "(b t) n c -> b (t n) c", t=num_frames) - x[:, :seq_len] = x[:, :seq_len] + residual_out + residual_out = rearrange(residual_out, "(b t) n c -> b (t n) c", t=num_frames) + x[:, :seq_len] = x[:, :seq_len] + residual_out * scale return x diff --git a/comfy/ldm/wan/model_wandancer.py b/comfy/ldm/wan/model_wandancer.py new file mode 100644 index 000000000..3caef6dc5 --- /dev/null +++ b/comfy/ldm/wan/model_wandancer.py @@ -0,0 +1,251 @@ +import torch +import torch.nn as nn +import comfy +from comfy.ldm.modules.attention import optimized_attention +from comfy.ldm.flux.math import apply_rope1 +from comfy.ldm.flux.layers import EmbedND + +from .model import AudioInjector_WAN, WanModel, MLPProj, Head, sinusoidal_embedding_1d + + +class MusicSelfAttention(nn.Module): + def __init__(self, dim, num_heads, device=None, dtype=None, operations=None): + assert dim % num_heads == 0 + super().__init__() + self.embed_dim = dim + self.num_heads = num_heads + self.head_dim = dim // num_heads + + self.q_proj = operations.Linear(dim, dim, device=device, dtype=dtype) + self.k_proj = operations.Linear(dim, dim, device=device, dtype=dtype) + self.v_proj = operations.Linear(dim, dim, device=device, dtype=dtype) + self.out_proj = operations.Linear(dim, dim, device=device, dtype=dtype) + + def forward(self, x, freqs): + b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim + + q = self.q_proj(x).view(b, s, n, d) + q = apply_rope1(q, freqs) + + k = self.k_proj(x).view(b, s, n, d) + k = apply_rope1(k, freqs) + + x = optimized_attention( + q.view(b, s, n * d), + k.view(b, s, n * d), + self.v_proj(x).view(b, s, n * d), + heads=self.num_heads, + ) + + return self.out_proj(x) + + +class MusicEncoderLayer(nn.Module): + def __init__(self, dim: int, num_heads: int, ffn_dim: int, device=None, dtype=None, operations=None): + super().__init__() + self.self_attn = MusicSelfAttention(dim, num_heads, device=device, dtype=dtype, operations=operations) + + self.linear1 = operations.Linear(dim, ffn_dim, device=device, dtype=dtype) + self.linear2 = operations.Linear(ffn_dim, dim, device=device, dtype=dtype) + + self.norm1 = operations.LayerNorm(dim, device=device, dtype=dtype) + self.norm2 = operations.LayerNorm(dim, device=device, dtype=dtype) + + def forward(self, x: torch.Tensor, freqs: torch.Tensor) -> torch.Tensor: + x = x + self.self_attn(self.norm1(x), freqs=freqs) + x = x + self.linear2(torch.nn.functional.gelu(self.linear1(self.norm2(x)))) # ffn + return x + + +class WanDancerModel(WanModel): + def __init__(self, + model_type='wandancer', + patch_size=(1, 2, 2), + text_len=512, + in_dim=16, + dim=5120, + ffn_dim=8192, + freq_dim=256, + text_dim=4096, + out_dim=16, + num_heads=16, + num_layers=40, + window_size=(-1, -1), + qk_norm=True, + cross_attn_norm=True, + eps=1e-6, + in_dim_ref_conv=None, + image_model=None, + device=None, dtype=None, operations=None, + audio_inject_layers=[0, 4, 8, 12, 16, 20, 24, 27], + music_dim = 256, + music_heads = 4, + music_feature_dim = 35, + music_latent_dim = 256 + ): + + super().__init__(model_type='i2v', patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, + num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, image_model=image_model, in_dim_ref_conv=in_dim_ref_conv, + device=device, dtype=dtype, operations=operations) + + self.dtype = dtype + operation_settings = {"operations": operations, "device": device, "dtype": dtype} + + self.patch_embedding_global = operations.Conv3d(in_dim, dim, kernel_size=patch_size, stride=patch_size, device=operation_settings.get("device"), dtype=torch.float32) + self.img_emb_refimage = MLPProj(1280, dim, operation_settings=operation_settings) + self.head_global = Head(dim, out_dim, patch_size, eps, operation_settings=operation_settings) + + self.music_injector = AudioInjector_WAN( + dim=self.dim, + num_heads=self.num_heads, + inject_layer=audio_inject_layers, + root_net=self, + enable_adain=False, + dtype=dtype, device=device, operations=operations + ) + + self.music_projection = operations.Linear(music_feature_dim, music_latent_dim, device=device, dtype=dtype) + self.music_encoder = nn.ModuleList([MusicEncoderLayer(dim=music_dim, num_heads=music_heads, ffn_dim=1024, device=device, dtype=dtype, operations=operations) for _ in range(2)]) + music_head_dim = music_dim // music_heads + self.music_rope_embedder = EmbedND(dim=music_head_dim, theta=10000.0, axes_dim=[music_head_dim]) + + def forward_orig(self, x, t, context, clip_fea=None, clip_fea_ref=None, freqs=None, audio_embed=None, fps=30, audio_inject_scale=1.0, transformer_options={}, **kwargs): + # embeddings + if int(fps + 0.5) != 30: + x = self.patch_embedding_global(x.float()).to(x.dtype) + else: + x = self.patch_embedding(x.float()).to(x.dtype) + + grid_sizes = x.shape[2:] + latent_frames = grid_sizes[0] + transformer_options["grid_sizes"] = grid_sizes + x = x.flatten(2).transpose(1, 2) + seq_len = x.size(1) + + # time embeddings + e = self.time_embedding(sinusoidal_embedding_1d(self.freq_dim, t.flatten()).to(dtype=x[0].dtype)) + e = e.reshape(t.shape[0], -1, e.shape[-1]) + e0 = self.time_projection(e).unflatten(2, (6, self.dim)) + + full_ref = None + if self.ref_conv is not None: # model has the weight, but this wasn't used in the original pipeline + full_ref = kwargs.get("reference_latent", None) + if full_ref is not None: + full_ref = self.ref_conv(full_ref).flatten(2).transpose(1, 2) + x = torch.concat((full_ref, x), dim=1) + + # context + context = self.text_embedding(context) + + audio_emb = None + if audio_embed is not None: # encode music feature，[1, frame_num, 35] -> [1, F*8, dim] + music_feature = self.music_projection(audio_embed) + + music_seq_len = music_feature.shape[1] + music_ids = torch.arange(music_seq_len, device=music_feature.device, dtype=music_feature.dtype).reshape(1, -1, 1) # create 1D position IDs + music_freqs = self.music_rope_embedder(music_ids).movedim(1, 2) + + # apply encoder layers + for layer in self.music_encoder: + music_feature = layer(music_feature, music_freqs) + + # interpolate + audio_emb = torch.nn.functional.interpolate(music_feature.unsqueeze(1), size=(latent_frames * 8, self.dim), mode='bilinear').squeeze(1) + + context_img_len = 0 + if self.img_emb is not None and clip_fea is not None: + context_clip = self.img_emb(clip_fea) # bs x 257 x dim + context = torch.cat([context_clip, context], dim=1) + context_img_len += clip_fea.shape[-2] + if self.img_emb_refimage is not None and clip_fea_ref is not None: + context_clip_ref = self.img_emb_refimage(clip_fea_ref) + context = torch.cat([context_clip_ref, context], dim=1) + context_img_len += clip_fea_ref.shape[-2] + + patches_replace = transformer_options.get("patches_replace", {}) + blocks_replace = patches_replace.get("dit", {}) + transformer_options["total_blocks"] = len(self.blocks) + transformer_options["block_type"] = "double" + for i, block in enumerate(self.blocks): + transformer_options["block_index"] = i + if ("double_block", i) in blocks_replace: + def block_wrap(args): + out = {} + out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], context_img_len=context_img_len, transformer_options=args["transformer_options"]) + return out + out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs, "transformer_options": transformer_options}, {"original_block": block_wrap}) + x = out["img"] + else: + x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len, transformer_options=transformer_options) + if audio_emb is not None: + x = self.music_injector(x, i, audio_emb, audio_emb_global=None, seq_len=seq_len, scale=audio_inject_scale) + + # head + if int(fps + 0.5) != 30: + x = self.head_global(x, e) + else: + x = self.head(x, e) + + if full_ref is not None: + x = x[:, full_ref.shape[1]:] + + # unpatchify + x = self.unpatchify(x, grid_sizes) + return x + + def _forward(self, x, timestep, context, clip_fea=None, time_dim_concat=None, transformer_options={}, clip_fea_ref=None, fps=30, audio_inject_scale=1.0, **kwargs): + bs, c, t, h, w = x.shape + x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size) + + t_len = t + if time_dim_concat is not None: + time_dim_concat = comfy.ldm.common_dit.pad_to_patch_size(time_dim_concat, self.patch_size) + x = torch.cat([x, time_dim_concat], dim=2) + t_len = x.shape[2] + + freqs = self.rope_encode(t_len, h, w, device=x.device, dtype=x.dtype, fps=fps, transformer_options=transformer_options) + return self.forward_orig(x, timestep, context, clip_fea=clip_fea, clip_fea_ref=clip_fea_ref, freqs=freqs, fps=fps, audio_inject_scale=audio_inject_scale, transformer_options=transformer_options, **kwargs)[:, :, :t, :h, :w] + + def rope_encode(self, t, h, w, t_start=0, steps_t=None, steps_h=None, steps_w=None, fps=30, device=None, dtype=None, transformer_options={}): + 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]) + w_len = ((w + (patch_size[2] // 2)) // patch_size[2]) + + if steps_t is None: + steps_t = t_len + if steps_h is None: + steps_h = h_len + if steps_w is None: + steps_w = w_len + + h_start = 0 + w_start = 0 + rope_options = transformer_options.get("rope_options", None) + if rope_options is not None: + t_len = (t_len - 1.0) * rope_options.get("scale_t", 1.0) + 1.0 + h_len = (h_len - 1.0) * rope_options.get("scale_y", 1.0) + 1.0 + w_len = (w_len - 1.0) * rope_options.get("scale_x", 1.0) + 1.0 + + t_start += rope_options.get("shift_t", 0.0) + h_start += rope_options.get("shift_y", 0.0) + w_start += rope_options.get("shift_x", 0.0) + + img_ids = torch.zeros((steps_t, steps_h, steps_w, 3), device=device, dtype=dtype) + + if int(fps + 0.5) != 30: + time_scale = 30.0 / fps # how many time units each frame represents relative to 30fps + positions_new = torch.arange(steps_t, device=device, dtype=dtype) * time_scale + t_start + total_frames_at_30fps = int(time_scale * steps_t + 0.5) + positions_new[-1] = t_start + (total_frames_at_30fps - 1) + + img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + positions_new.reshape(-1, 1, 1) + else: + img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(t_start, t_start + (t_len - 1), steps=steps_t, device=device, dtype=dtype).reshape(-1, 1, 1) + + img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(h_start, h_start + (h_len - 1), steps=steps_h, device=device, dtype=dtype).reshape(1, -1, 1) + img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(w_start, w_start + (w_len - 1), steps=steps_w, device=device, dtype=dtype).reshape(1, 1, -1) + img_ids = img_ids.reshape(1, -1, img_ids.shape[-1]) + + freqs = self.rope_embedder(img_ids).movedim(1, 2) + return freqs diff --git a/comfy/lora.py b/comfy/lora.py index e4337c729..f11e26ec9 100644 --- a/comfy/lora.py +++ b/comfy/lora.py @@ -17,6 +17,7 @@ """ from __future__ import annotations +import comfy.memory_management import comfy.utils import comfy.model_management import comfy.model_base @@ -96,12 +97,14 @@ def load_lora(lora, to_load, log_missing=True): def model_lora_keys_clip(model, key_map={}): sdk = model.state_dict().keys() + prefix_set = set() for k in sdk: if k.endswith(".weight"): key_map["text_encoders.{}".format(k[:-len(".weight")])] = k #generic lora format without any weird key names tp = k.find(".transformer.") #also map without wrapper prefix for composite text encoder models if tp > 0 and not k.startswith("clip_"): key_map["text_encoders.{}".format(k[tp + 1:-len(".weight")])] = k + prefix_set.add(k.split('.')[0]) text_model_lora_key = "lora_te_text_model_encoder_layers_{}_{}" clip_l_present = False @@ -162,6 +165,13 @@ def model_lora_keys_clip(model, key_map={}): lora_key = "lora_te1_{}".format(l_key.replace(".", "_")) key_map[lora_key] = k + if len(prefix_set) == 1: + full_prefix = "{}.transformer.model.".format(next(iter(prefix_set))) # kohya anima and maybe other single TE models that use a single llama arch based te + for k in sdk: + if k.endswith(".weight"): + if k.startswith(full_prefix): + l_key = k[len(full_prefix):-len(".weight")] + key_map["lora_te_{}".format(l_key.replace(".", "_"))] = k k = "clip_g.transformer.text_projection.weight" if k in sdk: @@ -473,3 +483,17 @@ def calculate_weight(patches, weight, key, intermediate_dtype=torch.float32, ori weight = old_weight return weight + +def prefetch_prepared_value(value, allocate_buffer, stream): + if isinstance(value, torch.Tensor): + dest = allocate_buffer(comfy.memory_management.vram_aligned_size(value)) + comfy.model_management.cast_to_gathered([value], dest, non_blocking=True, stream=stream) + return comfy.memory_management.interpret_gathered_like([value], dest)[0] + elif isinstance(value, weight_adapter.WeightAdapterBase): + return type(value)(value.loaded_keys, prefetch_prepared_value(value.weights, allocate_buffer, stream)) + elif isinstance(value, tuple): + return tuple(prefetch_prepared_value(item, allocate_buffer, stream) for item in value) + elif isinstance(value, list): + return [prefetch_prepared_value(item, allocate_buffer, stream) for item in value] + + return value diff --git a/comfy/model_base.py b/comfy/model_base.py index 50dab5782..c22705655 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -42,6 +42,8 @@ import comfy.ldm.cosmos.predict2 import comfy.ldm.lumina.model import comfy.ldm.wan.model 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.hidream.model import comfy.ldm.chroma.model @@ -56,6 +58,8 @@ import comfy.ldm.cogvideo.model import comfy.ldm.rt_detr.rtdetr_v4 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.model_management import comfy.patcher_extension @@ -214,6 +218,11 @@ class BaseModel(torch.nn.Module): if "latent_shapes" in extra_conds: xc = utils.unpack_latents(xc, extra_conds.pop("latent_shapes")) + transformer_options = transformer_options.copy() + transformer_options["prefetch_dynamic_vbars"] = ( + self.current_patcher is not None and self.current_patcher.is_dynamic() + ) + model_output = self.diffusion_model(xc, t, context=context, control=control, transformer_options=transformer_options, **extra_conds) if len(model_output) > 1 and not torch.is_tensor(model_output): model_output, _ = utils.pack_latents(model_output) @@ -1360,6 +1369,13 @@ class WAN21(BaseModel): return out +class WAN21_CausalAR(WAN21): + def __init__(self, model_config, model_type=ModelType.FLOW, device=None): + super(WAN21, self).__init__(model_config, model_type, device=device, + unet_model=comfy.ldm.wan.ar_model.CausalWanModel) + self.image_to_video = False + + class WAN21_Vace(WAN21): 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.VaceWanModel) @@ -1586,6 +1602,30 @@ class WAN21_SCAIL(WAN21): return out +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) + self.image_to_video = image_to_video + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + audio_embed = kwargs.get("audio_embed", None) + if audio_embed is not None: + out['audio_embed'] = comfy.conds.CONDRegular(audio_embed) + + clip_vision_output_ref = kwargs.get("clip_vision_output_ref", None) + if clip_vision_output_ref is not None: + out['clip_fea_ref'] = comfy.conds.CONDRegular(clip_vision_output_ref.penultimate_hidden_states) + + fps = kwargs.get("fps", None) + if fps is not None: + out['fps'] = comfy.conds.CONDRegular(torch.FloatTensor([fps])) + + audio_inject_scale = kwargs.get("audio_inject_scale", None) + if audio_inject_scale is not None: + out['audio_inject_scale'] = comfy.conds.CONDRegular(torch.FloatTensor([audio_inject_scale])) + return out + class Hunyuan3Dv2(BaseModel): def __init__(self, model_config, model_type=ModelType.FLOW, device=None): super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hunyuan3d.model.Hunyuan3Dv2) @@ -1636,6 +1676,39 @@ class HiDream(BaseModel): out['image_cond'] = comfy.conds.CONDNoiseShape(self.process_latent_in(image_cond)) return out +class HiDreamO1(BaseModel): + """HiDream-O1-Image: pixel-space DiT (no VAE). Refs from HiDreamO1ReferenceImages and tokens from the stub TE flow through + extra_conds; the heavy preprocessing lives in comfy.ldm.hidream_o1.conditioning.""" + PATCH_SIZE = 32 + + def __init__(self, model_config, model_type=ModelType.FLOW, device=None): + super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hidream_o1.model.HiDreamO1Transformer) + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + text_input_ids = kwargs.get("text_input_ids", None) + noise = kwargs.get("noise", None) + if text_input_ids is None or noise is None: + return out + + # handle area conds + area = kwargs.get("area", None) + if area is not None: + crop_h = min(noise.shape[-2] - area[2], area[0]) + crop_w = min(noise.shape[-1] - area[3], area[1]) + noise = torch.empty((noise.shape[0], 3, crop_h, crop_w), dtype=noise.dtype, device=noise.device) + + conds = build_extra_conds( + text_input_ids, noise, + ref_images=kwargs.get("reference_latents", None), + target_patch_size=self.PATCH_SIZE, + ) + for k, v in conds.items(): + # ar_len is a Python int (precomputed to avoid a GPU sync in forward). + cls = comfy.conds.CONDConstant if k == "ar_len" else comfy.conds.CONDRegular + out[k] = cls(v) + return out + class Chroma(Flux): def __init__(self, model_config, model_type=ModelType.FLUX, device=None, unet_model=comfy.ldm.chroma.model.Chroma): super().__init__(model_config, model_type, device=device, unet_model=unet_model) diff --git a/comfy/model_detection.py b/comfy/model_detection.py index d9b67dcdf..bc0b933bc 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -572,6 +572,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["model_type"] = "animate" 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: + dit_config["model_type"] = "wandancer" else: if '{}img_emb.proj.0.bias'.format(key_prefix) in state_dict_keys: dit_config["model_type"] = "i2v" @@ -618,6 +620,9 @@ 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 '{}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"} + if '{}caption_projection.0.linear.weight'.format(key_prefix) in state_dict_keys: # HiDream dit_config = {} dit_config["image_model"] = "hidream" diff --git a/comfy/model_management.py b/comfy/model_management.py index 95af40012..21738a4c7 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -31,6 +31,7 @@ from contextlib import nullcontext import comfy.memory_management import comfy.utils import comfy.quant_ops +import comfy_aimdo.vram_buffer class VRAMState(Enum): DISABLED = 0 #No vram present: no need to move models to vram @@ -112,10 +113,6 @@ if args.directml is not None: # torch_directml.disable_tiled_resources(True) lowvram_available = False #TODO: need to find a way to get free memory in directml before this can be enabled by default. -try: - import intel_extension_for_pytorch as ipex # noqa: F401 -except: - pass try: _ = torch.xpu.device_count() @@ -583,9 +580,6 @@ class LoadedModel: real_model = self.model.model - if is_intel_xpu() and not args.disable_ipex_optimize and 'ipex' in globals() and real_model is not None: - with torch.no_grad(): - real_model = ipex.optimize(real_model.eval(), inplace=True, graph_mode=True, concat_linear=True) self.real_model = weakref.ref(real_model) self.model_finalizer = weakref.finalize(real_model, cleanup_models) @@ -727,13 +721,15 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu else: minimum_memory_required = max(inference_memory, minimum_memory_required + extra_reserved_memory()) - models_temp = set() + # Order-preserving dedup. A plain set() would randomize iteration order across runs + models_temp = {} for m in models: - models_temp.add(m) + models_temp[m] = None for mm in m.model_patches_models(): - models_temp.add(mm) + models_temp[mm] = None - models = models_temp + models = list(models_temp) + models.reverse() models_to_load = [] @@ -1182,6 +1178,10 @@ stream_counters = {} STREAM_CAST_BUFFERS = {} LARGEST_CASTED_WEIGHT = (None, 0) +STREAM_AIMDO_CAST_BUFFERS = {} +LARGEST_AIMDO_CASTED_WEIGHT = (None, 0) + +DEFAULT_AIMDO_CAST_BUFFER_RESERVATION_SIZE = 16 * 1024 ** 3 def get_cast_buffer(offload_stream, device, size, ref): global LARGEST_CASTED_WEIGHT @@ -1215,13 +1215,26 @@ def get_cast_buffer(offload_stream, device, size, ref): return cast_buffer +def get_aimdo_cast_buffer(offload_stream, device): + cast_buffer = STREAM_AIMDO_CAST_BUFFERS.get(offload_stream, None) + if cast_buffer is None: + cast_buffer = comfy_aimdo.vram_buffer.VRAMBuffer(DEFAULT_AIMDO_CAST_BUFFER_RESERVATION_SIZE, device.index) + STREAM_AIMDO_CAST_BUFFERS[offload_stream] = cast_buffer + + return cast_buffer def reset_cast_buffers(): global LARGEST_CASTED_WEIGHT + global LARGEST_AIMDO_CASTED_WEIGHT + LARGEST_CASTED_WEIGHT = (None, 0) - for offload_stream in STREAM_CAST_BUFFERS: - offload_stream.synchronize() + LARGEST_AIMDO_CASTED_WEIGHT = (None, 0) + for offload_stream in set(STREAM_CAST_BUFFERS) | set(STREAM_AIMDO_CAST_BUFFERS): + if offload_stream is not None: + offload_stream.synchronize() synchronize() + STREAM_CAST_BUFFERS.clear() + STREAM_AIMDO_CAST_BUFFERS.clear() soft_empty_cache() def get_offload_stream(device): @@ -1581,10 +1594,7 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True, ma return False if is_intel_xpu(): - if torch_version_numeric < (2, 3): - return True - else: - return torch.xpu.get_device_properties(device).has_fp16 + return torch.xpu.get_device_properties(device).has_fp16 if is_ascend_npu(): return True @@ -1650,10 +1660,7 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma return False if is_intel_xpu(): - if torch_version_numeric < (2, 3): - return True - else: - return torch.xpu.is_bf16_supported() + return torch.xpu.is_bf16_supported() if is_ascend_npu(): return True @@ -1784,6 +1791,7 @@ def soft_empty_cache(force=False): if cpu_state == CPUState.MPS: torch.mps.empty_cache() elif is_intel_xpu(): + torch.xpu.synchronize() torch.xpu.empty_cache() elif is_ascend_npu(): torch.npu.empty_cache() diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index e259aed63..4f9d8403e 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -26,6 +26,7 @@ import uuid from typing import Callable, Optional import torch +import tqdm import comfy.float import comfy.hooks @@ -121,9 +122,20 @@ class LowVramPatch: self.patches = patches self.convert_func = convert_func # TODO: remove self.set_func = set_func + self.prepared_patches = None + + def prepare(self, allocate_buffer, stream): + self.prepared_patches = [ + (patch[0], comfy.lora.prefetch_prepared_value(patch[1], allocate_buffer, stream), patch[2], patch[3], patch[4]) + for patch in self.patches[self.key] + ] + + def clear_prepared(self): + self.prepared_patches = None def __call__(self, weight): - return comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=weight.dtype) + patches = self.prepared_patches if self.prepared_patches is not None else self.patches[self.key] + return comfy.lora.calculate_weight(patches, weight, self.key, intermediate_dtype=weight.dtype) LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR = 2 @@ -230,6 +242,37 @@ class LazyCastingParam(torch.nn.Parameter): return self.model.patch_weight_to_device(self.key, device_to=self.model.load_device, return_weight=True).to("cpu") +class LazyCastingQuantizedParam: + def __init__(self, model, key): + self.model = model + self.key = key + self.cpu_state_dict = None + + def state_dict_tensor(self, state_dict_key): + if self.cpu_state_dict is None: + weight = self.model.patch_weight_to_device(self.key, device_to=self.model.load_device, return_weight=True) + self.cpu_state_dict = {k: v.to("cpu") for k, v in weight.state_dict(self.key).items()} + return self.cpu_state_dict[state_dict_key] + + +class LazyCastingParamPiece(torch.nn.Parameter): + def __new__(cls, caster, state_dict_key, tensor): + return super().__new__(cls, tensor) + + def __init__(self, caster, state_dict_key, tensor): + self.caster = caster + self.state_dict_key = state_dict_key + + @property + def device(self): + return CustomTorchDevice + + def to(self, *args, **kwargs): + caster = self.caster + del self.caster + return caster.state_dict_tensor(self.state_dict_key) + + class ModelPatcher: def __init__(self, model, load_device, offload_device, size=0, weight_inplace_update=False): self.size = size @@ -1450,21 +1493,45 @@ class ModelPatcher: self.unpatch_hooks() self.clear_cached_hook_weights() - def state_dict_for_saving(self, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None): - unet_state_dict = self.model.diffusion_model.state_dict() - for k, v in unet_state_dict.items(): + def model_state_dict_for_saving(self, model=None, prefix=""): + if model is None: + model = self.model + + original_state_dict = model.state_dict() + output_state_dict = {} + keys = list(original_state_dict) + while len(keys) > 0: + k = keys.pop(0) + v = original_state_dict[k] op_keys = k.rsplit('.', 1) if (len(op_keys) < 2) or op_keys[1] not in ["weight", "bias"]: + output_state_dict[k] = v continue try: - op = comfy.utils.get_attr(self.model.diffusion_model, op_keys[0]) + op = comfy.utils.get_attr(model, op_keys[0]) except: + output_state_dict[k] = v continue if not op or not hasattr(op, "comfy_cast_weights") or \ (hasattr(op, "comfy_patched_weights") and op.comfy_patched_weights == True): + output_state_dict[k] = v continue - key = "diffusion_model." + k - unet_state_dict[k] = LazyCastingParam(self, key, comfy.utils.get_attr(self.model, key)) + key = prefix + k + weight = comfy.utils.get_attr(self.model, key) + if isinstance(weight, QuantizedTensor) and k in original_state_dict: + qt_state_dict = weight.state_dict(k) + caster = LazyCastingQuantizedParam(self, key) + for group_key in (x for x in qt_state_dict if x in original_state_dict): + if group_key in keys: + keys.remove(group_key) + output_state_dict.pop(group_key, "") + output_state_dict[group_key] = LazyCastingParamPiece(caster, prefix + group_key, original_state_dict[group_key]) + continue + output_state_dict[k] = LazyCastingParam(self, key, weight) + return output_state_dict + + def state_dict_for_saving(self, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None): + unet_state_dict = self.model_state_dict_for_saving(self.model.diffusion_model, "diffusion_model.") return self.model.state_dict_for_saving(unet_state_dict, clip_state_dict=clip_state_dict, vae_state_dict=vae_state_dict, clip_vision_state_dict=clip_vision_state_dict) def __del__(self): @@ -1640,7 +1707,11 @@ class ModelPatcherDynamic(ModelPatcher): self.model.model_loaded_weight_memory += casted_buf.numel() * casted_buf.element_size() force_load_stat = f" Force pre-loaded {len(self.backup)} weights: {self.model.model_loaded_weight_memory // 1024} KB." if len(self.backup) > 0 else "" - logging.info(f"Model {self.model.__class__.__name__} prepared for dynamic VRAM loading. {allocated_size // (1024 ** 2)}MB Staged. {num_patches} patches attached.{force_load_stat}") + log_key = (self.patches_uuid, allocated_size, num_patches, len(self.backup), self.model.model_loaded_weight_memory) + in_loop = bool(getattr(tqdm.tqdm, "_instances", None)) + level = logging.DEBUG if in_loop and getattr(self, "_last_prepare_log_key", None) == log_key else logging.INFO + self._last_prepare_log_key = log_key + logging.log(level, f"Model {self.model.__class__.__name__} prepared for dynamic VRAM loading. {allocated_size // (1024 ** 2)}MB Staged. {num_patches} patches attached.{force_load_stat}") self.model.device = device_to self.model.current_weight_patches_uuid = self.patches_uuid diff --git a/comfy/model_prefetch.py b/comfy/model_prefetch.py new file mode 100644 index 000000000..72e11dec6 --- /dev/null +++ b/comfy/model_prefetch.py @@ -0,0 +1,66 @@ +import comfy_aimdo.model_vbar +import comfy.model_management +import comfy.ops + +PREFETCH_QUEUES = [] + +def cleanup_prefetched_modules(comfy_modules): + for s in comfy_modules: + prefetch = getattr(s, "_prefetch", None) + if prefetch is None: + continue + for param_key in ("weight", "bias"): + lowvram_fn = getattr(s, param_key + "_lowvram_function", None) + if lowvram_fn is not None: + lowvram_fn.clear_prepared() + if prefetch["signature"] is not None: + comfy_aimdo.model_vbar.vbar_unpin(s._v) + delattr(s, "_prefetch") + +def cleanup_prefetch_queues(): + global PREFETCH_QUEUES + + for queue in PREFETCH_QUEUES: + for entry in queue: + if entry is None or not isinstance(entry, tuple): + continue + _, prefetch_state = entry + comfy_modules = prefetch_state[1] + if comfy_modules is not None: + cleanup_prefetched_modules(comfy_modules) + PREFETCH_QUEUES = [] + +def prefetch_queue_pop(queue, device, module): + if queue is None: + return + + consumed = queue.pop(0) + if consumed is not None: + offload_stream, prefetch_state = consumed + if offload_stream is not None: + offload_stream.wait_stream(comfy.model_management.current_stream(device)) + _, comfy_modules = prefetch_state + if comfy_modules is not None: + cleanup_prefetched_modules(comfy_modules) + + prefetch = queue[0] + if prefetch is not None: + comfy_modules = [] + for s in prefetch.modules(): + if hasattr(s, "_v"): + comfy_modules.append(s) + + offload_stream = comfy.ops.cast_modules_with_vbar(comfy_modules, None, device, None, True) + comfy.model_management.sync_stream(device, offload_stream) + queue[0] = (offload_stream, (prefetch, comfy_modules)) + +def make_prefetch_queue(queue, device, transformer_options): + if (not transformer_options.get("prefetch_dynamic_vbars", False) + or comfy.model_management.NUM_STREAMS == 0 + or comfy.model_management.is_device_cpu(device) + or not comfy.model_management.device_supports_non_blocking(device)): + return None + + queue = [None] + queue + [None] + PREFETCH_QUEUES.append(queue) + return queue diff --git a/comfy/model_sampling.py b/comfy/model_sampling.py index cf2b5db5f..5af336e76 100644 --- a/comfy/model_sampling.py +++ b/comfy/model_sampling.py @@ -93,7 +93,8 @@ class CONST: def noise_scaling(self, sigma, noise, latent_image, max_denoise=False): sigma = reshape_sigma(sigma, noise.ndim) - return sigma * noise + (1.0 - sigma) * latent_image + s = getattr(self, "noise_scale", 1.0) + return sigma * (s * noise) + (1.0 - sigma) * latent_image def inverse_noise_scaling(self, sigma, latent): sigma = reshape_sigma(sigma, latent.ndim) @@ -288,7 +289,11 @@ class ModelSamplingDiscreteFlow(torch.nn.Module): else: sampling_settings = {} - self.set_parameters(shift=sampling_settings.get("shift", 1.0), multiplier=sampling_settings.get("multiplier", 1000)) + self.set_noise_scale(sampling_settings.get("noise_scale", 1.0)) + self.set_parameters( + shift=sampling_settings.get("shift", 1.0), + multiplier=sampling_settings.get("multiplier", 1000), + ) def set_parameters(self, shift=1.0, timesteps=1000, multiplier=1000): self.shift = shift @@ -296,6 +301,9 @@ class ModelSamplingDiscreteFlow(torch.nn.Module): ts = self.sigma((torch.arange(1, timesteps + 1, 1) / timesteps) * multiplier) self.register_buffer('sigmas', ts) + def set_noise_scale(self, noise_scale): + self.noise_scale = float(noise_scale) + @property def sigma_min(self): return self.sigmas[0] diff --git a/comfy/ops.py b/comfy/ops.py index 050f7cda0..f9456854b 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -86,38 +86,61 @@ def materialize_meta_param(s, param_keys): setattr(s, param_key, torch.nn.Parameter(torch.zeros(param.shape, dtype=param.dtype), requires_grad=param.requires_grad)) -def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype, want_requant): - #vbar doesn't support CPU weights, but some custom nodes have weird paths - #that might switch the layer to the CPU and expect it to work. We have to take - #a clone conservatively as we are mmapped and some SFT files are packed misaligned - #If you are a custom node author reading this, please move your layer to the GPU - #or declare your ModelPatcher as CPU in the first place. - if comfy.model_management.is_device_cpu(device): - materialize_meta_param(s, ["weight", "bias"]) - weight = s.weight.to(dtype=dtype, copy=True) - if isinstance(weight, QuantizedTensor): - weight = weight.dequantize() - bias = None - if s.bias is not None: - bias = s.bias.to(dtype=bias_dtype, copy=True) - return weight, bias, (None, None, None) - +# FIXME: add n=1 cache hit fast path +def cast_modules_with_vbar(comfy_modules, dtype, device, bias_dtype, non_blocking): offload_stream = None - xfer_dest = None + cast_buffer = None + cast_buffer_offset = 0 + + def ensure_offload_stream(module, required_size, check_largest): + nonlocal offload_stream + nonlocal cast_buffer + + if offload_stream is None: + offload_stream = comfy.model_management.get_offload_stream(device) + if offload_stream is None or not check_largest or len(comfy_modules) != 1: + return + + current_size = 0 if cast_buffer is None else cast_buffer.size() + if current_size < required_size and module is comfy.model_management.LARGEST_AIMDO_CASTED_WEIGHT[0]: + offload_stream = comfy.model_management.get_offload_stream(device) + cast_buffer = None + if required_size > comfy.model_management.LARGEST_AIMDO_CASTED_WEIGHT[1]: + comfy.model_management.LARGEST_AIMDO_CASTED_WEIGHT = (module, required_size) + + def get_cast_buffer(buffer_size): + nonlocal offload_stream + nonlocal cast_buffer + nonlocal cast_buffer_offset + + if buffer_size == 0: + return None + + if offload_stream is None: + return torch.empty((buffer_size,), dtype=torch.uint8, device=device) + + cast_buffer = comfy.model_management.get_aimdo_cast_buffer(offload_stream, device) + buffer = comfy_aimdo.torch.aimdo_to_tensor(cast_buffer.get(buffer_size, cast_buffer_offset), device) + cast_buffer_offset += buffer_size + return buffer + + 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) + prefetch = { + "signature": signature, + "resident": resident, + } - signature = comfy_aimdo.model_vbar.vbar_fault(s._v) - resident = comfy_aimdo.model_vbar.vbar_signature_compare(signature, s._v_signature) - if signature is not None: if resident: - weight = s._v_weight - bias = s._v_bias - else: - xfer_dest = comfy_aimdo.torch.aimdo_to_tensor(s._v, device) + s._prefetch = prefetch + continue - if not resident: materialize_meta_param(s, ["weight", "bias"]) + xfer_dest = comfy_aimdo.torch.aimdo_to_tensor(s._v, device) if signature is not None else None cast_geometry = comfy.memory_management.tensors_to_geometries([ s.weight, s.bias ]) cast_dest = None + needs_cast = False xfer_source = [ s.weight, s.bias ] @@ -129,22 +152,15 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu if data is None: continue if data.dtype != geometry.dtype: + needs_cast = True cast_dest = xfer_dest - if cast_dest is None: - cast_dest = torch.empty((comfy.memory_management.vram_aligned_size(cast_geometry),), dtype=torch.uint8, device=device) xfer_dest = None break dest_size = comfy.memory_management.vram_aligned_size(xfer_source) - offload_stream = comfy.model_management.get_offload_stream(device) - if xfer_dest is None and offload_stream is not None: - xfer_dest = comfy.model_management.get_cast_buffer(offload_stream, device, dest_size, s) - if xfer_dest is None: - offload_stream = comfy.model_management.get_offload_stream(device) - xfer_dest = comfy.model_management.get_cast_buffer(offload_stream, device, dest_size, s) + ensure_offload_stream(s, dest_size if xfer_dest is None else 0, True) if xfer_dest is None: - xfer_dest = torch.empty((dest_size,), dtype=torch.uint8, device=device) - offload_stream = None + xfer_dest = get_cast_buffer(dest_size) if signature is None and pin is None: comfy.pinned_memory.pin_memory(s) @@ -157,27 +173,54 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu xfer_source = [ pin ] #send it over comfy.model_management.cast_to_gathered(xfer_source, xfer_dest, non_blocking=non_blocking, stream=offload_stream) - comfy.model_management.sync_stream(device, offload_stream) - if cast_dest is not None: + for param_key in ("weight", "bias"): + lowvram_fn = getattr(s, param_key + "_lowvram_function", None) + if lowvram_fn is not None: + ensure_offload_stream(s, cast_buffer_offset, False) + lowvram_fn.prepare(lambda size: get_cast_buffer(size), offload_stream) + + prefetch["xfer_dest"] = xfer_dest + prefetch["cast_dest"] = cast_dest + prefetch["cast_geometry"] = cast_geometry + prefetch["needs_cast"] = needs_cast + s._prefetch = prefetch + + return offload_stream + + +def resolve_cast_module_with_vbar(s, dtype, device, bias_dtype, compute_dtype, want_requant): + + prefetch = getattr(s, "_prefetch", None) + + if prefetch["resident"]: + weight = s._v_weight + bias = s._v_bias + else: + xfer_dest = prefetch["xfer_dest"] + if prefetch["needs_cast"]: + cast_dest = prefetch["cast_dest"] if prefetch["cast_dest"] is not None else torch.empty((comfy.memory_management.vram_aligned_size(prefetch["cast_geometry"]),), dtype=torch.uint8, device=device) for pre_cast, post_cast in zip(comfy.memory_management.interpret_gathered_like([s.weight, s.bias ], xfer_dest), - comfy.memory_management.interpret_gathered_like(cast_geometry, cast_dest)): + comfy.memory_management.interpret_gathered_like(prefetch["cast_geometry"], cast_dest)): if post_cast is not None: post_cast.copy_(pre_cast) xfer_dest = cast_dest - params = comfy.memory_management.interpret_gathered_like(cast_geometry, xfer_dest) + params = comfy.memory_management.interpret_gathered_like(prefetch["cast_geometry"], xfer_dest) weight = params[0] bias = params[1] - if signature is not None: + if prefetch["signature"] is not None: s._v_weight = weight s._v_bias = bias - s._v_signature=signature + s._v_signature = prefetch["signature"] def post_cast(s, param_key, x, dtype, resident, update_weight): lowvram_fn = getattr(s, param_key + "_lowvram_function", None) fns = getattr(s, param_key + "_function", []) + if x is None: + return None + orig = x def to_dequant(tensor, dtype): @@ -205,14 +248,15 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu x = f(x) return x - update_weight = signature is not None + update_weight = prefetch["signature"] is not None + weight = post_cast(s, "weight", weight, dtype, prefetch["resident"], update_weight) + if bias is not None: + bias = post_cast(s, "bias", bias, bias_dtype, prefetch["resident"], update_weight) - weight = post_cast(s, "weight", weight, dtype, resident, update_weight) - if s.bias is not None: - bias = post_cast(s, "bias", bias, bias_dtype, resident, update_weight) + if prefetch["signature"] is not None: + prefetch["resident"] = True - #FIXME: weird offload return protocol - return weight, bias, (offload_stream, device if signature is not None else None, None) + return weight, bias def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, offloadable=False, compute_dtype=None, want_requant=False): @@ -230,10 +274,46 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of if device is None: device = input.device + def format_return(result, offloadable): + weight, bias, offload_stream = result + return (weight, bias, offload_stream) if offloadable else (weight, bias) + non_blocking = comfy.model_management.device_supports_non_blocking(device) if hasattr(s, "_v"): - return cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype, want_requant) + + #vbar doesn't support CPU weights, but some custom nodes have weird paths + #that might switch the layer to the CPU and expect it to work. We have to take + #a clone conservatively as we are mmapped and some SFT files are packed misaligned + #If you are a custom node author reading this, please move your layer to the GPU + #or declare your ModelPatcher as CPU in the first place. + if comfy.model_management.is_device_cpu(device): + materialize_meta_param(s, ["weight", "bias"]) + weight = s.weight.to(dtype=dtype, copy=True) + if isinstance(weight, QuantizedTensor): + weight = weight.dequantize() + bias = s.bias.to(dtype=bias_dtype, copy=True) if s.bias is not None else None + return format_return((weight, bias, (None, None, None)), offloadable) + + prefetched = hasattr(s, "_prefetch") + offload_stream = None + offload_device = None + if not prefetched: + offload_stream = cast_modules_with_vbar([s], dtype, device, bias_dtype, non_blocking) + comfy.model_management.sync_stream(device, offload_stream) + + weight, bias = resolve_cast_module_with_vbar(s, dtype, device, bias_dtype, compute_dtype, want_requant) + + if not prefetched: + if getattr(s, "_prefetch")["signature"] is not None: + offload_device = device + for param_key in ("weight", "bias"): + lowvram_fn = getattr(s, param_key + "_lowvram_function", None) + if lowvram_fn is not None: + lowvram_fn.clear_prepared() + delattr(s, "_prefetch") + return format_return((weight, bias, (offload_stream, offload_device, None)), offloadable) + if offloadable and (device != s.weight.device or (s.bias is not None and device != s.bias.device)): @@ -280,11 +360,7 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of for f in s.weight_function: weight = f(weight) - if offloadable: - return weight, bias, (offload_stream, weight_a, bias_a) - else: - #Legacy function signature - return weight, bias + return format_return((weight, bias, (offload_stream, weight_a, bias_a)), offloadable) def uncast_bias_weight(s, weight, bias, offload_stream): @@ -486,6 +562,25 @@ class disable_weight_init: else: return super().forward(*args, **kwargs) + class BatchNorm2d(torch.nn.BatchNorm2d, CastWeightBiasOp): + def reset_parameters(self): + return None + + def forward_comfy_cast_weights(self, input): + weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True) + running_mean = self.running_mean.to(device=input.device, dtype=weight.dtype) if self.running_mean is not None else None + running_var = self.running_var.to(device=input.device, dtype=weight.dtype) if self.running_var is not None else None + x = torch.nn.functional.batch_norm(input, running_mean, running_var, weight, bias, self.training, self.momentum, self.eps) + uncast_bias_weight(self, weight, bias, offload_stream) + return x + + def forward(self, *args, **kwargs): + run_every_op() + if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0: + return self.forward_comfy_cast_weights(*args, **kwargs) + else: + return super().forward(*args, **kwargs) + class LayerNorm(torch.nn.LayerNorm, CastWeightBiasOp): def reset_parameters(self): return None @@ -673,6 +768,9 @@ class manual_cast(disable_weight_init): class Conv3d(disable_weight_init.Conv3d): comfy_cast_weights = True + class BatchNorm2d(disable_weight_init.BatchNorm2d): + comfy_cast_weights = True + class GroupNorm(disable_weight_init.GroupNorm): comfy_cast_weights = True @@ -1173,6 +1271,94 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec self._buffers[key] = fn(buf) return self + class Embedding(manual_cast.Embedding): + 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: + layer_conf = json.loads(layer_conf.numpy().tobytes()) + + # 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: + 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] + + scale_key = f"{prefix}weight_scale" + scale = state_dict.pop(scale_key, None) + if scale is not None: + scale = scale.float() + manually_loaded_keys.append(scale_key) + + params = layout_cls.Params( + scale=scale if scale is not None else torch.ones((), dtype=torch.float32), + orig_dtype=MixedPrecisionOps._compute_dtype, + orig_shape=(self.num_embeddings, self.embedding_dim), + ) + self.weight = torch.nn.Parameter( + QuantizedTensor(weight.to(dtype=qconfig["storage_t"]), qconfig["comfy_tensor_layout"], params), + requires_grad=False) + + 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) + + 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 + + def forward_comfy_cast_weights(self, input, out_dtype=None): + weight = self.weight + + # Optimized path: lookup in fp8, dequantize only the selected rows. + if isinstance(weight, QuantizedTensor) and len(self.weight_function) == 0: + qdata, _, offload_stream = cast_bias_weight(self, device=input.device, dtype=weight.dtype, offloadable=True) + if isinstance(qdata, QuantizedTensor): + scale = qdata._params.scale + qdata = qdata._qdata + else: + scale = None + + x = torch.nn.functional.embedding( + input, qdata, self.padding_idx, self.max_norm, + self.norm_type, self.scale_grad_by_freq, self.sparse) + uncast_bias_weight(self, qdata, None, offload_stream) + target_dtype = out_dtype if out_dtype is not None else weight._params.orig_dtype + x = x.to(dtype=target_dtype) + if scale is not None and scale != 1.0: + x = x * scale.to(dtype=target_dtype) + return x + + # Fallback for non-quantized or weight_function (LoRA) case + return super().forward_comfy_cast_weights(input, out_dtype=out_dtype) + return MixedPrecisionOps def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, model_config=None): @@ -1190,6 +1376,7 @@ def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_ if not fp8_compute: disabled.add("float8_e4m3fn") disabled.add("float8_e5m2") + logging.info("Native ops: {} {}".format(", ".join(QUANT_ALGOS.keys() - disabled), ", emulated ops: {}".format(", ".join(disabled)) if len(disabled) > 0 else "")) return mixed_precision_ops(model_config.quant_config, compute_dtype, disabled=disabled) if ( diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index 42ee08fb2..b90bcfd25 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -1,6 +1,8 @@ import torch import logging +from comfy.cli_args import args + try: import comfy_kitchen as ck from comfy_kitchen.tensor import ( @@ -21,7 +23,15 @@ try: ck.registry.disable("cuda") logging.warning("WARNING: You need pytorch with cu130 or higher to use optimized CUDA operations.") - ck.registry.disable("triton") + if args.enable_triton_backend: + try: + import triton + logging.info("Found triton %s. Enabling comfy-kitchen triton backend.", triton.__version__) + except ImportError as e: + logging.error(f"Failed to import triton, Error: {e}, the comfy-kitchen triton backend will not be available.") + ck.registry.disable("triton") + else: + ck.registry.disable("triton") for k, v in ck.list_backends().items(): logging.info(f"Found comfy_kitchen backend {k}: {v}") except ImportError as e: diff --git a/comfy/rmsnorm.py b/comfy/rmsnorm.py index ab7cf14fa..e54be98d6 100644 --- a/comfy/rmsnorm.py +++ b/comfy/rmsnorm.py @@ -3,6 +3,7 @@ import comfy.model_management RMSNorm = torch.nn.RMSNorm +# Note: torch's fused F.rms_norm is faster but produces slightly different output than manual implementations (rsqrt/reduction rounding). def rms_norm(x, weight=None, eps=1e-6): if weight is None: return torch.nn.functional.rms_norm(x, (x.shape[-1],), eps=eps) diff --git a/comfy/sample.py b/comfy/sample.py index 653829582..2be0cae5f 100644 --- a/comfy/sample.py +++ b/comfy/sample.py @@ -37,11 +37,12 @@ def prepare_noise(latent_image, seed, noise_inds=None): return noises -def fix_empty_latent_channels(model, latent_image, downscale_ratio_spacial=None): +def fix_empty_latent_channels(model, latent_image, downscale_ratio_spacial=None, downscale_ratio_temporal=None): if latent_image.is_nested: return latent_image latent_format = model.get_model_object("latent_format") #Resize the empty latent image so it has the right number of channels - if torch.count_nonzero(latent_image) == 0: + is_empty = torch.count_nonzero(latent_image) == 0 + if is_empty: if latent_format.latent_channels != latent_image.shape[1]: latent_image = comfy.utils.repeat_to_batch_size(latent_image, latent_format.latent_channels, dim=1) if downscale_ratio_spacial is not None: @@ -51,6 +52,13 @@ def fix_empty_latent_channels(model, latent_image, downscale_ratio_spacial=None) if latent_format.latent_dimensions == 3 and latent_image.ndim == 4: latent_image = latent_image.unsqueeze(2) + + if is_empty and downscale_ratio_temporal is not None: + if downscale_ratio_temporal != latent_format.temporal_downscale_ratio: + ratio = downscale_ratio_temporal / latent_format.temporal_downscale_ratio + new_t = max(1, round(latent_image.shape[2] * ratio)) + latent_image = comfy.utils.repeat_to_batch_size(latent_image, new_t, dim=2) + return latent_image def prepare_sampling(model, noise_shape, positive, negative, noise_mask): diff --git a/comfy/sampler_helpers.py b/comfy/sampler_helpers.py index bbba09e26..3782fd2d5 100644 --- a/comfy/sampler_helpers.py +++ b/comfy/sampler_helpers.py @@ -89,7 +89,8 @@ def get_additional_models(conds, dtype): gligen += get_models_from_cond(conds[k], "gligen") add_models += get_models_from_cond(conds[k], "additional_models") - control_nets = set(cnets) + # Order-preserving dedup. A plain set() would randomize iteration order across runs + control_nets = list(dict.fromkeys(cnets)) inference_memory = 0 control_models = [] diff --git a/comfy/sd.py b/comfy/sd.py index ee66490f5..2443353a4 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -65,6 +65,8 @@ import comfy.text_encoders.ace15 import comfy.text_encoders.longcat_image import comfy.text_encoders.qwen35 import comfy.text_encoders.ernie +import comfy.text_encoders.gemma4 +import comfy.text_encoders.cogvideo import comfy.model_patcher import comfy.lora @@ -77,7 +79,7 @@ import comfy.latent_formats import comfy.ldm.flux.redux -def load_lora_for_models(model, clip, lora, strength_model, strength_clip): +def load_lora_for_models(model, clip, lora, strength_model, strength_clip, lora_metadata=None): key_map = {} if model is not None: key_map = comfy.lora.model_lora_keys_unet(model.model, key_map) @@ -89,6 +91,8 @@ def load_lora_for_models(model, clip, lora, strength_model, strength_clip): if model is not None: new_modelpatcher = model.clone() k = new_modelpatcher.add_patches(loaded, strength_model) + if lora_metadata: + new_modelpatcher.set_attachments("lora_metadata", lora_metadata) else: k = () new_modelpatcher = None @@ -96,6 +100,8 @@ def load_lora_for_models(model, clip, lora, strength_model, strength_clip): if clip is not None: new_clip = clip.clone() k1 = new_clip.add_patches(loaded, strength_clip) + if lora_metadata: + new_clip.patcher.set_attachments("lora_metadata", lora_metadata) else: k1 = () new_clip = None @@ -237,7 +243,8 @@ class CLIP: model_management.archive_model_dtypes(self.cond_stage_model) self.tokenizer = tokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data) - ModelPatcher = comfy.model_patcher.ModelPatcher if disable_dynamic else comfy.model_patcher.CoreModelPatcher + te_disable_dynamic = disable_dynamic or getattr(self.cond_stage_model, "disable_offload", False) + ModelPatcher = comfy.model_patcher.ModelPatcher if te_disable_dynamic else comfy.model_patcher.CoreModelPatcher self.patcher = ModelPatcher(self.cond_stage_model, load_device=load_device, offload_device=offload_device) #Match torch.float32 hardcode upcast in TE implemention self.patcher.set_model_compute_dtype(torch.float32) @@ -416,6 +423,13 @@ class CLIP: sd_clip[k] = sd_tokenizer[k] return sd_clip + def state_dict_for_saving(self): + sd_clip = self.patcher.model_state_dict_for_saving() + sd_tokenizer = self.tokenizer.state_dict() + for k in sd_tokenizer: + sd_clip[k] = sd_tokenizer[k] + return sd_clip + def load_model(self, tokens={}): memory_used = 0 if hasattr(self.cond_stage_model, "memory_estimation_function"): @@ -774,6 +788,7 @@ class VAE: self.latent_channels = 3 self.latent_dim = 2 self.output_channels = 3 + self.disable_offload = True elif "vocoder.activation_post.downsample.lowpass.filter" in sd: #MMAudio VAE sample_rate = 16000 if sample_rate == 16000: @@ -1223,6 +1238,7 @@ class CLIPType(Enum): NEWBIE = 24 FLUX2 = 25 LONGCAT_IMAGE = 26 + COGVIDEOX = 27 @@ -1271,6 +1287,9 @@ class TEModel(Enum): QWEN35_9B = 26 QWEN35_27B = 27 MINISTRAL_3_3B = 28 + GEMMA_4_E4B = 29 + GEMMA_4_E2B = 30 + GEMMA_4_31B = 31 def detect_te_model(sd): @@ -1296,6 +1315,12 @@ def detect_te_model(sd): return TEModel.BYT5_SMALL_GLYPH return TEModel.T5_BASE if 'model.layers.0.post_feedforward_layernorm.weight' in sd: + if 'model.layers.59.self_attn.q_norm.weight' in sd: + return TEModel.GEMMA_4_31B + if 'model.layers.41.self_attn.q_norm.weight' in sd and 'model.layers.47.self_attn.q_norm.weight' not in sd: + return TEModel.GEMMA_4_E4B + if 'model.layers.34.self_attn.q_norm.weight' in sd and 'model.layers.41.self_attn.q_norm.weight' not in sd: + return TEModel.GEMMA_4_E2B if 'model.layers.47.self_attn.q_norm.weight' in sd: return TEModel.GEMMA_3_12B if 'model.layers.0.self_attn.q_norm.weight' in sd: @@ -1418,6 +1443,9 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**t5xxl_detect(clip_data), clip_l=False, clip_g=False, t5=True, llama=False, dtype_llama=None) clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer + elif clip_type == CLIPType.COGVIDEOX: + clip_target.clip = comfy.text_encoders.cogvideo.cogvideo_te(**t5xxl_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.cogvideo.CogVideoXTokenizer else: #CLIPType.MOCHI clip_target.clip = comfy.text_encoders.genmo.mochi_te(**t5xxl_detect(clip_data)) clip_target.tokenizer = comfy.text_encoders.genmo.MochiT5Tokenizer @@ -1435,6 +1463,13 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip else: clip_target.clip = comfy.text_encoders.sa_t5.SAT5Model clip_target.tokenizer = comfy.text_encoders.sa_t5.SAT5Tokenizer + elif te_model in (TEModel.GEMMA_4_E4B, TEModel.GEMMA_4_E2B, TEModel.GEMMA_4_31B): + variant = {TEModel.GEMMA_4_E4B: comfy.text_encoders.gemma4.Gemma4_E4B, + TEModel.GEMMA_4_E2B: comfy.text_encoders.gemma4.Gemma4_E2B, + TEModel.GEMMA_4_31B: comfy.text_encoders.gemma4.Gemma4_31B}[te_model] + clip_target.clip = comfy.text_encoders.gemma4.gemma4_te(**llama_detect(clip_data), model_class=variant) + 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 @@ -1880,7 +1915,7 @@ def save_checkpoint(output_path, model, clip=None, vae=None, clip_vision=None, m load_models = [model] if clip is not None: load_models.append(clip.load_model()) - clip_sd = clip.get_sd() + clip_sd = clip.state_dict_for_saving() vae_sd = None if vae is not None: vae_sd = vae.get_sd() diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 92d0305c5..1e4434fd5 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -28,6 +28,7 @@ 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 from . import supported_models_base from . import latent_formats @@ -1167,6 +1168,25 @@ class WAN21_T2V(supported_models_base.BASE): t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}umt5xxl.transformer.".format(pref)) return supported_models_base.ClipTarget(comfy.text_encoders.wan.WanT5Tokenizer, comfy.text_encoders.wan.te(**t5_detect)) +class WAN21_CausalAR_T2V(WAN21_T2V): + unet_config = { + "image_model": "wan2.1", + "model_type": "t2v", + "causal_ar": True, + } + + sampling_settings = { + "shift": 5.0, + } + + def __init__(self, unet_config): + super().__init__(unet_config) + self.unet_config.pop("causal_ar", None) + + def get_model(self, state_dict, prefix="", device=None): + return model_base.WAN21_CausalAR(self, device=device) + + class WAN21_I2V(WAN21_T2V): unet_config = { "image_model": "wan2.1", @@ -1294,6 +1314,37 @@ class WAN21_SCAIL(WAN21_T2V): out = model_base.WAN21_SCAIL(self, image_to_video=False, device=device) return out +class WAN22_WanDancer(WAN21_T2V): + unet_config = { + "image_model": "wan2.1", + "model_type": "wandancer", + "in_dim": 36, + } + + def __init__(self, unet_config): + super().__init__(unet_config) + self.memory_usage_factor = 1.8 + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.WAN22_WanDancer(self, image_to_video=True, device=device) + return out + + def process_unet_state_dict(self, state_dict): + out_sd = {} + for k in list(state_dict.keys()): + # split music_encoder in_proj into q_proj, k_proj, v_proj + if "music_encoder" in k and "self_attn.in_proj" in k: + suffix = "weight" if k.endswith("weight") else "bias" + tensor = state_dict[k] + d = tensor.shape[0] // 3 + prefix = k.replace(f"in_proj_{suffix}", "") + out_sd[f"{prefix}q_proj.{suffix}"] = tensor[:d] + out_sd[f"{prefix}k_proj.{suffix}"] = tensor[d:2*d] + out_sd[f"{prefix}v_proj.{suffix}"] = tensor[2*d:] + else: + out_sd[k] = state_dict[k] + return out_sd + class Hunyuan3Dv2(supported_models_base.BASE): unet_config = { "image_model": "hunyuan3d2", @@ -1381,6 +1432,50 @@ class HiDream(supported_models_base.BASE): def clip_target(self, state_dict={}): return None # TODO +class HiDreamO1(supported_models_base.BASE): + unet_config = { + "image_model": "hidream_o1", + } + + sampling_settings = { + "shift": 3.0, + "noise_scale": 8.0, + } + + latent_format = latent_formats.HiDreamO1Pixel + memory_usage_factor = 0.033 + # fp16 not supported: LM MLP down_proj activations fp16 overflow, causing NaNs + supported_inference_dtypes = [torch.bfloat16, torch.float32] + + vae_key_prefix = ["vae."] + text_encoder_key_prefix = ["text_encoders."] + + optimizations = {"fp8": False} + + def get_model(self, state_dict, prefix="", device=None): + return model_base.HiDreamO1(self, device=device) + + def process_unet_state_dict(self, state_dict): + # Drop unused Qwen3-VL deepstack merger weights; upstream discards them at inference. + for key in list(state_dict.keys()): + if "visual.deepstack_merger_list" in key: + del state_dict[key] + return state_dict + + def process_vae_state_dict(self, state_dict): + # Pixel-space model: inject sentinel so VAE construction picks PixelspaceConversionVAE. + return {"pixel_space_vae": torch.tensor(1.0)} + + def process_clip_state_dict(self, state_dict): + # Tokenizer-only TE: inject sentinel so load_state_dict_guess_config triggers CLIP init. + return {"_hidream_o1_te_sentinel": torch.zeros(1)} + + def clip_target(self, state_dict={}): + return supported_models_base.ClipTarget( + comfy.text_encoders.hidream_o1.HiDreamO1Tokenizer, + comfy.text_encoders.hidream_o1.HiDreamO1TE, + ) + class Chroma(supported_models_base.BASE): unet_config = { "image_model": "chroma", @@ -1853,6 +1948,14 @@ class CogVideoX_T2V(supported_models_base.BASE): vae_key_prefix = ["vae."] text_encoder_key_prefix = ["text_encoders."] + def __init__(self, unet_config): + # 2b-class (dim=1920, heads=30) uses scale_factor=1.15258426. + # 5b-class (dim=3072, heads=48) — incl. CogVideoX-5b, 1.5-5B, and + # Fun-V1.5 inpainting — uses scale_factor=0.7 per vae/config.json. + if unet_config.get("num_attention_heads", 0) >= 48: + self.latent_format = latent_formats.CogVideoX1_5 + super().__init__(unet_config) + def get_model(self, state_dict, prefix="", device=None): # CogVideoX 1.5 (patch_size_t=2) has different training base dimensions for RoPE if self.unet_config.get("patch_size_t") is not None: @@ -1879,6 +1982,104 @@ class CogVideoX_I2V(CogVideoX_T2V): out = model_base.CogVideoX(self, image_to_video=True, device=device) return out -models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, LongCatImage, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImagePixelSpace, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, WAN21_FlowRVS, WAN21_SCAIL, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, ACEStep15, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima, RT_DETR_v4, ErnieImage, SAM3, SAM31, CogVideoX_I2V, CogVideoX_T2V] +class CogVideoX_Inpaint(CogVideoX_T2V): + unet_config = { + "image_model": "cogvideox", + "in_channels": 48, + } -models += [SVD_img2vid] + def get_model(self, state_dict, prefix="", device=None): + if self.unet_config.get("patch_size_t") is not None: + self.unet_config.setdefault("sample_height", 96) + self.unet_config.setdefault("sample_width", 170) + self.unet_config.setdefault("sample_frames", 81) + out = model_base.CogVideoX(self, image_to_video=True, device=device) + return out + + +models = [ + LotusD, + Stable_Zero123, + SD15_instructpix2pix, + SD15, + SD20, + SD21UnclipL, + SD21UnclipH, + SDXL_instructpix2pix, + SDXLRefiner, + SDXL, + SSD1B, + KOALA_700M, + KOALA_1B, + Segmind_Vega, + SD_X4Upscaler, + Stable_Cascade_C, + Stable_Cascade_B, + SV3D_u, + SV3D_p, + SD3, + StableAudio, + AuraFlow, + PixArtAlpha, + PixArtSigma, + HunyuanDiT, + HunyuanDiT1, + FluxInpaint, + Flux, + LongCatImage, + FluxSchnell, + GenmoMochi, + LTXV, + LTXAV, + HunyuanVideo15_SR_Distilled, + HunyuanVideo15, + HunyuanImage21Refiner, + HunyuanImage21, + HunyuanVideoSkyreelsI2V, + HunyuanVideoI2V, + HunyuanVideo, + CosmosT2V, + CosmosI2V, + CosmosT2IPredict2, + CosmosI2VPredict2, + ZImagePixelSpace, + ZImage, + Lumina2, + WAN22_T2V, + WAN21_CausalAR_T2V, + WAN21_T2V, + WAN21_I2V, + WAN21_FunControl2V, + WAN21_Vace, + WAN21_Camera, + WAN22_Camera, + WAN22_S2V, + WAN21_HuMo, + WAN22_Animate, + WAN21_FlowRVS, + WAN21_SCAIL, + WAN22_WanDancer, + Hunyuan3Dv2mini, + Hunyuan3Dv2, + Hunyuan3Dv2_1, + HiDream, + HiDreamO1, + Chroma, + ChromaRadiance, + ACEStep, + ACEStep15, + Omnigen2, + QwenImage, + Flux2, + Kandinsky5Image, + Kandinsky5, + Anima, + RT_DETR_v4, + ErnieImage, + SAM3, + SAM31, + CogVideoX_Inpaint, + CogVideoX_I2V, + CogVideoX_T2V, + SVD_img2vid, +] diff --git a/comfy/text_encoders/cogvideo.py b/comfy/text_encoders/cogvideo.py index f1e8e3f5d..b97310709 100644 --- a/comfy/text_encoders/cogvideo.py +++ b/comfy/text_encoders/cogvideo.py @@ -1,6 +1,48 @@ import comfy.text_encoders.sd3_clip +from comfy import sd1_clip class CogVideoXT5Tokenizer(comfy.text_encoders.sd3_clip.T5XXLTokenizer): + """Inner T5 tokenizer for CogVideoX. + + CogVideoX was trained with T5 embeddings padded to 226 tokens (not 77 like SD3). + Used both directly by supported_models.CogVideoX_T2V.clip_target (paired with + the raw T5XXLModel) and by the CogVideoXTokenizer outer wrapper below. + """ def __init__(self, embedding_directory=None, tokenizer_data={}): super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, min_length=226) + + +class CogVideoXTokenizer(sd1_clip.SD1Tokenizer): + """Outer tokenizer wrapper for CLIPLoader (type="cogvideox").""" + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, + clip_name="t5xxl", tokenizer=CogVideoXT5Tokenizer) + + +class CogVideoXT5XXL(sd1_clip.SD1ClipModel): + """Outer T5XXL model wrapper for CLIPLoader (type="cogvideox"). + + Wraps the raw T5XXL model in the SD1ClipModel interface so that CLIP.__init__ + (which reads self.dtypes) works correctly. The inner model is the standard + sd3_clip.T5XXLModel (no attention_mask change needed for CogVideoX). + """ + def __init__(self, device="cpu", dtype=None, model_options={}): + super().__init__(device=device, dtype=dtype, name="t5xxl", + clip_model=comfy.text_encoders.sd3_clip.T5XXLModel, + model_options=model_options) + + +def cogvideo_te(dtype_t5=None, t5_quantization_metadata=None): + """Factory that returns a CogVideoXT5XXL class configured with the detected + T5 dtype and optional quantization metadata, for use in load_text_encoder_state_dicts. + """ + class CogVideoXTEModel_(CogVideoXT5XXL): + def __init__(self, device="cpu", dtype=None, model_options={}): + if t5_quantization_metadata is not None: + model_options = model_options.copy() + model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata + if dtype_t5 is not None: + dtype = dtype_t5 + super().__init__(device=device, dtype=dtype, model_options=model_options) + return CogVideoXTEModel_ diff --git a/comfy/text_encoders/gemma4.py b/comfy/text_encoders/gemma4.py new file mode 100644 index 000000000..f050061ed --- /dev/null +++ b/comfy/text_encoders/gemma4.py @@ -0,0 +1,1298 @@ +import torch +import torch.nn as nn +import numpy as np +from dataclasses import dataclass +import math + +from comfy import sd1_clip +import comfy.model_management +from comfy.ldm.modules.attention import optimized_attention_for_device +from comfy.rmsnorm import rms_norm +from comfy.text_encoders.llama import RMSNorm, MLP, BaseLlama, BaseGenerate, _make_scaled_embedding + + +# Intentional minor divergences from transformers -reference implementation: +# - Embedding sqrt(hidden_size) scale applied as a Python scalar (full precision) instead of dtype-matched buffer tensor. +# - RMSNorm uses torch fused F.rms_norm, very slight numerical differences, but considerably faster +# - Input image and audio resizing/resampling slightly different numerically + + +GEMMA4_VISION_CONFIG = {"hidden_size": 768, "image_size": 896, "intermediate_size": 3072, "num_attention_heads": 12, "num_hidden_layers": 16, "patch_size": 16, "head_dim": 64, "rms_norm_eps": 1e-6, "position_embedding_size": 10240, "pooling_kernel_size": 3} +GEMMA4_VISION_31B_CONFIG = {"hidden_size": 1152, "image_size": 896, "intermediate_size": 4304, "num_attention_heads": 16, "num_hidden_layers": 27, "patch_size": 16, "head_dim": 72, "rms_norm_eps": 1e-6, "position_embedding_size": 10240, "pooling_kernel_size": 3} +GEMMA4_AUDIO_CONFIG = {"hidden_size": 1024, "num_hidden_layers": 12, "num_attention_heads": 8, "intermediate_size": 4096, "conv_kernel_size": 5, "attention_chunk_size": 12, "attention_context_left": 13, "attention_context_right": 0, "attention_logit_cap": 50.0, "output_proj_dims": 1536, "rms_norm_eps": 1e-6, "residual_weight": 0.5} + +@dataclass +class Gemma4Config: + vocab_size: int = 262144 + hidden_size: int = 2560 + intermediate_size: int = 10240 + num_hidden_layers: int = 42 + num_attention_heads: int = 8 + num_key_value_heads: int = 2 + max_position_embeddings: int = 131072 + rms_norm_eps: float = 1e-6 + rope_theta = [1000000.0, 10000.0] + transformer_type: str = "gemma4" + head_dim = 256 + global_head_dim = 512 + rms_norm_add = False + mlp_activation = "gelu_pytorch_tanh" + qkv_bias = False + rope_dims = None + q_norm = "gemma3" + k_norm = "gemma3" + sliding_attention = [512, 512, 512, 512, 512, False] + rope_scale = None + partial_rotary_factor: float = 0.25 + final_norm: bool = True + lm_head: bool = False + final_logit_softcapping: float = 30.0 + hidden_size_per_layer_input: int = 256 + num_kv_shared_layers: int = 18 + use_double_wide_mlp: bool = False + stop_tokens = [1, 50, 106] + vision_config = GEMMA4_VISION_CONFIG + audio_config = GEMMA4_AUDIO_CONFIG + mm_tokens_per_image = 280 + +@dataclass +class Gemma4_E2B_Config(Gemma4Config): + hidden_size: int = 1536 + intermediate_size: int = 6144 + num_hidden_layers: int = 35 + num_key_value_heads: int = 1 + sliding_attention = [512, 512, 512, 512, False] + num_kv_shared_layers: int = 20 + use_double_wide_mlp: bool = True + +@dataclass +class Gemma4_31B_Config(Gemma4Config): + hidden_size: int = 5376 + intermediate_size: int = 21504 + num_hidden_layers: int = 60 + num_attention_heads: int = 32 + num_key_value_heads: int = 16 + sliding_attention = [1024, 1024, 1024, 1024, 1024, False] + hidden_size_per_layer_input: int = 0 + num_kv_shared_layers: int = 0 + audio_config = None + vision_config = GEMMA4_VISION_31B_CONFIG + + +# unfused RoPE as addcmul_ RoPE diverges from reference code +def _apply_rotary_pos_emb(x, freqs_cis): + cos, sin = freqs_cis[0], freqs_cis[1] + half = x.shape[-1] // 2 + out = x * cos + out[..., :half] -= x[..., half:] * sin[..., :half] + out[..., half:] += x[..., :half] * sin[..., half:] + return out + +class Gemma4Attention(nn.Module): + def __init__(self, config, head_dim, device=None, dtype=None, ops=None): + super().__init__() + self.num_heads = config.num_attention_heads + self.num_kv_heads = config.num_key_value_heads + self.hidden_size = config.hidden_size + self.head_dim = head_dim + self.inner_size = self.num_heads * head_dim + + self.q_proj = ops.Linear(config.hidden_size, self.inner_size, bias=config.qkv_bias, device=device, dtype=dtype) + self.k_proj = ops.Linear(config.hidden_size, self.num_kv_heads * head_dim, bias=config.qkv_bias, device=device, dtype=dtype) + self.v_proj = ops.Linear(config.hidden_size, self.num_kv_heads * head_dim, bias=config.qkv_bias, device=device, dtype=dtype) + self.o_proj = ops.Linear(self.inner_size, config.hidden_size, bias=False, device=device, dtype=dtype) + + self.q_norm = None + self.k_norm = None + if config.q_norm == "gemma3": + self.q_norm = RMSNorm(head_dim, eps=config.rms_norm_eps, device=device, dtype=dtype) + if config.k_norm == "gemma3": + self.k_norm = RMSNorm(head_dim, eps=config.rms_norm_eps, device=device, dtype=dtype) + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask=None, + freqs_cis=None, + past_key_value=None, + sliding_window=None, + shared_kv=None, + ): + batch_size, seq_length, _ = hidden_states.shape + + xq = self.q_proj(hidden_states) + xq = xq.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2) + if self.q_norm is not None: + xq = self.q_norm(xq) + + if shared_kv is not None: + xk, xv = shared_kv + # Apply RoPE to Q only (K already has RoPE from source layer) + xq = _apply_rotary_pos_emb(xq, freqs_cis) + present_key_value = None + shareable_kv = None + else: + xk = self.k_proj(hidden_states).view(batch_size, seq_length, self.num_kv_heads, self.head_dim) + xv = self.v_proj(hidden_states).view(batch_size, seq_length, self.num_kv_heads, self.head_dim) + if self.k_norm is not None: + xk = self.k_norm(xk) + xv = rms_norm(xv) + xk = xk.transpose(1, 2) + xv = xv.transpose(1, 2) + xq = _apply_rotary_pos_emb(xq, freqs_cis) + xk = _apply_rotary_pos_emb(xk, freqs_cis) + + present_key_value = None + if past_key_value is not None: + cumulative_len = 0 + if len(past_key_value) > 0: + past_key, past_value, cumulative_len = past_key_value + xk = torch.cat((past_key, xk), dim=2) + xv = torch.cat((past_value, xv), dim=2) + new_cumulative = cumulative_len + seq_length + if sliding_window is not None and xk.shape[2] > sliding_window - 1: + cache_k = xk[:, :, -(sliding_window - 1):] + cache_v = xv[:, :, -(sliding_window - 1):] + else: + cache_k = xk + cache_v = xv + present_key_value = (cache_k, cache_v, new_cumulative) + + # KV for sharing: full xk/xv that SDPA sees (not evicted cache) + shareable_kv = (xk, xv) + + # GQA: pass unexpanded KV with enable_gqa when no sliding mask, + # expand heads when sliding mask is present + # has to be done within SDPA itself to match the reference code, pre-scaling expansion causes numerical differences + expand_kv = (self.num_heads != self.num_kv_heads and + sliding_window is not None and + xk.shape[2] >= sliding_window) + if expand_kv: + xk = xk.repeat_interleave(self.num_heads // self.num_kv_heads, dim=1) + xv = xv.repeat_interleave(self.num_heads // self.num_kv_heads, dim=1) + gqa_kwargs = {} if expand_kv else ({"enable_gqa": True} if self.num_heads != self.num_kv_heads else {}) + output = optimized_attention_for_device(xq.device, mask=attention_mask is not None, small_input=True)(xq, xk, xv, self.num_heads, mask=attention_mask, skip_reshape=True, scale=1.0, **gqa_kwargs) + + return self.o_proj(output), present_key_value, shareable_kv + + +class TransformerBlockGemma4(nn.Module): + def __init__(self, config, index, device=None, dtype=None, ops=None): + super().__init__() + if config.sliding_attention is not None: + self.sliding_attention = config.sliding_attention[index % len(config.sliding_attention)] + else: + self.sliding_attention = False + + head_dim = config.head_dim if self.sliding_attention else config.global_head_dim + + self.self_attn = Gemma4Attention(config, head_dim=head_dim, device=device, dtype=dtype, ops=ops) + + num_kv_shared = config.num_kv_shared_layers + first_kv_shared = config.num_hidden_layers - num_kv_shared + mlp_size = config.intermediate_size * 2 if config.use_double_wide_mlp and index >= first_kv_shared else None + self.mlp = MLP(config, device=device, dtype=dtype, ops=ops, intermediate_size=mlp_size) + + 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.pre_feedforward_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, device=device, dtype=dtype) + self.post_feedforward_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, device=device, dtype=dtype) + + self.hidden_size_per_layer_input = config.hidden_size_per_layer_input + if self.hidden_size_per_layer_input: + self.per_layer_input_gate = ops.Linear(config.hidden_size, self.hidden_size_per_layer_input, bias=False, device=device, dtype=dtype) + self.per_layer_projection = ops.Linear(self.hidden_size_per_layer_input, config.hidden_size, bias=False, device=device, dtype=dtype) + self.post_per_layer_input_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, device=device, dtype=dtype) + self.register_buffer("layer_scalar", torch.ones(1, device=device, dtype=dtype)) + else: + self.layer_scalar = None + + def forward(self, x, attention_mask=None, freqs_cis=None, past_key_value=None, per_layer_input=None, shared_kv=None): + sliding_window = None + if self.sliding_attention: + sliding_window = self.sliding_attention + # For prefill > sliding window, add sliding window restriction to the causal mask. + if x.shape[1] > self.sliding_attention: + sw_mask = torch.zeros(x.shape[1], x.shape[1], dtype=x.dtype, device=x.device) + sw_mask.masked_fill_(torch.ones_like(sw_mask, dtype=torch.bool).tril_(-self.sliding_attention), torch.finfo(x.dtype).min) + attention_mask = attention_mask + sw_mask if attention_mask is not None else sw_mask + freqs_cis = freqs_cis[1] + else: + freqs_cis = freqs_cis[0] + + residual = x + x = self.input_layernorm(x) + x, present_key_value, shareable_kv = self.self_attn( + hidden_states=x, attention_mask=attention_mask, freqs_cis=freqs_cis, + past_key_value=past_key_value, sliding_window=sliding_window, shared_kv=shared_kv, + ) + x = self.post_attention_layernorm(x) + x = residual + x + + residual = x + x = self.pre_feedforward_layernorm(x) + x = self.mlp(x) + x = self.post_feedforward_layernorm(x) + x = residual + x + + if self.hidden_size_per_layer_input and per_layer_input is not None: + residual = x + x = self.per_layer_input_gate(x) + x = torch.nn.functional.gelu(x, approximate="tanh") + x = x * per_layer_input + x = self.per_layer_projection(x) + x = self.post_per_layer_input_norm(x) + x = residual + x + + if self.layer_scalar is not None: + x = x * self.layer_scalar + + return x, present_key_value, shareable_kv + + +class Gemma4Transformer(nn.Module): + def __init__(self, config, device=None, dtype=None, ops=None): + super().__init__() + self.config = config + + self.embed_tokens = _make_scaled_embedding(ops, config.vocab_size, config.hidden_size, config.hidden_size ** 0.5, device, dtype) + + self.layers = nn.ModuleList([ + TransformerBlockGemma4(config, index=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) if config.final_norm else None + + # Precompute RoPE inv_freq on CPU to match reference code's exact value + rope_angles_global = int(config.partial_rotary_factor * config.global_head_dim // 2) + nope_global = config.global_head_dim // 2 - rope_angles_global + global_inv = 1.0 / (config.rope_theta[0] ** (torch.arange(0, 2 * rope_angles_global, 2).float() / config.global_head_dim)) + if nope_global > 0: + global_inv = torch.cat([global_inv, torch.zeros(nope_global)]) + self.register_buffer("_global_inv_freq", global_inv, persistent=False) + + sliding_inv = 1.0 / (config.rope_theta[1] ** (torch.arange(0, config.head_dim, 2).float() / config.head_dim)) + self.register_buffer("_sliding_inv_freq", sliding_inv, persistent=False) + + # Per-layer input mechanism + self.hidden_size_per_layer_input = config.hidden_size_per_layer_input + if self.hidden_size_per_layer_input: + self.embed_tokens_per_layer = _make_scaled_embedding(ops, config.vocab_size, config.num_hidden_layers * self.hidden_size_per_layer_input, self.hidden_size_per_layer_input ** 0.5, device, dtype) + self.per_layer_model_projection = ops.Linear( + config.hidden_size, config.num_hidden_layers * self.hidden_size_per_layer_input, + bias=False, device=device, dtype=dtype) + self.per_layer_projection_norm = RMSNorm( + self.hidden_size_per_layer_input, eps=config.rms_norm_eps, + device=device, dtype=dtype) + + def get_past_len(self, past_key_values): + for kv in past_key_values: + if len(kv) >= 3: + return kv[2] + return 0 + + def _freqs_from_inv(self, inv_freq, position_ids, device, dtype): + """Compute cos/sin from stored inv_freq""" + inv_exp = inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(device) + pos_exp = position_ids[:, None, :].float() + freqs = (inv_exp @ pos_exp).transpose(1, 2) + emb = torch.cat((freqs, freqs), dim=-1) + return emb.cos().unsqueeze(1).to(dtype), emb.sin().unsqueeze(1).to(dtype) + + def compute_freqs_cis(self, position_ids, device, dtype=None): + global_freqs = self._freqs_from_inv(self._global_inv_freq, position_ids, device, dtype) + sliding_freqs = self._freqs_from_inv(self._sliding_inv_freq, position_ids, device, dtype) + return [global_freqs, sliding_freqs] + + def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, + final_layer_norm_intermediate=True, dtype=None, position_ids=None, embeds_info=None, + past_key_values=None, input_ids=None): + if embeds is not None: + x = embeds + else: + x = self.embed_tokens(x, out_dtype=dtype) + + seq_len = x.shape[1] + past_len = 0 + if past_key_values is not None and len(past_key_values) > 0: + past_len = self.get_past_len(past_key_values) + + if position_ids is None: + position_ids = torch.arange(past_len, past_len + seq_len, device=x.device).unsqueeze(0) + + freqs_cis = self.compute_freqs_cis(position_ids, x.device, dtype=x.dtype) + + mask = None + min_val = torch.finfo(x.dtype).min + if attention_mask is not None: + mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, seq_len, attention_mask.shape[-1]) + mask = mask.masked_fill(mask.to(torch.bool), min_val) + + if seq_len > 1: + causal_mask = torch.zeros(past_len + seq_len, past_len + seq_len, dtype=x.dtype, device=x.device) + causal_mask.masked_fill_(torch.ones_like(causal_mask, dtype=torch.bool).triu_(1), min_val) + mask = mask + causal_mask if mask is not None else causal_mask + + # Per-layer inputs + per_layer_inputs = None + if self.hidden_size_per_layer_input: + num_layers = self.config.num_hidden_layers + hpl = self.hidden_size_per_layer_input + per_layer_proj = self.per_layer_model_projection(x) * (1.0 / (self.config.hidden_size ** 0.5)) + per_layer_proj = self.per_layer_projection_norm(per_layer_proj.reshape(*x.shape[:-1], num_layers, hpl)) + if input_ids is not None and input_ids.shape[1] == x.shape[1]: + per_layer_emb = self.embed_tokens_per_layer(input_ids).reshape(*input_ids.shape, num_layers, hpl) + per_layer_inputs = (per_layer_proj + per_layer_emb) * (0.5 ** 0.5) + else: + per_layer_inputs = per_layer_proj + + # KV sharing: later layers reuse KV from the last non-shared sliding/global layer + num_kv_shared = self.config.num_kv_shared_layers + first_kv_shared = self.config.num_hidden_layers - num_kv_shared if num_kv_shared > 0 else self.config.num_hidden_layers + shared_sliding_kv = None # KV from last non-shared sliding layer + shared_global_kv = None # KV from last non-shared global layer + + intermediate = None + next_key_values = [] + for i, layer in enumerate(self.layers): + past_kv = past_key_values[i] if past_key_values is not None and len(past_key_values) > 0 else None + + layer_kwargs = {} + if per_layer_inputs is not None: + layer_kwargs['per_layer_input'] = per_layer_inputs[:, :, i, :] + + is_sliding = hasattr(layer, 'sliding_attention') and layer.sliding_attention + if i >= first_kv_shared and num_kv_shared > 0: + shared = shared_sliding_kv if is_sliding else shared_global_kv + if shared is not None: + layer_kwargs['shared_kv'] = shared + + x, current_kv, shareable_kv = layer(x=x, attention_mask=mask, freqs_cis=freqs_cis, past_key_value=past_kv, **layer_kwargs) + + next_key_values.append(current_kv if current_kv is not None else ()) + + # Only track the last sliding/global before the sharing boundary + if i < first_kv_shared and shareable_kv is not None: + if is_sliding: + shared_sliding_kv = shareable_kv + else: + shared_global_kv = shareable_kv + + if i == intermediate_output: + intermediate = x.clone() + + if self.norm is not None: + x = self.norm(x) + + if len(next_key_values) > 0: + return x, intermediate, next_key_values + return x, intermediate + + +class Gemma4Base(BaseLlama, BaseGenerate, torch.nn.Module): + """Common base for all Gemma4 variants: text model + vision.""" + def _init_model(self, config, dtype, device, operations): + self.num_layers = config.num_hidden_layers + self.model = Gemma4Transformer(config, device=device, dtype=dtype, ops=operations) + self.dtype = dtype + self.multi_modal_projector = Gemma4MultiModalProjector(config, dtype=dtype, device=device, ops=operations) + self.vision_model = Gemma4VisionEncoder(config.vision_config, dtype=dtype, device=device, ops=operations) + + def logits(self, x): + logits = super().logits(x) + cap = self.model.config.final_logit_softcapping + if cap: + logits = cap * torch.tanh(logits / cap) + return logits + + def init_kv_cache(self, batch, max_cache_len, device, execution_dtype): + past_key_values = [] + for _ in range(self.model.config.num_hidden_layers): + past_key_values.append(()) + return past_key_values + + def preprocess_embed(self, embed, device): + if embed["type"] == "image": + image = embed.pop("data").movedim(-1, 1) # [B, H, W, C] -> [B, C, H, W] + max_soft_tokens = embed.get("max_soft_tokens", None) + vision_out = self.vision_model(image.to(device, dtype=torch.float32), max_soft_tokens=max_soft_tokens) + return self.multi_modal_projector(vision_out), None + return None, None + + +class Gemma4AudioMixin: + """Adds audio support to a Gemma4 model.""" + def _init_audio(self, config, dtype, device, operations): + self.audio_model = Gemma4AudioEncoder(config.audio_config, dtype=dtype, device=device, ops=operations) + self.audio_projector = Gemma4AudioProjector({"audio_output_proj_dims": config.audio_config["output_proj_dims"], "text_hidden_size": config.hidden_size, "rms_norm_eps": config.rms_norm_eps}, dtype=dtype, device=device, ops=operations) + + def preprocess_embed(self, embed, device): + result, extra = super().preprocess_embed(embed, device) + if result is not None: + return result, extra + if embed["type"] == "audio": + audio = embed.pop("data").to(device, dtype=torch.float32) + audio_mask = embed.pop("mask", None) + if audio_mask is not None: + audio_mask = audio_mask.to(device) + audio_out = self.audio_model(audio, audio_mask=audio_mask) + return self.audio_projector(audio_out), None + return None, None + + +# Vision Encoder + +def _compute_vision_2d_rope(head_dim, pixel_position_ids, theta=100.0, device=None): + """Compute 2D RoPE for vision: separate frequencies for x and y dimensions. + + Args: + head_dim: dimension per head (e.g. 64) + pixel_position_ids: [batch, num_patches, 2] with (x, y) coords + theta: RoPE base frequency + Returns: + (cos, sin) each of shape [batch, num_patches, head_dim] + """ + rotary_dim_per_axis = head_dim // 2 + freq_indices = torch.arange(0, rotary_dim_per_axis, 2, device=device).float() + inv_freq = 1.0 / (theta ** (freq_indices / rotary_dim_per_axis)) + + all_cos, all_sin = [], [] + for i in range(2): # x and y + dim_positions = pixel_position_ids[:, :, i].float() # [batch, num_patches] + freqs = torch.einsum('bi,j->bij', dim_positions, inv_freq.to(device)) # [batch, num_patches, rotary_dim/2] + emb = torch.cat([freqs, freqs], dim=-1) # [batch, num_patches, rotary_dim] + all_cos.append(emb.cos()) + all_sin.append(emb.sin()) + + cos = torch.cat(all_cos, dim=-1).to(pixel_position_ids.device) # [batch, num_patches, head_dim] + sin = torch.cat(all_sin, dim=-1).to(pixel_position_ids.device) + return cos, sin + + +def _apply_vision_2d_rope(x, freqs): + """Apply 2D RoPE (multidimensional) to vision query/key states. + + Splits x and cos/sin into ndim=2 parts, applies 1D RoPE to each independently. + + x: [batch, heads, seq, head_dim] + freqs: (cos, sin) each [batch, seq, head_dim] + """ + cos = freqs[0].unsqueeze(1) # [batch, 1, seq, head_dim] + sin = freqs[1].unsqueeze(1) + half = x.shape[-1] // 2 + a = _apply_rotary_pos_emb(x[..., :half], (cos[..., :half], sin[..., :half])) + b = _apply_rotary_pos_emb(x[..., half:], (cos[..., half:], sin[..., half:])) + return torch.cat([a, b], dim=-1) + + +class ClippedLinear(nn.Module): + """Linear layer with activation clipping (from quantization-aware training). + + Stores input_max/min and output_max/min as buffers loaded from checkpoint. + """ + def __init__(self, in_features, out_features, bias=False, device=None, dtype=None, ops=None): + super().__init__() + self.linear = ops.Linear(in_features, out_features, bias=bias, device=device, dtype=dtype) + self.register_buffer('input_max', torch.tensor(float('inf'), device=device, dtype=dtype)) + self.register_buffer('input_min', torch.tensor(float('-inf'), device=device, dtype=dtype)) + self.register_buffer('output_max', torch.tensor(float('inf'), device=device, dtype=dtype)) + self.register_buffer('output_min', torch.tensor(float('-inf'), device=device, dtype=dtype)) + + @property + def weight(self): + return self.linear.weight + + def forward(self, x): + x = x.clamp(min=self.input_min, max=self.input_max) + x = self.linear(x) + return x.clamp_(min=self.output_min, max=self.output_max) + + +class Gemma4VisionMLP(nn.Module): + """SwiGLU MLP matching gate_proj/up_proj/down_proj structure.""" + def __init__(self, config, device=None, dtype=None, ops=None): + super().__init__() + hidden_size = config["hidden_size"] + intermediate_size = config["intermediate_size"] + self.gate_proj = ClippedLinear(hidden_size, intermediate_size, device=device, dtype=dtype, ops=ops) + self.up_proj = ClippedLinear(hidden_size, intermediate_size, device=device, dtype=dtype, ops=ops) + self.down_proj = ClippedLinear(intermediate_size, hidden_size, device=device, dtype=dtype, ops=ops) + + def forward(self, x): + return self.down_proj(torch.nn.functional.gelu(self.gate_proj(x), approximate="tanh") * self.up_proj(x)) + + +class Gemma4VisionAttention(nn.Module): + def __init__(self, config, device=None, dtype=None, ops=None): + super().__init__() + self.hidden_size = config["hidden_size"] + self.num_heads = config["num_attention_heads"] + self.head_dim = config.get("head_dim", self.hidden_size // self.num_heads) + + self.q_proj = ClippedLinear(self.hidden_size, self.num_heads * self.head_dim, device=device, dtype=dtype, ops=ops) + self.k_proj = ClippedLinear(self.hidden_size, self.num_heads * self.head_dim, device=device, dtype=dtype, ops=ops) + self.v_proj = ClippedLinear(self.hidden_size, self.num_heads * self.head_dim, device=device, dtype=dtype, ops=ops) + self.o_proj = ClippedLinear(self.num_heads * self.head_dim, self.hidden_size, device=device, dtype=dtype, ops=ops) + + self.q_norm = RMSNorm(self.head_dim, eps=config["rms_norm_eps"], device=device, dtype=dtype) + self.k_norm = RMSNorm(self.head_dim, eps=config["rms_norm_eps"], device=device, dtype=dtype) + + def forward(self, x, freqs, attention_mask=None): + batch_size, seq_length, _ = x.shape + + xq = self.q_proj(x).view(batch_size, seq_length, self.num_heads, self.head_dim) + xk = self.k_proj(x).view(batch_size, seq_length, self.num_heads, self.head_dim) + xv = self.v_proj(x).view(batch_size, seq_length, self.num_heads, self.head_dim) + + xq = self.q_norm(xq).transpose(1, 2) + xk = self.k_norm(xk).transpose(1, 2) + xv = rms_norm(xv) + + xq = _apply_vision_2d_rope(xq, freqs) + xk = _apply_vision_2d_rope(xk, freqs) + + xv = xv.to(xq.dtype).transpose(1, 2) + + output = optimized_attention_for_device(xq.device, mask=attention_mask is not None, small_input=True)(xq, xk, xv, self.num_heads, mask=attention_mask, skip_reshape=True, scale=1.0) + return self.o_proj(output) + + +class Gemma4VisionLayer(nn.Module): + def __init__(self, config, device=None, dtype=None, ops=None): + super().__init__() + self.self_attn = Gemma4VisionAttention(config, device=device, dtype=dtype, ops=ops) + self.mlp = Gemma4VisionMLP(config, device=device, dtype=dtype, ops=ops) + norm_kwargs = dict(eps=config["rms_norm_eps"], device=device, dtype=dtype) + hidden = config["hidden_size"] + self.input_layernorm = RMSNorm(hidden, **norm_kwargs) + self.post_attention_layernorm = RMSNorm(hidden, **norm_kwargs) + self.pre_feedforward_layernorm = RMSNorm(hidden, **norm_kwargs) + self.post_feedforward_layernorm = RMSNorm(hidden, **norm_kwargs) + + def forward(self, x, freqs, attention_mask=None): + residual = x + x = self.input_layernorm(x) + x = self.self_attn(x, freqs, attention_mask=attention_mask) + x = self.post_attention_layernorm(x) + x = residual + x + + residual = x + x = self.pre_feedforward_layernorm(x) + x = self.mlp(x) + x = self.post_feedforward_layernorm(x) + x = residual + x + return x + + +class Gemma4PatchEmbedder(nn.Module): + """Patch embedding with learned 2D position embeddings via one-hot lookup.""" + def __init__(self, config, device=None, dtype=None, ops=None): + super().__init__() + hidden_size = config["hidden_size"] + patch_size = config["patch_size"] + self.patch_size = patch_size + self.position_embedding_size = config.get("position_embedding_size", 10240) + + self.input_proj = ops.Linear(3 * patch_size * patch_size, hidden_size, bias=False, device=device, dtype=dtype) + self.position_embedding_table = nn.Parameter( + torch.empty(2, self.position_embedding_size, hidden_size, device=device, dtype=dtype) + ) + + def forward(self, patches, pixel_position_ids): + """ + patches: [B, num_patches, 3*patch_size²] in [0,1] range (normalized to [-1,1] inside, matching HF) + pixel_position_ids: [B, num_patches, 2] with (x,y) positions, (-1,-1) for padding + """ + hidden_states = self.input_proj((2.0 * (patches - 0.5)).to(self.input_proj.weight.dtype)) + + clamped_positions = pixel_position_ids.clamp(min=0) + pos_table = comfy.model_management.cast_to_device(self.position_embedding_table, hidden_states.device, hidden_states.dtype) + position_embeddings = pos_table[0][clamped_positions[..., 0]] + pos_table[1][clamped_positions[..., 1]] + + # Zero out position embeddings for padding patches (matching HF) + padding_positions = (pixel_position_ids == -1).all(dim=-1) + position_embeddings = torch.where(padding_positions.unsqueeze(-1), 0.0, position_embeddings) + + return hidden_states + position_embeddings + + +class Gemma4VisionEncoderLayers(nn.Module): + """Wrapper to produce state dict keys as encoder.layers.X.*""" + def __init__(self, config, dtype=None, device=None, ops=None): + super().__init__() + self.layers = nn.ModuleList([ + Gemma4VisionLayer(config, device=device, dtype=dtype, ops=ops) + for _ in range(config["num_hidden_layers"]) + ]) + + +class Gemma4VisionEncoder(nn.Module): + def __init__(self, config, dtype=None, device=None, ops=None): + super().__init__() + self.config = config + self.hidden_size = config["hidden_size"] + self.head_dim = config.get("head_dim", config["hidden_size"] // config["num_attention_heads"]) + self.patch_size = config["patch_size"] + self.pooling_kernel_size = config.get("pooling_kernel_size", 3) + self.root_hidden_size = self.hidden_size ** 0.5 + + self.patch_embedder = Gemma4PatchEmbedder(config, device=device, dtype=dtype, ops=ops) + self.encoder = Gemma4VisionEncoderLayers(config, dtype=dtype, device=device, ops=ops) + + def forward(self, pixel_values, max_soft_tokens=None): + """ + pixel_values: [B, C, H, W] in [0,1] range + max_soft_tokens: if provided, pad to max_soft_tokens * k² total patches + """ + batch_size, _, height, width = pixel_values.shape + ps = self.patch_size + k = self.pooling_kernel_size + patches_h, patches_w = height // ps, width // ps + num_patches = patches_h * patches_w + output_length = max_soft_tokens if max_soft_tokens is not None else num_patches // (k * k) + n_padding = output_length * k * k - num_patches + + # Patchify and build position grid + patches = pixel_values.reshape(batch_size, -1, patches_h, ps, patches_w, ps) + patches = patches.permute(0, 2, 4, 3, 5, 1).reshape(batch_size, num_patches, -1) + grid_y, grid_x = torch.meshgrid(torch.arange(patches_h, device=pixel_values.device), torch.arange(patches_w, device=pixel_values.device), indexing='ij') + position_ids = torch.stack([grid_x.flatten(), grid_y.flatten()], dim=-1).unsqueeze(0).expand(batch_size, -1, -1) + + # Append zero-pixel padding with (-1,-1) positions + if n_padding > 0: + patches = torch.cat([patches, patches.new_zeros(batch_size, n_padding, patches.shape[-1])], dim=1) + position_ids = torch.cat([position_ids, position_ids.new_full((batch_size, n_padding, 2), -1)], dim=1) + + padding = (position_ids == -1).all(dim=-1) + + # Embed, encode, pool + x = self.patch_embedder(patches, position_ids) + freqs = _compute_vision_2d_rope(self.head_dim, position_ids, device=pixel_values.device) + freqs = tuple(t.to(x.dtype) for t in freqs) + if n_padding > 0: + mask = padding.unsqueeze(1).unsqueeze(2).expand(-1, 1, position_ids.shape[1], -1) + mask = torch.zeros_like(mask, dtype=x.dtype).masked_fill_(mask, torch.finfo(x.dtype).min) + else: + mask = None + + for layer in self.encoder.layers: + x = layer(x, freqs, attention_mask=mask) + + if n_padding > 0: + x = x.masked_fill(padding.unsqueeze(-1), 0.0) + + # Average pool by spatial position + clamped = position_ids.clamp(min=0) + max_x = clamped[:, :, 0].max(dim=-1, keepdim=True)[0] + 1 + ki = torch.div(clamped, k, rounding_mode="floor") + ki = ki[:, :, 0] + (max_x // k) * ki[:, :, 1] + weights = torch.nn.functional.one_hot(ki.long(), output_length).float() / (k * k) + x = (weights.transpose(1, 2) @ x.float()).to(x.dtype) + + # Strip empty output tokens + valid_out = ~((weights == 0).all(dim=1)) + if valid_out.any() and not valid_out.all(): + x = x[:, valid_out[0]] if batch_size > 1 else x[valid_out].unsqueeze(0) + + return x * self.root_hidden_size + + +class Gemma4RMSNormProjector(nn.Module): + """Shared projector: parameterless RMSNorm → linear. Used for both vision and audio.""" + def __init__(self, in_dim, out_dim, dtype=None, device=None, ops=None): + super().__init__() + self.embedding_projection = ops.Linear(in_dim, out_dim, bias=False, device=device, dtype=dtype) + + def forward(self, x): + return self.embedding_projection(rms_norm(x)) + + +class Gemma4MultiModalProjector(Gemma4RMSNormProjector): + def __init__(self, config, dtype=None, device=None, ops=None): + super().__init__(config.vision_config["hidden_size"], config.hidden_size, dtype=dtype, device=device, ops=ops) + + +# Audio Encoder + +class Gemma4AudioConvSubsampler(nn.Module): + """2D convolution subsampling for audio features""" + def __init__(self, config, device=None, dtype=None, ops=None): + super().__init__() + eps = config["rms_norm_eps"] + self.layer0 = nn.ModuleDict({ + 'conv': ops.Conv2d(1, 128, kernel_size=3, stride=2, padding=1, bias=False, device=device, dtype=dtype), + 'norm': ops.LayerNorm(128, eps=eps, elementwise_affine=True, bias=False, device=device, dtype=dtype), + }) + self.layer1 = nn.ModuleDict({ + 'conv': ops.Conv2d(128, 32, kernel_size=3, stride=2, padding=1, bias=False, device=device, dtype=dtype), + 'norm': ops.LayerNorm(32, eps=eps, elementwise_affine=True, bias=False, device=device, dtype=dtype), + }) + # proj_input_dim = (128 // 4) * 32 = 1024 + self.input_proj_linear = ops.Linear(1024, config["hidden_size"], bias=False, device=device, dtype=dtype) + + def _conv_layer(self, x, layer, mask): + if mask is not None: + x = x * mask[:, None, :, None].to(x.device) + x = layer['conv'](x.to(layer['conv'].weight.dtype)) + x = torch.relu(layer['norm'](x.permute(0, 2, 3, 1)).permute(0, 3, 1, 2).contiguous()) + if mask is not None: + mask = mask[:, ::2] + return x, mask + + def forward(self, x, mask=None): + x = x.unsqueeze(1) + x, mask = self._conv_layer(x, self.layer0, mask) + x, mask = self._conv_layer(x, self.layer1, mask) + batch_size, _, seq_len, _ = x.shape + x = x.permute(0, 2, 3, 1).contiguous().reshape(batch_size, seq_len, -1) + return self.input_proj_linear(x), mask + + +class Gemma4AudioFeedForward(nn.Module): + """Conformer feed-forward with residual scaling.""" + def __init__(self, config, device=None, dtype=None, ops=None): + super().__init__() + hidden_size = config["hidden_size"] + intermediate_size = config.get("intermediate_size", hidden_size * 4) + self.pre_layer_norm = RMSNorm(hidden_size, eps=config["rms_norm_eps"], device=device, dtype=dtype) + self.ffw_layer_1 = ClippedLinear(hidden_size, intermediate_size, device=device, dtype=dtype, ops=ops) + self.ffw_layer_2 = ClippedLinear(intermediate_size, hidden_size, device=device, dtype=dtype, ops=ops) + self.post_layer_norm = RMSNorm(hidden_size, eps=config["rms_norm_eps"], device=device, dtype=dtype) + self.post_layer_scale = config.get("residual_weight", 0.5) + + def forward(self, x): + residual = x + x = self.pre_layer_norm(x) + x = torch.nn.functional.silu(self.ffw_layer_1(x)) + x = self.ffw_layer_2(x) + x = self.post_layer_norm(x) + x = x * self.post_layer_scale + return x + residual + + +class Gemma4AudioRelPositionalEncoding(nn.Module): + """Sinusoidal relative positional encoding for audio attention.""" + def __init__(self, config, device=None, dtype=None): + super().__init__() + hidden_size = config["hidden_size"] + context_left = config.get("attention_context_left", 13) + context_right = config.get("attention_context_right", 0) + self.chunk_size = config.get("attention_chunk_size", 12) + self.context_size = self.chunk_size + context_left - 1 + context_right + + num_timescales = hidden_size // 2 + log_inc = math.log(10000.0) / max(num_timescales - 1, 1) + inv_timescales = torch.exp(torch.arange(num_timescales) * -log_inc).to(dtype=dtype).unsqueeze(0).unsqueeze(0) + self.register_buffer("inv_timescales", inv_timescales, persistent=False) + + def forward(self, hidden_states): + positions = torch.arange(self.chunk_size, -1, -1, device=hidden_states.device).unsqueeze(-1) + scaled = positions * self.inv_timescales.to(device=hidden_states.device) + return torch.cat([torch.sin(scaled), torch.cos(scaled)], dim=-1).to(dtype=hidden_states.dtype) + + +class Gemma4AudioAttention(nn.Module): + """Chunked block attention with relative position bias and softcap.""" + def __init__(self, config, device=None, dtype=None, ops=None): + super().__init__() + self.hidden_size = config["hidden_size"] + self.num_heads = config["num_attention_heads"] + self.head_dim = self.hidden_size // self.num_heads + self.chunk_size = config.get("attention_chunk_size", 12) + self.max_past_horizon = config.get("attention_context_left", 13) - 1 + self.max_future_horizon = config.get("attention_context_right", 0) + self.context_size = self.chunk_size + self.max_past_horizon + self.max_future_horizon + + self.q_scale = (self.head_dim ** -0.5) / math.log(2) + self.k_scale = math.log(1 + math.e) / math.log(2) + self.register_buffer("softcap", torch.tensor(config.get("attention_logit_cap", 50.0), dtype=dtype), persistent=False) + + self.q_proj = ClippedLinear(self.hidden_size, self.hidden_size, device=device, dtype=dtype, ops=ops) + self.k_proj = ClippedLinear(self.hidden_size, self.hidden_size, device=device, dtype=dtype, ops=ops) + self.v_proj = ClippedLinear(self.hidden_size, self.hidden_size, device=device, dtype=dtype, ops=ops) + self.post = ClippedLinear(self.hidden_size, self.hidden_size, device=device, dtype=dtype, ops=ops) + self.per_dim_scale = nn.Parameter(torch.empty(self.head_dim, device=device, dtype=dtype)) + self.relative_k_proj = ops.Linear(self.hidden_size, self.hidden_size, bias=False, device=device, dtype=dtype) + + def _convert_to_block(self, x): + B, S, H, D = x.shape + num_blocks = (S + self.chunk_size - 1) // self.chunk_size + pad = num_blocks * self.chunk_size - S + x = torch.nn.functional.pad(x, (0, 0, 0, 0, 0, pad)) + return x.reshape(B, num_blocks, self.chunk_size, H, D).contiguous() + + def _extract_block_context(self, x): + x = torch.nn.functional.pad(x, (0, 0, 0, 0, self.max_past_horizon, self.max_future_horizon + self.chunk_size - 1)) + x = x.unfold(1, self.context_size, self.chunk_size) + return torch.movedim(x, -1, 2).contiguous() + + def _rel_shift(self, x): + B, H, NB, BS, PL = x.shape + CS = self.context_size + x = torch.nn.functional.pad(x, (0, CS + 1 - PL)) + x = x.view(B, H, NB, BS * (CS + 1)) + x = x[..., :BS * CS] + return x.view(B, H, NB, BS, CS) + + def _build_blocked_mask(self, seq_len, num_blocks, device, audio_mask=None): + """Build 5D boolean blocked attention mask (True=attend, False=mask)""" + q = torch.arange(seq_len, device=device) + dist = q[:, None] - q[None, :] + mask = (dist >= 0) & (dist < self.max_past_horizon) + if self.max_future_horizon > 0: + mask = mask | ((dist < 0) & ((-dist) < self.max_future_horizon)) + if audio_mask is not None: + mask = mask & audio_mask[0, None, :].bool() + m = mask[None, None] + # Reshape to blocked 5D matching reference code + p = num_blocks * self.chunk_size - seq_len + m = torch.nn.functional.pad(m, (0, p, 0, p), value=False) + m = m.reshape(1, 1, num_blocks, self.chunk_size, -1) + m = torch.nn.functional.pad(m, (self.max_past_horizon, self.max_future_horizon), value=False) + idx = (torch.arange(num_blocks, device=device) * self.chunk_size)[:, None] + torch.arange(self.context_size, device=device)[None, :] + return m.gather(-1, idx[None, None, :, None, :].expand(1, 1, -1, self.chunk_size, -1)) + + def forward(self, x, position_embeddings=None, attn_mask=None): + B, S, _ = x.shape + + q = self.q_proj(x).float().view(B, S, self.num_heads, self.head_dim) + k = self.k_proj(x).float().view(B, S, self.num_heads, self.head_dim) + v = self.v_proj(x).float().view(B, S, self.num_heads, self.head_dim) + + q = q * self.q_scale * torch.nn.functional.softplus(self.per_dim_scale) + k = k * self.k_scale + + q_blocks = self._convert_to_block(q) + k_context = self._extract_block_context(k) + v_context = self._extract_block_context(v) + num_blocks = q_blocks.shape[1] + + rel_k = self.relative_k_proj(position_embeddings).view(-1, self.num_heads, self.head_dim).to(q.dtype) + + queries = q_blocks.permute(0, 3, 1, 2, 4) # [B, H, NB, CS, D] + matrix_ac = queries @ k_context.permute(0, 3, 1, 4, 2) + + queries_flat = queries.reshape(B, self.num_heads, -1, self.head_dim) + matrix_bd = queries_flat @ rel_k.permute(1, 2, 0) + matrix_bd = matrix_bd.reshape(B, self.num_heads, num_blocks, self.chunk_size, -1) + matrix_bd = self._rel_shift(matrix_bd) + + attn_weights = matrix_ac + matrix_bd + attn_weights = torch.tanh(attn_weights / self.softcap) * self.softcap + + # Mask out invalid positions in chunk context (matching reference's masked_fill approach) + if attn_mask is None: + attn_mask = self._build_blocked_mask(S, num_blocks, x.device) + attn_weights = attn_weights.masked_fill(attn_mask.logical_not(), -1e9) + + attn_weights = torch.nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(v.dtype) + out = attn_weights @ v_context.permute(0, 3, 1, 2, 4) + out = out.permute(0, 2, 3, 1, 4).reshape(B, num_blocks * self.chunk_size, -1) + out = out[:, :S].contiguous() + return self.post(out.to(self.post.linear.weight.dtype)) + + +class Gemma4AudioLConv1d(nn.Module): + """Lightweight convolution with standard GLU.""" + def __init__(self, config, device=None, dtype=None, ops=None): + super().__init__() + hidden_size = config["hidden_size"] + conv_kernel_size = config.get("conv_kernel_size", 5) + self.pre_layer_norm = RMSNorm(hidden_size, eps=config["rms_norm_eps"], device=device, dtype=dtype) + self.linear_start = ClippedLinear(hidden_size, hidden_size * 2, device=device, dtype=dtype, ops=ops) + # Causal conv: left-pad only + self.depthwise_conv1d = ops.Conv1d(hidden_size, hidden_size, kernel_size=conv_kernel_size, padding=0, groups=hidden_size, bias=False, device=device, dtype=dtype) + self.conv_left_pad = conv_kernel_size - 1 # causal: pad left by kernel-1 + self.conv_norm = RMSNorm(hidden_size, eps=config["rms_norm_eps"], device=device, dtype=dtype) + self.linear_end = ClippedLinear(hidden_size, hidden_size, device=device, dtype=dtype, ops=ops) + + def forward(self, x): + residual = x + x = self.pre_layer_norm(x) + x = self.linear_start(x) + x = torch.nn.functional.glu(x, dim=-1) + x = x.transpose(1, 2) + x = torch.nn.functional.pad(x, (self.conv_left_pad, 0)) + x = self.depthwise_conv1d(x).transpose(1, 2) + x = self.conv_norm(x) + x = torch.nn.functional.silu(x) + x = self.linear_end(x) + return x + residual + + +class Gemma4AudioLayer(nn.Module): + """Conformer block: FFN1 -> Attention -> LConv -> FFN2.""" + def __init__(self, config, device=None, dtype=None, ops=None): + super().__init__() + self.feed_forward1 = Gemma4AudioFeedForward(config, device=device, dtype=dtype, ops=ops) + self.self_attn = Gemma4AudioAttention(config, device=device, dtype=dtype, ops=ops) + norm_kwargs = dict(eps=config["rms_norm_eps"], device=device, dtype=dtype) + hidden_size = config["hidden_size"] + self.norm_pre_attn = RMSNorm(hidden_size, **norm_kwargs) + self.norm_post_attn = RMSNorm(hidden_size, **norm_kwargs) + self.lconv1d = Gemma4AudioLConv1d(config, device=device, dtype=dtype, ops=ops) + self.feed_forward2 = Gemma4AudioFeedForward(config, device=device, dtype=dtype, ops=ops) + self.norm_out = RMSNorm(hidden_size, **norm_kwargs) + + def forward(self, x, position_embeddings=None, attn_mask=None): + x = self.feed_forward1(x) + + residual = x + x = self.norm_pre_attn(x) + x = self.self_attn(x, position_embeddings=position_embeddings, attn_mask=attn_mask) + x = self.norm_post_attn(x) + x = x + residual + + x = self.lconv1d(x) + x = self.feed_forward2(x) + + x = self.norm_out(x) + return x + + +class Gemma4AudioEncoder(nn.Module): + def __init__(self, config, dtype=None, device=None, ops=None): + super().__init__() + self.hidden_size = config["hidden_size"] + self.output_proj_dims = config.get("output_proj_dims", 1536) + + self.subsample_conv_projection = Gemma4AudioConvSubsampler(config, device=device, dtype=dtype, ops=ops) + self.rel_pos_enc = Gemma4AudioRelPositionalEncoding(config, device=device, dtype=dtype) + + self.layers = nn.ModuleList([ + Gemma4AudioLayer(config, device=device, dtype=dtype, ops=ops) + for _ in range(config["num_hidden_layers"]) + ]) + + self.output_proj = ops.Linear(self.hidden_size, self.output_proj_dims, bias=True, device=device, dtype=dtype) + + def forward(self, audio_features, audio_mask=None): + x, audio_mask = self.subsample_conv_projection(audio_features, audio_mask) + position_embeddings = self.rel_pos_enc(x) + + # Build blocked attention mask once for all layers + attn_mask = self.layers[0].self_attn._build_blocked_mask( + x.shape[1], (x.shape[1] + self.layers[0].self_attn.chunk_size - 1) // self.layers[0].self_attn.chunk_size, + x.device, audio_mask=audio_mask) + + for layer in self.layers: + x = layer(x, position_embeddings=position_embeddings, attn_mask=attn_mask) + + x = self.output_proj(x) + return x + + +class Gemma4AudioProjector(Gemma4RMSNormProjector): + def __init__(self, config, dtype=None, device=None, ops=None): + super().__init__(config.get("audio_output_proj_dims", 1536), config.get("text_hidden_size", 2560), dtype=dtype, device=device, ops=ops) + + +# Tokenizer and Wrappers + +class Gemma4_Tokenizer(): + tokenizer_json_data = None + + def state_dict(self): + if self.tokenizer_json_data is not None: + return {"tokenizer_json": self.tokenizer_json_data} + return {} + + def _extract_mel_spectrogram(self, waveform, sample_rate): + """Extract 128-bin log mel spectrogram. + Uses numpy for FFT/matmul/log to produce bit-identical results with reference code. + """ + # Mix to mono first, then resample to 16kHz + if waveform.dim() > 1 and waveform.shape[0] > 1: + waveform = waveform.mean(dim=0, keepdim=True) + if waveform.dim() == 1: + waveform = waveform.unsqueeze(0) + audio = waveform.squeeze(0).float().numpy() + if sample_rate != 16000: + # Use scipy's resample_poly with a high-quality FIR filter to get as close as possible to librosa's resampling (while still not full match) + from scipy.signal import resample_poly, firwin + from math import gcd + g = gcd(sample_rate, 16000) + up, down = 16000 // g, sample_rate // g + L = max(up, down) + h = firwin(160 * L + 1, 0.96 / L, window=('kaiser', 6.5)) + audio = resample_poly(audio, up, down, window=h).astype(np.float32) + n = len(audio) + + # Pad to multiple of 128, build sample-level mask + if n % 128 != 0: + audio = np.pad(audio, (0, 128 - n % 128)) + mask_raw = np.ones(len(audio), dtype=np.float32) + mask_raw[n:] = 0.0 + + # Semicausal padding: 160 zeros prepended + audio = np.pad(audio, (160, 0)) + mask_raw = np.pad(mask_raw, (160, 0)) + + # Extract 321-sample frames via stride tricks, drop last → 320 + nf = (len(audio) - 321) // 160 + 1 + strides = (audio.strides[0] * 160, audio.strides[0]) + frames = np.lib.stride_tricks.as_strided(audio, (nf, 321), strides)[..., :-1].copy() + + # Periodic Hann window, FFT magnitude, mel filterbank, log + window = (0.5 - 0.5 * np.cos(2 * np.pi * np.arange(320) / 320)).astype(np.float32) + magnitude = np.abs(np.fft.rfft(frames * window, n=512, axis=-1)) + mel_fb = self._build_mel_filterbank() + log_mel = np.log(np.matmul(magnitude, mel_fb) + np.float64(0.001)).astype(np.float32) + + # Frame mask: valid when last sample in window is real audio + mask = mask_raw[np.arange(nf) * 160 + 320].astype(bool) + log_mel = log_mel * mask[:, None] + return torch.from_numpy(log_mel), torch.from_numpy(mask) # [T, 128], [T] + + @staticmethod + def _build_mel_filterbank(): + """Build 128-bin HTK mel filterbank [257, 128] for 512-pt FFT at 16kHz.""" + mel_freqs = np.linspace(0.0, 2595.0 * np.log10(1.0 + 8000.0 / 700.0), 130) + filter_freqs = 700.0 * (10.0 ** (mel_freqs / 2595.0) - 1.0) + fft_freqs = np.linspace(0, 16000 // 2, 257) + filter_diff = np.diff(filter_freqs) + slopes = np.expand_dims(filter_freqs, 0) - np.expand_dims(fft_freqs, 1) + down_slopes = -slopes[:, :-2] / filter_diff[:-1] + up_slopes = slopes[:, 2:] / filter_diff[1:] + return np.maximum(np.zeros(1), np.minimum(down_slopes, up_slopes)) + + def tokenize_with_weights(self, text, return_word_ids=False, image=None, audio=None, video=None, llama_template=None, skip_template=True, thinking=False, **kwargs): + + # Process audio + audio_features = [] + if audio is not None: + waveform = audio["waveform"].squeeze(0) if hasattr(audio, "__getitem__") else audio + sample_rate = audio.get("sample_rate", 16000) if hasattr(audio, "get") else 16000 + mel, mel_mask = self._extract_mel_spectrogram(waveform, sample_rate) + audio_features = [(mel.unsqueeze(0), mel_mask.unsqueeze(0))] # ([1, T, 128], [1, T]) + + # Process image/video frames + is_video = video is not None + source = video if is_video else image + images = [] + if source is not None: + samples = source.movedim(-1, 1) # [B, C, H, W] + num_frames = samples.shape[0] + + # Subsample video to 1fps + if is_video: + fps = kwargs.get("fps", 24) + step = max(1, round(fps)) + indices = list(range(0, num_frames, step)) + if len(indices) == 0: + indices = [0] + samples = samples[indices] + num_frames = len(indices) + + h, w = samples.shape[2], samples.shape[3] + patch_size = 16 + pooling_k = 3 + max_soft_tokens = 70 if is_video else 280 # video uses smaller token budget per frame + max_patches = max_soft_tokens * pooling_k * pooling_k + target_px = max_patches * patch_size * patch_size + factor = (target_px / (h * w)) ** 0.5 + side_mult = pooling_k * patch_size + target_h = max(int(factor * h // side_mult) * side_mult, side_mult) + target_w = max(int(factor * w // side_mult) * side_mult, side_mult) + + import torchvision.transforms.functional as TVF + for i in range(num_frames): + # rescaling to match reference code + s = (samples[i].clamp(0, 1) * 255).to(torch.uint8) # [C, H, W] uint8 + if target_h != h or target_w != w: + s = TVF.resize(s, [target_h, target_w], interpolation=TVF.InterpolationMode.BICUBIC, antialias=True) + s = s.float() * (1.0 / 255.0) + images.append({"pixels": s.unsqueeze(0).movedim(1, -1)[:, :, :, :3], "max_soft_tokens": max_soft_tokens}) + + if text.startswith('<|turn>'): + skip_template = True + + if skip_template: + llama_text = text + else: + if llama_template is not None: + llama_text = llama_template.format(text) + else: + # Build template from modalities present + system = "<|turn>system\n<|think|>\n" if thinking else "" + media = "" + if len(images) > 0: + if is_video: + media += "\n\n" + for i in range(len(images)): + ts = f"{int(i // 60):02d}:{int(i % 60):02d}" + sep = "" if i == 0 else " " + media += f"{sep}{ts} <|image><|video|>" + media += "\n\n" + else: + media += "\n\n" + for i in range(len(images)): + if i > 0: + media += "\n\n\n\n" + media += "<|image><|image|>" + media += "\n\n" + if len(audio_features) > 0: + # Compute audio token count (always at 16kHz) + num_samples = int(waveform.shape[-1] * 16000 / sample_rate) if sample_rate != 16000 else waveform.shape[-1] + _fl = 320 # int(round(16000 * 20.0 / 1000.0)) + _hl = 160 # int(round(16000 * 10.0 / 1000.0)) + _nmel = (num_samples + _fl // 2 - (_fl + 1)) // _hl + 1 + _t = _nmel + for _ in range(2): + _t = (_t + 2 - 3) // 2 + 1 + n_audio_tokens = min(_t, 750) + media += "<|audio>" + "<|audio|>" * n_audio_tokens + "" + llama_text = f"{system}<|turn>user\n{media}{text}\n<|turn>model\n" + + text_tokens = super().tokenize_with_weights(llama_text, return_word_ids) + + def _replace_placeholders(token_list, token_id, embeds): + """Replace first placeholder with embed dict, remove remaining consecutive ones.""" + embed_idx = 0 + i = 0 + while i < len(token_list): + if token_list[i][0] == token_id and embed_idx < len(embeds): + token_list[i] = (embeds[embed_idx],) + token_list[i][1:] + embed_idx += 1 + i += 1 + while i < len(token_list) and token_list[i][0] == token_id: + token_list.pop(i) + else: + i += 1 + + if len(images) > 0: + img_token_id = 258884 if is_video else 258880 + img_embeds = [{"type": "image", "data": img["pixels"], "max_soft_tokens": img["max_soft_tokens"]} for img in images] + for r in text_tokens: + _replace_placeholders(r, img_token_id, img_embeds) + + if len(audio_features) > 0: + aud_embeds = [{"type": "audio", "data": mel, "mask": mask} for mel, mask in audio_features] + for r in text_tokens: + _replace_placeholders(r, 258881, aud_embeds) + + return text_tokens + + +class _Gemma4Tokenizer: + """Tokenizer using the tokenizers (Gemma4 doesn't come with sentencepiece model)""" + 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()) + 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)) + + +# Tokenizer +class Gemma4SDTokenizer(Gemma4_Tokenizer, sd1_clip.SDTokenizer): + embedding_size = 2560 + 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, pad_with_end=False, embedding_size=self.embedding_size, embedding_key='gemma4', tokenizer_class=_Gemma4Tokenizer, has_start_token=True, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_left=True, disable_weights=True, start_token=2, tokenizer_data=tokenizer_data) + + def decode(self, token_ids, **kwargs): + text = super().decode(token_ids, skip_special_tokens=False) + # Translate thinking channel markers to standard / tags + text = text.replace("<|channel>thought\n", "\n") + text = text.replace("", "") + # Strip remaining special tokens + text = text.replace("", "").replace("", "").strip() + return text + + +class Gemma4Tokenizer(sd1_clip.SD1Tokenizer): + tokenizer_class = Gemma4SDTokenizer + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="gemma4", tokenizer=self.tokenizer_class) + + +# Model wrappers +class Gemma4Model(sd1_clip.SDClipModel): + model_class = None + def __init__(self, device="cpu", layer="all", layer_idx=None, dtype=None, attention_mask=True, model_options={}): + self.dtypes = set() + self.dtypes.add(dtype) + 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=self.model_class, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) + + def process_tokens(self, tokens, device): + embeds, _, _, _ = super().process_tokens(tokens, device) + return embeds + + def generate(self, tokens, do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed, presence_penalty=0.0): + if isinstance(tokens, dict): + tokens = next(iter(tokens.values())) + tokens_only = [[t[0] for t in b] for b in tokens] + embeds, _, _, embeds_info = sd1_clip.SDClipModel.process_tokens(self, tokens_only, self.execution_device) + seq_len = embeds.shape[1] + ids = [0] * seq_len + expanded_idx = 0 + embed_map = {info["index"]: info["size"] for info in embeds_info} + for t in tokens_only[0]: + if expanded_idx in embed_map: + expanded_idx += embed_map[expanded_idx] + elif isinstance(t, int): + if expanded_idx < seq_len: + ids[expanded_idx] = t + expanded_idx += 1 + else: + expanded_idx += 1 + initial_token_ids = [ids] + input_ids = torch.tensor(initial_token_ids, device=self.execution_device) + return self.transformer.generate(embeds, do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed, initial_tokens=initial_token_ids[0], presence_penalty=presence_penalty, initial_input_ids=input_ids) + + +def gemma4_te(dtype_llama=None, llama_quantization_metadata=None, model_class=None): + clip_model = type('Gemma4Model_', (Gemma4Model,), {'model_class': model_class}) + class Gemma4TEModel_(sd1_clip.SD1ClipModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + if llama_quantization_metadata is not None: + model_options = model_options.copy() + model_options["quantization_metadata"] = llama_quantization_metadata + if dtype_llama is not None: + dtype = dtype_llama + super().__init__(device=device, dtype=dtype, name="gemma4", clip_model=clip_model, model_options=model_options) + return Gemma4TEModel_ + + +# Variants + +def _make_variant(config_cls): + audio = config_cls.audio_config is not None + bases = (Gemma4AudioMixin, Gemma4Base) if audio else (Gemma4Base,) + class Variant(*bases): + def __init__(self, config_dict, dtype, device, operations): + super().__init__() + self._init_model(config_cls(**config_dict), dtype, device, operations) + if audio: + self._init_audio(self.model.config, dtype, device, operations) + embedding_size = config_cls.hidden_size + if embedding_size != Gemma4SDTokenizer.embedding_size: + tok_cls = type('T', (Gemma4SDTokenizer,), {'embedding_size': embedding_size}) + class Tokenizer(Gemma4Tokenizer): + tokenizer_class = tok_cls + Variant.tokenizer = Tokenizer + else: + Variant.tokenizer = Gemma4Tokenizer + return Variant + +Gemma4_E4B = _make_variant(Gemma4Config) +Gemma4_E2B = _make_variant(Gemma4_E2B_Config) +Gemma4_31B = _make_variant(Gemma4_31B_Config) diff --git a/comfy/text_encoders/hidream_o1.py b/comfy/text_encoders/hidream_o1.py new file mode 100644 index 000000000..5d287b784 --- /dev/null +++ b/comfy/text_encoders/hidream_o1.py @@ -0,0 +1,119 @@ +"""HiDream-O1-Image tokenizer-only text encoder. + +The real Qwen3-VL backbone runs inside diffusion_model.* every step, so this +module just tokenizes the prompt into text_input_ids and emits them as +conditioning. Position ids / token_types / vinput_mask depend on target H/W +and are built later in model_base.HiDreamO1.extra_conds. +""" + +import os + +import torch +from transformers import Qwen2Tokenizer + +from comfy import sd1_clip + + +# Qwen3-VL special tokens +IM_START_ID = 151644 +IM_END_ID = 151645 +ASSISTANT_ID = 77091 +USER_ID = 872 +NEWLINE_ID = 198 +VISION_START_ID = 151652 +VISION_END_ID = 151653 +IMAGE_TOKEN_ID = 151655 +VIDEO_TOKEN_ID = 151656 +# HiDream-O1-specific tokens +BOI_TOKEN_ID = 151669 +BOR_TOKEN_ID = 151670 +EOR_TOKEN_ID = 151671 +BOT_TOKEN_ID = 151672 +TMS_TOKEN_ID = 151673 + + +class HiDreamO1QwenTokenizer(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_size=4096, + embedding_key="hidream_o1", + 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 HiDreamO1Tokenizer(sd1_clip.SD1Tokenizer): + """Wraps prompt in the upstream chat template ending with boi/tms markers. + Image tokens get spliced in at sample time once target H/W is known. + """ + + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__( + embedding_directory=embedding_directory, + tokenizer_data=tokenizer_data, + name="hidream_o1", + tokenizer=HiDreamO1QwenTokenizer, + ) + + def tokenize_with_weights(self, text, return_word_ids=False, **kwargs): + text_tokens_dict = super().tokenize_with_weights( + text, return_word_ids=return_word_ids, disable_weights=True, **kwargs + ) + text_tuples = text_tokens_dict["hidream_o1"][0] + text_tuples = [t for t in text_tuples if int(t[0]) != 151643] # strip pad + + # <|im_start|>user\n{text}<|im_end|>\n<|im_start|>assistant\n<|boi|><|tms|> + def tok(tid): + return (tid, 1.0) if not return_word_ids else (tid, 1.0, 0) + + prefix = [tok(IM_START_ID), tok(USER_ID), tok(NEWLINE_ID)] + suffix = [ + tok(IM_END_ID), tok(NEWLINE_ID), + tok(IM_START_ID), tok(ASSISTANT_ID), tok(NEWLINE_ID), + tok(BOI_TOKEN_ID), tok(TMS_TOKEN_ID), + ] + full = prefix + list(text_tuples) + suffix + return {"hidream_o1": [full]} + + +class HiDreamO1TE(torch.nn.Module): + """Passthrough TE: emits int token ids; the Qwen3-VL backbone in diffusion_model does the actual encoding.""" + + def __init__(self, device="cpu", dtype=None, model_options={}): + super().__init__() + self.dtypes = {torch.float32} + self.disable_offload = True # skips dynamic VRAM management for this zero-parameter module + self.device = torch.device("cpu") if device is None else torch.device(device) + + def encode_token_weights(self, token_weight_pairs): + tok_pairs = token_weight_pairs["hidream_o1"][0] + ids = [int(t[0]) for t in tok_pairs] + input_ids = torch.tensor([ids], dtype=torch.long) + # Surrogate keeps the cross_attn slot non-empty for CONDITIONING + # plumbing; the model reads text_input_ids out of `extra` instead. + cross_attn = input_ids.unsqueeze(-1).to(torch.float32) + extra = {"text_input_ids": input_ids} + return cross_attn, None, extra + + def load_sd(self, sd): + return [] + + def get_sd(self): + return {} + + def reset_clip_options(self): + pass + + def set_clip_options(self, options): + pass diff --git a/comfy/text_encoders/llama.py b/comfy/text_encoders/llama.py index 6ea8e36b1..5087228ca 100644 --- a/comfy/text_encoders/llama.py +++ b/comfy/text_encoders/llama.py @@ -397,7 +397,7 @@ class RMSNorm(nn.Module): -def precompute_freqs_cis(head_dim, position_ids, theta, rope_scale=None, rope_dims=None, device=None): +def precompute_freqs_cis(head_dim, position_ids, theta, rope_scale=None, rope_dims=None, device=None, interleaved_mrope=False): if not isinstance(theta, list): theta = [theta] @@ -415,16 +415,27 @@ def precompute_freqs_cis(head_dim, position_ids, theta, rope_scale=None, rope_di inv_freq_expanded = inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1) position_ids_expanded = position_ids[:, None, :].float() freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2) - emb = torch.cat((freqs, freqs), dim=-1) - cos = emb.cos() - sin = emb.sin() - if rope_dims is not None and position_ids.shape[0] > 1: - mrope_section = rope_dims * 2 - cos = torch.cat([m[i % 3] for i, m in enumerate(cos.split(mrope_section, dim=-1))], dim=-1).unsqueeze(0) - sin = torch.cat([m[i % 3] for i, m in enumerate(sin.split(mrope_section, dim=-1))], dim=-1).unsqueeze(0) + if rope_dims is not None and position_ids.shape[0] > 1 and interleaved_mrope: + # Qwen3-VL interleaved MRoPE: T-freqs by default, H/W replace every 3rd dim. + freqs_inter = freqs[0].clone() + for axis_idx, offset in ((1, 1), (2, 2)): + length = rope_dims[axis_idx] * 3 + idx = slice(offset, length, 3) + freqs_inter[..., idx] = freqs[axis_idx, ..., idx] + emb = torch.cat((freqs_inter, freqs_inter), dim=-1) + cos = emb.cos().unsqueeze(0) + sin = emb.sin().unsqueeze(0) else: - cos = cos.unsqueeze(1) - sin = sin.unsqueeze(1) + emb = torch.cat((freqs, freqs), dim=-1) + cos = emb.cos() + sin = emb.sin() + if rope_dims is not None and position_ids.shape[0] > 1: + mrope_section = rope_dims * 2 + cos = torch.cat([m[i % 3] for i, m in enumerate(cos.split(mrope_section, dim=-1))], dim=-1).unsqueeze(0) + sin = torch.cat([m[i % 3] for i, m in enumerate(sin.split(mrope_section, dim=-1))], dim=-1).unsqueeze(0) + else: + cos = cos.unsqueeze(1) + sin = sin.unsqueeze(1) sin_split = sin.shape[-1] // 2 out.append((cos, sin[..., : sin_split], -sin[..., sin_split :])) @@ -521,7 +532,7 @@ class Attention(nn.Module): else: present_key_value = (xk, xv, index + num_tokens) - if sliding_window is not None and xk.shape[2] > sliding_window: + if sliding_window is not None and xk.shape[2] > sliding_window and seq_length == 1: xk = xk[:, :, -sliding_window:] xv = xv[:, :, -sliding_window:] attention_mask = attention_mask[..., -sliding_window:] if attention_mask is not None else None @@ -533,12 +544,12 @@ class Attention(nn.Module): return self.o_proj(output), present_key_value class MLP(nn.Module): - def __init__(self, config: Llama2Config, device=None, dtype=None, ops: Any = None): + def __init__(self, config: Llama2Config, device=None, dtype=None, ops: Any = None, intermediate_size=None): super().__init__() - ops = ops or nn - self.gate_proj = ops.Linear(config.hidden_size, config.intermediate_size, bias=False, device=device, dtype=dtype) - self.up_proj = ops.Linear(config.hidden_size, config.intermediate_size, bias=False, device=device, dtype=dtype) - self.down_proj = ops.Linear(config.intermediate_size, config.hidden_size, bias=False, device=device, dtype=dtype) + intermediate_size = intermediate_size or config.intermediate_size + self.gate_proj = ops.Linear(config.hidden_size, intermediate_size, bias=False, device=device, dtype=dtype) + self.up_proj = ops.Linear(config.hidden_size, intermediate_size, bias=False, device=device, dtype=dtype) + self.down_proj = ops.Linear(intermediate_size, config.hidden_size, bias=False, device=device, dtype=dtype) if config.mlp_activation == "silu": self.activation = torch.nn.functional.silu elif config.mlp_activation == "gelu_pytorch_tanh": @@ -647,24 +658,25 @@ class TransformerBlockGemma2(nn.Module): return x, present_key_value +def _make_scaled_embedding(ops, vocab_size, hidden_size, scale, device, dtype): + class ScaledEmbedding(ops.Embedding): + def forward(self, input_ids, out_dtype=None): + return super().forward(input_ids, out_dtype=out_dtype) * scale + return ScaledEmbedding(vocab_size, hidden_size, device=device, dtype=dtype) + + class Llama2_(nn.Module): def __init__(self, config, device=None, dtype=None, ops=None): super().__init__() self.config = config self.vocab_size = config.vocab_size - self.embed_tokens = ops.Embedding( - config.vocab_size, - config.hidden_size, - device=device, - dtype=dtype - ) if self.config.transformer_type == "gemma2" or self.config.transformer_type == "gemma3": transformer = TransformerBlockGemma2 - self.normalize_in = True + self.embed_tokens = _make_scaled_embedding(ops, config.vocab_size, config.hidden_size, config.hidden_size ** 0.5, device, dtype) else: transformer = TransformerBlock - self.normalize_in = False + self.embed_tokens = ops.Embedding(config.vocab_size, config.hidden_size, device=device, dtype=dtype) self.layers = nn.ModuleList([ transformer(config, index=i, device=device, dtype=dtype, ops=ops) @@ -688,17 +700,15 @@ class Llama2_(nn.Module): self.config.rope_theta, self.config.rope_scale, self.config.rope_dims, + interleaved_mrope=getattr(self.config, "interleaved_mrope", False), device=device) - def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, position_ids=None, embeds_info=[], past_key_values=None): + def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, position_ids=None, embeds_info=[], past_key_values=None, input_ids=None): if embeds is not None: x = embeds else: x = self.embed_tokens(x, out_dtype=dtype) - if self.normalize_in: - x *= self.config.hidden_size ** 0.5 - seq_len = x.shape[1] past_len = 0 if past_key_values is not None and len(past_key_values) > 0: @@ -850,7 +860,7 @@ class BaseGenerate: torch.empty([batch, model_config.num_key_value_heads, max_cache_len, model_config.head_dim], device=device, dtype=execution_dtype), 0)) return past_key_values - def generate(self, embeds=None, do_sample=True, max_length=256, temperature=1.0, top_k=50, top_p=0.9, min_p=0.0, repetition_penalty=1.0, seed=42, stop_tokens=None, initial_tokens=[], execution_dtype=None, min_tokens=0, presence_penalty=0.0): + def generate(self, embeds=None, do_sample=True, max_length=256, temperature=1.0, top_k=50, top_p=0.9, min_p=0.0, repetition_penalty=1.0, seed=42, stop_tokens=None, initial_tokens=[], execution_dtype=None, min_tokens=0, presence_penalty=0.0, initial_input_ids=None): device = embeds.device if stop_tokens is None: @@ -875,14 +885,16 @@ class BaseGenerate: pbar = comfy.utils.ProgressBar(max_length) # Generation loop + current_input_ids = initial_input_ids for step in tqdm(range(max_length), desc="Generating tokens"): - x, _, past_key_values = self.model.forward(None, embeds=embeds, attention_mask=None, past_key_values=past_key_values) + x, _, past_key_values = self.model.forward(None, embeds=embeds, attention_mask=None, past_key_values=past_key_values, input_ids=current_input_ids) logits = self.logits(x)[:, -1] next_token = self.sample_token(logits, temperature, top_k, top_p, min_p, repetition_penalty, initial_tokens + generated_token_ids, generator, do_sample=do_sample, presence_penalty=presence_penalty) token_id = next_token[0].item() generated_token_ids.append(token_id) embeds = self.model.embed_tokens(next_token).to(execution_dtype) + current_input_ids = next_token if initial_input_ids is not None else None pbar.update(1) if token_id in stop_tokens: diff --git a/comfy/text_encoders/lt.py b/comfy/text_encoders/lt.py index 5aee1f4c0..bc5cbae28 100644 --- a/comfy/text_encoders/lt.py +++ b/comfy/text_encoders/lt.py @@ -93,8 +93,7 @@ class Gemma3_12BModel(sd1_clip.SDClipModel): def generate(self, tokens, do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed, presence_penalty): tokens_only = [[t[0] for t in b] for b in tokens] - embeds, _, _, embeds_info = self.process_tokens(tokens_only, self.execution_device) - comfy.utils.normalize_image_embeddings(embeds, embeds_info, self.transformer.model.config.hidden_size ** 0.5) + embeds, _, _, _ = self.process_tokens(tokens_only, self.execution_device) return self.transformer.generate(embeds, do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed, stop_tokens=[106], presence_penalty=presence_penalty) # 106 is class DualLinearProjection(torch.nn.Module): diff --git a/comfy/text_encoders/lumina2.py b/comfy/text_encoders/lumina2.py index 01ebdfabe..b1f1dbb9f 100644 --- a/comfy/text_encoders/lumina2.py +++ b/comfy/text_encoders/lumina2.py @@ -50,8 +50,7 @@ class Gemma3_4B_Vision_Model(sd1_clip.SDClipModel): 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.Gemma3_4B_Vision, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) def process_tokens(self, tokens, device): - embeds, _, _, embeds_info = super().process_tokens(tokens, device) - comfy.utils.normalize_image_embeddings(embeds, embeds_info, self.transformer.model.config.hidden_size ** 0.5) + embeds, _, _, _ = super().process_tokens(tokens, device) return embeds class LuminaModel(sd1_clip.SD1ClipModel): diff --git a/comfy/text_encoders/qwen35.py b/comfy/text_encoders/qwen35.py index ce9b07464..416ce9d18 100644 --- a/comfy/text_encoders/qwen35.py +++ b/comfy/text_encoders/qwen35.py @@ -408,8 +408,6 @@ class Qwen35Transformer(Llama2_): nn.Module.__init__(self) self.config = config self.vocab_size = config.vocab_size - self.normalize_in = False - self.embed_tokens = ops.Embedding(config.vocab_size, config.hidden_size, device=device, dtype=dtype) self.layers = nn.ModuleList([ Qwen35TransformerBlock(config, index=i, device=device, dtype=dtype, ops=ops) @@ -453,9 +451,8 @@ class Qwen35VisionPatchEmbed(nn.Module): self.proj = ops.Conv3d(self.in_channels, self.embed_dim, kernel_size=kernel_size, stride=kernel_size, bias=True, device=device, dtype=dtype) def forward(self, x): - target_dtype = self.proj.weight.dtype x = x.view(-1, self.in_channels, self.temporal_patch_size, self.patch_size, self.patch_size) - return self.proj(x.to(target_dtype)).view(-1, self.embed_dim) + return self.proj(x).view(-1, self.embed_dim) class Qwen35VisionMLP(nn.Module): @@ -653,7 +650,7 @@ class Qwen35VisionModel(nn.Module): x = self.patch_embed(x) pos_embeds = self.fast_pos_embed_interpolate(grid_thw).to(x.device) x = x + pos_embeds - rotary_pos_emb = self.rot_pos_emb(grid_thw) + rotary_pos_emb = self.rot_pos_emb(grid_thw).to(x.device) seq_len = x.shape[0] x = x.reshape(seq_len, -1) rotary_pos_emb = rotary_pos_emb.reshape(seq_len, -1) @@ -763,7 +760,7 @@ class Qwen35ImageTokenizer(sd1_clip.SD1Tokenizer): def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], prevent_empty_text=False, thinking=False, **kwargs): image = kwargs.get("image", None) if image is not None and len(images) == 0: - images = [image] + images = [image[i:i + 1] for i in range(image.shape[0])] skip_template = False if text.startswith('<|im_start|>'): @@ -774,13 +771,16 @@ class Qwen35ImageTokenizer(sd1_clip.SD1Tokenizer): if skip_template: llama_text = text else: - if llama_template is None: - if len(images) > 0: - llama_text = self.llama_template_images.format(text) - else: - llama_text = self.llama_template.format(text) + if llama_template is not None: + template = llama_template + elif len(images) == 0: + template = self.llama_template else: - llama_text = llama_template.format(text) + template = self.llama_template_images + if len(images) > 1: + vision_block = "<|vision_start|><|image_pad|><|vision_end|>" + template = template.replace(vision_block, vision_block * len(images), 1) + llama_text = template.format(text) if not thinking: llama_text += "\n\n" diff --git a/comfy/utils.py b/comfy/utils.py index 78c491b98..66682690a 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -1164,12 +1164,18 @@ def tiled_scale_multidim(samples, function, tile=(64, 64), overlap=8, upscale_am o = out o_d = out_div + ps_view = ps + mask_view = mask for d in range(dims): - o = o.narrow(d + 2, upscaled[d], mask.shape[d + 2]) - o_d = o_d.narrow(d + 2, upscaled[d], mask.shape[d + 2]) + l = min(ps_view.shape[d + 2], o.shape[d + 2] - upscaled[d]) + o = o.narrow(d + 2, upscaled[d], l) + o_d = o_d.narrow(d + 2, upscaled[d], l) + if l < ps_view.shape[d + 2]: + ps_view = ps_view.narrow(d + 2, 0, l) + mask_view = mask_view.narrow(d + 2, 0, l) - o.add_(ps * mask) - o_d.add_(mask) + o.add_(ps_view * mask_view) + o_d.add_(mask_view) if pbar is not None: pbar.update(1) @@ -1196,7 +1202,7 @@ def model_trange(*args, **kwargs): pbar.i1_time = time.time() pbar.set_postfix_str(" Model Initialization complete! ") elif pbar._i == 2: - #bring forward the effective start time based the the diff between first and second iteration + #bring forward the effective start time based the diff between first and second iteration #to attempt to remove load overhead from the final step rate estimate. pbar.start_t = pbar.i1_time - (time.time() - pbar.i1_time) pbar.set_postfix_str("") @@ -1390,7 +1396,7 @@ def convert_old_quants(state_dict, model_prefix="", metadata={}): k_out = "{}.weight_scale".format(layer) if layer is not None: - layer_conf = {"format": "float8_e4m3fn"} # TODO: check if anyone did some non e4m3fn scaled checkpoints + layer_conf = {"format": "float8_e4m3fn"} if full_precision_matrix_mult: layer_conf["full_precision_matrix_mult"] = full_precision_matrix_mult layers[layer] = layer_conf @@ -1446,10 +1452,3 @@ def deepcopy_list_dict(obj, memo=None): memo[obj_id] = res return res -def normalize_image_embeddings(embeds, embeds_info, scale_factor): - """Normalize image embeddings to match text embedding scale""" - for info in embeds_info: - if info.get("type") == "image": - start_idx = info["index"] - end_idx = start_idx + info["size"] - embeds[:, start_idx:end_idx, :] /= scale_factor diff --git a/comfy_api/feature_flags.py b/comfy_api/feature_flags.py index 9f6918315..adb5a3144 100644 --- a/comfy_api/feature_flags.py +++ b/comfy_api/feature_flags.py @@ -5,12 +5,95 @@ This module handles capability negotiation between frontend and backend, allowing graceful protocol evolution while maintaining backward compatibility. """ -from typing import Any +import logging +from typing import Any, TypedDict from comfy.cli_args import args + +class FeatureFlagInfo(TypedDict): + type: str + default: Any + description: str + + +# Registry of known CLI-settable feature flags. +# Launchers can query this via --list-feature-flags to discover valid flags. +CLI_FEATURE_FLAG_REGISTRY: dict[str, FeatureFlagInfo] = { + "show_signin_button": { + "type": "bool", + "default": False, + "description": "Show the sign-in button in the frontend even when not signed in", + }, +} + + +def _coerce_bool(v: str) -> bool: + """Strict bool coercion: only 'true'/'false' (case-insensitive). + + Anything else raises ValueError so the caller can warn and drop the flag, + rather than silently treating typos like 'ture' or 'yes' as False. + """ + lower = v.lower() + if lower == "true": + return True + if lower == "false": + return False + raise ValueError(f"expected 'true' or 'false', got {v!r}") + + +_COERCE_FNS: dict[str, Any] = { + "bool": _coerce_bool, + "int": lambda v: int(v), + "float": lambda v: float(v), +} + + +def _coerce_flag_value(key: str, raw_value: str) -> Any: + """Coerce a raw string value using the registry type, or keep as string. + + Returns the raw string if the key is unregistered or the type is unknown. + Raises ValueError/TypeError if the key is registered with a known type but + the value cannot be coerced; callers are expected to warn and drop the flag. + """ + info = CLI_FEATURE_FLAG_REGISTRY.get(key) + if info is None: + return raw_value + coerce = _COERCE_FNS.get(info["type"]) + if coerce is None: + return raw_value + return coerce(raw_value) + + +def _parse_cli_feature_flags() -> dict[str, Any]: + """Parse --feature-flag key=value pairs from CLI args into a dict. + + Items without '=' default to the value 'true' (bare flag form). + Flags whose value cannot be coerced to the registered type are dropped + with a warning, so a typo like '--feature-flag some_bool=ture' does not + silently take effect as the wrong value. + """ + result: dict[str, Any] = {} + for item in getattr(args, "feature_flag", []): + key, sep, raw_value = item.partition("=") + key = key.strip() + if not key: + continue + if not sep: + raw_value = "true" + try: + result[key] = _coerce_flag_value(key, raw_value.strip()) + except (ValueError, TypeError) as e: + info = CLI_FEATURE_FLAG_REGISTRY.get(key, {}) + logging.warning( + "Could not coerce --feature-flag %s=%r to %s (%s); dropping flag.", + key, raw_value.strip(), info.get("type", "?"), e, + ) + return result + + # Default server capabilities -SERVER_FEATURE_FLAGS: dict[str, Any] = { +_CORE_FEATURE_FLAGS: dict[str, Any] = { "supports_preview_metadata": True, "max_upload_size": args.max_upload_size * 1024 * 1024, # Convert MB to bytes "extension": {"manager": {"supports_v4": True}}, @@ -18,6 +101,11 @@ SERVER_FEATURE_FLAGS: dict[str, Any] = { "assets": args.enable_assets, } +# CLI-provided flags cannot overwrite core flags +_cli_flags = {k: v for k, v in _parse_cli_feature_flags().items() if k not in _CORE_FEATURE_FLAGS} + +SERVER_FEATURE_FLAGS: dict[str, Any] = {**_CORE_FEATURE_FLAGS, **_cli_flags} + def get_connection_feature( sockets_metadata: dict[str, dict[str, Any]], diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index 4942ed46c..5ed968960 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -17,6 +17,7 @@ if TYPE_CHECKING: from spandrel import ImageModelDescriptor from comfy.clip_vision import ClipVisionModel from comfy.clip_vision import Output as ClipVisionOutput_ + from comfy.bg_removal_model import BackgroundRemovalModel from comfy.controlnet import ControlNet from comfy.hooks import HookGroup, HookKeyframeGroup from comfy.model_patcher import ModelPatcher @@ -395,7 +396,6 @@ class Combo(ComfyTypeIO): @comfytype(io_type="COMBO") class MultiCombo(ComfyTypeI): '''Multiselect Combo input (dropdown for selecting potentially more than one value).''' - # TODO: something is wrong with the serialization, frontend does not recognize it as multiselect Type = list[str] class Input(Combo.Input): def __init__(self, id: str, options: list[str], display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, @@ -408,12 +408,14 @@ class MultiCombo(ComfyTypeI): self.default: list[str] def as_dict(self): - to_return = super().as_dict() | prune_dict({ - "multi_select": self.multiselect, - "placeholder": self.placeholder, - "chip": self.chip, + # Frontend expects `multi_select` to be an object config (not a boolean). + # Keep top-level `multiselect` from Combo.Input for backwards compatibility. + return super().as_dict() | prune_dict({ + "multi_select": prune_dict({ + "placeholder": self.placeholder, + "chip": self.chip, + }), }) - return to_return @comfytype(io_type="IMAGE") class Image(ComfyTypeIO): @@ -613,6 +615,11 @@ class Model(ComfyTypeIO): if TYPE_CHECKING: Type = ModelPatcher +@comfytype(io_type="BACKGROUND_REMOVAL") +class BackgroundRemoval(ComfyTypeIO): + if TYPE_CHECKING: + Type = BackgroundRemovalModel + @comfytype(io_type="CLIP_VISION") class ClipVision(ComfyTypeIO): if TYPE_CHECKING: @@ -2256,6 +2263,7 @@ __all__ = [ "ModelPatch", "ClipVision", "ClipVisionOutput", + "BackgroundRemoval", "AudioEncoder", "AudioEncoderOutput", "StyleModel", diff --git a/comfy_api/latest/_util/geometry_types.py b/comfy_api/latest/_util/geometry_types.py index b586fceb3..cdde60b10 100644 --- a/comfy_api/latest/_util/geometry_types.py +++ b/comfy_api/latest/_util/geometry_types.py @@ -12,9 +12,24 @@ class VOXEL: class MESH: - def __init__(self, vertices: torch.Tensor, faces: torch.Tensor): - self.vertices = vertices - self.faces = faces + def __init__(self, vertices: torch.Tensor, faces: torch.Tensor, + uvs: torch.Tensor | None = None, + vertex_colors: torch.Tensor | None = None, + texture: torch.Tensor | None = None, + vertex_counts: torch.Tensor | None = None, + face_counts: torch.Tensor | None = None): + + assert (vertex_counts is None) == (face_counts is None), \ + "vertex_counts and face_counts must be provided together (both or neither)" + self.vertices = vertices # vertices: (B, N, 3) + self.faces = faces # faces: (B, M, 3) + self.uvs = uvs # uvs: (B, N, 2) + self.vertex_colors = vertex_colors # vertex_colors: (B, N, 3 or 4) + self.texture = texture # texture: (B, H, W, 3) + # When vertices/faces are zero-padded to a common N/M across the batch (variable-size mesh batch), + # 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 class File3D: diff --git a/comfy_api_nodes/apis/anthropic.py b/comfy_api_nodes/apis/anthropic.py new file mode 100644 index 000000000..6cac537ea --- /dev/null +++ b/comfy_api_nodes/apis/anthropic.py @@ -0,0 +1,75 @@ +from enum import Enum +from typing import Literal + +from pydantic import BaseModel, Field + + +class AnthropicRole(str, Enum): + user = "user" + assistant = "assistant" + + +class AnthropicTextContent(BaseModel): + type: Literal["text"] = "text" + text: str = Field(...) + + +class AnthropicImageSourceBase64(BaseModel): + type: Literal["base64"] = "base64" + media_type: str = Field(..., description="MIME type of the image, e.g. image/png, image/jpeg") + data: str = Field(..., description="Base64-encoded image data") + + +class AnthropicImageSourceUrl(BaseModel): + type: Literal["url"] = "url" + url: str = Field(...) + + +class AnthropicImageContent(BaseModel): + type: Literal["image"] = "image" + source: AnthropicImageSourceBase64 | AnthropicImageSourceUrl = Field(...) + + +class AnthropicMessage(BaseModel): + role: AnthropicRole = Field(...) + content: list[AnthropicTextContent | AnthropicImageContent] = Field(...) + + +class AnthropicMessagesRequest(BaseModel): + model: str = Field(...) + messages: list[AnthropicMessage] = Field(...) + max_tokens: int = Field(..., ge=1) + system: str | None = Field(None, description="Top-level system prompt") + temperature: float | None = Field(None, ge=0.0, le=1.0) + 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) + + +class AnthropicResponseTextBlock(BaseModel): + type: Literal["text"] = "text" + text: str = Field(...) + + +class AnthropicCacheCreationUsage(BaseModel): + ephemeral_5m_input_tokens: int | None = Field(None) + ephemeral_1h_input_tokens: int | None = Field(None) + + +class AnthropicMessagesUsage(BaseModel): + input_tokens: int | None = Field(None) + output_tokens: int | None = Field(None) + cache_creation_input_tokens: int | None = Field(None) + cache_read_input_tokens: int | None = Field(None) + cache_creation: AnthropicCacheCreationUsage | None = Field(None) + + +class AnthropicMessagesResponse(BaseModel): + id: str | None = Field(None) + type: str | None = Field(None) + role: str | None = Field(None) + model: str | None = Field(None) + content: list[AnthropicResponseTextBlock] | 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/bria.py b/comfy_api_nodes/apis/bria.py index 8c496b56c..e08a519a8 100644 --- a/comfy_api_nodes/apis/bria.py +++ b/comfy_api_nodes/apis/bria.py @@ -23,7 +23,7 @@ class BriaEditImageRequest(BaseModel): None, description="Mask image (black and white). Black areas will be preserved, white areas will be edited. " "If omitted, the edit applies to the entire image. " - "The input image and the the input mask must be of the same size.", + "The input image and the input mask must be of the same size.", ) negative_prompt: str | None = Field(None) guidance_scale: float = Field(...) diff --git a/comfy_api_nodes/apis/bytedance.py b/comfy_api_nodes/apis/bytedance.py index c05bd6893..03f4c445b 100644 --- a/comfy_api_nodes/apis/bytedance.py +++ b/comfy_api_nodes/apis/bytedance.py @@ -198,6 +198,62 @@ RECOMMENDED_PRESETS_SEEDREAM_4 = [ ("Custom", None, None), ] +_PRESETS_SEEDREAM_1K = [ + ("(1K) 1024x1024 (1:1)", 1024, 1024), + ("(1K) 864x1152 (3:4)", 864, 1152), + ("(1K) 1152x864 (4:3)", 1152, 864), + ("(1K) 1312x736 (16:9)", 1312, 736), + ("(1K) 736x1312 (9:16)", 736, 1312), + ("(1K) 832x1248 (2:3)", 832, 1248), + ("(1K) 1248x832 (3:2)", 1248, 832), + ("(1K) 1568x672 (21:9)", 1568, 672), +] + +_PRESETS_SEEDREAM_2K = [ + ("(2K) 2048x2048 (1:1)", 2048, 2048), + ("(2K) 1728x2304 (3:4)", 1728, 2304), + ("(2K) 2304x1728 (4:3)", 2304, 1728), + ("(2K) 2848x1600 (16:9)", 2848, 1600), + ("(2K) 1600x2848 (9:16)", 1600, 2848), + ("(2K) 1664x2496 (2:3)", 1664, 2496), + ("(2K) 2496x1664 (3:2)", 2496, 1664), + ("(2K) 3136x1344 (21:9)", 3136, 1344), +] + +_PRESETS_SEEDREAM_3K = [ + ("(3K) 3072x3072 (1:1)", 3072, 3072), + ("(3K) 2592x3456 (3:4)", 2592, 3456), + ("(3K) 3456x2592 (4:3)", 3456, 2592), + ("(3K) 4096x2304 (16:9)", 4096, 2304), + ("(3K) 2304x4096 (9:16)", 2304, 4096), + ("(3K) 2496x3744 (2:3)", 2496, 3744), + ("(3K) 3744x2496 (3:2)", 3744, 2496), + ("(3K) 4704x2016 (21:9)", 4704, 2016), +] + +_PRESETS_SEEDREAM_4K = [ + ("(4K) 4096x4096 (1:1)", 4096, 4096), + ("(4K) 3520x4704 (3:4)", 3520, 4704), + ("(4K) 4704x3520 (4:3)", 4704, 3520), + ("(4K) 5504x3040 (16:9)", 5504, 3040), + ("(4K) 3040x5504 (9:16)", 3040, 5504), + ("(4K) 3328x4992 (2:3)", 3328, 4992), + ("(4K) 4992x3328 (3:2)", 4992, 3328), + ("(4K) 6240x2656 (21:9)", 6240, 2656), +] + +_CUSTOM_PRESET = [("Custom", None, None)] + +RECOMMENDED_PRESETS_SEEDREAM_5_LITE = ( + _PRESETS_SEEDREAM_2K + _PRESETS_SEEDREAM_3K + _PRESETS_SEEDREAM_4K + _CUSTOM_PRESET +) +RECOMMENDED_PRESETS_SEEDREAM_4_5 = ( + _PRESETS_SEEDREAM_2K + _PRESETS_SEEDREAM_4K + _CUSTOM_PRESET +) +RECOMMENDED_PRESETS_SEEDREAM_4_0 = ( + _PRESETS_SEEDREAM_1K + _PRESETS_SEEDREAM_2K + _PRESETS_SEEDREAM_4K + _CUSTOM_PRESET +) + # Seedance 2.0 reference video pixel count limits per model and output resolution. SEEDANCE2_REF_VIDEO_PIXEL_LIMITS = { "dreamina-seedance-2-0-260128": { diff --git a/comfy_api_nodes/apis/bytedance_llm.py b/comfy_api_nodes/apis/bytedance_llm.py new file mode 100644 index 000000000..654c875fc --- /dev/null +++ b/comfy_api_nodes/apis/bytedance_llm.py @@ -0,0 +1,101 @@ +"""Pydantic models for BytePlus ModelArk Responses API. + +See: https://docs.byteplus.com/en/docs/ModelArk/1585128 (request) + https://docs.byteplus.com/en/docs/ModelArk/1783703 (response) +""" + +from typing import Literal + +from pydantic import BaseModel, Field + + +class BytePlusInputText(BaseModel): + type: Literal["input_text"] = "input_text" + text: str = Field(...) + + +class BytePlusInputImage(BaseModel): + type: Literal["input_image"] = "input_image" + image_url: str = Field(..., description="Image URL or `data:image/...;base64,...` payload") + detail: str = Field("auto", description="One of high, low, auto") + + +class BytePlusInputVideo(BaseModel): + type: Literal["input_video"] = "input_video" + video_url: str = Field(..., description="Video URL or `data:video/...;base64,...` payload") + fps: float | None = Field(None, ge=0.2, le=5.0) + + +BytePlusMessageContent = BytePlusInputText | BytePlusInputImage | BytePlusInputVideo + + +class BytePlusInputMessage(BaseModel): + type: Literal["message"] = "message" + role: str = Field(..., description="One of user, system, assistant, developer") + content: list[BytePlusMessageContent] = Field(...) + + +class BytePlusResponseCreateRequest(BaseModel): + model: str = Field(...) + input: list[BytePlusInputMessage] = Field(...) + instructions: str | None = Field(None) + max_output_tokens: int | None = Field(None, ge=1) + temperature: float | None = Field(None, ge=0.0, le=2.0) + store: bool | None = Field(False) + stream: bool | None = Field(False) + + +class BytePlusOutputText(BaseModel): + type: Literal["output_text"] = "output_text" + text: str = Field(...) + + +class BytePlusOutputRefusal(BaseModel): + type: Literal["refusal"] = "refusal" + refusal: str = Field(...) + + +class BytePlusOutputContent(BaseModel): + type: str = Field(...) + text: str | None = Field(None) + refusal: str | None = Field(None) + + +class BytePlusOutputMessage(BaseModel): + type: str = Field(...) + id: str | None = Field(None) + role: str | None = Field(None) + status: str | None = Field(None) + content: list[BytePlusOutputContent] | None = Field(None) + + +class BytePlusInputTokensDetails(BaseModel): + cached_tokens: int | None = Field(None) + + +class BytePlusOutputTokensDetails(BaseModel): + reasoning_tokens: int | None = Field(None) + + +class BytePlusResponseUsage(BaseModel): + input_tokens: int | None = Field(None) + output_tokens: int | None = Field(None) + total_tokens: int | None = Field(None) + input_tokens_details: BytePlusInputTokensDetails | None = Field(None) + output_tokens_details: BytePlusOutputTokensDetails | None = Field(None) + + +class BytePlusResponseError(BaseModel): + code: str = Field(...) + message: str = Field(...) + + +class BytePlusResponseObject(BaseModel): + id: str | None = Field(None) + object: str | None = Field(None) + created_at: int | None = Field(None) + model: str | None = Field(None) + status: str | None = Field(None) + error: BytePlusResponseError | None = Field(None) + output: list[BytePlusOutputMessage] | None = Field(None) + usage: BytePlusResponseUsage | None = Field(None) diff --git a/comfy_api_nodes/apis/luma.py b/comfy_api_nodes/apis/luma.py index 632c4ab96..8c6db2022 100644 --- a/comfy_api_nodes/apis/luma.py +++ b/comfy_api_nodes/apis/luma.py @@ -1,15 +1,12 @@ from __future__ import annotations - -import torch - from enum import Enum from typing import Optional, Union +import torch from pydantic import BaseModel, Field, confloat - class LumaIO: LUMA_REF = "LUMA_REF" LUMA_CONCEPTS = "LUMA_CONCEPTS" @@ -183,13 +180,13 @@ class LumaAssets(BaseModel): class LumaImageRef(BaseModel): - '''Used for image gen''' + """Used for image gen""" url: str = Field(..., description='The URL of the image reference') weight: confloat(ge=0.0, le=1.0) = Field(..., description='The weight of the image reference') class LumaImageReference(BaseModel): - '''Used for video gen''' + """Used for video gen""" type: Optional[str] = Field('image', description='Input type, defaults to image') url: str = Field(..., description='The URL of the image') @@ -251,3 +248,32 @@ class LumaGeneration(BaseModel): assets: Optional[LumaAssets] = Field(None, description='The assets of the generation') model: str = Field(..., description='The model used for the generation') request: Union[LumaGenerationRequest, LumaImageGenerationRequest] = Field(..., description="The request used for the generation") + + +class Luma2ImageRef(BaseModel): + url: str | None = None + data: str | None = None + media_type: str | None = None + + +class Luma2GenerationRequest(BaseModel): + prompt: str = Field(..., min_length=1, max_length=6000) + model: str | None = None + type: str | None = None + aspect_ratio: str | None = None + style: str | None = None + output_format: str | None = None + web_search: bool | None = None + image_ref: list[Luma2ImageRef] | None = None + source: Luma2ImageRef | None = None + + +class Luma2Generation(BaseModel): + id: str | None = None + type: str | None = None + state: str | None = None + model: str | None = None + created_at: str | None = None + output: list[LumaImageReference] | None = None + failure_reason: str | None = None + failure_code: str | None = None diff --git a/comfy_api_nodes/apis/moonvalley.py b/comfy_api_nodes/apis/moonvalley.py deleted file mode 100644 index 7ec7a4ade..000000000 --- a/comfy_api_nodes/apis/moonvalley.py +++ /dev/null @@ -1,152 +0,0 @@ -from enum import Enum -from typing import Optional, Dict, Any - -from pydantic import BaseModel, Field, StrictBytes - - -class MoonvalleyPromptResponse(BaseModel): - error: Optional[Dict[str, Any]] = None - frame_conditioning: Optional[Dict[str, Any]] = None - id: Optional[str] = None - inference_params: Optional[Dict[str, Any]] = None - meta: Optional[Dict[str, Any]] = None - model_params: Optional[Dict[str, Any]] = None - output_url: Optional[str] = None - prompt_text: Optional[str] = None - status: Optional[str] = None - - -class MoonvalleyTextToVideoInferenceParams(BaseModel): - add_quality_guidance: Optional[bool] = Field( - True, description='Whether to add quality guidance' - ) - caching_coefficient: Optional[float] = Field( - 0.3, description='Caching coefficient for optimization' - ) - caching_cooldown: Optional[int] = Field( - 3, description='Number of caching cooldown steps' - ) - caching_warmup: Optional[int] = Field( - 3, description='Number of caching warmup steps' - ) - clip_value: Optional[float] = Field( - 3, description='CLIP value for generation control' - ) - conditioning_frame_index: Optional[int] = Field( - 0, description='Index of the conditioning frame' - ) - cooldown_steps: Optional[int] = Field( - 75, description='Number of cooldown steps (calculated based on num_frames)' - ) - fps: Optional[int] = Field( - 24, description='Frames per second of the generated video' - ) - guidance_scale: Optional[float] = Field( - 10, description='Guidance scale for generation control' - ) - height: Optional[int] = Field( - 1080, description='Height of the generated video in pixels' - ) - negative_prompt: Optional[str] = Field(None, description='Negative prompt text') - num_frames: Optional[int] = Field(64, description='Number of frames to generate') - seed: Optional[int] = Field( - None, description='Random seed for generation (default: random)' - ) - shift_value: Optional[float] = Field( - 3, description='Shift value for generation control' - ) - steps: Optional[int] = Field(80, description='Number of denoising steps') - use_guidance_schedule: Optional[bool] = Field( - True, description='Whether to use guidance scheduling' - ) - use_negative_prompts: Optional[bool] = Field( - False, description='Whether to use negative prompts' - ) - use_timestep_transform: Optional[bool] = Field( - True, description='Whether to use timestep transformation' - ) - warmup_steps: Optional[int] = Field( - 0, description='Number of warmup steps (calculated based on num_frames)' - ) - width: Optional[int] = Field( - 1920, description='Width of the generated video in pixels' - ) - - -class MoonvalleyTextToVideoRequest(BaseModel): - image_url: Optional[str] = None - inference_params: Optional[MoonvalleyTextToVideoInferenceParams] = None - prompt_text: Optional[str] = None - webhook_url: Optional[str] = None - - -class MoonvalleyUploadFileRequest(BaseModel): - file: Optional[StrictBytes] = None - - -class MoonvalleyUploadFileResponse(BaseModel): - access_url: Optional[str] = None - - -class MoonvalleyVideoToVideoInferenceParams(BaseModel): - add_quality_guidance: Optional[bool] = Field( - True, description='Whether to add quality guidance' - ) - caching_coefficient: Optional[float] = Field( - 0.3, description='Caching coefficient for optimization' - ) - caching_cooldown: Optional[int] = Field( - 3, description='Number of caching cooldown steps' - ) - caching_warmup: Optional[int] = Field( - 3, description='Number of caching warmup steps' - ) - clip_value: Optional[float] = Field( - 3, description='CLIP value for generation control' - ) - conditioning_frame_index: Optional[int] = Field( - 0, description='Index of the conditioning frame' - ) - cooldown_steps: Optional[int] = Field( - 36, description='Number of cooldown steps (calculated based on num_frames)' - ) - guidance_scale: Optional[float] = Field( - 15, description='Guidance scale for generation control' - ) - negative_prompt: Optional[str] = Field(None, description='Negative prompt text') - seed: Optional[int] = Field( - None, description='Random seed for generation (default: random)' - ) - shift_value: Optional[float] = Field( - 3, description='Shift value for generation control' - ) - steps: Optional[int] = Field(80, description='Number of denoising steps') - use_guidance_schedule: Optional[bool] = Field( - True, description='Whether to use guidance scheduling' - ) - use_negative_prompts: Optional[bool] = Field( - False, description='Whether to use negative prompts' - ) - use_timestep_transform: Optional[bool] = Field( - True, description='Whether to use timestep transformation' - ) - warmup_steps: Optional[int] = Field( - 24, description='Number of warmup steps (calculated based on num_frames)' - ) - - -class ControlType(str, Enum): - motion_control = 'motion_control' - pose_control = 'pose_control' - - -class MoonvalleyVideoToVideoRequest(BaseModel): - control_type: ControlType = Field( - ..., description='Supported types for video control' - ) - inference_params: Optional[MoonvalleyVideoToVideoInferenceParams] = None - prompt_text: str = Field(..., description='Describes the video to generate') - video_url: str = Field(..., description='Url to control video') - webhook_url: Optional[str] = Field( - None, description='Optional webhook URL for notifications' - ) diff --git a/comfy_api_nodes/apis/openai.py b/comfy_api_nodes/apis/openai.py index b85ef252b..bee75d639 100644 --- a/comfy_api_nodes/apis/openai.py +++ b/comfy_api_nodes/apis/openai.py @@ -56,14 +56,14 @@ class ModelResponseProperties(BaseModel): instructions: str | None = Field(None) max_output_tokens: int | None = Field(None) model: str | None = Field(None) - temperature: float | None = Field(1, description="Controls randomness in the response", ge=0.0, le=2.0) + temperature: float | None = Field(None, description="Controls randomness in the response", ge=0.0, le=2.0) top_p: float | None = Field( - 1, + None, description="Controls diversity of the response via nucleus sampling", ge=0.0, le=1.0, ) - truncation: str | None = Field("disabled", description="Allowed values: 'auto' or 'disabled'") + truncation: str | None = Field(None, description="Allowed values: 'auto' or 'disabled'") class ResponseProperties(BaseModel): diff --git a/comfy_api_nodes/apis/topaz.py b/comfy_api_nodes/apis/topaz.py index a9e6235a7..f91980e3d 100644 --- a/comfy_api_nodes/apis/topaz.py +++ b/comfy_api_nodes/apis/topaz.py @@ -1,4 +1,4 @@ -from typing import Optional, Union +from typing import Optional from pydantic import BaseModel, Field @@ -72,8 +72,11 @@ class VideoEnhancementFilter(BaseModel): grain: Optional[float] = Field(None, description="Grain after AI model processing") grainSize: Optional[float] = Field(None, description="Size of generated grain") recoverOriginalDetailValue: Optional[float] = Field(None, description="Source details into the output video") - creativity: Optional[str] = Field(None, description="Creativity level(high, low) for slc-1 only") + creativity: float | str | None = Field(None, description="slc-1/slp-2.5: enum (low/middle/high). ast-2: decimal 0.0-1.0.") isOptimizedMode: Optional[bool] = Field(None, description="Set to true for Starlight Creative (slc-1) only") + prompt: str | None = Field(None, description="Descriptive scene prompt (ast-2 only)") + sharp: float | None = Field(None, description="ast-2 pre-enhance sharpness") + realism: float | None = Field(None, description="ast-2 realism control") class OutputInformationVideo(BaseModel): @@ -90,7 +93,7 @@ class Overrides(BaseModel): class CreateVideoRequest(BaseModel): source: CreateVideoRequestSource = Field(...) - filters: list[Union[VideoFrameInterpolationFilter, VideoEnhancementFilter]] = Field(...) + filters: list[VideoFrameInterpolationFilter | VideoEnhancementFilter] = Field(...) output: OutputInformationVideo = Field(...) overrides: Overrides = Field(Overrides(isPaidDiffusion=True)) diff --git a/comfy_api_nodes/apis/tripo.py b/comfy_api_nodes/apis/tripo.py index ffaaa7dc1..bce6b0e89 100644 --- a/comfy_api_nodes/apis/tripo.py +++ b/comfy_api_nodes/apis/tripo.py @@ -1,10 +1,11 @@ -from __future__ import annotations from enum import Enum -from typing import Optional, List, Dict, Any, Union +from typing import Optional, 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' @@ -142,7 +143,7 @@ class TripoFileEmptyReference(BaseModel): pass class TripoFileReference(RootModel): - root: Union[TripoFileTokenReference, TripoUrlReference, TripoObjectReference, TripoFileEmptyReference] + root: TripoFileTokenReference | TripoUrlReference | TripoObjectReference | TripoFileEmptyReference class TripoGetStsTokenRequest(BaseModel): format: str = Field(..., description='The format of the image') @@ -183,7 +184,7 @@ class TripoImageToModelRequest(BaseModel): class TripoMultiviewToModelRequest(BaseModel): type: TripoTaskType = TripoTaskType.MULTIVIEW_TO_MODEL - files: List[TripoFileReference] = Field(..., description='The file references to convert to a 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') @@ -251,27 +252,13 @@ class TripoConvertModelRequest(BaseModel): 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') + 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') -class TripoTaskRequest(RootModel): - root: Union[ - TripoTextToModelRequest, - TripoImageToModelRequest, - TripoMultiviewToModelRequest, - TripoTextureModelRequest, - TripoRefineModelRequest, - TripoAnimatePrerigcheckRequest, - TripoAnimateRigRequest, - TripoAnimateRetargetRequest, - TripoStylizeModelRequest, - TripoConvertModelRequest - ] - 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') @@ -283,12 +270,13 @@ 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') + 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') + consumed_credit: int | None = Field(None) class TripoTaskResponse(BaseModel): code: int = Field(0, description='The response code') @@ -296,7 +284,7 @@ class TripoTaskResponse(BaseModel): class TripoGeneralResponse(BaseModel): code: int = Field(0, description='The response code') - data: Dict[str, str] = Field(..., description='The task ID data') + data: dict[str, str] = Field(..., description='The task ID data') class TripoBalanceData(BaseModel): balance: float = Field(..., description='The account balance') diff --git a/comfy_api_nodes/nodes_anthropic.py b/comfy_api_nodes/nodes_anthropic.py new file mode 100644 index 000000000..28dd70d4e --- /dev/null +++ b/comfy_api_nodes/nodes_anthropic.py @@ -0,0 +1,245 @@ +"""API Nodes for Anthropic Claude (Messages API). See: https://docs.anthropic.com/en/api/messages""" + +from typing_extensions import override + +from comfy_api.latest import IO, ComfyExtension, Input +from comfy_api_nodes.apis.anthropic import ( + AnthropicImageContent, + AnthropicImageSourceUrl, + AnthropicMessage, + AnthropicMessagesRequest, + AnthropicMessagesResponse, + AnthropicRole, + AnthropicTextContent, +) +from comfy_api_nodes.util import ( + ApiEndpoint, + get_number_of_images, + sync_op, + upload_images_to_comfyapi, + validate_string, +) + +ANTHROPIC_MESSAGES_ENDPOINT = "/proxy/anthropic/v1/messages" +ANTHROPIC_IMAGE_MAX_PIXELS = 1568 * 1568 +CLAUDE_MAX_IMAGES = 20 + +CLAUDE_MODELS: dict[str, str] = { + "Opus 4.7": "claude-opus-4-7", + "Opus 4.6": "claude-opus-4-6", + "Sonnet 4.6": "claude-sonnet-4-6", + "Sonnet 4.5": "claude-sonnet-4-5-20250929", + "Haiku 4.5": "claude-haiku-4-5-20251001", +} + + +def _claude_model_inputs(): + return [ + IO.Int.Input( + "max_tokens", + default=16000, + min=32, + max=32000, + tooltip="Maximum number of tokens to generate before stopping.", + advanced=True, + ), + IO.Float.Input( + "temperature", + default=1.0, + 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.", + advanced=True, + ), + ] + + +def _model_price_per_million(model: str) -> tuple[float, float] | None: + """Return (input_per_1M, output_per_1M) USD for a Claude model, or None if unknown.""" + if "opus-4-7" in model or "opus-4-6" in model or "opus-4-5" in model: + return 5.0, 25.0 + if "sonnet-4" in model: + return 3.0, 15.0 + if "haiku-4-5" in model: + return 1.0, 5.0 + return None + + +def calculate_tokens_price(response: AnthropicMessagesResponse) -> float | None: + """Compute approximate USD price from response usage. Server-side billing is authoritative.""" + if not response.usage or not response.model: + return None + rates = _model_price_per_million(response.model) + if rates is None: + return None + input_rate, output_rate = rates + input_tokens = response.usage.input_tokens or 0 + output_tokens = response.usage.output_tokens or 0 + cache_read = response.usage.cache_read_input_tokens or 0 + cache_5m = 0 + cache_1h = 0 + if response.usage.cache_creation: + cache_5m = response.usage.cache_creation.ephemeral_5m_input_tokens or 0 + cache_1h = response.usage.cache_creation.ephemeral_1h_input_tokens or 0 + total = ( + input_tokens * input_rate + + output_tokens * output_rate + + cache_read * input_rate * 0.1 + + cache_5m * input_rate * 1.25 + + cache_1h * input_rate * 2.0 + ) + return total / 1_000_000.0 + + +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) + + +async def _build_image_content_blocks( + cls: type[IO.ComfyNode], + image_tensors: list[Input.Image], +) -> list[AnthropicImageContent]: + urls = await upload_images_to_comfyapi( + cls, + image_tensors, + max_images=CLAUDE_MAX_IMAGES, + total_pixels=ANTHROPIC_IMAGE_MAX_PIXELS, + wait_label="Uploading reference images", + ) + return [AnthropicImageContent(source=AnthropicImageSourceUrl(url=url)) for url in urls] + + +class ClaudeNode(IO.ComfyNode): + """Generate text responses from an Anthropic Claude model.""" + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="ClaudeNode", + display_name="Anthropic Claude", + category="api node/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.", + inputs=[ + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Text input to the model.", + ), + IO.DynamicCombo.Input( + "model", + options=[IO.DynamicCombo.Option(label, _claude_model_inputs()) for label in CLAUDE_MODELS], + tooltip="The Claude model used to generate the response.", + ), + 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.", + ), + IO.Autogrow.Input( + "images", + template=IO.Autogrow.TemplateNames( + IO.Image.Input("image"), + names=[f"image_{i}" for i in range(1, CLAUDE_MAX_IMAGES + 1)], + min=0, + ), + tooltip=f"Optional image(s) to use as context for the model. Up to {CLAUDE_MAX_IMAGES} images.", + ), + 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, "opus") ? { + "type": "list_usd", + "usd": [0.005, 0.025], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } + : $contains($m, "sonnet") ? { + "type": "list_usd", + "usd": [0.003, 0.015], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } + : $contains($m, "haiku") ? { + "type": "list_usd", + "usd": [0.001, 0.005], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } + : {"type":"text", "text":"Token-based"} + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + prompt: str, + model: dict, + seed: int, + images: dict | None = None, + system_prompt: str = "", + ) -> 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"] + + 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: + raise ValueError(f"Up to {CLAUDE_MAX_IMAGES} images are supported per request.") + + content: list[AnthropicTextContent | AnthropicImageContent] = [] + if image_tensors: + content.extend(await _build_image_content_blocks(cls, image_tensors)) + content.append(AnthropicTextContent(text=prompt)) + + response = await sync_op( + cls, + ApiEndpoint(path=ANTHROPIC_MESSAGES_ENDPOINT, method="POST"), + response_model=AnthropicMessagesResponse, + data=AnthropicMessagesRequest( + model=CLAUDE_MODELS[model_label], + max_tokens=max_tokens, + messages=[AnthropicMessage(role=AnthropicRole.user, content=content)], + system=system_prompt or None, + temperature=temperature, + ), + price_extractor=calculate_tokens_price, + ) + return IO.NodeOutput(_get_text_from_response(response) or "Empty response from Claude model.") + + +class AnthropicExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ClaudeNode] + + +async def comfy_entrypoint() -> AnthropicExtension: + return AnthropicExtension() diff --git a/comfy_api_nodes/nodes_bfl.py b/comfy_api_nodes/nodes_bfl.py index 23590bf24..3f0ce29d8 100644 --- a/comfy_api_nodes/nodes_bfl.py +++ b/comfy_api_nodes/nodes_bfl.py @@ -596,6 +596,7 @@ class Flux2ProImageNode(IO.ComfyNode): depends_on=IO.PriceBadgeDepends(widgets=["width", "height"], inputs=["images"]), expr=cls.PRICE_BADGE_EXPR, ), + is_deprecated=True, ) @classmethod @@ -674,6 +675,175 @@ class Flux2MaxImageNode(Flux2ProImageNode): """ +_FLUX2_MODEL_ENDPOINTS = { + "Flux.2 [pro]": "/proxy/bfl/flux-2-pro/generate", + "Flux.2 [max]": "/proxy/bfl/flux-2-max/generate", +} + + +def _flux2_model_inputs(): + return [ + IO.Int.Input( + "width", + default=1024, + min=256, + max=2048, + step=32, + ), + IO.Int.Input( + "height", + default=768, + min=256, + max=2048, + step=32, + ), + IO.Autogrow.Input( + "images", + template=IO.Autogrow.TemplateNames( + IO.Image.Input("image"), + names=[f"image_{i}" for i in range(1, 9)], + min=0, + ), + tooltip="Optional reference image(s) for image-to-image generation. Up to 8 images.", + ), + ] + + +class Flux2ImageNode(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="Flux2ImageNode", + display_name="Flux.2 Image", + category="api node/image/BFL", + description="Generate images via Flux.2 [pro] or Flux.2 [max] from a prompt and optional reference images.", + inputs=[ + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Prompt for the image generation or edit", + ), + IO.DynamicCombo.Input( + "model", + options=[ + IO.DynamicCombo.Option("Flux.2 [pro]", _flux2_model_inputs()), + IO.DynamicCombo.Option("Flux.2 [max]", _flux2_model_inputs()), + ], + ), + 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( + depends_on=IO.PriceBadgeDepends( + widgets=["model", "model.width", "model.height"], + input_groups=["model.images"], + ), + expr=""" + ( + $isMax := widgets.model = "flux.2 [max]"; + $MP := 1024 * 1024; + $w := $lookup(widgets, "model.width"); + $h := $lookup(widgets, "model.height"); + $outMP := $max([1, $floor((($w * $h) + $MP - 1) / $MP)]); + $outputCost := $isMax + ? (0.07 + 0.03 * ($outMP - 1)) + : (0.03 + 0.015 * ($outMP - 1)); + $refMin := $isMax ? 0.03 : 0.015; + $refMax := $isMax ? 0.24 : 0.12; + $hasRefs := $lookup(inputGroups, "model.images") > 0; + $hasRefs + ? { + "type": "range_usd", + "min_usd": $outputCost + $refMin, + "max_usd": $outputCost + $refMax, + "format": { "approximate": true } + } + : {"type": "usd", "usd": $outputCost} + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + prompt: str, + model: dict, + seed: int, + ) -> IO.NodeOutput: + model_choice = model["model"] + endpoint = _FLUX2_MODEL_ENDPOINTS[model_choice] + width = model["width"] + height = model["height"] + images_dict = model.get("images") or {} + + image_tensors: list[Input.Image] = [t for t in images_dict.values() if t is not None] + n_images = sum(get_number_of_images(t) for t in image_tensors) + if n_images > 8: + raise ValueError("The current maximum number of supported images is 8.") + + flat_tensors: list[torch.Tensor] = [] + for tensor in image_tensors: + if len(tensor.shape) == 4: + flat_tensors.extend(tensor[i] for i in range(tensor.shape[0])) + else: + flat_tensors.append(tensor) + + reference_images: dict[str, str] = {} + for idx, tensor in enumerate(flat_tensors): + key_name = f"input_image_{idx + 1}" if idx else "input_image" + reference_images[key_name] = tensor_to_base64_string(tensor, total_pixels=2048 * 2048) + + initial_response = await sync_op( + cls, + ApiEndpoint(path=endpoint, method="POST"), + response_model=BFLFluxProGenerateResponse, + data=Flux2ProGenerateRequest( + prompt=prompt, + width=width, + height=height, + seed=seed, + **reference_images, + ), + ) + + 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 BFLExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: @@ -685,6 +855,7 @@ class BFLExtension(ComfyExtension): FluxProFillNode, Flux2ProImageNode, Flux2MaxImageNode, + Flux2ImageNode, ] diff --git a/comfy_api_nodes/nodes_bytedance.py b/comfy_api_nodes/nodes_bytedance.py index fee0ab888..d6b479336 100644 --- a/comfy_api_nodes/nodes_bytedance.py +++ b/comfy_api_nodes/nodes_bytedance.py @@ -10,6 +10,9 @@ from comfy_api.latest import IO, ComfyExtension, Input from comfy_api_nodes.apis.bytedance import ( RECOMMENDED_PRESETS, RECOMMENDED_PRESETS_SEEDREAM_4, + RECOMMENDED_PRESETS_SEEDREAM_4_0, + RECOMMENDED_PRESETS_SEEDREAM_4_5, + RECOMMENDED_PRESETS_SEEDREAM_5_LITE, SEEDANCE2_PRICE_PER_1K_TOKENS, SEEDANCE2_REF_VIDEO_PIXEL_LIMITS, VIDEO_TASKS_EXECUTION_TIME, @@ -68,6 +71,12 @@ SEEDREAM_MODELS = { "seedream-4-0-250828": "seedream-4-0-250828", } +SEEDREAM_PRESETS = { + "seedream-5-0-260128": RECOMMENDED_PRESETS_SEEDREAM_5_LITE, + "seedream-4-5-251128": RECOMMENDED_PRESETS_SEEDREAM_4_5, + "seedream-4-0-250828": RECOMMENDED_PRESETS_SEEDREAM_4_0, +} + # Long-running tasks endpoints(e.g., video) BYTEPLUS_TASK_ENDPOINT = "/proxy/byteplus/api/v3/contents/generations/tasks" BYTEPLUS_TASK_STATUS_ENDPOINT = "/proxy/byteplus/api/v3/contents/generations/tasks" # + /{task_id} @@ -562,6 +571,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode): ) """, ), + is_deprecated=True, ) @classmethod @@ -651,6 +661,226 @@ class ByteDanceSeedreamNode(IO.ComfyNode): return IO.NodeOutput(torch.cat([await download_url_to_image_tensor(i) for i in urls])) +def _seedream_model_inputs(*, max_ref_images: int, presets: list): + return [ + IO.Combo.Input( + "size_preset", + options=[label for label, _, _ in presets], + tooltip="Pick a recommended size. Select Custom to use the width and height below.", + ), + IO.Int.Input( + "width", + default=2048, + min=1024, + max=6240, + step=2, + tooltip="Custom width for image. Value is working only if `size_preset` is set to `Custom`", + ), + IO.Int.Input( + "height", + default=2048, + min=1024, + max=4992, + step=2, + tooltip="Custom height for image. Value is working only if `size_preset` is set to `Custom`", + ), + IO.Int.Input( + "max_images", + default=1, + min=1, + max=max_ref_images, + step=1, + display_mode=IO.NumberDisplay.number, + tooltip="Maximum number of images to generate. With 1, exactly one image is produced. " + "With >1, the model generates between 1 and max_images related images " + "(e.g., story scenes, character variations). " + "Total images (input + generated) cannot exceed 15.", + ), + IO.Autogrow.Input( + "images", + template=IO.Autogrow.TemplateNames( + IO.Image.Input("image"), + names=[f"image_{i}" for i in range(1, max_ref_images + 1)], + min=0, + ), + tooltip=f"Optional reference image(s) for image-to-image or multi-reference generation. " + f"Up to {max_ref_images} images.", + ), + IO.Boolean.Input( + "fail_on_partial", + default=False, + tooltip="If enabled, abort execution if any requested images are missing or return an error.", + advanced=True, + ), + ] + + +class ByteDanceSeedreamNodeV2(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="ByteDanceSeedreamNodeV2", + display_name="ByteDance Seedream 4.5 & 5.0", + category="api node/image/ByteDance", + description="Unified text-to-image generation and precise single-sentence editing at up to 4K resolution.", + inputs=[ + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Text prompt for creating or editing an image.", + ), + IO.DynamicCombo.Input( + "model", + options=[ + IO.DynamicCombo.Option( + "seedream 5.0 lite", + _seedream_model_inputs(max_ref_images=14, presets=RECOMMENDED_PRESETS_SEEDREAM_5_LITE), + ), + IO.DynamicCombo.Option( + "seedream-4-5-251128", + _seedream_model_inputs(max_ref_images=10, presets=RECOMMENDED_PRESETS_SEEDREAM_4_5), + ), + IO.DynamicCombo.Option( + "seedream-4-0-250828", + _seedream_model_inputs(max_ref_images=10, presets=RECOMMENDED_PRESETS_SEEDREAM_4_0), + ), + ], + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed to use for generation.", + ), + IO.Boolean.Input( + "watermark", + default=False, + tooltip='Whether to add an "AI generated" watermark to the image.', + advanced=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( + depends_on=IO.PriceBadgeDepends(widgets=["model"]), + expr=""" + ( + $price := $contains(widgets.model, "5.0 lite") ? 0.035 : + $contains(widgets.model, "4-5") ? 0.04 : 0.03; + { + "type":"usd", + "usd": $price, + "format": { "suffix":" x images/Run", "approximate": true } + } + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + prompt: str, + model: dict, + seed: int = 0, + watermark: bool = False, + ) -> IO.NodeOutput: + validate_string(prompt, strip_whitespace=True, min_length=1) + model_id = SEEDREAM_MODELS[model["model"]] + presets = SEEDREAM_PRESETS[model_id] + + size_preset = model.get("size_preset", presets[0][0]) + width = model.get("width", 2048) + height = model.get("height", 2048) + max_images = model.get("max_images", 1) + sequential_image_generation = "disabled" if max_images == 1 else "auto" + images_dict = model.get("images") or {} + fail_on_partial = model.get("fail_on_partial", False) + + w = h = None + for label, tw, th in presets: + if label == size_preset: + w, h = tw, th + break + if w is None or h is None: + w, h = width, height + + out_num_pixels = w * h + mp_provided = out_num_pixels / 1_000_000.0 + if ("seedream-4-5" in model_id or "seedream-5-0" in model_id) and out_num_pixels < 3686400: + raise ValueError( + f"Minimum image resolution for the selected model is 3.68MP, but {mp_provided:.2f}MP provided." + ) + if "seedream-4-0" in model_id and out_num_pixels < 921600: + raise ValueError( + f"Minimum image resolution that the selected model can generate is 0.92MP, " + f"but {mp_provided:.2f}MP provided." + ) + if out_num_pixels > 16_777_216: + raise ValueError( + f"Maximum image resolution for the selected model is 16.78MP, but {mp_provided:.2f}MP provided." + ) + + image_tensors: list[Input.Image] = [t for t in images_dict.values() if t is not None] + n_input_images = sum(get_number_of_images(t) for t in image_tensors) + max_num_of_images = 14 if model_id == "seedream-5-0-260128" else 10 + if n_input_images > max_num_of_images: + raise ValueError( + f"Maximum of {max_num_of_images} reference images are supported, but {n_input_images} received." + ) + if sequential_image_generation == "auto" and n_input_images + max_images > 15: + raise ValueError( + "The maximum number of generated images plus the number of reference images cannot exceed 15." + ) + + reference_images_urls: list[str] = [] + if image_tensors: + for tensor in image_tensors: + validate_image_aspect_ratio(tensor, (1, 3), (3, 1)) + reference_images_urls = await upload_images_to_comfyapi( + cls, + image_tensors, + max_images=n_input_images, + mime_type="image/png", + wait_label="Uploading reference images", + ) + + response = await sync_op( + cls, + ApiEndpoint(path=BYTEPLUS_IMAGE_ENDPOINT, method="POST"), + response_model=ImageTaskCreationResponse, + data=Seedream4TaskCreationRequest( + model=model_id, + prompt=prompt, + image=reference_images_urls, + size=f"{w}x{h}", + seed=seed, + sequential_image_generation=sequential_image_generation, + sequential_image_generation_options=Seedream4Options(max_images=max_images), + watermark=watermark, + ), + ) + if len(response.data) == 1: + return IO.NodeOutput(await download_url_to_image_tensor(get_image_url_from_response(response))) + urls = [str(d["url"]) for d in response.data if isinstance(d, dict) and "url" in d] + if fail_on_partial and len(urls) < len(response.data): + raise RuntimeError(f"Only {len(urls)} of {len(response.data)} images were generated before error.") + return IO.NodeOutput(torch.cat([await download_url_to_image_tensor(i) for i in urls])) + + class ByteDanceTextToVideoNode(IO.ComfyNode): @classmethod @@ -1271,7 +1501,7 @@ PRICE_BADGE_VIDEO = IO.PriceBadge( ) -def _seedance2_text_inputs(resolutions: list[str]): +def _seedance2_text_inputs(resolutions: list[str], default_ratio: str = "16:9"): return [ IO.String.Input( "prompt", @@ -1287,6 +1517,7 @@ def _seedance2_text_inputs(resolutions: list[str]): IO.Combo.Input( "ratio", options=["16:9", "4:3", "1:1", "3:4", "9:16", "21:9", "adaptive"], + default=default_ratio, tooltip="Aspect ratio of the output video.", ), IO.Int.Input( @@ -1403,7 +1634,6 @@ class ByteDance2TextToVideoNode(IO.ComfyNode): status_extractor=lambda r: r.status, price_extractor=_seedance2_price_extractor(model_id, has_video_input=False), poll_interval=9, - max_poll_attempts=180, ) return IO.NodeOutput(await download_url_to_video_output(response.content.video_url)) @@ -1421,8 +1651,14 @@ class ByteDance2FirstLastFrameNode(IO.ComfyNode): IO.DynamicCombo.Input( "model", options=[ - IO.DynamicCombo.Option("Seedance 2.0", _seedance2_text_inputs(["480p", "720p", "1080p"])), - IO.DynamicCombo.Option("Seedance 2.0 Fast", _seedance2_text_inputs(["480p", "720p"])), + IO.DynamicCombo.Option( + "Seedance 2.0", + _seedance2_text_inputs(["480p", "720p", "1080p"], default_ratio="adaptive"), + ), + IO.DynamicCombo.Option( + "Seedance 2.0 Fast", + _seedance2_text_inputs(["480p", "720p"], default_ratio="adaptive"), + ), ], tooltip="Seedance 2.0 for maximum quality; Seedance 2.0 Fast for speed optimization.", ), @@ -1585,14 +1821,13 @@ class ByteDance2FirstLastFrameNode(IO.ComfyNode): status_extractor=lambda r: r.status, price_extractor=_seedance2_price_extractor(model_id, has_video_input=False), poll_interval=9, - max_poll_attempts=180, ) return IO.NodeOutput(await download_url_to_video_output(response.content.video_url)) -def _seedance2_reference_inputs(resolutions: list[str]): +def _seedance2_reference_inputs(resolutions: list[str], default_ratio: str = "16:9"): return [ - *_seedance2_text_inputs(resolutions), + *_seedance2_text_inputs(resolutions, default_ratio=default_ratio), IO.Autogrow.Input( "reference_images", template=IO.Autogrow.TemplateNames( @@ -1670,8 +1905,14 @@ class ByteDance2ReferenceNode(IO.ComfyNode): IO.DynamicCombo.Input( "model", options=[ - IO.DynamicCombo.Option("Seedance 2.0", _seedance2_reference_inputs(["480p", "720p", "1080p"])), - IO.DynamicCombo.Option("Seedance 2.0 Fast", _seedance2_reference_inputs(["480p", "720p"])), + IO.DynamicCombo.Option( + "Seedance 2.0", + _seedance2_reference_inputs(["480p", "720p", "1080p"], default_ratio="adaptive"), + ), + IO.DynamicCombo.Option( + "Seedance 2.0 Fast", + _seedance2_reference_inputs(["480p", "720p"], default_ratio="adaptive"), + ), ], tooltip="Seedance 2.0 for maximum quality; Seedance 2.0 Fast for speed optimization.", ), @@ -1907,7 +2148,6 @@ class ByteDance2ReferenceNode(IO.ComfyNode): status_extractor=lambda r: r.status, price_extractor=_seedance2_price_extractor(model_id, has_video_input=has_video_input), poll_interval=9, - max_poll_attempts=180, ) return IO.NodeOutput(await download_url_to_video_output(response.content.video_url)) @@ -2095,6 +2335,7 @@ class ByteDanceExtension(ComfyExtension): return [ ByteDanceImageNode, ByteDanceSeedreamNode, + ByteDanceSeedreamNodeV2, ByteDanceTextToVideoNode, ByteDanceImageToVideoNode, ByteDanceFirstLastFrameNode, diff --git a/comfy_api_nodes/nodes_bytedance_llm.py b/comfy_api_nodes/nodes_bytedance_llm.py new file mode 100644 index 000000000..fa7fe370a --- /dev/null +++ b/comfy_api_nodes/nodes_bytedance_llm.py @@ -0,0 +1,271 @@ +"""API Nodes for ByteDance Seed LLM via the BytePlus ModelArk Responses API. + +See: https://docs.byteplus.com/en/docs/ModelArk/1585128 +""" + +from typing_extensions import override + +from comfy_api.latest import IO, ComfyExtension, Input +from comfy_api_nodes.apis.bytedance_llm import ( + BytePlusInputImage, + BytePlusInputMessage, + BytePlusInputText, + BytePlusInputVideo, + BytePlusMessageContent, + BytePlusResponseCreateRequest, + BytePlusResponseObject, +) +from comfy_api_nodes.util import ( + ApiEndpoint, + get_number_of_images, + sync_op, + upload_images_to_comfyapi, + upload_video_to_comfyapi, + validate_string, +) + +BYTEPLUS_RESPONSES_ENDPOINT = "/proxy/byteplus/api/v3/responses" +SEED_MAX_IMAGES = 20 +SEED_MAX_VIDEOS = 4 + +SEED_MODELS: dict[str, str] = { + "Seed 2.0 Pro": "seed-2-0-pro-260328", + "Seed 2.0 Lite": "seed-2-0-lite-260228", + "Seed 2.0 Mini": "seed-2-0-mini-260215", +} + +# USD per 1M tokens: (input, cache_hit_input, output) +_SEED_PRICES_PER_MILLION: dict[str, tuple[float, float, float]] = { + "seed-2-0-pro-260328": (0.50, 0.10, 3.00), + "seed-2-0-lite-260228": (0.25, 0.05, 2.00), + "seed-2-0-mini-260215": (0.10, 0.02, 0.40), +} + + +def _seed_model_inputs(max_images: int = SEED_MAX_IMAGES, max_videos: int = SEED_MAX_VIDEOS): + return [ + IO.Autogrow.Input( + "images", + template=IO.Autogrow.TemplateNames( + IO.Image.Input("image"), + names=[f"image_{i}" for i in range(1, max_images + 1)], + min=0, + ), + tooltip=f"Optional image(s) to use as context for the model. Up to {max_images} images.", + ), + IO.Autogrow.Input( + "videos", + template=IO.Autogrow.TemplateNames( + IO.Video.Input("video"), + names=[f"video_{i}" for i in range(1, max_videos + 1)], + min=0, + ), + tooltip=f"Optional video(s) to use as context for the model. Up to {max_videos} videos.", + ), + IO.Float.Input( + "temperature", + default=1.0, + min=0.0, + max=2.0, + step=0.01, + tooltip="Controls randomness. 0.0 is deterministic, higher values are more random.", + advanced=True, + ), + ] + + +def _calculate_price(model_id: str, response: BytePlusResponseObject) -> float | None: + """Compute approximate USD price from response usage.""" + if not response.usage: + return None + rates = _SEED_PRICES_PER_MILLION.get(model_id) + if rates is None: + return None + input_rate, cache_hit_rate, output_rate = rates + input_tokens = response.usage.input_tokens or 0 + output_tokens = response.usage.output_tokens or 0 + cached = 0 + if response.usage.input_tokens_details: + cached = response.usage.input_tokens_details.cached_tokens or 0 + fresh_input = max(0, input_tokens - cached) + total = fresh_input * input_rate + cached * cache_hit_rate + output_tokens * output_rate + return total / 1_000_000.0 + + +def _get_text_from_response(response: BytePlusResponseObject) -> str: + """Extract concatenated text from all assistant message output_text blocks.""" + if not response.output: + return "" + chunks: list[str] = [] + for item in response.output: + if item.type != "message" or not item.content: + continue + for block in item.content: + if block.type == "output_text" and block.text: + chunks.append(block.text) + elif block.type == "refusal" and block.refusal: + raise ValueError(f"Model refused to respond: {block.refusal}") + return "\n".join(chunks) + + +async def _build_image_content_blocks( + cls: type[IO.ComfyNode], + image_tensors: list[Input.Image], +) -> list[BytePlusInputImage]: + urls = await upload_images_to_comfyapi( + cls, + image_tensors, + max_images=SEED_MAX_IMAGES, + wait_label="Uploading reference images", + ) + return [BytePlusInputImage(image_url=url) for url in urls] + + +async def _build_video_content_blocks( + cls: type[IO.ComfyNode], + videos: list[Input.Video], +) -> list[BytePlusInputVideo]: + blocks: list[BytePlusInputVideo] = [] + 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(BytePlusInputVideo(video_url=url)) + return blocks + + +class ByteDanceSeedNode(IO.ComfyNode): + """Generate text responses from a ByteDance Seed 2.0 model.""" + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="ByteDanceSeedNode", + display_name="ByteDance Seed", + category="api node/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.", + inputs=[ + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Text input to the model.", + ), + IO.DynamicCombo.Input( + "model", + options=[IO.DynamicCombo.Option(label, _seed_model_inputs()) for label in SEED_MODELS], + tooltip="The Seed model used to generate the response.", + ), + 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.", + ), + 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, "mini") ? { + "type": "list_usd", + "usd": [0.00025, 0.0009], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } + : $contains($m, "lite") ? { + "type": "list_usd", + "usd": [0.0003, 0.002], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } + : $contains($m, "pro") ? { + "type": "list_usd", + "usd": [0.0005, 0.003], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } + : {"type":"text", "text":"Token-based"} + ) + """, + ), + ) + + @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_label = model["model"] + temperature = model["temperature"] + model_id = SEED_MODELS[model_label] + + image_tensors: list[Input.Image] = [t for t in (model.get("images") or {}).values() if t is not None] + if sum(get_number_of_images(t) for t in image_tensors) > SEED_MAX_IMAGES: + raise ValueError(f"Up to {SEED_MAX_IMAGES} images are supported per request.") + + video_inputs: list[Input.Video] = [v for v in (model.get("videos") or {}).values() if v is not None] + if len(video_inputs) > SEED_MAX_VIDEOS: + raise ValueError(f"Up to {SEED_MAX_VIDEOS} videos are supported per request.") + + content: list[BytePlusMessageContent] = [] + if image_tensors: + content.extend(await _build_image_content_blocks(cls, image_tensors)) + if video_inputs: + content.extend(await _build_video_content_blocks(cls, video_inputs)) + content.append(BytePlusInputText(text=prompt)) + + response = await sync_op( + cls, + ApiEndpoint(path=BYTEPLUS_RESPONSES_ENDPOINT, method="POST"), + response_model=BytePlusResponseObject, + data=BytePlusResponseCreateRequest( + model=model_id, + input=[BytePlusInputMessage(role="user", content=content)], + instructions=system_prompt or None, + temperature=temperature, + store=False, + stream=False, + ), + price_extractor=lambda r: _calculate_price(model_id, r), + ) + if response.error: + raise ValueError(f"Seed API error ({response.error.code}): {response.error.message}") + result = _get_text_from_response(response) + if not result: + raise ValueError("Empty response from Seed model.") + return IO.NodeOutput(result) + + +class ByteDanceLLMExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ByteDanceSeedNode] + + +async def comfy_entrypoint() -> ByteDanceLLMExtension: + return ByteDanceLLMExtension() diff --git a/comfy_api_nodes/nodes_gemini.py b/comfy_api_nodes/nodes_gemini.py index 2b77a022e..d18c958a8 100644 --- a/comfy_api_nodes/nodes_gemini.py +++ b/comfy_api_nodes/nodes_gemini.py @@ -83,13 +83,16 @@ class GeminiImageModel(str, Enum): async def create_image_parts( cls: type[IO.ComfyNode], - images: Input.Image, + images: Input.Image | list[Input.Image], image_limit: int = 0, ) -> list[GeminiPart]: image_parts: list[GeminiPart] = [] if image_limit < 0: raise ValueError("image_limit must be greater than or equal to 0 when creating Gemini image parts.") - total_images = get_number_of_images(images) + + # Accept either a single (possibly-batched) tensor or a list of them; share URL budget across all. + images_list: list[Input.Image] = images if isinstance(images, list) else [images] + total_images = sum(get_number_of_images(img) for img in images_list) if total_images <= 0: raise ValueError("No images provided to create_image_parts; at least one image is required.") @@ -98,10 +101,18 @@ async def create_image_parts( # Number of images we'll send as URLs (fileData) num_url_images = min(effective_max, 10) # Vertex API max number of image links + upload_kwargs: dict = {"wait_label": "Uploading reference images"} + if effective_max > num_url_images: + # Split path (e.g. 11+ images): suppress per-image counter to avoid a confusing dual-fraction label. + upload_kwargs = { + "wait_label": f"Uploading reference images ({num_url_images}+)", + "show_batch_index": False, + } reference_images_urls = await upload_images_to_comfyapi( cls, - images, + images_list, max_images=num_url_images, + **upload_kwargs, ) for reference_image_url in reference_images_urls: image_parts.append( @@ -112,15 +123,22 @@ async def create_image_parts( ) ) ) - for idx in range(num_url_images, effective_max): - image_parts.append( - GeminiPart( - inlineData=GeminiInlineData( - mimeType=GeminiMimeType.image_png, - data=tensor_to_base64_string(images[idx]), + if effective_max > num_url_images: + flat: list[torch.Tensor] = [] + for tensor in images_list: + if len(tensor.shape) == 4: + flat.extend(tensor[i] for i in range(tensor.shape[0])) + else: + flat.append(tensor) + for idx in range(num_url_images, effective_max): + image_parts.append( + GeminiPart( + inlineData=GeminiInlineData( + mimeType=GeminiMimeType.image_png, + data=tensor_to_base64_string(flat[idx]), + ) ) ) - ) return image_parts @@ -891,10 +909,6 @@ class GeminiNanoBanana2(IO.ComfyNode): "9:16", "16:9", "21:9", - # "1:4", - # "4:1", - # "8:1", - # "1:8", ], default="auto", tooltip="If set to 'auto', matches your input image's aspect ratio; " @@ -902,12 +916,7 @@ class GeminiNanoBanana2(IO.ComfyNode): ), IO.Combo.Input( "resolution", - options=[ - # "512px", - "1K", - "2K", - "4K", - ], + options=["1K", "2K", "4K"], tooltip="Target output resolution. For 2K/4K the native Gemini upscaler is used.", ), IO.Combo.Input( @@ -956,6 +965,7 @@ class GeminiNanoBanana2(IO.ComfyNode): ], is_api_node=True, price_badge=GEMINI_IMAGE_2_PRICE_BADGE, + is_deprecated=True, ) @classmethod @@ -1016,6 +1026,197 @@ class GeminiNanoBanana2(IO.ComfyNode): ) +def _nano_banana_2_v2_model_inputs(): + return [ + IO.Combo.Input( + "aspect_ratio", + options=[ + "auto", + "1:1", + "2:3", + "3:2", + "3:4", + "4:3", + "4:5", + "5:4", + "9:16", + "16:9", + "21:9", + "1:4", + "4:1", + "8:1", + "1:8", + ], + default="auto", + tooltip="If set to 'auto', matches your input image's aspect ratio; " + "if no image is provided, a 16:9 square is usually generated.", + ), + IO.Combo.Input( + "resolution", + options=["1K", "2K", "4K"], + tooltip="Target output resolution. For 2K/4K the native Gemini upscaler is used.", + ), + IO.Combo.Input( + "thinking_level", + options=["MINIMAL", "HIGH"], + ), + IO.Autogrow.Input( + "images", + template=IO.Autogrow.TemplateNames( + IO.Image.Input("image"), + names=[f"image_{i}" for i in range(1, 15)], + min=0, + ), + tooltip="Optional reference image(s). Up to 14 images total.", + ), + 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 Generate Content Input Files node.", + ), + ] + + +class GeminiNanoBanana2V2(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="GeminiNanoBanana2V2", + display_name="Nano Banana 2", + category="api node/image/Gemini", + description="Generate or edit images synchronously via Google Vertex API.", + inputs=[ + IO.String.Input( + "prompt", + multiline=True, + tooltip="Text prompt describing the image to generate or the edits to apply. " + "Include any constraints, styles, or details the model should follow.", + default="", + ), + IO.DynamicCombo.Input( + "model", + options=[ + IO.DynamicCombo.Option( + "Nano Banana 2 (Gemini 3.1 Flash Image)", + _nano_banana_2_v2_model_inputs(), + ), + ], + ), + IO.Int.Input( + "seed", + default=42, + min=0, + max=0xFFFFFFFFFFFFFFFF, + control_after_generate=True, + tooltip="When the seed is fixed to a specific value, the model makes a best effort to provide " + "the same response for repeated requests. Deterministic output isn't guaranteed. " + "Also, changing the model or parameter settings, such as the temperature, " + "can cause variations in the response even when you use the same seed value. " + "By default, a random seed value is used.", + ), + IO.Combo.Input( + "response_modalities", + options=["IMAGE", "IMAGE+TEXT"], + advanced=True, + ), + IO.String.Input( + "system_prompt", + multiline=True, + default=GEMINI_IMAGE_SYS_PROMPT, + optional=True, + tooltip="Foundational instructions that dictate an AI's behavior.", + advanced=True, + ), + ], + outputs=[ + IO.Image.Output(), + IO.String.Output(), + IO.Image.Output( + display_name="thought_image", + tooltip="First image from the model's thinking process. " + "Only available with thinking_level HIGH and IMAGE+TEXT modality.", + ), + ], + 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.resolution"]), + expr=""" + ( + $r := $lookup(widgets, "model.resolution"); + $prices := {"1k": 0.0696, "2k": 0.1014, "4k": 0.154}; + {"type":"usd","usd": $lookup($prices, $r), "format":{"suffix":"/Image","approximate":true}} + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + prompt: str, + model: dict, + seed: int, + response_modalities: str, + system_prompt: str = "", + ) -> IO.NodeOutput: + validate_string(prompt, strip_whitespace=True, min_length=1) + model_choice = model["model"] + if model_choice == "Nano Banana 2 (Gemini 3.1 Flash Image)": + model_id = "gemini-3.1-flash-image-preview" + else: + model_id = model_choice + + images = model.get("images") or {} + parts: list[GeminiPart] = [GeminiPart(text=prompt)] + if images: + image_tensors: list[Input.Image] = [t for t in images.values() if t is not None] + if image_tensors: + if sum(get_number_of_images(t) for t in image_tensors) > 14: + raise ValueError("The current maximum number of supported images is 14.") + parts.extend(await create_image_parts(cls, image_tensors)) + files = model.get("files") + if files is not None: + parts.extend(files) + + image_config = GeminiImageConfig(imageSize=model["resolution"]) + if model["aspect_ratio"] != "auto": + image_config.aspectRatio = model["aspect_ratio"] + + gemini_system_prompt = None + if system_prompt: + gemini_system_prompt = GeminiSystemInstructionContent(parts=[GeminiTextPart(text=system_prompt)], role=None) + + response = await sync_op( + cls, + ApiEndpoint(path=f"/proxy/vertexai/gemini/{model_id}", method="POST"), + data=GeminiImageGenerateContentRequest( + contents=[ + GeminiContent(role=GeminiRole.user, parts=parts), + ], + generationConfig=GeminiImageGenerationConfig( + responseModalities=(["IMAGE"] if response_modalities == "IMAGE" else ["TEXT", "IMAGE"]), + imageConfig=image_config, + thinkingConfig=GeminiThinkingConfig(thinkingLevel=model["thinking_level"]), + ), + systemInstruction=gemini_system_prompt, + ), + response_model=GeminiGenerateContentResponse, + price_extractor=calculate_tokens_price, + ) + return IO.NodeOutput( + await get_image_from_response(response), + get_text_from_response(response), + await get_image_from_response(response, thought=True), + ) + + class GeminiExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: @@ -1024,6 +1225,7 @@ class GeminiExtension(ComfyExtension): GeminiImage, GeminiImage2, GeminiNanoBanana2, + GeminiNanoBanana2V2, GeminiInputFiles, ] diff --git a/comfy_api_nodes/nodes_grok.py b/comfy_api_nodes/nodes_grok.py index f42d84616..a103f24ee 100644 --- a/comfy_api_nodes/nodes_grok.py +++ b/comfy_api_nodes/nodes_grok.py @@ -54,7 +54,12 @@ class GrokImageNode(IO.ComfyNode): inputs=[ IO.Combo.Input( "model", - options=["grok-imagine-image-pro", "grok-imagine-image", "grok-imagine-image-beta"], + options=[ + "grok-imagine-image-quality", + "grok-imagine-image-pro", + "grok-imagine-image", + "grok-imagine-image-beta", + ], ), IO.String.Input( "prompt", @@ -111,10 +116,12 @@ class GrokImageNode(IO.ComfyNode): ], is_api_node=True, price_badge=IO.PriceBadge( - depends_on=IO.PriceBadgeDepends(widgets=["model", "number_of_images"]), + depends_on=IO.PriceBadgeDepends(widgets=["model", "number_of_images", "resolution"]), expr=""" ( - $rate := $contains(widgets.model, "pro") ? 0.07 : 0.02; + $rate := widgets.model = "grok-imagine-image-quality" + ? (widgets.resolution = "1k" ? 0.05 : 0.07) + : ($contains(widgets.model, "pro") ? 0.07 : 0.02); {"type":"usd","usd": $rate * widgets.number_of_images} ) """, @@ -155,6 +162,61 @@ class GrokImageNode(IO.ComfyNode): ) +_GROK_IMAGE_EDIT_ASPECT_RATIO_OPTIONS = [ + "auto", + "1:1", + "2:3", + "3:2", + "3:4", + "4:3", + "9:16", + "16:9", + "9:19.5", + "19.5:9", + "9:20", + "20:9", + "1:2", + "2:1", +] + + +def _grok_image_edit_model_inputs(*, max_ref_images: int, with_aspect_ratio: bool): + inputs = [ + IO.Autogrow.Input( + "images", + template=IO.Autogrow.TemplateNames( + IO.Image.Input("image"), + names=[f"image_{i}" for i in range(1, max_ref_images + 1)], + min=1, + ), + tooltip=( + "Reference image to edit." + if max_ref_images == 1 + else f"Reference image(s) to edit. Up to {max_ref_images} images." + ), + ), + IO.Combo.Input("resolution", options=["1K", "2K"]), + IO.Int.Input( + "number_of_images", + default=1, + min=1, + max=10, + step=1, + tooltip="Number of edited images to generate", + display_mode=IO.NumberDisplay.number, + ), + ] + if with_aspect_ratio: + inputs.append( + IO.Combo.Input( + "aspect_ratio", + options=_GROK_IMAGE_EDIT_ASPECT_RATIO_OPTIONS, + tooltip="Only allowed when multiple images are connected.", + ) + ) + return inputs + + class GrokImageEditNode(IO.ComfyNode): @classmethod @@ -167,7 +229,12 @@ class GrokImageEditNode(IO.ComfyNode): inputs=[ IO.Combo.Input( "model", - options=["grok-imagine-image-pro", "grok-imagine-image", "grok-imagine-image-beta"], + options=[ + "grok-imagine-image-quality", + "grok-imagine-image-pro", + "grok-imagine-image", + "grok-imagine-image-beta", + ], ), IO.Image.Input("image", display_name="images"), IO.String.Input( @@ -228,14 +295,23 @@ class GrokImageEditNode(IO.ComfyNode): ], is_api_node=True, price_badge=IO.PriceBadge( - depends_on=IO.PriceBadgeDepends(widgets=["model", "number_of_images"]), + depends_on=IO.PriceBadgeDepends(widgets=["model", "number_of_images", "resolution"]), expr=""" ( - $rate := $contains(widgets.model, "pro") ? 0.07 : 0.02; - {"type":"usd","usd": 0.002 + $rate * widgets.number_of_images} + $isQualityModel := widgets.model = "grok-imagine-image-quality"; + $isPro := $contains(widgets.model, "pro"); + $rate := $isQualityModel + ? (widgets.resolution = "1k" ? 0.05 : 0.07) + : ($isPro ? 0.07 : 0.02); + $base := $isQualityModel ? 0.01 : 0.002; + $output := $rate * widgets.number_of_images; + $isPro + ? {"type":"usd","usd": $base + $output} + : {"type":"range_usd","min_usd": $base + $output, "max_usd": 3 * $base + $output} ) """, ), + is_deprecated=True, ) @classmethod @@ -283,6 +359,143 @@ class GrokImageEditNode(IO.ComfyNode): ) +class GrokImageEditNodeV2(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="GrokImageEditNodeV2", + display_name="Grok Image Edit", + category="api node/image/Grok", + description="Modify an existing image based on a text prompt", + inputs=[ + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="The text prompt used to generate the image", + ), + IO.DynamicCombo.Input( + "model", + options=[ + IO.DynamicCombo.Option( + "grok-imagine-image-quality", + _grok_image_edit_model_inputs(max_ref_images=3, with_aspect_ratio=True), + ), + IO.DynamicCombo.Option( + "grok-imagine-image-pro", + _grok_image_edit_model_inputs(max_ref_images=1, with_aspect_ratio=False), + ), + IO.DynamicCombo.Option( + "grok-imagine-image", + _grok_image_edit_model_inputs(max_ref_images=3, with_aspect_ratio=True), + ), + ], + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed to determine if node should re-run; " + "actual results are nondeterministic regardless of seed.", + ), + ], + 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.resolution", "model.number_of_images"], + ), + expr=""" + ( + $isQualityModel := widgets.model = "grok-imagine-image-quality"; + $isPro := $contains(widgets.model, "pro"); + $res := $lookup(widgets, "model.resolution"); + $n := $lookup(widgets, "model.number_of_images"); + $rate := $isQualityModel + ? ($res = "1k" ? 0.05 : 0.07) + : ($isPro ? 0.07 : 0.02); + $base := $isQualityModel ? 0.01 : 0.002; + $output := $rate * $n; + $isPro + ? {"type":"usd","usd": $base + $output} + : {"type":"range_usd","min_usd": $base + $output, "max_usd": 3 * $base + $output} + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + prompt: str, + model: dict, + seed: int, + ) -> IO.NodeOutput: + validate_string(prompt, strip_whitespace=True, min_length=1) + model_id = model["model"] + resolution = model["resolution"] + number_of_images = model["number_of_images"] + images_dict = model.get("images") or {} + aspect_ratio = model.get("aspect_ratio", "auto") + + image_tensors: list[Input.Image] = [t for t in images_dict.values() if t is not None] + n_images = sum(get_number_of_images(t) for t in image_tensors) + if n_images < 1: + raise ValueError("At least one image is required for editing.") + if model_id == "grok-imagine-image-pro" and n_images > 1: + raise ValueError("The pro model supports only 1 input image.") + if model_id != "grok-imagine-image-pro" and n_images > 3: + raise ValueError("A maximum of 3 input images is supported.") + if aspect_ratio != "auto" and n_images == 1: + raise ValueError( + "Custom aspect ratio is only allowed when multiple images are connected to the image input." + ) + + flat_tensors: list[torch.Tensor] = [] + for tensor in image_tensors: + if len(tensor.shape) == 4: + flat_tensors.extend(tensor[i] for i in range(tensor.shape[0])) + else: + flat_tensors.append(tensor) + + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/xai/v1/images/edits", method="POST"), + data=ImageEditRequest( + model=model_id, + images=[ + InputUrlObject(url=f"data:image/png;base64,{tensor_to_base64_string(i)}") for i in flat_tensors + ], + prompt=prompt, + resolution=resolution.lower(), + n=number_of_images, + seed=seed, + aspect_ratio=None if aspect_ratio == "auto" else aspect_ratio, + ), + response_model=ImageGenerationResponse, + price_extractor=_extract_grok_price, + ) + if len(response.data) == 1: + return IO.NodeOutput(await download_url_to_image_tensor(response.data[0].url)) + return IO.NodeOutput( + torch.cat( + [await download_url_to_image_tensor(i) for i in [str(d.url) for d in response.data if d.url]], + ) + ) + + class GrokVideoNode(IO.ComfyNode): @classmethod @@ -717,6 +930,7 @@ class GrokExtension(ComfyExtension): return [ GrokImageNode, GrokImageEditNode, + GrokImageEditNodeV2, GrokVideoNode, GrokVideoReferenceNode, GrokVideoEditNode, diff --git a/comfy_api_nodes/nodes_hitpaw.py b/comfy_api_nodes/nodes_hitpaw.py index 488080a74..bca5170e4 100644 --- a/comfy_api_nodes/nodes_hitpaw.py +++ b/comfy_api_nodes/nodes_hitpaw.py @@ -178,7 +178,6 @@ class HitPawGeneralImageEnhance(IO.ComfyNode): status_extractor=lambda x: x.data.status, price_extractor=lambda x: request_price, poll_interval=10.0, - max_poll_attempts=480, ) return IO.NodeOutput(await download_url_to_image_tensor(final_response.data.res_url)) @@ -324,7 +323,6 @@ class HitPawVideoEnhance(IO.ComfyNode): status_extractor=lambda x: x.data.status, price_extractor=lambda x: request_price, poll_interval=10.0, - max_poll_attempts=320, ) return IO.NodeOutput(await download_url_to_video_output(final_response.data.res_url)) diff --git a/comfy_api_nodes/nodes_kling.py b/comfy_api_nodes/nodes_kling.py index 709b3726c..7586f1816 100644 --- a/comfy_api_nodes/nodes_kling.py +++ b/comfy_api_nodes/nodes_kling.py @@ -276,7 +276,6 @@ async def finish_omni_video_task(cls: type[IO.ComfyNode], response: TaskStatusRe cls, ApiEndpoint(path=f"/proxy/kling/v1/videos/omni-video/{response.data.task_id}"), response_model=TaskStatusResponse, - max_poll_attempts=280, status_extractor=lambda r: (r.data.task_status if r.data else None), ) return IO.NodeOutput(await download_url_to_video_output(final_response.data.task_result.videos[0].url)) @@ -2788,11 +2787,15 @@ class MotionControl(IO.ComfyNode): ], is_api_node=True, price_badge=IO.PriceBadge( - depends_on=IO.PriceBadgeDepends(widgets=["mode"]), + depends_on=IO.PriceBadgeDepends(widgets=["mode", "model"]), expr=""" ( - $prices := {"std": 0.07, "pro": 0.112}; - {"type":"usd","usd": $lookup($prices, widgets.mode), "format":{"suffix":"/second"}} + $prices := { + "kling-v3": {"std": 0.126, "pro": 0.168}, + "kling-v2-6": {"std": 0.07, "pro": 0.112} + }; + $modelPrices := $lookup($prices, widgets.model); + {"type":"usd","usd": $lookup($modelPrices, widgets.mode), "format":{"suffix":"/second"}} ) """, ), @@ -3062,7 +3065,6 @@ class KlingVideoNode(IO.ComfyNode): cls, ApiEndpoint(path=poll_path), response_model=TaskStatusResponse, - max_poll_attempts=280, status_extractor=lambda r: (r.data.task_status if r.data else None), ) return IO.NodeOutput(await download_url_to_video_output(final_response.data.task_result.videos[0].url)) @@ -3188,7 +3190,6 @@ class KlingFirstLastFrameNode(IO.ComfyNode): cls, ApiEndpoint(path=f"/proxy/kling/v1/videos/image2video/{response.data.task_id}"), response_model=TaskStatusResponse, - max_poll_attempts=280, status_extractor=lambda r: (r.data.task_status if r.data else None), ) return IO.NodeOutput(await download_url_to_video_output(final_response.data.task_result.videos[0].url)) diff --git a/comfy_api_nodes/nodes_luma.py b/comfy_api_nodes/nodes_luma.py index 9ed6cd299..d92a7c382 100644 --- a/comfy_api_nodes/nodes_luma.py +++ b/comfy_api_nodes/nodes_luma.py @@ -1,10 +1,11 @@ -from typing import Optional - import torch from typing_extensions import override -from comfy_api.latest import IO, ComfyExtension +from comfy_api.latest import IO, ComfyExtension, Input from comfy_api_nodes.apis.luma import ( + Luma2Generation, + Luma2GenerationRequest, + Luma2ImageRef, LumaAspectRatio, LumaCharacterRef, LumaConceptChain, @@ -30,6 +31,7 @@ from comfy_api_nodes.util import ( download_url_to_video_output, poll_op, sync_op, + upload_image_to_comfyapi, upload_images_to_comfyapi, validate_string, ) @@ -212,9 +214,9 @@ class LumaImageGenerationNode(IO.ComfyNode): aspect_ratio: str, seed, style_image_weight: float, - image_luma_ref: Optional[LumaReferenceChain] = None, - style_image: Optional[torch.Tensor] = None, - character_image: Optional[torch.Tensor] = None, + image_luma_ref: LumaReferenceChain | None = None, + style_image: torch.Tensor | None = None, + character_image: torch.Tensor | None = None, ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=True, min_length=3) # handle image_luma_ref @@ -434,7 +436,7 @@ class LumaTextToVideoGenerationNode(IO.ComfyNode): duration: str, loop: bool, seed, - luma_concepts: Optional[LumaConceptChain] = None, + luma_concepts: LumaConceptChain | None = None, ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=False, min_length=3) duration = duration if model != LumaVideoModel.ray_1_6 else None @@ -533,7 +535,6 @@ class LumaImageToVideoGenerationNode(IO.ComfyNode): ], is_api_node=True, price_badge=PRICE_BADGE_VIDEO, - ) @classmethod @@ -644,6 +645,293 @@ PRICE_BADGE_VIDEO = IO.PriceBadge( ) +def _luma2_uni1_common_inputs(max_image_refs: int) -> list: + return [ + IO.Combo.Input( + "style", + options=["auto", "manga"], + default="auto", + tooltip="Style preset. 'auto' picks based on the prompt; " + "'manga' applies a manga/anime aesthetic and requires a portrait " + "aspect ratio (2:3, 9:16, 1:2, 1:3).", + ), + IO.Boolean.Input( + "web_search", + default=False, + tooltip="Search the web for visual references before generating.", + ), + IO.Autogrow.Input( + "image_ref", + template=IO.Autogrow.TemplateNames( + IO.Image.Input("image"), + names=[f"image_{i}" for i in range(1, max_image_refs + 1)], + min=0, + ), + optional=True, + tooltip=f"Up to {max_image_refs} reference images for style/content guidance.", + ), + ] + + +async def _luma2_upload_image_refs( + cls: type[IO.ComfyNode], + refs: dict | None, + max_count: int, +) -> list[Luma2ImageRef] | None: + if not refs: + return None + out: list[Luma2ImageRef] = [] + for key in refs: + url = await upload_image_to_comfyapi(cls, refs[key]) + out.append(Luma2ImageRef(url=url)) + if len(out) > max_count: + raise ValueError(f"Maximum {max_count} reference images are allowed.") + return out or None + + +async def _luma2_submit_and_poll( + cls: type[IO.ComfyNode], + request: Luma2GenerationRequest, +) -> Input.Image: + initial = await sync_op( + cls, + ApiEndpoint(path="/proxy/luma_2/generations", method="POST"), + response_model=Luma2Generation, + data=request, + ) + if not initial.id: + raise RuntimeError("Luma 2 API did not return a generation id.") + final = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/luma_2/generations/{initial.id}", method="GET"), + response_model=Luma2Generation, + status_extractor=lambda r: r.state, + progress_extractor=lambda r: None, + ) + if not final.output: + msg = final.failure_reason or "no output returned" + raise RuntimeError(f"Luma 2 generation failed: {msg}") + url = final.output[0].url + if not url: + raise RuntimeError("Luma 2 generation completed without an output URL.") + return await download_url_to_image_tensor(url) + + +class LumaImageNode(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="LumaImageNode2", + display_name="Luma UNI-1 Image", + category="api node/image/Luma", + description="Generate images from text using the Luma UNI-1 model.", + inputs=[ + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Text description of the desired image. 1–6000 characters.", + ), + IO.DynamicCombo.Input( + "model", + options=[ + IO.DynamicCombo.Option( + "uni-1", + [ + IO.Combo.Input( + "aspect_ratio", + options=[ + "auto", + "3:1", + "2:1", + "16:9", + "3:2", + "1:1", + "2:3", + "9:16", + "1:2", + "1:3", + ], + default="auto", + tooltip="Output image aspect ratio. 'auto' lets " + "the model pick based on the prompt.", + ), + *_luma2_uni1_common_inputs(max_image_refs=9), + ], + ), + IO.DynamicCombo.Option( + "uni-1-max", + [ + IO.Combo.Input( + "aspect_ratio", + options=[ + "auto", + "3:1", + "2:1", + "16:9", + "3:2", + "1:1", + "2:3", + "9:16", + "1:2", + "1:3", + ], + default="auto", + tooltip="Output image aspect ratio. 'auto' lets " + "the model pick based on the prompt.", + ), + *_luma2_uni1_common_inputs(max_image_refs=9), + ], + ), + ], + tooltip="Model to use for generation.", + ), + 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.", + ), + ], + 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"], input_groups=["model.image_ref"]), + expr=""" + ( + $m := widgets.model; + $refs := $lookup(inputGroups, "model.image_ref"); + $base := $m = "uni-1-max" ? 0.1 : 0.0404; + {"type":"usd","usd": $round($base + 0.003 * $refs, 4)} + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + prompt: str, + model: dict, + seed: int, + ) -> IO.NodeOutput: + validate_string(prompt, min_length=1, max_length=6000) + aspect_ratio = model["aspect_ratio"] + style = model["style"] + allowed_manga_ratios = {"2:3", "9:16", "1:2", "1:3"} + if style == "manga" and aspect_ratio != "auto" and aspect_ratio not in allowed_manga_ratios: + raise ValueError( + f"'manga' style requires a portrait aspect ratio " + f"({', '.join(sorted(allowed_manga_ratios))}) or 'auto'; got '{aspect_ratio}'." + ) + request = Luma2GenerationRequest( + prompt=prompt, + model=model["model"], + type="image", + aspect_ratio=aspect_ratio if aspect_ratio != "auto" else None, + style=style if style != "auto" else None, + output_format="png", + web_search=model["web_search"], + image_ref=await _luma2_upload_image_refs(cls, model.get("image_ref"), max_count=9), + ) + return IO.NodeOutput(await _luma2_submit_and_poll(cls, request)) + + +class LumaImageEditNode(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="LumaImageEditNode2", + display_name="Luma UNI-1 Image Edit", + category="api node/image/Luma", + description="Edit an existing image with a text prompt using the Luma UNI-1 model.", + inputs=[ + IO.Image.Input( + "source", + tooltip="Source image to edit.", + ), + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Description of the desired edit. 1–6000 characters.", + ), + IO.DynamicCombo.Input( + "model", + options=[ + IO.DynamicCombo.Option( + "uni-1", + _luma2_uni1_common_inputs(max_image_refs=8), + ), + IO.DynamicCombo.Option( + "uni-1-max", + _luma2_uni1_common_inputs(max_image_refs=8), + ), + ], + tooltip="Model to use for editing.", + ), + 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.", + ), + ], + 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"], input_groups=["model.image_ref"]), + expr=""" + ( + $m := widgets.model; + $refs := $lookup(inputGroups, "model.image_ref"); + $base := $m = "uni-1-max" ? 0.103 : 0.0434; + {"type":"usd","usd": $round($base + 0.003 * $refs, 4)} + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + source: Input.Image, + prompt: str, + model: dict, + seed: int, + ) -> IO.NodeOutput: + validate_string(prompt, min_length=1, max_length=6000) + request = Luma2GenerationRequest( + prompt=prompt, + model=model["model"], + type="image_edit", + source=Luma2ImageRef(url=await upload_image_to_comfyapi(cls, source)), + style=model["style"] if model["style"] != "auto" else None, + output_format="png", + web_search=model["web_search"], + image_ref=await _luma2_upload_image_refs(cls, model.get("image_ref"), max_count=8), + ) + return IO.NodeOutput(await _luma2_submit_and_poll(cls, request)) + + class LumaExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: @@ -654,6 +942,8 @@ class LumaExtension(ComfyExtension): LumaImageToVideoGenerationNode, LumaReferenceNode, LumaConceptsNode, + LumaImageNode, + LumaImageEditNode, ] diff --git a/comfy_api_nodes/nodes_magnific.py b/comfy_api_nodes/nodes_magnific.py index 0f53208d4..38b881fea 100644 --- a/comfy_api_nodes/nodes_magnific.py +++ b/comfy_api_nodes/nodes_magnific.py @@ -230,7 +230,6 @@ class MagnificImageUpscalerCreativeNode(IO.ComfyNode): status_extractor=lambda x: x.status, price_extractor=lambda _: price_usd, poll_interval=10.0, - max_poll_attempts=480, ) return IO.NodeOutput(await download_url_to_image_tensor(final_response.generated[0])) @@ -391,7 +390,6 @@ class MagnificImageUpscalerPreciseV2Node(IO.ComfyNode): status_extractor=lambda x: x.status, price_extractor=lambda _: price_usd, poll_interval=10.0, - max_poll_attempts=480, ) return IO.NodeOutput(await download_url_to_image_tensor(final_response.generated[0])) @@ -541,7 +539,6 @@ class MagnificImageStyleTransferNode(IO.ComfyNode): response_model=TaskResponse, status_extractor=lambda x: x.status, poll_interval=10.0, - max_poll_attempts=480, ) return IO.NodeOutput(await download_url_to_image_tensor(final_response.generated[0])) @@ -782,7 +779,6 @@ class MagnificImageRelightNode(IO.ComfyNode): response_model=TaskResponse, status_extractor=lambda x: x.status, poll_interval=10.0, - max_poll_attempts=480, ) return IO.NodeOutput(await download_url_to_image_tensor(final_response.generated[0])) @@ -924,7 +920,6 @@ class MagnificImageSkinEnhancerNode(IO.ComfyNode): response_model=TaskResponse, status_extractor=lambda x: x.status, poll_interval=10.0, - max_poll_attempts=480, ) return IO.NodeOutput(await download_url_to_image_tensor(final_response.generated[0])) diff --git a/comfy_api_nodes/nodes_moonvalley.py b/comfy_api_nodes/nodes_moonvalley.py deleted file mode 100644 index 78a230529..000000000 --- a/comfy_api_nodes/nodes_moonvalley.py +++ /dev/null @@ -1,534 +0,0 @@ -import logging - -from typing_extensions import override - -from comfy_api.latest import IO, ComfyExtension, Input -from comfy_api_nodes.apis.moonvalley import ( - MoonvalleyPromptResponse, - MoonvalleyTextToVideoInferenceParams, - MoonvalleyTextToVideoRequest, - MoonvalleyVideoToVideoInferenceParams, - MoonvalleyVideoToVideoRequest, -) -from comfy_api_nodes.util import ( - ApiEndpoint, - download_url_to_video_output, - poll_op, - sync_op, - trim_video, - upload_images_to_comfyapi, - upload_video_to_comfyapi, - validate_container_format_is_mp4, - validate_image_dimensions, - validate_string, -) - -API_UPLOADS_ENDPOINT = "/proxy/moonvalley/uploads" -API_PROMPTS_ENDPOINT = "/proxy/moonvalley/prompts" -API_VIDEO2VIDEO_ENDPOINT = "/proxy/moonvalley/prompts/video-to-video" -API_TXT2VIDEO_ENDPOINT = "/proxy/moonvalley/prompts/text-to-video" -API_IMG2VIDEO_ENDPOINT = "/proxy/moonvalley/prompts/image-to-video" - -MIN_WIDTH = 300 -MIN_HEIGHT = 300 - -MAX_WIDTH = 10000 -MAX_HEIGHT = 10000 - -MIN_VID_WIDTH = 300 -MIN_VID_HEIGHT = 300 - -MAX_VID_WIDTH = 10000 -MAX_VID_HEIGHT = 10000 - -MAX_VIDEO_SIZE = 1024 * 1024 * 1024 # 1 GB max for in-memory video processing - -MOONVALLEY_MAREY_MAX_PROMPT_LENGTH = 5000 - - -def is_valid_task_creation_response(response: MoonvalleyPromptResponse) -> bool: - """Verifies that the initial response contains a task ID.""" - return bool(response.id) - - -def validate_task_creation_response(response) -> None: - if not is_valid_task_creation_response(response): - error_msg = f"Moonvalley Marey API: Initial request failed. Code: {response.code}, Message: {response.message}, Data: {response}" - logging.error(error_msg) - raise RuntimeError(error_msg) - - -def validate_video_to_video_input(video: Input.Video) -> Input.Video: - """ - Validates and processes video input for Moonvalley Video-to-Video generation. - - Args: - video: Input video to validate - - Returns: - Validated and potentially trimmed video - - Raises: - ValueError: If video doesn't meet requirements - MoonvalleyApiError: If video duration is too short - """ - width, height = _get_video_dimensions(video) - _validate_video_dimensions(width, height) - validate_container_format_is_mp4(video) - - return _validate_and_trim_duration(video) - - -def _get_video_dimensions(video: Input.Video) -> tuple[int, int]: - """Extracts video dimensions with error handling.""" - try: - return video.get_dimensions() - except Exception as e: - logging.error("Error getting dimensions of video: %s", e) - raise ValueError(f"Cannot get video dimensions: {e}") from e - - -def _validate_video_dimensions(width: int, height: int) -> None: - """Validates video dimensions meet Moonvalley V2V requirements.""" - supported_resolutions = { - (1920, 1080), - (1080, 1920), - (1152, 1152), - (1536, 1152), - (1152, 1536), - } - - if (width, height) not in supported_resolutions: - supported_list = ", ".join([f"{w}x{h}" for w, h in sorted(supported_resolutions)]) - raise ValueError(f"Resolution {width}x{height} not supported. Supported: {supported_list}") - - -def _validate_and_trim_duration(video: Input.Video) -> Input.Video: - """Validates video duration and trims to 5 seconds if needed.""" - duration = video.get_duration() - _validate_minimum_duration(duration) - return _trim_if_too_long(video, duration) - - -def _validate_minimum_duration(duration: float) -> None: - """Ensures video is at least 5 seconds long.""" - if duration < 5: - raise ValueError("Input video must be at least 5 seconds long.") - - -def _trim_if_too_long(video: Input.Video, duration: float) -> Input.Video: - """Trims video to 5 seconds if longer.""" - if duration > 5: - return trim_video(video, 5) - return video - - -def parse_width_height_from_res(resolution: str): - # Accepts a string like "16:9 (1920 x 1080)" and returns width, height as a dict - res_map = { - "16:9 (1920 x 1080)": {"width": 1920, "height": 1080}, - "9:16 (1080 x 1920)": {"width": 1080, "height": 1920}, - "1:1 (1152 x 1152)": {"width": 1152, "height": 1152}, - "4:3 (1536 x 1152)": {"width": 1536, "height": 1152}, - "3:4 (1152 x 1536)": {"width": 1152, "height": 1536}, - # "21:9 (2560 x 1080)": {"width": 2560, "height": 1080}, - } - return res_map.get(resolution, {"width": 1920, "height": 1080}) - - -def parse_control_parameter(value): - control_map = { - "Motion Transfer": "motion_control", - "Canny": "canny_control", - "Pose Transfer": "pose_control", - "Depth": "depth_control", - } - return control_map.get(value, control_map["Motion Transfer"]) - - -async def get_response(cls: type[IO.ComfyNode], task_id: str) -> MoonvalleyPromptResponse: - return await poll_op( - cls, - ApiEndpoint(path=f"{API_PROMPTS_ENDPOINT}/{task_id}"), - response_model=MoonvalleyPromptResponse, - status_extractor=lambda r: (r.status if r and r.status else None), - poll_interval=16.0, - max_poll_attempts=240, - ) - - -class MoonvalleyImg2VideoNode(IO.ComfyNode): - - @classmethod - def define_schema(cls) -> IO.Schema: - return IO.Schema( - node_id="MoonvalleyImg2VideoNode", - display_name="Moonvalley Marey Image to Video", - category="api node/video/Moonvalley Marey", - description="Moonvalley Marey Image to Video Node", - inputs=[ - IO.Image.Input( - "image", - tooltip="The reference image used to generate the video", - ), - IO.String.Input( - "prompt", - multiline=True, - ), - IO.String.Input( - "negative_prompt", - multiline=True, - default=" gopro, bright, contrast, static, overexposed, vignette, " - "artifacts, still, noise, texture, scanlines, videogame, 360 camera, VR, transition, " - "flare, saturation, distorted, warped, wide angle, saturated, vibrant, glowing, " - "cross dissolve, cheesy, ugly hands, mutated hands, mutant, disfigured, extra fingers, " - "blown out, horrible, blurry, worst quality, bad, dissolve, melt, fade in, fade out, " - "wobbly, weird, low quality, plastic, stock footage, video camera, boring", - tooltip="Negative prompt text", - ), - IO.Combo.Input( - "resolution", - options=[ - "16:9 (1920 x 1080)", - "9:16 (1080 x 1920)", - "1:1 (1152 x 1152)", - "4:3 (1536 x 1152)", - "3:4 (1152 x 1536)", - # "21:9 (2560 x 1080)", - ], - default="16:9 (1920 x 1080)", - tooltip="Resolution of the output video", - ), - IO.Float.Input( - "prompt_adherence", - default=4.5, - min=1.0, - max=20.0, - step=1.0, - tooltip="Guidance scale for generation control", - ), - IO.Int.Input( - "seed", - default=9, - min=0, - max=4294967295, - step=1, - display_mode=IO.NumberDisplay.number, - tooltip="Random seed value", - control_after_generate=True, - ), - IO.Int.Input( - "steps", - default=80, - min=75, # steps should be greater or equal to cooldown_steps(75) + warmup_steps(0) - max=100, - step=1, - tooltip="Number of denoising steps", - ), - ], - 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( - depends_on=IO.PriceBadgeDepends(), - expr="""{"type":"usd","usd": 1.5}""", - ), - ) - - @classmethod - async def execute( - cls, - image: Input.Image, - prompt: str, - negative_prompt: str, - resolution: str, - prompt_adherence: float, - seed: int, - steps: int, - ) -> IO.NodeOutput: - validate_image_dimensions(image, min_width=300, min_height=300, max_height=MAX_HEIGHT, max_width=MAX_WIDTH) - validate_string(prompt, min_length=1, max_length=MOONVALLEY_MAREY_MAX_PROMPT_LENGTH) - validate_string(negative_prompt, field_name="negative_prompt", max_length=MOONVALLEY_MAREY_MAX_PROMPT_LENGTH) - width_height = parse_width_height_from_res(resolution) - - inference_params = MoonvalleyTextToVideoInferenceParams( - negative_prompt=negative_prompt, - steps=steps, - seed=seed, - guidance_scale=prompt_adherence, - width=width_height["width"], - height=width_height["height"], - use_negative_prompts=True, - ) - - # Get MIME type from tensor - assuming PNG format for image tensors - mime_type = "image/png" - image_url = (await upload_images_to_comfyapi(cls, image, max_images=1, mime_type=mime_type))[0] - task_creation_response = await sync_op( - cls, - endpoint=ApiEndpoint(path=API_IMG2VIDEO_ENDPOINT, method="POST"), - response_model=MoonvalleyPromptResponse, - data=MoonvalleyTextToVideoRequest( - image_url=image_url, prompt_text=prompt, inference_params=inference_params - ), - ) - validate_task_creation_response(task_creation_response) - final_response = await get_response(cls, task_creation_response.id) - video = await download_url_to_video_output(final_response.output_url) - return IO.NodeOutput(video) - - -class MoonvalleyVideo2VideoNode(IO.ComfyNode): - - @classmethod - def define_schema(cls) -> IO.Schema: - return IO.Schema( - node_id="MoonvalleyVideo2VideoNode", - display_name="Moonvalley Marey Video to Video", - category="api node/video/Moonvalley Marey", - description="", - inputs=[ - IO.String.Input( - "prompt", - multiline=True, - tooltip="Describes the video to generate", - ), - IO.String.Input( - "negative_prompt", - multiline=True, - default=" gopro, bright, contrast, static, overexposed, vignette, " - "artifacts, still, noise, texture, scanlines, videogame, 360 camera, VR, transition, " - "flare, saturation, distorted, warped, wide angle, saturated, vibrant, glowing, " - "cross dissolve, cheesy, ugly hands, mutated hands, mutant, disfigured, extra fingers, " - "blown out, horrible, blurry, worst quality, bad, dissolve, melt, fade in, fade out, " - "wobbly, weird, low quality, plastic, stock footage, video camera, boring", - tooltip="Negative prompt text", - ), - IO.Int.Input( - "seed", - default=9, - min=0, - max=4294967295, - step=1, - display_mode=IO.NumberDisplay.number, - tooltip="Random seed value", - control_after_generate=False, - ), - IO.Video.Input( - "video", - tooltip="The reference video used to generate the output video. Must be at least 5 seconds long. " - "Videos longer than 5s will be automatically trimmed. Only MP4 format supported.", - ), - IO.Combo.Input( - "control_type", - options=["Motion Transfer", "Pose Transfer"], - default="Motion Transfer", - optional=True, - ), - IO.Int.Input( - "motion_intensity", - default=100, - min=0, - max=100, - step=1, - tooltip="Only used if control_type is 'Motion Transfer'", - optional=True, - ), - IO.Int.Input( - "steps", - default=60, - min=60, # steps should be greater or equal to cooldown_steps(36) + warmup_steps(24) - max=100, - step=1, - display_mode=IO.NumberDisplay.number, - tooltip="Number of inference steps", - ), - ], - 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( - depends_on=IO.PriceBadgeDepends(), - expr="""{"type":"usd","usd": 2.25}""", - ), - ) - - @classmethod - async def execute( - cls, - prompt: str, - negative_prompt: str, - seed: int, - video: Input.Video | None = None, - control_type: str = "Motion Transfer", - motion_intensity: int | None = 100, - steps=60, - prompt_adherence=4.5, - ) -> IO.NodeOutput: - validated_video = validate_video_to_video_input(video) - video_url = await upload_video_to_comfyapi(cls, validated_video) - validate_string(prompt, min_length=1, max_length=MOONVALLEY_MAREY_MAX_PROMPT_LENGTH) - validate_string(negative_prompt, field_name="negative_prompt", max_length=MOONVALLEY_MAREY_MAX_PROMPT_LENGTH) - - # Only include motion_intensity for Motion Transfer - control_params = {} - if control_type == "Motion Transfer" and motion_intensity is not None: - control_params["motion_intensity"] = motion_intensity - - inference_params = MoonvalleyVideoToVideoInferenceParams( - negative_prompt=negative_prompt, - seed=seed, - control_params=control_params, - steps=steps, - guidance_scale=prompt_adherence, - ) - - task_creation_response = await sync_op( - cls, - endpoint=ApiEndpoint(path=API_VIDEO2VIDEO_ENDPOINT, method="POST"), - response_model=MoonvalleyPromptResponse, - data=MoonvalleyVideoToVideoRequest( - control_type=parse_control_parameter(control_type), - video_url=video_url, - prompt_text=prompt, - inference_params=inference_params, - ), - ) - validate_task_creation_response(task_creation_response) - final_response = await get_response(cls, task_creation_response.id) - return IO.NodeOutput(await download_url_to_video_output(final_response.output_url)) - - -class MoonvalleyTxt2VideoNode(IO.ComfyNode): - - @classmethod - def define_schema(cls) -> IO.Schema: - return IO.Schema( - node_id="MoonvalleyTxt2VideoNode", - display_name="Moonvalley Marey Text to Video", - category="api node/video/Moonvalley Marey", - description="", - inputs=[ - IO.String.Input( - "prompt", - multiline=True, - ), - IO.String.Input( - "negative_prompt", - multiline=True, - default=" gopro, bright, contrast, static, overexposed, vignette, " - "artifacts, still, noise, texture, scanlines, videogame, 360 camera, VR, transition, " - "flare, saturation, distorted, warped, wide angle, saturated, vibrant, glowing, " - "cross dissolve, cheesy, ugly hands, mutated hands, mutant, disfigured, extra fingers, " - "blown out, horrible, blurry, worst quality, bad, dissolve, melt, fade in, fade out, " - "wobbly, weird, low quality, plastic, stock footage, video camera, boring", - tooltip="Negative prompt text", - ), - IO.Combo.Input( - "resolution", - options=[ - "16:9 (1920 x 1080)", - "9:16 (1080 x 1920)", - "1:1 (1152 x 1152)", - "4:3 (1536 x 1152)", - "3:4 (1152 x 1536)", - "21:9 (2560 x 1080)", - ], - default="16:9 (1920 x 1080)", - tooltip="Resolution of the output video", - ), - IO.Float.Input( - "prompt_adherence", - default=4.0, - min=1.0, - max=20.0, - step=1.0, - tooltip="Guidance scale for generation control", - ), - IO.Int.Input( - "seed", - default=9, - min=0, - max=4294967295, - step=1, - display_mode=IO.NumberDisplay.number, - control_after_generate=True, - tooltip="Random seed value", - ), - IO.Int.Input( - "steps", - default=80, - min=75, # steps should be greater or equal to cooldown_steps(75) + warmup_steps(0) - max=100, - step=1, - tooltip="Inference steps", - ), - ], - 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( - depends_on=IO.PriceBadgeDepends(), - expr="""{"type":"usd","usd": 1.5}""", - ), - ) - - @classmethod - async def execute( - cls, - prompt: str, - negative_prompt: str, - resolution: str, - prompt_adherence: float, - seed: int, - steps: int, - ) -> IO.NodeOutput: - validate_string(prompt, min_length=1, max_length=MOONVALLEY_MAREY_MAX_PROMPT_LENGTH) - validate_string(negative_prompt, field_name="negative_prompt", max_length=MOONVALLEY_MAREY_MAX_PROMPT_LENGTH) - width_height = parse_width_height_from_res(resolution) - - inference_params = MoonvalleyTextToVideoInferenceParams( - negative_prompt=negative_prompt, - steps=steps, - seed=seed, - guidance_scale=prompt_adherence, - num_frames=128, - width=width_height["width"], - height=width_height["height"], - ) - - task_creation_response = await sync_op( - cls, - endpoint=ApiEndpoint(path=API_TXT2VIDEO_ENDPOINT, method="POST"), - response_model=MoonvalleyPromptResponse, - data=MoonvalleyTextToVideoRequest(prompt_text=prompt, inference_params=inference_params), - ) - validate_task_creation_response(task_creation_response) - final_response = await get_response(cls, task_creation_response.id) - return IO.NodeOutput(await download_url_to_video_output(final_response.output_url)) - - -class MoonvalleyExtension(ComfyExtension): - @override - async def get_node_list(self) -> list[type[IO.ComfyNode]]: - return [ - MoonvalleyImg2VideoNode, - MoonvalleyTxt2VideoNode, - MoonvalleyVideo2VideoNode, - ] - - -async def comfy_entrypoint() -> MoonvalleyExtension: - return MoonvalleyExtension() diff --git a/comfy_api_nodes/nodes_openai.py b/comfy_api_nodes/nodes_openai.py index 843681817..a5a188634 100644 --- a/comfy_api_nodes/nodes_openai.py +++ b/comfy_api_nodes/nodes_openai.py @@ -27,6 +27,7 @@ from comfy_api_nodes.util import ( ApiEndpoint, download_url_to_bytesio, downscale_image_tensor, + get_number_of_images, poll_op, sync_op, tensor_to_base64_string, @@ -39,16 +40,18 @@ STARTING_POINT_ID_PATTERN = r"" class SupportedOpenAIModel(str, Enum): - o4_mini = "o4-mini" - o1 = "o1" - o3 = "o3" - o1_pro = "o1-pro" - gpt_4_1 = "gpt-4.1" - gpt_4_1_mini = "gpt-4.1-mini" - gpt_4_1_nano = "gpt-4.1-nano" + gpt_5_5_pro = "gpt-5.5-pro" + gpt_5_5 = "gpt-5.5" gpt_5 = "gpt-5" gpt_5_mini = "gpt-5-mini" gpt_5_nano = "gpt-5-nano" + gpt_4_1 = "gpt-4.1" + gpt_4_1_mini = "gpt-4.1-mini" + gpt_4_1_nano = "gpt-4.1-nano" + o4_mini = "o4-mini" + o3 = "o3" + o1_pro = "o1-pro" + o1 = "o1" async def validate_and_cast_response(response, timeout: int = None) -> torch.Tensor: @@ -370,6 +373,7 @@ class OpenAIGPTImage1(IO.ComfyNode): display_name="OpenAI GPT Image 2", category="api node/image/OpenAI", description="Generates images synchronously via OpenAI's GPT Image endpoint.", + is_deprecated=True, inputs=[ IO.String.Input( "prompt", @@ -454,7 +458,6 @@ class OpenAIGPTImage1(IO.ComfyNode): step=16, tooltip="Used only when `size` is 'Custom'. Must be a multiple of 16 (GPT Image 2 only).", optional=True, - advanced=True, ), IO.Int.Input( "custom_height", @@ -464,7 +467,6 @@ class OpenAIGPTImage1(IO.ComfyNode): step=16, tooltip="Used only when `size` is 'Custom'. Must be a multiple of 16 (GPT Image 2 only).", optional=True, - advanced=True, ), ], outputs=[ @@ -640,6 +642,316 @@ class OpenAIGPTImage1(IO.ComfyNode): return IO.NodeOutput(await validate_and_cast_response(response)) +def _gpt_image_shared_inputs(): + """Inputs shared by all GPT Image models (quality + reference images + mask).""" + return [ + IO.Combo.Input( + "quality", + default="low", + options=["low", "medium", "high"], + tooltip="Image quality, affects cost and generation time.", + ), + 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 reference image(s) for image editing. Up to 16 images.", + ), + IO.Mask.Input( + "mask", + optional=True, + tooltip="Optional mask for inpainting (white areas will be replaced). " + "Requires exactly one reference image.", + ), + ] + + +def _gpt_image_legacy_model_inputs(): + """Per-model widget set for legacy gpt-image-1 / gpt-image-1.5 (4 base sizes, transparent bg allowed).""" + return [ + IO.Combo.Input( + "size", + default="auto", + options=["auto", "1024x1024", "1024x1536", "1536x1024"], + tooltip="Image size.", + ), + IO.Combo.Input( + "background", + default="auto", + options=["auto", "opaque", "transparent"], + tooltip="Return image with or without background.", + ), + *_gpt_image_shared_inputs(), + ] + + +class OpenAIGPTImageNodeV2(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="OpenAIGPTImageNodeV2", + display_name="OpenAI GPT Image 2", + category="api node/image/OpenAI", + description="Generates images via OpenAI's GPT Image endpoint.", + inputs=[ + IO.String.Input( + "prompt", + default="", + multiline=True, + tooltip="Text prompt for GPT Image", + ), + IO.DynamicCombo.Input( + "model", + options=[ + IO.DynamicCombo.Option( + "gpt-image-2", + [ + IO.Combo.Input( + "size", + default="auto", + options=[ + "auto", + "1024x1024", + "1024x1536", + "1536x1024", + "2048x2048", + "2048x1152", + "1152x2048", + "3840x2160", + "2160x3840", + "Custom", + ], + tooltip="Image size. Select 'Custom' to use the custom width and height.", + ), + IO.Int.Input( + "custom_width", + default=1024, + min=1024, + max=3840, + step=16, + tooltip="Used only when `size` is 'Custom'. Must be a multiple of 16.", + ), + IO.Int.Input( + "custom_height", + default=1024, + min=1024, + max=3840, + step=16, + tooltip="Used only when `size` is 'Custom'. Must be a multiple of 16.", + ), + IO.Combo.Input( + "background", + default="auto", + options=["auto", "opaque"], + tooltip="Return image with or without background.", + ), + *_gpt_image_shared_inputs(), + ], + ), + IO.DynamicCombo.Option("gpt-image-1.5", _gpt_image_legacy_model_inputs()), + IO.DynamicCombo.Option("gpt-image-1", _gpt_image_legacy_model_inputs()), + ], + ), + IO.Int.Input( + "n", + default=1, + min=1, + max=8, + step=1, + tooltip="How many images to generate", + display_mode=IO.NumberDisplay.number, + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="not implemented yet in backend", + ), + ], + 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.quality", "n"]), + expr=""" + ( + $ranges := { + "gpt-image-1": { + "low": [0.011, 0.02], + "medium": [0.042, 0.07], + "high": [0.167, 0.25] + }, + "gpt-image-1.5": { + "low": [0.009, 0.02], + "medium": [0.034, 0.062], + "high": [0.133, 0.22] + }, + "gpt-image-2": { + "low": [0.0048, 0.019], + "medium": [0.041, 0.168], + "high": [0.165, 0.67] + } + }; + $range := $lookup($lookup($ranges, widgets.model), $lookup(widgets, "model.quality")); + $nRaw := widgets.n; + $n := ($nRaw != null and $nRaw != 0) ? $nRaw : 1; + ($n = 1) + ? {"type":"range_usd","min_usd": $range[0], "max_usd": $range[1], "format": {"approximate": true}} + : { + "type":"range_usd", + "min_usd": $range[0] * $n, + "max_usd": $range[1] * $n, + "format": { "suffix": "/Run", "approximate": true } + } + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + prompt: str, + model: dict, + n: int, + seed: int, + ) -> IO.NodeOutput: + validate_string(prompt, strip_whitespace=False) + + model_id = model["model"] + size = model["size"] + background = model["background"] + quality = model["quality"] + custom_width = model.get("custom_width", 1024) + custom_height = model.get("custom_height", 1024) + + images_dict = model.get("images") or {} + image_tensors: list[Input.Image] = [t for t in images_dict.values() if t is not None] + n_images = sum(get_number_of_images(t) for t in image_tensors) + mask = model.get("mask") + + if mask is not None and n_images == 0: + raise ValueError("Cannot use a mask without an input image") + + if size == "Custom": + if custom_width % 16 != 0 or custom_height % 16 != 0: + raise ValueError( + f"Custom width and height must be multiples of 16, got {custom_width}x{custom_height}" + ) + if max(custom_width, custom_height) > 3840: + raise ValueError( + f"Custom resolution max edge must be <= 3840, got {custom_width}x{custom_height}" + ) + ratio = max(custom_width, custom_height) / min(custom_width, custom_height) + if ratio > 3: + raise ValueError( + f"Custom resolution aspect ratio must not exceed 3:1, got {custom_width}x{custom_height}" + ) + total_pixels = custom_width * custom_height + if not 655_360 <= total_pixels <= 8_294_400: + raise ValueError( + f"Custom resolution total pixels must be between 655,360 and 8,294,400, got {total_pixels}" + ) + size = f"{custom_width}x{custom_height}" + + if model_id == "gpt-image-1": + price_extractor = calculate_tokens_price_image_1 + elif model_id == "gpt-image-1.5": + price_extractor = calculate_tokens_price_image_1_5 + elif model_id == "gpt-image-2": + price_extractor = calculate_tokens_price_image_2_0 + else: + raise ValueError(f"Unknown model: {model_id}") + + if image_tensors: + flat: list[torch.Tensor] = [] + for tensor in image_tensors: + if len(tensor.shape) == 4: + flat.extend(tensor[i : i + 1] for i in range(tensor.shape[0])) + else: + flat.append(tensor.unsqueeze(0)) + + files = [] + for i, single_image in enumerate(flat): + scaled_image = downscale_image_tensor(single_image, total_pixels=2048 * 2048).squeeze() + image_np = (scaled_image.numpy() * 255).astype(np.uint8) + img = Image.fromarray(image_np) + img_byte_arr = BytesIO() + img.save(img_byte_arr, format="PNG") + img_byte_arr.seek(0) + + if len(flat) == 1: + files.append(("image", (f"image_{i}.png", img_byte_arr, "image/png"))) + else: + files.append(("image[]", (f"image_{i}.png", img_byte_arr, "image/png"))) + + if mask is not None: + if len(flat) != 1: + raise Exception("Cannot use a mask with multiple image") + ref_image = flat[0] + if mask.shape[1:] != ref_image.shape[1:-1]: + raise Exception("Mask and Image must be the same size") + _, height, width = mask.shape + rgba_mask = torch.zeros(height, width, 4, device="cpu") + rgba_mask[:, :, 3] = 1 - mask.squeeze().cpu() + scaled_mask = downscale_image_tensor( + rgba_mask.unsqueeze(0), total_pixels=2048 * 2048 + ).squeeze() + mask_np = (scaled_mask.numpy() * 255).astype(np.uint8) + mask_img = Image.fromarray(mask_np) + mask_img_byte_arr = BytesIO() + mask_img.save(mask_img_byte_arr, format="PNG") + mask_img_byte_arr.seek(0) + files.append(("mask", ("mask.png", mask_img_byte_arr, "image/png"))) + + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/openai/images/edits", method="POST"), + response_model=OpenAIImageGenerationResponse, + data=OpenAIImageEditRequest( + model=model_id, + prompt=prompt, + quality=quality, + background=background, + n=n, + size=size, + moderation="low", + ), + content_type="multipart/form-data", + files=files, + price_extractor=price_extractor, + ) + else: + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/openai/images/generations", method="POST"), + response_model=OpenAIImageGenerationResponse, + data=OpenAIImageGenerationRequest( + model=model_id, + prompt=prompt, + quality=quality, + background=background, + n=n, + size=size, + moderation="low", + ), + price_extractor=price_extractor, + ) + return IO.NodeOutput(await validate_and_cast_response(response)) + + class OpenAIChatNode(IO.ComfyNode): """ Node to generate text responses from an OpenAI model. @@ -741,6 +1053,16 @@ class OpenAIChatNode(IO.ComfyNode): "usd": [0.002, 0.008], "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } } + : $contains($m, "gpt-5.5-pro") ? { + "type": "list_usd", + "usd": [0.03, 0.18], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } + : $contains($m, "gpt-5.5") ? { + "type": "list_usd", + "usd": [0.005, 0.03], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } : $contains($m, "gpt-5-nano") ? { "type": "list_usd", "usd": [0.00005, 0.0004], @@ -989,6 +1311,7 @@ class OpenAIExtension(ComfyExtension): OpenAIDalle2, OpenAIDalle3, OpenAIGPTImage1, + OpenAIGPTImageNodeV2, OpenAIChatNode, OpenAIInputFiles, OpenAIChatConfig, diff --git a/comfy_api_nodes/nodes_quiver.py b/comfy_api_nodes/nodes_quiver.py index 28862e368..3269c0afe 100644 --- a/comfy_api_nodes/nodes_quiver.py +++ b/comfy_api_nodes/nodes_quiver.py @@ -143,7 +143,7 @@ class QuiverTextToSVGNode(IO.ComfyNode): if reference_images: references = [] for key in reference_images: - url = await upload_image_to_comfyapi(cls, reference_images[key]) + url = await upload_image_to_comfyapi(cls, reference_images[key], mime_type="image/png") references.append(QuiverImageObject(url=url)) if len(references) > 4: raise ValueError("Maximum 4 reference images are allowed.") @@ -252,7 +252,7 @@ class QuiverImageToSVGNode(IO.ComfyNode): model: dict, seed: int, ) -> IO.NodeOutput: - image_url = await upload_image_to_comfyapi(cls, image) + image_url = await upload_image_to_comfyapi(cls, image, mime_type="image/png") response = await sync_op( cls, diff --git a/comfy_api_nodes/nodes_sora.py b/comfy_api_nodes/nodes_sora.py index 4d9075dcf..c1d485188 100644 --- a/comfy_api_nodes/nodes_sora.py +++ b/comfy_api_nodes/nodes_sora.py @@ -33,7 +33,7 @@ class OpenAIVideoSora2(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="OpenAIVideoSora2", - display_name="OpenAI Sora - Video (Deprecated)", + display_name="OpenAI Sora - Video (DEPRECATED)", category="api node/video/Sora", description=( "OpenAI video and audio generation.\n\n" diff --git a/comfy_api_nodes/nodes_topaz.py b/comfy_api_nodes/nodes_topaz.py index b18b31af1..e79c16d3c 100644 --- a/comfy_api_nodes/nodes_topaz.py +++ b/comfy_api_nodes/nodes_topaz.py @@ -36,11 +36,15 @@ from comfy_api_nodes.util import ( ) UPSCALER_MODELS_MAP = { + "Astra 2": "ast-2", "Starlight (Astra) Fast": "slf-1", "Starlight (Astra) Creative": "slc-1", "Starlight Precise 2.5": "slp-2.5", } +AST2_MAX_FRAMES = 9000 +AST2_MAX_FRAMES_WITH_PROMPT = 450 + class TopazImageEnhance(IO.ComfyNode): @classmethod @@ -230,13 +234,20 @@ class TopazVideoEnhance(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="TopazVideoEnhance", - display_name="Topaz Video Enhance", + display_name="Topaz Video Enhance (Legacy)", category="api node/video/Topaz", description="Breathe new life into video with powerful upscaling and recovery technology.", inputs=[ IO.Video.Input("video"), IO.Boolean.Input("upscaler_enabled", default=True), - IO.Combo.Input("upscaler_model", options=list(UPSCALER_MODELS_MAP.keys())), + IO.Combo.Input( + "upscaler_model", + options=[ + "Starlight (Astra) Fast", + "Starlight (Astra) Creative", + "Starlight Precise 2.5", + ], + ), IO.Combo.Input("upscaler_resolution", options=["FullHD (1080p)", "4K (2160p)"]), IO.Combo.Input( "upscaler_creativity", @@ -304,6 +315,7 @@ class TopazVideoEnhance(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + is_deprecated=True, ) @classmethod @@ -453,7 +465,350 @@ class TopazVideoEnhance(IO.ComfyNode): progress_extractor=lambda x: getattr(x, "progress", 0), price_extractor=lambda x: (x.estimates.cost[0] * 0.08 if x.estimates and x.estimates.cost[0] else None), poll_interval=10.0, - max_poll_attempts=320, + ) + return IO.NodeOutput(await download_url_to_video_output(final_response.download.url)) + + +class TopazVideoEnhanceV2(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="TopazVideoEnhanceV2", + display_name="Topaz Video Enhance", + category="api node/video/Topaz", + description="Breathe new life into video with powerful upscaling and recovery technology.", + inputs=[ + IO.Video.Input("video"), + IO.DynamicCombo.Input( + "upscaler_model", + options=[ + IO.DynamicCombo.Option( + "Astra 2", + [ + IO.Combo.Input("upscaler_resolution", options=["FullHD (1080p)", "4K (2160p)"]), + IO.Float.Input( + "creativity", + default=0.5, + min=0.0, + max=1.0, + step=0.1, + display_mode=IO.NumberDisplay.slider, + tooltip="Creative strength of the upscale.", + ), + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Optional descriptive (not instructive) scene prompt." + f"Capping input at {AST2_MAX_FRAMES_WITH_PROMPT} frames (~15s @ 30fps) when set.", + ), + IO.Float.Input( + "sharp", + default=0.5, + min=0.0, + max=1.0, + step=0.01, + display_mode=IO.NumberDisplay.slider, + tooltip="Pre-enhance sharpness: " + "0.0=Gaussian blur, 0.5=passthrough (default), 1.0=USM sharpening.", + advanced=True, + ), + IO.Float.Input( + "realism", + default=0.0, + min=0.0, + max=1.0, + step=0.01, + display_mode=IO.NumberDisplay.slider, + tooltip="Pulls output toward photographic realism." + "Leave at 0 for the model default.", + advanced=True, + ), + ], + ), + IO.DynamicCombo.Option( + "Starlight (Astra) Fast", + [IO.Combo.Input("upscaler_resolution", options=["FullHD (1080p)", "4K (2160p)"]),], + ), + IO.DynamicCombo.Option( + "Starlight (Astra) Creative", + [ + IO.Combo.Input("upscaler_resolution", options=["FullHD (1080p)", "4K (2160p)"]), + IO.Combo.Input( + "creativity", + options=["low", "middle", "high"], + default="low", + tooltip="Creative strength of the upscale.", + ), + ], + ), + IO.DynamicCombo.Option( + "Starlight Precise 2.5", + [IO.Combo.Input("upscaler_resolution", options=["FullHD (1080p)", "4K (2160p)"])], + ), + IO.DynamicCombo.Option("Disabled", []), + ], + ), + IO.DynamicCombo.Input( + "interpolation_model", + options=[ + IO.DynamicCombo.Option("Disabled", []), + IO.DynamicCombo.Option( + "apo-8", + [ + IO.Int.Input( + "interpolation_frame_rate", + default=60, + min=15, + max=240, + display_mode=IO.NumberDisplay.number, + tooltip="Output frame rate.", + ), + IO.Int.Input( + "interpolation_slowmo", + default=1, + min=1, + max=16, + display_mode=IO.NumberDisplay.number, + tooltip="Slow-motion factor applied to the input video. " + "For example, 2 makes the output twice as slow and doubles the duration.", + advanced=True, + ), + IO.Boolean.Input( + "interpolation_duplicate", + default=False, + tooltip="Analyze the input for duplicate frames and remove them.", + advanced=True, + ), + IO.Float.Input( + "interpolation_duplicate_threshold", + default=0.01, + min=0.001, + max=0.1, + step=0.001, + display_mode=IO.NumberDisplay.number, + tooltip="Detection sensitivity for duplicate frames.", + advanced=True, + ), + ], + ), + ], + ), + IO.Combo.Input( + "dynamic_compression_level", + options=["Low", "Mid", "High"], + default="Low", + tooltip="CQP level.", + optional=True, + ), + ], + 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( + depends_on=IO.PriceBadgeDepends(widgets=[ + "upscaler_model", + "upscaler_model.upscaler_resolution", + "interpolation_model", + ]), + expr=""" + ( + $model := $lookup(widgets, "upscaler_model"); + $res := $lookup(widgets, "upscaler_model.upscaler_resolution"); + $interp := $lookup(widgets, "interpolation_model"); + $is4k := $contains($res, "4k"); + $hasInterp := $interp != "disabled"; + $rates := { + "starlight (astra) fast": {"hd": 0.43, "uhd": 0.85}, + "starlight precise 2.5": {"hd": 0.70, "uhd": 1.54}, + "astra 2": {"hd": 1.72, "uhd": 2.85}, + "starlight (astra) creative": {"hd": 2.25, "uhd": 3.99} + }; + $surcharge := $is4k ? 0.28 : 0.14; + $entry := $lookup($rates, $model); + $base := $is4k ? $entry.uhd : $entry.hd; + $hi := $base + ($hasInterp ? $surcharge : 0); + $model = "disabled" + ? {"type":"text","text":"Interpolation only"} + : ($hasInterp + ? {"type":"text","text":"~" & $string($base) & "–" & $string($hi) & " credits/src frame"} + : {"type":"text","text":"~" & $string($base) & " credits/src frame"}) + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + video: Input.Video, + upscaler_model: dict, + interpolation_model: dict, + dynamic_compression_level: str = "Low", + ) -> IO.NodeOutput: + upscaler_choice = upscaler_model["upscaler_model"] + interpolation_choice = interpolation_model["interpolation_model"] + if upscaler_choice == "Disabled" and interpolation_choice == "Disabled": + raise ValueError("There is nothing to do: both upscaling and interpolation are disabled.") + validate_container_format_is_mp4(video) + src_width, src_height = video.get_dimensions() + src_frame_rate = int(video.get_frame_rate()) + duration_sec = video.get_duration() + src_video_stream = video.get_stream_source() + target_width = src_width + target_height = src_height + target_frame_rate = src_frame_rate + filters = [] + if upscaler_choice != "Disabled": + if "1080p" in upscaler_model["upscaler_resolution"]: + target_pixel_p = 1080 + max_long_side = 1920 + else: + target_pixel_p = 2160 + max_long_side = 3840 + ar = src_width / src_height + if src_width >= src_height: + # Landscape or Square; Attempt to set height to target (e.g., 2160), calculate width + target_height = target_pixel_p + target_width = int(target_height * ar) + # Check if width exceeds standard bounds (for ultra-wide e.g., 21:9 ARs) + if target_width > max_long_side: + target_width = max_long_side + target_height = int(target_width / ar) + else: + # Portrait; Attempt to set width to target (e.g., 2160), calculate height + target_width = target_pixel_p + target_height = int(target_width / ar) + # Check if height exceeds standard bounds + if target_height > max_long_side: + target_height = max_long_side + target_width = int(target_height * ar) + if target_width % 2 != 0: + target_width += 1 + if target_height % 2 != 0: + target_height += 1 + model_id = UPSCALER_MODELS_MAP[upscaler_choice] + if model_id == "slc-1": + filters.append( + VideoEnhancementFilter( + model=model_id, + creativity=upscaler_model["creativity"], + isOptimizedMode=True, + ) + ) + elif model_id == "ast-2": + n_frames = video.get_frame_count() + ast2_prompt = (upscaler_model["prompt"] or "").strip() + if ast2_prompt and n_frames > AST2_MAX_FRAMES_WITH_PROMPT: + raise ValueError( + f"Astra 2 with a prompt is limited to {AST2_MAX_FRAMES_WITH_PROMPT} input frames " + f"(~15s @ 30fps); video has {n_frames}. Clear the prompt or shorten the clip." + ) + if n_frames > AST2_MAX_FRAMES: + raise ValueError(f"Astra 2 is limited to {AST2_MAX_FRAMES} input frames; video has {n_frames}.") + realism = upscaler_model["realism"] + filters.append( + VideoEnhancementFilter( + model=model_id, + creativity=upscaler_model["creativity"], + prompt=(ast2_prompt or None), + sharp=upscaler_model["sharp"], + realism=(realism if realism > 0 else None), + ) + ) + else: + filters.append(VideoEnhancementFilter(model=model_id)) + if interpolation_choice != "Disabled": + target_frame_rate = interpolation_model["interpolation_frame_rate"] + filters.append( + VideoFrameInterpolationFilter( + model=interpolation_choice, + slowmo=interpolation_model["interpolation_slowmo"], + fps=interpolation_model["interpolation_frame_rate"], + duplicate=interpolation_model["interpolation_duplicate"], + duplicate_threshold=interpolation_model["interpolation_duplicate_threshold"], + ), + ) + initial_res = await sync_op( + cls, + ApiEndpoint(path="/proxy/topaz/video/", method="POST"), + response_model=CreateVideoResponse, + data=CreateVideoRequest( + source=CreateVideoRequestSource( + container="mp4", + size=get_fs_object_size(src_video_stream), + duration=int(duration_sec), + frameCount=video.get_frame_count(), + frameRate=src_frame_rate, + resolution=Resolution(width=src_width, height=src_height), + ), + filters=filters, + output=OutputInformationVideo( + resolution=Resolution(width=target_width, height=target_height), + frameRate=target_frame_rate, + audioCodec="AAC", + audioTransfer="Copy", + dynamicCompressionLevel=dynamic_compression_level, + ), + ), + wait_label="Creating task", + final_label_on_success="Task created", + ) + upload_res = await sync_op( + cls, + ApiEndpoint( + path=f"/proxy/topaz/video/{initial_res.requestId}/accept", + method="PATCH", + ), + response_model=VideoAcceptResponse, + wait_label="Preparing upload", + final_label_on_success="Upload started", + ) + if len(upload_res.urls) > 1: + raise NotImplementedError( + "Large files are not currently supported. Please open an issue in the ComfyUI repository." + ) + async with aiohttp.ClientSession(headers={"Content-Type": "video/mp4"}) as session: + if isinstance(src_video_stream, BytesIO): + src_video_stream.seek(0) + async with session.put(upload_res.urls[0], data=src_video_stream, raise_for_status=True) as res: + upload_etag = res.headers["Etag"] + else: + with builtins.open(src_video_stream, "rb") as video_file: + async with session.put(upload_res.urls[0], data=video_file, raise_for_status=True) as res: + upload_etag = res.headers["Etag"] + await sync_op( + cls, + ApiEndpoint( + path=f"/proxy/topaz/video/{initial_res.requestId}/complete-upload", + method="PATCH", + ), + response_model=VideoCompleteUploadResponse, + data=VideoCompleteUploadRequest( + uploadResults=[ + VideoCompleteUploadRequestPart( + partNum=1, + eTag=upload_etag, + ), + ], + ), + wait_label="Finalizing upload", + final_label_on_success="Upload completed", + ) + final_response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/topaz/video/{initial_res.requestId}/status"), + response_model=VideoStatusResponse, + status_extractor=lambda x: x.status, + progress_extractor=lambda x: getattr(x, "progress", 0), + price_extractor=lambda x: (x.estimates.cost[0] * 0.08 if x.estimates and x.estimates.cost[0] else None), + poll_interval=10.0, ) return IO.NodeOutput(await download_url_to_video_output(final_response.download.url)) @@ -464,6 +819,7 @@ class TopazExtension(ComfyExtension): return [ TopazImageEnhance, TopazVideoEnhance, + TopazVideoEnhanceV2, ] diff --git a/comfy_api_nodes/nodes_tripo.py b/comfy_api_nodes/nodes_tripo.py index 9f4298dce..d6501dee4 100644 --- a/comfy_api_nodes/nodes_tripo.py +++ b/comfy_api_nodes/nodes_tripo.py @@ -60,6 +60,7 @@ async def poll_until_finished( ], status_extractor=lambda x: x.data.status, progress_extractor=lambda x: x.data.progress, + price_extractor=lambda x: x.data.consumed_credit * 0.01 if x.data.consumed_credit else None, estimated_duration=average_duration, ) if response_poll.data.status == TripoTaskStatus.SUCCESS: @@ -113,7 +114,6 @@ class TripoTextToModelNode(IO.ComfyNode): depends_on=IO.PriceBadgeDepends( widgets=[ "model_version", - "style", "texture", "pbr", "quad", @@ -124,20 +124,17 @@ class TripoTextToModelNode(IO.ComfyNode): expr=""" ( $isV14 := $contains(widgets.model_version,"v1.4"); - $style := widgets.style; - $hasStyle := ($style != "" and $style != "none"); + $isV3OrLater := $contains(widgets.model_version,"v3."); $withTexture := widgets.texture or widgets.pbr; $isHdTexture := (widgets.texture_quality = "detailed"); $isDetailedGeometry := (widgets.geometry_quality = "detailed"); - $baseCredits := - $isV14 ? 20 : ($withTexture ? 20 : 10); - $credits := - $baseCredits - + ($hasStyle ? 5 : 0) + $credits := $isV14 ? 20 : ( + ($withTexture ? 20 : 10) + (widgets.quad ? 5 : 0) + ($isHdTexture ? 10 : 0) - + ($isDetailedGeometry ? 20 : 0); - {"type":"usd","usd": $round($credits * 0.01, 2)} + + (($isDetailedGeometry and $isV3OrLater) ? 20 : 0) + ); + {"type":"usd","usd": $round($credits * 0.01, 2), "format": {"approximate": true}} ) """, ), @@ -239,7 +236,6 @@ class TripoImageToModelNode(IO.ComfyNode): depends_on=IO.PriceBadgeDepends( widgets=[ "model_version", - "style", "texture", "pbr", "quad", @@ -250,20 +246,17 @@ class TripoImageToModelNode(IO.ComfyNode): expr=""" ( $isV14 := $contains(widgets.model_version,"v1.4"); - $style := widgets.style; - $hasStyle := ($style != "" and $style != "none"); + $isV3OrLater := $contains(widgets.model_version,"v3."); $withTexture := widgets.texture or widgets.pbr; $isHdTexture := (widgets.texture_quality = "detailed"); $isDetailedGeometry := (widgets.geometry_quality = "detailed"); - $baseCredits := - $isV14 ? 30 : ($withTexture ? 30 : 20); - $credits := - $baseCredits - + ($hasStyle ? 5 : 0) + $credits := $isV14 ? 30 : ( + ($withTexture ? 30 : 20) + (widgets.quad ? 5 : 0) + ($isHdTexture ? 10 : 0) - + ($isDetailedGeometry ? 20 : 0); - {"type":"usd","usd": $round($credits * 0.01, 2)} + + (($isDetailedGeometry and $isV3OrLater) ? 20 : 0) + ); + {"type":"usd","usd": $round($credits * 0.01, 2), "format": {"approximate": true}} ) """, ), @@ -358,7 +351,7 @@ class TripoMultiviewToModelNode(IO.ComfyNode): "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.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=[ @@ -379,7 +372,6 @@ class TripoMultiviewToModelNode(IO.ComfyNode): "model_version", "texture", "pbr", - "quad", "texture_quality", "geometry_quality", ], @@ -387,17 +379,16 @@ class TripoMultiviewToModelNode(IO.ComfyNode): expr=""" ( $isV14 := $contains(widgets.model_version,"v1.4"); + $isV3OrLater := $contains(widgets.model_version,"v3."); $withTexture := widgets.texture or widgets.pbr; $isHdTexture := (widgets.texture_quality = "detailed"); $isDetailedGeometry := (widgets.geometry_quality = "detailed"); - $baseCredits := - $isV14 ? 30 : ($withTexture ? 30 : 20); - $credits := - $baseCredits - + (widgets.quad ? 5 : 0) + $credits := $isV14 ? 30 : ( + ($withTexture ? 30 : 20) + ($isHdTexture ? 10 : 0) - + ($isDetailedGeometry ? 20 : 0); - {"type":"usd","usd": $round($credits * 0.01, 2)} + + (($isDetailedGeometry and $isV3OrLater) ? 20 : 0) + ); + {"type":"usd","usd": $round($credits * 0.01, 2), "format": {"approximate": true}} ) """, ), @@ -457,7 +448,7 @@ class TripoMultiviewToModelNode(IO.ComfyNode): geometry_quality=geometry_quality, texture_alignment=texture_alignment, face_limit=face_limit if face_limit != -1 else None, - quad=quad, + quad=None, ), ) return await poll_until_finished(cls, response, average_duration=80) @@ -498,7 +489,7 @@ class TripoTextureNode(IO.ComfyNode): expr=""" ( $tq := widgets.texture_quality; - {"type":"usd","usd": ($contains($tq,"detailed") ? 0.2 : 0.1)} + {"type":"usd","usd": ($contains($tq,"detailed") ? 0.2 : 0.1), "format": {"approximate": true}} ) """, ), @@ -555,7 +546,7 @@ class TripoRefineNode(IO.ComfyNode): is_api_node=True, is_output_node=True, price_badge=IO.PriceBadge( - expr="""{"type":"usd","usd":0.3}""", + expr="""{"type":"usd","usd":0.3, "format": {"approximate": true}}""", ), ) @@ -592,7 +583,7 @@ class TripoRigNode(IO.ComfyNode): is_api_node=True, is_output_node=True, price_badge=IO.PriceBadge( - expr="""{"type":"usd","usd":0.25}""", + expr="""{"type":"usd","usd":0.25, "format": {"approximate": true}}""", ), ) @@ -652,7 +643,7 @@ class TripoRetargetNode(IO.ComfyNode): is_api_node=True, is_output_node=True, price_badge=IO.PriceBadge( - expr="""{"type":"usd","usd":0.1}""", + expr="""{"type":"usd","usd":0.1, "format": {"approximate": true}}""", ), ) @@ -761,19 +752,10 @@ class TripoConversionNode(IO.ComfyNode): "face_limit", "texture_size", "texture_format", - "force_symmetry", "flatten_bottom", "flatten_bottom_threshold", "pivot_to_center_bottom", "scale_factor", - "with_animation", - "pack_uv", - "bake", - "part_names", - "fbx_preset", - "export_vertex_colors", - "export_orientation", - "animate_in_place", ], ), expr=""" @@ -783,28 +765,16 @@ class TripoConversionNode(IO.ComfyNode): $flatThresh := (widgets.flatten_bottom_threshold != null) ? widgets.flatten_bottom_threshold : 0; $scale := (widgets.scale_factor != null) ? widgets.scale_factor : 1; $texFmt := (widgets.texture_format != "" ? widgets.texture_format : "jpeg"); - $part := widgets.part_names; - $fbx := (widgets.fbx_preset != "" ? widgets.fbx_preset : "blender"); - $orient := (widgets.export_orientation != "" ? widgets.export_orientation : "default"); $advanced := widgets.quad or - widgets.force_symmetry or widgets.flatten_bottom or widgets.pivot_to_center_bottom or - widgets.with_animation or - widgets.pack_uv or - widgets.bake or - widgets.export_vertex_colors or - widgets.animate_in_place or ($face != -1) or ($texSize != 4096) or ($flatThresh != 0) or ($scale != 1) or - ($texFmt != "jpeg") or - ($part != "") or - ($fbx != "blender") or - ($orient != "default"); - {"type":"usd","usd": ($advanced ? 0.1 : 0.05)} + ($texFmt != "jpeg"); + {"type":"usd","usd": ($advanced ? 0.1 : 0.05), "format": {"approximate": true}} ) """, ), diff --git a/comfy_api_nodes/nodes_vidu.py b/comfy_api_nodes/nodes_vidu.py index f04407eb5..8d90cefeb 100644 --- a/comfy_api_nodes/nodes_vidu.py +++ b/comfy_api_nodes/nodes_vidu.py @@ -38,7 +38,7 @@ async def execute_task( cls: type[IO.ComfyNode], vidu_endpoint: str, payload: TaskCreationRequest | TaskExtendCreationRequest | TaskMultiFrameCreationRequest, - max_poll_attempts: int = 320, + max_poll_attempts: int = 480, ) -> list[TaskResult]: task_creation_response = await sync_op( cls, @@ -1097,7 +1097,6 @@ class ViduExtendVideoNode(IO.ComfyNode): video_url=await upload_video_to_comfyapi(cls, video, wait_label="Uploading video"), images=[image_url] if image_url else None, ), - max_poll_attempts=480, ) return IO.NodeOutput(await download_url_to_video_output(results[0].url)) diff --git a/comfy_api_nodes/nodes_wan.py b/comfy_api_nodes/nodes_wan.py index 7d7466fb6..68061bb5c 100644 --- a/comfy_api_nodes/nodes_wan.py +++ b/comfy_api_nodes/nodes_wan.py @@ -818,7 +818,6 @@ class WanReferenceVideoApi(IO.ComfyNode): response_model=VideoTaskStatusResponse, status_extractor=lambda x: x.output.task_status, poll_interval=6, - max_poll_attempts=280, ) return IO.NodeOutput(await download_url_to_video_output(response.output.video_url)) diff --git a/comfy_api_nodes/nodes_wavespeed.py b/comfy_api_nodes/nodes_wavespeed.py index c59fafd3b..65e45f60a 100644 --- a/comfy_api_nodes/nodes_wavespeed.py +++ b/comfy_api_nodes/nodes_wavespeed.py @@ -84,7 +84,6 @@ class WavespeedFlashVSRNode(IO.ComfyNode): response_model=TaskResultResponse, status_extractor=lambda x: "failed" if x.data is None else x.data.status, poll_interval=10.0, - max_poll_attempts=480, ) if final_response.code != 200: raise ValueError( @@ -156,7 +155,6 @@ class WavespeedImageUpscaleNode(IO.ComfyNode): response_model=TaskResultResponse, status_extractor=lambda x: "failed" if x.data is None else x.data.status, poll_interval=10.0, - max_poll_attempts=480, ) if final_response.code != 200: raise ValueError( diff --git a/comfy_api_nodes/util/client.py b/comfy_api_nodes/util/client.py index b0cf97ae4..052301c33 100644 --- a/comfy_api_nodes/util/client.py +++ b/comfy_api_nodes/util/client.py @@ -19,6 +19,8 @@ from comfy import utils from comfy_api.latest import IO from server import PromptServer +from comfy.deploy_environment import get_deploy_environment + from . import request_logger from ._helpers import ( default_base_url, @@ -148,7 +150,7 @@ async def poll_op( queued_statuses: list[str | int] | None = None, data: BaseModel | None = None, poll_interval: float = 5.0, - max_poll_attempts: int = 160, + max_poll_attempts: int = 480, timeout_per_poll: float = 120.0, max_retries_per_poll: int = 10, retry_delay_per_poll: float = 1.0, @@ -254,7 +256,7 @@ async def poll_op_raw( queued_statuses: list[str | int] | None = None, data: dict[str, Any] | BaseModel | None = None, poll_interval: float = 5.0, - max_poll_attempts: int = 160, + max_poll_attempts: int = 480, timeout_per_poll: float = 120.0, max_retries_per_poll: int = 10, retry_delay_per_poll: float = 1.0, @@ -486,10 +488,30 @@ async def _diagnose_connectivity() -> dict[str, bool]: "api_accessible": False, } timeout = aiohttp.ClientTimeout(total=5.0) + + # Probe Google and Baidu in parallel: Google is blocked by the GFW in mainland China, so a Baidu probe is required + # to correctly detect that Chinese users with working internet do have working internet. + internet_probe_urls = ("https://www.google.com", "https://www.baidu.com") + async with aiohttp.ClientSession(timeout=timeout) as session: - with contextlib.suppress(ClientError, OSError): - async with session.get("https://www.google.com") as resp: - results["internet_accessible"] = resp.status < 500 + async def _probe(url: str) -> bool: + try: + async with session.get(url) as resp: + return resp.status < 500 + except (ClientError, OSError, asyncio.TimeoutError): + return False + + probe_tasks = [asyncio.create_task(_probe(u)) for u in internet_probe_urls] + try: + for fut in asyncio.as_completed(probe_tasks): + if await fut: + results["internet_accessible"] = True + break + finally: + for t in probe_tasks: + if not t.done(): + t.cancel() + await asyncio.gather(*probe_tasks, return_exceptions=True) if not results["internet_accessible"]: return results @@ -624,6 +646,7 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool): payload_headers = {"Accept": "*/*"} if expect_binary else {"Accept": "application/json"} if not parsed_url.scheme and not parsed_url.netloc: # is URL relative? payload_headers.update(get_auth_header(cfg.node_cls)) + payload_headers["Comfy-Env"] = get_deploy_environment() if cfg.endpoint.headers: payload_headers.update(cfg.endpoint.headers) diff --git a/comfy_extras/frame_interpolation_models/film_net.py b/comfy_extras/frame_interpolation_models/film_net.py index cf4f6e1e1..36bc79dc3 100644 --- a/comfy_extras/frame_interpolation_models/film_net.py +++ b/comfy_extras/frame_interpolation_models/film_net.py @@ -199,6 +199,9 @@ class FILMNet(nn.Module): def get_dtype(self): return self.extract.extract_sublevels.convs[0][0].conv.weight.dtype + def memory_used_forward(self, shape, dtype): + return 1700 * shape[1] * shape[2] * dtype.itemsize + def _build_warp_grids(self, H, W, device): """Pre-compute warp grids for all pyramid levels.""" if (H, W) in self._warp_grids: diff --git a/comfy_extras/frame_interpolation_models/ifnet.py b/comfy_extras/frame_interpolation_models/ifnet.py index 03cb34c50..ad6edbec9 100644 --- a/comfy_extras/frame_interpolation_models/ifnet.py +++ b/comfy_extras/frame_interpolation_models/ifnet.py @@ -74,6 +74,9 @@ class IFNet(nn.Module): def get_dtype(self): return self.encode.cnn0.weight.dtype + def memory_used_forward(self, shape, dtype): + return 300 * shape[1] * shape[2] * dtype.itemsize + def _build_warp_grids(self, H, W, device): if (H, W) in self._warp_grids: return diff --git a/comfy_extras/nodes_ace.py b/comfy_extras/nodes_ace.py index 1602add84..247d9ae8a 100644 --- a/comfy_extras/nodes_ace.py +++ b/comfy_extras/nodes_ace.py @@ -42,7 +42,7 @@ class TextEncodeAceStepAudio15(IO.ComfyNode): IO.Int.Input("bpm", default=120, min=10, max=300), IO.Float.Input("duration", default=120.0, min=0.0, max=2000.0, step=0.1), IO.Combo.Input("timesignature", options=['2', '3', '4', '6']), - IO.Combo.Input("language", options=["en", "ja", "zh", "es", "de", "fr", "pt", "ru", "it", "nl", "pl", "tr", "vi", "cs", "fa", "id", "ko", "uk", "hu", "ar", "sv", "ro", "el"]), + IO.Combo.Input("language", options=['ar', 'az', 'bg', 'bn', 'ca', 'cs', 'da', 'de', 'el', 'en', 'es', 'fa', 'fi', 'fr', 'he', 'hi', 'hr', 'ht', 'hu', 'id', 'is', 'it', 'ja', 'ko', 'la', 'lt', 'ms', 'ne', 'nl', 'no', 'pa', 'pl', 'pt', 'ro', 'ru', 'sa', 'sk', 'sr', 'sv', 'sw', 'ta', 'te', 'th', 'tl', 'tr', 'uk', 'ur', 'vi', 'yue', 'zh', 'unknown'], default='en'), IO.Combo.Input("keyscale", options=[f"{root} {quality}" for quality in ["major", "minor"] for root in ["C", "C#", "Db", "D", "D#", "Eb", "E", "F", "F#", "Gb", "G", "G#", "Ab", "A", "A#", "Bb", "B"]]), IO.Boolean.Input("generate_audio_codes", default=True, tooltip="Enable the LLM that generates audio codes. This can be slow but will increase the quality of the generated audio. Turn this off if you are giving the model an audio reference.", advanced=True), IO.Float.Input("cfg_scale", default=2.0, min=0.0, max=100.0, step=0.1, advanced=True), @@ -104,7 +104,7 @@ class EmptyAceStep15LatentAudio(IO.ComfyNode): def execute(cls, seconds, batch_size) -> IO.NodeOutput: length = round((seconds * 48000 / 1920)) latent = torch.zeros([batch_size, 64, length], device=comfy.model_management.intermediate_device(), dtype=comfy.model_management.intermediate_dtype()) - return IO.NodeOutput({"samples": latent, "type": "audio"}) + return IO.NodeOutput({"samples": latent, "type": "audio", "downscale_ratio_temporal": 1764}) class ReferenceAudio(IO.ComfyNode): @classmethod diff --git a/comfy_extras/nodes_advanced_samplers.py b/comfy_extras/nodes_advanced_samplers.py index 7f716cd76..20717ca38 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/custom_sampling/samplers", + category="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), @@ -86,13 +86,44 @@ def sample_euler_pp(model, x, sigmas, extra_args=None, callback=None, disable=No return x +class SamplerLCM(io.ComfyNode): + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="SamplerLCM", + category="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, + tooltip="Per-step noise multiplier at the first step (1.0 = match training)."), + io.Float.Input("s_noise_end", default=1.0, min=0.0, max=64.0, step=0.01, + tooltip="Per-step noise multiplier at the last step. Set equal to s_noise for a constant schedule."), + io.Float.Input("noise_clip_std", default=0.0, min=0.0, max=10.0, step=0.01, + tooltip="Clamp per-step noise to +/- N*std. 0 disables."), + ], + outputs=[io.Sampler.Output()], + ) + + @classmethod + def execute(cls, s_noise, s_noise_end, noise_clip_std) -> io.NodeOutput: + sampler = comfy.samplers.ksampler( + "lcm", + { + "s_noise": float(s_noise), + "s_noise_end": float(s_noise_end), + "noise_clip_std": float(noise_clip_std), + }, + ) + return io.NodeOutput(sampler) + + class SamplerEulerCFGpp(io.ComfyNode): @classmethod def define_schema(cls) -> io.Schema: return io.Schema( node_id="SamplerEulerCFGpp", display_name="SamplerEulerCFG++", - category="_for_testing", # "sampling/custom_sampling/samplers" + category="experimental", # "sampling/samplers" inputs=[ io.Combo.Input("version", options=["regular", "alternative"], advanced=True), ], @@ -114,6 +145,7 @@ class AdvancedSamplersExtension(ComfyExtension): async def get_node_list(self) -> list[type[io.ComfyNode]]: return [ SamplerLCMUpscale, + SamplerLCM, SamplerEulerCFGpp, ] diff --git a/comfy_extras/nodes_align_your_steps.py b/comfy_extras/nodes_align_your_steps.py index 4fc511d2c..307f41337 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/custom_sampling/schedulers", + category="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_ar_video.py b/comfy_extras/nodes_ar_video.py new file mode 100644 index 000000000..1a15facfa --- /dev/null +++ b/comfy_extras/nodes_ar_video.py @@ -0,0 +1,136 @@ +""" +ComfyUI nodes for autoregressive video generation (Causal Forcing, Self-Forcing, etc.). + - EmptyARVideoLatent: create 5D [B, C, T, H, W] video latent tensors + - SamplerARVideo: SAMPLER for the block-by-block autoregressive denoising loop + - ARVideoI2V: image-to-video conditioning for AR models (seeds KV cache with start image) +""" + +import torch +from typing_extensions import override + +import comfy.model_management +import comfy.samplers +import comfy.utils +from comfy_api.latest import ComfyExtension, io + + +class EmptyARVideoLatent(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="EmptyARVideoLatent", + category="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), + io.Int.Input("length", default=81, min=1, max=1024, step=4), + io.Int.Input("batch_size", default=1, min=1, max=64), + ], + outputs=[ + io.Latent.Output(display_name="LATENT"), + ], + ) + + @classmethod + def execute(cls, width, height, length, batch_size) -> io.NodeOutput: + lat_t = ((length - 1) // 4) + 1 + latent = torch.zeros( + [batch_size, 16, lat_t, height // 8, width // 8], + device=comfy.model_management.intermediate_device(), + ) + return io.NodeOutput({"samples": latent}) + + +class SamplerARVideo(io.ComfyNode): + """Sampler for autoregressive video models (Causal Forcing, Self-Forcing). + + All AR-loop parameters are owned by this node so they live in the workflow. + Add new widgets here as the AR sampler grows new options. + """ + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="SamplerARVideo", + display_name="Sampler AR Video", + category="sampling/samplers", + inputs=[ + io.Int.Input( + "num_frame_per_block", + default=1, min=1, max=64, + tooltip="Frames per autoregressive block. 1 = framewise, " + "3 = chunkwise. Must match the checkpoint's training mode.", + ), + ], + outputs=[io.Sampler.Output()], + ) + + @classmethod + def execute(cls, num_frame_per_block) -> io.NodeOutput: + extra_options = { + "num_frame_per_block": num_frame_per_block, + } + return io.NodeOutput(comfy.samplers.ksampler("ar_video", extra_options)) + + +class ARVideoI2V(io.ComfyNode): + """Image-to-video setup for AR video models (Causal Forcing, Self-Forcing). + + VAE-encodes the start image and stores it in the model's transformer_options + so that sample_ar_video can seed the KV cache before denoising. + Uses the same T2V model checkpoint -- no separate I2V architecture needed. + """ + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="ARVideoI2V", + category="conditioning/video_models", + inputs=[ + io.Model.Input("model"), + io.Vae.Input("vae"), + io.Image.Input("start_image"), + 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=1024, step=4), + io.Int.Input("batch_size", default=1, min=1, max=64), + ], + outputs=[ + io.Model.Output(display_name="MODEL"), + io.Latent.Output(display_name="LATENT"), + ], + ) + + @classmethod + def execute(cls, model, vae, start_image, width, height, length, batch_size) -> io.NodeOutput: + start_image = comfy.utils.common_upscale( + start_image[:1].movedim(-1, 1), width, height, "bilinear", "center" + ).movedim(1, -1) + + initial_latent = vae.encode(start_image[:, :, :, :3]) + + m = model.clone() + to = m.model_options.setdefault("transformer_options", {}) + ar_cfg = to.setdefault("ar_config", {}) + ar_cfg["initial_latent"] = initial_latent + + lat_t = ((length - 1) // 4) + 1 + latent = torch.zeros( + [batch_size, 16, lat_t, height // 8, width // 8], + device=comfy.model_management.intermediate_device(), + ) + return io.NodeOutput(m, {"samples": latent}) + + +class ARVideoExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + EmptyARVideoLatent, + SamplerARVideo, + ARVideoI2V, + ] + + +async def comfy_entrypoint() -> ARVideoExtension: + return ARVideoExtension() diff --git a/comfy_extras/nodes_attention_multiply.py b/comfy_extras/nodes_attention_multiply.py index 060a5c9be..f4ee6a689 100644 --- a/comfy_extras/nodes_attention_multiply.py +++ b/comfy_extras/nodes_attention_multiply.py @@ -25,7 +25,7 @@ class UNetSelfAttentionMultiply(io.ComfyNode): def define_schema(cls) -> io.Schema: return io.Schema( node_id="UNetSelfAttentionMultiply", - category="_for_testing/attention_experiments", + category="experimental/attention_experiments", inputs=[ io.Model.Input("model"), io.Float.Input("q", default=1.0, min=0.0, max=10.0, step=0.01, advanced=True), @@ -48,7 +48,7 @@ class UNetCrossAttentionMultiply(io.ComfyNode): def define_schema(cls) -> io.Schema: return io.Schema( node_id="UNetCrossAttentionMultiply", - category="_for_testing/attention_experiments", + category="experimental/attention_experiments", inputs=[ io.Model.Input("model"), io.Float.Input("q", default=1.0, min=0.0, max=10.0, step=0.01, advanced=True), @@ -72,7 +72,7 @@ class CLIPAttentionMultiply(io.ComfyNode): return io.Schema( node_id="CLIPAttentionMultiply", search_aliases=["clip attention scale", "text encoder attention"], - category="_for_testing/attention_experiments", + category="experimental/attention_experiments", inputs=[ io.Clip.Input("clip"), io.Float.Input("q", default=1.0, min=0.0, max=10.0, step=0.01, advanced=True), @@ -106,7 +106,7 @@ class UNetTemporalAttentionMultiply(io.ComfyNode): def define_schema(cls) -> io.Schema: return io.Schema( node_id="UNetTemporalAttentionMultiply", - category="_for_testing/attention_experiments", + category="experimental/attention_experiments", inputs=[ io.Model.Input("model"), io.Float.Input("self_structural", default=1.0, min=0.0, max=10.0, step=0.01, advanced=True), diff --git a/comfy_extras/nodes_audio.py b/comfy_extras/nodes_audio.py index 5f514716f..2d6b3c7ea 100644 --- a/comfy_extras/nodes_audio.py +++ b/comfy_extras/nodes_audio.py @@ -33,7 +33,7 @@ class EmptyLatentAudio(IO.ComfyNode): def execute(cls, seconds, batch_size) -> IO.NodeOutput: length = round((seconds * 44100 / 2048) / 2) * 2 latent = torch.zeros([batch_size, 64, length], device=comfy.model_management.intermediate_device()) - return IO.NodeOutput({"samples":latent, "type": "audio"}) + return IO.NodeOutput({"samples": latent, "type": "audio", "downscale_ratio_temporal": 2048}) generate = execute # TODO: remove @@ -82,6 +82,8 @@ class VAEEncodeAudio(IO.ComfyNode): @classmethod def execute(cls, vae, audio) -> IO.NodeOutput: + if audio is None: + raise ValueError("VAEEncodeAudio: input audio is None (source video may have no audio track).") sample_rate = audio["sample_rate"] vae_sample_rate = getattr(vae, "audio_sample_rate", 44100) if vae_sample_rate != sample_rate: @@ -171,6 +173,8 @@ class SaveAudio(IO.ComfyNode): @classmethod def execute(cls, audio, filename_prefix="ComfyUI", format="flac") -> IO.NodeOutput: + if audio is None: + raise ValueError("SaveAudio: input audio is None (source video may have no audio track).") return IO.NodeOutput( ui=UI.AudioSaveHelper.get_save_audio_ui(audio, filename_prefix=filename_prefix, cls=cls, format=format) ) @@ -198,6 +202,8 @@ class SaveAudioMP3(IO.ComfyNode): @classmethod def execute(cls, audio, filename_prefix="ComfyUI", format="mp3", quality="128k") -> IO.NodeOutput: + if audio is None: + raise ValueError("SaveAudioMP3: input audio is None (source video may have no audio track).") return IO.NodeOutput( ui=UI.AudioSaveHelper.get_save_audio_ui( audio, filename_prefix=filename_prefix, cls=cls, format=format, quality=quality @@ -226,6 +232,8 @@ class SaveAudioOpus(IO.ComfyNode): @classmethod def execute(cls, audio, filename_prefix="ComfyUI", format="opus", quality="V3") -> IO.NodeOutput: + if audio is None: + raise ValueError("SaveAudioOpus: input audio is None (source video may have no audio track).") return IO.NodeOutput( ui=UI.AudioSaveHelper.get_save_audio_ui( audio, filename_prefix=filename_prefix, cls=cls, format=format, quality=quality @@ -252,6 +260,8 @@ class PreviewAudio(IO.ComfyNode): @classmethod def execute(cls, audio) -> IO.NodeOutput: + if audio is None: + raise ValueError("PreviewAudio: input audio is None (source video may have no audio track).") return IO.NodeOutput(ui=UI.PreviewAudio(audio, cls=cls)) save_flac = execute # TODO: remove @@ -297,6 +307,7 @@ class LoadAudio(IO.ComfyNode): @classmethod def define_schema(cls): input_dir = folder_paths.get_input_directory() + os.makedirs(input_dir, exist_ok=True) files = folder_paths.filter_files_content_types(os.listdir(input_dir), ["audio", "video"]) return IO.Schema( node_id="LoadAudio", @@ -391,21 +402,26 @@ class TrimAudioDuration(IO.ComfyNode): @classmethod def execute(cls, audio, start_index, duration) -> IO.NodeOutput: + if audio is None: + return IO.NodeOutput(None) waveform = audio["waveform"] sample_rate = audio["sample_rate"] audio_length = waveform.shape[-1] + if audio_length == 0: + return IO.NodeOutput(audio) + if start_index < 0: start_frame = audio_length + int(round(start_index * sample_rate)) else: start_frame = int(round(start_index * sample_rate)) - start_frame = max(0, min(start_frame, audio_length - 1)) + start_frame = max(0, min(start_frame, audio_length)) end_frame = start_frame + int(round(duration * sample_rate)) end_frame = max(0, min(end_frame, audio_length)) if start_frame >= end_frame: - raise ValueError("AudioTrim: Start time must be less than end time and be within the audio length.") + raise ValueError("TrimAudioDuration: Start time must be less than end time and be within the audio length.") return IO.NodeOutput({"waveform": waveform[..., start_frame:end_frame], "sample_rate": sample_rate}) @@ -432,11 +448,13 @@ class SplitAudioChannels(IO.ComfyNode): @classmethod def execute(cls, audio) -> IO.NodeOutput: + if audio is None: + return IO.NodeOutput(None, None) waveform = audio["waveform"] sample_rate = audio["sample_rate"] if waveform.shape[1] != 2: - raise ValueError("AudioSplit: Input audio has only one channel.") + raise ValueError(f"AudioSplit: Input audio must be stereo (2 channels), got {waveform.shape[1]} channel(s).") left_channel = waveform[..., 0:1, :] right_channel = waveform[..., 1:2, :] @@ -464,6 +482,12 @@ class JoinAudioChannels(IO.ComfyNode): @classmethod def execute(cls, audio_left, audio_right) -> IO.NodeOutput: + if audio_left is None and audio_right is None: + return IO.NodeOutput(None) + if audio_left is None: + return IO.NodeOutput(audio_right) + if audio_right is None: + return IO.NodeOutput(audio_left) waveform_left = audio_left["waveform"] sample_rate_left = audio_left["sample_rate"] waveform_right = audio_right["waveform"] @@ -537,6 +561,12 @@ class AudioConcat(IO.ComfyNode): @classmethod def execute(cls, audio1, audio2, direction) -> IO.NodeOutput: + if audio1 is None and audio2 is None: + return IO.NodeOutput(None) + if audio1 is None: + return IO.NodeOutput(audio2) + if audio2 is None: + return IO.NodeOutput(audio1) waveform_1 = audio1["waveform"] waveform_2 = audio2["waveform"] sample_rate_1 = audio1["sample_rate"] @@ -584,6 +614,12 @@ class AudioMerge(IO.ComfyNode): @classmethod def execute(cls, audio1, audio2, merge_method) -> IO.NodeOutput: + if audio1 is None and audio2 is None: + return IO.NodeOutput(None) + if audio1 is None: + return IO.NodeOutput(audio2) + if audio2 is None: + return IO.NodeOutput(audio1) waveform_1 = audio1["waveform"] waveform_2 = audio2["waveform"] sample_rate_1 = audio1["sample_rate"] @@ -594,6 +630,9 @@ class AudioMerge(IO.ComfyNode): length_1 = waveform_1.shape[-1] length_2 = waveform_2.shape[-1] + if length_1 == 0 or length_2 == 0: + return IO.NodeOutput({"waveform": waveform_1, "sample_rate": output_sample_rate}) + if length_2 > length_1: logging.info(f"AudioMerge: Trimming audio2 from {length_2} to {length_1} samples to match audio1 length.") waveform_2 = waveform_2[..., :length_1] @@ -645,6 +684,8 @@ class AudioAdjustVolume(IO.ComfyNode): @classmethod def execute(cls, audio, volume) -> IO.NodeOutput: + if audio is None: + return IO.NodeOutput(None) if volume == 0: return IO.NodeOutput(audio) waveform = audio["waveform"] @@ -728,8 +769,14 @@ class AudioEqualizer3Band(IO.ComfyNode): @classmethod def execute(cls, audio, low_gain_dB, low_freq, mid_gain_dB, mid_freq, mid_q, high_gain_dB, high_freq) -> IO.NodeOutput: + if audio is None: + return IO.NodeOutput(None) waveform = audio["waveform"] sample_rate = audio["sample_rate"] + + if waveform.shape[-1] == 0: + return IO.NodeOutput(audio) + eq_waveform = waveform.clone() # 1. Apply Low Shelf (Bass) diff --git a/comfy_extras/nodes_audio_encoder.py b/comfy_extras/nodes_audio_encoder.py index 13aacd41a..6a85da89b 100644 --- a/comfy_extras/nodes_audio_encoder.py +++ b/comfy_extras/nodes_audio_encoder.py @@ -10,6 +10,7 @@ class AudioEncoderLoader(io.ComfyNode): def define_schema(cls) -> io.Schema: return io.Schema( node_id="AudioEncoderLoader", + display_name="Load Audio Encoder", category="loaders", inputs=[ io.Combo.Input( diff --git a/comfy_extras/nodes_bg_removal.py b/comfy_extras/nodes_bg_removal.py new file mode 100644 index 000000000..793fd802b --- /dev/null +++ b/comfy_extras/nodes_bg_removal.py @@ -0,0 +1,61 @@ +import folder_paths +from typing_extensions import override +from comfy_api.latest import ComfyExtension, IO +from comfy.bg_removal_model import load + + +class LoadBackgroundRemovalModel(IO.ComfyNode): + @classmethod + def define_schema(cls): + files = folder_paths.get_filename_list("background_removal") + return IO.Schema( + node_id="LoadBackgroundRemovalModel", + display_name="Load Background Removal Model", + category="loaders", + inputs=[ + IO.Combo.Input("bg_removal_name", options=sorted(files), tooltip="The model used to remove backgrounds from images"), + ], + outputs=[ + IO.BackgroundRemoval.Output("bg_model") + ] + ) + @classmethod + def execute(cls, bg_removal_name): + path = folder_paths.get_full_path_or_raise("background_removal", bg_removal_name) + bg = load(path) + if bg is None: + raise RuntimeError("ERROR: background model file is invalid and does not contain a valid background removal model.") + return IO.NodeOutput(bg) + +class RemoveBackground(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="RemoveBackground", + display_name="Remove Background", + 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") + ], + outputs=[ + IO.Mask.Output("mask", tooltip="Generated foreground mask") + ] + ) + @classmethod + def execute(cls, image, bg_removal_model): + mask = bg_removal_model.encode_image(image) + return IO.NodeOutput(mask) + +class BackgroundRemovalExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + LoadBackgroundRemovalModel, + RemoveBackground + ] + + +async def comfy_entrypoint() -> BackgroundRemovalExtension: + return BackgroundRemovalExtension() diff --git a/comfy_extras/nodes_camera_trajectory.py b/comfy_extras/nodes_camera_trajectory.py index e7efa29ba..34b78e81b 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="camera", + category="conditioning/video_models", inputs=[ io.Combo.Input( "camera_pose", diff --git a/comfy_extras/nodes_canny.py b/comfy_extras/nodes_canny.py index 648b4279d..462f6fea0 100644 --- a/comfy_extras/nodes_canny.py +++ b/comfy_extras/nodes_canny.py @@ -11,9 +11,9 @@ class Canny(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="Canny", - display_name="Canny", + display_name="Detect Edges (Canny)", search_aliases=["edge detection", "outline", "contour detection", "line art"], - category="image/preprocessors", + category="image/filters", essentials_category="Image Tools", inputs=[ io.Image.Input("image"), diff --git a/comfy_extras/nodes_compositing.py b/comfy_extras/nodes_compositing.py index 3bc9fccb3..8fcbe720e 100644 --- a/comfy_extras/nodes_compositing.py +++ b/comfy_extras/nodes_compositing.py @@ -111,7 +111,7 @@ class PorterDuffImageComposite(io.ComfyNode): node_id="PorterDuffImageComposite", search_aliases=["alpha composite", "blend modes", "layer blend", "transparency blend"], display_name="Porter-Duff Image Composite", - category="mask/compositing", + category="image/compositing", inputs=[ io.Image.Input("source"), io.Mask.Input("source_alpha"), @@ -168,7 +168,7 @@ class SplitImageWithAlpha(io.ComfyNode): node_id="SplitImageWithAlpha", search_aliases=["extract alpha", "separate transparency", "remove alpha"], display_name="Split Image with Alpha", - category="mask/compositing", + category="image/compositing", inputs=[ io.Image.Input("image"), ], @@ -192,7 +192,7 @@ class JoinImageWithAlpha(io.ComfyNode): node_id="JoinImageWithAlpha", search_aliases=["add transparency", "apply alpha", "composite alpha", "RGBA"], display_name="Join Image with Alpha", - category="mask/compositing", + category="image/compositing", inputs=[ io.Image.Input("image"), io.Mask.Input("alpha"), @@ -202,14 +202,11 @@ class JoinImageWithAlpha(io.ComfyNode): @classmethod def execute(cls, image: torch.Tensor, alpha: torch.Tensor) -> io.NodeOutput: - batch_size = min(len(image), len(alpha)) - out_images = [] - - alpha = 1.0 - resize_mask(alpha, image.shape[1:]) - for i in range(batch_size): - out_images.append(torch.cat((image[i][:,:,:3], alpha[i].unsqueeze(2)), dim=2)) - - return io.NodeOutput(torch.stack(out_images)) + batch_size = max(len(image), len(alpha)) + alpha = 1.0 - resize_mask(alpha.to(image), image.shape[1:]) + alpha = comfy.utils.repeat_to_batch_size(alpha, batch_size) + image = comfy.utils.repeat_to_batch_size(image, batch_size) + return io.NodeOutput(torch.cat((image[..., :3], alpha.unsqueeze(-1)), dim=-1)) class CompositingExtension(ComfyExtension): diff --git a/comfy_extras/nodes_cond.py b/comfy_extras/nodes_cond.py index 86426a780..b745a43af 100644 --- a/comfy_extras/nodes_cond.py +++ b/comfy_extras/nodes_cond.py @@ -8,7 +8,7 @@ class CLIPTextEncodeControlnet(io.ComfyNode): def define_schema(cls) -> io.Schema: return io.Schema( node_id="CLIPTextEncodeControlnet", - category="_for_testing/conditioning", + category="experimental/conditioning", inputs=[ io.Clip.Input("clip"), io.Conditioning.Input("conditioning"), @@ -35,7 +35,7 @@ class T5TokenizerOptions(io.ComfyNode): def define_schema(cls) -> io.Schema: return io.Schema( node_id="T5TokenizerOptions", - category="_for_testing/conditioning", + category="experimental/conditioning", inputs=[ io.Clip.Input("clip"), io.Int.Input("min_padding", default=0, min=0, max=10000, step=1, advanced=True), diff --git a/comfy_extras/nodes_context_windows.py b/comfy_extras/nodes_context_windows.py index 0e43f2e44..f7ca833dc 100644 --- a/comfy_extras/nodes_context_windows.py +++ b/comfy_extras/nodes_context_windows.py @@ -10,7 +10,7 @@ class ContextWindowsManualNode(io.ComfyNode): return io.Schema( node_id="ContextWindowsManual", display_name="Context Windows (Manual)", - category="context", + category="model_patches", description="Manually set context windows.", inputs=[ io.Model.Input("model", tooltip="The model to apply context windows to during sampling."), @@ -29,6 +29,7 @@ class ContextWindowsManualNode(io.ComfyNode): io.Boolean.Input("freenoise", default=False, tooltip="Whether to apply FreeNoise noise shuffling, improves window blending."), io.String.Input("cond_retain_index_list", default="", tooltip="List of latent indices to retain in the conditioning tensors for each window, for example setting this to '0' will use the initial start image for each window."), io.Boolean.Input("split_conds_to_windows", default=False, tooltip="Whether to split multiple conditionings (created by ConditionCombine) to each window based on region index."), + io.Boolean.Input("causal_window_fix", default=True, tooltip="Whether to add a causal fix frame to non-0-indexed context windows."), ], outputs=[ io.Model.Output(tooltip="The model with context windows applied during sampling."), @@ -38,7 +39,7 @@ class ContextWindowsManualNode(io.ComfyNode): @classmethod def execute(cls, model: io.Model.Type, context_length: int, context_overlap: int, context_schedule: str, context_stride: int, closed_loop: bool, fuse_method: str, dim: int, freenoise: bool, - cond_retain_index_list: list[int]=[], split_conds_to_windows: bool=False) -> io.Model: + cond_retain_index_list: list[int]=[], split_conds_to_windows: bool=False, causal_window_fix: bool=True) -> io.Model: model = model.clone() model.model_options["context_handler"] = comfy.context_windows.IndexListContextHandler( context_schedule=comfy.context_windows.get_matching_context_schedule(context_schedule), @@ -50,7 +51,8 @@ class ContextWindowsManualNode(io.ComfyNode): dim=dim, freenoise=freenoise, cond_retain_index_list=cond_retain_index_list, - split_conds_to_windows=split_conds_to_windows + split_conds_to_windows=split_conds_to_windows, + causal_window_fix=causal_window_fix, ) # make memory usage calculation only take into account the context window latents comfy.context_windows.create_prepare_sampling_wrapper(model) diff --git a/comfy_extras/nodes_custom_sampler.py b/comfy_extras/nodes_custom_sampler.py index 1e957c09b..10b56b91c 100644 --- a/comfy_extras/nodes_custom_sampler.py +++ b/comfy_extras/nodes_custom_sampler.py @@ -17,7 +17,7 @@ class BasicScheduler(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="BasicScheduler", - category="sampling/custom_sampling/schedulers", + category="sampling/schedulers", inputs=[ io.Model.Input("model"), io.Combo.Input("scheduler", options=comfy.samplers.SCHEDULER_NAMES), @@ -47,7 +47,7 @@ class KarrasScheduler(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="KarrasScheduler", - category="sampling/custom_sampling/schedulers", + category="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 +69,7 @@ class ExponentialScheduler(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="ExponentialScheduler", - category="sampling/custom_sampling/schedulers", + category="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 +90,7 @@ class PolyexponentialScheduler(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="PolyexponentialScheduler", - category="sampling/custom_sampling/schedulers", + category="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 +112,7 @@ class LaplaceScheduler(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LaplaceScheduler", - category="sampling/custom_sampling/schedulers", + category="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 +136,7 @@ class SDTurboScheduler(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SDTurboScheduler", - category="sampling/custom_sampling/schedulers", + category="sampling/schedulers", inputs=[ io.Model.Input("model"), io.Int.Input("steps", default=1, min=1, max=10), @@ -160,7 +160,7 @@ class BetaSamplingScheduler(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="BetaSamplingScheduler", - category="sampling/custom_sampling/schedulers", + category="sampling/schedulers", inputs=[ io.Model.Input("model"), io.Int.Input("steps", default=20, min=1, max=10000), @@ -182,7 +182,7 @@ class VPScheduler(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="VPScheduler", - category="sampling/custom_sampling/schedulers", + category="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 +204,7 @@ class SplitSigmas(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SplitSigmas", - category="sampling/custom_sampling/sigmas", + category="sampling/sigmas", inputs=[ io.Sigmas.Input("sigmas"), io.Int.Input("step", default=0, min=0, max=10000), @@ -228,7 +228,7 @@ class SplitSigmasDenoise(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SplitSigmasDenoise", - category="sampling/custom_sampling/sigmas", + category="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 +254,7 @@ class FlipSigmas(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="FlipSigmas", - category="sampling/custom_sampling/sigmas", + category="sampling/sigmas", inputs=[io.Sigmas.Input("sigmas")], outputs=[io.Sigmas.Output()] ) @@ -276,7 +276,7 @@ class SetFirstSigma(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SetFirstSigma", - category="sampling/custom_sampling/sigmas", + category="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 +298,7 @@ class ExtendIntermediateSigmas(io.ComfyNode): return io.Schema( node_id="ExtendIntermediateSigmas", search_aliases=["interpolate sigmas"], - category="sampling/custom_sampling/sigmas", + category="sampling/sigmas", inputs=[ io.Sigmas.Input("sigmas"), io.Int.Input("steps", default=2, min=1, max=100), @@ -351,7 +351,7 @@ class SamplingPercentToSigma(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SamplingPercentToSigma", - category="sampling/custom_sampling/sigmas", + category="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 +379,7 @@ class KSamplerSelect(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="KSamplerSelect", - category="sampling/custom_sampling/samplers", + category="sampling/samplers", inputs=[io.Combo.Input("sampler_name", options=comfy.samplers.SAMPLER_NAMES)], outputs=[io.Sampler.Output()] ) @@ -396,7 +396,7 @@ class SamplerDPMPP_3M_SDE(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SamplerDPMPP_3M_SDE", - category="sampling/custom_sampling/samplers", + category="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 +421,7 @@ class SamplerDPMPP_2M_SDE(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SamplerDPMPP_2M_SDE", - category="sampling/custom_sampling/samplers", + category="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 +448,7 @@ class SamplerDPMPP_SDE(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SamplerDPMPP_SDE", - category="sampling/custom_sampling/samplers", + category="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 +474,7 @@ class SamplerDPMPP_2S_Ancestral(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SamplerDPMPP_2S_Ancestral", - category="sampling/custom_sampling/samplers", + category="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 +494,7 @@ class SamplerEulerAncestral(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SamplerEulerAncestral", - category="sampling/custom_sampling/samplers", + category="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 +515,7 @@ class SamplerEulerAncestralCFGPP(io.ComfyNode): return io.Schema( node_id="SamplerEulerAncestralCFGPP", display_name="SamplerEulerAncestralCFG++", - category="sampling/custom_sampling/samplers", + category="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 +537,7 @@ class SamplerLMS(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SamplerLMS", - category="sampling/custom_sampling/samplers", + category="sampling/samplers", inputs=[io.Int.Input("order", default=4, min=1, max=100, advanced=True)], outputs=[io.Sampler.Output()] ) @@ -554,7 +554,7 @@ class SamplerDPMAdaptative(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SamplerDPMAdaptative", - category="sampling/custom_sampling/samplers", + category="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 +585,7 @@ class SamplerER_SDE(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SamplerER_SDE", - category="sampling/custom_sampling/samplers", + category="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 +623,7 @@ class SamplerSASolver(io.ComfyNode): return io.Schema( node_id="SamplerSASolver", search_aliases=["sde"], - category="sampling/custom_sampling/samplers", + category="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 +668,7 @@ class SamplerSEEDS2(io.ComfyNode): return io.Schema( node_id="SamplerSEEDS2", search_aliases=["sde", "exp heun"], - category="sampling/custom_sampling/samplers", + category="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), @@ -750,7 +750,7 @@ class SamplerCustom(io.ComfyNode): latent = latent_image latent_image = latent["samples"] latent = latent.copy() - latent_image = comfy.sample.fix_empty_latent_channels(model, latent_image, latent.get("downscale_ratio_spacial", None)) + latent_image = comfy.sample.fix_empty_latent_channels(model, latent_image, latent.get("downscale_ratio_spacial", None), latent.get("downscale_ratio_temporal", None)) latent["samples"] = latent_image if not add_noise: @@ -770,6 +770,7 @@ class SamplerCustom(io.ComfyNode): out = latent.copy() out.pop("downscale_ratio_spacial", None) + out.pop("downscale_ratio_temporal", None) out["samples"] = samples if "x0" in x0_output: x0_out = model.model.process_latent_out(x0_output["x0"].cpu()) @@ -793,7 +794,8 @@ class BasicGuider(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="BasicGuider", - category="sampling/custom_sampling/guiders", + display_name="Basic Guider", + category="sampling/guiders", inputs=[ io.Model.Input("model"), io.Conditioning.Input("conditioning"), @@ -814,7 +816,8 @@ class CFGGuider(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="CFGGuider", - category="sampling/custom_sampling/guiders", + display_name="CFG Guider", + category="sampling/guiders", inputs=[ io.Model.Input("model"), io.Conditioning.Input("positive"), @@ -868,7 +871,8 @@ class DualCFGGuider(io.ComfyNode): return io.Schema( node_id="DualCFGGuider", search_aliases=["dual prompt guidance"], - category="sampling/custom_sampling/guiders", + display_name="Dual CFG Guider", + category="sampling/guiders", inputs=[ io.Model.Input("model"), io.Conditioning.Input("cond1"), @@ -896,7 +900,7 @@ class DisableNoise(io.ComfyNode): return io.Schema( node_id="DisableNoise", search_aliases=["zero noise"], - category="sampling/custom_sampling/noise", + category="sampling/noise", inputs=[], outputs=[io.Noise.Output()] ) @@ -913,7 +917,7 @@ class RandomNoise(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="RandomNoise", - category="sampling/custom_sampling/noise", + category="sampling/noise", inputs=[io.Int.Input("noise_seed", default=0, min=0, max=0xffffffffffffffff, control_after_generate=True)], outputs=[io.Noise.Output()] ) @@ -949,7 +953,7 @@ class SamplerCustomAdvanced(io.ComfyNode): latent = latent_image latent_image = latent["samples"] latent = latent.copy() - latent_image = comfy.sample.fix_empty_latent_channels(guider.model_patcher, latent_image, latent.get("downscale_ratio_spacial", None)) + latent_image = comfy.sample.fix_empty_latent_channels(guider.model_patcher, latent_image, latent.get("downscale_ratio_spacial", None), latent.get("downscale_ratio_temporal", None)) latent["samples"] = latent_image noise_mask = None @@ -965,6 +969,7 @@ class SamplerCustomAdvanced(io.ComfyNode): out = latent.copy() out.pop("downscale_ratio_spacial", None) + out.pop("downscale_ratio_temporal", None) out["samples"] = samples if "x0" in x0_output: x0_out = guider.model_patcher.model.process_latent_out(x0_output["x0"].cpu()) @@ -984,7 +989,7 @@ class AddNoise(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="AddNoise", - category="_for_testing/custom_sampling/noise", + category="experimental/custom_sampling/noise", is_experimental=True, inputs=[ io.Model.Input("model"), @@ -1034,7 +1039,7 @@ class ManualSigmas(io.ComfyNode): return io.Schema( node_id="ManualSigmas", search_aliases=["custom noise schedule", "define sigmas"], - category="_for_testing/custom_sampling", + category="experimental/custom_sampling", is_experimental=True, inputs=[ io.String.Input("sigmas", default="1, 0.5", multiline=False) diff --git a/comfy_extras/nodes_differential_diffusion.py b/comfy_extras/nodes_differential_diffusion.py index 34ffb9a89..4fa61ad0e 100644 --- a/comfy_extras/nodes_differential_diffusion.py +++ b/comfy_extras/nodes_differential_diffusion.py @@ -13,7 +13,7 @@ class DifferentialDiffusion(io.ComfyNode): node_id="DifferentialDiffusion", search_aliases=["inpaint gradient", "variable denoise strength"], display_name="Differential Diffusion", - category="_for_testing", + category="experimental", inputs=[ io.Model.Input("model"), io.Float.Input( diff --git a/comfy_extras/nodes_flux.py b/comfy_extras/nodes_flux.py index 3a23c7d04..997f21c09 100644 --- a/comfy_extras/nodes_flux.py +++ b/comfy_extras/nodes_flux.py @@ -102,7 +102,7 @@ class FluxDisableGuidance(io.ComfyNode): append = execute # TODO: remove -PREFERED_KONTEXT_RESOLUTIONS = [ +PREFERRED_KONTEXT_RESOLUTIONS = [ (672, 1568), (688, 1504), (720, 1456), @@ -143,7 +143,7 @@ class FluxKontextImageScale(io.ComfyNode): width = image.shape[2] height = image.shape[1] aspect_ratio = width / height - _, width, height = min((abs(aspect_ratio - w / h), w, h) for w, h in PREFERED_KONTEXT_RESOLUTIONS) + _, width, height = min((abs(aspect_ratio - w / h), w, h) for w, h in PREFERRED_KONTEXT_RESOLUTIONS) image = comfy.utils.common_upscale(image.movedim(-1, 1), width, height, "lanczos", "center").movedim(1, -1) return io.NodeOutput(image) @@ -215,7 +215,7 @@ class Flux2Scheduler(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="Flux2Scheduler", - category="sampling/custom_sampling/schedulers", + category="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), @@ -263,7 +263,7 @@ class FluxKVCache(io.ComfyNode): node_id="FluxKVCache", display_name="Flux KV Cache", description="Enables KV Cache optimization for reference images on Flux family models.", - category="", + category="experimental", is_experimental=True, inputs=[ io.Model.Input("model", tooltip="The model to use KV Cache on."), diff --git a/comfy_extras/nodes_frame_interpolation.py b/comfy_extras/nodes_frame_interpolation.py index a3b00d36e..9dd34cfb8 100644 --- a/comfy_extras/nodes_frame_interpolation.py +++ b/comfy_extras/nodes_frame_interpolation.py @@ -37,7 +37,7 @@ class FrameInterpolationModelLoader(io.ComfyNode): model = cls._detect_and_load(sd) dtype = torch.float16 if model_management.should_use_fp16(model_management.get_torch_device()) else torch.float32 model.eval().to(dtype) - patcher = comfy.model_patcher.ModelPatcher( + patcher = comfy.model_patcher.CoreModelPatcher( model, load_device=model_management.get_torch_device(), offload_device=model_management.unet_offload_device(), @@ -78,7 +78,7 @@ class FrameInterpolate(io.ComfyNode): return io.Schema( node_id="FrameInterpolate", display_name="Frame Interpolate", - category="image/video", + category="video", search_aliases=["rife", "film", "frame interpolation", "slow motion", "interpolate frames", "vfi"], inputs=[ FrameInterpolationModel.Input("interp_model"), @@ -98,16 +98,13 @@ class FrameInterpolate(io.ComfyNode): if num_frames < 2 or multiplier < 2: return io.NodeOutput(images) - model_management.load_model_gpu(interp_model) device = interp_model.load_device dtype = interp_model.model_dtype() inference_model = interp_model.model - - # Free VRAM for inference activations (model weights + ~20x a single frame's worth) - H, W = images.shape[1], images.shape[2] - activation_mem = H * W * 3 * images.element_size() * 20 - model_management.free_memory(activation_mem, device) + activation_mem = inference_model.memory_used_forward(images.shape, dtype) + model_management.load_models_gpu([interp_model], memory_required=activation_mem) align = getattr(inference_model, "pad_align", 1) + H, W = images.shape[1], images.shape[2] # Prepare a single padded frame on device for determining output dimensions def prepare_frame(idx): diff --git a/comfy_extras/nodes_fresca.py b/comfy_extras/nodes_fresca.py index eab4f303f..173f42154 100644 --- a/comfy_extras/nodes_fresca.py +++ b/comfy_extras/nodes_fresca.py @@ -60,7 +60,7 @@ class FreSca(io.ComfyNode): node_id="FreSca", search_aliases=["frequency guidance"], display_name="FreSca", - category="_for_testing", + category="experimental", description="Applies frequency-dependent scaling to the guidance", inputs=[ io.Model.Input("model"), diff --git a/comfy_extras/nodes_gits.py b/comfy_extras/nodes_gits.py index d48483862..0b7666524 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/custom_sampling/schedulers", + category="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 new file mode 100644 index 000000000..f393745f6 --- /dev/null +++ b/comfy_extras/nodes_hidream_o1.py @@ -0,0 +1,256 @@ +from typing_extensions import override + +import torch + +import comfy.model_management +import comfy.patcher_extension +import node_helpers +from comfy_api.latest import ComfyExtension, io + + +class EmptyHiDreamO1LatentImage(io.ComfyNode): + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="EmptyHiDreamO1LatentImage", + display_name="Empty HiDream-O1 Latent Image", + category="latent/image", + description=( + "Empty pixel-space latent for HiDream-O1-Image. The model was " + "trained at ~4 megapixels; lower resolutions go off-distribution " + "and quality regresses noticeably. Trained resolutions: " + "2048x2048, 2304x1728, 1728x2304, 2560x1440, 1440x2560, " + "2496x1664, 1664x2496, 3104x1312, 1312x3104, 2304x1792, 1792x2304." + ), + inputs=[ + io.Int.Input(id="width", default=2048, min=64, max=4096, step=32), + io.Int.Input(id="height", default=2048, min=64, max=4096, step=32), + io.Int.Input(id="batch_size", default=1, min=1, max=64), + ], + outputs=[io.Latent().Output()], + ) + + @classmethod + def execute(cls, *, width: int, height: int, batch_size: int = 1) -> io.NodeOutput: + latent = torch.zeros( + (batch_size, 3, height, width), + device=comfy.model_management.intermediate_device(), + ) + return io.NodeOutput({"samples": latent}) + + +class HiDreamO1ReferenceImages(io.ComfyNode): + """Attach reference images to both positive and negative conditioning.""" + + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="HiDreamO1ReferenceImages", + display_name="HiDream-O1 Reference Images", + category="conditioning/image", + description=( + "Attach 1-10 reference images to conditioning, one for edit instruction" + "or multiple for subject-driven personalization." + ), + inputs=[ + io.Conditioning.Input(id="positive"), + io.Conditioning.Input(id="negative"), + io.Autogrow.Input( + "images", + template=io.Autogrow.TemplateNames( + io.Image.Input("image"), + names=[f"image_{i}" for i in range(1, 11)], + min=1, + ), + tooltip=("Reference images. 1 image = instruction edit; 2-10 images = multi reference." + ), + ), + ], + outputs=[ + io.Conditioning.Output(display_name="positive"), + io.Conditioning.Output(display_name="negative"), + ], + ) + + @classmethod + def execute(cls, *, positive, negative, images: io.Autogrow.Type) -> io.NodeOutput: + refs = [images[f"image_{i}"] for i in range(1, 11) if f"image_{i}" in images] + positive = node_helpers.conditioning_set_values(positive, {"reference_latents": refs}, append=True) + negative = node_helpers.conditioning_set_values(negative, {"reference_latents": refs}, append=True) + return io.NodeOutput(positive, negative) + + +class HiDreamO1PatchSeamSmoothing(io.ComfyNode): + PATCH_SIZE = 32 + EDGE_FEATHER = 4 + + # Shift presets per (pattern, N). 8-pass = 4-quadrant + 4 quarter-patch offsets. + SHIFTS_BY_PATTERN = { + ("single_shift", 2): [(0, 0), (16, 16)], + ("single_shift", 4): [(0, 0), (16, 0), (0, 16), (16, 16)], + ("single_shift", 8): [(0, 0), (16, 0), (0, 16), (16, 16), + (8, 8), (24, 8), (8, 24), (24, 24)], + ("symmetric", 2): [(-8, -8), (8, 8)], + ("symmetric", 4): [(-8, -8), (8, -8), (-8, 8), (8, 8)], + ("symmetric", 8): [(-12, -12), (4, -12), (-12, 4), (4, 4), + (-4, -4), (12, -4), (-4, 12), (12, 12)], + } + RAMP_LEVELS = { + "2": [2], + "4": [4], + "ramp_2_4": [2, 4], + "ramp_2_4_8": [2, 4, 8], + } + + @staticmethod + def _hann_tile(cy: int, cx: int, size: int = 32) -> torch.Tensor: + """size x size Hann tile peaking at (cy, cx) within a patch.""" + half = size // 2 + yy = torch.arange(size).view(size, 1) + xx = torch.arange(size).view(1, size) + dy = ((yy - cy + half) % size) - half + dx = ((xx - cx + half) % size) - half + return 0.25 * (1 + torch.cos(torch.pi * dy / half)) * (1 + torch.cos(torch.pi * dx / half)) + + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="HiDreamO1PatchSeamSmoothing", + display_name="HiDream-O1 Patch Seam Smoothing", + category="advanced/model", + is_experimental=True, + description=( + "Average the model output across multiple shifted patch-grid " + "positions during the late portion of sampling. Cancels seams." + ), + inputs=[ + io.Model.Input(id="model"), + io.Float.Input(id="start_percent", default=0.8, min=0.0, max=1.0, step=0.01, + tooltip="Sampling progress (0=start, 1=end) at which the blend turns ON.", + ), + io.Float.Input(id="end_percent", default=1.0, min=0.0, max=1.0, step=0.01, + tooltip="Sampling progress at which the blend turns OFF.", + ), + io.Combo.Input( + id="pattern", + options=["single_shift", "symmetric"], + default="single_shift", + tooltip="Shift layout. single_shift: one pass at the natural patch grid + others offset. symmetric: all passes off-grid, shifts split around origin.", + ), + io.Combo.Input( + id="passes", + options=["2", "4", "ramp_2_4", "ramp_2_4_8"], + default="2", + tooltip="Number of passes per gated step. 2/4 = fixed. ramp_*: pass count increases as sampling approaches end (more smoothing where seams are most visible).", + ), + io.Combo.Input( + id="blend", + options=["average", "window", "median"], + default="average", + tooltip="average: equal-weight mean. window: Hann-windowed weighting favoring each pass away from its patch boundaries. median: per-pixel median, rejects wraparound-outlier passes.", + ), + io.Float.Input(id="strength", default=1.0, min=0.0, max=1.0, step=0.01, + tooltip="Interpolation between the natural-grid pred (0) and the averaged result (1).", + ), + ], + outputs=[io.Model.Output()], + ) + + @classmethod + def execute(cls, *, model, start_percent: float, end_percent: float, pattern: str, passes: str, blend: str, strength: float) -> io.NodeOutput: + if strength <= 0.0 or end_percent <= start_percent: + return io.NodeOutput(model) + + P = cls.PATCH_SIZE + half = P // 2 + shift_levels = [cls.SHIFTS_BY_PATTERN[(pattern, n)] for n in cls.RAMP_LEVELS[passes]] + + if blend == "window": + window_tile_levels = [ + torch.stack([cls._hann_tile((half - sy) % P, (half - sx) % P, P) for sy, sx in lst], dim=0) + for lst in shift_levels + ] + else: + window_tile_levels = [None] * len(shift_levels) + + m = model.clone() + model_sampling = m.get_model_object("model_sampling") + multiplier = float(model_sampling.multiplier) + start_t = float(model_sampling.percent_to_sigma(start_percent)) * multiplier + end_t = float(model_sampling.percent_to_sigma(end_percent)) * multiplier + + edge_ramp_cache: dict = {} + + def get_edge_ramp(H: int, W: int, device, dtype) -> torch.Tensor: + key = (H, W, device, dtype) + cached = edge_ramp_cache.get(key) + if cached is not None: + return cached + feather = cls.EDGE_FEATHER + ys = torch.minimum(torch.arange(H, device=device, dtype=torch.float32), + (H - 1) - torch.arange(H, device=device, dtype=torch.float32)) + xs = torch.minimum(torch.arange(W, device=device, dtype=torch.float32), + (W - 1) - torch.arange(W, device=device, dtype=torch.float32)) + y_mask = ((ys - P) / feather).clamp(0, 1) + x_mask = ((xs - P) / feather).clamp(0, 1) + ramp = (y_mask[:, None] * x_mask[None, :]).to(dtype) + edge_ramp_cache[key] = ramp + return ramp + + def smoothing_wrapper(executor, *args, **kwargs): + x = args[0] + t = float(args[1][0]) + pred = executor(*args, **kwargs) + if not (end_t <= t <= start_t): + return pred + # Pick shift-level by sigma phase across the gated range. + if len(shift_levels) == 1: + level_idx = 0 + else: + phase = (start_t - t) / max(start_t - end_t, 1e-8) + level_idx = min(int(phase * len(shift_levels)), len(shift_levels) - 1) + shifts = shift_levels[level_idx] + window_tiles = window_tile_levels[level_idx] + + preds = [] + for sy, sx in shifts: + if sy == 0 and sx == 0: + preds.append(pred) + continue + x_rolled = torch.roll(x, shifts=(sy, sx), dims=(-2, -1)) + pred_rolled = executor(x_rolled, *args[1:], **kwargs) + preds.append(torch.roll(pred_rolled, shifts=(-sy, -sx), dims=(-2, -1))) + stacked = torch.stack(preds, dim=0) # (N, B, C, H, W) + _, _, _, H, W = stacked.shape + if blend == "window": + N = stacked.shape[0] + tiles = window_tiles.to(device=stacked.device, dtype=stacked.dtype) + w = tiles.repeat(1, H // P, W // P)[:, :H, :W] + sum_w = w.sum(dim=0, keepdim=True) + w = torch.where(sum_w < 1e-3, torch.full_like(w, 1.0 / N), w / sum_w.clamp(min=1e-8)) + avg = (stacked * w[:, None, None, :, :]).sum(dim=0) + elif blend == "median": + avg = torch.median(stacked, dim=0).values + else: + avg = stacked.mean(dim=0) + + # Mask out the P-px wraparound contamination strip at each edge. + mask = get_edge_ramp(H, W, pred.device, pred.dtype) + return pred * (1.0 - mask * strength) + avg * (mask * strength) + + m.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, "hidream_o1_patch_seam_smoothing", smoothing_wrapper) + return io.NodeOutput(m) + + +class HiDreamO1Extension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + EmptyHiDreamO1LatentImage, + HiDreamO1ReferenceImages, + HiDreamO1PatchSeamSmoothing, + ] + + +async def comfy_entrypoint() -> HiDreamO1Extension: + return HiDreamO1Extension() diff --git a/comfy_extras/nodes_hunyuan.py b/comfy_extras/nodes_hunyuan.py index 4ea93a499..9e4873be5 100644 --- a/comfy_extras/nodes_hunyuan.py +++ b/comfy_extras/nodes_hunyuan.py @@ -131,6 +131,8 @@ class HunyuanVideo15SuperResolution(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="HunyuanVideo15SuperResolution", + display_name="Hunyuan Video 1.5 Super Resolution", + category="conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), io.Conditioning.Input("negative"), @@ -381,6 +383,8 @@ class HunyuanRefinerLatent(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="HunyuanRefinerLatent", + display_name="Hunyuan Latent Refiner", + category="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 fa55ead59..403eb855b 100644 --- a/comfy_extras/nodes_hunyuan3d.py +++ b/comfy_extras/nodes_hunyuan3d.py @@ -1,12 +1,7 @@ import torch -import os -import json -import struct -import numpy as np from comfy.ldm.modules.diffusionmodules.mmdit import get_1d_sincos_pos_embed_from_grid_torch -import folder_paths import comfy.model_management -from comfy.cli_args import args +from comfy_extras.nodes_save_3d import pack_variable_mesh_batch from typing_extensions import override from comfy_api.latest import ComfyExtension, IO, Types from comfy_api.latest._util import MESH, VOXEL # only for backward compatibility if someone import it from this file (will be removed later) # noqa @@ -40,7 +35,7 @@ class Hunyuan3Dv2Conditioning(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="Hunyuan3Dv2Conditioning", - category="conditioning/video_models", + category="conditioning/3d_models", inputs=[ IO.ClipVisionOutput.Input("clip_vision_output"), ], @@ -65,7 +60,7 @@ class Hunyuan3Dv2ConditioningMultiView(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="Hunyuan3Dv2ConditioningMultiView", - category="conditioning/video_models", + category="conditioning/3d_models", inputs=[ IO.ClipVisionOutput.Input("front", optional=True), IO.ClipVisionOutput.Input("left", optional=True), @@ -424,6 +419,7 @@ class VoxelToMeshBasic(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="VoxelToMeshBasic", + display_name="Voxel to Mesh (Basic)", category="3d", inputs=[ IO.Voxel.Input("voxel"), @@ -443,7 +439,9 @@ class VoxelToMeshBasic(IO.ComfyNode): vertices.append(v) faces.append(f) - return IO.NodeOutput(Types.MESH(torch.stack(vertices), torch.stack(faces))) + if vertices and all(v.shape == vertices[0].shape for v in vertices) and all(f.shape == faces[0].shape for f in faces): + return IO.NodeOutput(Types.MESH(torch.stack(vertices), torch.stack(faces))) + return IO.NodeOutput(pack_variable_mesh_batch(vertices, faces)) decode = execute # TODO: remove @@ -453,6 +451,7 @@ class VoxelToMesh(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="VoxelToMesh", + display_name="Voxel to Mesh", category="3d", inputs=[ IO.Voxel.Input("voxel"), @@ -479,206 +478,13 @@ class VoxelToMesh(IO.ComfyNode): vertices.append(v) faces.append(f) - return IO.NodeOutput(Types.MESH(torch.stack(vertices), torch.stack(faces))) + if vertices and all(v.shape == vertices[0].shape for v in vertices) and all(f.shape == faces[0].shape for f in faces): + return IO.NodeOutput(Types.MESH(torch.stack(vertices), torch.stack(faces))) + return IO.NodeOutput(pack_variable_mesh_batch(vertices, faces)) decode = execute # TODO: remove -def save_glb(vertices, faces, filepath, metadata=None): - """ - Save PyTorch tensor vertices and faces as a GLB file without external dependencies. - - Parameters: - vertices: torch.Tensor of shape (N, 3) - The vertex coordinates - faces: torch.Tensor of shape (M, 3) - The face indices (triangle faces) - filepath: str - Output filepath (should end with .glb) - """ - - # Convert tensors to numpy arrays - vertices_np = vertices.cpu().numpy().astype(np.float32) - faces_np = faces.cpu().numpy().astype(np.uint32) - - vertices_buffer = vertices_np.tobytes() - indices_buffer = faces_np.tobytes() - - def pad_to_4_bytes(buffer): - padding_length = (4 - (len(buffer) % 4)) % 4 - return buffer + b'\x00' * padding_length - - vertices_buffer_padded = pad_to_4_bytes(vertices_buffer) - indices_buffer_padded = pad_to_4_bytes(indices_buffer) - - buffer_data = vertices_buffer_padded + indices_buffer_padded - - vertices_byte_length = len(vertices_buffer) - vertices_byte_offset = 0 - indices_byte_length = len(indices_buffer) - indices_byte_offset = len(vertices_buffer_padded) - - gltf = { - "asset": {"version": "2.0", "generator": "ComfyUI"}, - "buffers": [ - { - "byteLength": len(buffer_data) - } - ], - "bufferViews": [ - { - "buffer": 0, - "byteOffset": vertices_byte_offset, - "byteLength": vertices_byte_length, - "target": 34962 # ARRAY_BUFFER - }, - { - "buffer": 0, - "byteOffset": indices_byte_offset, - "byteLength": indices_byte_length, - "target": 34963 # ELEMENT_ARRAY_BUFFER - } - ], - "accessors": [ - { - "bufferView": 0, - "byteOffset": 0, - "componentType": 5126, # FLOAT - "count": len(vertices_np), - "type": "VEC3", - "max": vertices_np.max(axis=0).tolist(), - "min": vertices_np.min(axis=0).tolist() - }, - { - "bufferView": 1, - "byteOffset": 0, - "componentType": 5125, # UNSIGNED_INT - "count": faces_np.size, - "type": "SCALAR" - } - ], - "meshes": [ - { - "primitives": [ - { - "attributes": { - "POSITION": 0 - }, - "indices": 1, - "mode": 4 # TRIANGLES - } - ] - } - ], - "nodes": [ - { - "mesh": 0 - } - ], - "scenes": [ - { - "nodes": [0] - } - ], - "scene": 0 - } - - if metadata is not None: - gltf["asset"]["extras"] = metadata - - # Convert the JSON to bytes - gltf_json = json.dumps(gltf).encode('utf8') - - def pad_json_to_4_bytes(buffer): - padding_length = (4 - (len(buffer) % 4)) % 4 - return buffer + b' ' * padding_length - - gltf_json_padded = pad_json_to_4_bytes(gltf_json) - - # Create the GLB header - # Magic glTF - glb_header = struct.pack('<4sII', b'glTF', 2, 12 + 8 + len(gltf_json_padded) + 8 + len(buffer_data)) - - # Create JSON chunk header (chunk type 0) - json_chunk_header = struct.pack('