diff --git a/.ci/update_windows/update.py b/.ci/update_windows/update.py index 51a263203..fe646a6ed 100755 --- a/.ci/update_windows/update.py +++ b/.ci/update_windows/update.py @@ -53,6 +53,16 @@ try: repo.stash(ident) except KeyError: print("nothing to stash") # noqa: T201 +except: + print("Could not stash, cleaning index and trying again.") # noqa: T201 + repo.state_cleanup() + repo.index.read_tree(repo.head.peel().tree) + repo.index.write() + try: + repo.stash(ident) + except KeyError: + print("nothing to stash.") # noqa: T201 + backup_branch_name = 'backup_branch_{}'.format(datetime.today().strftime('%Y-%m-%d_%H_%M_%S')) print("creating backup branch: {}".format(backup_branch_name)) # noqa: T201 try: @@ -66,8 +76,10 @@ if branch is None: try: ref = repo.lookup_reference('refs/remotes/origin/master') except: - print("pulling.") # noqa: T201 - pull(repo) + print("fetching.") # noqa: T201 + for remote in repo.remotes: + if remote.name == "origin": + remote.fetch() ref = repo.lookup_reference('refs/remotes/origin/master') repo.checkout(ref) branch = repo.lookup_branch('master') @@ -149,3 +161,4 @@ try: shutil.copy(stable_update_script, stable_update_script_to) except: pass + diff --git a/.ci/windows_amd_base_files/README_VERY_IMPORTANT.txt b/.ci/windows_amd_base_files/README_VERY_IMPORTANT.txt index 96a500be2..2cbb00d99 100755 --- a/.ci/windows_amd_base_files/README_VERY_IMPORTANT.txt +++ b/.ci/windows_amd_base_files/README_VERY_IMPORTANT.txt @@ -1,5 +1,5 @@ -As of the time of writing this you need this preview driver for best results: -https://www.amd.com/en/resources/support-articles/release-notes/RN-AMDGPU-WINDOWS-PYTORCH-PREVIEW.html +As of the time of writing this you need this driver for best results: +https://www.amd.com/en/resources/support-articles/release-notes/RN-AMDGPU-WINDOWS-PYTORCH-7-1-1.html HOW TO RUN: @@ -25,3 +25,4 @@ In the ComfyUI directory you will find a file: extra_model_paths.yaml.example Rename this file to: extra_model_paths.yaml and edit it with your favorite text editor. + diff --git a/.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat b/.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat index ed00583b6..4501ef9a1 100644 --- a/.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat +++ b/.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat @@ -1,3 +1,3 @@ ..\python_embeded\python.exe -s ..\ComfyUI\main.py --windows-standalone-build --disable-api-nodes -echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. +echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. If you get a c10.dll error you need to install vc redist that you can find: https://aka.ms/vc14/vc_redist.x64.exe pause diff --git a/.ci/windows_nvidia_base_files/run_nvidia_gpu.bat b/.ci/windows_nvidia_base_files/run_nvidia_gpu.bat index 4898a424f..6487ac7ce 100755 --- a/.ci/windows_nvidia_base_files/run_nvidia_gpu.bat +++ b/.ci/windows_nvidia_base_files/run_nvidia_gpu.bat @@ -1,3 +1,3 @@ .\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build -echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. +echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. If you get a c10.dll error you need to install vc redist that you can find: https://aka.ms/vc14/vc_redist.x64.exe pause diff --git a/.ci/windows_nvidia_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat b/.ci/windows_nvidia_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat index 32611e4af..01c5bb33b 100644 --- a/.ci/windows_nvidia_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat +++ b/.ci/windows_nvidia_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat @@ -1,3 +1,3 @@ .\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --fast fp16_accumulation -echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. +echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. If you get a c10.dll error you need to install vc redist that you can find: https://aka.ms/vc14/vc_redist.x64.exe pause diff --git a/.github/workflows/release-stable-all.yml b/.github/workflows/release-stable-all.yml index f7de3a7c3..d72ece2ce 100644 --- a/.github/workflows/release-stable-all.yml +++ b/.github/workflows/release-stable-all.yml @@ -14,7 +14,7 @@ jobs: contents: "write" packages: "write" pull-requests: "read" - name: "Release NVIDIA Default (cu129)" + name: "Release NVIDIA Default (cu130)" uses: ./.github/workflows/stable-release.yml with: git_tag: ${{ inputs.git_tag }} @@ -65,11 +65,11 @@ jobs: contents: "write" packages: "write" pull-requests: "read" - name: "Release AMD ROCm 6.4.4" + name: "Release AMD ROCm 7.1.1" uses: ./.github/workflows/stable-release.yml with: git_tag: ${{ inputs.git_tag }} - cache_tag: "rocm644" + cache_tag: "rocm711" python_minor: "12" python_patch: "10" rel_name: "amd" diff --git a/.github/workflows/stable-release.yml b/.github/workflows/stable-release.yml index 28484a9d1..f501b7b31 100644 --- a/.github/workflows/stable-release.yml +++ b/.github/workflows/stable-release.yml @@ -117,7 +117,7 @@ jobs: ./python.exe get-pip.py ./python.exe -s -m pip install ../${{ inputs.cache_tag }}_python_deps/* - grep comfyui ../ComfyUI/requirements.txt > ./requirements_comfyui.txt + grep comfy ../ComfyUI/requirements.txt > ./requirements_comfyui.txt ./python.exe -s -m pip install -r requirements_comfyui.txt rm requirements_comfyui.txt diff --git a/.github/workflows/test-build.yml b/.github/workflows/test-build.yml index 419873ad8..9160242e9 100644 --- a/.github/workflows/test-build.yml +++ b/.github/workflows/test-build.yml @@ -18,7 +18,7 @@ jobs: strategy: fail-fast: false matrix: - python-version: ["3.9", "3.10", "3.11", "3.12", "3.13"] + python-version: ["3.10", "3.11", "3.12", "3.13", "3.14"] steps: - uses: actions/checkout@v4 - name: Set up Python ${{ matrix.python-version }} diff --git a/.github/workflows/test-ci.yml b/.github/workflows/test-ci.yml index 1660ec8e3..63df2dc3a 100644 --- a/.github/workflows/test-ci.yml +++ b/.github/workflows/test-ci.yml @@ -5,6 +5,7 @@ on: push: branches: - master + - release/** paths-ignore: - 'app/**' - 'input/**' @@ -19,6 +20,7 @@ jobs: test-stable: strategy: fail-fast: false + max-parallel: 1 # This forces sequential execution matrix: # os: [macos, linux, windows] # os: [macos, linux] @@ -73,6 +75,7 @@ jobs: test-unix-nightly: strategy: fail-fast: false + max-parallel: 1 # This forces sequential execution matrix: # os: [macos, linux] os: [linux] diff --git a/.github/workflows/test-execution.yml b/.github/workflows/test-execution.yml index 00ef07ebf..9012633d8 100644 --- a/.github/workflows/test-execution.yml +++ b/.github/workflows/test-execution.yml @@ -2,9 +2,9 @@ name: Execution Tests on: push: - branches: [ main, master ] + branches: [ main, master, release/** ] pull_request: - branches: [ main, master ] + branches: [ main, master, release/** ] jobs: test: diff --git a/.github/workflows/test-launch.yml b/.github/workflows/test-launch.yml index 1735fd83b..ef0d3f123 100644 --- a/.github/workflows/test-launch.yml +++ b/.github/workflows/test-launch.yml @@ -2,9 +2,9 @@ name: Test server launches without errors on: push: - branches: [ main, master ] + branches: [ main, master, release/** ] pull_request: - branches: [ main, master ] + branches: [ main, master, release/** ] jobs: test: @@ -32,7 +32,9 @@ jobs: working-directory: ComfyUI - name: Check for unhandled exceptions in server log run: | - if grep -qE "Exception|Error" console_output.log; then + grep -v "Found comfy_kitchen backend triton: {'available': False, 'disabled': True, 'unavailable_reason': \"ImportError: No module named 'triton'\", 'capabilities': \[\]}" console_output.log | grep -v "Found comfy_kitchen backend triton: {'available': False, 'disabled': False, 'unavailable_reason': \"ImportError: No module named 'triton'\", 'capabilities': \[\]}" > console_output_filtered.log + cat console_output_filtered.log + if grep -qE "Exception|Error" console_output_filtered.log; then echo "Unhandled exception/error found in server log." exit 1 fi diff --git a/.github/workflows/test-unit.yml b/.github/workflows/test-unit.yml index 00caf5b8a..d05179cd3 100644 --- a/.github/workflows/test-unit.yml +++ b/.github/workflows/test-unit.yml @@ -2,9 +2,9 @@ name: Unit Tests on: push: - branches: [ main, master ] + branches: [ main, master, release/** ] pull_request: - branches: [ main, master ] + branches: [ main, master, release/** ] jobs: test: diff --git a/.github/workflows/update-version.yml b/.github/workflows/update-version.yml index d9d488974..c2343cc39 100644 --- a/.github/workflows/update-version.yml +++ b/.github/workflows/update-version.yml @@ -6,6 +6,7 @@ on: - "pyproject.toml" branches: - master + - release/** jobs: update-version: diff --git a/CODEOWNERS b/CODEOWNERS index b7aca9b26..4d5448636 100644 --- a/CODEOWNERS +++ b/CODEOWNERS @@ -1,3 +1,2 @@ # Admins -* @comfyanonymous -* @kosinkadink +* @comfyanonymous @kosinkadink @guill diff --git a/README.md b/README.md index 28beec427..6d09758c0 100644 --- a/README.md +++ b/README.md @@ -67,6 +67,8 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith - [HiDream](https://comfyanonymous.github.io/ComfyUI_examples/hidream/) - [Qwen Image](https://comfyanonymous.github.io/ComfyUI_examples/qwen_image/) - [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/) - 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) @@ -79,6 +81,7 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith - [Hunyuan Video](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_video/) - [Wan 2.1](https://comfyanonymous.github.io/ComfyUI_examples/wan/) - [Wan 2.2](https://comfyanonymous.github.io/ComfyUI_examples/wan22/) + - [Hunyuan Video 1.5](https://docs.comfy.org/tutorials/video/hunyuan/hunyuan-video-1-5) - Audio Models - [Stable Audio](https://comfyanonymous.github.io/ComfyUI_examples/audio/) - [ACE Step](https://comfyanonymous.github.io/ComfyUI_examples/audio/) @@ -116,6 +119,9 @@ ComfyUI follows a weekly release cycle targeting Monday but this regularly chang 1. **[ComfyUI Core](https://github.com/comfyanonymous/ComfyUI)** - Releases a new stable version (e.g., v0.7.0) roughly every week. + - 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. - Commits outside of the stable release tags may be very unstable and break many custom nodes. - Serves as the foundation for the desktop release @@ -206,6 +212,8 @@ Python 3.14 works but you may encounter issues with the torch compile node. The Python 3.13 is very well supported. If you have trouble with some custom node dependencies on 3.13 you can try 3.12 +torch 2.4 and above is supported but some features might only work on newer versions. We generally recommend using the latest major version of pytorch unless it is less than 2 weeks old. + ### Instructions: Git clone this repo. @@ -317,6 +325,32 @@ For models compatible with Iluvatar Extension for PyTorch. Here's a step-by-step 1. Install the Iluvatar Corex Toolkit by adhering to the platform-specific instructions on the [Installation](https://support.iluvatar.com/#/DocumentCentre?id=1&nameCenter=2&productId=520117912052801536) 2. Launch ComfyUI by running `python main.py` + +## [ComfyUI-Manager](https://github.com/Comfy-Org/ComfyUI-Manager/tree/manager-v4) + +**ComfyUI-Manager** is an extension that allows you to easily install, update, and manage custom nodes for ComfyUI. + +### Setup + +1. Install the manager dependencies: + ```bash + pip install -r manager_requirements.txt + ``` + +2. Enable the manager with the `--enable-manager` flag when running ComfyUI: + ```bash + python main.py --enable-manager + ``` + +### Command Line Options + +| Flag | Description | +|------|-------------| +| `--enable-manager` | Enable ComfyUI-Manager | +| `--enable-manager-legacy-ui` | Use the legacy manager UI instead of the new UI (requires `--enable-manager`) | +| `--disable-manager-ui` | Disable the manager UI and endpoints while keeping background features like security checks and scheduled installation completion (requires `--enable-manager`) | + + # Running ```python main.py``` diff --git a/api_server/routes/internal/internal_routes.py b/api_server/routes/internal/internal_routes.py index 613b0f7c7..b224306da 100644 --- a/api_server/routes/internal/internal_routes.py +++ b/api_server/routes/internal/internal_routes.py @@ -58,8 +58,13 @@ class InternalRoutes: return web.json_response({"error": "Invalid directory type"}, status=400) directory = get_directory_by_type(directory_type) + + def is_visible_file(entry: os.DirEntry) -> bool: + """Filter out hidden files (e.g., .DS_Store on macOS).""" + return entry.is_file() and not entry.name.startswith('.') + sorted_files = sorted( - (entry for entry in os.scandir(directory) if entry.is_file()), + (entry for entry in os.scandir(directory) if is_visible_file(entry)), key=lambda entry: -entry.stat().st_mtime ) return web.json_response([entry.name for entry in sorted_files], status=200) diff --git a/app/frontend_management.py b/app/frontend_management.py index cce0c117d..bdaa85812 100644 --- a/app/frontend_management.py +++ b/app/frontend_management.py @@ -10,7 +10,8 @@ import importlib from dataclasses import dataclass from functools import cached_property from pathlib import Path -from typing import TypedDict, Optional +from typing import Dict, TypedDict, Optional +from aiohttp import web from importlib.metadata import version import requests @@ -257,7 +258,54 @@ comfyui-frontend-package is not installed. sys.exit(-1) @classmethod - def templates_path(cls) -> str: + def template_asset_map(cls) -> Optional[Dict[str, str]]: + """Return a mapping of template asset names to their absolute paths.""" + try: + from comfyui_workflow_templates import ( + get_asset_path, + iter_templates, + ) + except ImportError: + logging.error( + f""" +********** ERROR *********** + +comfyui-workflow-templates is not installed. + +{frontend_install_warning_message()} + +********** ERROR *********** +""".strip() + ) + return None + + try: + template_entries = list(iter_templates()) + except Exception as exc: + logging.error(f"Failed to enumerate workflow templates: {exc}") + return None + + asset_map: Dict[str, str] = {} + try: + for entry in template_entries: + for asset in entry.assets: + asset_map[asset.filename] = get_asset_path( + entry.template_id, asset.filename + ) + except Exception as exc: + logging.error(f"Failed to resolve template asset paths: {exc}") + return None + + if not asset_map: + logging.error("No workflow template assets found. Did the packages install correctly?") + return None + + return asset_map + + + @classmethod + def legacy_templates_path(cls) -> Optional[str]: + """Return the legacy templates directory shipped inside the meta package.""" try: import comfyui_workflow_templates @@ -276,6 +324,7 @@ comfyui-workflow-templates is not installed. ********** ERROR *********** """.strip() ) + return None @classmethod def embedded_docs_path(cls) -> str: @@ -392,3 +441,17 @@ comfyui-workflow-templates is not installed. logging.info("Falling back to the default frontend.") check_frontend_version() return cls.default_frontend_path() + @classmethod + def template_asset_handler(cls): + assets = cls.template_asset_map() + if not assets: + return None + + async def serve_template(request: web.Request) -> web.StreamResponse: + rel_path = request.match_info.get("path", "") + target = assets.get(rel_path) + if target is None: + raise web.HTTPNotFound() + return web.FileResponse(target) + + return serve_template diff --git a/app/model_manager.py b/app/model_manager.py index ab36bca74..f124d1117 100644 --- a/app/model_manager.py +++ b/app/model_manager.py @@ -44,7 +44,7 @@ class ModelFileManager: @routes.get("/experiment/models/{folder}") async def get_all_models(request): folder = request.match_info.get("folder", None) - if not folder in folder_paths.folder_names_and_paths: + if folder not in folder_paths.folder_names_and_paths: return web.Response(status=404) files = self.get_model_file_list(folder) return web.json_response(files) @@ -55,7 +55,7 @@ class ModelFileManager: path_index = int(request.match_info.get("path_index", None)) filename = request.match_info.get("filename", None) - if not folder_name in folder_paths.folder_names_and_paths: + if folder_name not in folder_paths.folder_names_and_paths: return web.Response(status=404) folders = folder_paths.folder_names_and_paths[folder_name] diff --git a/app/user_manager.py b/app/user_manager.py index a2d376c0c..e2c00dab2 100644 --- a/app/user_manager.py +++ b/app/user_manager.py @@ -59,6 +59,9 @@ class UserManager(): user = "default" if args.multi_user and "comfy-user" in request.headers: user = request.headers["comfy-user"] + # Block System Users (use same error message to prevent probing) + if user.startswith(folder_paths.SYSTEM_USER_PREFIX): + raise KeyError("Unknown user: " + user) if user not in self.users: raise KeyError("Unknown user: " + user) @@ -66,15 +69,16 @@ class UserManager(): return user def get_request_user_filepath(self, request, file, type="userdata", create_dir=True): - user_directory = folder_paths.get_user_directory() - if type == "userdata": - root_dir = user_directory + root_dir = folder_paths.get_user_directory() else: raise KeyError("Unknown filepath type:" + type) user = self.get_request_user_id(request) - path = user_root = os.path.abspath(os.path.join(root_dir, user)) + user_root = folder_paths.get_public_user_directory(user) + if user_root is None: + return None + path = user_root # prevent leaving /{type} if os.path.commonpath((root_dir, user_root)) != root_dir: @@ -101,7 +105,11 @@ class UserManager(): name = name.strip() if not name: raise ValueError("username not provided") + if name.startswith(folder_paths.SYSTEM_USER_PREFIX): + raise ValueError("System User prefix not allowed") user_id = re.sub("[^a-zA-Z0-9-_]+", '-', name) + if user_id.startswith(folder_paths.SYSTEM_USER_PREFIX): + raise ValueError("System User prefix not allowed") user_id = user_id + "_" + str(uuid.uuid4()) self.users[user_id] = name @@ -132,7 +140,10 @@ class UserManager(): if username in self.users.values(): return web.json_response({"error": "Duplicate username."}, status=400) - user_id = self.add_user(username) + try: + user_id = self.add_user(username) + except ValueError as e: + return web.json_response({"error": str(e)}, status=400) return web.json_response(user_id) @routes.get("/userdata") @@ -424,7 +435,7 @@ class UserManager(): return source dest = get_user_data_path(request, check_exists=False, param="dest") - if not isinstance(source, str): + if not isinstance(dest, str): return dest overwrite = request.query.get("overwrite", 'true') != "false" diff --git a/comfy/cldm/cldm.py b/comfy/cldm/cldm.py index ec01665e2..c93c2e909 100644 --- a/comfy/cldm/cldm.py +++ b/comfy/cldm/cldm.py @@ -413,7 +413,8 @@ class ControlNet(nn.Module): out_middle = [] if self.num_classes is not None: - assert y.shape[0] == x.shape[0] + if y is None: + raise ValueError("y is None, did you try using a controlnet for SDXL on SD1?") emb = emb + self.label_emb(y) h = x diff --git a/comfy/cli_args.py b/comfy/cli_args.py index 2f30b72d2..dae9a895d 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -97,6 +97,13 @@ class LatentPreviewMethod(enum.Enum): Latent2RGB = "latent2rgb" TAESD = "taesd" + @classmethod + def from_string(cls, value: str): + for member in cls: + if member.value == value: + return member + return None + parser.add_argument("--preview-method", type=LatentPreviewMethod, default=LatentPreviewMethod.NoPreviews, help="Default preview method for sampler nodes.", action=EnumAction) parser.add_argument("--preview-size", type=int, default=512, help="Sets the maximum preview size for sampler nodes.") @@ -121,6 +128,12 @@ upcast.add_argument("--force-upcast-attention", action="store_true", help="Force upcast.add_argument("--dont-upcast-attention", action="store_true", help="Disable all upcasting of attention. Should be unnecessary except for debugging.") +parser.add_argument("--enable-manager", action="store_true", help="Enable the ComfyUI-Manager feature.") +manager_group = parser.add_mutually_exclusive_group() +manager_group.add_argument("--disable-manager-ui", action="store_true", help="Disables only the ComfyUI-Manager UI and endpoints. Scheduled installations and similar background tasks will still operate.") +manager_group.add_argument("--enable-manager-legacy-ui", action="store_true", help="Enables the legacy UI of ComfyUI-Manager") + + 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.") @@ -131,7 +144,8 @@ vram_group.add_argument("--cpu", action="store_true", help="To use the CPU for e parser.add_argument("--reserve-vram", type=float, default=None, help="Set the amount of vram in GB you want to reserve for use by your OS/other software. By default some amount is reserved depending on your OS.") -parser.add_argument("--async-offload", action="store_true", help="Use async weight offloading.") +parser.add_argument("--async-offload", nargs='?', const=2, type=int, default=None, metavar="NUM_STREAMS", help="Use async weight offloading. An optional argument controls the amount of offload streams. Default is 2. Enabled by default on Nvidia.") +parser.add_argument("--disable-async-offload", action="store_true", help="Disable async weight offloading.") parser.add_argument("--force-non-blocking", action="store_true", help="Force ComfyUI to use non-blocking operations for all applicable tensors. This may improve performance on some non-Nvidia systems but can cause issues with some workflows.") @@ -160,13 +174,14 @@ parser.add_argument("--windows-standalone-build", action="store_true", help="Win parser.add_argument("--disable-metadata", action="store_true", help="Disable saving prompt metadata in files.") parser.add_argument("--disable-all-custom-nodes", action="store_true", help="Disable loading all custom nodes.") parser.add_argument("--whitelist-custom-nodes", type=str, nargs='+', default=[], help="Specify custom node folders to load even when --disable-all-custom-nodes is enabled.") -parser.add_argument("--disable-api-nodes", action="store_true", help="Disable loading all api nodes.") +parser.add_argument("--disable-api-nodes", action="store_true", help="Disable loading all api nodes. Also prevents the frontend from communicating with the internet.") parser.add_argument("--multi-user", action="store_true", help="Enables per-user storage.") parser.add_argument("--verbose", default='INFO', const='DEBUG', nargs="?", choices=['DEBUG', 'INFO', 'WARNING', 'ERROR', 'CRITICAL'], help='Set the logging level') parser.add_argument("--log-stdout", action="store_true", help="Send normal process output to stdout instead of stderr (default).") + # The default built-in provider hosted under web/ DEFAULT_VERSION_STRING = "comfyanonymous/ComfyUI@latest" diff --git a/comfy/clip_model.py b/comfy/clip_model.py index 7c0cadab5..e88872728 100644 --- a/comfy/clip_model.py +++ b/comfy/clip_model.py @@ -2,6 +2,25 @@ import torch from comfy.ldm.modules.attention import optimized_attention_for_device import comfy.ops +def clip_preprocess(image, size=224, mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711], crop=True): + image = image[:, :, :, :3] if image.shape[3] > 3 else image + mean = torch.tensor(mean, device=image.device, dtype=image.dtype) + std = torch.tensor(std, device=image.device, dtype=image.dtype) + image = image.movedim(-1, 1) + if not (image.shape[2] == size and image.shape[3] == size): + if crop: + scale = (size / min(image.shape[2], image.shape[3])) + scale_size = (round(scale * image.shape[2]), round(scale * image.shape[3])) + else: + scale_size = (size, size) + + image = torch.nn.functional.interpolate(image, size=scale_size, mode="bicubic", antialias=True) + h = (image.shape[2] - size)//2 + w = (image.shape[3] - size)//2 + image = image[:,:,h:h+size,w:w+size] + image = torch.clip((255. * image), 0, 255).round() / 255.0 + return (image - mean.view([3,1,1])) / std.view([3,1,1]) + class CLIPAttention(torch.nn.Module): def __init__(self, embed_dim, heads, dtype, device, operations): super().__init__() diff --git a/comfy/clip_vision.py b/comfy/clip_vision.py index 447b1ce4a..d5fc53497 100644 --- a/comfy/clip_vision.py +++ b/comfy/clip_vision.py @@ -1,6 +1,5 @@ from .utils import load_torch_file, transformers_convert, state_dict_prefix_replace import os -import torch import json import logging @@ -17,24 +16,7 @@ class Output: def __setitem__(self, key, item): setattr(self, key, item) -def clip_preprocess(image, size=224, mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711], crop=True): - image = image[:, :, :, :3] if image.shape[3] > 3 else image - mean = torch.tensor(mean, device=image.device, dtype=image.dtype) - std = torch.tensor(std, device=image.device, dtype=image.dtype) - image = image.movedim(-1, 1) - if not (image.shape[2] == size and image.shape[3] == size): - if crop: - scale = (size / min(image.shape[2], image.shape[3])) - scale_size = (round(scale * image.shape[2]), round(scale * image.shape[3])) - else: - scale_size = (size, size) - - image = torch.nn.functional.interpolate(image, size=scale_size, mode="bicubic", antialias=True) - h = (image.shape[2] - size)//2 - w = (image.shape[3] - size)//2 - image = image[:,:,h:h+size,w:w+size] - image = torch.clip((255. * image), 0, 255).round() / 255.0 - return (image - mean.view([3,1,1])) / std.view([3,1,1]) +clip_preprocess = comfy.clip_model.clip_preprocess # Prevent some stuff from breaking, TODO: remove eventually IMAGE_ENCODERS = { "clip_vision_model": comfy.clip_model.CLIPVisionModelProjection, @@ -73,7 +55,7 @@ class ClipVisionModel(): def encode_image(self, image, crop=True): comfy.model_management.load_model_gpu(self.patcher) - pixel_values = clip_preprocess(image.to(self.load_device), size=self.image_size, mean=self.image_mean, std=self.image_std, crop=crop).float() + 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=crop).float() out = self.model(pixel_values=pixel_values, intermediate_output='all' if self.return_all_hidden_states else -2) outputs = Output() diff --git a/comfy/context_windows.py b/comfy/context_windows.py index 041f380f9..2f82d51da 100644 --- a/comfy/context_windows.py +++ b/comfy/context_windows.py @@ -51,32 +51,43 @@ class ContextHandlerABC(ABC): class IndexListContextWindow(ContextWindowABC): - def __init__(self, index_list: list[int], dim: int=0): + def __init__(self, index_list: list[int], dim: int=0, total_frames: int=0): self.index_list = index_list self.context_length = len(index_list) self.dim = dim + self.total_frames = total_frames + self.center_ratio = (min(index_list) + max(index_list)) / (2 * total_frames) - def get_tensor(self, full: torch.Tensor, device=None, dim=None) -> torch.Tensor: + def get_tensor(self, full: torch.Tensor, device=None, dim=None, retain_index_list=[]) -> torch.Tensor: if dim is None: dim = self.dim if dim == 0 and full.shape[dim] == 1: return full - idx = [slice(None)] * dim + [self.index_list] - return full[idx].to(device) + idx = tuple([slice(None)] * dim + [self.index_list]) + window = full[idx] + if retain_index_list: + idx = tuple([slice(None)] * dim + [retain_index_list]) + window[idx] = full[idx] + return window.to(device) def add_window(self, full: torch.Tensor, to_add: torch.Tensor, dim=None) -> torch.Tensor: if dim is None: dim = self.dim - idx = [slice(None)] * dim + [self.index_list] + idx = tuple([slice(None)] * dim + [self.index_list]) full[idx] += to_add return full + def get_region_index(self, num_regions: int) -> int: + region_idx = int(self.center_ratio * num_regions) + return min(max(region_idx, 0), num_regions - 1) + class IndexListCallbacks: EVALUATE_CONTEXT_WINDOWS = "evaluate_context_windows" COMBINE_CONTEXT_WINDOW_RESULTS = "combine_context_window_results" EXECUTE_START = "execute_start" EXECUTE_CLEANUP = "execute_cleanup" + RESIZE_COND_ITEM = "resize_cond_item" def init_callbacks(self): return {} @@ -94,7 +105,8 @@ 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=False, dim=0): + 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): self.context_schedule = context_schedule self.fuse_method = fuse_method self.context_length = context_length @@ -103,13 +115,18 @@ class IndexListContextHandler(ContextHandlerABC): self.closed_loop = closed_loop self.dim = dim self._step = 0 + 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.callbacks = {} def should_use_context(self, model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]) -> bool: # for now, assume first dim is batch - should have stored on BaseModel in actual implementation if x_in.size(self.dim) > self.context_length: - logging.info(f"Using context windows {self.context_length} for {x_in.size(self.dim)} frames.") + logging.info(f"Using context windows {self.context_length} with overlap {self.context_overlap} for {x_in.size(self.dim)} frames.") + if self.cond_retain_index_list: + logging.info(f"Retaining original cond for indexes: {self.cond_retain_index_list}") return True return False @@ -123,6 +140,11 @@ class IndexListContextHandler(ContextHandlerABC): return None # reuse or resize cond items to match context requirements resized_cond = [] + # if multiple conds, split based on primary region + if self.split_conds_to_windows and len(cond_in) > 1: + region = window.get_region_index(len(cond_in)) + logging.info(f"Splitting conds to windows; using region {region} for window {window.index_list[0]}-{window.index_list[-1]} with center ratio {window.center_ratio:.3f}") + cond_in = [cond_in[region]] # cond object is a list containing a dict - outer list is irrelevant, so just loop through it for actual_cond in cond_in: resized_actual_cond = actual_cond.copy() @@ -145,13 +167,38 @@ class IndexListContextHandler(ContextHandlerABC): new_cond_item = cond_item.copy() # when in dictionary, look for tensors and CONDCrossAttn [comfy/conds.py] (has cond attr that is a tensor) for cond_key, cond_value in new_cond_item.items(): + # Allow callbacks to handle custom conditioning items + handled = False + for callback in comfy.patcher_extension.get_all_callbacks( + IndexListCallbacks.RESIZE_COND_ITEM, self.callbacks + ): + result = callback(cond_key, cond_value, window, x_in, device, new_cond_item) + if result is not None: + new_cond_item[cond_key] = result + handled = True + break + if handled: + continue if isinstance(cond_value, torch.Tensor): - if cond_value.ndim < self.dim and cond_value.size(0) == x_in.size(self.dim): + if (self.dim < cond_value.ndim and cond_value(self.dim) == x_in.size(self.dim)) or \ + (cond_value.ndim < self.dim and cond_value.size(0) == x_in.size(self.dim)): new_cond_item[cond_key] = window.get_tensor(cond_value, device) + # Handle audio_embed (temporal dim is 1) + elif cond_key == "audio_embed" and hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor): + audio_cond = cond_value.cond + if audio_cond.ndim > 1 and audio_cond.size(1) == x_in.size(self.dim): + new_cond_item[cond_key] = cond_value._copy_with(window.get_tensor(audio_cond, device, dim=1)) + # Handle vace_context (temporal dim is 3) + elif cond_key == "vace_context" and hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor): + vace_cond = cond_value.cond + if vace_cond.ndim >= 4 and vace_cond.size(3) == x_in.size(self.dim): + sliced_vace = window.get_tensor(vace_cond, device, dim=3, retain_index_list=self.cond_retain_index_list) + new_cond_item[cond_key] = cond_value._copy_with(sliced_vace) # if has cond that is a Tensor, check if needs to be subset elif hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor): - if cond_value.cond.ndim < self.dim and cond_value.cond.size(0) == x_in.size(self.dim): - new_cond_item[cond_key] = cond_value._copy_with(window.get_tensor(cond_value.cond, device)) + if (self.dim < cond_value.cond.ndim and cond_value.cond.size(self.dim) == x_in.size(self.dim)) or \ + (cond_value.cond.ndim < self.dim and cond_value.cond.size(0) == x_in.size(self.dim)): + new_cond_item[cond_key] = cond_value._copy_with(window.get_tensor(cond_value.cond, device, retain_index_list=self.cond_retain_index_list)) elif cond_key == "num_video_frames": # for SVD new_cond_item[cond_key] = cond_value._copy_with(cond_value.cond) new_cond_item[cond_key].cond = window.context_length @@ -164,7 +211,7 @@ class IndexListContextHandler(ContextHandlerABC): return resized_cond def set_step(self, timestep: torch.Tensor, model_options: dict[str]): - mask = torch.isclose(model_options["transformer_options"]["sample_sigmas"], timestep, rtol=0.0001) + mask = torch.isclose(model_options["transformer_options"]["sample_sigmas"], timestep[0], rtol=0.0001) matches = torch.nonzero(mask) if torch.numel(matches) == 0: raise Exception("No sample_sigmas matched current timestep; something went wrong.") @@ -173,7 +220,7 @@ class IndexListContextHandler(ContextHandlerABC): def get_context_windows(self, model: BaseModel, x_in: torch.Tensor, model_options: dict[str]) -> list[IndexListContextWindow]: full_length = x_in.size(self.dim) # TODO: choose dim based on model context_windows = self.context_schedule.func(full_length, self, model_options) - context_windows = [IndexListContextWindow(window, dim=self.dim) for window in context_windows] + context_windows = [IndexListContextWindow(window, dim=self.dim, total_frames=full_length) for window in context_windows] return context_windows def execute(self, calc_cond_batch: Callable, model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]): @@ -250,8 +297,8 @@ class IndexListContextHandler(ContextHandlerABC): prev_weight = (bias_total / (bias_total + bias)) new_weight = (bias / (bias_total + bias)) # account for dims of tensors - idx_window = [slice(None)] * self.dim + [idx] - pos_window = [slice(None)] * self.dim + [pos] + idx_window = tuple([slice(None)] * self.dim + [idx]) + pos_window = tuple([slice(None)] * self.dim + [pos]) # apply new values conds_final[i][idx_window] = conds_final[i][idx_window] * prev_weight + sub_conds_out[i][pos_window] * new_weight biases_final[i][idx] = bias_total + bias @@ -287,6 +334,28 @@ def create_prepare_sampling_wrapper(model: ModelPatcher): ) +def _sampler_sample_wrapper(executor, guider, sigmas, extra_args, callback, noise, *args, **kwargs): + model_options = extra_args.get("model_options", None) + if model_options is None: + raise Exception("model_options not found in sampler_sample_wrapper; this should never happen, something went wrong.") + handler: IndexListContextHandler = model_options.get("context_handler", None) + if handler is None: + raise Exception("context_handler not found in sampler_sample_wrapper; this should never happen, something went wrong.") + if not handler.freenoise: + return executor(guider, sigmas, extra_args, callback, noise, *args, **kwargs) + noise = apply_freenoise(noise, handler.dim, handler.context_length, handler.context_overlap, extra_args["seed"]) + + return executor(guider, sigmas, extra_args, callback, noise, *args, **kwargs) + + +def create_sampler_sample_wrapper(model: ModelPatcher): + model.add_wrapper_with_key( + comfy.patcher_extension.WrappersMP.SAMPLER_SAMPLE, + "ContextWindows_sampler_sample", + _sampler_sample_wrapper + ) + + def match_weights_to_dim(weights: list[float], x_in: torch.Tensor, dim: int, device=None) -> torch.Tensor: total_dims = len(x_in.shape) weights_tensor = torch.Tensor(weights).to(device=device) @@ -538,3 +607,29 @@ def shift_window_to_end(window: list[int], num_frames: int): for i in range(len(window)): # 2) add end_delta to each val to slide windows to end window[i] = window[i] + end_delta + + +# https://github.com/Kosinkadink/ComfyUI-AnimateDiff-Evolved/blob/90fb1331201a4b29488089e4fbffc0d82cc6d0a9/animatediff/sample_settings.py#L465 +def apply_freenoise(noise: torch.Tensor, dim: int, context_length: int, context_overlap: int, seed: int): + logging.info("Context windows: Applying FreeNoise") + generator = torch.Generator(device='cpu').manual_seed(seed) + latent_video_length = noise.shape[dim] + delta = context_length - context_overlap + + for start_idx in range(0, latent_video_length - context_length, delta): + place_idx = start_idx + context_length + + actual_delta = min(delta, latent_video_length - place_idx) + if actual_delta <= 0: + break + + list_idx = torch.randperm(actual_delta, generator=generator, device='cpu') + start_idx + + source_slice = [slice(None)] * noise.ndim + source_slice[dim] = list_idx + target_slice = [slice(None)] * noise.ndim + target_slice[dim] = slice(place_idx, place_idx + actual_delta) + + noise[tuple(target_slice)] = noise[tuple(source_slice)] + + return noise diff --git a/comfy/hooks.py b/comfy/hooks.py index 9d0731072..1a76c7ba4 100644 --- a/comfy/hooks.py +++ b/comfy/hooks.py @@ -527,7 +527,8 @@ class HookKeyframeGroup: if self._current_keyframe.get_effective_guarantee_steps(max_sigma) > 0: break # if eval_c is outside the percent range, stop looking further - else: break + else: + break # update steps current context is used self._current_used_steps += 1 # update current timestep this was performed on diff --git a/comfy/k_diffusion/sampling.py b/comfy/k_diffusion/sampling.py index 0e2cda291..0949dee44 100644 --- a/comfy/k_diffusion/sampling.py +++ b/comfy/k_diffusion/sampling.py @@ -74,6 +74,9 @@ def get_ancestral_step(sigma_from, sigma_to, eta=1.): def default_noise_sampler(x, seed=None): if seed is not None: + if x.device == torch.device("cpu"): + seed += 1 + generator = torch.Generator(device=x.device) generator.manual_seed(seed) else: @@ -1557,10 +1560,13 @@ def sample_er_sde(model, x, sigmas, extra_args=None, callback=None, disable=None @torch.no_grad() -def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=0.5): +def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=0.5, solver_type="phi_1"): """SEEDS-2 - Stochastic Explicit Exponential Derivative-free Solvers (VP Data Prediction) stage 2. arXiv: https://arxiv.org/abs/2305.14267 (NeurIPS 2023) """ + if solver_type not in {"phi_1", "phi_2"}: + raise ValueError("solver_type must be 'phi_1' or 'phi_2'") + 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 @@ -1600,8 +1606,14 @@ def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=Non denoised_2 = model(x_2, sigma_s_1 * s_in, **extra_args) # Step 2 - denoised_d = torch.lerp(denoised, denoised_2, fac) - x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * ei_h_phi_1(-h_eta) * denoised_d + if solver_type == "phi_1": + denoised_d = torch.lerp(denoised, denoised_2, fac) + x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * ei_h_phi_1(-h_eta) * denoised_d + elif solver_type == "phi_2": + b2 = ei_h_phi_2(-h_eta) / r + b1 = ei_h_phi_1(-h_eta) - b2 + x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * (b1 * denoised + b2 * denoised_2) + if inject_noise: segment_factor = (r - 1) * h * eta sde_noise = sde_noise * segment_factor.exp() @@ -1609,6 +1621,17 @@ def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=Non x = x + sde_noise * sigmas[i + 1] * s_noise return x +@torch.no_grad() +def sample_exp_heun_2_x0(model, x, sigmas, extra_args=None, callback=None, disable=None, solver_type="phi_2"): + """Deterministic exponential Heun second order method in data prediction (x0) and logSNR time.""" + return sample_seeds_2(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=0.0, s_noise=0.0, noise_sampler=None, r=1.0, solver_type=solver_type) + + +@torch.no_grad() +def sample_exp_heun_2_x0_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type="phi_2"): + """Stochastic exponential Heun second order method in data prediction (x0) and logSNR time.""" + return sample_seeds_2(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, r=1.0, solver_type=solver_type) + @torch.no_grad() def sample_seeds_3(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r_1=1./3, r_2=2./3): @@ -1756,7 +1779,7 @@ def sample_sa_solver(model, x, sigmas, extra_args=None, callback=None, disable=F # Predictor if sigmas[i + 1] == 0: # Denoising step - x = denoised + x_pred = denoised else: tau_t = tau_func(sigmas[i + 1]) curr_lambdas = lambdas[i - predictor_order_used + 1:i + 1] @@ -1777,7 +1800,7 @@ def sample_sa_solver(model, x, sigmas, extra_args=None, callback=None, disable=F if tau_t > 0 and s_noise > 0: noise = noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * tau_t ** 2 * h).expm1().neg().sqrt() * s_noise x_pred = x_pred + noise - return x + return x_pred @torch.no_grad() diff --git a/comfy/latent_formats.py b/comfy/latent_formats.py index 77e642a94..cb4f52ce1 100644 --- a/comfy/latent_formats.py +++ b/comfy/latent_formats.py @@ -6,6 +6,7 @@ class LatentFormat: latent_dimensions = 2 latent_rgb_factors = None latent_rgb_factors_bias = None + latent_rgb_factors_reshape = None taesd_decoder_name = None def process_in(self, latent): @@ -178,6 +179,54 @@ class Flux(SD3): def process_out(self, latent): return (latent / self.scale_factor) + self.shift_factor +class Flux2(LatentFormat): + latent_channels = 128 + + def __init__(self): + self.latent_rgb_factors =[ + [0.0058, 0.0113, 0.0073], + [0.0495, 0.0443, 0.0836], + [-0.0099, 0.0096, 0.0644], + [0.2144, 0.3009, 0.3652], + [0.0166, -0.0039, -0.0054], + [0.0157, 0.0103, -0.0160], + [-0.0398, 0.0902, -0.0235], + [-0.0052, 0.0095, 0.0109], + [-0.3527, -0.2712, -0.1666], + [-0.0301, -0.0356, -0.0180], + [-0.0107, 0.0078, 0.0013], + [0.0746, 0.0090, -0.0941], + [0.0156, 0.0169, 0.0070], + [-0.0034, -0.0040, -0.0114], + [0.0032, 0.0181, 0.0080], + [-0.0939, -0.0008, 0.0186], + [0.0018, 0.0043, 0.0104], + [0.0284, 0.0056, -0.0127], + [-0.0024, -0.0022, -0.0030], + [0.1207, -0.0026, 0.0065], + [0.0128, 0.0101, 0.0142], + [0.0137, -0.0072, -0.0007], + [0.0095, 0.0092, -0.0059], + [0.0000, -0.0077, -0.0049], + [-0.0465, -0.0204, -0.0312], + [0.0095, 0.0012, -0.0066], + [0.0290, -0.0034, 0.0025], + [0.0220, 0.0169, -0.0048], + [-0.0332, -0.0457, -0.0468], + [-0.0085, 0.0389, 0.0609], + [-0.0076, 0.0003, -0.0043], + [-0.0111, -0.0460, -0.0614], + ] + + self.latent_rgb_factors_bias = [-0.0329, -0.0718, -0.0851] + self.latent_rgb_factors_reshape = lambda t: t.reshape(t.shape[0], 32, 2, 2, t.shape[-2], t.shape[-1]).permute(0, 1, 4, 2, 5, 3).reshape(t.shape[0], 32, t.shape[-2] * 2, t.shape[-1] * 2) + + def process_in(self, latent): + return latent + + def process_out(self, latent): + return latent + class Mochi(LatentFormat): latent_channels = 12 latent_dimensions = 3 @@ -358,6 +407,11 @@ class LTXV(LatentFormat): self.latent_rgb_factors_bias = [-0.0571, -0.1657, -0.2512] +class LTXAV(LTXV): + def __init__(self): + self.latent_rgb_factors = None + self.latent_rgb_factors_bias = None + class HunyuanVideo(LatentFormat): latent_channels = 16 latent_dimensions = 3 @@ -382,6 +436,7 @@ class HunyuanVideo(LatentFormat): ] latent_rgb_factors_bias = [ 0.0259, -0.0192, -0.0761] + taesd_decoder_name = "taehv" class Cosmos1CV8x8x8(LatentFormat): latent_channels = 16 @@ -445,7 +500,7 @@ class Wan21(LatentFormat): ]).view(1, self.latent_channels, 1, 1, 1) - self.taesd_decoder_name = None #TODO + self.taesd_decoder_name = "lighttaew2_1" def process_in(self, latent): latents_mean = self.latents_mean.to(latent.device, latent.dtype) @@ -516,6 +571,7 @@ class Wan22(Wan21): def __init__(self): self.scale_factor = 1.0 + self.taesd_decoder_name = "lighttaew2_2" self.latents_mean = torch.tensor([ -0.2289, -0.0052, -0.1323, -0.2339, -0.2799, 0.0174, 0.1838, 0.1557, -0.1382, 0.0542, 0.2813, 0.0891, 0.1570, -0.0098, 0.0375, -0.1825, @@ -611,6 +667,67 @@ class HunyuanImage21Refiner(LatentFormat): latent_dimensions = 3 scale_factor = 1.03682 + def process_in(self, latent): + out = latent * self.scale_factor + out = torch.cat((out[:, :, :1], out), dim=2) + out = out.permute(0, 2, 1, 3, 4) + b, f_times_2, c, h, w = out.shape + out = out.reshape(b, f_times_2 // 2, 2 * c, h, w) + out = out.permute(0, 2, 1, 3, 4).contiguous() + return out + + def process_out(self, latent): + z = latent / self.scale_factor + z = z.permute(0, 2, 1, 3, 4) + b, f, c, h, w = z.shape + z = z.reshape(b, f, 2, c // 2, h, w) + z = z.permute(0, 1, 2, 3, 4, 5).reshape(b, f * 2, c // 2, h, w) + z = z.permute(0, 2, 1, 3, 4) + z = z[:, :, 1:] + return z + +class HunyuanVideo15(LatentFormat): + latent_rgb_factors = [ + [ 0.0568, -0.0521, -0.0131], + [ 0.0014, 0.0735, 0.0326], + [ 0.0186, 0.0531, -0.0138], + [-0.0031, 0.0051, 0.0288], + [ 0.0110, 0.0556, 0.0432], + [-0.0041, -0.0023, -0.0485], + [ 0.0530, 0.0413, 0.0253], + [ 0.0283, 0.0251, 0.0339], + [ 0.0277, -0.0372, -0.0093], + [ 0.0393, 0.0944, 0.1131], + [ 0.0020, 0.0251, 0.0037], + [-0.0017, 0.0012, 0.0234], + [ 0.0468, 0.0436, 0.0203], + [ 0.0354, 0.0439, -0.0233], + [ 0.0090, 0.0123, 0.0346], + [ 0.0382, 0.0029, 0.0217], + [ 0.0261, -0.0300, 0.0030], + [-0.0088, -0.0220, -0.0283], + [-0.0272, -0.0121, -0.0363], + [-0.0664, -0.0622, 0.0144], + [ 0.0414, 0.0479, 0.0529], + [ 0.0355, 0.0612, -0.0247], + [ 0.0147, 0.0264, 0.0174], + [ 0.0438, 0.0038, 0.0542], + [ 0.0431, -0.0573, -0.0033], + [-0.0162, -0.0211, -0.0406], + [-0.0487, -0.0295, -0.0393], + [ 0.0005, -0.0109, 0.0253], + [ 0.0296, 0.0591, 0.0353], + [ 0.0119, 0.0181, -0.0306], + [-0.0085, -0.0362, 0.0229], + [ 0.0005, -0.0106, 0.0242] + ] + + latent_rgb_factors_bias = [ 0.0456, -0.0202, -0.0644] + latent_channels = 32 + latent_dimensions = 3 + scale_factor = 1.03682 + taesd_decoder_name = "lighttaehy1_5" + class Hunyuan3Dv2(LatentFormat): latent_channels = 64 latent_dimensions = 1 diff --git a/comfy/ldm/chroma/model.py b/comfy/ldm/chroma/model.py index 67bf70eb1..2e8ef0687 100644 --- a/comfy/ldm/chroma/model.py +++ b/comfy/ldm/chroma/model.py @@ -40,7 +40,8 @@ class ChromaParams: out_dim: int hidden_dim: int n_layers: int - + txt_ids_dims: list + vec_in_dim: int @@ -179,7 +180,10 @@ class Chroma(nn.Module): pe = self.pe_embedder(ids) blocks_replace = patches_replace.get("dit", {}) + transformer_options["total_blocks"] = len(self.double_blocks) + transformer_options["block_type"] = "double" for i, block in enumerate(self.double_blocks): + transformer_options["block_index"] = i if i not in self.skip_mmdit: double_mod = ( self.get_modulations(mod_vectors, "double_img", idx=i), @@ -222,7 +226,10 @@ class Chroma(nn.Module): img = torch.cat((txt, img), 1) + transformer_options["total_blocks"] = len(self.single_blocks) + transformer_options["block_type"] = "single" for i, block in enumerate(self.single_blocks): + transformer_options["block_index"] = i if i not in self.skip_dit: single_mod = self.get_modulations(mod_vectors, "single", idx=i) if ("single_block", i) in blocks_replace: diff --git a/comfy/ldm/chroma_radiance/model.py b/comfy/ldm/chroma_radiance/model.py index e643b4414..4fb56165e 100644 --- a/comfy/ldm/chroma_radiance/model.py +++ b/comfy/ldm/chroma_radiance/model.py @@ -37,7 +37,7 @@ class ChromaRadianceParams(ChromaParams): nerf_final_head_type: str # None means use the same dtype as the model. nerf_embedder_dtype: Optional[torch.dtype] - + use_x0: bool class ChromaRadiance(Chroma): """ @@ -159,6 +159,9 @@ class ChromaRadiance(Chroma): self.skip_dit = [] self.lite = False + if params.use_x0: + self.register_buffer("__x0__", torch.tensor([])) + @property def _nerf_final_layer(self) -> nn.Module: if self.params.nerf_final_head_type == "linear": @@ -267,7 +270,7 @@ class ChromaRadiance(Chroma): bad_keys = tuple( k for k, v in overrides.items() - if type(v) != type(getattr(params, k)) and (v is not None or k not in nullable_keys) + if not isinstance(v, type(getattr(params, k))) and (v is not None or k not in nullable_keys) ) if bad_keys: e = f"Invalid value(s) in transformer_options chroma_radiance_options: {', '.join(bad_keys)}" @@ -276,6 +279,12 @@ class ChromaRadiance(Chroma): params_dict |= overrides return params.__class__(**params_dict) + def _apply_x0_residual(self, predicted, noisy, timesteps): + + # non zero during training to prevent 0 div + eps = 0.0 + return (noisy - predicted) / (timesteps.view(-1,1,1,1) + eps) + def _forward( self, x: Tensor, @@ -316,4 +325,11 @@ class ChromaRadiance(Chroma): transformer_options, attn_mask=kwargs.get("attention_mask", None), ) - return self.forward_nerf(img, img_out, params)[:, :, :h, :w] + + out = self.forward_nerf(img, img_out, params)[:, :, :h, :w] + + # If x0 variant → v-pred, just return this instead + if hasattr(self, "__x0__"): + out = self._apply_x0_residual(out, img, timestep) + return out + diff --git a/comfy/ldm/flux/layers.py b/comfy/ldm/flux/layers.py index 23150a712..60f2bdae2 100644 --- a/comfy/ldm/flux/layers.py +++ b/comfy/ldm/flux/layers.py @@ -48,15 +48,44 @@ def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 10 return embedding class MLPEmbedder(nn.Module): - def __init__(self, in_dim: int, hidden_dim: int, dtype=None, device=None, operations=None): + def __init__(self, in_dim: int, hidden_dim: int, bias=True, dtype=None, device=None, operations=None): super().__init__() - self.in_layer = operations.Linear(in_dim, hidden_dim, bias=True, dtype=dtype, device=device) + self.in_layer = operations.Linear(in_dim, hidden_dim, bias=bias, dtype=dtype, device=device) self.silu = nn.SiLU() - self.out_layer = operations.Linear(hidden_dim, hidden_dim, bias=True, dtype=dtype, device=device) + self.out_layer = operations.Linear(hidden_dim, hidden_dim, bias=bias, dtype=dtype, device=device) def forward(self, x: Tensor) -> Tensor: return self.out_layer(self.silu(self.in_layer(x))) +class YakMLP(nn.Module): + def __init__(self, hidden_size: int, intermediate_size: int, dtype=None, device=None, operations=None): + super().__init__() + self.hidden_size = hidden_size + self.intermediate_size = intermediate_size + self.gate_proj = operations.Linear(self.hidden_size, self.intermediate_size, bias=True, dtype=dtype, device=device) + self.up_proj = operations.Linear(self.hidden_size, self.intermediate_size, bias=True, dtype=dtype, device=device) + self.down_proj = operations.Linear(self.intermediate_size, self.hidden_size, bias=True, dtype=dtype, device=device) + self.act_fn = nn.SiLU() + + def forward(self, x: Tensor) -> Tensor: + down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x)) + return down_proj + +def build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=False, yak_mlp=False, dtype=None, device=None, operations=None): + if yak_mlp: + return YakMLP(hidden_size, mlp_hidden_dim, dtype=dtype, device=device, operations=operations) + if mlp_silu_act: + return nn.Sequential( + operations.Linear(hidden_size, mlp_hidden_dim * 2, bias=False, dtype=dtype, device=device), + SiLUActivation(), + operations.Linear(mlp_hidden_dim, hidden_size, bias=False, dtype=dtype, device=device), + ) + else: + return nn.Sequential( + operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device), + nn.GELU(approximate="tanh"), + operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device), + ) class RMSNorm(torch.nn.Module): def __init__(self, dim: int, dtype=None, device=None, operations=None): @@ -80,14 +109,14 @@ class QKNorm(torch.nn.Module): class SelfAttention(nn.Module): - def __init__(self, dim: int, num_heads: int = 8, qkv_bias: bool = False, dtype=None, device=None, operations=None): + def __init__(self, dim: int, num_heads: int = 8, qkv_bias: bool = False, proj_bias: bool = True, dtype=None, device=None, operations=None): super().__init__() self.num_heads = num_heads head_dim = dim // num_heads self.qkv = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device) self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations) - self.proj = operations.Linear(dim, dim, dtype=dtype, device=device) + self.proj = operations.Linear(dim, dim, bias=proj_bias, dtype=dtype, device=device) @dataclass @@ -98,11 +127,11 @@ class ModulationOut: class Modulation(nn.Module): - def __init__(self, dim: int, double: bool, dtype=None, device=None, operations=None): + def __init__(self, dim: int, double: bool, bias=True, dtype=None, device=None, operations=None): super().__init__() self.is_double = double self.multiplier = 6 if double else 3 - self.lin = operations.Linear(dim, self.multiplier * dim, bias=True, dtype=dtype, device=device) + self.lin = operations.Linear(dim, self.multiplier * dim, bias=bias, dtype=dtype, device=device) def forward(self, vec: Tensor) -> tuple: if vec.ndim == 2: @@ -129,8 +158,18 @@ def apply_mod(tensor, m_mult, m_add=None, modulation_dims=None): return tensor +class SiLUActivation(nn.Module): + def __init__(self): + super().__init__() + self.gate_fn = nn.SiLU() + + def forward(self, x: Tensor) -> Tensor: + x1, x2 = x.chunk(2, dim=-1) + return self.gate_fn(x1) * x2 + + class DoubleStreamBlock(nn.Module): - def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, modulation=True, dtype=None, device=None, operations=None): + def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, modulation=True, mlp_silu_act=False, proj_bias=True, yak_mlp=False, dtype=None, device=None, operations=None): super().__init__() mlp_hidden_dim = int(hidden_size * mlp_ratio) @@ -142,27 +181,22 @@ class DoubleStreamBlock(nn.Module): self.img_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations) self.img_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations) + self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, proj_bias=proj_bias, dtype=dtype, device=device, operations=operations) self.img_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - self.img_mlp = nn.Sequential( - operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device), - nn.GELU(approximate="tanh"), - operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device), - ) + + self.img_mlp = build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=mlp_silu_act, yak_mlp=yak_mlp, dtype=dtype, device=device, operations=operations) if self.modulation: self.txt_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations) self.txt_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations) + self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, proj_bias=proj_bias, dtype=dtype, device=device, operations=operations) self.txt_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - self.txt_mlp = nn.Sequential( - operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device), - nn.GELU(approximate="tanh"), - operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device), - ) + + self.txt_mlp = build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=mlp_silu_act, yak_mlp=yak_mlp, dtype=dtype, device=device, operations=operations) + self.flipped_img_txt = flipped_img_txt def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims_img=None, modulation_dims_txt=None, transformer_options={}): @@ -246,6 +280,9 @@ class SingleStreamBlock(nn.Module): mlp_ratio: float = 4.0, qk_scale: float = None, modulation=True, + mlp_silu_act=False, + bias=True, + yak_mlp=False, dtype=None, device=None, operations=None @@ -257,17 +294,29 @@ class SingleStreamBlock(nn.Module): self.scale = qk_scale or head_dim**-0.5 self.mlp_hidden_dim = int(hidden_size * mlp_ratio) + + self.mlp_hidden_dim_first = self.mlp_hidden_dim + self.yak_mlp = yak_mlp + if mlp_silu_act: + self.mlp_hidden_dim_first = int(hidden_size * mlp_ratio * 2) + self.mlp_act = SiLUActivation() + else: + self.mlp_act = nn.GELU(approximate="tanh") + + if self.yak_mlp: + self.mlp_hidden_dim_first *= 2 + self.mlp_act = nn.SiLU() + # qkv and mlp_in - self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim, dtype=dtype, device=device) + self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim_first, bias=bias, dtype=dtype, device=device) # proj and mlp_out - self.linear2 = operations.Linear(hidden_size + self.mlp_hidden_dim, hidden_size, dtype=dtype, device=device) + self.linear2 = operations.Linear(hidden_size + self.mlp_hidden_dim, hidden_size, bias=bias, dtype=dtype, device=device) self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations) self.hidden_size = hidden_size self.pre_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - self.mlp_act = nn.GELU(approximate="tanh") if modulation: self.modulation = Modulation(hidden_size, double=False, dtype=dtype, device=device, operations=operations) else: @@ -279,7 +328,7 @@ class SingleStreamBlock(nn.Module): else: mod = vec - qkv, mlp = torch.split(self.linear1(apply_mod(self.pre_norm(x), (1 + mod.scale), mod.shift, modulation_dims)), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1) + qkv, mlp = torch.split(self.linear1(apply_mod(self.pre_norm(x), (1 + mod.scale), mod.shift, modulation_dims)), [3 * self.hidden_size, self.mlp_hidden_dim_first], dim=-1) q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) del qkv @@ -289,7 +338,10 @@ class SingleStreamBlock(nn.Module): attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options) del q, k, v # compute activation in mlp stream, cat again and run second linear layer - mlp = self.mlp_act(mlp) + if self.yak_mlp: + mlp = self.mlp_act(mlp[..., self.mlp_hidden_dim_first // 2:]) * mlp[..., :self.mlp_hidden_dim_first // 2] + else: + mlp = self.mlp_act(mlp) output = self.linear2(torch.cat((attn, mlp), 2)) x += apply_mod(output, mod.gate, None, modulation_dims) if x.dtype == torch.float16: @@ -298,11 +350,11 @@ class SingleStreamBlock(nn.Module): class LastLayer(nn.Module): - def __init__(self, hidden_size: int, patch_size: int, out_channels: int, dtype=None, device=None, operations=None): + def __init__(self, hidden_size: int, patch_size: int, out_channels: int, bias=True, dtype=None, device=None, operations=None): super().__init__() self.norm_final = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - self.linear = operations.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True, dtype=dtype, device=device) - self.adaLN_modulation = nn.Sequential(nn.SiLU(), operations.Linear(hidden_size, 2 * hidden_size, bias=True, dtype=dtype, device=device)) + self.linear = operations.Linear(hidden_size, patch_size * patch_size * out_channels, bias=bias, dtype=dtype, device=device) + self.adaLN_modulation = nn.Sequential(nn.SiLU(), operations.Linear(hidden_size, 2 * hidden_size, bias=bias, dtype=dtype, device=device)) def forward(self, x: Tensor, vec: Tensor, modulation_dims=None) -> Tensor: if vec.ndim == 2: diff --git a/comfy/ldm/flux/math.py b/comfy/ldm/flux/math.py index 158420290..f9597de5b 100644 --- a/comfy/ldm/flux/math.py +++ b/comfy/ldm/flux/math.py @@ -4,15 +4,16 @@ from torch import Tensor from comfy.ldm.modules.attention import optimized_attention import comfy.model_management +import logging def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transformer_options={}) -> Tensor: - q, k = apply_rope(q, k, pe) + if pe is not None: + q, k = apply_rope(q, k, pe) heads = q.shape[1] x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask, transformer_options=transformer_options) return x - def rope(pos: Tensor, dim: int, theta: int) -> Tensor: assert dim % 2 == 0 if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu() or comfy.model_management.is_directml_enabled(): @@ -27,13 +28,20 @@ def rope(pos: Tensor, dim: int, theta: int) -> Tensor: out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2) return out.to(dtype=torch.float32, device=pos.device) -def apply_rope1(x: Tensor, freqs_cis: Tensor): - x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2) - x_out = freqs_cis[..., 0] * x_[..., 0] - x_out.addcmul_(freqs_cis[..., 1], x_[..., 1]) +try: + import comfy.quant_ops + apply_rope = comfy.quant_ops.ck.apply_rope + apply_rope1 = comfy.quant_ops.ck.apply_rope1 +except: + logging.warning("No comfy kitchen, using old apply_rope functions.") + def apply_rope1(x: Tensor, freqs_cis: Tensor): + x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2) - return x_out.reshape(*x.shape).type_as(x) + x_out = freqs_cis[..., 0] * x_[..., 0] + x_out.addcmul_(freqs_cis[..., 1], x_[..., 1]) -def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor): - return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis) + return x_out.reshape(*x.shape).type_as(x) + + def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor): + return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis) diff --git a/comfy/ldm/flux/model.py b/comfy/ldm/flux/model.py index b9d36f202..f40c2a7a9 100644 --- a/comfy/ldm/flux/model.py +++ b/comfy/ldm/flux/model.py @@ -15,6 +15,8 @@ from .layers import ( MLPEmbedder, SingleStreamBlock, timestep_embedding, + Modulation, + RMSNorm ) @dataclass @@ -33,6 +35,14 @@ class FluxParams: patch_size: int qkv_bias: bool guidance_embed: bool + txt_ids_dims: list + global_modulation: bool = False + mlp_silu_act: bool = False + ops_bias: bool = True + default_ref_method: str = "offset" + ref_index_scale: float = 1.0 + yak_mlp: bool = False + txt_norm: bool = False class Flux(nn.Module): @@ -58,13 +68,22 @@ class Flux(nn.Module): self.hidden_size = params.hidden_size self.num_heads = params.num_heads self.pe_embedder = EmbedND(dim=pe_dim, theta=params.theta, axes_dim=params.axes_dim) - self.img_in = operations.Linear(self.in_channels, self.hidden_size, bias=True, dtype=dtype, device=device) - self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations) - self.vector_in = MLPEmbedder(params.vec_in_dim, self.hidden_size, dtype=dtype, device=device, operations=operations) + self.img_in = operations.Linear(self.in_channels, self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device) + self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device, operations=operations) + if params.vec_in_dim is not None: + self.vector_in = MLPEmbedder(params.vec_in_dim, self.hidden_size, dtype=dtype, device=device, operations=operations) + else: + self.vector_in = None + self.guidance_in = ( - MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations) if params.guidance_embed else nn.Identity() + MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device, operations=operations) if params.guidance_embed else nn.Identity() ) - self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, dtype=dtype, device=device) + self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device) + + if params.txt_norm: + self.txt_norm = RMSNorm(params.context_in_dim, dtype=dtype, device=device, operations=operations) + else: + self.txt_norm = None self.double_blocks = nn.ModuleList( [ @@ -73,6 +92,10 @@ class Flux(nn.Module): self.num_heads, mlp_ratio=params.mlp_ratio, qkv_bias=params.qkv_bias, + modulation=params.global_modulation is False, + mlp_silu_act=params.mlp_silu_act, + proj_bias=params.ops_bias, + yak_mlp=params.yak_mlp, dtype=dtype, device=device, operations=operations ) for _ in range(params.depth) @@ -81,13 +104,30 @@ class Flux(nn.Module): self.single_blocks = nn.ModuleList( [ - SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, dtype=dtype, device=device, operations=operations) + SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, modulation=params.global_modulation is False, mlp_silu_act=params.mlp_silu_act, bias=params.ops_bias, yak_mlp=params.yak_mlp, dtype=dtype, device=device, operations=operations) for _ in range(params.depth_single_blocks) ] ) if final_layer: - self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels, dtype=dtype, device=device, operations=operations) + self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels, bias=params.ops_bias, dtype=dtype, device=device, operations=operations) + + if params.global_modulation: + self.double_stream_modulation_img = Modulation( + self.hidden_size, + double=True, + bias=False, + dtype=dtype, device=device, operations=operations + ) + self.double_stream_modulation_txt = Modulation( + self.hidden_size, + double=True, + bias=False, + dtype=dtype, device=device, operations=operations + ) + self.single_stream_modulation = Modulation( + self.hidden_size, double=False, bias=False, dtype=dtype, device=device, operations=operations + ) def forward_orig( self, @@ -103,9 +143,6 @@ class Flux(nn.Module): attn_mask: Tensor = None, ) -> Tensor: - if y is None: - y = torch.zeros((img.shape[0], self.params.vec_in_dim), device=img.device, dtype=img.dtype) - patches = transformer_options.get("patches", {}) patches_replace = transformer_options.get("patches_replace", {}) if img.ndim != 3 or txt.ndim != 3: @@ -118,9 +155,19 @@ class Flux(nn.Module): if guidance is not None: vec = vec + self.guidance_in(timestep_embedding(guidance, 256).to(img.dtype)) - vec = vec + self.vector_in(y[:, :self.params.vec_in_dim]) + if self.vector_in is not None: + if y is None: + y = torch.zeros((img.shape[0], self.params.vec_in_dim), device=img.device, dtype=img.dtype) + vec = vec + self.vector_in(y[:, :self.params.vec_in_dim]) + + if self.txt_norm is not None: + txt = self.txt_norm(txt) txt = self.txt_in(txt) + vec_orig = vec + if self.params.global_modulation: + vec = (self.double_stream_modulation_img(vec_orig), self.double_stream_modulation_txt(vec_orig)) + if "post_input" in patches: for p in patches["post_input"]: out = p({"img": img, "txt": txt, "img_ids": img_ids, "txt_ids": txt_ids}) @@ -136,7 +183,10 @@ class Flux(nn.Module): pe = None blocks_replace = patches_replace.get("dit", {}) + transformer_options["total_blocks"] = len(self.double_blocks) + transformer_options["block_type"] = "double" for i, block in enumerate(self.double_blocks): + transformer_options["block_index"] = i if ("double_block", i) in blocks_replace: def block_wrap(args): out = {} @@ -177,7 +227,13 @@ class Flux(nn.Module): img = torch.cat((txt, img), 1) + if self.params.global_modulation: + vec, _ = self.single_stream_modulation(vec_orig) + + transformer_options["total_blocks"] = len(self.single_blocks) + transformer_options["block_type"] = "single" for i, block in enumerate(self.single_blocks): + transformer_options["block_index"] = i if ("single_block", i) in blocks_replace: def block_wrap(args): out = {} @@ -207,7 +263,7 @@ class Flux(nn.Module): img = img[:, txt.shape[1] :, ...] - img = self.final_layer(img, vec) # (N, T, patch_size ** 2 * out_channels) + img = self.final_layer(img, vec_orig) # (N, T, patch_size ** 2 * out_channels) return img def process_img(self, x, index=0, h_offset=0, w_offset=0, transformer_options={}): @@ -234,10 +290,10 @@ class Flux(nn.Module): h_offset += rope_options.get("shift_y", 0.0) w_offset += rope_options.get("shift_x", 0.0) - img_ids = torch.zeros((steps_h, steps_w, 3), device=x.device, dtype=x.dtype) + img_ids = torch.zeros((steps_h, steps_w, len(self.params.axes_dim)), device=x.device, dtype=torch.float32) img_ids[:, :, 0] = img_ids[:, :, 1] + index - img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=steps_h, device=x.device, dtype=x.dtype).unsqueeze(1) - img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=steps_w, device=x.device, dtype=x.dtype).unsqueeze(0) + img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=steps_h, device=x.device, dtype=torch.float32).unsqueeze(1) + img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=steps_w, device=x.device, dtype=torch.float32).unsqueeze(0) return img, repeat(img_ids, "h w c -> b (h w) c", b=bs) def forward(self, x, timestep, context, y=None, guidance=None, ref_latents=None, control=None, transformer_options={}, **kwargs): @@ -259,10 +315,10 @@ class Flux(nn.Module): h = 0 w = 0 index = 0 - ref_latents_method = kwargs.get("ref_latents_method", "offset") + ref_latents_method = kwargs.get("ref_latents_method", self.params.default_ref_method) for ref in ref_latents: if ref_latents_method == "index": - index += 1 + index += self.params.ref_index_scale h_offset = 0 w_offset = 0 elif ref_latents_method == "uxo": @@ -286,7 +342,12 @@ class Flux(nn.Module): img = torch.cat([img, kontext], dim=1) img_ids = torch.cat([img_ids, kontext_ids], dim=1) - txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype) + txt_ids = torch.zeros((bs, context.shape[1], len(self.params.axes_dim)), device=x.device, dtype=torch.float32) + + if len(self.params.txt_ids_dims) > 0: + for i in self.params.txt_ids_dims: + txt_ids[:, :, i] = torch.linspace(0, context.shape[1] - 1, steps=context.shape[1], device=x.device, dtype=torch.float32) + out = self.forward_orig(img, img_ids, context, txt_ids, timestep, y, guidance, control, transformer_options, attn_mask=kwargs.get("attention_mask", None)) out = out[:, :img_tokens] - return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=2, pw=2)[:,:,:h_orig,:w_orig] + return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=self.patch_size, pw=self.patch_size)[:,:,:h_orig,:w_orig] diff --git a/comfy/ldm/hunyuan_video/model.py b/comfy/ldm/hunyuan_video/model.py index 5132e6c07..55ab550f8 100644 --- a/comfy/ldm/hunyuan_video/model.py +++ b/comfy/ldm/hunyuan_video/model.py @@ -6,7 +6,6 @@ import comfy.ldm.flux.layers import comfy.ldm.modules.diffusionmodules.mmdit from comfy.ldm.modules.attention import optimized_attention - from dataclasses import dataclass from einops import repeat @@ -42,6 +41,9 @@ class HunyuanVideoParams: guidance_embed: bool byt5: bool meanflow: bool + use_cond_type_embedding: bool + vision_in_dim: int + meanflow_sum: bool class SelfAttentionRef(nn.Module): @@ -157,7 +159,10 @@ class TokenRefiner(nn.Module): t = self.t_embedder(timestep_embedding(timesteps, 256, time_factor=1.0).to(x.dtype)) # m = mask.float().unsqueeze(-1) # c = (x.float() * m).sum(dim=1) / m.sum(dim=1) #TODO: the following works when the x.shape is the same length as the tokens but might break otherwise - c = x.sum(dim=1) / x.shape[1] + if x.dtype == torch.float16: + c = x.float().sum(dim=1) / x.shape[1] + else: + c = x.sum(dim=1) / x.shape[1] c = t + self.c_embedder(c.to(x.dtype)) x = self.input_embedder(x) @@ -196,11 +201,15 @@ class HunyuanVideo(nn.Module): def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, operations=None, **kwargs): super().__init__() self.dtype = dtype + operation_settings = {"operations": operations, "device": device, "dtype": dtype} + params = HunyuanVideoParams(**kwargs) self.params = params self.patch_size = params.patch_size self.in_channels = params.in_channels self.out_channels = params.out_channels + self.use_cond_type_embedding = params.use_cond_type_embedding + self.vision_in_dim = params.vision_in_dim if params.hidden_size % params.num_heads != 0: raise ValueError( f"Hidden size {params.hidden_size} must be divisible by num_heads {params.num_heads}" @@ -266,6 +275,18 @@ class HunyuanVideo(nn.Module): if final_layer: self.final_layer = LastLayer(self.hidden_size, self.patch_size[-1], self.out_channels, dtype=dtype, device=device, operations=operations) + # HunyuanVideo 1.5 specific modules + if self.vision_in_dim is not None: + from comfy.ldm.wan.model import MLPProj + self.vision_in = MLPProj(in_dim=self.vision_in_dim, out_dim=self.hidden_size, operation_settings=operation_settings) + else: + self.vision_in = None + if self.use_cond_type_embedding: + # 0: text_encoder feature 1: byt5 feature 2: vision_encoder feature + self.cond_type_embedding = nn.Embedding(3, self.hidden_size) + else: + self.cond_type_embedding = None + def forward_orig( self, img: Tensor, @@ -276,6 +297,7 @@ class HunyuanVideo(nn.Module): timesteps: Tensor, y: Tensor = None, txt_byt5=None, + clip_fea=None, guidance: Tensor = None, guiding_frame_index=None, ref_latent=None, @@ -296,7 +318,7 @@ class HunyuanVideo(nn.Module): timesteps_r = transformer_options['sample_sigmas'][w[0] + 1] timesteps_r = timesteps_r.unsqueeze(0).to(device=timesteps.device, dtype=timesteps.dtype) vec_r = self.time_r_in(timestep_embedding(timesteps_r, 256, time_factor=1000.0).to(img.dtype)) - vec = (vec + vec_r) / 2 + vec = (vec + vec_r) if self.params.meanflow_sum else (vec + vec_r) / 2 if ref_latent is not None: ref_latent_ids = self.img_ids(ref_latent) @@ -331,12 +353,31 @@ class HunyuanVideo(nn.Module): txt = self.txt_in(txt, timesteps, txt_mask, transformer_options=transformer_options) + if self.cond_type_embedding is not None: + self.cond_type_embedding.to(txt.device) + cond_emb = self.cond_type_embedding(torch.zeros_like(txt[:, :, 0], device=txt.device, dtype=torch.long)) + txt = txt + cond_emb.to(txt.dtype) + if self.byt5_in is not None and txt_byt5 is not None: txt_byt5 = self.byt5_in(txt_byt5) + if self.cond_type_embedding is not None: + cond_emb = self.cond_type_embedding(torch.ones_like(txt_byt5[:, :, 0], device=txt_byt5.device, dtype=torch.long)) + txt_byt5 = txt_byt5 + cond_emb.to(txt_byt5.dtype) + txt = torch.cat((txt_byt5, txt), dim=1) # byt5 first for HunyuanVideo1.5 + else: + txt = torch.cat((txt, txt_byt5), dim=1) txt_byt5_ids = torch.zeros((txt_ids.shape[0], txt_byt5.shape[1], txt_ids.shape[-1]), device=txt_ids.device, dtype=txt_ids.dtype) - txt = torch.cat((txt, txt_byt5), dim=1) txt_ids = torch.cat((txt_ids, txt_byt5_ids), dim=1) + if clip_fea is not None: + txt_vision_states = self.vision_in(clip_fea) + if self.cond_type_embedding is not None: + cond_emb = self.cond_type_embedding(2 * torch.ones_like(txt_vision_states[:, :, 0], dtype=torch.long, device=txt_vision_states.device)) + txt_vision_states = txt_vision_states + cond_emb + txt = torch.cat((txt_vision_states.to(txt.dtype), txt), dim=1) + extra_txt_ids = torch.zeros((txt_ids.shape[0], txt_vision_states.shape[1], txt_ids.shape[-1]), device=txt_ids.device, dtype=txt_ids.dtype) + txt_ids = torch.cat((txt_ids, extra_txt_ids), dim=1) + ids = torch.cat((img_ids, txt_ids), dim=1) pe = self.pe_embedder(ids) @@ -349,7 +390,10 @@ class HunyuanVideo(nn.Module): attn_mask = None blocks_replace = patches_replace.get("dit", {}) + transformer_options["total_blocks"] = len(self.double_blocks) + transformer_options["block_type"] = "double" for i, block in enumerate(self.double_blocks): + transformer_options["block_index"] = i if ("double_block", i) in blocks_replace: def block_wrap(args): out = {} @@ -371,7 +415,10 @@ class HunyuanVideo(nn.Module): img = torch.cat((img, txt), 1) + transformer_options["total_blocks"] = len(self.single_blocks) + transformer_options["block_type"] = "single" for i, block in enumerate(self.single_blocks): + transformer_options["block_index"] = i if ("single_block", i) in blocks_replace: def block_wrap(args): out = {} @@ -430,14 +477,14 @@ class HunyuanVideo(nn.Module): img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0) return repeat(img_ids, "h w c -> b (h w) c", b=bs) - def forward(self, x, timestep, context, y=None, txt_byt5=None, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, disable_time_r=False, control=None, transformer_options={}, **kwargs): + def forward(self, x, timestep, context, y=None, txt_byt5=None, clip_fea=None, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, disable_time_r=False, control=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, timestep, context, y, txt_byt5, guidance, attention_mask, guiding_frame_index, ref_latent, disable_time_r, control, transformer_options, **kwargs) + ).execute(x, timestep, context, y, txt_byt5, clip_fea, guidance, attention_mask, guiding_frame_index, ref_latent, disable_time_r, control, transformer_options, **kwargs) - def _forward(self, x, timestep, context, y=None, txt_byt5=None, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, disable_time_r=False, control=None, transformer_options={}, **kwargs): + def _forward(self, x, timestep, context, y=None, txt_byt5=None, clip_fea=None, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, disable_time_r=False, control=None, transformer_options={}, **kwargs): bs = x.shape[0] if len(self.patch_size) == 3: img_ids = self.img_ids(x) @@ -445,5 +492,5 @@ class HunyuanVideo(nn.Module): else: img_ids = self.img_ids_2d(x) txt_ids = torch.zeros((bs, context.shape[1], 2), device=x.device, dtype=x.dtype) - out = self.forward_orig(x, img_ids, context, txt_ids, attention_mask, timestep, y, txt_byt5, guidance, guiding_frame_index, ref_latent, disable_time_r=disable_time_r, control=control, transformer_options=transformer_options) + out = self.forward_orig(x, img_ids, context, txt_ids, attention_mask, timestep, y, txt_byt5, clip_fea, guidance, guiding_frame_index, ref_latent, disable_time_r=disable_time_r, control=control, transformer_options=transformer_options) return out diff --git a/comfy/ldm/hunyuan_video/upsampler.py b/comfy/ldm/hunyuan_video/upsampler.py new file mode 100644 index 000000000..d9e76922f --- /dev/null +++ b/comfy/ldm/hunyuan_video/upsampler.py @@ -0,0 +1,122 @@ +import torch +import torch.nn as nn +import torch.nn.functional as F +from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, VideoConv3d +from comfy.ldm.hunyuan_video.vae_refiner import RMS_norm +import model_management +import model_patcher + +class SRResidualCausalBlock3D(nn.Module): + def __init__(self, channels: int): + super().__init__() + self.block = nn.Sequential( + VideoConv3d(channels, channels, kernel_size=3), + nn.SiLU(inplace=True), + VideoConv3d(channels, channels, kernel_size=3), + nn.SiLU(inplace=True), + VideoConv3d(channels, channels, kernel_size=3), + ) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + return x + self.block(x) + +class SRModel3DV2(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + hidden_channels: int = 64, + num_blocks: int = 6, + global_residual: bool = False, + ): + super().__init__() + self.in_conv = VideoConv3d(in_channels, hidden_channels, kernel_size=3) + self.blocks = nn.ModuleList([SRResidualCausalBlock3D(hidden_channels) for _ in range(num_blocks)]) + self.out_conv = VideoConv3d(hidden_channels, out_channels, kernel_size=3) + self.global_residual = bool(global_residual) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + residual = x + y = self.in_conv(x) + for blk in self.blocks: + y = blk(y) + y = self.out_conv(y) + if self.global_residual and (y.shape == residual.shape): + y = y + residual + return y + + +class Upsampler(nn.Module): + def __init__( + self, + z_channels: int, + out_channels: int, + block_out_channels: tuple[int, ...], + num_res_blocks: int = 2, + ): + super().__init__() + self.num_res_blocks = num_res_blocks + self.block_out_channels = block_out_channels + self.z_channels = z_channels + + ch = block_out_channels[0] + self.conv_in = VideoConv3d(z_channels, ch, kernel_size=3) + + self.up = nn.ModuleList() + + for i, tgt in enumerate(block_out_channels): + stage = nn.Module() + stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt, + out_channels=tgt, + temb_channels=0, + conv_shortcut=False, + conv_op=VideoConv3d, norm_op=RMS_norm) + for j in range(num_res_blocks + 1)]) + ch = tgt + self.up.append(stage) + + self.norm_out = RMS_norm(ch) + self.conv_out = VideoConv3d(ch, out_channels, kernel_size=3) + + def forward(self, z): + """ + Args: + z: (B, C, T, H, W) + target_shape: (H, W) + """ + # z to block_in + repeats = self.block_out_channels[0] // (self.z_channels) + x = self.conv_in(z) + z.repeat_interleave(repeats=repeats, dim=1) + + # upsampling + for stage in self.up: + for blk in stage.block: + x = blk(x) + + out = self.conv_out(F.silu(self.norm_out(x))) + return out + +UPSAMPLERS = { + "720p": SRModel3DV2, + "1080p": Upsampler, +} + +class HunyuanVideo15SRModel(): + def __init__(self, model_type, config): + self.load_device = model_management.vae_device() + offload_device = model_management.vae_offload_device() + self.dtype = model_management.vae_dtype(self.load_device) + self.model_class = UPSAMPLERS.get(model_type) + self.model = self.model_class(**config).eval() + + self.patcher = model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device) + + def load_sd(self, sd): + return self.model.load_state_dict(sd, strict=True) + + def get_sd(self): + return self.model.state_dict() + + def resample_latent(self, latent): + model_management.load_model_gpu(self.patcher) + return self.model(latent.to(self.load_device)) diff --git a/comfy/ldm/hunyuan_video/vae_refiner.py b/comfy/ldm/hunyuan_video/vae_refiner.py index c2a0b507d..ddf77cd0e 100644 --- a/comfy/ldm/hunyuan_video/vae_refiner.py +++ b/comfy/ldm/hunyuan_video/vae_refiner.py @@ -1,11 +1,13 @@ import torch import torch.nn as nn import torch.nn.functional as F -from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, AttnBlock, VideoConv3d, Normalize +from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, AttnBlock, CarriedConv3d, Normalize, conv_carry_causal_3d, torch_cat_if_needed import comfy.ops import comfy.ldm.models.autoencoder +import comfy.model_management ops = comfy.ops.disable_weight_init + class RMS_norm(nn.Module): def __init__(self, dim): super().__init__() @@ -14,10 +16,10 @@ class RMS_norm(nn.Module): self.gamma = nn.Parameter(torch.empty(shape)) def forward(self, x): - return F.normalize(x, dim=1) * self.scale * self.gamma + return F.normalize(x, dim=1) * self.scale * comfy.model_management.cast_to(self.gamma, dtype=x.dtype, device=x.device) class DnSmpl(nn.Module): - def __init__(self, ic, oc, tds=True, refiner_vae=True, op=VideoConv3d): + def __init__(self, ic, oc, tds, refiner_vae, op): super().__init__() fct = 2 * 2 * 2 if tds else 1 * 2 * 2 assert oc % fct == 0 @@ -27,11 +29,12 @@ class DnSmpl(nn.Module): self.tds = tds self.gs = fct * ic // oc - def forward(self, x): + def forward(self, x, conv_carry_in=None, conv_carry_out=None): r1 = 2 if self.tds else 1 - h = self.conv(x) + h = conv_carry_causal_3d([x], self.conv, conv_carry_in, conv_carry_out) + + if self.tds and self.refiner_vae and conv_carry_in is None: - if self.tds and self.refiner_vae: hf = h[:, :, :1, :, :] b, c, f, ht, wd = hf.shape hf = hf.reshape(b, c, f, ht // 2, 2, wd // 2, 2) @@ -39,14 +42,7 @@ class DnSmpl(nn.Module): hf = hf.reshape(b, 2 * 2 * c, f, ht // 2, wd // 2) hf = torch.cat([hf, hf], dim=1) - hn = h[:, :, 1:, :, :] - b, c, frms, ht, wd = hn.shape - nf = frms // r1 - hn = hn.reshape(b, c, nf, r1, ht // 2, 2, wd // 2, 2) - hn = hn.permute(0, 3, 5, 7, 1, 2, 4, 6) - hn = hn.reshape(b, r1 * 2 * 2 * c, nf, ht // 2, wd // 2) - - h = torch.cat([hf, hn], dim=2) + h = h[:, :, 1:, :, :] xf = x[:, :, :1, :, :] b, ci, f, ht, wd = xf.shape @@ -54,38 +50,36 @@ class DnSmpl(nn.Module): xf = xf.permute(0, 4, 6, 1, 2, 3, 5) xf = xf.reshape(b, 2 * 2 * ci, f, ht // 2, wd // 2) B, C, T, H, W = xf.shape - xf = xf.view(B, h.shape[1], self.gs // 2, T, H, W).mean(dim=2) + xf = xf.view(B, hf.shape[1], self.gs // 2, T, H, W).mean(dim=2) - xn = x[:, :, 1:, :, :] - b, ci, frms, ht, wd = xn.shape - nf = frms // r1 - xn = xn.reshape(b, ci, nf, r1, ht // 2, 2, wd // 2, 2) - xn = xn.permute(0, 3, 5, 7, 1, 2, 4, 6) - xn = xn.reshape(b, r1 * 2 * 2 * ci, nf, ht // 2, wd // 2) - B, C, T, H, W = xn.shape - xn = xn.view(B, h.shape[1], self.gs, T, H, W).mean(dim=2) - sc = torch.cat([xf, xn], dim=2) - else: - b, c, frms, ht, wd = h.shape + x = x[:, :, 1:, :, :] - nf = frms // r1 - h = h.reshape(b, c, nf, r1, ht // 2, 2, wd // 2, 2) - h = h.permute(0, 3, 5, 7, 1, 2, 4, 6) - h = h.reshape(b, r1 * 2 * 2 * c, nf, ht // 2, wd // 2) + if h.shape[2] == 0: + return hf + xf - b, ci, frms, ht, wd = x.shape - nf = frms // r1 - sc = x.reshape(b, ci, nf, r1, ht // 2, 2, wd // 2, 2) - sc = sc.permute(0, 3, 5, 7, 1, 2, 4, 6) - sc = sc.reshape(b, r1 * 2 * 2 * ci, nf, ht // 2, wd // 2) - B, C, T, H, W = sc.shape - sc = sc.view(B, h.shape[1], self.gs, T, H, W).mean(dim=2) + b, c, frms, ht, wd = h.shape + nf = frms // r1 + h = h.reshape(b, c, nf, r1, ht // 2, 2, wd // 2, 2) + h = h.permute(0, 3, 5, 7, 1, 2, 4, 6) + h = h.reshape(b, r1 * 2 * 2 * c, nf, ht // 2, wd // 2) - return h + sc + b, ci, frms, ht, wd = x.shape + nf = frms // r1 + x = x.reshape(b, ci, nf, r1, ht // 2, 2, wd // 2, 2) + x = x.permute(0, 3, 5, 7, 1, 2, 4, 6) + x = x.reshape(b, r1 * 2 * 2 * ci, nf, ht // 2, wd // 2) + B, C, T, H, W = x.shape + x = x.view(B, h.shape[1], self.gs, T, H, W).mean(dim=2) + + if self.tds and self.refiner_vae and conv_carry_in is None: + h = torch.cat([hf, h], dim=2) + x = torch.cat([xf, x], dim=2) + + return h + x class UpSmpl(nn.Module): - def __init__(self, ic, oc, tus=True, refiner_vae=True, op=VideoConv3d): + def __init__(self, ic, oc, tus, refiner_vae, op): super().__init__() fct = 2 * 2 * 2 if tus else 1 * 2 * 2 self.conv = op(ic, oc * fct, kernel_size=3, stride=1, padding=1) @@ -94,11 +88,11 @@ class UpSmpl(nn.Module): self.tus = tus self.rp = fct * oc // ic - def forward(self, x): + def forward(self, x, conv_carry_in=None, conv_carry_out=None): r1 = 2 if self.tus else 1 - h = self.conv(x) + h = conv_carry_causal_3d([x], self.conv, conv_carry_in, conv_carry_out) - if self.tus and self.refiner_vae: + if self.tus and self.refiner_vae and conv_carry_in is None: hf = h[:, :, :1, :, :] b, c, f, ht, wd = hf.shape nc = c // (2 * 2) @@ -107,14 +101,7 @@ class UpSmpl(nn.Module): hf = hf.reshape(b, nc, f, ht * 2, wd * 2) hf = hf[:, : hf.shape[1] // 2] - hn = h[:, :, 1:, :, :] - b, c, frms, ht, wd = hn.shape - nc = c // (r1 * 2 * 2) - hn = hn.reshape(b, r1, 2, 2, nc, frms, ht, wd) - hn = hn.permute(0, 4, 5, 1, 6, 2, 7, 3) - hn = hn.reshape(b, nc, frms * r1, ht * 2, wd * 2) - - h = torch.cat([hf, hn], dim=2) + h = h[:, :, 1:, :, :] xf = x[:, :, :1, :, :] b, ci, f, ht, wd = xf.shape @@ -125,29 +112,26 @@ class UpSmpl(nn.Module): xf = xf.permute(0, 3, 4, 5, 1, 6, 2) xf = xf.reshape(b, nc, f, ht * 2, wd * 2) - xn = x[:, :, 1:, :, :] - xn = xn.repeat_interleave(repeats=self.rp, dim=1) - b, c, frms, ht, wd = xn.shape - nc = c // (r1 * 2 * 2) - xn = xn.reshape(b, r1, 2, 2, nc, frms, ht, wd) - xn = xn.permute(0, 4, 5, 1, 6, 2, 7, 3) - xn = xn.reshape(b, nc, frms * r1, ht * 2, wd * 2) - sc = torch.cat([xf, xn], dim=2) - else: - b, c, frms, ht, wd = h.shape - nc = c // (r1 * 2 * 2) - h = h.reshape(b, r1, 2, 2, nc, frms, ht, wd) - h = h.permute(0, 4, 5, 1, 6, 2, 7, 3) - h = h.reshape(b, nc, frms * r1, ht * 2, wd * 2) + x = x[:, :, 1:, :, :] - sc = x.repeat_interleave(repeats=self.rp, dim=1) - b, c, frms, ht, wd = sc.shape - nc = c // (r1 * 2 * 2) - sc = sc.reshape(b, r1, 2, 2, nc, frms, ht, wd) - sc = sc.permute(0, 4, 5, 1, 6, 2, 7, 3) - sc = sc.reshape(b, nc, frms * r1, ht * 2, wd * 2) + b, c, frms, ht, wd = h.shape + nc = c // (r1 * 2 * 2) + h = h.reshape(b, r1, 2, 2, nc, frms, ht, wd) + h = h.permute(0, 4, 5, 1, 6, 2, 7, 3) + h = h.reshape(b, nc, frms * r1, ht * 2, wd * 2) - return h + sc + x = x.repeat_interleave(repeats=self.rp, dim=1) + b, c, frms, ht, wd = x.shape + nc = c // (r1 * 2 * 2) + x = x.reshape(b, r1, 2, 2, nc, frms, ht, wd) + x = x.permute(0, 4, 5, 1, 6, 2, 7, 3) + x = x.reshape(b, nc, frms * r1, ht * 2, wd * 2) + + if self.tus and self.refiner_vae and conv_carry_in is None: + h = torch.cat([hf, h], dim=2) + x = torch.cat([xf, x], dim=2) + + return h + x class Encoder(nn.Module): def __init__(self, in_channels, z_channels, block_out_channels, num_res_blocks, @@ -160,7 +144,7 @@ class Encoder(nn.Module): self.refiner_vae = refiner_vae if self.refiner_vae: - conv_op = VideoConv3d + conv_op = CarriedConv3d norm_op = RMS_norm else: conv_op = ops.Conv3d @@ -188,9 +172,9 @@ class Encoder(nn.Module): self.down.append(stage) self.mid = nn.Module() - self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op) + self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op) self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=norm_op) - self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op) + self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op) self.norm_out = norm_op(ch) self.conv_out = conv_op(ch, z_channels << 1, 3, 1, 1) @@ -201,31 +185,48 @@ class Encoder(nn.Module): if not self.refiner_vae and x.shape[2] == 1: x = x.expand(-1, -1, self.ffactor_temporal, -1, -1) - x = self.conv_in(x) + if self.refiner_vae: + xl = [x[:, :, :1, :, :]] + if x.shape[2] > self.ffactor_temporal: + xl += torch.split(x[:, :, 1: 1 + ((x.shape[2] - 1) // self.ffactor_temporal) * self.ffactor_temporal, :, :], self.ffactor_temporal * 2, dim=2) + x = xl + else: + x = [x] + out = [] - for stage in self.down: - for blk in stage.block: - x = blk(x) - if hasattr(stage, 'downsample'): - x = stage.downsample(x) + conv_carry_in = None - x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(x))) + for i, x1 in enumerate(x): + conv_carry_out = [] + if i == len(x) - 1: + conv_carry_out = None + + x1 = [ x1 ] + x1 = conv_carry_causal_3d(x1, self.conv_in, conv_carry_in, conv_carry_out) + + for stage in self.down: + for blk in stage.block: + x1 = blk(x1, None, conv_carry_in, conv_carry_out) + if hasattr(stage, 'downsample'): + x1 = stage.downsample(x1, conv_carry_in, conv_carry_out) + + out.append(x1) + conv_carry_in = conv_carry_out + + out = torch_cat_if_needed(out, dim=2) + + x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(out))) + del out b, c, t, h, w = x.shape grp = c // (self.z_channels << 1) skip = x.view(b, c // grp, grp, t, h, w).mean(2) - out = self.conv_out(F.silu(self.norm_out(x))) + skip + out = conv_carry_causal_3d([F.silu(self.norm_out(x))], self.conv_out) + skip if self.refiner_vae: out = self.regul(out)[0] - out = torch.cat((out[:, :, :1], out), dim=2) - out = out.permute(0, 2, 1, 3, 4) - b, f_times_2, c, h, w = out.shape - out = out.reshape(b, f_times_2 // 2, 2 * c, h, w) - out = out.permute(0, 2, 1, 3, 4).contiguous() - return out class Decoder(nn.Module): @@ -239,7 +240,7 @@ class Decoder(nn.Module): self.refiner_vae = refiner_vae if self.refiner_vae: - conv_op = VideoConv3d + conv_op = CarriedConv3d norm_op = RMS_norm else: conv_op = ops.Conv3d @@ -249,9 +250,9 @@ class Decoder(nn.Module): self.conv_in = conv_op(z_channels, ch, kernel_size=3, stride=1, padding=1) self.mid = nn.Module() - self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op) + self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op) self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=norm_op) - self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op) + self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op) self.up = nn.ModuleList() depth = (ffactor_spatial >> 1).bit_length() @@ -275,27 +276,38 @@ class Decoder(nn.Module): self.conv_out = conv_op(ch, out_channels, 3, stride=1, padding=1) def forward(self, z): - if self.refiner_vae: - z = z.permute(0, 2, 1, 3, 4) - b, f, c, h, w = z.shape - z = z.reshape(b, f, 2, c // 2, h, w) - z = z.permute(0, 1, 2, 3, 4, 5).reshape(b, f * 2, c // 2, h, w) - z = z.permute(0, 2, 1, 3, 4) - z = z[:, :, 1:] - - x = self.conv_in(z) + z.repeat_interleave(self.block_out_channels[0] // self.z_channels, 1) + x = conv_carry_causal_3d([z], self.conv_in) + z.repeat_interleave(self.block_out_channels[0] // self.z_channels, 1) x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(x))) - for stage in self.up: - for blk in stage.block: - x = blk(x) - if hasattr(stage, 'upsample'): - x = stage.upsample(x) + if self.refiner_vae: + x = torch.split(x, 2, dim=2) + else: + x = [ x ] + out = [] - out = self.conv_out(F.silu(self.norm_out(x))) + conv_carry_in = None + + for i, x1 in enumerate(x): + conv_carry_out = [] + if i == len(x) - 1: + conv_carry_out = None + for stage in self.up: + for blk in stage.block: + x1 = blk(x1, None, conv_carry_in, conv_carry_out) + if hasattr(stage, 'upsample'): + x1 = stage.upsample(x1, conv_carry_in, conv_carry_out) + + x1 = [ F.silu(self.norm_out(x1)) ] + x1 = conv_carry_causal_3d(x1, self.conv_out, conv_carry_in, conv_carry_out) + out.append(x1) + conv_carry_in = conv_carry_out + del x + + out = torch_cat_if_needed(out, dim=2) if not self.refiner_vae: if z.shape[-3] == 1: out = out[:, :, -1:] return out + diff --git a/comfy/ldm/kandinsky5/model.py b/comfy/ldm/kandinsky5/model.py new file mode 100644 index 000000000..1509de2f8 --- /dev/null +++ b/comfy/ldm/kandinsky5/model.py @@ -0,0 +1,413 @@ +import torch +from torch import nn +import math + +import comfy.ldm.common_dit +from comfy.ldm.modules.attention import optimized_attention +from comfy.ldm.flux.math import apply_rope1 +from comfy.ldm.flux.layers import EmbedND + +def attention(q, k, v, heads, transformer_options={}): + return optimized_attention( + q.transpose(1, 2), + k.transpose(1, 2), + v.transpose(1, 2), + heads=heads, + skip_reshape=True, + transformer_options=transformer_options + ) + +def apply_scale_shift_norm(norm, x, scale, shift): + return torch.addcmul(shift, norm(x), scale + 1.0) + +def apply_gate_sum(x, out, gate): + return torch.addcmul(x, gate, out) + +def get_shift_scale_gate(params): + shift, scale, gate = torch.chunk(params, 3, dim=-1) + return tuple(x.unsqueeze(1) for x in (shift, scale, gate)) + +def get_freqs(dim, max_period=10000.0): + return torch.exp(-math.log(max_period) * torch.arange(start=0, end=dim, dtype=torch.float32) / dim) + + +class TimeEmbeddings(nn.Module): + def __init__(self, model_dim, time_dim, max_period=10000.0, operation_settings=None): + super().__init__() + assert model_dim % 2 == 0 + self.model_dim = model_dim + self.max_period = max_period + self.register_buffer("freqs", get_freqs(model_dim // 2, max_period), persistent=False) + operations = operation_settings.get("operations") + self.in_layer = operations.Linear(model_dim, time_dim, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.activation = nn.SiLU() + self.out_layer = operations.Linear(time_dim, time_dim, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + + def forward(self, timestep, dtype): + args = torch.outer(timestep, self.freqs.to(device=timestep.device)) + time_embed = torch.cat([torch.cos(args), torch.sin(args)], dim=-1).to(dtype) + time_embed = self.out_layer(self.activation(self.in_layer(time_embed))) + return time_embed + + +class TextEmbeddings(nn.Module): + def __init__(self, text_dim, model_dim, operation_settings=None): + super().__init__() + operations = operation_settings.get("operations") + self.in_layer = operations.Linear(text_dim, model_dim, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.norm = operations.LayerNorm(model_dim, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + + def forward(self, text_embed): + text_embed = self.in_layer(text_embed) + return self.norm(text_embed).type_as(text_embed) + + +class VisualEmbeddings(nn.Module): + def __init__(self, visual_dim, model_dim, patch_size, operation_settings=None): + super().__init__() + self.patch_size = patch_size + operations = operation_settings.get("operations") + self.in_layer = operations.Linear(visual_dim, model_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + + def forward(self, x): + x = x.movedim(1, -1) # B C T H W -> B T H W C + B, T, H, W, dim = x.shape + pt, ph, pw = self.patch_size + + x = x.view( + B, + T // pt, pt, + H // ph, ph, + W // pw, pw, + dim, + ).permute(0, 1, 3, 5, 2, 4, 6, 7).flatten(4, 7) + + return self.in_layer(x) + + +class Modulation(nn.Module): + def __init__(self, time_dim, model_dim, num_params, operation_settings=None): + super().__init__() + self.activation = nn.SiLU() + self.out_layer = operation_settings.get("operations").Linear(time_dim, num_params * model_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + + def forward(self, x): + return self.out_layer(self.activation(x)) + + +class SelfAttention(nn.Module): + def __init__(self, num_channels, head_dim, operation_settings=None): + super().__init__() + assert num_channels % head_dim == 0 + self.num_heads = num_channels // head_dim + self.head_dim = head_dim + + operations = operation_settings.get("operations") + self.to_query = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.to_key = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.to_value = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.query_norm = operations.RMSNorm(head_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.key_norm = operations.RMSNorm(head_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + + self.out_layer = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.num_chunks = 2 + + def _compute_qk(self, x, freqs, proj_fn, norm_fn): + result = proj_fn(x).view(*x.shape[:-1], self.num_heads, -1) + return apply_rope1(norm_fn(result), freqs) + + def _forward(self, x, freqs, transformer_options={}): + q = self._compute_qk(x, freqs, self.to_query, self.query_norm) + k = self._compute_qk(x, freqs, self.to_key, self.key_norm) + v = self.to_value(x).view(*x.shape[:-1], self.num_heads, -1) + out = attention(q, k, v, self.num_heads, transformer_options=transformer_options) + return self.out_layer(out) + + def _forward_chunked(self, x, freqs, transformer_options={}): + def process_chunks(proj_fn, norm_fn): + x_chunks = torch.chunk(x, self.num_chunks, dim=1) + freqs_chunks = torch.chunk(freqs, self.num_chunks, dim=1) + chunks = [] + for x_chunk, freqs_chunk in zip(x_chunks, freqs_chunks): + chunks.append(self._compute_qk(x_chunk, freqs_chunk, proj_fn, norm_fn)) + return torch.cat(chunks, dim=1) + + q = process_chunks(self.to_query, self.query_norm) + k = process_chunks(self.to_key, self.key_norm) + v = self.to_value(x).view(*x.shape[:-1], self.num_heads, -1) + out = attention(q, k, v, self.num_heads, transformer_options=transformer_options) + return self.out_layer(out) + + def forward(self, x, freqs, transformer_options={}): + if x.shape[1] > 8192: + return self._forward_chunked(x, freqs, transformer_options=transformer_options) + else: + return self._forward(x, freqs, transformer_options=transformer_options) + + +class CrossAttention(SelfAttention): + def get_qkv(self, x, context): + q = self.to_query(x).view(*x.shape[:-1], self.num_heads, -1) + k = self.to_key(context).view(*context.shape[:-1], self.num_heads, -1) + v = self.to_value(context).view(*context.shape[:-1], self.num_heads, -1) + return q, k, v + + def forward(self, x, context, transformer_options={}): + q, k, v = self.get_qkv(x, context) + out = attention(self.query_norm(q), self.key_norm(k), v, self.num_heads, transformer_options=transformer_options) + return self.out_layer(out) + + +class FeedForward(nn.Module): + def __init__(self, dim, ff_dim, operation_settings=None): + super().__init__() + operations = operation_settings.get("operations") + self.in_layer = operations.Linear(dim, ff_dim, bias=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.activation = nn.GELU() + self.out_layer = operations.Linear(ff_dim, dim, bias=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.num_chunks = 4 + + def _forward(self, x): + return self.out_layer(self.activation(self.in_layer(x))) + + def _forward_chunked(self, x): + chunks = torch.chunk(x, self.num_chunks, dim=1) + output_chunks = [] + for chunk in chunks: + output_chunks.append(self._forward(chunk)) + return torch.cat(output_chunks, dim=1) + + def forward(self, x): + if x.shape[1] > 8192: + return self._forward_chunked(x) + else: + return self._forward(x) + + +class OutLayer(nn.Module): + def __init__(self, model_dim, time_dim, visual_dim, patch_size, operation_settings=None): + super().__init__() + self.patch_size = patch_size + self.modulation = Modulation(time_dim, model_dim, 2, operation_settings=operation_settings) + operations = operation_settings.get("operations") + self.norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.out_layer = operations.Linear(model_dim, math.prod(patch_size) * visual_dim, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + + def forward(self, visual_embed, time_embed): + B, T, H, W, _ = visual_embed.shape + shift, scale = torch.chunk(self.modulation(time_embed), 2, dim=-1) + scale = scale[:, None, None, None, :] + shift = shift[:, None, None, None, :] + visual_embed = apply_scale_shift_norm(self.norm, visual_embed, scale, shift) + x = self.out_layer(visual_embed) + + out_dim = x.shape[-1] // (self.patch_size[0] * self.patch_size[1] * self.patch_size[2]) + x = x.view( + B, T, H, W, + out_dim, + self.patch_size[0], self.patch_size[1], self.patch_size[2] + ) + return x.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(2, 3).flatten(3, 4).flatten(4, 5) + + +class TransformerEncoderBlock(nn.Module): + def __init__(self, model_dim, time_dim, ff_dim, head_dim, operation_settings=None): + super().__init__() + self.text_modulation = Modulation(time_dim, model_dim, 6, operation_settings=operation_settings) + operations = operation_settings.get("operations") + + self.self_attention_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.self_attention = SelfAttention(model_dim, head_dim, operation_settings=operation_settings) + + self.feed_forward_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.feed_forward = FeedForward(model_dim, ff_dim, operation_settings=operation_settings) + + def forward(self, x, time_embed, freqs, transformer_options={}): + self_attn_params, ff_params = torch.chunk(self.text_modulation(time_embed), 2, dim=-1) + shift, scale, gate = get_shift_scale_gate(self_attn_params) + out = apply_scale_shift_norm(self.self_attention_norm, x, scale, shift) + out = self.self_attention(out, freqs, transformer_options=transformer_options) + x = apply_gate_sum(x, out, gate) + + shift, scale, gate = get_shift_scale_gate(ff_params) + out = apply_scale_shift_norm(self.feed_forward_norm, x, scale, shift) + out = self.feed_forward(out) + x = apply_gate_sum(x, out, gate) + return x + + +class TransformerDecoderBlock(nn.Module): + def __init__(self, model_dim, time_dim, ff_dim, head_dim, operation_settings=None): + super().__init__() + self.visual_modulation = Modulation(time_dim, model_dim, 9, operation_settings=operation_settings) + + operations = operation_settings.get("operations") + self.self_attention_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.self_attention = SelfAttention(model_dim, head_dim, operation_settings=operation_settings) + + self.cross_attention_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.cross_attention = CrossAttention(model_dim, head_dim, operation_settings=operation_settings) + + self.feed_forward_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.feed_forward = FeedForward(model_dim, ff_dim, operation_settings=operation_settings) + + def forward(self, visual_embed, text_embed, time_embed, freqs, transformer_options={}): + self_attn_params, cross_attn_params, ff_params = torch.chunk(self.visual_modulation(time_embed), 3, dim=-1) + # self attention + shift, scale, gate = get_shift_scale_gate(self_attn_params) + visual_out = apply_scale_shift_norm(self.self_attention_norm, visual_embed, scale, shift) + visual_out = self.self_attention(visual_out, freqs, transformer_options=transformer_options) + visual_embed = apply_gate_sum(visual_embed, visual_out, gate) + # cross attention + shift, scale, gate = get_shift_scale_gate(cross_attn_params) + visual_out = apply_scale_shift_norm(self.cross_attention_norm, visual_embed, scale, shift) + visual_out = self.cross_attention(visual_out, text_embed, transformer_options=transformer_options) + visual_embed = apply_gate_sum(visual_embed, visual_out, gate) + # feed forward + shift, scale, gate = get_shift_scale_gate(ff_params) + visual_out = apply_scale_shift_norm(self.feed_forward_norm, visual_embed, scale, shift) + visual_out = self.feed_forward(visual_out) + visual_embed = apply_gate_sum(visual_embed, visual_out, gate) + return visual_embed + + +class Kandinsky5(nn.Module): + def __init__( + self, + in_visual_dim=16, out_visual_dim=16, in_text_dim=3584, in_text_dim2=768, time_dim=512, + model_dim=1792, ff_dim=7168, visual_embed_dim=132, patch_size=(1, 2, 2), num_text_blocks=2, num_visual_blocks=32, + axes_dims=(16, 24, 24), rope_scale_factor=(1.0, 2.0, 2.0), + dtype=None, device=None, operations=None, **kwargs + ): + super().__init__() + head_dim = sum(axes_dims) + self.rope_scale_factor = rope_scale_factor + self.in_visual_dim = in_visual_dim + self.model_dim = model_dim + self.patch_size = patch_size + self.visual_embed_dim = visual_embed_dim + self.dtype = dtype + self.device = device + operation_settings = {"operations": operations, "device": device, "dtype": dtype} + + self.time_embeddings = TimeEmbeddings(model_dim, time_dim, operation_settings=operation_settings) + self.text_embeddings = TextEmbeddings(in_text_dim, model_dim, operation_settings=operation_settings) + self.pooled_text_embeddings = TextEmbeddings(in_text_dim2, time_dim, operation_settings=operation_settings) + self.visual_embeddings = VisualEmbeddings(visual_embed_dim, model_dim, patch_size, operation_settings=operation_settings) + + self.text_transformer_blocks = nn.ModuleList( + [TransformerEncoderBlock(model_dim, time_dim, ff_dim, head_dim, operation_settings=operation_settings) for _ in range(num_text_blocks)] + ) + + self.visual_transformer_blocks = nn.ModuleList( + [TransformerDecoderBlock(model_dim, time_dim, ff_dim, head_dim, operation_settings=operation_settings) for _ in range(num_visual_blocks)] + ) + + self.out_layer = OutLayer(model_dim, time_dim, out_visual_dim, patch_size, operation_settings=operation_settings) + + self.rope_embedder_3d = EmbedND(dim=head_dim, theta=10000.0, axes_dim=axes_dims) + self.rope_embedder_1d = EmbedND(dim=head_dim, theta=10000.0, axes_dim=[head_dim]) + + def rope_encode_1d(self, seq_len, seq_start=0, steps=None, device=None, dtype=None, transformer_options={}): + steps = seq_len if steps is None else steps + seq_ids = torch.linspace(seq_start, seq_start + (seq_len - 1), steps=steps, device=device, dtype=dtype) + seq_ids = seq_ids.reshape(-1, 1).unsqueeze(0) # Shape: (1, steps, 1) + freqs = self.rope_embedder_1d(seq_ids).movedim(1, 2) + return freqs + + def rope_encode_3d(self, t, h, w, t_start=0, steps_t=None, steps_h=None, steps_w=None, 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) + else: + rope_scale_factor = self.rope_scale_factor + if self.model_dim == 4096: # pro video model uses different rope scaling at higher resolutions + if h * w >= 14080: + rope_scale_factor = (1.0, 3.16, 3.16) + + t_len = (t_len - 1.0) / rope_scale_factor[0] + 1.0 + h_len = (h_len - 1.0) / rope_scale_factor[1] + 1.0 + w_len = (w_len - 1.0) / rope_scale_factor[2] + 1.0 + + img_ids = torch.zeros((steps_t, steps_h, steps_w, 3), device=device, dtype=dtype) + 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_3d(img_ids).movedim(1, 2) + return freqs + + def forward_orig(self, x, timestep, context, y, freqs, freqs_text, transformer_options={}, **kwargs): + patches_replace = transformer_options.get("patches_replace", {}) + context = self.text_embeddings(context) + time_embed = self.time_embeddings(timestep, x.dtype) + self.pooled_text_embeddings(y) + + for block in self.text_transformer_blocks: + context = block(context, time_embed, freqs_text, transformer_options=transformer_options) + + visual_embed = self.visual_embeddings(x) + visual_shape = visual_embed.shape[:-1] + visual_embed = visual_embed.flatten(1, -2) + + blocks_replace = patches_replace.get("dit", {}) + transformer_options["total_blocks"] = len(self.visual_transformer_blocks) + transformer_options["block_type"] = "double" + for i, block in enumerate(self.visual_transformer_blocks): + transformer_options["block_index"] = i + if ("double_block", i) in blocks_replace: + def block_wrap(args): + return block(x=args["x"], context=args["context"], time_embed=args["time_embed"], freqs=args["freqs"], transformer_options=args.get("transformer_options")) + visual_embed = blocks_replace[("double_block", i)]({"x": visual_embed, "context": context, "time_embed": time_embed, "freqs": freqs, "transformer_options": transformer_options}, {"original_block": block_wrap})["x"] + else: + visual_embed = block(visual_embed, context, time_embed, freqs=freqs, transformer_options=transformer_options) + + visual_embed = visual_embed.reshape(*visual_shape, -1) + return self.out_layer(visual_embed, time_embed) + + def _forward(self, x, timestep, context, y, time_dim_replace=None, transformer_options={}, **kwargs): + original_dims = x.ndim + if original_dims == 4: + x = x.unsqueeze(2) + bs, c, t_len, h, w = x.shape + x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size) + + if time_dim_replace is not None: + time_dim_replace = comfy.ldm.common_dit.pad_to_patch_size(time_dim_replace, self.patch_size) + x[:, :time_dim_replace.shape[1], :time_dim_replace.shape[2]] = time_dim_replace + + freqs = self.rope_encode_3d(t_len, h, w, device=x.device, dtype=x.dtype, transformer_options=transformer_options) + freqs_text = self.rope_encode_1d(context.shape[1], device=x.device, dtype=x.dtype, transformer_options=transformer_options) + + out = self.forward_orig(x, timestep, context, y, freqs, freqs_text, transformer_options=transformer_options, **kwargs) + if original_dims == 4: + out = out.squeeze(2) + return out + + def forward(self, x, timestep, context, y, time_dim_replace=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, timestep, context, y, time_dim_replace=time_dim_replace, transformer_options=transformer_options, **kwargs) diff --git a/comfy/ldm/lightricks/av_model.py b/comfy/ldm/lightricks/av_model.py new file mode 100644 index 000000000..759535501 --- /dev/null +++ b/comfy/ldm/lightricks/av_model.py @@ -0,0 +1,837 @@ +from typing import Tuple +import torch +import torch.nn as nn +from comfy.ldm.lightricks.model import ( + CrossAttention, + FeedForward, + AdaLayerNormSingle, + PixArtAlphaTextProjection, + LTXVModel, +) +from comfy.ldm.lightricks.symmetric_patchifier import AudioPatchifier +import comfy.ldm.common_dit + +class BasicAVTransformerBlock(nn.Module): + def __init__( + self, + v_dim, + a_dim, + v_heads, + a_heads, + vd_head, + ad_head, + v_context_dim=None, + a_context_dim=None, + attn_precision=None, + dtype=None, + device=None, + operations=None, + ): + super().__init__() + + self.attn_precision = attn_precision + + self.attn1 = CrossAttention( + query_dim=v_dim, + heads=v_heads, + dim_head=vd_head, + context_dim=None, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) + self.audio_attn1 = CrossAttention( + query_dim=a_dim, + heads=a_heads, + dim_head=ad_head, + context_dim=None, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) + + self.attn2 = CrossAttention( + query_dim=v_dim, + context_dim=v_context_dim, + heads=v_heads, + dim_head=vd_head, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) + self.audio_attn2 = CrossAttention( + query_dim=a_dim, + context_dim=a_context_dim, + heads=a_heads, + dim_head=ad_head, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) + + # Q: Video, K,V: Audio + self.audio_to_video_attn = CrossAttention( + query_dim=v_dim, + context_dim=a_dim, + heads=a_heads, + dim_head=ad_head, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) + + # Q: Audio, K,V: Video + self.video_to_audio_attn = CrossAttention( + query_dim=a_dim, + context_dim=v_dim, + heads=a_heads, + dim_head=ad_head, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) + + self.ff = FeedForward( + v_dim, dim_out=v_dim, glu=True, dtype=dtype, device=device, operations=operations + ) + self.audio_ff = FeedForward( + a_dim, dim_out=a_dim, glu=True, dtype=dtype, device=device, operations=operations + ) + + self.scale_shift_table = nn.Parameter(torch.empty(6, v_dim, device=device, dtype=dtype)) + self.audio_scale_shift_table = nn.Parameter( + torch.empty(6, a_dim, device=device, dtype=dtype) + ) + + self.scale_shift_table_a2v_ca_audio = nn.Parameter( + torch.empty(5, a_dim, device=device, dtype=dtype) + ) + self.scale_shift_table_a2v_ca_video = nn.Parameter( + torch.empty(5, v_dim, device=device, dtype=dtype) + ) + + def get_ada_values( + self, scale_shift_table: torch.Tensor, batch_size: int, timestep: torch.Tensor, indices: slice = slice(None, None) + ): + num_ada_params = scale_shift_table.shape[0] + + ada_values = ( + scale_shift_table[indices].unsqueeze(0).unsqueeze(0).to(device=timestep.device, dtype=timestep.dtype) + + timestep.reshape(batch_size, timestep.shape[1], num_ada_params, -1)[:, :, indices, :] + ).unbind(dim=2) + return ada_values + + def get_av_ca_ada_values( + self, + scale_shift_table: torch.Tensor, + batch_size: int, + scale_shift_timestep: torch.Tensor, + gate_timestep: torch.Tensor, + num_scale_shift_values: int = 4, + ): + scale_shift_ada_values = self.get_ada_values( + scale_shift_table[:num_scale_shift_values, :], + batch_size, + scale_shift_timestep, + ) + gate_ada_values = self.get_ada_values( + scale_shift_table[num_scale_shift_values:, :], + batch_size, + gate_timestep, + ) + + scale_shift_chunks = [t.squeeze(2) for t in scale_shift_ada_values] + gate_ada_values = [t.squeeze(2) for t in gate_ada_values] + + return (*scale_shift_chunks, *gate_ada_values) + + def forward( + self, + x: Tuple[torch.Tensor, torch.Tensor], + v_context=None, + a_context=None, + attention_mask=None, + v_timestep=None, + a_timestep=None, + v_pe=None, + a_pe=None, + v_cross_pe=None, + a_cross_pe=None, + v_cross_scale_shift_timestep=None, + a_cross_scale_shift_timestep=None, + v_cross_gate_timestep=None, + a_cross_gate_timestep=None, + transformer_options=None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + run_vx = transformer_options.get("run_vx", True) + run_ax = transformer_options.get("run_ax", True) + + vx, ax = x + run_ax = run_ax and ax.numel() > 0 + run_a2v = run_vx and transformer_options.get("a2v_cross_attn", True) and ax.numel() > 0 + run_v2a = run_ax and transformer_options.get("v2a_cross_attn", True) + + if run_vx: + vshift_msa, vscale_msa, vgate_msa = ( + self.get_ada_values(self.scale_shift_table, vx.shape[0], v_timestep, slice(0, 3)) + ) + + norm_vx = comfy.ldm.common_dit.rms_norm(vx) * (1 + vscale_msa) + vshift_msa + vx += self.attn1(norm_vx, pe=v_pe, transformer_options=transformer_options) * vgate_msa + vx += self.attn2( + comfy.ldm.common_dit.rms_norm(vx), + context=v_context, + mask=attention_mask, + transformer_options=transformer_options, + ) + + del vshift_msa, vscale_msa, vgate_msa + + if run_ax: + ashift_msa, ascale_msa, agate_msa = ( + self.get_ada_values(self.audio_scale_shift_table, ax.shape[0], a_timestep, slice(0, 3)) + ) + + norm_ax = comfy.ldm.common_dit.rms_norm(ax) * (1 + ascale_msa) + ashift_msa + ax += ( + self.audio_attn1(norm_ax, pe=a_pe, transformer_options=transformer_options) + * agate_msa + ) + ax += self.audio_attn2( + comfy.ldm.common_dit.rms_norm(ax), + context=a_context, + mask=attention_mask, + transformer_options=transformer_options, + ) + + del ashift_msa, ascale_msa, agate_msa + + # Audio - Video cross attention. + if run_a2v or run_v2a: + # norm3 + vx_norm3 = comfy.ldm.common_dit.rms_norm(vx) + ax_norm3 = comfy.ldm.common_dit.rms_norm(ax) + + ( + scale_ca_audio_hidden_states_a2v, + shift_ca_audio_hidden_states_a2v, + scale_ca_audio_hidden_states_v2a, + shift_ca_audio_hidden_states_v2a, + gate_out_v2a, + ) = self.get_av_ca_ada_values( + self.scale_shift_table_a2v_ca_audio, + ax.shape[0], + a_cross_scale_shift_timestep, + a_cross_gate_timestep, + ) + + ( + scale_ca_video_hidden_states_a2v, + shift_ca_video_hidden_states_a2v, + scale_ca_video_hidden_states_v2a, + shift_ca_video_hidden_states_v2a, + gate_out_a2v, + ) = self.get_av_ca_ada_values( + self.scale_shift_table_a2v_ca_video, + vx.shape[0], + v_cross_scale_shift_timestep, + v_cross_gate_timestep, + ) + + if run_a2v: + vx_scaled = ( + vx_norm3 * (1 + scale_ca_video_hidden_states_a2v) + + shift_ca_video_hidden_states_a2v + ) + ax_scaled = ( + ax_norm3 * (1 + scale_ca_audio_hidden_states_a2v) + + shift_ca_audio_hidden_states_a2v + ) + vx += ( + self.audio_to_video_attn( + vx_scaled, + context=ax_scaled, + pe=v_cross_pe, + k_pe=a_cross_pe, + transformer_options=transformer_options, + ) + * gate_out_a2v + ) + + del gate_out_a2v + del scale_ca_video_hidden_states_a2v,\ + shift_ca_video_hidden_states_a2v,\ + scale_ca_audio_hidden_states_a2v,\ + shift_ca_audio_hidden_states_a2v,\ + + if run_v2a: + ax_scaled = ( + ax_norm3 * (1 + scale_ca_audio_hidden_states_v2a) + + shift_ca_audio_hidden_states_v2a + ) + vx_scaled = ( + vx_norm3 * (1 + scale_ca_video_hidden_states_v2a) + + shift_ca_video_hidden_states_v2a + ) + ax += ( + self.video_to_audio_attn( + ax_scaled, + context=vx_scaled, + pe=a_cross_pe, + k_pe=v_cross_pe, + transformer_options=transformer_options, + ) + * gate_out_v2a + ) + + del gate_out_v2a + del scale_ca_video_hidden_states_v2a,\ + shift_ca_video_hidden_states_v2a,\ + scale_ca_audio_hidden_states_v2a,\ + shift_ca_audio_hidden_states_v2a + + if run_vx: + vshift_mlp, vscale_mlp, vgate_mlp = ( + self.get_ada_values(self.scale_shift_table, vx.shape[0], v_timestep, slice(3, None)) + ) + + vx_scaled = comfy.ldm.common_dit.rms_norm(vx) * (1 + vscale_mlp) + vshift_mlp + vx += self.ff(vx_scaled) * vgate_mlp + del vshift_mlp, vscale_mlp, vgate_mlp + + if run_ax: + ashift_mlp, ascale_mlp, agate_mlp = ( + self.get_ada_values(self.audio_scale_shift_table, ax.shape[0], a_timestep, slice(3, None)) + ) + + ax_scaled = comfy.ldm.common_dit.rms_norm(ax) * (1 + ascale_mlp) + ashift_mlp + ax += self.audio_ff(ax_scaled) * agate_mlp + + del ashift_mlp, ascale_mlp, agate_mlp + + + return vx, ax + + +class LTXAVModel(LTXVModel): + """LTXAV model for audio-video generation.""" + + def __init__( + self, + in_channels=128, + audio_in_channels=128, + cross_attention_dim=4096, + audio_cross_attention_dim=2048, + attention_head_dim=128, + audio_attention_head_dim=64, + num_attention_heads=32, + audio_num_attention_heads=32, + caption_channels=3840, + num_layers=48, + positional_embedding_theta=10000.0, + positional_embedding_max_pos=[20, 2048, 2048], + audio_positional_embedding_max_pos=[20], + causal_temporal_positioning=False, + vae_scale_factors=(8, 32, 32), + use_middle_indices_grid=False, + timestep_scale_multiplier=1000.0, + av_ca_timestep_scale_multiplier=1.0, + dtype=None, + device=None, + operations=None, + **kwargs, + ): + # Store audio-specific parameters + self.audio_in_channels = audio_in_channels + self.audio_cross_attention_dim = audio_cross_attention_dim + self.audio_attention_head_dim = audio_attention_head_dim + self.audio_num_attention_heads = audio_num_attention_heads + self.audio_positional_embedding_max_pos = audio_positional_embedding_max_pos + + # Calculate audio dimensions + self.audio_inner_dim = audio_num_attention_heads * audio_attention_head_dim + self.audio_out_channels = audio_in_channels + + # Audio-specific constants + self.num_audio_channels = 8 + self.audio_frequency_bins = 16 + + self.av_ca_timestep_scale_multiplier = av_ca_timestep_scale_multiplier + + super().__init__( + in_channels=in_channels, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + num_attention_heads=num_attention_heads, + caption_channels=caption_channels, + num_layers=num_layers, + positional_embedding_theta=positional_embedding_theta, + positional_embedding_max_pos=positional_embedding_max_pos, + causal_temporal_positioning=causal_temporal_positioning, + vae_scale_factors=vae_scale_factors, + use_middle_indices_grid=use_middle_indices_grid, + timestep_scale_multiplier=timestep_scale_multiplier, + dtype=dtype, + device=device, + operations=operations, + **kwargs, + ) + + def _init_model_components(self, device, dtype, **kwargs): + """Initialize LTXAV-specific components.""" + # Audio-specific projections + self.audio_patchify_proj = self.operations.Linear( + self.audio_in_channels, self.audio_inner_dim, bias=True, dtype=dtype, device=device + ) + + # Audio-specific AdaLN + self.audio_adaln_single = AdaLayerNormSingle( + self.audio_inner_dim, + use_additional_conditions=False, + dtype=dtype, + device=device, + operations=self.operations, + ) + + num_scale_shift_values = 4 + self.av_ca_video_scale_shift_adaln_single = AdaLayerNormSingle( + self.inner_dim, + use_additional_conditions=False, + embedding_coefficient=num_scale_shift_values, + dtype=dtype, + device=device, + operations=self.operations, + ) + self.av_ca_a2v_gate_adaln_single = AdaLayerNormSingle( + self.inner_dim, + use_additional_conditions=False, + embedding_coefficient=1, + dtype=dtype, + device=device, + operations=self.operations, + ) + self.av_ca_audio_scale_shift_adaln_single = AdaLayerNormSingle( + self.audio_inner_dim, + use_additional_conditions=False, + embedding_coefficient=num_scale_shift_values, + dtype=dtype, + device=device, + operations=self.operations, + ) + self.av_ca_v2a_gate_adaln_single = AdaLayerNormSingle( + self.audio_inner_dim, + use_additional_conditions=False, + embedding_coefficient=1, + dtype=dtype, + device=device, + operations=self.operations, + ) + + # Audio caption projection + self.audio_caption_projection = PixArtAlphaTextProjection( + in_features=self.caption_channels, + hidden_size=self.audio_inner_dim, + dtype=dtype, + device=device, + operations=self.operations, + ) + + def _init_transformer_blocks(self, device, dtype, **kwargs): + """Initialize transformer blocks for LTXAV.""" + self.transformer_blocks = nn.ModuleList( + [ + BasicAVTransformerBlock( + v_dim=self.inner_dim, + a_dim=self.audio_inner_dim, + v_heads=self.num_attention_heads, + a_heads=self.audio_num_attention_heads, + vd_head=self.attention_head_dim, + ad_head=self.audio_attention_head_dim, + v_context_dim=self.cross_attention_dim, + a_context_dim=self.audio_cross_attention_dim, + dtype=dtype, + device=device, + operations=self.operations, + ) + for _ in range(self.num_layers) + ] + ) + + def _init_output_components(self, device, dtype): + """Initialize output components for LTXAV.""" + # Video output components + super()._init_output_components(device, dtype) + # Audio output components + self.audio_scale_shift_table = nn.Parameter( + torch.empty(2, self.audio_inner_dim, dtype=dtype, device=device) + ) + self.audio_norm_out = self.operations.LayerNorm( + self.audio_inner_dim, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device + ) + self.audio_proj_out = self.operations.Linear( + self.audio_inner_dim, self.audio_out_channels, dtype=dtype, device=device + ) + self.a_patchifier = AudioPatchifier(1, start_end=True) + + def separate_audio_and_video_latents(self, x, audio_length): + """Separate audio and video latents from combined input.""" + # vx = x[:, : self.in_channels] + # ax = x[:, self.in_channels :] + # + # ax = ax.reshape(ax.shape[0], -1) + # ax = ax[:, : audio_length * self.num_audio_channels * self.audio_frequency_bins] + # + # ax = ax.reshape( + # ax.shape[0], self.num_audio_channels, audio_length, self.audio_frequency_bins + # ) + + vx = x[0] + ax = x[1] if len(x) > 1 else torch.zeros( + (vx.shape[0], self.num_audio_channels, 0, self.audio_frequency_bins), + device=vx.device, dtype=vx.dtype + ) + return vx, ax + + def recombine_audio_and_video_latents(self, vx, ax, target_shape=None): + if ax.numel() == 0: + return vx + else: + return [vx, ax] + """Recombine audio and video latents for output.""" + # if ax.device != vx.device or ax.dtype != vx.dtype: + # logging.warning("Audio and video latents are on different devices or dtypes.") + # ax = ax.to(device=vx.device, dtype=vx.dtype) + # logging.warning(f"Audio audio latent moved to device: {ax.device}, dtype: {ax.dtype}") + # + # ax = ax.reshape(ax.shape[0], -1) + # # pad to f x h x w of the video latents + # divisor = vx.shape[-1] * vx.shape[-2] * vx.shape[-3] + # if target_shape is None: + # repetitions = math.ceil(ax.shape[-1] / divisor) + # else: + # repetitions = target_shape[1] - vx.shape[1] + # padded_len = repetitions * divisor + # ax = F.pad(ax, (0, padded_len - ax.shape[-1])) + # ax = ax.reshape(ax.shape[0], -1, vx.shape[-3], vx.shape[-2], vx.shape[-1]) + # return torch.cat([vx, ax], dim=1) + + def _process_input(self, x, keyframe_idxs, denoise_mask, **kwargs): + """Process input for LTXAV - separate audio and video, then patchify.""" + audio_length = kwargs.get("audio_length", 0) + # Separate audio and video latents + vx, ax = self.separate_audio_and_video_latents(x, audio_length) + [vx, v_pixel_coords, additional_args] = super()._process_input( + vx, keyframe_idxs, denoise_mask, **kwargs + ) + + ax, a_latent_coords = self.a_patchifier.patchify(ax) + ax = self.audio_patchify_proj(ax) + + # additional_args.update({"av_orig_shape": list(x.shape)}) + return [vx, ax], [v_pixel_coords, a_latent_coords], additional_args + + 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 = timestep * self.timestep_scale_multiplier + v_timestep, v_embedded_timestep = self.adaln_single( + timestep.flatten(), + {"resolution": None, "aspect_ratio": None}, + batch_size=batch_size, + hidden_dtype=hidden_dtype, + ) + + # Second dimension is 1 or number of tokens (if timestep_per_token) + v_timestep = v_timestep.view(batch_size, -1, v_timestep.shape[-1]) + v_embedded_timestep = v_embedded_timestep.view( + batch_size, -1, v_embedded_timestep.shape[-1] + ) + + # Prepare audio timestep + a_timestep = kwargs.get("a_timestep") + if a_timestep is not None: + a_timestep = a_timestep * self.timestep_scale_multiplier + av_ca_factor = self.av_ca_timestep_scale_multiplier / self.timestep_scale_multiplier + + av_ca_audio_scale_shift_timestep, _ = self.av_ca_audio_scale_shift_adaln_single( + a_timestep.flatten(), + {"resolution": None, "aspect_ratio": None}, + batch_size=batch_size, + hidden_dtype=hidden_dtype, + ) + av_ca_video_scale_shift_timestep, _ = self.av_ca_video_scale_shift_adaln_single( + timestep.flatten(), + {"resolution": None, "aspect_ratio": None}, + batch_size=batch_size, + hidden_dtype=hidden_dtype, + ) + av_ca_a2v_gate_noise_timestep, _ = self.av_ca_a2v_gate_adaln_single( + timestep.flatten() * av_ca_factor, + {"resolution": None, "aspect_ratio": None}, + batch_size=batch_size, + hidden_dtype=hidden_dtype, + ) + av_ca_v2a_gate_noise_timestep, _ = self.av_ca_v2a_gate_adaln_single( + a_timestep.flatten() * av_ca_factor, + {"resolution": None, "aspect_ratio": None}, + batch_size=batch_size, + hidden_dtype=hidden_dtype, + ) + + a_timestep, a_embedded_timestep = self.audio_adaln_single( + a_timestep.flatten(), + {"resolution": None, "aspect_ratio": None}, + batch_size=batch_size, + hidden_dtype=hidden_dtype, + ) + a_timestep = a_timestep.view(batch_size, -1, a_timestep.shape[-1]) + a_embedded_timestep = a_embedded_timestep.view( + batch_size, -1, a_embedded_timestep.shape[-1] + ) + cross_av_timestep_ss = [ + av_ca_audio_scale_shift_timestep, + av_ca_video_scale_shift_timestep, + av_ca_a2v_gate_noise_timestep, + av_ca_v2a_gate_noise_timestep, + ] + cross_av_timestep_ss = list( + [t.view(batch_size, -1, t.shape[-1]) for t in cross_av_timestep_ss] + ) + else: + a_timestep = timestep + a_embedded_timestep = kwargs.get("embedded_timestep") + cross_av_timestep_ss = [] + + return [v_timestep, a_timestep, cross_av_timestep_ss], [ + v_embedded_timestep, + a_embedded_timestep, + ] + + def _prepare_context(self, context, batch_size, x, attention_mask=None): + vx = x[0] + ax = x[1] + v_context, a_context = torch.split( + context, int(context.shape[-1] / 2), len(context.shape) - 1 + ) + + v_context, attention_mask = super()._prepare_context( + v_context, batch_size, vx, attention_mask + ) + if self.audio_caption_projection is not None: + a_context = self.audio_caption_projection(a_context) + a_context = a_context.view(batch_size, -1, ax.shape[-1]) + + return [v_context, a_context], attention_mask + + def _prepare_positional_embeddings(self, pixel_coords, frame_rate, x_dtype): + v_pixel_coords = pixel_coords[0] + v_pe = super()._prepare_positional_embeddings(v_pixel_coords, frame_rate, x_dtype) + + a_latent_coords = pixel_coords[1] + a_pe = self._precompute_freqs_cis( + a_latent_coords, + dim=self.audio_inner_dim, + out_dtype=x_dtype, + max_pos=self.audio_positional_embedding_max_pos, + use_middle_indices_grid=self.use_middle_indices_grid, + num_attention_heads=self.audio_num_attention_heads, + ) + + # calculate positional embeddings for the middle of the token duration, to use in av cross attention layers. + max_pos = max( + self.positional_embedding_max_pos[0], self.audio_positional_embedding_max_pos[0] + ) + v_pixel_coords = v_pixel_coords.to(torch.float32) + v_pixel_coords[:, 0] = v_pixel_coords[:, 0] * (1.0 / frame_rate) + av_cross_video_freq_cis = self._precompute_freqs_cis( + v_pixel_coords[:, 0:1, :], + dim=self.audio_cross_attention_dim, + out_dtype=x_dtype, + max_pos=[max_pos], + use_middle_indices_grid=True, + num_attention_heads=self.audio_num_attention_heads, + ) + av_cross_audio_freq_cis = self._precompute_freqs_cis( + a_latent_coords[:, 0:1, :], + dim=self.audio_cross_attention_dim, + out_dtype=x_dtype, + max_pos=[max_pos], + use_middle_indices_grid=True, + num_attention_heads=self.audio_num_attention_heads, + ) + + return [(v_pe, av_cross_video_freq_cis), (a_pe, av_cross_audio_freq_cis)] + + def _process_transformer_blocks( + self, x, context, attention_mask, timestep, pe, transformer_options={}, **kwargs + ): + vx = x[0] + ax = x[1] + v_context = context[0] + a_context = context[1] + v_timestep = timestep[0] + a_timestep = timestep[1] + v_pe, av_cross_video_freq_cis = pe[0] + a_pe, av_cross_audio_freq_cis = pe[1] + + ( + av_ca_audio_scale_shift_timestep, + av_ca_video_scale_shift_timestep, + av_ca_a2v_gate_noise_timestep, + av_ca_v2a_gate_noise_timestep, + ) = timestep[2] + + """Process transformer blocks for LTXAV.""" + patches_replace = transformer_options.get("patches_replace", {}) + blocks_replace = patches_replace.get("dit", {}) + + # Process transformer blocks + for i, block in enumerate(self.transformer_blocks): + if ("double_block", i) in blocks_replace: + + def block_wrap(args): + out = {} + out["img"] = block( + args["img"], + v_context=args["v_context"], + a_context=args["a_context"], + attention_mask=args["attention_mask"], + v_timestep=args["v_timestep"], + a_timestep=args["a_timestep"], + v_pe=args["v_pe"], + a_pe=args["a_pe"], + v_cross_pe=args["v_cross_pe"], + a_cross_pe=args["a_cross_pe"], + v_cross_scale_shift_timestep=args["v_cross_scale_shift_timestep"], + a_cross_scale_shift_timestep=args["a_cross_scale_shift_timestep"], + v_cross_gate_timestep=args["v_cross_gate_timestep"], + a_cross_gate_timestep=args["a_cross_gate_timestep"], + transformer_options=args["transformer_options"], + ) + return out + + out = blocks_replace[("double_block", i)]( + { + "img": (vx, ax), + "v_context": v_context, + "a_context": a_context, + "attention_mask": attention_mask, + "v_timestep": v_timestep, + "a_timestep": a_timestep, + "v_pe": v_pe, + "a_pe": a_pe, + "v_cross_pe": av_cross_video_freq_cis, + "a_cross_pe": av_cross_audio_freq_cis, + "v_cross_scale_shift_timestep": av_ca_video_scale_shift_timestep, + "a_cross_scale_shift_timestep": av_ca_audio_scale_shift_timestep, + "v_cross_gate_timestep": av_ca_a2v_gate_noise_timestep, + "a_cross_gate_timestep": av_ca_v2a_gate_noise_timestep, + "transformer_options": transformer_options, + }, + {"original_block": block_wrap}, + ) + vx, ax = out["img"] + else: + vx, ax = block( + (vx, ax), + v_context=v_context, + a_context=a_context, + attention_mask=attention_mask, + v_timestep=v_timestep, + a_timestep=a_timestep, + v_pe=v_pe, + a_pe=a_pe, + v_cross_pe=av_cross_video_freq_cis, + a_cross_pe=av_cross_audio_freq_cis, + v_cross_scale_shift_timestep=av_ca_video_scale_shift_timestep, + a_cross_scale_shift_timestep=av_ca_audio_scale_shift_timestep, + v_cross_gate_timestep=av_ca_a2v_gate_noise_timestep, + a_cross_gate_timestep=av_ca_v2a_gate_noise_timestep, + transformer_options=transformer_options, + ) + + return [vx, ax] + + def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs): + vx = x[0] + ax = x[1] + v_embedded_timestep = embedded_timestep[0] + a_embedded_timestep = embedded_timestep[1] + vx = super()._process_output(vx, v_embedded_timestep, keyframe_idxs, **kwargs) + + # Process audio output + a_scale_shift_values = ( + self.audio_scale_shift_table[None, None].to(device=a_embedded_timestep.device, dtype=a_embedded_timestep.dtype) + + a_embedded_timestep[:, :, None] + ) + a_shift, a_scale = a_scale_shift_values[:, :, 0], a_scale_shift_values[:, :, 1] + + ax = self.audio_norm_out(ax) + ax = ax * (1 + a_scale) + a_shift + ax = self.audio_proj_out(ax) + + # Unpatchify audio + ax = self.a_patchifier.unpatchify( + ax, channels=self.num_audio_channels, freq=self.audio_frequency_bins + ) + + # Recombine audio and video + original_shape = kwargs.get("av_orig_shape") + return self.recombine_audio_and_video_latents(vx, ax, original_shape) + + def forward( + self, + x, + timestep, + context, + attention_mask=None, + frame_rate=25, + transformer_options={}, + keyframe_idxs=None, + **kwargs, + ): + """ + Forward pass for LTXAV model. + + Args: + x: Combined audio-video input tensor + timestep: Tuple of (video_timestep, audio_timestep) or single timestep + context: Context tensor (e.g., text embeddings) + attention_mask: Attention mask tensor + frame_rate: Frame rate for temporal processing + transformer_options: Additional options for transformer blocks + keyframe_idxs: Keyframe indices for temporal processing + **kwargs: Additional keyword arguments including audio_length + + Returns: + Combined audio-video output tensor + """ + # Handle timestep format + if isinstance(timestep, (tuple, list)) and len(timestep) == 2: + v_timestep, a_timestep = timestep + kwargs["a_timestep"] = a_timestep + timestep = v_timestep + else: + kwargs["a_timestep"] = timestep + + # Call parent forward method + return super().forward( + x, + timestep, + context, + attention_mask, + frame_rate, + transformer_options, + keyframe_idxs, + **kwargs, + ) diff --git a/comfy/ldm/lightricks/embeddings_connector.py b/comfy/ldm/lightricks/embeddings_connector.py new file mode 100644 index 000000000..06f5ada89 --- /dev/null +++ b/comfy/ldm/lightricks/embeddings_connector.py @@ -0,0 +1,305 @@ +import math +from typing import Optional + +import comfy.ldm.common_dit +import torch +from comfy.ldm.lightricks.model import ( + CrossAttention, + FeedForward, + generate_freq_grid_np, + interleaved_freqs_cis, + split_freqs_cis, +) +from torch import nn + + +class BasicTransformerBlock1D(nn.Module): + r""" + A basic Transformer block. + + Parameters: + + dim (`int`): The number of channels in the input and output. + num_attention_heads (`int`): The number of heads to use for multi-head attention. + attention_head_dim (`int`): The number of channels in each head. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. + attention_bias (: + obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter. + upcast_attention (`bool`, *optional*): + Whether to upcast the attention computation to float32. This is useful for mixed precision training. + norm_elementwise_affine (`bool`, *optional*, defaults to `True`): + Whether to use learnable elementwise affine parameters for normalization. + standardization_norm (`str`, *optional*, defaults to `"layer_norm"`): The type of pre-normalization to use. Can be `"layer_norm"` or `"rms_norm"`. + norm_eps (`float`, *optional*, defaults to 1e-5): Epsilon value for normalization layers. + qk_norm (`str`, *optional*, defaults to None): + Set to 'layer_norm' or `rms_norm` to perform query and key normalization. + final_dropout (`bool` *optional*, defaults to False): + Whether to apply a final dropout after the last feed-forward layer. + ff_inner_dim (`int`, *optional*): Dimension of the inner feed-forward layer. If not provided, defaults to `dim * 4`. + ff_bias (`bool`, *optional*, defaults to `True`): Whether to use bias in the feed-forward layer. + attention_out_bias (`bool`, *optional*, defaults to `True`): Whether to use bias in the attention output layer. + use_rope (`bool`, *optional*, defaults to `False`): Whether to use Rotary Position Embeddings (RoPE). + ffn_dim_mult (`int`, *optional*, defaults to 4): Multiplier for the inner dimension of the feed-forward layer. + """ + + def __init__( + self, + dim, + n_heads, + d_head, + context_dim=None, + attn_precision=None, + dtype=None, + device=None, + operations=None, + ): + super().__init__() + + # Define 3 blocks. Each block has its own normalization layer. + # 1. Self-Attn + self.attn1 = CrossAttention( + query_dim=dim, + heads=n_heads, + dim_head=d_head, + context_dim=None, + dtype=dtype, + device=device, + operations=operations, + ) + + # 3. Feed-forward + self.ff = FeedForward( + dim, + dim_out=dim, + glu=True, + dtype=dtype, + device=device, + operations=operations, + ) + + def forward(self, hidden_states, attention_mask=None, pe=None) -> torch.FloatTensor: + + # Notice that normalization is always applied before the real computation in the following blocks. + + # 1. Normalization Before Self-Attention + norm_hidden_states = comfy.ldm.common_dit.rms_norm(hidden_states) + + norm_hidden_states = norm_hidden_states.squeeze(1) + + # 2. Self-Attention + attn_output = self.attn1(norm_hidden_states, mask=attention_mask, pe=pe) + + hidden_states = attn_output + hidden_states + if hidden_states.ndim == 4: + hidden_states = hidden_states.squeeze(1) + + # 3. Normalization before Feed-Forward + norm_hidden_states = comfy.ldm.common_dit.rms_norm(hidden_states) + + # 4. Feed-forward + ff_output = self.ff(norm_hidden_states) + + hidden_states = ff_output + hidden_states + if hidden_states.ndim == 4: + hidden_states = hidden_states.squeeze(1) + + return hidden_states + + +class Embeddings1DConnector(nn.Module): + _supports_gradient_checkpointing = True + + def __init__( + self, + in_channels=128, + cross_attention_dim=2048, + attention_head_dim=128, + num_attention_heads=30, + num_layers=2, + positional_embedding_theta=10000.0, + positional_embedding_max_pos=[4096], + causal_temporal_positioning=False, + num_learnable_registers: Optional[int] = 128, + dtype=None, + device=None, + operations=None, + split_rope=False, + double_precision_rope=False, + **kwargs, + ): + super().__init__() + self.dtype = dtype + self.out_channels = in_channels + self.num_attention_heads = num_attention_heads + self.inner_dim = num_attention_heads * attention_head_dim + self.causal_temporal_positioning = causal_temporal_positioning + self.positional_embedding_theta = positional_embedding_theta + self.positional_embedding_max_pos = positional_embedding_max_pos + self.split_rope = split_rope + self.double_precision_rope = double_precision_rope + self.transformer_1d_blocks = nn.ModuleList( + [ + BasicTransformerBlock1D( + self.inner_dim, + num_attention_heads, + attention_head_dim, + context_dim=cross_attention_dim, + dtype=dtype, + device=device, + operations=operations, + ) + for _ in range(num_layers) + ] + ) + + inner_dim = num_attention_heads * attention_head_dim + self.num_learnable_registers = num_learnable_registers + if self.num_learnable_registers: + self.learnable_registers = nn.Parameter( + torch.rand( + self.num_learnable_registers, inner_dim, dtype=dtype, device=device + ) + * 2.0 + - 1.0 + ) + + def get_fractional_positions(self, indices_grid): + fractional_positions = torch.stack( + [ + indices_grid[:, i] / self.positional_embedding_max_pos[i] + for i in range(1) + ], + dim=-1, + ) + return fractional_positions + + def precompute_freqs(self, indices_grid, spacing): + source_dtype = indices_grid.dtype + dtype = ( + torch.float32 + if source_dtype in (torch.bfloat16, torch.float16) + else source_dtype + ) + + fractional_positions = self.get_fractional_positions(indices_grid) + indices = ( + generate_freq_grid_np( + self.positional_embedding_theta, + indices_grid.shape[1], + self.inner_dim, + ) + if self.double_precision_rope + else self.generate_freq_grid(spacing, dtype, fractional_positions.device) + ).to(device=fractional_positions.device) + + if spacing == "exp_2": + freqs = ( + (indices * fractional_positions.unsqueeze(-1)) + .transpose(-1, -2) + .flatten(2) + ) + else: + freqs = ( + (indices * (fractional_positions.unsqueeze(-1) * 2 - 1)) + .transpose(-1, -2) + .flatten(2) + ) + return freqs + + def generate_freq_grid(self, spacing, dtype, device): + dim = self.inner_dim + theta = self.positional_embedding_theta + n_pos_dims = 1 + n_elem = 2 * n_pos_dims # 2 for cos and sin e.g. x 3 = 6 + start = 1 + end = theta + + if spacing == "exp": + indices = theta ** (torch.arange(0, dim, n_elem, device="cpu", dtype=torch.float32) / (dim - n_elem)) + indices = indices.to(dtype=dtype, device=device) + elif spacing == "exp_2": + indices = 1.0 / theta ** (torch.arange(0, dim, n_elem, device=device) / dim) + indices = indices.to(dtype=dtype) + elif spacing == "linear": + indices = torch.linspace( + start, end, dim // n_elem, device=device, dtype=dtype + ) + elif spacing == "sqrt": + indices = torch.linspace( + start**2, end**2, dim // n_elem, device=device, dtype=dtype + ).sqrt() + + indices = indices * math.pi / 2 + + return indices + + def precompute_freqs_cis(self, indices_grid, spacing="exp"): + dim = self.inner_dim + n_elem = 2 # 2 because of cos and sin + freqs = self.precompute_freqs(indices_grid, spacing) + if self.split_rope: + expected_freqs = dim // 2 + current_freqs = freqs.shape[-1] + pad_size = expected_freqs - current_freqs + cos_freq, sin_freq = split_freqs_cis( + freqs, pad_size, self.num_attention_heads + ) + else: + cos_freq, sin_freq = interleaved_freqs_cis(freqs, dim % n_elem) + return cos_freq.to(self.dtype), sin_freq.to(self.dtype), self.split_rope + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + ): + """ + The [`Transformer2DModel`] forward method. + + Args: + hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous): + Input `hidden_states`. + indices_grid (`torch.LongTensor` of shape `(batch size, 3, num latent pixels)`): + attention_mask ( `torch.Tensor`, *optional*): + An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask + is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large + negative values to the attention scores corresponding to "discard" tokens. + Returns: + If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a + `tuple` where the first element is the sample tensor. + """ + # 1. Input + + if self.num_learnable_registers: + num_registers_duplications = math.ceil( + max(1024, hidden_states.shape[1]) / self.num_learnable_registers + ) + learnable_registers = torch.tile( + self.learnable_registers.to(hidden_states), (num_registers_duplications, 1) + ) + + hidden_states = torch.cat((hidden_states, learnable_registers[hidden_states.shape[1]:].unsqueeze(0).repeat(hidden_states.shape[0], 1, 1)), dim=1) + + if attention_mask is not None: + attention_mask = torch.zeros([1, 1, 1, hidden_states.shape[1]], dtype=attention_mask.dtype, device=attention_mask.device) + + indices_grid = torch.arange( + hidden_states.shape[1], dtype=torch.float32, device=hidden_states.device + ) + indices_grid = indices_grid[None, None, :] + freqs_cis = self.precompute_freqs_cis(indices_grid) + + # 2. Blocks + for block_idx, block in enumerate(self.transformer_1d_blocks): + hidden_states = block( + hidden_states, attention_mask=attention_mask, pe=freqs_cis + ) + + # 3. Output + # if self.output_scale is not None: + # hidden_states = hidden_states / self.output_scale + + hidden_states = comfy.ldm.common_dit.rms_norm(hidden_states) + + return hidden_states, attention_mask diff --git a/comfy/ldm/lightricks/latent_upsampler.py b/comfy/ldm/lightricks/latent_upsampler.py new file mode 100644 index 000000000..78ed7653f --- /dev/null +++ b/comfy/ldm/lightricks/latent_upsampler.py @@ -0,0 +1,292 @@ +from typing import Optional, Tuple +import torch +import torch.nn as nn +import torch.nn.functional as F +from einops import rearrange + + +def _rational_for_scale(scale: float) -> Tuple[int, int]: + mapping = {0.75: (3, 4), 1.5: (3, 2), 2.0: (2, 1), 4.0: (4, 1)} + if float(scale) not in mapping: + raise ValueError( + f"Unsupported spatial_scale {scale}. Choose from {list(mapping.keys())}" + ) + return mapping[float(scale)] + + +class PixelShuffleND(nn.Module): + def __init__(self, dims, upscale_factors=(2, 2, 2)): + super().__init__() + assert dims in [1, 2, 3], "dims must be 1, 2, or 3" + self.dims = dims + self.upscale_factors = upscale_factors + + def forward(self, x): + if self.dims == 3: + return rearrange( + x, + "b (c p1 p2 p3) d h w -> b c (d p1) (h p2) (w p3)", + p1=self.upscale_factors[0], + p2=self.upscale_factors[1], + p3=self.upscale_factors[2], + ) + elif self.dims == 2: + return rearrange( + x, + "b (c p1 p2) h w -> b c (h p1) (w p2)", + p1=self.upscale_factors[0], + p2=self.upscale_factors[1], + ) + elif self.dims == 1: + return rearrange( + x, + "b (c p1) f h w -> b c (f p1) h w", + p1=self.upscale_factors[0], + ) + + +class BlurDownsample(nn.Module): + """ + Anti-aliased spatial downsampling by integer stride using a fixed separable binomial kernel. + Applies only on H,W. Works for dims=2 or dims=3 (per-frame). + """ + + def __init__(self, dims: int, stride: int): + super().__init__() + assert dims in (2, 3) + assert stride >= 1 and isinstance(stride, int) + self.dims = dims + self.stride = stride + + # 5x5 separable binomial kernel [1,4,6,4,1] (outer product), normalized + k = torch.tensor([1.0, 4.0, 6.0, 4.0, 1.0]) + k2d = k[:, None] @ k[None, :] + k2d = (k2d / k2d.sum()).float() # shape (5,5) + self.register_buffer("kernel", k2d[None, None, :, :]) # (1,1,5,5) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + if self.stride == 1: + return x + + def _apply_2d(x2d: torch.Tensor) -> torch.Tensor: + # x2d: (B, C, H, W) + B, C, H, W = x2d.shape + weight = self.kernel.expand(C, 1, 5, 5) # depthwise + x2d = F.conv2d( + x2d, weight=weight, bias=None, stride=self.stride, padding=2, groups=C + ) + return x2d + + if self.dims == 2: + return _apply_2d(x) + else: + # dims == 3: apply per-frame on H,W + b, c, f, h, w = x.shape + x = rearrange(x, "b c f h w -> (b f) c h w") + x = _apply_2d(x) + h2, w2 = x.shape[-2:] + x = rearrange(x, "(b f) c h w -> b c f h w", b=b, f=f, h=h2, w=w2) + return x + + +class SpatialRationalResampler(nn.Module): + """ + Fully-learned rational spatial scaling: up by 'num' via PixelShuffle, then anti-aliased + downsample by 'den' using fixed blur + stride. Operates on H,W only. + + For dims==3, work per-frame for spatial scaling (temporal axis untouched). + """ + + def __init__(self, mid_channels: int, scale: float): + super().__init__() + self.scale = float(scale) + self.num, self.den = _rational_for_scale(self.scale) + self.conv = nn.Conv2d( + mid_channels, (self.num**2) * mid_channels, kernel_size=3, padding=1 + ) + self.pixel_shuffle = PixelShuffleND(2, upscale_factors=(self.num, self.num)) + self.blur_down = BlurDownsample(dims=2, stride=self.den) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + b, c, f, h, w = x.shape + x = rearrange(x, "b c f h w -> (b f) c h w") + x = self.conv(x) + x = self.pixel_shuffle(x) + x = self.blur_down(x) + x = rearrange(x, "(b f) c h w -> b c f h w", b=b, f=f) + return x + + +class ResBlock(nn.Module): + def __init__( + self, channels: int, mid_channels: Optional[int] = None, dims: int = 3 + ): + super().__init__() + if mid_channels is None: + mid_channels = channels + + Conv = nn.Conv2d if dims == 2 else nn.Conv3d + + self.conv1 = Conv(channels, mid_channels, kernel_size=3, padding=1) + self.norm1 = nn.GroupNorm(32, mid_channels) + self.conv2 = Conv(mid_channels, channels, kernel_size=3, padding=1) + self.norm2 = nn.GroupNorm(32, channels) + self.activation = nn.SiLU() + + def forward(self, x: torch.Tensor) -> torch.Tensor: + residual = x + x = self.conv1(x) + x = self.norm1(x) + x = self.activation(x) + x = self.conv2(x) + x = self.norm2(x) + x = self.activation(x + residual) + return x + + +class LatentUpsampler(nn.Module): + """ + Model to spatially upsample VAE latents. + + Args: + in_channels (`int`): Number of channels in the input latent + mid_channels (`int`): Number of channels in the middle layers + num_blocks_per_stage (`int`): Number of ResBlocks to use in each stage (pre/post upsampling) + dims (`int`): Number of dimensions for convolutions (2 or 3) + spatial_upsample (`bool`): Whether to spatially upsample the latent + temporal_upsample (`bool`): Whether to temporally upsample the latent + """ + + def __init__( + self, + in_channels: int = 128, + mid_channels: int = 512, + num_blocks_per_stage: int = 4, + dims: int = 3, + spatial_upsample: bool = True, + temporal_upsample: bool = False, + spatial_scale: float = 2.0, + rational_resampler: bool = False, + ): + super().__init__() + + self.in_channels = in_channels + self.mid_channels = mid_channels + self.num_blocks_per_stage = num_blocks_per_stage + self.dims = dims + self.spatial_upsample = spatial_upsample + self.temporal_upsample = temporal_upsample + self.spatial_scale = float(spatial_scale) + self.rational_resampler = rational_resampler + + Conv = nn.Conv2d if dims == 2 else nn.Conv3d + + self.initial_conv = Conv(in_channels, mid_channels, kernel_size=3, padding=1) + self.initial_norm = nn.GroupNorm(32, mid_channels) + self.initial_activation = nn.SiLU() + + self.res_blocks = nn.ModuleList( + [ResBlock(mid_channels, dims=dims) for _ in range(num_blocks_per_stage)] + ) + + if spatial_upsample and temporal_upsample: + self.upsampler = nn.Sequential( + nn.Conv3d(mid_channels, 8 * mid_channels, kernel_size=3, padding=1), + PixelShuffleND(3), + ) + elif spatial_upsample: + if rational_resampler: + self.upsampler = SpatialRationalResampler( + mid_channels=mid_channels, scale=self.spatial_scale + ) + else: + self.upsampler = nn.Sequential( + nn.Conv2d(mid_channels, 4 * mid_channels, kernel_size=3, padding=1), + PixelShuffleND(2), + ) + elif temporal_upsample: + self.upsampler = nn.Sequential( + nn.Conv3d(mid_channels, 2 * mid_channels, kernel_size=3, padding=1), + PixelShuffleND(1), + ) + else: + raise ValueError( + "Either spatial_upsample or temporal_upsample must be True" + ) + + self.post_upsample_res_blocks = nn.ModuleList( + [ResBlock(mid_channels, dims=dims) for _ in range(num_blocks_per_stage)] + ) + + self.final_conv = Conv(mid_channels, in_channels, kernel_size=3, padding=1) + + def forward(self, latent: torch.Tensor) -> torch.Tensor: + b, c, f, h, w = latent.shape + + if self.dims == 2: + x = rearrange(latent, "b c f h w -> (b f) c h w") + x = self.initial_conv(x) + x = self.initial_norm(x) + x = self.initial_activation(x) + + for block in self.res_blocks: + x = block(x) + + x = self.upsampler(x) + + for block in self.post_upsample_res_blocks: + x = block(x) + + x = self.final_conv(x) + x = rearrange(x, "(b f) c h w -> b c f h w", b=b, f=f) + else: + x = self.initial_conv(latent) + x = self.initial_norm(x) + x = self.initial_activation(x) + + for block in self.res_blocks: + x = block(x) + + if self.temporal_upsample: + x = self.upsampler(x) + x = x[:, :, 1:, :, :] + else: + if isinstance(self.upsampler, SpatialRationalResampler): + x = self.upsampler(x) + else: + x = rearrange(x, "b c f h w -> (b f) c h w") + x = self.upsampler(x) + x = rearrange(x, "(b f) c h w -> b c f h w", b=b, f=f) + + for block in self.post_upsample_res_blocks: + x = block(x) + + x = self.final_conv(x) + + return x + + @classmethod + def from_config(cls, config): + return cls( + in_channels=config.get("in_channels", 4), + mid_channels=config.get("mid_channels", 128), + num_blocks_per_stage=config.get("num_blocks_per_stage", 4), + dims=config.get("dims", 2), + spatial_upsample=config.get("spatial_upsample", True), + temporal_upsample=config.get("temporal_upsample", False), + spatial_scale=config.get("spatial_scale", 2.0), + rational_resampler=config.get("rational_resampler", False), + ) + + def config(self): + return { + "_class_name": "LatentUpsampler", + "in_channels": self.in_channels, + "mid_channels": self.mid_channels, + "num_blocks_per_stage": self.num_blocks_per_stage, + "dims": self.dims, + "spatial_upsample": self.spatial_upsample, + "temporal_upsample": self.temporal_upsample, + "spatial_scale": self.spatial_scale, + "rational_resampler": self.rational_resampler, + } diff --git a/comfy/ldm/lightricks/model.py b/comfy/ldm/lightricks/model.py index 593f7940f..d61e19d6e 100644 --- a/comfy/ldm/lightricks/model.py +++ b/comfy/ldm/lightricks/model.py @@ -1,13 +1,47 @@ +from abc import ABC, abstractmethod +from enum import Enum +import functools +import math +from typing import Dict, Optional, Tuple + +from einops import rearrange +import numpy as np import torch from torch import nn import comfy.patcher_extension import comfy.ldm.modules.attention import comfy.ldm.common_dit -import math -from typing import Dict, Optional, Tuple from .symmetric_patchifier import SymmetricPatchifier, latent_to_pixel_coords -from comfy.ldm.flux.math import apply_rope1 + +def _log_base(x, base): + return np.log(x) / np.log(base) + +class LTXRopeType(str, Enum): + INTERLEAVED = "interleaved" + SPLIT = "split" + + KEY = "rope_type" + + @classmethod + def from_dict(cls, kwargs, default=None): + if default is None: + default = cls.INTERLEAVED + return cls(kwargs.get(cls.KEY, default)) + + +class LTXFrequenciesPrecision(str, Enum): + FLOAT32 = "float32" + FLOAT64 = "float64" + + KEY = "frequencies_precision" + + @classmethod + def from_dict(cls, kwargs, default=None): + if default is None: + default = cls.FLOAT32 + return cls(kwargs.get(cls.KEY, default)) + def get_timestep_embedding( timesteps: torch.Tensor, @@ -39,9 +73,7 @@ def get_timestep_embedding( assert len(timesteps.shape) == 1, "Timesteps should be a 1d-array" half_dim = embedding_dim // 2 - exponent = -math.log(max_period) * torch.arange( - start=0, end=half_dim, dtype=torch.float32, device=timesteps.device - ) + exponent = -math.log(max_period) * torch.arange(start=0, end=half_dim, dtype=torch.float32, device=timesteps.device) exponent = exponent / (half_dim - downscale_freq_shift) emb = torch.exp(exponent) @@ -73,7 +105,9 @@ class TimestepEmbedding(nn.Module): post_act_fn: Optional[str] = None, cond_proj_dim=None, sample_proj_bias=True, - dtype=None, device=None, operations=None, + dtype=None, + device=None, + operations=None, ): super().__init__() @@ -90,7 +124,9 @@ class TimestepEmbedding(nn.Module): time_embed_dim_out = out_dim else: time_embed_dim_out = time_embed_dim - self.linear_2 = operations.Linear(time_embed_dim, time_embed_dim_out, sample_proj_bias, dtype=dtype, device=device) + self.linear_2 = operations.Linear( + time_embed_dim, time_embed_dim_out, sample_proj_bias, dtype=dtype, device=device + ) if post_act_fn is None: self.post_act = None @@ -139,12 +175,22 @@ class PixArtAlphaCombinedTimestepSizeEmbeddings(nn.Module): https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L164C9-L168C29 """ - def __init__(self, embedding_dim, size_emb_dim, use_additional_conditions: bool = False, dtype=None, device=None, operations=None): + def __init__( + self, + embedding_dim, + size_emb_dim, + use_additional_conditions: bool = False, + dtype=None, + device=None, + operations=None, + ): super().__init__() self.outdim = size_emb_dim self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0) - self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim, dtype=dtype, device=device, operations=operations) + self.timestep_embedder = TimestepEmbedding( + in_channels=256, time_embed_dim=embedding_dim, dtype=dtype, device=device, operations=operations + ) def forward(self, timestep, resolution, aspect_ratio, batch_size, hidden_dtype): timesteps_proj = self.time_proj(timestep) @@ -163,15 +209,22 @@ class AdaLayerNormSingle(nn.Module): use_additional_conditions (`bool`): To use additional conditions for normalization or not. """ - def __init__(self, embedding_dim: int, use_additional_conditions: bool = False, dtype=None, device=None, operations=None): + def __init__( + self, embedding_dim: int, embedding_coefficient: int = 6, use_additional_conditions: bool = False, dtype=None, device=None, operations=None + ): super().__init__() self.emb = PixArtAlphaCombinedTimestepSizeEmbeddings( - embedding_dim, size_emb_dim=embedding_dim // 3, use_additional_conditions=use_additional_conditions, dtype=dtype, device=device, operations=operations + embedding_dim, + size_emb_dim=embedding_dim // 3, + use_additional_conditions=use_additional_conditions, + dtype=dtype, + device=device, + operations=operations, ) self.silu = nn.SiLU() - self.linear = operations.Linear(embedding_dim, 6 * embedding_dim, bias=True, dtype=dtype, device=device) + self.linear = operations.Linear(embedding_dim, embedding_coefficient * embedding_dim, bias=True, dtype=dtype, device=device) def forward( self, @@ -185,6 +238,7 @@ class AdaLayerNormSingle(nn.Module): embedded_timestep = self.emb(timestep, **added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_dtype) return self.linear(self.silu(embedded_timestep)), embedded_timestep + class PixArtAlphaTextProjection(nn.Module): """ Projects caption embeddings. Also handles dropout for classifier-free guidance. @@ -192,18 +246,24 @@ class PixArtAlphaTextProjection(nn.Module): Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/nets/PixArt_blocks.py """ - def __init__(self, in_features, hidden_size, out_features=None, act_fn="gelu_tanh", dtype=None, device=None, operations=None): + def __init__( + self, in_features, hidden_size, out_features=None, act_fn="gelu_tanh", dtype=None, device=None, operations=None + ): super().__init__() if out_features is None: out_features = hidden_size - self.linear_1 = operations.Linear(in_features=in_features, out_features=hidden_size, bias=True, dtype=dtype, device=device) + self.linear_1 = operations.Linear( + in_features=in_features, out_features=hidden_size, bias=True, dtype=dtype, device=device + ) if act_fn == "gelu_tanh": self.act_1 = nn.GELU(approximate="tanh") elif act_fn == "silu": self.act_1 = nn.SiLU() else: raise ValueError(f"Unknown activation function: {act_fn}") - self.linear_2 = operations.Linear(in_features=hidden_size, out_features=out_features, bias=True, dtype=dtype, device=device) + self.linear_2 = operations.Linear( + in_features=hidden_size, out_features=out_features, bias=True, dtype=dtype, device=device + ) def forward(self, caption): hidden_states = self.linear_1(caption) @@ -222,23 +282,68 @@ class GELU_approx(nn.Module): class FeedForward(nn.Module): - def __init__(self, dim, dim_out, mult=4, glu=False, dropout=0., dtype=None, device=None, operations=None): + def __init__(self, dim, dim_out, mult=4, glu=False, dropout=0.0, dtype=None, device=None, operations=None): super().__init__() inner_dim = int(dim * mult) project_in = GELU_approx(dim, inner_dim, dtype=dtype, device=device, operations=operations) self.net = nn.Sequential( - project_in, - nn.Dropout(dropout), - operations.Linear(inner_dim, dim_out, dtype=dtype, device=device) + project_in, nn.Dropout(dropout), operations.Linear(inner_dim, dim_out, dtype=dtype, device=device) ) def forward(self, x): return self.net(x) +def apply_rotary_emb(input_tensor, freqs_cis): + cos_freqs, sin_freqs = freqs_cis[0], freqs_cis[1] + split_pe = freqs_cis[2] if len(freqs_cis) > 2 else False + return ( + apply_split_rotary_emb(input_tensor, cos_freqs, sin_freqs) + if split_pe else + apply_interleaved_rotary_emb(input_tensor, cos_freqs, sin_freqs) + ) + +def apply_interleaved_rotary_emb(input_tensor, cos_freqs, sin_freqs): # TODO: remove duplicate funcs and pick the best/fastest one + t_dup = rearrange(input_tensor, "... (d r) -> ... d r", r=2) + t1, t2 = t_dup.unbind(dim=-1) + t_dup = torch.stack((-t2, t1), dim=-1) + input_tensor_rot = rearrange(t_dup, "... d r -> ... (d r)") + + out = input_tensor * cos_freqs + input_tensor_rot * sin_freqs + + return out + +def apply_split_rotary_emb(input_tensor, cos, sin): + needs_reshape = False + if input_tensor.ndim != 4 and cos.ndim == 4: + B, H, T, _ = cos.shape + input_tensor = input_tensor.reshape(B, T, H, -1).swapaxes(1, 2) + needs_reshape = True + split_input = rearrange(input_tensor, "... (d r) -> ... d r", d=2) + first_half_input = split_input[..., :1, :] + second_half_input = split_input[..., 1:, :] + output = split_input * cos.unsqueeze(-2) + first_half_output = output[..., :1, :] + second_half_output = output[..., 1:, :] + first_half_output.addcmul_(-sin.unsqueeze(-2), second_half_input) + second_half_output.addcmul_(sin.unsqueeze(-2), first_half_input) + output = rearrange(output, "... d r -> ... (d r)") + return output.swapaxes(1, 2).reshape(B, T, -1) if needs_reshape else output + class CrossAttention(nn.Module): - def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., attn_precision=None, dtype=None, device=None, operations=None): + def __init__( + self, + query_dim, + context_dim=None, + heads=8, + dim_head=64, + dropout=0.0, + attn_precision=None, + dtype=None, + device=None, + operations=None, + ): super().__init__() inner_dim = dim_head * heads context_dim = query_dim if context_dim is None else context_dim @@ -254,9 +359,11 @@ class CrossAttention(nn.Module): self.to_k = operations.Linear(context_dim, inner_dim, bias=True, dtype=dtype, device=device) self.to_v = operations.Linear(context_dim, inner_dim, bias=True, dtype=dtype, device=device) - self.to_out = nn.Sequential(operations.Linear(inner_dim, query_dim, dtype=dtype, device=device), nn.Dropout(dropout)) + self.to_out = nn.Sequential( + operations.Linear(inner_dim, query_dim, dtype=dtype, device=device), nn.Dropout(dropout) + ) - def forward(self, x, context=None, mask=None, pe=None, transformer_options={}): + def forward(self, x, context=None, mask=None, pe=None, k_pe=None, transformer_options={}): q = self.to_q(x) context = x if context is None else context k = self.to_k(context) @@ -266,8 +373,8 @@ class CrossAttention(nn.Module): k = self.k_norm(k) if pe is not None: - q = apply_rope1(q.unsqueeze(1), pe).squeeze(1) - k = apply_rope1(k.unsqueeze(1), pe).squeeze(1) + q = apply_rotary_emb(q, pe) + k = apply_rotary_emb(k, pe if k_pe is None else k_pe) 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) @@ -277,14 +384,34 @@ class CrossAttention(nn.Module): class BasicTransformerBlock(nn.Module): - def __init__(self, dim, n_heads, d_head, context_dim=None, attn_precision=None, dtype=None, device=None, operations=None): + def __init__( + self, dim, n_heads, d_head, context_dim=None, attn_precision=None, dtype=None, device=None, operations=None + ): super().__init__() self.attn_precision = attn_precision - self.attn1 = CrossAttention(query_dim=dim, heads=n_heads, dim_head=d_head, context_dim=None, attn_precision=self.attn_precision, dtype=dtype, device=device, operations=operations) + self.attn1 = CrossAttention( + query_dim=dim, + heads=n_heads, + dim_head=d_head, + context_dim=None, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) self.ff = FeedForward(dim, dim_out=dim, glu=True, dtype=dtype, device=device, operations=operations) - self.attn2 = CrossAttention(query_dim=dim, context_dim=context_dim, heads=n_heads, dim_head=d_head, attn_precision=self.attn_precision, dtype=dtype, device=device, operations=operations) + self.attn2 = CrossAttention( + query_dim=dim, + context_dim=context_dim, + heads=n_heads, + dim_head=d_head, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) self.scale_shift_table = nn.Parameter(torch.empty(6, dim, device=device, dtype=dtype)) @@ -306,116 +433,446 @@ class BasicTransformerBlock(nn.Module): return x def get_fractional_positions(indices_grid, max_pos): + n_pos_dims = indices_grid.shape[1] + assert n_pos_dims == len(max_pos), f'Number of position dimensions ({n_pos_dims}) must match max_pos length ({len(max_pos)})' fractional_positions = torch.stack( - [ - indices_grid[:, i] / max_pos[i] - for i in range(3) - ], - dim=-1, + [indices_grid[:, i] / max_pos[i] for i in range(n_pos_dims)], + axis=-1, ) return fractional_positions -def precompute_freqs_cis(indices_grid, dim, out_dtype, theta=10000.0, max_pos=[20, 2048, 2048]): - dtype = torch.float32 - device = indices_grid.device +@functools.lru_cache(maxsize=5) +def generate_freq_grid_np(positional_embedding_theta, positional_embedding_max_pos_count, inner_dim, _ = None): + theta = positional_embedding_theta + start = 1 + end = theta + + n_elem = 2 * positional_embedding_max_pos_count + pow_indices = np.power( + theta, + np.linspace( + _log_base(start, theta), + _log_base(end, theta), + inner_dim // n_elem, + dtype=np.float64, + ), + ) + return torch.tensor(pow_indices * math.pi / 2, dtype=torch.float32) + +def generate_freq_grid_pytorch(positional_embedding_theta, positional_embedding_max_pos_count, inner_dim, device): + theta = positional_embedding_theta + start = 1 + end = theta + n_elem = 2 * positional_embedding_max_pos_count + + indices = theta ** ( + torch.linspace( + math.log(start, theta), + math.log(end, theta), + inner_dim // n_elem, + device=device, + dtype=torch.float32, + ) + ) + indices = indices.to(dtype=torch.float32) + + indices = indices * math.pi / 2 + + return indices + +def generate_freqs(indices, indices_grid, max_pos, use_middle_indices_grid): + if use_middle_indices_grid: + assert(len(indices_grid.shape) == 4 and indices_grid.shape[-1] ==2) + indices_grid_start, indices_grid_end = indices_grid[..., 0], indices_grid[..., 1] + indices_grid = (indices_grid_start + indices_grid_end) / 2.0 + elif len(indices_grid.shape) == 4: + indices_grid = indices_grid[..., 0] # Get fractional positions and compute frequency indices fractional_positions = get_fractional_positions(indices_grid, max_pos) - indices = theta ** torch.linspace(0, 1, dim // 6, device=device, dtype=dtype) * math.pi / 2 + indices = indices.to(device=fractional_positions.device) - # Compute frequencies and apply cos/sin - freqs = (indices * (fractional_positions.unsqueeze(-1) * 2 - 1)).transpose(-1, -2).flatten(2) - cos_vals = freqs.cos().repeat_interleave(2, dim=-1) - sin_vals = freqs.sin().repeat_interleave(2, dim=-1) + freqs = ( + (indices * (fractional_positions.unsqueeze(-1) * 2 - 1)) + .transpose(-1, -2) + .flatten(2) + ) + return freqs - # Pad if dim is not divisible by 6 - if dim % 6 != 0: - padding_size = dim % 6 - cos_vals = torch.cat([torch.ones_like(cos_vals[:, :, :padding_size]), cos_vals], dim=-1) - sin_vals = torch.cat([torch.zeros_like(sin_vals[:, :, :padding_size]), sin_vals], dim=-1) +def interleaved_freqs_cis(freqs, pad_size): + cos_freq = freqs.cos().repeat_interleave(2, dim=-1) + sin_freq = freqs.sin().repeat_interleave(2, dim=-1) + if pad_size != 0: + cos_padding = torch.ones_like(cos_freq[:, :, : pad_size]) + sin_padding = torch.zeros_like(cos_freq[:, :, : pad_size]) + cos_freq = torch.cat([cos_padding, cos_freq], dim=-1) + sin_freq = torch.cat([sin_padding, sin_freq], dim=-1) + return cos_freq, sin_freq - # Reshape and extract one value per pair (since repeat_interleave duplicates each value) - cos_vals = cos_vals.reshape(*cos_vals.shape[:2], -1, 2)[..., 0].to(out_dtype) # [B, N, dim//2] - sin_vals = sin_vals.reshape(*sin_vals.shape[:2], -1, 2)[..., 0].to(out_dtype) # [B, N, dim//2] +def split_freqs_cis(freqs, pad_size, num_attention_heads): + cos_freq = freqs.cos() + sin_freq = freqs.sin() - # Build rotation matrix [[cos, -sin], [sin, cos]] and add heads dimension - freqs_cis = torch.stack([ - torch.stack([cos_vals, -sin_vals], dim=-1), - torch.stack([sin_vals, cos_vals], dim=-1) - ], dim=-2).unsqueeze(1) # [B, 1, N, dim//2, 2, 2] + if pad_size != 0: + cos_padding = torch.ones_like(cos_freq[:, :, :pad_size]) + sin_padding = torch.zeros_like(sin_freq[:, :, :pad_size]) - return freqs_cis + cos_freq = torch.concatenate([cos_padding, cos_freq], axis=-1) + sin_freq = torch.concatenate([sin_padding, sin_freq], axis=-1) + # Reshape freqs to be compatible with multi-head attention + B , T, half_HD = cos_freq.shape -class LTXVModel(torch.nn.Module): - def __init__(self, - in_channels=128, - cross_attention_dim=2048, - attention_head_dim=64, - num_attention_heads=32, + cos_freq = cos_freq.reshape(B, T, num_attention_heads, half_HD // num_attention_heads) + sin_freq = sin_freq.reshape(B, T, num_attention_heads, half_HD // num_attention_heads) - caption_channels=4096, - num_layers=28, + cos_freq = torch.swapaxes(cos_freq, 1, 2) # (B,H,T,D//2) + sin_freq = torch.swapaxes(sin_freq, 1, 2) # (B,H,T,D//2) + return cos_freq, sin_freq +class LTXBaseModel(torch.nn.Module, ABC): + """ + Abstract base class for LTX models (Lightricks Transformer models). - positional_embedding_theta=10000.0, - positional_embedding_max_pos=[20, 2048, 2048], - causal_temporal_positioning=False, - vae_scale_factors=(8, 32, 32), - dtype=None, device=None, operations=None, **kwargs): + This class defines the common interface and shared functionality for all LTX models, + including LTXV (video) and LTXAV (audio-video) variants. + """ + + def __init__( + self, + in_channels: int, + cross_attention_dim: int, + attention_head_dim: int, + num_attention_heads: int, + caption_channels: int, + num_layers: int, + positional_embedding_theta: float = 10000.0, + positional_embedding_max_pos: list = [20, 2048, 2048], + causal_temporal_positioning: bool = False, + vae_scale_factors: tuple = (8, 32, 32), + use_middle_indices_grid=False, + timestep_scale_multiplier = 1000.0, + dtype=None, + device=None, + operations=None, + **kwargs, + ): super().__init__() self.generator = None self.vae_scale_factors = vae_scale_factors + self.use_middle_indices_grid = use_middle_indices_grid self.dtype = dtype - self.out_channels = in_channels - self.inner_dim = num_attention_heads * attention_head_dim + self.in_channels = in_channels + self.cross_attention_dim = cross_attention_dim + self.attention_head_dim = attention_head_dim + self.num_attention_heads = num_attention_heads + self.caption_channels = caption_channels + self.num_layers = num_layers + self.positional_embedding_theta = positional_embedding_theta + self.positional_embedding_max_pos = positional_embedding_max_pos + self.split_positional_embedding = LTXRopeType.from_dict(kwargs) + self.freq_grid_generator = ( + generate_freq_grid_np if LTXFrequenciesPrecision.from_dict(kwargs) == LTXFrequenciesPrecision.FLOAT64 + else generate_freq_grid_pytorch + ) self.causal_temporal_positioning = causal_temporal_positioning + self.operations = operations + self.timestep_scale_multiplier = timestep_scale_multiplier - self.patchify_proj = operations.Linear(in_channels, self.inner_dim, bias=True, dtype=dtype, device=device) + # Common dimensions + self.inner_dim = num_attention_heads * attention_head_dim + self.out_channels = in_channels + + # Initialize common components + self._init_common_components(device, dtype) + + # Initialize model-specific components + self._init_model_components(device, dtype, **kwargs) + + # Initialize transformer blocks + self._init_transformer_blocks(device, dtype, **kwargs) + + # Initialize output components + self._init_output_components(device, dtype) + + def _init_common_components(self, device, dtype): + """Initialize components common to all LTX models + - patchify_proj: Linear projection for patchifying input + - adaln_single: AdaLN layer for timestep embedding + - caption_projection: Linear projection for caption embedding + """ + self.patchify_proj = self.operations.Linear( + self.in_channels, self.inner_dim, bias=True, dtype=dtype, device=device + ) self.adaln_single = AdaLayerNormSingle( - self.inner_dim, use_additional_conditions=False, dtype=dtype, device=device, operations=operations + self.inner_dim, use_additional_conditions=False, dtype=dtype, device=device, operations=self.operations ) - # self.adaln_single.linear = operations.Linear(self.inner_dim, 4 * self.inner_dim, bias=True, dtype=dtype, device=device) - self.caption_projection = PixArtAlphaTextProjection( - in_features=caption_channels, hidden_size=self.inner_dim, dtype=dtype, device=device, operations=operations + in_features=self.caption_channels, + hidden_size=self.inner_dim, + dtype=dtype, + device=device, + operations=self.operations, ) + @abstractmethod + def _init_model_components(self, device, dtype, **kwargs): + """Initialize model-specific components. Must be implemented by subclasses.""" + pass + + @abstractmethod + def _init_transformer_blocks(self, device, dtype, **kwargs): + """Initialize transformer blocks. Must be implemented by subclasses.""" + pass + + @abstractmethod + def _init_output_components(self, device, dtype): + """Initialize output components. Must be implemented by subclasses.""" + pass + + @abstractmethod + def _process_input(self, x, keyframe_idxs, denoise_mask, **kwargs): + """Process input data. Must be implemented by subclasses.""" + pass + + @abstractmethod + def _process_transformer_blocks(self, x, context, attention_mask, timestep, pe, **kwargs): + """Process transformer blocks. Must be implemented by subclasses.""" + pass + + @abstractmethod + def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs): + """Process output data. Must be implemented by subclasses.""" + pass + + def _prepare_timestep(self, timestep, batch_size, hidden_dtype, **kwargs): + """Prepare timestep embeddings.""" + grid_mask = kwargs.get("grid_mask", None) + if grid_mask is not None: + timestep = timestep[:, grid_mask] + + timestep = timestep * self.timestep_scale_multiplier + timestep, embedded_timestep = self.adaln_single( + timestep.flatten(), + {"resolution": None, "aspect_ratio": None}, + batch_size=batch_size, + hidden_dtype=hidden_dtype, + ) + + # Second dimension is 1 or number of tokens (if timestep_per_token) + timestep = timestep.view(batch_size, -1, timestep.shape[-1]) + embedded_timestep = embedded_timestep.view(batch_size, -1, embedded_timestep.shape[-1]) + + return timestep, embedded_timestep + + def _prepare_context(self, context, batch_size, x, attention_mask=None): + """Prepare context for transformer blocks.""" + if self.caption_projection is not None: + context = self.caption_projection(context) + context = context.view(batch_size, -1, x.shape[-1]) + + return context, attention_mask + + def _precompute_freqs_cis( + self, + indices_grid, + dim, + out_dtype, + theta=10000.0, + max_pos=[20, 2048, 2048], + use_middle_indices_grid=False, + num_attention_heads=32, + ): + split_mode = self.split_positional_embedding == LTXRopeType.SPLIT + indices = self.freq_grid_generator(theta, indices_grid.shape[1], dim, indices_grid.device) + freqs = generate_freqs(indices, indices_grid, max_pos, use_middle_indices_grid) + + if split_mode: + expected_freqs = dim // 2 + current_freqs = freqs.shape[-1] + pad_size = expected_freqs - current_freqs + cos_freq, sin_freq = split_freqs_cis(freqs, pad_size, num_attention_heads) + else: + # 2 because of cos and sin by 3 for (t, x, y), 1 for temporal only + n_elem = 2 * indices_grid.shape[1] + cos_freq, sin_freq = interleaved_freqs_cis(freqs, dim % n_elem) + return cos_freq.to(out_dtype), sin_freq.to(out_dtype), split_mode + + def _prepare_positional_embeddings(self, pixel_coords, frame_rate, x_dtype): + """Prepare positional embeddings.""" + fractional_coords = pixel_coords.to(torch.float32) + fractional_coords[:, 0] = fractional_coords[:, 0] * (1.0 / frame_rate) + pe = self._precompute_freqs_cis( + fractional_coords, + dim=self.inner_dim, + out_dtype=x_dtype, + max_pos=self.positional_embedding_max_pos, + use_middle_indices_grid=self.use_middle_indices_grid, + num_attention_heads=self.num_attention_heads, + ) + return pe + + def _prepare_attention_mask(self, attention_mask, x_dtype): + """Prepare attention mask.""" + if attention_mask is not None and not torch.is_floating_point(attention_mask): + attention_mask = (attention_mask - 1).to(x_dtype).reshape( + (attention_mask.shape[0], 1, -1, attention_mask.shape[-1]) + ) * torch.finfo(x_dtype).max + return attention_mask + + def forward( + self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, denoise_mask=None, **kwargs + ): + """ + Forward pass for LTX models. + + Args: + x: Input tensor + timestep: Timestep tensor + context: Context tensor (e.g., text embeddings) + attention_mask: Attention mask tensor + frame_rate: Frame rate for temporal processing + transformer_options: Additional options for transformer blocks + keyframe_idxs: Keyframe indices for temporal processing + **kwargs: Additional keyword arguments + + Returns: + Processed output tensor + """ + return comfy.patcher_extension.WrapperExecutor.new_class_executor( + self._forward, + self, + comfy.patcher_extension.get_all_wrappers( + comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options + ), + ).execute(x, timestep, context, attention_mask, frame_rate, transformer_options, keyframe_idxs, denoise_mask=denoise_mask, **kwargs) + + def _forward( + self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, denoise_mask=None, **kwargs + ): + """ + Internal forward pass for LTX models. + + Args: + x: Input tensor + timestep: Timestep tensor + context: Context tensor (e.g., text embeddings) + attention_mask: Attention mask tensor + frame_rate: Frame rate for temporal processing + transformer_options: Additional options for transformer blocks + keyframe_idxs: Keyframe indices for temporal processing + **kwargs: Additional keyword arguments + + Returns: + Processed output tensor + """ + if isinstance(x, list): + input_dtype = x[0].dtype + batch_size = x[0].shape[0] + else: + input_dtype = x.dtype + batch_size = x.shape[0] + # Process input + merged_args = {**transformer_options, **kwargs} + x, pixel_coords, additional_args = self._process_input(x, keyframe_idxs, denoise_mask, **merged_args) + merged_args.update(additional_args) + + # Prepare timestep and context + timestep, embedded_timestep = self._prepare_timestep(timestep, batch_size, input_dtype, **merged_args) + context, attention_mask = self._prepare_context(context, batch_size, x, attention_mask) + + # Prepare attention mask and positional embeddings + attention_mask = self._prepare_attention_mask(attention_mask, input_dtype) + pe = self._prepare_positional_embeddings(pixel_coords, frame_rate, input_dtype) + + # Process transformer blocks + x = self._process_transformer_blocks( + x, context, attention_mask, timestep, pe, transformer_options=transformer_options, **merged_args + ) + + # Process output + x = self._process_output(x, embedded_timestep, keyframe_idxs, **merged_args) + return x + + +class LTXVModel(LTXBaseModel): + """LTXV model for video generation.""" + + def __init__( + self, + in_channels=128, + cross_attention_dim=2048, + attention_head_dim=64, + num_attention_heads=32, + caption_channels=4096, + num_layers=28, + positional_embedding_theta=10000.0, + positional_embedding_max_pos=[20, 2048, 2048], + causal_temporal_positioning=False, + vae_scale_factors=(8, 32, 32), + use_middle_indices_grid=False, + timestep_scale_multiplier = 1000.0, + dtype=None, + device=None, + operations=None, + **kwargs, + ): + super().__init__( + in_channels=in_channels, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + num_attention_heads=num_attention_heads, + caption_channels=caption_channels, + num_layers=num_layers, + positional_embedding_theta=positional_embedding_theta, + positional_embedding_max_pos=positional_embedding_max_pos, + causal_temporal_positioning=causal_temporal_positioning, + vae_scale_factors=vae_scale_factors, + use_middle_indices_grid=use_middle_indices_grid, + timestep_scale_multiplier=timestep_scale_multiplier, + dtype=dtype, + device=device, + operations=operations, + **kwargs, + ) + + def _init_model_components(self, device, dtype, **kwargs): + """Initialize LTXV-specific components.""" + # No additional components needed for LTXV beyond base class + pass + + def _init_transformer_blocks(self, device, dtype, **kwargs): + """Initialize transformer blocks for LTXV.""" self.transformer_blocks = nn.ModuleList( [ BasicTransformerBlock( self.inner_dim, - num_attention_heads, - attention_head_dim, - context_dim=cross_attention_dim, - # attn_precision=attn_precision, - dtype=dtype, device=device, operations=operations + self.num_attention_heads, + self.attention_head_dim, + context_dim=self.cross_attention_dim, + dtype=dtype, + device=device, + operations=self.operations, ) - for d in range(num_layers) + for _ in range(self.num_layers) ] ) + def _init_output_components(self, device, dtype): + """Initialize output components for LTXV.""" self.scale_shift_table = nn.Parameter(torch.empty(2, self.inner_dim, dtype=dtype, device=device)) - self.norm_out = operations.LayerNorm(self.inner_dim, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - self.proj_out = operations.Linear(self.inner_dim, self.out_channels, dtype=dtype, device=device) - - self.patchifier = SymmetricPatchifier(1) - - def forward(self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, **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, timestep, context, attention_mask, frame_rate, transformer_options, keyframe_idxs, **kwargs) - - def _forward(self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, **kwargs): - patches_replace = transformer_options.get("patches_replace", {}) - - orig_shape = list(x.shape) + self.norm_out = self.operations.LayerNorm( + self.inner_dim, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device + ) + self.proj_out = self.operations.Linear(self.inner_dim, self.out_channels, dtype=dtype, device=device) + self.patchifier = SymmetricPatchifier(1, start_end=True) + def _process_input(self, x, keyframe_idxs, denoise_mask, **kwargs): + """Process input for LTXV.""" + additional_args = {"orig_shape": list(x.shape)} x, latent_coords = self.patchifier.patchify(x) pixel_coords = latent_to_pixel_coords( latent_coords=latent_coords, @@ -423,44 +880,30 @@ class LTXVModel(torch.nn.Module): causal_fix=self.causal_temporal_positioning, ) + grid_mask = None if keyframe_idxs is not None: - pixel_coords[:, :, -keyframe_idxs.shape[2]:] = keyframe_idxs + additional_args.update({ "orig_patchified_shape": list(x.shape)}) + denoise_mask = self.patchifier.patchify(denoise_mask)[0] + grid_mask = ~torch.any(denoise_mask < 0, dim=-1)[0] + additional_args.update({"grid_mask": grid_mask}) + x = x[:, grid_mask, :] + pixel_coords = pixel_coords[:, :, grid_mask, ...] - fractional_coords = pixel_coords.to(torch.float32) - fractional_coords[:, 0] = fractional_coords[:, 0] * (1.0 / frame_rate) + kf_grid_mask = grid_mask[-keyframe_idxs.shape[2]:] + keyframe_idxs = keyframe_idxs[..., kf_grid_mask, :] + pixel_coords[:, :, -keyframe_idxs.shape[2]:, :] = keyframe_idxs x = self.patchify_proj(x) - timestep = timestep * 1000.0 - - if attention_mask is not None and not torch.is_floating_point(attention_mask): - attention_mask = (attention_mask - 1).to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])) * torch.finfo(x.dtype).max - - pe = precompute_freqs_cis(fractional_coords, dim=self.inner_dim, out_dtype=x.dtype) - - batch_size = x.shape[0] - timestep, embedded_timestep = self.adaln_single( - timestep.flatten(), - {"resolution": None, "aspect_ratio": None}, - batch_size=batch_size, - hidden_dtype=x.dtype, - ) - # Second dimension is 1 or number of tokens (if timestep_per_token) - timestep = timestep.view(batch_size, -1, timestep.shape[-1]) - embedded_timestep = embedded_timestep.view( - batch_size, -1, embedded_timestep.shape[-1] - ) - - # 2. Blocks - if self.caption_projection is not None: - batch_size = x.shape[0] - context = self.caption_projection(context) - context = context.view( - batch_size, -1, x.shape[-1] - ) + return x, pixel_coords, additional_args + def _process_transformer_blocks(self, x, context, attention_mask, timestep, pe, transformer_options={}, **kwargs): + """Process transformer blocks for LTXV.""" + patches_replace = transformer_options.get("patches_replace", {}) blocks_replace = patches_replace.get("dit", {}) + for i, block in enumerate(self.transformer_blocks): if ("double_block", i) in blocks_replace: + def block_wrap(args): out = {} out["img"] = block(args["img"], context=args["txt"], attention_mask=args["attention_mask"], timestep=args["vec"], pe=args["pe"], transformer_options=args["transformer_options"]) @@ -478,16 +921,28 @@ class LTXVModel(torch.nn.Module): transformer_options=transformer_options, ) - # 3. Output + return x + + def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs): + """Process output for LTXV.""" + # Apply scale-shift modulation scale_shift_values = ( self.scale_shift_table[None, None].to(device=x.device, dtype=x.dtype) + embedded_timestep[:, :, None] ) shift, scale = scale_shift_values[:, :, 0], scale_shift_values[:, :, 1] + x = self.norm_out(x) - # Modulation - x = torch.addcmul(x, x, scale).add_(shift) + x = x * (1 + scale) + shift x = self.proj_out(x) + if keyframe_idxs is not None: + grid_mask = kwargs["grid_mask"] + orig_patchified_shape = kwargs["orig_patchified_shape"] + full_x = torch.zeros(orig_patchified_shape, dtype=x.dtype, device=x.device) + full_x[:, grid_mask, :] = x + x = full_x + # Unpatchify to restore original dimensions + orig_shape = kwargs["orig_shape"] x = self.patchifier.unpatchify( latents=x, output_height=orig_shape[3], diff --git a/comfy/ldm/lightricks/symmetric_patchifier.py b/comfy/ldm/lightricks/symmetric_patchifier.py index 4b9972b9f..8f9a41186 100644 --- a/comfy/ldm/lightricks/symmetric_patchifier.py +++ b/comfy/ldm/lightricks/symmetric_patchifier.py @@ -21,20 +21,23 @@ def latent_to_pixel_coords( Returns: Tensor: A tensor of pixel coordinates corresponding to the input latent coordinates. """ + shape = [1] * latent_coords.ndim + shape[1] = -1 pixel_coords = ( latent_coords - * torch.tensor(scale_factors, device=latent_coords.device)[None, :, None] + * torch.tensor(scale_factors, device=latent_coords.device).view(*shape) ) if causal_fix: # Fix temporal scale for first frame to 1 due to causality - pixel_coords[:, 0] = (pixel_coords[:, 0] + 1 - scale_factors[0]).clamp(min=0) + pixel_coords[:, 0, ...] = (pixel_coords[:, 0, ...] + 1 - scale_factors[0]).clamp(min=0) return pixel_coords class Patchifier(ABC): - def __init__(self, patch_size: int): + def __init__(self, patch_size: int, start_end: bool=False): super().__init__() self._patch_size = (1, patch_size, patch_size) + self.start_end = start_end @abstractmethod def patchify( @@ -71,11 +74,23 @@ class Patchifier(ABC): torch.arange(0, latent_width, self._patch_size[2], device=device), indexing="ij", ) - latent_sample_coords = torch.stack(latent_sample_coords, dim=0) - latent_coords = latent_sample_coords.unsqueeze(0).repeat(batch_size, 1, 1, 1, 1) - latent_coords = rearrange( - latent_coords, "b c f h w -> b c (f h w)", b=batch_size + latent_sample_coords_start = torch.stack(latent_sample_coords, dim=0) + delta = torch.tensor(self._patch_size, device=latent_sample_coords_start.device, dtype=latent_sample_coords_start.dtype)[:, None, None, None] + latent_sample_coords_end = latent_sample_coords_start + delta + + latent_sample_coords_start = latent_sample_coords_start.unsqueeze(0).repeat(batch_size, 1, 1, 1, 1) + latent_sample_coords_start = rearrange( + latent_sample_coords_start, "b c f h w -> b c (f h w)", b=batch_size ) + if self.start_end: + latent_sample_coords_end = latent_sample_coords_end.unsqueeze(0).repeat(batch_size, 1, 1, 1, 1) + latent_sample_coords_end = rearrange( + latent_sample_coords_end, "b c f h w -> b c (f h w)", b=batch_size + ) + + latent_coords = torch.stack((latent_sample_coords_start, latent_sample_coords_end), dim=-1) + else: + latent_coords = latent_sample_coords_start return latent_coords @@ -115,3 +130,61 @@ class SymmetricPatchifier(Patchifier): q=self._patch_size[2], ) return latents + + +class AudioPatchifier(Patchifier): + def __init__(self, patch_size: int, + sample_rate=16000, + hop_length=160, + audio_latent_downsample_factor=4, + is_causal=True, + start_end=False, + shift = 0 + ): + super().__init__(patch_size, start_end=start_end) + self.hop_length = hop_length + self.sample_rate = sample_rate + self.audio_latent_downsample_factor = audio_latent_downsample_factor + self.is_causal = is_causal + self.shift = shift + + def copy_with_shift(self, shift): + return AudioPatchifier( + self.patch_size, self.sample_rate, self.hop_length, self.audio_latent_downsample_factor, + self.is_causal, self.start_end, shift + ) + + def _get_audio_latent_time_in_sec(self, start_latent, end_latent: int, dtype: torch.dtype, device=torch.device): + audio_latent_frame = torch.arange(start_latent, end_latent, dtype=dtype, device=device) + audio_mel_frame = audio_latent_frame * self.audio_latent_downsample_factor + if self.is_causal: + audio_mel_frame = (audio_mel_frame + 1 - self.audio_latent_downsample_factor).clip(min=0) + return audio_mel_frame * self.hop_length / self.sample_rate + + + def patchify(self, audio_latents: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: + # audio_latents: (batch, channels, time, freq) + b, _, t, _ = audio_latents.shape + audio_latents = rearrange( + audio_latents, + "b c t f -> b t (c f)", + ) + + audio_latents_start_timings = self._get_audio_latent_time_in_sec(self.shift, t + self.shift, torch.float32, audio_latents.device) + audio_latents_start_timings = audio_latents_start_timings.unsqueeze(0).expand(b, -1).unsqueeze(1) + + if self.start_end: + audio_latents_end_timings = self._get_audio_latent_time_in_sec(self.shift + 1, t + self.shift + 1, torch.float32, audio_latents.device) + audio_latents_end_timings = audio_latents_end_timings.unsqueeze(0).expand(b, -1).unsqueeze(1) + + audio_latents_timings = torch.stack([audio_latents_start_timings, audio_latents_end_timings], dim=-1) + else: + audio_latents_timings = audio_latents_start_timings + return audio_latents, audio_latents_timings + + def unpatchify(self, audio_latents: torch.Tensor, channels: int, freq: int) -> torch.Tensor: + # audio_latents: (batch, time, freq * channels) + audio_latents = rearrange( + audio_latents, "b t (c f) -> b c t f", c=channels, f=freq + ) + return audio_latents diff --git a/comfy/ldm/lightricks/vae/audio_vae.py b/comfy/ldm/lightricks/vae/audio_vae.py new file mode 100644 index 000000000..a9111d3bd --- /dev/null +++ b/comfy/ldm/lightricks/vae/audio_vae.py @@ -0,0 +1,286 @@ +import json +from dataclasses import dataclass +import math +import torch +import torchaudio + +import comfy.model_management +import comfy.model_patcher +import comfy.utils as utils +from comfy.ldm.mmaudio.vae.distributions import DiagonalGaussianDistribution +from comfy.ldm.lightricks.symmetric_patchifier import AudioPatchifier +from comfy.ldm.lightricks.vae.causal_audio_autoencoder import ( + CausalityAxis, + CausalAudioAutoencoder, +) +from comfy.ldm.lightricks.vocoders.vocoder import Vocoder + +LATENT_DOWNSAMPLE_FACTOR = 4 + + +@dataclass(frozen=True) +class AudioVAEComponentConfig: + """Container for model component configuration extracted from metadata.""" + + autoencoder: dict + vocoder: dict + + @classmethod + def from_metadata(cls, metadata: dict) -> "AudioVAEComponentConfig": + assert metadata is not None and "config" in metadata, "Metadata is required for audio VAE" + + raw_config = metadata["config"] + if isinstance(raw_config, str): + parsed_config = json.loads(raw_config) + else: + parsed_config = raw_config + + audio_config = parsed_config.get("audio_vae") + vocoder_config = parsed_config.get("vocoder") + + assert audio_config is not None, "Audio VAE config is required for audio VAE" + assert vocoder_config is not None, "Vocoder config is required for audio VAE" + + return cls(autoencoder=audio_config, vocoder=vocoder_config) + + +class ModelDeviceManager: + """Manages device placement and GPU residency for the composed model.""" + + def __init__(self, module: torch.nn.Module): + load_device = comfy.model_management.get_torch_device() + offload_device = comfy.model_management.vae_offload_device() + self.patcher = comfy.model_patcher.ModelPatcher(module, load_device, offload_device) + + def ensure_model_loaded(self) -> None: + comfy.model_management.free_memory( + self.patcher.model_size(), + self.patcher.load_device, + ) + comfy.model_management.load_model_gpu(self.patcher) + + def move_to_load_device(self, tensor: torch.Tensor) -> torch.Tensor: + return tensor.to(self.patcher.load_device) + + @property + def load_device(self): + return self.patcher.load_device + + +class AudioLatentNormalizer: + """Applies per-channel statistics in patch space and restores original layout.""" + + def __init__(self, patchfier: AudioPatchifier, statistics_processor: torch.nn.Module): + self.patchifier = patchfier + self.statistics = statistics_processor + + def normalize(self, latents: torch.Tensor) -> torch.Tensor: + channels = latents.shape[1] + freq = latents.shape[3] + patched, _ = self.patchifier.patchify(latents) + normalized = self.statistics.normalize(patched) + return self.patchifier.unpatchify(normalized, channels=channels, freq=freq) + + def denormalize(self, latents: torch.Tensor) -> torch.Tensor: + channels = latents.shape[1] + freq = latents.shape[3] + patched, _ = self.patchifier.patchify(latents) + denormalized = self.statistics.un_normalize(patched) + return self.patchifier.unpatchify(denormalized, channels=channels, freq=freq) + + +class AudioPreprocessor: + """Prepares raw waveforms for the autoencoder by matching training conditions.""" + + def __init__(self, target_sample_rate: int, mel_bins: int, mel_hop_length: int, n_fft: int): + self.target_sample_rate = target_sample_rate + self.mel_bins = mel_bins + self.mel_hop_length = mel_hop_length + self.n_fft = n_fft + + def resample(self, waveform: torch.Tensor, source_rate: int) -> torch.Tensor: + if source_rate == self.target_sample_rate: + return waveform + return torchaudio.functional.resample(waveform, source_rate, self.target_sample_rate) + + @staticmethod + def normalize_amplitude( + waveform: torch.Tensor, max_amplitude: float = 0.5, eps: float = 1e-5 + ) -> torch.Tensor: + waveform = waveform - waveform.mean(dim=2, keepdim=True) + peak = torch.max(torch.abs(waveform)) + eps + scale = peak.clamp(max=max_amplitude) / peak + return waveform * scale + + def waveform_to_mel( + self, waveform: torch.Tensor, waveform_sample_rate: int, device + ) -> torch.Tensor: + waveform = self.resample(waveform, waveform_sample_rate) + waveform = self.normalize_amplitude(waveform) + + mel_transform = torchaudio.transforms.MelSpectrogram( + sample_rate=self.target_sample_rate, + n_fft=self.n_fft, + win_length=self.n_fft, + hop_length=self.mel_hop_length, + f_min=0.0, + f_max=self.target_sample_rate / 2.0, + n_mels=self.mel_bins, + window_fn=torch.hann_window, + center=True, + pad_mode="reflect", + power=1.0, + mel_scale="slaney", + norm="slaney", + ).to(device) + + mel = mel_transform(waveform) + mel = torch.log(torch.clamp(mel, min=1e-5)) + return mel.permute(0, 1, 3, 2).contiguous() + + +class AudioVAE(torch.nn.Module): + """High-level Audio VAE wrapper exposing encode and decode entry points.""" + + def __init__(self, state_dict: dict, metadata: dict): + super().__init__() + + component_config = AudioVAEComponentConfig.from_metadata(metadata) + + vae_sd = utils.state_dict_prefix_replace(state_dict, {"audio_vae.": ""}, filter_keys=True) + vocoder_sd = utils.state_dict_prefix_replace(state_dict, {"vocoder.": ""}, filter_keys=True) + + self.autoencoder = CausalAudioAutoencoder(config=component_config.autoencoder) + self.vocoder = Vocoder(config=component_config.vocoder) + + self.autoencoder.load_state_dict(vae_sd, strict=False) + self.vocoder.load_state_dict(vocoder_sd, strict=False) + + autoencoder_config = self.autoencoder.get_config() + self.normalizer = AudioLatentNormalizer( + AudioPatchifier( + patch_size=1, + audio_latent_downsample_factor=LATENT_DOWNSAMPLE_FACTOR, + sample_rate=autoencoder_config["sampling_rate"], + hop_length=autoencoder_config["mel_hop_length"], + is_causal=autoencoder_config["is_causal"], + ), + self.autoencoder.per_channel_statistics, + ) + + self.preprocessor = AudioPreprocessor( + target_sample_rate=autoencoder_config["sampling_rate"], + mel_bins=autoencoder_config["mel_bins"], + mel_hop_length=autoencoder_config["mel_hop_length"], + n_fft=autoencoder_config["n_fft"], + ) + + self.device_manager = ModelDeviceManager(self) + + def encode(self, audio: dict) -> torch.Tensor: + """Encode a waveform dictionary into normalized latent tensors.""" + + waveform = audio["waveform"] + waveform_sample_rate = audio["sample_rate"] + input_device = waveform.device + # Ensure that Audio VAE is loaded on the correct device. + self.device_manager.ensure_model_loaded() + + waveform = self.device_manager.move_to_load_device(waveform) + expected_channels = self.autoencoder.encoder.in_channels + if waveform.shape[1] != expected_channels: + raise ValueError( + f"Input audio must have {expected_channels} channels, got {waveform.shape[1]}" + ) + + mel_spec = self.preprocessor.waveform_to_mel( + waveform, waveform_sample_rate, device=self.device_manager.load_device + ) + + latents = self.autoencoder.encode(mel_spec) + posterior = DiagonalGaussianDistribution(latents) + latent_mode = posterior.mode() + + normalized = self.normalizer.normalize(latent_mode) + return normalized.to(input_device) + + def decode(self, latents: torch.Tensor) -> torch.Tensor: + """Decode normalized latent tensors into an audio waveform.""" + original_shape = latents.shape + + # Ensure that Audio VAE is loaded on the correct device. + self.device_manager.ensure_model_loaded() + + latents = self.device_manager.move_to_load_device(latents) + latents = self.normalizer.denormalize(latents) + + target_shape = self.target_shape_from_latents(original_shape) + mel_spec = self.autoencoder.decode(latents, target_shape=target_shape) + + waveform = self.run_vocoder(mel_spec) + return self.device_manager.move_to_load_device(waveform) + + def target_shape_from_latents(self, latents_shape): + batch, _, time, _ = latents_shape + target_length = time * LATENT_DOWNSAMPLE_FACTOR + if self.autoencoder.causality_axis != CausalityAxis.NONE: + target_length -= LATENT_DOWNSAMPLE_FACTOR - 1 + return ( + batch, + self.autoencoder.decoder.out_ch, + target_length, + self.autoencoder.mel_bins, + ) + + def num_of_latents_from_frames(self, frames_number: int, frame_rate: int) -> int: + return math.ceil((float(frames_number) / frame_rate) * self.latents_per_second) + + def run_vocoder(self, mel_spec: torch.Tensor) -> torch.Tensor: + audio_channels = self.autoencoder.decoder.out_ch + vocoder_input = mel_spec.transpose(2, 3) + + if audio_channels == 1: + vocoder_input = vocoder_input.squeeze(1) + elif audio_channels != 2: + raise ValueError(f"Unsupported audio_channels: {audio_channels}") + + return self.vocoder(vocoder_input) + + @property + def sample_rate(self) -> int: + return int(self.autoencoder.sampling_rate) + + @property + def mel_hop_length(self) -> int: + return int(self.autoencoder.mel_hop_length) + + @property + def mel_bins(self) -> int: + return int(self.autoencoder.mel_bins) + + @property + def latent_channels(self) -> int: + return int(self.autoencoder.decoder.z_channels) + + @property + def latent_frequency_bins(self) -> int: + return int(self.mel_bins // LATENT_DOWNSAMPLE_FACTOR) + + @property + def latents_per_second(self) -> float: + return self.sample_rate / self.mel_hop_length / LATENT_DOWNSAMPLE_FACTOR + + @property + def output_sample_rate(self) -> int: + output_rate = getattr(self.vocoder, "output_sample_rate", None) + if output_rate is not None: + return int(output_rate) + upsample_factor = getattr(self.vocoder, "upsample_factor", None) + if upsample_factor is None: + raise AttributeError( + "Vocoder is missing upsample_factor; cannot infer output sample rate" + ) + return int(self.sample_rate * upsample_factor / self.mel_hop_length) + + def memory_required(self, input_shape): + return self.device_manager.patcher.model_size() diff --git a/comfy/ldm/lightricks/vae/causal_audio_autoencoder.py b/comfy/ldm/lightricks/vae/causal_audio_autoencoder.py new file mode 100644 index 000000000..f12b9bb53 --- /dev/null +++ b/comfy/ldm/lightricks/vae/causal_audio_autoencoder.py @@ -0,0 +1,909 @@ +from __future__ import annotations +import torch +from torch import nn +from torch.nn import functional as F +from typing import Optional +from enum import Enum +from .pixel_norm import PixelNorm +import comfy.ops +import logging + +ops = comfy.ops.disable_weight_init + + +class StringConvertibleEnum(Enum): + """ + Base enum class that provides string-to-enum conversion functionality. + + This mixin adds a str_to_enum() class method that handles conversion from + strings, None, or existing enum instances with case-insensitive matching. + """ + + @classmethod + def str_to_enum(cls, value): + """ + Convert a string, enum instance, or None to the appropriate enum member. + + Args: + value: Can be an enum instance of this class, a string, or None + + Returns: + Enum member of this class + + Raises: + ValueError: If the value cannot be converted to a valid enum member + """ + # Already an enum instance of this class + if isinstance(value, cls): + return value + + # None maps to NONE member if it exists + if value is None: + if hasattr(cls, "NONE"): + return cls.NONE + raise ValueError(f"{cls.__name__} does not have a NONE member to map None to") + + # String conversion (case-insensitive) + if isinstance(value, str): + value_lower = value.lower() + + # Try to match against enum values + for member in cls: + # Handle members with None values + if member.value is None: + if value_lower == "none": + return member + # Handle members with string values + elif isinstance(member.value, str) and member.value.lower() == value_lower: + return member + + # Build helpful error message with valid values + valid_values = [] + for member in cls: + if member.value is None: + valid_values.append("none") + elif isinstance(member.value, str): + valid_values.append(member.value) + + raise ValueError(f"Invalid {cls.__name__} string: '{value}'. " f"Valid values are: {valid_values}") + + raise ValueError( + f"Cannot convert type {type(value).__name__} to {cls.__name__} enum. " + f"Expected string, None, or {cls.__name__} instance." + ) + + +class AttentionType(StringConvertibleEnum): + """Enum for specifying the attention mechanism type.""" + + VANILLA = "vanilla" + LINEAR = "linear" + NONE = "none" + + +class CausalityAxis(StringConvertibleEnum): + """Enum for specifying the causality axis in causal convolutions.""" + + NONE = None + WIDTH = "width" + HEIGHT = "height" + WIDTH_COMPATIBILITY = "width-compatibility" + + +def Normalize(in_channels, *, num_groups=32, normtype="group"): + if normtype == "group": + return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True) + elif normtype == "pixel": + return PixelNorm(dim=1, eps=1e-6) + else: + raise ValueError(f"Invalid normalization type: {normtype}") + + +class CausalConv2d(nn.Module): + """ + A causal 2D convolution. + + This layer ensures that the output at time `t` only depends on inputs + at time `t` and earlier. It achieves this by applying asymmetric padding + to the time dimension (width) before the convolution. + """ + + def __init__( + self, + in_channels, + out_channels, + kernel_size, + stride=1, + dilation=1, + groups=1, + bias=True, + causality_axis: CausalityAxis = CausalityAxis.HEIGHT, + ): + super().__init__() + + self.causality_axis = causality_axis + + # Ensure kernel_size and dilation are tuples + kernel_size = nn.modules.utils._pair(kernel_size) + dilation = nn.modules.utils._pair(dilation) + + # Calculate padding dimensions + pad_h = (kernel_size[0] - 1) * dilation[0] + pad_w = (kernel_size[1] - 1) * dilation[1] + + # The padding tuple for F.pad is (pad_left, pad_right, pad_top, pad_bottom) + match self.causality_axis: + case CausalityAxis.NONE: + self.padding = (pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2) + case CausalityAxis.WIDTH | CausalityAxis.WIDTH_COMPATIBILITY: + self.padding = (pad_w, 0, pad_h // 2, pad_h - pad_h // 2) + case CausalityAxis.HEIGHT: + self.padding = (pad_w // 2, pad_w - pad_w // 2, pad_h, 0) + case _: + raise ValueError(f"Invalid causality_axis: {causality_axis}") + + # The internal convolution layer uses no padding, as we handle it manually + self.conv = ops.Conv2d( + in_channels, + out_channels, + kernel_size, + stride=stride, + padding=0, + dilation=dilation, + groups=groups, + bias=bias, + ) + + def forward(self, x): + # Apply causal padding before convolution + x = F.pad(x, self.padding) + return self.conv(x) + + +def make_conv2d( + in_channels, + out_channels, + kernel_size, + stride=1, + padding=None, + dilation=1, + groups=1, + bias=True, + causality_axis: Optional[CausalityAxis] = None, +): + """ + Create a 2D convolution layer that can be either causal or non-causal. + + Args: + in_channels: Number of input channels + out_channels: Number of output channels + kernel_size: Size of the convolution kernel + stride: Convolution stride + padding: Padding (if None, will be calculated based on causal flag) + dilation: Dilation rate + groups: Number of groups for grouped convolution + bias: Whether to use bias + causality_axis: Dimension along which to apply causality. + + Returns: + Either a regular Conv2d or CausalConv2d layer + """ + if causality_axis is not None: + # For causal convolution, padding is handled internally by CausalConv2d + return CausalConv2d(in_channels, out_channels, kernel_size, stride, dilation, groups, bias, causality_axis) + else: + # For non-causal convolution, use symmetric padding if not specified + if padding is None: + if isinstance(kernel_size, int): + padding = kernel_size // 2 + else: + padding = tuple(k // 2 for k in kernel_size) + return ops.Conv2d( + in_channels, + out_channels, + kernel_size, + stride, + padding, + dilation, + groups, + bias, + ) + + +class Upsample(nn.Module): + def __init__(self, in_channels, with_conv, causality_axis: CausalityAxis = CausalityAxis.HEIGHT): + super().__init__() + self.with_conv = with_conv + self.causality_axis = causality_axis + if self.with_conv: + self.conv = make_conv2d(in_channels, in_channels, kernel_size=3, stride=1, causality_axis=causality_axis) + + def forward(self, x): + x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest") + if self.with_conv: + x = self.conv(x) + # Drop FIRST element in the causal axis to undo encoder's padding, while keeping the length 1 + 2 * n. + # For example, if the input is [0, 1, 2], after interpolation, the output is [0, 0, 1, 1, 2, 2]. + # The causal convolution will pad the first element as [-, -, 0, 0, 1, 1, 2, 2], + # So the output elements rely on the following windows: + # 0: [-,-,0] + # 1: [-,0,0] + # 2: [0,0,1] + # 3: [0,1,1] + # 4: [1,1,2] + # 5: [1,2,2] + # Notice that the first and second elements in the output rely only on the first element in the input, + # while all other elements rely on two elements in the input. + # So we can drop the first element to undo the padding (rather than the last element). + # This is a no-op for non-causal convolutions. + match self.causality_axis: + case CausalityAxis.NONE: + pass # x remains unchanged + case CausalityAxis.HEIGHT: + x = x[:, :, 1:, :] + case CausalityAxis.WIDTH: + x = x[:, :, :, 1:] + case CausalityAxis.WIDTH_COMPATIBILITY: + pass # x remains unchanged + case _: + raise ValueError(f"Invalid causality_axis: {self.causality_axis}") + + return x + + +class Downsample(nn.Module): + """ + A downsampling layer that can use either a strided convolution + or average pooling. Supports standard and causal padding for the + convolutional mode. + """ + + def __init__(self, in_channels, with_conv, causality_axis: CausalityAxis = CausalityAxis.WIDTH): + super().__init__() + self.with_conv = with_conv + self.causality_axis = causality_axis + + if self.causality_axis != CausalityAxis.NONE and not self.with_conv: + raise ValueError("causality is only supported when `with_conv=True`.") + + if self.with_conv: + # Do time downsampling here + # no asymmetric padding in torch conv, must do it ourselves + self.conv = ops.Conv2d(in_channels, in_channels, kernel_size=3, stride=2, padding=0) + + def forward(self, x): + if self.with_conv: + # (pad_left, pad_right, pad_top, pad_bottom) + match self.causality_axis: + case CausalityAxis.NONE: + pad = (0, 1, 0, 1) + case CausalityAxis.WIDTH: + pad = (2, 0, 0, 1) + case CausalityAxis.HEIGHT: + pad = (0, 1, 2, 0) + case CausalityAxis.WIDTH_COMPATIBILITY: + pad = (1, 0, 0, 1) + case _: + raise ValueError(f"Invalid causality_axis: {self.causality_axis}") + + x = torch.nn.functional.pad(x, pad, mode="constant", value=0) + x = self.conv(x) + else: + # This branch is only taken if with_conv=False, which implies causality_axis is NONE. + x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2) + + return x + + +class ResnetBlock(nn.Module): + def __init__( + self, + *, + in_channels, + out_channels=None, + conv_shortcut=False, + dropout, + temb_channels=512, + norm_type="group", + causality_axis: CausalityAxis = CausalityAxis.HEIGHT, + ): + super().__init__() + self.causality_axis = causality_axis + + if self.causality_axis != CausalityAxis.NONE and norm_type == "group": + raise ValueError("Causal ResnetBlock with GroupNorm is not supported.") + self.in_channels = in_channels + out_channels = in_channels if out_channels is None else out_channels + self.out_channels = out_channels + self.use_conv_shortcut = conv_shortcut + + self.norm1 = Normalize(in_channels, normtype=norm_type) + self.non_linearity = nn.SiLU() + self.conv1 = make_conv2d(in_channels, out_channels, kernel_size=3, stride=1, causality_axis=causality_axis) + if temb_channels > 0: + self.temb_proj = ops.Linear(temb_channels, out_channels) + self.norm2 = Normalize(out_channels, normtype=norm_type) + self.dropout = torch.nn.Dropout(dropout) + self.conv2 = make_conv2d(out_channels, out_channels, kernel_size=3, stride=1, causality_axis=causality_axis) + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + self.conv_shortcut = make_conv2d( + in_channels, out_channels, kernel_size=3, stride=1, causality_axis=causality_axis + ) + else: + self.nin_shortcut = make_conv2d( + in_channels, out_channels, kernel_size=1, stride=1, causality_axis=causality_axis + ) + + def forward(self, x, temb): + h = x + h = self.norm1(h) + h = self.non_linearity(h) + h = self.conv1(h) + + if temb is not None: + h = h + self.temb_proj(self.non_linearity(temb))[:, :, None, None] + + h = self.norm2(h) + h = self.non_linearity(h) + h = self.dropout(h) + h = self.conv2(h) + + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + x = self.conv_shortcut(x) + else: + x = self.nin_shortcut(x) + + return x + h + + +class AttnBlock(nn.Module): + def __init__(self, in_channels, norm_type="group"): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels, normtype=norm_type) + self.q = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0) + self.k = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0) + self.v = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0) + self.proj_out = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0) + + def forward(self, x): + h_ = x + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + b, c, h, w = q.shape + q = q.reshape(b, c, h * w).contiguous() + q = q.permute(0, 2, 1).contiguous() # b,hw,c + k = k.reshape(b, c, h * w).contiguous() # b,c,hw + w_ = torch.bmm(q, k).contiguous() # b,hw,hw w[b,i,j]=sum_c q[b,i,c]k[b,c,j] + w_ = w_ * (int(c) ** (-0.5)) + w_ = torch.nn.functional.softmax(w_, dim=2) + + # attend to values + v = v.reshape(b, c, h * w).contiguous() + w_ = w_.permute(0, 2, 1).contiguous() # b,hw,hw (first hw of k, second of q) + h_ = torch.bmm(v, w_).contiguous() # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j] + h_ = h_.reshape(b, c, h, w).contiguous() + + h_ = self.proj_out(h_) + + return x + h_ + + +def make_attn(in_channels, attn_type="vanilla", norm_type="group"): + # Convert string to enum if needed + attn_type = AttentionType.str_to_enum(attn_type) + + if attn_type != AttentionType.NONE: + logging.info(f"making attention of type '{attn_type.value}' with {in_channels} in_channels") + else: + logging.info(f"making identity attention with {in_channels} in_channels") + + match attn_type: + case AttentionType.VANILLA: + return AttnBlock(in_channels, norm_type=norm_type) + case AttentionType.NONE: + return nn.Identity(in_channels) + case AttentionType.LINEAR: + raise NotImplementedError(f"Attention type {attn_type.value} is not supported yet.") + case _: + raise ValueError(f"Unknown attention type: {attn_type}") + + +class Encoder(nn.Module): + def __init__( + self, + *, + ch, + out_ch, + ch_mult=(1, 2, 4, 8), + num_res_blocks, + attn_resolutions, + dropout=0.0, + resamp_with_conv=True, + in_channels, + resolution, + z_channels, + double_z=True, + attn_type="vanilla", + mid_block_add_attention=True, + norm_type="group", + causality_axis=CausalityAxis.WIDTH.value, + **ignore_kwargs, + ): + super().__init__() + self.ch = ch + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + self.z_channels = z_channels + self.double_z = double_z + self.norm_type = norm_type + # Convert string to enum if needed (for config loading) + causality_axis = CausalityAxis.str_to_enum(causality_axis) + self.attn_type = AttentionType.str_to_enum(attn_type) + + # downsampling + self.conv_in = make_conv2d( + in_channels, + self.ch, + kernel_size=3, + stride=1, + causality_axis=causality_axis, + ) + + self.non_linearity = nn.SiLU() + + curr_res = resolution + in_ch_mult = (1,) + tuple(ch_mult) + self.in_ch_mult = in_ch_mult + self.down = nn.ModuleList() + + for i_level in range(self.num_resolutions): + block = nn.ModuleList() + attn = nn.ModuleList() + block_in = ch * in_ch_mult[i_level] + block_out = ch * ch_mult[i_level] + + for _ in range(self.num_res_blocks): + block.append( + ResnetBlock( + in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout, + norm_type=self.norm_type, + causality_axis=causality_axis, + ) + ) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type)) + + down = nn.Module() + down.block = block + down.attn = attn + if i_level != self.num_resolutions - 1: + down.downsample = Downsample(block_in, resamp_with_conv, causality_axis=causality_axis) + curr_res = curr_res // 2 + self.down.append(down) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock( + in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + norm_type=self.norm_type, + causality_axis=causality_axis, + ) + if mid_block_add_attention: + self.mid.attn_1 = make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type) + else: + self.mid.attn_1 = nn.Identity() + self.mid.block_2 = ResnetBlock( + in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + norm_type=self.norm_type, + causality_axis=causality_axis, + ) + + # end + self.norm_out = Normalize(block_in, normtype=self.norm_type) + self.conv_out = make_conv2d( + block_in, + 2 * z_channels if double_z else z_channels, + kernel_size=3, + stride=1, + causality_axis=causality_axis, + ) + + def forward(self, x): + """ + Forward pass through the encoder. + + Args: + x: Input tensor of shape [batch, channels, time, n_mels] + + Returns: + Encoded latent representation + """ + feature_maps = [self.conv_in(x)] + + # Process each resolution level (from high to low resolution) + for resolution_level in range(self.num_resolutions): + # Apply residual blocks at current resolution level + for block_idx in range(self.num_res_blocks): + # Apply ResNet block with optional timestep embedding + current_features = self.down[resolution_level].block[block_idx](feature_maps[-1], temb=None) + + # Apply attention if configured for this resolution level + if len(self.down[resolution_level].attn) > 0: + current_features = self.down[resolution_level].attn[block_idx](current_features) + + # Store processed features + feature_maps.append(current_features) + + # Downsample spatial dimensions (except at the final resolution level) + if resolution_level != self.num_resolutions - 1: + downsampled_features = self.down[resolution_level].downsample(feature_maps[-1]) + feature_maps.append(downsampled_features) + + # === MIDDLE PROCESSING PHASE === + # Take the lowest resolution features for middle processing + bottleneck_features = feature_maps[-1] + + # Apply first middle ResNet block + bottleneck_features = self.mid.block_1(bottleneck_features, temb=None) + + # Apply middle attention block + bottleneck_features = self.mid.attn_1(bottleneck_features) + + # Apply second middle ResNet block + bottleneck_features = self.mid.block_2(bottleneck_features, temb=None) + + # === OUTPUT PHASE === + # Normalize the bottleneck features + output_features = self.norm_out(bottleneck_features) + + # Apply non-linearity (SiLU activation) + output_features = self.non_linearity(output_features) + + # Final convolution to produce latent representation + # [batch, channels, time, n_mels] -> [batch, 2 * z_channels if double_z else z_channels, time, n_mels] + return self.conv_out(output_features) + + +class Decoder(nn.Module): + def __init__( + self, + *, + ch, + out_ch, + ch_mult=(1, 2, 4, 8), + num_res_blocks, + attn_resolutions, + dropout=0.0, + resamp_with_conv=True, + in_channels, + resolution, + z_channels, + give_pre_end=False, + tanh_out=False, + attn_type="vanilla", + mid_block_add_attention=True, + norm_type="group", + causality_axis=CausalityAxis.WIDTH.value, + **ignorekwargs, + ): + super().__init__() + self.ch = ch + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + self.out_ch = out_ch + self.give_pre_end = give_pre_end + self.tanh_out = tanh_out + self.norm_type = norm_type + self.z_channels = z_channels + # Convert string to enum if needed (for config loading) + causality_axis = CausalityAxis.str_to_enum(causality_axis) + self.attn_type = AttentionType.str_to_enum(attn_type) + + # compute block_in and curr_res at lowest res + block_in = ch * ch_mult[self.num_resolutions - 1] + curr_res = resolution // 2 ** (self.num_resolutions - 1) + self.z_shape = (1, z_channels, curr_res, curr_res) + + # z to block_in + self.conv_in = make_conv2d(z_channels, block_in, kernel_size=3, stride=1, causality_axis=causality_axis) + + self.non_linearity = nn.SiLU() + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock( + in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + norm_type=self.norm_type, + causality_axis=causality_axis, + ) + if mid_block_add_attention: + self.mid.attn_1 = make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type) + else: + self.mid.attn_1 = nn.Identity() + self.mid.block_2 = ResnetBlock( + in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + norm_type=self.norm_type, + causality_axis=causality_axis, + ) + + # upsampling + self.up = nn.ModuleList() + for i_level in reversed(range(self.num_resolutions)): + block = nn.ModuleList() + attn = nn.ModuleList() + block_out = ch * ch_mult[i_level] + for _ in range(self.num_res_blocks + 1): + block.append( + ResnetBlock( + in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout, + norm_type=self.norm_type, + causality_axis=causality_axis, + ) + ) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type)) + up = nn.Module() + up.block = block + up.attn = attn + if i_level != 0: + up.upsample = Upsample(block_in, resamp_with_conv, causality_axis=causality_axis) + curr_res = curr_res * 2 + self.up.insert(0, up) # prepend to get consistent order + + # end + self.norm_out = Normalize(block_in, normtype=self.norm_type) + self.conv_out = make_conv2d(block_in, out_ch, kernel_size=3, stride=1, causality_axis=causality_axis) + + def _adjust_output_shape(self, decoded_output, target_shape): + """ + Adjust output shape to match target dimensions for variable-length audio. + + This function handles the common case where decoded audio spectrograms need to be + resized to match a specific target shape. + + Args: + decoded_output: Tensor of shape (batch, channels, time, frequency) + target_shape: Target shape tuple (batch, channels, time, frequency) + + Returns: + Tensor adjusted to match target_shape exactly + """ + # Current output shape: (batch, channels, time, frequency) + _, _, current_time, current_freq = decoded_output.shape + _, target_channels, target_time, target_freq = target_shape + + # Step 1: Crop first to avoid exceeding target dimensions + decoded_output = decoded_output[ + :, :target_channels, : min(current_time, target_time), : min(current_freq, target_freq) + ] + + # Step 2: Calculate padding needed for time and frequency dimensions + time_padding_needed = target_time - decoded_output.shape[2] + freq_padding_needed = target_freq - decoded_output.shape[3] + + # Step 3: Apply padding if needed + if time_padding_needed > 0 or freq_padding_needed > 0: + # PyTorch padding format: (pad_left, pad_right, pad_top, pad_bottom) + # For audio: pad_left/right = frequency, pad_top/bottom = time + padding = ( + 0, + max(freq_padding_needed, 0), # frequency padding (left, right) + 0, + max(time_padding_needed, 0), # time padding (top, bottom) + ) + decoded_output = F.pad(decoded_output, padding) + + # Step 4: Final safety crop to ensure exact target shape + decoded_output = decoded_output[:, :target_channels, :target_time, :target_freq] + + return decoded_output + + def get_config(self): + return { + "ch": self.ch, + "out_ch": self.out_ch, + "ch_mult": self.ch_mult, + "num_res_blocks": self.num_res_blocks, + "in_channels": self.in_channels, + "resolution": self.resolution, + "z_channels": self.z_channels, + } + + def forward(self, latent_features, target_shape=None): + """ + Decode latent features back to audio spectrograms. + + Args: + latent_features: Encoded latent representation of shape (batch, channels, height, width) + target_shape: Optional target output shape (batch, channels, time, frequency) + If provided, output will be cropped/padded to match this shape + + Returns: + Reconstructed audio spectrogram of shape (batch, channels, time, frequency) + """ + assert target_shape is not None, "Target shape is required for CausalAudioAutoencoder Decoder" + + # Transform latent features to decoder's internal feature dimension + hidden_features = self.conv_in(latent_features) + + # Middle processing + hidden_features = self.mid.block_1(hidden_features, temb=None) + hidden_features = self.mid.attn_1(hidden_features) + hidden_features = self.mid.block_2(hidden_features, temb=None) + + # Upsampling + # Progressively increase spatial resolution from lowest to highest + for resolution_level in reversed(range(self.num_resolutions)): + # Apply residual blocks at current resolution level + for block_index in range(self.num_res_blocks + 1): + hidden_features = self.up[resolution_level].block[block_index](hidden_features, temb=None) + + if len(self.up[resolution_level].attn) > 0: + hidden_features = self.up[resolution_level].attn[block_index](hidden_features) + + if resolution_level != 0: + hidden_features = self.up[resolution_level].upsample(hidden_features) + + # Output + if self.give_pre_end: + # Return intermediate features before final processing (for debugging/analysis) + decoded_output = hidden_features + else: + # Standard output path: normalize, activate, and convert to output channels + # Final normalization layer + hidden_features = self.norm_out(hidden_features) + + # Apply SiLU (Swish) activation function + hidden_features = self.non_linearity(hidden_features) + + # Final convolution to map to output channels (typically 2 for stereo audio) + decoded_output = self.conv_out(hidden_features) + + # Optional tanh activation to bound output values to [-1, 1] range + if self.tanh_out: + decoded_output = torch.tanh(decoded_output) + + # Adjust shape for audio data + if target_shape is not None: + decoded_output = self._adjust_output_shape(decoded_output, target_shape) + + return decoded_output + + +class processor(nn.Module): + def __init__(self): + super().__init__() + self.register_buffer("std-of-means", torch.empty(128)) + self.register_buffer("mean-of-means", torch.empty(128)) + + def un_normalize(self, x): + return (x * self.get_buffer("std-of-means").to(x)) + self.get_buffer("mean-of-means").to(x) + + def normalize(self, x): + return (x - self.get_buffer("mean-of-means").to(x)) / self.get_buffer("std-of-means").to(x) + + +class CausalAudioAutoencoder(nn.Module): + def __init__(self, config=None): + super().__init__() + + if config is None: + config = self._guess_config() + + # Extract encoder and decoder configs from the new format + model_config = config.get("model", {}).get("params", {}) + variables_config = config.get("variables", {}) + + self.sampling_rate = variables_config.get( + "sampling_rate", + model_config.get("sampling_rate", config.get("sampling_rate", 16000)), + ) + encoder_config = model_config.get("encoder", model_config.get("ddconfig", {})) + decoder_config = model_config.get("decoder", encoder_config) + + # Load mel spectrogram parameters + self.mel_bins = encoder_config.get("mel_bins", 64) + self.mel_hop_length = model_config.get("preprocessing", {}).get("stft", {}).get("hop_length", 160) + self.n_fft = model_config.get("preprocessing", {}).get("stft", {}).get("filter_length", 1024) + + # Store causality configuration at VAE level (not just in encoder internals) + causality_axis_value = encoder_config.get("causality_axis", CausalityAxis.WIDTH.value) + self.causality_axis = CausalityAxis.str_to_enum(causality_axis_value) + self.is_causal = self.causality_axis == CausalityAxis.HEIGHT + + self.encoder = Encoder(**encoder_config) + self.decoder = Decoder(**decoder_config) + + self.per_channel_statistics = processor() + + def _guess_config(self): + encoder_config = { + # Required parameters - based on ltx-video-av-1679000 model metadata + "ch": 128, + "out_ch": 8, + "ch_mult": [1, 2, 4], # Based on metadata: [1, 2, 4] not [1, 2, 4, 8] + "num_res_blocks": 2, + "attn_resolutions": [], # Based on metadata: empty list, no attention + "dropout": 0.0, + "resamp_with_conv": True, + "in_channels": 2, # stereo + "resolution": 256, + "z_channels": 8, + "double_z": True, + "attn_type": "vanilla", + "mid_block_add_attention": False, # Based on metadata: false + "norm_type": "pixel", + "causality_axis": "height", # Based on metadata + "mel_bins": 64, # Based on metadata: mel_bins = 64 + } + + decoder_config = { + # Inherits encoder config, can override specific params + **encoder_config, + "out_ch": 2, # Stereo audio output (2 channels) + "give_pre_end": False, + "tanh_out": False, + } + + config = { + "_class_name": "CausalAudioAutoencoder", + "sampling_rate": 16000, + "model": { + "params": { + "encoder": encoder_config, + "decoder": decoder_config, + } + }, + } + + return config + + def get_config(self): + return { + "sampling_rate": self.sampling_rate, + "mel_bins": self.mel_bins, + "mel_hop_length": self.mel_hop_length, + "n_fft": self.n_fft, + "causality_axis": self.causality_axis.value, + "is_causal": self.is_causal, + } + + def encode(self, x): + return self.encoder(x) + + def decode(self, x, target_shape=None): + return self.decoder(x, target_shape=target_shape) diff --git a/comfy/ldm/lightricks/vocoders/vocoder.py b/comfy/ldm/lightricks/vocoders/vocoder.py new file mode 100644 index 000000000..b1f15f2c5 --- /dev/null +++ b/comfy/ldm/lightricks/vocoders/vocoder.py @@ -0,0 +1,213 @@ +import torch +import torch.nn.functional as F +import torch.nn as nn +import comfy.ops +import numpy as np + +ops = comfy.ops.disable_weight_init + +LRELU_SLOPE = 0.1 + +def get_padding(kernel_size, dilation=1): + return int((kernel_size * dilation - dilation) / 2) + + +class ResBlock1(torch.nn.Module): + def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)): + super(ResBlock1, self).__init__() + self.convs1 = nn.ModuleList( + [ + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=dilation[0], + padding=get_padding(kernel_size, dilation[0]), + ), + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=dilation[1], + padding=get_padding(kernel_size, dilation[1]), + ), + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=dilation[2], + padding=get_padding(kernel_size, dilation[2]), + ), + ] + ) + + self.convs2 = nn.ModuleList( + [ + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=1, + padding=get_padding(kernel_size, 1), + ), + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=1, + padding=get_padding(kernel_size, 1), + ), + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=1, + padding=get_padding(kernel_size, 1), + ), + ] + ) + + def forward(self, x): + for c1, c2 in zip(self.convs1, self.convs2): + xt = F.leaky_relu(x, LRELU_SLOPE) + xt = c1(xt) + xt = F.leaky_relu(xt, LRELU_SLOPE) + xt = c2(xt) + x = xt + x + return x + + +class ResBlock2(torch.nn.Module): + def __init__(self, channels, kernel_size=3, dilation=(1, 3)): + super(ResBlock2, self).__init__() + self.convs = nn.ModuleList( + [ + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=dilation[0], + padding=get_padding(kernel_size, dilation[0]), + ), + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=dilation[1], + padding=get_padding(kernel_size, dilation[1]), + ), + ] + ) + + def forward(self, x): + for c in self.convs: + xt = F.leaky_relu(x, LRELU_SLOPE) + xt = c(xt) + x = xt + x + return x + + +class Vocoder(torch.nn.Module): + """ + Vocoder model for synthesizing audio from spectrograms, based on: https://github.com/jik876/hifi-gan. + + """ + + def __init__(self, config=None): + super(Vocoder, self).__init__() + + if config is None: + config = self.get_default_config() + + resblock_kernel_sizes = config.get("resblock_kernel_sizes", [3, 7, 11]) + upsample_rates = config.get("upsample_rates", [6, 5, 2, 2, 2]) + upsample_kernel_sizes = config.get("upsample_kernel_sizes", [16, 15, 8, 4, 4]) + resblock_dilation_sizes = config.get("resblock_dilation_sizes", [[1, 3, 5], [1, 3, 5], [1, 3, 5]]) + upsample_initial_channel = config.get("upsample_initial_channel", 1024) + stereo = config.get("stereo", True) + resblock = config.get("resblock", "1") + + self.output_sample_rate = config.get("output_sample_rate") + self.num_kernels = len(resblock_kernel_sizes) + self.num_upsamples = len(upsample_rates) + in_channels = 128 if stereo else 64 + self.conv_pre = ops.Conv1d(in_channels, upsample_initial_channel, 7, 1, padding=3) + resblock_class = ResBlock1 if resblock == "1" else ResBlock2 + + self.ups = nn.ModuleList() + for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)): + self.ups.append( + ops.ConvTranspose1d( + upsample_initial_channel // (2**i), + upsample_initial_channel // (2 ** (i + 1)), + k, + u, + padding=(k - u) // 2, + ) + ) + + self.resblocks = nn.ModuleList() + for i in range(len(self.ups)): + ch = upsample_initial_channel // (2 ** (i + 1)) + for _, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)): + self.resblocks.append(resblock_class(ch, k, d)) + + out_channels = 2 if stereo else 1 + self.conv_post = ops.Conv1d(ch, out_channels, 7, 1, padding=3) + + self.upsample_factor = np.prod([self.ups[i].stride[0] for i in range(len(self.ups))]) + + def get_default_config(self): + """Generate default configuration for the vocoder.""" + + config = { + "resblock_kernel_sizes": [3, 7, 11], + "upsample_rates": [6, 5, 2, 2, 2], + "upsample_kernel_sizes": [16, 15, 8, 4, 4], + "resblock_dilation_sizes": [[1, 3, 5], [1, 3, 5], [1, 3, 5]], + "upsample_initial_channel": 1024, + "stereo": True, + "resblock": "1", + } + + return config + + def forward(self, x): + """ + Forward pass of the vocoder. + + Args: + x: Input spectrogram tensor. Can be: + - 3D: (batch_size, channels, time_steps) for mono + - 4D: (batch_size, 2, channels, time_steps) for stereo + + Returns: + Audio tensor of shape (batch_size, out_channels, audio_length) + """ + if x.dim() == 4: # stereo + assert x.shape[1] == 2, "Input must have 2 channels for stereo" + x = torch.cat((x[:, 0, :, :], x[:, 1, :, :]), dim=1) + x = self.conv_pre(x) + for i in range(self.num_upsamples): + x = F.leaky_relu(x, LRELU_SLOPE) + x = self.ups[i](x) + xs = None + for j in range(self.num_kernels): + if xs is None: + xs = self.resblocks[i * self.num_kernels + j](x) + else: + xs += self.resblocks[i * self.num_kernels + j](x) + x = xs / self.num_kernels + x = F.leaky_relu(x) + x = self.conv_post(x) + x = torch.tanh(x) + + return x diff --git a/comfy/ldm/lumina/controlnet.py b/comfy/ldm/lumina/controlnet.py new file mode 100644 index 000000000..8e2de7977 --- /dev/null +++ b/comfy/ldm/lumina/controlnet.py @@ -0,0 +1,160 @@ +import torch +from torch import nn + +from .model import JointTransformerBlock + +class ZImageControlTransformerBlock(JointTransformerBlock): + def __init__( + self, + layer_id: int, + dim: int, + n_heads: int, + n_kv_heads: int, + multiple_of: int, + ffn_dim_multiplier: float, + norm_eps: float, + qk_norm: bool, + modulation=True, + block_id=0, + operation_settings=None, + ): + super().__init__(layer_id, dim, n_heads, n_kv_heads, multiple_of, ffn_dim_multiplier, norm_eps, qk_norm, modulation, z_image_modulation=True, operation_settings=operation_settings) + self.block_id = block_id + if block_id == 0: + self.before_proj = operation_settings.get("operations").Linear(self.dim, self.dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.after_proj = operation_settings.get("operations").Linear(self.dim, self.dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + + def forward(self, c, x, **kwargs): + if self.block_id == 0: + c = self.before_proj(c) + x + c = super().forward(c, **kwargs) + c_skip = self.after_proj(c) + return c_skip, c + +class ZImage_Control(torch.nn.Module): + def __init__( + self, + dim: int = 3840, + n_heads: int = 30, + n_kv_heads: int = 30, + multiple_of: int = 256, + ffn_dim_multiplier: float = (8.0 / 3.0), + norm_eps: float = 1e-5, + qk_norm: bool = True, + n_control_layers=6, + control_in_dim=16, + additional_in_dim=0, + broken=False, + refiner_control=False, + dtype=None, + device=None, + operations=None, + **kwargs + ): + super().__init__() + operation_settings = {"operations": operations, "device": device, "dtype": dtype} + + self.broken = broken + self.additional_in_dim = additional_in_dim + self.control_in_dim = control_in_dim + n_refiner_layers = 2 + self.n_control_layers = n_control_layers + self.control_layers = nn.ModuleList( + [ + ZImageControlTransformerBlock( + i, + dim, + n_heads, + n_kv_heads, + multiple_of, + ffn_dim_multiplier, + norm_eps, + qk_norm, + block_id=i, + operation_settings=operation_settings, + ) + for i in range(self.n_control_layers) + ] + ) + + all_x_embedder = {} + patch_size = 2 + f_patch_size = 1 + x_embedder = operations.Linear(f_patch_size * patch_size * patch_size * (self.control_in_dim + self.additional_in_dim), dim, bias=True, device=device, dtype=dtype) + all_x_embedder[f"{patch_size}-{f_patch_size}"] = x_embedder + + self.refiner_control = refiner_control + + self.control_all_x_embedder = nn.ModuleDict(all_x_embedder) + if self.refiner_control: + self.control_noise_refiner = nn.ModuleList( + [ + ZImageControlTransformerBlock( + layer_id, + dim, + n_heads, + n_kv_heads, + multiple_of, + ffn_dim_multiplier, + norm_eps, + qk_norm, + block_id=layer_id, + operation_settings=operation_settings, + ) + for layer_id in range(n_refiner_layers) + ] + ) + else: + self.control_noise_refiner = nn.ModuleList( + [ + JointTransformerBlock( + layer_id, + dim, + n_heads, + n_kv_heads, + multiple_of, + ffn_dim_multiplier, + norm_eps, + qk_norm, + modulation=True, + z_image_modulation=True, + operation_settings=operation_settings, + ) + for layer_id in range(n_refiner_layers) + ] + ) + + def forward(self, cap_feats, control_context, x_freqs_cis, adaln_input): + patch_size = 2 + f_patch_size = 1 + pH = pW = patch_size + B, C, H, W = control_context.shape + control_context = self.control_all_x_embedder[f"{patch_size}-{f_patch_size}"](control_context.view(B, C, H // pH, pH, W // pW, pW).permute(0, 2, 4, 3, 5, 1).flatten(3).flatten(1, 2)) + + x_attn_mask = None + if not self.refiner_control: + for layer in self.control_noise_refiner: + control_context = layer(control_context, x_attn_mask, x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input) + + return control_context + + def forward_noise_refiner_block(self, layer_id, control_context, x, x_attn_mask, x_freqs_cis, adaln_input): + if self.refiner_control: + if self.broken: + if layer_id == 0: + return self.control_layers[layer_id](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input) + if layer_id > 0: + out = None + for i in range(1, len(self.control_layers)): + o, control_context = self.control_layers[i](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input) + if out is None: + out = o + + return (out, control_context) + else: + return self.control_noise_refiner[layer_id](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input) + else: + return (None, control_context) + + def forward_control_block(self, layer_id, control_context, x, x_attn_mask, x_freqs_cis, adaln_input): + return self.control_layers[layer_id](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input) diff --git a/comfy/ldm/lumina/model.py b/comfy/ldm/lumina/model.py index b4494a51d..afbab2ac7 100644 --- a/comfy/ldm/lumina/model.py +++ b/comfy/ldm/lumina/model.py @@ -11,6 +11,7 @@ import comfy.ldm.common_dit from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder from comfy.ldm.modules.attention import optimized_attention_masked from comfy.ldm.flux.layers import EmbedND +from comfy.ldm.flux.math import apply_rope import comfy.patcher_extension @@ -21,6 +22,10 @@ def modulate(x, scale): # Core NextDiT Model # ############################################################################# +def clamp_fp16(x): + if x.dtype == torch.float16: + return torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504) + return x class JointAttention(nn.Module): """Multi-head attention module.""" @@ -31,6 +36,7 @@ class JointAttention(nn.Module): n_heads: int, n_kv_heads: Optional[int], qk_norm: bool, + out_bias: bool = False, operation_settings={}, ): """ @@ -59,7 +65,7 @@ class JointAttention(nn.Module): self.out = operation_settings.get("operations").Linear( n_heads * self.head_dim, dim, - bias=False, + bias=out_bias, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"), ) @@ -70,35 +76,6 @@ class JointAttention(nn.Module): else: self.q_norm = self.k_norm = nn.Identity() - @staticmethod - def apply_rotary_emb( - x_in: torch.Tensor, - freqs_cis: torch.Tensor, - ) -> torch.Tensor: - """ - Apply rotary embeddings to input tensors using the given frequency - tensor. - - This function applies rotary embeddings to the given query 'xq' and - key 'xk' tensors using the provided frequency tensor 'freqs_cis'. The - input tensors are reshaped as complex numbers, and the frequency tensor - is reshaped for broadcasting compatibility. The resulting tensors - contain rotary embeddings and are returned as real tensors. - - Args: - x_in (torch.Tensor): Query or Key tensor to apply rotary embeddings. - freqs_cis (torch.Tensor): Precomputed frequency tensor for complex - exponentials. - - Returns: - Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor - and key tensor with rotary embeddings. - """ - - t_ = x_in.reshape(*x_in.shape[:-1], -1, 1, 2) - t_out = freqs_cis[..., 0] * t_[..., 0] + freqs_cis[..., 1] * t_[..., 1] - return t_out.reshape(*x_in.shape) - def forward( self, x: torch.Tensor, @@ -134,8 +111,7 @@ class JointAttention(nn.Module): xq = self.q_norm(xq) xk = self.k_norm(xk) - xq = JointAttention.apply_rotary_emb(xq, freqs_cis=freqs_cis) - xk = JointAttention.apply_rotary_emb(xk, freqs_cis=freqs_cis) + xq, xk = apply_rope(xq, xk, freqs_cis) n_rep = self.n_local_heads // self.n_local_kv_heads if n_rep >= 1: @@ -197,7 +173,7 @@ class FeedForward(nn.Module): # @torch.compile def _forward_silu_gating(self, x1, x3): - return F.silu(x1) * x3 + return clamp_fp16(F.silu(x1) * x3) def forward(self, x): return self.w2(self._forward_silu_gating(self.w1(x), self.w3(x))) @@ -215,6 +191,8 @@ class JointTransformerBlock(nn.Module): norm_eps: float, qk_norm: bool, modulation=True, + z_image_modulation=False, + attn_out_bias=False, operation_settings={}, ) -> None: """ @@ -235,10 +213,10 @@ class JointTransformerBlock(nn.Module): super().__init__() self.dim = dim self.head_dim = dim // n_heads - self.attention = JointAttention(dim, n_heads, n_kv_heads, qk_norm, operation_settings=operation_settings) + self.attention = JointAttention(dim, n_heads, n_kv_heads, qk_norm, out_bias=attn_out_bias, operation_settings=operation_settings) self.feed_forward = FeedForward( dim=dim, - hidden_dim=4 * dim, + hidden_dim=dim, multiple_of=multiple_of, ffn_dim_multiplier=ffn_dim_multiplier, operation_settings=operation_settings, @@ -252,16 +230,27 @@ class JointTransformerBlock(nn.Module): self.modulation = modulation if modulation: - self.adaLN_modulation = nn.Sequential( - nn.SiLU(), - operation_settings.get("operations").Linear( - min(dim, 1024), - 4 * dim, - bias=True, - device=operation_settings.get("device"), - dtype=operation_settings.get("dtype"), - ), - ) + if z_image_modulation: + self.adaLN_modulation = nn.Sequential( + operation_settings.get("operations").Linear( + min(dim, 256), + 4 * dim, + bias=True, + device=operation_settings.get("device"), + dtype=operation_settings.get("dtype"), + ), + ) + else: + self.adaLN_modulation = nn.Sequential( + nn.SiLU(), + operation_settings.get("operations").Linear( + min(dim, 1024), + 4 * dim, + bias=True, + device=operation_settings.get("device"), + dtype=operation_settings.get("dtype"), + ), + ) def forward( self, @@ -288,27 +277,27 @@ class JointTransformerBlock(nn.Module): scale_msa, gate_msa, scale_mlp, gate_mlp = self.adaLN_modulation(adaln_input).chunk(4, dim=1) x = x + gate_msa.unsqueeze(1).tanh() * self.attention_norm2( - self.attention( + clamp_fp16(self.attention( modulate(self.attention_norm1(x), scale_msa), x_mask, freqs_cis, transformer_options=transformer_options, - ) + )) ) x = x + gate_mlp.unsqueeze(1).tanh() * self.ffn_norm2( - self.feed_forward( + clamp_fp16(self.feed_forward( modulate(self.ffn_norm1(x), scale_mlp), - ) + )) ) else: assert adaln_input is None x = x + self.attention_norm2( - self.attention( + clamp_fp16(self.attention( self.attention_norm1(x), x_mask, freqs_cis, transformer_options=transformer_options, - ) + )) ) x = x + self.ffn_norm2( self.feed_forward( @@ -323,7 +312,7 @@ class FinalLayer(nn.Module): The final layer of NextDiT. """ - def __init__(self, hidden_size, patch_size, out_channels, operation_settings={}): + def __init__(self, hidden_size, patch_size, out_channels, z_image_modulation=False, operation_settings={}): super().__init__() self.norm_final = operation_settings.get("operations").LayerNorm( hidden_size, @@ -340,10 +329,15 @@ class FinalLayer(nn.Module): dtype=operation_settings.get("dtype"), ) + if z_image_modulation: + min_mod = 256 + else: + min_mod = 1024 + self.adaLN_modulation = nn.Sequential( nn.SiLU(), operation_settings.get("operations").Linear( - min(hidden_size, 1024), + min(hidden_size, min_mod), hidden_size, bias=True, device=operation_settings.get("device"), @@ -373,12 +367,17 @@ class NextDiT(nn.Module): n_heads: int = 32, n_kv_heads: Optional[int] = None, multiple_of: int = 256, - ffn_dim_multiplier: Optional[float] = None, + ffn_dim_multiplier: float = 4.0, norm_eps: float = 1e-5, qk_norm: bool = False, cap_feat_dim: int = 5120, axes_dims: List[int] = (16, 56, 56), axes_lens: List[int] = (1, 512, 512), + rope_theta=10000.0, + z_image_modulation=False, + time_scale=1.0, + pad_tokens_multiple=None, + clip_text_dim=None, image_model=None, device=None, dtype=None, @@ -390,6 +389,8 @@ class NextDiT(nn.Module): self.in_channels = in_channels self.out_channels = in_channels self.patch_size = patch_size + self.time_scale = time_scale + self.pad_tokens_multiple = pad_tokens_multiple self.x_embedder = operation_settings.get("operations").Linear( in_features=patch_size * patch_size * in_channels, @@ -411,6 +412,7 @@ class NextDiT(nn.Module): norm_eps, qk_norm, modulation=True, + z_image_modulation=z_image_modulation, operation_settings=operation_settings, ) for layer_id in range(n_refiner_layers) @@ -434,7 +436,7 @@ class NextDiT(nn.Module): ] ) - self.t_embedder = TimestepEmbedder(min(dim, 1024), **operation_settings) + self.t_embedder = TimestepEmbedder(min(dim, 1024), output_size=256 if z_image_modulation else None, **operation_settings) self.cap_embedder = nn.Sequential( operation_settings.get("operations").RMSNorm(cap_feat_dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")), operation_settings.get("operations").Linear( @@ -446,6 +448,31 @@ class NextDiT(nn.Module): ), ) + self.clip_text_pooled_proj = None + + if clip_text_dim is not None: + self.clip_text_dim = clip_text_dim + self.clip_text_pooled_proj = nn.Sequential( + operation_settings.get("operations").RMSNorm(clip_text_dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")), + operation_settings.get("operations").Linear( + clip_text_dim, + clip_text_dim, + bias=True, + device=operation_settings.get("device"), + dtype=operation_settings.get("dtype"), + ), + ) + self.time_text_embed = nn.Sequential( + nn.SiLU(), + operation_settings.get("operations").Linear( + min(dim, 1024) + clip_text_dim, + min(dim, 1024), + bias=True, + device=operation_settings.get("device"), + dtype=operation_settings.get("dtype"), + ), + ) + self.layers = nn.ModuleList( [ JointTransformerBlock( @@ -457,18 +484,25 @@ class NextDiT(nn.Module): ffn_dim_multiplier, norm_eps, qk_norm, + z_image_modulation=z_image_modulation, + attn_out_bias=False, operation_settings=operation_settings, ) for layer_id in range(n_layers) ] ) - self.norm_final = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) - self.final_layer = FinalLayer(dim, patch_size, self.out_channels, operation_settings=operation_settings) + # This norm final is in the lumina 2.0 code but isn't actually used for anything. + # self.norm_final = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.final_layer = FinalLayer(dim, patch_size, self.out_channels, z_image_modulation=z_image_modulation, operation_settings=operation_settings) + + if self.pad_tokens_multiple is not None: + self.x_pad_token = nn.Parameter(torch.empty((1, dim), device=device, dtype=dtype)) + self.cap_pad_token = nn.Parameter(torch.empty((1, dim), device=device, dtype=dtype)) assert (dim // n_heads) == sum(axes_dims) self.axes_dims = axes_dims self.axes_lens = axes_lens - self.rope_embedder = EmbedND(dim=dim // n_heads, theta=10000.0, axes_dim=axes_dims) + self.rope_embedder = EmbedND(dim=dim // n_heads, theta=rope_theta, axes_dim=axes_dims) self.dim = dim self.n_heads = n_heads @@ -503,108 +537,63 @@ class NextDiT(nn.Module): bsz = len(x) pH = pW = self.patch_size device = x[0].device - dtype = x[0].dtype + orig_x = x - if cap_mask is not None: - l_effective_cap_len = cap_mask.sum(dim=1).tolist() - else: - l_effective_cap_len = [num_tokens] * bsz + if self.pad_tokens_multiple is not None: + pad_extra = (-cap_feats.shape[1]) % self.pad_tokens_multiple + cap_feats = torch.cat((cap_feats, self.cap_pad_token.to(device=cap_feats.device, dtype=cap_feats.dtype, copy=True).unsqueeze(0).repeat(cap_feats.shape[0], pad_extra, 1)), dim=1) - if cap_mask is not None and not torch.is_floating_point(cap_mask): - cap_mask = (cap_mask - 1).to(dtype) * torch.finfo(dtype).max + cap_pos_ids = torch.zeros(bsz, cap_feats.shape[1], 3, dtype=torch.float32, device=device) + cap_pos_ids[:, :, 0] = torch.arange(cap_feats.shape[1], dtype=torch.float32, device=device) + 1.0 - img_sizes = [(img.size(1), img.size(2)) for img in x] - l_effective_img_len = [(H // pH) * (W // pW) for (H, W) in img_sizes] + B, C, H, W = x.shape + x = self.x_embedder(x.view(B, C, H // pH, pH, W // pW, pW).permute(0, 2, 4, 3, 5, 1).flatten(3).flatten(1, 2)) - max_seq_len = max( - (cap_len+img_len for cap_len, img_len in zip(l_effective_cap_len, l_effective_img_len)) - ) - max_cap_len = max(l_effective_cap_len) - max_img_len = max(l_effective_img_len) + rope_options = transformer_options.get("rope_options", None) + h_scale = 1.0 + w_scale = 1.0 + h_start = 0 + w_start = 0 + if rope_options is not None: + h_scale = rope_options.get("scale_y", 1.0) + w_scale = rope_options.get("scale_x", 1.0) - position_ids = torch.zeros(bsz, max_seq_len, 3, dtype=torch.float32, device=device) + h_start = rope_options.get("shift_y", 0.0) + w_start = rope_options.get("shift_x", 0.0) - for i in range(bsz): - cap_len = l_effective_cap_len[i] - img_len = l_effective_img_len[i] - H, W = img_sizes[i] - H_tokens, W_tokens = H // pH, W // pW - assert H_tokens * W_tokens == img_len + H_tokens, W_tokens = H // pH, W // pW + x_pos_ids = torch.zeros((bsz, x.shape[1], 3), dtype=torch.float32, device=device) + x_pos_ids[:, :, 0] = cap_feats.shape[1] + 1 + x_pos_ids[:, :, 1] = (torch.arange(H_tokens, dtype=torch.float32, device=device) * h_scale + h_start).view(-1, 1).repeat(1, W_tokens).flatten() + x_pos_ids[:, :, 2] = (torch.arange(W_tokens, dtype=torch.float32, device=device) * w_scale + w_start).view(1, -1).repeat(H_tokens, 1).flatten() - rope_options = transformer_options.get("rope_options", None) - h_scale = 1.0 - w_scale = 1.0 - h_start = 0 - w_start = 0 - if rope_options is not None: - h_scale = rope_options.get("scale_y", 1.0) - w_scale = rope_options.get("scale_x", 1.0) + if self.pad_tokens_multiple is not None: + pad_extra = (-x.shape[1]) % self.pad_tokens_multiple + x = torch.cat((x, self.x_pad_token.to(device=x.device, dtype=x.dtype, copy=True).unsqueeze(0).repeat(x.shape[0], pad_extra, 1)), dim=1) + x_pos_ids = torch.nn.functional.pad(x_pos_ids, (0, 0, 0, pad_extra)) - h_start = rope_options.get("shift_y", 0.0) - w_start = rope_options.get("shift_x", 0.0) + freqs_cis = self.rope_embedder(torch.cat((cap_pos_ids, x_pos_ids), dim=1)).movedim(1, 2) - position_ids[i, :cap_len, 0] = torch.arange(cap_len, dtype=torch.float32, device=device) - position_ids[i, cap_len:cap_len+img_len, 0] = cap_len - row_ids = (torch.arange(H_tokens, dtype=torch.float32, device=device) * h_scale + h_start).view(-1, 1).repeat(1, W_tokens).flatten() - col_ids = (torch.arange(W_tokens, dtype=torch.float32, device=device) * w_scale + w_start).view(1, -1).repeat(H_tokens, 1).flatten() - position_ids[i, cap_len:cap_len+img_len, 1] = row_ids - position_ids[i, cap_len:cap_len+img_len, 2] = col_ids - - freqs_cis = self.rope_embedder(position_ids).movedim(1, 2).to(dtype) - - # build freqs_cis for cap and image individually - cap_freqs_cis_shape = list(freqs_cis.shape) - # cap_freqs_cis_shape[1] = max_cap_len - cap_freqs_cis_shape[1] = cap_feats.shape[1] - cap_freqs_cis = torch.zeros(*cap_freqs_cis_shape, device=device, dtype=freqs_cis.dtype) - - img_freqs_cis_shape = list(freqs_cis.shape) - img_freqs_cis_shape[1] = max_img_len - img_freqs_cis = torch.zeros(*img_freqs_cis_shape, device=device, dtype=freqs_cis.dtype) - - for i in range(bsz): - cap_len = l_effective_cap_len[i] - img_len = l_effective_img_len[i] - cap_freqs_cis[i, :cap_len] = freqs_cis[i, :cap_len] - img_freqs_cis[i, :img_len] = freqs_cis[i, cap_len:cap_len+img_len] + patches = transformer_options.get("patches", {}) # refine context for layer in self.context_refiner: - cap_feats = layer(cap_feats, cap_mask, cap_freqs_cis, transformer_options=transformer_options) + cap_feats = layer(cap_feats, cap_mask, freqs_cis[:, :cap_pos_ids.shape[1]], transformer_options=transformer_options) - # refine image - flat_x = [] - for i in range(bsz): - img = x[i] - C, H, W = img.size() - img = img.view(C, H // pH, pH, W // pW, pW).permute(1, 3, 2, 4, 0).flatten(2).flatten(0, 1) - flat_x.append(img) - x = flat_x - padded_img_embed = torch.zeros(bsz, max_img_len, x[0].shape[-1], device=device, dtype=x[0].dtype) - padded_img_mask = torch.zeros(bsz, max_img_len, dtype=dtype, device=device) - for i in range(bsz): - padded_img_embed[i, :l_effective_img_len[i]] = x[i] - padded_img_mask[i, l_effective_img_len[i]:] = -torch.finfo(dtype).max - - padded_img_embed = self.x_embedder(padded_img_embed) - padded_img_mask = padded_img_mask.unsqueeze(1) - for layer in self.noise_refiner: - padded_img_embed = layer(padded_img_embed, padded_img_mask, img_freqs_cis, t, transformer_options=transformer_options) - - if cap_mask is not None: - mask = torch.zeros(bsz, max_seq_len, dtype=dtype, device=device) - mask[:, :max_cap_len] = cap_mask[:, :max_cap_len] - else: - mask = None - - padded_full_embed = torch.zeros(bsz, max_seq_len, self.dim, device=device, dtype=x[0].dtype) - for i in range(bsz): - cap_len = l_effective_cap_len[i] - img_len = l_effective_img_len[i] - - padded_full_embed[i, :cap_len] = cap_feats[i, :cap_len] - padded_full_embed[i, cap_len:cap_len+img_len] = padded_img_embed[i, :img_len] + padded_img_mask = None + x_input = x + for i, layer in enumerate(self.noise_refiner): + x = layer(x, padded_img_mask, freqs_cis[:, cap_pos_ids.shape[1]:], t, transformer_options=transformer_options) + if "noise_refiner" in patches: + for p in patches["noise_refiner"]: + out = p({"img": x, "img_input": x_input, "txt": cap_feats, "pe": freqs_cis[:, cap_pos_ids.shape[1]:], "vec": t, "x": orig_x, "block_index": i, "transformer_options": transformer_options, "block_type": "noise_refiner"}) + if "img" in out: + x = out["img"] + padded_full_embed = torch.cat((cap_feats, x), dim=1) + mask = None + img_sizes = [(H, W)] * bsz + l_effective_cap_len = [cap_feats.shape[1]] * bsz return padded_full_embed, mask, img_sizes, l_effective_cap_len, freqs_cis def forward(self, x, timesteps, context, num_tokens, attention_mask=None, **kwargs): @@ -615,7 +604,7 @@ class NextDiT(nn.Module): ).execute(x, timesteps, context, num_tokens, attention_mask, **kwargs) # def forward(self, x, t, cap_feats, cap_mask): - def _forward(self, x, timesteps, context, num_tokens, attention_mask=None, **kwargs): + def _forward(self, x, timesteps, context, num_tokens, attention_mask=None, transformer_options={}, **kwargs): t = 1.0 - timesteps cap_feats = context cap_mask = attention_mask @@ -627,21 +616,41 @@ class NextDiT(nn.Module): y: (N,) tensor of text tokens/features """ - t = self.t_embedder(t, dtype=x.dtype) # (N, D) + t = self.t_embedder(t * self.time_scale, dtype=x.dtype) # (N, D) adaln_input = t cap_feats = self.cap_embedder(cap_feats) # (N, L, D) # todo check if able to batchify w.o. redundant compute - transformer_options = kwargs.get("transformer_options", {}) + if self.clip_text_pooled_proj is not None: + pooled = kwargs.get("clip_text_pooled", None) + if pooled is not None: + pooled = self.clip_text_pooled_proj(pooled) + else: + pooled = torch.zeros((x.shape[0], self.clip_text_dim), device=x.device, dtype=x.dtype) + + adaln_input = self.time_text_embed(torch.cat((t, pooled), dim=-1)) + + patches = transformer_options.get("patches", {}) x_is_tensor = isinstance(x, torch.Tensor) - x, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, t, num_tokens, transformer_options=transformer_options) - freqs_cis = freqs_cis.to(x.device) + img, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, adaln_input, num_tokens, transformer_options=transformer_options) + freqs_cis = freqs_cis.to(img.device) - for layer in self.layers: - x = layer(x, mask, freqs_cis, adaln_input, transformer_options=transformer_options) + transformer_options["total_blocks"] = len(self.layers) + transformer_options["block_type"] = "double" + img_input = img + for i, layer in enumerate(self.layers): + transformer_options["block_index"] = i + img = layer(img, mask, freqs_cis, adaln_input, transformer_options=transformer_options) + if "double_block" in patches: + for p in patches["double_block"]: + out = p({"img": img[:, cap_size[0]:], "img_input": img_input[:, cap_size[0]:], "txt": img[:, :cap_size[0]], "pe": freqs_cis[:, cap_size[0]:], "vec": adaln_input, "x": x, "block_index": i, "transformer_options": transformer_options}) + if "img" in out: + img[:, cap_size[0]:] = out["img"] + if "txt" in out: + img[:, :cap_size[0]] = out["txt"] - x = self.final_layer(x, adaln_input) - x = self.unpatchify(x, img_size, cap_size, return_tensor=x_is_tensor)[:,:,:h,:w] + img = self.final_layer(img, adaln_input) + img = self.unpatchify(img, img_size, cap_size, return_tensor=x_is_tensor)[:, :, :h, :w] - return -x + return -img diff --git a/comfy/ldm/models/autoencoder.py b/comfy/ldm/models/autoencoder.py index 611d36a1b..4f50810dc 100644 --- a/comfy/ldm/models/autoencoder.py +++ b/comfy/ldm/models/autoencoder.py @@ -9,6 +9,8 @@ from comfy.ldm.modules.distributions.distributions import DiagonalGaussianDistri from comfy.ldm.util import get_obj_from_str, instantiate_from_config from comfy.ldm.modules.ema import LitEma import comfy.ops +from einops import rearrange +import comfy.model_management class DiagonalGaussianRegularizer(torch.nn.Module): def __init__(self, sample: bool = False): @@ -179,6 +181,21 @@ class AutoencodingEngineLegacy(AutoencodingEngine): self.post_quant_conv = conv_op(embed_dim, ddconfig["z_channels"], 1) self.embed_dim = embed_dim + if ddconfig.get("batch_norm_latent", False): + self.bn_eps = 1e-4 + self.bn_momentum = 0.1 + self.ps = [2, 2] + self.bn = torch.nn.BatchNorm2d(math.prod(self.ps) * ddconfig["z_channels"], + eps=self.bn_eps, + momentum=self.bn_momentum, + affine=False, + track_running_stats=True, + ) + self.bn.eval() + else: + self.bn = None + + def get_autoencoder_params(self) -> list: params = super().get_autoencoder_params() return params @@ -201,11 +218,36 @@ class AutoencodingEngineLegacy(AutoencodingEngine): z = torch.cat(z, 0) z, reg_log = self.regularization(z) + + if self.bn is not None: + z = rearrange(z, + "... c (i pi) (j pj) -> ... (c pi pj) i j", + pi=self.ps[0], + pj=self.ps[1], + ) + + z = torch.nn.functional.batch_norm(z, + comfy.model_management.cast_to(self.bn.running_mean, dtype=z.dtype, device=z.device), + comfy.model_management.cast_to(self.bn.running_var, dtype=z.dtype, device=z.device), + momentum=self.bn_momentum, + eps=self.bn_eps) + if return_reg_log: return z, reg_log return z def decode(self, z: torch.Tensor, **decoder_kwargs) -> torch.Tensor: + if self.bn is not None: + s = torch.sqrt(comfy.model_management.cast_to(self.bn.running_var.view(1, -1, 1, 1), dtype=z.dtype, device=z.device) + self.bn_eps) + m = comfy.model_management.cast_to(self.bn.running_mean.view(1, -1, 1, 1), dtype=z.dtype, device=z.device) + z = z * s + m + z = rearrange( + z, + "... (c pi pj) i j -> ... c (i pi) (j pj)", + pi=self.ps[0], + pj=self.ps[1], + ) + if self.max_batch_size is None: dec = self.post_quant_conv(z) dec = self.decoder(dec, **decoder_kwargs) diff --git a/comfy/ldm/modules/attention.py b/comfy/ldm/modules/attention.py index 7437e0567..ccf690945 100644 --- a/comfy/ldm/modules/attention.py +++ b/comfy/ldm/modules/attention.py @@ -30,6 +30,13 @@ except ImportError as e: raise e exit(-1) +SAGE_ATTENTION3_IS_AVAILABLE = False +try: + from sageattn3 import sageattn3_blackwell + SAGE_ATTENTION3_IS_AVAILABLE = True +except ImportError: + pass + FLASH_ATTENTION_IS_AVAILABLE = False try: from flash_attn import flash_attn_func @@ -517,6 +524,7 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha @wrap_attn def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs): + exception_fallback = False if skip_reshape: b, _, _, dim_head = q.shape tensor_layout = "HND" @@ -541,6 +549,8 @@ def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape= out = sageattn(q, k, v, attn_mask=mask, is_causal=False, tensor_layout=tensor_layout) except Exception as e: logging.error("Error running sage attention: {}, using pytorch attention instead.".format(e)) + exception_fallback = True + if exception_fallback: if tensor_layout == "NHD": q, k, v = map( lambda t: t.transpose(1, 2), @@ -560,6 +570,93 @@ def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape= out = out.reshape(b, -1, heads * dim_head) return out +@wrap_attn +def attention3_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs): + exception_fallback = False + if (q.device.type != "cuda" or + q.dtype not in (torch.float16, torch.bfloat16) or + mask is not None): + return attention_pytorch( + q, k, v, heads, + mask=mask, + attn_precision=attn_precision, + skip_reshape=skip_reshape, + skip_output_reshape=skip_output_reshape, + **kwargs + ) + + if skip_reshape: + B, H, L, D = q.shape + if H != heads: + return attention_pytorch( + q, k, v, heads, + mask=mask, + attn_precision=attn_precision, + skip_reshape=True, + skip_output_reshape=skip_output_reshape, + **kwargs + ) + q_s, k_s, v_s = q, k, v + N = q.shape[2] + dim_head = D + else: + B, N, inner_dim = q.shape + if inner_dim % heads != 0: + return attention_pytorch( + q, k, v, heads, + mask=mask, + attn_precision=attn_precision, + skip_reshape=False, + skip_output_reshape=skip_output_reshape, + **kwargs + ) + dim_head = inner_dim // heads + + if dim_head >= 256 or N <= 1024: + return attention_pytorch( + q, k, v, heads, + mask=mask, + attn_precision=attn_precision, + skip_reshape=skip_reshape, + skip_output_reshape=skip_output_reshape, + **kwargs + ) + + if not skip_reshape: + q_s, k_s, v_s = map( + lambda t: t.view(B, -1, heads, dim_head).permute(0, 2, 1, 3).contiguous(), + (q, k, v), + ) + B, H, L, D = q_s.shape + + try: + out = sageattn3_blackwell(q_s, k_s, v_s, is_causal=False) + except Exception as e: + exception_fallback = True + logging.error("Error running SageAttention3: %s, falling back to pytorch attention.", e) + + if exception_fallback: + if not skip_reshape: + del q_s, k_s, v_s + return attention_pytorch( + q, k, v, heads, + mask=mask, + attn_precision=attn_precision, + skip_reshape=False, + skip_output_reshape=skip_output_reshape, + **kwargs + ) + + if skip_reshape: + if not skip_output_reshape: + out = out.permute(0, 2, 1, 3).reshape(B, L, H * D) + else: + if skip_output_reshape: + pass + else: + out = out.permute(0, 2, 1, 3).reshape(B, L, H * D) + + return out try: @torch.library.custom_op("flash_attention::flash_attn", mutates_args=()) @@ -647,6 +744,8 @@ optimized_attention_masked = optimized_attention # register core-supported attention functions if SAGE_ATTENTION_IS_AVAILABLE: register_attention_function("sage", attention_sage) +if SAGE_ATTENTION3_IS_AVAILABLE: + register_attention_function("sage3", attention3_sage) if FLASH_ATTENTION_IS_AVAILABLE: register_attention_function("flash", attention_flash) if model_management.xformers_enabled(): diff --git a/comfy/ldm/modules/diffusionmodules/mmdit.py b/comfy/ldm/modules/diffusionmodules/mmdit.py index 42f406f1a..0dc8fe789 100644 --- a/comfy/ldm/modules/diffusionmodules/mmdit.py +++ b/comfy/ldm/modules/diffusionmodules/mmdit.py @@ -211,12 +211,14 @@ class TimestepEmbedder(nn.Module): Embeds scalar timesteps into vector representations. """ - def __init__(self, hidden_size, frequency_embedding_size=256, dtype=None, device=None, operations=None): + def __init__(self, hidden_size, frequency_embedding_size=256, output_size=None, dtype=None, device=None, operations=None): super().__init__() + if output_size is None: + output_size = hidden_size self.mlp = nn.Sequential( operations.Linear(frequency_embedding_size, hidden_size, bias=True, dtype=dtype, device=device), nn.SiLU(), - operations.Linear(hidden_size, hidden_size, bias=True, dtype=dtype, device=device), + operations.Linear(hidden_size, output_size, bias=True, dtype=dtype, device=device), ) self.frequency_embedding_size = frequency_embedding_size diff --git a/comfy/ldm/modules/diffusionmodules/model.py b/comfy/ldm/modules/diffusionmodules/model.py index 4245eedca..1ae3ef034 100644 --- a/comfy/ldm/modules/diffusionmodules/model.py +++ b/comfy/ldm/modules/diffusionmodules/model.py @@ -13,6 +13,12 @@ if model_management.xformers_enabled_vae(): import xformers import xformers.ops +def torch_cat_if_needed(xl, dim): + if len(xl) > 1: + return torch.cat(xl, dim) + else: + return xl[0] + def get_timestep_embedding(timesteps, embedding_dim): """ This matches the implementation in Denoising Diffusion Probabilistic Models: @@ -43,6 +49,37 @@ def Normalize(in_channels, num_groups=32): return ops.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True) +class CarriedConv3d(nn.Module): + def __init__(self, n_channels, out_channels, kernel_size, stride=1, dilation=1, padding=0, **kwargs): + super().__init__() + self.conv = ops.Conv3d(n_channels, out_channels, kernel_size, stride=stride, dilation=dilation, **kwargs) + + def forward(self, x): + return self.conv(x) + + +def conv_carry_causal_3d(xl, op, conv_carry_in=None, conv_carry_out=None): + + x = xl[0] + xl.clear() + + if isinstance(op, CarriedConv3d): + if conv_carry_in is None: + x = torch.nn.functional.pad(x, (1, 1, 1, 1, 2, 0), mode = 'replicate') + else: + carry_len = conv_carry_in[0].shape[2] + x = torch.nn.functional.pad(x, (1, 1, 1, 1, 2 - carry_len, 0), mode = 'replicate') + x = torch.cat([conv_carry_in.pop(0), x], dim=2) + + if conv_carry_out is not None: + to_push = x[:, :, -2:, :, :].clone() + conv_carry_out.append(to_push) + + out = op(x) + + return out + + class VideoConv3d(nn.Module): def __init__(self, n_channels, out_channels, kernel_size, stride=1, dilation=1, padding_mode='replicate', padding=1, **kwargs): super().__init__() @@ -89,29 +126,24 @@ class Upsample(nn.Module): stride=1, padding=1) - def forward(self, x): + def forward(self, x, conv_carry_in=None, conv_carry_out=None): scale_factor = self.scale_factor if isinstance(scale_factor, (int, float)): scale_factor = (scale_factor,) * (x.ndim - 2) if x.ndim == 5 and scale_factor[0] > 1.0: - t = x.shape[2] - if t > 1: - a, b = x.split((1, t - 1), dim=2) - del x - b = interpolate_up(b, scale_factor) - else: - a = x - - a = interpolate_up(a.squeeze(2), scale_factor=scale_factor[1:]).unsqueeze(2) - if t > 1: - x = torch.cat((a, b), dim=2) - else: - x = a + results = [] + if conv_carry_in is None: + first = x[:, :, :1, :, :] + results.append(interpolate_up(first.squeeze(2), scale_factor=scale_factor[1:]).unsqueeze(2)) + x = x[:, :, 1:, :, :] + if x.shape[2] > 0: + results.append(interpolate_up(x, scale_factor)) + x = torch_cat_if_needed(results, dim=2) else: x = interpolate_up(x, scale_factor) if self.with_conv: - x = self.conv(x) + x = conv_carry_causal_3d([x], self.conv, conv_carry_in, conv_carry_out) return x @@ -127,17 +159,20 @@ class Downsample(nn.Module): stride=stride, padding=0) - def forward(self, x): + def forward(self, x, conv_carry_in=None, conv_carry_out=None): if self.with_conv: - if x.ndim == 4: + if isinstance(self.conv, CarriedConv3d): + x = conv_carry_causal_3d([x], self.conv, conv_carry_in, conv_carry_out) + elif x.ndim == 4: pad = (0, 1, 0, 1) mode = "constant" x = torch.nn.functional.pad(x, pad, mode=mode, value=0) + x = self.conv(x) elif x.ndim == 5: pad = (1, 1, 1, 1, 2, 0) mode = "replicate" x = torch.nn.functional.pad(x, pad, mode=mode) - x = self.conv(x) + x = self.conv(x) else: x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2) return x @@ -183,23 +218,23 @@ class ResnetBlock(nn.Module): stride=1, padding=0) - def forward(self, x, temb=None): + def forward(self, x, temb=None, conv_carry_in=None, conv_carry_out=None): h = x h = self.norm1(h) - h = self.swish(h) - h = self.conv1(h) + h = [ self.swish(h) ] + h = conv_carry_causal_3d(h, self.conv1, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out) if temb is not None: h = h + self.temb_proj(self.swish(temb))[:,:,None,None] h = self.norm2(h) h = self.swish(h) - h = self.dropout(h) - h = self.conv2(h) + h = [ self.dropout(h) ] + h = conv_carry_causal_3d(h, self.conv2, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out) if self.in_channels != self.out_channels: if self.use_conv_shortcut: - x = self.conv_shortcut(x) + x = conv_carry_causal_3d([x], self.conv_shortcut, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out) else: x = self.nin_shortcut(x) @@ -279,6 +314,7 @@ def pytorch_attention(q, k, v): orig_shape = q.shape B = orig_shape[0] C = orig_shape[1] + oom_fallback = False q, k, v = map( lambda t: t.view(B, 1, C, -1).transpose(2, 3).contiguous(), (q, k, v), @@ -289,6 +325,8 @@ def pytorch_attention(q, k, v): out = out.transpose(2, 3).reshape(orig_shape) except model_management.OOM_EXCEPTION: logging.warning("scaled_dot_product_attention OOMed: switched to slice attention") + oom_fallback = True + if oom_fallback: out = slice_attention(q.view(B, -1, C), k.view(B, -1, C).transpose(1, 2), v.view(B, -1, C).transpose(1, 2)).reshape(orig_shape) return out @@ -356,7 +394,8 @@ class Model(nn.Module): attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, resolution, use_timestep=True, use_linear_attn=False, attn_type="vanilla"): super().__init__() - if use_linear_attn: attn_type = "linear" + if use_linear_attn: + attn_type = "linear" self.ch = ch self.temb_ch = self.ch*4 self.num_resolutions = len(ch_mult) @@ -510,16 +549,22 @@ class Encoder(nn.Module): conv3d=False, time_compress=None, **ignore_kwargs): super().__init__() - if use_linear_attn: attn_type = "linear" + if use_linear_attn: + attn_type = "linear" self.ch = ch self.temb_ch = 0 self.num_resolutions = len(ch_mult) self.num_res_blocks = num_res_blocks self.resolution = resolution self.in_channels = in_channels + self.carried = False if conv3d: - conv_op = VideoConv3d + if not attn_resolutions: + conv_op = CarriedConv3d + self.carried = True + else: + conv_op = VideoConv3d mid_attn_conv_op = ops.Conv3d else: conv_op = ops.Conv2d @@ -532,6 +577,7 @@ class Encoder(nn.Module): stride=1, padding=1) + self.time_compress = 1 curr_res = resolution in_ch_mult = (1,)+tuple(ch_mult) self.in_ch_mult = in_ch_mult @@ -558,10 +604,15 @@ class Encoder(nn.Module): if time_compress is not None: if (self.num_resolutions - 1 - i_level) > math.log2(time_compress): stride = (1, 2, 2) + else: + self.time_compress *= 2 down.downsample = Downsample(block_in, resamp_with_conv, stride=stride, conv_op=conv_op) curr_res = curr_res // 2 self.down.append(down) + if time_compress is not None: + self.time_compress = time_compress + # middle self.mid = nn.Module() self.mid.block_1 = ResnetBlock(in_channels=block_in, @@ -587,15 +638,42 @@ class Encoder(nn.Module): def forward(self, x): # timestep embedding temb = None - # downsampling - h = self.conv_in(x) - for i_level in range(self.num_resolutions): - for i_block in range(self.num_res_blocks): - h = self.down[i_level].block[i_block](h, temb) - if len(self.down[i_level].attn) > 0: - h = self.down[i_level].attn[i_block](h) - if i_level != self.num_resolutions-1: - h = self.down[i_level].downsample(h) + + if self.carried: + xl = [x[:, :, :1, :, :]] + if x.shape[2] > self.time_compress: + tc = self.time_compress + xl += torch.split(x[:, :, 1: 1 + ((x.shape[2] - 1) // tc) * tc, :, :], tc * 2, dim = 2) + x = xl + else: + x = [x] + out = [] + + conv_carry_in = None + + for i, x1 in enumerate(x): + conv_carry_out = [] + if i == len(x) - 1: + conv_carry_out = None + + # downsampling + x1 = [ x1 ] + h1 = conv_carry_causal_3d(x1, self.conv_in, conv_carry_in, conv_carry_out) + + for i_level in range(self.num_resolutions): + for i_block in range(self.num_res_blocks): + h1 = self.down[i_level].block[i_block](h1, temb, conv_carry_in, conv_carry_out) + if len(self.down[i_level].attn) > 0: + assert i == 0 #carried should not happen if attn exists + h1 = self.down[i_level].attn[i_block](h1) + if i_level != self.num_resolutions-1: + h1 = self.down[i_level].downsample(h1, conv_carry_in, conv_carry_out) + + out.append(h1) + conv_carry_in = conv_carry_out + + h = torch_cat_if_needed(out, dim=2) + del out # middle h = self.mid.block_1(h, temb) @@ -604,15 +682,15 @@ class Encoder(nn.Module): # end h = self.norm_out(h) - h = nonlinearity(h) - h = self.conv_out(h) + h = [ nonlinearity(h) ] + h = conv_carry_causal_3d(h, self.conv_out) return h class Decoder(nn.Module): def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, - resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False, + resolution, z_channels, tanh_out=False, use_linear_attn=False, conv_out_op=ops.Conv2d, resnet_op=ResnetBlock, attn_op=AttnBlock, @@ -626,12 +704,18 @@ class Decoder(nn.Module): self.num_res_blocks = num_res_blocks self.resolution = resolution self.in_channels = in_channels - self.give_pre_end = give_pre_end self.tanh_out = tanh_out + self.carried = False if conv3d: - conv_op = VideoConv3d - conv_out_op = VideoConv3d + if not attn_resolutions and resnet_op == ResnetBlock: + conv_op = CarriedConv3d + conv_out_op = CarriedConv3d + self.carried = True + else: + conv_op = VideoConv3d + conv_out_op = VideoConv3d + mid_attn_conv_op = ops.Conv3d else: conv_op = ops.Conv2d @@ -706,29 +790,43 @@ class Decoder(nn.Module): temb = None # z to block_in - h = self.conv_in(z) + h = conv_carry_causal_3d([z], self.conv_in) # middle h = self.mid.block_1(h, temb, **kwargs) h = self.mid.attn_1(h, **kwargs) h = self.mid.block_2(h, temb, **kwargs) + if self.carried: + h = torch.split(h, 2, dim=2) + else: + h = [ h ] + out = [] + + conv_carry_in = None + # upsampling - for i_level in reversed(range(self.num_resolutions)): - for i_block in range(self.num_res_blocks+1): - h = self.up[i_level].block[i_block](h, temb, **kwargs) - if len(self.up[i_level].attn) > 0: - h = self.up[i_level].attn[i_block](h, **kwargs) - if i_level != 0: - h = self.up[i_level].upsample(h) + for i, h1 in enumerate(h): + conv_carry_out = [] + if i == len(h) - 1: + conv_carry_out = None + for i_level in reversed(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks+1): + h1 = self.up[i_level].block[i_block](h1, temb, conv_carry_in, conv_carry_out, **kwargs) + if len(self.up[i_level].attn) > 0: + assert i == 0 #carried should not happen if attn exists + h1 = self.up[i_level].attn[i_block](h1, **kwargs) + if i_level != 0: + h1 = self.up[i_level].upsample(h1, conv_carry_in, conv_carry_out) - # end - if self.give_pre_end: - return h + h1 = self.norm_out(h1) + h1 = [ nonlinearity(h1) ] + h1 = conv_carry_causal_3d(h1, self.conv_out, conv_carry_in, conv_carry_out) + if self.tanh_out: + h1 = torch.tanh(h1) + out.append(h1) + conv_carry_in = conv_carry_out - h = self.norm_out(h) - h = nonlinearity(h) - h = self.conv_out(h, **kwargs) - if self.tanh_out: - h = torch.tanh(h) - return h + out = torch_cat_if_needed(out, dim=2) + + return out diff --git a/comfy/ldm/modules/ema.py b/comfy/ldm/modules/ema.py index bded25019..96ee6e895 100644 --- a/comfy/ldm/modules/ema.py +++ b/comfy/ldm/modules/ema.py @@ -45,7 +45,7 @@ class LitEma(nn.Module): shadow_params[sname] = shadow_params[sname].type_as(m_param[key]) shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key])) else: - assert not key in self.m_name2s_name + assert key not in self.m_name2s_name def copy_to(self, model): m_param = dict(model.named_parameters()) @@ -54,7 +54,7 @@ class LitEma(nn.Module): if m_param[key].requires_grad: m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data) else: - assert not key in self.m_name2s_name + assert key not in self.m_name2s_name def store(self, parameters): """ diff --git a/comfy/ldm/qwen_image/model.py b/comfy/ldm/qwen_image/model.py index 427ea19c1..00c597535 100644 --- a/comfy/ldm/qwen_image/model.py +++ b/comfy/ldm/qwen_image/model.py @@ -61,7 +61,7 @@ def apply_rotary_emb(x, freqs_cis): class QwenTimestepProjEmbeddings(nn.Module): - def __init__(self, embedding_dim, pooled_projection_dim, dtype=None, device=None, operations=None): + def __init__(self, embedding_dim, pooled_projection_dim, use_additional_t_cond=False, dtype=None, device=None, operations=None): super().__init__() self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0, scale=1000) self.timestep_embedder = TimestepEmbedding( @@ -72,9 +72,19 @@ class QwenTimestepProjEmbeddings(nn.Module): operations=operations ) - def forward(self, timestep, hidden_states): + self.use_additional_t_cond = use_additional_t_cond + if self.use_additional_t_cond: + self.addition_t_embedding = operations.Embedding(2, embedding_dim, device=device, dtype=dtype) + + def forward(self, timestep, hidden_states, addition_t_cond=None): timesteps_proj = self.time_proj(timestep) timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_states.dtype)) + + if self.use_additional_t_cond: + if addition_t_cond is None: + addition_t_cond = torch.zeros((timesteps_emb.shape[0]), device=timesteps_emb.device, dtype=torch.long) + timesteps_emb += self.addition_t_embedding(addition_t_cond, out_dtype=timesteps_emb.dtype) + return timesteps_emb @@ -218,9 +228,24 @@ class QwenImageTransformerBlock(nn.Module): operations=operations, ) - def _modulate(self, x: torch.Tensor, mod_params: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: + def _apply_gate(self, x, y, gate, timestep_zero_index=None): + if timestep_zero_index is not None: + return y + torch.cat((x[:, :timestep_zero_index] * gate[0], x[:, timestep_zero_index:] * gate[1]), dim=1) + else: + return torch.addcmul(y, gate, x) + + def _modulate(self, x: torch.Tensor, mod_params: torch.Tensor, timestep_zero_index=None) -> Tuple[torch.Tensor, torch.Tensor]: shift, scale, gate = torch.chunk(mod_params, 3, dim=-1) - return torch.addcmul(shift.unsqueeze(1), x, 1 + scale.unsqueeze(1)), gate.unsqueeze(1) + if timestep_zero_index is not None: + actual_batch = shift.size(0) // 2 + shift, shift_0 = shift[:actual_batch], shift[actual_batch:] + scale, scale_0 = scale[:actual_batch], scale[actual_batch:] + gate, gate_0 = gate[:actual_batch], gate[actual_batch:] + reg = torch.addcmul(shift.unsqueeze(1), x[:, :timestep_zero_index], 1 + scale.unsqueeze(1)) + zero = torch.addcmul(shift_0.unsqueeze(1), x[:, timestep_zero_index:], 1 + scale_0.unsqueeze(1)) + return torch.cat((reg, zero), dim=1), (gate.unsqueeze(1), gate_0.unsqueeze(1)) + else: + return torch.addcmul(shift.unsqueeze(1), x, 1 + scale.unsqueeze(1)), gate.unsqueeze(1) def forward( self, @@ -229,14 +254,19 @@ class QwenImageTransformerBlock(nn.Module): encoder_hidden_states_mask: torch.Tensor, temb: torch.Tensor, image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, + timestep_zero_index=None, transformer_options={}, ) -> Tuple[torch.Tensor, torch.Tensor]: img_mod_params = self.img_mod(temb) + + if timestep_zero_index is not None: + temb = temb.chunk(2, dim=0)[0] + txt_mod_params = self.txt_mod(temb) img_mod1, img_mod2 = img_mod_params.chunk(2, dim=-1) txt_mod1, txt_mod2 = txt_mod_params.chunk(2, dim=-1) - img_modulated, img_gate1 = self._modulate(self.img_norm1(hidden_states), img_mod1) + img_modulated, img_gate1 = self._modulate(self.img_norm1(hidden_states), img_mod1, timestep_zero_index) del img_mod1 txt_modulated, txt_gate1 = self._modulate(self.txt_norm1(encoder_hidden_states), txt_mod1) del txt_mod1 @@ -251,15 +281,15 @@ class QwenImageTransformerBlock(nn.Module): del img_modulated del txt_modulated - hidden_states = hidden_states + img_gate1 * img_attn_output + hidden_states = self._apply_gate(img_attn_output, hidden_states, img_gate1, timestep_zero_index) encoder_hidden_states = encoder_hidden_states + txt_gate1 * txt_attn_output del img_attn_output del txt_attn_output del img_gate1 del txt_gate1 - img_modulated2, img_gate2 = self._modulate(self.img_norm2(hidden_states), img_mod2) - hidden_states = torch.addcmul(hidden_states, img_gate2, self.img_mlp(img_modulated2)) + img_modulated2, img_gate2 = self._modulate(self.img_norm2(hidden_states), img_mod2, timestep_zero_index) + hidden_states = self._apply_gate(self.img_mlp(img_modulated2), hidden_states, img_gate2, timestep_zero_index) txt_modulated2, txt_gate2 = self._modulate(self.txt_norm2(encoder_hidden_states), txt_mod2) encoder_hidden_states = torch.addcmul(encoder_hidden_states, txt_gate2, self.txt_mlp(txt_modulated2)) @@ -300,10 +330,11 @@ class QwenImageTransformer2DModel(nn.Module): num_attention_heads: int = 24, joint_attention_dim: int = 3584, pooled_projection_dim: int = 768, - guidance_embeds: bool = False, axes_dims_rope: Tuple[int, int, int] = (16, 56, 56), + default_ref_method="index", image_model=None, final_layer=True, + use_additional_t_cond=False, dtype=None, device=None, operations=None, @@ -314,12 +345,14 @@ class QwenImageTransformer2DModel(nn.Module): self.in_channels = in_channels self.out_channels = out_channels or in_channels self.inner_dim = num_attention_heads * attention_head_dim + self.default_ref_method = default_ref_method self.pe_embedder = EmbedND(dim=attention_head_dim, theta=10000, axes_dim=list(axes_dims_rope)) self.time_text_embed = QwenTimestepProjEmbeddings( embedding_dim=self.inner_dim, pooled_projection_dim=pooled_projection_dim, + use_additional_t_cond=use_additional_t_cond, dtype=dtype, device=device, operations=operations @@ -341,6 +374,9 @@ class QwenImageTransformer2DModel(nn.Module): for _ in range(num_layers) ]) + if self.default_ref_method == "index_timestep_zero": + self.register_buffer("__index_timestep_zero__", torch.tensor([])) + if final_layer: self.norm_out = LastLayer(self.inner_dim, self.inner_dim, dtype=dtype, device=device, operations=operations) self.proj_out = operations.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True, dtype=dtype, device=device) @@ -350,27 +386,33 @@ class QwenImageTransformer2DModel(nn.Module): patch_size = self.patch_size hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (1, self.patch_size, self.patch_size)) orig_shape = hidden_states.shape - hidden_states = hidden_states.view(orig_shape[0], orig_shape[1], orig_shape[-2] // 2, 2, orig_shape[-1] // 2, 2) - hidden_states = hidden_states.permute(0, 2, 4, 1, 3, 5) - hidden_states = hidden_states.reshape(orig_shape[0], (orig_shape[-2] // 2) * (orig_shape[-1] // 2), orig_shape[1] * 4) + hidden_states = hidden_states.view(orig_shape[0], orig_shape[1], orig_shape[-3], orig_shape[-2] // 2, 2, orig_shape[-1] // 2, 2) + hidden_states = hidden_states.permute(0, 2, 3, 5, 1, 4, 6) + hidden_states = hidden_states.reshape(orig_shape[0], orig_shape[-3] * (orig_shape[-2] // 2) * (orig_shape[-1] // 2), orig_shape[1] * 4) + t_len = t h_len = ((h + (patch_size // 2)) // patch_size) w_len = ((w + (patch_size // 2)) // patch_size) h_offset = ((h_offset + (patch_size // 2)) // patch_size) w_offset = ((w_offset + (patch_size // 2)) // patch_size) - img_ids = torch.zeros((h_len, w_len, 3), device=x.device) - img_ids[:, :, 0] = img_ids[:, :, 1] + index - img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1) - (h_len // 2) - img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0) - (w_len // 2) - return hidden_states, repeat(img_ids, "h w c -> b (h w) c", b=bs), orig_shape + img_ids = torch.zeros((t_len, h_len, w_len, 3), device=x.device) - def forward(self, x, timestep, context, attention_mask=None, guidance=None, ref_latents=None, transformer_options={}, **kwargs): + if t_len > 1: + img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(0, t_len - 1, steps=t_len, device=x.device, dtype=x.dtype).unsqueeze(1).unsqueeze(1) + else: + img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + index + + img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1).unsqueeze(0) - (h_len // 2) + img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0).unsqueeze(0) - (w_len // 2) + return hidden_states, repeat(img_ids, "t h w c -> b (t h w) c", b=bs), orig_shape + + def forward(self, x, timestep, context, attention_mask=None, ref_latents=None, additional_t_cond=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, timestep, context, attention_mask, guidance, ref_latents, transformer_options, **kwargs) + ).execute(x, timestep, context, attention_mask, ref_latents, additional_t_cond, transformer_options, **kwargs) def _forward( self, @@ -378,8 +420,8 @@ class QwenImageTransformer2DModel(nn.Module): timesteps, context, attention_mask=None, - guidance: torch.Tensor = None, ref_latents=None, + additional_t_cond=None, transformer_options={}, control=None, **kwargs @@ -391,16 +433,24 @@ class QwenImageTransformer2DModel(nn.Module): hidden_states, img_ids, orig_shape = self.process_img(x) num_embeds = hidden_states.shape[1] + timestep_zero_index = None if ref_latents is not None: h = 0 w = 0 index = 0 - index_ref_method = kwargs.get("ref_latents_method", "index") == "index" + ref_method = kwargs.get("ref_latents_method", self.default_ref_method) + index_ref_method = (ref_method == "index") or (ref_method == "index_timestep_zero") + negative_ref_method = ref_method == "negative_index" + timestep_zero = ref_method == "index_timestep_zero" for ref in ref_latents: if index_ref_method: index += 1 h_offset = 0 w_offset = 0 + elif negative_ref_method: + index -= 1 + h_offset = 0 + w_offset = 0 else: index = 1 h_offset = 0 @@ -415,6 +465,10 @@ class QwenImageTransformer2DModel(nn.Module): kontext, kontext_ids, _ = self.process_img(ref, index=index, h_offset=h_offset, w_offset=w_offset) hidden_states = torch.cat([hidden_states, kontext], dim=1) img_ids = torch.cat([img_ids, kontext_ids], dim=1) + if timestep_zero: + if index > 0: + timestep = torch.cat([timestep, timestep * 0], dim=0) + timestep_zero_index = num_embeds txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size) // 2, ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size) // 2)) txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3) @@ -426,24 +480,20 @@ class QwenImageTransformer2DModel(nn.Module): encoder_hidden_states = self.txt_norm(encoder_hidden_states) encoder_hidden_states = self.txt_in(encoder_hidden_states) - if guidance is not None: - guidance = guidance * 1000 - - temb = ( - self.time_text_embed(timestep, hidden_states) - if guidance is None - else self.time_text_embed(timestep, guidance, hidden_states) - ) + temb = self.time_text_embed(timestep, hidden_states, additional_t_cond) patches_replace = transformer_options.get("patches_replace", {}) patches = transformer_options.get("patches", {}) blocks_replace = patches_replace.get("dit", {}) + transformer_options["total_blocks"] = len(self.transformer_blocks) + transformer_options["block_type"] = "double" for i, block in enumerate(self.transformer_blocks): + transformer_options["block_index"] = i if ("double_block", i) in blocks_replace: def block_wrap(args): out = {} - out["txt"], out["img"] = block(hidden_states=args["img"], encoder_hidden_states=args["txt"], encoder_hidden_states_mask=encoder_hidden_states_mask, temb=args["vec"], image_rotary_emb=args["pe"], transformer_options=args["transformer_options"]) + out["txt"], out["img"] = block(hidden_states=args["img"], encoder_hidden_states=args["txt"], encoder_hidden_states_mask=encoder_hidden_states_mask, temb=args["vec"], image_rotary_emb=args["pe"], timestep_zero_index=timestep_zero_index, transformer_options=args["transformer_options"]) return out out = blocks_replace[("double_block", i)]({"img": hidden_states, "txt": encoder_hidden_states, "vec": temb, "pe": image_rotary_emb, "transformer_options": transformer_options}, {"original_block": block_wrap}) hidden_states = out["img"] @@ -455,6 +505,7 @@ class QwenImageTransformer2DModel(nn.Module): encoder_hidden_states_mask=encoder_hidden_states_mask, temb=temb, image_rotary_emb=image_rotary_emb, + timestep_zero_index=timestep_zero_index, transformer_options=transformer_options, ) @@ -471,9 +522,12 @@ class QwenImageTransformer2DModel(nn.Module): if add is not None: hidden_states[:, :add.shape[1]] += add + if timestep_zero_index is not None: + temb = temb.chunk(2, dim=0)[0] + hidden_states = self.norm_out(hidden_states, temb) hidden_states = self.proj_out(hidden_states) - hidden_states = hidden_states[:, :num_embeds].view(orig_shape[0], orig_shape[-2] // 2, orig_shape[-1] // 2, orig_shape[1], 2, 2) - hidden_states = hidden_states.permute(0, 3, 1, 4, 2, 5) + hidden_states = hidden_states[:, :num_embeds].view(orig_shape[0], orig_shape[-3], orig_shape[-2] // 2, orig_shape[-1] // 2, orig_shape[1], 2, 2) + hidden_states = hidden_states.permute(0, 4, 1, 2, 5, 3, 6) return hidden_states.reshape(orig_shape)[:, :, :, :x.shape[-2], :x.shape[-1]] diff --git a/comfy/ldm/util.py b/comfy/ldm/util.py index 30b4b4721..304936ff4 100644 --- a/comfy/ldm/util.py +++ b/comfy/ldm/util.py @@ -71,7 +71,7 @@ def count_params(model, verbose=False): def instantiate_from_config(config): - if not "target" in config: + if "target" not in config: if config == '__is_first_stage__': return None elif config == "__is_unconditional__": diff --git a/comfy/ldm/wan/model.py b/comfy/ldm/wan/model.py index a9d5e10d9..4216ce831 100644 --- a/comfy/ldm/wan/model.py +++ b/comfy/ldm/wan/model.py @@ -568,7 +568,10 @@ class WanModel(torch.nn.Module): 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 = {} @@ -763,7 +766,10 @@ class VaceWanModel(WanModel): 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 = {} @@ -862,7 +868,10 @@ class CameraWanModel(WanModel): 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 = {} @@ -1326,16 +1335,19 @@ class WanModel_S2V(WanModel): 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"]) + out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], transformer_options=args["transformer_options"]) return out - out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap}) + 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) + x = block(x, e=e0, freqs=freqs, context=context, transformer_options=transformer_options) if audio_emb is not None: x = self.audio_injector(x, i, audio_emb, audio_emb_global, seq_len) # head @@ -1574,7 +1586,10 @@ class HumoWanModel(WanModel): 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 = {} diff --git a/comfy/ldm/wan/model_animate.py b/comfy/ldm/wan/model_animate.py index 7c87835d4..84d7adec4 100644 --- a/comfy/ldm/wan/model_animate.py +++ b/comfy/ldm/wan/model_animate.py @@ -523,7 +523,10 @@ class AnimateWanModel(WanModel): 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 = {} diff --git a/comfy/ldm/wan/vae.py b/comfy/ldm/wan/vae.py index ccbb25822..08315f1a8 100644 --- a/comfy/ldm/wan/vae.py +++ b/comfy/ldm/wan/vae.py @@ -227,6 +227,7 @@ class Encoder3d(nn.Module): def __init__(self, dim=128, z_dim=4, + input_channels=3, dim_mult=[1, 2, 4, 4], num_res_blocks=2, attn_scales=[], @@ -245,7 +246,7 @@ class Encoder3d(nn.Module): scale = 1.0 # init block - self.conv1 = CausalConv3d(3, dims[0], 3, padding=1) + self.conv1 = CausalConv3d(input_channels, dims[0], 3, padding=1) # downsample blocks downsamples = [] @@ -331,6 +332,7 @@ class Decoder3d(nn.Module): def __init__(self, dim=128, z_dim=4, + output_channels=3, dim_mult=[1, 2, 4, 4], num_res_blocks=2, attn_scales=[], @@ -378,7 +380,7 @@ class Decoder3d(nn.Module): # output blocks self.head = nn.Sequential( RMS_norm(out_dim, images=False), nn.SiLU(), - CausalConv3d(out_dim, 3, 3, padding=1)) + CausalConv3d(out_dim, output_channels, 3, padding=1)) def forward(self, x, feat_cache=None, feat_idx=[0]): ## conv1 @@ -449,6 +451,7 @@ class WanVAE(nn.Module): num_res_blocks=2, attn_scales=[], temperal_downsample=[True, True, False], + image_channels=3, dropout=0.0): super().__init__() self.dim = dim @@ -460,11 +463,11 @@ class WanVAE(nn.Module): self.temperal_upsample = temperal_downsample[::-1] # modules - self.encoder = Encoder3d(dim, z_dim * 2, dim_mult, num_res_blocks, + self.encoder = Encoder3d(dim, z_dim * 2, image_channels, dim_mult, num_res_blocks, attn_scales, self.temperal_downsample, dropout) self.conv1 = CausalConv3d(z_dim * 2, z_dim * 2, 1) self.conv2 = CausalConv3d(z_dim, z_dim, 1) - self.decoder = Decoder3d(dim, z_dim, dim_mult, num_res_blocks, + self.decoder = Decoder3d(dim, z_dim, image_channels, dim_mult, num_res_blocks, attn_scales, self.temperal_upsample, dropout) def encode(self, x): diff --git a/comfy/lora.py b/comfy/lora.py index 36d26293a..2ed0acb9d 100644 --- a/comfy/lora.py +++ b/comfy/lora.py @@ -313,6 +313,23 @@ def model_lora_keys_unet(model, key_map={}): key_map["transformer.{}".format(key_lora)] = k key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = k #SimpleTuner lycoris format + if isinstance(model, comfy.model_base.Lumina2): + diffusers_keys = comfy.utils.z_image_to_diffusers(model.model_config.unet_config, output_prefix="diffusion_model.") + for k in diffusers_keys: + if k.endswith(".weight"): + to = diffusers_keys[k] + key_lora = k[:-len(".weight")] + key_map["diffusion_model.{}".format(key_lora)] = to + key_map["transformer.{}".format(key_lora)] = to + key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = to + + if isinstance(model, comfy.model_base.Kandinsky5): + for k in sdk: + if k.startswith("diffusion_model.") and k.endswith(".weight"): + key_lora = k[len("diffusion_model."):-len(".weight")] + key_map["{}".format(key_lora)] = k + key_map["transformer.{}".format(key_lora)] = k + return key_map diff --git a/comfy/model_base.py b/comfy/model_base.py index 7c788d085..49efd700b 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -20,6 +20,7 @@ import comfy.ldm.hunyuan3dv2_1 import comfy.ldm.hunyuan3dv2_1.hunyuandit import torch import logging +import comfy.ldm.lightricks.av_model from comfy.ldm.modules.diffusionmodules.openaimodel import UNetModel, Timestep from comfy.ldm.cascade.stage_c import StageC from comfy.ldm.cascade.stage_b import StageB @@ -47,6 +48,7 @@ import comfy.ldm.chroma_radiance.model import comfy.ldm.ace.model import comfy.ldm.omnigen.omnigen2 import comfy.ldm.qwen_image.model +import comfy.ldm.kandinsky5.model import comfy.model_management import comfy.patcher_extension @@ -134,7 +136,7 @@ class BaseModel(torch.nn.Module): if not unet_config.get("disable_unet_model_creation", False): if model_config.custom_operations is None: fp8 = model_config.optimizations.get("fp8", False) - operations = comfy.ops.pick_operations(unet_config.get("dtype", None), self.manual_cast_dtype, fp8_optimizations=fp8, scaled_fp8=model_config.scaled_fp8, model_config=model_config) + operations = comfy.ops.pick_operations(unet_config.get("dtype", None), self.manual_cast_dtype, fp8_optimizations=fp8, model_config=model_config) else: operations = model_config.custom_operations self.diffusion_model = unet_model(**unet_config, device=device, operations=operations) @@ -329,18 +331,6 @@ class BaseModel(torch.nn.Module): extra_sds.append(self.model_config.process_clip_vision_state_dict_for_saving(clip_vision_state_dict)) unet_state_dict = self.diffusion_model.state_dict() - - if self.model_config.scaled_fp8 is not None: - unet_state_dict["scaled_fp8"] = torch.tensor([], dtype=self.model_config.scaled_fp8) - - # Save mixed precision metadata - if hasattr(self.model_config, 'layer_quant_config') and self.model_config.layer_quant_config: - metadata = { - "format_version": "1.0", - "layers": self.model_config.layer_quant_config - } - unet_state_dict["_quantization_metadata"] = metadata - unet_state_dict = self.model_config.process_unet_state_dict_for_saving(unet_state_dict) if self.model_type == ModelType.V_PREDICTION: @@ -898,12 +888,13 @@ class Flux(BaseModel): attention_mask = kwargs.get("attention_mask", None) if attention_mask is not None: shape = kwargs["noise"].shape - mask_ref_size = kwargs["attention_mask_img_shape"] - # the model will pad to the patch size, and then divide - # essentially dividing and rounding up - (h_tok, w_tok) = (math.ceil(shape[2] / self.diffusion_model.patch_size), math.ceil(shape[3] / self.diffusion_model.patch_size)) - attention_mask = utils.upscale_dit_mask(attention_mask, mask_ref_size, (h_tok, w_tok)) - out['attention_mask'] = comfy.conds.CONDRegular(attention_mask) + mask_ref_size = kwargs.get("attention_mask_img_shape", None) + if mask_ref_size is not None: + # the model will pad to the patch size, and then divide + # essentially dividing and rounding up + (h_tok, w_tok) = (math.ceil(shape[2] / self.diffusion_model.patch_size), math.ceil(shape[3] / self.diffusion_model.patch_size)) + attention_mask = utils.upscale_dit_mask(attention_mask, mask_ref_size, (h_tok, w_tok)) + out['attention_mask'] = comfy.conds.CONDRegular(attention_mask) guidance = kwargs.get("guidance", 3.5) if guidance is not None: @@ -925,9 +916,19 @@ class Flux(BaseModel): out = {} ref_latents = kwargs.get("reference_latents", None) if ref_latents is not None: - out['ref_latents'] = list([1, 16, sum(map(lambda a: math.prod(a.size()), ref_latents)) // 16]) + out['ref_latents'] = list([1, 16, sum(map(lambda a: math.prod(a.size()[2:]), ref_latents))]) return out +class Flux2(Flux): + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + cross_attn = kwargs.get("cross_attn", None) + if cross_attn is not None: + target_text_len = 512 + if cross_attn.shape[1] < target_text_len: + cross_attn = torch.nn.functional.pad(cross_attn, (0, 0, target_text_len - cross_attn.shape[1], 0)) + out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) + return out class GenmoMochi(BaseModel): def __init__(self, model_config, model_type=ModelType.FLOW, device=None): @@ -946,7 +947,7 @@ class GenmoMochi(BaseModel): class LTXV(BaseModel): def __init__(self, model_config, model_type=ModelType.FLUX, device=None): - super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lightricks.model.LTXVModel) #TODO + super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lightricks.model.LTXVModel) def extra_conds(self, **kwargs): out = super().extra_conds(**kwargs) @@ -977,6 +978,60 @@ class LTXV(BaseModel): def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs): return latent_image +class LTXAV(BaseModel): + def __init__(self, model_config, model_type=ModelType.FLUX, device=None): + super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lightricks.av_model.LTXAVModel) #TODO + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + attention_mask = kwargs.get("attention_mask", None) + if attention_mask is not None: + out['attention_mask'] = comfy.conds.CONDRegular(attention_mask) + cross_attn = kwargs.get("cross_attn", None) + if cross_attn is not None: + out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) + + out['frame_rate'] = comfy.conds.CONDConstant(kwargs.get("frame_rate", 25)) + + denoise_mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None)) + + audio_denoise_mask = None + if denoise_mask is not None and "latent_shapes" in kwargs: + denoise_mask = utils.unpack_latents(denoise_mask, kwargs["latent_shapes"]) + if len(denoise_mask) > 1: + audio_denoise_mask = denoise_mask[1] + denoise_mask = denoise_mask[0] + + if denoise_mask is not None: + out["denoise_mask"] = comfy.conds.CONDRegular(denoise_mask) + + if audio_denoise_mask is not None: + out["audio_denoise_mask"] = comfy.conds.CONDRegular(audio_denoise_mask) + + keyframe_idxs = kwargs.get("keyframe_idxs", None) + if keyframe_idxs is not None: + out['keyframe_idxs'] = comfy.conds.CONDRegular(keyframe_idxs) + + latent_shapes = kwargs.get("latent_shapes", None) + if latent_shapes is not None: + out['latent_shapes'] = comfy.conds.CONDConstant(latent_shapes) + + return out + + def process_timestep(self, timestep, x, denoise_mask=None, audio_denoise_mask=None, **kwargs): + v_timestep = timestep + a_timestep = timestep + + if denoise_mask is not None: + v_timestep = self.diffusion_model.patchifier.patchify(((denoise_mask) * timestep.view([timestep.shape[0]] + [1] * (denoise_mask.ndim - 1)))[:, :1])[0] + if audio_denoise_mask is not None: + a_timestep = self.diffusion_model.a_patchifier.patchify(((audio_denoise_mask) * timestep.view([timestep.shape[0]] + [1] * (audio_denoise_mask.ndim - 1)))[:, :1, :, :1])[0] + + return v_timestep, a_timestep + + def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs): + return latent_image + class HunyuanVideo(BaseModel): def __init__(self, model_config, model_type=ModelType.FLOW, device=None): super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hunyuan_video.model.HunyuanVideo) @@ -1103,9 +1158,17 @@ class Lumina2(BaseModel): if torch.numel(attention_mask) != attention_mask.sum(): out['attention_mask'] = comfy.conds.CONDRegular(attention_mask) out['num_tokens'] = comfy.conds.CONDConstant(max(1, torch.sum(attention_mask).item())) + cross_attn = kwargs.get("cross_attn", None) if cross_attn is not None: out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) + if 'num_tokens' not in out: + out['num_tokens'] = comfy.conds.CONDConstant(cross_attn.shape[1]) + + clip_text_pooled = kwargs.get("pooled_output", None) # NewBie + if clip_text_pooled is not None: + out['clip_text_pooled'] = comfy.conds.CONDRegular(clip_text_pooled) + return out class WAN21(BaseModel): @@ -1536,3 +1599,140 @@ class HunyuanImage21Refiner(HunyuanImage21): out = super().extra_conds(**kwargs) out['disable_time_r'] = comfy.conds.CONDConstant(True) return out + +class HunyuanVideo15(HunyuanVideo): + def __init__(self, model_config, model_type=ModelType.FLOW, device=None): + super().__init__(model_config, model_type, device=device) + + def concat_cond(self, **kwargs): + noise = kwargs.get("noise", None) + extra_channels = self.diffusion_model.img_in.proj.weight.shape[1] - noise.shape[1] - 1 #noise 32 img cond 32 + mask 1 + if extra_channels == 0: + return None + + image = kwargs.get("concat_latent_image", None) + device = kwargs["device"] + + if image is None: + shape_image = list(noise.shape) + shape_image[1] = extra_channels + image = torch.zeros(shape_image, dtype=noise.dtype, layout=noise.layout, device=noise.device) + else: + latent_dim = self.latent_format.latent_channels + image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center") + for i in range(0, image.shape[1], latent_dim): + image[:, i: i + latent_dim] = self.process_latent_in(image[:, i: i + latent_dim]) + image = utils.resize_to_batch_size(image, noise.shape[0]) + + mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None)) + if mask is None: + mask = torch.zeros_like(noise)[:, :1] + else: + mask = 1.0 - mask + mask = utils.common_upscale(mask.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center") + if mask.shape[-3] < noise.shape[-3]: + mask = torch.nn.functional.pad(mask, (0, 0, 0, 0, 0, noise.shape[-3] - mask.shape[-3]), mode='constant', value=0) + mask = utils.resize_to_batch_size(mask, noise.shape[0]) + + return torch.cat((image, mask), dim=1) + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + attention_mask = kwargs.get("attention_mask", None) + if attention_mask is not None: + if torch.numel(attention_mask) != attention_mask.sum(): + out['attention_mask'] = comfy.conds.CONDRegular(attention_mask) + cross_attn = kwargs.get("cross_attn", None) + if cross_attn is not None: + out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) + + conditioning_byt5small = kwargs.get("conditioning_byt5small", None) + if conditioning_byt5small is not None: + out['txt_byt5'] = comfy.conds.CONDRegular(conditioning_byt5small) + + guidance = kwargs.get("guidance", 6.0) + if guidance is not None: + out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([guidance])) + + clip_vision_output = kwargs.get("clip_vision_output", None) + if clip_vision_output is not None: + out['clip_fea'] = comfy.conds.CONDRegular(clip_vision_output.last_hidden_state) + + return out + +class HunyuanVideo15_SR_Distilled(HunyuanVideo15): + def __init__(self, model_config, model_type=ModelType.FLOW, device=None): + super().__init__(model_config, model_type, device=device) + + def concat_cond(self, **kwargs): + noise = kwargs.get("noise", None) + image = kwargs.get("concat_latent_image", None) + noise_augmentation = kwargs.get("noise_augmentation", 0.0) + device = kwargs["device"] + + if image is None: + image = torch.zeros([noise.shape[0], noise.shape[1] * 2 + 2, noise.shape[-3], noise.shape[-2], noise.shape[-1]], device=comfy.model_management.intermediate_device()) + else: + image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center") + #image = self.process_latent_in(image) # scaling wasn't applied in reference code + image = utils.resize_to_batch_size(image, noise.shape[0]) + lq_image_slice = slice(noise.shape[1] + 1, 2 * noise.shape[1] + 1) + if noise_augmentation > 0: + generator = torch.Generator(device="cpu") + generator.manual_seed(kwargs.get("seed", 0) - 10) + noise = torch.randn(image[:, lq_image_slice].shape, generator=generator, dtype=image.dtype, device="cpu").to(image.device) + image[:, lq_image_slice] = noise_augmentation * noise + min(1.0 - noise_augmentation, 0.75) * image[:, lq_image_slice] + else: + image[:, lq_image_slice] = 0.75 * image[:, lq_image_slice] + return image + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + out['disable_time_r'] = comfy.conds.CONDConstant(False) + return out + +class Kandinsky5(BaseModel): + def __init__(self, model_config, model_type=ModelType.FLOW, device=None): + super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.kandinsky5.model.Kandinsky5) + + def encode_adm(self, **kwargs): + return kwargs["pooled_output"] + + def concat_cond(self, **kwargs): + noise = kwargs.get("noise", None) + device = kwargs["device"] + image = torch.zeros_like(noise) + + mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None)) + if mask is None: + mask = torch.zeros_like(noise)[:, :1] + else: + mask = 1.0 - mask + mask = utils.common_upscale(mask.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center") + if mask.shape[-3] < noise.shape[-3]: + mask = torch.nn.functional.pad(mask, (0, 0, 0, 0, 0, noise.shape[-3] - mask.shape[-3]), mode='constant', value=0) + mask = utils.resize_to_batch_size(mask, noise.shape[0]) + + return torch.cat((image, mask), dim=1) + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + attention_mask = kwargs.get("attention_mask", None) + if attention_mask is not None: + out['attention_mask'] = comfy.conds.CONDRegular(attention_mask) + cross_attn = kwargs.get("cross_attn", None) + if cross_attn is not None: + out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) + + time_dim_replace = kwargs.get("time_dim_replace", None) + if time_dim_replace is not None: + out['time_dim_replace'] = comfy.conds.CONDRegular(self.process_latent_in(time_dim_replace)) + + return out + +class Kandinsky5Image(Kandinsky5): + def __init__(self, model_config, model_type=ModelType.FLOW, device=None): + super().__init__(model_config, model_type, device=device) + + def concat_cond(self, **kwargs): + return None diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 3142a7fc3..0853b3aec 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -6,20 +6,6 @@ import math import logging import torch - -def detect_layer_quantization(metadata): - quant_key = "_quantization_metadata" - if metadata is not None and quant_key in metadata: - quant_metadata = metadata.pop(quant_key) - quant_metadata = json.loads(quant_metadata) - if isinstance(quant_metadata, dict) and "layers" in quant_metadata: - logging.info(f"Found quantization metadata (version {quant_metadata.get('format_version', 'unknown')})") - return quant_metadata["layers"] - else: - raise ValueError("Invalid quantization metadata format") - return None - - def count_blocks(state_dict_keys, prefix_string): count = 0 while True: @@ -186,30 +172,73 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): guidance_keys = list(filter(lambda a: a.startswith("{}guidance_in.".format(key_prefix)), state_dict_keys)) dit_config["guidance_embed"] = len(guidance_keys) > 0 + + # HunyuanVideo 1.5 + if '{}cond_type_embedding.weight'.format(key_prefix) in state_dict_keys: + dit_config["use_cond_type_embedding"] = True + else: + dit_config["use_cond_type_embedding"] = False + if '{}vision_in.proj.0.weight'.format(key_prefix) in state_dict_keys: + dit_config["vision_in_dim"] = state_dict['{}vision_in.proj.0.weight'.format(key_prefix)].shape[0] + dit_config["meanflow_sum"] = True + else: + dit_config["vision_in_dim"] = None + dit_config["meanflow_sum"] = False return dit_config if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys and ('{}img_in.weight'.format(key_prefix) in state_dict_keys or f"{key_prefix}distilled_guidance_layer.norms.0.scale" in state_dict_keys): #Flux, Chroma or Chroma Radiance (has no img_in.weight) dit_config = {} - dit_config["image_model"] = "flux" + if '{}double_stream_modulation_img.lin.weight'.format(key_prefix) in state_dict_keys: + dit_config["image_model"] = "flux2" + dit_config["axes_dim"] = [32, 32, 32, 32] + dit_config["num_heads"] = 48 + dit_config["mlp_ratio"] = 3.0 + dit_config["theta"] = 2000 + dit_config["out_channels"] = 128 + dit_config["global_modulation"] = True + dit_config["mlp_silu_act"] = True + dit_config["qkv_bias"] = False + dit_config["ops_bias"] = False + dit_config["default_ref_method"] = "index" + dit_config["ref_index_scale"] = 10.0 + dit_config["txt_ids_dims"] = [3] + patch_size = 1 + else: + dit_config["image_model"] = "flux" + dit_config["axes_dim"] = [16, 56, 56] + dit_config["num_heads"] = 24 + dit_config["mlp_ratio"] = 4.0 + dit_config["theta"] = 10000 + dit_config["out_channels"] = 16 + dit_config["qkv_bias"] = True + dit_config["txt_ids_dims"] = [] + patch_size = 2 + dit_config["in_channels"] = 16 - patch_size = 2 + dit_config["hidden_size"] = 3072 + dit_config["context_in_dim"] = 4096 + dit_config["patch_size"] = patch_size in_key = "{}img_in.weight".format(key_prefix) if in_key in state_dict_keys: - dit_config["in_channels"] = state_dict[in_key].shape[1] // (patch_size * patch_size) - dit_config["out_channels"] = 16 + w = state_dict[in_key] + dit_config["in_channels"] = w.shape[1] // (patch_size * patch_size) + dit_config["hidden_size"] = w.shape[0] + + txt_in_key = "{}txt_in.weight".format(key_prefix) + if txt_in_key in state_dict_keys: + w = state_dict[txt_in_key] + dit_config["context_in_dim"] = w.shape[1] + dit_config["hidden_size"] = w.shape[0] + vec_in_key = '{}vector_in.in_layer.weight'.format(key_prefix) if vec_in_key in state_dict_keys: dit_config["vec_in_dim"] = state_dict[vec_in_key].shape[1] - dit_config["context_in_dim"] = 4096 - dit_config["hidden_size"] = 3072 - dit_config["mlp_ratio"] = 4.0 - dit_config["num_heads"] = 24 + else: + dit_config["vec_in_dim"] = None + dit_config["depth"] = count_blocks(state_dict_keys, '{}double_blocks.'.format(key_prefix) + '{}.') dit_config["depth_single_blocks"] = count_blocks(state_dict_keys, '{}single_blocks.'.format(key_prefix) + '{}.') - dit_config["axes_dim"] = [16, 56, 56] - dit_config["theta"] = 10000 - dit_config["qkv_bias"] = True if '{}distilled_guidance_layer.0.norms.0.scale'.format(key_prefix) in state_dict_keys or '{}distilled_guidance_layer.norms.0.scale'.format(key_prefix) in state_dict_keys: #Chroma dit_config["image_model"] = "chroma" dit_config["in_channels"] = 64 @@ -230,8 +259,17 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["nerf_tile_size"] = 512 dit_config["nerf_final_head_type"] = "conv" if f"{key_prefix}nerf_final_layer_conv.norm.scale" in state_dict_keys else "linear" dit_config["nerf_embedder_dtype"] = torch.float32 + if "{}__x0__".format(key_prefix) in state_dict_keys: # x0 pred + dit_config["use_x0"] = True + else: + dit_config["use_x0"] = False else: dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys + dit_config["yak_mlp"] = '{}double_blocks.0.img_mlp.gate_proj.weight'.format(key_prefix) in state_dict_keys + dit_config["txt_norm"] = "{}txt_norm.scale".format(key_prefix) in state_dict_keys + if dit_config["yak_mlp"] and dit_config["txt_norm"]: # Ovis model + dit_config["txt_ids_dims"] = [1, 2] + return dit_config if '{}t5_yproj.weight'.format(key_prefix) in state_dict_keys: #Genmo mochi preview @@ -267,7 +305,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): if '{}adaln_single.emb.timestep_embedder.linear_1.bias'.format(key_prefix) in state_dict_keys: #Lightricks ltxv dit_config = {} - dit_config["image_model"] = "ltxv" + dit_config["image_model"] = "ltxav" if f'{key_prefix}audio_adaln_single.linear.weight' in state_dict_keys else "ltxv" dit_config["num_layers"] = count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.') shape = state_dict['{}transformer_blocks.0.attn2.to_k.weight'.format(key_prefix)].shape dit_config["attention_head_dim"] = shape[0] // 32 @@ -378,14 +416,35 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["image_model"] = "lumina2" dit_config["patch_size"] = 2 dit_config["in_channels"] = 16 - dit_config["dim"] = 2304 - dit_config["cap_feat_dim"] = state_dict['{}cap_embedder.1.weight'.format(key_prefix)].shape[1] + w = state_dict['{}cap_embedder.1.weight'.format(key_prefix)] + dit_config["dim"] = w.shape[0] + dit_config["cap_feat_dim"] = w.shape[1] dit_config["n_layers"] = count_blocks(state_dict_keys, '{}layers.'.format(key_prefix) + '{}.') - dit_config["n_heads"] = 24 - dit_config["n_kv_heads"] = 8 dit_config["qk_norm"] = True - dit_config["axes_dims"] = [32, 32, 32] - dit_config["axes_lens"] = [300, 512, 512] + + if dit_config["dim"] == 2304: # Original Lumina 2 + dit_config["n_heads"] = 24 + dit_config["n_kv_heads"] = 8 + dit_config["axes_dims"] = [32, 32, 32] + dit_config["axes_lens"] = [300, 512, 512] + dit_config["rope_theta"] = 10000.0 + dit_config["ffn_dim_multiplier"] = 4.0 + ctd_weight = state_dict.get('{}clip_text_pooled_proj.0.weight'.format(key_prefix), None) + if ctd_weight is not None: # NewBie + dit_config["clip_text_dim"] = ctd_weight.shape[0] + # NewBie also sets axes_lens = [1024, 512, 512] but it's not used in ComfyUI + elif dit_config["dim"] == 3840: # Z image + dit_config["n_heads"] = 30 + dit_config["n_kv_heads"] = 30 + dit_config["axes_dims"] = [32, 48, 48] + dit_config["axes_lens"] = [1536, 512, 512] + dit_config["rope_theta"] = 256.0 + dit_config["ffn_dim_multiplier"] = (8.0 / 3.0) + dit_config["z_image_modulation"] = True + dit_config["time_scale"] = 1000.0 + if '{}cap_pad_token'.format(key_prefix) in state_dict_keys: + dit_config["pad_tokens_multiple"] = 32 + return dit_config if '{}head.modulation'.format(key_prefix) in state_dict_keys: # Wan 2.1 @@ -560,6 +619,29 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["image_model"] = "qwen_image" dit_config["in_channels"] = state_dict['{}img_in.weight'.format(key_prefix)].shape[1] dit_config["num_layers"] = count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.') + if "{}__index_timestep_zero__".format(key_prefix) in state_dict_keys: # 2511 + dit_config["default_ref_method"] = "index_timestep_zero" + if "{}time_text_embed.addition_t_embedding.weight".format(key_prefix) in state_dict_keys: # Layered + dit_config["use_additional_t_cond"] = True + dit_config["default_ref_method"] = "negative_index" + return dit_config + + if '{}visual_transformer_blocks.0.cross_attention.key_norm.weight'.format(key_prefix) in state_dict_keys: # Kandinsky 5 + dit_config = {} + model_dim = state_dict['{}visual_embeddings.in_layer.bias'.format(key_prefix)].shape[0] + dit_config["model_dim"] = model_dim + if model_dim in [4096, 2560]: # pro video and lite image + dit_config["axes_dims"] = (32, 48, 48) + if model_dim == 2560: # lite image + dit_config["rope_scale_factor"] = (1.0, 1.0, 1.0) + elif model_dim == 1792: # lite video + dit_config["axes_dims"] = (16, 24, 24) + dit_config["time_dim"] = state_dict['{}time_embeddings.in_layer.bias'.format(key_prefix)].shape[0] + dit_config["image_model"] = "kandinsky5" + dit_config["ff_dim"] = state_dict['{}visual_transformer_blocks.0.feed_forward.in_layer.weight'.format(key_prefix)].shape[0] + dit_config["visual_embed_dim"] = state_dict['{}visual_embeddings.in_layer.weight'.format(key_prefix)].shape[1] + dit_config["num_text_blocks"] = count_blocks(state_dict_keys, '{}text_transformer_blocks.'.format(key_prefix) + '{}.') + dit_config["num_visual_blocks"] = count_blocks(state_dict_keys, '{}visual_transformer_blocks.'.format(key_prefix) + '{}.') return dit_config if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys: @@ -704,22 +786,11 @@ def model_config_from_unet(state_dict, unet_key_prefix, use_base_if_no_match=Fal if model_config is None and use_base_if_no_match: model_config = comfy.supported_models_base.BASE(unet_config) - scaled_fp8_key = "{}scaled_fp8".format(unet_key_prefix) - if scaled_fp8_key in state_dict: - scaled_fp8_weight = state_dict.pop(scaled_fp8_key) - model_config.scaled_fp8 = scaled_fp8_weight.dtype - if model_config.scaled_fp8 == torch.float32: - model_config.scaled_fp8 = torch.float8_e4m3fn - if scaled_fp8_weight.nelement() == 2: - model_config.optimizations["fp8"] = False - else: - model_config.optimizations["fp8"] = True - # Detect per-layer quantization (mixed precision) - layer_quant_config = detect_layer_quantization(metadata) - if layer_quant_config: - model_config.layer_quant_config = layer_quant_config - logging.info(f"Detected mixed precision quantization: {len(layer_quant_config)} layers quantized") + quant_config = comfy.utils.detect_layer_quantization(state_dict, unet_key_prefix) + if quant_config: + model_config.quant_config = quant_config + logging.info("Detected mixed precision quantization") return model_config diff --git a/comfy/model_management.py b/comfy/model_management.py index f0ccfc5dd..906350a47 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -26,6 +26,7 @@ import importlib import platform import weakref import gc +import os class VRAMState(Enum): DISABLED = 0 #No vram present: no need to move models to vram @@ -334,13 +335,15 @@ except: SUPPORT_FP8_OPS = args.supports_fp8_compute AMD_RDNA2_AND_OLDER_ARCH = ["gfx1030", "gfx1031", "gfx1010", "gfx1011", "gfx1012", "gfx906", "gfx900", "gfx803"] +AMD_ENABLE_MIOPEN_ENV = 'COMFYUI_ENABLE_MIOPEN' try: if is_amd(): arch = torch.cuda.get_device_properties(get_torch_device()).gcnArchName if not (any((a in arch) for a in AMD_RDNA2_AND_OLDER_ARCH)): - torch.backends.cudnn.enabled = False # Seems to improve things a lot on AMD - logging.info("Set: torch.backends.cudnn.enabled = False for better AMD performance.") + if os.getenv(AMD_ENABLE_MIOPEN_ENV) != '1': + torch.backends.cudnn.enabled = False # Seems to improve things a lot on AMD + logging.info("Set: torch.backends.cudnn.enabled = False for better AMD performance.") try: rocm_version = tuple(map(int, str(torch.version.hip).split(".")[:2])) @@ -454,7 +457,7 @@ def module_size(module): sd = module.state_dict() for k in sd: t = sd[k] - module_mem += t.nelement() * t.element_size() + module_mem += t.nbytes return module_mem class LoadedModel: @@ -690,7 +693,7 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu loaded_memory = loaded_model.model_loaded_memory() current_free_mem = get_free_memory(torch_dev) + loaded_memory - lowvram_model_memory = max(128 * 1024 * 1024, (current_free_mem - minimum_memory_required), min(current_free_mem * MIN_WEIGHT_MEMORY_RATIO, current_free_mem - minimum_inference_memory())) + lowvram_model_memory = max(0, (current_free_mem - minimum_memory_required), min(current_free_mem * MIN_WEIGHT_MEMORY_RATIO, current_free_mem - minimum_inference_memory())) lowvram_model_memory = lowvram_model_memory - loaded_memory if lowvram_model_memory == 0: @@ -1013,9 +1016,18 @@ def force_channels_last(): STREAMS = {} -NUM_STREAMS = 1 -if args.async_offload: - NUM_STREAMS = 2 +NUM_STREAMS = 0 +if args.async_offload is not None: + NUM_STREAMS = args.async_offload +else: + # Enable by default on Nvidia and AMD + if is_nvidia() or is_amd(): + NUM_STREAMS = 2 + +if args.disable_async_offload: + NUM_STREAMS = 0 + +if NUM_STREAMS > 0: logging.info("Using async weight offloading with {} streams".format(NUM_STREAMS)) def current_stream(device): @@ -1031,7 +1043,10 @@ def current_stream(device): stream_counters = {} def get_offload_stream(device): stream_counter = stream_counters.get(device, 0) - if NUM_STREAMS <= 1: + if NUM_STREAMS == 0: + return None + + if torch.compiler.is_compiling(): return None if device in STREAMS: @@ -1044,7 +1059,9 @@ def get_offload_stream(device): elif is_device_cuda(device): ss = [] for k in range(NUM_STREAMS): - ss.append(torch.cuda.Stream(device=device, priority=0)) + s1 = torch.cuda.Stream(device=device, priority=0) + s1.as_context = torch.cuda.stream + ss.append(s1) STREAMS[device] = ss s = ss[stream_counter] stream_counters[device] = stream_counter @@ -1052,7 +1069,9 @@ def get_offload_stream(device): elif is_device_xpu(device): ss = [] for k in range(NUM_STREAMS): - ss.append(torch.xpu.Stream(device=device, priority=0)) + s1 = torch.xpu.Stream(device=device, priority=0) + s1.as_context = torch.xpu.stream + ss.append(s1) STREAMS[device] = ss s = ss[stream_counter] stream_counters[device] = stream_counter @@ -1070,12 +1089,19 @@ def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, str if dtype is None or weight.dtype == dtype: return weight if stream is not None: - with stream: + wf_context = stream + if hasattr(wf_context, "as_context"): + wf_context = wf_context.as_context(stream) + with wf_context: return weight.to(dtype=dtype, copy=copy) return weight.to(dtype=dtype, copy=copy) + if stream is not None: - with stream: + wf_context = stream + if hasattr(wf_context, "as_context"): + wf_context = wf_context.as_context(stream) + with wf_context: r = torch.empty_like(weight, dtype=dtype, device=device) r.copy_(weight, non_blocking=non_blocking) else: @@ -1099,13 +1125,24 @@ if not args.disable_pinned_memory: MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.95 logging.info("Enabled pinned memory {}".format(MAX_PINNED_MEMORY // (1024 * 1024))) +PINNING_ALLOWED_TYPES = set(["Parameter", "QuantizedTensor"]) + +def discard_cuda_async_error(): + try: + a = torch.tensor([1], dtype=torch.uint8, device=get_torch_device()) + b = torch.tensor([1], dtype=torch.uint8, device=get_torch_device()) + _ = a + b + torch.cuda.synchronize() + except torch.AcceleratorError: + #Dump it! We already know about it from the synchronous return + pass def pin_memory(tensor): global TOTAL_PINNED_MEMORY if MAX_PINNED_MEMORY <= 0: return False - if type(tensor) is not torch.nn.parameter.Parameter: + if type(tensor).__name__ not in PINNING_ALLOWED_TYPES: return False if not is_device_cpu(tensor.device): @@ -1120,15 +1157,21 @@ def pin_memory(tensor): if not tensor.is_contiguous(): return False - size = tensor.numel() * tensor.element_size() + size = tensor.nbytes if (TOTAL_PINNED_MEMORY + size) > MAX_PINNED_MEMORY: return False ptr = tensor.data_ptr() + if ptr == 0: + return False + if torch.cuda.cudart().cudaHostRegister(ptr, size, 1) == 0: PINNED_MEMORY[ptr] = size TOTAL_PINNED_MEMORY += size return True + else: + logging.warning("Pin error.") + discard_cuda_async_error() return False @@ -1141,7 +1184,7 @@ def unpin_memory(tensor): return False ptr = tensor.data_ptr() - size = tensor.numel() * tensor.element_size() + size = tensor.nbytes size_stored = PINNED_MEMORY.get(ptr, None) if size_stored is None: @@ -1157,6 +1200,9 @@ def unpin_memory(tensor): if len(PINNED_MEMORY) == 0: TOTAL_PINNED_MEMORY = 0 return True + else: + logging.warning("Unpin error.") + discard_cuda_async_error() return False @@ -1459,6 +1505,16 @@ def supports_fp8_compute(device=None): return True +def supports_nvfp4_compute(device=None): + if not is_nvidia(): + return False + + props = torch.cuda.get_device_properties(device) + if props.major < 10: + return False + + return True + def extended_fp16_support(): # TODO: check why some models work with fp16 on newer torch versions but not on older if torch_version_numeric < (2, 7): @@ -1466,6 +1522,20 @@ def extended_fp16_support(): return True +LORA_COMPUTE_DTYPES = {} +def lora_compute_dtype(device): + dtype = LORA_COMPUTE_DTYPES.get(device, None) + if dtype is not None: + return dtype + + if should_use_fp16(device): + dtype = torch.float16 + else: + dtype = torch.float32 + + LORA_COMPUTE_DTYPES[device] = dtype + return dtype + def soft_empty_cache(force=False): global cpu_state if cpu_state == CPUState.MPS: @@ -1483,6 +1553,10 @@ def soft_empty_cache(force=False): def unload_all_models(): free_memory(1e30, get_torch_device()) +def debug_memory_summary(): + if is_amd() or is_nvidia(): + return torch.cuda.memory.memory_summary() + return "" #TODO: might be cleaner to put this somewhere else import threading diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index cf1b0d441..93d26c690 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -35,6 +35,7 @@ import comfy.model_management import comfy.patcher_extension import comfy.utils from comfy.comfy_types import UnetWrapperFunction +from comfy.quant_ops import QuantizedTensor from comfy.patcher_extension import CallbacksMP, PatcherInjection, WrappersMP @@ -126,27 +127,23 @@ class LowVramPatch: def __init__(self, key, patches, convert_func=None, set_func=None): self.key = key self.patches = patches - self.convert_func = convert_func + self.convert_func = convert_func # TODO: remove self.set_func = set_func def __call__(self, weight): - intermediate_dtype = weight.dtype - if self.convert_func is not None: - weight = self.convert_func(weight.to(dtype=torch.float32, copy=True), inplace=True) + return comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=weight.dtype) - if intermediate_dtype not in [torch.float32, torch.float16, torch.bfloat16]: #intermediate_dtype has to be one that is supported in math ops - intermediate_dtype = torch.float32 - out = comfy.lora.calculate_weight(self.patches[self.key], weight.to(intermediate_dtype), self.key, intermediate_dtype=intermediate_dtype) - if self.set_func is None: - return comfy.float.stochastic_rounding(out, weight.dtype, seed=string_to_seed(self.key)) - else: - return self.set_func(out, seed=string_to_seed(self.key), return_weight=True) +LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR = 2 - out = comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=intermediate_dtype) - if self.set_func is not None: - return self.set_func(out, seed=string_to_seed(self.key), return_weight=True).to(dtype=intermediate_dtype) - else: - return out +def low_vram_patch_estimate_vram(model, key): + weight, set_func, convert_func = get_key_weight(model, key) + if weight is None: + return 0 + model_dtype = getattr(model, "manual_cast_dtype", torch.float32) + if model_dtype is None: + model_dtype = weight.dtype + + return weight.numel() * model_dtype.itemsize * LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR def get_key_weight(model, key): set_func = None @@ -231,7 +228,6 @@ class ModelPatcher: self.object_patches_backup = {} self.weight_wrapper_patches = {} self.model_options = {"transformer_options":{}} - self.model_size() self.load_device = load_device self.offload_device = offload_device self.weight_inplace_update = weight_inplace_update @@ -270,6 +266,9 @@ class ModelPatcher: if not hasattr(self.model, 'current_weight_patches_uuid'): self.model.current_weight_patches_uuid = None + if not hasattr(self.model, 'model_offload_buffer_memory'): + self.model.model_offload_buffer_memory = 0 + def model_size(self): if self.size > 0: return self.size @@ -286,7 +285,7 @@ class ModelPatcher: return self.model.lowvram_patch_counter def clone(self): - n = self.__class__(self.model, self.load_device, self.offload_device, self.size, weight_inplace_update=self.weight_inplace_update) + n = self.__class__(self.model, self.load_device, self.offload_device, self.model_size(), weight_inplace_update=self.weight_inplace_update) n.patches = {} for k in self.patches: n.patches[k] = self.patches[k][:] @@ -455,6 +454,9 @@ class ModelPatcher: def set_model_post_input_patch(self, patch): self.set_model_patch(patch, "post_input") + def set_model_noise_refiner_patch(self, patch): + self.set_model_patch(patch, "noise_refiner") + def set_model_rope_options(self, scale_x, shift_x, scale_y, shift_y, scale_t, shift_t, **kwargs): rope_options = self.model_options["transformer_options"].get("rope_options", {}) rope_options["scale_x"] = scale_x @@ -619,10 +621,11 @@ class ModelPatcher: if key not in self.backup: self.backup[key] = collections.namedtuple('Dimension', ['weight', 'inplace_update'])(weight.to(device=self.offload_device, copy=inplace_update), inplace_update) + temp_dtype = comfy.model_management.lora_compute_dtype(device_to) if device_to is not None: - temp_weight = comfy.model_management.cast_to_device(weight, device_to, torch.float32, copy=True) + temp_weight = comfy.model_management.cast_to_device(weight, device_to, temp_dtype, copy=True) else: - temp_weight = weight.to(torch.float32, copy=True) + temp_weight = weight.to(temp_dtype, copy=True) if convert_func is not None: temp_weight = convert_func(temp_weight, inplace=True) @@ -663,7 +666,22 @@ class ModelPatcher: skip = True # skip random weights in non leaf modules break if not skip and (hasattr(m, "comfy_cast_weights") or len(params) > 0): - loading.append((comfy.model_management.module_size(m), n, m, params)) + module_mem = comfy.model_management.module_size(m) + module_offload_mem = module_mem + if hasattr(m, "comfy_cast_weights"): + def check_module_offload_mem(key): + if key in self.patches: + return low_vram_patch_estimate_vram(self.model, key) + model_dtype = getattr(self.model, "manual_cast_dtype", None) + weight, _, _ = get_key_weight(self.model, key) + if model_dtype is None or weight is None: + return 0 + if (weight.dtype != model_dtype or isinstance(weight, QuantizedTensor)): + return weight.numel() * model_dtype.itemsize + return 0 + module_offload_mem += check_module_offload_mem("{}.weight".format(n)) + module_offload_mem += check_module_offload_mem("{}.bias".format(n)) + loading.append((module_offload_mem, module_mem, n, m, params)) return loading def load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False, full_load=False): @@ -677,20 +695,22 @@ class ModelPatcher: load_completely = [] offloaded = [] + offload_buffer = 0 loading.sort(reverse=True) - for x in loading: - n = x[1] - m = x[2] - params = x[3] - module_mem = x[0] + for i, x in enumerate(loading): + module_offload_mem, module_mem, n, m, params = x lowvram_weight = False + potential_offload = max(offload_buffer, module_offload_mem + sum([ x1[1] for x1 in loading[i+1:i+1+comfy.model_management.NUM_STREAMS]])) + lowvram_fits = mem_counter + module_mem + potential_offload < lowvram_model_memory + weight_key = "{}.weight".format(n) bias_key = "{}.bias".format(n) if not full_load and hasattr(m, "comfy_cast_weights"): - if mem_counter + module_mem >= lowvram_model_memory: + if not lowvram_fits: + offload_buffer = potential_offload lowvram_weight = True lowvram_counter += 1 lowvram_mem_counter += module_mem @@ -724,9 +744,11 @@ class ModelPatcher: if hasattr(m, "comfy_cast_weights"): wipe_lowvram_weight(m) - if full_load or mem_counter + module_mem < lowvram_model_memory: + if full_load or lowvram_fits: mem_counter += module_mem load_completely.append((module_mem, n, m, params)) + else: + offload_buffer = potential_offload if cast_weight and hasattr(m, "comfy_cast_weights"): m.prev_comfy_cast_weights = m.comfy_cast_weights @@ -753,6 +775,8 @@ class ModelPatcher: key = "{}.{}".format(n, param) self.unpin_weight(key) self.patch_weight_to_device(key, device_to=device_to) + if comfy.model_management.is_device_cuda(device_to): + torch.cuda.synchronize() logging.debug("lowvram: loaded module regularly {} {}".format(n, m)) m.comfy_patched_weights = True @@ -767,7 +791,7 @@ class ModelPatcher: self.pin_weight_to_device("{}.{}".format(n, param)) if lowvram_counter > 0: - logging.info("loaded partially; {:.2f} MB usable, {:.2f} MB loaded, {:.2f} MB offloaded, lowvram patches: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), patch_counter)) + logging.info("loaded partially; {:.2f} MB usable, {:.2f} MB loaded, {:.2f} MB offloaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), offload_buffer / (1024 * 1024), patch_counter)) self.model.model_lowvram = True else: logging.info("loaded completely; {:.2f} MB usable, {:.2f} MB loaded, full load: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), full_load)) @@ -779,6 +803,7 @@ class ModelPatcher: self.model.lowvram_patch_counter += patch_counter self.model.device = device_to self.model.model_loaded_weight_memory = mem_counter + self.model.model_offload_buffer_memory = offload_buffer self.model.current_weight_patches_uuid = self.patches_uuid for callback in self.get_all_callbacks(CallbacksMP.ON_LOAD): @@ -832,6 +857,7 @@ class ModelPatcher: self.model.to(device_to) self.model.device = device_to self.model.model_loaded_weight_memory = 0 + self.model.model_offload_buffer_memory = 0 for m in self.model.modules(): if hasattr(m, "comfy_patched_weights"): @@ -850,13 +876,18 @@ class ModelPatcher: patch_counter = 0 unload_list = self._load_list() unload_list.sort() + + offload_buffer = self.model.model_offload_buffer_memory + if len(unload_list) > 0: + NS = comfy.model_management.NUM_STREAMS + offload_weight_factor = [ min(offload_buffer / (NS + 1), unload_list[0][1]) ] * NS + for unload in unload_list: - if memory_to_free < memory_freed: + if memory_to_free + offload_buffer - self.model.model_offload_buffer_memory < memory_freed: break - module_mem = unload[0] - n = unload[1] - m = unload[2] - params = unload[3] + module_offload_mem, module_mem, n, m, params = unload + + potential_offload = module_offload_mem + sum(offload_weight_factor) lowvram_possible = hasattr(m, "comfy_cast_weights") if hasattr(m, "comfy_patched_weights") and m.comfy_patched_weights == True: @@ -902,20 +933,25 @@ class ModelPatcher: patch_counter += 1 cast_weight = True - if cast_weight: + if cast_weight and hasattr(m, "comfy_cast_weights"): m.prev_comfy_cast_weights = m.comfy_cast_weights m.comfy_cast_weights = True m.comfy_patched_weights = False memory_freed += module_mem + offload_buffer = max(offload_buffer, potential_offload) + offload_weight_factor.append(module_mem) + offload_weight_factor.pop(0) logging.debug("freed {}".format(n)) for param in params: self.pin_weight_to_device("{}.{}".format(n, param)) + self.model.model_lowvram = True self.model.lowvram_patch_counter += patch_counter self.model.model_loaded_weight_memory -= memory_freed - logging.info("loaded partially: {:.2f} MB loaded, lowvram patches: {}".format(self.model.model_loaded_weight_memory / (1024 * 1024), self.model.lowvram_patch_counter)) + self.model.model_offload_buffer_memory = offload_buffer + logging.info("Unloaded partially: {:.2f} MB freed, {:.2f} MB remains loaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(memory_freed / (1024 * 1024), self.model.model_loaded_weight_memory / (1024 * 1024), offload_buffer / (1024 * 1024), self.model.lowvram_patch_counter)) return memory_freed def partially_load(self, device_to, extra_memory=0, force_patch_weights=False): diff --git a/comfy/ops.py b/comfy/ops.py index 2a90a5ba2..cd536e22d 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -22,7 +22,7 @@ import comfy.model_management from comfy.cli_args import args, PerformanceFeature import comfy.float import comfy.rmsnorm -import contextlib +import json def run_every_op(): if torch.compiler.is_compiling(): @@ -58,7 +58,8 @@ except (ModuleNotFoundError, TypeError): NVIDIA_MEMORY_CONV_BUG_WORKAROUND = False try: if comfy.model_management.is_nvidia(): - if torch.backends.cudnn.version() >= 91002 and comfy.model_management.torch_version_numeric >= (2, 9) and comfy.model_management.torch_version_numeric <= (2, 10): + cudnn_version = torch.backends.cudnn.version() + if (cudnn_version >= 91002 and cudnn_version < 91500) and comfy.model_management.torch_version_numeric >= (2, 9) and comfy.model_management.torch_version_numeric <= (2, 10): #TODO: change upper bound version once it's fixed' NVIDIA_MEMORY_CONV_BUG_WORKAROUND = True logging.info("working around nvidia conv3d memory bug.") @@ -78,7 +79,7 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of if input is not None: if dtype is None: if isinstance(input, QuantizedTensor): - dtype = input._layout_params["orig_dtype"] + dtype = input.params.orig_dtype else: dtype = input.dtype if bias_dtype is None: @@ -92,11 +93,6 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of else: offload_stream = None - if offload_stream is not None: - wf_context = offload_stream - else: - wf_context = contextlib.nullcontext() - non_blocking = comfy.model_management.device_supports_non_blocking(device) weight_has_function = len(s.weight_function) > 0 @@ -108,20 +104,24 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of if s.bias is not None: bias = comfy.model_management.cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=bias_has_function, stream=offload_stream) - if bias_has_function: - with wf_context: - for f in s.bias_function: - bias = f(bias) + comfy.model_management.sync_stream(device, offload_stream) + + bias_a = bias + weight_a = weight + + if s.bias is not None: + for f in s.bias_function: + bias = f(bias) if weight_has_function or weight.dtype != dtype: - with wf_context: - weight = weight.to(dtype=dtype) - for f in s.weight_function: - weight = f(weight) + weight = weight.to(dtype=dtype) + if isinstance(weight, QuantizedTensor): + weight = weight.dequantize() + for f in s.weight_function: + weight = f(weight) - comfy.model_management.sync_stream(device, offload_stream) if offloadable: - return weight, bias, offload_stream + return weight, bias, (offload_stream, weight_a, bias_a) else: #Legacy function signature return weight, bias @@ -130,13 +130,16 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of def uncast_bias_weight(s, weight, bias, offload_stream): if offload_stream is None: return - if weight is not None: - device = weight.device + os, weight_a, bias_a = offload_stream + if os is None: + return + if weight_a is not None: + device = weight_a.device else: - if bias is None: + if bias_a is None: return - device = bias.device - offload_stream.wait_stream(comfy.model_management.current_stream(device)) + device = bias_a.device + os.wait_stream(comfy.model_management.current_stream(device)) class CastWeightBiasOp: @@ -409,36 +412,34 @@ def fp8_linear(self, input): return None input_dtype = input.dtype + input_shape = input.shape + tensor_3d = input.ndim == 3 - if input.ndim == 3 or input.ndim == 2: - w, bias, offload_stream = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype, offloadable=True) + if tensor_3d: + input = input.reshape(-1, input_shape[2]) - scale_weight = self.scale_weight - scale_input = self.scale_input - if scale_weight is None: - scale_weight = torch.ones((), device=input.device, dtype=torch.float32) - else: - scale_weight = scale_weight.to(input.device) + if input.ndim != 2: + return None + w, bias, offload_stream = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype, offloadable=True) + scale_weight = torch.ones((), device=input.device, dtype=torch.float32) - if scale_input is None: - scale_input = torch.ones((), device=input.device, dtype=torch.float32) - input = torch.clamp(input, min=-448, max=448, out=input) - layout_params_weight = {'scale': scale_input, 'orig_dtype': input_dtype} - quantized_input = QuantizedTensor(input.to(dtype).contiguous(), "TensorCoreFP8Layout", layout_params_weight) - else: - scale_input = scale_input.to(input.device) - quantized_input = QuantizedTensor.from_float(input, "TensorCoreFP8Layout", scale=scale_input, dtype=dtype) + scale_input = torch.ones((), device=input.device, dtype=torch.float32) + input = torch.clamp(input, min=-448, max=448, out=input) + input_fp8 = input.to(dtype).contiguous() + layout_params_input = TensorCoreFP8Layout.Params(scale=scale_input, orig_dtype=input_dtype, orig_shape=tuple(input_fp8.shape)) + quantized_input = QuantizedTensor(input_fp8, "TensorCoreFP8Layout", layout_params_input) - # Wrap weight in QuantizedTensor - this enables unified dispatch - # Call F.linear - __torch_dispatch__ routes to fp8_linear handler in quant_ops.py! - layout_params_weight = {'scale': scale_weight, 'orig_dtype': input_dtype} - quantized_weight = QuantizedTensor(w, "TensorCoreFP8Layout", layout_params_weight) - o = torch.nn.functional.linear(quantized_input, quantized_weight, bias) + # Wrap weight in QuantizedTensor - this enables unified dispatch + # Call F.linear - __torch_dispatch__ routes to fp8_linear handler in quant_ops.py! + layout_params_weight = TensorCoreFP8Layout.Params(scale=scale_weight, orig_dtype=input_dtype, orig_shape=tuple(w.shape)) + quantized_weight = QuantizedTensor(w, "TensorCoreFP8Layout", layout_params_weight) + o = torch.nn.functional.linear(quantized_input, quantized_weight, bias) - uncast_bias_weight(self, w, bias, offload_stream) - return o + uncast_bias_weight(self, w, bias, offload_stream) + if tensor_3d: + o = o.reshape((input_shape[0], input_shape[1], w.shape[0])) - return None + return o class fp8_ops(manual_cast): class Linear(manual_cast.Linear): @@ -448,7 +449,7 @@ class fp8_ops(manual_cast): return None def forward_comfy_cast_weights(self, input): - if not self.training: + if len(self.weight_function) == 0 and len(self.bias_function) == 0: try: out = fp8_linear(self, input) if out is not None: @@ -461,59 +462,6 @@ class fp8_ops(manual_cast): uncast_bias_weight(self, weight, bias, offload_stream) return x -def scaled_fp8_ops(fp8_matrix_mult=False, scale_input=False, override_dtype=None): - logging.info("Using scaled fp8: fp8 matrix mult: {}, scale input: {}".format(fp8_matrix_mult, scale_input)) - class scaled_fp8_op(manual_cast): - class Linear(manual_cast.Linear): - def __init__(self, *args, **kwargs): - if override_dtype is not None: - kwargs['dtype'] = override_dtype - super().__init__(*args, **kwargs) - - def reset_parameters(self): - if not hasattr(self, 'scale_weight'): - self.scale_weight = torch.nn.parameter.Parameter(data=torch.ones((), device=self.weight.device, dtype=torch.float32), requires_grad=False) - - if not scale_input: - self.scale_input = None - - if not hasattr(self, 'scale_input'): - self.scale_input = torch.nn.parameter.Parameter(data=torch.ones((), device=self.weight.device, dtype=torch.float32), requires_grad=False) - return None - - def forward_comfy_cast_weights(self, input): - if fp8_matrix_mult: - out = fp8_linear(self, input) - if out is not None: - return out - - weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True) - - if weight.numel() < input.numel(): #TODO: optimize - x = torch.nn.functional.linear(input, weight * self.scale_weight.to(device=weight.device, dtype=weight.dtype), bias) - else: - x = torch.nn.functional.linear(input * self.scale_weight.to(device=weight.device, dtype=weight.dtype), weight, bias) - uncast_bias_weight(self, weight, bias, offload_stream) - return x - - def convert_weight(self, weight, inplace=False, **kwargs): - if inplace: - weight *= self.scale_weight.to(device=weight.device, dtype=weight.dtype) - return weight - else: - return weight * self.scale_weight.to(device=weight.device, dtype=weight.dtype) - - def set_weight(self, weight, inplace_update=False, seed=None, return_weight=False, **kwargs): - weight = comfy.float.stochastic_rounding(weight / self.scale_weight.to(device=weight.device, dtype=weight.dtype), self.weight.dtype, seed=seed) - if return_weight: - return weight - if inplace_update: - self.weight.data.copy_(weight) - else: - self.weight = torch.nn.Parameter(weight, requires_grad=False) - - return scaled_fp8_op - CUBLAS_IS_AVAILABLE = False try: from cublas_ops import CublasLinear @@ -537,119 +485,249 @@ if CUBLAS_IS_AVAILABLE: # ============================================================================== # Mixed Precision Operations # ============================================================================== -from .quant_ops import QuantizedTensor, QUANT_ALGOS +from .quant_ops import ( + QuantizedTensor, + QUANT_ALGOS, + TensorCoreFP8Layout, + get_layout_class, +) -class MixedPrecisionOps(disable_weight_init): - _layer_quant_config = {} - _compute_dtype = torch.bfloat16 - class Linear(torch.nn.Module, CastWeightBiasOp): - def __init__( - self, - in_features: int, - out_features: int, - bias: bool = True, - device=None, - dtype=None, - ) -> None: - super().__init__() +def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_precision_mm=False, disabled=[]): + class MixedPrecisionOps(manual_cast): + _quant_config = quant_config + _compute_dtype = compute_dtype + _full_precision_mm = full_precision_mm + _disabled = disabled - self.factory_kwargs = {"device": device, "dtype": MixedPrecisionOps._compute_dtype} - # self.factory_kwargs = {"device": device, "dtype": dtype} + class Linear(torch.nn.Module, CastWeightBiasOp): + def __init__( + self, + in_features: int, + out_features: int, + bias: bool = True, + device=None, + dtype=None, + ) -> None: + super().__init__() - self.in_features = in_features - self.out_features = out_features - if bias: - self.bias = torch.nn.Parameter(torch.empty(out_features, **self.factory_kwargs)) - else: - self.register_parameter("bias", None) + self.factory_kwargs = {"device": device, "dtype": MixedPrecisionOps._compute_dtype} + # self.factory_kwargs = {"device": device, "dtype": dtype} - self.tensor_class = None + self.in_features = in_features + self.out_features = out_features + if bias: + self.bias = torch.nn.Parameter(torch.empty(out_features, **self.factory_kwargs)) + else: + self.register_parameter("bias", None) - def reset_parameters(self): - return None + self.tensor_class = None + self._full_precision_mm = MixedPrecisionOps._full_precision_mm + self._full_precision_mm_config = False - def _load_from_state_dict(self, state_dict, prefix, local_metadata, - strict, missing_keys, unexpected_keys, error_msgs): + def reset_parameters(self): + return None - device = self.factory_kwargs["device"] - layer_name = prefix.rstrip('.') - weight_key = f"{prefix}weight" - weight = state_dict.pop(weight_key, None) - if weight is None: - raise ValueError(f"Missing weight for layer {layer_name}") + def _load_scale_param(self, state_dict, prefix, param_name, device, manually_loaded_keys, dtype=None): + key = f"{prefix}{param_name}" + value = state_dict.pop(key, None) + if value is not None: + value = value.to(device=device) + if dtype is not None: + value = value.view(dtype=dtype) + manually_loaded_keys.append(key) + return value - manually_loaded_keys = [weight_key] + def _load_from_state_dict(self, state_dict, prefix, local_metadata, + strict, missing_keys, unexpected_keys, error_msgs): - if layer_name not in MixedPrecisionOps._layer_quant_config: - self.weight = torch.nn.Parameter(weight.to(device=device, dtype=MixedPrecisionOps._compute_dtype), requires_grad=False) - else: - quant_format = MixedPrecisionOps._layer_quant_config[layer_name].get("format", None) - if quant_format is None: - raise ValueError(f"Unknown quantization format for layer {layer_name}") + device = self.factory_kwargs["device"] + layer_name = prefix.rstrip('.') + weight_key = f"{prefix}weight" + weight = state_dict.pop(weight_key, None) + if weight is None: + raise ValueError(f"Missing weight for layer {layer_name}") - qconfig = QUANT_ALGOS[quant_format] - self.layout_type = qconfig["comfy_tensor_layout"] + manually_loaded_keys = [weight_key] - weight_scale_key = f"{prefix}weight_scale" - layout_params = { - 'scale': state_dict.pop(weight_scale_key, None), - 'orig_dtype': MixedPrecisionOps._compute_dtype, - 'block_size': qconfig.get("group_size", None), - } - if layout_params['scale'] is not None: - manually_loaded_keys.append(weight_scale_key) + layer_conf = state_dict.pop(f"{prefix}comfy_quant", None) + if layer_conf is not None: + layer_conf = json.loads(layer_conf.numpy().tobytes()) - self.weight = torch.nn.Parameter( - QuantizedTensor(weight.to(device=device), self.layout_type, layout_params), - requires_grad=False - ) + if layer_conf is None: + self.weight = torch.nn.Parameter(weight.to(device=device, dtype=MixedPrecisionOps._compute_dtype), requires_grad=False) + else: + self.quant_format = layer_conf.get("format", None) + self._full_precision_mm_config = layer_conf.get("full_precision_matrix_mult", False) + if not self._full_precision_mm: + self._full_precision_mm = self._full_precision_mm_config - for param_name in qconfig["parameters"]: - param_key = f"{prefix}{param_name}" - _v = state_dict.pop(param_key, None) - if _v is None: + if self.quant_format in MixedPrecisionOps._disabled: + self._full_precision_mm = True + + if self.quant_format is None: + raise ValueError(f"Unknown quantization format for layer {layer_name}") + + qconfig = QUANT_ALGOS[self.quant_format] + self.layout_type = qconfig["comfy_tensor_layout"] + layout_cls = get_layout_class(self.layout_type) + + # Load format-specific parameters + if self.quant_format in ["float8_e4m3fn", "float8_e5m2"]: + # FP8: single tensor scale + scale = self._load_scale_param(state_dict, prefix, "weight_scale", device, manually_loaded_keys) + + params = layout_cls.Params( + scale=scale, + orig_dtype=MixedPrecisionOps._compute_dtype, + orig_shape=(self.out_features, self.in_features), + ) + + elif self.quant_format == "nvfp4": + # NVFP4: tensor_scale (weight_scale_2) + block_scale (weight_scale) + tensor_scale = self._load_scale_param(state_dict, prefix, "weight_scale_2", device, manually_loaded_keys) + block_scale = self._load_scale_param(state_dict, prefix, "weight_scale", device, manually_loaded_keys, + dtype=torch.float8_e4m3fn) + + if tensor_scale is None or block_scale is None: + raise ValueError(f"Missing NVFP4 scales for layer {layer_name}") + + params = layout_cls.Params( + scale=tensor_scale, + block_scale=block_scale, + orig_dtype=MixedPrecisionOps._compute_dtype, + orig_shape=(self.out_features, self.in_features), + ) + else: + raise ValueError(f"Unsupported quantization format: {self.quant_format}") + + self.weight = torch.nn.Parameter( + QuantizedTensor(weight.to(device=device, dtype=qconfig["storage_t"]), self.layout_type, params), + requires_grad=False + ) + + for param_name in qconfig["parameters"]: + if param_name in {"weight_scale", "weight_scale_2"}: + continue # Already handled above + + param_key = f"{prefix}{param_name}" + _v = state_dict.pop(param_key, None) + if _v is None: + continue + self.register_parameter(param_name, torch.nn.Parameter(_v.to(device=device), requires_grad=False)) + manually_loaded_keys.append(param_key) + + super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) + + for key in manually_loaded_keys: + if key in missing_keys: + missing_keys.remove(key) + + def state_dict(self, *args, destination=None, prefix="", **kwargs): + sd = super().state_dict(*args, destination=destination, prefix=prefix, **kwargs) + if isinstance(self.weight, QuantizedTensor): + layout_cls = self.weight._layout_cls + + # Check if it's any FP8 variant (E4M3 or E5M2) + if layout_cls in ("TensorCoreFP8E4M3Layout", "TensorCoreFP8E5M2Layout", "TensorCoreFP8Layout"): + sd["{}weight_scale".format(prefix)] = self.weight._params.scale + elif layout_cls == "TensorCoreNVFP4Layout": + sd["{}weight_scale_2".format(prefix)] = self.weight._params.scale + sd["{}weight_scale".format(prefix)] = self.weight._params.block_scale + + quant_conf = {"format": self.quant_format} + if self._full_precision_mm_config: + quant_conf["full_precision_matrix_mult"] = True + sd["{}comfy_quant".format(prefix)] = torch.tensor(list(json.dumps(quant_conf).encode('utf-8')), dtype=torch.uint8) + return sd + + def _forward(self, input, weight, bias): + return torch.nn.functional.linear(input, weight, bias) + + def forward_comfy_cast_weights(self, input): + weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True) + x = self._forward(input, weight, bias) + uncast_bias_weight(self, weight, bias, offload_stream) + return x + + def forward(self, input, *args, **kwargs): + run_every_op() + + input_shape = input.shape + tensor_3d = input.ndim == 3 + + if self._full_precision_mm or self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0: + return self.forward_comfy_cast_weights(input, *args, **kwargs) + + if (getattr(self, 'layout_type', None) is not None and + not isinstance(input, QuantizedTensor)): + + # Reshape 3D tensors to 2D for quantization (needed for NVFP4 and others) + if tensor_3d: + input = input.reshape(-1, input_shape[2]) + + if input.ndim != 2: + # Fall back to comfy_cast_weights for non-2D tensors + return self.forward_comfy_cast_weights(input.reshape(input_shape), *args, **kwargs) + + # dtype is now implicit in the layout class + input = QuantizedTensor.from_float(input, self.layout_type, scale=getattr(self, 'input_scale', None)) + + output = self._forward(input, self.weight, self.bias) + + # Reshape output back to 3D if input was 3D + if tensor_3d: + output = output.reshape((input_shape[0], input_shape[1], self.weight.shape[0])) + + return output + + def convert_weight(self, weight, inplace=False, **kwargs): + if isinstance(weight, QuantizedTensor): + return weight.dequantize() + else: + return weight + + def set_weight(self, weight, inplace_update=False, seed=None, return_weight=False, **kwargs): + if getattr(self, 'layout_type', None) is not None: + # dtype is now implicit in the layout class + weight = QuantizedTensor.from_float(weight, self.layout_type, scale="recalculate", stochastic_rounding=seed, inplace_ops=True) + else: + weight = weight.to(self.weight.dtype) + if return_weight: + return weight + + assert inplace_update is False # TODO: eventually remove the inplace_update stuff + self.weight = torch.nn.Parameter(weight, requires_grad=False) + + def _apply(self, fn, recurse=True): # This is to get torch.compile + moving weights to another device working + if recurse: + for module in self.children(): + module._apply(fn) + + for key, param in self._parameters.items(): + if param is None: continue - setattr(self, param_name, torch.nn.Parameter(_v.to(device=device), requires_grad=False)) - manually_loaded_keys.append(param_key) + self.register_parameter(key, torch.nn.Parameter(fn(param), requires_grad=False)) + for key, buf in self._buffers.items(): + if buf is not None: + self._buffers[key] = fn(buf) + return self - super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) + return MixedPrecisionOps - for key in manually_loaded_keys: - if key in missing_keys: - missing_keys.remove(key) +def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, model_config=None): + fp8_compute = comfy.model_management.supports_fp8_compute(load_device) # TODO: if we support more ops this needs to be more granular + nvfp4_compute = comfy.model_management.supports_nvfp4_compute(load_device) - def _forward(self, input, weight, bias): - return torch.nn.functional.linear(input, weight, bias) - - def forward_comfy_cast_weights(self, input): - weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True) - x = self._forward(input, weight, bias) - uncast_bias_weight(self, weight, bias, offload_stream) - return x - - def forward(self, input, *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(input, *args, **kwargs) - if (getattr(self, 'layout_type', None) is not None and - getattr(self, 'input_scale', None) is not None and - not isinstance(input, QuantizedTensor)): - input = QuantizedTensor.from_float(input, self.layout_type, scale=self.input_scale, dtype=self.weight.dtype) - return self._forward(input, self.weight, self.bias) - - -def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, scaled_fp8=None, model_config=None): - if model_config and hasattr(model_config, 'layer_quant_config') and model_config.layer_quant_config: - MixedPrecisionOps._layer_quant_config = model_config.layer_quant_config - MixedPrecisionOps._compute_dtype = compute_dtype - logging.info(f"Using mixed precision operations: {len(model_config.layer_quant_config)} quantized layers") - return MixedPrecisionOps - - fp8_compute = comfy.model_management.supports_fp8_compute(load_device) - if scaled_fp8 is not None: - return scaled_fp8_ops(fp8_matrix_mult=fp8_compute and fp8_optimizations, scale_input=fp8_optimizations, override_dtype=scaled_fp8) + if model_config and hasattr(model_config, 'quant_config') and model_config.quant_config: + logging.info("Using mixed precision operations") + disabled = set() + if not nvfp4_compute: + disabled.add("nvfp4") + if not fp8_compute: + disabled.add("float8_e4m3fn") + disabled.add("float8_e5m2") + return mixed_precision_ops(model_config.quant_config, compute_dtype, disabled=disabled) if ( fp8_compute and diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index 1d058bece..5a17bc6f5 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -1,545 +1,140 @@ import torch import logging -from typing import Tuple, Dict -_LAYOUT_REGISTRY = {} -_GENERIC_UTILS = {} +try: + import comfy_kitchen as ck + from comfy_kitchen.tensor import ( + QuantizedTensor, + QuantizedLayout, + TensorCoreFP8Layout as _CKFp8Layout, + TensorCoreNVFP4Layout, # Direct import, no wrapper needed + register_layout_op, + register_layout_class, + get_layout_class, + ) + _CK_AVAILABLE = True + if torch.version.cuda is None: + ck.registry.disable("cuda") + else: + cuda_version = tuple(map(int, str(torch.version.cuda).split('.'))) + if cuda_version < (13,): + ck.registry.disable("cuda") + 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: + logging.error(f"Failed to import comfy_kitchen, Error: {e}, fp8 and fp4 support will not be available.") + _CK_AVAILABLE = False -def register_layout_op(torch_op, layout_type): - """ - Decorator to register a layout-specific operation handler. - Args: - torch_op: PyTorch operation (e.g., torch.ops.aten.linear.default) - layout_type: Layout class (e.g., TensorCoreFP8Layout) - Example: - @register_layout_op(torch.ops.aten.linear.default, TensorCoreFP8Layout) - def fp8_linear(func, args, kwargs): - # FP8-specific linear implementation - ... - """ - def decorator(handler_func): - if torch_op not in _LAYOUT_REGISTRY: - _LAYOUT_REGISTRY[torch_op] = {} - _LAYOUT_REGISTRY[torch_op][layout_type] = handler_func - return handler_func - return decorator + class QuantizedTensor: + pass + class _CKFp8Layout: + pass -def register_generic_util(torch_op): - """ - Decorator to register a generic utility that works for all layouts. - Args: - torch_op: PyTorch operation (e.g., torch.ops.aten.detach.default) + class TensorCoreNVFP4Layout: + pass - Example: - @register_generic_util(torch.ops.aten.detach.default) - def generic_detach(func, args, kwargs): - # Works for any layout - ... - """ - def decorator(handler_func): - _GENERIC_UTILS[torch_op] = handler_func - return handler_func - return decorator + def register_layout_class(name, cls): + pass + def get_layout_class(name): + return None -def _get_layout_from_args(args): - for arg in args: - if isinstance(arg, QuantizedTensor): - return arg._layout_type - elif isinstance(arg, (list, tuple)): - for item in arg: - if isinstance(item, QuantizedTensor): - return item._layout_type - return None - - -def _move_layout_params_to_device(params, device): - new_params = {} - for k, v in params.items(): - if isinstance(v, torch.Tensor): - new_params[k] = v.to(device=device) - else: - new_params[k] = v - return new_params - - -def _copy_layout_params(params): - new_params = {} - for k, v in params.items(): - if isinstance(v, torch.Tensor): - new_params[k] = v.clone() - else: - new_params[k] = v - return new_params - -def _copy_layout_params_inplace(src, dst, non_blocking=False): - for k, v in src.items(): - if isinstance(v, torch.Tensor): - dst[k].copy_(v, non_blocking=non_blocking) - else: - dst[k] = v - -class QuantizedLayout: - """ - Base class for quantization layouts. - - A layout encapsulates the format-specific logic for quantization/dequantization - and provides a uniform interface for extracting raw tensors needed for computation. - - New quantization formats should subclass this and implement the required methods. - """ - @classmethod - def quantize(cls, tensor, **kwargs) -> Tuple[torch.Tensor, Dict]: - raise NotImplementedError(f"{cls.__name__} must implement quantize()") - - @staticmethod - def dequantize(qdata, **layout_params) -> torch.Tensor: - raise NotImplementedError("TensorLayout must implement dequantize()") - - @classmethod - def get_plain_tensors(cls, qtensor) -> torch.Tensor: - raise NotImplementedError(f"{cls.__name__} must implement get_plain_tensors()") - - -class QuantizedTensor(torch.Tensor): - """ - Universal quantized tensor that works with any layout. - - This tensor subclass uses a pluggable layout system to support multiple - quantization formats (FP8, INT4, INT8, etc.) without code duplication. - - The layout_type determines format-specific behavior, while common operations - (detach, clone, to) are handled generically. - - Attributes: - _qdata: The quantized tensor data - _layout_type: Layout class (e.g., TensorCoreFP8Layout) - _layout_params: Dict with layout-specific params (scale, zero_point, etc.) - """ - - @staticmethod - def __new__(cls, qdata, layout_type, layout_params): - """ - Create a quantized tensor. - - Args: - qdata: The quantized data tensor - layout_type: Layout class (subclass of QuantizedLayout) - layout_params: Dict with layout-specific parameters - """ - return torch.Tensor._make_wrapper_subclass(cls, qdata.shape, device=qdata.device, dtype=qdata.dtype, requires_grad=False) - - def __init__(self, qdata, layout_type, layout_params): - self._qdata = qdata - self._layout_type = layout_type - self._layout_params = layout_params - - def __repr__(self): - layout_name = self._layout_type - param_str = ", ".join(f"{k}={v}" for k, v in list(self._layout_params.items())[:2]) - return f"QuantizedTensor(shape={self.shape}, layout={layout_name}, {param_str})" - - @property - def layout_type(self): - return self._layout_type - - def __tensor_flatten__(self): - """ - Tensor flattening protocol for proper device movement. - """ - inner_tensors = ["_qdata"] - ctx = { - "layout_type": self._layout_type, - } - - tensor_params = {} - non_tensor_params = {} - for k, v in self._layout_params.items(): - if isinstance(v, torch.Tensor): - tensor_params[k] = v - else: - non_tensor_params[k] = v - - ctx["tensor_param_keys"] = list(tensor_params.keys()) - ctx["non_tensor_params"] = non_tensor_params - - for k, v in tensor_params.items(): - attr_name = f"_layout_param_{k}" - object.__setattr__(self, attr_name, v) - inner_tensors.append(attr_name) - - return inner_tensors, ctx - - @staticmethod - def __tensor_unflatten__(inner_tensors, ctx, outer_size, outer_stride): - """ - Tensor unflattening protocol for proper device movement. - Reconstructs the QuantizedTensor after device movement. - """ - layout_type = ctx["layout_type"] - layout_params = dict(ctx["non_tensor_params"]) - - for key in ctx["tensor_param_keys"]: - attr_name = f"_layout_param_{key}" - layout_params[key] = inner_tensors[attr_name] - - return QuantizedTensor(inner_tensors["_qdata"], layout_type, layout_params) - - @classmethod - def from_float(cls, tensor, layout_type, **quantize_kwargs) -> 'QuantizedTensor': - qdata, layout_params = LAYOUTS[layout_type].quantize(tensor, **quantize_kwargs) - return cls(qdata, layout_type, layout_params) - - def dequantize(self) -> torch.Tensor: - return LAYOUTS[self._layout_type].dequantize(self._qdata, **self._layout_params) - - @classmethod - def __torch_dispatch__(cls, func, types, args=(), kwargs=None): - kwargs = kwargs or {} - - # Step 1: Check generic utilities first (detach, clone, to, etc.) - if func in _GENERIC_UTILS: - return _GENERIC_UTILS[func](func, args, kwargs) - - # Step 2: Check layout-specific handlers (linear, matmul, etc.) - layout_type = _get_layout_from_args(args) - if layout_type and func in _LAYOUT_REGISTRY: - handler = _LAYOUT_REGISTRY[func].get(layout_type) - if handler: - return handler(func, args, kwargs) - - # Step 3: Fallback to dequantization - if isinstance(args[0] if args else None, QuantizedTensor): - logging.info(f"QuantizedTensor: Unhandled operation {func}, falling back to dequantization. kwargs={kwargs}") - return cls._dequant_and_fallback(func, args, kwargs) - - @classmethod - def _dequant_and_fallback(cls, func, args, kwargs): - def dequant_arg(arg): - if isinstance(arg, QuantizedTensor): - return arg.dequantize() - elif isinstance(arg, (list, tuple)): - return type(arg)(dequant_arg(a) for a in arg) - return arg - - new_args = dequant_arg(args) - new_kwargs = dequant_arg(kwargs) - return func(*new_args, **new_kwargs) - +import comfy.float # ============================================================================== -# Generic Utilities (Layout-Agnostic Operations) +# FP8 Layouts with Comfy-Specific Extensions # ============================================================================== -def _create_transformed_qtensor(qt, transform_fn): - new_data = transform_fn(qt._qdata) - new_params = _copy_layout_params(qt._layout_params) - return QuantizedTensor(new_data, qt._layout_type, new_params) +class _TensorCoreFP8LayoutBase(_CKFp8Layout): + FP8_DTYPE = None # Must be overridden in subclass - -def _handle_device_transfer(qt, target_device, target_dtype=None, target_layout=None, op_name="to"): - if target_dtype is not None and target_dtype != qt.dtype: - logging.warning( - f"QuantizedTensor: dtype conversion requested to {target_dtype}, " - f"but not supported for quantized tensors. Ignoring dtype." - ) - - if target_layout is not None and target_layout != torch.strided: - logging.warning( - f"QuantizedTensor: layout change requested to {target_layout}, " - f"but not supported. Ignoring layout." - ) - - # Handle device transfer - current_device = qt._qdata.device - if target_device is not None: - # Normalize device for comparison - if isinstance(target_device, str): - target_device = torch.device(target_device) - if isinstance(current_device, str): - current_device = torch.device(current_device) - - if target_device != current_device: - logging.debug(f"QuantizedTensor.{op_name}: Moving from {current_device} to {target_device}") - new_q_data = qt._qdata.to(device=target_device) - new_params = _move_layout_params_to_device(qt._layout_params, target_device) - new_qt = QuantizedTensor(new_q_data, qt._layout_type, new_params) - logging.debug(f"QuantizedTensor.{op_name}: Created new tensor on {target_device}") - return new_qt - - logging.debug(f"QuantizedTensor.{op_name}: No device change needed, returning original") - return qt - - -@register_generic_util(torch.ops.aten.detach.default) -def generic_detach(func, args, kwargs): - """Detach operation - creates a detached copy of the quantized tensor.""" - qt = args[0] - if isinstance(qt, QuantizedTensor): - return _create_transformed_qtensor(qt, lambda x: x.detach()) - return func(*args, **kwargs) - - -@register_generic_util(torch.ops.aten.clone.default) -def generic_clone(func, args, kwargs): - """Clone operation - creates a deep copy of the quantized tensor.""" - qt = args[0] - if isinstance(qt, QuantizedTensor): - return _create_transformed_qtensor(qt, lambda x: x.clone()) - return func(*args, **kwargs) - - -@register_generic_util(torch.ops.aten._to_copy.default) -def generic_to_copy(func, args, kwargs): - """Device/dtype transfer operation - handles .to(device) calls.""" - qt = args[0] - if isinstance(qt, QuantizedTensor): - return _handle_device_transfer( - qt, - target_device=kwargs.get('device', None), - target_dtype=kwargs.get('dtype', None), - op_name="_to_copy" - ) - return func(*args, **kwargs) - - -@register_generic_util(torch.ops.aten.to.dtype_layout) -def generic_to_dtype_layout(func, args, kwargs): - """Handle .to(device) calls using the dtype_layout variant.""" - qt = args[0] - if isinstance(qt, QuantizedTensor): - return _handle_device_transfer( - qt, - target_device=kwargs.get('device', None), - target_dtype=kwargs.get('dtype', None), - target_layout=kwargs.get('layout', None), - op_name="to" - ) - return func(*args, **kwargs) - - -@register_generic_util(torch.ops.aten.copy_.default) -def generic_copy_(func, args, kwargs): - qt_dest = args[0] - src = args[1] - non_blocking = args[2] if len(args) > 2 else False - if isinstance(qt_dest, QuantizedTensor): - if isinstance(src, QuantizedTensor): - # Copy from another quantized tensor - qt_dest._qdata.copy_(src._qdata, non_blocking=non_blocking) - qt_dest._layout_type = src._layout_type - _copy_layout_params_inplace(src._layout_params, qt_dest._layout_params, non_blocking=non_blocking) - else: - # Copy from regular tensor - just copy raw data - qt_dest._qdata.copy_(src) - return qt_dest - return func(*args, **kwargs) - - -@register_generic_util(torch.ops.aten._has_compatible_shallow_copy_type.default) -def generic_has_compatible_shallow_copy_type(func, args, kwargs): - return True - - -@register_generic_util(torch.ops.aten.empty_like.default) -def generic_empty_like(func, args, kwargs): - """Empty_like operation - creates an empty tensor with the same quantized structure.""" - qt = args[0] - if isinstance(qt, QuantizedTensor): - # Create empty tensor with same shape and dtype as the quantized data - hp_dtype = kwargs.pop('dtype', qt._layout_params["orig_dtype"]) - new_qdata = torch.empty_like(qt._qdata, **kwargs) - - # Handle device transfer for layout params - target_device = kwargs.get('device', new_qdata.device) - new_params = _move_layout_params_to_device(qt._layout_params, target_device) - - # Update orig_dtype if dtype is specified - new_params['orig_dtype'] = hp_dtype - - return QuantizedTensor(new_qdata, qt._layout_type, new_params) - return func(*args, **kwargs) - -# ============================================================================== -# FP8 Layout + Operation Handlers -# ============================================================================== -class TensorCoreFP8Layout(QuantizedLayout): - """ - Storage format: - - qdata: FP8 tensor (torch.float8_e4m3fn or torch.float8_e5m2) - - scale: Scalar tensor (float32) for dequantization - - orig_dtype: Original dtype before quantization (for casting back) - """ @classmethod - def quantize(cls, tensor, scale=None, dtype=torch.float8_e4m3fn): + def quantize(cls, tensor, scale=None, stochastic_rounding=0, inplace_ops=False): + if cls.FP8_DTYPE is None: + raise NotImplementedError(f"{cls.__name__} must define FP8_DTYPE") + orig_dtype = tensor.dtype + orig_shape = tuple(tensor.shape) + + if isinstance(scale, str) and scale == "recalculate": + scale = torch.amax(tensor.abs()).to(dtype=torch.float32) / torch.finfo(cls.FP8_DTYPE).max + if tensor.dtype not in [torch.float32, torch.bfloat16]: # Prevent scale from being too small + tensor_info = torch.finfo(tensor.dtype) + scale = (1.0 / torch.clamp((1.0 / scale), min=tensor_info.min, max=tensor_info.max)) if scale is None: - scale = torch.amax(tensor.abs()) / torch.finfo(dtype).max - + scale = torch.ones((), device=tensor.device, dtype=torch.float32) if not isinstance(scale, torch.Tensor): - scale = torch.tensor(scale) - scale = scale.to(device=tensor.device, dtype=torch.float32) + scale = torch.tensor(scale, device=tensor.device, dtype=torch.float32) - tensor_scaled = tensor * (1.0 / scale).to(tensor.dtype) - # TODO: uncomment this if it's actually needed because the clamp has a small performance penality' - # lp_amax = torch.finfo(dtype).max - # torch.clamp(tensor_scaled, min=-lp_amax, max=lp_amax, out=tensor_scaled) - qdata = tensor_scaled.to(dtype, memory_format=torch.contiguous_format) + if stochastic_rounding > 0: + if inplace_ops: + tensor *= (1.0 / scale).to(tensor.dtype) + else: + tensor = tensor * (1.0 / scale).to(tensor.dtype) + qdata = comfy.float.stochastic_rounding(tensor, dtype=cls.FP8_DTYPE, seed=stochastic_rounding) + else: + qdata = ck.quantize_per_tensor_fp8(tensor, scale, cls.FP8_DTYPE) - layout_params = { - 'scale': scale, - 'orig_dtype': orig_dtype - } - return qdata, layout_params + params = cls.Params(scale=scale.float(), orig_dtype=orig_dtype, orig_shape=orig_shape) + return qdata, params - @staticmethod - def dequantize(qdata, scale, orig_dtype, **kwargs): - plain_tensor = torch.ops.aten._to_copy.default(qdata, dtype=orig_dtype) - return plain_tensor * scale - @classmethod - def get_plain_tensors(cls, qtensor): - return qtensor._qdata, qtensor._layout_params['scale'] +class TensorCoreFP8E4M3Layout(_TensorCoreFP8LayoutBase): + FP8_DTYPE = torch.float8_e4m3fn + + +class TensorCoreFP8E5M2Layout(_TensorCoreFP8LayoutBase): + FP8_DTYPE = torch.float8_e5m2 + + +# Backward compatibility alias - default to E4M3 +TensorCoreFP8Layout = TensorCoreFP8E4M3Layout + + +# ============================================================================== +# Registry +# ============================================================================== + +register_layout_class("TensorCoreFP8Layout", TensorCoreFP8Layout) +register_layout_class("TensorCoreFP8E4M3Layout", TensorCoreFP8E4M3Layout) +register_layout_class("TensorCoreFP8E5M2Layout", TensorCoreFP8E5M2Layout) +register_layout_class("TensorCoreNVFP4Layout", TensorCoreNVFP4Layout) QUANT_ALGOS = { "float8_e4m3fn": { "storage_t": torch.float8_e4m3fn, "parameters": {"weight_scale", "input_scale"}, - "comfy_tensor_layout": "TensorCoreFP8Layout", + "comfy_tensor_layout": "TensorCoreFP8E4M3Layout", + }, + "float8_e5m2": { + "storage_t": torch.float8_e5m2, + "parameters": {"weight_scale", "input_scale"}, + "comfy_tensor_layout": "TensorCoreFP8E5M2Layout", + }, + "nvfp4": { + "storage_t": torch.uint8, + "parameters": {"weight_scale", "weight_scale_2", "input_scale"}, + "comfy_tensor_layout": "TensorCoreNVFP4Layout", + "group_size": 16, }, } -LAYOUTS = { - "TensorCoreFP8Layout": TensorCoreFP8Layout, -} +# ============================================================================== +# Re-exports for backward compatibility +# ============================================================================== -@register_layout_op(torch.ops.aten.linear.default, "TensorCoreFP8Layout") -def fp8_linear(func, args, kwargs): - input_tensor = args[0] - weight = args[1] - bias = args[2] if len(args) > 2 else None - - if isinstance(input_tensor, QuantizedTensor) and isinstance(weight, QuantizedTensor): - plain_input, scale_a = TensorCoreFP8Layout.get_plain_tensors(input_tensor) - plain_weight, scale_b = TensorCoreFP8Layout.get_plain_tensors(weight) - - out_dtype = kwargs.get("out_dtype") - if out_dtype is None: - out_dtype = input_tensor._layout_params['orig_dtype'] - - weight_t = plain_weight.t() - - tensor_2d = False - if len(plain_input.shape) == 2: - tensor_2d = True - plain_input = plain_input.unsqueeze(1) - - input_shape = plain_input.shape - if len(input_shape) != 3: - return None - - try: - output = torch._scaled_mm( - plain_input.reshape(-1, input_shape[2]).contiguous(), - weight_t, - bias=bias, - scale_a=scale_a, - scale_b=scale_b, - out_dtype=out_dtype, - ) - - if isinstance(output, tuple): # TODO: remove when we drop support for torch 2.4 - output = output[0] - - if not tensor_2d: - output = output.reshape((-1, input_shape[1], weight.shape[0])) - - if output.dtype in [torch.float8_e4m3fn, torch.float8_e5m2]: - output_scale = scale_a * scale_b - output_params = { - 'scale': output_scale, - 'orig_dtype': input_tensor._layout_params['orig_dtype'] - } - return QuantizedTensor(output, "TensorCoreFP8Layout", output_params) - else: - return output - - except Exception as e: - raise RuntimeError(f"FP8 _scaled_mm failed, falling back to dequantization: {e}") - - # Case 2: DQ Fallback - if isinstance(weight, QuantizedTensor): - weight = weight.dequantize() - if isinstance(input_tensor, QuantizedTensor): - input_tensor = input_tensor.dequantize() - - return torch.nn.functional.linear(input_tensor, weight, bias) - -def fp8_mm_(input_tensor, weight, bias=None, out_dtype=None): - if out_dtype is None: - out_dtype = input_tensor._layout_params['orig_dtype'] - - plain_input, scale_a = TensorCoreFP8Layout.get_plain_tensors(input_tensor) - plain_weight, scale_b = TensorCoreFP8Layout.get_plain_tensors(weight) - - output = torch._scaled_mm( - plain_input.contiguous(), - plain_weight, - bias=bias, - scale_a=scale_a, - scale_b=scale_b, - out_dtype=out_dtype, - ) - - if isinstance(output, tuple): # TODO: remove when we drop support for torch 2.4 - output = output[0] - return output - -@register_layout_op(torch.ops.aten.addmm.default, "TensorCoreFP8Layout") -def fp8_addmm(func, args, kwargs): - input_tensor = args[1] - weight = args[2] - bias = args[0] - - if isinstance(input_tensor, QuantizedTensor) and isinstance(weight, QuantizedTensor): - return fp8_mm_(input_tensor, weight, bias=bias, out_dtype=kwargs.get("out_dtype", None)) - - a = list(args) - if isinstance(args[0], QuantizedTensor): - a[0] = args[0].dequantize() - if isinstance(args[1], QuantizedTensor): - a[1] = args[1].dequantize() - if isinstance(args[2], QuantizedTensor): - a[2] = args[2].dequantize() - - return func(*a, **kwargs) - -@register_layout_op(torch.ops.aten.mm.default, "TensorCoreFP8Layout") -def fp8_mm(func, args, kwargs): - input_tensor = args[0] - weight = args[1] - - if isinstance(input_tensor, QuantizedTensor) and isinstance(weight, QuantizedTensor): - return fp8_mm_(input_tensor, weight, bias=None, out_dtype=kwargs.get("out_dtype", None)) - - a = list(args) - if isinstance(args[0], QuantizedTensor): - a[0] = args[0].dequantize() - if isinstance(args[1], QuantizedTensor): - a[1] = args[1].dequantize() - return func(*a, **kwargs) - -@register_layout_op(torch.ops.aten.view.default, "TensorCoreFP8Layout") -@register_layout_op(torch.ops.aten.t.default, "TensorCoreFP8Layout") -def fp8_func(func, args, kwargs): - input_tensor = args[0] - if isinstance(input_tensor, QuantizedTensor): - plain_input, scale_a = TensorCoreFP8Layout.get_plain_tensors(input_tensor) - ar = list(args) - ar[0] = plain_input - return QuantizedTensor(func(*ar, **kwargs), "TensorCoreFP8Layout", input_tensor._layout_params) - return func(*args, **kwargs) +__all__ = [ + "QuantizedTensor", + "QuantizedLayout", + "TensorCoreFP8Layout", + "TensorCoreFP8E4M3Layout", + "TensorCoreFP8E5M2Layout", + "TensorCoreNVFP4Layout", + "QUANT_ALGOS", + "register_layout_op", +] diff --git a/comfy/sampler_helpers.py b/comfy/sampler_helpers.py index e46971afb..9134e6d71 100644 --- a/comfy/sampler_helpers.py +++ b/comfy/sampler_helpers.py @@ -122,20 +122,20 @@ def estimate_memory(model, noise_shape, conds): minimum_memory_required = model.model.memory_required([noise_shape[0]] + list(noise_shape[1:]), cond_shapes=cond_shapes_min) return memory_required, minimum_memory_required -def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None): +def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False): executor = comfy.patcher_extension.WrapperExecutor.new_executor( _prepare_sampling, comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.PREPARE_SAMPLING, model_options, is_model_options=True) ) - return executor.execute(model, noise_shape, conds, model_options=model_options) + return executor.execute(model, noise_shape, conds, model_options=model_options, force_full_load=force_full_load) -def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None): +def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False): real_model: BaseModel = None models, inference_memory = get_additional_models(conds, model.model_dtype()) models += get_additional_models_from_model_options(model_options) models += model.get_nested_additional_models() # TODO: does this require inference_memory update? memory_required, minimum_memory_required = estimate_memory(model, noise_shape, conds) - comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required + inference_memory, minimum_memory_required=minimum_memory_required + inference_memory) + comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required + inference_memory, minimum_memory_required=minimum_memory_required + inference_memory, force_full_load=force_full_load) real_model = model.model return real_model, conds, models diff --git a/comfy/samplers.py b/comfy/samplers.py index fa4640842..1989ef107 100755 --- a/comfy/samplers.py +++ b/comfy/samplers.py @@ -720,7 +720,7 @@ class Sampler: sigma = float(sigmas[0]) return math.isclose(max_sigma, sigma, rel_tol=1e-05) or sigma > max_sigma -KSAMPLER_NAMES = ["euler", "euler_cfg_pp", "euler_ancestral", "euler_ancestral_cfg_pp", "heun", "heunpp2","dpm_2", "dpm_2_ancestral", +KSAMPLER_NAMES = ["euler", "euler_cfg_pp", "euler_ancestral", "euler_ancestral_cfg_pp", "heun", "heunpp2", "exp_heun_2_x0", "exp_heun_2_x0_sde", "dpm_2", "dpm_2_ancestral", "lms", "dpm_fast", "dpm_adaptive", "dpmpp_2s_ancestral", "dpmpp_2s_ancestral_cfg_pp", "dpmpp_sde", "dpmpp_sde_gpu", "dpmpp_2m", "dpmpp_2m_cfg_pp", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_2m_sde_heun", "dpmpp_2m_sde_heun_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm", "ipndm", "ipndm_v", "deis", "res_multistep", "res_multistep_cfg_pp", "res_multistep_ancestral", "res_multistep_ancestral_cfg_pp", @@ -984,9 +984,6 @@ class CFGGuider: self.inner_model, self.conds, self.loaded_models = comfy.sampler_helpers.prepare_sampling(self.model_patcher, noise.shape, self.conds, self.model_options) device = self.model_patcher.load_device - if denoise_mask is not None: - denoise_mask = comfy.sampler_helpers.prepare_mask(denoise_mask, noise.shape, device) - noise = noise.to(device) latent_image = latent_image.to(device) sigmas = sigmas.to(device) @@ -1013,6 +1010,24 @@ class CFGGuider: else: latent_shapes = [latent_image.shape] + if denoise_mask is not None: + if denoise_mask.is_nested: + denoise_masks = denoise_mask.unbind() + denoise_masks = denoise_masks[:len(latent_shapes)] + else: + denoise_masks = [denoise_mask] + + for i in range(len(denoise_masks), len(latent_shapes)): + denoise_masks.append(torch.ones(latent_shapes[i])) + + for i in range(len(denoise_masks)): + denoise_masks[i] = comfy.sampler_helpers.prepare_mask(denoise_masks[i], latent_shapes[i], self.model_patcher.load_device) + + if len(denoise_masks) > 1: + denoise_mask, _ = comfy.utils.pack_latents(denoise_masks) + else: + denoise_mask = denoise_masks[0] + self.conds = {} for k in self.original_conds: self.conds[k] = list(map(lambda a: a.copy(), self.original_conds[k])) diff --git a/comfy/sd.py b/comfy/sd.py index 9e5ebbf15..32157e18b 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -52,6 +52,11 @@ import comfy.text_encoders.ace import comfy.text_encoders.omnigen2 import comfy.text_encoders.qwen_image import comfy.text_encoders.hunyuan_image +import comfy.text_encoders.z_image +import comfy.text_encoders.ovis +import comfy.text_encoders.kandinsky5 +import comfy.text_encoders.jina_clip_2 +import comfy.text_encoders.newbie import comfy.model_patcher import comfy.lora @@ -59,6 +64,8 @@ import comfy.lora_convert import comfy.hooks import comfy.t2i_adapter.adapter import comfy.taesd.taesd +import comfy.taesd.taehv +import comfy.latent_formats import comfy.ldm.flux.redux @@ -94,7 +101,7 @@ def load_lora_for_models(model, clip, lora, strength_model, strength_clip): class CLIP: - def __init__(self, target=None, embedding_directory=None, no_init=False, tokenizer_data={}, parameters=0, model_options={}): + def __init__(self, target=None, embedding_directory=None, no_init=False, tokenizer_data={}, parameters=0, state_dict=[], model_options={}): if no_init: return params = target.params.copy() @@ -122,9 +129,32 @@ class CLIP: self.tokenizer = tokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data) self.patcher = comfy.model_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) self.patcher.hook_mode = comfy.hooks.EnumHookMode.MinVram self.patcher.is_clip = True self.apply_hooks_to_conds = None + if len(state_dict) > 0: + if isinstance(state_dict, list): + for c in state_dict: + m, u = self.load_sd(c) + if len(m) > 0: + logging.warning("clip missing: {}".format(m)) + + if len(u) > 0: + logging.debug("clip unexpected: {}".format(u)) + else: + m, u = self.load_sd(state_dict, full_model=True) + if len(m) > 0: + m_filter = list(filter(lambda a: ".logit_scale" not in a and ".transformer.text_projection.weight" not in a, m)) + if len(m_filter) > 0: + logging.warning("clip missing: {}".format(m)) + else: + logging.debug("clip missing: {}".format(m)) + + if len(u) > 0: + logging.debug("clip unexpected {}:".format(u)) + if params['device'] == load_device: model_management.load_models_gpu([self.patcher], force_full_load=True) self.layer_idx = None @@ -189,6 +219,7 @@ class CLIP: self.cond_stage_model.set_clip_options({"projected_pooled": False}) self.load_model() + self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device}) all_hooks.reset() self.patcher.patch_hooks(None) if show_pbar: @@ -236,6 +267,7 @@ class CLIP: self.cond_stage_model.set_clip_options({"projected_pooled": False}) self.load_model() + self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device}) o = self.cond_stage_model.encode_token_weights(tokens) cond, pooled = o[:2] if return_dict: @@ -291,6 +323,7 @@ class VAE: self.latent_channels = 4 self.latent_dim = 2 self.output_channels = 3 + self.pad_channel_value = None self.process_input = lambda image: image * 2.0 - 1.0 self.process_output = lambda image: torch.clamp((image + 1.0) / 2.0, min=0.0, max=1.0) self.working_dtypes = [torch.bfloat16, torch.float32] @@ -356,7 +389,7 @@ class VAE: self.memory_used_encode = lambda shape, dtype: (700 * shape[2] * shape[3]) * model_management.dtype_size(dtype) self.memory_used_decode = lambda shape, dtype: (700 * shape[2] * shape[3] * 32 * 32) * model_management.dtype_size(dtype) - elif sd['decoder.conv_in.weight'].shape[1] == 32: + elif sd['decoder.conv_in.weight'].shape[1] == 32 and sd['decoder.conv_in.weight'].ndim == 5: ddconfig = {"block_out_channels": [128, 256, 512, 1024, 1024], "in_channels": 3, "out_channels": 3, "num_res_blocks": 2, "ffactor_spatial": 16, "ffactor_temporal": 4, "downsample_match_channel": True, "upsample_match_channel": True, "refiner_vae": False} self.latent_channels = ddconfig['z_channels'] = sd["decoder.conv_in.weight"].shape[1] self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32] @@ -382,6 +415,17 @@ class VAE: self.upscale_ratio = 4 self.latent_channels = ddconfig['z_channels'] = sd["decoder.conv_in.weight"].shape[1] + if 'decoder.post_quant_conv.weight' in sd: + sd = comfy.utils.state_dict_prefix_replace(sd, {"decoder.post_quant_conv.": "post_quant_conv.", "encoder.quant_conv.": "quant_conv."}) + + if 'bn.running_mean' in sd: + ddconfig["batch_norm_latent"] = True + self.downscale_ratio *= 2 + self.upscale_ratio *= 2 + self.latent_channels *= 4 + old_memory_used_decode = self.memory_used_decode + self.memory_used_decode = lambda shape, dtype: old_memory_used_decode(shape, dtype) * 4.0 + if 'post_quant_conv.weight' in sd: self.first_stage_model = AutoencoderKL(ddconfig=ddconfig, embed_dim=sd['post_quant_conv.weight'].shape[1]) else: @@ -394,6 +438,7 @@ class VAE: self.memory_used_decode = lambda shape, dtype: (1000 * shape[2] * 2048) * model_management.dtype_size(dtype) self.latent_channels = 64 self.output_channels = 2 + self.pad_channel_value = "replicate" self.upscale_ratio = 2048 self.downscale_ratio = 2048 self.latent_dim = 1 @@ -441,20 +486,20 @@ class VAE: elif "decoder.conv_in.conv.weight" in sd and sd['decoder.conv_in.conv.weight'].shape[1] == 32: ddconfig = {"block_out_channels": [128, 256, 512, 1024, 1024], "in_channels": 3, "out_channels": 3, "num_res_blocks": 2, "ffactor_spatial": 16, "ffactor_temporal": 4, "downsample_match_channel": True, "upsample_match_channel": True} ddconfig['z_channels'] = sd["decoder.conv_in.conv.weight"].shape[1] - self.latent_channels = 64 + self.latent_channels = 32 self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 16, 16) self.upscale_index_formula = (4, 16, 16) self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 16, 16) self.downscale_index_formula = (4, 16, 16) self.latent_dim = 3 - self.not_video = True + self.not_video = False self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32] self.first_stage_model = AutoencodingEngine(regularizer_config={'target': "comfy.ldm.models.autoencoder.EmptyRegularizer"}, encoder_config={'target': "comfy.ldm.hunyuan_video.vae_refiner.Encoder", 'params': ddconfig}, decoder_config={'target': "comfy.ldm.hunyuan_video.vae_refiner.Decoder", 'params': ddconfig}) - self.memory_used_encode = lambda shape, dtype: (1400 * shape[-2] * shape[-1]) * model_management.dtype_size(dtype) - self.memory_used_decode = lambda shape, dtype: (1400 * shape[-3] * shape[-2] * shape[-1] * 16 * 16) * model_management.dtype_size(dtype) + self.memory_used_encode = lambda shape, dtype: (1400 * 9 * shape[-2] * shape[-1]) * model_management.dtype_size(dtype) + self.memory_used_decode = lambda shape, dtype: (3600 * 4 * shape[-2] * shape[-1] * 16 * 16) * model_management.dtype_size(dtype) elif "decoder.conv_in.conv.weight" in sd: ddconfig = {'double_z': True, 'z_channels': 4, 'resolution': 256, 'in_channels': 3, 'out_ch': 3, 'ch': 128, 'ch_mult': [1, 2, 4, 4], 'num_res_blocks': 2, 'attn_resolutions': [], 'dropout': 0.0} ddconfig["conv3d"] = True @@ -466,8 +511,10 @@ class VAE: self.latent_dim = 3 self.latent_channels = ddconfig['z_channels'] = sd["decoder.conv_in.conv.weight"].shape[1] self.first_stage_model = AutoencoderKL(ddconfig=ddconfig, embed_dim=sd['post_quant_conv.weight'].shape[1]) - self.memory_used_decode = lambda shape, dtype: (1500 * shape[2] * shape[3] * shape[4] * (4 * 8 * 8)) * model_management.dtype_size(dtype) - self.memory_used_encode = lambda shape, dtype: (900 * max(shape[2], 2) * shape[3] * shape[4]) * model_management.dtype_size(dtype) + #This is likely to significantly over-estimate with single image or low frame counts as the + #implementation is able to completely skip caching. Rework if used as an image only VAE + self.memory_used_decode = lambda shape, dtype: (2800 * min(8, ((shape[2] - 1) * 4) + 1) * shape[3] * shape[4] * (8 * 8)) * model_management.dtype_size(dtype) + self.memory_used_encode = lambda shape, dtype: (1400 * min(9, shape[2]) * shape[3] * shape[4]) * model_management.dtype_size(dtype) self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32] elif "decoder.unpatcher3d.wavelets" in sd: self.upscale_ratio = (lambda a: max(0, a * 8 - 7), 8, 8) @@ -496,17 +543,22 @@ class VAE: self.memory_used_encode = lambda shape, dtype: 3300 * shape[3] * shape[4] * model_management.dtype_size(dtype) self.memory_used_decode = lambda shape, dtype: 8000 * shape[3] * shape[4] * (16 * 16) * model_management.dtype_size(dtype) else: # Wan 2.1 VAE + dim = sd["decoder.head.0.gamma"].shape[0] self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 8, 8) self.upscale_index_formula = (4, 8, 8) self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 8, 8) self.downscale_index_formula = (4, 8, 8) self.latent_dim = 3 self.latent_channels = 16 - ddconfig = {"dim": 96, "z_dim": self.latent_channels, "dim_mult": [1, 2, 4, 4], "num_res_blocks": 2, "attn_scales": [], "temperal_downsample": [False, True, True], "dropout": 0.0} + self.output_channels = sd["encoder.conv1.weight"].shape[1] + self.pad_channel_value = 1.0 + ddconfig = {"dim": dim, "z_dim": self.latent_channels, "dim_mult": [1, 2, 4, 4], "num_res_blocks": 2, "attn_scales": [], "temperal_downsample": [False, True, True], "image_channels": self.output_channels, "dropout": 0.0} self.first_stage_model = comfy.ldm.wan.vae.WanVAE(**ddconfig) self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32] - self.memory_used_encode = lambda shape, dtype: 6000 * shape[3] * shape[4] * model_management.dtype_size(dtype) - self.memory_used_decode = lambda shape, dtype: 7000 * shape[3] * shape[4] * (8 * 8) * model_management.dtype_size(dtype) + self.memory_used_encode = lambda shape, dtype: (1500 if shape[2]<=4 else 6000) * shape[3] * shape[4] * model_management.dtype_size(dtype) + self.memory_used_decode = lambda shape, dtype: (2200 if shape[2]<=4 else 7000) * shape[3] * shape[4] * (8*8) * model_management.dtype_size(dtype) + + # Hunyuan 3d v2 2.0 & 2.1 elif "geo_decoder.cross_attn_decoder.ln_1.bias" in sd: @@ -536,6 +588,7 @@ class VAE: self.memory_used_decode = lambda shape, dtype: (shape[2] * shape[3] * 87000) * model_management.dtype_size(dtype) self.latent_channels = 8 self.output_channels = 2 + self.pad_channel_value = "replicate" self.upscale_ratio = 4096 self.downscale_ratio = 4096 self.latent_dim = 2 @@ -572,6 +625,35 @@ class VAE: self.process_input = lambda audio: audio self.working_dtypes = [torch.float32] self.crop_input = False + elif "decoder.22.bias" in sd: # taehv, taew and lighttae + self.latent_channels = sd["decoder.1.weight"].shape[1] + self.latent_dim = 3 + self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 16, 16) + self.upscale_index_formula = (4, 16, 16) + self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 16, 16) + self.downscale_index_formula = (4, 16, 16) + if self.latent_channels == 48: # Wan 2.2 + self.first_stage_model = comfy.taesd.taehv.TAEHV(latent_channels=self.latent_channels, latent_format=None) # taehv doesn't need scaling + self.process_input = lambda image: (_ for _ in ()).throw(NotImplementedError("This light tae doesn't support encoding currently")) + self.process_output = lambda image: image + self.memory_used_decode = lambda shape, dtype: (1800 * (max(1, (shape[-3] ** 0.7 * 0.1)) * shape[-2] * shape[-1] * 16 * 16) * model_management.dtype_size(dtype)) + elif self.latent_channels == 32 and sd["decoder.22.bias"].shape[0] == 12: # lighttae_hv15 + self.first_stage_model = comfy.taesd.taehv.TAEHV(latent_channels=self.latent_channels, latent_format=comfy.latent_formats.HunyuanVideo15) + self.process_input = lambda image: (_ for _ in ()).throw(NotImplementedError("This light tae doesn't support encoding currently")) + self.memory_used_decode = lambda shape, dtype: (1200 * (max(1, (shape[-3] ** 0.7 * 0.05)) * shape[-2] * shape[-1] * 32 * 32) * model_management.dtype_size(dtype)) + else: + if sd["decoder.1.weight"].dtype == torch.float16: # taehv currently only available in float16, so assume it's not lighttaew2_1 as otherwise state dicts are identical + latent_format=comfy.latent_formats.HunyuanVideo + else: + latent_format=None # lighttaew2_1 doesn't need scaling + self.first_stage_model = comfy.taesd.taehv.TAEHV(latent_channels=self.latent_channels, latent_format=latent_format) + self.process_input = self.process_output = lambda image: image + self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 8, 8) + self.upscale_index_formula = (4, 8, 8) + self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 8, 8) + self.downscale_index_formula = (4, 8, 8) + self.memory_used_encode = lambda shape, dtype: (700 * (max(1, (shape[-3] ** 0.66 * 0.11)) * shape[-2] * shape[-1]) * model_management.dtype_size(dtype)) + self.memory_used_decode = lambda shape, dtype: (50 * (max(1, (shape[-3] ** 0.65 * 0.26)) * shape[-2] * shape[-1] * 32 * 32) * model_management.dtype_size(dtype)) else: logging.warning("WARNING: No VAE weights detected, VAE not initalized.") self.first_stage_model = None @@ -615,17 +697,28 @@ class VAE: raise RuntimeError("ERROR: VAE is invalid: None\n\nIf the VAE is from a checkpoint loader node your checkpoint does not contain a valid VAE.") def vae_encode_crop_pixels(self, pixels): - if not self.crop_input: - return pixels + if self.crop_input: + downscale_ratio = self.spacial_compression_encode() - downscale_ratio = self.spacial_compression_encode() + dims = pixels.shape[1:-1] + for d in range(len(dims)): + x = (dims[d] // downscale_ratio) * downscale_ratio + x_offset = (dims[d] % downscale_ratio) // 2 + if x != dims[d]: + pixels = pixels.narrow(d + 1, x_offset, x) - dims = pixels.shape[1:-1] - for d in range(len(dims)): - x = (dims[d] // downscale_ratio) * downscale_ratio - x_offset = (dims[d] % downscale_ratio) // 2 - if x != dims[d]: - pixels = pixels.narrow(d + 1, x_offset, x) + if pixels.shape[-1] > self.output_channels: + pixels = pixels[..., :self.output_channels] + elif pixels.shape[-1] < self.output_channels: + if self.pad_channel_value is not None: + if isinstance(self.pad_channel_value, str): + mode = self.pad_channel_value + value = None + else: + mode = "constant" + value = self.pad_channel_value + + pixels = torch.nn.functional.pad(pixels, (0, self.output_channels - pixels.shape[-1]), mode=mode, value=value) return pixels def decode_tiled_(self, samples, tile_x=64, tile_y=64, overlap = 16): @@ -696,6 +789,8 @@ class VAE: self.throw_exception_if_invalid() pixel_samples = None do_tile = False + if self.latent_dim == 2 and samples_in.ndim == 5: + samples_in = samples_in[:, :, 0] try: memory_used = self.memory_used_decode(samples_in.shape, self.vae_dtype) model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload) @@ -911,12 +1006,20 @@ class CLIPType(Enum): OMNIGEN2 = 17 QWEN_IMAGE = 18 HUNYUAN_IMAGE = 19 + HUNYUAN_VIDEO_15 = 20 + OVIS = 21 + KANDINSKY5 = 22 + KANDINSKY5_IMAGE = 23 + NEWBIE = 24 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}): clip_data = [] for p in ckpt_paths: - clip_data.append(comfy.utils.load_torch_file(p, safe_load=True)) + sd, metadata = comfy.utils.load_torch_file(p, safe_load=True, return_metadata=True) + if model_options.get("custom_operations", None) is None: + sd, metadata = comfy.utils.convert_old_quants(sd, model_prefix="", metadata=metadata) + clip_data.append(sd) return load_text_encoder_state_dicts(clip_data, embedding_directory=embedding_directory, clip_type=clip_type, model_options=model_options) @@ -934,6 +1037,13 @@ class TEModel(Enum): QWEN25_7B = 11 BYT5_SMALL_GLYPH = 12 GEMMA_3_4B = 13 + MISTRAL3_24B = 14 + MISTRAL3_24B_PRUNED_FLUX2 = 15 + QWEN3_4B = 16 + QWEN3_2B = 17 + GEMMA_3_12B = 18 + JINA_CLIP_2 = 19 + def detect_te_model(sd): if "text_model.encoder.layers.30.mlp.fc1.weight" in sd: @@ -942,6 +1052,8 @@ def detect_te_model(sd): return TEModel.CLIP_H if "text_model.encoder.layers.0.mlp.fc1.weight" in sd: return TEModel.CLIP_L + if "model.encoder.layers.0.mixer.Wqkv.weight" in sd: + return TEModel.JINA_CLIP_2 if "encoder.block.23.layer.1.DenseReluDense.wi_1.weight" in sd: weight = sd["encoder.block.23.layer.1.DenseReluDense.wi_1.weight"] if weight.shape[-1] == 4096: @@ -956,6 +1068,8 @@ 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.47.self_attn.q_norm.weight' in sd: + return TEModel.GEMMA_3_12B if 'model.layers.0.self_attn.q_norm.weight' in sd: return TEModel.GEMMA_3_4B return TEModel.GEMMA_2_2B @@ -966,6 +1080,18 @@ def detect_te_model(sd): if weight.shape[0] == 512: return TEModel.QWEN25_7B if "model.layers.0.post_attention_layernorm.weight" in sd: + weight = sd['model.layers.0.post_attention_layernorm.weight'] + if 'model.layers.0.self_attn.q_norm.weight' in sd: + if weight.shape[0] == 2560: + return TEModel.QWEN3_4B + elif weight.shape[0] == 2048: + return TEModel.QWEN3_2B + if weight.shape[0] == 5120: + if "model.layers.39.post_attention_layernorm.weight" in sd: + return TEModel.MISTRAL3_24B + else: + return TEModel.MISTRAL3_24B_PRUNED_FLUX2 + return TEModel.LLAMA3_8 return None @@ -1015,7 +1141,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip clip_target.clip = comfy.text_encoders.sd3_clip.sd3_clip(clip_l=False, clip_g=True, t5=False) clip_target.tokenizer = comfy.text_encoders.sd3_clip.SD3Tokenizer elif clip_type == CLIPType.HIDREAM: - clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=False, clip_g=True, t5=False, llama=False, dtype_t5=None, dtype_llama=None, t5xxl_scaled_fp8=None, llama_scaled_fp8=None) + clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=False, clip_g=True, t5=False, llama=False, dtype_t5=None, dtype_llama=None) clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer else: clip_target.clip = sdxl_clip.SDXLRefinerClipModel @@ -1039,7 +1165,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None) elif clip_type == CLIPType.HIDREAM: 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, llama_scaled_fp8=None) + clip_l=False, clip_g=False, t5=True, llama=False, dtype_llama=None) clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer else: #CLIPType.MOCHI clip_target.clip = comfy.text_encoders.genmo.mochi_te(**t5xxl_detect(clip_data)) @@ -1068,7 +1194,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None) elif te_model == TEModel.LLAMA3_8: clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**llama_detect(clip_data), - clip_l=False, clip_g=False, t5=False, llama=True, dtype_t5=None, t5xxl_scaled_fp8=None) + clip_l=False, clip_g=False, t5=False, llama=True, dtype_t5=None) clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer elif te_model == TEModel.QWEN25_3B: clip_target.clip = comfy.text_encoders.omnigen2.te(**llama_detect(clip_data)) @@ -1080,13 +1206,26 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip else: clip_target.clip = comfy.text_encoders.qwen_image.te(**llama_detect(clip_data)) clip_target.tokenizer = comfy.text_encoders.qwen_image.QwenImageTokenizer + elif te_model == TEModel.MISTRAL3_24B or te_model == TEModel.MISTRAL3_24B_PRUNED_FLUX2: + clip_target.clip = comfy.text_encoders.flux.flux2_te(**llama_detect(clip_data), pruned=te_model == TEModel.MISTRAL3_24B_PRUNED_FLUX2) + clip_target.tokenizer = comfy.text_encoders.flux.Flux2Tokenizer + tokenizer_data["tekken_model"] = clip_data[0].get("tekken_model", None) + elif te_model == TEModel.QWEN3_4B: + clip_target.clip = comfy.text_encoders.z_image.te(**llama_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.z_image.ZImageTokenizer + elif te_model == TEModel.QWEN3_2B: + clip_target.clip = comfy.text_encoders.ovis.te(**llama_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.ovis.OvisTokenizer + elif te_model == TEModel.JINA_CLIP_2: + clip_target.clip = comfy.text_encoders.jina_clip_2.JinaClip2TextModelWrapper + clip_target.tokenizer = comfy.text_encoders.jina_clip_2.JinaClip2TokenizerWrapper else: # clip_l if clip_type == CLIPType.SD3: clip_target.clip = comfy.text_encoders.sd3_clip.sd3_clip(clip_l=True, clip_g=False, t5=False) clip_target.tokenizer = comfy.text_encoders.sd3_clip.SD3Tokenizer elif clip_type == CLIPType.HIDREAM: - clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=True, clip_g=False, t5=False, llama=False, dtype_t5=None, dtype_llama=None, t5xxl_scaled_fp8=None, llama_scaled_fp8=None) + clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=True, clip_g=False, t5=False, llama=False, dtype_t5=None, dtype_llama=None) clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer else: clip_target.clip = sd1_clip.SD1ClipModel @@ -1126,6 +1265,30 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip elif clip_type == CLIPType.HUNYUAN_IMAGE: clip_target.clip = comfy.text_encoders.hunyuan_image.te(**llama_detect(clip_data)) clip_target.tokenizer = comfy.text_encoders.hunyuan_image.HunyuanImageTokenizer + elif clip_type == CLIPType.HUNYUAN_VIDEO_15: + clip_target.clip = comfy.text_encoders.hunyuan_image.te(**llama_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.hunyuan_video.HunyuanVideo15Tokenizer + elif clip_type == CLIPType.KANDINSKY5: + clip_target.clip = comfy.text_encoders.kandinsky5.te(**llama_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.kandinsky5.Kandinsky5Tokenizer + elif clip_type == CLIPType.KANDINSKY5_IMAGE: + clip_target.clip = comfy.text_encoders.kandinsky5.te(**llama_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.kandinsky5.Kandinsky5TokenizerImage + elif clip_type == CLIPType.LTXV: + clip_target.clip = comfy.text_encoders.lt.ltxav_te(**llama_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.lt.LTXAVGemmaTokenizer + tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None) + elif clip_type == CLIPType.NEWBIE: + clip_target.clip = comfy.text_encoders.newbie.te(**llama_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.newbie.NewBieTokenizer + if "model.layers.0.self_attn.q_norm.weight" in clip_data[0]: + clip_data_gemma = clip_data[0] + clip_data_jina = clip_data[1] + else: + clip_data_gemma = clip_data[1] + clip_data_jina = clip_data[0] + tokenizer_data["gemma_spiece_model"] = clip_data_gemma.get("spiece_model", None) + tokenizer_data["jina_spiece_model"] = clip_data_jina.get("spiece_model", None) else: clip_target.clip = sdxl_clip.SDXLClipModel clip_target.tokenizer = sdxl_clip.SDXLTokenizer @@ -1141,14 +1304,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip parameters += comfy.utils.calculate_parameters(c) tokenizer_data, model_options = comfy.text_encoders.long_clipl.model_options_long_clip(c, tokenizer_data, model_options) - clip = CLIP(clip_target, embedding_directory=embedding_directory, parameters=parameters, tokenizer_data=tokenizer_data, model_options=model_options) - for c in clip_data: - m, u = clip.load_sd(c) - if len(m) > 0: - logging.warning("clip missing: {}".format(m)) - - if len(u) > 0: - logging.debug("clip unexpected: {}".format(u)) + clip = CLIP(clip_target, embedding_directory=embedding_directory, parameters=parameters, tokenizer_data=tokenizer_data, state_dict=clip_data, model_options=model_options) return clip def load_gligen(ckpt_path): @@ -1207,6 +1363,10 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c weight_dtype = comfy.utils.weight_dtype(sd, diffusion_model_prefix) load_device = model_management.get_torch_device() + custom_operations = model_options.get("custom_operations", None) + if custom_operations is None: + sd, metadata = comfy.utils.convert_old_quants(sd, diffusion_model_prefix, metadata=metadata) + model_config = model_detection.model_config_from_unet(sd, diffusion_model_prefix, metadata=metadata) if model_config is None: logging.warning("Warning, This is not a checkpoint file, trying to load it as a diffusion model only.") @@ -1215,18 +1375,22 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c return None return (diffusion_model, None, VAE(sd={}), None) # The VAE object is there to throw an exception if it's actually used' - unet_weight_dtype = list(model_config.supported_inference_dtypes) - if model_config.scaled_fp8 is not None: + if model_config.quant_config is not None: weight_dtype = None - model_config.custom_operations = model_options.get("custom_operations", None) + if custom_operations is not None: + model_config.custom_operations = custom_operations + unet_dtype = model_options.get("dtype", model_options.get("weight_dtype", None)) if unet_dtype is None: unet_dtype = model_management.unet_dtype(model_params=parameters, supported_dtypes=unet_weight_dtype, weight_dtype=weight_dtype) - manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes) + if model_config.quant_config is not None: + manual_cast_dtype = model_management.unet_manual_cast(None, load_device, model_config.supported_inference_dtypes) + else: + manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes) model_config.set_inference_dtype(unet_dtype, manual_cast_dtype) if model_config.clip_vision_prefix is not None: @@ -1244,22 +1408,33 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c vae = VAE(sd=vae_sd, metadata=metadata) if output_clip: + if te_model_options.get("custom_operations", None) is None: + scaled_fp8_list = [] + for k in list(sd.keys()): # Convert scaled fp8 to mixed ops + if k.endswith(".scaled_fp8"): + scaled_fp8_list.append(k[:-len("scaled_fp8")]) + + if len(scaled_fp8_list) > 0: + out_sd = {} + for k in sd: + skip = False + for pref in scaled_fp8_list: + skip = skip or k.startswith(pref) + if not skip: + out_sd[k] = sd[k] + + for pref in scaled_fp8_list: + quant_sd, qmetadata = comfy.utils.convert_old_quants(sd, pref, metadata={}) + for k in quant_sd: + out_sd[k] = quant_sd[k] + sd = out_sd + clip_target = model_config.clip_target(state_dict=sd) if clip_target is not None: clip_sd = model_config.process_clip_state_dict(sd) if len(clip_sd) > 0: parameters = comfy.utils.calculate_parameters(clip_sd) - clip = CLIP(clip_target, embedding_directory=embedding_directory, tokenizer_data=clip_sd, parameters=parameters, model_options=te_model_options) - m, u = clip.load_sd(clip_sd, full_model=True) - if len(m) > 0: - m_filter = list(filter(lambda a: ".logit_scale" not in a and ".transformer.text_projection.weight" not in a, m)) - if len(m_filter) > 0: - logging.warning("clip missing: {}".format(m)) - else: - logging.debug("clip missing: {}".format(m)) - - if len(u) > 0: - logging.debug("clip unexpected {}:".format(u)) + clip = CLIP(clip_target, embedding_directory=embedding_directory, tokenizer_data=clip_sd, parameters=parameters, state_dict=clip_sd, model_options=te_model_options) else: logging.warning("no CLIP/text encoder weights in checkpoint, the text encoder model will not be loaded.") @@ -1306,6 +1481,9 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None): if len(temp_sd) > 0: sd = temp_sd + custom_operations = model_options.get("custom_operations", None) + if custom_operations is None: + sd, metadata = comfy.utils.convert_old_quants(sd, "", metadata=metadata) parameters = comfy.utils.calculate_parameters(sd) weight_dtype = comfy.utils.weight_dtype(sd) @@ -1336,7 +1514,7 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None): offload_device = model_management.unet_offload_device() unet_weight_dtype = list(model_config.supported_inference_dtypes) - if model_config.scaled_fp8 is not None: + if model_config.quant_config is not None: weight_dtype = None if dtype is None: @@ -1344,12 +1522,15 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None): else: unet_dtype = dtype - if model_config.layer_quant_config is not None: + if model_config.quant_config is not None: manual_cast_dtype = model_management.unet_manual_cast(None, load_device, model_config.supported_inference_dtypes) else: manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes) model_config.set_inference_dtype(unet_dtype, manual_cast_dtype) - model_config.custom_operations = model_options.get("custom_operations", model_config.custom_operations) + + if custom_operations is not None: + model_config.custom_operations = custom_operations + if model_options.get("fp8_optimizations", False): model_config.optimizations["fp8"] = True @@ -1388,6 +1569,9 @@ def save_checkpoint(output_path, model, clip=None, vae=None, clip_vision=None, m if vae is not None: vae_sd = vae.get_sd() + if metadata is None: + metadata = {} + model_management.load_models_gpu(load_models, force_patch_weights=True) clip_vision_sd = clip_vision.get_sd() if clip_vision is not None else None sd = model.model.state_dict_for_saving(clip_sd, vae_sd, clip_vision_sd) diff --git a/comfy/sd1_clip.py b/comfy/sd1_clip.py index 3066de2d7..c512ca5d0 100644 --- a/comfy/sd1_clip.py +++ b/comfy/sd1_clip.py @@ -90,7 +90,6 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=True, enable_attention_masks=False, zero_out_masked=False, return_projected_pooled=True, return_attention_masks=False, model_options={}): # clip-vit-base-patch32 super().__init__() - assert layer in self.LAYERS if textmodel_json_config is None: textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_clip_config.json") @@ -108,19 +107,17 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): config[k] = v operations = model_options.get("custom_operations", None) - scaled_fp8 = None + quant_config = model_options.get("quantization_metadata", None) if operations is None: - scaled_fp8 = model_options.get("scaled_fp8", None) - if scaled_fp8 is not None: - operations = comfy.ops.scaled_fp8_ops(fp8_matrix_mult=False, override_dtype=scaled_fp8) + if quant_config is not None: + operations = comfy.ops.mixed_precision_ops(quant_config, dtype, full_precision_mm=True) + logging.info("Using MixedPrecisionOps for text encoder") else: operations = comfy.ops.manual_cast self.operations = operations self.transformer = model_class(config, dtype, device, self.operations) - if scaled_fp8 is not None: - self.transformer.scaled_fp8 = torch.nn.Parameter(torch.tensor([], dtype=scaled_fp8)) self.num_layers = self.transformer.num_layers @@ -138,6 +135,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): self.layer_norm_hidden_state = layer_norm_hidden_state self.return_projected_pooled = return_projected_pooled self.return_attention_masks = return_attention_masks + self.execution_device = None if layer == "hidden": assert layer_idx is not None @@ -154,7 +152,8 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): def set_clip_options(self, options): layer_idx = options.get("layer", self.layer_idx) self.return_projected_pooled = options.get("projected_pooled", self.return_projected_pooled) - if self.layer == "all": + self.execution_device = options.get("execution_device", self.execution_device) + if isinstance(self.layer, list) or self.layer == "all": pass elif layer_idx is None or abs(layer_idx) > self.num_layers: self.layer = "last" @@ -166,6 +165,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): self.layer = self.options_default[0] self.layer_idx = self.options_default[1] self.return_projected_pooled = self.options_default[2] + self.execution_device = None def process_tokens(self, tokens, device): end_token = self.special_tokens.get("end", None) @@ -249,14 +249,20 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): return torch.cat(embeds_out), torch.tensor(attention_masks, device=device, dtype=torch.long), num_tokens, embeds_info def forward(self, tokens): - device = self.transformer.get_input_embeddings().weight.device + if self.execution_device is None: + device = self.transformer.get_input_embeddings().weight.device + else: + device = self.execution_device + embeds, attention_mask, num_tokens, embeds_info = self.process_tokens(tokens, device) attention_mask_model = None if self.enable_attention_masks: attention_mask_model = attention_mask - if self.layer == "all": + if isinstance(self.layer, list): + intermediate_output = self.layer + elif self.layer == "all": intermediate_output = "all" else: intermediate_output = self.layer_idx @@ -460,7 +466,7 @@ def load_embed(embedding_name, embedding_directory, embedding_size, embed_key=No return embed_out class SDTokenizer: - def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=True, min_length=None, pad_token=None, end_token=None, min_padding=None, pad_left=False, tokenizer_data={}, tokenizer_args={}): + def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=True, min_length=None, pad_token=None, end_token=None, min_padding=None, pad_left=False, disable_weights=False, tokenizer_data={}, tokenizer_args={}): if tokenizer_path is None: tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_tokenizer") self.tokenizer = tokenizer_class.from_pretrained(tokenizer_path, **tokenizer_args) @@ -507,6 +513,8 @@ class SDTokenizer: self.embedding_size = embedding_size self.embedding_key = embedding_key + self.disable_weights = disable_weights + def _try_get_embedding(self, embedding_name:str): ''' Takes a potential embedding name and tries to retrieve it. @@ -541,7 +549,7 @@ class SDTokenizer: min_padding = tokenizer_options.get("{}_min_padding".format(self.embedding_key), self.min_padding) text = escape_important(text) - if kwargs.get("disable_weights", False): + if kwargs.get("disable_weights", self.disable_weights): parsed_weights = [(text, 1.0)] else: parsed_weights = token_weights(text, 1.0) diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 4064bdae1..ee9a79001 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -21,11 +21,14 @@ import comfy.text_encoders.ace import comfy.text_encoders.omnigen2 import comfy.text_encoders.qwen_image import comfy.text_encoders.hunyuan_image +import comfy.text_encoders.kandinsky5 +import comfy.text_encoders.z_image from . import supported_models_base from . import latent_formats from . import diffusers_convert +import comfy.model_management class SD15(supported_models_base.BASE): unet_config = { @@ -539,7 +542,7 @@ class SD3(supported_models_base.BASE): unet_extra_config = {} latent_format = latent_formats.SD3 - memory_usage_factor = 1.2 + memory_usage_factor = 1.6 text_encoder_key_prefix = ["text_encoders."] @@ -741,6 +744,37 @@ class FluxSchnell(Flux): out = model_base.Flux(self, model_type=model_base.ModelType.FLOW, device=device) return out +class Flux2(Flux): + unet_config = { + "image_model": "flux2", + } + + sampling_settings = { + "shift": 2.02, + } + + unet_extra_config = {} + latent_format = latent_formats.Flux2 + + supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32] + + vae_key_prefix = ["vae."] + text_encoder_key_prefix = ["text_encoders."] + + def __init__(self, unet_config): + super().__init__(unet_config) + self.memory_usage_factor = self.memory_usage_factor * (2.0 * 2.0) * 2.36 + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.Flux2(self, device=device) + return out + + def clip_target(self, state_dict={}): + return None # TODO + pref = self.text_encoder_key_prefix[0] + t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref)) + return supported_models_base.ClipTarget(comfy.text_encoders.flux.FluxTokenizer, comfy.text_encoders.flux.flux_clip(**t5_detect)) + class GenmoMochi(supported_models_base.BASE): unet_config = { "image_model": "mochi_preview", @@ -802,6 +836,21 @@ class LTXV(supported_models_base.BASE): t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref)) return supported_models_base.ClipTarget(comfy.text_encoders.lt.LTXVT5Tokenizer, comfy.text_encoders.lt.ltxv_te(**t5_detect)) +class LTXAV(LTXV): + unet_config = { + "image_model": "ltxav", + } + + latent_format = latent_formats.LTXAV + + def __init__(self, unet_config): + super().__init__(unet_config) + self.memory_usage_factor = 0.055 # TODO + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.LTXAV(self, device=device) + return out + class HunyuanVideo(supported_models_base.BASE): unet_config = { "image_model": "hunyuan_video", @@ -932,7 +981,7 @@ class CosmosT2IPredict2(supported_models_base.BASE): def __init__(self, unet_config): super().__init__(unet_config) - self.memory_usage_factor = (unet_config.get("model_channels", 2048) / 2048) * 0.9 + self.memory_usage_factor = (unet_config.get("model_channels", 2048) / 2048) * 0.95 def get_model(self, state_dict, prefix="", device=None): out = model_base.CosmosPredict2(self, device=device) @@ -963,7 +1012,7 @@ class Lumina2(supported_models_base.BASE): "shift": 6.0, } - memory_usage_factor = 1.2 + memory_usage_factor = 1.4 unet_extra_config = {} latent_format = latent_formats.Flux @@ -982,6 +1031,32 @@ class Lumina2(supported_models_base.BASE): hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}gemma2_2b.transformer.".format(pref)) return supported_models_base.ClipTarget(comfy.text_encoders.lumina2.LuminaTokenizer, comfy.text_encoders.lumina2.te(**hunyuan_detect)) +class ZImage(Lumina2): + unet_config = { + "image_model": "lumina2", + "dim": 3840, + } + + sampling_settings = { + "multiplier": 1.0, + "shift": 3.0, + } + + memory_usage_factor = 2.0 + + supported_inference_dtypes = [torch.bfloat16, torch.float32] + + def __init__(self, unet_config): + super().__init__(unet_config) + if comfy.model_management.extended_fp16_support(): + self.supported_inference_dtypes = self.supported_inference_dtypes.copy() + self.supported_inference_dtypes.insert(1, torch.float16) + + def clip_target(self, state_dict={}): + pref = self.text_encoder_key_prefix[0] + hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_4b.transformer.".format(pref)) + return supported_models_base.ClipTarget(comfy.text_encoders.z_image.ZImageTokenizer, comfy.text_encoders.z_image.te(**hunyuan_detect)) + class WAN21_T2V(supported_models_base.BASE): unet_config = { "image_model": "wan2.1", @@ -1236,7 +1311,7 @@ class ChromaRadiance(Chroma): latent_format = comfy.latent_formats.ChromaRadiance # Pixel-space model, no spatial compression for model input. - memory_usage_factor = 0.038 + memory_usage_factor = 0.044 def get_model(self, state_dict, prefix="", device=None): return model_base.ChromaRadiance(self, device=device) @@ -1279,7 +1354,7 @@ class Omnigen2(supported_models_base.BASE): "shift": 2.6, } - memory_usage_factor = 1.65 #TODO + memory_usage_factor = 1.95 #TODO unet_extra_config = {} latent_format = latent_formats.Flux @@ -1344,7 +1419,7 @@ class HunyuanImage21(HunyuanVideo): latent_format = latent_formats.HunyuanImage21 - memory_usage_factor = 7.7 + memory_usage_factor = 8.7 supported_inference_dtypes = [torch.bfloat16, torch.float32] @@ -1374,6 +1449,108 @@ class HunyuanImage21Refiner(HunyuanVideo): out = model_base.HunyuanImage21Refiner(self, 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, FluxSchnell, GenmoMochi, LTXV, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage] +class HunyuanVideo15(HunyuanVideo): + unet_config = { + "image_model": "hunyuan_video", + "vision_in_dim": 1152, + } + + sampling_settings = { + "shift": 7.0, + } + memory_usage_factor = 4.0 #TODO + supported_inference_dtypes = [torch.float16, torch.bfloat16, torch.float32] + + latent_format = latent_formats.HunyuanVideo15 + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.HunyuanVideo15(self, device=device) + return out + + def clip_target(self, state_dict={}): + pref = self.text_encoder_key_prefix[0] + hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref)) + return supported_models_base.ClipTarget(comfy.text_encoders.hunyuan_video.HunyuanVideo15Tokenizer, comfy.text_encoders.hunyuan_image.te(**hunyuan_detect)) + + +class HunyuanVideo15_SR_Distilled(HunyuanVideo): + unet_config = { + "image_model": "hunyuan_video", + "vision_in_dim": 1152, + "in_channels": 98, + } + + sampling_settings = { + "shift": 2.0, + } + memory_usage_factor = 4.0 #TODO + supported_inference_dtypes = [torch.float16, torch.bfloat16, torch.float32] + + latent_format = latent_formats.HunyuanVideo15 + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.HunyuanVideo15_SR_Distilled(self, device=device) + return out + + def clip_target(self, state_dict={}): + pref = self.text_encoder_key_prefix[0] + hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref)) + return supported_models_base.ClipTarget(comfy.text_encoders.hunyuan_video.HunyuanVideo15Tokenizer, comfy.text_encoders.hunyuan_image.te(**hunyuan_detect)) + + +class Kandinsky5(supported_models_base.BASE): + unet_config = { + "image_model": "kandinsky5", + } + + sampling_settings = { + "shift": 10.0, + } + + unet_extra_config = {} + latent_format = latent_formats.HunyuanVideo + + memory_usage_factor = 1.25 #TODO + + supported_inference_dtypes = [torch.bfloat16, torch.float32] + + vae_key_prefix = ["vae."] + text_encoder_key_prefix = ["text_encoders."] + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.Kandinsky5(self, device=device) + return out + + def clip_target(self, state_dict={}): + pref = self.text_encoder_key_prefix[0] + hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref)) + return supported_models_base.ClipTarget(comfy.text_encoders.kandinsky5.Kandinsky5Tokenizer, comfy.text_encoders.kandinsky5.te(**hunyuan_detect)) + + +class Kandinsky5Image(Kandinsky5): + unet_config = { + "image_model": "kandinsky5", + "model_dim": 2560, + "visual_embed_dim": 64, + } + + sampling_settings = { + "shift": 3.0, + } + + latent_format = latent_formats.Flux + memory_usage_factor = 1.25 #TODO + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.Kandinsky5Image(self, device=device) + return out + + def clip_target(self, state_dict={}): + pref = self.text_encoder_key_prefix[0] + hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref)) + return supported_models_base.ClipTarget(comfy.text_encoders.kandinsky5.Kandinsky5TokenizerImage, comfy.text_encoders.kandinsky5.te(**hunyuan_detect)) + + +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, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5] models += [SVD_img2vid] diff --git a/comfy/supported_models_base.py b/comfy/supported_models_base.py index e4bd74514..0e7a829ba 100644 --- a/comfy/supported_models_base.py +++ b/comfy/supported_models_base.py @@ -17,6 +17,7 @@ """ import torch +import logging from . import model_base from . import utils from . import latent_formats @@ -49,8 +50,7 @@ class BASE: manual_cast_dtype = None custom_operations = None - scaled_fp8 = None - layer_quant_config = None # Per-layer quantization configuration for mixed precision + quant_config = None # quantization configuration for mixed precision optimizations = {"fp8": False} @classmethod @@ -118,3 +118,7 @@ class BASE: def set_inference_dtype(self, dtype, manual_cast_dtype): self.unet_config['dtype'] = dtype self.manual_cast_dtype = manual_cast_dtype + + def __getattr__(self, name): + logging.warning("\nWARNING, you accessed {} from the model config object which doesn't exist. Please fix your code.\n".format(name)) + return None diff --git a/comfy/taesd/taehv.py b/comfy/taesd/taehv.py new file mode 100644 index 000000000..0e5f9a378 --- /dev/null +++ b/comfy/taesd/taehv.py @@ -0,0 +1,173 @@ +# Tiny AutoEncoder for HunyuanVideo and WanVideo https://github.com/madebyollin/taehv + +import torch +import torch.nn as nn +import torch.nn.functional as F +from tqdm.auto import tqdm +from collections import namedtuple, deque + +import comfy.ops +operations=comfy.ops.disable_weight_init + +DecoderResult = namedtuple("DecoderResult", ("frame", "memory")) +TWorkItem = namedtuple("TWorkItem", ("input_tensor", "block_index")) + +def conv(n_in, n_out, **kwargs): + return operations.Conv2d(n_in, n_out, 3, padding=1, **kwargs) + +class Clamp(nn.Module): + def forward(self, x): + return torch.tanh(x / 3) * 3 + +class MemBlock(nn.Module): + def __init__(self, n_in, n_out, act_func): + super().__init__() + self.conv = nn.Sequential(conv(n_in * 2, n_out), act_func, conv(n_out, n_out), act_func, conv(n_out, n_out)) + self.skip = operations.Conv2d(n_in, n_out, 1, bias=False) if n_in != n_out else nn.Identity() + self.act = act_func + def forward(self, x, past): + return self.act(self.conv(torch.cat([x, past], 1)) + self.skip(x)) + +class TPool(nn.Module): + def __init__(self, n_f, stride): + super().__init__() + self.stride = stride + self.conv = operations.Conv2d(n_f*stride,n_f, 1, bias=False) + def forward(self, x): + _NT, C, H, W = x.shape + return self.conv(x.reshape(-1, self.stride * C, H, W)) + +class TGrow(nn.Module): + def __init__(self, n_f, stride): + super().__init__() + self.stride = stride + self.conv = operations.Conv2d(n_f, n_f*stride, 1, bias=False) + def forward(self, x): + _NT, C, H, W = x.shape + x = self.conv(x) + return x.reshape(-1, C, H, W) + +def apply_model_with_memblocks(model, x, parallel, show_progress_bar): + + B, T, C, H, W = x.shape + if parallel: + x = x.reshape(B*T, C, H, W) + # parallel over input timesteps, iterate over blocks + for b in tqdm(model, disable=not show_progress_bar): + if isinstance(b, MemBlock): + BT, C, H, W = x.shape + T = BT // B + _x = x.reshape(B, T, C, H, W) + mem = F.pad(_x, (0,0,0,0,0,0,1,0), value=0)[:,:T].reshape(x.shape) + x = b(x, mem) + else: + x = b(x) + BT, C, H, W = x.shape + T = BT // B + x = x.view(B, T, C, H, W) + else: + out = [] + work_queue = deque([TWorkItem(xt, 0) for t, xt in enumerate(x.reshape(B, T * C, H, W).chunk(T, dim=1))]) + progress_bar = tqdm(range(T), disable=not show_progress_bar) + mem = [None] * len(model) + while work_queue: + xt, i = work_queue.popleft() + if i == 0: + progress_bar.update(1) + if i == len(model): + out.append(xt) + del xt + else: + b = model[i] + if isinstance(b, MemBlock): + if mem[i] is None: + xt_new = b(xt, xt * 0) + mem[i] = xt.detach().clone() + else: + xt_new = b(xt, mem[i]) + mem[i] = xt.detach().clone() + del xt + work_queue.appendleft(TWorkItem(xt_new, i+1)) + elif isinstance(b, TPool): + if mem[i] is None: + mem[i] = [] + mem[i].append(xt.detach().clone()) + if len(mem[i]) == b.stride: + B, C, H, W = xt.shape + xt = b(torch.cat(mem[i], 1).view(B*b.stride, C, H, W)) + mem[i] = [] + work_queue.appendleft(TWorkItem(xt, i+1)) + elif isinstance(b, TGrow): + xt = b(xt) + NT, C, H, W = xt.shape + for xt_next in reversed(xt.view(B, b.stride*C, H, W).chunk(b.stride, 1)): + work_queue.appendleft(TWorkItem(xt_next, i+1)) + del xt + else: + xt = b(xt) + work_queue.appendleft(TWorkItem(xt, i+1)) + progress_bar.close() + x = torch.stack(out, 1) + return x + + +class TAEHV(nn.Module): + def __init__(self, latent_channels, parallel=False, decoder_time_upscale=(True, True), decoder_space_upscale=(True, True, True), latent_format=None, show_progress_bar=True): + super().__init__() + self.image_channels = 3 + self.patch_size = 1 + self.latent_channels = latent_channels + self.parallel = parallel + self.latent_format = latent_format + self.show_progress_bar = show_progress_bar + self.process_in = latent_format().process_in if latent_format is not None else (lambda x: x) + self.process_out = latent_format().process_out if latent_format is not None else (lambda x: x) + if self.latent_channels in [48, 32]: # Wan 2.2 and HunyuanVideo1.5 + self.patch_size = 2 + if self.latent_channels == 32: # HunyuanVideo1.5 + act_func = nn.LeakyReLU(0.2, inplace=True) + else: # HunyuanVideo, Wan 2.1 + act_func = nn.ReLU(inplace=True) + + self.encoder = nn.Sequential( + conv(self.image_channels*self.patch_size**2, 64), act_func, + TPool(64, 2), conv(64, 64, stride=2, bias=False), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), + TPool(64, 2), conv(64, 64, stride=2, bias=False), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), + TPool(64, 1), conv(64, 64, stride=2, bias=False), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), + conv(64, self.latent_channels), + ) + n_f = [256, 128, 64, 64] + self.frames_to_trim = 2**sum(decoder_time_upscale) - 1 + self.decoder = nn.Sequential( + Clamp(), conv(self.latent_channels, n_f[0]), act_func, + MemBlock(n_f[0], n_f[0], act_func), MemBlock(n_f[0], n_f[0], act_func), MemBlock(n_f[0], n_f[0], act_func), nn.Upsample(scale_factor=2 if decoder_space_upscale[0] else 1), TGrow(n_f[0], 1), conv(n_f[0], n_f[1], bias=False), + MemBlock(n_f[1], n_f[1], act_func), MemBlock(n_f[1], n_f[1], act_func), MemBlock(n_f[1], n_f[1], act_func), nn.Upsample(scale_factor=2 if decoder_space_upscale[1] else 1), TGrow(n_f[1], 2 if decoder_time_upscale[0] else 1), conv(n_f[1], n_f[2], bias=False), + MemBlock(n_f[2], n_f[2], act_func), MemBlock(n_f[2], n_f[2], act_func), MemBlock(n_f[2], n_f[2], act_func), nn.Upsample(scale_factor=2 if decoder_space_upscale[2] else 1), TGrow(n_f[2], 2 if decoder_time_upscale[1] else 1), conv(n_f[2], n_f[3], bias=False), + act_func, conv(n_f[3], self.image_channels*self.patch_size**2), + ) + @property + def show_progress_bar(self): + return self._show_progress_bar + + @show_progress_bar.setter + def show_progress_bar(self, value): + self._show_progress_bar = value + + def encode(self, x, **kwargs): + if self.patch_size > 1: + x = F.pixel_unshuffle(x, self.patch_size) + x = x.movedim(2, 1) # [B, C, T, H, W] -> [B, T, C, H, W] + if x.shape[1] % 4 != 0: + # pad at end to multiple of 4 + n_pad = 4 - x.shape[1] % 4 + padding = x[:, -1:].repeat_interleave(n_pad, dim=1) + x = torch.cat([x, padding], 1) + x = apply_model_with_memblocks(self.encoder, x, self.parallel, self.show_progress_bar).movedim(2, 1) + return self.process_out(x) + + def decode(self, x, **kwargs): + x = self.process_in(x).movedim(2, 1) # [B, C, T, H, W] -> [B, T, C, H, W] + x = apply_model_with_memblocks(self.decoder, x, self.parallel, self.show_progress_bar) + if self.patch_size > 1: + x = F.pixel_shuffle(x, self.patch_size) + return x[:, self.frames_to_trim:].movedim(2, 1) diff --git a/comfy/text_encoders/cosmos.py b/comfy/text_encoders/cosmos.py index a1adb5242..448381fa9 100644 --- a/comfy/text_encoders/cosmos.py +++ b/comfy/text_encoders/cosmos.py @@ -7,10 +7,10 @@ from transformers import T5TokenizerFast class T5XXLModel(sd1_clip.SDClipModel): def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}): textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_old_config_xxl.json") - t5xxl_scaled_fp8 = model_options.get("t5xxl_scaled_fp8", None) - if t5xxl_scaled_fp8 is not None: + t5xxl_quantization_metadata = model_options.get("t5xxl_quantization_metadata", None) + if t5xxl_quantization_metadata is not None: model_options = model_options.copy() - model_options["scaled_fp8"] = t5xxl_scaled_fp8 + model_options["quantization_metadata"] = t5xxl_quantization_metadata super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, zero_out_masked=attention_mask, model_options=model_options) @@ -30,12 +30,12 @@ class CosmosT5Tokenizer(sd1_clip.SD1Tokenizer): super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5xxl", tokenizer=T5XXLTokenizer) -def te(dtype_t5=None, t5xxl_scaled_fp8=None): +def te(dtype_t5=None, t5_quantization_metadata=None): class CosmosTEModel_(CosmosT5XXL): def __init__(self, device="cpu", dtype=None, model_options={}): - if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options: + if t5_quantization_metadata is not None: model_options = model_options.copy() - model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8 + model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata if dtype is None: dtype = dtype_t5 super().__init__(device=device, dtype=dtype, model_options=model_options) diff --git a/comfy/text_encoders/flux.py b/comfy/text_encoders/flux.py index d61ef6668..21d93d757 100644 --- a/comfy/text_encoders/flux.py +++ b/comfy/text_encoders/flux.py @@ -1,10 +1,13 @@ from comfy import sd1_clip import comfy.text_encoders.t5 import comfy.text_encoders.sd3_clip +import comfy.text_encoders.llama import comfy.model_management -from transformers import T5TokenizerFast +from transformers import T5TokenizerFast, LlamaTokenizerFast import torch import os +import json +import base64 class T5XXLTokenizer(sd1_clip.SDTokenizer): def __init__(self, embedding_directory=None, tokenizer_data={}): @@ -60,11 +63,112 @@ class FluxClipModel(torch.nn.Module): else: return self.t5xxl.load_sd(sd) -def flux_clip(dtype_t5=None, t5xxl_scaled_fp8=None): +def flux_clip(dtype_t5=None, t5_quantization_metadata=None): class FluxClipModel_(FluxClipModel): def __init__(self, device="cpu", dtype=None, model_options={}): - if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options: + if t5_quantization_metadata is not None: model_options = model_options.copy() - model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8 + model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata super().__init__(dtype_t5=dtype_t5, device=device, dtype=dtype, model_options=model_options) return FluxClipModel_ + +def load_mistral_tokenizer(data): + if torch.is_tensor(data): + data = data.numpy().tobytes() + + try: + from transformers.integrations.mistral import MistralConverter + except ModuleNotFoundError: + from transformers.models.pixtral.convert_pixtral_weights_to_hf import MistralConverter + + mistral_vocab = json.loads(data) + + special_tokens = {} + vocab = {} + + max_vocab = mistral_vocab["config"]["default_vocab_size"] + max_vocab -= len(mistral_vocab["special_tokens"]) + + for w in mistral_vocab["vocab"]: + r = w["rank"] + if r >= max_vocab: + continue + + vocab[base64.b64decode(w["token_bytes"])] = r + + for w in mistral_vocab["special_tokens"]: + if "token_bytes" in w: + special_tokens[base64.b64decode(w["token_bytes"])] = w["rank"] + else: + special_tokens[w["token_str"]] = w["rank"] + + all_special = [] + for v in special_tokens: + all_special.append(v) + + special_tokens.update(vocab) + vocab = special_tokens + return {"tokenizer_object": MistralConverter(vocab=vocab, additional_special_tokens=all_special).converted(), "legacy": False} + +class MistralTokenizerClass: + @staticmethod + def from_pretrained(path, **kwargs): + return LlamaTokenizerFast(**kwargs) + +class Mistral3Tokenizer(sd1_clip.SDTokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + self.tekken_data = tokenizer_data.get("tekken_model", None) + super().__init__("", pad_with_end=False, embedding_size=5120, embedding_key='mistral3_24b', tokenizer_class=MistralTokenizerClass, has_end_token=False, pad_to_max_length=False, pad_token=11, max_length=99999999, min_length=1, pad_left=True, tokenizer_args=load_mistral_tokenizer(self.tekken_data), tokenizer_data=tokenizer_data) + + def state_dict(self): + return {"tekken_model": self.tekken_data} + +class Flux2Tokenizer(sd1_clip.SD1Tokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="mistral3_24b", tokenizer=Mistral3Tokenizer) + self.llama_template = '[SYSTEM_PROMPT]You are an AI that reasons about image descriptions. You give structured responses focusing on object relationships, object\nattribution and actions without speculation.[/SYSTEM_PROMPT][INST]{}[/INST]' + + def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, **kwargs): + if llama_template is None: + llama_text = self.llama_template.format(text) + else: + llama_text = llama_template.format(text) + + tokens = super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, disable_weights=True, **kwargs) + return tokens + +class Mistral3_24BModel(sd1_clip.SDClipModel): + def __init__(self, device="cpu", layer=[10, 20, 30], layer_idx=None, dtype=None, attention_mask=True, model_options={}): + textmodel_json_config = {} + num_layers = model_options.get("num_layers", None) + if num_layers is not None: + textmodel_json_config["num_hidden_layers"] = num_layers + if num_layers < 40: + textmodel_json_config["final_norm"] = False + super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 1, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Mistral3Small24B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) + +class Flux2TEModel(sd1_clip.SD1ClipModel): + def __init__(self, device="cpu", dtype=None, model_options={}, name="mistral3_24b", clip_model=Mistral3_24BModel): + super().__init__(device=device, dtype=dtype, name=name, clip_model=clip_model, model_options=model_options) + + def encode_token_weights(self, token_weight_pairs): + out, pooled, extra = super().encode_token_weights(token_weight_pairs) + + out = torch.stack((out[:, 0], out[:, 1], out[:, 2]), dim=1) + out = out.movedim(1, 2) + out = out.reshape(out.shape[0], out.shape[1], -1) + return out, pooled, extra + +def flux2_te(dtype_llama=None, llama_quantization_metadata=None, pruned=False): + class Flux2TEModel_(Flux2TEModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + if dtype_llama is not None: + dtype = dtype_llama + if llama_quantization_metadata is not None: + model_options = model_options.copy() + model_options["quantization_metadata"] = llama_quantization_metadata + if pruned: + model_options = model_options.copy() + model_options["num_layers"] = 30 + super().__init__(device=device, dtype=dtype, model_options=model_options) + return Flux2TEModel_ diff --git a/comfy/text_encoders/genmo.py b/comfy/text_encoders/genmo.py index 9dcf190a2..5daea8135 100644 --- a/comfy/text_encoders/genmo.py +++ b/comfy/text_encoders/genmo.py @@ -26,12 +26,12 @@ class MochiT5Tokenizer(sd1_clip.SD1Tokenizer): super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5xxl", tokenizer=T5XXLTokenizer) -def mochi_te(dtype_t5=None, t5xxl_scaled_fp8=None): +def mochi_te(dtype_t5=None, t5_quantization_metadata=None): class MochiTEModel_(MochiT5XXL): def __init__(self, device="cpu", dtype=None, model_options={}): - if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options: + if t5_quantization_metadata is not None: model_options = model_options.copy() - model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8 + model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata if dtype is None: dtype = dtype_t5 super().__init__(device=device, dtype=dtype, model_options=model_options) diff --git a/comfy/text_encoders/hidream.py b/comfy/text_encoders/hidream.py index dbcf52784..600b34480 100644 --- a/comfy/text_encoders/hidream.py +++ b/comfy/text_encoders/hidream.py @@ -142,14 +142,14 @@ class HiDreamTEModel(torch.nn.Module): return self.llama.load_sd(sd) -def hidream_clip(clip_l=True, clip_g=True, t5=True, llama=True, dtype_t5=None, dtype_llama=None, t5xxl_scaled_fp8=None, llama_scaled_fp8=None): +def hidream_clip(clip_l=True, clip_g=True, t5=True, llama=True, dtype_t5=None, dtype_llama=None, t5_quantization_metadata=None, llama_quantization_metadata=None): class HiDreamTEModel_(HiDreamTEModel): def __init__(self, device="cpu", dtype=None, model_options={}): - if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options: + if t5_quantization_metadata is not None: model_options = model_options.copy() - model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8 - if llama_scaled_fp8 is not None and "llama_scaled_fp8" not in model_options: + model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata + if llama_quantization_metadata is not None: model_options = model_options.copy() - model_options["llama_scaled_fp8"] = llama_scaled_fp8 + model_options["llama_quantization_metadata"] = llama_quantization_metadata super().__init__(clip_l=clip_l, clip_g=clip_g, t5=t5, llama=llama, dtype_t5=dtype_t5, dtype_llama=dtype_llama, device=device, dtype=dtype, model_options=model_options) return HiDreamTEModel_ diff --git a/comfy/text_encoders/hunyuan_image.py b/comfy/text_encoders/hunyuan_image.py index ff04726e1..cd198036c 100644 --- a/comfy/text_encoders/hunyuan_image.py +++ b/comfy/text_encoders/hunyuan_image.py @@ -40,10 +40,10 @@ class HunyuanImageTokenizer(QwenImageTokenizer): class Qwen25_7BVLIModel(sd1_clip.SDClipModel): def __init__(self, device="cpu", layer="hidden", layer_idx=-3, dtype=None, attention_mask=True, model_options={}): - llama_scaled_fp8 = model_options.get("qwen_scaled_fp8", None) - if llama_scaled_fp8 is not None: + llama_quantization_metadata = model_options.get("llama_quantization_metadata", None) + if llama_quantization_metadata is not None: model_options = model_options.copy() - model_options["scaled_fp8"] = llama_scaled_fp8 + model_options["quantization_metadata"] = llama_quantization_metadata super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen25_7BVLI, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) @@ -91,12 +91,12 @@ class HunyuanImageTEModel(QwenImageTEModel): else: return super().load_sd(sd) -def te(byt5=True, dtype_llama=None, llama_scaled_fp8=None): +def te(byt5=True, dtype_llama=None, llama_quantization_metadata=None): class QwenImageTEModel_(HunyuanImageTEModel): def __init__(self, device="cpu", dtype=None, model_options={}): - if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options: + if llama_quantization_metadata is not None: model_options = model_options.copy() - model_options["qwen_scaled_fp8"] = llama_scaled_fp8 + model_options["llama_quantization_metadata"] = llama_quantization_metadata if dtype_llama is not None: dtype = dtype_llama super().__init__(byt5=byt5, device=device, dtype=dtype, model_options=model_options) diff --git a/comfy/text_encoders/hunyuan_video.py b/comfy/text_encoders/hunyuan_video.py index b02148b33..a9a6c525e 100644 --- a/comfy/text_encoders/hunyuan_video.py +++ b/comfy/text_encoders/hunyuan_video.py @@ -1,11 +1,12 @@ from comfy import sd1_clip import comfy.model_management import comfy.text_encoders.llama +from .hunyuan_image import HunyuanImageTokenizer from transformers import LlamaTokenizerFast import torch import os import numbers - +import comfy.utils def llama_detect(state_dict, prefix=""): out = {} @@ -13,9 +14,9 @@ def llama_detect(state_dict, prefix=""): if t5_key in state_dict: out["dtype_llama"] = state_dict[t5_key].dtype - scaled_fp8_key = "{}scaled_fp8".format(prefix) - if scaled_fp8_key in state_dict: - out["llama_scaled_fp8"] = state_dict[scaled_fp8_key].dtype + quant = comfy.utils.detect_layer_quantization(state_dict, prefix) + if quant is not None: + out["llama_quantization_metadata"] = quant return out @@ -27,10 +28,10 @@ class LLAMA3Tokenizer(sd1_clip.SDTokenizer): class LLAMAModel(sd1_clip.SDClipModel): def __init__(self, device="cpu", layer="hidden", layer_idx=-3, dtype=None, attention_mask=True, model_options={}, special_tokens={"start": 128000, "pad": 128258}): - llama_scaled_fp8 = model_options.get("llama_scaled_fp8", None) - if llama_scaled_fp8 is not None: + llama_quantization_metadata = model_options.get("llama_quantization_metadata", None) + if llama_quantization_metadata is not None: model_options = model_options.copy() - model_options["scaled_fp8"] = llama_scaled_fp8 + model_options["quantization_metadata"] = llama_quantization_metadata textmodel_json_config = {} vocab_size = model_options.get("vocab_size", None) @@ -73,6 +74,14 @@ class HunyuanVideoTokenizer: return {} +class HunyuanVideo15Tokenizer(HunyuanImageTokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data) + self.llama_template = "<|im_start|>system\nYou are a helpful assistant. Describe the video by detailing the following aspects:\n1. The main content and theme of the video.\n2. The color, shape, size, texture, quantity, text, and spatial relationships of the objects.\n3. Actions, events, behaviors temporal relationships, physical movement changes of the objects.\n4. background environment, light, style and atmosphere.\n5. camera angles, movements, and transitions used in the video.<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n" + + def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs): + return super().tokenize_with_weights(text, return_word_ids, prevent_empty_text=True, **kwargs) + class HunyuanVideoClipModel(torch.nn.Module): def __init__(self, dtype_llama=None, device="cpu", dtype=None, model_options={}): super().__init__() @@ -149,11 +158,11 @@ class HunyuanVideoClipModel(torch.nn.Module): return self.llama.load_sd(sd) -def hunyuan_video_clip(dtype_llama=None, llama_scaled_fp8=None): +def hunyuan_video_clip(dtype_llama=None, llama_quantization_metadata=None): class HunyuanVideoClipModel_(HunyuanVideoClipModel): def __init__(self, device="cpu", dtype=None, model_options={}): - if llama_scaled_fp8 is not None and "llama_scaled_fp8" not in model_options: + if llama_quantization_metadata is not None: model_options = model_options.copy() - model_options["llama_scaled_fp8"] = llama_scaled_fp8 + model_options["llama_quantization_metadata"] = llama_quantization_metadata super().__init__(dtype_llama=dtype_llama, device=device, dtype=dtype, model_options=model_options) return HunyuanVideoClipModel_ diff --git a/comfy/text_encoders/jina_clip_2.py b/comfy/text_encoders/jina_clip_2.py new file mode 100644 index 000000000..0cffb6d16 --- /dev/null +++ b/comfy/text_encoders/jina_clip_2.py @@ -0,0 +1,219 @@ +# Jina CLIP v2 and Jina Embeddings v3 both use their modified XLM-RoBERTa architecture. Reference implementation: +# Jina CLIP v2 (both text and vision): https://huggingface.co/jinaai/jina-clip-implementation/blob/39e6a55ae971b59bea6e44675d237c99762e7ee2/modeling_clip.py +# Jina XLM-RoBERTa (text only): http://huggingface.co/jinaai/xlm-roberta-flash-implementation/blob/2b6bc3f30750b3a9648fe9b63448c09920efe9be/modeling_xlm_roberta.py + +from dataclasses import dataclass + +import torch +from torch import nn as nn +from torch.nn import functional as F + +import comfy.model_management +import comfy.ops +from comfy import sd1_clip +from .spiece_tokenizer import SPieceTokenizer + +class JinaClip2Tokenizer(sd1_clip.SDTokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + tokenizer = tokenizer_data.get("spiece_model", None) + # The official NewBie uses max_length=8000, but Jina Embeddings v3 actually supports 8192 + super().__init__(tokenizer, pad_with_end=False, embedding_size=1024, embedding_key='jina_clip_2', tokenizer_class=SPieceTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=False, max_length=8192, min_length=1, pad_token=1, end_token=2, tokenizer_args={"add_bos": True, "add_eos": True}, tokenizer_data=tokenizer_data) + + def state_dict(self): + return {"spiece_model": self.tokenizer.serialize_model()} + +class JinaClip2TokenizerWrapper(sd1_clip.SD1Tokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, tokenizer=JinaClip2Tokenizer, name="jina_clip_2") + +# https://huggingface.co/jinaai/jina-embeddings-v3/blob/343dbf534c76fe845f304fa5c2d1fd87e1e78918/config.json +@dataclass +class XLMRobertaConfig: + vocab_size: int = 250002 + type_vocab_size: int = 1 + hidden_size: int = 1024 + num_hidden_layers: int = 24 + num_attention_heads: int = 16 + rotary_emb_base: float = 20000.0 + intermediate_size: int = 4096 + hidden_act: str = "gelu" + hidden_dropout_prob: float = 0.1 + attention_probs_dropout_prob: float = 0.1 + layer_norm_eps: float = 1e-05 + bos_token_id: int = 0 + eos_token_id: int = 2 + pad_token_id: int = 1 + +class XLMRobertaEmbeddings(nn.Module): + def __init__(self, config, device=None, dtype=None, ops=None): + super().__init__() + embed_dim = config.hidden_size + self.word_embeddings = ops.Embedding(config.vocab_size, embed_dim, padding_idx=config.pad_token_id, device=device, dtype=dtype) + self.token_type_embeddings = ops.Embedding(config.type_vocab_size, embed_dim, device=device, dtype=dtype) + + def forward(self, input_ids=None, embeddings=None): + if input_ids is not None and embeddings is None: + embeddings = self.word_embeddings(input_ids) + + if embeddings is not None: + token_type_ids = torch.zeros(embeddings.shape[1], device=embeddings.device, dtype=torch.int32) + token_type_embeddings = self.token_type_embeddings(token_type_ids) + embeddings = embeddings + token_type_embeddings + return embeddings + +class RotaryEmbedding(nn.Module): + def __init__(self, dim, base, device=None): + super().__init__() + inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, device=device, dtype=torch.float32) / dim)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached = None + self._sin_cached = None + + def _update_cos_sin_cache(self, seqlen, device=None, dtype=None): + if seqlen > self._seq_len_cached or self._cos_cached is None or self._cos_cached.device != device or self._cos_cached.dtype != dtype: + self._seq_len_cached = seqlen + t = torch.arange(seqlen, device=device, dtype=torch.float32) + freqs = torch.outer(t, self.inv_freq.to(device=t.device)) + emb = torch.cat((freqs, freqs), dim=-1) + self._cos_cached = emb.cos().to(dtype) + self._sin_cached = emb.sin().to(dtype) + + def forward(self, q, k): + batch, seqlen, heads, head_dim = q.shape + self._update_cos_sin_cache(seqlen, device=q.device, dtype=q.dtype) + + cos = self._cos_cached[:seqlen].view(1, seqlen, 1, head_dim) + sin = self._sin_cached[:seqlen].view(1, seqlen, 1, head_dim) + + def rotate_half(x): + size = x.shape[-1] // 2 + x1, x2 = x[..., :size], x[..., size:] + return torch.cat((-x2, x1), dim=-1) + + q_embed = (q * cos) + (rotate_half(q) * sin) + k_embed = (k * cos) + (rotate_half(k) * sin) + return q_embed, k_embed + +class MHA(nn.Module): + def __init__(self, config, device=None, dtype=None, ops=None): + super().__init__() + embed_dim = config.hidden_size + self.num_heads = config.num_attention_heads + self.head_dim = embed_dim // config.num_attention_heads + + self.rotary_emb = RotaryEmbedding(self.head_dim, config.rotary_emb_base, device=device) + self.Wqkv = ops.Linear(embed_dim, 3 * embed_dim, device=device, dtype=dtype) + self.out_proj = ops.Linear(embed_dim, embed_dim, device=device, dtype=dtype) + + def forward(self, x, mask=None, optimized_attention=None): + qkv = self.Wqkv(x) + batch_size, seq_len, _ = qkv.shape + qkv = qkv.view(batch_size, seq_len, 3, self.num_heads, self.head_dim) + q, k, v = qkv.unbind(2) + + q, k = self.rotary_emb(q, k) + + # NHD -> HND + q = q.transpose(1, 2) + k = k.transpose(1, 2) + v = v.transpose(1, 2) + + out = optimized_attention(q, k, v, heads=self.num_heads, mask=mask, skip_reshape=True) + return self.out_proj(out) + +class MLP(nn.Module): + def __init__(self, config, device=None, dtype=None, ops=None): + super().__init__() + self.fc1 = ops.Linear(config.hidden_size, config.intermediate_size, device=device, dtype=dtype) + self.activation = F.gelu + self.fc2 = ops.Linear(config.intermediate_size, config.hidden_size, device=device, dtype=dtype) + + def forward(self, x): + x = self.fc1(x) + x = self.activation(x) + x = self.fc2(x) + return x + +class Block(nn.Module): + def __init__(self, config, device=None, dtype=None, ops=None): + super().__init__() + self.mixer = MHA(config, device=device, dtype=dtype, ops=ops) + self.dropout1 = nn.Dropout(config.hidden_dropout_prob) + self.norm1 = ops.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, device=device, dtype=dtype) + self.mlp = MLP(config, device=device, dtype=dtype, ops=ops) + self.dropout2 = nn.Dropout(config.hidden_dropout_prob) + self.norm2 = ops.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, device=device, dtype=dtype) + + def forward(self, hidden_states, mask=None, optimized_attention=None): + mixer_out = self.mixer(hidden_states, mask=mask, optimized_attention=optimized_attention) + hidden_states = self.norm1(self.dropout1(mixer_out) + hidden_states) + mlp_out = self.mlp(hidden_states) + hidden_states = self.norm2(self.dropout2(mlp_out) + hidden_states) + return hidden_states + +class XLMRobertaEncoder(nn.Module): + def __init__(self, config, device=None, dtype=None, ops=None): + super().__init__() + self.layers = nn.ModuleList([Block(config, device=device, dtype=dtype, ops=ops) for _ in range(config.num_hidden_layers)]) + + def forward(self, hidden_states, attention_mask=None): + optimized_attention = comfy.ldm.modules.attention.optimized_attention_for_device(hidden_states.device, mask=attention_mask is not None, small_input=True) + for layer in self.layers: + hidden_states = layer(hidden_states, mask=attention_mask, optimized_attention=optimized_attention) + return hidden_states + +class XLMRobertaModel_(nn.Module): + def __init__(self, config, device=None, dtype=None, ops=None): + super().__init__() + self.embeddings = XLMRobertaEmbeddings(config, device=device, dtype=dtype, ops=ops) + self.emb_ln = ops.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, device=device, dtype=dtype) + self.emb_drop = nn.Dropout(config.hidden_dropout_prob) + self.encoder = XLMRobertaEncoder(config, device=device, dtype=dtype, ops=ops) + + def forward(self, input_ids, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, embeds_info=[]): + x = self.embeddings(input_ids=input_ids, embeddings=embeds) + x = self.emb_ln(x) + x = self.emb_drop(x) + + mask = None + 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])) + mask = mask.masked_fill(mask.to(torch.bool), -torch.finfo(x.dtype).max) + + sequence_output = self.encoder(x, attention_mask=mask) + + # Mean pool, see https://huggingface.co/jinaai/jina-clip-implementation/blob/39e6a55ae971b59bea6e44675d237c99762e7ee2/hf_model.py + pooled_output = None + if attention_mask is None: + pooled_output = sequence_output.mean(dim=1) + else: + attention_mask = attention_mask.to(sequence_output.dtype) + pooled_output = (sequence_output * attention_mask.unsqueeze(-1)).sum(dim=1) / attention_mask.sum(dim=-1, keepdim=True) + + # Intermediate output is not yet implemented, use None for placeholder + return sequence_output, None, pooled_output + +class XLMRobertaModel(nn.Module): + def __init__(self, config_dict, dtype, device, operations): + super().__init__() + self.config = XLMRobertaConfig(**config_dict) + self.model = XLMRobertaModel_(self.config, device=device, dtype=dtype, ops=operations) + self.num_layers = self.config.num_hidden_layers + + def get_input_embeddings(self): + return self.model.embeddings.word_embeddings + + def set_input_embeddings(self, embeddings): + self.model.embeddings.word_embeddings = embeddings + + def forward(self, *args, **kwargs): + return self.model(*args, **kwargs) + +class JinaClip2TextModel(sd1_clip.SDClipModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + super().__init__(device=device, dtype=dtype, textmodel_json_config={}, model_class=XLMRobertaModel, special_tokens={"start": 0, "end": 2, "pad": 1}, enable_attention_masks=True, return_attention_masks=True, model_options=model_options) + +class JinaClip2TextModelWrapper(sd1_clip.SD1ClipModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + super().__init__(device=device, dtype=dtype, clip_model=JinaClip2TextModel, name="jina_clip_2", model_options=model_options) diff --git a/comfy/text_encoders/kandinsky5.py b/comfy/text_encoders/kandinsky5.py new file mode 100644 index 000000000..be086458c --- /dev/null +++ b/comfy/text_encoders/kandinsky5.py @@ -0,0 +1,68 @@ +from comfy import sd1_clip +from .qwen_image import QwenImageTokenizer, QwenImageTEModel +from .llama import Qwen25_7BVLI + + +class Kandinsky5Tokenizer(QwenImageTokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data) + self.llama_template = "<|im_start|>system\nYou are a prompt engineer. Describe the video in detail.\nDescribe how the camera moves or shakes, describe the zoom and view angle, whether it follows the objects.\nDescribe the location of the video, main characters or objects and their action.\nDescribe the dynamism of the video and presented actions.\nName the visual style of the video: whether it is a professional footage, user generated content, some kind of animation, video game or screen content.\nDescribe the visual effects, postprocessing and transitions if they are presented in the video.\nPay attention to the order of key actions shown in the scene.<|im_end|>\n<|im_start|>user\n{}<|im_end|>" + self.clip_l = sd1_clip.SDTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data) + + def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs): + out = super().tokenize_with_weights(text, return_word_ids, **kwargs) + out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids, **kwargs) + + return out + + +class Kandinsky5TokenizerImage(Kandinsky5Tokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data) + self.llama_template = "<|im_start|>system\nYou are a promt engineer. Describe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>" + + +class Qwen25_7BVLIModel(sd1_clip.SDClipModel): + def __init__(self, device="cpu", layer="hidden", layer_idx=-1, dtype=None, attention_mask=True, model_options={}): + llama_quantization_metadata = model_options.get("llama_quantization_metadata", None) + if llama_quantization_metadata is not None: + model_options = model_options.copy() + model_options["quantization_metadata"] = llama_quantization_metadata + super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=Qwen25_7BVLI, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) + + +class Kandinsky5TEModel(QwenImageTEModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + super(QwenImageTEModel, self).__init__(device=device, dtype=dtype, name="qwen25_7b", clip_model=Qwen25_7BVLIModel, model_options=model_options) + self.clip_l = sd1_clip.SDClipModel(device=device, dtype=dtype, return_projected_pooled=False, model_options=model_options) + + def encode_token_weights(self, token_weight_pairs): + cond, p, extra = super().encode_token_weights(token_weight_pairs, template_end=-1) + l_out, l_pooled = self.clip_l.encode_token_weights(token_weight_pairs["l"]) + + return cond, l_pooled, extra + + def set_clip_options(self, options): + super().set_clip_options(options) + self.clip_l.set_clip_options(options) + + def reset_clip_options(self): + super().reset_clip_options() + self.clip_l.reset_clip_options() + + def load_sd(self, sd): + if "text_model.encoder.layers.1.mlp.fc1.weight" in sd: + return self.clip_l.load_sd(sd) + else: + return super().load_sd(sd) + +def te(dtype_llama=None, llama_quantization_metadata=None): + class Kandinsky5TEModel_(Kandinsky5TEModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + if llama_quantization_metadata is not None: + model_options = model_options.copy() + model_options["llama_quantization_metadata"] = llama_quantization_metadata + if dtype_llama is not None: + dtype = dtype_llama + super().__init__(device=device, dtype=dtype, model_options=model_options) + return Kandinsky5TEModel_ diff --git a/comfy/text_encoders/llama.py b/comfy/text_encoders/llama.py index c050759fe..76731576b 100644 --- a/comfy/text_encoders/llama.py +++ b/comfy/text_encoders/llama.py @@ -3,13 +3,12 @@ import torch.nn as nn from dataclasses import dataclass from typing import Optional, Any import math -import logging from comfy.ldm.modules.attention import optimized_attention_for_device import comfy.model_management import comfy.ldm.common_dit +import comfy.clip_model -import comfy.model_management from . import qwen_vl @dataclass @@ -32,6 +31,29 @@ class Llama2Config: q_norm = None k_norm = None rope_scale = None + final_norm: bool = True + +@dataclass +class Mistral3Small24BConfig: + vocab_size: int = 131072 + hidden_size: int = 5120 + intermediate_size: int = 32768 + num_hidden_layers: int = 40 + num_attention_heads: int = 32 + num_key_value_heads: int = 8 + max_position_embeddings: int = 8192 + rms_norm_eps: float = 1e-5 + rope_theta: float = 1000000000.0 + transformer_type: str = "llama" + head_dim = 128 + rms_norm_add = False + mlp_activation = "silu" + qkv_bias = False + rope_dims = None + q_norm = None + k_norm = None + rope_scale = None + final_norm: bool = True @dataclass class Qwen25_3BConfig: @@ -53,6 +75,51 @@ class Qwen25_3BConfig: q_norm = None k_norm = None rope_scale = None + final_norm: bool = True + +@dataclass +class Qwen3_4BConfig: + vocab_size: int = 151936 + hidden_size: int = 2560 + intermediate_size: int = 9728 + num_hidden_layers: int = 36 + num_attention_heads: int = 32 + num_key_value_heads: int = 8 + max_position_embeddings: int = 40960 + rms_norm_eps: float = 1e-6 + rope_theta: float = 1000000.0 + transformer_type: str = "llama" + head_dim = 128 + rms_norm_add = False + mlp_activation = "silu" + qkv_bias = False + rope_dims = None + q_norm = "gemma3" + k_norm = "gemma3" + rope_scale = None + final_norm: bool = True + +@dataclass +class Ovis25_2BConfig: + vocab_size: int = 151936 + hidden_size: int = 2048 + intermediate_size: int = 6144 + num_hidden_layers: int = 28 + num_attention_heads: int = 16 + num_key_value_heads: int = 8 + max_position_embeddings: int = 40960 + rms_norm_eps: float = 1e-6 + rope_theta: float = 1000000.0 + transformer_type: str = "llama" + head_dim = 128 + rms_norm_add = False + mlp_activation = "silu" + qkv_bias = False + rope_dims = None + q_norm = "gemma3" + k_norm = "gemma3" + rope_scale = None + final_norm: bool = True @dataclass class Qwen25_7BVLI_Config: @@ -74,6 +141,7 @@ class Qwen25_7BVLI_Config: q_norm = None k_norm = None rope_scale = None + final_norm: bool = True @dataclass class Gemma2_2B_Config: @@ -96,6 +164,7 @@ class Gemma2_2B_Config: k_norm = None sliding_attention = None rope_scale = None + final_norm: bool = True @dataclass class Gemma3_4B_Config: @@ -107,7 +176,7 @@ class Gemma3_4B_Config: num_key_value_heads: int = 4 max_position_embeddings: int = 131072 rms_norm_eps: float = 1e-6 - rope_theta = [10000.0, 1000000.0] + rope_theta = [1000000.0, 10000.0] transformer_type: str = "gemma3" head_dim = 256 rms_norm_add = True @@ -116,8 +185,34 @@ class Gemma3_4B_Config: rope_dims = None q_norm = "gemma3" k_norm = "gemma3" - sliding_attention = [False, False, False, False, False, 1024] - rope_scale = [1.0, 8.0] + sliding_attention = [1024, 1024, 1024, 1024, 1024, False] + rope_scale = [8.0, 1.0] + final_norm: bool = True + +@dataclass +class Gemma3_12B_Config: + vocab_size: int = 262208 + hidden_size: int = 3840 + intermediate_size: int = 15360 + num_hidden_layers: int = 48 + num_attention_heads: int = 16 + num_key_value_heads: int = 8 + max_position_embeddings: int = 131072 + rms_norm_eps: float = 1e-6 + rope_theta = [1000000.0, 10000.0] + transformer_type: str = "gemma3" + head_dim = 256 + rms_norm_add = True + mlp_activation = "gelu_pytorch_tanh" + qkv_bias = False + rope_dims = None + q_norm = "gemma3" + k_norm = "gemma3" + sliding_attention = [1024, 1024, 1024, 1024, 1024, False] + rope_scale = [8.0, 1.0] + final_norm: bool = True + vision_config = {"num_channels": 3, "hidden_act": "gelu_pytorch_tanh", "hidden_size": 1152, "image_size": 896, "intermediate_size": 4304, "model_type": "siglip_vision_model", "num_attention_heads": 16, "num_hidden_layers": 27, "patch_size": 14} + mm_tokens_per_image = 256 class RMSNorm(nn.Module): def __init__(self, dim: int, eps: float = 1e-5, add=False, device=None, dtype=None): @@ -299,7 +394,7 @@ class TransformerBlockGemma2(nn.Module): self.pre_feedforward_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype) self.post_feedforward_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype) - if config.sliding_attention is not None: # TODO: implement. (Not that necessary since models are trained on less than 1024 tokens) + if config.sliding_attention is not None: self.sliding_attention = config.sliding_attention[index % len(config.sliding_attention)] else: self.sliding_attention = False @@ -316,7 +411,12 @@ class TransformerBlockGemma2(nn.Module): if self.transformer_type == 'gemma3': if self.sliding_attention: if x.shape[1] > self.sliding_attention: - logging.warning("Warning: sliding attention not implemented, results may be incorrect") + sliding_mask = torch.full((x.shape[1], x.shape[1]), float("-inf"), device=x.device, dtype=x.dtype) + sliding_mask.tril_(diagonal=-self.sliding_attention) + if attention_mask is not None: + attention_mask = attention_mask + sliding_mask + else: + attention_mask = sliding_mask freqs_cis = freqs_cis[1] else: freqs_cis = freqs_cis[0] @@ -366,7 +466,12 @@ class Llama2_(nn.Module): transformer(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, add=config.rms_norm_add, device=device, dtype=dtype) + + if config.final_norm: + self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype) + else: + self.norm = None + # self.lm_head = ops.Linear(config.hidden_size, config.vocab_size, bias=False, device=device, dtype=dtype) 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=[]): @@ -402,8 +507,12 @@ class Llama2_(nn.Module): intermediate = None all_intermediate = None + only_layers = None if intermediate_output is not None: - if intermediate_output == "all": + if isinstance(intermediate_output, list): + all_intermediate = [] + only_layers = set(intermediate_output) + elif intermediate_output == "all": all_intermediate = [] intermediate_output = None elif intermediate_output < 0: @@ -411,7 +520,8 @@ class Llama2_(nn.Module): for i, layer in enumerate(self.layers): if all_intermediate is not None: - all_intermediate.append(x.unsqueeze(1).clone()) + if only_layers is None or (i in only_layers): + all_intermediate.append(x.unsqueeze(1).clone()) x = layer( x=x, attention_mask=mask, @@ -421,18 +531,56 @@ class Llama2_(nn.Module): if i == intermediate_output: intermediate = x.clone() - x = self.norm(x) + if self.norm is not None: + x = self.norm(x) + if all_intermediate is not None: - all_intermediate.append(x.unsqueeze(1).clone()) + if only_layers is None or ((i + 1) in only_layers): + all_intermediate.append(x.unsqueeze(1).clone()) if all_intermediate is not None: intermediate = torch.cat(all_intermediate, dim=1) - if intermediate is not None and final_layer_norm_intermediate: + if intermediate is not None and final_layer_norm_intermediate and self.norm is not None: intermediate = self.norm(intermediate) return x, intermediate + +class Gemma3MultiModalProjector(torch.nn.Module): + def __init__(self, config, dtype, device, operations): + super().__init__() + + self.mm_input_projection_weight = nn.Parameter( + torch.empty(config.vision_config["hidden_size"], config.hidden_size, device=device, dtype=dtype) + ) + + self.mm_soft_emb_norm = RMSNorm(config.vision_config["hidden_size"], eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype) + + self.patches_per_image = int(config.vision_config["image_size"] // config.vision_config["patch_size"]) + self.tokens_per_side = int(config.mm_tokens_per_image**0.5) + self.kernel_size = self.patches_per_image // self.tokens_per_side + self.avg_pool = nn.AvgPool2d(kernel_size=self.kernel_size, stride=self.kernel_size) + + def forward(self, vision_outputs: torch.Tensor): + batch_size, _, seq_length = vision_outputs.shape + + reshaped_vision_outputs = vision_outputs.transpose(1, 2) + reshaped_vision_outputs = reshaped_vision_outputs.reshape( + batch_size, seq_length, self.patches_per_image, self.patches_per_image + ) + reshaped_vision_outputs = reshaped_vision_outputs.contiguous() + + pooled_vision_outputs = self.avg_pool(reshaped_vision_outputs) + pooled_vision_outputs = pooled_vision_outputs.flatten(2) + pooled_vision_outputs = pooled_vision_outputs.transpose(1, 2) + + normed_vision_outputs = self.mm_soft_emb_norm(pooled_vision_outputs) + + projected_vision_outputs = torch.matmul(normed_vision_outputs, comfy.model_management.cast_to_device(self.mm_input_projection_weight, device=normed_vision_outputs.device, dtype=normed_vision_outputs.dtype)) + return projected_vision_outputs.type_as(vision_outputs) + + class BaseLlama: def get_input_embeddings(self): return self.model.embed_tokens @@ -453,6 +601,15 @@ class Llama2(BaseLlama, torch.nn.Module): self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) self.dtype = dtype +class Mistral3Small24B(BaseLlama, torch.nn.Module): + def __init__(self, config_dict, dtype, device, operations): + super().__init__() + config = Mistral3Small24BConfig(**config_dict) + self.num_layers = config.num_hidden_layers + + self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) + self.dtype = dtype + class Qwen25_3B(BaseLlama, torch.nn.Module): def __init__(self, config_dict, dtype, device, operations): super().__init__() @@ -462,6 +619,24 @@ class Qwen25_3B(BaseLlama, torch.nn.Module): self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) self.dtype = dtype +class Qwen3_4B(BaseLlama, torch.nn.Module): + def __init__(self, config_dict, dtype, device, operations): + super().__init__() + config = Qwen3_4BConfig(**config_dict) + self.num_layers = config.num_hidden_layers + + self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) + self.dtype = dtype + +class Ovis25_2B(BaseLlama, torch.nn.Module): + def __init__(self, config_dict, dtype, device, operations): + super().__init__() + config = Ovis25_2BConfig(**config_dict) + self.num_layers = config.num_hidden_layers + + self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) + self.dtype = dtype + class Qwen25_7BVLI(BaseLlama, torch.nn.Module): def __init__(self, config_dict, dtype, device, operations): super().__init__() @@ -522,3 +697,21 @@ class Gemma3_4B(BaseLlama, torch.nn.Module): self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) self.dtype = dtype + +class Gemma3_12B(BaseLlama, torch.nn.Module): + def __init__(self, config_dict, dtype, device, operations): + super().__init__() + config = Gemma3_12B_Config(**config_dict) + self.num_layers = config.num_hidden_layers + + self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) + self.multi_modal_projector = Gemma3MultiModalProjector(config, dtype, device, operations) + self.vision_model = comfy.clip_model.CLIPVision(config.vision_config, dtype, device, operations) + self.dtype = dtype + self.image_size = config.vision_config["image_size"] + + def preprocess_embed(self, embed, device): + if embed["type"] == "image": + image = comfy.clip_model.clip_preprocess(embed["data"], size=self.image_size, mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], crop=True) + return self.multi_modal_projector(self.vision_model(image.to(device, dtype=torch.float32))[0]), None + return None, None diff --git a/comfy/text_encoders/lt.py b/comfy/text_encoders/lt.py index 48ea67e67..130ebaeae 100644 --- a/comfy/text_encoders/lt.py +++ b/comfy/text_encoders/lt.py @@ -1,7 +1,11 @@ from comfy import sd1_clip import os from transformers import T5TokenizerFast +from .spiece_tokenizer import SPieceTokenizer import comfy.text_encoders.genmo +from comfy.ldm.lightricks.embeddings_connector import Embeddings1DConnector +import torch +import comfy.utils class T5XXLTokenizer(sd1_clip.SDTokenizer): def __init__(self, embedding_directory=None, tokenizer_data={}): @@ -16,3 +20,112 @@ class LTXVT5Tokenizer(sd1_clip.SD1Tokenizer): def ltxv_te(*args, **kwargs): return comfy.text_encoders.genmo.mochi_te(*args, **kwargs) + + +class Gemma3_12BTokenizer(sd1_clip.SDTokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + tokenizer = tokenizer_data.get("spiece_model", None) + super().__init__(tokenizer, pad_with_end=False, embedding_size=3840, embedding_key='gemma3_12b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False}, tokenizer_data=tokenizer_data) + + def state_dict(self): + return {"spiece_model": self.tokenizer.serialize_model()} + +class LTXAVGemmaTokenizer(sd1_clip.SD1Tokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="gemma3_12b", tokenizer=Gemma3_12BTokenizer) + +class Gemma3_12BModel(sd1_clip.SDClipModel): + def __init__(self, device="cpu", layer="all", layer_idx=None, dtype=None, attention_mask=True, model_options={}): + llama_quantization_metadata = model_options.get("llama_quantization_metadata", None) + if llama_quantization_metadata is not None: + model_options = model_options.copy() + model_options["quantization_metadata"] = llama_quantization_metadata + + super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_12B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) + + def tokenize_with_weights(self, text, return_word_ids=False, llama_template="{}", image_embeds=None, **kwargs): + text = llama_template.format(text) + text_tokens = super().tokenize_with_weights(text, return_word_ids) + embed_count = 0 + for k in text_tokens: + tt = text_tokens[k] + for r in tt: + for i in range(len(r)): + if r[i][0] == 262144: + if image_embeds is not None and embed_count < image_embeds.shape[0]: + r[i] = ({"type": "embedding", "data": image_embeds[embed_count], "original_type": "image"},) + r[i][1:] + embed_count += 1 + return text_tokens + +class LTXAVTEModel(torch.nn.Module): + def __init__(self, dtype_llama=None, device="cpu", dtype=None, model_options={}): + super().__init__() + self.dtypes = set() + self.dtypes.add(dtype) + + self.gemma3_12b = Gemma3_12BModel(device=device, dtype=dtype_llama, model_options=model_options, layer="all", layer_idx=None) + self.dtypes.add(dtype_llama) + + operations = self.gemma3_12b.operations # TODO + self.text_embedding_projection = operations.Linear(3840 * 49, 3840, bias=False, dtype=dtype, device=device) + + self.audio_embeddings_connector = Embeddings1DConnector( + split_rope=True, + double_precision_rope=True, + dtype=dtype, + device=device, + operations=operations, + ) + + self.video_embeddings_connector = Embeddings1DConnector( + split_rope=True, + double_precision_rope=True, + dtype=dtype, + device=device, + operations=operations, + ) + + def set_clip_options(self, options): + self.execution_device = options.get("execution_device", self.execution_device) + self.gemma3_12b.set_clip_options(options) + + def reset_clip_options(self): + self.gemma3_12b.reset_clip_options() + self.execution_device = None + + def encode_token_weights(self, token_weight_pairs): + token_weight_pairs = token_weight_pairs["gemma3_12b"] + + out, pooled, extra = self.gemma3_12b.encode_token_weights(token_weight_pairs) + out_device = out.device + out = out.movedim(1, -1).to(self.execution_device) + out = 8.0 * (out - out.mean(dim=(1, 2), keepdim=True)) / (out.amax(dim=(1, 2), keepdim=True) - out.amin(dim=(1, 2), keepdim=True) + 1e-6) + out = out.reshape((out.shape[0], out.shape[1], -1)) + out = self.text_embedding_projection(out) + out_vid = self.video_embeddings_connector(out)[0] + out_audio = self.audio_embeddings_connector(out)[0] + out = torch.concat((out_vid, out_audio), dim=-1) + + return out.to(out_device), pooled + + def load_sd(self, sd): + if "model.layers.47.self_attn.q_norm.weight" in sd: + return self.gemma3_12b.load_sd(sd) + else: + sdo = comfy.utils.state_dict_prefix_replace(sd, {"text_embedding_projection.aggregate_embed.weight": "text_embedding_projection.weight", "model.diffusion_model.video_embeddings_connector.": "video_embeddings_connector.", "model.diffusion_model.audio_embeddings_connector.": "audio_embeddings_connector."}, filter_keys=True) + if len(sdo) == 0: + sdo = sd + + return self.load_state_dict(sdo, strict=False) + + +def ltxav_te(dtype_llama=None, llama_quantization_metadata=None): + class LTXAVTEModel_(LTXAVTEModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + if llama_quantization_metadata is not None: + model_options = model_options.copy() + model_options["llama_quantization_metadata"] = llama_quantization_metadata + if dtype_llama is not None: + dtype = dtype_llama + super().__init__(dtype_llama=dtype_llama, device=device, dtype=dtype, model_options=model_options) + return LTXAVTEModel_ diff --git a/comfy/text_encoders/lumina2.py b/comfy/text_encoders/lumina2.py index fd986e2c1..b29a7cc87 100644 --- a/comfy/text_encoders/lumina2.py +++ b/comfy/text_encoders/lumina2.py @@ -14,7 +14,7 @@ class Gemma2BTokenizer(sd1_clip.SDTokenizer): class Gemma3_4BTokenizer(sd1_clip.SDTokenizer): def __init__(self, embedding_directory=None, tokenizer_data={}): tokenizer = tokenizer_data.get("spiece_model", None) - super().__init__(tokenizer, pad_with_end=False, embedding_size=2560, embedding_key='gemma3_4b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False}, tokenizer_data=tokenizer_data) + super().__init__(tokenizer, pad_with_end=False, embedding_size=2560, embedding_key='gemma3_4b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False}, disable_weights=True, tokenizer_data=tokenizer_data) def state_dict(self): return {"spiece_model": self.tokenizer.serialize_model()} @@ -33,6 +33,11 @@ class Gemma2_2BModel(sd1_clip.SDClipModel): class Gemma3_4BModel(sd1_clip.SDClipModel): def __init__(self, device="cpu", layer="hidden", layer_idx=-2, dtype=None, attention_mask=True, model_options={}): + llama_quantization_metadata = model_options.get("llama_quantization_metadata", None) + if llama_quantization_metadata is not None: + model_options = model_options.copy() + model_options["quantization_metadata"] = llama_quantization_metadata + super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_4B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) class LuminaModel(sd1_clip.SD1ClipModel): @@ -40,7 +45,7 @@ class LuminaModel(sd1_clip.SD1ClipModel): super().__init__(device=device, dtype=dtype, name=name, clip_model=clip_model, model_options=model_options) -def te(dtype_llama=None, llama_scaled_fp8=None, model_type="gemma2_2b"): +def te(dtype_llama=None, llama_quantization_metadata=None, model_type="gemma2_2b"): if model_type == "gemma2_2b": model = Gemma2_2BModel elif model_type == "gemma3_4b": @@ -48,9 +53,9 @@ def te(dtype_llama=None, llama_scaled_fp8=None, model_type="gemma2_2b"): class LuminaTEModel_(LuminaModel): def __init__(self, device="cpu", dtype=None, model_options={}): - if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options: + if llama_quantization_metadata is not None: model_options = model_options.copy() - model_options["scaled_fp8"] = llama_scaled_fp8 + model_options["quantization_metadata"] = llama_quantization_metadata if dtype_llama is not None: dtype = dtype_llama super().__init__(device=device, dtype=dtype, name=model_type, model_options=model_options, clip_model=model) diff --git a/comfy/text_encoders/newbie.py b/comfy/text_encoders/newbie.py new file mode 100644 index 000000000..db2324576 --- /dev/null +++ b/comfy/text_encoders/newbie.py @@ -0,0 +1,62 @@ +import torch + +import comfy.model_management +import comfy.text_encoders.jina_clip_2 +import comfy.text_encoders.lumina2 + +class NewBieTokenizer: + def __init__(self, embedding_directory=None, tokenizer_data={}): + self.gemma = comfy.text_encoders.lumina2.Gemma3_4BTokenizer(embedding_directory=embedding_directory, tokenizer_data={"spiece_model": tokenizer_data["gemma_spiece_model"]}) + self.jina = comfy.text_encoders.jina_clip_2.JinaClip2Tokenizer(embedding_directory=embedding_directory, tokenizer_data={"spiece_model": tokenizer_data["jina_spiece_model"]}) + + def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs): + out = {} + out["gemma"] = self.gemma.tokenize_with_weights(text, return_word_ids, **kwargs) + out["jina"] = self.jina.tokenize_with_weights(text, return_word_ids, **kwargs) + return out + + def untokenize(self, token_weight_pair): + raise NotImplementedError + + def state_dict(self): + return {} + +class NewBieTEModel(torch.nn.Module): + def __init__(self, dtype_gemma=None, device="cpu", dtype=None, model_options={}): + super().__init__() + dtype_gemma = comfy.model_management.pick_weight_dtype(dtype_gemma, dtype, device) + self.gemma = comfy.text_encoders.lumina2.Gemma3_4BModel(device=device, dtype=dtype_gemma, model_options=model_options) + self.jina = comfy.text_encoders.jina_clip_2.JinaClip2TextModel(device=device, dtype=dtype, model_options=model_options) + self.dtypes = {dtype, dtype_gemma} + + def set_clip_options(self, options): + self.gemma.set_clip_options(options) + self.jina.set_clip_options(options) + + def reset_clip_options(self): + self.gemma.reset_clip_options() + self.jina.reset_clip_options() + + def encode_token_weights(self, token_weight_pairs): + token_weight_pairs_gemma = token_weight_pairs["gemma"] + token_weight_pairs_jina = token_weight_pairs["jina"] + + gemma_out, gemma_pooled, gemma_extra = self.gemma.encode_token_weights(token_weight_pairs_gemma) + jina_out, jina_pooled, jina_extra = self.jina.encode_token_weights(token_weight_pairs_jina) + + return gemma_out, jina_pooled, gemma_extra + + def load_sd(self, sd): + if "model.layers.0.self_attn.q_norm.weight" in sd: + return self.gemma.load_sd(sd) + else: + return self.jina.load_sd(sd) + +def te(dtype_llama=None, llama_quantization_metadata=None): + class NewBieTEModel_(NewBieTEModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + if llama_quantization_metadata is not None: + model_options = model_options.copy() + model_options["llama_quantization_metadata"] = llama_quantization_metadata + super().__init__(dtype_gemma=dtype_llama, device=device, dtype=dtype, model_options=model_options) + return NewBieTEModel_ diff --git a/comfy/text_encoders/omnigen2.py b/comfy/text_encoders/omnigen2.py index 1a01b2dd4..50aa4121f 100644 --- a/comfy/text_encoders/omnigen2.py +++ b/comfy/text_encoders/omnigen2.py @@ -32,12 +32,12 @@ class Omnigen2Model(sd1_clip.SD1ClipModel): super().__init__(device=device, dtype=dtype, name="qwen25_3b", clip_model=Qwen25_3BModel, model_options=model_options) -def te(dtype_llama=None, llama_scaled_fp8=None): +def te(dtype_llama=None, llama_quantization_metadata=None): class Omnigen2TEModel_(Omnigen2Model): def __init__(self, device="cpu", dtype=None, model_options={}): - if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options: + if llama_quantization_metadata is not None: model_options = model_options.copy() - model_options["scaled_fp8"] = llama_scaled_fp8 + model_options["quantization_metadata"] = llama_quantization_metadata if dtype_llama is not None: dtype = dtype_llama super().__init__(device=device, dtype=dtype, model_options=model_options) diff --git a/comfy/text_encoders/ovis.py b/comfy/text_encoders/ovis.py new file mode 100644 index 000000000..5754424d2 --- /dev/null +++ b/comfy/text_encoders/ovis.py @@ -0,0 +1,66 @@ +from transformers import Qwen2Tokenizer +import comfy.text_encoders.llama +from comfy import sd1_clip +import os +import torch +import numbers + +class Qwen3Tokenizer(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=2048, embedding_key='qwen3_2b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=284, pad_token=151643, tokenizer_data=tokenizer_data) + + +class OvisTokenizer(sd1_clip.SD1Tokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen3_2b", tokenizer=Qwen3Tokenizer) + self.llama_template = "<|im_start|>user\nDescribe the image by detailing the color, quantity, text, shape, size, texture, spatial relationships of the objects and background: {}<|im_end|>\n<|im_start|>assistant\n\n\n\n\n" + + def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, **kwargs): + if llama_template is None: + llama_text = self.llama_template.format(text) + else: + llama_text = llama_template.format(text) + + tokens = super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, disable_weights=True, **kwargs) + return tokens + +class Ovis25_2BModel(sd1_clip.SDClipModel): + def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}): + super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Ovis25_2B, enable_attention_masks=attention_mask, return_attention_masks=False, zero_out_masked=True, model_options=model_options) + + +class OvisTEModel(sd1_clip.SD1ClipModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + super().__init__(device=device, dtype=dtype, name="qwen3_2b", clip_model=Ovis25_2BModel, model_options=model_options) + + def encode_token_weights(self, token_weight_pairs, template_end=-1): + out, pooled = super().encode_token_weights(token_weight_pairs) + tok_pairs = token_weight_pairs["qwen3_2b"][0] + count_im_start = 0 + if template_end == -1: + for i, v in enumerate(tok_pairs): + elem = v[0] + if not torch.is_tensor(elem): + if isinstance(elem, numbers.Integral): + if elem == 4004 and count_im_start < 1: + template_end = i + count_im_start += 1 + + if out.shape[1] > (template_end + 1): + if tok_pairs[template_end + 1][0] == 25: + template_end += 1 + + out = out[:, template_end:] + return out, pooled, {} + + +def te(dtype_llama=None, llama_quantization_metadata=None): + class OvisTEModel_(OvisTEModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + if dtype_llama is not None: + dtype = dtype_llama + if llama_quantization_metadata is not None: + model_options["quantization_metadata"] = llama_quantization_metadata + super().__init__(device=device, dtype=dtype, model_options=model_options) + return OvisTEModel_ diff --git a/comfy/text_encoders/pixart_t5.py b/comfy/text_encoders/pixart_t5.py index 5f383de07..e5e5f18be 100644 --- a/comfy/text_encoders/pixart_t5.py +++ b/comfy/text_encoders/pixart_t5.py @@ -30,12 +30,12 @@ class PixArtTokenizer(sd1_clip.SD1Tokenizer): def __init__(self, embedding_directory=None, tokenizer_data={}): super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5xxl", tokenizer=T5XXLTokenizer) -def pixart_te(dtype_t5=None, t5xxl_scaled_fp8=None): +def pixart_te(dtype_t5=None, t5_quantization_metadata=None): class PixArtTEModel_(PixArtT5XXL): def __init__(self, device="cpu", dtype=None, model_options={}): - if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options: + if t5_quantization_metadata is not None: model_options = model_options.copy() - model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8 + model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata if dtype is None: dtype = dtype_t5 super().__init__(device=device, dtype=dtype, model_options=model_options) diff --git a/comfy/text_encoders/qwen25_tokenizer/tokenizer_config.json b/comfy/text_encoders/qwen25_tokenizer/tokenizer_config.json index 67688e82c..df5b5d7fe 100644 --- a/comfy/text_encoders/qwen25_tokenizer/tokenizer_config.json +++ b/comfy/text_encoders/qwen25_tokenizer/tokenizer_config.json @@ -179,36 +179,36 @@ "special": false }, "151665": { - "content": "<|img|>", + "content": "", "lstrip": false, "normalized": false, "rstrip": false, "single_word": false, - "special": true + "special": false }, "151666": { - "content": "<|endofimg|>", + "content": "", "lstrip": false, "normalized": false, "rstrip": false, "single_word": false, - "special": true + "special": false }, "151667": { - "content": "<|meta|>", + "content": "", "lstrip": false, "normalized": false, "rstrip": false, "single_word": false, - "special": true + "special": false }, "151668": { - "content": "<|endofmeta|>", + "content": "", "lstrip": false, "normalized": false, "rstrip": false, "single_word": false, - "special": true + "special": false } }, "additional_special_tokens": [ diff --git a/comfy/text_encoders/qwen_image.py b/comfy/text_encoders/qwen_image.py index 40fa67937..5c14dec23 100644 --- a/comfy/text_encoders/qwen_image.py +++ b/comfy/text_encoders/qwen_image.py @@ -17,12 +17,14 @@ class QwenImageTokenizer(sd1_clip.SD1Tokenizer): self.llama_template = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n" self.llama_template_images = "<|im_start|>system\nDescribe the key features of the input image (color, shape, size, texture, objects, background), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate.<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>{}<|im_end|>\n<|im_start|>assistant\n" - def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], **kwargs): + def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], prevent_empty_text=False, **kwargs): skip_template = False if text.startswith('<|im_start|>'): skip_template = True if text.startswith('<|start_header_id|>'): skip_template = True + if prevent_empty_text and text == '': + text = ' ' if skip_template: llama_text = text @@ -83,12 +85,12 @@ class QwenImageTEModel(sd1_clip.SD1ClipModel): return out, pooled, extra -def te(dtype_llama=None, llama_scaled_fp8=None): +def te(dtype_llama=None, llama_quantization_metadata=None): class QwenImageTEModel_(QwenImageTEModel): def __init__(self, device="cpu", dtype=None, model_options={}): - if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options: + if llama_quantization_metadata is not None: model_options = model_options.copy() - model_options["scaled_fp8"] = llama_scaled_fp8 + model_options["quantization_metadata"] = llama_quantization_metadata if dtype_llama is not None: dtype = dtype_llama super().__init__(device=device, dtype=dtype, model_options=model_options) diff --git a/comfy/text_encoders/sd3_clip.py b/comfy/text_encoders/sd3_clip.py index ff5d412db..8b153c72b 100644 --- a/comfy/text_encoders/sd3_clip.py +++ b/comfy/text_encoders/sd3_clip.py @@ -6,14 +6,15 @@ import torch import os import comfy.model_management import logging +import comfy.utils class T5XXLModel(sd1_clip.SDClipModel): def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=False, model_options={}): textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_config_xxl.json") - t5xxl_scaled_fp8 = model_options.get("t5xxl_scaled_fp8", None) - if t5xxl_scaled_fp8 is not None: + t5xxl_quantization_metadata = model_options.get("t5xxl_quantization_metadata", None) + if t5xxl_quantization_metadata is not None: model_options = model_options.copy() - model_options["scaled_fp8"] = t5xxl_scaled_fp8 + model_options["quantization_metadata"] = t5xxl_quantization_metadata model_options = {**model_options, "model_name": "t5xxl"} super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) @@ -25,9 +26,9 @@ def t5_xxl_detect(state_dict, prefix=""): if t5_key in state_dict: out["dtype_t5"] = state_dict[t5_key].dtype - scaled_fp8_key = "{}scaled_fp8".format(prefix) - if scaled_fp8_key in state_dict: - out["t5xxl_scaled_fp8"] = state_dict[scaled_fp8_key].dtype + quant = comfy.utils.detect_layer_quantization(state_dict, prefix) + if quant is not None: + out["t5_quantization_metadata"] = quant return out @@ -156,11 +157,11 @@ class SD3ClipModel(torch.nn.Module): else: return self.t5xxl.load_sd(sd) -def sd3_clip(clip_l=True, clip_g=True, t5=True, dtype_t5=None, t5xxl_scaled_fp8=None, t5_attention_mask=False): +def sd3_clip(clip_l=True, clip_g=True, t5=True, dtype_t5=None, t5_quantization_metadata=None, t5_attention_mask=False): class SD3ClipModel_(SD3ClipModel): def __init__(self, device="cpu", dtype=None, model_options={}): - if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options: + if t5_quantization_metadata is not None: model_options = model_options.copy() - model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8 + model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata super().__init__(clip_l=clip_l, clip_g=clip_g, t5=t5, dtype_t5=dtype_t5, t5_attention_mask=t5_attention_mask, device=device, dtype=dtype, model_options=model_options) return SD3ClipModel_ diff --git a/comfy/text_encoders/wan.py b/comfy/text_encoders/wan.py index d50fa4b28..164a57edd 100644 --- a/comfy/text_encoders/wan.py +++ b/comfy/text_encoders/wan.py @@ -25,12 +25,12 @@ class WanT5Model(sd1_clip.SD1ClipModel): def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs): super().__init__(device=device, dtype=dtype, model_options=model_options, name="umt5xxl", clip_model=UMT5XXlModel, **kwargs) -def te(dtype_t5=None, t5xxl_scaled_fp8=None): +def te(dtype_t5=None, t5_quantization_metadata=None): class WanTEModel(WanT5Model): def __init__(self, device="cpu", dtype=None, model_options={}): - if t5xxl_scaled_fp8 is not None and "scaled_fp8" not in model_options: + if t5_quantization_metadata is not None: model_options = model_options.copy() - model_options["scaled_fp8"] = t5xxl_scaled_fp8 + model_options["quantization_metadata"] = t5_quantization_metadata if dtype_t5 is not None: dtype = dtype_t5 super().__init__(device=device, dtype=dtype, model_options=model_options) diff --git a/comfy/text_encoders/z_image.py b/comfy/text_encoders/z_image.py new file mode 100644 index 000000000..19adde0b7 --- /dev/null +++ b/comfy/text_encoders/z_image.py @@ -0,0 +1,45 @@ +from transformers import Qwen2Tokenizer +import comfy.text_encoders.llama +from comfy import sd1_clip +import os + +class Qwen3Tokenizer(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=2560, embedding_key='qwen3_4b', 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 ZImageTokenizer(sd1_clip.SD1Tokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen3_4b", tokenizer=Qwen3Tokenizer) + self.llama_template = "<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n" + + def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, **kwargs): + if llama_template is None: + llama_text = self.llama_template.format(text) + else: + llama_text = llama_template.format(text) + + tokens = super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, disable_weights=True, **kwargs) + return tokens + + +class Qwen3_4BModel(sd1_clip.SDClipModel): + def __init__(self, device="cpu", layer="hidden", layer_idx=-2, dtype=None, attention_mask=True, model_options={}): + super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_4B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) + + +class ZImageTEModel(sd1_clip.SD1ClipModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + super().__init__(device=device, dtype=dtype, name="qwen3_4b", clip_model=Qwen3_4BModel, model_options=model_options) + + +def te(dtype_llama=None, llama_quantization_metadata=None): + class ZImageTEModel_(ZImageTEModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + if dtype_llama is not None: + dtype = dtype_llama + if llama_quantization_metadata is not None: + model_options["quantization_metadata"] = llama_quantization_metadata + super().__init__(device=device, dtype=dtype, model_options=model_options) + return ZImageTEModel_ diff --git a/comfy/utils.py b/comfy/utils.py index 4bd281057..ffa98c9b1 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -29,6 +29,7 @@ import itertools from torch.nn.functional import interpolate from einops import rearrange from comfy.cli_args import args +import json MMAP_TORCH_FILES = args.mmap_torch_files DISABLE_MMAP = args.disable_mmap @@ -52,7 +53,7 @@ if hasattr(torch.serialization, "add_safe_globals"): # TODO: this was added in ALWAYS_SAFE_LOAD = True logging.info("Checkpoint files will always be loaded safely.") else: - logging.info("Warning, you are using an old pytorch version and some ckpt/pt files might be loaded unsafely. Upgrading to 2.4 or above is recommended.") + logging.warning("Warning, you are using an old pytorch version and some ckpt/pt files might be loaded unsafely. Upgrading to 2.4 or above is recommended as older versions of pytorch are no longer supported.") def load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False): if device is None: @@ -675,6 +676,72 @@ def flux_to_diffusers(mmdit_config, output_prefix=""): return key_map +def z_image_to_diffusers(mmdit_config, output_prefix=""): + n_layers = mmdit_config.get("n_layers", 0) + hidden_size = mmdit_config.get("dim", 0) + n_context_refiner = mmdit_config.get("n_refiner_layers", 2) + n_noise_refiner = mmdit_config.get("n_refiner_layers", 2) + key_map = {} + + def add_block_keys(prefix_from, prefix_to, has_adaln=True): + for end in ("weight", "bias"): + k = "{}.attention.".format(prefix_from) + qkv = "{}.attention.qkv.{}".format(prefix_to, end) + key_map["{}to_q.{}".format(k, end)] = (qkv, (0, 0, hidden_size)) + key_map["{}to_k.{}".format(k, end)] = (qkv, (0, hidden_size, hidden_size)) + key_map["{}to_v.{}".format(k, end)] = (qkv, (0, hidden_size * 2, hidden_size)) + + block_map = { + "attention.norm_q.weight": "attention.q_norm.weight", + "attention.norm_k.weight": "attention.k_norm.weight", + "attention.to_out.0.weight": "attention.out.weight", + "attention.to_out.0.bias": "attention.out.bias", + "attention_norm1.weight": "attention_norm1.weight", + "attention_norm2.weight": "attention_norm2.weight", + "feed_forward.w1.weight": "feed_forward.w1.weight", + "feed_forward.w2.weight": "feed_forward.w2.weight", + "feed_forward.w3.weight": "feed_forward.w3.weight", + "ffn_norm1.weight": "ffn_norm1.weight", + "ffn_norm2.weight": "ffn_norm2.weight", + } + if has_adaln: + block_map["adaLN_modulation.0.weight"] = "adaLN_modulation.0.weight" + block_map["adaLN_modulation.0.bias"] = "adaLN_modulation.0.bias" + for k, v in block_map.items(): + key_map["{}.{}".format(prefix_from, k)] = "{}.{}".format(prefix_to, v) + + for i in range(n_layers): + add_block_keys("layers.{}".format(i), "{}layers.{}".format(output_prefix, i)) + + for i in range(n_context_refiner): + add_block_keys("context_refiner.{}".format(i), "{}context_refiner.{}".format(output_prefix, i)) + + for i in range(n_noise_refiner): + add_block_keys("noise_refiner.{}".format(i), "{}noise_refiner.{}".format(output_prefix, i)) + + MAP_BASIC = [ + ("final_layer.linear.weight", "all_final_layer.2-1.linear.weight"), + ("final_layer.linear.bias", "all_final_layer.2-1.linear.bias"), + ("final_layer.adaLN_modulation.1.weight", "all_final_layer.2-1.adaLN_modulation.1.weight"), + ("final_layer.adaLN_modulation.1.bias", "all_final_layer.2-1.adaLN_modulation.1.bias"), + ("x_embedder.weight", "all_x_embedder.2-1.weight"), + ("x_embedder.bias", "all_x_embedder.2-1.bias"), + ("x_pad_token", "x_pad_token"), + ("cap_embedder.0.weight", "cap_embedder.0.weight"), + ("cap_embedder.1.weight", "cap_embedder.1.weight"), + ("cap_embedder.1.bias", "cap_embedder.1.bias"), + ("cap_pad_token", "cap_pad_token"), + ("t_embedder.mlp.0.weight", "t_embedder.mlp.0.weight"), + ("t_embedder.mlp.0.bias", "t_embedder.mlp.0.bias"), + ("t_embedder.mlp.2.weight", "t_embedder.mlp.2.weight"), + ("t_embedder.mlp.2.bias", "t_embedder.mlp.2.bias"), + ] + + for c, diffusers in MAP_BASIC: + key_map[diffusers] = "{}{}".format(output_prefix, c) + + return key_map + def repeat_to_batch_size(tensor, batch_size, dim=0): if tensor.shape[dim] > batch_size: return tensor.narrow(dim, 0, batch_size) @@ -736,12 +803,17 @@ def safetensors_header(safetensors_path, max_size=100*1024*1024): return None return f.read(length_of_header) +ATTR_UNSET={} + def set_attr(obj, attr, value): attrs = attr.split(".") for name in attrs[:-1]: obj = getattr(obj, name) - prev = getattr(obj, attrs[-1]) - setattr(obj, attrs[-1], value) + prev = getattr(obj, attrs[-1], ATTR_UNSET) + if value is ATTR_UNSET: + delattr(obj, attrs[-1]) + else: + setattr(obj, attrs[-1], value) return prev def set_attr_param(obj, attr, value): @@ -1126,5 +1198,72 @@ def unpack_latents(combined_latent, latent_shapes): combined_latent = combined_latent[:, :, cut:] output_tensors.append(tens.reshape([tens.shape[0]] + list(shape)[1:])) else: - output_tensors = combined_latent + output_tensors = [combined_latent] return output_tensors + +def detect_layer_quantization(state_dict, prefix): + for k in state_dict: + if k.startswith(prefix) and k.endswith(".comfy_quant"): + logging.info("Found quantization metadata version 1") + return {"mixed_ops": True} + return None + +def convert_old_quants(state_dict, model_prefix="", metadata={}): + if metadata is None: + metadata = {} + + quant_metadata = None + if "_quantization_metadata" not in metadata: + scaled_fp8_key = "{}scaled_fp8".format(model_prefix) + + if scaled_fp8_key in state_dict: + scaled_fp8_weight = state_dict[scaled_fp8_key] + scaled_fp8_dtype = scaled_fp8_weight.dtype + if scaled_fp8_dtype == torch.float32: + scaled_fp8_dtype = torch.float8_e4m3fn + + if scaled_fp8_weight.nelement() == 2: + full_precision_matrix_mult = True + else: + full_precision_matrix_mult = False + + out_sd = {} + layers = {} + for k in list(state_dict.keys()): + if k == scaled_fp8_key: + continue + if not k.startswith(model_prefix): + out_sd[k] = state_dict[k] + continue + k_out = k + w = state_dict.pop(k) + layer = None + if k_out.endswith(".scale_weight"): + layer = k_out[:-len(".scale_weight")] + 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 + if full_precision_matrix_mult: + layer_conf["full_precision_matrix_mult"] = full_precision_matrix_mult + layers[layer] = layer_conf + + if k_out.endswith(".scale_input"): + layer = k_out[:-len(".scale_input")] + k_out = "{}.input_scale".format(layer) + if w.item() == 1.0: + continue + + out_sd[k_out] = w + + state_dict = out_sd + quant_metadata = {"layers": layers} + else: + quant_metadata = json.loads(metadata["_quantization_metadata"]) + + if quant_metadata is not None: + layers = quant_metadata["layers"] + for k, v in layers.items(): + state_dict["{}.comfy_quant".format(k)] = torch.tensor(list(json.dumps(v).encode('utf-8')), dtype=torch.uint8) + + return state_dict, metadata diff --git a/comfy/weight_adapter/lora.py b/comfy/weight_adapter/lora.py index 4db004e50..3cc60bb1b 100644 --- a/comfy/weight_adapter/lora.py +++ b/comfy/weight_adapter/lora.py @@ -194,6 +194,7 @@ class LoRAAdapter(WeightAdapterBase): lora_diff = torch.mm( mat1.flatten(start_dim=1), mat2.flatten(start_dim=1) ).reshape(weight.shape) + del mat1, mat2 if dora_scale is not None: weight = weight_decompose( dora_scale, diff --git a/comfy_api/feature_flags.py b/comfy_api/feature_flags.py index 0d4389a6e..de167f037 100644 --- a/comfy_api/feature_flags.py +++ b/comfy_api/feature_flags.py @@ -5,19 +5,20 @@ This module handles capability negotiation between frontend and backend, allowing graceful protocol evolution while maintaining backward compatibility. """ -from typing import Any, Dict +from typing import Any from comfy.cli_args import args # Default server capabilities -SERVER_FEATURE_FLAGS: Dict[str, Any] = { +SERVER_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}}, } def get_connection_feature( - sockets_metadata: Dict[str, Dict[str, Any]], + sockets_metadata: dict[str, dict[str, Any]], sid: str, feature_name: str, default: Any = False @@ -41,7 +42,7 @@ def get_connection_feature( def supports_feature( - sockets_metadata: Dict[str, Dict[str, Any]], + sockets_metadata: dict[str, dict[str, Any]], sid: str, feature_name: str ) -> bool: @@ -59,7 +60,7 @@ def supports_feature( return get_connection_feature(sockets_metadata, sid, feature_name, False) is True -def get_server_features() -> Dict[str, Any]: +def get_server_features() -> dict[str, Any]: """ Get the server's feature flags. diff --git a/comfy_api/internal/api_registry.py b/comfy_api/internal/api_registry.py index 7e3375cf6..2b1cb016a 100644 --- a/comfy_api/internal/api_registry.py +++ b/comfy_api/internal/api_registry.py @@ -1,4 +1,4 @@ -from typing import Type, List, NamedTuple +from typing import NamedTuple from comfy_api.internal.singleton import ProxiedSingleton from packaging import version as packaging_version @@ -10,7 +10,7 @@ class ComfyAPIBase(ProxiedSingleton): class ComfyAPIWithVersion(NamedTuple): version: str - api_class: Type[ComfyAPIBase] + api_class: type[ComfyAPIBase] def parse_version(version_str: str) -> packaging_version.Version: @@ -23,16 +23,16 @@ def parse_version(version_str: str) -> packaging_version.Version: return packaging_version.parse(version_str) -registered_versions: List[ComfyAPIWithVersion] = [] +registered_versions: list[ComfyAPIWithVersion] = [] -def register_versions(versions: List[ComfyAPIWithVersion]): +def register_versions(versions: list[ComfyAPIWithVersion]): versions.sort(key=lambda x: parse_version(x.version)) global registered_versions registered_versions = versions -def get_all_versions() -> List[ComfyAPIWithVersion]: +def get_all_versions() -> list[ComfyAPIWithVersion]: """ Returns a list of all registered ComfyAPI versions. """ diff --git a/comfy_api/internal/async_to_sync.py b/comfy_api/internal/async_to_sync.py index f5f805a62..c9b0576e1 100644 --- a/comfy_api/internal/async_to_sync.py +++ b/comfy_api/internal/async_to_sync.py @@ -8,7 +8,7 @@ import os import textwrap import threading from enum import Enum -from typing import Optional, Type, get_origin, get_args +from typing import Optional, get_origin, get_args, get_type_hints class TypeTracker: @@ -193,7 +193,7 @@ class AsyncToSyncConverter: return result_container["result"] @classmethod - def create_sync_class(cls, async_class: Type, thread_pool_size=10) -> Type: + def create_sync_class(cls, async_class: type, thread_pool_size=10) -> type: """ Creates a new class with synchronous versions of all async methods. @@ -220,11 +220,18 @@ class AsyncToSyncConverter: self._async_instance = async_class(*args, **kwargs) # Handle annotated class attributes (like execution: Execution) - # Get all annotations from the class hierarchy - all_annotations = {} - for base_class in reversed(inspect.getmro(async_class)): - if hasattr(base_class, "__annotations__"): - all_annotations.update(base_class.__annotations__) + # Get all annotations from the class hierarchy and resolve string annotations + try: + # get_type_hints resolves string annotations to actual type objects + # This handles classes using 'from __future__ import annotations' + all_annotations = get_type_hints(async_class) + except Exception: + # Fallback to raw annotations if get_type_hints fails + # (e.g., for undefined forward references) + all_annotations = {} + for base_class in reversed(inspect.getmro(async_class)): + if hasattr(base_class, "__annotations__"): + all_annotations.update(base_class.__annotations__) # For each annotated attribute, check if it needs to be created or wrapped for attr_name, attr_type in all_annotations.items(): @@ -556,7 +563,7 @@ class AsyncToSyncConverter: @classmethod def _generate_imports( - cls, async_class: Type, type_tracker: TypeTracker + cls, async_class: type, type_tracker: TypeTracker ) -> list[str]: """Generate import statements for the stub file.""" imports = [] @@ -621,19 +628,23 @@ class AsyncToSyncConverter: return imports @classmethod - def _get_class_attributes(cls, async_class: Type) -> list[tuple[str, Type]]: + def _get_class_attributes(cls, async_class: type) -> list[tuple[str, type]]: """Extract class attributes that are classes themselves.""" class_attributes = [] + # Get resolved type hints to handle string annotations + try: + type_hints = get_type_hints(async_class) + except Exception: + type_hints = {} + # Look for class attributes that are classes for name, attr in sorted(inspect.getmembers(async_class)): if isinstance(attr, type) and not name.startswith("_"): class_attributes.append((name, attr)) - elif ( - hasattr(async_class, "__annotations__") - and name in async_class.__annotations__ - ): - annotation = async_class.__annotations__[name] + elif name in type_hints: + # Use resolved type hint instead of raw annotation + annotation = type_hints[name] if isinstance(annotation, type): class_attributes.append((name, annotation)) @@ -643,7 +654,7 @@ class AsyncToSyncConverter: def _generate_inner_class_stub( cls, name: str, - attr: Type, + attr: type, indent: str = " ", type_tracker: Optional[TypeTracker] = None, ) -> list[str]: @@ -771,7 +782,7 @@ class AsyncToSyncConverter: return processed @classmethod - def generate_stub_file(cls, async_class: Type, sync_class: Type) -> None: + def generate_stub_file(cls, async_class: type, sync_class: type) -> None: """ Generate a .pyi stub file for the sync class to help IDEs with type checking. """ @@ -908,11 +919,15 @@ class AsyncToSyncConverter: attribute_mappings = {} # First check annotations for typed attributes (including from parent classes) - # Collect all annotations from the class hierarchy - all_annotations = {} - for base_class in reversed(inspect.getmro(async_class)): - if hasattr(base_class, "__annotations__"): - all_annotations.update(base_class.__annotations__) + # Resolve string annotations to actual types + try: + all_annotations = get_type_hints(async_class) + except Exception: + # Fallback to raw annotations + all_annotations = {} + for base_class in reversed(inspect.getmro(async_class)): + if hasattr(base_class, "__annotations__"): + all_annotations.update(base_class.__annotations__) for attr_name, attr_type in sorted(all_annotations.items()): for class_name, class_type in class_attributes: @@ -973,7 +988,7 @@ class AsyncToSyncConverter: logging.error(traceback.format_exc()) -def create_sync_class(async_class: Type, thread_pool_size=10) -> Type: +def create_sync_class(async_class: type, thread_pool_size=10) -> type: """ Creates a sync version of an async class diff --git a/comfy_api/internal/singleton.py b/comfy_api/internal/singleton.py index 75f16f98e..d89380262 100644 --- a/comfy_api/internal/singleton.py +++ b/comfy_api/internal/singleton.py @@ -1,4 +1,4 @@ -from typing import Type, TypeVar +from typing import TypeVar class SingletonMetaclass(type): T = TypeVar("T", bound="SingletonMetaclass") @@ -11,13 +11,13 @@ class SingletonMetaclass(type): ) return cls._instances[cls] - def inject_instance(cls: Type[T], instance: T) -> None: + def inject_instance(cls: type[T], instance: T) -> None: assert cls not in SingletonMetaclass._instances, ( "Cannot inject instance after first instantiation" ) SingletonMetaclass._instances[cls] = instance - def get_instance(cls: Type[T], *args, **kwargs) -> T: + def get_instance(cls: type[T], *args, **kwargs) -> T: """ Gets the singleton instance of the class, creating it if it doesn't exist. """ diff --git a/comfy_api/latest/__init__.py b/comfy_api/latest/__init__.py index b7a3fa9c1..b0fa14ff6 100644 --- a/comfy_api/latest/__init__.py +++ b/comfy_api/latest/__init__.py @@ -1,16 +1,15 @@ from __future__ import annotations from abc import ABC, abstractmethod -from typing import Type, TYPE_CHECKING +from typing import TYPE_CHECKING from comfy_api.internal import ComfyAPIBase from comfy_api.internal.singleton import ProxiedSingleton from comfy_api.internal.async_to_sync import create_sync_class -from comfy_api.latest._input import ImageInput, AudioInput, MaskInput, LatentInput, VideoInput -from comfy_api.latest._input_impl import VideoFromFile, VideoFromComponents -from comfy_api.latest._util import VideoCodec, VideoContainer, VideoComponents -from . import _io as io -from . import _ui as ui -# from comfy_api.latest._resources import _RESOURCES as resources #noqa: F401 +from ._input import ImageInput, AudioInput, MaskInput, LatentInput, VideoInput +from ._input_impl import VideoFromFile, VideoFromComponents +from ._util import VideoCodec, VideoContainer, VideoComponents, MESH, VOXEL +from . import _io_public as io +from . import _ui_public as ui from comfy_execution.utils import get_executing_context from comfy_execution.progress import get_progress_state, PreviewImageTuple from PIL import Image @@ -80,7 +79,7 @@ class ComfyExtension(ABC): async def on_load(self) -> None: """ Called when an extension is loaded. - This should be used to initialize any global resources neeeded by the extension. + This should be used to initialize any global resources needed by the extension. """ @abstractmethod @@ -104,6 +103,8 @@ class Types: VideoCodec = VideoCodec VideoContainer = VideoContainer VideoComponents = VideoComponents + MESH = MESH + VOXEL = VOXEL ComfyAPI = ComfyAPI_latest @@ -111,7 +112,7 @@ ComfyAPI = ComfyAPI_latest if TYPE_CHECKING: import comfy_api.latest.generated.ComfyAPISyncStub # type: ignore - ComfyAPISync: Type[comfy_api.latest.generated.ComfyAPISyncStub.ComfyAPISyncStub] + ComfyAPISync: type[comfy_api.latest.generated.ComfyAPISyncStub.ComfyAPISyncStub] ComfyAPISync = create_sync_class(ComfyAPI_latest) # create new aliases for io and ui diff --git a/comfy_api/latest/_input/basic_types.py b/comfy_api/latest/_input/basic_types.py index 245c6cbb1..d73deabd2 100644 --- a/comfy_api/latest/_input/basic_types.py +++ b/comfy_api/latest/_input/basic_types.py @@ -1,5 +1,5 @@ import torch -from typing import TypedDict, List, Optional +from typing import TypedDict, Optional ImageInput = torch.Tensor """ @@ -39,4 +39,4 @@ class LatentInput(TypedDict): Optional noise mask tensor in the same format as samples. """ - batch_index: Optional[List[int]] + batch_index: Optional[list[int]] diff --git a/comfy_api/latest/_input/video_types.py b/comfy_api/latest/_input/video_types.py index a335df4d0..e634a0311 100644 --- a/comfy_api/latest/_input/video_types.py +++ b/comfy_api/latest/_input/video_types.py @@ -1,9 +1,10 @@ from __future__ import annotations from abc import ABC, abstractmethod +from fractions import Fraction from typing import Optional, Union, IO import io import av -from comfy_api.util import VideoContainer, VideoCodec, VideoComponents +from .._util import VideoContainer, VideoCodec, VideoComponents class VideoInput(ABC): """ @@ -72,6 +73,33 @@ class VideoInput(ABC): frame_count = components.images.shape[0] return float(frame_count / components.frame_rate) + def get_frame_count(self) -> int: + """ + Returns the number of frames in the video. + + Default implementation uses :meth:`get_components`, which may require + loading all frames into memory. File-based implementations should + override this method and use container/stream metadata instead. + + Returns: + Total number of frames as an integer. + """ + return int(self.get_components().images.shape[0]) + + def get_frame_rate(self) -> Fraction: + """ + Returns the frame rate of the video. + + Default implementation materializes the video into memory via + `get_components()`. Subclasses that can inspect the underlying + container (e.g. `VideoFromFile`) should override this with a more + efficient implementation. + + Returns: + Frame rate as a Fraction. + """ + return self.get_components().frame_rate + def get_container_format(self) -> str: """ Returns the container format of the video (e.g., 'mp4', 'mov', 'avi'). diff --git a/comfy_api/latest/_input_impl/video_types.py b/comfy_api/latest/_input_impl/video_types.py index f646504c8..ea35c6062 100644 --- a/comfy_api/latest/_input_impl/video_types.py +++ b/comfy_api/latest/_input_impl/video_types.py @@ -3,14 +3,14 @@ from av.container import InputContainer from av.subtitles.stream import SubtitleStream from fractions import Fraction from typing import Optional -from comfy_api.latest._input import AudioInput, VideoInput +from .._input import AudioInput, VideoInput import av import io import json import numpy as np import math import torch -from comfy_api.latest._util import VideoContainer, VideoCodec, VideoComponents +from .._util import VideoContainer, VideoCodec, VideoComponents def container_to_output_format(container_format: str | None) -> str | None: @@ -121,6 +121,71 @@ class VideoFromFile(VideoInput): raise ValueError(f"Could not determine duration for file '{self.__file}'") + def get_frame_count(self) -> int: + """ + Returns the number of frames in the video without materializing them as + torch tensors. + """ + if isinstance(self.__file, io.BytesIO): + self.__file.seek(0) + + with av.open(self.__file, mode="r") as container: + video_stream = self._get_first_video_stream(container) + # 1. Prefer the frames field if available + if video_stream.frames and video_stream.frames > 0: + return int(video_stream.frames) + + # 2. Try to estimate from duration and average_rate using only metadata + if container.duration is not None and video_stream.average_rate: + duration_seconds = float(container.duration / av.time_base) + estimated_frames = int(round(duration_seconds * float(video_stream.average_rate))) + if estimated_frames > 0: + return estimated_frames + + if ( + getattr(video_stream, "duration", None) is not None + and getattr(video_stream, "time_base", None) is not None + and video_stream.average_rate + ): + duration_seconds = float(video_stream.duration * video_stream.time_base) + estimated_frames = int(round(duration_seconds * float(video_stream.average_rate))) + if estimated_frames > 0: + return estimated_frames + + # 3. Last resort: decode frames and count them (streaming) + frame_count = 0 + container.seek(0) + for packet in container.demux(video_stream): + for _ in packet.decode(): + frame_count += 1 + + if frame_count == 0: + raise ValueError(f"Could not determine frame count for file '{self.__file}'") + return frame_count + + def get_frame_rate(self) -> Fraction: + """ + Returns the average frame rate of the video using container metadata + without decoding all frames. + """ + if isinstance(self.__file, io.BytesIO): + self.__file.seek(0) + + with av.open(self.__file, mode="r") as container: + video_stream = self._get_first_video_stream(container) + # Preferred: use PyAV's average_rate (usually already a Fraction-like) + if video_stream.average_rate: + return Fraction(video_stream.average_rate) + + # Fallback: estimate from frames + duration if available + if video_stream.frames and container.duration: + duration_seconds = float(container.duration / av.time_base) + if duration_seconds > 0: + return Fraction(video_stream.frames / duration_seconds).limit_denominator() + + # Last resort: match get_components_internal default + return Fraction(1) + def get_container_format(self) -> str: """ Returns the container format of the video (e.g., 'mp4', 'mov', 'avi'). @@ -238,6 +303,13 @@ class VideoFromFile(VideoInput): packet.stream = stream_map[packet.stream] output_container.mux(packet) + def _get_first_video_stream(self, container: InputContainer): + video_stream = next((s for s in container.streams if s.type == "video"), None) + if video_stream is None: + raise ValueError(f"No video stream found in file '{self.__file}'") + return video_stream + + class VideoFromComponents(VideoInput): """ Class representing video input from tensors. @@ -264,7 +336,10 @@ class VideoFromComponents(VideoInput): raise ValueError("Only MP4 format is supported for now") if codec != VideoCodec.AUTO and codec != VideoCodec.H264: raise ValueError("Only H264 codec is supported for now") - with av.open(path, mode='w', options={'movflags': 'use_metadata_tags'}) as output: + extra_kwargs = {} + if isinstance(format, VideoContainer) and format != VideoContainer.AUTO: + extra_kwargs["format"] = format.value + with av.open(path, mode='w', options={'movflags': 'use_metadata_tags'}, **extra_kwargs) as output: # Add metadata before writing any streams if metadata is not None: for key, value in metadata.items(): diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index 0b701260f..764fa8b2b 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -4,7 +4,8 @@ import copy import inspect from abc import ABC, abstractmethod from collections import Counter -from dataclasses import asdict, dataclass +from collections.abc import Iterable +from dataclasses import asdict, dataclass, field from enum import Enum from typing import Any, Callable, Literal, TypedDict, TypeVar, TYPE_CHECKING from typing_extensions import NotRequired, final @@ -25,10 +26,9 @@ if TYPE_CHECKING: from comfy_api.input import VideoInput from comfy_api.internal import (_ComfyNodeInternal, _NodeOutputInternal, classproperty, copy_class, first_real_override, is_class, prune_dict, shallow_clone_class) -from comfy_api.latest._resources import Resources, ResourcesLocal from comfy_execution.graph_utils import ExecutionBlocker +from ._util import MESH, VOXEL, SVG as _SVG -# from comfy_extras.nodes_images import SVG as SVG_ # NOTE: needs to be moved before can be imported due to circular reference class FolderType(str, Enum): input = "input" @@ -75,16 +75,6 @@ class NumberDisplay(str, Enum): slider = "slider" -class _StringIOType(str): - def __ne__(self, value: object) -> bool: - if self == "*" or value == "*": - return False - if not isinstance(value, str): - return True - a = frozenset(self.split(",")) - b = frozenset(value.split(",")) - return not (b.issubset(a) or a.issubset(b)) - class _ComfyType(ABC): Type = Any io_type: str = None @@ -124,8 +114,7 @@ def comfytype(io_type: str, **kwargs): new_cls.__module__ = cls.__module__ new_cls.__doc__ = cls.__doc__ # assign ComfyType attributes, if needed - # NOTE: use __ne__ trick for io_type (see node_typing.IO.__ne__ for details) - new_cls.io_type = _StringIOType(io_type) + new_cls.io_type = io_type if hasattr(new_cls, "Input") and new_cls.Input is not None: new_cls.Input.Parent = new_cls if hasattr(new_cls, "Output") and new_cls.Output is not None: @@ -149,6 +138,9 @@ class _IO_V3: def __init__(self): pass + def validate(self): + pass + @property def io_type(self): return self.Parent.io_type @@ -161,7 +153,7 @@ class Input(_IO_V3): ''' Base class for a V3 Input. ''' - def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None): + def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None, raw_link: bool=None): super().__init__() self.id = id self.display_name = display_name @@ -169,6 +161,7 @@ class Input(_IO_V3): self.tooltip = tooltip self.lazy = lazy self.extra_dict = extra_dict if extra_dict is not None else {} + self.rawLink = raw_link def as_dict(self): return prune_dict({ @@ -176,10 +169,14 @@ class Input(_IO_V3): "optional": self.optional, "tooltip": self.tooltip, "lazy": self.lazy, + "rawLink": self.rawLink, }) | prune_dict(self.extra_dict) def get_io_type(self): - return _StringIOType(self.io_type) + return self.io_type + + def get_all(self) -> list[Input]: + return [self] class WidgetInput(Input): ''' @@ -187,8 +184,8 @@ class WidgetInput(Input): ''' def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, default: Any=None, - socketless: bool=None, widget_type: str=None, force_input: bool=None, extra_dict=None): - super().__init__(id, display_name, optional, tooltip, lazy, extra_dict) + socketless: bool=None, widget_type: str=None, force_input: bool=None, extra_dict=None, raw_link: bool=None): + super().__init__(id, display_name, optional, tooltip, lazy, extra_dict, raw_link) self.default = default self.socketless = socketless self.widget_type = widget_type @@ -210,13 +207,14 @@ class Output(_IO_V3): def __init__(self, id: str=None, display_name: str=None, tooltip: str=None, is_output_list=False): self.id = id - self.display_name = display_name + self.display_name = display_name if display_name else id self.tooltip = tooltip self.is_output_list = is_output_list def as_dict(self): + display_name = self.display_name if self.display_name else self.id return prune_dict({ - "display_name": self.display_name, + "display_name": display_name, "tooltip": self.tooltip, "is_output_list": self.is_output_list, }) @@ -244,8 +242,8 @@ class Boolean(ComfyTypeIO): '''Boolean input.''' def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, default: bool=None, label_on: str=None, label_off: str=None, - socketless: bool=None, force_input: bool=None): - super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input) + socketless: bool=None, force_input: bool=None, extra_dict=None, raw_link: bool=None): + super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input, extra_dict, raw_link) self.label_on = label_on self.label_off = label_off self.default: bool @@ -264,8 +262,8 @@ class Int(ComfyTypeIO): '''Integer input.''' def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, default: int=None, min: int=None, max: int=None, step: int=None, control_after_generate: bool=None, - display_mode: NumberDisplay=None, socketless: bool=None, force_input: bool=None): - super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input) + display_mode: NumberDisplay=None, socketless: bool=None, force_input: bool=None, extra_dict=None, raw_link: bool=None): + super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input, extra_dict, raw_link) self.min = min self.max = max self.step = step @@ -290,8 +288,8 @@ class Float(ComfyTypeIO): '''Float input.''' def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, default: float=None, min: float=None, max: float=None, step: float=None, round: float=None, - display_mode: NumberDisplay=None, socketless: bool=None, force_input: bool=None): - super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input) + display_mode: NumberDisplay=None, socketless: bool=None, force_input: bool=None, extra_dict=None, raw_link: bool=None): + super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input, extra_dict, raw_link) self.min = min self.max = max self.step = step @@ -316,8 +314,8 @@ class String(ComfyTypeIO): '''String input.''' def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, multiline=False, placeholder: str=None, default: str=None, dynamic_prompts: bool=None, - socketless: bool=None, force_input: bool=None): - super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input) + socketless: bool=None, force_input: bool=None, extra_dict=None, raw_link: bool=None): + super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input, extra_dict, raw_link) self.multiline = multiline self.placeholder = placeholder self.dynamic_prompts = dynamic_prompts @@ -350,12 +348,14 @@ class Combo(ComfyTypeIO): image_folder: FolderType=None, remote: RemoteOptions=None, socketless: bool=None, + extra_dict=None, + raw_link: bool=None, ): if isinstance(options, type) and issubclass(options, Enum): options = [v.value for v in options] if isinstance(default, Enum): default = default.value - super().__init__(id, display_name, optional, tooltip, lazy, default, socketless) + super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, None, extra_dict, raw_link) self.multiselect = False self.options = options self.control_after_generate = control_after_generate @@ -379,10 +379,6 @@ class Combo(ComfyTypeIO): super().__init__(id, display_name, tooltip, is_output_list) self.options = options if options is not None else [] - @property - def io_type(self): - return self.options - @comfytype(io_type="COMBO") class MultiCombo(ComfyTypeI): '''Multiselect Combo input (dropdown for selecting potentially more than one value).''' @@ -391,8 +387,8 @@ class MultiCombo(ComfyTypeI): class Input(Combo.Input): def __init__(self, id: str, options: list[str], display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, default: list[str]=None, placeholder: str=None, chip: bool=None, control_after_generate: bool=None, - socketless: bool=None): - super().__init__(id, options, display_name, optional, tooltip, lazy, default, control_after_generate, socketless=socketless) + socketless: bool=None, extra_dict=None, raw_link: bool=None): + super().__init__(id, options, display_name, optional, tooltip, lazy, default, control_after_generate, socketless=socketless, extra_dict=extra_dict, raw_link=raw_link) self.multiselect = True self.placeholder = placeholder self.chip = chip @@ -425,9 +421,9 @@ class Webcam(ComfyTypeIO): Type = str def __init__( self, id: str, display_name: str=None, optional=False, - tooltip: str=None, lazy: bool=None, default: str=None, socketless: bool=None + tooltip: str=None, lazy: bool=None, default: str=None, socketless: bool=None, extra_dict=None, raw_link: bool=None ): - super().__init__(id, display_name, optional, tooltip, lazy, default, socketless) + super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, None, extra_dict, raw_link) @comfytype(io_type="MASK") @@ -560,6 +556,8 @@ class Conditioning(ComfyTypeIO): '''Used by WAN Camera.''' time_dim_concat: NotRequired[torch.Tensor] '''Used by WAN Phantom Subject.''' + time_dim_replace: NotRequired[torch.Tensor] + '''Used by Kandinsky5 I2V.''' CondList = list[tuple[torch.Tensor, PooledDict]] Type = CondList @@ -628,6 +626,10 @@ class UpscaleModel(ComfyTypeIO): if TYPE_CHECKING: Type = ImageModelDescriptor +@comfytype(io_type="LATENT_UPSCALE_MODEL") +class LatentUpscaleModel(ComfyTypeIO): + Type = Any + @comfytype(io_type="AUDIO") class Audio(ComfyTypeIO): class AudioDict(TypedDict): @@ -642,7 +644,7 @@ class Video(ComfyTypeIO): @comfytype(io_type="SVG") class SVG(ComfyTypeIO): - Type = Any # TODO: SVG class is defined in comfy_extras/nodes_images.py, causing circular reference; should be moved to somewhere else before referenced directly in v3 + Type = _SVG @comfytype(io_type="LORA_MODEL") class LoraModel(ComfyTypeIO): @@ -656,11 +658,11 @@ class LossMap(ComfyTypeIO): @comfytype(io_type="VOXEL") class Voxel(ComfyTypeIO): - Type = Any # TODO: VOXEL class is defined in comfy_extras/nodes_hunyuan3d.py; should be moved to somewhere else before referenced directly in v3 + Type = VOXEL @comfytype(io_type="MESH") class Mesh(ComfyTypeIO): - Type = Any # TODO: MESH class is defined in comfy_extras/nodes_hunyuan3d.py; should be moved to somewhere else before referenced directly in v3 + Type = MESH @comfytype(io_type="HOOKS") class Hooks(ComfyTypeIO): @@ -760,6 +762,13 @@ class AudioEncoder(ComfyTypeIO): class AudioEncoderOutput(ComfyTypeIO): Type = Any +@comfytype(io_type="TRACKS") +class Tracks(ComfyTypeIO): + class TrackDict(TypedDict): + track_path: torch.Tensor + track_visibility: torch.Tensor + Type = TrackDict + @comfytype(io_type="COMFY_MULTITYPED_V3") class MultiType: Type = Any @@ -767,7 +776,7 @@ class MultiType: ''' Input that permits more than one input type; if `id` is an instance of `ComfyType.Input`, then that input will be used to create a widget (if applicable) with overridden values. ''' - def __init__(self, id: str | Input, types: list[type[_ComfyType] | _ComfyType], display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None): + def __init__(self, id: str | Input, types: list[type[_ComfyType] | _ComfyType], display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None, raw_link: bool=None): # if id is an Input, then use that Input with overridden values self.input_override = None if isinstance(id, Input): @@ -780,7 +789,7 @@ class MultiType: # if is a widget input, make sure widget_type is set appropriately if isinstance(self.input_override, WidgetInput): self.input_override.widget_type = self.input_override.get_io_type() - super().__init__(id, display_name, optional, tooltip, lazy, extra_dict) + super().__init__(id, display_name, optional, tooltip, lazy, extra_dict, raw_link) self._io_types = types @property @@ -809,115 +818,326 @@ class MultiType: else: return super().as_dict() +@comfytype(io_type="COMFY_MATCHTYPE_V3") +class MatchType(ComfyTypeIO): + class Template: + def __init__(self, template_id: str, allowed_types: _ComfyType | list[_ComfyType] = AnyType): + self.template_id = template_id + # account for syntactic sugar + if not isinstance(allowed_types, Iterable): + allowed_types = [allowed_types] + for t in allowed_types: + if not isinstance(t, type): + if not isinstance(t, _ComfyType): + raise ValueError(f"Allowed types must be a ComfyType or a list of ComfyTypes, got {t.__class__.__name__}") + else: + if not issubclass(t, _ComfyType): + raise ValueError(f"Allowed types must be a ComfyType or a list of ComfyTypes, got {t.__name__}") + self.allowed_types = allowed_types + + def as_dict(self): + return { + "template_id": self.template_id, + "allowed_types": ",".join([t.io_type for t in self.allowed_types]), + } + + class Input(Input): + def __init__(self, id: str, template: MatchType.Template, + display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None, raw_link: bool=None): + super().__init__(id, display_name, optional, tooltip, lazy, extra_dict, raw_link) + self.template = template + + def as_dict(self): + return super().as_dict() | prune_dict({ + "template": self.template.as_dict(), + }) + + class Output(Output): + def __init__(self, template: MatchType.Template, id: str=None, display_name: str=None, tooltip: str=None, + is_output_list=False): + if not id and not display_name: + display_name = "MATCHTYPE" + super().__init__(id, display_name, tooltip, is_output_list) + self.template = template + + def as_dict(self): + return super().as_dict() | prune_dict({ + "template": self.template.as_dict(), + }) + class DynamicInput(Input, ABC): ''' Abstract class for dynamic input registration. ''' - @abstractmethod - def get_dynamic(self) -> list[Input]: - ... + pass + class DynamicOutput(Output, ABC): ''' Abstract class for dynamic output registration. ''' - def __init__(self, id: str=None, display_name: str=None, tooltip: str=None, - is_output_list=False): - super().__init__(id, display_name, tooltip, is_output_list) + pass - @abstractmethod - def get_dynamic(self) -> list[Output]: - ... +def handle_prefix(prefix_list: list[str] | None, id: str | None = None) -> list[str]: + if prefix_list is None: + prefix_list = [] + if id is not None: + prefix_list = prefix_list + [id] + return prefix_list + +def finalize_prefix(prefix_list: list[str] | None, id: str | None = None) -> str: + assert not (prefix_list is None and id is None) + if prefix_list is None: + return id + elif id is not None: + prefix_list = prefix_list + [id] + return ".".join(prefix_list) @comfytype(io_type="COMFY_AUTOGROW_V3") -class AutogrowDynamic(ComfyTypeI): - Type = list[Any] - class Input(DynamicInput): - def __init__(self, id: str, template_input: Input, min: int=1, max: int=None, - display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None): - super().__init__(id, display_name, optional, tooltip, lazy, extra_dict) - self.template_input = template_input - if min is not None: - assert(min >= 1) - if max is not None: - assert(max >= 1) +class Autogrow(ComfyTypeI): + Type = dict[str, Any] + _MaxNames = 100 # NOTE: max 100 names for sanity + + class _AutogrowTemplate: + def __init__(self, input: Input): + # dynamic inputs are not allowed as the template input + assert(not isinstance(input, DynamicInput)) + self.input = copy.copy(input) + if isinstance(self.input, WidgetInput): + self.input.force_input = True + self.names: list[str] = [] + self.cached_inputs = {} + + def _create_input(self, input: Input, name: str): + new_input = copy.copy(self.input) + new_input.id = name + return new_input + + def _create_cached_inputs(self): + for name in self.names: + self.cached_inputs[name] = self._create_input(self.input, name) + + def get_all(self) -> list[Input]: + return list(self.cached_inputs.values()) + + def as_dict(self): + return prune_dict({ + "input": create_input_dict_v1([self.input]), + }) + + def validate(self): + self.input.validate() + + class TemplatePrefix(_AutogrowTemplate): + def __init__(self, input: Input, prefix: str, min: int=1, max: int=10): + super().__init__(input) + self.prefix = prefix + assert(min >= 0) + assert(max >= 1) + assert(max <= Autogrow._MaxNames) self.min = min self.max = max + self.names = [f"{self.prefix}{i}" for i in range(self.max)] + self._create_cached_inputs() - def get_dynamic(self) -> list[Input]: - curr_count = 1 - new_inputs = [] - for i in range(self.min): - new_input = copy.copy(self.template_input) - new_input.id = f"{new_input.id}{curr_count}_${self.id}_ag$" - if new_input.display_name is not None: - new_input.display_name = f"{new_input.display_name}{curr_count}" - new_input.optional = self.optional or new_input.optional - if isinstance(self.template_input, WidgetInput): - new_input.force_input = True - new_inputs.append(new_input) - curr_count += 1 - # pretend to expand up to max - for i in range(curr_count-1, self.max): - new_input = copy.copy(self.template_input) - new_input.id = f"{new_input.id}{curr_count}_${self.id}_ag$" - if new_input.display_name is not None: - new_input.display_name = f"{new_input.display_name}{curr_count}" - new_input.optional = True - if isinstance(self.template_input, WidgetInput): - new_input.force_input = True - new_inputs.append(new_input) - curr_count += 1 - return new_inputs + def as_dict(self): + return super().as_dict() | prune_dict({ + "prefix": self.prefix, + "min": self.min, + "max": self.max, + }) + + class TemplateNames(_AutogrowTemplate): + def __init__(self, input: Input, names: list[str], min: int=1): + super().__init__(input) + self.names = names[:Autogrow._MaxNames] + assert(min >= 0) + self.min = min + self._create_cached_inputs() + + def as_dict(self): + return super().as_dict() | prune_dict({ + "names": self.names, + "min": self.min, + }) -@comfytype(io_type="COMFY_COMBODYNAMIC_V3") -class ComboDynamic(ComfyTypeI): class Input(DynamicInput): - def __init__(self, id: str): - pass + def __init__(self, id: str, template: Autogrow.TemplatePrefix | Autogrow.TemplateNames, + display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None): + super().__init__(id, display_name, optional, tooltip, lazy, extra_dict) + self.template = template -@comfytype(io_type="COMFY_MATCHTYPE_V3") -class MatchType(ComfyTypeIO): - class Template: - def __init__(self, template_id: str, allowed_types: _ComfyType | list[_ComfyType]): - self.template_id = template_id - self.allowed_types = [allowed_types] if isinstance(allowed_types, _ComfyType) else allowed_types + def as_dict(self): + return super().as_dict() | prune_dict({ + "template": self.template.as_dict(), + }) + + def get_all(self) -> list[Input]: + return [self] + self.template.get_all() + + def validate(self): + self.template.validate() + + @staticmethod + def _expand_schema_for_dynamic(out_dict: dict[str, Any], live_inputs: dict[str, Any], value: tuple[str, dict[str, Any]], input_type: str, curr_prefix: list[str] | None): + # NOTE: purposely do not include self in out_dict; instead use only the template inputs + # need to figure out names based on template type + is_names = ("names" in value[1]["template"]) + is_prefix = ("prefix" in value[1]["template"]) + input = value[1]["template"]["input"] + if is_names: + min = value[1]["template"]["min"] + names = value[1]["template"]["names"] + max = len(names) + elif is_prefix: + prefix = value[1]["template"]["prefix"] + min = value[1]["template"]["min"] + max = value[1]["template"]["max"] + names = [f"{prefix}{i}" for i in range(max)] + # need to create a new input based on the contents of input + template_input = None + for _, dict_input in input.items(): + # for now, get just the first value from dict_input + template_input = list(dict_input.values())[0] + new_dict = {} + for i, name in enumerate(names): + expected_id = finalize_prefix(curr_prefix, name) + if expected_id in live_inputs: + # required + if i < min: + type_dict = new_dict.setdefault("required", {}) + # optional + else: + type_dict = new_dict.setdefault("optional", {}) + type_dict[name] = template_input + parse_class_inputs(out_dict, live_inputs, new_dict, curr_prefix) + +@comfytype(io_type="COMFY_DYNAMICCOMBO_V3") +class DynamicCombo(ComfyTypeI): + Type = dict[str, Any] + + class Option: + def __init__(self, key: str, inputs: list[Input]): + self.key = key + self.inputs = inputs def as_dict(self): return { - "template_id": self.template_id, - "allowed_types": "".join(t.io_type for t in self.allowed_types), + "key": self.key, + "inputs": create_input_dict_v1(self.inputs), } class Input(DynamicInput): - def __init__(self, id: str, template: MatchType.Template, + def __init__(self, id: str, options: list[DynamicCombo.Option], display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None): super().__init__(id, display_name, optional, tooltip, lazy, extra_dict) - self.template = template + self.options = options - def get_dynamic(self) -> list[Input]: - return [self] + def get_all(self) -> list[Input]: + return [self] + [input for option in self.options for input in option.inputs] def as_dict(self): return super().as_dict() | prune_dict({ - "template": self.template.as_dict(), + "options": [o.as_dict() for o in self.options], }) - class Output(DynamicOutput): - def __init__(self, id: str, template: MatchType.Template, display_name: str=None, tooltip: str=None, - is_output_list=False): - super().__init__(id, display_name, tooltip, is_output_list) - self.template = template + def validate(self): + # make sure all nested inputs are validated + for option in self.options: + for input in option.inputs: + input.validate() - def get_dynamic(self) -> list[Output]: - return [self] + @staticmethod + def _expand_schema_for_dynamic(out_dict: dict[str, Any], live_inputs: dict[str, Any], value: tuple[str, dict[str, Any]], input_type: str, curr_prefix: list[str] | None): + finalized_id = finalize_prefix(curr_prefix) + if finalized_id in live_inputs: + key = live_inputs[finalized_id] + selected_option = None + # get options from dict + options: list[dict[str, str | dict[str, Any]]] = value[1]["options"] + for option in options: + if option["key"] == key: + selected_option = option + break + if selected_option is not None: + parse_class_inputs(out_dict, live_inputs, selected_option["inputs"], curr_prefix) + # add self to inputs + out_dict[input_type][finalized_id] = value + out_dict["dynamic_paths"][finalized_id] = finalize_prefix(curr_prefix, curr_prefix[-1]) + +@comfytype(io_type="COMFY_DYNAMICSLOT_V3") +class DynamicSlot(ComfyTypeI): + Type = dict[str, Any] + + class Input(DynamicInput): + def __init__(self, slot: Input, inputs: list[Input], + display_name: str=None, tooltip: str=None, lazy: bool=None, extra_dict=None): + assert(not isinstance(slot, DynamicInput)) + self.slot = copy.copy(slot) + self.slot.display_name = slot.display_name if slot.display_name is not None else display_name + optional = True + self.slot.tooltip = slot.tooltip if slot.tooltip is not None else tooltip + self.slot.lazy = slot.lazy if slot.lazy is not None else lazy + self.slot.extra_dict = slot.extra_dict if slot.extra_dict is not None else extra_dict + super().__init__(slot.id, self.slot.display_name, optional, self.slot.tooltip, self.slot.lazy, self.slot.extra_dict) + self.inputs = inputs + self.force_input = None + # force widget inputs to have no widgets, otherwise this would be awkward + if isinstance(self.slot, WidgetInput): + self.force_input = True + self.slot.force_input = True + + def get_all(self) -> list[Input]: + return [self.slot] + self.inputs def as_dict(self): return super().as_dict() | prune_dict({ - "template": self.template.as_dict(), + "slotType": str(self.slot.get_io_type()), + "inputs": create_input_dict_v1(self.inputs), + "forceInput": self.force_input, }) + def validate(self): + self.slot.validate() + for input in self.inputs: + input.validate() + + @staticmethod + def _expand_schema_for_dynamic(out_dict: dict[str, Any], live_inputs: dict[str, Any], value: tuple[str, dict[str, Any]], input_type: str, curr_prefix: list[str] | None): + finalized_id = finalize_prefix(curr_prefix) + if finalized_id in live_inputs: + inputs = value[1]["inputs"] + parse_class_inputs(out_dict, live_inputs, inputs, curr_prefix) + # add self to inputs + out_dict[input_type][finalized_id] = value + out_dict["dynamic_paths"][finalized_id] = finalize_prefix(curr_prefix, curr_prefix[-1]) + +DYNAMIC_INPUT_LOOKUP: dict[str, Callable[[dict[str, Any], dict[str, Any], tuple[str, dict[str, Any]], str, list[str] | None], None]] = {} +def register_dynamic_input_func(io_type: str, func: Callable[[dict[str, Any], dict[str, Any], tuple[str, dict[str, Any]], str, list[str] | None], None]): + DYNAMIC_INPUT_LOOKUP[io_type] = func + +def get_dynamic_input_func(io_type: str) -> Callable[[dict[str, Any], dict[str, Any], tuple[str, dict[str, Any]], str, list[str] | None], None]: + return DYNAMIC_INPUT_LOOKUP[io_type] + +def setup_dynamic_input_funcs(): + # Autogrow.Input + register_dynamic_input_func(Autogrow.io_type, Autogrow._expand_schema_for_dynamic) + # DynamicCombo.Input + register_dynamic_input_func(DynamicCombo.io_type, DynamicCombo._expand_schema_for_dynamic) + # DynamicSlot.Input + register_dynamic_input_func(DynamicSlot.io_type, DynamicSlot._expand_schema_for_dynamic) + +if len(DYNAMIC_INPUT_LOOKUP) == 0: + setup_dynamic_input_funcs() + +class V3Data(TypedDict): + hidden_inputs: dict[str, Any] + 'Dictionary where the keys are the hidden input ids and the values are the values of the hidden inputs.' + dynamic_paths: dict[str, Any] + 'Dictionary where the keys are the input ids and the values dictate how to turn the inputs into a nested dictionary.' + create_dynamic_tuple: bool + 'When True, the value of the dynamic input will be in the format (value, path_key).' class HiddenHolder: def __init__(self, unique_id: str, prompt: Any, @@ -953,6 +1173,10 @@ class HiddenHolder: api_key_comfy_org=d.get(Hidden.api_key_comfy_org, None), ) + @classmethod + def from_v3_data(cls, v3_data: V3Data | None) -> HiddenHolder: + return cls.from_dict(v3_data["hidden_inputs"] if v3_data else None) + class Hidden(str, Enum): ''' Enumerator for requesting hidden variables in nodes. @@ -979,6 +1203,7 @@ class NodeInfoV1: output_is_list: list[bool]=None output_name: list[str]=None output_tooltips: list[str]=None + output_matchtypes: list[str]=None name: str=None display_name: str=None description: str=None @@ -1014,9 +1239,9 @@ class Schema: """Display name of node.""" category: str = "sd" """The category of the node, as per the "Add Node" menu.""" - inputs: list[Input]=None - outputs: list[Output]=None - hidden: list[Hidden]=None + inputs: list[Input] = field(default_factory=list) + outputs: list[Output] = field(default_factory=list) + hidden: list[Hidden] = field(default_factory=list) description: str="" """Node description, shown as a tooltip when hovering over the node.""" is_input_list: bool = False @@ -1056,60 +1281,57 @@ class Schema: '''Validate the schema: - verify ids on inputs and outputs are unique - both internally and in relation to each other ''' - input_ids = [i.id for i in self.inputs] if self.inputs is not None else [] - output_ids = [o.id for o in self.outputs] if self.outputs is not None else [] + nested_inputs: list[Input] = [] + for input in self.inputs: + if not isinstance(input, DynamicInput): + nested_inputs.extend(input.get_all()) + input_ids = [i.id for i in nested_inputs] + output_ids = [o.id for o in self.outputs] input_set = set(input_ids) output_set = set(output_ids) - issues = [] + issues: list[str] = [] # verify ids are unique per list if len(input_set) != len(input_ids): issues.append(f"Input ids must be unique, but {[item for item, count in Counter(input_ids).items() if count > 1]} are not.") if len(output_set) != len(output_ids): issues.append(f"Output ids must be unique, but {[item for item, count in Counter(output_ids).items() if count > 1]} are not.") - # verify ids are unique between lists - intersection = input_set & output_set - if len(intersection) > 0: - issues.append(f"Ids must be unique between inputs and outputs, but {intersection} are not.") if len(issues) > 0: raise ValueError("\n".join(issues)) + # validate inputs and outputs + for input in self.inputs: + input.validate() + for output in self.outputs: + output.validate() def finalize(self): """Add hidden based on selected schema options, and give outputs without ids default ids.""" + # ensure inputs, outputs, and hidden are lists + if self.inputs is None: + self.inputs = [] + if self.outputs is None: + self.outputs = [] + if self.hidden is None: + self.hidden = [] # if is an api_node, will need key-related hidden if self.is_api_node: - if self.hidden is None: - self.hidden = [] if Hidden.auth_token_comfy_org not in self.hidden: self.hidden.append(Hidden.auth_token_comfy_org) if Hidden.api_key_comfy_org not in self.hidden: self.hidden.append(Hidden.api_key_comfy_org) # if is an output_node, will need prompt and extra_pnginfo if self.is_output_node: - if self.hidden is None: - self.hidden = [] if Hidden.prompt not in self.hidden: self.hidden.append(Hidden.prompt) if Hidden.extra_pnginfo not in self.hidden: self.hidden.append(Hidden.extra_pnginfo) # give outputs without ids default ids - if self.outputs is not None: - for i, output in enumerate(self.outputs): - if output.id is None: - output.id = f"_{i}_{output.io_type}_" + for i, output in enumerate(self.outputs): + if output.id is None: + output.id = f"_{i}_{output.io_type}_" def get_v1_info(self, cls) -> NodeInfoV1: # get V1 inputs - input = { - "required": {} - } - if self.inputs: - for i in self.inputs: - if isinstance(i, DynamicInput): - dynamic_inputs = i.get_dynamic() - for d in dynamic_inputs: - add_to_dict_v1(d, input) - else: - add_to_dict_v1(i, input) + input = create_input_dict_v1(self.inputs) if self.hidden: for hidden in self.hidden: input.setdefault("hidden", {})[hidden.name] = (hidden.value,) @@ -1118,12 +1340,24 @@ class Schema: output_is_list = [] output_name = [] output_tooltips = [] + output_matchtypes = [] + any_matchtypes = False if self.outputs: for o in self.outputs: output.append(o.io_type) output_is_list.append(o.is_output_list) output_name.append(o.display_name if o.display_name else o.io_type) output_tooltips.append(o.tooltip if o.tooltip else None) + # special handling for MatchType + if isinstance(o, MatchType.Output): + output_matchtypes.append(o.template.template_id) + any_matchtypes = True + else: + output_matchtypes.append(None) + + # clear out lists that are all None + if not any_matchtypes: + output_matchtypes = None info = NodeInfoV1( input=input, @@ -1132,6 +1366,7 @@ class Schema: output_is_list=output_is_list, output_name=output_name, output_tooltips=output_tooltips, + output_matchtypes=output_matchtypes, name=self.node_id, display_name=self.display_name, category=self.category, @@ -1176,17 +1411,84 @@ class Schema: ) return info +def get_finalized_class_inputs(d: dict[str, Any], live_inputs: dict[str, Any], include_hidden=False) -> tuple[dict[str, Any], V3Data]: + out_dict = { + "required": {}, + "optional": {}, + "dynamic_paths": {}, + } + d = d.copy() + # ignore hidden for parsing + hidden = d.pop("hidden", None) + parse_class_inputs(out_dict, live_inputs, d) + if hidden is not None and include_hidden: + out_dict["hidden"] = hidden + v3_data = {} + dynamic_paths = out_dict.pop("dynamic_paths", None) + if dynamic_paths is not None: + v3_data["dynamic_paths"] = dynamic_paths + return out_dict, hidden, v3_data -def add_to_dict_v1(i: Input, input: dict): +def parse_class_inputs(out_dict: dict[str, Any], live_inputs: dict[str, Any], curr_dict: dict[str, Any], curr_prefix: list[str] | None=None) -> None: + for input_type, inner_d in curr_dict.items(): + for id, value in inner_d.items(): + io_type = value[0] + if io_type in DYNAMIC_INPUT_LOOKUP: + # dynamic inputs need to be handled with lookup functions + dynamic_input_func = get_dynamic_input_func(io_type) + new_prefix = handle_prefix(curr_prefix, id) + dynamic_input_func(out_dict, live_inputs, value, input_type, new_prefix) + else: + # non-dynamic inputs get directly transferred + finalized_id = finalize_prefix(curr_prefix, id) + out_dict[input_type][finalized_id] = value + if curr_prefix: + out_dict["dynamic_paths"][finalized_id] = finalized_id + +def create_input_dict_v1(inputs: list[Input]) -> dict: + input = { + "required": {} + } + for i in inputs: + add_to_dict_v1(i, input) + return input + +def add_to_dict_v1(i: Input, d: dict): key = "optional" if i.optional else "required" as_dict = i.as_dict() # for v1, we don't want to include the optional key as_dict.pop("optional", None) - input.setdefault(key, {})[i.id] = (i.get_io_type(), as_dict) + d.setdefault(key, {})[i.id] = (i.get_io_type(), as_dict) def add_to_dict_v3(io: Input | Output, d: dict): d[io.id] = (io.get_io_type(), io.as_dict()) +def build_nested_inputs(values: dict[str, Any], v3_data: V3Data): + paths = v3_data.get("dynamic_paths", None) + if paths is None: + return values + values = values.copy() + result = {} + + create_tuple = v3_data.get("create_dynamic_tuple", False) + + for key, path in paths.items(): + parts = path.split(".") + current = result + + for i, p in enumerate(parts): + is_last = (i == len(parts) - 1) + + if is_last: + value = values.pop(key, None) + if create_tuple: + value = (value, key) + current[p] = value + else: + current = current.setdefault(p, {}) + + values.update(result) + return values class _ComfyNodeBaseInternal(_ComfyNodeInternal): @@ -1196,7 +1498,6 @@ class _ComfyNodeBaseInternal(_ComfyNodeInternal): SCHEMA = None # filled in during execution - resources: Resources = None hidden: HiddenHolder = None @classmethod @@ -1243,7 +1544,6 @@ class _ComfyNodeBaseInternal(_ComfyNodeInternal): return [name for name in kwargs if kwargs[name] is None] def __init__(self): - self.local_resources: ResourcesLocal = None self.__class__.VALIDATE_CLASS() @classmethod @@ -1306,12 +1606,12 @@ class _ComfyNodeBaseInternal(_ComfyNodeInternal): @final @classmethod - def PREPARE_CLASS_CLONE(cls, hidden_inputs: dict) -> type[ComfyNode]: + def PREPARE_CLASS_CLONE(cls, v3_data: V3Data | None) -> type[ComfyNode]: """Creates clone of real node class to prevent monkey-patching.""" c_type: type[ComfyNode] = cls if is_class(cls) else type(cls) type_clone: type[ComfyNode] = shallow_clone_class(c_type) # set hidden - type_clone.hidden = HiddenHolder.from_dict(hidden_inputs) + type_clone.hidden = HiddenHolder.from_v3_data(v3_data) return type_clone @final @@ -1428,15 +1728,10 @@ class _ComfyNodeBaseInternal(_ComfyNodeInternal): @final @classmethod - def INPUT_TYPES(cls, include_hidden=True, return_schema=False) -> dict[str, dict] | tuple[dict[str, dict], Schema]: + def INPUT_TYPES(cls) -> dict[str, dict]: schema = cls.FINALIZE_SCHEMA() info = schema.get_v1_info(cls) - input = info.input - if not include_hidden: - input.pop("hidden", None) - if return_schema: - return input, schema - return input + return info.input @final @classmethod @@ -1508,7 +1803,7 @@ class ComfyNode(_ComfyNodeBaseInternal): raise NotImplementedError @classmethod - def validate_inputs(cls, **kwargs) -> bool: + def validate_inputs(cls, **kwargs) -> bool | str: """Optionally, define this function to validate inputs; equivalent to V1's VALIDATE_INPUTS.""" raise NotImplementedError @@ -1555,7 +1850,7 @@ class NodeOutput(_NodeOutputInternal): return self.args if len(self.args) > 0 else None @classmethod - def from_dict(cls, data: dict[str, Any]) -> "NodeOutput": + def from_dict(cls, data: dict[str, Any]) -> NodeOutput: args = () ui = None expand = None @@ -1568,7 +1863,7 @@ class NodeOutput(_NodeOutputInternal): ui = data["ui"] if "expand" in data: expand = data["expand"] - return cls(args=args, ui=ui, expand=expand) + return cls(*args, ui=ui, expand=expand) def __getitem__(self, index) -> Any: return self.args[index] @@ -1623,6 +1918,7 @@ __all__ = [ "StyleModel", "Gligen", "UpscaleModel", + "LatentUpscaleModel", "Audio", "Video", "SVG", @@ -1646,6 +1942,11 @@ __all__ = [ "SEGS", "AnyType", "MultiType", + "Tracks", + # Dynamic Types + "MatchType", + "DynamicCombo", + "Autogrow", # Other classes "HiddenHolder", "Hidden", @@ -1656,4 +1957,5 @@ __all__ = [ "NodeOutput", "add_to_dict_v1", "add_to_dict_v3", + "V3Data", ] diff --git a/comfy_api/latest/_io_public.py b/comfy_api/latest/_io_public.py new file mode 100644 index 000000000..43c7680f3 --- /dev/null +++ b/comfy_api/latest/_io_public.py @@ -0,0 +1 @@ +from ._io import * # noqa: F403 diff --git a/comfy_api/latest/_resources.py b/comfy_api/latest/_resources.py deleted file mode 100644 index a6bdda972..000000000 --- a/comfy_api/latest/_resources.py +++ /dev/null @@ -1,72 +0,0 @@ -from __future__ import annotations -import comfy.utils -import folder_paths -import logging -from abc import ABC, abstractmethod -from typing import Any -import torch - -class ResourceKey(ABC): - Type = Any - def __init__(self): - ... - -class TorchDictFolderFilename(ResourceKey): - '''Key for requesting a torch file via file_name from a folder category.''' - Type = dict[str, torch.Tensor] - def __init__(self, folder_name: str, file_name: str): - self.folder_name = folder_name - self.file_name = file_name - - def __hash__(self): - return hash((self.folder_name, self.file_name)) - - def __eq__(self, other: object) -> bool: - if not isinstance(other, TorchDictFolderFilename): - return False - return self.folder_name == other.folder_name and self.file_name == other.file_name - - def __str__(self): - return f"{self.folder_name} -> {self.file_name}" - -class Resources(ABC): - def __init__(self): - ... - - @abstractmethod - def get(self, key: ResourceKey, default: Any=...) -> Any: - pass - -class ResourcesLocal(Resources): - def __init__(self): - super().__init__() - self.local_resources: dict[ResourceKey, Any] = {} - - def get(self, key: ResourceKey, default: Any=...) -> Any: - cached = self.local_resources.get(key, None) - if cached is not None: - logging.info(f"Using cached resource '{key}'") - return cached - logging.info(f"Loading resource '{key}'") - to_return = None - if isinstance(key, TorchDictFolderFilename): - if default is ...: - to_return = comfy.utils.load_torch_file(folder_paths.get_full_path_or_raise(key.folder_name, key.file_name), safe_load=True) - else: - full_path = folder_paths.get_full_path(key.folder_name, key.file_name) - if full_path is not None: - to_return = comfy.utils.load_torch_file(full_path, safe_load=True) - - if to_return is not None: - self.local_resources[key] = to_return - return to_return - if default is not ...: - return default - raise Exception(f"Unsupported resource key type: {type(key)}") - - -class _RESOURCES: - ResourceKey = ResourceKey - TorchDictFolderFilename = TorchDictFolderFilename - Resources = Resources - ResourcesLocal = ResourcesLocal diff --git a/comfy_api/latest/_ui.py b/comfy_api/latest/_ui.py index b0bbabe2a..e238cdf3c 100644 --- a/comfy_api/latest/_ui.py +++ b/comfy_api/latest/_ui.py @@ -3,8 +3,8 @@ from __future__ import annotations import json import os import random +import uuid from io import BytesIO -from typing import Type import av import numpy as np @@ -21,7 +21,7 @@ import folder_paths # used for image preview from comfy.cli_args import args -from comfy_api.latest._io import ComfyNode, FolderType, Image, _UIOutput +from ._io import ComfyNode, FolderType, Image, _UIOutput class SavedResult(dict): @@ -82,7 +82,7 @@ class ImageSaveHelper: return PILImage.fromarray(np.clip(255.0 * image_tensor.cpu().numpy(), 0, 255).astype(np.uint8)) @staticmethod - def _create_png_metadata(cls: Type[ComfyNode] | None) -> PngInfo | None: + def _create_png_metadata(cls: type[ComfyNode] | None) -> PngInfo | None: """Creates a PngInfo object with prompt and extra_pnginfo.""" if args.disable_metadata or cls is None or not cls.hidden: return None @@ -95,7 +95,7 @@ class ImageSaveHelper: return metadata @staticmethod - def _create_animated_png_metadata(cls: Type[ComfyNode] | None) -> PngInfo | None: + def _create_animated_png_metadata(cls: type[ComfyNode] | None) -> PngInfo | None: """Creates a PngInfo object with prompt and extra_pnginfo for animated PNGs (APNG).""" if args.disable_metadata or cls is None or not cls.hidden: return None @@ -120,7 +120,7 @@ class ImageSaveHelper: return metadata @staticmethod - def _create_webp_metadata(pil_image: PILImage.Image, cls: Type[ComfyNode] | None) -> PILImage.Exif: + def _create_webp_metadata(pil_image: PILImage.Image, cls: type[ComfyNode] | None) -> PILImage.Exif: """Creates EXIF metadata bytes for WebP images.""" exif_data = pil_image.getexif() if args.disable_metadata or cls is None or cls.hidden is None: @@ -136,7 +136,7 @@ class ImageSaveHelper: @staticmethod def save_images( - images, filename_prefix: str, folder_type: FolderType, cls: Type[ComfyNode] | None, compress_level = 4, + images, filename_prefix: str, folder_type: FolderType, cls: type[ComfyNode] | None, compress_level = 4, ) -> list[SavedResult]: """Saves a batch of images as individual PNG files.""" full_output_folder, filename, counter, subfolder, _ = folder_paths.get_save_image_path( @@ -154,7 +154,7 @@ class ImageSaveHelper: return results @staticmethod - def get_save_images_ui(images, filename_prefix: str, cls: Type[ComfyNode] | None, compress_level=4) -> SavedImages: + def get_save_images_ui(images, filename_prefix: str, cls: type[ComfyNode] | None, compress_level=4) -> SavedImages: """Saves a batch of images and returns a UI object for the node output.""" return SavedImages( ImageSaveHelper.save_images( @@ -168,7 +168,7 @@ class ImageSaveHelper: @staticmethod def save_animated_png( - images, filename_prefix: str, folder_type: FolderType, cls: Type[ComfyNode] | None, fps: float, compress_level: int + images, filename_prefix: str, folder_type: FolderType, cls: type[ComfyNode] | None, fps: float, compress_level: int ) -> SavedResult: """Saves a batch of images as a single animated PNG.""" full_output_folder, filename, counter, subfolder, _ = folder_paths.get_save_image_path( @@ -190,7 +190,7 @@ class ImageSaveHelper: @staticmethod def get_save_animated_png_ui( - images, filename_prefix: str, cls: Type[ComfyNode] | None, fps: float, compress_level: int + images, filename_prefix: str, cls: type[ComfyNode] | None, fps: float, compress_level: int ) -> SavedImages: """Saves an animated PNG and returns a UI object for the node output.""" result = ImageSaveHelper.save_animated_png( @@ -208,7 +208,7 @@ class ImageSaveHelper: images, filename_prefix: str, folder_type: FolderType, - cls: Type[ComfyNode] | None, + cls: type[ComfyNode] | None, fps: float, lossless: bool, quality: int, @@ -237,7 +237,7 @@ class ImageSaveHelper: def get_save_animated_webp_ui( images, filename_prefix: str, - cls: Type[ComfyNode] | None, + cls: type[ComfyNode] | None, fps: float, lossless: bool, quality: int, @@ -266,7 +266,7 @@ class AudioSaveHelper: audio: dict, filename_prefix: str, folder_type: FolderType, - cls: Type[ComfyNode] | None, + cls: type[ComfyNode] | None, format: str = "flac", quality: str = "128k", ) -> list[SavedResult]: @@ -318,9 +318,10 @@ class AudioSaveHelper: for key, value in metadata.items(): output_container.metadata[key] = value + layout = "mono" if waveform.shape[0] == 1 else "stereo" # Set up the output stream with appropriate properties if format == "opus": - out_stream = output_container.add_stream("libopus", rate=sample_rate) + out_stream = output_container.add_stream("libopus", rate=sample_rate, layout=layout) if quality == "64k": out_stream.bit_rate = 64000 elif quality == "96k": @@ -332,7 +333,7 @@ class AudioSaveHelper: elif quality == "320k": out_stream.bit_rate = 320000 elif format == "mp3": - out_stream = output_container.add_stream("libmp3lame", rate=sample_rate) + out_stream = output_container.add_stream("libmp3lame", rate=sample_rate, layout=layout) if quality == "V0": # TODO i would really love to support V3 and V5 but there doesn't seem to be a way to set the qscale level, the property below is a bool out_stream.codec_context.qscale = 1 @@ -341,12 +342,12 @@ class AudioSaveHelper: elif quality == "320k": out_stream.bit_rate = 320000 else: # format == "flac": - out_stream = output_container.add_stream("flac", rate=sample_rate) + out_stream = output_container.add_stream("flac", rate=sample_rate, layout=layout) frame = av.AudioFrame.from_ndarray( waveform.movedim(0, 1).reshape(1, -1).float().numpy(), format="flt", - layout="mono" if waveform.shape[0] == 1 else "stereo", + layout=layout, ) frame.sample_rate = sample_rate frame.pts = 0 @@ -370,7 +371,7 @@ class AudioSaveHelper: @staticmethod def get_save_audio_ui( - audio, filename_prefix: str, cls: Type[ComfyNode] | None, format: str = "flac", quality: str = "128k", + audio, filename_prefix: str, cls: type[ComfyNode] | None, format: str = "flac", quality: str = "128k", ) -> SavedAudios: """Save and instantly wrap for UI.""" return SavedAudios( @@ -386,7 +387,7 @@ class AudioSaveHelper: class PreviewImage(_UIOutput): - def __init__(self, image: Image.Type, animated: bool = False, cls: Type[ComfyNode] = None, **kwargs): + def __init__(self, image: Image.Type, animated: bool = False, cls: type[ComfyNode] = None, **kwargs): self.values = ImageSaveHelper.save_images( image, filename_prefix="ComfyUI_temp_" + ''.join(random.choice("abcdefghijklmnopqrstupvxyz") for _ in range(5)), @@ -410,7 +411,7 @@ class PreviewMask(PreviewImage): class PreviewAudio(_UIOutput): - def __init__(self, audio: dict, cls: Type[ComfyNode] = None, **kwargs): + def __init__(self, audio: dict, cls: type[ComfyNode] = None, **kwargs): self.values = AudioSaveHelper.save_audio( audio, filename_prefix="ComfyUI_temp_" + "".join(random.choice("abcdefghijklmnopqrstuvwxyz") for _ in range(5)), @@ -436,9 +437,19 @@ class PreviewUI3D(_UIOutput): def __init__(self, model_file, camera_info, **kwargs): self.model_file = model_file self.camera_info = camera_info + self.bg_image_path = None + bg_image = kwargs.get("bg_image", None) + if bg_image is not None: + img_array = (bg_image[0].cpu().numpy() * 255).astype(np.uint8) + img = PILImage.fromarray(img_array) + temp_dir = folder_paths.get_temp_directory() + filename = f"bg_{uuid.uuid4().hex}.png" + bg_image_path = os.path.join(temp_dir, filename) + img.save(bg_image_path, compress_level=1) + self.bg_image_path = f"temp/{filename}" def as_dict(self): - return {"result": [self.model_file, self.camera_info]} + return {"result": [self.model_file, self.camera_info, self.bg_image_path]} class PreviewText(_UIOutput): diff --git a/comfy_api/latest/_ui_public.py b/comfy_api/latest/_ui_public.py new file mode 100644 index 000000000..85b11d78b --- /dev/null +++ b/comfy_api/latest/_ui_public.py @@ -0,0 +1 @@ +from ._ui import * # noqa: F403 diff --git a/comfy_api/latest/_util/__init__.py b/comfy_api/latest/_util/__init__.py index 9019c46db..6313eb01b 100644 --- a/comfy_api/latest/_util/__init__.py +++ b/comfy_api/latest/_util/__init__.py @@ -1,8 +1,13 @@ from .video_types import VideoContainer, VideoCodec, VideoComponents +from .geometry_types import VOXEL, MESH +from .image_types import SVG __all__ = [ # Utility Types "VideoContainer", "VideoCodec", "VideoComponents", + "VOXEL", + "MESH", + "SVG", ] diff --git a/comfy_api/latest/_util/geometry_types.py b/comfy_api/latest/_util/geometry_types.py new file mode 100644 index 000000000..385122778 --- /dev/null +++ b/comfy_api/latest/_util/geometry_types.py @@ -0,0 +1,12 @@ +import torch + + +class VOXEL: + def __init__(self, data: torch.Tensor): + self.data = data + + +class MESH: + def __init__(self, vertices: torch.Tensor, faces: torch.Tensor): + self.vertices = vertices + self.faces = faces diff --git a/comfy_api/latest/_util/image_types.py b/comfy_api/latest/_util/image_types.py new file mode 100644 index 000000000..f031ed426 --- /dev/null +++ b/comfy_api/latest/_util/image_types.py @@ -0,0 +1,18 @@ +from io import BytesIO + + +class SVG: + """Stores SVG representations via a list of BytesIO objects.""" + + def __init__(self, data: list[BytesIO]): + self.data = data + + def combine(self, other: 'SVG') -> 'SVG': + return SVG(self.data + other.data) + + @staticmethod + def combine_all(svgs: list['SVG']) -> 'SVG': + all_svgs_list: list[BytesIO] = [] + for svg_item in svgs: + all_svgs_list.extend(svg_item.data) + return SVG(all_svgs_list) diff --git a/comfy_api/latest/_util/video_types.py b/comfy_api/latest/_util/video_types.py index c3e3d8e3a..fd3b5a510 100644 --- a/comfy_api/latest/_util/video_types.py +++ b/comfy_api/latest/_util/video_types.py @@ -3,7 +3,7 @@ from dataclasses import dataclass from enum import Enum from fractions import Fraction from typing import Optional -from comfy_api.latest._input import ImageInput, AudioInput +from .._input import ImageInput, AudioInput class VideoCodec(str, Enum): AUTO = "auto" diff --git a/comfy_api/v0_0_2/__init__.py b/comfy_api/v0_0_2/__init__.py index de0f95001..c4fa1d971 100644 --- a/comfy_api/v0_0_2/__init__.py +++ b/comfy_api/v0_0_2/__init__.py @@ -6,7 +6,7 @@ from comfy_api.latest import ( ) from typing import Type, TYPE_CHECKING from comfy_api.internal.async_to_sync import create_sync_class -from comfy_api.latest import io, ui, ComfyExtension #noqa: F401 +from comfy_api.latest import io, ui, IO, UI, ComfyExtension #noqa: F401 class ComfyAPIAdapter_v0_0_2(ComfyAPI_latest): @@ -42,4 +42,8 @@ __all__ = [ "InputImpl", "Types", "ComfyExtension", + "io", + "IO", + "ui", + "UI", ] diff --git a/comfy_api/version_list.py b/comfy_api/version_list.py index 7cb1871d5..be6e1db66 100644 --- a/comfy_api/version_list.py +++ b/comfy_api/version_list.py @@ -2,9 +2,8 @@ from comfy_api.latest import ComfyAPI_latest from comfy_api.v0_0_2 import ComfyAPIAdapter_v0_0_2 from comfy_api.v0_0_1 import ComfyAPIAdapter_v0_0_1 from comfy_api.internal import ComfyAPIBase -from typing import List, Type -supported_versions: List[Type[ComfyAPIBase]] = [ +supported_versions: list[type[ComfyAPIBase]] = [ ComfyAPI_latest, ComfyAPIAdapter_v0_0_2, ComfyAPIAdapter_v0_0_1, diff --git a/comfy_api_nodes/apis/bfl_api.py b/comfy_api_nodes/apis/bfl_api.py index 0fc8c0607..d8d3557b3 100644 --- a/comfy_api_nodes/apis/bfl_api.py +++ b/comfy_api_nodes/apis/bfl_api.py @@ -70,6 +70,29 @@ class BFLFluxProGenerateRequest(BaseModel): # ) +class Flux2ProGenerateRequest(BaseModel): + prompt: str = Field(...) + width: int = Field(1024, description="Must be a multiple of 32.") + height: int = Field(768, description="Must be a multiple of 32.") + seed: int | None = Field(None) + prompt_upsampling: bool | None = Field(None) + input_image: str | None = Field(None, description="Base64 encoded image for image-to-image generation") + input_image_2: str | None = Field(None, description="Base64 encoded image for image-to-image generation") + input_image_3: str | None = Field(None, description="Base64 encoded image for image-to-image generation") + input_image_4: str | None = Field(None, description="Base64 encoded image for image-to-image generation") + input_image_5: str | None = Field(None, description="Base64 encoded image for image-to-image generation") + input_image_6: str | None = Field(None, description="Base64 encoded image for image-to-image generation") + input_image_7: str | None = Field(None, description="Base64 encoded image for image-to-image generation") + input_image_8: str | None = Field(None, description="Base64 encoded image for image-to-image generation") + input_image_9: str | None = Field(None, description="Base64 encoded image for image-to-image generation") + safety_tolerance: int | None = Field( + 5, description="Tolerance level for input and output moderation. Value 0 being most strict.", ge=0, le=5 + ) + output_format: str | None = Field( + "png", description="Output format for the generated image. Can be 'jpeg' or 'png'." + ) + + class BFLFluxKontextProGenerateRequest(BaseModel): prompt: str = Field(..., description='The text prompt for what you wannt to edit.') input_image: Optional[str] = Field(None, description='Image to edit in base64 format') @@ -109,8 +132,9 @@ class BFLFluxProUltraGenerateRequest(BaseModel): class BFLFluxProGenerateResponse(BaseModel): - id: str = Field(..., description='The unique identifier for the generation task.') - polling_url: str = Field(..., description='URL to poll for the generation result.') + id: str = Field(..., description="The unique identifier for the generation task.") + polling_url: str = Field(..., description="URL to poll for the generation result.") + cost: float | None = Field(None, description="Price in cents") class BFLStatus(str, Enum): diff --git a/comfy_api_nodes/apis/bytedance_api.py b/comfy_api_nodes/apis/bytedance_api.py new file mode 100644 index 000000000..b8c2f618b --- /dev/null +++ b/comfy_api_nodes/apis/bytedance_api.py @@ -0,0 +1,144 @@ +from typing import Literal + +from pydantic import BaseModel, Field + + +class Text2ImageTaskCreationRequest(BaseModel): + model: str = Field(...) + prompt: str = Field(...) + response_format: str | None = Field("url") + size: str | None = Field(None) + seed: int | None = Field(0, ge=0, le=2147483647) + guidance_scale: float | None = Field(..., ge=1.0, le=10.0) + watermark: bool | None = Field(False) + + +class Image2ImageTaskCreationRequest(BaseModel): + model: str = Field(...) + prompt: str = Field(...) + response_format: str | None = Field("url") + image: str = Field(..., description="Base64 encoded string or image URL") + size: str | None = Field("adaptive") + seed: int | None = Field(..., ge=0, le=2147483647) + guidance_scale: float | None = Field(..., ge=1.0, le=10.0) + watermark: bool | None = Field(False) + + +class Seedream4Options(BaseModel): + max_images: int = Field(15) + + +class Seedream4TaskCreationRequest(BaseModel): + model: str = Field(...) + prompt: str = Field(...) + response_format: str = Field("url") + image: list[str] | None = Field(None, description="Image URLs") + size: str = Field(...) + seed: int = Field(..., ge=0, le=2147483647) + sequential_image_generation: str = Field("disabled") + sequential_image_generation_options: Seedream4Options = Field(Seedream4Options(max_images=15)) + watermark: bool = Field(False) + + +class ImageTaskCreationResponse(BaseModel): + model: str = Field(...) + created: int = Field(..., description="Unix timestamp (in seconds) indicating time when the request was created.") + data: list = Field([], description="Contains information about the generated image(s).") + error: dict = Field({}, description="Contains `code` and `message` fields in case of error.") + + +class TaskTextContent(BaseModel): + type: str = Field("text") + text: str = Field(...) + + +class TaskImageContentUrl(BaseModel): + url: str = Field(...) + + +class TaskImageContent(BaseModel): + type: str = Field("image_url") + image_url: TaskImageContentUrl = Field(...) + role: Literal["first_frame", "last_frame", "reference_image"] | None = Field(None) + + +class Text2VideoTaskCreationRequest(BaseModel): + model: str = Field(...) + content: list[TaskTextContent] = Field(..., min_length=1) + + +class Image2VideoTaskCreationRequest(BaseModel): + model: str = Field(...) + content: list[TaskTextContent | TaskImageContent] = Field(..., min_length=2) + + +class TaskCreationResponse(BaseModel): + id: str = Field(...) + + +class TaskStatusError(BaseModel): + code: str = Field(...) + message: str = Field(...) + + +class TaskStatusResult(BaseModel): + video_url: str = Field(...) + + +class TaskStatusResponse(BaseModel): + id: str = Field(...) + model: str = Field(...) + status: Literal["queued", "running", "cancelled", "succeeded", "failed"] = Field(...) + error: TaskStatusError | None = Field(None) + content: TaskStatusResult | None = Field(None) + + +RECOMMENDED_PRESETS = [ + ("1024x1024 (1:1)", 1024, 1024), + ("864x1152 (3:4)", 864, 1152), + ("1152x864 (4:3)", 1152, 864), + ("1280x720 (16:9)", 1280, 720), + ("720x1280 (9:16)", 720, 1280), + ("832x1248 (2:3)", 832, 1248), + ("1248x832 (3:2)", 1248, 832), + ("1512x648 (21:9)", 1512, 648), + ("2048x2048 (1:1)", 2048, 2048), + ("Custom", None, None), +] + +RECOMMENDED_PRESETS_SEEDREAM_4 = [ + ("2048x2048 (1:1)", 2048, 2048), + ("2304x1728 (4:3)", 2304, 1728), + ("1728x2304 (3:4)", 1728, 2304), + ("2560x1440 (16:9)", 2560, 1440), + ("1440x2560 (9:16)", 1440, 2560), + ("2496x1664 (3:2)", 2496, 1664), + ("1664x2496 (2:3)", 1664, 2496), + ("3024x1296 (21:9)", 3024, 1296), + ("4096x4096 (1:1)", 4096, 4096), + ("Custom", None, None), +] + +# The time in this dictionary are given for 10 seconds duration. +VIDEO_TASKS_EXECUTION_TIME = { + "seedance-1-0-lite-t2v-250428": { + "480p": 40, + "720p": 60, + "1080p": 90, + }, + "seedance-1-0-lite-i2v-250428": { + "480p": 40, + "720p": 60, + "1080p": 90, + }, + "seedance-1-0-pro-250528": { + "480p": 70, + "720p": 85, + "1080p": 115, + }, + "seedance-1-0-pro-fast-251015": { + "480p": 50, + "720p": 65, + "1080p": 100, + }, +} diff --git a/comfy_api_nodes/apis/gemini_api.py b/comfy_api_nodes/apis/gemini_api.py index 2bf28bf93..d81337dae 100644 --- a/comfy_api_nodes/apis/gemini_api.py +++ b/comfy_api_nodes/apis/gemini_api.py @@ -1,22 +1,229 @@ -from typing import Optional +from datetime import date +from enum import Enum +from typing import Any -from comfy_api_nodes.apis import GeminiGenerationConfig, GeminiContent, GeminiSafetySetting, GeminiSystemInstructionContent, GeminiTool, GeminiVideoMetadata -from pydantic import BaseModel +from pydantic import BaseModel, Field + + +class GeminiSafetyCategory(str, Enum): + HARM_CATEGORY_SEXUALLY_EXPLICIT = "HARM_CATEGORY_SEXUALLY_EXPLICIT" + HARM_CATEGORY_HATE_SPEECH = "HARM_CATEGORY_HATE_SPEECH" + HARM_CATEGORY_HARASSMENT = "HARM_CATEGORY_HARASSMENT" + HARM_CATEGORY_DANGEROUS_CONTENT = "HARM_CATEGORY_DANGEROUS_CONTENT" + + +class GeminiSafetyThreshold(str, Enum): + OFF = "OFF" + BLOCK_NONE = "BLOCK_NONE" + BLOCK_LOW_AND_ABOVE = "BLOCK_LOW_AND_ABOVE" + BLOCK_MEDIUM_AND_ABOVE = "BLOCK_MEDIUM_AND_ABOVE" + BLOCK_ONLY_HIGH = "BLOCK_ONLY_HIGH" + + +class GeminiSafetySetting(BaseModel): + category: GeminiSafetyCategory + threshold: GeminiSafetyThreshold + + +class GeminiRole(str, Enum): + user = "user" + model = "model" + + +class GeminiMimeType(str, Enum): + application_pdf = "application/pdf" + audio_mpeg = "audio/mpeg" + audio_mp3 = "audio/mp3" + audio_wav = "audio/wav" + image_png = "image/png" + image_jpeg = "image/jpeg" + image_webp = "image/webp" + text_plain = "text/plain" + video_mov = "video/mov" + video_mpeg = "video/mpeg" + video_mp4 = "video/mp4" + video_mpg = "video/mpg" + video_avi = "video/avi" + video_wmv = "video/wmv" + video_mpegps = "video/mpegps" + video_flv = "video/flv" + + +class GeminiInlineData(BaseModel): + data: str | None = Field( + None, + description="The base64 encoding of the image, PDF, or video to include inline in the prompt. " + "When including media inline, you must also specify the media type (mimeType) of the data. Size limit: 20MB", + ) + mimeType: GeminiMimeType | None = Field(None) + + +class GeminiFileData(BaseModel): + fileUri: str | None = Field(None) + mimeType: GeminiMimeType | None = Field(None) + + +class GeminiPart(BaseModel): + inlineData: GeminiInlineData | None = Field(None) + fileData: GeminiFileData | None = Field(None) + text: str | None = Field(None) + + +class GeminiTextPart(BaseModel): + text: str | None = Field(None) + + +class GeminiContent(BaseModel): + parts: list[GeminiPart] = Field([]) + role: GeminiRole = Field(..., examples=["user"]) + + +class GeminiSystemInstructionContent(BaseModel): + parts: list[GeminiTextPart] = Field( + ..., + description="A list of ordered parts that make up a single message. " + "Different parts may have different IANA MIME types.", + ) + role: GeminiRole | None = Field(..., description="The role field of systemInstruction may be ignored.") + + +class GeminiFunctionDeclaration(BaseModel): + description: str | None = Field(None) + name: str = Field(...) + parameters: dict[str, Any] = Field(..., description="JSON schema for the function parameters") + + +class GeminiTool(BaseModel): + functionDeclarations: list[GeminiFunctionDeclaration] | None = Field(None) + + +class GeminiOffset(BaseModel): + nanos: int | None = Field(None, ge=0, le=999999999) + seconds: int | None = Field(None, ge=-315576000000, le=315576000000) + + +class GeminiVideoMetadata(BaseModel): + endOffset: GeminiOffset | None = Field(None) + startOffset: GeminiOffset | None = Field(None) + + +class GeminiGenerationConfig(BaseModel): + maxOutputTokens: int | None = Field(None, ge=16, le=8192) + seed: int | None = Field(None) + stopSequences: list[str] | None = Field(None) + temperature: float | None = Field(None, ge=0.0, le=2.0) + topK: int | None = Field(None, ge=1) + topP: float | None = Field(None, ge=0.0, le=1.0) class GeminiImageConfig(BaseModel): - aspectRatio: Optional[str] = None + aspectRatio: str | None = Field(None) + imageSize: str | None = Field(None) class GeminiImageGenerationConfig(GeminiGenerationConfig): - responseModalities: Optional[list[str]] = None - imageConfig: Optional[GeminiImageConfig] = None + responseModalities: list[str] | None = Field(None) + imageConfig: GeminiImageConfig | None = Field(None) class GeminiImageGenerateContentRequest(BaseModel): - contents: list[GeminiContent] - generationConfig: Optional[GeminiImageGenerationConfig] = None - safetySettings: Optional[list[GeminiSafetySetting]] = None - systemInstruction: Optional[GeminiSystemInstructionContent] = None - tools: Optional[list[GeminiTool]] = None - videoMetadata: Optional[GeminiVideoMetadata] = None + contents: list[GeminiContent] = Field(...) + generationConfig: GeminiImageGenerationConfig | None = Field(None) + safetySettings: list[GeminiSafetySetting] | None = Field(None) + systemInstruction: GeminiSystemInstructionContent | None = Field(None) + tools: list[GeminiTool] | None = Field(None) + videoMetadata: GeminiVideoMetadata | None = Field(None) + uploadImagesToStorage: bool = Field(True) + + +class GeminiGenerateContentRequest(BaseModel): + contents: list[GeminiContent] = Field(...) + generationConfig: GeminiGenerationConfig | None = Field(None) + safetySettings: list[GeminiSafetySetting] | None = Field(None) + systemInstruction: GeminiSystemInstructionContent | None = Field(None) + tools: list[GeminiTool] | None = Field(None) + videoMetadata: GeminiVideoMetadata | None = Field(None) + + +class Modality(str, Enum): + MODALITY_UNSPECIFIED = "MODALITY_UNSPECIFIED" + TEXT = "TEXT" + IMAGE = "IMAGE" + VIDEO = "VIDEO" + AUDIO = "AUDIO" + DOCUMENT = "DOCUMENT" + + +class ModalityTokenCount(BaseModel): + modality: Modality | None = None + tokenCount: int | None = Field(None, description="Number of tokens for the given modality.") + + +class Probability(str, Enum): + NEGLIGIBLE = "NEGLIGIBLE" + LOW = "LOW" + MEDIUM = "MEDIUM" + HIGH = "HIGH" + UNKNOWN = "UNKNOWN" + + +class GeminiSafetyRating(BaseModel): + category: GeminiSafetyCategory | None = None + probability: Probability | None = Field( + None, + description="The probability that the content violates the specified safety category", + ) + + +class GeminiCitation(BaseModel): + authors: list[str] | None = None + endIndex: int | None = None + license: str | None = None + publicationDate: date | None = None + startIndex: int | None = None + title: str | None = None + uri: str | None = None + + +class GeminiCitationMetadata(BaseModel): + citations: list[GeminiCitation] | None = None + + +class GeminiCandidate(BaseModel): + citationMetadata: GeminiCitationMetadata | None = None + content: GeminiContent | None = None + finishReason: str | None = None + safetyRatings: list[GeminiSafetyRating] | None = None + + +class GeminiPromptFeedback(BaseModel): + blockReason: str | None = None + blockReasonMessage: str | None = None + safetyRatings: list[GeminiSafetyRating] | None = None + + +class GeminiUsageMetadata(BaseModel): + cachedContentTokenCount: int | None = Field( + None, + description="Output only. Number of tokens in the cached part in the input (the cached content).", + ) + candidatesTokenCount: int | None = Field(None, description="Number of tokens in the response(s).") + candidatesTokensDetails: list[ModalityTokenCount] | None = Field( + None, description="Breakdown of candidate tokens by modality." + ) + promptTokenCount: int | None = Field( + None, + description="Number of tokens in the request. When cachedContent is set, this is still the total effective prompt size meaning this includes the number of tokens in the cached content.", + ) + promptTokensDetails: list[ModalityTokenCount] | None = Field( + None, description="Breakdown of prompt tokens by modality." + ) + thoughtsTokenCount: int | None = Field(None, description="Number of tokens present in thoughts output.") + toolUsePromptTokenCount: int | None = Field(None, description="Number of tokens present in tool-use prompt(s).") + + +class GeminiGenerateContentResponse(BaseModel): + candidates: list[GeminiCandidate] | None = Field(None) + promptFeedback: GeminiPromptFeedback | None = Field(None) + usageMetadata: GeminiUsageMetadata | None = Field(None) + modelVersion: str | None = Field(None) diff --git a/comfy_api_nodes/apis/kling_api.py b/comfy_api_nodes/apis/kling_api.py new file mode 100644 index 000000000..bf54ede3e --- /dev/null +++ b/comfy_api_nodes/apis/kling_api.py @@ -0,0 +1,113 @@ +from pydantic import BaseModel, Field + + +class OmniProText2VideoRequest(BaseModel): + model_name: str = Field(..., description="kling-video-o1") + aspect_ratio: str = Field(..., description="'16:9', '9:16' or '1:1'") + duration: str = Field(..., description="'5' or '10'") + prompt: str = Field(...) + mode: str = Field("pro") + + +class OmniParamImage(BaseModel): + image_url: str = Field(...) + type: str | None = Field(None, description="Can be 'first_frame' or 'end_frame'") + + +class OmniParamVideo(BaseModel): + video_url: str = Field(...) + refer_type: str | None = Field(..., description="Can be 'base' or 'feature'") + keep_original_sound: str = Field(..., description="'yes' or 'no'") + + +class OmniProFirstLastFrameRequest(BaseModel): + model_name: str = Field(..., description="kling-video-o1") + image_list: list[OmniParamImage] = Field(..., min_length=1, max_length=7) + duration: str = Field(..., description="'5' or '10'") + prompt: str = Field(...) + mode: str = Field("pro") + + +class OmniProReferences2VideoRequest(BaseModel): + model_name: str = Field(..., description="kling-video-o1") + aspect_ratio: str | None = Field(..., description="'16:9', '9:16' or '1:1'") + image_list: list[OmniParamImage] | None = Field( + None, max_length=7, description="Max length 4 when video is present." + ) + video_list: list[OmniParamVideo] | None = Field(None, max_length=1) + duration: str | None = Field(..., description="From 3 to 10.") + prompt: str = Field(...) + mode: str = Field("pro") + + +class TaskStatusVideoResult(BaseModel): + duration: str | None = Field(None, description="Total video duration") + id: str | None = Field(None, description="Generated video ID") + url: str | None = Field(None, description="URL for generated video") + + +class TaskStatusImageResult(BaseModel): + index: int = Field(..., description="Image Number，0-9") + url: str = Field(..., description="URL for generated image") + + +class TaskStatusResults(BaseModel): + videos: list[TaskStatusVideoResult] | None = Field(None) + images: list[TaskStatusImageResult] | None = Field(None) + + +class TaskStatusResponseData(BaseModel): + created_at: int | None = Field(None, description="Task creation time") + updated_at: int | None = Field(None, description="Task update time") + task_status: str | None = None + task_status_msg: str | None = Field(None, description="Additional failure reason. Only for polling endpoint.") + task_id: str | None = Field(None, description="Task ID") + task_result: TaskStatusResults | None = Field(None) + + +class TaskStatusResponse(BaseModel): + code: int | None = Field(None, description="Error code") + message: str | None = Field(None, description="Error message") + request_id: str | None = Field(None, description="Request ID") + data: TaskStatusResponseData | None = Field(None) + + +class OmniImageParamImage(BaseModel): + image: str = Field(...) + + +class OmniProImageRequest(BaseModel): + model_name: str = Field(..., description="kling-image-o1") + resolution: str = Field(..., description="'1k' or '2k'") + aspect_ratio: str | None = Field(...) + prompt: str = Field(...) + mode: str = Field("pro") + n: int | None = Field(1, le=9) + image_list: list[OmniImageParamImage] | None = Field(..., max_length=10) + + +class TextToVideoWithAudioRequest(BaseModel): + model_name: str = Field(..., description="kling-v2-6") + aspect_ratio: str = Field(..., description="'16:9', '9:16' or '1:1'") + duration: str = Field(..., description="'5' or '10'") + prompt: str = Field(...) + mode: str = Field("pro") + sound: str = Field(..., description="'on' or 'off'") + + +class ImageToVideoWithAudioRequest(BaseModel): + model_name: str = Field(..., description="kling-v2-6") + image: str = Field(...) + duration: str = Field(..., description="'5' or '10'") + prompt: str = Field(...) + mode: str = Field("pro") + sound: str = Field(..., description="'on' or 'off'") + + +class MotionControlRequest(BaseModel): + prompt: str = Field(...) + image_url: str = Field(...) + video_url: str = Field(...) + keep_original_sound: str = Field(...) + character_orientation: str = Field(...) + mode: str = Field(..., description="'pro' or 'std'") diff --git a/comfy_api_nodes/apis/openai_api.py b/comfy_api_nodes/apis/openai_api.py new file mode 100644 index 000000000..ae5bb2673 --- /dev/null +++ b/comfy_api_nodes/apis/openai_api.py @@ -0,0 +1,52 @@ +from pydantic import BaseModel, Field + + +class Datum2(BaseModel): + b64_json: str | None = Field(None, description="Base64 encoded image data") + revised_prompt: str | None = Field(None, description="Revised prompt") + url: str | None = Field(None, description="URL of the image") + + +class InputTokensDetails(BaseModel): + image_tokens: int | None = None + text_tokens: int | None = None + + +class Usage(BaseModel): + input_tokens: int | None = None + input_tokens_details: InputTokensDetails | None = None + output_tokens: int | None = None + total_tokens: int | None = None + + +class OpenAIImageGenerationResponse(BaseModel): + data: list[Datum2] | None = None + usage: Usage | None = None + + +class OpenAIImageEditRequest(BaseModel): + background: str | None = Field(None, description="Background transparency") + model: str = Field(...) + moderation: str | None = Field(None) + n: int | None = Field(None, description="The number of images to generate") + output_compression: int | None = Field(None, description="Compression level for JPEG or WebP (0-100)") + output_format: str | None = Field(None) + prompt: str = Field(...) + quality: str | None = Field(None, description="Size of the image (e.g., 1024x1024, 1536x1024, auto)") + size: str | None = Field(None, description="Size of the output image") + + +class OpenAIImageGenerationRequest(BaseModel): + background: str | None = Field(None, description="Background transparency") + model: str | None = Field(None) + moderation: str | None = Field(None) + n: int | None = Field( + None, + description="The number of images to generate.", + ) + output_compression: int | None = Field(None, description="Compression level for JPEG or WebP (0-100)") + output_format: str | None = Field(None) + prompt: str = Field(...) + quality: str | None = Field(None, description="The quality of the generated image") + size: str | None = Field(None, description="Size of the image (e.g., 1024x1024, 1536x1024, auto)") + style: str | None = Field(None, description="Style of the image (only for dall-e-3)") diff --git a/comfy_api_nodes/apis/pika_api.py b/comfy_api_nodes/apis/pika_api.py deleted file mode 100644 index 232558cd7..000000000 --- a/comfy_api_nodes/apis/pika_api.py +++ /dev/null @@ -1,100 +0,0 @@ -from typing import Optional -from enum import Enum -from pydantic import BaseModel, Field - - -class Pikaffect(str, Enum): - Cake_ify = "Cake-ify" - Crumble = "Crumble" - Crush = "Crush" - Decapitate = "Decapitate" - Deflate = "Deflate" - Dissolve = "Dissolve" - Explode = "Explode" - Eye_pop = "Eye-pop" - Inflate = "Inflate" - Levitate = "Levitate" - Melt = "Melt" - Peel = "Peel" - Poke = "Poke" - Squish = "Squish" - Ta_da = "Ta-da" - Tear = "Tear" - - -class PikaBodyGenerate22C2vGenerate22PikascenesPost(BaseModel): - aspectRatio: Optional[float] = Field(None, description='Aspect ratio (width / height)') - duration: Optional[int] = Field(5) - ingredientsMode: str = Field(...) - negativePrompt: Optional[str] = Field(None) - promptText: Optional[str] = Field(None) - resolution: Optional[str] = Field('1080p') - seed: Optional[int] = Field(None) - - -class PikaGenerateResponse(BaseModel): - video_id: str = Field(...) - - -class PikaBodyGenerate22I2vGenerate22I2vPost(BaseModel): - duration: Optional[int] = 5 - negativePrompt: Optional[str] = Field(None) - promptText: Optional[str] = Field(None) - resolution: Optional[str] = '1080p' - seed: Optional[int] = Field(None) - - -class PikaBodyGenerate22KeyframeGenerate22PikaframesPost(BaseModel): - duration: Optional[int] = Field(None, ge=5, le=10) - negativePrompt: Optional[str] = Field(None) - promptText: str = Field(...) - resolution: Optional[str] = '1080p' - seed: Optional[int] = Field(None) - - -class PikaBodyGenerate22T2vGenerate22T2vPost(BaseModel): - aspectRatio: Optional[float] = Field( - 1.7777777777777777, - description='Aspect ratio (width / height)', - ge=0.4, - le=2.5, - ) - duration: Optional[int] = 5 - negativePrompt: Optional[str] = Field(None) - promptText: str = Field(...) - resolution: Optional[str] = '1080p' - seed: Optional[int] = Field(None) - - -class PikaBodyGeneratePikadditionsGeneratePikadditionsPost(BaseModel): - negativePrompt: Optional[str] = Field(None) - promptText: Optional[str] = Field(None) - seed: Optional[int] = Field(None) - - -class PikaBodyGeneratePikaffectsGeneratePikaffectsPost(BaseModel): - negativePrompt: Optional[str] = Field(None) - pikaffect: Optional[str] = None - promptText: Optional[str] = Field(None) - seed: Optional[int] = Field(None) - - -class PikaBodyGeneratePikaswapsGeneratePikaswapsPost(BaseModel): - negativePrompt: Optional[str] = Field(None) - promptText: Optional[str] = Field(None) - seed: Optional[int] = Field(None) - modifyRegionRoi: Optional[str] = Field(None) - - -class PikaStatusEnum(str, Enum): - queued = "queued" - started = "started" - finished = "finished" - failed = "failed" - - -class PikaVideoResponse(BaseModel): - id: str = Field(...) - progress: Optional[int] = Field(None) - status: PikaStatusEnum - url: Optional[str] = Field(None) diff --git a/comfy_api_nodes/apis/topaz_api.py b/comfy_api_nodes/apis/topaz_api.py new file mode 100644 index 000000000..4d9e62e72 --- /dev/null +++ b/comfy_api_nodes/apis/topaz_api.py @@ -0,0 +1,133 @@ +from typing import Optional, Union + +from pydantic import BaseModel, Field + + +class ImageEnhanceRequest(BaseModel): + model: str = Field("Reimagine") + output_format: str = Field("jpeg") + subject_detection: str = Field("All") + face_enhancement: bool = Field(True) + face_enhancement_creativity: float = Field(0, description="Is ignored if face_enhancement is false") + face_enhancement_strength: float = Field(0.8, description="Is ignored if face_enhancement is false") + source_url: str = Field(...) + output_width: Optional[int] = Field(None) + output_height: Optional[int] = Field(None) + crop_to_fill: bool = Field(False) + prompt: Optional[str] = Field(None, description="Text prompt for creative upscaling guidance") + creativity: int = Field(3, description="Creativity settings range from 1 to 9") + face_preservation: str = Field("true", description="To preserve the identity of characters") + color_preservation: str = Field("true", description="To preserve the original color") + + +class ImageAsyncTaskResponse(BaseModel): + process_id: str = Field(...) + + +class ImageStatusResponse(BaseModel): + process_id: str = Field(...) + status: str = Field(...) + progress: Optional[int] = Field(None) + credits: int = Field(...) + + +class ImageDownloadResponse(BaseModel): + download_url: str = Field(...) + expiry: int = Field(...) + + +class Resolution(BaseModel): + width: int = Field(...) + height: int = Field(...) + + +class CreateCreateVideoRequestSource(BaseModel): + container: str = Field(...) + size: int = Field(..., description="Size of the video file in bytes") + duration: int = Field(..., description="Duration of the video file in seconds") + frameCount: int = Field(..., description="Total number of frames in the video") + frameRate: int = Field(...) + resolution: Resolution = Field(...) + + +class VideoFrameInterpolationFilter(BaseModel): + model: str = Field(...) + slowmo: Optional[int] = Field(None) + fps: int = Field(...) + duplicate: bool = Field(...) + duplicate_threshold: float = Field(...) + + +class VideoEnhancementFilter(BaseModel): + model: str = Field(...) + auto: Optional[str] = Field(None, description="Auto, Manual, Relative") + focusFixLevel: Optional[str] = Field(None, description="Downscales video input for correction of blurred subjects") + compression: Optional[float] = Field(None, description="Strength of compression recovery") + details: Optional[float] = Field(None, description="Amount of detail reconstruction") + prenoise: Optional[float] = Field(None, description="Amount of noise to add to input to reduce over-smoothing") + noise: Optional[float] = Field(None, description="Amount of noise reduction") + halo: Optional[float] = Field(None, description="Amount of halo reduction") + preblur: Optional[float] = Field(None, description="Anti-aliasing and deblurring strength") + blur: Optional[float] = Field(None, description="Amount of sharpness applied") + 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") + isOptimizedMode: Optional[bool] = Field(None, description="Set to true for Starlight Creative (slc-1) only") + + +class OutputInformationVideo(BaseModel): + resolution: Resolution = Field(...) + frameRate: int = Field(...) + audioCodec: Optional[str] = Field(..., description="Required if audioTransfer is Copy or Convert") + audioTransfer: str = Field(..., description="Copy, Convert, None") + dynamicCompressionLevel: str = Field(..., description="Low, Mid, High") + + +class Overrides(BaseModel): + isPaidDiffusion: bool = Field(True) + + +class CreateVideoRequest(BaseModel): + source: CreateCreateVideoRequestSource = Field(...) + filters: list[Union[VideoFrameInterpolationFilter, VideoEnhancementFilter]] = Field(...) + output: OutputInformationVideo = Field(...) + overrides: Overrides = Field(Overrides(isPaidDiffusion=True)) + + +class CreateVideoResponse(BaseModel): + requestId: str = Field(...) + + +class VideoAcceptResponse(BaseModel): + uploadId: str = Field(...) + urls: list[str] = Field(...) + + +class VideoCompleteUploadRequestPart(BaseModel): + partNum: int = Field(...) + eTag: str = Field(...) + + +class VideoCompleteUploadRequest(BaseModel): + uploadResults: list[VideoCompleteUploadRequestPart] = Field(...) + + +class VideoCompleteUploadResponse(BaseModel): + message: str = Field(..., description="Confirmation message") + + +class VideoStatusResponseEstimates(BaseModel): + cost: list[int] = Field(...) + + +class VideoStatusResponseDownloadUrl(BaseModel): + url: str = Field(...) + + +class VideoStatusResponse(BaseModel): + status: str = Field(...) + estimates: Optional[VideoStatusResponseEstimates] = Field(None) + progress: Optional[float] = Field(None) + message: Optional[str] = Field("") + download: Optional[VideoStatusResponseDownloadUrl] = Field(None) diff --git a/comfy_api_nodes/apis/tripo_api.py b/comfy_api_nodes/apis/tripo_api.py index 713260e2a..ffaaa7dc1 100644 --- a/comfy_api_nodes/apis/tripo_api.py +++ b/comfy_api_nodes/apis/tripo_api.py @@ -5,11 +5,17 @@ from typing import Optional, List, Dict, Any, Union from pydantic import BaseModel, Field, RootModel class TripoModelVersion(str, Enum): + v3_0_20250812 = 'v3.0-20250812' v2_5_20250123 = 'v2.5-20250123' v2_0_20240919 = 'v2.0-20240919' v1_4_20240625 = 'v1.4-20240625' +class TripoGeometryQuality(str, Enum): + standard = 'standard' + detailed = 'detailed' + + class TripoTextureQuality(str, Enum): standard = 'standard' detailed = 'detailed' @@ -61,14 +67,20 @@ class TripoSpec(str, Enum): class TripoAnimation(str, Enum): IDLE = "preset:idle" WALK = "preset:walk" + RUN = "preset:run" + DIVE = "preset:dive" CLIMB = "preset:climb" JUMP = "preset:jump" - RUN = "preset:run" SLASH = "preset:slash" SHOOT = "preset:shoot" HURT = "preset:hurt" FALL = "preset:fall" TURN = "preset:turn" + QUADRUPED_WALK = "preset:quadruped:walk" + HEXAPOD_WALK = "preset:hexapod:walk" + OCTOPOD_WALK = "preset:octopod:walk" + SERPENTINE_MARCH = "preset:serpentine:march" + AQUATIC_MARCH = "preset:aquatic:march" class TripoStylizeStyle(str, Enum): LEGO = "lego" @@ -105,6 +117,11 @@ class TripoTaskStatus(str, Enum): BANNED = "banned" EXPIRED = "expired" +class TripoFbxPreset(str, Enum): + BLENDER = "blender" + MIXAMO = "mixamo" + _3DSMAX = "3dsmax" + class TripoFileTokenReference(BaseModel): type: Optional[str] = Field(None, description='The type of the reference') file_token: str @@ -142,6 +159,7 @@ class TripoTextToModelRequest(BaseModel): model_seed: Optional[int] = Field(None, description='The seed for the model') texture_seed: Optional[int] = Field(None, description='The seed for the texture') texture_quality: Optional[TripoTextureQuality] = TripoTextureQuality.standard + geometry_quality: Optional[TripoGeometryQuality] = TripoGeometryQuality.standard style: Optional[TripoStyle] = None auto_size: Optional[bool] = Field(False, description='Whether to auto-size the model') quad: Optional[bool] = Field(False, description='Whether to apply quad to the generated model') @@ -156,6 +174,7 @@ class TripoImageToModelRequest(BaseModel): model_seed: Optional[int] = Field(None, description='The seed for the model') texture_seed: Optional[int] = Field(None, description='The seed for the texture') texture_quality: Optional[TripoTextureQuality] = TripoTextureQuality.standard + geometry_quality: Optional[TripoGeometryQuality] = TripoGeometryQuality.standard texture_alignment: Optional[TripoTextureAlignment] = Field(TripoTextureAlignment.ORIGINAL_IMAGE, description='The texture alignment method') style: Optional[TripoStyle] = Field(None, description='The style to apply to the generated model') auto_size: Optional[bool] = Field(False, description='Whether to auto-size the model') @@ -173,6 +192,7 @@ class TripoMultiviewToModelRequest(BaseModel): model_seed: Optional[int] = Field(None, description='The seed for the model') texture_seed: Optional[int] = Field(None, description='The seed for the texture') texture_quality: Optional[TripoTextureQuality] = TripoTextureQuality.standard + geometry_quality: Optional[TripoGeometryQuality] = TripoGeometryQuality.standard texture_alignment: Optional[TripoTextureAlignment] = TripoTextureAlignment.ORIGINAL_IMAGE auto_size: Optional[bool] = Field(False, description='Whether to auto-size the model') orientation: Optional[TripoOrientation] = Field(TripoOrientation.DEFAULT, description='The orientation for the model') @@ -219,14 +239,24 @@ class TripoConvertModelRequest(BaseModel): type: TripoTaskType = Field(TripoTaskType.CONVERT_MODEL, description='Type of task') format: TripoConvertFormat = Field(..., description='The format to convert to') original_model_task_id: str = Field(..., description='The task ID of the original model') - quad: Optional[bool] = Field(False, description='Whether to apply quad to the model') - force_symmetry: Optional[bool] = Field(False, description='Whether to force symmetry') - face_limit: Optional[int] = Field(10000, description='The number of faces to limit the conversion to') - flatten_bottom: Optional[bool] = Field(False, description='Whether to flatten the bottom of the model') - flatten_bottom_threshold: Optional[float] = Field(0.01, description='The threshold for flattening the bottom') - texture_size: Optional[int] = Field(4096, description='The size of the texture') + quad: Optional[bool] = Field(None, description='Whether to apply quad to the model') + force_symmetry: Optional[bool] = Field(None, description='Whether to force symmetry') + face_limit: Optional[int] = Field(None, description='The number of faces to limit the conversion to') + flatten_bottom: Optional[bool] = Field(None, description='Whether to flatten the bottom of the model') + flatten_bottom_threshold: Optional[float] = Field(None, description='The threshold for flattening the bottom') + texture_size: Optional[int] = Field(None, description='The size of the texture') texture_format: Optional[TripoTextureFormat] = Field(TripoTextureFormat.JPEG, description='The format of the texture') - pivot_to_center_bottom: Optional[bool] = Field(False, description='Whether to pivot to the center bottom') + pivot_to_center_bottom: Optional[bool] = Field(None, description='Whether to pivot to the center bottom') + scale_factor: Optional[float] = Field(None, description='The scale factor for the model') + with_animation: Optional[bool] = Field(None, description='Whether to include animations') + pack_uv: Optional[bool] = Field(None, description='Whether to pack the UVs') + bake: Optional[bool] = Field(None, description='Whether to bake the model') + part_names: Optional[List[str]] = Field(None, description='The names of the parts to include') + fbx_preset: Optional[TripoFbxPreset] = Field(None, description='The preset for the FBX export') + export_vertex_colors: Optional[bool] = Field(None, description='Whether to export the vertex colors') + export_orientation: Optional[TripoOrientation] = Field(None, description='The orientation for the export') + animate_in_place: Optional[bool] = Field(None, description='Whether to animate in place') + class TripoTaskRequest(RootModel): root: Union[ diff --git a/comfy_api_nodes/apis/veo_api.py b/comfy_api_nodes/apis/veo_api.py index a55137afb..23ca725b7 100644 --- a/comfy_api_nodes/apis/veo_api.py +++ b/comfy_api_nodes/apis/veo_api.py @@ -1,34 +1,21 @@ -from typing import Optional, Union -from enum import Enum +from typing import Optional from pydantic import BaseModel, Field -class Image2(BaseModel): - bytesBase64Encoded: str - gcsUri: Optional[str] = None - mimeType: Optional[str] = None +class VeoRequestInstanceImage(BaseModel): + bytesBase64Encoded: str | None = Field(None) + gcsUri: str | None = Field(None) + mimeType: str | None = Field(None) -class Image3(BaseModel): - bytesBase64Encoded: Optional[str] = None - gcsUri: str - mimeType: Optional[str] = None - - -class Instance1(BaseModel): - image: Optional[Union[Image2, Image3]] = Field( - None, description='Optional image to guide video generation' - ) +class VeoRequestInstance(BaseModel): + image: VeoRequestInstanceImage | None = Field(None) + lastFrame: VeoRequestInstanceImage | None = Field(None) prompt: str = Field(..., description='Text description of the video') -class PersonGeneration1(str, Enum): - ALLOW = 'ALLOW' - BLOCK = 'BLOCK' - - -class Parameters1(BaseModel): +class VeoRequestParameters(BaseModel): aspectRatio: Optional[str] = Field(None, examples=['16:9']) durationSeconds: Optional[int] = None enhancePrompt: Optional[bool] = None @@ -37,17 +24,18 @@ class Parameters1(BaseModel): description='Generate audio for the video. Only supported by veo 3 models.', ) negativePrompt: Optional[str] = None - personGeneration: Optional[PersonGeneration1] = None + personGeneration: str | None = Field(None, description="ALLOW or BLOCK") sampleCount: Optional[int] = None seed: Optional[int] = None storageUri: Optional[str] = Field( None, description='Optional Cloud Storage URI to upload the video' ) + resolution: str | None = Field(None) class VeoGenVidRequest(BaseModel): - instances: Optional[list[Instance1]] = None - parameters: Optional[Parameters1] = None + instances: list[VeoRequestInstance] | None = Field(None) + parameters: VeoRequestParameters | None = Field(None) class VeoGenVidResponse(BaseModel): @@ -97,7 +85,7 @@ class Response1(BaseModel): raiMediaFilteredReasons: Optional[list[str]] = Field( None, description='Reasons why media was filtered by responsible AI policies' ) - videos: Optional[list[Video]] = None + videos: Optional[list[Video]] = Field(None) class VeoGenVidPollResponse(BaseModel): diff --git a/comfy_api_nodes/nodes_bfl.py b/comfy_api_nodes/nodes_bfl.py index 1740fb377..ce077d6b3 100644 --- a/comfy_api_nodes/nodes_bfl.py +++ b/comfy_api_nodes/nodes_bfl.py @@ -1,23 +1,22 @@ -from inspect import cleandoc -from typing import Optional - import torch +from pydantic import BaseModel 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.bfl_api import ( BFLFluxExpandImageRequest, BFLFluxFillImageRequest, BFLFluxKontextProGenerateRequest, - BFLFluxProGenerateRequest, BFLFluxProGenerateResponse, BFLFluxProUltraGenerateRequest, BFLFluxStatusResponse, BFLStatus, + Flux2ProGenerateRequest, ) from comfy_api_nodes.util import ( ApiEndpoint, download_url_to_image_tensor, + get_number_of_images, poll_op, resize_mask_to_image, sync_op, @@ -27,7 +26,7 @@ from comfy_api_nodes.util import ( ) -def convert_mask_to_image(mask: torch.Tensor): +def convert_mask_to_image(mask: Input.Image): """ Make mask have the expected amount of dims (4) and channels (3) to be recognized as an image. """ @@ -37,9 +36,6 @@ def convert_mask_to_image(mask: torch.Tensor): class FluxProUltraImageNode(IO.ComfyNode): - """ - Generates images using Flux Pro 1.1 Ultra via api based on prompt and resolution. - """ @classmethod def define_schema(cls) -> IO.Schema: @@ -47,7 +43,7 @@ class FluxProUltraImageNode(IO.ComfyNode): node_id="FluxProUltraImageNode", display_name="Flux 1.1 [pro] Ultra Image", category="api node/image/BFL", - description=cleandoc(cls.__doc__ or ""), + description="Generates images using Flux Pro 1.1 Ultra via api based on prompt and resolution.", inputs=[ IO.String.Input( "prompt", @@ -116,7 +112,7 @@ class FluxProUltraImageNode(IO.ComfyNode): prompt_upsampling: bool = False, raw: bool = False, seed: int = 0, - image_prompt: Optional[torch.Tensor] = None, + image_prompt: Input.Image | None = None, image_prompt_strength: float = 0.1, ) -> IO.NodeOutput: if image_prompt is None: @@ -154,9 +150,6 @@ class FluxProUltraImageNode(IO.ComfyNode): class FluxKontextProImageNode(IO.ComfyNode): - """ - Edits images using Flux.1 Kontext [pro] via api based on prompt and aspect ratio. - """ @classmethod def define_schema(cls) -> IO.Schema: @@ -164,7 +157,7 @@ class FluxKontextProImageNode(IO.ComfyNode): node_id=cls.NODE_ID, display_name=cls.DISPLAY_NAME, category="api node/image/BFL", - description=cleandoc(cls.__doc__ or ""), + description="Edits images using Flux.1 Kontext [pro] via api based on prompt and aspect ratio.", inputs=[ IO.String.Input( "prompt", @@ -230,7 +223,7 @@ class FluxKontextProImageNode(IO.ComfyNode): aspect_ratio: str, guidance: float, steps: int, - input_image: Optional[torch.Tensor] = None, + input_image: Input.Image | None = None, seed=0, prompt_upsampling=False, ) -> IO.NodeOutput: @@ -270,138 +263,14 @@ class FluxKontextProImageNode(IO.ComfyNode): class FluxKontextMaxImageNode(FluxKontextProImageNode): - """ - Edits images using Flux.1 Kontext [max] via api based on prompt and aspect ratio. - """ - DESCRIPTION = cleandoc(__doc__ or "") + DESCRIPTION = "Edits images using Flux.1 Kontext [max] via api based on prompt and aspect ratio." BFL_PATH = "/proxy/bfl/flux-kontext-max/generate" NODE_ID = "FluxKontextMaxImageNode" DISPLAY_NAME = "Flux.1 Kontext [max] Image" -class FluxProImageNode(IO.ComfyNode): - """ - Generates images synchronously based on prompt and resolution. - """ - - @classmethod - def define_schema(cls) -> IO.Schema: - return IO.Schema( - node_id="FluxProImageNode", - display_name="Flux 1.1 [pro] Image", - category="api node/image/BFL", - description=cleandoc(cls.__doc__ or ""), - inputs=[ - IO.String.Input( - "prompt", - multiline=True, - default="", - tooltip="Prompt for the image generation", - ), - IO.Boolean.Input( - "prompt_upsampling", - default=False, - tooltip="Whether to perform upsampling on the prompt. " - "If active, automatically modifies the prompt for more creative generation, " - "but results are nondeterministic (same seed will not produce exactly the same result).", - ), - IO.Int.Input( - "width", - default=1024, - min=256, - max=1440, - step=32, - ), - IO.Int.Input( - "height", - default=768, - min=256, - max=1440, - step=32, - ), - IO.Int.Input( - "seed", - default=0, - min=0, - max=0xFFFFFFFFFFFFFFFF, - control_after_generate=True, - tooltip="The random seed used for creating the noise.", - ), - IO.Image.Input( - "image_prompt", - optional=True, - ), - # "image_prompt_strength": ( - # IO.FLOAT, - # { - # "default": 0.1, - # "min": 0.0, - # "max": 1.0, - # "step": 0.01, - # "tooltip": "Blend between the prompt and the image prompt.", - # }, - # ), - ], - 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, - ) - - @classmethod - async def execute( - cls, - prompt: str, - prompt_upsampling, - width: int, - height: int, - seed=0, - image_prompt=None, - # image_prompt_strength=0.1, - ) -> IO.NodeOutput: - image_prompt = image_prompt if image_prompt is None else tensor_to_base64_string(image_prompt) - initial_response = await sync_op( - cls, - ApiEndpoint( - path="/proxy/bfl/flux-pro-1.1/generate", - method="POST", - ), - response_model=BFLFluxProGenerateResponse, - data=BFLFluxProGenerateRequest( - prompt=prompt, - prompt_upsampling=prompt_upsampling, - width=width, - height=height, - seed=seed, - image_prompt=image_prompt, - ), - ) - 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, - 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 FluxProExpandNode(IO.ComfyNode): - """ - Outpaints image based on prompt. - """ @classmethod def define_schema(cls) -> IO.Schema: @@ -409,7 +278,7 @@ class FluxProExpandNode(IO.ComfyNode): node_id="FluxProExpandNode", display_name="Flux.1 Expand Image", category="api node/image/BFL", - description=cleandoc(cls.__doc__ or ""), + description="Outpaints image based on prompt.", inputs=[ IO.Image.Input("image"), IO.String.Input( @@ -488,7 +357,7 @@ class FluxProExpandNode(IO.ComfyNode): @classmethod async def execute( cls, - image: torch.Tensor, + image: Input.Image, prompt: str, prompt_upsampling: bool, top: int, @@ -535,9 +404,6 @@ class FluxProExpandNode(IO.ComfyNode): class FluxProFillNode(IO.ComfyNode): - """ - Inpaints image based on mask and prompt. - """ @classmethod def define_schema(cls) -> IO.Schema: @@ -545,7 +411,7 @@ class FluxProFillNode(IO.ComfyNode): node_id="FluxProFillNode", display_name="Flux.1 Fill Image", category="api node/image/BFL", - description=cleandoc(cls.__doc__ or ""), + description="Inpaints image based on mask and prompt.", inputs=[ IO.Image.Input("image"), IO.Mask.Input("mask"), @@ -597,8 +463,8 @@ class FluxProFillNode(IO.ComfyNode): @classmethod async def execute( cls, - image: torch.Tensor, - mask: torch.Tensor, + image: Input.Image, + mask: Input.Image, prompt: str, prompt_upsampling: bool, steps: int, @@ -640,16 +506,136 @@ class FluxProFillNode(IO.ComfyNode): return IO.NodeOutput(await download_url_to_image_tensor(response.result["sample"])) +class Flux2ProImageNode(IO.ComfyNode): + + NODE_ID = "Flux2ProImageNode" + DISPLAY_NAME = "Flux.2 [pro] Image" + API_ENDPOINT = "/proxy/bfl/flux-2-pro/generate" + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id=cls.NODE_ID, + display_name=cls.DISPLAY_NAME, + category="api node/image/BFL", + description="Generates images synchronously based on prompt and resolution.", + inputs=[ + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Prompt for the image generation or edit", + ), + 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.Int.Input( + "seed", + default=0, + min=0, + max=0xFFFFFFFFFFFFFFFF, + control_after_generate=True, + tooltip="The random seed used for creating the noise.", + ), + IO.Boolean.Input( + "prompt_upsampling", + default=True, + tooltip="Whether to perform upsampling on the prompt. " + "If active, automatically modifies the prompt for more creative generation.", + ), + IO.Image.Input("images", optional=True, tooltip="Up to 9 images to be used as references."), + ], + 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, + ) + + @classmethod + async def execute( + cls, + prompt: str, + width: int, + height: int, + seed: int, + prompt_upsampling: bool, + images: Input.Image | None = None, + ) -> IO.NodeOutput: + reference_images = {} + if images is not None: + if get_number_of_images(images) > 9: + raise ValueError("The current maximum number of supported images is 9.") + for image_index in range(images.shape[0]): + key_name = f"input_image_{image_index + 1}" if image_index else "input_image" + reference_images[key_name] = tensor_to_base64_string(images[image_index], total_pixels=2048 * 2048) + initial_response = await sync_op( + cls, + ApiEndpoint(path=cls.API_ENDPOINT, method="POST"), + response_model=BFLFluxProGenerateResponse, + data=Flux2ProGenerateRequest( + prompt=prompt, + width=width, + height=height, + seed=seed, + prompt_upsampling=prompt_upsampling, + **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 Flux2MaxImageNode(Flux2ProImageNode): + + NODE_ID = "Flux2MaxImageNode" + DISPLAY_NAME = "Flux.2 [max] Image" + API_ENDPOINT = "/proxy/bfl/flux-2-max/generate" + + class BFLExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: return [ FluxProUltraImageNode, - # FluxProImageNode, FluxKontextProImageNode, FluxKontextMaxImageNode, FluxProExpandNode, FluxProFillNode, + Flux2ProImageNode, + Flux2MaxImageNode, ] diff --git a/comfy_api_nodes/nodes_bytedance.py b/comfy_api_nodes/nodes_bytedance.py index caced471e..d4a2cfae6 100644 --- a/comfy_api_nodes/nodes_bytedance.py +++ b/comfy_api_nodes/nodes_bytedance.py @@ -1,13 +1,27 @@ import logging import math -from enum import Enum -from typing import Literal, Optional, Union import torch -from pydantic import BaseModel, Field 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.bytedance_api import ( + RECOMMENDED_PRESETS, + RECOMMENDED_PRESETS_SEEDREAM_4, + VIDEO_TASKS_EXECUTION_TIME, + Image2ImageTaskCreationRequest, + Image2VideoTaskCreationRequest, + ImageTaskCreationResponse, + Seedream4Options, + Seedream4TaskCreationRequest, + TaskCreationResponse, + TaskImageContent, + TaskImageContentUrl, + TaskStatusResponse, + TaskTextContent, + Text2ImageTaskCreationRequest, + Text2VideoTaskCreationRequest, +) from comfy_api_nodes.util import ( ApiEndpoint, download_url_to_image_tensor, @@ -29,162 +43,6 @@ BYTEPLUS_TASK_ENDPOINT = "/proxy/byteplus/api/v3/contents/generations/tasks" BYTEPLUS_TASK_STATUS_ENDPOINT = "/proxy/byteplus/api/v3/contents/generations/tasks" # + /{task_id} -class Text2ImageModelName(str, Enum): - seedream_3 = "seedream-3-0-t2i-250415" - - -class Image2ImageModelName(str, Enum): - seededit_3 = "seededit-3-0-i2i-250628" - - -class Text2VideoModelName(str, Enum): - seedance_1_pro = "seedance-1-0-pro-250528" - seedance_1_lite = "seedance-1-0-lite-t2v-250428" - - -class Image2VideoModelName(str, Enum): - """note(August 31): Pro model only supports FirstFrame: https://docs.byteplus.com/en/docs/ModelArk/1520757""" - - seedance_1_pro = "seedance-1-0-pro-250528" - seedance_1_lite = "seedance-1-0-lite-i2v-250428" - - -class Text2ImageTaskCreationRequest(BaseModel): - model: Text2ImageModelName = Text2ImageModelName.seedream_3 - prompt: str = Field(...) - response_format: Optional[str] = Field("url") - size: Optional[str] = Field(None) - seed: Optional[int] = Field(0, ge=0, le=2147483647) - guidance_scale: Optional[float] = Field(..., ge=1.0, le=10.0) - watermark: Optional[bool] = Field(True) - - -class Image2ImageTaskCreationRequest(BaseModel): - model: Image2ImageModelName = Image2ImageModelName.seededit_3 - prompt: str = Field(...) - response_format: Optional[str] = Field("url") - image: str = Field(..., description="Base64 encoded string or image URL") - size: Optional[str] = Field("adaptive") - seed: Optional[int] = Field(..., ge=0, le=2147483647) - guidance_scale: Optional[float] = Field(..., ge=1.0, le=10.0) - watermark: Optional[bool] = Field(True) - - -class Seedream4Options(BaseModel): - max_images: int = Field(15) - - -class Seedream4TaskCreationRequest(BaseModel): - model: str = Field("seedream-4-0-250828") - prompt: str = Field(...) - response_format: str = Field("url") - image: Optional[list[str]] = Field(None, description="Image URLs") - size: str = Field(...) - seed: int = Field(..., ge=0, le=2147483647) - sequential_image_generation: str = Field("disabled") - sequential_image_generation_options: Seedream4Options = Field(Seedream4Options(max_images=15)) - watermark: bool = Field(True) - - -class ImageTaskCreationResponse(BaseModel): - model: str = Field(...) - created: int = Field(..., description="Unix timestamp (in seconds) indicating time when the request was created.") - data: list = Field([], description="Contains information about the generated image(s).") - error: dict = Field({}, description="Contains `code` and `message` fields in case of error.") - - -class TaskTextContent(BaseModel): - type: str = Field("text") - text: str = Field(...) - - -class TaskImageContentUrl(BaseModel): - url: str = Field(...) - - -class TaskImageContent(BaseModel): - type: str = Field("image_url") - image_url: TaskImageContentUrl = Field(...) - role: Optional[Literal["first_frame", "last_frame", "reference_image"]] = Field(None) - - -class Text2VideoTaskCreationRequest(BaseModel): - model: Text2VideoModelName = Text2VideoModelName.seedance_1_pro - content: list[TaskTextContent] = Field(..., min_length=1) - - -class Image2VideoTaskCreationRequest(BaseModel): - model: Image2VideoModelName = Image2VideoModelName.seedance_1_pro - content: list[Union[TaskTextContent, TaskImageContent]] = Field(..., min_length=2) - - -class TaskCreationResponse(BaseModel): - id: str = Field(...) - - -class TaskStatusError(BaseModel): - code: str = Field(...) - message: str = Field(...) - - -class TaskStatusResult(BaseModel): - video_url: str = Field(...) - - -class TaskStatusResponse(BaseModel): - id: str = Field(...) - model: str = Field(...) - status: Literal["queued", "running", "cancelled", "succeeded", "failed"] = Field(...) - error: Optional[TaskStatusError] = Field(None) - content: Optional[TaskStatusResult] = Field(None) - - -RECOMMENDED_PRESETS = [ - ("1024x1024 (1:1)", 1024, 1024), - ("864x1152 (3:4)", 864, 1152), - ("1152x864 (4:3)", 1152, 864), - ("1280x720 (16:9)", 1280, 720), - ("720x1280 (9:16)", 720, 1280), - ("832x1248 (2:3)", 832, 1248), - ("1248x832 (3:2)", 1248, 832), - ("1512x648 (21:9)", 1512, 648), - ("2048x2048 (1:1)", 2048, 2048), - ("Custom", None, None), -] - -RECOMMENDED_PRESETS_SEEDREAM_4 = [ - ("2048x2048 (1:1)", 2048, 2048), - ("2304x1728 (4:3)", 2304, 1728), - ("1728x2304 (3:4)", 1728, 2304), - ("2560x1440 (16:9)", 2560, 1440), - ("1440x2560 (9:16)", 1440, 2560), - ("2496x1664 (3:2)", 2496, 1664), - ("1664x2496 (2:3)", 1664, 2496), - ("3024x1296 (21:9)", 3024, 1296), - ("4096x4096 (1:1)", 4096, 4096), - ("Custom", None, None), -] - -# The time in this dictionary are given for 10 seconds duration. -VIDEO_TASKS_EXECUTION_TIME = { - "seedance-1-0-lite-t2v-250428": { - "480p": 40, - "720p": 60, - "1080p": 90, - }, - "seedance-1-0-lite-i2v-250428": { - "480p": 40, - "720p": 60, - "1080p": 90, - }, - "seedance-1-0-pro-250528": { - "480p": 70, - "720p": 85, - "1080p": 115, - }, -} - - def get_image_url_from_response(response: ImageTaskCreationResponse) -> str: if response.error: error_msg = f"ByteDance request failed. Code: {response.error['code']}, message: {response.error['message']}" @@ -194,13 +52,6 @@ def get_image_url_from_response(response: ImageTaskCreationResponse) -> str: return response.data[0]["url"] -def get_video_url_from_task_status(response: TaskStatusResponse) -> Union[str, None]: - """Returns the video URL from the task status response if it exists.""" - if hasattr(response, "content") and response.content: - return response.content.video_url - return None - - class ByteDanceImageNode(IO.ComfyNode): @classmethod @@ -211,12 +62,7 @@ class ByteDanceImageNode(IO.ComfyNode): category="api node/image/ByteDance", description="Generate images using ByteDance models via api based on prompt", inputs=[ - IO.Combo.Input( - "model", - options=Text2ImageModelName, - default=Text2ImageModelName.seedream_3, - tooltip="Model name", - ), + IO.Combo.Input("model", options=["seedream-3-0-t2i-250415"]), IO.String.Input( "prompt", multiline=True, @@ -266,7 +112,7 @@ class ByteDanceImageNode(IO.ComfyNode): ), IO.Boolean.Input( "watermark", - default=True, + default=False, tooltip='Whether to add an "AI generated" watermark to the image', optional=True, ), @@ -335,12 +181,7 @@ class ByteDanceImageEditNode(IO.ComfyNode): category="api node/image/ByteDance", description="Edit images using ByteDance models via api based on prompt", inputs=[ - IO.Combo.Input( - "model", - options=Image2ImageModelName, - default=Image2ImageModelName.seededit_3, - tooltip="Model name", - ), + IO.Combo.Input("model", options=["seededit-3-0-i2i-250628"]), IO.Image.Input( "image", tooltip="The base image to edit", @@ -374,7 +215,7 @@ class ByteDanceImageEditNode(IO.ComfyNode): ), IO.Boolean.Input( "watermark", - default=True, + default=False, tooltip='Whether to add an "AI generated" watermark to the image', optional=True, ), @@ -388,13 +229,14 @@ class ByteDanceImageEditNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + is_deprecated=True, ) @classmethod async def execute( cls, model: str, - image: torch.Tensor, + image: Input.Image, prompt: str, seed: int, guidance_scale: float, @@ -428,13 +270,13 @@ class ByteDanceSeedreamNode(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="ByteDanceSeedreamNode", - display_name="ByteDance Seedream 4", + display_name="ByteDance Seedream 4.5", category="api node/image/ByteDance", description="Unified text-to-image generation and precise single-sentence editing at up to 4K resolution.", inputs=[ IO.Combo.Input( "model", - options=["seedream-4-0-250828"], + options=["seedream-4-5-251128", "seedream-4-0-250828"], tooltip="Model name", ), IO.String.Input( @@ -459,7 +301,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode): default=2048, min=1024, max=4096, - step=64, + step=8, tooltip="Custom width for image. Value is working only if `size_preset` is set to `Custom`", optional=True, ), @@ -468,7 +310,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode): default=2048, min=1024, max=4096, - step=64, + step=8, tooltip="Custom height for image. Value is working only if `size_preset` is set to `Custom`", optional=True, ), @@ -505,7 +347,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode): ), IO.Boolean.Input( "watermark", - default=True, + default=False, tooltip='Whether to add an "AI generated" watermark to the image.', optional=True, ), @@ -532,14 +374,14 @@ class ByteDanceSeedreamNode(IO.ComfyNode): cls, model: str, prompt: str, - image: torch.Tensor = None, + image: Input.Image | None = None, size_preset: str = RECOMMENDED_PRESETS_SEEDREAM_4[0][0], width: int = 2048, height: int = 2048, sequential_image_generation: str = "disabled", max_images: int = 1, seed: int = 0, - watermark: bool = True, + watermark: bool = False, fail_on_partial: bool = True, ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=True, min_length=1) @@ -555,6 +397,18 @@ class ByteDanceSeedreamNode(IO.ComfyNode): raise ValueError( f"Custom size out of range: {w}x{h}. " "Both width and height must be between 1024 and 4096 pixels." ) + out_num_pixels = w * h + mp_provided = out_num_pixels / 1_000_000.0 + if "seedream-4-5" in model and out_num_pixels < 3686400: + raise ValueError( + f"Minimum image resolution that Seedream 4.5 can generate is 3.68MP, " + f"but {mp_provided:.2f}MP provided." + ) + if "seedream-4-0" in model 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." + ) n_input_images = get_number_of_images(image) if image is not None else 0 if n_input_images > 10: raise ValueError(f"Maximum of 10 reference images are supported, but {n_input_images} received.") @@ -607,9 +461,8 @@ class ByteDanceTextToVideoNode(IO.ComfyNode): inputs=[ IO.Combo.Input( "model", - options=Text2VideoModelName, - default=Text2VideoModelName.seedance_1_pro, - tooltip="Model name", + options=["seedance-1-0-pro-250528", "seedance-1-0-lite-t2v-250428", "seedance-1-0-pro-fast-251015"], + default="seedance-1-0-pro-fast-251015", ), IO.String.Input( "prompt", @@ -655,7 +508,7 @@ class ByteDanceTextToVideoNode(IO.ComfyNode): ), IO.Boolean.Input( "watermark", - default=True, + default=False, tooltip='Whether to add an "AI generated" watermark to the video.', optional=True, ), @@ -714,9 +567,8 @@ class ByteDanceImageToVideoNode(IO.ComfyNode): inputs=[ IO.Combo.Input( "model", - options=Image2VideoModelName, - default=Image2VideoModelName.seedance_1_pro, - tooltip="Model name", + options=["seedance-1-0-pro-250528", "seedance-1-0-lite-t2v-250428", "seedance-1-0-pro-fast-251015"], + default="seedance-1-0-pro-fast-251015", ), IO.String.Input( "prompt", @@ -766,7 +618,7 @@ class ByteDanceImageToVideoNode(IO.ComfyNode): ), IO.Boolean.Input( "watermark", - default=True, + default=False, tooltip='Whether to add an "AI generated" watermark to the video.', optional=True, ), @@ -787,7 +639,7 @@ class ByteDanceImageToVideoNode(IO.ComfyNode): cls, model: str, prompt: str, - image: torch.Tensor, + image: Input.Image, resolution: str, aspect_ratio: str, duration: int, @@ -833,9 +685,8 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode): inputs=[ IO.Combo.Input( "model", - options=[model.value for model in Image2VideoModelName], - default=Image2VideoModelName.seedance_1_lite.value, - tooltip="Model name", + options=["seedance-1-0-pro-250528", "seedance-1-0-lite-i2v-250428"], + default="seedance-1-0-lite-i2v-250428", ), IO.String.Input( "prompt", @@ -889,7 +740,7 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode): ), IO.Boolean.Input( "watermark", - default=True, + default=False, tooltip='Whether to add an "AI generated" watermark to the video.', optional=True, ), @@ -910,8 +761,8 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode): cls, model: str, prompt: str, - first_frame: torch.Tensor, - last_frame: torch.Tensor, + first_frame: Input.Image, + last_frame: Input.Image, resolution: str, aspect_ratio: str, duration: int, @@ -968,9 +819,8 @@ class ByteDanceImageReferenceNode(IO.ComfyNode): inputs=[ IO.Combo.Input( "model", - options=[Image2VideoModelName.seedance_1_lite.value], - default=Image2VideoModelName.seedance_1_lite.value, - tooltip="Model name", + options=["seedance-1-0-pro-250528", "seedance-1-0-lite-i2v-250428"], + default="seedance-1-0-lite-i2v-250428", ), IO.String.Input( "prompt", @@ -1013,7 +863,7 @@ class ByteDanceImageReferenceNode(IO.ComfyNode): ), IO.Boolean.Input( "watermark", - default=True, + default=False, tooltip='Whether to add an "AI generated" watermark to the video.', optional=True, ), @@ -1034,7 +884,7 @@ class ByteDanceImageReferenceNode(IO.ComfyNode): cls, model: str, prompt: str, - images: torch.Tensor, + images: Input.Image, resolution: str, aspect_ratio: str, duration: int, @@ -1069,8 +919,8 @@ class ByteDanceImageReferenceNode(IO.ComfyNode): async def process_video_task( cls: type[IO.ComfyNode], - payload: Union[Text2VideoTaskCreationRequest, Image2VideoTaskCreationRequest], - estimated_duration: Optional[int], + payload: Text2VideoTaskCreationRequest | Image2VideoTaskCreationRequest, + estimated_duration: int | None, ) -> IO.NodeOutput: initial_response = await sync_op( cls, @@ -1085,7 +935,7 @@ async def process_video_task( estimated_duration=estimated_duration, response_model=TaskStatusResponse, ) - return IO.NodeOutput(await download_url_to_video_output(get_video_url_from_task_status(response))) + return IO.NodeOutput(await download_url_to_video_output(response.content.video_url)) def raise_if_text_params(prompt: str, text_params: list[str]) -> None: diff --git a/comfy_api_nodes/nodes_gemini.py b/comfy_api_nodes/nodes_gemini.py index 67f2469ad..e8ed7e797 100644 --- a/comfy_api_nodes/nodes_gemini.py +++ b/comfy_api_nodes/nodes_gemini.py @@ -3,49 +3,56 @@ API Nodes for Gemini Multimodal LLM Usage via Remote API See: https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/inference """ -from __future__ import annotations - import base64 -import json import os -import time -import uuid from enum import Enum from io import BytesIO -from typing import Literal, Optional +from typing import Literal import torch from typing_extensions import override import folder_paths -from comfy_api.latest import IO, ComfyExtension, Input -from comfy_api.util import VideoCodec, VideoContainer -from comfy_api_nodes.apis import ( +from comfy_api.latest import IO, ComfyExtension, Input, Types +from comfy_api_nodes.apis.gemini_api import ( GeminiContent, + GeminiFileData, GeminiGenerateContentRequest, GeminiGenerateContentResponse, - GeminiInlineData, - GeminiMimeType, - GeminiPart, -) -from comfy_api_nodes.apis.gemini_api import ( GeminiImageConfig, GeminiImageGenerateContentRequest, GeminiImageGenerationConfig, + GeminiInlineData, + GeminiMimeType, + GeminiPart, + GeminiRole, + GeminiSystemInstructionContent, + GeminiTextPart, + Modality, ) from comfy_api_nodes.util import ( ApiEndpoint, audio_to_base64_string, bytesio_to_image_tensor, + download_url_to_image_tensor, + get_number_of_images, sync_op, tensor_to_base64_string, + upload_images_to_comfyapi, validate_string, video_to_base64_string, ) -from server import PromptServer GEMINI_BASE_ENDPOINT = "/proxy/vertexai/gemini" GEMINI_MAX_INPUT_FILE_SIZE = 20 * 1024 * 1024 # 20 MB +GEMINI_IMAGE_SYS_PROMPT = ( + "You are an expert image-generation engine. You must ALWAYS produce an image.\n" + "Interpret all user input—regardless of " + "format, intent, or abstraction—as literal visual directives for image composition.\n" + "If a prompt is conversational or lacks specific visual details, " + "you must creatively invent a concrete visual scenario that depicts the concept.\n" + "Prioritize generating the visual representation above any text, formatting, or conversational requests." +) class GeminiModel(str, Enum): @@ -57,6 +64,7 @@ class GeminiModel(str, Enum): gemini_2_5_flash_preview_04_17 = "gemini-2.5-flash-preview-04-17" gemini_2_5_pro = "gemini-2.5-pro" gemini_2_5_flash = "gemini-2.5-flash" + gemini_3_0_pro = "gemini-3-pro-preview" class GeminiImageModel(str, Enum): @@ -68,24 +76,43 @@ class GeminiImageModel(str, Enum): gemini_2_5_flash_image = "gemini-2.5-flash-image" -def create_image_parts(image_input: torch.Tensor) -> list[GeminiPart]: - """ - Convert image tensor input to Gemini API compatible parts. - - Args: - image_input: Batch of image tensors from ComfyUI. - - Returns: - List of GeminiPart objects containing the encoded images. - """ +async def create_image_parts( + cls: type[IO.ComfyNode], + images: Input.Image, + image_limit: int = 0, +) -> list[GeminiPart]: image_parts: list[GeminiPart] = [] - for image_index in range(image_input.shape[0]): - image_as_b64 = tensor_to_base64_string(image_input[image_index].unsqueeze(0)) + 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) + if total_images <= 0: + raise ValueError("No images provided to create_image_parts; at least one image is required.") + + # If image_limit == 0 --> use all images; otherwise clamp to image_limit. + effective_max = total_images if image_limit == 0 else min(total_images, image_limit) + + # Number of images we'll send as URLs (fileData) + num_url_images = min(effective_max, 10) # Vertex API max number of image links + reference_images_urls = await upload_images_to_comfyapi( + cls, + images, + max_images=num_url_images, + ) + for reference_image_url in reference_images_urls: + image_parts.append( + GeminiPart( + fileData=GeminiFileData( + mimeType=GeminiMimeType.image_png, + fileUri=reference_image_url, + ) + ) + ) + for idx in range(num_url_images, effective_max): image_parts.append( GeminiPart( inlineData=GeminiInlineData( mimeType=GeminiMimeType.image_png, - data=image_as_b64, + data=tensor_to_base64_string(images[idx]), ) ) ) @@ -103,11 +130,23 @@ def get_parts_by_type(response: GeminiGenerateContentResponse, part_type: Litera Returns: List of response parts matching the requested type. """ + if response.candidates is None: + if response.promptFeedback and response.promptFeedback.blockReason: + feedback = response.promptFeedback + raise ValueError( + f"Gemini API blocked the request. Reason: {feedback.blockReason} ({feedback.blockReasonMessage})" + ) + raise ValueError( + "Gemini API returned no response candidates. If you are using the `IMAGE` modality, " + "try changing it to `IMAGE+TEXT` to view the model's reasoning and understand why image generation failed." + ) parts = [] for part in response.candidates[0].content.parts: - if part_type == "text" and hasattr(part, "text") and part.text: + if part_type == "text" and part.text: parts.append(part) - elif hasattr(part, "inlineData") and part.inlineData and part.inlineData.mimeType == part_type: + elif part.inlineData and part.inlineData.mimeType == part_type: + parts.append(part) + elif part.fileData and part.fileData.mimeType == part_type: parts.append(part) # Skip parts that don't match the requested type return parts @@ -127,18 +166,65 @@ def get_text_from_response(response: GeminiGenerateContentResponse) -> str: return "\n".join([part.text for part in parts]) -def get_image_from_response(response: GeminiGenerateContentResponse) -> torch.Tensor: - image_tensors: list[torch.Tensor] = [] +async def get_image_from_response(response: GeminiGenerateContentResponse) -> Input.Image: + image_tensors: list[Input.Image] = [] parts = get_parts_by_type(response, "image/png") for part in parts: - image_data = base64.b64decode(part.inlineData.data) - returned_image = bytesio_to_image_tensor(BytesIO(image_data)) + if part.inlineData: + image_data = base64.b64decode(part.inlineData.data) + returned_image = bytesio_to_image_tensor(BytesIO(image_data)) + else: + returned_image = await download_url_to_image_tensor(part.fileData.fileUri) image_tensors.append(returned_image) if len(image_tensors) == 0: return torch.zeros((1, 1024, 1024, 4)) return torch.cat(image_tensors, dim=0) +def calculate_tokens_price(response: GeminiGenerateContentResponse) -> float | None: + if not response.modelVersion: + return None + # Define prices (Cost per 1,000,000 tokens), see https://cloud.google.com/vertex-ai/generative-ai/pricing + if response.modelVersion in ("gemini-2.5-pro-preview-05-06", "gemini-2.5-pro"): + input_tokens_price = 1.25 + output_text_tokens_price = 10.0 + output_image_tokens_price = 0.0 + elif response.modelVersion in ( + "gemini-2.5-flash-preview-04-17", + "gemini-2.5-flash", + ): + input_tokens_price = 0.30 + output_text_tokens_price = 2.50 + output_image_tokens_price = 0.0 + elif response.modelVersion in ( + "gemini-2.5-flash-image-preview", + "gemini-2.5-flash-image", + ): + input_tokens_price = 0.30 + output_text_tokens_price = 2.50 + output_image_tokens_price = 30.0 + elif response.modelVersion == "gemini-3-pro-preview": + input_tokens_price = 2 + output_text_tokens_price = 12.0 + output_image_tokens_price = 0.0 + elif response.modelVersion == "gemini-3-pro-image-preview": + input_tokens_price = 2 + output_text_tokens_price = 12.0 + output_image_tokens_price = 120.0 + else: + return None + final_price = response.usageMetadata.promptTokenCount * input_tokens_price + if response.usageMetadata.candidatesTokensDetails: + for i in response.usageMetadata.candidatesTokensDetails: + if i.modality == Modality.IMAGE: + final_price += output_image_tokens_price * i.tokenCount # for Nano Banana models + else: + final_price += output_text_tokens_price * i.tokenCount + if response.usageMetadata.thoughtsTokenCount: + final_price += output_text_tokens_price * response.usageMetadata.thoughtsTokenCount + return final_price / 1_000_000.0 + + class GeminiNode(IO.ComfyNode): """ Node to generate text responses from a Gemini model. @@ -206,6 +292,13 @@ class GeminiNode(IO.ComfyNode): tooltip="Optional file(s) to use as context for the model. " "Accepts inputs from the Gemini Generate Content Input Files node.", ), + IO.String.Input( + "system_prompt", + multiline=True, + default="", + optional=True, + tooltip="Foundational instructions that dictate an AI's behavior.", + ), ], outputs=[ IO.String.Output(), @@ -222,7 +315,9 @@ class GeminiNode(IO.ComfyNode): def create_video_parts(cls, video_input: Input.Video) -> list[GeminiPart]: """Convert video input to Gemini API compatible parts.""" - base_64_string = video_to_base64_string(video_input, container_format=VideoContainer.MP4, codec=VideoCodec.H264) + base_64_string = video_to_base64_string( + video_input, container_format=Types.VideoContainer.MP4, codec=Types.VideoCodec.H264 + ) return [ GeminiPart( inlineData=GeminiInlineData( @@ -272,10 +367,11 @@ class GeminiNode(IO.ComfyNode): prompt: str, model: str, seed: int, - images: Optional[torch.Tensor] = None, - audio: Optional[Input.Audio] = None, - video: Optional[Input.Video] = None, - files: Optional[list[GeminiPart]] = None, + images: Input.Image | None = None, + audio: Input.Audio | None = None, + video: Input.Video | None = None, + files: list[GeminiPart] | None = None, + system_prompt: str = "", ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=False) @@ -284,8 +380,7 @@ class GeminiNode(IO.ComfyNode): # Add other modal parts if images is not None: - image_parts = create_image_parts(images) - parts.extend(image_parts) + parts.extend(await create_image_parts(cls, images)) if audio is not None: parts.extend(cls.create_audio_parts(audio)) if video is not None: @@ -293,46 +388,27 @@ class GeminiNode(IO.ComfyNode): if files is not None: parts.extend(files) - # Create response + gemini_system_prompt = None + if system_prompt: + gemini_system_prompt = GeminiSystemInstructionContent(parts=[GeminiTextPart(text=system_prompt)], role=None) + response = await sync_op( cls, endpoint=ApiEndpoint(path=f"{GEMINI_BASE_ENDPOINT}/{model}", method="POST"), data=GeminiGenerateContentRequest( contents=[ GeminiContent( - role="user", + role=GeminiRole.user, parts=parts, ) - ] + ], + systemInstruction=gemini_system_prompt, ), response_model=GeminiGenerateContentResponse, + price_extractor=calculate_tokens_price, ) - # Get result output output_text = get_text_from_response(response) - if output_text: - # Not a true chat history like the OpenAI Chat node. It is emulated so the frontend can show a copy button. - render_spec = { - "node_id": cls.hidden.unique_id, - "component": "ChatHistoryWidget", - "props": { - "history": json.dumps( - [ - { - "prompt": prompt, - "response": output_text, - "response_id": str(uuid.uuid4()), - "timestamp": time.time(), - } - ] - ), - }, - } - PromptServer.instance.send_sync( - "display_component", - render_spec, - ) - return IO.NodeOutput(output_text or "Empty response from Gemini model...") @@ -406,7 +482,7 @@ class GeminiInputFiles(IO.ComfyNode): ) @classmethod - def execute(cls, file: str, GEMINI_INPUT_FILES: Optional[list[GeminiPart]] = None) -> IO.NodeOutput: + def execute(cls, file: str, GEMINI_INPUT_FILES: list[GeminiPart] | None = None) -> IO.NodeOutput: """Loads and formats input files for Gemini API.""" if GEMINI_INPUT_FILES is None: GEMINI_INPUT_FILES = [] @@ -421,7 +497,7 @@ class GeminiImage(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="GeminiImageNode", - display_name="Google Gemini Image", + display_name="Nano Banana (Google Gemini Image)", category="api node/image/Gemini", description="Edit images synchronously via Google API.", inputs=[ @@ -469,6 +545,20 @@ class GeminiImage(IO.ComfyNode): "or otherwise generates 1:1 squares.", optional=True, ), + IO.Combo.Input( + "response_modalities", + options=["IMAGE+TEXT", "IMAGE"], + tooltip="Choose 'IMAGE' for image-only output, or " + "'IMAGE+TEXT' to return both the generated image and a text response.", + optional=True, + ), + IO.String.Input( + "system_prompt", + multiline=True, + default=GEMINI_IMAGE_SYS_PROMPT, + optional=True, + tooltip="Foundational instructions that dictate an AI's behavior.", + ), ], outputs=[ IO.Image.Output(), @@ -488,9 +578,11 @@ class GeminiImage(IO.ComfyNode): prompt: str, model: str, seed: int, - images: Optional[torch.Tensor] = None, - files: Optional[list[GeminiPart]] = None, + images: Input.Image | None = None, + files: list[GeminiPart] | None = None, aspect_ratio: str = "auto", + response_modalities: str = "IMAGE+TEXT", + system_prompt: str = "", ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=True, min_length=1) parts: list[GeminiPart] = [GeminiPart(text=prompt)] @@ -500,53 +592,164 @@ class GeminiImage(IO.ComfyNode): image_config = GeminiImageConfig(aspectRatio=aspect_ratio) if images is not None: - image_parts = create_image_parts(images) - parts.extend(image_parts) + parts.extend(await create_image_parts(cls, images)) if files is not None: parts.extend(files) + gemini_system_prompt = None + if system_prompt: + gemini_system_prompt = GeminiSystemInstructionContent(parts=[GeminiTextPart(text=system_prompt)], role=None) + response = await sync_op( cls, - endpoint=ApiEndpoint(path=f"{GEMINI_BASE_ENDPOINT}/{model}", method="POST"), + ApiEndpoint(path=f"/proxy/vertexai/gemini/{model}", method="POST"), data=GeminiImageGenerateContentRequest( contents=[ - GeminiContent(role="user", parts=parts), + GeminiContent(role=GeminiRole.user, parts=parts), ], generationConfig=GeminiImageGenerationConfig( - responseModalities=["TEXT", "IMAGE"], + responseModalities=(["IMAGE"] if response_modalities == "IMAGE" else ["TEXT", "IMAGE"]), imageConfig=None if aspect_ratio == "auto" else image_config, ), + 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)) + + +class GeminiImage2(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="GeminiImage2Node", + display_name="Nano Banana Pro (Google Gemini Image)", + 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.Combo.Input( + "model", + options=["gemini-3-pro-image-preview"], + ), + 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( + "aspect_ratio", + options=["auto", "1:1", "2:3", "3:2", "3:4", "4:3", "4:5", "5:4", "9:16", "16:9", "21:9"], + 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( + "response_modalities", + options=["IMAGE+TEXT", "IMAGE"], + tooltip="Choose 'IMAGE' for image-only output, or " + "'IMAGE+TEXT' to return both the generated image and a text response.", + ), + IO.Image.Input( + "images", + optional=True, + tooltip="Optional reference image(s). " + "To include multiple images, use the Batch Images node (up to 14).", + ), + 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.", + ), + IO.String.Input( + "system_prompt", + multiline=True, + default=GEMINI_IMAGE_SYS_PROMPT, + optional=True, + tooltip="Foundational instructions that dictate an AI's behavior.", + ), + ], + outputs=[ + IO.Image.Output(), + IO.String.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, ) - output_image = get_image_from_response(response) - output_text = get_text_from_response(response) - if output_text: - # Not a true chat history like the OpenAI Chat node. It is emulated so the frontend can show a copy button. - render_spec = { - "node_id": cls.hidden.unique_id, - "component": "ChatHistoryWidget", - "props": { - "history": json.dumps( - [ - { - "prompt": prompt, - "response": output_text, - "response_id": str(uuid.uuid4()), - "timestamp": time.time(), - } - ] - ), - }, - } - PromptServer.instance.send_sync( - "display_component", - render_spec, - ) + @classmethod + async def execute( + cls, + prompt: str, + model: str, + seed: int, + aspect_ratio: str, + resolution: str, + response_modalities: str, + images: Input.Image | None = None, + files: list[GeminiPart] | None = None, + system_prompt: str = "", + ) -> IO.NodeOutput: + validate_string(prompt, strip_whitespace=True, min_length=1) - output_text = output_text or "Empty response from Gemini model..." - return IO.NodeOutput(output_image, output_text) + parts: list[GeminiPart] = [GeminiPart(text=prompt)] + if images is not None: + if get_number_of_images(images) > 14: + raise ValueError("The current maximum number of supported images is 14.") + parts.extend(await create_image_parts(cls, images)) + if files is not None: + parts.extend(files) + + image_config = GeminiImageConfig(imageSize=resolution) + if aspect_ratio != "auto": + image_config.aspectRatio = 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}", 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, + ), + 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)) class GeminiExtension(ComfyExtension): @@ -555,6 +758,7 @@ class GeminiExtension(ComfyExtension): return [ GeminiNode, GeminiImage, + GeminiImage2, GeminiInputFiles, ] diff --git a/comfy_api_nodes/nodes_kling.py b/comfy_api_nodes/nodes_kling.py index 7b23e9cf9..9c707a339 100644 --- a/comfy_api_nodes/nodes_kling.py +++ b/comfy_api_nodes/nodes_kling.py @@ -4,15 +4,14 @@ For source of truth on the allowed permutations of request fields, please refere - [Compatibility Table](https://app.klingai.com/global/dev/document-api/apiReference/model/skillsMap) """ -from __future__ import annotations -from typing import Optional, TypeVar -import math import logging - -from typing_extensions import override +import math +import re import torch +from typing_extensions import override +from comfy_api.latest import IO, ComfyExtension, Input, InputImpl from comfy_api_nodes.apis import ( KlingCameraControl, KlingCameraConfig, @@ -50,25 +49,36 @@ from comfy_api_nodes.apis import ( KlingCharacterEffectModelName, KlingSingleImageEffectModelName, ) +from comfy_api_nodes.apis.kling_api import ( + ImageToVideoWithAudioRequest, + MotionControlRequest, + OmniImageParamImage, + OmniParamImage, + OmniParamVideo, + OmniProFirstLastFrameRequest, + OmniProImageRequest, + OmniProReferences2VideoRequest, + OmniProText2VideoRequest, + TaskStatusResponse, + TextToVideoWithAudioRequest, +) from comfy_api_nodes.util import ( - validate_image_dimensions, + ApiEndpoint, + download_url_to_image_tensor, + download_url_to_video_output, + get_number_of_images, + poll_op, + sync_op, + tensor_to_base64_string, + upload_audio_to_comfyapi, + upload_images_to_comfyapi, + upload_video_to_comfyapi, validate_image_aspect_ratio, + validate_image_dimensions, + validate_string, validate_video_dimensions, validate_video_duration, - tensor_to_base64_string, - validate_string, - upload_audio_to_comfyapi, - download_url_to_image_tensor, - upload_video_to_comfyapi, - download_url_to_video_output, - sync_op, - ApiEndpoint, - poll_op, ) -from comfy_api.input_impl import VideoFromFile -from comfy_api.input.basic_types import AudioInput -from comfy_api.input.video_types import VideoInput -from comfy_api.latest import ComfyExtension, IO KLING_API_VERSION = "v1" PATH_TEXT_TO_VIDEO = f"/proxy/kling/{KLING_API_VERSION}/videos/text2video" @@ -94,14 +104,8 @@ AVERAGE_DURATION_IMAGE_GEN = 32 AVERAGE_DURATION_VIDEO_EFFECTS = 320 AVERAGE_DURATION_VIDEO_EXTEND = 320 -R = TypeVar("R") - MODE_TEXT2VIDEO = { - "standard mode / 5s duration / kling-v1": ("std", "5", "kling-v1"), - "standard mode / 10s duration / kling-v1": ("std", "10", "kling-v1"), - "pro mode / 5s duration / kling-v1": ("pro", "5", "kling-v1"), - "pro mode / 10s duration / kling-v1": ("pro", "10", "kling-v1"), "standard mode / 5s duration / kling-v1-6": ("std", "5", "kling-v1-6"), "standard mode / 10s duration / kling-v1-6": ("std", "10", "kling-v1-6"), "pro mode / 5s duration / kling-v2-master": ("pro", "5", "kling-v2-master"), @@ -122,14 +126,14 @@ See: [Kling API Docs Capability Map](https://app.klingai.com/global/dev/document MODE_START_END_FRAME = { - "standard mode / 5s duration / kling-v1": ("std", "5", "kling-v1"), - "pro mode / 5s duration / kling-v1": ("pro", "5", "kling-v1"), "pro mode / 5s duration / kling-v1-5": ("pro", "5", "kling-v1-5"), "pro mode / 10s duration / kling-v1-5": ("pro", "10", "kling-v1-5"), "pro mode / 5s duration / kling-v1-6": ("pro", "5", "kling-v1-6"), "pro mode / 10s duration / kling-v1-6": ("pro", "10", "kling-v1-6"), "pro mode / 5s duration / kling-v2-1": ("pro", "5", "kling-v2-1"), "pro mode / 10s duration / kling-v2-1": ("pro", "10", "kling-v2-1"), + "pro mode / 5s duration / kling-v2-5-turbo": ("pro", "5", "kling-v2-5-turbo"), + "pro mode / 10s duration / kling-v2-5-turbo": ("pro", "10", "kling-v2-5-turbo"), } """ Returns a mapping of mode strings to their corresponding (mode, duration, model_name) tuples. @@ -206,6 +210,50 @@ VOICES_CONFIG = { } +def normalize_omni_prompt_references(prompt: str) -> str: + """ + Rewrites Kling Omni-style placeholders used in the app, like: + + @image, @image1, @image2, ... @imageN + @video, @video1, @video2, ... @videoN + + into the API-compatible form: + + <<>>, <<>>, ... + <<>>, <<>>, ... + + This is a UX shim for ComfyUI so users can type the same syntax as in the Kling app. + """ + if not prompt: + return prompt + + def _image_repl(match): + return f"<<>>" + + def _video_repl(match): + return f"<<>>" + + # (? and not @imageFoo + prompt = re.sub(r"(?\d*)(?!\w)", _image_repl, prompt) + return re.sub(r"(?\d*)(?!\w)", _video_repl, prompt) + + +async def finish_omni_video_task(cls: type[IO.ComfyNode], response: TaskStatusResponse) -> IO.NodeOutput: + if response.code: + raise RuntimeError( + f"Kling request failed. Code: {response.code}, Message: {response.message}, Data: {response.data}" + ) + final_response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/kling/v1/videos/omni-video/{response.data.task_id}"), + response_model=TaskStatusResponse, + status_extractor=lambda r: (r.data.task_status if r.data else None), + max_poll_attempts=160, + ) + return IO.NodeOutput(await download_url_to_video_output(final_response.data.task_result.videos[0].url)) + + def is_valid_camera_control_configs(configs: list[float]) -> bool: """Verifies that at least one camera control configuration is non-zero.""" return any(not math.isclose(value, 0.0) for value in configs) @@ -296,7 +344,7 @@ def get_video_from_response(response) -> KlingVideoResult: return video -def get_video_url_from_response(response) -> Optional[str]: +def get_video_url_from_response(response) -> str | None: """Returns the first video url from the Kling video generation task result. Will not raise an error if the response is not valid. """ @@ -315,7 +363,7 @@ def get_images_from_response(response) -> list[KlingImageResult]: return images -def get_images_urls_from_response(response) -> Optional[str]: +def get_images_urls_from_response(response) -> str | None: """Returns the list of image urls from the Kling image generation task result. Will not raise an error if the response is not valid. If there is only one image, returns the url as a string. If there are multiple images, returns a list of urls. """ @@ -349,7 +397,7 @@ async def execute_text2video( model_mode: str, duration: str, aspect_ratio: str, - camera_control: Optional[KlingCameraControl] = None, + camera_control: KlingCameraControl | None = None, ) -> IO.NodeOutput: validate_prompts(prompt, negative_prompt, MAX_PROMPT_LENGTH_T2V) task_creation_response = await sync_op( @@ -394,8 +442,8 @@ async def execute_image2video( model_mode: str, aspect_ratio: str, duration: str, - camera_control: Optional[KlingCameraControl] = None, - end_frame: Optional[torch.Tensor] = None, + camera_control: KlingCameraControl | None = None, + end_frame: torch.Tensor | None = None, ) -> IO.NodeOutput: validate_prompts(prompt, negative_prompt, MAX_PROMPT_LENGTH_I2V) validate_input_image(start_frame) @@ -432,12 +480,12 @@ async def execute_image2video( task_id = task_creation_response.data.task_id final_response = await poll_op( - cls, - ApiEndpoint(path=f"{PATH_IMAGE_TO_VIDEO}/{task_id}"), - response_model=KlingImage2VideoResponse, - estimated_duration=AVERAGE_DURATION_I2V, - status_extractor=lambda r: (r.data.task_status.value if r.data and r.data.task_status else None), - ) + cls, + ApiEndpoint(path=f"{PATH_IMAGE_TO_VIDEO}/{task_id}"), + response_model=KlingImage2VideoResponse, + estimated_duration=AVERAGE_DURATION_I2V, + status_extractor=lambda r: (r.data.task_status.value if r.data and r.data.task_status else None), + ) validate_video_result_response(final_response) video = get_video_from_response(final_response) @@ -451,9 +499,9 @@ async def execute_video_effect( model_name: str, duration: KlingVideoGenDuration, image_1: torch.Tensor, - image_2: Optional[torch.Tensor] = None, - model_mode: Optional[KlingVideoGenMode] = None, -) -> tuple[VideoFromFile, str, str]: + image_2: torch.Tensor | None = None, + model_mode: KlingVideoGenMode | None = None, +) -> tuple[InputImpl.VideoFromFile, str, str]: if dual_character: request_input_field = KlingDualCharacterEffectInput( model_name=model_name, @@ -499,13 +547,13 @@ async def execute_video_effect( async def execute_lipsync( cls: type[IO.ComfyNode], - video: VideoInput, - audio: Optional[AudioInput] = None, - voice_language: Optional[str] = None, - model_mode: Optional[str] = None, - text: Optional[str] = None, - voice_speed: Optional[float] = None, - voice_id: Optional[str] = None, + video: Input.Video, + audio: Input.Audio | None = None, + voice_language: str | None = None, + model_mode: str | None = None, + text: str | None = None, + voice_speed: float | None = None, + voice_id: str | None = None, ) -> IO.NodeOutput: if text: validate_string(text, field_name="Text", max_length=MAX_PROMPT_LENGTH_LIP_SYNC) @@ -518,7 +566,9 @@ async def execute_lipsync( # Upload the audio file to Comfy API and get download URL if audio: - audio_url = await upload_audio_to_comfyapi(cls, audio) + audio_url = await upload_audio_to_comfyapi( + cls, audio, container_format="mp3", codec_name="libmp3lame", mime_type="audio/mpeg", filename="output.mp3" + ) logging.info("Uploaded audio to Comfy API. URL: %s", audio_url) else: audio_url = None @@ -699,7 +749,7 @@ class KlingTextToVideoNode(IO.ComfyNode): IO.Combo.Input( "mode", options=modes, - default=modes[4], + default=modes[8], tooltip="The configuration to use for the video generation following the format: mode / duration / model_name.", ), ], @@ -738,6 +788,493 @@ class KlingTextToVideoNode(IO.ComfyNode): ) +class OmniProTextToVideoNode(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="KlingOmniProTextToVideoNode", + display_name="Kling Omni Text to Video (Pro)", + category="api node/video/Kling", + description="Use text prompts to generate videos with the latest Kling model.", + inputs=[ + IO.Combo.Input("model_name", options=["kling-video-o1"]), + IO.String.Input( + "prompt", + multiline=True, + tooltip="A text prompt describing the video content. " + "This can include both positive and negative descriptions.", + ), + IO.Combo.Input("aspect_ratio", options=["16:9", "9:16", "1:1"]), + IO.Combo.Input("duration", options=[5, 10]), + IO.Combo.Input("resolution", options=["1080p", "720p"], 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, + ) + + @classmethod + async def execute( + cls, + model_name: str, + prompt: str, + aspect_ratio: str, + duration: int, + resolution: str = "1080p", + ) -> IO.NodeOutput: + validate_string(prompt, min_length=1, max_length=2500) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"), + response_model=TaskStatusResponse, + data=OmniProText2VideoRequest( + model_name=model_name, + prompt=prompt, + aspect_ratio=aspect_ratio, + duration=str(duration), + mode="pro" if resolution == "1080p" else "std", + ), + ) + return await finish_omni_video_task(cls, response) + + +class OmniProFirstLastFrameNode(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="KlingOmniProFirstLastFrameNode", + display_name="Kling Omni First-Last-Frame to Video (Pro)", + category="api node/video/Kling", + description="Use a start frame, an optional end frame, or reference images with the latest Kling model.", + inputs=[ + IO.Combo.Input("model_name", options=["kling-video-o1"]), + IO.String.Input( + "prompt", + multiline=True, + tooltip="A text prompt describing the video content. " + "This can include both positive and negative descriptions.", + ), + IO.Int.Input("duration", default=5, min=3, max=10, display_mode=IO.NumberDisplay.slider), + IO.Image.Input("first_frame"), + IO.Image.Input( + "end_frame", + optional=True, + tooltip="An optional end frame for the video. " + "This cannot be used simultaneously with 'reference_images'.", + ), + IO.Image.Input( + "reference_images", + optional=True, + tooltip="Up to 6 additional reference images.", + ), + IO.Combo.Input("resolution", options=["1080p", "720p"], 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, + ) + + @classmethod + async def execute( + cls, + model_name: str, + prompt: str, + duration: int, + first_frame: Input.Image, + end_frame: Input.Image | None = None, + reference_images: Input.Image | None = None, + resolution: str = "1080p", + ) -> IO.NodeOutput: + prompt = normalize_omni_prompt_references(prompt) + validate_string(prompt, min_length=1, max_length=2500) + if end_frame is not None and reference_images is not None: + raise ValueError("The 'end_frame' input cannot be used simultaneously with 'reference_images'.") + if duration not in (5, 10) and end_frame is None and reference_images is None: + raise ValueError( + "Duration is only supported for 5 or 10 seconds if there is no end frame or reference images." + ) + validate_image_dimensions(first_frame, min_width=300, min_height=300) + validate_image_aspect_ratio(first_frame, (1, 2.5), (2.5, 1)) + image_list: list[OmniParamImage] = [ + OmniParamImage( + image_url=(await upload_images_to_comfyapi(cls, first_frame, wait_label="Uploading first frame"))[0], + type="first_frame", + ) + ] + if end_frame is not None: + validate_image_dimensions(end_frame, min_width=300, min_height=300) + validate_image_aspect_ratio(end_frame, (1, 2.5), (2.5, 1)) + image_list.append( + OmniParamImage( + image_url=(await upload_images_to_comfyapi(cls, end_frame, wait_label="Uploading end frame"))[0], + type="end_frame", + ) + ) + if reference_images is not None: + if get_number_of_images(reference_images) > 6: + raise ValueError("The maximum number of reference images allowed is 6.") + for i in reference_images: + validate_image_dimensions(i, min_width=300, min_height=300) + validate_image_aspect_ratio(i, (1, 2.5), (2.5, 1)) + for i in await upload_images_to_comfyapi(cls, reference_images, wait_label="Uploading reference frame(s)"): + image_list.append(OmniParamImage(image_url=i)) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"), + response_model=TaskStatusResponse, + data=OmniProFirstLastFrameRequest( + model_name=model_name, + prompt=prompt, + duration=str(duration), + image_list=image_list, + mode="pro" if resolution == "1080p" else "std", + ), + ) + return await finish_omni_video_task(cls, response) + + +class OmniProImageToVideoNode(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="KlingOmniProImageToVideoNode", + display_name="Kling Omni Image to Video (Pro)", + category="api node/video/Kling", + description="Use up to 7 reference images to generate a video with the latest Kling model.", + inputs=[ + IO.Combo.Input("model_name", options=["kling-video-o1"]), + IO.String.Input( + "prompt", + multiline=True, + tooltip="A text prompt describing the video content. " + "This can include both positive and negative descriptions.", + ), + IO.Combo.Input("aspect_ratio", options=["16:9", "9:16", "1:1"]), + IO.Int.Input("duration", default=3, min=3, max=10, display_mode=IO.NumberDisplay.slider), + IO.Image.Input( + "reference_images", + tooltip="Up to 7 reference images.", + ), + IO.Combo.Input("resolution", options=["1080p", "720p"], 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, + ) + + @classmethod + async def execute( + cls, + model_name: str, + prompt: str, + aspect_ratio: str, + duration: int, + reference_images: Input.Image, + resolution: str = "1080p", + ) -> IO.NodeOutput: + prompt = normalize_omni_prompt_references(prompt) + validate_string(prompt, min_length=1, max_length=2500) + if get_number_of_images(reference_images) > 7: + raise ValueError("The maximum number of reference images is 7.") + for i in reference_images: + validate_image_dimensions(i, min_width=300, min_height=300) + validate_image_aspect_ratio(i, (1, 2.5), (2.5, 1)) + image_list: list[OmniParamImage] = [] + for i in await upload_images_to_comfyapi(cls, reference_images, wait_label="Uploading reference image"): + image_list.append(OmniParamImage(image_url=i)) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"), + response_model=TaskStatusResponse, + data=OmniProReferences2VideoRequest( + model_name=model_name, + prompt=prompt, + aspect_ratio=aspect_ratio, + duration=str(duration), + image_list=image_list, + mode="pro" if resolution == "1080p" else "std", + ), + ) + return await finish_omni_video_task(cls, response) + + +class OmniProVideoToVideoNode(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="KlingOmniProVideoToVideoNode", + display_name="Kling Omni Video to Video (Pro)", + category="api node/video/Kling", + description="Use a video and up to 4 reference images to generate a video with the latest Kling model.", + inputs=[ + IO.Combo.Input("model_name", options=["kling-video-o1"]), + IO.String.Input( + "prompt", + multiline=True, + tooltip="A text prompt describing the video content. " + "This can include both positive and negative descriptions.", + ), + IO.Combo.Input("aspect_ratio", options=["16:9", "9:16", "1:1"]), + IO.Int.Input("duration", default=3, min=3, max=10, display_mode=IO.NumberDisplay.slider), + IO.Video.Input("reference_video", tooltip="Video to use as a reference."), + IO.Boolean.Input("keep_original_sound", default=True), + IO.Image.Input( + "reference_images", + tooltip="Up to 4 additional reference images.", + optional=True, + ), + IO.Combo.Input("resolution", options=["1080p", "720p"], 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, + ) + + @classmethod + async def execute( + cls, + model_name: str, + prompt: str, + aspect_ratio: str, + duration: int, + reference_video: Input.Video, + keep_original_sound: bool, + reference_images: Input.Image | None = None, + resolution: str = "1080p", + ) -> IO.NodeOutput: + prompt = normalize_omni_prompt_references(prompt) + validate_string(prompt, min_length=1, max_length=2500) + validate_video_duration(reference_video, min_duration=3.0, max_duration=10.05) + validate_video_dimensions(reference_video, min_width=720, min_height=720, max_width=2160, max_height=2160) + image_list: list[OmniParamImage] = [] + if reference_images is not None: + if get_number_of_images(reference_images) > 4: + raise ValueError("The maximum number of reference images allowed with a video input is 4.") + for i in reference_images: + validate_image_dimensions(i, min_width=300, min_height=300) + validate_image_aspect_ratio(i, (1, 2.5), (2.5, 1)) + for i in await upload_images_to_comfyapi(cls, reference_images, wait_label="Uploading reference image"): + image_list.append(OmniParamImage(image_url=i)) + video_list = [ + OmniParamVideo( + video_url=await upload_video_to_comfyapi(cls, reference_video, wait_label="Uploading reference video"), + refer_type="feature", + keep_original_sound="yes" if keep_original_sound else "no", + ) + ] + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"), + response_model=TaskStatusResponse, + data=OmniProReferences2VideoRequest( + model_name=model_name, + prompt=prompt, + aspect_ratio=aspect_ratio, + duration=str(duration), + image_list=image_list if image_list else None, + video_list=video_list, + mode="pro" if resolution == "1080p" else "std", + ), + ) + return await finish_omni_video_task(cls, response) + + +class OmniProEditVideoNode(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="KlingOmniProEditVideoNode", + display_name="Kling Omni Edit Video (Pro)", + category="api node/video/Kling", + description="Edit an existing video with the latest model from Kling.", + inputs=[ + IO.Combo.Input("model_name", options=["kling-video-o1"]), + IO.String.Input( + "prompt", + multiline=True, + tooltip="A text prompt describing the video content. " + "This can include both positive and negative descriptions.", + ), + IO.Video.Input("video", tooltip="Video for editing. The output video length will be the same."), + IO.Boolean.Input("keep_original_sound", default=True), + IO.Image.Input( + "reference_images", + tooltip="Up to 4 additional reference images.", + optional=True, + ), + IO.Combo.Input("resolution", options=["1080p", "720p"], 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, + ) + + @classmethod + async def execute( + cls, + model_name: str, + prompt: str, + video: Input.Video, + keep_original_sound: bool, + reference_images: Input.Image | None = None, + resolution: str = "1080p", + ) -> IO.NodeOutput: + prompt = normalize_omni_prompt_references(prompt) + validate_string(prompt, min_length=1, max_length=2500) + validate_video_duration(video, min_duration=3.0, max_duration=10.05) + validate_video_dimensions(video, min_width=720, min_height=720, max_width=2160, max_height=2160) + image_list: list[OmniParamImage] = [] + if reference_images is not None: + if get_number_of_images(reference_images) > 4: + raise ValueError("The maximum number of reference images allowed with a video input is 4.") + for i in reference_images: + validate_image_dimensions(i, min_width=300, min_height=300) + validate_image_aspect_ratio(i, (1, 2.5), (2.5, 1)) + for i in await upload_images_to_comfyapi(cls, reference_images, wait_label="Uploading reference image"): + image_list.append(OmniParamImage(image_url=i)) + video_list = [ + OmniParamVideo( + video_url=await upload_video_to_comfyapi(cls, video, wait_label="Uploading base video"), + refer_type="base", + keep_original_sound="yes" if keep_original_sound else "no", + ) + ] + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"), + response_model=TaskStatusResponse, + data=OmniProReferences2VideoRequest( + model_name=model_name, + prompt=prompt, + aspect_ratio=None, + duration=None, + image_list=image_list if image_list else None, + video_list=video_list, + mode="pro" if resolution == "1080p" else "std", + ), + ) + return await finish_omni_video_task(cls, response) + + +class OmniProImageNode(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="KlingOmniProImageNode", + display_name="Kling Omni Image (Pro)", + category="api node/image/Kling", + description="Create or edit images with the latest model from Kling.", + inputs=[ + IO.Combo.Input("model_name", options=["kling-image-o1"]), + IO.String.Input( + "prompt", + multiline=True, + tooltip="A text prompt describing the image content. " + "This can include both positive and negative descriptions.", + ), + IO.Combo.Input("resolution", options=["1K", "2K"]), + IO.Combo.Input( + "aspect_ratio", + options=["16:9", "9:16", "1:1", "4:3", "3:4", "3:2", "2:3", "21:9"], + ), + IO.Image.Input( + "reference_images", + tooltip="Up to 10 additional reference images.", + optional=True, + ), + ], + outputs=[ + IO.Image.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, + model_name: str, + prompt: str, + resolution: str, + aspect_ratio: str, + reference_images: Input.Image | None = None, + ) -> IO.NodeOutput: + prompt = normalize_omni_prompt_references(prompt) + validate_string(prompt, min_length=1, max_length=2500) + image_list: list[OmniImageParamImage] = [] + if reference_images is not None: + if get_number_of_images(reference_images) > 10: + raise ValueError("The maximum number of reference images is 10.") + for i in reference_images: + validate_image_dimensions(i, min_width=300, min_height=300) + validate_image_aspect_ratio(i, (1, 2.5), (2.5, 1)) + for i in await upload_images_to_comfyapi(cls, reference_images, wait_label="Uploading reference image"): + image_list.append(OmniImageParamImage(image=i)) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/kling/v1/images/omni-image", method="POST"), + response_model=TaskStatusResponse, + data=OmniProImageRequest( + model_name=model_name, + prompt=prompt, + resolution=resolution.lower(), + aspect_ratio=aspect_ratio, + image_list=image_list if image_list else None, + ), + ) + if response.code: + raise RuntimeError( + f"Kling request failed. Code: {response.code}, Message: {response.message}, Data: {response.data}" + ) + final_response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/kling/v1/images/omni-image/{response.data.task_id}"), + response_model=TaskStatusResponse, + status_extractor=lambda r: (r.data.task_status if r.data else None), + ) + return IO.NodeOutput(await download_url_to_image_tensor(final_response.data.task_result.images[0].url)) + + class KlingCameraControlT2VNode(IO.ComfyNode): """ Kling Text to Video Camera Control Node. This node is a text to video node, but it supports controlling the camera. @@ -785,7 +1322,7 @@ class KlingCameraControlT2VNode(IO.ComfyNode): negative_prompt: str, cfg_scale: float, aspect_ratio: str, - camera_control: Optional[KlingCameraControl] = None, + camera_control: KlingCameraControl | None = None, ) -> IO.NodeOutput: return await execute_text2video( cls, @@ -807,9 +1344,8 @@ class KlingImage2VideoNode(IO.ComfyNode): def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="KlingImage2VideoNode", - display_name="Kling Image to Video", + display_name="Kling Image(First Frame) to Video", category="api node/video/Kling", - description="Kling Image to Video Node", inputs=[ IO.Image.Input("start_frame", tooltip="The reference image used to generate the video."), IO.String.Input("prompt", multiline=True, tooltip="Positive text prompt"), @@ -852,8 +1388,8 @@ class KlingImage2VideoNode(IO.ComfyNode): mode: str, aspect_ratio: str, duration: str, - camera_control: Optional[KlingCameraControl] = None, - end_frame: Optional[torch.Tensor] = None, + camera_control: KlingCameraControl | None = None, + end_frame: torch.Tensor | None = None, ) -> IO.NodeOutput: return await execute_image2video( cls, @@ -963,15 +1499,11 @@ class KlingStartEndFrameNode(IO.ComfyNode): IO.String.Input("prompt", multiline=True, tooltip="Positive text prompt"), IO.String.Input("negative_prompt", multiline=True, tooltip="Negative text prompt"), IO.Float.Input("cfg_scale", default=0.5, min=0.0, max=1.0), - IO.Combo.Input( - "aspect_ratio", - options=[i.value for i in KlingVideoGenAspectRatio], - default="16:9", - ), + IO.Combo.Input("aspect_ratio", options=["16:9", "9:16", "1:1"]), IO.Combo.Input( "mode", options=modes, - default=modes[2], + default=modes[6], tooltip="The configuration to use for the video generation following the format: mode / duration / model_name.", ), ], @@ -1168,7 +1700,10 @@ class KlingSingleImageVideoEffectNode(IO.ComfyNode): category="api node/video/Kling", description="Achieve different special effects when generating a video based on the effect_scene.", inputs=[ - IO.Image.Input("image", tooltip=" Reference Image. URL or Base64 encoded string (without data:image prefix). File size cannot exceed 10MB, resolution not less than 300*300px, aspect ratio between 1:2.5 ~ 2.5:1"), + IO.Image.Input( + "image", + tooltip=" Reference Image. URL or Base64 encoded string (without data:image prefix). File size cannot exceed 10MB, resolution not less than 300*300px, aspect ratio between 1:2.5 ~ 2.5:1", + ), IO.Combo.Input( "effect_scene", options=[i.value for i in KlingSingleImageEffectsScene], @@ -1252,8 +1787,8 @@ class KlingLipSyncAudioToVideoNode(IO.ComfyNode): @classmethod async def execute( cls, - video: VideoInput, - audio: AudioInput, + video: Input.Video, + audio: Input.Audio, voice_language: str, ) -> IO.NodeOutput: return await execute_lipsync( @@ -1312,7 +1847,7 @@ class KlingLipSyncTextToVideoNode(IO.ComfyNode): @classmethod async def execute( cls, - video: VideoInput, + video: Input.Video, text: str, voice: str, voice_speed: float, @@ -1431,7 +1966,7 @@ class KlingImageGenerationNode(IO.ComfyNode): IO.Combo.Input( "model_name", options=[i.value for i in KlingImageGenModelName], - default="kling-v1", + default="kling-v2", ), IO.Combo.Input( "aspect_ratio", @@ -1469,7 +2004,7 @@ class KlingImageGenerationNode(IO.ComfyNode): human_fidelity: float, n: int, aspect_ratio: KlingImageGenAspectRatio, - image: Optional[torch.Tensor] = None, + image: torch.Tensor | None = None, ) -> IO.NodeOutput: validate_string(prompt, field_name="prompt", min_length=1, max_length=MAX_PROMPT_LENGTH_IMAGE_GEN) validate_string(negative_prompt, field_name="negative_prompt", max_length=MAX_PROMPT_LENGTH_IMAGE_GEN) @@ -1514,6 +2049,221 @@ class KlingImageGenerationNode(IO.ComfyNode): return IO.NodeOutput(await image_result_to_node_output(images)) +class TextToVideoWithAudio(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="KlingTextToVideoWithAudio", + display_name="Kling Text to Video with Audio", + category="api node/video/Kling", + inputs=[ + IO.Combo.Input("model_name", options=["kling-v2-6"]), + IO.String.Input("prompt", multiline=True, tooltip="Positive text prompt."), + IO.Combo.Input("mode", options=["pro"]), + IO.Combo.Input("aspect_ratio", options=["16:9", "9:16", "1:1"]), + IO.Combo.Input("duration", options=[5, 10]), + IO.Boolean.Input("generate_audio", default=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, + ) + + @classmethod + async def execute( + cls, + model_name: str, + prompt: str, + mode: str, + aspect_ratio: str, + duration: int, + generate_audio: bool, + ) -> IO.NodeOutput: + validate_string(prompt, min_length=1, max_length=2500) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/kling/v1/videos/text2video", method="POST"), + response_model=TaskStatusResponse, + data=TextToVideoWithAudioRequest( + model_name=model_name, + prompt=prompt, + mode=mode, + aspect_ratio=aspect_ratio, + duration=str(duration), + sound="on" if generate_audio else "off", + ), + ) + if response.code: + raise RuntimeError( + f"Kling request failed. Code: {response.code}, Message: {response.message}, Data: {response.data}" + ) + final_response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/kling/v1/videos/text2video/{response.data.task_id}"), + response_model=TaskStatusResponse, + 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)) + + +class ImageToVideoWithAudio(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="KlingImageToVideoWithAudio", + display_name="Kling Image(First Frame) to Video with Audio", + category="api node/video/Kling", + inputs=[ + IO.Combo.Input("model_name", options=["kling-v2-6"]), + IO.Image.Input("start_frame"), + IO.String.Input("prompt", multiline=True, tooltip="Positive text prompt."), + IO.Combo.Input("mode", options=["pro"]), + IO.Combo.Input("duration", options=[5, 10]), + IO.Boolean.Input("generate_audio", default=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, + ) + + @classmethod + async def execute( + cls, + model_name: str, + start_frame: Input.Image, + prompt: str, + mode: str, + duration: int, + generate_audio: bool, + ) -> IO.NodeOutput: + validate_string(prompt, min_length=1, max_length=2500) + validate_image_dimensions(start_frame, min_width=300, min_height=300) + validate_image_aspect_ratio(start_frame, (1, 2.5), (2.5, 1)) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/kling/v1/videos/image2video", method="POST"), + response_model=TaskStatusResponse, + data=ImageToVideoWithAudioRequest( + model_name=model_name, + image=(await upload_images_to_comfyapi(cls, start_frame))[0], + prompt=prompt, + mode=mode, + duration=str(duration), + sound="on" if generate_audio else "off", + ), + ) + if response.code: + raise RuntimeError( + f"Kling request failed. Code: {response.code}, Message: {response.message}, Data: {response.data}" + ) + final_response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/kling/v1/videos/image2video/{response.data.task_id}"), + response_model=TaskStatusResponse, + 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)) + + +class MotionControl(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="KlingMotionControl", + display_name="Kling Motion Control", + category="api node/video/Kling", + inputs=[ + IO.String.Input("prompt", multiline=True), + IO.Image.Input("reference_image"), + IO.Video.Input( + "reference_video", + tooltip="Motion reference video used to drive movement/expression.\n" + "Duration limits depend on character_orientation:\n" + " - image: 3–10s (max 10s)\n" + " - video: 3–30s (max 30s)", + ), + IO.Boolean.Input("keep_original_sound", default=True), + IO.Combo.Input( + "character_orientation", + options=["video", "image"], + tooltip="Controls where the character's facing/orientation comes from.\n" + "video: movements, expressions, camera moves, and orientation " + "follow the motion reference video (other details via prompt).\n" + "image: movements and expressions still follow the motion reference video, " + "but the character orientation matches the reference image (camera/other details via prompt).", + ), + IO.Combo.Input("mode", options=["pro", "std"]), + ], + 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, + ) + + @classmethod + async def execute( + cls, + prompt: str, + reference_image: Input.Image, + reference_video: Input.Video, + keep_original_sound: bool, + character_orientation: str, + mode: str, + ) -> IO.NodeOutput: + validate_string(prompt, max_length=2500) + validate_image_dimensions(reference_image, min_width=340, min_height=340) + validate_image_aspect_ratio(reference_image, (1, 2.5), (2.5, 1)) + if character_orientation == "image": + validate_video_duration(reference_video, min_duration=3, max_duration=10) + else: + validate_video_duration(reference_video, min_duration=3, max_duration=30) + validate_video_dimensions(reference_video, min_width=340, min_height=340, max_width=3850, max_height=3850) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/kling/v1/videos/motion-control", method="POST"), + response_model=TaskStatusResponse, + data=MotionControlRequest( + prompt=prompt, + image_url=(await upload_images_to_comfyapi(cls, reference_image))[0], + video_url=await upload_video_to_comfyapi(cls, reference_video), + keep_original_sound="yes" if keep_original_sound else "no", + character_orientation=character_orientation, + mode=mode, + ), + ) + if response.code: + raise RuntimeError( + f"Kling request failed. Code: {response.code}, Message: {response.message}, Data: {response.data}" + ) + final_response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/kling/v1/videos/motion-control/{response.data.task_id}"), + response_model=TaskStatusResponse, + 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)) + + class KlingExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: @@ -1531,6 +2281,15 @@ class KlingExtension(ComfyExtension): KlingImageGenerationNode, KlingSingleImageVideoEffectNode, KlingDualCharacterVideoEffectNode, + OmniProTextToVideoNode, + OmniProFirstLastFrameNode, + OmniProImageToVideoNode, + OmniProVideoToVideoNode, + OmniProEditVideoNode, + OmniProImageNode, + TextToVideoWithAudio, + ImageToVideoWithAudio, + MotionControl, ] diff --git a/comfy_api_nodes/nodes_ltxv.py b/comfy_api_nodes/nodes_ltxv.py index 0b757a62b..7e61560dc 100644 --- a/comfy_api_nodes/nodes_ltxv.py +++ b/comfy_api_nodes/nodes_ltxv.py @@ -1,12 +1,9 @@ from io import BytesIO -from typing import Optional -import torch from pydantic import BaseModel, Field from typing_extensions import override -from comfy_api.input_impl import VideoFromFile -from comfy_api.latest import IO, ComfyExtension +from comfy_api.latest import IO, ComfyExtension, Input, InputImpl from comfy_api_nodes.util import ( ApiEndpoint, get_number_of_images, @@ -26,9 +23,9 @@ class ExecuteTaskRequest(BaseModel): model: str = Field(...) duration: int = Field(...) resolution: str = Field(...) - fps: Optional[int] = Field(25) - generate_audio: Optional[bool] = Field(True) - image_uri: Optional[str] = Field(None) + fps: int | None = Field(25) + generate_audio: bool | None = Field(True) + image_uri: str | None = Field(None) class TextToVideoNode(IO.ComfyNode): @@ -103,7 +100,7 @@ class TextToVideoNode(IO.ComfyNode): as_binary=True, max_retries=1, ) - return IO.NodeOutput(VideoFromFile(BytesIO(response))) + return IO.NodeOutput(InputImpl.VideoFromFile(BytesIO(response))) class ImageToVideoNode(IO.ComfyNode): @@ -153,7 +150,7 @@ class ImageToVideoNode(IO.ComfyNode): @classmethod async def execute( cls, - image: torch.Tensor, + image: Input.Image, model: str, prompt: str, duration: int, @@ -183,7 +180,7 @@ class ImageToVideoNode(IO.ComfyNode): as_binary=True, max_retries=1, ) - return IO.NodeOutput(VideoFromFile(BytesIO(response))) + return IO.NodeOutput(InputImpl.VideoFromFile(BytesIO(response))) class LtxvApiExtension(ComfyExtension): diff --git a/comfy_api_nodes/nodes_moonvalley.py b/comfy_api_nodes/nodes_moonvalley.py index 7c31d95b3..2771e4790 100644 --- a/comfy_api_nodes/nodes_moonvalley.py +++ b/comfy_api_nodes/nodes_moonvalley.py @@ -1,11 +1,8 @@ import logging -from typing import Optional -import torch from typing_extensions import override -from comfy_api.input import VideoInput -from comfy_api.latest import IO, ComfyExtension +from comfy_api.latest import IO, ComfyExtension, Input from comfy_api_nodes.apis import ( MoonvalleyPromptResponse, MoonvalleyTextToVideoInferenceParams, @@ -61,7 +58,7 @@ def validate_task_creation_response(response) -> None: raise RuntimeError(error_msg) -def validate_video_to_video_input(video: VideoInput) -> VideoInput: +def validate_video_to_video_input(video: Input.Video) -> Input.Video: """ Validates and processes video input for Moonvalley Video-to-Video generation. @@ -82,7 +79,7 @@ def validate_video_to_video_input(video: VideoInput) -> VideoInput: return _validate_and_trim_duration(video) -def _get_video_dimensions(video: VideoInput) -> tuple[int, int]: +def _get_video_dimensions(video: Input.Video) -> tuple[int, int]: """Extracts video dimensions with error handling.""" try: return video.get_dimensions() @@ -106,7 +103,7 @@ def _validate_video_dimensions(width: int, height: int) -> None: raise ValueError(f"Resolution {width}x{height} not supported. Supported: {supported_list}") -def _validate_and_trim_duration(video: VideoInput) -> VideoInput: +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) @@ -119,7 +116,7 @@ def _validate_minimum_duration(duration: float) -> None: raise ValueError("Input video must be at least 5 seconds long.") -def _trim_if_too_long(video: VideoInput, duration: float) -> VideoInput: +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) @@ -241,7 +238,7 @@ class MoonvalleyImg2VideoNode(IO.ComfyNode): @classmethod async def execute( cls, - image: torch.Tensor, + image: Input.Image, prompt: str, negative_prompt: str, resolution: str, @@ -362,9 +359,9 @@ class MoonvalleyVideo2VideoNode(IO.ComfyNode): prompt: str, negative_prompt: str, seed: int, - video: Optional[VideoInput] = None, + video: Input.Video | None = None, control_type: str = "Motion Transfer", - motion_intensity: Optional[int] = 100, + motion_intensity: int | None = 100, steps=33, prompt_adherence=4.5, ) -> IO.NodeOutput: diff --git a/comfy_api_nodes/nodes_openai.py b/comfy_api_nodes/nodes_openai.py index acf35d276..a6205a34f 100644 --- a/comfy_api_nodes/nodes_openai.py +++ b/comfy_api_nodes/nodes_openai.py @@ -1,51 +1,45 @@ -from io import BytesIO -from typing import Optional, Union -import json +import base64 import os -import time -import uuid from enum import Enum -from inspect import cleandoc +from io import BytesIO + import numpy as np import torch from PIL import Image -from server import PromptServer -import folder_paths -import base64 -from comfy_api.latest import IO, ComfyExtension from typing_extensions import override - +import folder_paths +from comfy_api.latest import IO, ComfyExtension, Input from comfy_api_nodes.apis import ( - OpenAIImageGenerationRequest, - OpenAIImageEditRequest, - OpenAIImageGenerationResponse, - OpenAICreateResponse, - OpenAIResponse, CreateModelResponseProperties, - Item, - OutputContent, - InputImageContent, Detail, - InputTextContent, - InputMessage, - InputMessageContentList, InputContent, InputFileContent, + InputImageContent, + InputMessage, + InputMessageContentList, + InputTextContent, + Item, + OpenAICreateResponse, + OpenAIResponse, + OutputContent, +) +from comfy_api_nodes.apis.openai_api import ( + OpenAIImageEditRequest, + OpenAIImageGenerationRequest, + OpenAIImageGenerationResponse, ) - from comfy_api_nodes.util import ( - downscale_image_tensor, - download_url_to_bytesio, - validate_string, - tensor_to_base64_string, ApiEndpoint, - sync_op, + download_url_to_bytesio, + downscale_image_tensor, poll_op, + sync_op, + tensor_to_base64_string, text_filepath_to_data_uri, + validate_string, ) - RESPONSES_ENDPOINT = "/proxy/openai/v1/responses" STARTING_POINT_ID_PATTERN = r"" @@ -103,9 +97,6 @@ async def validate_and_cast_response(response, timeout: int = None) -> torch.Ten class OpenAIDalle2(IO.ComfyNode): - """ - Generates images synchronously via OpenAI's DALL·E 2 endpoint. - """ @classmethod def define_schema(cls): @@ -113,7 +104,7 @@ class OpenAIDalle2(IO.ComfyNode): node_id="OpenAIDalle2", display_name="OpenAI DALL·E 2", category="api node/image/OpenAI", - description=cleandoc(cls.__doc__ or ""), + description="Generates images synchronously via OpenAI's DALL·E 2 endpoint.", inputs=[ IO.String.Input( "prompt", @@ -239,9 +230,6 @@ class OpenAIDalle2(IO.ComfyNode): class OpenAIDalle3(IO.ComfyNode): - """ - Generates images synchronously via OpenAI's DALL·E 3 endpoint. - """ @classmethod def define_schema(cls): @@ -249,7 +237,7 @@ class OpenAIDalle3(IO.ComfyNode): node_id="OpenAIDalle3", display_name="OpenAI DALL·E 3", category="api node/image/OpenAI", - description=cleandoc(cls.__doc__ or ""), + description="Generates images synchronously via OpenAI's DALL·E 3 endpoint.", inputs=[ IO.String.Input( "prompt", @@ -331,10 +319,16 @@ class OpenAIDalle3(IO.ComfyNode): return IO.NodeOutput(await validate_and_cast_response(response)) +def calculate_tokens_price_image_1(response: OpenAIImageGenerationResponse) -> float | None: + # https://platform.openai.com/docs/pricing + return ((response.usage.input_tokens * 10.0) + (response.usage.output_tokens * 40.0)) / 1_000_000.0 + + +def calculate_tokens_price_image_1_5(response: OpenAIImageGenerationResponse) -> float | None: + return ((response.usage.input_tokens * 8.0) + (response.usage.output_tokens * 32.0)) / 1_000_000.0 + + class OpenAIGPTImage1(IO.ComfyNode): - """ - Generates images synchronously via OpenAI's GPT Image 1 endpoint. - """ @classmethod def define_schema(cls): @@ -342,13 +336,13 @@ class OpenAIGPTImage1(IO.ComfyNode): node_id="OpenAIGPTImage1", display_name="OpenAI GPT Image 1", category="api node/image/OpenAI", - description=cleandoc(cls.__doc__ or ""), + description="Generates images synchronously via OpenAI's GPT Image 1 endpoint.", inputs=[ IO.String.Input( "prompt", default="", multiline=True, - tooltip="Text prompt for GPT Image 1", + tooltip="Text prompt for GPT Image", ), IO.Int.Input( "seed", @@ -370,8 +364,8 @@ class OpenAIGPTImage1(IO.ComfyNode): ), IO.Combo.Input( "background", - default="opaque", - options=["opaque", "transparent"], + default="auto", + options=["auto", "opaque", "transparent"], tooltip="Return image with or without background", optional=True, ), @@ -402,6 +396,11 @@ class OpenAIGPTImage1(IO.ComfyNode): tooltip="Optional mask for inpainting (white areas will be replaced)", optional=True, ), + IO.Combo.Input( + "model", + options=["gpt-image-1", "gpt-image-1.5"], + optional=True, + ), ], outputs=[ IO.Image.Output(), @@ -417,32 +416,34 @@ class OpenAIGPTImage1(IO.ComfyNode): @classmethod async def execute( cls, - prompt, - seed=0, - quality="low", - background="opaque", - image=None, - mask=None, - n=1, - size="1024x1024", + prompt: str, + seed: int = 0, + quality: str = "low", + background: str = "opaque", + image: Input.Image | None = None, + mask: Input.Image | None = None, + n: int = 1, + size: str = "1024x1024", + model: str = "gpt-image-1", ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=False) - model = "gpt-image-1" - path = "/proxy/openai/images/generations" - content_type = "application/json" - request_class = OpenAIImageGenerationRequest - files = [] + + if mask is not None and image is None: + raise ValueError("Cannot use a mask without an input image") + + if model == "gpt-image-1": + price_extractor = calculate_tokens_price_image_1 + elif model == "gpt-image-1.5": + price_extractor = calculate_tokens_price_image_1_5 + else: + raise ValueError(f"Unknown model: {model}") if image is not None: - path = "/proxy/openai/images/edits" - request_class = OpenAIImageEditRequest - content_type = "multipart/form-data" - + files = [] batch_size = image.shape[0] - for i in range(batch_size): - single_image = image[i : i + 1] - scaled_image = downscale_image_tensor(single_image).squeeze() + single_image = image[i: i + 1] + 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) @@ -455,44 +456,59 @@ class OpenAIGPTImage1(IO.ComfyNode): else: files.append(("image[]", (f"image_{i}.png", img_byte_arr, "image/png"))) - if mask is not None: - if image is None: - raise Exception("Cannot use a mask without an input image") - if image.shape[0] != 1: - raise Exception("Cannot use a mask with multiple image") - if mask.shape[1:] != image.shape[1:-1]: - raise Exception("Mask and Image must be the same size") - batch, height, width = mask.shape - rgba_mask = torch.zeros(height, width, 4, device="cpu") - rgba_mask[:, :, 3] = 1 - mask.squeeze().cpu() + if mask is not None: + if image.shape[0] != 1: + raise Exception("Cannot use a mask with multiple image") + if mask.shape[1:] != 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)).squeeze() + 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"))) - - # Build the operation - response = await sync_op( - cls, - ApiEndpoint(path=path, method="POST"), - response_model=OpenAIImageGenerationResponse, - data=request_class( - model=model, - prompt=prompt, - quality=quality, - background=background, - n=n, - seed=seed, - size=size, - ), - files=files if files else None, - content_type=content_type, - ) + 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, + prompt=prompt, + quality=quality, + background=background, + n=n, + seed=seed, + 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, + prompt=prompt, + quality=quality, + background=background, + n=n, + seed=seed, + size=size, + moderation="low", + ), + price_extractor=price_extractor, + ) return IO.NodeOutput(await validate_and_cast_response(response)) @@ -587,11 +603,11 @@ class OpenAIChatNode(IO.ComfyNode): def create_input_message_contents( cls, prompt: str, - image: Optional[torch.Tensor] = None, - files: Optional[list[InputFileContent]] = None, + image: torch.Tensor | None = None, + files: list[InputFileContent] | None = None, ) -> InputMessageContentList: """Create a list of input message contents from prompt and optional image.""" - content_list: list[Union[InputContent, InputTextContent, InputImageContent, InputFileContent]] = [ + content_list: list[InputContent | InputTextContent | InputImageContent | InputFileContent] = [ InputTextContent(text=prompt, type="input_text"), ] if image is not None: @@ -617,9 +633,9 @@ class OpenAIChatNode(IO.ComfyNode): prompt: str, persist_context: bool = False, model: SupportedOpenAIModel = SupportedOpenAIModel.gpt_5.value, - images: Optional[torch.Tensor] = None, - files: Optional[list[InputFileContent]] = None, - advanced_options: Optional[CreateModelResponseProperties] = None, + images: torch.Tensor | None = None, + files: list[InputFileContent] | None = None, + advanced_options: CreateModelResponseProperties | None = None, ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=False) @@ -660,30 +676,7 @@ class OpenAIChatNode(IO.ComfyNode): status_extractor=lambda response: response.status, completed_statuses=["incomplete", "completed"] ) - output_text = cls.get_text_from_message_content(cls.get_message_content_from_response(result_response)) - - # Update history - render_spec = { - "node_id": cls.hidden.unique_id, - "component": "ChatHistoryWidget", - "props": { - "history": json.dumps( - [ - { - "prompt": prompt, - "response": output_text, - "response_id": str(uuid.uuid4()), - "timestamp": time.time(), - } - ] - ), - }, - } - PromptServer.instance.send_sync( - "display_component", - render_spec, - ) - return IO.NodeOutput(output_text) + return IO.NodeOutput(cls.get_text_from_message_content(cls.get_message_content_from_response(result_response))) class OpenAIInputFiles(IO.ComfyNode): @@ -790,8 +783,8 @@ class OpenAIChatConfig(IO.ComfyNode): def execute( cls, truncation: bool, - instructions: Optional[str] = None, - max_output_tokens: Optional[int] = None, + instructions: str | None = None, + max_output_tokens: int | None = None, ) -> IO.NodeOutput: """ Configure advanced options for the OpenAI Chat Node. diff --git a/comfy_api_nodes/nodes_pika.py b/comfy_api_nodes/nodes_pika.py deleted file mode 100644 index 51148211b..000000000 --- a/comfy_api_nodes/nodes_pika.py +++ /dev/null @@ -1,568 +0,0 @@ -""" -Pika x ComfyUI API Nodes - -Pika API docs: https://pika-827374fb.mintlify.app/api-reference -""" -from __future__ import annotations - -from io import BytesIO -import logging -from typing import Optional - -import torch - -from typing_extensions import override -from comfy_api.latest import ComfyExtension, IO -from comfy_api.input_impl.video_types import VideoCodec, VideoContainer, VideoInput -from comfy_api_nodes.apis import pika_api as pika_defs -from comfy_api_nodes.util import ( - validate_string, - download_url_to_video_output, - tensor_to_bytesio, - ApiEndpoint, - sync_op, - poll_op, -) - - -PATH_PIKADDITIONS = "/proxy/pika/generate/pikadditions" -PATH_PIKASWAPS = "/proxy/pika/generate/pikaswaps" -PATH_PIKAFFECTS = "/proxy/pika/generate/pikaffects" - -PIKA_API_VERSION = "2.2" -PATH_TEXT_TO_VIDEO = f"/proxy/pika/generate/{PIKA_API_VERSION}/t2v" -PATH_IMAGE_TO_VIDEO = f"/proxy/pika/generate/{PIKA_API_VERSION}/i2v" -PATH_PIKAFRAMES = f"/proxy/pika/generate/{PIKA_API_VERSION}/pikaframes" -PATH_PIKASCENES = f"/proxy/pika/generate/{PIKA_API_VERSION}/pikascenes" - -PATH_VIDEO_GET = "/proxy/pika/videos" - - -async def execute_task( - task_id: str, - cls: type[IO.ComfyNode], -) -> IO.NodeOutput: - final_response: pika_defs.PikaVideoResponse = await poll_op( - cls, - ApiEndpoint(path=f"{PATH_VIDEO_GET}/{task_id}"), - response_model=pika_defs.PikaVideoResponse, - status_extractor=lambda response: (response.status.value if response.status else None), - progress_extractor=lambda response: (response.progress if hasattr(response, "progress") else None), - estimated_duration=60, - max_poll_attempts=240, - ) - if not final_response.url: - error_msg = f"Pika task {task_id} succeeded but no video data found in response:\n{final_response}" - logging.error(error_msg) - raise Exception(error_msg) - video_url = final_response.url - logging.info("Pika task %s succeeded. Video URL: %s", task_id, video_url) - return IO.NodeOutput(await download_url_to_video_output(video_url)) - - -def get_base_inputs_types() -> list[IO.Input]: - """Get the base required inputs types common to all Pika nodes.""" - return [ - IO.String.Input("prompt_text", multiline=True), - IO.String.Input("negative_prompt", multiline=True), - IO.Int.Input("seed", min=0, max=0xFFFFFFFF, control_after_generate=True), - IO.Combo.Input("resolution", options=["1080p", "720p"], default="1080p"), - IO.Combo.Input("duration", options=[5, 10], default=5), - ] - - -class PikaImageToVideo(IO.ComfyNode): - """Pika 2.2 Image to Video Node.""" - - @classmethod - def define_schema(cls) -> IO.Schema: - return IO.Schema( - node_id="PikaImageToVideoNode2_2", - display_name="Pika Image to Video", - description="Sends an image and prompt to the Pika API v2.2 to generate a video.", - category="api node/video/Pika", - inputs=[ - IO.Image.Input("image", tooltip="The image to convert to video"), - *get_base_inputs_types(), - ], - 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, - ) - - @classmethod - async def execute( - cls, - image: torch.Tensor, - prompt_text: str, - negative_prompt: str, - seed: int, - resolution: str, - duration: int, - ) -> IO.NodeOutput: - image_bytes_io = tensor_to_bytesio(image) - pika_files = {"image": ("image.png", image_bytes_io, "image/png")} - pika_request_data = pika_defs.PikaBodyGenerate22I2vGenerate22I2vPost( - promptText=prompt_text, - negativePrompt=negative_prompt, - seed=seed, - resolution=resolution, - duration=duration, - ) - initial_operation = await sync_op( - cls, - ApiEndpoint(path=PATH_IMAGE_TO_VIDEO, method="POST"), - response_model=pika_defs.PikaGenerateResponse, - data=pika_request_data, - files=pika_files, - content_type="multipart/form-data", - ) - return await execute_task(initial_operation.video_id, cls) - - -class PikaTextToVideoNode(IO.ComfyNode): - """Pika Text2Video v2.2 Node.""" - - @classmethod - def define_schema(cls) -> IO.Schema: - return IO.Schema( - node_id="PikaTextToVideoNode2_2", - display_name="Pika Text to Video", - description="Sends a text prompt to the Pika API v2.2 to generate a video.", - category="api node/video/Pika", - inputs=[ - *get_base_inputs_types(), - IO.Float.Input( - "aspect_ratio", - step=0.001, - min=0.4, - max=2.5, - default=1.7777777777777777, - tooltip="Aspect ratio (width / height)", - ) - ], - 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, - ) - - @classmethod - async def execute( - cls, - prompt_text: str, - negative_prompt: str, - seed: int, - resolution: str, - duration: int, - aspect_ratio: float, - ) -> IO.NodeOutput: - initial_operation = await sync_op( - cls, - ApiEndpoint(path=PATH_TEXT_TO_VIDEO, method="POST"), - response_model=pika_defs.PikaGenerateResponse, - data=pika_defs.PikaBodyGenerate22T2vGenerate22T2vPost( - promptText=prompt_text, - negativePrompt=negative_prompt, - seed=seed, - resolution=resolution, - duration=duration, - aspectRatio=aspect_ratio, - ), - content_type="application/x-www-form-urlencoded", - ) - return await execute_task(initial_operation.video_id, cls) - - -class PikaScenes(IO.ComfyNode): - """PikaScenes v2.2 Node.""" - - @classmethod - def define_schema(cls) -> IO.Schema: - return IO.Schema( - node_id="PikaScenesV2_2", - display_name="Pika Scenes (Video Image Composition)", - description="Combine your images to create a video with the objects in them. Upload multiple images as ingredients and generate a high-quality video that incorporates all of them.", - category="api node/video/Pika", - inputs=[ - *get_base_inputs_types(), - IO.Combo.Input( - "ingredients_mode", - options=["creative", "precise"], - default="creative", - ), - IO.Float.Input( - "aspect_ratio", - step=0.001, - min=0.4, - max=2.5, - default=1.7777777777777777, - tooltip="Aspect ratio (width / height)", - ), - IO.Image.Input( - "image_ingredient_1", - optional=True, - tooltip="Image that will be used as ingredient to create a video.", - ), - IO.Image.Input( - "image_ingredient_2", - optional=True, - tooltip="Image that will be used as ingredient to create a video.", - ), - IO.Image.Input( - "image_ingredient_3", - optional=True, - tooltip="Image that will be used as ingredient to create a video.", - ), - IO.Image.Input( - "image_ingredient_4", - optional=True, - tooltip="Image that will be used as ingredient to create a video.", - ), - IO.Image.Input( - "image_ingredient_5", - optional=True, - tooltip="Image that will be used as ingredient to create a video.", - ), - ], - 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, - ) - - @classmethod - async def execute( - cls, - prompt_text: str, - negative_prompt: str, - seed: int, - resolution: str, - duration: int, - ingredients_mode: str, - aspect_ratio: float, - image_ingredient_1: Optional[torch.Tensor] = None, - image_ingredient_2: Optional[torch.Tensor] = None, - image_ingredient_3: Optional[torch.Tensor] = None, - image_ingredient_4: Optional[torch.Tensor] = None, - image_ingredient_5: Optional[torch.Tensor] = None, - ) -> IO.NodeOutput: - all_image_bytes_io = [] - for image in [ - image_ingredient_1, - image_ingredient_2, - image_ingredient_3, - image_ingredient_4, - image_ingredient_5, - ]: - if image is not None: - all_image_bytes_io.append(tensor_to_bytesio(image)) - - pika_files = [ - ("images", (f"image_{i}.png", image_bytes_io, "image/png")) - for i, image_bytes_io in enumerate(all_image_bytes_io) - ] - - pika_request_data = pika_defs.PikaBodyGenerate22C2vGenerate22PikascenesPost( - ingredientsMode=ingredients_mode, - promptText=prompt_text, - negativePrompt=negative_prompt, - seed=seed, - resolution=resolution, - duration=duration, - aspectRatio=aspect_ratio, - ) - initial_operation = await sync_op( - cls, - ApiEndpoint(path=PATH_PIKASCENES, method="POST"), - response_model=pika_defs.PikaGenerateResponse, - data=pika_request_data, - files=pika_files, - content_type="multipart/form-data", - ) - - return await execute_task(initial_operation.video_id, cls) - - -class PikAdditionsNode(IO.ComfyNode): - """Pika Pikadditions Node. Add an image into a video.""" - - @classmethod - def define_schema(cls) -> IO.Schema: - return IO.Schema( - node_id="Pikadditions", - display_name="Pikadditions (Video Object Insertion)", - description="Add any object or image into your video. Upload a video and specify what you'd like to add to create a seamlessly integrated result.", - category="api node/video/Pika", - inputs=[ - IO.Video.Input("video", tooltip="The video to add an image to."), - IO.Image.Input("image", tooltip="The image to add to the video."), - IO.String.Input("prompt_text", multiline=True), - IO.String.Input("negative_prompt", multiline=True), - IO.Int.Input( - "seed", - min=0, - max=0xFFFFFFFF, - control_after_generate=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, - ) - - @classmethod - async def execute( - cls, - video: VideoInput, - image: torch.Tensor, - prompt_text: str, - negative_prompt: str, - seed: int, - ) -> IO.NodeOutput: - video_bytes_io = BytesIO() - video.save_to(video_bytes_io, format=VideoContainer.MP4, codec=VideoCodec.H264) - video_bytes_io.seek(0) - - image_bytes_io = tensor_to_bytesio(image) - pika_files = { - "video": ("video.mp4", video_bytes_io, "video/mp4"), - "image": ("image.png", image_bytes_io, "image/png"), - } - pika_request_data = pika_defs.PikaBodyGeneratePikadditionsGeneratePikadditionsPost( - promptText=prompt_text, - negativePrompt=negative_prompt, - seed=seed, - ) - initial_operation = await sync_op( - cls, - ApiEndpoint(path=PATH_PIKADDITIONS, method="POST"), - response_model=pika_defs.PikaGenerateResponse, - data=pika_request_data, - files=pika_files, - content_type="multipart/form-data", - ) - - return await execute_task(initial_operation.video_id, cls) - - -class PikaSwapsNode(IO.ComfyNode): - """Pika Pikaswaps Node.""" - - @classmethod - def define_schema(cls) -> IO.Schema: - return IO.Schema( - node_id="Pikaswaps", - display_name="Pika Swaps (Video Object Replacement)", - description="Swap out any object or region of your video with a new image or object. Define areas to replace either with a mask or coordinates.", - category="api node/video/Pika", - inputs=[ - IO.Video.Input("video", tooltip="The video to swap an object in."), - IO.Image.Input( - "image", - tooltip="The image used to replace the masked object in the video.", - optional=True, - ), - IO.Mask.Input( - "mask", - tooltip="Use the mask to define areas in the video to replace.", - optional=True, - ), - IO.String.Input("prompt_text", multiline=True, optional=True), - IO.String.Input("negative_prompt", multiline=True, optional=True), - IO.Int.Input("seed", min=0, max=0xFFFFFFFF, control_after_generate=True, optional=True), - IO.String.Input( - "region_to_modify", - multiline=True, - optional=True, - tooltip="Plaintext description of the object / region to modify.", - ), - ], - 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, - ) - - @classmethod - async def execute( - cls, - video: VideoInput, - image: Optional[torch.Tensor] = None, - mask: Optional[torch.Tensor] = None, - prompt_text: str = "", - negative_prompt: str = "", - seed: int = 0, - region_to_modify: str = "", - ) -> IO.NodeOutput: - video_bytes_io = BytesIO() - video.save_to(video_bytes_io, format=VideoContainer.MP4, codec=VideoCodec.H264) - video_bytes_io.seek(0) - pika_files = { - "video": ("video.mp4", video_bytes_io, "video/mp4"), - } - if mask is not None: - pika_files["modifyRegionMask"] = ("mask.png", tensor_to_bytesio(mask), "image/png") - if image is not None: - pika_files["image"] = ("image.png", tensor_to_bytesio(image), "image/png") - - pika_request_data = pika_defs.PikaBodyGeneratePikaswapsGeneratePikaswapsPost( - promptText=prompt_text, - negativePrompt=negative_prompt, - seed=seed, - modifyRegionRoi=region_to_modify if region_to_modify else None, - ) - initial_operation = await sync_op( - cls, - ApiEndpoint(path=PATH_PIKASWAPS, method="POST"), - response_model=pika_defs.PikaGenerateResponse, - data=pika_request_data, - files=pika_files, - content_type="multipart/form-data", - ) - return await execute_task(initial_operation.video_id, cls) - - -class PikaffectsNode(IO.ComfyNode): - """Pika Pikaffects Node.""" - - @classmethod - def define_schema(cls) -> IO.Schema: - return IO.Schema( - node_id="Pikaffects", - display_name="Pikaffects (Video Effects)", - description="Generate a video with a specific Pikaffect. Supported Pikaffects: Cake-ify, Crumble, Crush, Decapitate, Deflate, Dissolve, Explode, Eye-pop, Inflate, Levitate, Melt, Peel, Poke, Squish, Ta-da, Tear", - category="api node/video/Pika", - inputs=[ - IO.Image.Input("image", tooltip="The reference image to apply the Pikaffect to."), - IO.Combo.Input( - "pikaffect", options=pika_defs.Pikaffect, default="Cake-ify" - ), - IO.String.Input("prompt_text", multiline=True), - IO.String.Input("negative_prompt", multiline=True), - IO.Int.Input("seed", min=0, max=0xFFFFFFFF, control_after_generate=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, - ) - - @classmethod - async def execute( - cls, - image: torch.Tensor, - pikaffect: str, - prompt_text: str, - negative_prompt: str, - seed: int, - ) -> IO.NodeOutput: - initial_operation = await sync_op( - cls, - ApiEndpoint(path=PATH_PIKAFFECTS, method="POST"), - response_model=pika_defs.PikaGenerateResponse, - data=pika_defs.PikaBodyGeneratePikaffectsGeneratePikaffectsPost( - pikaffect=pikaffect, - promptText=prompt_text, - negativePrompt=negative_prompt, - seed=seed, - ), - files={"image": ("image.png", tensor_to_bytesio(image), "image/png")}, - content_type="multipart/form-data", - ) - return await execute_task(initial_operation.video_id, cls) - - -class PikaStartEndFrameNode(IO.ComfyNode): - """PikaFrames v2.2 Node.""" - - @classmethod - def define_schema(cls) -> IO.Schema: - return IO.Schema( - node_id="PikaStartEndFrameNode2_2", - display_name="Pika Start and End Frame to Video", - description="Generate a video by combining your first and last frame. Upload two images to define the start and end points, and let the AI create a smooth transition between them.", - category="api node/video/Pika", - inputs=[ - IO.Image.Input("image_start", tooltip="The first image to combine."), - IO.Image.Input("image_end", tooltip="The last image to combine."), - *get_base_inputs_types(), - ], - 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, - ) - - @classmethod - async def execute( - cls, - image_start: torch.Tensor, - image_end: torch.Tensor, - prompt_text: str, - negative_prompt: str, - seed: int, - resolution: str, - duration: int, - ) -> IO.NodeOutput: - validate_string(prompt_text, field_name="prompt_text", min_length=1) - pika_files = [ - ("keyFrames", ("image_start.png", tensor_to_bytesio(image_start), "image/png")), - ("keyFrames", ("image_end.png", tensor_to_bytesio(image_end), "image/png")), - ] - initial_operation = await sync_op( - cls, - ApiEndpoint(path=PATH_PIKAFRAMES, method="POST"), - response_model=pika_defs.PikaGenerateResponse, - data=pika_defs.PikaBodyGenerate22KeyframeGenerate22PikaframesPost( - promptText=prompt_text, - negativePrompt=negative_prompt, - seed=seed, - resolution=resolution, - duration=duration, - ), - files=pika_files, - content_type="multipart/form-data", - ) - return await execute_task(initial_operation.video_id, cls) - - -class PikaApiNodesExtension(ComfyExtension): - @override - async def get_node_list(self) -> list[type[IO.ComfyNode]]: - return [ - PikaImageToVideo, - PikaTextToVideoNode, - PikaScenes, - PikAdditionsNode, - PikaSwapsNode, - PikaffectsNode, - PikaStartEndFrameNode, - ] - - -async def comfy_entrypoint() -> PikaApiNodesExtension: - return PikaApiNodesExtension() diff --git a/comfy_api_nodes/nodes_runway.py b/comfy_api_nodes/nodes_runway.py index 2fdafbbfe..3c55039c9 100644 --- a/comfy_api_nodes/nodes_runway.py +++ b/comfy_api_nodes/nodes_runway.py @@ -11,12 +11,11 @@ User Guides: """ -from typing import Union, Optional -from typing_extensions import override from enum import Enum -import torch +from typing_extensions import override +from comfy_api.latest import IO, ComfyExtension, Input, InputImpl from comfy_api_nodes.apis import ( RunwayImageToVideoRequest, RunwayImageToVideoResponse, @@ -44,8 +43,6 @@ from comfy_api_nodes.util import ( sync_op, poll_op, ) -from comfy_api.input_impl import VideoFromFile -from comfy_api.latest import ComfyExtension, IO PATH_IMAGE_TO_VIDEO = "/proxy/runway/image_to_video" PATH_TEXT_TO_IMAGE = "/proxy/runway/text_to_image" @@ -80,7 +77,7 @@ class RunwayGen3aAspectRatio(str, Enum): field_1280_768 = "1280:768" -def get_video_url_from_task_status(response: TaskStatusResponse) -> Union[str, None]: +def get_video_url_from_task_status(response: TaskStatusResponse) -> str | None: """Returns the video URL from the task status response if it exists.""" if hasattr(response, "output") and len(response.output) > 0: return response.output[0] @@ -89,13 +86,13 @@ def get_video_url_from_task_status(response: TaskStatusResponse) -> Union[str, N def extract_progress_from_task_status( response: TaskStatusResponse, -) -> Union[float, None]: +) -> float | None: if hasattr(response, "progress") and response.progress is not None: return response.progress * 100 return None -def get_image_url_from_task_status(response: TaskStatusResponse) -> Union[str, None]: +def get_image_url_from_task_status(response: TaskStatusResponse) -> str | None: """Returns the image URL from the task status response if it exists.""" if hasattr(response, "output") and len(response.output) > 0: return response.output[0] @@ -103,7 +100,7 @@ def get_image_url_from_task_status(response: TaskStatusResponse) -> Union[str, N async def get_response( - cls: type[IO.ComfyNode], task_id: str, estimated_duration: Optional[int] = None + cls: type[IO.ComfyNode], task_id: str, estimated_duration: int | None = None ) -> TaskStatusResponse: """Poll the task status until it is finished then get the response.""" return await poll_op( @@ -119,8 +116,8 @@ async def get_response( async def generate_video( cls: type[IO.ComfyNode], request: RunwayImageToVideoRequest, - estimated_duration: Optional[int] = None, -) -> VideoFromFile: + estimated_duration: int | None = None, +) -> InputImpl.VideoFromFile: initial_response = await sync_op( cls, endpoint=ApiEndpoint(path=PATH_IMAGE_TO_VIDEO, method="POST"), @@ -193,7 +190,7 @@ class RunwayImageToVideoNodeGen3a(IO.ComfyNode): async def execute( cls, prompt: str, - start_frame: torch.Tensor, + start_frame: Input.Image, duration: str, ratio: str, seed: int, @@ -283,7 +280,7 @@ class RunwayImageToVideoNodeGen4(IO.ComfyNode): async def execute( cls, prompt: str, - start_frame: torch.Tensor, + start_frame: Input.Image, duration: str, ratio: str, seed: int, @@ -381,8 +378,8 @@ class RunwayFirstLastFrameNode(IO.ComfyNode): async def execute( cls, prompt: str, - start_frame: torch.Tensor, - end_frame: torch.Tensor, + start_frame: Input.Image, + end_frame: Input.Image, duration: str, ratio: str, seed: int, @@ -467,7 +464,7 @@ class RunwayTextToImageNode(IO.ComfyNode): cls, prompt: str, ratio: str, - reference_image: Optional[torch.Tensor] = None, + reference_image: Input.Image | None = None, ) -> IO.NodeOutput: validate_string(prompt, min_length=1) diff --git a/comfy_api_nodes/nodes_topaz.py b/comfy_api_nodes/nodes_topaz.py new file mode 100644 index 000000000..b04575ad8 --- /dev/null +++ b/comfy_api_nodes/nodes_topaz.py @@ -0,0 +1,439 @@ +import builtins +from io import BytesIO + +import aiohttp +import torch +from typing_extensions import override + +from comfy_api.latest import IO, ComfyExtension, Input +from comfy_api_nodes.apis import topaz_api +from comfy_api_nodes.util import ( + ApiEndpoint, + download_url_to_image_tensor, + download_url_to_video_output, + get_fs_object_size, + get_number_of_images, + poll_op, + sync_op, + upload_images_to_comfyapi, + validate_container_format_is_mp4, +) + +UPSCALER_MODELS_MAP = { + "Starlight (Astra) Fast": "slf-1", + "Starlight (Astra) Creative": "slc-1", +} + + +class TopazImageEnhance(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="TopazImageEnhance", + display_name="Topaz Image Enhance", + category="api node/image/Topaz", + description="Industry-standard upscaling and image enhancement.", + inputs=[ + IO.Combo.Input("model", options=["Reimagine"]), + IO.Image.Input("image"), + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Optional text prompt for creative upscaling guidance.", + optional=True, + ), + IO.Combo.Input( + "subject_detection", + options=["All", "Foreground", "Background"], + optional=True, + ), + IO.Boolean.Input( + "face_enhancement", + default=True, + optional=True, + tooltip="Enhance faces (if present) during processing.", + ), + IO.Float.Input( + "face_enhancement_creativity", + default=0.0, + min=0.0, + max=1.0, + step=0.01, + display_mode=IO.NumberDisplay.number, + optional=True, + tooltip="Set the creativity level for face enhancement.", + ), + IO.Float.Input( + "face_enhancement_strength", + default=1.0, + min=0.0, + max=1.0, + step=0.01, + display_mode=IO.NumberDisplay.number, + optional=True, + tooltip="Controls how sharp enhanced faces are relative to the background.", + ), + IO.Boolean.Input( + "crop_to_fill", + default=False, + optional=True, + tooltip="By default, the image is letterboxed when the output aspect ratio differs. " + "Enable to crop the image to fill the output dimensions.", + ), + IO.Int.Input( + "output_width", + default=0, + min=0, + max=32000, + step=1, + display_mode=IO.NumberDisplay.number, + optional=True, + tooltip="Zero value means to calculate automatically (usually it will be original size or output_height if specified).", + ), + IO.Int.Input( + "output_height", + default=0, + min=0, + max=32000, + step=1, + display_mode=IO.NumberDisplay.number, + optional=True, + tooltip="Zero value means to output in the same height as original or output width.", + ), + IO.Int.Input( + "creativity", + default=3, + min=1, + max=9, + step=1, + display_mode=IO.NumberDisplay.slider, + optional=True, + ), + IO.Boolean.Input( + "face_preservation", + default=True, + optional=True, + tooltip="Preserve subjects' facial identity.", + ), + IO.Boolean.Input( + "color_preservation", + default=True, + optional=True, + tooltip="Preserve the original colors.", + ), + ], + 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, + ) + + @classmethod + async def execute( + cls, + model: str, + image: torch.Tensor, + prompt: str = "", + subject_detection: str = "All", + face_enhancement: bool = True, + face_enhancement_creativity: float = 1.0, + face_enhancement_strength: float = 0.8, + crop_to_fill: bool = False, + output_width: int = 0, + output_height: int = 0, + creativity: int = 3, + face_preservation: bool = True, + color_preservation: bool = True, + ) -> IO.NodeOutput: + if get_number_of_images(image) != 1: + raise ValueError("Only one input image is supported.") + download_url = await upload_images_to_comfyapi(cls, image, max_images=1, mime_type="image/png") + initial_response = await sync_op( + cls, + ApiEndpoint(path="/proxy/topaz/image/v1/enhance-gen/async", method="POST"), + response_model=topaz_api.ImageAsyncTaskResponse, + data=topaz_api.ImageEnhanceRequest( + model=model, + prompt=prompt, + subject_detection=subject_detection, + face_enhancement=face_enhancement, + face_enhancement_creativity=face_enhancement_creativity, + face_enhancement_strength=face_enhancement_strength, + crop_to_fill=crop_to_fill, + output_width=output_width if output_width else None, + output_height=output_height if output_height else None, + creativity=creativity, + face_preservation=str(face_preservation).lower(), + color_preservation=str(color_preservation).lower(), + source_url=download_url[0], + output_format="png", + ), + content_type="multipart/form-data", + ) + + await poll_op( + cls, + poll_endpoint=ApiEndpoint(path=f"/proxy/topaz/image/v1/status/{initial_response.process_id}"), + response_model=topaz_api.ImageStatusResponse, + status_extractor=lambda x: x.status, + progress_extractor=lambda x: getattr(x, "progress", 0), + price_extractor=lambda x: x.credits * 0.08, + poll_interval=8.0, + max_poll_attempts=160, + estimated_duration=60, + ) + + results = await sync_op( + cls, + ApiEndpoint(path=f"/proxy/topaz/image/v1/download/{initial_response.process_id}"), + response_model=topaz_api.ImageDownloadResponse, + monitor_progress=False, + ) + return IO.NodeOutput(await download_url_to_image_tensor(results.download_url)) + + +class TopazVideoEnhance(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="TopazVideoEnhance", + 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.Boolean.Input("upscaler_enabled", default=True), + IO.Combo.Input("upscaler_model", options=list(UPSCALER_MODELS_MAP.keys())), + IO.Combo.Input("upscaler_resolution", options=["FullHD (1080p)", "4K (2160p)"]), + IO.Combo.Input( + "upscaler_creativity", + options=["low", "middle", "high"], + default="low", + tooltip="Creativity level (applies only to Starlight (Astra) Creative).", + optional=True, + ), + IO.Boolean.Input("interpolation_enabled", default=False, optional=True), + IO.Combo.Input("interpolation_model", options=["apo-8"], default="apo-8", optional=True), + 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.", + optional=True, + ), + IO.Int.Input( + "interpolation_frame_rate", + default=60, + min=15, + max=240, + display_mode=IO.NumberDisplay.number, + tooltip="Output frame rate.", + optional=True, + ), + IO.Boolean.Input( + "interpolation_duplicate", + default=False, + tooltip="Analyze the input for duplicate frames and remove them.", + optional=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.", + optional=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, + ) + + @classmethod + async def execute( + cls, + video: Input.Video, + upscaler_enabled: bool, + upscaler_model: str, + upscaler_resolution: str, + upscaler_creativity: str = "low", + interpolation_enabled: bool = False, + interpolation_model: str = "apo-8", + interpolation_slowmo: int = 1, + interpolation_frame_rate: int = 60, + interpolation_duplicate: bool = False, + interpolation_duplicate_threshold: float = 0.01, + dynamic_compression_level: str = "Low", + ) -> IO.NodeOutput: + if upscaler_enabled is False and interpolation_enabled is False: + 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_enabled: + if "1080p" in 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 + filters.append( + topaz_api.VideoEnhancementFilter( + model=UPSCALER_MODELS_MAP[upscaler_model], + creativity=(upscaler_creativity if UPSCALER_MODELS_MAP[upscaler_model] == "slc-1" else None), + isOptimizedMode=(True if UPSCALER_MODELS_MAP[upscaler_model] == "slc-1" else None), + ), + ) + if interpolation_enabled: + target_frame_rate = interpolation_frame_rate + filters.append( + topaz_api.VideoFrameInterpolationFilter( + model=interpolation_model, + slowmo=interpolation_slowmo, + fps=interpolation_frame_rate, + duplicate=interpolation_duplicate, + duplicate_threshold=interpolation_duplicate_threshold, + ), + ) + initial_res = await sync_op( + cls, + ApiEndpoint(path="/proxy/topaz/video/", method="POST"), + response_model=topaz_api.CreateVideoResponse, + data=topaz_api.CreateVideoRequest( + source=topaz_api.CreateCreateVideoRequestSource( + container="mp4", + size=get_fs_object_size(src_video_stream), + duration=int(duration_sec), + frameCount=video.get_frame_count(), + frameRate=src_frame_rate, + resolution=topaz_api.Resolution(width=src_width, height=src_height), + ), + filters=filters, + output=topaz_api.OutputInformationVideo( + resolution=topaz_api.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=topaz_api.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=topaz_api.VideoCompleteUploadResponse, + data=topaz_api.VideoCompleteUploadRequest( + uploadResults=[ + topaz_api.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=topaz_api.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, + max_poll_attempts=320, + ) + return IO.NodeOutput(await download_url_to_video_output(final_response.download.url)) + + +class TopazExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + TopazImageEnhance, + TopazVideoEnhance, + ] + + +async def comfy_entrypoint() -> TopazExtension: + return TopazExtension() diff --git a/comfy_api_nodes/nodes_tripo.py b/comfy_api_nodes/nodes_tripo.py index 697100ff2..e72f8e96a 100644 --- a/comfy_api_nodes/nodes_tripo.py +++ b/comfy_api_nodes/nodes_tripo.py @@ -102,8 +102,9 @@ class TripoTextToModelNode(IO.ComfyNode): IO.Int.Input("model_seed", default=42, optional=True), IO.Int.Input("texture_seed", default=42, optional=True), IO.Combo.Input("texture_quality", default="standard", options=["standard", "detailed"], optional=True), - IO.Int.Input("face_limit", default=-1, min=-1, max=500000, optional=True), + IO.Int.Input("face_limit", default=-1, min=-1, max=2000000, optional=True), IO.Boolean.Input("quad", default=False, optional=True), + IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True), ], outputs=[ IO.String.Output(display_name="model_file"), @@ -131,6 +132,7 @@ class TripoTextToModelNode(IO.ComfyNode): model_seed: Optional[int] = None, texture_seed: Optional[int] = None, texture_quality: Optional[str] = None, + geometry_quality: Optional[str] = None, face_limit: Optional[int] = None, quad: Optional[bool] = None, ) -> IO.NodeOutput: @@ -153,7 +155,8 @@ class TripoTextToModelNode(IO.ComfyNode): model_seed=model_seed, texture_seed=texture_seed, texture_quality=texture_quality, - face_limit=face_limit, + face_limit=face_limit if face_limit != -1 else None, + geometry_quality=geometry_quality, auto_size=True, quad=quad, ), @@ -194,6 +197,7 @@ class TripoImageToModelNode(IO.ComfyNode): ), IO.Int.Input("face_limit", default=-1, min=-1, max=500000, optional=True), IO.Boolean.Input("quad", default=False, optional=True), + IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True), ], outputs=[ IO.String.Output(display_name="model_file"), @@ -220,6 +224,7 @@ class TripoImageToModelNode(IO.ComfyNode): orientation=None, texture_seed: Optional[int] = None, texture_quality: Optional[str] = None, + geometry_quality: Optional[str] = None, texture_alignment: Optional[str] = None, face_limit: Optional[int] = None, quad: Optional[bool] = None, @@ -246,10 +251,11 @@ class TripoImageToModelNode(IO.ComfyNode): pbr=pbr, model_seed=model_seed, orientation=orientation, + geometry_quality=geometry_quality, texture_alignment=texture_alignment, texture_seed=texture_seed, texture_quality=texture_quality, - face_limit=face_limit, + face_limit=face_limit if face_limit != -1 else None, auto_size=True, quad=quad, ), @@ -295,6 +301,7 @@ class TripoMultiviewToModelNode(IO.ComfyNode): ), IO.Int.Input("face_limit", default=-1, min=-1, max=500000, optional=True), IO.Boolean.Input("quad", default=False, optional=True), + IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True), ], outputs=[ IO.String.Output(display_name="model_file"), @@ -323,6 +330,7 @@ class TripoMultiviewToModelNode(IO.ComfyNode): model_seed: Optional[int] = None, texture_seed: Optional[int] = None, texture_quality: Optional[str] = None, + geometry_quality: Optional[str] = None, texture_alignment: Optional[str] = None, face_limit: Optional[int] = None, quad: Optional[bool] = None, @@ -359,8 +367,9 @@ class TripoMultiviewToModelNode(IO.ComfyNode): model_seed=model_seed, texture_seed=texture_seed, texture_quality=texture_quality, + geometry_quality=geometry_quality, texture_alignment=texture_alignment, - face_limit=face_limit, + face_limit=face_limit if face_limit != -1 else None, quad=quad, ), ) @@ -508,6 +517,8 @@ class TripoRetargetNode(IO.ComfyNode): options=[ "preset:idle", "preset:walk", + "preset:run", + "preset:dive", "preset:climb", "preset:jump", "preset:slash", @@ -515,6 +526,11 @@ class TripoRetargetNode(IO.ComfyNode): "preset:hurt", "preset:fall", "preset:turn", + "preset:quadruped:walk", + "preset:hexapod:walk", + "preset:octopod:walk", + "preset:serpentine:march", + "preset:aquatic:march" ], ), ], @@ -563,7 +579,7 @@ class TripoConversionNode(IO.ComfyNode): "face_limit", default=-1, min=-1, - max=500000, + max=2000000, optional=True, ), IO.Int.Input( @@ -579,6 +595,40 @@ class TripoConversionNode(IO.ComfyNode): default="JPEG", optional=True, ), + IO.Boolean.Input("force_symmetry", default=False, optional=True), + IO.Boolean.Input("flatten_bottom", default=False, optional=True), + IO.Float.Input( + "flatten_bottom_threshold", + default=0.0, + min=0.0, + max=1.0, + optional=True, + ), + IO.Boolean.Input("pivot_to_center_bottom", default=False, optional=True), + IO.Float.Input( + "scale_factor", + default=1.0, + min=0.0, + optional=True, + ), + IO.Boolean.Input("with_animation", default=False, optional=True), + IO.Boolean.Input("pack_uv", default=False, optional=True), + IO.Boolean.Input("bake", default=False, optional=True), + IO.String.Input("part_names", default="", optional=True), # comma-separated list + IO.Combo.Input( + "fbx_preset", + options=["blender", "mixamo", "3dsmax"], + default="blender", + optional=True, + ), + IO.Boolean.Input("export_vertex_colors", default=False, optional=True), + IO.Combo.Input( + "export_orientation", + options=["align_image", "default"], + default="default", + optional=True, + ), + IO.Boolean.Input("animate_in_place", default=False, optional=True), ], outputs=[], hidden=[ @@ -604,12 +654,31 @@ class TripoConversionNode(IO.ComfyNode): original_model_task_id, format: str, quad: bool, + force_symmetry: bool, face_limit: int, + flatten_bottom: bool, + flatten_bottom_threshold: float, texture_size: int, texture_format: str, + pivot_to_center_bottom: bool, + scale_factor: float, + with_animation: bool, + pack_uv: bool, + bake: bool, + part_names: str, + fbx_preset: str, + export_vertex_colors: bool, + export_orientation: str, + animate_in_place: bool, ) -> IO.NodeOutput: if not original_model_task_id: raise RuntimeError("original_model_task_id is required") + + # Parse part_names from comma-separated string to list + part_names_list = None + if part_names and part_names.strip(): + part_names_list = [name.strip() for name in part_names.split(',') if name.strip()] + response = await sync_op( cls, endpoint=ApiEndpoint(path="/proxy/tripo/v2/openapi/task", method="POST"), @@ -618,9 +687,22 @@ class TripoConversionNode(IO.ComfyNode): original_model_task_id=original_model_task_id, format=format, quad=quad if quad else None, + force_symmetry=force_symmetry if force_symmetry else None, face_limit=face_limit if face_limit != -1 else None, + flatten_bottom=flatten_bottom if flatten_bottom else None, + flatten_bottom_threshold=flatten_bottom_threshold if flatten_bottom_threshold != 0.0 else None, texture_size=texture_size if texture_size != 4096 else None, texture_format=texture_format if texture_format != "JPEG" else None, + pivot_to_center_bottom=pivot_to_center_bottom if pivot_to_center_bottom else None, + scale_factor=scale_factor if scale_factor != 1.0 else None, + with_animation=with_animation if with_animation else None, + pack_uv=pack_uv if pack_uv else None, + bake=bake if bake else None, + part_names=part_names_list, + fbx_preset=fbx_preset if fbx_preset != "blender" else None, + export_vertex_colors=export_vertex_colors if export_vertex_colors else None, + export_orientation=export_orientation if export_orientation != "default" else None, + animate_in_place=animate_in_place if animate_in_place else None, ), ) return await poll_until_finished(cls, response, average_duration=30) diff --git a/comfy_api_nodes/nodes_veo2.py b/comfy_api_nodes/nodes_veo2.py index d37e9e9b4..13a6bfd91 100644 --- a/comfy_api_nodes/nodes_veo2.py +++ b/comfy_api_nodes/nodes_veo2.py @@ -3,13 +3,15 @@ from io import BytesIO from typing_extensions import override -from comfy_api.input_impl.video_types import VideoFromFile -from comfy_api.latest import IO, ComfyExtension +from comfy_api.latest import IO, ComfyExtension, Input, InputImpl from comfy_api_nodes.apis.veo_api import ( VeoGenVidPollRequest, VeoGenVidPollResponse, VeoGenVidRequest, VeoGenVidResponse, + VeoRequestInstance, + VeoRequestInstanceImage, + VeoRequestParameters, ) from comfy_api_nodes.util import ( ApiEndpoint, @@ -166,6 +168,8 @@ class VeoVideoGenerationNode(IO.ComfyNode): # Only add generateAudio for Veo 3 models if model.find("veo-2.0") == -1: parameters["generateAudio"] = generate_audio + # force "enhance_prompt" to True for Veo3 models + parameters["enhancePrompt"] = True initial_response = await sync_op( cls, @@ -228,7 +232,7 @@ class VeoVideoGenerationNode(IO.ComfyNode): # Check if video is provided as base64 or URL if hasattr(video, "bytesBase64Encoded") and video.bytesBase64Encoded: - return IO.NodeOutput(VideoFromFile(BytesIO(base64.b64decode(video.bytesBase64Encoded)))) + return IO.NodeOutput(InputImpl.VideoFromFile(BytesIO(base64.b64decode(video.bytesBase64Encoded)))) if hasattr(video, "gcsUri") and video.gcsUri: return IO.NodeOutput(await download_url_to_video_output(video.gcsUri)) @@ -289,7 +293,7 @@ class Veo3VideoGenerationNode(VeoVideoGenerationNode): IO.Boolean.Input( "enhance_prompt", default=True, - tooltip="Whether to enhance the prompt with AI assistance", + tooltip="This parameter is deprecated and ignored.", optional=True, ), IO.Combo.Input( @@ -346,12 +350,163 @@ class Veo3VideoGenerationNode(VeoVideoGenerationNode): ) +class Veo3FirstLastFrameNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="Veo3FirstLastFrameNode", + display_name="Google Veo 3 First-Last-Frame to Video", + category="api node/video/Veo", + description="Generate video using prompt and first and last frames.", + inputs=[ + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Text description of the video", + ), + IO.String.Input( + "negative_prompt", + multiline=True, + default="", + tooltip="Negative text prompt to guide what to avoid in the video", + ), + IO.Combo.Input("resolution", options=["720p", "1080p"]), + IO.Combo.Input( + "aspect_ratio", + options=["16:9", "9:16"], + default="16:9", + tooltip="Aspect ratio of the output video", + ), + IO.Int.Input( + "duration", + default=8, + min=4, + max=8, + step=2, + display_mode=IO.NumberDisplay.slider, + tooltip="Duration of the output video in seconds", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=0xFFFFFFFF, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed for video generation", + ), + IO.Image.Input("first_frame", tooltip="Start frame"), + IO.Image.Input("last_frame", tooltip="End frame"), + IO.Combo.Input( + "model", + options=["veo-3.1-generate", "veo-3.1-fast-generate"], + default="veo-3.1-fast-generate", + ), + IO.Boolean.Input( + "generate_audio", + default=True, + tooltip="Generate audio for the video.", + ), + ], + 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, + ) + + @classmethod + async def execute( + cls, + prompt: str, + negative_prompt: str, + resolution: str, + aspect_ratio: str, + duration: int, + seed: int, + first_frame: Input.Image, + last_frame: Input.Image, + model: str, + generate_audio: bool, + ): + model = MODELS_MAP[model] + initial_response = await sync_op( + cls, + ApiEndpoint(path=f"/proxy/veo/{model}/generate", method="POST"), + response_model=VeoGenVidResponse, + data=VeoGenVidRequest( + instances=[ + VeoRequestInstance( + prompt=prompt, + image=VeoRequestInstanceImage( + bytesBase64Encoded=tensor_to_base64_string(first_frame), mimeType="image/png" + ), + lastFrame=VeoRequestInstanceImage( + bytesBase64Encoded=tensor_to_base64_string(last_frame), mimeType="image/png" + ), + ), + ], + parameters=VeoRequestParameters( + aspectRatio=aspect_ratio, + personGeneration="ALLOW", + durationSeconds=duration, + enhancePrompt=True, # cannot be False for Veo3 + seed=seed, + generateAudio=generate_audio, + negativePrompt=negative_prompt, + resolution=resolution, + ), + ), + ) + poll_response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/veo/{model}/poll", method="POST"), + response_model=VeoGenVidPollResponse, + status_extractor=lambda r: "completed" if r.done else "pending", + data=VeoGenVidPollRequest( + operationName=initial_response.name, + ), + poll_interval=5.0, + estimated_duration=AVERAGE_DURATION_VIDEO_GEN, + ) + + if poll_response.error: + raise Exception(f"Veo API error: {poll_response.error.message} (code: {poll_response.error.code})") + + response = poll_response.response + filtered_count = response.raiMediaFilteredCount + if filtered_count: + reasons = response.raiMediaFilteredReasons or [] + reason_part = f": {reasons[0]}" if reasons else "" + raise Exception( + f"Content blocked by Google's Responsible AI filters{reason_part} " + f"({filtered_count} video{'s' if filtered_count != 1 else ''} filtered)." + ) + + if response.videos: + video = response.videos[0] + if video.bytesBase64Encoded: + return IO.NodeOutput(InputImpl.VideoFromFile(BytesIO(base64.b64decode(video.bytesBase64Encoded)))) + if video.gcsUri: + return IO.NodeOutput(await download_url_to_video_output(video.gcsUri)) + raise Exception("Video returned but no data or URL was provided") + raise Exception("Video generation completed but no video was returned") + + class VeoExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: return [ VeoVideoGenerationNode, Veo3VideoGenerationNode, + Veo3FirstLastFrameNode, ] diff --git a/comfy_api_nodes/nodes_wan.py b/comfy_api_nodes/nodes_wan.py index 2aab3c2ff..3e04786a9 100644 --- a/comfy_api_nodes/nodes_wan.py +++ b/comfy_api_nodes/nodes_wan.py @@ -1,7 +1,5 @@ import re -from typing import Optional -import torch from pydantic import BaseModel, Field from typing_extensions import override @@ -15,32 +13,40 @@ from comfy_api_nodes.util import ( poll_op, sync_op, tensor_to_base64_string, + upload_video_to_comfyapi, validate_audio_duration, + validate_video_duration, ) class Text2ImageInputField(BaseModel): prompt: str = Field(...) - negative_prompt: Optional[str] = Field(None) + negative_prompt: str | None = Field(None) class Image2ImageInputField(BaseModel): prompt: str = Field(...) - negative_prompt: Optional[str] = Field(None) + negative_prompt: str | None = Field(None) images: list[str] = Field(..., min_length=1, max_length=2) class Text2VideoInputField(BaseModel): prompt: str = Field(...) - negative_prompt: Optional[str] = Field(None) - audio_url: Optional[str] = Field(None) + negative_prompt: str | None = Field(None) + audio_url: str | None = Field(None) class Image2VideoInputField(BaseModel): prompt: str = Field(...) - negative_prompt: Optional[str] = Field(None) + negative_prompt: str | None = Field(None) img_url: str = Field(...) - audio_url: Optional[str] = Field(None) + audio_url: str | None = Field(None) + + +class Reference2VideoInputField(BaseModel): + prompt: str = Field(...) + negative_prompt: str | None = Field(None) + reference_video_urls: list[str] = Field(...) class Txt2ImageParametersField(BaseModel): @@ -48,32 +54,42 @@ class Txt2ImageParametersField(BaseModel): n: int = Field(1, description="Number of images to generate.") # we support only value=1 seed: int = Field(..., ge=0, le=2147483647) prompt_extend: bool = Field(True) - watermark: bool = Field(True) + watermark: bool = Field(False) class Image2ImageParametersField(BaseModel): - size: Optional[str] = Field(None) + size: str | None = Field(None) n: int = Field(1, description="Number of images to generate.") # we support only value=1 seed: int = Field(..., ge=0, le=2147483647) - watermark: bool = Field(True) + watermark: bool = Field(False) class Text2VideoParametersField(BaseModel): size: str = Field(...) seed: int = Field(..., ge=0, le=2147483647) - duration: int = Field(5, ge=5, le=10) + duration: int = Field(5, ge=5, le=15) prompt_extend: bool = Field(True) - watermark: bool = Field(True) - audio: bool = Field(False, description="Should be audio generated automatically") + watermark: bool = Field(False) + audio: bool = Field(False, description="Whether to generate audio automatically.") + shot_type: str = Field("single") class Image2VideoParametersField(BaseModel): resolution: str = Field(...) seed: int = Field(..., ge=0, le=2147483647) - duration: int = Field(5, ge=5, le=10) + duration: int = Field(5, ge=5, le=15) prompt_extend: bool = Field(True) - watermark: bool = Field(True) - audio: bool = Field(False, description="Should be audio generated automatically") + watermark: bool = Field(False) + audio: bool = Field(False, description="Whether to generate audio automatically.") + shot_type: str = Field("single") + + +class Reference2VideoParametersField(BaseModel): + size: str = Field(...) + duration: int = Field(5, ge=5, le=15) + shot_type: str = Field("single") + seed: int = Field(..., ge=0, le=2147483647) + watermark: bool = Field(False) class Text2ImageTaskCreationRequest(BaseModel): @@ -100,45 +116,51 @@ class Image2VideoTaskCreationRequest(BaseModel): parameters: Image2VideoParametersField = Field(...) +class Reference2VideoTaskCreationRequest(BaseModel): + model: str = Field(...) + input: Reference2VideoInputField = Field(...) + parameters: Reference2VideoParametersField = Field(...) + + class TaskCreationOutputField(BaseModel): task_id: str = Field(...) task_status: str = Field(...) class TaskCreationResponse(BaseModel): - output: Optional[TaskCreationOutputField] = Field(None) + output: TaskCreationOutputField | None = Field(None) request_id: str = Field(...) - code: Optional[str] = Field(None, description="The error code of the failed request.") - message: Optional[str] = Field(None, description="Details of the failed request.") + code: str | None = Field(None, description="Error code for the failed request.") + message: str | None = Field(None, description="Details about the failed request.") class TaskResult(BaseModel): - url: Optional[str] = Field(None) - code: Optional[str] = Field(None) - message: Optional[str] = Field(None) + url: str | None = Field(None) + code: str | None = Field(None) + message: str | None = Field(None) class ImageTaskStatusOutputField(TaskCreationOutputField): task_id: str = Field(...) task_status: str = Field(...) - results: Optional[list[TaskResult]] = Field(None) + results: list[TaskResult] | None = Field(None) class VideoTaskStatusOutputField(TaskCreationOutputField): task_id: str = Field(...) task_status: str = Field(...) - video_url: Optional[str] = Field(None) - code: Optional[str] = Field(None) - message: Optional[str] = Field(None) + video_url: str | None = Field(None) + code: str | None = Field(None) + message: str | None = Field(None) class ImageTaskStatusResponse(BaseModel): - output: Optional[ImageTaskStatusOutputField] = Field(None) + output: ImageTaskStatusOutputField | None = Field(None) request_id: str = Field(...) class VideoTaskStatusResponse(BaseModel): - output: Optional[VideoTaskStatusOutputField] = Field(None) + output: VideoTaskStatusOutputField | None = Field(None) request_id: str = Field(...) @@ -152,7 +174,7 @@ class WanTextToImageApi(IO.ComfyNode): node_id="WanTextToImageApi", display_name="Wan Text to Image", category="api node/image/Wan", - description="Generates image based on text prompt.", + description="Generates an image based on a text prompt.", inputs=[ IO.Combo.Input( "model", @@ -164,13 +186,13 @@ class WanTextToImageApi(IO.ComfyNode): "prompt", multiline=True, default="", - tooltip="Prompt used to describe the elements and visual features, supports English/Chinese.", + tooltip="Prompt describing the elements and visual features. Supports English and Chinese.", ), IO.String.Input( "negative_prompt", multiline=True, default="", - tooltip="Negative text prompt to guide what to avoid.", + tooltip="Negative prompt describing what to avoid.", optional=True, ), IO.Int.Input( @@ -208,8 +230,8 @@ class WanTextToImageApi(IO.ComfyNode): ), IO.Boolean.Input( "watermark", - default=True, - tooltip='Whether to add an "AI generated" watermark to the result.', + default=False, + tooltip="Whether to add an AI-generated watermark to the result.", optional=True, ), ], @@ -234,7 +256,7 @@ class WanTextToImageApi(IO.ComfyNode): height: int = 1024, seed: int = 0, prompt_extend: bool = True, - watermark: bool = True, + watermark: bool = False, ): initial_response = await sync_op( cls, @@ -252,7 +274,7 @@ class WanTextToImageApi(IO.ComfyNode): ), ) if not initial_response.output: - raise Exception(f"Unknown error occurred: {initial_response.code} - {initial_response.message}") + raise Exception(f"An unknown error occurred: {initial_response.code} - {initial_response.message}") response = await poll_op( cls, ApiEndpoint(path=f"/proxy/wan/api/v1/tasks/{initial_response.output.task_id}"), @@ -272,7 +294,7 @@ class WanImageToImageApi(IO.ComfyNode): display_name="Wan Image to Image", category="api node/image/Wan", description="Generates an image from one or two input images and a text prompt. " - "The output image is currently fixed at 1.6 MP; its aspect ratio matches the input image(s).", + "The output image is currently fixed at 1.6 MP, and its aspect ratio matches the input image(s).", inputs=[ IO.Combo.Input( "model", @@ -282,19 +304,19 @@ class WanImageToImageApi(IO.ComfyNode): ), IO.Image.Input( "image", - tooltip="Single-image editing or multi-image fusion, maximum 2 images.", + tooltip="Single-image editing or multi-image fusion. Maximum 2 images.", ), IO.String.Input( "prompt", multiline=True, default="", - tooltip="Prompt used to describe the elements and visual features, supports English/Chinese.", + tooltip="Prompt describing the elements and visual features. Supports English and Chinese.", ), IO.String.Input( "negative_prompt", multiline=True, default="", - tooltip="Negative text prompt to guide what to avoid.", + tooltip="Negative prompt describing what to avoid.", optional=True, ), # redo this later as an optional combo of recommended resolutions @@ -327,8 +349,8 @@ class WanImageToImageApi(IO.ComfyNode): ), IO.Boolean.Input( "watermark", - default=True, - tooltip='Whether to add an "AI generated" watermark to the result.', + default=False, + tooltip="Whether to add an AI-generated watermark to the result.", optional=True, ), ], @@ -347,17 +369,17 @@ class WanImageToImageApi(IO.ComfyNode): async def execute( cls, model: str, - image: torch.Tensor, + image: Input.Image, prompt: str, negative_prompt: str = "", # width: int = 1024, # height: int = 1024, seed: int = 0, - watermark: bool = True, + watermark: bool = False, ): n_images = get_number_of_images(image) if n_images not in (1, 2): - raise ValueError(f"Expected 1 or 2 input images, got {n_images}.") + raise ValueError(f"Expected 1 or 2 input images, but got {n_images}.") images = [] for i in image: images.append("data:image/png;base64," + tensor_to_base64_string(i, total_pixels=4096 * 4096)) @@ -376,7 +398,7 @@ class WanImageToImageApi(IO.ComfyNode): ), ) if not initial_response.output: - raise Exception(f"Unknown error occurred: {initial_response.code} - {initial_response.message}") + raise Exception(f"An unknown error occurred: {initial_response.code} - {initial_response.message}") response = await poll_op( cls, ApiEndpoint(path=f"/proxy/wan/api/v1/tasks/{initial_response.output.task_id}"), @@ -395,25 +417,25 @@ class WanTextToVideoApi(IO.ComfyNode): node_id="WanTextToVideoApi", display_name="Wan Text to Video", category="api node/video/Wan", - description="Generates video based on text prompt.", + description="Generates a video based on a text prompt.", inputs=[ IO.Combo.Input( "model", - options=["wan2.5-t2v-preview"], - default="wan2.5-t2v-preview", + options=["wan2.5-t2v-preview", "wan2.6-t2v"], + default="wan2.6-t2v", tooltip="Model to use.", ), IO.String.Input( "prompt", multiline=True, default="", - tooltip="Prompt used to describe the elements and visual features, supports English/Chinese.", + tooltip="Prompt describing the elements and visual features. Supports English and Chinese.", ), IO.String.Input( "negative_prompt", multiline=True, default="", - tooltip="Negative text prompt to guide what to avoid.", + tooltip="Negative prompt describing what to avoid.", optional=True, ), IO.Combo.Input( @@ -433,23 +455,23 @@ class WanTextToVideoApi(IO.ComfyNode): "1080p: 4:3 (1632x1248)", "1080p: 3:4 (1248x1632)", ], - default="480p: 1:1 (624x624)", + default="720p: 1:1 (960x960)", optional=True, ), IO.Int.Input( "duration", default=5, min=5, - max=10, + max=15, step=5, display_mode=IO.NumberDisplay.number, - tooltip="Available durations: 5 and 10 seconds", + tooltip="A 15-second duration is available only for the Wan 2.6 model.", optional=True, ), IO.Audio.Input( "audio", optional=True, - tooltip="Audio must contain a clear, loud voice, without extraneous noise, background music.", + tooltip="Audio must contain a clear, loud voice, without extraneous noise or background music.", ), IO.Int.Input( "seed", @@ -466,7 +488,7 @@ class WanTextToVideoApi(IO.ComfyNode): "generate_audio", default=False, optional=True, - tooltip="If there is no audio input, generate audio automatically.", + tooltip="If no audio input is provided, generate audio automatically.", ), IO.Boolean.Input( "prompt_extend", @@ -476,8 +498,16 @@ class WanTextToVideoApi(IO.ComfyNode): ), IO.Boolean.Input( "watermark", - default=True, - tooltip='Whether to add an "AI generated" watermark to the result.', + default=False, + tooltip="Whether to add an AI-generated watermark to the result.", + optional=True, + ), + IO.Combo.Input( + "shot_type", + options=["single", "multi"], + tooltip="Specifies the shot type for the generated video, that is, whether the video is a " + "single continuous shot or multiple shots with cuts. " + "This parameter takes effect only when prompt_extend is True.", optional=True, ), ], @@ -498,14 +528,19 @@ class WanTextToVideoApi(IO.ComfyNode): model: str, prompt: str, negative_prompt: str = "", - size: str = "480p: 1:1 (624x624)", + size: str = "720p: 1:1 (960x960)", duration: int = 5, - audio: Optional[Input.Audio] = None, + audio: Input.Audio | None = None, seed: int = 0, generate_audio: bool = False, prompt_extend: bool = True, - watermark: bool = True, + watermark: bool = False, + shot_type: str = "single", ): + if "480p" in size and model == "wan2.6-t2v": + raise ValueError("The Wan 2.6 model does not support 480p.") + if duration == 15 and model == "wan2.5-t2v-preview": + raise ValueError("A 15-second duration is supported only by the Wan 2.6 model.") width, height = RES_IN_PARENS.search(size).groups() audio_url = None if audio is not None: @@ -526,11 +561,12 @@ class WanTextToVideoApi(IO.ComfyNode): audio=generate_audio, prompt_extend=prompt_extend, watermark=watermark, + shot_type=shot_type, ), ), ) if not initial_response.output: - raise Exception(f"Unknown error occurred: {initial_response.code} - {initial_response.message}") + raise Exception(f"An unknown error occurred: {initial_response.code} - {initial_response.message}") response = await poll_op( cls, ApiEndpoint(path=f"/proxy/wan/api/v1/tasks/{initial_response.output.task_id}"), @@ -549,12 +585,12 @@ class WanImageToVideoApi(IO.ComfyNode): node_id="WanImageToVideoApi", display_name="Wan Image to Video", category="api node/video/Wan", - description="Generates video based on the first frame and text prompt.", + description="Generates a video from the first frame and a text prompt.", inputs=[ IO.Combo.Input( "model", - options=["wan2.5-i2v-preview"], - default="wan2.5-i2v-preview", + options=["wan2.5-i2v-preview", "wan2.6-i2v"], + default="wan2.6-i2v", tooltip="Model to use.", ), IO.Image.Input( @@ -564,13 +600,13 @@ class WanImageToVideoApi(IO.ComfyNode): "prompt", multiline=True, default="", - tooltip="Prompt used to describe the elements and visual features, supports English/Chinese.", + tooltip="Prompt describing the elements and visual features. Supports English and Chinese.", ), IO.String.Input( "negative_prompt", multiline=True, default="", - tooltip="Negative text prompt to guide what to avoid.", + tooltip="Negative prompt describing what to avoid.", optional=True, ), IO.Combo.Input( @@ -580,23 +616,23 @@ class WanImageToVideoApi(IO.ComfyNode): "720P", "1080P", ], - default="480P", + default="720P", optional=True, ), IO.Int.Input( "duration", default=5, min=5, - max=10, + max=15, step=5, display_mode=IO.NumberDisplay.number, - tooltip="Available durations: 5 and 10 seconds", + tooltip="Duration 15 available only for WAN2.6 model.", optional=True, ), IO.Audio.Input( "audio", optional=True, - tooltip="Audio must contain a clear, loud voice, without extraneous noise, background music.", + tooltip="Audio must contain a clear, loud voice, without extraneous noise or background music.", ), IO.Int.Input( "seed", @@ -613,7 +649,7 @@ class WanImageToVideoApi(IO.ComfyNode): "generate_audio", default=False, optional=True, - tooltip="If there is no audio input, generate audio automatically.", + tooltip="If no audio input is provided, generate audio automatically.", ), IO.Boolean.Input( "prompt_extend", @@ -623,8 +659,16 @@ class WanImageToVideoApi(IO.ComfyNode): ), IO.Boolean.Input( "watermark", - default=True, - tooltip='Whether to add an "AI generated" watermark to the result.', + default=False, + tooltip="Whether to add an AI-generated watermark to the result.", + optional=True, + ), + IO.Combo.Input( + "shot_type", + options=["single", "multi"], + tooltip="Specifies the shot type for the generated video, that is, whether the video is a " + "single continuous shot or multiple shots with cuts. " + "This parameter takes effect only when prompt_extend is True.", optional=True, ), ], @@ -643,19 +687,24 @@ class WanImageToVideoApi(IO.ComfyNode): async def execute( cls, model: str, - image: torch.Tensor, + image: Input.Image, prompt: str, negative_prompt: str = "", - resolution: str = "480P", + resolution: str = "720P", duration: int = 5, - audio: Optional[Input.Audio] = None, + audio: Input.Audio | None = None, seed: int = 0, generate_audio: bool = False, prompt_extend: bool = True, - watermark: bool = True, + watermark: bool = False, + shot_type: str = "single", ): if get_number_of_images(image) != 1: raise ValueError("Exactly one input image is required.") + if "480P" in resolution and model == "wan2.6-i2v": + raise ValueError("The Wan 2.6 model does not support 480P.") + if duration == 15 and model == "wan2.5-i2v-preview": + raise ValueError("A 15-second duration is supported only by the Wan 2.6 model.") image_url = "data:image/png;base64," + tensor_to_base64_string(image, total_pixels=2000 * 2000) audio_url = None if audio is not None: @@ -677,11 +726,12 @@ class WanImageToVideoApi(IO.ComfyNode): audio=generate_audio, prompt_extend=prompt_extend, watermark=watermark, + shot_type=shot_type, ), ), ) if not initial_response.output: - raise Exception(f"Unknown error occurred: {initial_response.code} - {initial_response.message}") + raise Exception(f"An unknown error occurred: {initial_response.code} - {initial_response.message}") response = await poll_op( cls, ApiEndpoint(path=f"/proxy/wan/api/v1/tasks/{initial_response.output.task_id}"), @@ -693,6 +743,143 @@ class WanImageToVideoApi(IO.ComfyNode): return IO.NodeOutput(await download_url_to_video_output(response.output.video_url)) +class WanReferenceVideoApi(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="WanReferenceVideoApi", + display_name="Wan Reference to Video", + category="api node/video/Wan", + description="Use the character and voice from input videos, combined with a prompt, " + "to generate a new video that maintains character consistency.", + inputs=[ + IO.Combo.Input("model", options=["wan2.6-r2v"]), + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Prompt describing the elements and visual features. Supports English and Chinese. " + "Use identifiers such as `character1` and `character2` to refer to the reference characters.", + ), + IO.String.Input( + "negative_prompt", + multiline=True, + default="", + tooltip="Negative prompt describing what to avoid.", + ), + IO.Autogrow.Input( + "reference_videos", + template=IO.Autogrow.TemplateNames( + IO.Video.Input("reference_video"), + names=["character1", "character2", "character3"], + min=1, + ), + ), + IO.Combo.Input( + "size", + options=[ + "720p: 1:1 (960x960)", + "720p: 16:9 (1280x720)", + "720p: 9:16 (720x1280)", + "720p: 4:3 (1088x832)", + "720p: 3:4 (832x1088)", + "1080p: 1:1 (1440x1440)", + "1080p: 16:9 (1920x1080)", + "1080p: 9:16 (1080x1920)", + "1080p: 4:3 (1632x1248)", + "1080p: 3:4 (1248x1632)", + ], + ), + IO.Int.Input( + "duration", + default=5, + min=5, + max=10, + step=5, + display_mode=IO.NumberDisplay.slider, + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + ), + IO.Combo.Input( + "shot_type", + options=["single", "multi"], + tooltip="Specifies the shot type for the generated video, that is, whether the video is a " + "single continuous shot or multiple shots with cuts.", + ), + IO.Boolean.Input( + "watermark", + default=False, + tooltip="Whether to add an AI-generated watermark to the result.", + ), + ], + 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, + ) + + @classmethod + async def execute( + cls, + model: str, + prompt: str, + negative_prompt: str, + reference_videos: IO.Autogrow.Type, + size: str, + duration: int, + seed: int, + shot_type: str, + watermark: bool, + ): + reference_video_urls = [] + for i in reference_videos: + validate_video_duration(reference_videos[i], min_duration=2, max_duration=30) + for i in reference_videos: + reference_video_urls.append(await upload_video_to_comfyapi(cls, reference_videos[i])) + width, height = RES_IN_PARENS.search(size).groups() + initial_response = await sync_op( + cls, + ApiEndpoint(path="/proxy/wan/api/v1/services/aigc/video-generation/video-synthesis", method="POST"), + response_model=TaskCreationResponse, + data=Reference2VideoTaskCreationRequest( + model=model, + input=Reference2VideoInputField( + prompt=prompt, negative_prompt=negative_prompt, reference_video_urls=reference_video_urls + ), + parameters=Reference2VideoParametersField( + size=f"{width}*{height}", + duration=duration, + shot_type=shot_type, + watermark=watermark, + seed=seed, + ), + ), + ) + if not initial_response.output: + raise Exception(f"An unknown error occurred: {initial_response.code} - {initial_response.message}") + response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/wan/api/v1/tasks/{initial_response.output.task_id}"), + 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)) + + class WanApiExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: @@ -701,6 +888,7 @@ class WanApiExtension(ComfyExtension): WanImageToImageApi, WanTextToVideoApi, WanImageToVideoApi, + WanReferenceVideoApi, ] diff --git a/comfy_api_nodes/util/__init__.py b/comfy_api_nodes/util/__init__.py index 21013b591..4cc22abfb 100644 --- a/comfy_api_nodes/util/__init__.py +++ b/comfy_api_nodes/util/__init__.py @@ -36,6 +36,7 @@ from .upload_helpers import ( upload_video_to_comfyapi, ) from .validation_utils import ( + get_image_dimensions, get_number_of_images, validate_aspect_ratio_string, validate_audio_duration, @@ -46,6 +47,7 @@ from .validation_utils import ( validate_string, validate_video_dimensions, validate_video_duration, + validate_video_frame_count, ) __all__ = [ @@ -82,6 +84,7 @@ __all__ = [ "trim_video", "video_to_base64_string", # Validation utilities + "get_image_dimensions", "get_number_of_images", "validate_aspect_ratio_string", "validate_audio_duration", @@ -92,6 +95,7 @@ __all__ = [ "validate_string", "validate_video_dimensions", "validate_video_duration", + "validate_video_frame_count", # Misc functions "get_fs_object_size", ] diff --git a/comfy_api_nodes/util/_helpers.py b/comfy_api_nodes/util/_helpers.py index 328fe5227..648defe3d 100644 --- a/comfy_api_nodes/util/_helpers.py +++ b/comfy_api_nodes/util/_helpers.py @@ -1,9 +1,12 @@ import asyncio import contextlib import os +import re import time +from collections.abc import Callable from io import BytesIO -from typing import Callable, Optional, Union + +from yarl import URL from comfy.cli_args import args from comfy.model_management import processing_interrupted @@ -11,6 +14,9 @@ from comfy_api.latest import IO from .common_exceptions import ProcessingInterrupted +_HAS_PCT_ESC = re.compile(r"%[0-9A-Fa-f]{2}") # any % followed by 2 hex digits +_HAS_BAD_PCT = re.compile(r"%(?![0-9A-Fa-f]{2})") # any % not followed by 2 hex digits + def is_processing_interrupted() -> bool: """Return True if user/runtime requested interruption.""" @@ -35,12 +41,12 @@ def default_base_url() -> str: async def sleep_with_interrupt( seconds: float, - node_cls: Optional[type[IO.ComfyNode]], - label: Optional[str] = None, - start_ts: Optional[float] = None, - estimated_total: Optional[int] = None, + node_cls: type[IO.ComfyNode] | None, + label: str | None = None, + start_ts: float | None = None, + estimated_total: int | None = None, *, - display_callback: Optional[Callable[[type[IO.ComfyNode], str, int, Optional[int]], None]] = None, + display_callback: Callable[[type[IO.ComfyNode], str, int, int | None], None] | None = None, ): """ Sleep in 1s slices while: @@ -65,7 +71,21 @@ def mimetype_to_extension(mime_type: str) -> str: return mime_type.split("/")[-1].lower() -def get_fs_object_size(path_or_object: Union[str, BytesIO]) -> int: +def get_fs_object_size(path_or_object: str | BytesIO) -> int: if isinstance(path_or_object, str): return os.path.getsize(path_or_object) return len(path_or_object.getvalue()) + + +def to_aiohttp_url(url: str) -> URL: + """If `url` appears to be already percent-encoded (contains at least one valid %HH + escape and no malformed '%' sequences) and contains no raw whitespace/control + characters preserve the original encoding byte-for-byte (important for signed/presigned URLs). + Otherwise, return `URL(url)` and allow yarl to normalize/quote as needed.""" + if any(c.isspace() for c in url) or any(ord(c) < 0x20 for c in url): + # Avoid encoded=True if URL contains raw whitespace/control chars + return URL(url) + if _HAS_PCT_ESC.search(url) and not _HAS_BAD_PCT.search(url): + # Preserve encoding only if it appears pre-encoded AND has no invalid % sequences + return URL(url, encoded=True) + return URL(url) diff --git a/comfy_api_nodes/util/client.py b/comfy_api_nodes/util/client.py index 2d5dcd648..f372ec7b5 100644 --- a/comfy_api_nodes/util/client.py +++ b/comfy_api_nodes/util/client.py @@ -4,10 +4,11 @@ import json import logging import time import uuid +from collections.abc import Callable, Iterable from dataclasses import dataclass from enum import Enum from io import BytesIO -from typing import Any, Callable, Iterable, Literal, Optional, Type, TypeVar, Union +from typing import Any, Literal, TypeVar from urllib.parse import urljoin, urlparse import aiohttp @@ -37,8 +38,8 @@ class ApiEndpoint: path: str, method: Literal["GET", "POST", "PUT", "DELETE", "PATCH"] = "GET", *, - query_params: Optional[dict[str, Any]] = None, - headers: Optional[dict[str, str]] = None, + query_params: dict[str, Any] | None = None, + headers: dict[str, str] | None = None, ): self.path = path self.method = method @@ -52,17 +53,18 @@ class _RequestConfig: endpoint: ApiEndpoint timeout: float content_type: str - data: Optional[dict[str, Any]] - files: Optional[Union[dict[str, Any], list[tuple[str, Any]]]] - multipart_parser: Optional[Callable] + data: dict[str, Any] | None + files: dict[str, Any] | list[tuple[str, Any]] | None + multipart_parser: Callable | None max_retries: int retry_delay: float retry_backoff: float wait_label: str = "Waiting" monitor_progress: bool = True - estimated_total: Optional[int] = None - final_label_on_success: Optional[str] = "Completed" - progress_origin_ts: Optional[float] = None + estimated_total: int | None = None + final_label_on_success: str | None = "Completed" + progress_origin_ts: float | None = None + price_extractor: Callable[[dict[str, Any]], float | None] | None = None @dataclass @@ -70,40 +72,42 @@ class _PollUIState: started: float status_label: str = "Queued" is_queued: bool = True - price: Optional[float] = None - estimated_duration: Optional[int] = None + price: float | None = None + estimated_duration: int | None = None base_processing_elapsed: float = 0.0 # sum of completed active intervals - active_since: Optional[float] = None # start time of current active interval (None if queued) + active_since: float | None = None # start time of current active interval (None if queued) _RETRY_STATUS = {408, 429, 500, 502, 503, 504} -COMPLETED_STATUSES = ["succeeded", "succeed", "success", "completed", "finished", "done"] -FAILED_STATUSES = ["cancelled", "canceled", "fail", "failed", "error"] -QUEUED_STATUSES = ["created", "queued", "queueing", "submitted"] +COMPLETED_STATUSES = ["succeeded", "succeed", "success", "completed", "finished", "done", "complete"] +FAILED_STATUSES = ["cancelled", "canceled", "canceling", "fail", "failed", "error"] +QUEUED_STATUSES = ["created", "queued", "queueing", "submitted", "initializing"] async def sync_op( cls: type[IO.ComfyNode], endpoint: ApiEndpoint, *, - response_model: Type[M], - data: Optional[BaseModel] = None, - files: Optional[Union[dict[str, Any], list[tuple[str, Any]]]] = None, + response_model: type[M], + price_extractor: Callable[[M | Any], float | None] | None = None, + data: BaseModel | None = None, + files: dict[str, Any] | list[tuple[str, Any]] | None = None, content_type: str = "application/json", timeout: float = 3600.0, - multipart_parser: Optional[Callable] = None, + multipart_parser: Callable | None = None, max_retries: int = 3, retry_delay: float = 1.0, retry_backoff: float = 2.0, wait_label: str = "Waiting for server", - estimated_duration: Optional[int] = None, - final_label_on_success: Optional[str] = "Completed", - progress_origin_ts: Optional[float] = None, + estimated_duration: int | None = None, + final_label_on_success: str | None = "Completed", + progress_origin_ts: float | None = None, monitor_progress: bool = True, ) -> M: raw = await sync_op_raw( cls, endpoint, + price_extractor=_wrap_model_extractor(response_model, price_extractor), data=data, files=files, content_type=content_type, @@ -128,22 +132,22 @@ async def poll_op( cls: type[IO.ComfyNode], poll_endpoint: ApiEndpoint, *, - response_model: Type[M], - status_extractor: Callable[[M], Optional[Union[str, int]]], - progress_extractor: Optional[Callable[[M], Optional[int]]] = None, - price_extractor: Optional[Callable[[M], Optional[float]]] = None, - completed_statuses: Optional[list[Union[str, int]]] = None, - failed_statuses: Optional[list[Union[str, int]]] = None, - queued_statuses: Optional[list[Union[str, int]]] = None, - data: Optional[BaseModel] = None, + response_model: type[M], + status_extractor: Callable[[M | Any], str | int | None], + progress_extractor: Callable[[M | Any], int | None] | None = None, + price_extractor: Callable[[M | Any], float | None] | None = None, + completed_statuses: list[str | int] | None = None, + failed_statuses: list[str | int] | None = None, + queued_statuses: list[str | int] | None = None, + data: BaseModel | None = None, poll_interval: float = 5.0, max_poll_attempts: int = 120, timeout_per_poll: float = 120.0, max_retries_per_poll: int = 3, retry_delay_per_poll: float = 1.0, retry_backoff_per_poll: float = 2.0, - estimated_duration: Optional[int] = None, - cancel_endpoint: Optional[ApiEndpoint] = None, + estimated_duration: int | None = None, + cancel_endpoint: ApiEndpoint | None = None, cancel_timeout: float = 10.0, ) -> M: raw = await poll_op_raw( @@ -175,21 +179,22 @@ async def sync_op_raw( cls: type[IO.ComfyNode], endpoint: ApiEndpoint, *, - data: Optional[Union[dict[str, Any], BaseModel]] = None, - files: Optional[Union[dict[str, Any], list[tuple[str, Any]]]] = None, + price_extractor: Callable[[dict[str, Any]], float | None] | None = None, + data: dict[str, Any] | BaseModel | None = None, + files: dict[str, Any] | list[tuple[str, Any]] | None = None, content_type: str = "application/json", timeout: float = 3600.0, - multipart_parser: Optional[Callable] = None, + multipart_parser: Callable | None = None, max_retries: int = 3, retry_delay: float = 1.0, retry_backoff: float = 2.0, wait_label: str = "Waiting for server", - estimated_duration: Optional[int] = None, + estimated_duration: int | None = None, as_binary: bool = False, - final_label_on_success: Optional[str] = "Completed", - progress_origin_ts: Optional[float] = None, + final_label_on_success: str | None = "Completed", + progress_origin_ts: float | None = None, monitor_progress: bool = True, -) -> Union[dict[str, Any], bytes]: +) -> dict[str, Any] | bytes: """ Make a single network request. - If as_binary=False (default): returns JSON dict (or {'_raw': ''} if non-JSON). @@ -216,6 +221,7 @@ async def sync_op_raw( estimated_total=estimated_duration, final_label_on_success=final_label_on_success, progress_origin_ts=progress_origin_ts, + price_extractor=price_extractor, ) return await _request_base(cfg, expect_binary=as_binary) @@ -224,21 +230,21 @@ async def poll_op_raw( cls: type[IO.ComfyNode], poll_endpoint: ApiEndpoint, *, - status_extractor: Callable[[dict[str, Any]], Optional[Union[str, int]]], - progress_extractor: Optional[Callable[[dict[str, Any]], Optional[int]]] = None, - price_extractor: Optional[Callable[[dict[str, Any]], Optional[float]]] = None, - completed_statuses: Optional[list[Union[str, int]]] = None, - failed_statuses: Optional[list[Union[str, int]]] = None, - queued_statuses: Optional[list[Union[str, int]]] = None, - data: Optional[Union[dict[str, Any], BaseModel]] = None, + status_extractor: Callable[[dict[str, Any]], str | int | None], + progress_extractor: Callable[[dict[str, Any]], int | None] | None = None, + price_extractor: Callable[[dict[str, Any]], float | None] | None = None, + completed_statuses: list[str | int] | None = None, + failed_statuses: list[str | int] | None = None, + queued_statuses: list[str | int] | None = None, + data: dict[str, Any] | BaseModel | None = None, poll_interval: float = 5.0, max_poll_attempts: int = 120, timeout_per_poll: float = 120.0, max_retries_per_poll: int = 3, retry_delay_per_poll: float = 1.0, retry_backoff_per_poll: float = 2.0, - estimated_duration: Optional[int] = None, - cancel_endpoint: Optional[ApiEndpoint] = None, + estimated_duration: int | None = None, + cancel_endpoint: ApiEndpoint | None = None, cancel_timeout: float = 10.0, ) -> dict[str, Any]: """ @@ -256,7 +262,7 @@ async def poll_op_raw( consumed_attempts = 0 # counts only non-queued polls progress_bar = utils.ProgressBar(100) if progress_extractor else None - last_progress: Optional[int] = None + last_progress: int | None = None state = _PollUIState(started=started, estimated_duration=estimated_duration) stop_ticker = asyncio.Event() @@ -415,16 +421,18 @@ async def poll_op_raw( def _display_text( node_cls: type[IO.ComfyNode], - text: Optional[str], + text: str | None, *, - status: Optional[Union[str, int]] = None, - price: Optional[float] = None, + status: str | int | None = None, + price: float | None = None, ) -> None: display_lines: list[str] = [] if status: display_lines.append(f"Status: {status.capitalize() if isinstance(status, str) else status}") if price is not None: - display_lines.append(f"Price: ${float(price):,.4f}") + p = f"{float(price) * 211:,.1f}".rstrip("0").rstrip(".") + if p != "0": + display_lines.append(f"Price: {p} credits") if text is not None: display_lines.append(text) if display_lines: @@ -433,13 +441,13 @@ def _display_text( def _display_time_progress( node_cls: type[IO.ComfyNode], - status: Optional[Union[str, int]], + status: str | int | None, elapsed_seconds: int, - estimated_total: Optional[int] = None, + estimated_total: int | None = None, *, - price: Optional[float] = None, - is_queued: Optional[bool] = None, - processing_elapsed_seconds: Optional[int] = None, + price: float | None = None, + is_queued: bool | None = None, + processing_elapsed_seconds: int | None = None, ) -> None: if estimated_total is not None and estimated_total > 0 and is_queued is False: pe = processing_elapsed_seconds if processing_elapsed_seconds is not None else elapsed_seconds @@ -481,7 +489,7 @@ def _unpack_tuple(t: tuple) -> tuple[str, Any, str]: raise ValueError("files tuple must be (filename, file[, content_type])") -def _merge_params(endpoint_params: dict[str, Any], method: str, data: Optional[dict[str, Any]]) -> dict[str, Any]: +def _merge_params(endpoint_params: dict[str, Any], method: str, data: dict[str, Any] | None) -> dict[str, Any]: params = dict(endpoint_params or {}) if method.upper() == "GET" and data: for k, v in data.items(): @@ -527,9 +535,9 @@ def _generate_operation_id(method: str, path: str, attempt: int) -> str: def _snapshot_request_body_for_logging( content_type: str, method: str, - data: Optional[dict[str, Any]], - files: Optional[Union[dict[str, Any], list[tuple[str, Any]]]], -) -> Optional[Union[dict[str, Any], str]]: + data: dict[str, Any] | None, + files: dict[str, Any] | list[tuple[str, Any]] | None, +) -> dict[str, Any] | str | None: if method.upper() == "GET": return None if content_type == "multipart/form-data": @@ -579,12 +587,13 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool): attempt = 0 delay = cfg.retry_delay operation_succeeded: bool = False - final_elapsed_seconds: Optional[int] = None + final_elapsed_seconds: int | None = None + extracted_price: float | None = None while True: attempt += 1 stop_event = asyncio.Event() - monitor_task: Optional[asyncio.Task] = None - sess: Optional[aiohttp.ClientSession] = None + monitor_task: asyncio.Task | None = None + sess: aiohttp.ClientSession | None = None operation_id = _generate_operation_id(method, cfg.endpoint.path, attempt) logging.debug("[DEBUG] HTTP %s %s (attempt %d)", method, url, attempt) @@ -767,6 +776,8 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool): except json.JSONDecodeError: payload = {"_raw": text} response_content_to_log = payload if isinstance(payload, dict) else text + with contextlib.suppress(Exception): + extracted_price = cfg.price_extractor(payload) if cfg.price_extractor else None operation_succeeded = True final_elapsed_seconds = int(time.monotonic() - start_time) try: @@ -871,13 +882,13 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool): else int(time.monotonic() - start_time) ), estimated_total=cfg.estimated_total, - price=None, + price=extracted_price, is_queued=False, processing_elapsed_seconds=final_elapsed_seconds, ) -def _validate_or_raise(response_model: Type[M], payload: Any) -> M: +def _validate_or_raise(response_model: type[M], payload: Any) -> M: try: return response_model.model_validate(payload) except Exception as e: @@ -892,9 +903,9 @@ def _validate_or_raise(response_model: Type[M], payload: Any) -> M: def _wrap_model_extractor( - response_model: Type[M], - extractor: Optional[Callable[[M], Any]], -) -> Optional[Callable[[dict[str, Any]], Any]]: + response_model: type[M], + extractor: Callable[[M], Any] | None, +) -> Callable[[dict[str, Any]], Any] | None: """Wrap a typed extractor so it can be used by the dict-based poller. Validates the dict into `response_model` before invoking `extractor`. Uses a small per-wrapper cache keyed by `id(dict)` to avoid re-validating @@ -919,10 +930,10 @@ def _wrap_model_extractor( return _wrapped -def _normalize_statuses(values: Optional[Iterable[Union[str, int]]]) -> set[Union[str, int]]: +def _normalize_statuses(values: Iterable[str | int] | None) -> set[str | int]: if not values: return set() - out: set[Union[str, int]] = set() + out: set[str | int] = set() for v in values: nv = _normalize_status_value(v) if nv is not None: @@ -930,7 +941,7 @@ def _normalize_statuses(values: Optional[Iterable[Union[str, int]]]) -> set[Unio return out -def _normalize_status_value(val: Union[str, int, None]) -> Union[str, int, None]: +def _normalize_status_value(val: str | int | None) -> str | int | None: if isinstance(val, str): return val.strip().lower() return val diff --git a/comfy_api_nodes/util/conversions.py b/comfy_api_nodes/util/conversions.py index 971dc57de..d64239c86 100644 --- a/comfy_api_nodes/util/conversions.py +++ b/comfy_api_nodes/util/conversions.py @@ -4,7 +4,6 @@ import math import mimetypes import uuid from io import BytesIO -from typing import Optional import av import numpy as np @@ -12,8 +11,7 @@ import torch from PIL import Image from comfy.utils import common_upscale -from comfy_api.latest import Input, InputImpl -from comfy_api.util import VideoCodec, VideoContainer +from comfy_api.latest import Input, InputImpl, Types from ._helpers import mimetype_to_extension @@ -57,7 +55,7 @@ def image_tensor_pair_to_batch(image1: torch.Tensor, image2: torch.Tensor) -> to def tensor_to_bytesio( image: torch.Tensor, - name: Optional[str] = None, + name: str | None = None, total_pixels: int = 2048 * 2048, mime_type: str = "image/png", ) -> BytesIO: @@ -131,7 +129,7 @@ def pil_to_bytesio(img: Image.Image, mime_type: str = "image/png") -> BytesIO: return img_byte_arr -def downscale_image_tensor(image, total_pixels=1536 * 1024) -> torch.Tensor: +def downscale_image_tensor(image: torch.Tensor, total_pixels: int = 1536 * 1024) -> torch.Tensor: """Downscale input image tensor to roughly the specified total pixels.""" samples = image.movedim(-1, 1) total = int(total_pixels) @@ -177,8 +175,8 @@ def audio_to_base64_string(audio: Input.Audio, container_format: str = "mp4", co def video_to_base64_string( video: Input.Video, - container_format: VideoContainer = None, - codec: VideoCodec = None + container_format: Types.VideoContainer | None = None, + codec: Types.VideoCodec | None = None, ) -> str: """ Converts a video input to a base64 string. @@ -189,12 +187,11 @@ def video_to_base64_string( codec: Optional codec to use (defaults to video.codec if available) """ video_bytes_io = BytesIO() - - # Use provided format/codec if specified, otherwise use video's own if available - format_to_use = container_format if container_format is not None else getattr(video, 'container', VideoContainer.MP4) - codec_to_use = codec if codec is not None else getattr(video, 'codec', VideoCodec.H264) - - video.save_to(video_bytes_io, format=format_to_use, codec=codec_to_use) + video.save_to( + video_bytes_io, + format=container_format or getattr(video, "container", Types.VideoContainer.MP4), + codec=codec or getattr(video, "codec", Types.VideoCodec.H264), + ) video_bytes_io.seek(0) return base64.b64encode(video_bytes_io.getvalue()).decode("utf-8") diff --git a/comfy_api_nodes/util/download_helpers.py b/comfy_api_nodes/util/download_helpers.py index 14207dc68..4668d14a9 100644 --- a/comfy_api_nodes/util/download_helpers.py +++ b/comfy_api_nodes/util/download_helpers.py @@ -3,15 +3,15 @@ import contextlib import uuid from io import BytesIO from pathlib import Path -from typing import IO, Optional, Union +from typing import IO from urllib.parse import urljoin, urlparse import aiohttp import torch from aiohttp.client_exceptions import ClientError, ContentTypeError -from comfy_api.input_impl import VideoFromFile from comfy_api.latest import IO as COMFY_IO +from comfy_api.latest import InputImpl from . import request_logger from ._helpers import ( @@ -19,6 +19,7 @@ from ._helpers import ( get_auth_header, is_processing_interrupted, sleep_with_interrupt, + to_aiohttp_url, ) from .client import _diagnose_connectivity from .common_exceptions import ApiServerError, LocalNetworkError, ProcessingInterrupted @@ -29,9 +30,9 @@ _RETRY_STATUS = {408, 429, 500, 502, 503, 504} async def download_url_to_bytesio( url: str, - dest: Optional[Union[BytesIO, IO[bytes], str, Path]], + dest: BytesIO | IO[bytes] | str | Path | None, *, - timeout: Optional[float] = None, + timeout: float | None = None, max_retries: int = 5, retry_delay: float = 1.0, retry_backoff: float = 2.0, @@ -71,10 +72,10 @@ async def download_url_to_bytesio( is_path_sink = isinstance(dest, (str, Path)) fhandle = None - session: Optional[aiohttp.ClientSession] = None - stop_evt: Optional[asyncio.Event] = None - monitor_task: Optional[asyncio.Task] = None - req_task: Optional[asyncio.Task] = None + session: aiohttp.ClientSession | None = None + stop_evt: asyncio.Event | None = None + monitor_task: asyncio.Task | None = None + req_task: asyncio.Task | None = None try: with contextlib.suppress(Exception): @@ -94,7 +95,7 @@ async def download_url_to_bytesio( monitor_task = asyncio.create_task(_monitor()) - req_task = asyncio.create_task(session.get(url, headers=headers)) + req_task = asyncio.create_task(session.get(to_aiohttp_url(url), headers=headers)) done, pending = await asyncio.wait({req_task, monitor_task}, return_when=asyncio.FIRST_COMPLETED) if monitor_task in done and req_task in pending: @@ -234,11 +235,11 @@ async def download_url_to_video_output( timeout: float = None, max_retries: int = 5, cls: type[COMFY_IO.ComfyNode] = None, -) -> VideoFromFile: +) -> InputImpl.VideoFromFile: """Downloads a video from a URL and returns a `VIDEO` output.""" result = BytesIO() await download_url_to_bytesio(video_url, result, timeout=timeout, max_retries=max_retries, cls=cls) - return VideoFromFile(result) + return InputImpl.VideoFromFile(result) async def download_url_as_bytesio( diff --git a/comfy_api_nodes/util/request_logger.py b/comfy_api_nodes/util/request_logger.py index ac52e2eab..e0cb4428d 100644 --- a/comfy_api_nodes/util/request_logger.py +++ b/comfy_api_nodes/util/request_logger.py @@ -1,5 +1,3 @@ -from __future__ import annotations - import datetime import hashlib import json diff --git a/comfy_api_nodes/util/upload_helpers.py b/comfy_api_nodes/util/upload_helpers.py index 632450d9b..f1ed7fe9c 100644 --- a/comfy_api_nodes/util/upload_helpers.py +++ b/comfy_api_nodes/util/upload_helpers.py @@ -4,15 +4,13 @@ import logging import time import uuid from io import BytesIO -from typing import Optional, Union from urllib.parse import urlparse import aiohttp import torch from pydantic import BaseModel, Field -from comfy_api.latest import IO, Input -from comfy_api.util import VideoCodec, VideoContainer +from comfy_api.latest import IO, Input, Types from . import request_logger from ._helpers import is_processing_interrupted, sleep_with_interrupt @@ -32,7 +30,7 @@ from .conversions import ( class UploadRequest(BaseModel): file_name: str = Field(..., description="Filename to upload") - content_type: Optional[str] = Field( + content_type: str | None = Field( None, description="Mime type of the file. For example: image/png, image/jpeg, video/mp4, etc.", ) @@ -48,22 +46,30 @@ async def upload_images_to_comfyapi( image: torch.Tensor, *, max_images: int = 8, - mime_type: Optional[str] = None, - wait_label: Optional[str] = "Uploading", + mime_type: str | None = None, + wait_label: str | None = "Uploading", + show_batch_index: bool = True, ) -> list[str]: """ Uploads images to ComfyUI API and returns download URLs. To upload multiple images, stack them in the batch dimension first. """ - # if batch, try to upload each file if max_images is greater than 0 + # if batched, try to upload each file if max_images is greater than 0 download_urls: list[str] = [] is_batch = len(image.shape) > 3 batch_len = image.shape[0] if is_batch else 1 + num_to_upload = min(batch_len, max_images) + batch_start_ts = time.monotonic() - for idx in range(min(batch_len, max_images)): + for idx in range(num_to_upload): tensor = image[idx] if is_batch else image img_io = tensor_to_bytesio(tensor, mime_type=mime_type) - url = await upload_file_to_comfyapi(cls, img_io, img_io.name, mime_type, wait_label) + + effective_label = wait_label + if wait_label and show_batch_index and num_to_upload > 1: + effective_label = f"{wait_label} ({idx + 1}/{num_to_upload})" + + url = await upload_file_to_comfyapi(cls, img_io, img_io.name, mime_type, effective_label, batch_start_ts) download_urls.append(url) return download_urls @@ -92,9 +98,10 @@ async def upload_video_to_comfyapi( cls: type[IO.ComfyNode], video: Input.Video, *, - container: VideoContainer = VideoContainer.MP4, - codec: VideoCodec = VideoCodec.H264, - max_duration: Optional[int] = None, + container: Types.VideoContainer = Types.VideoContainer.MP4, + codec: Types.VideoCodec = Types.VideoCodec.H264, + max_duration: int | None = None, + wait_label: str | None = "Uploading", ) -> str: """ Uploads a single video to ComfyUI API and returns its download URL. @@ -112,22 +119,23 @@ async def upload_video_to_comfyapi( raise ValueError(f"Could not verify video duration from source: {e}") from e upload_mime_type = f"video/{container.value.lower()}" - filename = f"uploaded_video.{container.value.lower()}" + filename = f"{uuid.uuid4()}.{container.value.lower()}" # Convert VideoInput to BytesIO using specified container/codec video_bytes_io = BytesIO() video.save_to(video_bytes_io, format=container, codec=codec) video_bytes_io.seek(0) - return await upload_file_to_comfyapi(cls, video_bytes_io, filename, upload_mime_type) + return await upload_file_to_comfyapi(cls, video_bytes_io, filename, upload_mime_type, wait_label) async def upload_file_to_comfyapi( cls: type[IO.ComfyNode], file_bytes_io: BytesIO, filename: str, - upload_mime_type: Optional[str], - wait_label: Optional[str] = "Uploading", + upload_mime_type: str | None, + wait_label: str | None = "Uploading", + progress_origin_ts: float | None = None, ) -> str: """Uploads a single file to ComfyUI API and returns its download URL.""" if upload_mime_type is None: @@ -148,6 +156,7 @@ async def upload_file_to_comfyapi( file_bytes_io, content_type=upload_mime_type, wait_label=wait_label, + progress_origin_ts=progress_origin_ts, ) return create_resp.download_url @@ -155,27 +164,18 @@ async def upload_file_to_comfyapi( async def upload_file( cls: type[IO.ComfyNode], upload_url: str, - file: Union[BytesIO, str], + file: BytesIO | str, *, - content_type: Optional[str] = None, + content_type: str | None = None, max_retries: int = 3, retry_delay: float = 1.0, retry_backoff: float = 2.0, - wait_label: Optional[str] = None, + wait_label: str | None = None, + progress_origin_ts: float | None = None, ) -> None: """ Upload a file to a signed URL (e.g., S3 pre-signed PUT) with retries, Comfy progress display, and interruption. - Args: - cls: Node class (provides auth context + UI progress hooks). - upload_url: Pre-signed PUT URL. - file: BytesIO or path string. - content_type: Explicit MIME type. If None, we *suppress* Content-Type. - max_retries: Maximum retry attempts. - retry_delay: Initial delay in seconds. - retry_backoff: Exponential backoff factor. - wait_label: Progress label shown in Comfy UI. - Raises: ProcessingInterrupted, LocalNetworkError, ApiServerError, Exception """ @@ -198,7 +198,7 @@ async def upload_file( attempt = 0 delay = retry_delay - start_ts = time.monotonic() + start_ts = progress_origin_ts if progress_origin_ts is not None else time.monotonic() op_uuid = uuid.uuid4().hex[:8] while True: attempt += 1 @@ -218,7 +218,7 @@ async def upload_file( return monitor_task = asyncio.create_task(_monitor()) - sess: Optional[aiohttp.ClientSession] = None + sess: aiohttp.ClientSession | None = None try: try: request_logger.log_request_response( diff --git a/comfy_api_nodes/util/validation_utils.py b/comfy_api_nodes/util/validation_utils.py index ec7006aed..f01edea96 100644 --- a/comfy_api_nodes/util/validation_utils.py +++ b/comfy_api_nodes/util/validation_utils.py @@ -1,9 +1,7 @@ import logging -from typing import Optional import torch -from comfy_api.input.video_types import VideoInput from comfy_api.latest import Input @@ -18,10 +16,10 @@ def get_image_dimensions(image: torch.Tensor) -> tuple[int, int]: def validate_image_dimensions( image: torch.Tensor, - min_width: Optional[int] = None, - max_width: Optional[int] = None, - min_height: Optional[int] = None, - max_height: Optional[int] = None, + min_width: int | None = None, + max_width: int | None = None, + min_height: int | None = None, + max_height: int | None = None, ): height, width = get_image_dimensions(image) @@ -37,8 +35,8 @@ def validate_image_dimensions( def validate_image_aspect_ratio( image: torch.Tensor, - min_ratio: Optional[tuple[float, float]] = None, # e.g. (1, 4) - max_ratio: Optional[tuple[float, float]] = None, # e.g. (4, 1) + min_ratio: tuple[float, float] | None = None, # e.g. (1, 4) + max_ratio: tuple[float, float] | None = None, # e.g. (4, 1) *, strict: bool = True, # True -> (min, max); False -> [min, max] ) -> float: @@ -54,8 +52,8 @@ def validate_image_aspect_ratio( def validate_images_aspect_ratio_closeness( first_image: torch.Tensor, second_image: torch.Tensor, - min_rel: float, # e.g. 0.8 - max_rel: float, # e.g. 1.25 + min_rel: float, # e.g. 0.8 + max_rel: float, # e.g. 1.25 *, strict: bool = False, # True -> (min, max); False -> [min, max] ) -> float: @@ -84,8 +82,8 @@ def validate_images_aspect_ratio_closeness( def validate_aspect_ratio_string( aspect_ratio: str, - min_ratio: Optional[tuple[float, float]] = None, # e.g. (1, 4) - max_ratio: Optional[tuple[float, float]] = None, # e.g. (4, 1) + min_ratio: tuple[float, float] | None = None, # e.g. (1, 4) + max_ratio: tuple[float, float] | None = None, # e.g. (4, 1) *, strict: bool = False, # True -> (min, max); False -> [min, max] ) -> float: @@ -97,10 +95,10 @@ def validate_aspect_ratio_string( def validate_video_dimensions( video: Input.Video, - min_width: Optional[int] = None, - max_width: Optional[int] = None, - min_height: Optional[int] = None, - max_height: Optional[int] = None, + min_width: int | None = None, + max_width: int | None = None, + min_height: int | None = None, + max_height: int | None = None, ): try: width, height = video.get_dimensions() @@ -120,8 +118,8 @@ def validate_video_dimensions( def validate_video_duration( video: Input.Video, - min_duration: Optional[float] = None, - max_duration: Optional[float] = None, + min_duration: float | None = None, + max_duration: float | None = None, ): try: duration = video.get_duration() @@ -136,6 +134,23 @@ def validate_video_duration( raise ValueError(f"Video duration must be at most {max_duration}s, got {duration}s") +def validate_video_frame_count( + video: Input.Video, + min_frame_count: int | None = None, + max_frame_count: int | None = None, +): + try: + frame_count = video.get_frame_count() + except Exception as e: + logging.error("Error getting frame count of video: %s", e) + return + + if min_frame_count is not None and min_frame_count > frame_count: + raise ValueError(f"Video frame count must be at least {min_frame_count}, got {frame_count}") + if max_frame_count is not None and frame_count > max_frame_count: + raise ValueError(f"Video frame count must be at most {max_frame_count}, got {frame_count}") + + def get_number_of_images(images): if isinstance(images, torch.Tensor): return images.shape[0] if images.ndim >= 4 else 1 @@ -144,8 +159,8 @@ def get_number_of_images(images): def validate_audio_duration( audio: Input.Audio, - min_duration: Optional[float] = None, - max_duration: Optional[float] = None, + min_duration: float | None = None, + max_duration: float | None = None, ) -> None: sr = int(audio["sample_rate"]) dur = int(audio["waveform"].shape[-1]) / sr @@ -177,7 +192,7 @@ def validate_string( ) -def validate_container_format_is_mp4(video: VideoInput) -> None: +def validate_container_format_is_mp4(video: Input.Video) -> None: """Validates video container format is MP4.""" container_format = video.get_container_format() if container_format not in ["mp4", "mov,mp4,m4a,3gp,3g2,mj2"]: @@ -194,8 +209,8 @@ def _ratio_from_tuple(r: tuple[float, float]) -> float: def _assert_ratio_bounds( ar: float, *, - min_ratio: Optional[tuple[float, float]] = None, - max_ratio: Optional[tuple[float, float]] = None, + min_ratio: tuple[float, float] | None = None, + max_ratio: tuple[float, float] | None = None, strict: bool = True, ) -> None: """Validate a numeric aspect ratio against optional min/max ratio bounds.""" diff --git a/comfy_execution/graph.py b/comfy_execution/graph.py index 0d811e354..9d170b16e 100644 --- a/comfy_execution/graph.py +++ b/comfy_execution/graph.py @@ -97,6 +97,11 @@ def get_input_info( extra_info = input_info[1] else: extra_info = {} + # if input_type is a list, it is a Combo defined in outdated format; convert it. + # NOTE: uncomment this when we are confident old format going away won't cause too much trouble. + # if isinstance(input_type, list): + # extra_info["options"] = input_type + # input_type = IO.Combo.io_type return input_type, input_category, extra_info class TopologicalSort: @@ -202,15 +207,15 @@ class ExecutionList(TopologicalSort): return self.output_cache.get(node_id) is not None def cache_link(self, from_node_id, to_node_id): - if not to_node_id in self.execution_cache: + if to_node_id not in self.execution_cache: self.execution_cache[to_node_id] = {} self.execution_cache[to_node_id][from_node_id] = self.output_cache.get(from_node_id) - if not from_node_id in self.execution_cache_listeners: + if from_node_id not in self.execution_cache_listeners: self.execution_cache_listeners[from_node_id] = set() self.execution_cache_listeners[from_node_id].add(to_node_id) def get_cache(self, from_node_id, to_node_id): - if not to_node_id in self.execution_cache: + if to_node_id not in self.execution_cache: return None value = self.execution_cache[to_node_id].get(from_node_id) if value is None: diff --git a/comfy_execution/jobs.py b/comfy_execution/jobs.py new file mode 100644 index 000000000..59fb49357 --- /dev/null +++ b/comfy_execution/jobs.py @@ -0,0 +1,291 @@ +""" +Job utilities for the /api/jobs endpoint. +Provides normalization and helper functions for job status tracking. +""" + +from typing import Optional + +from comfy_api.internal import prune_dict + + +class JobStatus: + """Job status constants.""" + PENDING = 'pending' + IN_PROGRESS = 'in_progress' + COMPLETED = 'completed' + FAILED = 'failed' + + ALL = [PENDING, IN_PROGRESS, COMPLETED, FAILED] + + +# Media types that can be previewed in the frontend +PREVIEWABLE_MEDIA_TYPES = frozenset({'images', 'video', 'audio'}) + +# 3D file extensions for preview fallback (no dedicated media_type exists) +THREE_D_EXTENSIONS = frozenset({'.obj', '.fbx', '.gltf', '.glb'}) + + +def _extract_job_metadata(extra_data: dict) -> tuple[Optional[int], Optional[str]]: + """Extract create_time and workflow_id from extra_data. + + Returns: + tuple: (create_time, workflow_id) + """ + create_time = extra_data.get('create_time') + extra_pnginfo = extra_data.get('extra_pnginfo', {}) + workflow_id = extra_pnginfo.get('workflow', {}).get('id') + return create_time, workflow_id + + +def is_previewable(media_type: str, item: dict) -> bool: + """ + Check if an output item is previewable. + Matches frontend logic in ComfyUI_frontend/src/stores/queueStore.ts + Maintains backwards compatibility with existing logic. + + Priority: + 1. media_type is 'images', 'video', or 'audio' + 2. format field starts with 'video/' or 'audio/' + 3. filename has a 3D extension (.obj, .fbx, .gltf, .glb) + """ + if media_type in PREVIEWABLE_MEDIA_TYPES: + return True + + # Check format field (MIME type). + # Maintains backwards compatibility with how custom node outputs are handled in the frontend. + fmt = item.get('format', '') + if fmt and (fmt.startswith('video/') or fmt.startswith('audio/')): + return True + + # Check for 3D files by extension + filename = item.get('filename', '').lower() + if any(filename.endswith(ext) for ext in THREE_D_EXTENSIONS): + return True + + return False + + +def normalize_queue_item(item: tuple, status: str) -> dict: + """Convert queue item tuple to unified job dict. + + Expects item with sensitive data already removed (5 elements). + """ + priority, prompt_id, _, extra_data, _ = item + create_time, workflow_id = _extract_job_metadata(extra_data) + + return prune_dict({ + 'id': prompt_id, + 'status': status, + 'priority': priority, + 'create_time': create_time, + 'outputs_count': 0, + 'workflow_id': workflow_id, + }) + + +def normalize_history_item(prompt_id: str, history_item: dict, include_outputs: bool = False) -> dict: + """Convert history item dict to unified job dict. + + History items have sensitive data already removed (prompt tuple has 5 elements). + """ + prompt_tuple = history_item['prompt'] + priority, _, prompt, extra_data, _ = prompt_tuple + create_time, workflow_id = _extract_job_metadata(extra_data) + + status_info = history_item.get('status', {}) + status_str = status_info.get('status_str') if status_info else None + if status_str == 'success': + status = JobStatus.COMPLETED + elif status_str == 'error': + status = JobStatus.FAILED + else: + status = JobStatus.COMPLETED + + outputs = history_item.get('outputs', {}) + outputs_count, preview_output = get_outputs_summary(outputs) + + execution_error = None + execution_start_time = None + execution_end_time = None + if status_info: + messages = status_info.get('messages', []) + for entry in messages: + if isinstance(entry, (list, tuple)) and len(entry) >= 2: + event_name, event_data = entry[0], entry[1] + if isinstance(event_data, dict): + if event_name == 'execution_start': + execution_start_time = event_data.get('timestamp') + elif event_name in ('execution_success', 'execution_error', 'execution_interrupted'): + execution_end_time = event_data.get('timestamp') + if event_name == 'execution_error': + execution_error = event_data + + job = prune_dict({ + 'id': prompt_id, + 'status': status, + 'priority': priority, + 'create_time': create_time, + 'execution_start_time': execution_start_time, + 'execution_end_time': execution_end_time, + 'execution_error': execution_error, + 'outputs_count': outputs_count, + 'preview_output': preview_output, + 'workflow_id': workflow_id, + }) + + if include_outputs: + job['outputs'] = outputs + job['execution_status'] = status_info + job['workflow'] = { + 'prompt': prompt, + 'extra_data': extra_data, + } + + return job + + +def get_outputs_summary(outputs: dict) -> tuple[int, Optional[dict]]: + """ + Count outputs and find preview in a single pass. + Returns (outputs_count, preview_output). + + Preview priority (matching frontend): + 1. type="output" with previewable media + 2. Any previewable media + """ + count = 0 + preview_output = None + fallback_preview = None + + for node_id, node_outputs in outputs.items(): + if not isinstance(node_outputs, dict): + continue + for media_type, items in node_outputs.items(): + # 'animated' is a boolean flag, not actual output items + if media_type == 'animated' or not isinstance(items, list): + continue + + for item in items: + if not isinstance(item, dict): + continue + count += 1 + + if preview_output is None and is_previewable(media_type, item): + enriched = { + **item, + 'nodeId': node_id, + 'mediaType': media_type + } + if item.get('type') == 'output': + preview_output = enriched + elif fallback_preview is None: + fallback_preview = enriched + + return count, preview_output or fallback_preview + + +def apply_sorting(jobs: list[dict], sort_by: str, sort_order: str) -> list[dict]: + """Sort jobs list by specified field and order.""" + reverse = (sort_order == 'desc') + + if sort_by == 'execution_duration': + def get_sort_key(job): + start = job.get('execution_start_time', 0) + end = job.get('execution_end_time', 0) + return end - start if end and start else 0 + else: + def get_sort_key(job): + return job.get('create_time', 0) + + return sorted(jobs, key=get_sort_key, reverse=reverse) + + +def get_job(prompt_id: str, running: list, queued: list, history: dict) -> Optional[dict]: + """ + Get a single job by prompt_id from history or queue. + + Args: + prompt_id: The prompt ID to look up + running: List of currently running queue items + queued: List of pending queue items + history: Dict of history items keyed by prompt_id + + Returns: + Job dict with full details, or None if not found + """ + if prompt_id in history: + return normalize_history_item(prompt_id, history[prompt_id], include_outputs=True) + + for item in running: + if item[1] == prompt_id: + return normalize_queue_item(item, JobStatus.IN_PROGRESS) + + for item in queued: + if item[1] == prompt_id: + return normalize_queue_item(item, JobStatus.PENDING) + + return None + + +def get_all_jobs( + running: list, + queued: list, + history: dict, + status_filter: Optional[list[str]] = None, + workflow_id: Optional[str] = None, + sort_by: str = "created_at", + sort_order: str = "desc", + limit: Optional[int] = None, + offset: int = 0 +) -> tuple[list[dict], int]: + """ + Get all jobs (running, pending, completed) with filtering and sorting. + + Args: + running: List of currently running queue items + queued: List of pending queue items + history: Dict of history items keyed by prompt_id + status_filter: List of statuses to include (from JobStatus.ALL) + workflow_id: Filter by workflow ID + sort_by: Field to sort by ('created_at', 'execution_duration') + sort_order: 'asc' or 'desc' + limit: Maximum number of items to return + offset: Number of items to skip + + Returns: + tuple: (jobs_list, total_count) + """ + jobs = [] + + if status_filter is None: + status_filter = JobStatus.ALL + + if JobStatus.IN_PROGRESS in status_filter: + for item in running: + jobs.append(normalize_queue_item(item, JobStatus.IN_PROGRESS)) + + if JobStatus.PENDING in status_filter: + for item in queued: + jobs.append(normalize_queue_item(item, JobStatus.PENDING)) + + include_completed = JobStatus.COMPLETED in status_filter + include_failed = JobStatus.FAILED in status_filter + if include_completed or include_failed: + for prompt_id, history_item in history.items(): + is_failed = history_item.get('status', {}).get('status_str') == 'error' + if (is_failed and include_failed) or (not is_failed and include_completed): + jobs.append(normalize_history_item(prompt_id, history_item)) + + if workflow_id: + jobs = [j for j in jobs if j.get('workflow_id') == workflow_id] + + jobs = apply_sorting(jobs, sort_by, sort_order) + + total_count = len(jobs) + + if offset > 0: + jobs = jobs[offset:] + if limit is not None: + jobs = jobs[:limit] + + return (jobs, total_count) diff --git a/comfy_execution/validation.py b/comfy_execution/validation.py index cec105fc9..e73624bd1 100644 --- a/comfy_execution/validation.py +++ b/comfy_execution/validation.py @@ -1,4 +1,5 @@ from __future__ import annotations +from comfy_api.latest import IO def validate_node_input( @@ -20,9 +21,24 @@ def validate_node_input( """ # If the types are exactly the same, we can return immediately # Use pre-union behaviour: inverse of `__ne__` + # NOTE: this lets legacy '*' Any types work that override the __ne__ method of the str class. if not received_type != input_type: return True + # If one of the types is '*', we can return True immediately; this is the 'Any' type. + if received_type == IO.AnyType.io_type or input_type == IO.AnyType.io_type: + return True + + # If the received type or input_type is a MatchType, we can return True immediately; + # validation for this is handled by the frontend + if received_type == IO.MatchType.io_type or input_type == IO.MatchType.io_type: + return True + + # This accounts for some custom nodes that output lists of options as the type; + # if we ever want to break them on purpose, this can be removed + if isinstance(received_type, list) and input_type == IO.Combo.io_type: + return True + # Not equal, and not strings if not isinstance(received_type, str) or not isinstance(input_type, str): return False @@ -31,6 +47,10 @@ def validate_node_input( received_types = set(t.strip() for t in received_type.split(",")) input_types = set(t.strip() for t in input_type.split(",")) + # If any of the types is '*', we can return True immediately; this is the 'Any' type. + if IO.AnyType.io_type in received_types or IO.AnyType.io_type in input_types: + return True + if strict: # In strict mode, all received types must be in the input types return received_types.issubset(input_types) diff --git a/comfy_extras/nodes_apg.py b/comfy_extras/nodes_apg.py index f27ae7da8..b9df2dcc9 100644 --- a/comfy_extras/nodes_apg.py +++ b/comfy_extras/nodes_apg.py @@ -55,7 +55,8 @@ class APG(io.ComfyNode): def pre_cfg_function(args): nonlocal running_avg, prev_sigma - if len(args["conds_out"]) == 1: return args["conds_out"] + if len(args["conds_out"]) == 1: + return args["conds_out"] cond = args["conds_out"][0] uncond = args["conds_out"][1] diff --git a/comfy_extras/nodes_audio.py b/comfy_extras/nodes_audio.py index 2ed7e0b22..94ad5e8a8 100644 --- a/comfy_extras/nodes_audio.py +++ b/comfy_extras/nodes_audio.py @@ -6,65 +6,80 @@ import torch import comfy.model_management import folder_paths import os -import io -import json -import random import hashlib import node_helpers import logging -from comfy.cli_args import args -from comfy.comfy_types import FileLocator +from typing_extensions import override +from comfy_api.latest import ComfyExtension, IO, UI -class EmptyLatentAudio: - def __init__(self): - self.device = comfy.model_management.intermediate_device() +class EmptyLatentAudio(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="EmptyLatentAudio", + display_name="Empty Latent Audio", + category="latent/audio", + inputs=[ + IO.Float.Input("seconds", default=47.6, min=1.0, max=1000.0, step=0.1), + IO.Int.Input( + "batch_size", default=1, min=1, max=4096, tooltip="The number of latent images in the batch." + ), + ], + outputs=[IO.Latent.Output()], + ) @classmethod - def INPUT_TYPES(s): - return {"required": {"seconds": ("FLOAT", {"default": 47.6, "min": 1.0, "max": 1000.0, "step": 0.1}), - "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096, "tooltip": "The number of latent images in the batch."}), - }} - RETURN_TYPES = ("LATENT",) - FUNCTION = "generate" - - CATEGORY = "latent/audio" - - def generate(self, seconds, batch_size): + def execute(cls, seconds, batch_size) -> IO.NodeOutput: length = round((seconds * 44100 / 2048) / 2) * 2 - latent = torch.zeros([batch_size, 64, length], device=self.device) - return ({"samples":latent, "type": "audio"}, ) + latent = torch.zeros([batch_size, 64, length], device=comfy.model_management.intermediate_device()) + return IO.NodeOutput({"samples":latent, "type": "audio"}) -class ConditioningStableAudio: + generate = execute # TODO: remove + + +class ConditioningStableAudio(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": {"positive": ("CONDITIONING", ), - "negative": ("CONDITIONING", ), - "seconds_start": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 1000.0, "step": 0.1}), - "seconds_total": ("FLOAT", {"default": 47.0, "min": 0.0, "max": 1000.0, "step": 0.1}), - }} + def define_schema(cls): + return IO.Schema( + node_id="ConditioningStableAudio", + category="conditioning", + inputs=[ + IO.Conditioning.Input("positive"), + IO.Conditioning.Input("negative"), + IO.Float.Input("seconds_start", default=0.0, min=0.0, max=1000.0, step=0.1), + IO.Float.Input("seconds_total", default=47.0, min=0.0, max=1000.0, step=0.1), + ], + outputs=[ + IO.Conditioning.Output(display_name="positive"), + IO.Conditioning.Output(display_name="negative"), + ], + ) - RETURN_TYPES = ("CONDITIONING","CONDITIONING") - RETURN_NAMES = ("positive", "negative") - - FUNCTION = "append" - - CATEGORY = "conditioning" - - def append(self, positive, negative, seconds_start, seconds_total): + @classmethod + def execute(cls, positive, negative, seconds_start, seconds_total) -> IO.NodeOutput: positive = node_helpers.conditioning_set_values(positive, {"seconds_start": seconds_start, "seconds_total": seconds_total}) negative = node_helpers.conditioning_set_values(negative, {"seconds_start": seconds_start, "seconds_total": seconds_total}) - return (positive, negative) + return IO.NodeOutput(positive, negative) -class VAEEncodeAudio: + append = execute # TODO: remove + + +class VAEEncodeAudio(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { "audio": ("AUDIO", ), "vae": ("VAE", )}} - RETURN_TYPES = ("LATENT",) - FUNCTION = "encode" + def define_schema(cls): + return IO.Schema( + node_id="VAEEncodeAudio", + display_name="VAE Encode Audio", + category="latent/audio", + inputs=[ + IO.Audio.Input("audio"), + IO.Vae.Input("vae"), + ], + outputs=[IO.Latent.Output()], + ) - CATEGORY = "latent/audio" - - def encode(self, vae, audio): + @classmethod + def execute(cls, vae, audio) -> IO.NodeOutput: sample_rate = audio["sample_rate"] if 44100 != sample_rate: waveform = torchaudio.functional.resample(audio["waveform"], sample_rate, 44100) @@ -72,213 +87,134 @@ class VAEEncodeAudio: waveform = audio["waveform"] t = vae.encode(waveform.movedim(1, -1)) - return ({"samples":t}, ) + return IO.NodeOutput({"samples":t}) -class VAEDecodeAudio: + encode = execute # TODO: remove + + +class VAEDecodeAudio(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { "samples": ("LATENT", ), "vae": ("VAE", )}} - RETURN_TYPES = ("AUDIO",) - FUNCTION = "decode" + def define_schema(cls): + return IO.Schema( + node_id="VAEDecodeAudio", + display_name="VAE Decode Audio", + category="latent/audio", + inputs=[ + IO.Latent.Input("samples"), + IO.Vae.Input("vae"), + ], + outputs=[IO.Audio.Output()], + ) - CATEGORY = "latent/audio" - - def decode(self, vae, samples): + @classmethod + def execute(cls, vae, samples) -> IO.NodeOutput: audio = vae.decode(samples["samples"]).movedim(-1, 1) std = torch.std(audio, dim=[1,2], keepdim=True) * 5.0 std[std < 1.0] = 1.0 audio /= std - return ({"waveform": audio, "sample_rate": 44100}, ) + return IO.NodeOutput({"waveform": audio, "sample_rate": 44100 if "sample_rate" not in samples else samples["sample_rate"]}) + + decode = execute # TODO: remove -def save_audio(self, audio, filename_prefix="ComfyUI", format="flac", prompt=None, extra_pnginfo=None, quality="128k"): - - filename_prefix += self.prefix_append - full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix, self.output_dir) - results: list[FileLocator] = [] - - # Prepare metadata dictionary - metadata = {} - if not args.disable_metadata: - if prompt is not None: - metadata["prompt"] = json.dumps(prompt) - if extra_pnginfo is not None: - for x in extra_pnginfo: - metadata[x] = json.dumps(extra_pnginfo[x]) - - # Opus supported sample rates - OPUS_RATES = [8000, 12000, 16000, 24000, 48000] - - for (batch_number, waveform) in enumerate(audio["waveform"].cpu()): - filename_with_batch_num = filename.replace("%batch_num%", str(batch_number)) - file = f"{filename_with_batch_num}_{counter:05}_.{format}" - output_path = os.path.join(full_output_folder, file) - - # Use original sample rate initially - sample_rate = audio["sample_rate"] - - # Handle Opus sample rate requirements - if format == "opus": - if sample_rate > 48000: - sample_rate = 48000 - elif sample_rate not in OPUS_RATES: - # Find the next highest supported rate - for rate in sorted(OPUS_RATES): - if rate > sample_rate: - sample_rate = rate - break - if sample_rate not in OPUS_RATES: # Fallback if still not supported - sample_rate = 48000 - - # Resample if necessary - if sample_rate != audio["sample_rate"]: - waveform = torchaudio.functional.resample(waveform, audio["sample_rate"], sample_rate) - - # Create output with specified format - output_buffer = io.BytesIO() - output_container = av.open(output_buffer, mode='w', format=format) - - # Set metadata on the container - for key, value in metadata.items(): - output_container.metadata[key] = value - - layout = 'mono' if waveform.shape[0] == 1 else 'stereo' - # Set up the output stream with appropriate properties - if format == "opus": - out_stream = output_container.add_stream("libopus", rate=sample_rate, layout=layout) - if quality == "64k": - out_stream.bit_rate = 64000 - elif quality == "96k": - out_stream.bit_rate = 96000 - elif quality == "128k": - out_stream.bit_rate = 128000 - elif quality == "192k": - out_stream.bit_rate = 192000 - elif quality == "320k": - out_stream.bit_rate = 320000 - elif format == "mp3": - out_stream = output_container.add_stream("libmp3lame", rate=sample_rate, layout=layout) - if quality == "V0": - #TODO i would really love to support V3 and V5 but there doesn't seem to be a way to set the qscale level, the property below is a bool - out_stream.codec_context.qscale = 1 - elif quality == "128k": - out_stream.bit_rate = 128000 - elif quality == "320k": - out_stream.bit_rate = 320000 - else: #format == "flac": - out_stream = output_container.add_stream("flac", rate=sample_rate, layout=layout) - - frame = av.AudioFrame.from_ndarray(waveform.movedim(0, 1).reshape(1, -1).float().numpy(), format='flt', layout=layout) - frame.sample_rate = sample_rate - frame.pts = 0 - output_container.mux(out_stream.encode(frame)) - - # Flush encoder - output_container.mux(out_stream.encode(None)) - - # Close containers - output_container.close() - - # Write the output to file - output_buffer.seek(0) - with open(output_path, 'wb') as f: - f.write(output_buffer.getbuffer()) - - results.append({ - "filename": file, - "subfolder": subfolder, - "type": self.type - }) - counter += 1 - - return { "ui": { "audio": results } } - -class SaveAudio: - def __init__(self): - self.output_dir = folder_paths.get_output_directory() - self.type = "output" - self.prefix_append = "" +class SaveAudio(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="SaveAudio", + display_name="Save Audio (FLAC)", + category="audio", + inputs=[ + IO.Audio.Input("audio"), + IO.String.Input("filename_prefix", default="audio/ComfyUI"), + ], + hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo], + is_output_node=True, + ) @classmethod - def INPUT_TYPES(s): - return {"required": { "audio": ("AUDIO", ), - "filename_prefix": ("STRING", {"default": "audio/ComfyUI"}), - }, - "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"}, - } + def execute(cls, audio, filename_prefix="ComfyUI", format="flac") -> IO.NodeOutput: + return IO.NodeOutput( + ui=UI.AudioSaveHelper.get_save_audio_ui(audio, filename_prefix=filename_prefix, cls=cls, format=format) + ) - RETURN_TYPES = () - FUNCTION = "save_flac" + save_flac = execute # TODO: remove - OUTPUT_NODE = True - CATEGORY = "audio" - - def save_flac(self, audio, filename_prefix="ComfyUI", format="flac", prompt=None, extra_pnginfo=None): - return save_audio(self, audio, filename_prefix, format, prompt, extra_pnginfo) - -class SaveAudioMP3: - def __init__(self): - self.output_dir = folder_paths.get_output_directory() - self.type = "output" - self.prefix_append = "" +class SaveAudioMP3(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="SaveAudioMP3", + display_name="Save Audio (MP3)", + category="audio", + inputs=[ + IO.Audio.Input("audio"), + IO.String.Input("filename_prefix", default="audio/ComfyUI"), + IO.Combo.Input("quality", options=["V0", "128k", "320k"], default="V0"), + ], + hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo], + is_output_node=True, + ) @classmethod - def INPUT_TYPES(s): - return {"required": { "audio": ("AUDIO", ), - "filename_prefix": ("STRING", {"default": "audio/ComfyUI"}), - "quality": (["V0", "128k", "320k"], {"default": "V0"}), - }, - "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"}, - } + def execute(cls, audio, filename_prefix="ComfyUI", format="mp3", quality="128k") -> IO.NodeOutput: + return IO.NodeOutput( + ui=UI.AudioSaveHelper.get_save_audio_ui( + audio, filename_prefix=filename_prefix, cls=cls, format=format, quality=quality + ) + ) - RETURN_TYPES = () - FUNCTION = "save_mp3" + save_mp3 = execute # TODO: remove - OUTPUT_NODE = True - CATEGORY = "audio" - - def save_mp3(self, audio, filename_prefix="ComfyUI", format="mp3", prompt=None, extra_pnginfo=None, quality="128k"): - return save_audio(self, audio, filename_prefix, format, prompt, extra_pnginfo, quality) - -class SaveAudioOpus: - def __init__(self): - self.output_dir = folder_paths.get_output_directory() - self.type = "output" - self.prefix_append = "" +class SaveAudioOpus(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="SaveAudioOpus", + display_name="Save Audio (Opus)", + category="audio", + inputs=[ + IO.Audio.Input("audio"), + IO.String.Input("filename_prefix", default="audio/ComfyUI"), + IO.Combo.Input("quality", options=["64k", "96k", "128k", "192k", "320k"], default="128k"), + ], + hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo], + is_output_node=True, + ) @classmethod - def INPUT_TYPES(s): - return {"required": { "audio": ("AUDIO", ), - "filename_prefix": ("STRING", {"default": "audio/ComfyUI"}), - "quality": (["64k", "96k", "128k", "192k", "320k"], {"default": "128k"}), - }, - "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"}, - } + def execute(cls, audio, filename_prefix="ComfyUI", format="opus", quality="V3") -> IO.NodeOutput: + return IO.NodeOutput( + ui=UI.AudioSaveHelper.get_save_audio_ui( + audio, filename_prefix=filename_prefix, cls=cls, format=format, quality=quality + ) + ) - RETURN_TYPES = () - FUNCTION = "save_opus" + save_opus = execute # TODO: remove - OUTPUT_NODE = True - CATEGORY = "audio" - - def save_opus(self, audio, filename_prefix="ComfyUI", format="opus", prompt=None, extra_pnginfo=None, quality="V3"): - return save_audio(self, audio, filename_prefix, format, prompt, extra_pnginfo, quality) - -class PreviewAudio(SaveAudio): - def __init__(self): - self.output_dir = folder_paths.get_temp_directory() - self.type = "temp" - self.prefix_append = "_temp_" + ''.join(random.choice("abcdefghijklmnopqrstupvxyz") for x in range(5)) +class PreviewAudio(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="PreviewAudio", + display_name="Preview Audio", + category="audio", + inputs=[ + IO.Audio.Input("audio"), + ], + hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo], + is_output_node=True, + ) @classmethod - def INPUT_TYPES(s): - return {"required": - {"audio": ("AUDIO", ), }, - "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"}, - } + def execute(cls, audio) -> IO.NodeOutput: + return IO.NodeOutput(ui=UI.PreviewAudio(audio, cls=cls)) + + save_flac = execute # TODO: remove + def f32_pcm(wav: torch.Tensor) -> torch.Tensor: """Convert audio to float 32 bits PCM format.""" @@ -316,26 +252,30 @@ def load(filepath: str) -> tuple[torch.Tensor, int]: wav = f32_pcm(wav) return wav, sr -class LoadAudio: +class LoadAudio(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): + def define_schema(cls): input_dir = folder_paths.get_input_directory() files = folder_paths.filter_files_content_types(os.listdir(input_dir), ["audio", "video"]) - return {"required": {"audio": (sorted(files), {"audio_upload": True})}} + return IO.Schema( + node_id="LoadAudio", + display_name="Load Audio", + category="audio", + inputs=[ + IO.Combo.Input("audio", upload=IO.UploadType.audio, options=sorted(files)), + ], + outputs=[IO.Audio.Output()], + ) - CATEGORY = "audio" - - RETURN_TYPES = ("AUDIO", ) - FUNCTION = "load" - - def load(self, audio): + @classmethod + def execute(cls, audio) -> IO.NodeOutput: audio_path = folder_paths.get_annotated_filepath(audio) waveform, sample_rate = load(audio_path) audio = {"waveform": waveform.unsqueeze(0), "sample_rate": sample_rate} - return (audio, ) + return IO.NodeOutput(audio) @classmethod - def IS_CHANGED(s, audio): + def fingerprint_inputs(cls, audio): image_path = folder_paths.get_annotated_filepath(audio) m = hashlib.sha256() with open(image_path, 'rb') as f: @@ -343,46 +283,69 @@ class LoadAudio: return m.digest().hex() @classmethod - def VALIDATE_INPUTS(s, audio): + def validate_inputs(cls, audio): if not folder_paths.exists_annotated_filepath(audio): return "Invalid audio file: {}".format(audio) return True -class RecordAudio: + load = execute # TODO: remove + + +class RecordAudio(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": {"audio": ("AUDIO_RECORD", {})}} + def define_schema(cls): + return IO.Schema( + node_id="RecordAudio", + display_name="Record Audio", + category="audio", + inputs=[ + IO.Custom("AUDIO_RECORD").Input("audio"), + ], + outputs=[IO.Audio.Output()], + ) - CATEGORY = "audio" - - RETURN_TYPES = ("AUDIO", ) - FUNCTION = "load" - - def load(self, audio): + @classmethod + def execute(cls, audio) -> IO.NodeOutput: audio_path = folder_paths.get_annotated_filepath(audio) waveform, sample_rate = load(audio_path) audio = {"waveform": waveform.unsqueeze(0), "sample_rate": sample_rate} - return (audio, ) + return IO.NodeOutput(audio) + + load = execute # TODO: remove -class TrimAudioDuration: +class TrimAudioDuration(IO.ComfyNode): @classmethod - def INPUT_TYPES(cls): - return { - "required": { - "audio": ("AUDIO",), - "start_index": ("FLOAT", {"default": 0.0, "min": -0xffffffffffffffff, "max": 0xffffffffffffffff, "step": 0.01, "tooltip": "Start time in seconds, can be negative to count from the end (supports sub-seconds)."}), - "duration": ("FLOAT", {"default": 60.0, "min": 0.0, "step": 0.01, "tooltip": "Duration in seconds"}), - }, - } + def define_schema(cls): + return IO.Schema( + node_id="TrimAudioDuration", + display_name="Trim Audio Duration", + description="Trim audio tensor into chosen time range.", + category="audio", + inputs=[ + IO.Audio.Input("audio"), + IO.Float.Input( + "start_index", + default=0.0, + min=-0xffffffffffffffff, + max=0xffffffffffffffff, + step=0.01, + tooltip="Start time in seconds, can be negative to count from the end (supports sub-seconds).", + ), + IO.Float.Input( + "duration", + default=60.0, + min=0.0, + step=0.01, + tooltip="Duration in seconds", + ), + ], + outputs=[IO.Audio.Output()], + ) - FUNCTION = "trim" - RETURN_TYPES = ("AUDIO",) - CATEGORY = "audio" - DESCRIPTION = "Trim audio tensor into chosen time range." - - def trim(self, audio, start_index, duration): + @classmethod + def execute(cls, audio, start_index, duration) -> IO.NodeOutput: waveform = audio["waveform"] sample_rate = audio["sample_rate"] audio_length = waveform.shape[-1] @@ -399,23 +362,30 @@ class TrimAudioDuration: if start_frame >= end_frame: raise ValueError("AudioTrim: Start time must be less than end time and be within the audio length.") - return ({"waveform": waveform[..., start_frame:end_frame], "sample_rate": sample_rate},) + return IO.NodeOutput({"waveform": waveform[..., start_frame:end_frame], "sample_rate": sample_rate}) + + trim = execute # TODO: remove -class SplitAudioChannels: +class SplitAudioChannels(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { - "audio": ("AUDIO",), - }} + def define_schema(cls): + return IO.Schema( + node_id="SplitAudioChannels", + display_name="Split Audio Channels", + description="Separates the audio into left and right channels.", + category="audio", + inputs=[ + IO.Audio.Input("audio"), + ], + outputs=[ + IO.Audio.Output(display_name="left"), + IO.Audio.Output(display_name="right"), + ], + ) - RETURN_TYPES = ("AUDIO", "AUDIO") - RETURN_NAMES = ("left", "right") - FUNCTION = "separate" - CATEGORY = "audio" - DESCRIPTION = "Separates the audio into left and right channels." - - def separate(self, audio): + @classmethod + def execute(cls, audio) -> IO.NodeOutput: waveform = audio["waveform"] sample_rate = audio["sample_rate"] @@ -425,7 +395,9 @@ class SplitAudioChannels: left_channel = waveform[..., 0:1, :] right_channel = waveform[..., 1:2, :] - return ({"waveform": left_channel, "sample_rate": sample_rate}, {"waveform": right_channel, "sample_rate": sample_rate}) + return IO.NodeOutput({"waveform": left_channel, "sample_rate": sample_rate}, {"waveform": right_channel, "sample_rate": sample_rate}) + + separate = execute # TODO: remove def match_audio_sample_rates(waveform_1, sample_rate_1, waveform_2, sample_rate_2): @@ -443,21 +415,29 @@ def match_audio_sample_rates(waveform_1, sample_rate_1, waveform_2, sample_rate_ return waveform_1, waveform_2, output_sample_rate -class AudioConcat: +class AudioConcat(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { - "audio1": ("AUDIO",), - "audio2": ("AUDIO",), - "direction": (['after', 'before'], {"default": 'after', "tooltip": "Whether to append audio2 after or before audio1."}), - }} + def define_schema(cls): + return IO.Schema( + node_id="AudioConcat", + display_name="Audio Concat", + description="Concatenates the audio1 to audio2 in the specified direction.", + category="audio", + inputs=[ + IO.Audio.Input("audio1"), + IO.Audio.Input("audio2"), + IO.Combo.Input( + "direction", + options=['after', 'before'], + default="after", + tooltip="Whether to append audio2 after or before audio1.", + ) + ], + outputs=[IO.Audio.Output()], + ) - RETURN_TYPES = ("AUDIO",) - FUNCTION = "concat" - CATEGORY = "audio" - DESCRIPTION = "Concatenates the audio1 to audio2 in the specified direction." - - def concat(self, audio1, audio2, direction): + @classmethod + def execute(cls, audio1, audio2, direction) -> IO.NodeOutput: waveform_1 = audio1["waveform"] waveform_2 = audio2["waveform"] sample_rate_1 = audio1["sample_rate"] @@ -477,26 +457,33 @@ class AudioConcat: elif direction == 'before': concatenated_audio = torch.cat((waveform_2, waveform_1), dim=2) - return ({"waveform": concatenated_audio, "sample_rate": output_sample_rate},) + return IO.NodeOutput({"waveform": concatenated_audio, "sample_rate": output_sample_rate}) + + concat = execute # TODO: remove -class AudioMerge: +class AudioMerge(IO.ComfyNode): @classmethod - def INPUT_TYPES(cls): - return { - "required": { - "audio1": ("AUDIO",), - "audio2": ("AUDIO",), - "merge_method": (["add", "mean", "subtract", "multiply"], {"tooltip": "The method used to combine the audio waveforms."}), - }, - } + def define_schema(cls): + return IO.Schema( + node_id="AudioMerge", + display_name="Audio Merge", + description="Combine two audio tracks by overlaying their waveforms.", + category="audio", + inputs=[ + IO.Audio.Input("audio1"), + IO.Audio.Input("audio2"), + IO.Combo.Input( + "merge_method", + options=["add", "mean", "subtract", "multiply"], + tooltip="The method used to combine the audio waveforms.", + ) + ], + outputs=[IO.Audio.Output()], + ) - FUNCTION = "merge" - RETURN_TYPES = ("AUDIO",) - CATEGORY = "audio" - DESCRIPTION = "Combine two audio tracks by overlaying their waveforms." - - def merge(self, audio1, audio2, merge_method): + @classmethod + def execute(cls, audio1, audio2, merge_method) -> IO.NodeOutput: waveform_1 = audio1["waveform"] waveform_2 = audio2["waveform"] sample_rate_1 = audio1["sample_rate"] @@ -530,85 +517,110 @@ class AudioMerge: if max_val > 1.0: waveform = waveform / max_val - return ({"waveform": waveform, "sample_rate": output_sample_rate},) + return IO.NodeOutput({"waveform": waveform, "sample_rate": output_sample_rate}) + + merge = execute # TODO: remove -class AudioAdjustVolume: +class AudioAdjustVolume(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { - "audio": ("AUDIO",), - "volume": ("INT", {"default": 1.0, "min": -100, "max": 100, "tooltip": "Volume adjustment in decibels (dB). 0 = no change, +6 = double, -6 = half, etc"}), - }} + def define_schema(cls): + return IO.Schema( + node_id="AudioAdjustVolume", + display_name="Audio Adjust Volume", + category="audio", + inputs=[ + IO.Audio.Input("audio"), + IO.Int.Input( + "volume", + default=1, + min=-100, + max=100, + tooltip="Volume adjustment in decibels (dB). 0 = no change, +6 = double, -6 = half, etc", + ) + ], + outputs=[IO.Audio.Output()], + ) - RETURN_TYPES = ("AUDIO",) - FUNCTION = "adjust_volume" - CATEGORY = "audio" - - def adjust_volume(self, audio, volume): + @classmethod + def execute(cls, audio, volume) -> IO.NodeOutput: if volume == 0: - return (audio,) + return IO.NodeOutput(audio) waveform = audio["waveform"] sample_rate = audio["sample_rate"] gain = 10 ** (volume / 20) waveform = waveform * gain - return ({"waveform": waveform, "sample_rate": sample_rate},) + return IO.NodeOutput({"waveform": waveform, "sample_rate": sample_rate}) + + adjust_volume = execute # TODO: remove -class EmptyAudio: +class EmptyAudio(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { - "duration": ("FLOAT", {"default": 60.0, "min": 0.0, "max": 0xffffffffffffffff, "step": 0.01, "tooltip": "Duration of the empty audio clip in seconds"}), - "sample_rate": ("INT", {"default": 44100, "tooltip": "Sample rate of the empty audio clip."}), - "channels": ("INT", {"default": 2, "min": 1, "max": 2, "tooltip": "Number of audio channels (1 for mono, 2 for stereo)."}), - }} + def define_schema(cls): + return IO.Schema( + node_id="EmptyAudio", + display_name="Empty Audio", + category="audio", + inputs=[ + IO.Float.Input( + "duration", + default=60.0, + min=0.0, + max=0xffffffffffffffff, + step=0.01, + tooltip="Duration of the empty audio clip in seconds", + ), + IO.Int.Input( + "sample_rate", + default=44100, + tooltip="Sample rate of the empty audio clip.", + min=1, + max=192000, + ), + IO.Int.Input( + "channels", + default=2, + min=1, + max=2, + tooltip="Number of audio channels (1 for mono, 2 for stereo).", + ), + ], + outputs=[IO.Audio.Output()], + ) - RETURN_TYPES = ("AUDIO",) - FUNCTION = "create_empty_audio" - CATEGORY = "audio" - - def create_empty_audio(self, duration, sample_rate, channels): + @classmethod + def execute(cls, duration, sample_rate, channels) -> IO.NodeOutput: num_samples = int(round(duration * sample_rate)) waveform = torch.zeros((1, channels, num_samples), dtype=torch.float32) - return ({"waveform": waveform, "sample_rate": sample_rate},) + return IO.NodeOutput({"waveform": waveform, "sample_rate": sample_rate}) + + create_empty_audio = execute # TODO: remove -NODE_CLASS_MAPPINGS = { - "EmptyLatentAudio": EmptyLatentAudio, - "VAEEncodeAudio": VAEEncodeAudio, - "VAEDecodeAudio": VAEDecodeAudio, - "SaveAudio": SaveAudio, - "SaveAudioMP3": SaveAudioMP3, - "SaveAudioOpus": SaveAudioOpus, - "LoadAudio": LoadAudio, - "PreviewAudio": PreviewAudio, - "ConditioningStableAudio": ConditioningStableAudio, - "RecordAudio": RecordAudio, - "TrimAudioDuration": TrimAudioDuration, - "SplitAudioChannels": SplitAudioChannels, - "AudioConcat": AudioConcat, - "AudioMerge": AudioMerge, - "AudioAdjustVolume": AudioAdjustVolume, - "EmptyAudio": EmptyAudio, -} +class AudioExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + EmptyLatentAudio, + VAEEncodeAudio, + VAEDecodeAudio, + SaveAudio, + SaveAudioMP3, + SaveAudioOpus, + LoadAudio, + PreviewAudio, + ConditioningStableAudio, + RecordAudio, + TrimAudioDuration, + SplitAudioChannels, + AudioConcat, + AudioMerge, + AudioAdjustVolume, + EmptyAudio, + ] -NODE_DISPLAY_NAME_MAPPINGS = { - "EmptyLatentAudio": "Empty Latent Audio", - "VAEEncodeAudio": "VAE Encode Audio", - "VAEDecodeAudio": "VAE Decode Audio", - "PreviewAudio": "Preview Audio", - "LoadAudio": "Load Audio", - "SaveAudio": "Save Audio (FLAC)", - "SaveAudioMP3": "Save Audio (MP3)", - "SaveAudioOpus": "Save Audio (Opus)", - "RecordAudio": "Record Audio", - "TrimAudioDuration": "Trim Audio Duration", - "SplitAudioChannels": "Split Audio Channels", - "AudioConcat": "Audio Concat", - "AudioMerge": "Audio Merge", - "AudioAdjustVolume": "Audio Adjust Volume", - "EmptyAudio": "Empty Audio", -} +async def comfy_entrypoint() -> AudioExtension: + return AudioExtension() diff --git a/comfy_extras/nodes_context_windows.py b/comfy_extras/nodes_context_windows.py index 1c3d9e697..3799a9004 100644 --- a/comfy_extras/nodes_context_windows.py +++ b/comfy_extras/nodes_context_windows.py @@ -26,6 +26,9 @@ class ContextWindowsManualNode(io.ComfyNode): io.Boolean.Input("closed_loop", default=False, tooltip="Whether to close the context window loop; only applicable to looped schedules."), io.Combo.Input("fuse_method", options=comfy.context_windows.ContextFuseMethods.LIST_STATIC, default=comfy.context_windows.ContextFuseMethods.PYRAMID, tooltip="The method to use to fuse the context windows."), io.Int.Input("dim", min=0, max=5, default=0, tooltip="The dimension to apply the context windows to."), + 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."), ], outputs=[ io.Model.Output(tooltip="The model with context windows applied during sampling."), @@ -34,7 +37,8 @@ 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) -> io.Model: + 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: model = model.clone() model.model_options["context_handler"] = comfy.context_windows.IndexListContextHandler( context_schedule=comfy.context_windows.get_matching_context_schedule(context_schedule), @@ -43,9 +47,15 @@ class ContextWindowsManualNode(io.ComfyNode): context_overlap=context_overlap, context_stride=context_stride, closed_loop=closed_loop, - dim=dim) + dim=dim, + freenoise=freenoise, + cond_retain_index_list=cond_retain_index_list, + split_conds_to_windows=split_conds_to_windows + ) # make memory usage calculation only take into account the context window latents comfy.context_windows.create_prepare_sampling_wrapper(model) + if freenoise: # no other use for this wrapper at this time + comfy.context_windows.create_sampler_sample_wrapper(model) return io.NodeOutput(model) class WanContextWindowsManualNode(ContextWindowsManualNode): @@ -68,14 +78,18 @@ class WanContextWindowsManualNode(ContextWindowsManualNode): io.Int.Input("context_stride", min=1, default=1, tooltip="The stride of the context window; only applicable to uniform schedules."), io.Boolean.Input("closed_loop", default=False, tooltip="Whether to close the context window loop; only applicable to looped schedules."), io.Combo.Input("fuse_method", options=comfy.context_windows.ContextFuseMethods.LIST_STATIC, default=comfy.context_windows.ContextFuseMethods.PYRAMID, tooltip="The method to use to fuse the context windows."), + 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."), ] return schema @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) -> io.Model: + 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, freenoise: bool, + cond_retain_index_list: list[int]=[], split_conds_to_windows: bool=False) -> io.Model: context_length = max(((context_length - 1) // 4) + 1, 1) # at least length 1 context_overlap = max(((context_overlap - 1) // 4) + 1, 0) # at least overlap 0 - return super().execute(model, context_length, context_overlap, context_schedule, context_stride, closed_loop, fuse_method, dim=2) + return super().execute(model, context_length, context_overlap, context_schedule, context_stride, closed_loop, fuse_method, dim=2, freenoise=freenoise, cond_retain_index_list=cond_retain_index_list, split_conds_to_windows=split_conds_to_windows) class ContextWindowsExtension(ComfyExtension): diff --git a/comfy_extras/nodes_custom_sampler.py b/comfy_extras/nodes_custom_sampler.py index d011f433b..f19adf4b9 100644 --- a/comfy_extras/nodes_custom_sampler.py +++ b/comfy_extras/nodes_custom_sampler.py @@ -3,272 +3,313 @@ import comfy.samplers import comfy.sample from comfy.k_diffusion import sampling as k_diffusion_sampling from comfy.k_diffusion import sa_solver -from comfy.comfy_types import IO, ComfyNodeABC, InputTypeDict import latent_preview import torch import comfy.utils import node_helpers +from typing_extensions import override +from comfy_api.latest import ComfyExtension, io +import re -class BasicScheduler: +class BasicScheduler(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"model": ("MODEL",), - "scheduler": (comfy.samplers.SCHEDULER_NAMES, ), - "steps": ("INT", {"default": 20, "min": 1, "max": 10000}), - "denoise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}), - } - } - RETURN_TYPES = ("SIGMAS",) - CATEGORY = "sampling/custom_sampling/schedulers" + def define_schema(cls): + return io.Schema( + node_id="BasicScheduler", + category="sampling/custom_sampling/schedulers", + inputs=[ + io.Model.Input("model"), + io.Combo.Input("scheduler", options=comfy.samplers.SCHEDULER_NAMES), + io.Int.Input("steps", default=20, min=1, max=10000), + io.Float.Input("denoise", default=1.0, min=0.0, max=1.0, step=0.01), + ], + outputs=[io.Sigmas.Output()] + ) - FUNCTION = "get_sigmas" - - def get_sigmas(self, model, scheduler, steps, denoise): + @classmethod + def execute(cls, model, scheduler, steps, denoise) -> io.NodeOutput: total_steps = steps if denoise < 1.0: if denoise <= 0.0: - return (torch.FloatTensor([]),) + return io.NodeOutput(torch.FloatTensor([])) total_steps = int(steps/denoise) sigmas = comfy.samplers.calculate_sigmas(model.get_model_object("model_sampling"), scheduler, total_steps).cpu() sigmas = sigmas[-(steps + 1):] - return (sigmas, ) + return io.NodeOutput(sigmas) + + get_sigmas = execute -class KarrasScheduler: +class KarrasScheduler(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"steps": ("INT", {"default": 20, "min": 1, "max": 10000}), - "sigma_max": ("FLOAT", {"default": 14.614642, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), - "sigma_min": ("FLOAT", {"default": 0.0291675, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), - "rho": ("FLOAT", {"default": 7.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - } - } - RETURN_TYPES = ("SIGMAS",) - CATEGORY = "sampling/custom_sampling/schedulers" + def define_schema(cls): + return io.Schema( + node_id="KarrasScheduler", + category="sampling/custom_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), + io.Float.Input("sigma_min", default=0.0291675, min=0.0, max=5000.0, step=0.01, round=False), + io.Float.Input("rho", default=7.0, min=0.0, max=100.0, step=0.01, round=False), + ], + outputs=[io.Sigmas.Output()] + ) - FUNCTION = "get_sigmas" - - def get_sigmas(self, steps, sigma_max, sigma_min, rho): + @classmethod + def execute(cls, steps, sigma_max, sigma_min, rho) -> io.NodeOutput: sigmas = k_diffusion_sampling.get_sigmas_karras(n=steps, sigma_min=sigma_min, sigma_max=sigma_max, rho=rho) - return (sigmas, ) + return io.NodeOutput(sigmas) -class ExponentialScheduler: + get_sigmas = execute + +class ExponentialScheduler(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"steps": ("INT", {"default": 20, "min": 1, "max": 10000}), - "sigma_max": ("FLOAT", {"default": 14.614642, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), - "sigma_min": ("FLOAT", {"default": 0.0291675, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), - } - } - RETURN_TYPES = ("SIGMAS",) - CATEGORY = "sampling/custom_sampling/schedulers" + def define_schema(cls): + return io.Schema( + node_id="ExponentialScheduler", + category="sampling/custom_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), + io.Float.Input("sigma_min", default=0.0291675, min=0.0, max=5000.0, step=0.01, round=False), + ], + outputs=[io.Sigmas.Output()] + ) - FUNCTION = "get_sigmas" - - def get_sigmas(self, steps, sigma_max, sigma_min): + @classmethod + def execute(cls, steps, sigma_max, sigma_min) -> io.NodeOutput: sigmas = k_diffusion_sampling.get_sigmas_exponential(n=steps, sigma_min=sigma_min, sigma_max=sigma_max) - return (sigmas, ) + return io.NodeOutput(sigmas) -class PolyexponentialScheduler: + get_sigmas = execute + +class PolyexponentialScheduler(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"steps": ("INT", {"default": 20, "min": 1, "max": 10000}), - "sigma_max": ("FLOAT", {"default": 14.614642, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), - "sigma_min": ("FLOAT", {"default": 0.0291675, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), - "rho": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - } - } - RETURN_TYPES = ("SIGMAS",) - CATEGORY = "sampling/custom_sampling/schedulers" + def define_schema(cls): + return io.Schema( + node_id="PolyexponentialScheduler", + category="sampling/custom_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), + io.Float.Input("sigma_min", default=0.0291675, min=0.0, max=5000.0, step=0.01, round=False), + io.Float.Input("rho", default=1.0, min=0.0, max=100.0, step=0.01, round=False), + ], + outputs=[io.Sigmas.Output()] + ) - FUNCTION = "get_sigmas" - - def get_sigmas(self, steps, sigma_max, sigma_min, rho): + @classmethod + def execute(cls, steps, sigma_max, sigma_min, rho) -> io.NodeOutput: sigmas = k_diffusion_sampling.get_sigmas_polyexponential(n=steps, sigma_min=sigma_min, sigma_max=sigma_max, rho=rho) - return (sigmas, ) + return io.NodeOutput(sigmas) -class LaplaceScheduler: + get_sigmas = execute + +class LaplaceScheduler(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"steps": ("INT", {"default": 20, "min": 1, "max": 10000}), - "sigma_max": ("FLOAT", {"default": 14.614642, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), - "sigma_min": ("FLOAT", {"default": 0.0291675, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), - "mu": ("FLOAT", {"default": 0.0, "min": -10.0, "max": 10.0, "step":0.1, "round": False}), - "beta": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 10.0, "step":0.1, "round": False}), - } - } - RETURN_TYPES = ("SIGMAS",) - CATEGORY = "sampling/custom_sampling/schedulers" + def define_schema(cls): + return io.Schema( + node_id="LaplaceScheduler", + category="sampling/custom_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), + io.Float.Input("sigma_min", default=0.0291675, min=0.0, max=5000.0, step=0.01, round=False), + io.Float.Input("mu", default=0.0, min=-10.0, max=10.0, step=0.1, round=False), + io.Float.Input("beta", default=0.5, min=0.0, max=10.0, step=0.1, round=False), + ], + outputs=[io.Sigmas.Output()] + ) - FUNCTION = "get_sigmas" - - def get_sigmas(self, steps, sigma_max, sigma_min, mu, beta): + @classmethod + def execute(cls, steps, sigma_max, sigma_min, mu, beta) -> io.NodeOutput: sigmas = k_diffusion_sampling.get_sigmas_laplace(n=steps, sigma_min=sigma_min, sigma_max=sigma_max, mu=mu, beta=beta) - return (sigmas, ) + return io.NodeOutput(sigmas) + + get_sigmas = execute -class SDTurboScheduler: +class SDTurboScheduler(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"model": ("MODEL",), - "steps": ("INT", {"default": 1, "min": 1, "max": 10}), - "denoise": ("FLOAT", {"default": 1.0, "min": 0, "max": 1.0, "step": 0.01}), - } - } - RETURN_TYPES = ("SIGMAS",) - CATEGORY = "sampling/custom_sampling/schedulers" + def define_schema(cls): + return io.Schema( + node_id="SDTurboScheduler", + category="sampling/custom_sampling/schedulers", + inputs=[ + io.Model.Input("model"), + io.Int.Input("steps", default=1, min=1, max=10), + io.Float.Input("denoise", default=1.0, min=0, max=1.0, step=0.01), + ], + outputs=[io.Sigmas.Output()] + ) - FUNCTION = "get_sigmas" - - def get_sigmas(self, model, steps, denoise): + @classmethod + def execute(cls, model, steps, denoise) -> io.NodeOutput: start_step = 10 - int(10 * denoise) timesteps = torch.flip(torch.arange(1, 11) * 100 - 1, (0,))[start_step:start_step + steps] sigmas = model.get_model_object("model_sampling").sigma(timesteps) sigmas = torch.cat([sigmas, sigmas.new_zeros([1])]) - return (sigmas, ) + return io.NodeOutput(sigmas) -class BetaSamplingScheduler: + get_sigmas = execute + +class BetaSamplingScheduler(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"model": ("MODEL",), - "steps": ("INT", {"default": 20, "min": 1, "max": 10000}), - "alpha": ("FLOAT", {"default": 0.6, "min": 0.0, "max": 50.0, "step":0.01, "round": False}), - "beta": ("FLOAT", {"default": 0.6, "min": 0.0, "max": 50.0, "step":0.01, "round": False}), - } - } - RETURN_TYPES = ("SIGMAS",) - CATEGORY = "sampling/custom_sampling/schedulers" + def define_schema(cls): + return io.Schema( + node_id="BetaSamplingScheduler", + category="sampling/custom_sampling/schedulers", + inputs=[ + io.Model.Input("model"), + io.Int.Input("steps", default=20, min=1, max=10000), + io.Float.Input("alpha", default=0.6, min=0.0, max=50.0, step=0.01, round=False), + io.Float.Input("beta", default=0.6, min=0.0, max=50.0, step=0.01, round=False), + ], + outputs=[io.Sigmas.Output()] + ) - FUNCTION = "get_sigmas" - - def get_sigmas(self, model, steps, alpha, beta): + @classmethod + def execute(cls, model, steps, alpha, beta) -> io.NodeOutput: sigmas = comfy.samplers.beta_scheduler(model.get_model_object("model_sampling"), steps, alpha=alpha, beta=beta) - return (sigmas, ) + return io.NodeOutput(sigmas) -class VPScheduler: + get_sigmas = execute + +class VPScheduler(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"steps": ("INT", {"default": 20, "min": 1, "max": 10000}), - "beta_d": ("FLOAT", {"default": 19.9, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), #TODO: fix default values - "beta_min": ("FLOAT", {"default": 0.1, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), - "eps_s": ("FLOAT", {"default": 0.001, "min": 0.0, "max": 1.0, "step":0.0001, "round": False}), - } - } - RETURN_TYPES = ("SIGMAS",) - CATEGORY = "sampling/custom_sampling/schedulers" + def define_schema(cls): + return io.Schema( + node_id="VPScheduler", + category="sampling/custom_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), #TODO: fix default values + io.Float.Input("beta_min", default=0.1, min=0.0, max=5000.0, step=0.01, round=False), + io.Float.Input("eps_s", default=0.001, min=0.0, max=1.0, step=0.0001, round=False), + ], + outputs=[io.Sigmas.Output()] + ) - FUNCTION = "get_sigmas" - - def get_sigmas(self, steps, beta_d, beta_min, eps_s): + @classmethod + def execute(cls, steps, beta_d, beta_min, eps_s) -> io.NodeOutput: sigmas = k_diffusion_sampling.get_sigmas_vp(n=steps, beta_d=beta_d, beta_min=beta_min, eps_s=eps_s) - return (sigmas, ) + return io.NodeOutput(sigmas) -class SplitSigmas: + get_sigmas = execute + +class SplitSigmas(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"sigmas": ("SIGMAS", ), - "step": ("INT", {"default": 0, "min": 0, "max": 10000}), - } - } - RETURN_TYPES = ("SIGMAS","SIGMAS") - RETURN_NAMES = ("high_sigmas", "low_sigmas") - CATEGORY = "sampling/custom_sampling/sigmas" + def define_schema(cls): + return io.Schema( + node_id="SplitSigmas", + category="sampling/custom_sampling/sigmas", + inputs=[ + io.Sigmas.Input("sigmas"), + io.Int.Input("step", default=0, min=0, max=10000), + ], + outputs=[ + io.Sigmas.Output(display_name="high_sigmas"), + io.Sigmas.Output(display_name="low_sigmas"), + ] + ) - FUNCTION = "get_sigmas" - - def get_sigmas(self, sigmas, step): + @classmethod + def execute(cls, sigmas, step) -> io.NodeOutput: sigmas1 = sigmas[:step + 1] sigmas2 = sigmas[step:] - return (sigmas1, sigmas2) + return io.NodeOutput(sigmas1, sigmas2) -class SplitSigmasDenoise: + get_sigmas = execute + +class SplitSigmasDenoise(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"sigmas": ("SIGMAS", ), - "denoise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}), - } - } - RETURN_TYPES = ("SIGMAS","SIGMAS") - RETURN_NAMES = ("high_sigmas", "low_sigmas") - CATEGORY = "sampling/custom_sampling/sigmas" + def define_schema(cls): + return io.Schema( + node_id="SplitSigmasDenoise", + category="sampling/custom_sampling/sigmas", + inputs=[ + io.Sigmas.Input("sigmas"), + io.Float.Input("denoise", default=1.0, min=0.0, max=1.0, step=0.01), + ], + outputs=[ + io.Sigmas.Output(display_name="high_sigmas"), + io.Sigmas.Output(display_name="low_sigmas"), + ] + ) - FUNCTION = "get_sigmas" - - def get_sigmas(self, sigmas, denoise): + @classmethod + def execute(cls, sigmas, denoise) -> io.NodeOutput: steps = max(sigmas.shape[-1] - 1, 0) total_steps = round(steps * denoise) sigmas1 = sigmas[:-(total_steps)] sigmas2 = sigmas[-(total_steps + 1):] - return (sigmas1, sigmas2) + return io.NodeOutput(sigmas1, sigmas2) -class FlipSigmas: + get_sigmas = execute + +class FlipSigmas(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"sigmas": ("SIGMAS", ), - } - } - RETURN_TYPES = ("SIGMAS",) - CATEGORY = "sampling/custom_sampling/sigmas" + def define_schema(cls): + return io.Schema( + node_id="FlipSigmas", + category="sampling/custom_sampling/sigmas", + inputs=[io.Sigmas.Input("sigmas")], + outputs=[io.Sigmas.Output()] + ) - FUNCTION = "get_sigmas" - - def get_sigmas(self, sigmas): + @classmethod + def execute(cls, sigmas) -> io.NodeOutput: if len(sigmas) == 0: - return (sigmas,) + return io.NodeOutput(sigmas) sigmas = sigmas.flip(0) if sigmas[0] == 0: sigmas[0] = 0.0001 - return (sigmas,) + return io.NodeOutput(sigmas) -class SetFirstSigma: + get_sigmas = execute + +class SetFirstSigma(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"sigmas": ("SIGMAS", ), - "sigma": ("FLOAT", {"default": 136.0, "min": 0.0, "max": 20000.0, "step": 0.001, "round": False}), - } - } - RETURN_TYPES = ("SIGMAS",) - CATEGORY = "sampling/custom_sampling/sigmas" + def define_schema(cls): + return io.Schema( + node_id="SetFirstSigma", + category="sampling/custom_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), + ], + outputs=[io.Sigmas.Output()] + ) - FUNCTION = "set_first_sigma" - - def set_first_sigma(self, sigmas, sigma): + @classmethod + def execute(cls, sigmas, sigma) -> io.NodeOutput: sigmas = sigmas.clone() sigmas[0] = sigma - return (sigmas, ) + return io.NodeOutput(sigmas) -class ExtendIntermediateSigmas: + set_first_sigma = execute + +class ExtendIntermediateSigmas(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"sigmas": ("SIGMAS", ), - "steps": ("INT", {"default": 2, "min": 1, "max": 100}), - "start_at_sigma": ("FLOAT", {"default": -1.0, "min": -1.0, "max": 20000.0, "step": 0.01, "round": False}), - "end_at_sigma": ("FLOAT", {"default": 12.0, "min": 0.0, "max": 20000.0, "step": 0.01, "round": False}), - "spacing": (['linear', 'cosine', 'sine'],), - } - } - RETURN_TYPES = ("SIGMAS",) - CATEGORY = "sampling/custom_sampling/sigmas" + def define_schema(cls): + return io.Schema( + node_id="ExtendIntermediateSigmas", + category="sampling/custom_sampling/sigmas", + inputs=[ + io.Sigmas.Input("sigmas"), + io.Int.Input("steps", default=2, min=1, max=100), + io.Float.Input("start_at_sigma", default=-1.0, min=-1.0, max=20000.0, step=0.01, round=False), + io.Float.Input("end_at_sigma", default=12.0, min=0.0, max=20000.0, step=0.01, round=False), + io.Combo.Input("spacing", options=['linear', 'cosine', 'sine']), + ], + outputs=[io.Sigmas.Output()] + ) - FUNCTION = "extend" - - def extend(self, sigmas: torch.Tensor, steps: int, start_at_sigma: float, end_at_sigma: float, spacing: str): + @classmethod + def execute(cls, sigmas: torch.Tensor, steps: int, start_at_sigma: float, end_at_sigma: float, spacing: str) -> io.NodeOutput: if start_at_sigma < 0: start_at_sigma = float("inf") @@ -299,27 +340,27 @@ class ExtendIntermediateSigmas: extended_sigmas = torch.FloatTensor(extended_sigmas) - return (extended_sigmas,) + return io.NodeOutput(extended_sigmas) + + extend = execute -class SamplingPercentToSigma: +class SamplingPercentToSigma(io.ComfyNode): @classmethod - def INPUT_TYPES(cls) -> InputTypeDict: - return { - "required": { - "model": (IO.MODEL, {}), - "sampling_percent": (IO.FLOAT, {"default": 0.0, "min": 0.0, "max": 1.0, "step": 0.0001}), - "return_actual_sigma": (IO.BOOLEAN, {"default": False, "tooltip": "Return the actual sigma value instead of the value used for interval checks.\nThis only affects results at 0.0 and 1.0."}), - } - } + def define_schema(cls): + return io.Schema( + node_id="SamplingPercentToSigma", + category="sampling/custom_sampling/sigmas", + inputs=[ + io.Model.Input("model"), + io.Float.Input("sampling_percent", default=0.0, min=0.0, max=1.0, step=0.0001), + io.Boolean.Input("return_actual_sigma", default=False, tooltip="Return the actual sigma value instead of the value used for interval checks.\nThis only affects results at 0.0 and 1.0."), + ], + outputs=[io.Float.Output(display_name="sigma_value")] + ) - RETURN_TYPES = (IO.FLOAT,) - RETURN_NAMES = ("sigma_value",) - CATEGORY = "sampling/custom_sampling/sigmas" - - FUNCTION = "get_sigma" - - def get_sigma(self, model, sampling_percent, return_actual_sigma): + @classmethod + def execute(cls, model, sampling_percent, return_actual_sigma) -> io.NodeOutput: model_sampling = model.get_model_object("model_sampling") sigma_val = model_sampling.percent_to_sigma(sampling_percent) if return_actual_sigma: @@ -327,212 +368,234 @@ class SamplingPercentToSigma: sigma_val = model_sampling.sigma_max.item() elif sampling_percent == 1.0: sigma_val = model_sampling.sigma_min.item() - return (sigma_val,) + return io.NodeOutput(sigma_val) + + get_sigma = execute -class KSamplerSelect: +class KSamplerSelect(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"sampler_name": (comfy.samplers.SAMPLER_NAMES, ), - } - } - RETURN_TYPES = ("SAMPLER",) - CATEGORY = "sampling/custom_sampling/samplers" + def define_schema(cls): + return io.Schema( + node_id="KSamplerSelect", + category="sampling/custom_sampling/samplers", + inputs=[io.Combo.Input("sampler_name", options=comfy.samplers.SAMPLER_NAMES)], + outputs=[io.Sampler.Output()] + ) - FUNCTION = "get_sampler" - - def get_sampler(self, sampler_name): + @classmethod + def execute(cls, sampler_name) -> io.NodeOutput: sampler = comfy.samplers.sampler_object(sampler_name) - return (sampler, ) + return io.NodeOutput(sampler) -class SamplerDPMPP_3M_SDE: + get_sampler = execute + +class SamplerDPMPP_3M_SDE(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"eta": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "s_noise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "noise_device": (['gpu', 'cpu'], ), - } - } - RETURN_TYPES = ("SAMPLER",) - CATEGORY = "sampling/custom_sampling/samplers" + def define_schema(cls): + return io.Schema( + node_id="SamplerDPMPP_3M_SDE", + category="sampling/custom_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), + io.Combo.Input("noise_device", options=['gpu', 'cpu']), + ], + outputs=[io.Sampler.Output()] + ) - FUNCTION = "get_sampler" - - def get_sampler(self, eta, s_noise, noise_device): + @classmethod + def execute(cls, eta, s_noise, noise_device) -> io.NodeOutput: if noise_device == 'cpu': sampler_name = "dpmpp_3m_sde" else: sampler_name = "dpmpp_3m_sde_gpu" sampler = comfy.samplers.ksampler(sampler_name, {"eta": eta, "s_noise": s_noise}) - return (sampler, ) + return io.NodeOutput(sampler) -class SamplerDPMPP_2M_SDE: + get_sampler = execute + +class SamplerDPMPP_2M_SDE(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"solver_type": (['midpoint', 'heun'], ), - "eta": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "s_noise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "noise_device": (['gpu', 'cpu'], ), - } - } - RETURN_TYPES = ("SAMPLER",) - CATEGORY = "sampling/custom_sampling/samplers" + def define_schema(cls): + return io.Schema( + node_id="SamplerDPMPP_2M_SDE", + category="sampling/custom_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), + io.Float.Input("s_noise", default=1.0, min=0.0, max=100.0, step=0.01, round=False), + io.Combo.Input("noise_device", options=['gpu', 'cpu']), + ], + outputs=[io.Sampler.Output()] + ) - FUNCTION = "get_sampler" - - def get_sampler(self, solver_type, eta, s_noise, noise_device): + @classmethod + def execute(cls, solver_type, eta, s_noise, noise_device) -> io.NodeOutput: if noise_device == 'cpu': sampler_name = "dpmpp_2m_sde" else: sampler_name = "dpmpp_2m_sde_gpu" sampler = comfy.samplers.ksampler(sampler_name, {"eta": eta, "s_noise": s_noise, "solver_type": solver_type}) - return (sampler, ) + return io.NodeOutput(sampler) + + get_sampler = execute -class SamplerDPMPP_SDE: +class SamplerDPMPP_SDE(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"eta": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "s_noise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "r": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "noise_device": (['gpu', 'cpu'], ), - } - } - RETURN_TYPES = ("SAMPLER",) - CATEGORY = "sampling/custom_sampling/samplers" + def define_schema(cls): + return io.Schema( + node_id="SamplerDPMPP_SDE", + category="sampling/custom_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), + io.Float.Input("r", default=0.5, min=0.0, max=100.0, step=0.01, round=False), + io.Combo.Input("noise_device", options=['gpu', 'cpu']), + ], + outputs=[io.Sampler.Output()] + ) - FUNCTION = "get_sampler" - - def get_sampler(self, eta, s_noise, r, noise_device): + @classmethod + def execute(cls, eta, s_noise, r, noise_device) -> io.NodeOutput: if noise_device == 'cpu': sampler_name = "dpmpp_sde" else: sampler_name = "dpmpp_sde_gpu" sampler = comfy.samplers.ksampler(sampler_name, {"eta": eta, "s_noise": s_noise, "r": r}) - return (sampler, ) + return io.NodeOutput(sampler) -class SamplerDPMPP_2S_Ancestral: + get_sampler = execute + +class SamplerDPMPP_2S_Ancestral(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"eta": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "s_noise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - } - } - RETURN_TYPES = ("SAMPLER",) - CATEGORY = "sampling/custom_sampling/samplers" + def define_schema(cls): + return io.Schema( + node_id="SamplerDPMPP_2S_Ancestral", + category="sampling/custom_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), + ], + outputs=[io.Sampler.Output()] + ) - FUNCTION = "get_sampler" - - def get_sampler(self, eta, s_noise): + @classmethod + def execute(cls, eta, s_noise) -> io.NodeOutput: sampler = comfy.samplers.ksampler("dpmpp_2s_ancestral", {"eta": eta, "s_noise": s_noise}) - return (sampler, ) + return io.NodeOutput(sampler) -class SamplerEulerAncestral: + get_sampler = execute + +class SamplerEulerAncestral(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"eta": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "s_noise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - } - } - RETURN_TYPES = ("SAMPLER",) - CATEGORY = "sampling/custom_sampling/samplers" + def define_schema(cls): + return io.Schema( + node_id="SamplerEulerAncestral", + category="sampling/custom_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), + ], + outputs=[io.Sampler.Output()] + ) - FUNCTION = "get_sampler" - - def get_sampler(self, eta, s_noise): + @classmethod + def execute(cls, eta, s_noise) -> io.NodeOutput: sampler = comfy.samplers.ksampler("euler_ancestral", {"eta": eta, "s_noise": s_noise}) - return (sampler, ) + return io.NodeOutput(sampler) -class SamplerEulerAncestralCFGPP: + get_sampler = execute + +class SamplerEulerAncestralCFGPP(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "eta": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step":0.01, "round": False}), - "s_noise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step":0.01, "round": False}), - }} - RETURN_TYPES = ("SAMPLER",) - CATEGORY = "sampling/custom_sampling/samplers" + def define_schema(cls): + return io.Schema( + node_id="SamplerEulerAncestralCFGPP", + display_name="SamplerEulerAncestralCFG++", + category="sampling/custom_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), + ], + outputs=[io.Sampler.Output()] + ) - FUNCTION = "get_sampler" - - def get_sampler(self, eta, s_noise): + @classmethod + def execute(cls, eta, s_noise) -> io.NodeOutput: sampler = comfy.samplers.ksampler( "euler_ancestral_cfg_pp", {"eta": eta, "s_noise": s_noise}) - return (sampler, ) + return io.NodeOutput(sampler) -class SamplerLMS: + get_sampler = execute + +class SamplerLMS(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"order": ("INT", {"default": 4, "min": 1, "max": 100}), - } - } - RETURN_TYPES = ("SAMPLER",) - CATEGORY = "sampling/custom_sampling/samplers" + def define_schema(cls): + return io.Schema( + node_id="SamplerLMS", + category="sampling/custom_sampling/samplers", + inputs=[io.Int.Input("order", default=4, min=1, max=100)], + outputs=[io.Sampler.Output()] + ) - FUNCTION = "get_sampler" - - def get_sampler(self, order): + @classmethod + def execute(cls, order) -> io.NodeOutput: sampler = comfy.samplers.ksampler("lms", {"order": order}) - return (sampler, ) + return io.NodeOutput(sampler) -class SamplerDPMAdaptative: + get_sampler = execute + +class SamplerDPMAdaptative(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"order": ("INT", {"default": 3, "min": 2, "max": 3}), - "rtol": ("FLOAT", {"default": 0.05, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "atol": ("FLOAT", {"default": 0.0078, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "h_init": ("FLOAT", {"default": 0.05, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "pcoeff": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "icoeff": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "dcoeff": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "accept_safety": ("FLOAT", {"default": 0.81, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "eta": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "s_noise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - } - } - RETURN_TYPES = ("SAMPLER",) - CATEGORY = "sampling/custom_sampling/samplers" + def define_schema(cls): + return io.Schema( + node_id="SamplerDPMAdaptative", + category="sampling/custom_sampling/samplers", + inputs=[ + io.Int.Input("order", default=3, min=2, max=3), + io.Float.Input("rtol", default=0.05, min=0.0, max=100.0, step=0.01, round=False), + io.Float.Input("atol", default=0.0078, min=0.0, max=100.0, step=0.01, round=False), + io.Float.Input("h_init", default=0.05, min=0.0, max=100.0, step=0.01, round=False), + io.Float.Input("pcoeff", default=0.0, min=0.0, max=100.0, step=0.01, round=False), + io.Float.Input("icoeff", default=1.0, min=0.0, max=100.0, step=0.01, round=False), + io.Float.Input("dcoeff", default=0.0, min=0.0, max=100.0, step=0.01, round=False), + io.Float.Input("accept_safety", default=0.81, min=0.0, max=100.0, step=0.01, round=False), + io.Float.Input("eta", default=0.0, 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), + ], + outputs=[io.Sampler.Output()] + ) - FUNCTION = "get_sampler" - - def get_sampler(self, order, rtol, atol, h_init, pcoeff, icoeff, dcoeff, accept_safety, eta, s_noise): + @classmethod + def execute(cls, order, rtol, atol, h_init, pcoeff, icoeff, dcoeff, accept_safety, eta, s_noise) -> io.NodeOutput: sampler = comfy.samplers.ksampler("dpm_adaptive", {"order": order, "rtol": rtol, "atol": atol, "h_init": h_init, "pcoeff": pcoeff, "icoeff": icoeff, "dcoeff": dcoeff, "accept_safety": accept_safety, "eta": eta, "s_noise":s_noise }) - return (sampler, ) + return io.NodeOutput(sampler) + + get_sampler = execute -class SamplerER_SDE(ComfyNodeABC): +class SamplerER_SDE(io.ComfyNode): @classmethod - def INPUT_TYPES(cls) -> InputTypeDict: - return { - "required": { - "solver_type": (IO.COMBO, {"options": ["ER-SDE", "Reverse-time SDE", "ODE"]}), - "max_stage": (IO.INT, {"default": 3, "min": 1, "max": 3}), - "eta": ( - IO.FLOAT, - {"default": 1.0, "min": 0.0, "max": 100.0, "step": 0.01, "round": False, "tooltip": "Stochastic strength of reverse-time SDE.\nWhen eta=0, it reduces to deterministic ODE. This setting doesn't apply to ER-SDE solver type."}, - ), - "s_noise": (IO.FLOAT, {"default": 1.0, "min": 0.0, "max": 100.0, "step": 0.01, "round": False}), - } - } + def define_schema(cls): + return io.Schema( + node_id="SamplerER_SDE", + category="sampling/custom_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), + io.Float.Input("eta", default=1.0, min=0.0, max=100.0, step=0.01, round=False, tooltip="Stochastic strength of reverse-time SDE.\nWhen eta=0, it reduces to deterministic ODE. This setting doesn't apply to ER-SDE solver type."), + io.Float.Input("s_noise", default=1.0, min=0.0, max=100.0, step=0.01, round=False), + ], + outputs=[io.Sampler.Output()] + ) - RETURN_TYPES = (IO.SAMPLER,) - CATEGORY = "sampling/custom_sampling/samplers" - - FUNCTION = "get_sampler" - - def get_sampler(self, solver_type, max_stage, eta, s_noise): + @classmethod + def execute(cls, solver_type, max_stage, eta, s_noise) -> io.NodeOutput: if solver_type == "ODE" or (solver_type == "Reverse-time SDE" and eta == 0): eta = 0 s_noise = 0 @@ -548,32 +611,33 @@ class SamplerER_SDE(ComfyNodeABC): sampler_name = "er_sde" sampler = comfy.samplers.ksampler(sampler_name, {"s_noise": s_noise, "noise_scaler": noise_scaler, "max_stage": max_stage}) - return (sampler,) + return io.NodeOutput(sampler) + + get_sampler = execute -class SamplerSASolver(ComfyNodeABC): +class SamplerSASolver(io.ComfyNode): @classmethod - def INPUT_TYPES(cls) -> InputTypeDict: - return { - "required": { - "model": (IO.MODEL, {}), - "eta": (IO.FLOAT, {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01, "round": False},), - "sde_start_percent": (IO.FLOAT, {"default": 0.2, "min": 0.0, "max": 1.0, "step": 0.001},), - "sde_end_percent": (IO.FLOAT, {"default": 0.8, "min": 0.0, "max": 1.0, "step": 0.001},), - "s_noise": (IO.FLOAT, {"default": 1.0, "min": 0.0, "max": 100.0, "step": 0.01, "round": False},), - "predictor_order": (IO.INT, {"default": 3, "min": 1, "max": 6}), - "corrector_order": (IO.INT, {"default": 4, "min": 0, "max": 6}), - "use_pece": (IO.BOOLEAN, {}), - "simple_order_2": (IO.BOOLEAN, {}), - } - } + def define_schema(cls): + return io.Schema( + node_id="SamplerSASolver", + category="sampling/custom_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), + io.Float.Input("sde_start_percent", default=0.2, min=0.0, max=1.0, step=0.001), + io.Float.Input("sde_end_percent", default=0.8, min=0.0, max=1.0, step=0.001), + io.Float.Input("s_noise", default=1.0, min=0.0, max=100.0, step=0.01, round=False), + io.Int.Input("predictor_order", default=3, min=1, max=6), + io.Int.Input("corrector_order", default=4, min=0, max=6), + io.Boolean.Input("use_pece"), + io.Boolean.Input("simple_order_2"), + ], + outputs=[io.Sampler.Output()] + ) - RETURN_TYPES = (IO.SAMPLER,) - CATEGORY = "sampling/custom_sampling/samplers" - - FUNCTION = "get_sampler" - - def get_sampler(self, model, eta, sde_start_percent, sde_end_percent, s_noise, predictor_order, corrector_order, use_pece, simple_order_2): + @classmethod + def execute(cls, model, eta, sde_start_percent, sde_end_percent, s_noise, predictor_order, corrector_order, use_pece, simple_order_2) -> io.NodeOutput: model_sampling = model.get_model_object("model_sampling") start_sigma = model_sampling.percent_to_sigma(sde_start_percent) end_sigma = model_sampling.percent_to_sigma(sde_end_percent) @@ -591,7 +655,43 @@ class SamplerSASolver(ComfyNodeABC): "simple_order_2": simple_order_2, }, ) - return (sampler,) + return io.NodeOutput(sampler) + + get_sampler = execute + + +class SamplerSEEDS2(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="SamplerSEEDS2", + category="sampling/custom_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"), + io.Float.Input("s_noise", default=1.0, min=0.0, max=100.0, step=0.01, round=False, tooltip="SDE noise multiplier"), + io.Float.Input("r", default=0.5, min=0.01, max=1.0, step=0.01, round=False, tooltip="Relative step size for the intermediate stage (c2 node)"), + ], + outputs=[io.Sampler.Output()], + description=( + "This sampler node can represent multiple samplers:\n\n" + "seeds_2\n" + "- default setting\n\n" + "exp_heun_2_x0\n" + "- solver_type=phi_2, r=1.0, eta=0.0\n\n" + "exp_heun_2_x0_sde\n" + "- solver_type=phi_2, r=1.0, eta=1.0, s_noise=1.0" + ) + ) + + @classmethod + def execute(cls, solver_type, eta, s_noise, r) -> io.NodeOutput: + sampler_name = "seeds_2" + sampler = comfy.samplers.ksampler( + sampler_name, + {"eta": eta, "s_noise": s_noise, "r": r, "solver_type": solver_type}, + ) + return io.NodeOutput(sampler) class Noise_EmptyNoise: @@ -612,30 +712,31 @@ class Noise_RandomNoise: batch_inds = input_latent["batch_index"] if "batch_index" in input_latent else None return comfy.sample.prepare_noise(latent_image, self.seed, batch_inds) -class SamplerCustom: +class SamplerCustom(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"model": ("MODEL",), - "add_noise": ("BOOLEAN", {"default": True}), - "noise_seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff, "control_after_generate": True}), - "cfg": ("FLOAT", {"default": 8.0, "min": 0.0, "max": 100.0, "step":0.1, "round": 0.01}), - "positive": ("CONDITIONING", ), - "negative": ("CONDITIONING", ), - "sampler": ("SAMPLER", ), - "sigmas": ("SIGMAS", ), - "latent_image": ("LATENT", ), - } - } + def define_schema(cls): + return io.Schema( + node_id="SamplerCustom", + category="sampling/custom_sampling", + inputs=[ + io.Model.Input("model"), + io.Boolean.Input("add_noise", default=True), + io.Int.Input("noise_seed", default=0, min=0, max=0xffffffffffffffff, control_after_generate=True), + io.Float.Input("cfg", default=8.0, min=0.0, max=100.0, step=0.1, round=0.01), + io.Conditioning.Input("positive"), + io.Conditioning.Input("negative"), + io.Sampler.Input("sampler"), + io.Sigmas.Input("sigmas"), + io.Latent.Input("latent_image"), + ], + outputs=[ + io.Latent.Output(display_name="output"), + io.Latent.Output(display_name="denoised_output"), + ] + ) - RETURN_TYPES = ("LATENT","LATENT") - RETURN_NAMES = ("output", "denoised_output") - - FUNCTION = "sample" - - CATEGORY = "sampling/custom_sampling" - - def sample(self, model, add_noise, noise_seed, cfg, positive, negative, sampler, sigmas, latent_image): + @classmethod + def execute(cls, model, add_noise, noise_seed, cfg, positive, negative, sampler, sigmas, latent_image) -> io.NodeOutput: latent = latent_image latent_image = latent["samples"] latent = latent.copy() @@ -660,56 +761,66 @@ class SamplerCustom: out = latent.copy() out["samples"] = samples if "x0" in x0_output: + x0_out = model.model.process_latent_out(x0_output["x0"].cpu()) + if samples.is_nested: + latent_shapes = [x.shape for x in samples.unbind()] + x0_out = comfy.nested_tensor.NestedTensor(comfy.utils.unpack_latents(x0_out, latent_shapes)) out_denoised = latent.copy() - out_denoised["samples"] = model.model.process_latent_out(x0_output["x0"].cpu()) + out_denoised["samples"] = x0_out else: out_denoised = out - return (out, out_denoised) + return io.NodeOutput(out, out_denoised) + + sample = execute class Guider_Basic(comfy.samplers.CFGGuider): def set_conds(self, positive): self.inner_set_conds({"positive": positive}) -class BasicGuider: +class BasicGuider(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"model": ("MODEL",), - "conditioning": ("CONDITIONING", ), - } - } + def define_schema(cls): + return io.Schema( + node_id="BasicGuider", + category="sampling/custom_sampling/guiders", + inputs=[ + io.Model.Input("model"), + io.Conditioning.Input("conditioning"), + ], + outputs=[io.Guider.Output()] + ) - RETURN_TYPES = ("GUIDER",) - - FUNCTION = "get_guider" - CATEGORY = "sampling/custom_sampling/guiders" - - def get_guider(self, model, conditioning): + @classmethod + def execute(cls, model, conditioning) -> io.NodeOutput: guider = Guider_Basic(model) guider.set_conds(conditioning) - return (guider,) + return io.NodeOutput(guider) -class CFGGuider: + get_guider = execute + +class CFGGuider(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"model": ("MODEL",), - "positive": ("CONDITIONING", ), - "negative": ("CONDITIONING", ), - "cfg": ("FLOAT", {"default": 8.0, "min": 0.0, "max": 100.0, "step":0.1, "round": 0.01}), - } - } + def define_schema(cls): + return io.Schema( + node_id="CFGGuider", + category="sampling/custom_sampling/guiders", + inputs=[ + io.Model.Input("model"), + io.Conditioning.Input("positive"), + io.Conditioning.Input("negative"), + io.Float.Input("cfg", default=8.0, min=0.0, max=100.0, step=0.1, round=0.01), + ], + outputs=[io.Guider.Output()] + ) - RETURN_TYPES = ("GUIDER",) - - FUNCTION = "get_guider" - CATEGORY = "sampling/custom_sampling/guiders" - - def get_guider(self, model, positive, negative, cfg): + @classmethod + def execute(cls, model, positive, negative, cfg) -> io.NodeOutput: guider = comfy.samplers.CFGGuider(model) guider.set_conds(positive, negative) guider.set_cfg(cfg) - return (guider,) + return io.NodeOutput(guider) + + get_guider = execute class Guider_DualCFG(comfy.samplers.CFGGuider): def set_cfg(self, cfg1, cfg2, nested=False): @@ -740,84 +851,88 @@ class Guider_DualCFG(comfy.samplers.CFGGuider): out = comfy.samplers.calc_cond_batch(self.inner_model, [negative_cond, middle_cond, positive_cond], x, timestep, model_options) return comfy.samplers.cfg_function(self.inner_model, out[1], out[0], self.cfg2, x, timestep, model_options=model_options, cond=middle_cond, uncond=negative_cond) + (out[2] - out[1]) * self.cfg1 -class DualCFGGuider: +class DualCFGGuider(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"model": ("MODEL",), - "cond1": ("CONDITIONING", ), - "cond2": ("CONDITIONING", ), - "negative": ("CONDITIONING", ), - "cfg_conds": ("FLOAT", {"default": 8.0, "min": 0.0, "max": 100.0, "step":0.1, "round": 0.01}), - "cfg_cond2_negative": ("FLOAT", {"default": 8.0, "min": 0.0, "max": 100.0, "step":0.1, "round": 0.01}), - "style": (["regular", "nested"],), - } - } + def define_schema(cls): + return io.Schema( + node_id="DualCFGGuider", + category="sampling/custom_sampling/guiders", + inputs=[ + io.Model.Input("model"), + io.Conditioning.Input("cond1"), + io.Conditioning.Input("cond2"), + io.Conditioning.Input("negative"), + io.Float.Input("cfg_conds", default=8.0, min=0.0, max=100.0, step=0.1, round=0.01), + io.Float.Input("cfg_cond2_negative", default=8.0, min=0.0, max=100.0, step=0.1, round=0.01), + io.Combo.Input("style", options=["regular", "nested"]), + ], + outputs=[io.Guider.Output()] + ) - RETURN_TYPES = ("GUIDER",) - - FUNCTION = "get_guider" - CATEGORY = "sampling/custom_sampling/guiders" - - def get_guider(self, model, cond1, cond2, negative, cfg_conds, cfg_cond2_negative, style): + @classmethod + def execute(cls, model, cond1, cond2, negative, cfg_conds, cfg_cond2_negative, style) -> io.NodeOutput: guider = Guider_DualCFG(model) guider.set_conds(cond1, cond2, negative) guider.set_cfg(cfg_conds, cfg_cond2_negative, nested=(style == "nested")) - return (guider,) + return io.NodeOutput(guider) -class DisableNoise: + get_guider = execute + +class DisableNoise(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required":{ - } - } + def define_schema(cls): + return io.Schema( + node_id="DisableNoise", + category="sampling/custom_sampling/noise", + inputs=[], + outputs=[io.Noise.Output()] + ) - RETURN_TYPES = ("NOISE",) - FUNCTION = "get_noise" - CATEGORY = "sampling/custom_sampling/noise" - - def get_noise(self): - return (Noise_EmptyNoise(),) - - -class RandomNoise(DisableNoise): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "noise_seed": ("INT", { - "default": 0, - "min": 0, - "max": 0xffffffffffffffff, - "control_after_generate": True, - }), - } - } + def execute(cls) -> io.NodeOutput: + return io.NodeOutput(Noise_EmptyNoise()) - def get_noise(self, noise_seed): - return (Noise_RandomNoise(noise_seed),) + get_noise = execute -class SamplerCustomAdvanced: +class RandomNoise(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"noise": ("NOISE", ), - "guider": ("GUIDER", ), - "sampler": ("SAMPLER", ), - "sigmas": ("SIGMAS", ), - "latent_image": ("LATENT", ), - } - } + def define_schema(cls): + return io.Schema( + node_id="RandomNoise", + category="sampling/custom_sampling/noise", + inputs=[io.Int.Input("noise_seed", default=0, min=0, max=0xffffffffffffffff, control_after_generate=True)], + outputs=[io.Noise.Output()] + ) - RETURN_TYPES = ("LATENT","LATENT") - RETURN_NAMES = ("output", "denoised_output") + @classmethod + def execute(cls, noise_seed) -> io.NodeOutput: + return io.NodeOutput(Noise_RandomNoise(noise_seed)) - FUNCTION = "sample" + get_noise = execute - CATEGORY = "sampling/custom_sampling" - def sample(self, noise, guider, sampler, sigmas, latent_image): +class SamplerCustomAdvanced(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="SamplerCustomAdvanced", + category="sampling/custom_sampling", + inputs=[ + io.Noise.Input("noise"), + io.Guider.Input("guider"), + io.Sampler.Input("sampler"), + io.Sigmas.Input("sigmas"), + io.Latent.Input("latent_image"), + ], + outputs=[ + io.Latent.Output(display_name="output"), + io.Latent.Output(display_name="denoised_output"), + ] + ) + + @classmethod + def execute(cls, noise, guider, sampler, sigmas, latent_image) -> io.NodeOutput: latent = latent_image latent_image = latent["samples"] latent = latent.copy() @@ -838,32 +953,40 @@ class SamplerCustomAdvanced: out = latent.copy() out["samples"] = samples if "x0" in x0_output: + x0_out = guider.model_patcher.model.process_latent_out(x0_output["x0"].cpu()) + if samples.is_nested: + latent_shapes = [x.shape for x in samples.unbind()] + x0_out = comfy.nested_tensor.NestedTensor(comfy.utils.unpack_latents(x0_out, latent_shapes)) out_denoised = latent.copy() - out_denoised["samples"] = guider.model_patcher.model.process_latent_out(x0_output["x0"].cpu()) + out_denoised["samples"] = x0_out else: out_denoised = out - return (out, out_denoised) + return io.NodeOutput(out, out_denoised) -class AddNoise: + sample = execute + +class AddNoise(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"model": ("MODEL",), - "noise": ("NOISE", ), - "sigmas": ("SIGMAS", ), - "latent_image": ("LATENT", ), - } - } + def define_schema(cls): + return io.Schema( + node_id="AddNoise", + category="_for_testing/custom_sampling/noise", + is_experimental=True, + inputs=[ + io.Model.Input("model"), + io.Noise.Input("noise"), + io.Sigmas.Input("sigmas"), + io.Latent.Input("latent_image"), + ], + outputs=[ + io.Latent.Output(), + ] + ) - RETURN_TYPES = ("LATENT",) - - FUNCTION = "add_noise" - - CATEGORY = "_for_testing/custom_sampling/noise" - - def add_noise(self, model, noise, sigmas, latent_image): + @classmethod + def execute(cls, model, noise, sigmas, latent_image) -> io.NodeOutput: if len(sigmas) == 0: - return latent_image + return io.NodeOutput(latent_image) latent = latent_image latent_image = latent["samples"] @@ -887,46 +1010,71 @@ class AddNoise: out = latent.copy() out["samples"] = noisy - return (out,) + return io.NodeOutput(out) + + add_noise = execute + +class ManualSigmas(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="ManualSigmas", + category="_for_testing/custom_sampling", + is_experimental=True, + inputs=[ + io.String.Input("sigmas", default="1, 0.5", multiline=False) + ], + outputs=[io.Sigmas.Output()] + ) + + @classmethod + def execute(cls, sigmas) -> io.NodeOutput: + sigmas = re.findall(r"[-+]?(?:\d*\.*\d+)", sigmas) + sigmas = [float(i) for i in sigmas] + sigmas = torch.FloatTensor(sigmas) + return io.NodeOutput(sigmas) + +class CustomSamplersExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + SamplerCustom, + BasicScheduler, + KarrasScheduler, + ExponentialScheduler, + PolyexponentialScheduler, + LaplaceScheduler, + VPScheduler, + BetaSamplingScheduler, + SDTurboScheduler, + KSamplerSelect, + SamplerEulerAncestral, + SamplerEulerAncestralCFGPP, + SamplerLMS, + SamplerDPMPP_3M_SDE, + SamplerDPMPP_2M_SDE, + SamplerDPMPP_SDE, + SamplerDPMPP_2S_Ancestral, + SamplerDPMAdaptative, + SamplerER_SDE, + SamplerSASolver, + SamplerSEEDS2, + SplitSigmas, + SplitSigmasDenoise, + FlipSigmas, + SetFirstSigma, + ExtendIntermediateSigmas, + SamplingPercentToSigma, + CFGGuider, + DualCFGGuider, + BasicGuider, + RandomNoise, + DisableNoise, + AddNoise, + SamplerCustomAdvanced, + ManualSigmas, + ] -NODE_CLASS_MAPPINGS = { - "SamplerCustom": SamplerCustom, - "BasicScheduler": BasicScheduler, - "KarrasScheduler": KarrasScheduler, - "ExponentialScheduler": ExponentialScheduler, - "PolyexponentialScheduler": PolyexponentialScheduler, - "LaplaceScheduler": LaplaceScheduler, - "VPScheduler": VPScheduler, - "BetaSamplingScheduler": BetaSamplingScheduler, - "SDTurboScheduler": SDTurboScheduler, - "KSamplerSelect": KSamplerSelect, - "SamplerEulerAncestral": SamplerEulerAncestral, - "SamplerEulerAncestralCFGPP": SamplerEulerAncestralCFGPP, - "SamplerLMS": SamplerLMS, - "SamplerDPMPP_3M_SDE": SamplerDPMPP_3M_SDE, - "SamplerDPMPP_2M_SDE": SamplerDPMPP_2M_SDE, - "SamplerDPMPP_SDE": SamplerDPMPP_SDE, - "SamplerDPMPP_2S_Ancestral": SamplerDPMPP_2S_Ancestral, - "SamplerDPMAdaptative": SamplerDPMAdaptative, - "SamplerER_SDE": SamplerER_SDE, - "SamplerSASolver": SamplerSASolver, - "SplitSigmas": SplitSigmas, - "SplitSigmasDenoise": SplitSigmasDenoise, - "FlipSigmas": FlipSigmas, - "SetFirstSigma": SetFirstSigma, - "ExtendIntermediateSigmas": ExtendIntermediateSigmas, - "SamplingPercentToSigma": SamplingPercentToSigma, - - "CFGGuider": CFGGuider, - "DualCFGGuider": DualCFGGuider, - "BasicGuider": BasicGuider, - "RandomNoise": RandomNoise, - "DisableNoise": DisableNoise, - "AddNoise": AddNoise, - "SamplerCustomAdvanced": SamplerCustomAdvanced, -} - -NODE_DISPLAY_NAME_MAPPINGS = { - "SamplerEulerAncestralCFGPP": "SamplerEulerAncestralCFG++", -} +async def comfy_entrypoint() -> CustomSamplersExtension: + return CustomSamplersExtension() diff --git a/comfy_extras/nodes_dataset.py b/comfy_extras/nodes_dataset.py new file mode 100644 index 000000000..5ef851bd0 --- /dev/null +++ b/comfy_extras/nodes_dataset.py @@ -0,0 +1,1529 @@ +import logging +import os +import json + +import numpy as np +import torch +from PIL import Image +from typing_extensions import override + +import folder_paths +import node_helpers +from comfy_api.latest import ComfyExtension, io + + +def load_and_process_images(image_files, input_dir): + """Utility function to load and process a list of images. + + Args: + image_files: List of image filenames + input_dir: Base directory containing the images + resize_method: How to handle images of different sizes ("None", "Stretch", "Crop", "Pad") + + Returns: + torch.Tensor: Batch of processed images + """ + if not image_files: + raise ValueError("No valid images found in input") + + output_images = [] + + for file in image_files: + image_path = os.path.join(input_dir, file) + img = node_helpers.pillow(Image.open, image_path) + + if img.mode == "I": + img = img.point(lambda i: i * (1 / 255)) + img = img.convert("RGB") + img_array = np.array(img).astype(np.float32) / 255.0 + img_tensor = torch.from_numpy(img_array)[None,] + output_images.append(img_tensor) + + return output_images + + +class LoadImageDataSetFromFolderNode(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LoadImageDataSetFromFolder", + display_name="Load Image Dataset from Folder", + category="dataset", + is_experimental=True, + inputs=[ + io.Combo.Input( + "folder", + options=folder_paths.get_input_subfolders(), + tooltip="The folder to load images from.", + ) + ], + outputs=[ + io.Image.Output( + display_name="images", + is_output_list=True, + tooltip="List of loaded images", + ) + ], + ) + + @classmethod + def execute(cls, folder): + sub_input_dir = os.path.join(folder_paths.get_input_directory(), folder) + valid_extensions = [".png", ".jpg", ".jpeg", ".webp"] + image_files = [ + f + for f in os.listdir(sub_input_dir) + if any(f.lower().endswith(ext) for ext in valid_extensions) + ] + output_tensor = load_and_process_images(image_files, sub_input_dir) + return io.NodeOutput(output_tensor) + + +class LoadImageTextDataSetFromFolderNode(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LoadImageTextDataSetFromFolder", + display_name="Load Image and Text Dataset from Folder", + category="dataset", + is_experimental=True, + inputs=[ + io.Combo.Input( + "folder", + options=folder_paths.get_input_subfolders(), + tooltip="The folder to load images from.", + ) + ], + outputs=[ + io.Image.Output( + display_name="images", + is_output_list=True, + tooltip="List of loaded images", + ), + io.String.Output( + display_name="texts", + is_output_list=True, + tooltip="List of text captions", + ), + ], + ) + + @classmethod + def execute(cls, folder): + logging.info(f"Loading images from folder: {folder}") + + sub_input_dir = os.path.join(folder_paths.get_input_directory(), folder) + valid_extensions = [".png", ".jpg", ".jpeg", ".webp"] + + image_files = [] + for item in os.listdir(sub_input_dir): + path = os.path.join(sub_input_dir, item) + if any(item.lower().endswith(ext) for ext in valid_extensions): + image_files.append(path) + elif os.path.isdir(path): + # Support kohya-ss/sd-scripts folder structure + repeat = 1 + if item.split("_")[0].isdigit(): + repeat = int(item.split("_")[0]) + image_files.extend( + [ + os.path.join(path, f) + for f in os.listdir(path) + if any(f.lower().endswith(ext) for ext in valid_extensions) + ] + * repeat + ) + + caption_file_path = [ + f.replace(os.path.splitext(f)[1], ".txt") for f in image_files + ] + captions = [] + for caption_file in caption_file_path: + caption_path = os.path.join(sub_input_dir, caption_file) + if os.path.exists(caption_path): + with open(caption_path, "r", encoding="utf-8") as f: + caption = f.read().strip() + captions.append(caption) + else: + captions.append("") + + output_tensor = load_and_process_images(image_files, sub_input_dir) + + logging.info(f"Loaded {len(output_tensor)} images from {sub_input_dir}.") + return io.NodeOutput(output_tensor, captions) + + +def save_images_to_folder(image_list, output_dir, prefix="image"): + """Utility function to save a list of image tensors to disk. + + Args: + image_list: List of image tensors (each [1, H, W, C] or [H, W, C] or [C, H, W]) + output_dir: Directory to save images to + prefix: Filename prefix + + Returns: + List of saved filenames + """ + os.makedirs(output_dir, exist_ok=True) + saved_files = [] + + for idx, img_tensor in enumerate(image_list): + # Handle different tensor shapes + if isinstance(img_tensor, torch.Tensor): + # Remove batch dimension if present [1, H, W, C] -> [H, W, C] + if img_tensor.dim() == 4 and img_tensor.shape[0] == 1: + img_tensor = img_tensor.squeeze(0) + + # If tensor is [C, H, W], permute to [H, W, C] + if img_tensor.dim() == 3 and img_tensor.shape[0] in [1, 3, 4]: + if ( + img_tensor.shape[0] <= 4 + and img_tensor.shape[1] > 4 + and img_tensor.shape[2] > 4 + ): + img_tensor = img_tensor.permute(1, 2, 0) + + # Convert to numpy and scale to 0-255 + img_array = img_tensor.cpu().numpy() + img_array = np.clip(img_array * 255.0, 0, 255).astype(np.uint8) + + # Convert to PIL Image + img = Image.fromarray(img_array) + else: + raise ValueError(f"Expected torch.Tensor, got {type(img_tensor)}") + + # Save image + filename = f"{prefix}_{idx:05d}.png" + filepath = os.path.join(output_dir, filename) + img.save(filepath) + saved_files.append(filename) + + return saved_files + + +class SaveImageDataSetToFolderNode(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="SaveImageDataSetToFolder", + display_name="Save Image Dataset to Folder", + category="dataset", + is_experimental=True, + is_output_node=True, + is_input_list=True, # Receive images as list + inputs=[ + io.Image.Input("images", tooltip="List of images to save."), + io.String.Input( + "folder_name", + default="dataset", + tooltip="Name of the folder to save images to (inside output directory).", + ), + io.String.Input( + "filename_prefix", + default="image", + tooltip="Prefix for saved image filenames.", + ), + ], + outputs=[], + ) + + @classmethod + def execute(cls, images, folder_name, filename_prefix): + # Extract scalar values + folder_name = folder_name[0] + filename_prefix = filename_prefix[0] + + output_dir = os.path.join(folder_paths.get_output_directory(), folder_name) + saved_files = save_images_to_folder(images, output_dir, filename_prefix) + + logging.info(f"Saved {len(saved_files)} images to {output_dir}.") + return io.NodeOutput() + + +class SaveImageTextDataSetToFolderNode(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="SaveImageTextDataSetToFolder", + display_name="Save Image and Text Dataset to Folder", + category="dataset", + is_experimental=True, + is_output_node=True, + is_input_list=True, # Receive both images and texts as lists + inputs=[ + io.Image.Input("images", tooltip="List of images to save."), + io.String.Input("texts", tooltip="List of text captions to save."), + io.String.Input( + "folder_name", + default="dataset", + tooltip="Name of the folder to save images to (inside output directory).", + ), + io.String.Input( + "filename_prefix", + default="image", + tooltip="Prefix for saved image filenames.", + ), + ], + outputs=[], + ) + + @classmethod + def execute(cls, images, texts, folder_name, filename_prefix): + # Extract scalar values + folder_name = folder_name[0] + filename_prefix = filename_prefix[0] + + output_dir = os.path.join(folder_paths.get_output_directory(), folder_name) + saved_files = save_images_to_folder(images, output_dir, filename_prefix) + + # Save captions + for idx, (filename, caption) in enumerate(zip(saved_files, texts)): + caption_filename = filename.replace(".png", ".txt") + caption_path = os.path.join(output_dir, caption_filename) + with open(caption_path, "w", encoding="utf-8") as f: + f.write(caption) + + logging.info(f"Saved {len(saved_files)} images and captions to {output_dir}.") + return io.NodeOutput() + + +# ========== Helper Functions for Transform Nodes ========== + + +def tensor_to_pil(img_tensor): + """Convert tensor to PIL Image.""" + if img_tensor.dim() == 4 and img_tensor.shape[0] == 1: + img_tensor = img_tensor.squeeze(0) + img_array = (img_tensor.cpu().numpy() * 255).clip(0, 255).astype(np.uint8) + return Image.fromarray(img_array) + + +def pil_to_tensor(img): + """Convert PIL Image to tensor.""" + img_array = np.array(img).astype(np.float32) / 255.0 + return torch.from_numpy(img_array)[None,] + + +# ========== Base Classes for Transform Nodes ========== + + +class ImageProcessingNode(io.ComfyNode): + """Base class for image processing nodes that operate on images. + + Child classes should set: + node_id: Unique node identifier (required) + display_name: Display name (optional, defaults to node_id) + description: Node description (optional) + extra_inputs: List of additional io.Input objects beyond "images" (optional) + is_group_process: None (auto-detect), True (group), or False (individual) (optional) + is_output_list: True (list output) or False (single output) (optional, default True) + + Child classes must implement ONE of: + _process(cls, image, **kwargs) -> tensor (for single-item processing) + _group_process(cls, images, **kwargs) -> list[tensor] (for group processing) + """ + + node_id = None + display_name = None + description = None + extra_inputs = [] + is_group_process = None # None = auto-detect, True/False = explicit + is_output_list = None # None = auto-detect based on processing mode + + @classmethod + def _detect_processing_mode(cls): + """Detect whether this node uses group or individual processing. + + Returns: + bool: True if group processing, False if individual processing + """ + # Explicit setting takes precedence + if cls.is_group_process is not None: + return cls.is_group_process + + # Check which method is overridden by looking at the defining class in MRO + base_class = ImageProcessingNode + + # Find which class in MRO defines _process + process_definer = None + for klass in cls.__mro__: + if "_process" in klass.__dict__: + process_definer = klass + break + + # Find which class in MRO defines _group_process + group_definer = None + for klass in cls.__mro__: + if "_group_process" in klass.__dict__: + group_definer = klass + break + + # Check what was overridden (not defined in base class) + has_process = process_definer is not None and process_definer is not base_class + has_group = group_definer is not None and group_definer is not base_class + + if has_process and has_group: + raise ValueError( + f"{cls.__name__}: Cannot override both _process and _group_process. " + "Override only one, or set is_group_process explicitly." + ) + if not has_process and not has_group: + raise ValueError( + f"{cls.__name__}: Must override either _process or _group_process" + ) + + return has_group + + @classmethod + def define_schema(cls): + if cls.node_id is None: + raise NotImplementedError(f"{cls.__name__} must set node_id class variable") + + is_group = cls._detect_processing_mode() + + # Auto-detect is_output_list if not explicitly set + # Single processing: False (backend collects results into list) + # Group processing: True by default (can be False for single-output nodes) + output_is_list = ( + cls.is_output_list if cls.is_output_list is not None else is_group + ) + + inputs = [ + io.Image.Input( + "images", + tooltip=( + "List of images to process." if is_group else "Image to process." + ), + ) + ] + inputs.extend(cls.extra_inputs) + + return io.Schema( + node_id=cls.node_id, + display_name=cls.display_name or cls.node_id, + category="dataset/image", + is_experimental=True, + is_input_list=is_group, # True for group, False for individual + inputs=inputs, + outputs=[ + io.Image.Output( + display_name="images", + is_output_list=output_is_list, + tooltip="Processed images", + ) + ], + ) + + @classmethod + def execute(cls, images, **kwargs): + """Execute the node. Routes to _process or _group_process based on mode.""" + is_group = cls._detect_processing_mode() + + # Extract scalar values from lists for parameters + params = {} + for k, v in kwargs.items(): + if isinstance(v, list) and len(v) == 1: + params[k] = v[0] + else: + params[k] = v + + if is_group: + # Group processing: images is list, call _group_process + result = cls._group_process(images, **params) + else: + # Individual processing: images is single item, call _process + result = cls._process(images, **params) + + return io.NodeOutput(result) + + @classmethod + def _process(cls, image, **kwargs): + """Override this method for single-item processing. + + Args: + image: tensor - Single image tensor + **kwargs: Additional parameters (already extracted from lists) + + Returns: + tensor - Processed image + """ + raise NotImplementedError(f"{cls.__name__} must implement _process method") + + @classmethod + def _group_process(cls, images, **kwargs): + """Override this method for group processing. + + Args: + images: list[tensor] - List of image tensors + **kwargs: Additional parameters (already extracted from lists) + + Returns: + list[tensor] - Processed images + """ + raise NotImplementedError( + f"{cls.__name__} must implement _group_process method" + ) + + +class TextProcessingNode(io.ComfyNode): + """Base class for text processing nodes that operate on texts. + + Child classes should set: + node_id: Unique node identifier (required) + display_name: Display name (optional, defaults to node_id) + description: Node description (optional) + extra_inputs: List of additional io.Input objects beyond "texts" (optional) + is_group_process: None (auto-detect), True (group), or False (individual) (optional) + is_output_list: True (list output) or False (single output) (optional, default True) + + Child classes must implement ONE of: + _process(cls, text, **kwargs) -> str (for single-item processing) + _group_process(cls, texts, **kwargs) -> list[str] (for group processing) + """ + + node_id = None + display_name = None + description = None + extra_inputs = [] + is_group_process = None # None = auto-detect, True/False = explicit + is_output_list = None # None = auto-detect based on processing mode + + @classmethod + def _detect_processing_mode(cls): + """Detect whether this node uses group or individual processing. + + Returns: + bool: True if group processing, False if individual processing + """ + # Explicit setting takes precedence + if cls.is_group_process is not None: + return cls.is_group_process + + # Check which method is overridden by looking at the defining class in MRO + base_class = TextProcessingNode + + # Find which class in MRO defines _process + process_definer = None + for klass in cls.__mro__: + if "_process" in klass.__dict__: + process_definer = klass + break + + # Find which class in MRO defines _group_process + group_definer = None + for klass in cls.__mro__: + if "_group_process" in klass.__dict__: + group_definer = klass + break + + # Check what was overridden (not defined in base class) + has_process = process_definer is not None and process_definer is not base_class + has_group = group_definer is not None and group_definer is not base_class + + if has_process and has_group: + raise ValueError( + f"{cls.__name__}: Cannot override both _process and _group_process. " + "Override only one, or set is_group_process explicitly." + ) + if not has_process and not has_group: + raise ValueError( + f"{cls.__name__}: Must override either _process or _group_process" + ) + + return has_group + + @classmethod + def define_schema(cls): + if cls.node_id is None: + raise NotImplementedError(f"{cls.__name__} must set node_id class variable") + + is_group = cls._detect_processing_mode() + + inputs = [ + io.String.Input( + "texts", + tooltip="List of texts to process." if is_group else "Text to process.", + ) + ] + inputs.extend(cls.extra_inputs) + + return io.Schema( + node_id=cls.node_id, + display_name=cls.display_name or cls.node_id, + category="dataset/text", + is_experimental=True, + is_input_list=is_group, # True for group, False for individual + inputs=inputs, + outputs=[ + io.String.Output( + display_name="texts", + is_output_list=cls.is_output_list, + tooltip="Processed texts", + ) + ], + ) + + @classmethod + def execute(cls, texts, **kwargs): + """Execute the node. Routes to _process or _group_process based on mode.""" + is_group = cls._detect_processing_mode() + + # Extract scalar values from lists for parameters + params = {} + for k, v in kwargs.items(): + if isinstance(v, list) and len(v) == 1: + params[k] = v[0] + else: + params[k] = v + + if is_group: + # Group processing: texts is list, call _group_process + result = cls._group_process(texts, **params) + else: + # Individual processing: texts is single item, call _process + result = cls._process(texts, **params) + + # Wrap result based on is_output_list + if cls.is_output_list: + # Result should already be a list (or will be for individual) + return io.NodeOutput(result if is_group else [result]) + else: + # Single output - wrap in list for NodeOutput + return io.NodeOutput([result]) + + @classmethod + def _process(cls, text, **kwargs): + """Override this method for single-item processing. + + Args: + text: str - Single text string + **kwargs: Additional parameters (already extracted from lists) + + Returns: + str - Processed text + """ + raise NotImplementedError(f"{cls.__name__} must implement _process method") + + @classmethod + def _group_process(cls, texts, **kwargs): + """Override this method for group processing. + + Args: + texts: list[str] - List of text strings + **kwargs: Additional parameters (already extracted from lists) + + Returns: + list[str] - Processed texts + """ + raise NotImplementedError( + f"{cls.__name__} must implement _group_process method" + ) + + +# ========== Image Transform Nodes ========== + + +class ResizeImagesByShorterEdgeNode(ImageProcessingNode): + node_id = "ResizeImagesByShorterEdge" + display_name = "Resize Images by Shorter Edge" + description = "Resize images so that the shorter edge matches the specified length while preserving aspect ratio." + extra_inputs = [ + io.Int.Input( + "shorter_edge", + default=512, + min=1, + max=8192, + tooltip="Target length for the shorter edge.", + ), + ] + + @classmethod + def _process(cls, image, shorter_edge): + img = tensor_to_pil(image) + w, h = img.size + if w < h: + new_w = shorter_edge + new_h = int(h * (shorter_edge / w)) + else: + new_h = shorter_edge + new_w = int(w * (shorter_edge / h)) + img = img.resize((new_w, new_h), Image.Resampling.LANCZOS) + return pil_to_tensor(img) + + +class ResizeImagesByLongerEdgeNode(ImageProcessingNode): + node_id = "ResizeImagesByLongerEdge" + display_name = "Resize Images by Longer Edge" + description = "Resize images so that the longer edge matches the specified length while preserving aspect ratio." + extra_inputs = [ + io.Int.Input( + "longer_edge", + default=1024, + min=1, + max=8192, + tooltip="Target length for the longer edge.", + ), + ] + + @classmethod + def _process(cls, image, longer_edge): + resized_images = [] + for image_i in image: + img = tensor_to_pil(image_i) + w, h = img.size + if w > h: + new_w = longer_edge + new_h = int(h * (longer_edge / w)) + else: + new_h = longer_edge + new_w = int(w * (longer_edge / h)) + img = img.resize((new_w, new_h), Image.Resampling.LANCZOS) + resized_images.append(pil_to_tensor(img)) + return torch.cat(resized_images, dim=0) + + +class CenterCropImagesNode(ImageProcessingNode): + node_id = "CenterCropImages" + display_name = "Center Crop Images" + description = "Center crop all images to the specified dimensions." + extra_inputs = [ + io.Int.Input("width", default=512, min=1, max=8192, tooltip="Crop width."), + io.Int.Input("height", default=512, min=1, max=8192, tooltip="Crop height."), + ] + + @classmethod + def _process(cls, image, width, height): + img = tensor_to_pil(image) + left = max(0, (img.width - width) // 2) + top = max(0, (img.height - height) // 2) + right = min(img.width, left + width) + bottom = min(img.height, top + height) + img = img.crop((left, top, right, bottom)) + return pil_to_tensor(img) + + +class RandomCropImagesNode(ImageProcessingNode): + node_id = "RandomCropImages" + display_name = "Random Crop Images" + description = ( + "Randomly crop all images to the specified dimensions (for data augmentation)." + ) + extra_inputs = [ + io.Int.Input("width", default=512, min=1, max=8192, tooltip="Crop width."), + io.Int.Input("height", default=512, min=1, max=8192, tooltip="Crop height."), + io.Int.Input( + "seed", default=0, min=0, max=0xFFFFFFFFFFFFFFFF, tooltip="Random seed." + ), + ] + + @classmethod + def _process(cls, image, width, height, seed): + np.random.seed(seed % (2**32 - 1)) + img = tensor_to_pil(image) + max_left = max(0, img.width - width) + max_top = max(0, img.height - height) + left = np.random.randint(0, max_left + 1) if max_left > 0 else 0 + top = np.random.randint(0, max_top + 1) if max_top > 0 else 0 + right = min(img.width, left + width) + bottom = min(img.height, top + height) + img = img.crop((left, top, right, bottom)) + return pil_to_tensor(img) + + +class NormalizeImagesNode(ImageProcessingNode): + node_id = "NormalizeImages" + display_name = "Normalize Images" + description = "Normalize images using mean and standard deviation." + extra_inputs = [ + io.Float.Input( + "mean", + default=0.5, + min=0.0, + max=1.0, + tooltip="Mean value for normalization.", + ), + io.Float.Input( + "std", + default=0.5, + min=0.001, + max=1.0, + tooltip="Standard deviation for normalization.", + ), + ] + + @classmethod + def _process(cls, image, mean, std): + return (image - mean) / std + + +class AdjustBrightnessNode(ImageProcessingNode): + node_id = "AdjustBrightness" + display_name = "Adjust Brightness" + description = "Adjust brightness of all images." + extra_inputs = [ + io.Float.Input( + "factor", + default=1.0, + min=0.0, + max=2.0, + tooltip="Brightness factor. 1.0 = no change, <1.0 = darker, >1.0 = brighter.", + ), + ] + + @classmethod + def _process(cls, image, factor): + return (image * factor).clamp(0.0, 1.0) + + +class AdjustContrastNode(ImageProcessingNode): + node_id = "AdjustContrast" + display_name = "Adjust Contrast" + description = "Adjust contrast of all images." + extra_inputs = [ + io.Float.Input( + "factor", + default=1.0, + min=0.0, + max=2.0, + tooltip="Contrast factor. 1.0 = no change, <1.0 = less contrast, >1.0 = more contrast.", + ), + ] + + @classmethod + def _process(cls, image, factor): + return ((image - 0.5) * factor + 0.5).clamp(0.0, 1.0) + + +class ShuffleDatasetNode(ImageProcessingNode): + node_id = "ShuffleDataset" + display_name = "Shuffle Image Dataset" + description = "Randomly shuffle the order of images in the dataset." + is_group_process = True # Requires full list to shuffle + extra_inputs = [ + io.Int.Input( + "seed", default=0, min=0, max=0xFFFFFFFFFFFFFFFF, tooltip="Random seed." + ), + ] + + @classmethod + def _group_process(cls, images, seed): + np.random.seed(seed % (2**32 - 1)) + indices = np.random.permutation(len(images)) + return [images[i] for i in indices] + + +class ShuffleImageTextDatasetNode(io.ComfyNode): + """Special node that shuffles both images and texts together.""" + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="ShuffleImageTextDataset", + display_name="Shuffle Image-Text Dataset", + category="dataset/image", + is_experimental=True, + is_input_list=True, + inputs=[ + io.Image.Input("images", tooltip="List of images to shuffle."), + io.String.Input("texts", tooltip="List of texts to shuffle."), + io.Int.Input( + "seed", + default=0, + min=0, + max=0xFFFFFFFFFFFFFFFF, + tooltip="Random seed.", + ), + ], + outputs=[ + io.Image.Output( + display_name="images", + is_output_list=True, + tooltip="Shuffled images", + ), + io.String.Output( + display_name="texts", is_output_list=True, tooltip="Shuffled texts" + ), + ], + ) + + @classmethod + def execute(cls, images, texts, seed): + seed = seed[0] # Extract scalar + np.random.seed(seed % (2**32 - 1)) + indices = np.random.permutation(len(images)) + shuffled_images = [images[i] for i in indices] + shuffled_texts = [texts[i] for i in indices] + return io.NodeOutput(shuffled_images, shuffled_texts) + + +# ========== Text Transform Nodes ========== + + +class TextToLowercaseNode(TextProcessingNode): + node_id = "TextToLowercase" + display_name = "Text to Lowercase" + description = "Convert all texts to lowercase." + + @classmethod + def _process(cls, text): + return text.lower() + + +class TextToUppercaseNode(TextProcessingNode): + node_id = "TextToUppercase" + display_name = "Text to Uppercase" + description = "Convert all texts to uppercase." + + @classmethod + def _process(cls, text): + return text.upper() + + +class TruncateTextNode(TextProcessingNode): + node_id = "TruncateText" + display_name = "Truncate Text" + description = "Truncate all texts to a maximum length." + extra_inputs = [ + io.Int.Input( + "max_length", default=77, min=1, max=10000, tooltip="Maximum text length." + ), + ] + + @classmethod + def _process(cls, text, max_length): + return text[:max_length] + + +class AddTextPrefixNode(TextProcessingNode): + node_id = "AddTextPrefix" + display_name = "Add Text Prefix" + description = "Add a prefix to all texts." + extra_inputs = [ + io.String.Input("prefix", default="", tooltip="Prefix to add."), + ] + + @classmethod + def _process(cls, text, prefix): + return prefix + text + + +class AddTextSuffixNode(TextProcessingNode): + node_id = "AddTextSuffix" + display_name = "Add Text Suffix" + description = "Add a suffix to all texts." + extra_inputs = [ + io.String.Input("suffix", default="", tooltip="Suffix to add."), + ] + + @classmethod + def _process(cls, text, suffix): + return text + suffix + + +class ReplaceTextNode(TextProcessingNode): + node_id = "ReplaceText" + display_name = "Replace Text" + description = "Replace text in all texts." + extra_inputs = [ + io.String.Input("find", default="", tooltip="Text to find."), + io.String.Input("replace", default="", tooltip="Text to replace with."), + ] + + @classmethod + def _process(cls, text, find, replace): + return text.replace(find, replace) + + +class StripWhitespaceNode(TextProcessingNode): + node_id = "StripWhitespace" + display_name = "Strip Whitespace" + description = "Strip leading and trailing whitespace from all texts." + + @classmethod + def _process(cls, text): + return text.strip() + + +# ========== Group Processing Example Nodes ========== + + +class ImageDeduplicationNode(ImageProcessingNode): + """Remove duplicate or very similar images from the dataset using perceptual hashing.""" + + node_id = "ImageDeduplication" + display_name = "Image Deduplication" + description = "Remove duplicate or very similar images from the dataset." + is_group_process = True # Requires full list to compare images + extra_inputs = [ + io.Float.Input( + "similarity_threshold", + default=0.95, + min=0.0, + max=1.0, + tooltip="Similarity threshold (0-1). Higher means more similar. Images above this threshold are considered duplicates.", + ), + ] + + @classmethod + def _group_process(cls, images, similarity_threshold): + """Remove duplicate images using perceptual hashing.""" + if len(images) == 0: + return [] + + # Compute simple perceptual hash for each image + def compute_hash(img_tensor): + """Compute a simple perceptual hash by resizing to 8x8 and comparing to average.""" + img = tensor_to_pil(img_tensor) + # Resize to 8x8 + img_small = img.resize((8, 8), Image.Resampling.LANCZOS).convert("L") + # Get pixels + pixels = list(img_small.getdata()) + # Compute average + avg = sum(pixels) / len(pixels) + # Create hash (1 if above average, 0 otherwise) + hash_bits = "".join("1" if p > avg else "0" for p in pixels) + return hash_bits + + def hamming_distance(hash1, hash2): + """Compute Hamming distance between two hash strings.""" + return sum(c1 != c2 for c1, c2 in zip(hash1, hash2)) + + # Compute hashes for all images + hashes = [compute_hash(img) for img in images] + + # Find duplicates + keep_indices = [] + for i in range(len(images)): + is_duplicate = False + for j in keep_indices: + # Compare hashes + distance = hamming_distance(hashes[i], hashes[j]) + similarity = 1.0 - (distance / 64.0) # 64 bits total + if similarity >= similarity_threshold: + is_duplicate = True + logging.info( + f"Image {i} is similar to image {j} (similarity: {similarity:.3f}), skipping" + ) + break + + if not is_duplicate: + keep_indices.append(i) + + # Return only unique images + unique_images = [images[i] for i in keep_indices] + logging.info( + f"Deduplication: kept {len(unique_images)} out of {len(images)} images" + ) + return unique_images + + +class ImageGridNode(ImageProcessingNode): + """Combine multiple images into a single grid/collage.""" + + node_id = "ImageGrid" + display_name = "Image Grid" + description = "Arrange multiple images into a grid layout." + is_group_process = True # Requires full list to create grid + is_output_list = False # Outputs single grid image + extra_inputs = [ + io.Int.Input( + "columns", + default=4, + min=1, + max=20, + tooltip="Number of columns in the grid.", + ), + io.Int.Input( + "cell_width", + default=256, + min=32, + max=2048, + tooltip="Width of each cell in the grid.", + ), + io.Int.Input( + "cell_height", + default=256, + min=32, + max=2048, + tooltip="Height of each cell in the grid.", + ), + io.Int.Input( + "padding", default=4, min=0, max=50, tooltip="Padding between images." + ), + ] + + @classmethod + def _group_process(cls, images, columns, cell_width, cell_height, padding): + """Arrange images into a grid.""" + if len(images) == 0: + raise ValueError("Cannot create grid from empty image list") + + # Calculate grid dimensions + num_images = len(images) + rows = (num_images + columns - 1) // columns # Ceiling division + + # Calculate total grid size + grid_width = columns * cell_width + (columns - 1) * padding + grid_height = rows * cell_height + (rows - 1) * padding + + # Create blank grid + grid = Image.new("RGB", (grid_width, grid_height), (0, 0, 0)) + + # Place images + for idx, img_tensor in enumerate(images): + row = idx // columns + col = idx % columns + + # Convert to PIL and resize to cell size + img = tensor_to_pil(img_tensor) + img = img.resize((cell_width, cell_height), Image.Resampling.LANCZOS) + + # Calculate position + x = col * (cell_width + padding) + y = row * (cell_height + padding) + + # Paste into grid + grid.paste(img, (x, y)) + + logging.info( + f"Created {columns}x{rows} grid with {num_images} images ({grid_width}x{grid_height})" + ) + return pil_to_tensor(grid) + + +class MergeImageListsNode(ImageProcessingNode): + """Merge multiple image lists into a single list.""" + + node_id = "MergeImageLists" + display_name = "Merge Image Lists" + description = "Concatenate multiple image lists into one." + is_group_process = True # Receives images as list + + @classmethod + def _group_process(cls, images): + """Simply return the images list (already merged by input handling).""" + # When multiple list inputs are connected, they're concatenated + # For now, this is a simple pass-through + logging.info(f"Merged image list contains {len(images)} images") + return images + + +class MergeTextListsNode(TextProcessingNode): + """Merge multiple text lists into a single list.""" + + node_id = "MergeTextLists" + display_name = "Merge Text Lists" + description = "Concatenate multiple text lists into one." + is_group_process = True # Receives texts as list + + @classmethod + def _group_process(cls, texts): + """Simply return the texts list (already merged by input handling).""" + # When multiple list inputs are connected, they're concatenated + # For now, this is a simple pass-through + logging.info(f"Merged text list contains {len(texts)} texts") + return texts + + +# ========== Training Dataset Nodes ========== + + +class ResolutionBucket(io.ComfyNode): + """Bucket latents and conditions by resolution for efficient batch training.""" + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="ResolutionBucket", + display_name="Resolution Bucket", + category="dataset", + is_experimental=True, + is_input_list=True, + inputs=[ + io.Latent.Input( + "latents", + tooltip="List of latent dicts to bucket by resolution.", + ), + io.Conditioning.Input( + "conditioning", + tooltip="List of conditioning lists (must match latents length).", + ), + ], + outputs=[ + io.Latent.Output( + display_name="latents", + is_output_list=True, + tooltip="List of batched latent dicts, one per resolution bucket.", + ), + io.Conditioning.Output( + display_name="conditioning", + is_output_list=True, + tooltip="List of condition lists, one per resolution bucket.", + ), + ], + ) + + @classmethod + def execute(cls, latents, conditioning): + # latents: list[{"samples": tensor}] where tensor is (B, C, H, W), typically B=1 + # conditioning: list[list[cond]] + + # Validate lengths match + if len(latents) != len(conditioning): + raise ValueError( + f"Number of latents ({len(latents)}) does not match number of conditions ({len(conditioning)})." + ) + + # Flatten latents and conditions to individual samples + flat_latents = [] # list of (C, H, W) tensors + flat_conditions = [] # list of condition lists + + for latent_dict, cond in zip(latents, conditioning): + samples = latent_dict["samples"] # (B, C, H, W) + batch_size = samples.shape[0] + + # cond is a list of conditions with length == batch_size + for i in range(batch_size): + flat_latents.append(samples[i]) # (C, H, W) + flat_conditions.append(cond[i]) # single condition + + # Group by resolution (H, W) + buckets = {} # (H, W) -> {"latents": list, "conditions": list} + + for latent, cond in zip(flat_latents, flat_conditions): + # latent shape is (..., H, W) (B, C, H, W) or (B, T, C, H ,W) + h, w = latent.shape[-2], latent.shape[-1] + key = (h, w) + + if key not in buckets: + buckets[key] = {"latents": [], "conditions": []} + + buckets[key]["latents"].append(latent) + buckets[key]["conditions"].append(cond) + + # Convert buckets to output format + output_latents = [] # list[{"samples": tensor}] where tensor is (Bi, ..., H, W) + output_conditions = [] # list[list[cond]] where each inner list has Bi conditions + + for (h, w), bucket_data in buckets.items(): + # Stack latents into batch: list of (..., H, W) -> (Bi, ..., H, W) + stacked_latents = torch.stack(bucket_data["latents"], dim=0) + output_latents.append({"samples": stacked_latents}) + + # Conditions stay as list of condition lists + output_conditions.append(bucket_data["conditions"]) + + logging.info( + f"Resolution bucket ({h}x{w}): {len(bucket_data['latents'])} samples" + ) + + logging.info(f"Created {len(buckets)} resolution buckets from {len(flat_latents)} samples") + return io.NodeOutput(output_latents, output_conditions) + + +class MakeTrainingDataset(io.ComfyNode): + """Encode images with VAE and texts with CLIP to create a training dataset.""" + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="MakeTrainingDataset", + display_name="Make Training Dataset", + category="dataset", + is_experimental=True, + is_input_list=True, # images and texts as lists + inputs=[ + io.Image.Input("images", tooltip="List of images to encode."), + io.Vae.Input( + "vae", tooltip="VAE model for encoding images to latents." + ), + io.Clip.Input( + "clip", tooltip="CLIP model for encoding text to conditioning." + ), + io.String.Input( + "texts", + optional=True, + tooltip="List of text captions. Can be length n (matching images), 1 (repeated for all), or omitted (uses empty string).", + ), + ], + outputs=[ + io.Latent.Output( + display_name="latents", + is_output_list=True, + tooltip="List of latent dicts", + ), + io.Conditioning.Output( + display_name="conditioning", + is_output_list=True, + tooltip="List of conditioning lists", + ), + ], + ) + + @classmethod + def execute(cls, images, vae, clip, texts=None): + # Extract scalars (vae and clip are single values wrapped in lists) + vae = vae[0] + clip = clip[0] + + # Handle text list + num_images = len(images) + + if texts is None or len(texts) == 0: + # Treat as [""] for unconditional training + texts = [""] + + if len(texts) == 1 and num_images > 1: + # Repeat single text for all images + texts = texts * num_images + elif len(texts) != num_images: + raise ValueError( + f"Number of texts ({len(texts)}) does not match number of images ({num_images}). " + f"Text list should have length {num_images}, 1, or 0." + ) + + # Encode images with VAE + logging.info(f"Encoding {num_images} images with VAE...") + latents_list = [] # list[{"samples": tensor}] + for img_tensor in images: + # img_tensor is [1, H, W, 3] + latent_tensor = vae.encode(img_tensor[:, :, :, :3]) + latents_list.append({"samples": latent_tensor}) + + # Encode texts with CLIP + logging.info(f"Encoding {len(texts)} texts with CLIP...") + conditioning_list = [] # list[list[cond]] + for text in texts: + if text == "": + cond = clip.encode_from_tokens_scheduled(clip.tokenize("")) + else: + tokens = clip.tokenize(text) + cond = clip.encode_from_tokens_scheduled(tokens) + conditioning_list.append(cond) + + logging.info( + f"Created dataset with {len(latents_list)} latents and {len(conditioning_list)} conditioning." + ) + return io.NodeOutput(latents_list, conditioning_list) + + +class SaveTrainingDataset(io.ComfyNode): + """Save encoded training dataset (latents + conditioning) to disk.""" + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="SaveTrainingDataset", + display_name="Save Training Dataset", + category="dataset", + is_experimental=True, + is_output_node=True, + is_input_list=True, # Receive lists + inputs=[ + io.Latent.Input( + "latents", + tooltip="List of latent dicts from MakeTrainingDataset.", + ), + io.Conditioning.Input( + "conditioning", + tooltip="List of conditioning lists from MakeTrainingDataset.", + ), + io.String.Input( + "folder_name", + default="training_dataset", + tooltip="Name of folder to save dataset (inside output directory).", + ), + io.Int.Input( + "shard_size", + default=1000, + min=1, + max=100000, + tooltip="Number of samples per shard file.", + ), + ], + outputs=[], + ) + + @classmethod + def execute(cls, latents, conditioning, folder_name, shard_size): + # Extract scalars + folder_name = folder_name[0] + shard_size = shard_size[0] + + # latents: list[{"samples": tensor}] + # conditioning: list[list[cond]] + + # Validate lengths match + if len(latents) != len(conditioning): + raise ValueError( + f"Number of latents ({len(latents)}) does not match number of conditions ({len(conditioning)}). " + f"Something went wrong in dataset preparation." + ) + + # Create output directory + output_dir = os.path.join(folder_paths.get_output_directory(), folder_name) + os.makedirs(output_dir, exist_ok=True) + + # Prepare data pairs + num_samples = len(latents) + num_shards = (num_samples + shard_size - 1) // shard_size # Ceiling division + + logging.info( + f"Saving {num_samples} samples to {num_shards} shards in {output_dir}..." + ) + + # Save data in shards + for shard_idx in range(num_shards): + start_idx = shard_idx * shard_size + end_idx = min(start_idx + shard_size, num_samples) + + # Get shard data (list of latent dicts and conditioning lists) + shard_data = { + "latents": latents[start_idx:end_idx], + "conditioning": conditioning[start_idx:end_idx], + } + + # Save shard + shard_filename = f"shard_{shard_idx:04d}.pkl" + shard_path = os.path.join(output_dir, shard_filename) + + with open(shard_path, "wb") as f: + torch.save(shard_data, f) + + logging.info( + f"Saved shard {shard_idx + 1}/{num_shards}: {shard_filename} ({end_idx - start_idx} samples)" + ) + + # Save metadata + metadata = { + "num_samples": num_samples, + "num_shards": num_shards, + "shard_size": shard_size, + } + metadata_path = os.path.join(output_dir, "metadata.json") + with open(metadata_path, "w") as f: + json.dump(metadata, f, indent=2) + + logging.info(f"Successfully saved {num_samples} samples to {output_dir}.") + return io.NodeOutput() + + +class LoadTrainingDataset(io.ComfyNode): + """Load encoded training dataset from disk.""" + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LoadTrainingDataset", + display_name="Load Training Dataset", + category="dataset", + is_experimental=True, + inputs=[ + io.String.Input( + "folder_name", + default="training_dataset", + tooltip="Name of folder containing the saved dataset (inside output directory).", + ), + ], + outputs=[ + io.Latent.Output( + display_name="latents", + is_output_list=True, + tooltip="List of latent dicts", + ), + io.Conditioning.Output( + display_name="conditioning", + is_output_list=True, + tooltip="List of conditioning lists", + ), + ], + ) + + @classmethod + def execute(cls, folder_name): + # Get dataset directory + dataset_dir = os.path.join(folder_paths.get_output_directory(), folder_name) + + if not os.path.exists(dataset_dir): + raise ValueError(f"Dataset directory not found: {dataset_dir}") + + # Find all shard files + shard_files = sorted( + [ + f + for f in os.listdir(dataset_dir) + if f.startswith("shard_") and f.endswith(".pkl") + ] + ) + + if not shard_files: + raise ValueError(f"No shard files found in {dataset_dir}") + + logging.info(f"Loading {len(shard_files)} shards from {dataset_dir}...") + + # Load all shards + all_latents = [] # list[{"samples": tensor}] + all_conditioning = [] # list[list[cond]] + + for shard_file in shard_files: + shard_path = os.path.join(dataset_dir, shard_file) + + with open(shard_path, "rb") as f: + shard_data = torch.load(f) + + all_latents.extend(shard_data["latents"]) + all_conditioning.extend(shard_data["conditioning"]) + + logging.info(f"Loaded {shard_file}: {len(shard_data['latents'])} samples") + + logging.info( + f"Successfully loaded {len(all_latents)} samples from {dataset_dir}." + ) + return io.NodeOutput(all_latents, all_conditioning) + + +# ========== Extension Setup ========== + + +class DatasetExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + # Data loading/saving nodes + LoadImageDataSetFromFolderNode, + LoadImageTextDataSetFromFolderNode, + SaveImageDataSetToFolderNode, + SaveImageTextDataSetToFolderNode, + # Image transform nodes + ResizeImagesByShorterEdgeNode, + ResizeImagesByLongerEdgeNode, + CenterCropImagesNode, + RandomCropImagesNode, + NormalizeImagesNode, + AdjustBrightnessNode, + AdjustContrastNode, + ShuffleDatasetNode, + ShuffleImageTextDatasetNode, + # Text transform nodes + TextToLowercaseNode, + TextToUppercaseNode, + TruncateTextNode, + AddTextPrefixNode, + AddTextSuffixNode, + ReplaceTextNode, + StripWhitespaceNode, + # Group processing examples + ImageDeduplicationNode, + ImageGridNode, + MergeImageListsNode, + MergeTextListsNode, + # Training dataset nodes + MakeTrainingDataset, + SaveTrainingDataset, + LoadTrainingDataset, + ResolutionBucket, + ] + + +async def comfy_entrypoint() -> DatasetExtension: + return DatasetExtension() diff --git a/comfy_extras/nodes_flux.py b/comfy_extras/nodes_flux.py index ce1b2e89f..12c8ed3e6 100644 --- a/comfy_extras/nodes_flux.py +++ b/comfy_extras/nodes_flux.py @@ -2,7 +2,10 @@ import node_helpers import comfy.utils from typing_extensions import override from comfy_api.latest import ComfyExtension, io - +import comfy.model_management +import torch +import math +import nodes class CLIPTextEncodeFlux(io.ComfyNode): @classmethod @@ -30,6 +33,27 @@ class CLIPTextEncodeFlux(io.ComfyNode): encode = execute # TODO: remove +class EmptyFlux2LatentImage(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="EmptyFlux2LatentImage", + display_name="Empty Flux 2 Latent", + category="latent", + inputs=[ + io.Int.Input("width", default=1024, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("height", default=1024, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("batch_size", default=1, min=1, max=4096), + ], + outputs=[ + io.Latent.Output(), + ], + ) + + @classmethod + def execute(cls, width, height, batch_size=1) -> io.NodeOutput: + latent = torch.zeros([batch_size, 128, height // 16, width // 16], device=comfy.model_management.intermediate_device()) + return io.NodeOutput({"samples": latent}) class FluxGuidance(io.ComfyNode): @classmethod @@ -130,12 +154,13 @@ class FluxKontextMultiReferenceLatentMethod(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="FluxKontextMultiReferenceLatentMethod", + display_name="Edit Model Reference Method", category="advanced/conditioning/flux", inputs=[ io.Conditioning.Input("conditioning"), io.Combo.Input( "reference_latents_method", - options=["offset", "index", "uxo/uno"], + options=["offset", "index", "uxo/uno", "index_timestep_zero"], ), ], outputs=[ @@ -154,6 +179,58 @@ class FluxKontextMultiReferenceLatentMethod(io.ComfyNode): append = execute # TODO: remove +def generalized_time_snr_shift(t, mu: float, sigma: float): + return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma) + + +def compute_empirical_mu(image_seq_len: int, num_steps: int) -> float: + a1, b1 = 8.73809524e-05, 1.89833333 + a2, b2 = 0.00016927, 0.45666666 + + if image_seq_len > 4300: + mu = a2 * image_seq_len + b2 + return float(mu) + + m_200 = a2 * image_seq_len + b2 + m_10 = a1 * image_seq_len + b1 + + a = (m_200 - m_10) / 190.0 + b = m_200 - 200.0 * a + mu = a * num_steps + b + + return float(mu) + + +def get_schedule(num_steps: int, image_seq_len: int) -> list[float]: + mu = compute_empirical_mu(image_seq_len, num_steps) + timesteps = torch.linspace(1, 0, num_steps + 1) + timesteps = generalized_time_snr_shift(timesteps, mu, 1.0) + return timesteps + + +class Flux2Scheduler(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="Flux2Scheduler", + category="sampling/custom_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), + io.Int.Input("height", default=1024, min=16, max=nodes.MAX_RESOLUTION, step=1), + ], + outputs=[ + io.Sigmas.Output(), + ], + ) + + @classmethod + def execute(cls, steps, width, height) -> io.NodeOutput: + seq_len = (width * height / (16 * 16)) + sigmas = get_schedule(steps, round(seq_len)) + return io.NodeOutput(sigmas) + + class FluxExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[io.ComfyNode]]: @@ -163,6 +240,8 @@ class FluxExtension(ComfyExtension): FluxDisableGuidance, FluxKontextImageScale, FluxKontextMultiReferenceLatentMethod, + EmptyFlux2LatentImage, + Flux2Scheduler, ] diff --git a/comfy_extras/nodes_freelunch.py b/comfy_extras/nodes_freelunch.py index e3ac58447..3429b731e 100644 --- a/comfy_extras/nodes_freelunch.py +++ b/comfy_extras/nodes_freelunch.py @@ -2,6 +2,8 @@ import torch import logging +from typing_extensions import override +from comfy_api.latest import ComfyExtension, IO def Fourier_filter(x, threshold, scale): # FFT @@ -22,21 +24,26 @@ def Fourier_filter(x, threshold, scale): return x_filtered.to(x.dtype) -class FreeU: +class FreeU(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { "model": ("MODEL",), - "b1": ("FLOAT", {"default": 1.1, "min": 0.0, "max": 10.0, "step": 0.01}), - "b2": ("FLOAT", {"default": 1.2, "min": 0.0, "max": 10.0, "step": 0.01}), - "s1": ("FLOAT", {"default": 0.9, "min": 0.0, "max": 10.0, "step": 0.01}), - "s2": ("FLOAT", {"default": 0.2, "min": 0.0, "max": 10.0, "step": 0.01}), - }} - RETURN_TYPES = ("MODEL",) - FUNCTION = "patch" + def define_schema(cls): + return IO.Schema( + node_id="FreeU", + category="model_patches/unet", + inputs=[ + IO.Model.Input("model"), + IO.Float.Input("b1", default=1.1, min=0.0, max=10.0, step=0.01), + IO.Float.Input("b2", default=1.2, min=0.0, max=10.0, step=0.01), + IO.Float.Input("s1", default=0.9, min=0.0, max=10.0, step=0.01), + IO.Float.Input("s2", default=0.2, min=0.0, max=10.0, step=0.01), + ], + outputs=[ + IO.Model.Output(), + ], + ) - CATEGORY = "model_patches/unet" - - def patch(self, model, b1, b2, s1, s2): + @classmethod + def execute(cls, model, b1, b2, s1, s2) -> IO.NodeOutput: model_channels = model.model.model_config.unet_config["model_channels"] scale_dict = {model_channels * 4: (b1, s1), model_channels * 2: (b2, s2)} on_cpu_devices = {} @@ -59,23 +66,31 @@ class FreeU: m = model.clone() m.set_model_output_block_patch(output_block_patch) - return (m, ) + return IO.NodeOutput(m) -class FreeU_V2: + patch = execute # TODO: remove + + +class FreeU_V2(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { "model": ("MODEL",), - "b1": ("FLOAT", {"default": 1.3, "min": 0.0, "max": 10.0, "step": 0.01}), - "b2": ("FLOAT", {"default": 1.4, "min": 0.0, "max": 10.0, "step": 0.01}), - "s1": ("FLOAT", {"default": 0.9, "min": 0.0, "max": 10.0, "step": 0.01}), - "s2": ("FLOAT", {"default": 0.2, "min": 0.0, "max": 10.0, "step": 0.01}), - }} - RETURN_TYPES = ("MODEL",) - FUNCTION = "patch" + def define_schema(cls): + return IO.Schema( + node_id="FreeU_V2", + category="model_patches/unet", + inputs=[ + IO.Model.Input("model"), + IO.Float.Input("b1", default=1.3, min=0.0, max=10.0, step=0.01), + IO.Float.Input("b2", default=1.4, min=0.0, max=10.0, step=0.01), + IO.Float.Input("s1", default=0.9, min=0.0, max=10.0, step=0.01), + IO.Float.Input("s2", default=0.2, min=0.0, max=10.0, step=0.01), + ], + outputs=[ + IO.Model.Output(), + ], + ) - CATEGORY = "model_patches/unet" - - def patch(self, model, b1, b2, s1, s2): + @classmethod + def execute(cls, model, b1, b2, s1, s2) -> IO.NodeOutput: model_channels = model.model.model_config.unet_config["model_channels"] scale_dict = {model_channels * 4: (b1, s1), model_channels * 2: (b2, s2)} on_cpu_devices = {} @@ -105,9 +120,19 @@ class FreeU_V2: m = model.clone() m.set_model_output_block_patch(output_block_patch) - return (m, ) + return IO.NodeOutput(m) -NODE_CLASS_MAPPINGS = { - "FreeU": FreeU, - "FreeU_V2": FreeU_V2, -} + patch = execute # TODO: remove + + +class FreelunchExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + FreeU, + FreeU_V2, + ] + + +async def comfy_entrypoint() -> FreelunchExtension: + return FreelunchExtension() diff --git a/comfy_extras/nodes_hunyuan.py b/comfy_extras/nodes_hunyuan.py index f7c34d059..ceff657d3 100644 --- a/comfy_extras/nodes_hunyuan.py +++ b/comfy_extras/nodes_hunyuan.py @@ -4,7 +4,10 @@ import torch import comfy.model_management from typing_extensions import override from comfy_api.latest import ComfyExtension, io - +from comfy.ldm.hunyuan_video.upsampler import HunyuanVideo15SRModel +from comfy.ldm.lightricks.latent_upsampler import LatentUpsampler +import folder_paths +import json class CLIPTextEncodeHunyuanDiT(io.ComfyNode): @classmethod @@ -37,6 +40,7 @@ class EmptyHunyuanLatentVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="EmptyHunyuanLatentVideo", + display_name="Empty HunyuanVideo 1.0 Latent", category="latent/video", inputs=[ io.Int.Input("width", default=848, min=16, max=nodes.MAX_RESOLUTION, step=16), @@ -57,6 +61,203 @@ class EmptyHunyuanLatentVideo(io.ComfyNode): generate = execute # TODO: remove +class EmptyHunyuanVideo15Latent(EmptyHunyuanLatentVideo): + @classmethod + def define_schema(cls): + schema = super().define_schema() + schema.node_id = "EmptyHunyuanVideo15Latent" + schema.display_name = "Empty HunyuanVideo 1.5 Latent" + return schema + + @classmethod + def execute(cls, width, height, length, batch_size=1) -> io.NodeOutput: + # Using scale factor of 16 instead of 8 + latent = torch.zeros([batch_size, 32, ((length - 1) // 4) + 1, height // 16, width // 16], device=comfy.model_management.intermediate_device()) + return io.NodeOutput({"samples": latent}) + + +class HunyuanVideo15ImageToVideo(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="HunyuanVideo15ImageToVideo", + category="conditioning/video_models", + inputs=[ + io.Conditioning.Input("positive"), + io.Conditioning.Input("negative"), + io.Vae.Input("vae"), + io.Int.Input("width", default=848, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("length", default=33, min=1, max=nodes.MAX_RESOLUTION, step=4), + io.Int.Input("batch_size", default=1, min=1, max=4096), + io.Image.Input("start_image", optional=True), + io.ClipVisionOutput.Input("clip_vision_output", optional=True), + ], + outputs=[ + io.Conditioning.Output(display_name="positive"), + io.Conditioning.Output(display_name="negative"), + io.Latent.Output(display_name="latent"), + ], + ) + + @classmethod + def execute(cls, positive, negative, vae, width, height, length, batch_size, start_image=None, clip_vision_output=None) -> io.NodeOutput: + latent = torch.zeros([batch_size, 32, ((length - 1) // 4) + 1, height // 16, width // 16], device=comfy.model_management.intermediate_device()) + + if start_image is not None: + start_image = comfy.utils.common_upscale(start_image[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1) + + encoded = vae.encode(start_image[:, :, :, :3]) + concat_latent_image = torch.zeros((latent.shape[0], 32, latent.shape[2], latent.shape[3], latent.shape[4]), device=comfy.model_management.intermediate_device()) + concat_latent_image[:, :, :encoded.shape[2], :, :] = encoded + + mask = torch.ones((1, 1, latent.shape[2], concat_latent_image.shape[-2], concat_latent_image.shape[-1]), device=start_image.device, dtype=start_image.dtype) + mask[:, :, :((start_image.shape[0] - 1) // 4) + 1] = 0.0 + + positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": concat_latent_image, "concat_mask": mask}) + negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": concat_latent_image, "concat_mask": mask}) + + if clip_vision_output is not None: + positive = node_helpers.conditioning_set_values(positive, {"clip_vision_output": clip_vision_output}) + negative = node_helpers.conditioning_set_values(negative, {"clip_vision_output": clip_vision_output}) + + out_latent = {} + out_latent["samples"] = latent + return io.NodeOutput(positive, negative, out_latent) + + +class HunyuanVideo15SuperResolution(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="HunyuanVideo15SuperResolution", + inputs=[ + io.Conditioning.Input("positive"), + io.Conditioning.Input("negative"), + io.Vae.Input("vae", optional=True), + io.Image.Input("start_image", optional=True), + io.ClipVisionOutput.Input("clip_vision_output", optional=True), + io.Latent.Input("latent"), + io.Float.Input("noise_augmentation", default=0.70, min=0.0, max=1.0, step=0.01), + + ], + outputs=[ + io.Conditioning.Output(display_name="positive"), + io.Conditioning.Output(display_name="negative"), + io.Latent.Output(display_name="latent"), + ], + ) + + @classmethod + def execute(cls, positive, negative, latent, noise_augmentation, vae=None, start_image=None, clip_vision_output=None) -> io.NodeOutput: + in_latent = latent["samples"] + in_channels = in_latent.shape[1] + cond_latent = torch.zeros([in_latent.shape[0], in_channels * 2 + 2, in_latent.shape[-3], in_latent.shape[-2], in_latent.shape[-1]], device=comfy.model_management.intermediate_device()) + cond_latent[:, in_channels + 1 : 2 * in_channels + 1] = in_latent + cond_latent[:, 2 * in_channels + 1] = 1 + if start_image is not None: + start_image = comfy.utils.common_upscale(start_image.movedim(-1, 1), in_latent.shape[-1] * 16, in_latent.shape[-2] * 16, "bilinear", "center").movedim(1, -1) + encoded = vae.encode(start_image[:, :, :, :3]) + cond_latent[:, :in_channels, :encoded.shape[2], :, :] = encoded + cond_latent[:, in_channels + 1, 0] = 1 + + positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": cond_latent, "noise_augmentation": noise_augmentation}) + negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": cond_latent, "noise_augmentation": noise_augmentation}) + if clip_vision_output is not None: + positive = node_helpers.conditioning_set_values(positive, {"clip_vision_output": clip_vision_output}) + negative = node_helpers.conditioning_set_values(negative, {"clip_vision_output": clip_vision_output}) + + return io.NodeOutput(positive, negative, latent) + + +class LatentUpscaleModelLoader(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LatentUpscaleModelLoader", + display_name="Load Latent Upscale Model", + category="loaders", + inputs=[ + io.Combo.Input("model_name", options=folder_paths.get_filename_list("latent_upscale_models")), + ], + outputs=[ + io.LatentUpscaleModel.Output(), + ], + ) + + @classmethod + def execute(cls, model_name) -> io.NodeOutput: + model_path = folder_paths.get_full_path_or_raise("latent_upscale_models", model_name) + sd, metadata = comfy.utils.load_torch_file(model_path, safe_load=True, return_metadata=True) + + if "blocks.0.block.0.conv.weight" in sd: + config = { + "in_channels": sd["in_conv.conv.weight"].shape[1], + "out_channels": sd["out_conv.conv.weight"].shape[0], + "hidden_channels": sd["in_conv.conv.weight"].shape[0], + "num_blocks": len([k for k in sd.keys() if k.startswith("blocks.") and k.endswith(".block.0.conv.weight")]), + "global_residual": False, + } + model_type = "720p" + model = HunyuanVideo15SRModel(model_type, config) + model.load_sd(sd) + elif "up.0.block.0.conv1.conv.weight" in sd: + sd = {key.replace("nin_shortcut", "nin_shortcut.conv", 1): value for key, value in sd.items()} + config = { + "z_channels": sd["conv_in.conv.weight"].shape[1], + "out_channels": sd["conv_out.conv.weight"].shape[0], + "block_out_channels": tuple(sd[f"up.{i}.block.0.conv1.conv.weight"].shape[0] for i in range(len([k for k in sd.keys() if k.startswith("up.") and k.endswith(".block.0.conv1.conv.weight")]))), + } + model_type = "1080p" + model = HunyuanVideo15SRModel(model_type, config) + model.load_sd(sd) + elif "post_upsample_res_blocks.0.conv2.bias" in sd: + config = json.loads(metadata["config"]) + model = LatentUpsampler.from_config(config).to(dtype=comfy.model_management.vae_dtype(allowed_dtypes=[torch.bfloat16, torch.float32])) + model.load_state_dict(sd) + + return io.NodeOutput(model) + + +class HunyuanVideo15LatentUpscaleWithModel(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="HunyuanVideo15LatentUpscaleWithModel", + display_name="Hunyuan Video 15 Latent Upscale With Model", + category="latent", + inputs=[ + io.LatentUpscaleModel.Input("model"), + io.Latent.Input("samples"), + io.Combo.Input("upscale_method", options=["nearest-exact", "bilinear", "area", "bicubic", "bislerp"], default="bilinear"), + io.Int.Input("width", default=1280, min=0, max=16384, step=8), + io.Int.Input("height", default=720, min=0, max=16384, step=8), + io.Combo.Input("crop", options=["disabled", "center"]), + ], + outputs=[ + io.Latent.Output(), + ], + ) + + @classmethod + def execute(cls, model, samples, upscale_method, width, height, crop) -> io.NodeOutput: + if width == 0 and height == 0: + return io.NodeOutput(samples) + else: + if width == 0: + height = max(64, height) + width = max(64, round(samples["samples"].shape[-1] * height / samples["samples"].shape[-2])) + elif height == 0: + width = max(64, width) + height = max(64, round(samples["samples"].shape[-2] * width / samples["samples"].shape[-1])) + else: + width = max(64, width) + height = max(64, height) + s = comfy.utils.common_upscale(samples["samples"], width // 16, height // 16, upscale_method, crop) + s = model.resample_latent(s) + return io.NodeOutput({"samples": s.cpu().float()}) + + PROMPT_TEMPLATE_ENCODE_VIDEO_I2V = ( "<|start_header_id|>system<|end_header_id|>\n\n\nDescribe the video by detailing the following aspects according to the reference image: " "1. The main content and theme of the video." @@ -210,6 +411,11 @@ class HunyuanExtension(ComfyExtension): CLIPTextEncodeHunyuanDiT, TextEncodeHunyuanVideo_ImageToVideo, EmptyHunyuanLatentVideo, + EmptyHunyuanVideo15Latent, + HunyuanVideo15ImageToVideo, + HunyuanVideo15SuperResolution, + HunyuanVideo15LatentUpscaleWithModel, + LatentUpscaleModelLoader, HunyuanImageToVideo, EmptyHunyuanImageLatent, HunyuanRefinerLatent, diff --git a/comfy_extras/nodes_hunyuan3d.py b/comfy_extras/nodes_hunyuan3d.py index f6e71e0a8..adca14f62 100644 --- a/comfy_extras/nodes_hunyuan3d.py +++ b/comfy_extras/nodes_hunyuan3d.py @@ -7,63 +7,79 @@ from comfy.ldm.modules.diffusionmodules.mmdit import get_1d_sincos_pos_embed_fro import folder_paths import comfy.model_management from comfy.cli_args import args +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 -class EmptyLatentHunyuan3Dv2: + +class EmptyLatentHunyuan3Dv2(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "resolution": ("INT", {"default": 3072, "min": 1, "max": 8192}), - "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096, "tooltip": "The number of latent images in the batch."}), - } - } + def define_schema(cls): + return IO.Schema( + node_id="EmptyLatentHunyuan3Dv2", + category="latent/3d", + inputs=[ + IO.Int.Input("resolution", default=3072, min=1, max=8192), + IO.Int.Input("batch_size", default=1, min=1, max=4096, tooltip="The number of latent images in the batch."), + ], + outputs=[ + IO.Latent.Output(), + ] + ) - RETURN_TYPES = ("LATENT",) - FUNCTION = "generate" - - CATEGORY = "latent/3d" - - def generate(self, resolution, batch_size): + @classmethod + def execute(cls, resolution, batch_size) -> IO.NodeOutput: latent = torch.zeros([batch_size, 64, resolution], device=comfy.model_management.intermediate_device()) - return ({"samples": latent, "type": "hunyuan3dv2"}, ) + return IO.NodeOutput({"samples": latent, "type": "hunyuan3dv2"}) -class Hunyuan3Dv2Conditioning: + generate = execute # TODO: remove + + +class Hunyuan3Dv2Conditioning(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": {"clip_vision_output": ("CLIP_VISION_OUTPUT",), - }} + def define_schema(cls): + return IO.Schema( + node_id="Hunyuan3Dv2Conditioning", + category="conditioning/video_models", + inputs=[ + IO.ClipVisionOutput.Input("clip_vision_output"), + ], + outputs=[ + IO.Conditioning.Output(display_name="positive"), + IO.Conditioning.Output(display_name="negative"), + ] + ) - RETURN_TYPES = ("CONDITIONING", "CONDITIONING") - RETURN_NAMES = ("positive", "negative") - - FUNCTION = "encode" - - CATEGORY = "conditioning/video_models" - - def encode(self, clip_vision_output): + @classmethod + def execute(cls, clip_vision_output) -> IO.NodeOutput: embeds = clip_vision_output.last_hidden_state positive = [[embeds, {}]] negative = [[torch.zeros_like(embeds), {}]] - return (positive, negative) + return IO.NodeOutput(positive, negative) + + encode = execute # TODO: remove -class Hunyuan3Dv2ConditioningMultiView: +class Hunyuan3Dv2ConditioningMultiView(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": {}, - "optional": {"front": ("CLIP_VISION_OUTPUT",), - "left": ("CLIP_VISION_OUTPUT",), - "back": ("CLIP_VISION_OUTPUT",), - "right": ("CLIP_VISION_OUTPUT",), }} + def define_schema(cls): + return IO.Schema( + node_id="Hunyuan3Dv2ConditioningMultiView", + category="conditioning/video_models", + inputs=[ + IO.ClipVisionOutput.Input("front", optional=True), + IO.ClipVisionOutput.Input("left", optional=True), + IO.ClipVisionOutput.Input("back", optional=True), + IO.ClipVisionOutput.Input("right", optional=True), + ], + outputs=[ + IO.Conditioning.Output(display_name="positive"), + IO.Conditioning.Output(display_name="negative"), + ] + ) - RETURN_TYPES = ("CONDITIONING", "CONDITIONING") - RETURN_NAMES = ("positive", "negative") - - FUNCTION = "encode" - - CATEGORY = "conditioning/video_models" - - def encode(self, front=None, left=None, back=None, right=None): + @classmethod + def execute(cls, front=None, left=None, back=None, right=None) -> IO.NodeOutput: all_embeds = [front, left, back, right] out = [] pos_embeds = None @@ -76,29 +92,35 @@ class Hunyuan3Dv2ConditioningMultiView: embeds = torch.cat(out, dim=1) positive = [[embeds, {}]] negative = [[torch.zeros_like(embeds), {}]] - return (positive, negative) + return IO.NodeOutput(positive, negative) + + encode = execute # TODO: remove -class VOXEL: - def __init__(self, data): - self.data = data - -class VAEDecodeHunyuan3D: +class VAEDecodeHunyuan3D(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": {"samples": ("LATENT", ), - "vae": ("VAE", ), - "num_chunks": ("INT", {"default": 8000, "min": 1000, "max": 500000}), - "octree_resolution": ("INT", {"default": 256, "min": 16, "max": 512}), - }} - RETURN_TYPES = ("VOXEL",) - FUNCTION = "decode" + def define_schema(cls): + return IO.Schema( + node_id="VAEDecodeHunyuan3D", + category="latent/3d", + inputs=[ + IO.Latent.Input("samples"), + IO.Vae.Input("vae"), + IO.Int.Input("num_chunks", default=8000, min=1000, max=500000), + IO.Int.Input("octree_resolution", default=256, min=16, max=512), + ], + outputs=[ + IO.Voxel.Output(), + ] + ) - CATEGORY = "latent/3d" + @classmethod + def execute(cls, vae, samples, num_chunks, octree_resolution) -> IO.NodeOutput: + voxels = Types.VOXEL(vae.decode(samples["samples"], vae_options={"num_chunks": num_chunks, "octree_resolution": octree_resolution})) + return IO.NodeOutput(voxels) + + decode = execute # TODO: remove - def decode(self, vae, samples, num_chunks, octree_resolution): - voxels = VOXEL(vae.decode(samples["samples"], vae_options={"num_chunks": num_chunks, "octree_resolution": octree_resolution})) - return (voxels, ) def voxel_to_mesh(voxels, threshold=0.5, device=None): if device is None: @@ -396,24 +418,24 @@ def voxel_to_mesh_surfnet(voxels, threshold=0.5, device=None): return final_vertices, faces -class MESH: - def __init__(self, vertices, faces): - self.vertices = vertices - self.faces = faces - -class VoxelToMeshBasic: +class VoxelToMeshBasic(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": {"voxel": ("VOXEL", ), - "threshold": ("FLOAT", {"default": 0.6, "min": -1.0, "max": 1.0, "step": 0.01}), - }} - RETURN_TYPES = ("MESH",) - FUNCTION = "decode" + def define_schema(cls): + return IO.Schema( + node_id="VoxelToMeshBasic", + category="3d", + inputs=[ + IO.Voxel.Input("voxel"), + IO.Float.Input("threshold", default=0.6, min=-1.0, max=1.0, step=0.01), + ], + outputs=[ + IO.Mesh.Output(), + ] + ) - CATEGORY = "3d" - - def decode(self, voxel, threshold): + @classmethod + def execute(cls, voxel, threshold) -> IO.NodeOutput: vertices = [] faces = [] for x in voxel.data: @@ -421,21 +443,29 @@ class VoxelToMeshBasic: vertices.append(v) faces.append(f) - return (MESH(torch.stack(vertices), torch.stack(faces)), ) + return IO.NodeOutput(Types.MESH(torch.stack(vertices), torch.stack(faces))) -class VoxelToMesh: + decode = execute # TODO: remove + + +class VoxelToMesh(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": {"voxel": ("VOXEL", ), - "algorithm": (["surface net", "basic"], ), - "threshold": ("FLOAT", {"default": 0.6, "min": -1.0, "max": 1.0, "step": 0.01}), - }} - RETURN_TYPES = ("MESH",) - FUNCTION = "decode" + def define_schema(cls): + return IO.Schema( + node_id="VoxelToMesh", + category="3d", + inputs=[ + IO.Voxel.Input("voxel"), + IO.Combo.Input("algorithm", options=["surface net", "basic"]), + IO.Float.Input("threshold", default=0.6, min=-1.0, max=1.0, step=0.01), + ], + outputs=[ + IO.Mesh.Output(), + ] + ) - CATEGORY = "3d" - - def decode(self, voxel, algorithm, threshold): + @classmethod + def execute(cls, voxel, algorithm, threshold) -> IO.NodeOutput: vertices = [] faces = [] @@ -449,7 +479,9 @@ class VoxelToMesh: vertices.append(v) faces.append(f) - return (MESH(torch.stack(vertices), torch.stack(faces)), ) + return IO.NodeOutput(Types.MESH(torch.stack(vertices), torch.stack(faces))) + + decode = execute # TODO: remove def save_glb(vertices, faces, filepath, metadata=None): @@ -581,31 +613,32 @@ def save_glb(vertices, faces, filepath, metadata=None): return filepath -class SaveGLB: +class SaveGLB(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": {"mesh": ("MESH", ), - "filename_prefix": ("STRING", {"default": "mesh/ComfyUI"}), }, - "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"}, } + def define_schema(cls): + return IO.Schema( + node_id="SaveGLB", + category="3d", + is_output_node=True, + inputs=[ + IO.Mesh.Input("mesh"), + IO.String.Input("filename_prefix", default="mesh/ComfyUI"), + ], + hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo] + ) - RETURN_TYPES = () - FUNCTION = "save" - - OUTPUT_NODE = True - - CATEGORY = "3d" - - def save(self, mesh, filename_prefix, prompt=None, extra_pnginfo=None): + @classmethod + def execute(cls, mesh, filename_prefix) -> IO.NodeOutput: full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix, folder_paths.get_output_directory()) results = [] metadata = {} if not args.disable_metadata: - if prompt is not None: - metadata["prompt"] = json.dumps(prompt) - if extra_pnginfo is not None: - for x in extra_pnginfo: - metadata[x] = json.dumps(extra_pnginfo[x]) + if cls.hidden.prompt is not None: + metadata["prompt"] = json.dumps(cls.hidden.prompt) + if cls.hidden.extra_pnginfo is not None: + for x in cls.hidden.extra_pnginfo: + metadata[x] = json.dumps(cls.hidden.extra_pnginfo[x]) for i in range(mesh.vertices.shape[0]): f = f"{filename}_{counter:05}_.glb" @@ -616,15 +649,22 @@ class SaveGLB: "type": "output" }) counter += 1 - return {"ui": {"3d": results}} + return IO.NodeOutput(ui={"3d": results}) -NODE_CLASS_MAPPINGS = { - "EmptyLatentHunyuan3Dv2": EmptyLatentHunyuan3Dv2, - "Hunyuan3Dv2Conditioning": Hunyuan3Dv2Conditioning, - "Hunyuan3Dv2ConditioningMultiView": Hunyuan3Dv2ConditioningMultiView, - "VAEDecodeHunyuan3D": VAEDecodeHunyuan3D, - "VoxelToMeshBasic": VoxelToMeshBasic, - "VoxelToMesh": VoxelToMesh, - "SaveGLB": SaveGLB, -} +class Hunyuan3dExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + EmptyLatentHunyuan3Dv2, + Hunyuan3Dv2Conditioning, + Hunyuan3Dv2ConditioningMultiView, + VAEDecodeHunyuan3D, + VoxelToMeshBasic, + VoxelToMesh, + SaveGLB, + ] + + +async def comfy_entrypoint() -> Hunyuan3dExtension: + return Hunyuan3dExtension() diff --git a/comfy_extras/nodes_images.py b/comfy_extras/nodes_images.py index 392aea32c..ce21caade 100644 --- a/comfy_extras/nodes_images.py +++ b/comfy_extras/nodes_images.py @@ -2,280 +2,231 @@ from __future__ import annotations import nodes import folder_paths -from comfy.cli_args import args -from PIL import Image -from PIL.PngImagePlugin import PngInfo - -import numpy as np import json import os import re -from io import BytesIO -from inspect import cleandoc import torch import comfy.utils -from comfy.comfy_types import FileLocator, IO from server import PromptServer +from comfy_api.latest import ComfyExtension, IO, UI +from typing_extensions import override + +SVG = IO.SVG.Type # TODO: temporary solution for backward compatibility, will be removed later. MAX_RESOLUTION = nodes.MAX_RESOLUTION -class ImageCrop: +class ImageCrop(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { "image": ("IMAGE",), - "width": ("INT", {"default": 512, "min": 1, "max": MAX_RESOLUTION, "step": 1}), - "height": ("INT", {"default": 512, "min": 1, "max": MAX_RESOLUTION, "step": 1}), - "x": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - "y": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - }} - RETURN_TYPES = ("IMAGE",) - FUNCTION = "crop" + def define_schema(cls): + return IO.Schema( + node_id="ImageCrop", + display_name="Image Crop", + category="image/transform", + inputs=[ + IO.Image.Input("image"), + IO.Int.Input("width", default=512, min=1, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("height", default=512, min=1, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("x", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("y", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + ], + outputs=[IO.Image.Output()], + ) - CATEGORY = "image/transform" - - def crop(self, image, width, height, x, y): + @classmethod + def execute(cls, image, width, height, x, y) -> IO.NodeOutput: x = min(x, image.shape[2] - 1) y = min(y, image.shape[1] - 1) to_x = width + x to_y = height + y img = image[:,y:to_y, x:to_x, :] - return (img,) + return IO.NodeOutput(img) -class RepeatImageBatch: + crop = execute # TODO: remove + + +class RepeatImageBatch(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { "image": ("IMAGE",), - "amount": ("INT", {"default": 1, "min": 1, "max": 4096}), - }} - RETURN_TYPES = ("IMAGE",) - FUNCTION = "repeat" + def define_schema(cls): + return IO.Schema( + node_id="RepeatImageBatch", + category="image/batch", + inputs=[ + IO.Image.Input("image"), + IO.Int.Input("amount", default=1, min=1, max=4096), + ], + outputs=[IO.Image.Output()], + ) - CATEGORY = "image/batch" - - def repeat(self, image, amount): + @classmethod + def execute(cls, image, amount) -> IO.NodeOutput: s = image.repeat((amount, 1,1,1)) - return (s,) + return IO.NodeOutput(s) -class ImageFromBatch: + repeat = execute # TODO: remove + + +class ImageFromBatch(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { "image": ("IMAGE",), - "batch_index": ("INT", {"default": 0, "min": 0, "max": 4095}), - "length": ("INT", {"default": 1, "min": 1, "max": 4096}), - }} - RETURN_TYPES = ("IMAGE",) - FUNCTION = "frombatch" + def define_schema(cls): + return IO.Schema( + node_id="ImageFromBatch", + category="image/batch", + inputs=[ + IO.Image.Input("image"), + IO.Int.Input("batch_index", default=0, min=0, max=4095), + IO.Int.Input("length", default=1, min=1, max=4096), + ], + outputs=[IO.Image.Output()], + ) - CATEGORY = "image/batch" - - def frombatch(self, image, batch_index, length): + @classmethod + def execute(cls, image, batch_index, length) -> IO.NodeOutput: s_in = image batch_index = min(s_in.shape[0] - 1, batch_index) length = min(s_in.shape[0] - batch_index, length) s = s_in[batch_index:batch_index + length].clone() - return (s,) + return IO.NodeOutput(s) + + frombatch = execute # TODO: remove -class ImageAddNoise: +class ImageAddNoise(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { "image": ("IMAGE",), - "seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff, "control_after_generate": True, "tooltip": "The random seed used for creating the noise."}), - "strength": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 1.0, "step": 0.01}), - }} - RETURN_TYPES = ("IMAGE",) - FUNCTION = "repeat" + def define_schema(cls): + return IO.Schema( + node_id="ImageAddNoise", + category="image", + inputs=[ + IO.Image.Input("image"), + IO.Int.Input( + "seed", + default=0, + min=0, + max=0xFFFFFFFFFFFFFFFF, + control_after_generate=True, + tooltip="The random seed used for creating the noise.", + ), + IO.Float.Input("strength", default=0.5, min=0.0, max=1.0, step=0.01), + ], + outputs=[IO.Image.Output()], + ) - CATEGORY = "image" - - def repeat(self, image, seed, strength): + @classmethod + def execute(cls, image, seed, strength) -> IO.NodeOutput: generator = torch.manual_seed(seed) s = torch.clip((image + strength * torch.randn(image.size(), generator=generator, device="cpu").to(image)), min=0.0, max=1.0) - return (s,) + return IO.NodeOutput(s) -class SaveAnimatedWEBP: - def __init__(self): - self.output_dir = folder_paths.get_output_directory() - self.type = "output" - self.prefix_append = "" + repeat = execute # TODO: remove - methods = {"default": 4, "fastest": 0, "slowest": 6} - @classmethod - def INPUT_TYPES(s): - return {"required": - {"images": ("IMAGE", ), - "filename_prefix": ("STRING", {"default": "ComfyUI"}), - "fps": ("FLOAT", {"default": 6.0, "min": 0.01, "max": 1000.0, "step": 0.01}), - "lossless": ("BOOLEAN", {"default": True}), - "quality": ("INT", {"default": 80, "min": 0, "max": 100}), - "method": (list(s.methods.keys()),), - # "num_frames": ("INT", {"default": 0, "min": 0, "max": 8192}), - }, - "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"}, - } - RETURN_TYPES = () - FUNCTION = "save_images" - - OUTPUT_NODE = True - - CATEGORY = "image/animation" - - def save_images(self, images, fps, filename_prefix, lossless, quality, method, num_frames=0, prompt=None, extra_pnginfo=None): - method = self.methods.get(method) - filename_prefix += self.prefix_append - full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix, self.output_dir, images[0].shape[1], images[0].shape[0]) - results: list[FileLocator] = [] - pil_images = [] - for image in images: - i = 255. * image.cpu().numpy() - img = Image.fromarray(np.clip(i, 0, 255).astype(np.uint8)) - pil_images.append(img) - - metadata = pil_images[0].getexif() - if not args.disable_metadata: - if prompt is not None: - metadata[0x0110] = "prompt:{}".format(json.dumps(prompt)) - if extra_pnginfo is not None: - inital_exif = 0x010f - for x in extra_pnginfo: - metadata[inital_exif] = "{}:{}".format(x, json.dumps(extra_pnginfo[x])) - inital_exif -= 1 - - if num_frames == 0: - num_frames = len(pil_images) - - c = len(pil_images) - for i in range(0, c, num_frames): - file = f"{filename}_{counter:05}_.webp" - pil_images[i].save(os.path.join(full_output_folder, file), save_all=True, duration=int(1000.0/fps), append_images=pil_images[i + 1:i + num_frames], exif=metadata, lossless=lossless, quality=quality, method=method) - results.append({ - "filename": file, - "subfolder": subfolder, - "type": self.type - }) - counter += 1 - - animated = num_frames != 1 - return { "ui": { "images": results, "animated": (animated,) } } - -class SaveAnimatedPNG: - def __init__(self): - self.output_dir = folder_paths.get_output_directory() - self.type = "output" - self.prefix_append = "" +class SaveAnimatedWEBP(IO.ComfyNode): + COMPRESS_METHODS = {"default": 4, "fastest": 0, "slowest": 6} @classmethod - def INPUT_TYPES(s): - return {"required": - {"images": ("IMAGE", ), - "filename_prefix": ("STRING", {"default": "ComfyUI"}), - "fps": ("FLOAT", {"default": 6.0, "min": 0.01, "max": 1000.0, "step": 0.01}), - "compress_level": ("INT", {"default": 4, "min": 0, "max": 9}) - }, - "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"}, - } + def define_schema(cls): + return IO.Schema( + node_id="SaveAnimatedWEBP", + category="image/animation", + inputs=[ + IO.Image.Input("images"), + IO.String.Input("filename_prefix", default="ComfyUI"), + IO.Float.Input("fps", default=6.0, min=0.01, max=1000.0, step=0.01), + IO.Boolean.Input("lossless", default=True), + IO.Int.Input("quality", default=80, min=0, max=100), + IO.Combo.Input("method", options=list(cls.COMPRESS_METHODS.keys())), + # "num_frames": ("INT", {"default": 0, "min": 0, "max": 8192}), + ], + hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo], + is_output_node=True, + ) - RETURN_TYPES = () - FUNCTION = "save_images" + @classmethod + def execute(cls, images, fps, filename_prefix, lossless, quality, method, num_frames=0) -> IO.NodeOutput: + return IO.NodeOutput( + ui=UI.ImageSaveHelper.get_save_animated_webp_ui( + images=images, + filename_prefix=filename_prefix, + cls=cls, + fps=fps, + lossless=lossless, + quality=quality, + method=cls.COMPRESS_METHODS.get(method) + ) + ) - OUTPUT_NODE = True - - CATEGORY = "image/animation" - - def save_images(self, images, fps, compress_level, filename_prefix="ComfyUI", prompt=None, extra_pnginfo=None): - filename_prefix += self.prefix_append - full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix, self.output_dir, images[0].shape[1], images[0].shape[0]) - results = list() - pil_images = [] - for image in images: - i = 255. * image.cpu().numpy() - img = Image.fromarray(np.clip(i, 0, 255).astype(np.uint8)) - pil_images.append(img) - - metadata = None - if not args.disable_metadata: - metadata = PngInfo() - if prompt is not None: - metadata.add(b"comf", "prompt".encode("latin-1", "strict") + b"\0" + json.dumps(prompt).encode("latin-1", "strict"), after_idat=True) - if extra_pnginfo is not None: - for x in extra_pnginfo: - metadata.add(b"comf", x.encode("latin-1", "strict") + b"\0" + json.dumps(extra_pnginfo[x]).encode("latin-1", "strict"), after_idat=True) - - file = f"{filename}_{counter:05}_.png" - pil_images[0].save(os.path.join(full_output_folder, file), pnginfo=metadata, compress_level=compress_level, save_all=True, duration=int(1000.0/fps), append_images=pil_images[1:]) - results.append({ - "filename": file, - "subfolder": subfolder, - "type": self.type - }) - - return { "ui": { "images": results, "animated": (True,)} } - -class SVG: - """ - Stores SVG representations via a list of BytesIO objects. - """ - def __init__(self, data: list[BytesIO]): - self.data = data - - def combine(self, other: 'SVG') -> 'SVG': - return SVG(self.data + other.data) - - @staticmethod - def combine_all(svgs: list['SVG']) -> 'SVG': - all_svgs_list: list[BytesIO] = [] - for svg_item in svgs: - all_svgs_list.extend(svg_item.data) - return SVG(all_svgs_list) + save_images = execute # TODO: remove -class ImageStitch: +class SaveAnimatedPNG(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="SaveAnimatedPNG", + category="image/animation", + inputs=[ + IO.Image.Input("images"), + IO.String.Input("filename_prefix", default="ComfyUI"), + IO.Float.Input("fps", default=6.0, min=0.01, max=1000.0, step=0.01), + IO.Int.Input("compress_level", default=4, min=0, max=9), + ], + hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo], + is_output_node=True, + ) + + @classmethod + def execute(cls, images, fps, compress_level, filename_prefix="ComfyUI") -> IO.NodeOutput: + return IO.NodeOutput( + ui=UI.ImageSaveHelper.get_save_animated_png_ui( + images=images, + filename_prefix=filename_prefix, + cls=cls, + fps=fps, + compress_level=compress_level, + ) + ) + + save_images = execute # TODO: remove + + +class ImageStitch(IO.ComfyNode): """Upstreamed from https://github.com/kijai/ComfyUI-KJNodes""" @classmethod - def INPUT_TYPES(s): - return { - "required": { - "image1": ("IMAGE",), - "direction": (["right", "down", "left", "up"], {"default": "right"}), - "match_image_size": ("BOOLEAN", {"default": True}), - "spacing_width": ( - "INT", - {"default": 0, "min": 0, "max": 1024, "step": 2}, - ), - "spacing_color": ( - ["white", "black", "red", "green", "blue"], - {"default": "white"}, - ), - }, - "optional": { - "image2": ("IMAGE",), - }, - } + def define_schema(cls): + return IO.Schema( + node_id="ImageStitch", + display_name="Image Stitch", + description="Stitches image2 to image1 in the specified direction.\n" + "If image2 is not provided, returns image1 unchanged.\n" + "Optional spacing can be added between images.", + category="image/transform", + inputs=[ + IO.Image.Input("image1"), + IO.Combo.Input("direction", options=["right", "down", "left", "up"], default="right"), + IO.Boolean.Input("match_image_size", default=True), + IO.Int.Input("spacing_width", default=0, min=0, max=1024, step=2), + IO.Combo.Input("spacing_color", options=["white", "black", "red", "green", "blue"], default="white"), + IO.Image.Input("image2", optional=True), + ], + outputs=[IO.Image.Output()], + ) - RETURN_TYPES = ("IMAGE",) - FUNCTION = "stitch" - CATEGORY = "image/transform" - DESCRIPTION = """ -Stitches image2 to image1 in the specified direction. -If image2 is not provided, returns image1 unchanged. -Optional spacing can be added between images. -""" - - def stitch( - self, + @classmethod + def execute( + cls, image1, direction, match_image_size, spacing_width, spacing_color, image2=None, - ): + ) -> IO.NodeOutput: if image2 is None: - return (image1,) + return IO.NodeOutput(image1) # Handle batch size differences if image1.shape[0] != image2.shape[0]: @@ -412,36 +363,30 @@ Optional spacing can be added between images. images.insert(1, spacing) concat_dim = 2 if direction in ["left", "right"] else 1 - return (torch.cat(images, dim=concat_dim),) + return IO.NodeOutput(torch.cat(images, dim=concat_dim)) + + stitch = execute # TODO: remove + + +class ResizeAndPadImage(IO.ComfyNode): -class ResizeAndPadImage: @classmethod - def INPUT_TYPES(cls): - return { - "required": { - "image": ("IMAGE",), - "target_width": ("INT", { - "default": 512, - "min": 1, - "max": MAX_RESOLUTION, - "step": 1 - }), - "target_height": ("INT", { - "default": 512, - "min": 1, - "max": MAX_RESOLUTION, - "step": 1 - }), - "padding_color": (["white", "black"],), - "interpolation": (["area", "bicubic", "nearest-exact", "bilinear", "lanczos"],), - } - } + def define_schema(cls): + return IO.Schema( + node_id="ResizeAndPadImage", + category="image/transform", + inputs=[ + IO.Image.Input("image"), + IO.Int.Input("target_width", default=512, min=1, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("target_height", default=512, min=1, max=nodes.MAX_RESOLUTION, step=1), + IO.Combo.Input("padding_color", options=["white", "black"]), + IO.Combo.Input("interpolation", options=["area", "bicubic", "nearest-exact", "bilinear", "lanczos"]), + ], + outputs=[IO.Image.Output()], + ) - RETURN_TYPES = ("IMAGE",) - FUNCTION = "resize_and_pad" - CATEGORY = "image/transform" - - def resize_and_pad(self, image, target_width, target_height, padding_color, interpolation): + @classmethod + def execute(cls, image, target_width, target_height, padding_color, interpolation) -> IO.NodeOutput: batch_size, orig_height, orig_width, channels = image.shape scale_w = target_width / orig_width @@ -469,52 +414,47 @@ class ResizeAndPadImage: padded[:, :, y_offset:y_offset + new_height, x_offset:x_offset + new_width] = resized output = padded.permute(0, 2, 3, 1) - return (output,) + return IO.NodeOutput(output) -class SaveSVGNode: - """ - Save SVG files on disk. - """ + resize_and_pad = execute # TODO: remove - def __init__(self): - self.output_dir = folder_paths.get_output_directory() - self.type = "output" - self.prefix_append = "" - RETURN_TYPES = () - DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value - FUNCTION = "save_svg" - CATEGORY = "image/save" # Changed - OUTPUT_NODE = True +class SaveSVGNode(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "svg": ("SVG",), # Changed - "filename_prefix": ("STRING", {"default": "svg/ComfyUI", "tooltip": "The prefix for the file to save. This may include formatting information such as %date:yyyy-MM-dd% or %Empty Latent Image.width% to include values from nodes."}) - }, - "hidden": { - "prompt": "PROMPT", - "extra_pnginfo": "EXTRA_PNGINFO" - } - } + def define_schema(cls): + return IO.Schema( + node_id="SaveSVGNode", + description="Save SVG files on disk.", + category="image/save", + inputs=[ + IO.SVG.Input("svg"), + IO.String.Input( + "filename_prefix", + default="svg/ComfyUI", + tooltip="The prefix for the file to save. This may include formatting information such as %date:yyyy-MM-dd% or %Empty Latent Image.width% to include values from nodes.", + ), + ], + hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo], + is_output_node=True, + ) - def save_svg(self, svg: SVG, filename_prefix="svg/ComfyUI", prompt=None, extra_pnginfo=None): - filename_prefix += self.prefix_append - full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix, self.output_dir) - results = list() + @classmethod + def execute(cls, svg: IO.SVG.Type, filename_prefix="svg/ComfyUI") -> IO.NodeOutput: + full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix, folder_paths.get_output_directory()) + results: list[UI.SavedResult] = [] # Prepare metadata JSON metadata_dict = {} - if prompt is not None: - metadata_dict["prompt"] = prompt - if extra_pnginfo is not None: - metadata_dict.update(extra_pnginfo) + if cls.hidden.prompt is not None: + metadata_dict["prompt"] = cls.hidden.prompt + if cls.hidden.extra_pnginfo is not None: + metadata_dict.update(cls.hidden.extra_pnginfo) # Convert metadata to JSON string metadata_json = json.dumps(metadata_dict, indent=2) if metadata_dict else None + for batch_number, svg_bytes in enumerate(svg.data): filename_with_batch_num = filename.replace("%batch_num%", str(batch_number)) file = f"{filename_with_batch_num}_{counter:05}_.svg" @@ -544,57 +484,64 @@ class SaveSVGNode: with open(os.path.join(full_output_folder, file), 'wb') as svg_file: svg_file.write(svg_content.encode('utf-8')) - results.append({ - "filename": file, - "subfolder": subfolder, - "type": self.type - }) + results.append(UI.SavedResult(filename=file, subfolder=subfolder, type=IO.FolderType.output)) counter += 1 - return { "ui": { "images": results } } + return IO.NodeOutput(ui={"images": results}) -class GetImageSize: + save_svg = execute # TODO: remove + + +class GetImageSize(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "image": (IO.IMAGE,), - }, - "hidden": { - "unique_id": "UNIQUE_ID", - } - } + def define_schema(cls): + return IO.Schema( + node_id="GetImageSize", + display_name="Get Image Size", + description="Returns width and height of the image, and passes it through unchanged.", + category="image", + inputs=[ + IO.Image.Input("image"), + ], + outputs=[ + IO.Int.Output(display_name="width"), + IO.Int.Output(display_name="height"), + IO.Int.Output(display_name="batch_size"), + ], + hidden=[IO.Hidden.unique_id], + ) - RETURN_TYPES = (IO.INT, IO.INT, IO.INT) - RETURN_NAMES = ("width", "height", "batch_size") - FUNCTION = "get_size" - - CATEGORY = "image" - DESCRIPTION = """Returns width and height of the image, and passes it through unchanged.""" - - def get_size(self, image, unique_id=None) -> tuple[int, int]: + @classmethod + def execute(cls, image) -> IO.NodeOutput: height = image.shape[1] width = image.shape[2] batch_size = image.shape[0] # Send progress text to display size on the node - if unique_id: - PromptServer.instance.send_progress_text(f"width: {width}, height: {height}\n batch size: {batch_size}", unique_id) + if cls.hidden.unique_id: + PromptServer.instance.send_progress_text(f"width: {width}, height: {height}\n batch size: {batch_size}", cls.hidden.unique_id) - return width, height, batch_size + return IO.NodeOutput(width, height, batch_size) + + get_size = execute # TODO: remove + + +class ImageRotate(IO.ComfyNode): -class ImageRotate: @classmethod - def INPUT_TYPES(s): - return {"required": { "image": (IO.IMAGE,), - "rotation": (["none", "90 degrees", "180 degrees", "270 degrees"],), - }} - RETURN_TYPES = (IO.IMAGE,) - FUNCTION = "rotate" + def define_schema(cls): + return IO.Schema( + node_id="ImageRotate", + category="image/transform", + inputs=[ + IO.Image.Input("image"), + IO.Combo.Input("rotation", options=["none", "90 degrees", "180 degrees", "270 degrees"]), + ], + outputs=[IO.Image.Output()], + ) - CATEGORY = "image/transform" - - def rotate(self, image, rotation): + @classmethod + def execute(cls, image, rotation) -> IO.NodeOutput: rotate_by = 0 if rotation.startswith("90"): rotate_by = 1 @@ -604,41 +551,57 @@ class ImageRotate: rotate_by = 3 image = torch.rot90(image, k=rotate_by, dims=[2, 1]) - return (image,) + return IO.NodeOutput(image) + + rotate = execute # TODO: remove + + +class ImageFlip(IO.ComfyNode): -class ImageFlip: @classmethod - def INPUT_TYPES(s): - return {"required": { "image": (IO.IMAGE,), - "flip_method": (["x-axis: vertically", "y-axis: horizontally"],), - }} - RETURN_TYPES = (IO.IMAGE,) - FUNCTION = "flip" + def define_schema(cls): + return IO.Schema( + node_id="ImageFlip", + category="image/transform", + inputs=[ + IO.Image.Input("image"), + IO.Combo.Input("flip_method", options=["x-axis: vertically", "y-axis: horizontally"]), + ], + outputs=[IO.Image.Output()], + ) - CATEGORY = "image/transform" - - def flip(self, image, flip_method): + @classmethod + def execute(cls, image, flip_method) -> IO.NodeOutput: if flip_method.startswith("x"): image = torch.flip(image, dims=[1]) elif flip_method.startswith("y"): image = torch.flip(image, dims=[2]) - return (image,) + return IO.NodeOutput(image) -class ImageScaleToMaxDimension: - upscale_methods = ["area", "lanczos", "bilinear", "nearest-exact", "bilinear", "bicubic"] + flip = execute # TODO: remove + + +class ImageScaleToMaxDimension(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": {"image": ("IMAGE",), - "upscale_method": (s.upscale_methods,), - "largest_size": ("INT", {"default": 512, "min": 0, "max": MAX_RESOLUTION, "step": 1})}} - RETURN_TYPES = ("IMAGE",) - FUNCTION = "upscale" + def define_schema(cls): + return IO.Schema( + node_id="ImageScaleToMaxDimension", + category="image/upscaling", + inputs=[ + IO.Image.Input("image"), + IO.Combo.Input( + "upscale_method", + options=["area", "lanczos", "bilinear", "nearest-exact", "bilinear", "bicubic"], + ), + IO.Int.Input("largest_size", default=512, min=0, max=MAX_RESOLUTION, step=1), + ], + outputs=[IO.Image.Output()], + ) - CATEGORY = "image/upscaling" - - def upscale(self, image, upscale_method, largest_size): + @classmethod + def execute(cls, image, upscale_method, largest_size) -> IO.NodeOutput: height = image.shape[1] width = image.shape[2] @@ -655,20 +618,30 @@ class ImageScaleToMaxDimension: samples = image.movedim(-1, 1) s = comfy.utils.common_upscale(samples, width, height, upscale_method, "disabled") s = s.movedim(1, -1) - return (s,) + return IO.NodeOutput(s) -NODE_CLASS_MAPPINGS = { - "ImageCrop": ImageCrop, - "RepeatImageBatch": RepeatImageBatch, - "ImageFromBatch": ImageFromBatch, - "ImageAddNoise": ImageAddNoise, - "SaveAnimatedWEBP": SaveAnimatedWEBP, - "SaveAnimatedPNG": SaveAnimatedPNG, - "SaveSVGNode": SaveSVGNode, - "ImageStitch": ImageStitch, - "ResizeAndPadImage": ResizeAndPadImage, - "GetImageSize": GetImageSize, - "ImageRotate": ImageRotate, - "ImageFlip": ImageFlip, - "ImageScaleToMaxDimension": ImageScaleToMaxDimension, -} + upscale = execute # TODO: remove + + +class ImagesExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + ImageCrop, + RepeatImageBatch, + ImageFromBatch, + ImageAddNoise, + SaveAnimatedWEBP, + SaveAnimatedPNG, + SaveSVGNode, + ImageStitch, + ResizeAndPadImage, + GetImageSize, + ImageRotate, + ImageFlip, + ImageScaleToMaxDimension, + ] + + +async def comfy_entrypoint() -> ImagesExtension: + return ImagesExtension() diff --git a/comfy_extras/nodes_kandinsky5.py b/comfy_extras/nodes_kandinsky5.py new file mode 100644 index 000000000..9cb234be1 --- /dev/null +++ b/comfy_extras/nodes_kandinsky5.py @@ -0,0 +1,136 @@ +import nodes +import node_helpers +import torch +import comfy.model_management +import comfy.utils + +from typing_extensions import override +from comfy_api.latest import ComfyExtension, io + + +class Kandinsky5ImageToVideo(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="Kandinsky5ImageToVideo", + category="conditioning/video_models", + inputs=[ + io.Conditioning.Input("positive"), + io.Conditioning.Input("negative"), + io.Vae.Input("vae"), + io.Int.Input("width", default=768, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("height", default=512, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("length", default=121, min=1, max=nodes.MAX_RESOLUTION, step=4), + io.Int.Input("batch_size", default=1, min=1, max=4096), + io.Image.Input("start_image", optional=True), + ], + outputs=[ + io.Conditioning.Output(display_name="positive"), + io.Conditioning.Output(display_name="negative"), + io.Latent.Output(display_name="latent", tooltip="Empty video latent"), + io.Latent.Output(display_name="cond_latent", tooltip="Clean encoded start images, used to replace the noisy start of the model output latents"), + ], + ) + + @classmethod + def execute(cls, positive, negative, vae, width, height, length, batch_size, start_image=None) -> io.NodeOutput: + latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device()) + cond_latent_out = {} + if start_image is not None: + start_image = comfy.utils.common_upscale(start_image[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1) + encoded = vae.encode(start_image[:, :, :, :3]) + cond_latent_out["samples"] = encoded + + mask = torch.ones((1, 1, latent.shape[2], latent.shape[-2], latent.shape[-1]), device=start_image.device, dtype=start_image.dtype) + mask[:, :, :((start_image.shape[0] - 1) // 4) + 1] = 0.0 + + positive = node_helpers.conditioning_set_values(positive, {"time_dim_replace": encoded, "concat_mask": mask}) + negative = node_helpers.conditioning_set_values(negative, {"time_dim_replace": encoded, "concat_mask": mask}) + + out_latent = {} + out_latent["samples"] = latent + return io.NodeOutput(positive, negative, out_latent, cond_latent_out) + + +def adaptive_mean_std_normalization(source, reference, clump_mean_low=0.3, clump_mean_high=0.35, clump_std_low=0.35, clump_std_high=0.5): + source_mean = source.mean(dim=(1, 3, 4), keepdim=True) # mean over C, H, W + source_std = source.std(dim=(1, 3, 4), keepdim=True) # std over C, H, W + + reference_mean = torch.clamp(reference.mean(), source_mean - clump_mean_low, source_mean + clump_mean_high) + reference_std = torch.clamp(reference.std(), source_std - clump_std_low, source_std + clump_std_high) + + # normalization + normalized = (source - source_mean) / (source_std + 1e-8) + normalized = normalized * reference_std + reference_mean + + return normalized + + +class NormalizeVideoLatentStart(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="NormalizeVideoLatentStart", + category="conditioning/video_models", + description="Normalizes the initial frames of a video latent to match the mean and standard deviation of subsequent reference frames. Helps reduce differences between the starting frames and the rest of the video.", + inputs=[ + io.Latent.Input("latent"), + io.Int.Input("start_frame_count", default=4, min=1, max=nodes.MAX_RESOLUTION, step=1, tooltip="Number of latent frames to normalize, counted from the start"), + io.Int.Input("reference_frame_count", default=5, min=1, max=nodes.MAX_RESOLUTION, step=1, tooltip="Number of latent frames after the start frames to use as reference"), + ], + outputs=[ + io.Latent.Output(display_name="latent"), + ], + ) + + @classmethod + def execute(cls, latent, start_frame_count, reference_frame_count) -> io.NodeOutput: + if latent["samples"].shape[2] <= 1: + return io.NodeOutput(latent) + s = latent.copy() + samples = latent["samples"].clone() + + first_frames = samples[:, :, :start_frame_count] + reference_frames_data = samples[:, :, start_frame_count:start_frame_count+min(reference_frame_count, samples.shape[2]-1)] + normalized_first_frames = adaptive_mean_std_normalization(first_frames, reference_frames_data) + + samples[:, :, :start_frame_count] = normalized_first_frames + s["samples"] = samples + return io.NodeOutput(s) + + +class CLIPTextEncodeKandinsky5(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="CLIPTextEncodeKandinsky5", + category="advanced/conditioning/kandinsky5", + inputs=[ + io.Clip.Input("clip"), + io.String.Input("clip_l", multiline=True, dynamic_prompts=True), + io.String.Input("qwen25_7b", multiline=True, dynamic_prompts=True), + ], + outputs=[ + io.Conditioning.Output(), + ], + ) + + @classmethod + def execute(cls, clip, clip_l, qwen25_7b) -> io.NodeOutput: + tokens = clip.tokenize(clip_l) + tokens["qwen25_7b"] = clip.tokenize(qwen25_7b)["qwen25_7b"] + + return io.NodeOutput(clip.encode_from_tokens_scheduled(tokens)) + + +class Kandinsky5Extension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + Kandinsky5ImageToVideo, + NormalizeVideoLatentStart, + CLIPTextEncodeKandinsky5, + ] + +async def comfy_entrypoint() -> Kandinsky5Extension: + return Kandinsky5Extension() diff --git a/comfy_extras/nodes_latent.py b/comfy_extras/nodes_latent.py index d2df07ff9..9ba1c4ba8 100644 --- a/comfy_extras/nodes_latent.py +++ b/comfy_extras/nodes_latent.py @@ -4,7 +4,8 @@ import torch import nodes from typing_extensions import override from comfy_api.latest import ComfyExtension, io - +import logging +import math def reshape_latent_to(target_shape, latent, repeat_batch=True): if latent.shape[1:] != target_shape[1:]: @@ -207,12 +208,54 @@ class LatentCut(io.ComfyNode): samples_out["samples"] = torch.narrow(s1, dim, index, amount) return io.NodeOutput(samples_out) +class LatentCutToBatch(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LatentCutToBatch", + category="latent/advanced", + inputs=[ + io.Latent.Input("samples"), + io.Combo.Input("dim", options=["t", "x", "y"]), + io.Int.Input("slice_size", default=1, min=1, max=nodes.MAX_RESOLUTION, step=1), + ], + outputs=[ + io.Latent.Output(), + ], + ) + + @classmethod + def execute(cls, samples, dim, slice_size) -> io.NodeOutput: + samples_out = samples.copy() + + s1 = samples["samples"] + + if "x" in dim: + dim = s1.ndim - 1 + elif "y" in dim: + dim = s1.ndim - 2 + elif "t" in dim: + dim = s1.ndim - 3 + + if dim < 2: + return io.NodeOutput(samples) + + s = s1.movedim(dim, 1) + if s.shape[1] < slice_size: + slice_size = s.shape[1] + elif s.shape[1] % slice_size != 0: + s = s[:, :math.floor(s.shape[1] / slice_size) * slice_size] + new_shape = [-1, slice_size] + list(s.shape[2:]) + samples_out["samples"] = s.reshape(new_shape).movedim(1, dim) + return io.NodeOutput(samples_out) + class LatentBatch(io.ComfyNode): @classmethod def define_schema(cls): return io.Schema( node_id="LatentBatch", category="latent/batch", + is_deprecated=True, inputs=[ io.Latent.Input("samples1"), io.Latent.Input("samples2"), @@ -388,6 +431,42 @@ class LatentOperationSharpen(io.ComfyNode): return luminance * sharpened return io.NodeOutput(sharpen) +class ReplaceVideoLatentFrames(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="ReplaceVideoLatentFrames", + category="latent/batch", + inputs=[ + io.Latent.Input("destination", tooltip="The destination latent where frames will be replaced."), + io.Latent.Input("source", optional=True, tooltip="The source latent providing frames to insert into the destination latent. If not provided, the destination latent is returned unchanged."), + io.Int.Input("index", default=0, min=-nodes.MAX_RESOLUTION, max=nodes.MAX_RESOLUTION, step=1, tooltip="The starting latent frame index in the destination latent where the source latent frames will be placed. Negative values count from the end."), + ], + outputs=[ + io.Latent.Output(), + ], + ) + + @classmethod + def execute(cls, destination, index, source=None) -> io.NodeOutput: + if source is None: + return io.NodeOutput(destination) + dest_frames = destination["samples"].shape[2] + source_frames = source["samples"].shape[2] + if index < 0: + index = dest_frames + index + if index > dest_frames: + logging.warning(f"ReplaceVideoLatentFrames: Index {index} is out of bounds for destination latent frames {dest_frames}.") + return io.NodeOutput(destination) + if index + source_frames > dest_frames: + logging.warning(f"ReplaceVideoLatentFrames: Source latent frames {source_frames} do not fit within destination latent frames {dest_frames} at the specified index {index}.") + return io.NodeOutput(destination) + s = source.copy() + s_source = source["samples"] + s_destination = destination["samples"].clone() + s_destination[:, :, index:index + s_source.shape[2]] = s_source + s["samples"] = s_destination + return io.NodeOutput(s) class LatentExtension(ComfyExtension): @override @@ -399,12 +478,14 @@ class LatentExtension(ComfyExtension): LatentInterpolate, LatentConcat, LatentCut, + LatentCutToBatch, LatentBatch, LatentBatchSeedBehavior, LatentApplyOperation, LatentApplyOperationCFG, LatentOperationTonemapReinhard, LatentOperationSharpen, + ReplaceVideoLatentFrames ] diff --git a/comfy_extras/nodes_load_3d.py b/comfy_extras/nodes_load_3d.py index 899608149..545588ef8 100644 --- a/comfy_extras/nodes_load_3d.py +++ b/comfy_extras/nodes_load_3d.py @@ -2,8 +2,8 @@ import nodes import folder_paths import os -from comfy.comfy_types import IO -from comfy_api.input_impl import VideoFromFile +from typing_extensions import override +from comfy_api.latest import IO, ComfyExtension, InputImpl, UI from pathlib import Path @@ -11,9 +11,9 @@ from pathlib import Path def normalize_path(path): return path.replace('\\', '/') -class Load3D(): +class Load3D(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): + def define_schema(cls): input_dir = os.path.join(folder_paths.get_input_directory(), "3d") os.makedirs(input_dir, exist_ok=True) @@ -26,157 +26,84 @@ class Load3D(): for file_path in input_path.rglob("*") if file_path.suffix.lower() in {'.gltf', '.glb', '.obj', '.fbx', '.stl'} ] + return IO.Schema( + node_id="Load3D", + display_name="Load 3D & Animation", + category="3d", + is_experimental=True, + inputs=[ + IO.Combo.Input("model_file", options=sorted(files), upload=IO.UploadType.model), + IO.Load3D.Input("image"), + IO.Int.Input("width", default=1024, min=1, max=4096, step=1), + IO.Int.Input("height", default=1024, min=1, max=4096, step=1), + ], + outputs=[ + IO.Image.Output(display_name="image"), + IO.Mask.Output(display_name="mask"), + IO.String.Output(display_name="mesh_path"), + IO.Image.Output(display_name="normal"), + IO.Load3DCamera.Output(display_name="camera_info"), + IO.Video.Output(display_name="recording_video"), + ], + ) - return {"required": { - "model_file": (sorted(files), {"file_upload": True}), - "image": ("LOAD_3D", {}), - "width": ("INT", {"default": 1024, "min": 1, "max": 4096, "step": 1}), - "height": ("INT", {"default": 1024, "min": 1, "max": 4096, "step": 1}), - }} - - RETURN_TYPES = ("IMAGE", "MASK", "STRING", "IMAGE", "IMAGE", "LOAD3D_CAMERA", IO.VIDEO) - RETURN_NAMES = ("image", "mask", "mesh_path", "normal", "lineart", "camera_info", "recording_video") - - FUNCTION = "process" - EXPERIMENTAL = True - - CATEGORY = "3d" - - def process(self, model_file, image, **kwargs): + @classmethod + def execute(cls, model_file, image, **kwargs) -> IO.NodeOutput: image_path = folder_paths.get_annotated_filepath(image['image']) mask_path = folder_paths.get_annotated_filepath(image['mask']) normal_path = folder_paths.get_annotated_filepath(image['normal']) - lineart_path = folder_paths.get_annotated_filepath(image['lineart']) load_image_node = nodes.LoadImage() output_image, ignore_mask = load_image_node.load_image(image=image_path) ignore_image, output_mask = load_image_node.load_image(image=mask_path) normal_image, ignore_mask2 = load_image_node.load_image(image=normal_path) - lineart_image, ignore_mask3 = load_image_node.load_image(image=lineart_path) video = None if image['recording'] != "": recording_video_path = folder_paths.get_annotated_filepath(image['recording']) - video = VideoFromFile(recording_video_path) + video = InputImpl.VideoFromFile(recording_video_path) - return output_image, output_mask, model_file, normal_image, lineart_image, image['camera_info'], video + return IO.NodeOutput(output_image, output_mask, model_file, normal_image, image['camera_info'], video) -class Load3DAnimation(): + process = execute # TODO: remove + + +class Preview3D(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - input_dir = os.path.join(folder_paths.get_input_directory(), "3d") + def define_schema(cls): + return IO.Schema( + node_id="Preview3D", + display_name="Preview 3D & Animation", + category="3d", + is_experimental=True, + is_output_node=True, + inputs=[ + IO.String.Input("model_file", default="", multiline=False), + IO.Load3DCamera.Input("camera_info", optional=True), + IO.Image.Input("bg_image", optional=True), + ], + outputs=[], + ) - os.makedirs(input_dir, exist_ok=True) + @classmethod + def execute(cls, model_file, **kwargs) -> IO.NodeOutput: + camera_info = kwargs.get("camera_info", None) + bg_image = kwargs.get("bg_image", None) + return IO.NodeOutput(ui=UI.PreviewUI3D(model_file, camera_info, bg_image=bg_image)) - input_path = Path(input_dir) - base_path = Path(folder_paths.get_input_directory()) + process = execute # TODO: remove - files = [ - normalize_path(str(file_path.relative_to(base_path))) - for file_path in input_path.rglob("*") - if file_path.suffix.lower() in {'.gltf', '.glb', '.fbx'} + +class Load3DExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + Load3D, + Preview3D, ] - return {"required": { - "model_file": (sorted(files), {"file_upload": True}), - "image": ("LOAD_3D_ANIMATION", {}), - "width": ("INT", {"default": 1024, "min": 1, "max": 4096, "step": 1}), - "height": ("INT", {"default": 1024, "min": 1, "max": 4096, "step": 1}), - }} - RETURN_TYPES = ("IMAGE", "MASK", "STRING", "IMAGE", "LOAD3D_CAMERA", IO.VIDEO) - RETURN_NAMES = ("image", "mask", "mesh_path", "normal", "camera_info", "recording_video") - - FUNCTION = "process" - EXPERIMENTAL = True - - CATEGORY = "3d" - - def process(self, model_file, image, **kwargs): - image_path = folder_paths.get_annotated_filepath(image['image']) - mask_path = folder_paths.get_annotated_filepath(image['mask']) - normal_path = folder_paths.get_annotated_filepath(image['normal']) - - load_image_node = nodes.LoadImage() - output_image, ignore_mask = load_image_node.load_image(image=image_path) - ignore_image, output_mask = load_image_node.load_image(image=mask_path) - normal_image, ignore_mask2 = load_image_node.load_image(image=normal_path) - - video = None - - if image['recording'] != "": - recording_video_path = folder_paths.get_annotated_filepath(image['recording']) - - video = VideoFromFile(recording_video_path) - - return output_image, output_mask, model_file, normal_image, image['camera_info'], video - -class Preview3D(): - @classmethod - def INPUT_TYPES(s): - return {"required": { - "model_file": ("STRING", {"default": "", "multiline": False}), - }, - "optional": { - "camera_info": ("LOAD3D_CAMERA", {}) - }} - - OUTPUT_NODE = True - RETURN_TYPES = () - - CATEGORY = "3d" - - FUNCTION = "process" - EXPERIMENTAL = True - - def process(self, model_file, **kwargs): - camera_info = kwargs.get("camera_info", None) - - return { - "ui": { - "result": [model_file, camera_info] - } - } - -class Preview3DAnimation(): - @classmethod - def INPUT_TYPES(s): - return {"required": { - "model_file": ("STRING", {"default": "", "multiline": False}), - }, - "optional": { - "camera_info": ("LOAD3D_CAMERA", {}) - }} - - OUTPUT_NODE = True - RETURN_TYPES = () - - CATEGORY = "3d" - - FUNCTION = "process" - EXPERIMENTAL = True - - def process(self, model_file, **kwargs): - camera_info = kwargs.get("camera_info", None) - - return { - "ui": { - "result": [model_file, camera_info] - } - } - -NODE_CLASS_MAPPINGS = { - "Load3D": Load3D, - "Load3DAnimation": Load3DAnimation, - "Preview3D": Preview3D, - "Preview3DAnimation": Preview3DAnimation -} - -NODE_DISPLAY_NAME_MAPPINGS = { - "Load3D": "Load 3D", - "Load3DAnimation": "Load 3D - Animation", - "Preview3D": "Preview 3D", - "Preview3DAnimation": "Preview 3D - Animation" -} +async def comfy_entrypoint() -> Load3DExtension: + return Load3DExtension() diff --git a/comfy_extras/nodes_logic.py b/comfy_extras/nodes_logic.py new file mode 100644 index 000000000..eb888316a --- /dev/null +++ b/comfy_extras/nodes_logic.py @@ -0,0 +1,268 @@ +from __future__ import annotations +from typing import TypedDict +from typing_extensions import override +from comfy_api.latest import ComfyExtension, io +from comfy_api.latest import _io + +# sentinel for missing inputs +MISSING = object() + + +class SwitchNode(io.ComfyNode): + @classmethod + def define_schema(cls): + template = io.MatchType.Template("switch") + return io.Schema( + node_id="ComfySwitchNode", + display_name="Switch", + category="logic", + is_experimental=True, + inputs=[ + io.Boolean.Input("switch"), + io.MatchType.Input("on_false", template=template, lazy=True), + io.MatchType.Input("on_true", template=template, lazy=True), + ], + outputs=[ + io.MatchType.Output(template=template, display_name="output"), + ], + ) + + @classmethod + def check_lazy_status(cls, switch, on_false=None, on_true=None): + if switch and on_true is None: + return ["on_true"] + if not switch and on_false is None: + return ["on_false"] + + @classmethod + def execute(cls, switch, on_true, on_false) -> io.NodeOutput: + return io.NodeOutput(on_true if switch else on_false) + + +class SoftSwitchNode(io.ComfyNode): + @classmethod + def define_schema(cls): + template = io.MatchType.Template("switch") + return io.Schema( + node_id="ComfySoftSwitchNode", + display_name="Soft Switch", + category="logic", + is_experimental=True, + inputs=[ + io.Boolean.Input("switch"), + io.MatchType.Input("on_false", template=template, lazy=True, optional=True), + io.MatchType.Input("on_true", template=template, lazy=True, optional=True), + ], + outputs=[ + io.MatchType.Output(template=template, display_name="output"), + ], + ) + + @classmethod + def check_lazy_status(cls, switch, on_false=MISSING, on_true=MISSING): + # We use MISSING instead of None, as None is passed for connected-but-unevaluated inputs. + # This trick allows us to ignore the value of the switch and still be able to run execute(). + + # One of the inputs may be missing, in which case we need to evaluate the other input + if on_false is MISSING: + return ["on_true"] + if on_true is MISSING: + return ["on_false"] + # Normal lazy switch operation + if switch and on_true is None: + return ["on_true"] + if not switch and on_false is None: + return ["on_false"] + + @classmethod + def validate_inputs(cls, switch, on_false=MISSING, on_true=MISSING): + # This check happens before check_lazy_status(), so we can eliminate the case where + # both inputs are missing. + if on_false is MISSING and on_true is MISSING: + return "At least one of on_false or on_true must be connected to Switch node" + return True + + @classmethod + def execute(cls, switch, on_true=MISSING, on_false=MISSING) -> io.NodeOutput: + if on_true is MISSING: + return io.NodeOutput(on_false) + if on_false is MISSING: + return io.NodeOutput(on_true) + return io.NodeOutput(on_true if switch else on_false) + + +class CustomComboNode(io.ComfyNode): + """ + Frontend node that allows user to write their own options for a combo. + This is here to make sure the node has a backend-representation to avoid some annoyances. + """ + @classmethod + def define_schema(cls): + return io.Schema( + node_id="CustomCombo", + display_name="Custom Combo", + category="utils", + is_experimental=True, + inputs=[io.Combo.Input("choice", options=[])], + outputs=[io.String.Output()] + ) + + @classmethod + def validate_inputs(cls, choice: io.Combo.Type) -> bool: + # NOTE: DO NOT DO THIS unless you want to skip validation entirely on the node's inputs. + # I am doing that here because the widgets (besides the combo dropdown) on this node are fully frontend defined. + # I need to skip checking that the chosen combo option is in the options list, since those are defined by the user. + return True + + @classmethod + def execute(cls, choice: io.Combo.Type) -> io.NodeOutput: + return io.NodeOutput(choice) + + +class DCTestNode(io.ComfyNode): + class DCValues(TypedDict): + combo: str + string: str + integer: int + image: io.Image.Type + subcombo: dict[str] + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="DCTestNode", + display_name="DCTest", + category="logic", + is_output_node=True, + inputs=[io.DynamicCombo.Input("combo", options=[ + io.DynamicCombo.Option("option1", [io.String.Input("string")]), + io.DynamicCombo.Option("option2", [io.Int.Input("integer")]), + io.DynamicCombo.Option("option3", [io.Image.Input("image")]), + io.DynamicCombo.Option("option4", [ + io.DynamicCombo.Input("subcombo", options=[ + io.DynamicCombo.Option("opt1", [io.Float.Input("float_x"), io.Float.Input("float_y")]), + io.DynamicCombo.Option("opt2", [io.Mask.Input("mask1", optional=True)]), + ]) + ])] + )], + outputs=[io.AnyType.Output()], + ) + + @classmethod + def execute(cls, combo: DCValues) -> io.NodeOutput: + combo_val = combo["combo"] + if combo_val == "option1": + return io.NodeOutput(combo["string"]) + elif combo_val == "option2": + return io.NodeOutput(combo["integer"]) + elif combo_val == "option3": + return io.NodeOutput(combo["image"]) + elif combo_val == "option4": + return io.NodeOutput(f"{combo['subcombo']}") + else: + raise ValueError(f"Invalid combo: {combo_val}") + + +class AutogrowNamesTestNode(io.ComfyNode): + @classmethod + def define_schema(cls): + template = _io.Autogrow.TemplateNames(input=io.Float.Input("float"), names=["a", "b", "c"]) + return io.Schema( + node_id="AutogrowNamesTestNode", + display_name="AutogrowNamesTest", + category="logic", + inputs=[ + _io.Autogrow.Input("autogrow", template=template) + ], + outputs=[io.String.Output()], + ) + + @classmethod + def execute(cls, autogrow: _io.Autogrow.Type) -> io.NodeOutput: + vals = list(autogrow.values()) + combined = ",".join([str(x) for x in vals]) + return io.NodeOutput(combined) + +class AutogrowPrefixTestNode(io.ComfyNode): + @classmethod + def define_schema(cls): + template = _io.Autogrow.TemplatePrefix(input=io.Float.Input("float"), prefix="float", min=1, max=10) + return io.Schema( + node_id="AutogrowPrefixTestNode", + display_name="AutogrowPrefixTest", + category="logic", + inputs=[ + _io.Autogrow.Input("autogrow", template=template) + ], + outputs=[io.String.Output()], + ) + + @classmethod + def execute(cls, autogrow: _io.Autogrow.Type) -> io.NodeOutput: + vals = list(autogrow.values()) + combined = ",".join([str(x) for x in vals]) + return io.NodeOutput(combined) + +class ComboOutputTestNode(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="ComboOptionTestNode", + display_name="ComboOptionTest", + category="logic", + inputs=[io.Combo.Input("combo", options=["option1", "option2", "option3"]), + io.Combo.Input("combo2", options=["option4", "option5", "option6"])], + outputs=[io.Combo.Output(), io.Combo.Output()], + ) + + @classmethod + def execute(cls, combo: io.Combo.Type, combo2: io.Combo.Type) -> io.NodeOutput: + return io.NodeOutput(combo, combo2) + +class ConvertStringToComboNode(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="ConvertStringToComboNode", + display_name="Convert String to Combo", + category="logic", + inputs=[io.String.Input("string")], + outputs=[io.Combo.Output()], + ) + + @classmethod + def execute(cls, string: str) -> io.NodeOutput: + return io.NodeOutput(string) + +class InvertBooleanNode(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="InvertBooleanNode", + display_name="Invert Boolean", + category="logic", + inputs=[io.Boolean.Input("boolean")], + outputs=[io.Boolean.Output()], + ) + + @classmethod + def execute(cls, boolean: bool) -> io.NodeOutput: + return io.NodeOutput(not boolean) + +class LogicExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + SwitchNode, + CustomComboNode, + # SoftSwitchNode, + # ConvertStringToComboNode, + # DCTestNode, + # AutogrowNamesTestNode, + # AutogrowPrefixTestNode, + # ComboOutputTestNode, + # InvertBooleanNode, + ] + +async def comfy_entrypoint() -> LogicExtension: + return LogicExtension() diff --git a/comfy_extras/nodes_lt.py b/comfy_extras/nodes_lt.py index 50da5f4eb..b91a22309 100644 --- a/comfy_extras/nodes_lt.py +++ b/comfy_extras/nodes_lt.py @@ -81,6 +81,59 @@ class LTXVImgToVideo(io.ComfyNode): generate = execute # TODO: remove +class LTXVImgToVideoInplace(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LTXVImgToVideoInplace", + category="conditioning/video_models", + inputs=[ + io.Vae.Input("vae"), + io.Image.Input("image"), + io.Latent.Input("latent"), + io.Float.Input("strength", default=1.0, min=0.0, max=1.0), + io.Boolean.Input("bypass", default=False, tooltip="Bypass the conditioning.") + ], + outputs=[ + io.Latent.Output(display_name="latent"), + ], + ) + + @classmethod + def execute(cls, vae, image, latent, strength, bypass=False) -> io.NodeOutput: + if bypass: + return (latent,) + + samples = latent["samples"] + _, height_scale_factor, width_scale_factor = ( + vae.downscale_index_formula + ) + + batch, _, latent_frames, latent_height, latent_width = samples.shape + width = latent_width * width_scale_factor + height = latent_height * height_scale_factor + + if image.shape[1] != height or image.shape[2] != width: + pixels = comfy.utils.common_upscale(image.movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1) + else: + pixels = image + encode_pixels = pixels[:, :, :, :3] + t = vae.encode(encode_pixels) + + samples[:, :, :t.shape[2]] = t + + conditioning_latent_frames_mask = torch.ones( + (batch, 1, latent_frames, 1, 1), + dtype=torch.float32, + device=samples.device, + ) + conditioning_latent_frames_mask[:, :, :t.shape[2]] = 1.0 - strength + + return io.NodeOutput({"samples": samples, "noise_mask": conditioning_latent_frames_mask}) + + generate = execute # TODO: remove + + def conditioning_get_any_value(conditioning, key, default=None): for t in conditioning: if key in t[1]: @@ -106,12 +159,12 @@ def get_keyframe_idxs(cond): keyframe_idxs = conditioning_get_any_value(cond, "keyframe_idxs", None) if keyframe_idxs is None: return None, 0 - num_keyframes = torch.unique(keyframe_idxs[:, 0]).shape[0] + # keyframe_idxs contains start/end positions (last dimension), checking for unqiue values only for start + num_keyframes = torch.unique(keyframe_idxs[:, 0, :, 0]).shape[0] return keyframe_idxs, num_keyframes class LTXVAddGuide(io.ComfyNode): - NUM_PREFIX_FRAMES = 2 - PATCHIFIER = SymmetricPatchifier(1) + PATCHIFIER = SymmetricPatchifier(1, start_end=True) @classmethod def define_schema(cls): @@ -182,26 +235,35 @@ class LTXVAddGuide(io.ComfyNode): return node_helpers.conditioning_set_values(cond, {"keyframe_idxs": keyframe_idxs}) @classmethod - def append_keyframe(cls, positive, negative, frame_idx, latent_image, noise_mask, guiding_latent, strength, scale_factors): - _, latent_idx = cls.get_latent_index( - cond=positive, - latent_length=latent_image.shape[2], - guide_length=guiding_latent.shape[2], - frame_idx=frame_idx, - scale_factors=scale_factors, - ) - noise_mask[:, :, latent_idx:latent_idx + guiding_latent.shape[2]] = 1.0 + def append_keyframe(cls, positive, negative, frame_idx, latent_image, noise_mask, guiding_latent, strength, scale_factors, guide_mask=None, in_channels=128): + if latent_image.shape[1] != in_channels or guiding_latent.shape[1] != in_channels: + raise ValueError("Adding guide to a combined AV latent is not supported.") positive = cls.add_keyframe_index(positive, frame_idx, guiding_latent, scale_factors) negative = cls.add_keyframe_index(negative, frame_idx, guiding_latent, scale_factors) - mask = torch.full( - (noise_mask.shape[0], 1, guiding_latent.shape[2], noise_mask.shape[3], noise_mask.shape[4]), - 1.0 - strength, - dtype=noise_mask.dtype, - device=noise_mask.device, - ) + if guide_mask is not None: + target_h = max(noise_mask.shape[3], guide_mask.shape[3]) + target_w = max(noise_mask.shape[4], guide_mask.shape[4]) + if noise_mask.shape[3] == 1 or noise_mask.shape[4] == 1: + noise_mask = noise_mask.expand(-1, -1, -1, target_h, target_w) + + if guide_mask.shape[3] == 1 or guide_mask.shape[4] == 1: + guide_mask = guide_mask.expand(-1, -1, -1, target_h, target_w) + mask = guide_mask - strength + else: + mask = torch.full( + (noise_mask.shape[0], 1, guiding_latent.shape[2], noise_mask.shape[3], noise_mask.shape[4]), + 1.0 - strength, + dtype=noise_mask.dtype, + device=noise_mask.device, + ) + # This solves audio video combined latent case where latent_image has audio latent concatenated + # in channel dimension with video latent. The solution is to pad guiding latent accordingly. + if latent_image.shape[1] > guiding_latent.shape[1]: + pad_len = latent_image.shape[1] - guiding_latent.shape[1] + guiding_latent = torch.nn.functional.pad(guiding_latent, pad=(0, 0, 0, 0, 0, 0, 0, pad_len), value=0) latent_image = torch.cat([latent_image, guiding_latent], dim=2) noise_mask = torch.cat([noise_mask, mask], dim=2) return positive, negative, latent_image, noise_mask @@ -238,33 +300,17 @@ class LTXVAddGuide(io.ComfyNode): frame_idx, latent_idx = cls.get_latent_index(positive, latent_length, len(image), frame_idx, scale_factors) assert latent_idx + t.shape[2] <= latent_length, "Conditioning frames exceed the length of the latent sequence." - num_prefix_frames = min(cls.NUM_PREFIX_FRAMES, t.shape[2]) - positive, negative, latent_image, noise_mask = cls.append_keyframe( positive, negative, frame_idx, latent_image, noise_mask, - t[:, :, :num_prefix_frames], + t, strength, scale_factors, ) - latent_idx += num_prefix_frames - - t = t[:, :, num_prefix_frames:] - if t.shape[2] == 0: - return io.NodeOutput(positive, negative, {"samples": latent_image, "noise_mask": noise_mask}) - - latent_image, noise_mask = cls.replace_latent_frames( - latent_image, - noise_mask, - t, - latent_idx, - strength, - ) - return io.NodeOutput(positive, negative, {"samples": latent_image, "noise_mask": noise_mask}) generate = execute # TODO: remove @@ -507,18 +553,90 @@ class LTXVPreprocess(io.ComfyNode): preprocess = execute # TODO: remove + +import comfy.nested_tensor +class LTXVConcatAVLatent(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LTXVConcatAVLatent", + category="latent/video/ltxv", + inputs=[ + io.Latent.Input("video_latent"), + io.Latent.Input("audio_latent"), + ], + outputs=[ + io.Latent.Output(display_name="latent"), + ], + ) + + @classmethod + def execute(cls, video_latent, audio_latent) -> io.NodeOutput: + output = {} + output.update(video_latent) + output.update(audio_latent) + video_noise_mask = video_latent.get("noise_mask", None) + audio_noise_mask = audio_latent.get("noise_mask", None) + + if video_noise_mask is not None or audio_noise_mask is not None: + if video_noise_mask is None: + video_noise_mask = torch.ones_like(video_latent["samples"]) + if audio_noise_mask is None: + audio_noise_mask = torch.ones_like(audio_latent["samples"]) + output["noise_mask"] = comfy.nested_tensor.NestedTensor((video_noise_mask, audio_noise_mask)) + + output["samples"] = comfy.nested_tensor.NestedTensor((video_latent["samples"], audio_latent["samples"])) + + return io.NodeOutput(output) + + +class LTXVSeparateAVLatent(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LTXVSeparateAVLatent", + category="latent/video/ltxv", + description="LTXV Separate AV Latent", + inputs=[ + io.Latent.Input("av_latent"), + ], + outputs=[ + io.Latent.Output(display_name="video_latent"), + io.Latent.Output(display_name="audio_latent"), + ], + ) + + @classmethod + def execute(cls, av_latent) -> io.NodeOutput: + latents = av_latent["samples"].unbind() + video_latent = av_latent.copy() + video_latent["samples"] = latents[0] + audio_latent = av_latent.copy() + audio_latent["samples"] = latents[1] + if "noise_mask" in av_latent: + masks = av_latent["noise_mask"] + if masks is not None: + masks = masks.unbind() + video_latent["noise_mask"] = masks[0] + audio_latent["noise_mask"] = masks[1] + return io.NodeOutput(video_latent, audio_latent) + + class LtxvExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[io.ComfyNode]]: return [ EmptyLTXVLatentVideo, LTXVImgToVideo, + LTXVImgToVideoInplace, ModelSamplingLTXV, LTXVConditioning, LTXVScheduler, LTXVAddGuide, LTXVPreprocess, LTXVCropGuides, + LTXVConcatAVLatent, + LTXVSeparateAVLatent, ] diff --git a/comfy_extras/nodes_lt_audio.py b/comfy_extras/nodes_lt_audio.py new file mode 100644 index 000000000..26b0160d2 --- /dev/null +++ b/comfy_extras/nodes_lt_audio.py @@ -0,0 +1,216 @@ +import folder_paths +import comfy.utils +import comfy.model_management +import torch + +from comfy.ldm.lightricks.vae.audio_vae import AudioVAE +from comfy_api.latest import ComfyExtension, io + + +class LTXVAudioVAELoader(io.ComfyNode): + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="LTXVAudioVAELoader", + display_name="LTXV Audio VAE Loader", + category="audio", + inputs=[ + io.Combo.Input( + "ckpt_name", + options=folder_paths.get_filename_list("checkpoints"), + tooltip="Audio VAE checkpoint to load.", + ) + ], + outputs=[io.Vae.Output(display_name="Audio VAE")], + ) + + @classmethod + def execute(cls, ckpt_name: str) -> io.NodeOutput: + ckpt_path = folder_paths.get_full_path_or_raise("checkpoints", ckpt_name) + sd, metadata = comfy.utils.load_torch_file(ckpt_path, return_metadata=True) + return io.NodeOutput(AudioVAE(sd, metadata)) + + +class LTXVAudioVAEEncode(io.ComfyNode): + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="LTXVAudioVAEEncode", + display_name="LTXV Audio VAE Encode", + category="audio", + inputs=[ + io.Audio.Input("audio", tooltip="The audio to be encoded."), + io.Vae.Input( + id="audio_vae", + display_name="Audio VAE", + tooltip="The Audio VAE model to use for encoding.", + ), + ], + outputs=[io.Latent.Output(display_name="Audio Latent")], + ) + + @classmethod + def execute(cls, audio, audio_vae: AudioVAE) -> io.NodeOutput: + audio_latents = audio_vae.encode(audio) + return io.NodeOutput( + { + "samples": audio_latents, + "sample_rate": int(audio_vae.sample_rate), + "type": "audio", + } + ) + + +class LTXVAudioVAEDecode(io.ComfyNode): + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="LTXVAudioVAEDecode", + display_name="LTXV Audio VAE Decode", + category="audio", + inputs=[ + io.Latent.Input("samples", tooltip="The latent to be decoded."), + io.Vae.Input( + id="audio_vae", + display_name="Audio VAE", + tooltip="The Audio VAE model used for decoding the latent.", + ), + ], + outputs=[io.Audio.Output(display_name="Audio")], + ) + + @classmethod + def execute(cls, samples, audio_vae: AudioVAE) -> io.NodeOutput: + audio_latent = samples["samples"] + if audio_latent.is_nested: + audio_latent = audio_latent.unbind()[-1] + audio = audio_vae.decode(audio_latent).to(audio_latent.device) + output_audio_sample_rate = audio_vae.output_sample_rate + return io.NodeOutput( + { + "waveform": audio, + "sample_rate": int(output_audio_sample_rate), + } + ) + + +class LTXVEmptyLatentAudio(io.ComfyNode): + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="LTXVEmptyLatentAudio", + display_name="LTXV Empty Latent Audio", + category="latent/audio", + inputs=[ + io.Int.Input( + "frames_number", + default=97, + min=1, + max=1000, + step=1, + display_mode=io.NumberDisplay.number, + tooltip="Number of frames.", + ), + io.Int.Input( + "frame_rate", + default=25, + min=1, + max=1000, + step=1, + display_mode=io.NumberDisplay.number, + tooltip="Number of frames per second.", + ), + io.Int.Input( + "batch_size", + default=1, + min=1, + max=4096, + display_mode=io.NumberDisplay.number, + tooltip="The number of latent audio samples in the batch.", + ), + io.Vae.Input( + id="audio_vae", + display_name="Audio VAE", + tooltip="The Audio VAE model to get configuration from.", + ), + ], + outputs=[io.Latent.Output(display_name="Latent")], + ) + + @classmethod + def execute( + cls, + frames_number: int, + frame_rate: int, + batch_size: int, + audio_vae: AudioVAE, + ) -> io.NodeOutput: + """Generate empty audio latents matching the reference pipeline structure.""" + + assert audio_vae is not None, "Audio VAE model is required" + + z_channels = audio_vae.latent_channels + audio_freq = audio_vae.latent_frequency_bins + sampling_rate = int(audio_vae.sample_rate) + + num_audio_latents = audio_vae.num_of_latents_from_frames(frames_number, frame_rate) + + audio_latents = torch.zeros( + (batch_size, z_channels, num_audio_latents, audio_freq), + device=comfy.model_management.intermediate_device(), + ) + + return io.NodeOutput( + { + "samples": audio_latents, + "sample_rate": sampling_rate, + "type": "audio", + } + ) + + +class LTXAVTextEncoderLoader(io.ComfyNode): + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="LTXAVTextEncoderLoader", + display_name="LTXV Audio Text Encoder Loader", + category="advanced/loaders", + description="[Recipes]\n\nltxav: gemma 3 12B", + inputs=[ + io.Combo.Input( + "text_encoder", + options=folder_paths.get_filename_list("text_encoders"), + ), + io.Combo.Input( + "ckpt_name", + options=folder_paths.get_filename_list("checkpoints"), + ) + ], + outputs=[io.Clip.Output()], + ) + + @classmethod + def execute(cls, text_encoder, ckpt_name, device="default"): + clip_type = comfy.sd.CLIPType.LTXV + + clip_path1 = folder_paths.get_full_path_or_raise("text_encoders", text_encoder) + clip_path2 = folder_paths.get_full_path_or_raise("checkpoints", ckpt_name) + + clip = comfy.sd.load_clip(ckpt_paths=[clip_path1, clip_path2], embedding_directory=folder_paths.get_folder_paths("embeddings"), clip_type=clip_type) + return io.NodeOutput(clip) + + +class LTXVAudioExtension(ComfyExtension): + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + LTXVAudioVAELoader, + LTXVAudioVAEEncode, + LTXVAudioVAEDecode, + LTXVEmptyLatentAudio, + LTXAVTextEncoderLoader, + ] + + +async def comfy_entrypoint() -> ComfyExtension: + return LTXVAudioExtension() diff --git a/comfy_extras/nodes_lt_upsampler.py b/comfy_extras/nodes_lt_upsampler.py new file mode 100644 index 000000000..f99ba13fb --- /dev/null +++ b/comfy_extras/nodes_lt_upsampler.py @@ -0,0 +1,75 @@ +from comfy import model_management +import math + +class LTXVLatentUpsampler: + """ + Upsamples a video latent by a factor of 2. + """ + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "samples": ("LATENT",), + "upscale_model": ("LATENT_UPSCALE_MODEL",), + "vae": ("VAE",), + } + } + + RETURN_TYPES = ("LATENT",) + FUNCTION = "upsample_latent" + CATEGORY = "latent/video" + EXPERIMENTAL = True + + def upsample_latent( + self, + samples: dict, + upscale_model, + vae, + ) -> tuple: + """ + Upsample the input latent using the provided model. + + Args: + samples (dict): Input latent samples + upscale_model (LatentUpsampler): Loaded upscale model + vae: VAE model for normalization + auto_tiling (bool): Whether to automatically tile the input for processing + + Returns: + tuple: Tuple containing the upsampled latent + """ + device = model_management.get_torch_device() + memory_required = model_management.module_size(upscale_model) + + model_dtype = next(upscale_model.parameters()).dtype + latents = samples["samples"] + input_dtype = latents.dtype + + memory_required += math.prod(latents.shape) * 3000.0 # TODO: more accurate + model_management.free_memory(memory_required, device) + + try: + upscale_model.to(device) # TODO: use the comfy model management system. + + latents = latents.to(dtype=model_dtype, device=device) + + """Upsample latents without tiling.""" + latents = vae.first_stage_model.per_channel_statistics.un_normalize(latents) + upsampled_latents = upscale_model(latents) + finally: + upscale_model.cpu() + + upsampled_latents = vae.first_stage_model.per_channel_statistics.normalize( + upsampled_latents + ) + upsampled_latents = upsampled_latents.to(dtype=input_dtype, device=model_management.intermediate_device()) + return_dict = samples.copy() + return_dict["samples"] = upsampled_latents + return_dict.pop("noise_mask", None) + return (return_dict,) + + +NODE_CLASS_MAPPINGS = { + "LTXVLatentUpsampler": LTXVLatentUpsampler, +} diff --git a/comfy_extras/nodes_mahiro.py b/comfy_extras/nodes_mahiro.py index 07b3353f4..6459ca8c1 100644 --- a/comfy_extras/nodes_mahiro.py +++ b/comfy_extras/nodes_mahiro.py @@ -10,7 +10,7 @@ class Mahiro(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="Mahiro", - display_name="Mahiro is so cute that she deserves a better guidance function!! (。・ω・。)", + display_name="Mahiro CFG", category="_for_testing", description="Modify the guidance to scale more on the 'direction' of the positive prompt rather than the difference between the negative prompt.", inputs=[ diff --git a/comfy_extras/nodes_mask.py b/comfy_extras/nodes_mask.py index a5e405008..290e6f55e 100644 --- a/comfy_extras/nodes_mask.py +++ b/comfy_extras/nodes_mask.py @@ -3,11 +3,10 @@ import scipy.ndimage import torch import comfy.utils import node_helpers -import folder_paths -import random +from typing_extensions import override +from comfy_api.latest import ComfyExtension, IO, UI import nodes -from nodes import MAX_RESOLUTION def composite(destination, source, x, y, mask = None, multiplier = 8, resize_source = False): source = source.to(destination.device) @@ -46,202 +45,213 @@ def composite(destination, source, x, y, mask = None, multiplier = 8, resize_sou destination[..., top:bottom, left:right] = source_portion + destination_portion return destination -class LatentCompositeMasked: +class LatentCompositeMasked(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "destination": ("LATENT",), - "source": ("LATENT",), - "x": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 8}), - "y": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 8}), - "resize_source": ("BOOLEAN", {"default": False}), - }, - "optional": { - "mask": ("MASK",), - } - } - RETURN_TYPES = ("LATENT",) - FUNCTION = "composite" + def define_schema(cls): + return IO.Schema( + node_id="LatentCompositeMasked", + category="latent", + inputs=[ + IO.Latent.Input("destination"), + IO.Latent.Input("source"), + IO.Int.Input("x", default=0, min=0, max=nodes.MAX_RESOLUTION, step=8), + IO.Int.Input("y", default=0, min=0, max=nodes.MAX_RESOLUTION, step=8), + IO.Boolean.Input("resize_source", default=False), + IO.Mask.Input("mask", optional=True), + ], + outputs=[IO.Latent.Output()], + ) - CATEGORY = "latent" - - def composite(self, destination, source, x, y, resize_source, mask = None): + @classmethod + def execute(cls, destination, source, x, y, resize_source, mask = None) -> IO.NodeOutput: output = destination.copy() destination = destination["samples"].clone() source = source["samples"] output["samples"] = composite(destination, source, x, y, mask, 8, resize_source) - return (output,) + return IO.NodeOutput(output) -class ImageCompositeMasked: + composite = execute # TODO: remove + + +class ImageCompositeMasked(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "destination": ("IMAGE",), - "source": ("IMAGE",), - "x": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - "y": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - "resize_source": ("BOOLEAN", {"default": False}), - }, - "optional": { - "mask": ("MASK",), - } - } - RETURN_TYPES = ("IMAGE",) - FUNCTION = "composite" + def define_schema(cls): + return IO.Schema( + node_id="ImageCompositeMasked", + category="image", + inputs=[ + IO.Image.Input("destination"), + IO.Image.Input("source"), + IO.Int.Input("x", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("y", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + IO.Boolean.Input("resize_source", default=False), + IO.Mask.Input("mask", optional=True), + ], + outputs=[IO.Image.Output()], + ) - CATEGORY = "image" - - def composite(self, destination, source, x, y, resize_source, mask = None): + @classmethod + def execute(cls, destination, source, x, y, resize_source, mask = None) -> IO.NodeOutput: destination, source = node_helpers.image_alpha_fix(destination, source) destination = destination.clone().movedim(-1, 1) output = composite(destination, source.movedim(-1, 1), x, y, mask, 1, resize_source).movedim(1, -1) - return (output,) + return IO.NodeOutput(output) -class MaskToImage: + composite = execute # TODO: remove + + +class MaskToImage(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "mask": ("MASK",), - } - } + def define_schema(cls): + return IO.Schema( + node_id="MaskToImage", + display_name="Convert Mask to Image", + category="mask", + inputs=[ + IO.Mask.Input("mask"), + ], + outputs=[IO.Image.Output()], + ) - CATEGORY = "mask" - - RETURN_TYPES = ("IMAGE",) - FUNCTION = "mask_to_image" - - def mask_to_image(self, mask): + @classmethod + def execute(cls, mask) -> IO.NodeOutput: result = mask.reshape((-1, 1, mask.shape[-2], mask.shape[-1])).movedim(1, -1).expand(-1, -1, -1, 3) - return (result,) + return IO.NodeOutput(result) -class ImageToMask: + mask_to_image = execute # TODO: remove + + +class ImageToMask(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "image": ("IMAGE",), - "channel": (["red", "green", "blue", "alpha"],), - } - } + def define_schema(cls): + return IO.Schema( + node_id="ImageToMask", + display_name="Convert Image to Mask", + category="mask", + inputs=[ + IO.Image.Input("image"), + IO.Combo.Input("channel", options=["red", "green", "blue", "alpha"]), + ], + outputs=[IO.Mask.Output()], + ) - CATEGORY = "mask" - - RETURN_TYPES = ("MASK",) - FUNCTION = "image_to_mask" - - def image_to_mask(self, image, channel): + @classmethod + def execute(cls, image, channel) -> IO.NodeOutput: channels = ["red", "green", "blue", "alpha"] mask = image[:, :, :, channels.index(channel)] - return (mask,) + return IO.NodeOutput(mask) -class ImageColorToMask: + image_to_mask = execute # TODO: remove + + +class ImageColorToMask(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "image": ("IMAGE",), - "color": ("INT", {"default": 0, "min": 0, "max": 0xFFFFFF, "step": 1, "display": "color"}), - } - } + def define_schema(cls): + return IO.Schema( + node_id="ImageColorToMask", + category="mask", + inputs=[ + IO.Image.Input("image"), + IO.Int.Input("color", default=0, min=0, max=0xFFFFFF, step=1, display_mode=IO.NumberDisplay.number), + ], + outputs=[IO.Mask.Output()], + ) - CATEGORY = "mask" - - RETURN_TYPES = ("MASK",) - FUNCTION = "image_to_mask" - - def image_to_mask(self, image, color): + @classmethod + def execute(cls, image, color) -> IO.NodeOutput: temp = (torch.clamp(image, 0, 1.0) * 255.0).round().to(torch.int) temp = torch.bitwise_left_shift(temp[:,:,:,0], 16) + torch.bitwise_left_shift(temp[:,:,:,1], 8) + temp[:,:,:,2] mask = torch.where(temp == color, 1.0, 0).float() - return (mask,) + return IO.NodeOutput(mask) -class SolidMask: + image_to_mask = execute # TODO: remove + + +class SolidMask(IO.ComfyNode): @classmethod - def INPUT_TYPES(cls): - return { - "required": { - "value": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}), - "width": ("INT", {"default": 512, "min": 1, "max": MAX_RESOLUTION, "step": 1}), - "height": ("INT", {"default": 512, "min": 1, "max": MAX_RESOLUTION, "step": 1}), - } - } + def define_schema(cls): + return IO.Schema( + node_id="SolidMask", + category="mask", + inputs=[ + IO.Float.Input("value", default=1.0, min=0.0, max=1.0, step=0.01), + IO.Int.Input("width", default=512, min=1, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("height", default=512, min=1, max=nodes.MAX_RESOLUTION, step=1), + ], + outputs=[IO.Mask.Output()], + ) - CATEGORY = "mask" - - RETURN_TYPES = ("MASK",) - - FUNCTION = "solid" - - def solid(self, value, width, height): + @classmethod + def execute(cls, value, width, height) -> IO.NodeOutput: out = torch.full((1, height, width), value, dtype=torch.float32, device="cpu") - return (out,) + return IO.NodeOutput(out) -class InvertMask: + solid = execute # TODO: remove + + +class InvertMask(IO.ComfyNode): @classmethod - def INPUT_TYPES(cls): - return { - "required": { - "mask": ("MASK",), - } - } + def define_schema(cls): + return IO.Schema( + node_id="InvertMask", + category="mask", + inputs=[ + IO.Mask.Input("mask"), + ], + outputs=[IO.Mask.Output()], + ) - CATEGORY = "mask" - - RETURN_TYPES = ("MASK",) - - FUNCTION = "invert" - - def invert(self, mask): + @classmethod + def execute(cls, mask) -> IO.NodeOutput: out = 1.0 - mask - return (out,) + return IO.NodeOutput(out) -class CropMask: + invert = execute # TODO: remove + + +class CropMask(IO.ComfyNode): @classmethod - def INPUT_TYPES(cls): - return { - "required": { - "mask": ("MASK",), - "x": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - "y": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - "width": ("INT", {"default": 512, "min": 1, "max": MAX_RESOLUTION, "step": 1}), - "height": ("INT", {"default": 512, "min": 1, "max": MAX_RESOLUTION, "step": 1}), - } - } + def define_schema(cls): + return IO.Schema( + node_id="CropMask", + category="mask", + inputs=[ + IO.Mask.Input("mask"), + IO.Int.Input("x", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("y", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("width", default=512, min=1, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("height", default=512, min=1, max=nodes.MAX_RESOLUTION, step=1), + ], + outputs=[IO.Mask.Output()], + ) - CATEGORY = "mask" - - RETURN_TYPES = ("MASK",) - - FUNCTION = "crop" - - def crop(self, mask, x, y, width, height): + @classmethod + def execute(cls, mask, x, y, width, height) -> IO.NodeOutput: mask = mask.reshape((-1, mask.shape[-2], mask.shape[-1])) out = mask[:, y:y + height, x:x + width] - return (out,) + return IO.NodeOutput(out) -class MaskComposite: + crop = execute # TODO: remove + + +class MaskComposite(IO.ComfyNode): @classmethod - def INPUT_TYPES(cls): - return { - "required": { - "destination": ("MASK",), - "source": ("MASK",), - "x": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - "y": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - "operation": (["multiply", "add", "subtract", "and", "or", "xor"],), - } - } + def define_schema(cls): + return IO.Schema( + node_id="MaskComposite", + category="mask", + inputs=[ + IO.Mask.Input("destination"), + IO.Mask.Input("source"), + IO.Int.Input("x", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("y", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + IO.Combo.Input("operation", options=["multiply", "add", "subtract", "and", "or", "xor"]), + ], + outputs=[IO.Mask.Output()], + ) - CATEGORY = "mask" - - RETURN_TYPES = ("MASK",) - - FUNCTION = "combine" - - def combine(self, destination, source, x, y, operation): + @classmethod + def execute(cls, destination, source, x, y, operation) -> IO.NodeOutput: output = destination.reshape((-1, destination.shape[-2], destination.shape[-1])).clone() source = source.reshape((-1, source.shape[-2], source.shape[-1])) @@ -267,28 +277,29 @@ class MaskComposite: output = torch.clamp(output, 0.0, 1.0) - return (output,) + return IO.NodeOutput(output) -class FeatherMask: + combine = execute # TODO: remove + + +class FeatherMask(IO.ComfyNode): @classmethod - def INPUT_TYPES(cls): - return { - "required": { - "mask": ("MASK",), - "left": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - "top": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - "right": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - "bottom": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - } - } + def define_schema(cls): + return IO.Schema( + node_id="FeatherMask", + category="mask", + inputs=[ + IO.Mask.Input("mask"), + IO.Int.Input("left", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("top", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("right", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("bottom", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + ], + outputs=[IO.Mask.Output()], + ) - CATEGORY = "mask" - - RETURN_TYPES = ("MASK",) - - FUNCTION = "feather" - - def feather(self, mask, left, top, right, bottom): + @classmethod + def execute(cls, mask, left, top, right, bottom) -> IO.NodeOutput: output = mask.reshape((-1, mask.shape[-2], mask.shape[-1])).clone() left = min(left, output.shape[-1]) @@ -312,26 +323,28 @@ class FeatherMask: feather_rate = (y + 1) / bottom output[:, -y, :] *= feather_rate - return (output,) + return IO.NodeOutput(output) -class GrowMask: + feather = execute # TODO: remove + + +class GrowMask(IO.ComfyNode): @classmethod - def INPUT_TYPES(cls): - return { - "required": { - "mask": ("MASK",), - "expand": ("INT", {"default": 0, "min": -MAX_RESOLUTION, "max": MAX_RESOLUTION, "step": 1}), - "tapered_corners": ("BOOLEAN", {"default": True}), - }, - } + def define_schema(cls): + return IO.Schema( + node_id="GrowMask", + display_name="Grow Mask", + category="mask", + inputs=[ + IO.Mask.Input("mask"), + IO.Int.Input("expand", default=0, min=-nodes.MAX_RESOLUTION, max=nodes.MAX_RESOLUTION, step=1), + IO.Boolean.Input("tapered_corners", default=True), + ], + outputs=[IO.Mask.Output()], + ) - CATEGORY = "mask" - - RETURN_TYPES = ("MASK",) - - FUNCTION = "expand_mask" - - def expand_mask(self, mask, expand, tapered_corners): + @classmethod + def execute(cls, mask, expand, tapered_corners) -> IO.NodeOutput: c = 0 if tapered_corners else 1 kernel = np.array([[c, 1, c], [1, 1, 1], @@ -347,69 +360,74 @@ class GrowMask: output = scipy.ndimage.grey_dilation(output, footprint=kernel) output = torch.from_numpy(output) out.append(output) - return (torch.stack(out, dim=0),) + return IO.NodeOutput(torch.stack(out, dim=0)) -class ThresholdMask: + expand_mask = execute # TODO: remove + + +class ThresholdMask(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "mask": ("MASK",), - "value": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 1.0, "step": 0.01}), - } - } + def define_schema(cls): + return IO.Schema( + node_id="ThresholdMask", + category="mask", + inputs=[ + IO.Mask.Input("mask"), + IO.Float.Input("value", default=0.5, min=0.0, max=1.0, step=0.01), + ], + outputs=[IO.Mask.Output()], + ) - CATEGORY = "mask" - - RETURN_TYPES = ("MASK",) - FUNCTION = "image_to_mask" - - def image_to_mask(self, mask, value): + @classmethod + def execute(cls, mask, value) -> IO.NodeOutput: mask = (mask > value).float() - return (mask,) + return IO.NodeOutput(mask) + + image_to_mask = execute # TODO: remove + # Mask Preview - original implement from # https://github.com/cubiq/ComfyUI_essentials/blob/9d9f4bedfc9f0321c19faf71855e228c93bd0dc9/mask.py#L81 # upstream requested in https://github.com/Kosinkadink/rfcs/blob/main/rfcs/0000-corenodes.md#preview-nodes -class MaskPreview(nodes.SaveImage): - def __init__(self): - self.output_dir = folder_paths.get_temp_directory() - self.type = "temp" - self.prefix_append = "_temp_" + ''.join(random.choice("abcdefghijklmnopqrstupvxyz") for x in range(5)) - self.compress_level = 4 +class MaskPreview(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="MaskPreview", + display_name="Preview Mask", + category="mask", + description="Saves the input images to your ComfyUI output directory.", + inputs=[ + IO.Mask.Input("mask"), + ], + hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo], + is_output_node=True, + ) @classmethod - def INPUT_TYPES(s): - return { - "required": {"mask": ("MASK",), }, - "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"}, - } - - FUNCTION = "execute" - CATEGORY = "mask" - - def execute(self, mask, filename_prefix="ComfyUI", prompt=None, extra_pnginfo=None): - preview = mask.reshape((-1, 1, mask.shape[-2], mask.shape[-1])).movedim(1, -1).expand(-1, -1, -1, 3) - return self.save_images(preview, filename_prefix, prompt, extra_pnginfo) + def execute(cls, mask, filename_prefix="ComfyUI") -> IO.NodeOutput: + return IO.NodeOutput(ui=UI.PreviewMask(mask)) -NODE_CLASS_MAPPINGS = { - "LatentCompositeMasked": LatentCompositeMasked, - "ImageCompositeMasked": ImageCompositeMasked, - "MaskToImage": MaskToImage, - "ImageToMask": ImageToMask, - "ImageColorToMask": ImageColorToMask, - "SolidMask": SolidMask, - "InvertMask": InvertMask, - "CropMask": CropMask, - "MaskComposite": MaskComposite, - "FeatherMask": FeatherMask, - "GrowMask": GrowMask, - "ThresholdMask": ThresholdMask, - "MaskPreview": MaskPreview -} +class MaskExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + LatentCompositeMasked, + ImageCompositeMasked, + MaskToImage, + ImageToMask, + ImageColorToMask, + SolidMask, + InvertMask, + CropMask, + MaskComposite, + FeatherMask, + GrowMask, + ThresholdMask, + MaskPreview, + ] -NODE_DISPLAY_NAME_MAPPINGS = { - "ImageToMask": "Convert Image to Mask", - "MaskToImage": "Convert Mask to Image", -} + +async def comfy_entrypoint() -> MaskExtension: + return MaskExtension() diff --git a/comfy_extras/nodes_model_downscale.py b/comfy_extras/nodes_model_downscale.py index f7ca9699d..dec2ae841 100644 --- a/comfy_extras/nodes_model_downscale.py +++ b/comfy_extras/nodes_model_downscale.py @@ -53,11 +53,6 @@ class PatchModelAddDownscale(io.ComfyNode): return io.NodeOutput(m) -NODE_DISPLAY_NAME_MAPPINGS = { - # Sampling - "PatchModelAddDownscale": "", -} - class ModelDownscaleExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[io.ComfyNode]]: diff --git a/comfy_extras/nodes_model_patch.py b/comfy_extras/nodes_model_patch.py index 783c59b6b..1355b3c93 100644 --- a/comfy_extras/nodes_model_patch.py +++ b/comfy_extras/nodes_model_patch.py @@ -6,6 +6,7 @@ import comfy.ops import comfy.model_management import comfy.ldm.common_dit import comfy.latent_formats +import comfy.ldm.lumina.controlnet class BlockWiseControlBlock(torch.nn.Module): @@ -189,6 +190,35 @@ class SigLIPMultiFeatProjModel(torch.nn.Module): return embedding +def z_image_convert(sd): + replace_keys = {".attention.to_out.0.bias": ".attention.out.bias", + ".attention.norm_k.weight": ".attention.k_norm.weight", + ".attention.norm_q.weight": ".attention.q_norm.weight", + ".attention.to_out.0.weight": ".attention.out.weight" + } + + out_sd = {} + for k in sorted(sd.keys()): + w = sd[k] + + k_out = k + if k_out.endswith(".attention.to_k.weight"): + cc = [w] + continue + if k_out.endswith(".attention.to_q.weight"): + cc = [w] + cc + continue + if k_out.endswith(".attention.to_v.weight"): + cc = cc + [w] + w = torch.cat(cc, dim=0) + k_out = k_out.replace(".attention.to_v.weight", ".attention.qkv.weight") + + for r, rr in replace_keys.items(): + k_out = k_out.replace(r, rr) + out_sd[k_out] = w + + return out_sd + class ModelPatchLoader: @classmethod def INPUT_TYPES(s): @@ -211,6 +241,18 @@ class ModelPatchLoader: elif 'feature_embedder.mid_layer_norm.bias' in sd: sd = comfy.utils.state_dict_prefix_replace(sd, {"feature_embedder.": ""}, filter_keys=True) model = SigLIPMultiFeatProjModel(device=comfy.model_management.unet_offload_device(), dtype=dtype, operations=comfy.ops.manual_cast) + elif 'control_all_x_embedder.2-1.weight' in sd: # alipai z image fun controlnet + sd = z_image_convert(sd) + config = {} + if 'control_layers.14.adaLN_modulation.0.weight' in sd: + config['n_control_layers'] = 15 + config['additional_in_dim'] = 17 + config['refiner_control'] = True + ref_weight = sd.get("control_noise_refiner.0.after_proj.weight", None) + if ref_weight is not None: + if torch.count_nonzero(ref_weight) == 0: + config['broken'] = True + model = comfy.ldm.lumina.controlnet.ZImage_Control(device=comfy.model_management.unet_offload_device(), dtype=dtype, operations=comfy.ops.manual_cast, **config) model.load_state_dict(sd) model = comfy.model_patcher.ModelPatcher(model, load_device=comfy.model_management.get_torch_device(), offload_device=comfy.model_management.unet_offload_device()) @@ -263,6 +305,129 @@ class DiffSynthCnetPatch: def models(self): return [self.model_patch] +class ZImageControlPatch: + def __init__(self, model_patch, vae, image, strength, inpaint_image=None, mask=None): + self.model_patch = model_patch + self.vae = vae + self.image = image + self.inpaint_image = inpaint_image + self.mask = mask + self.strength = strength + self.is_inpaint = self.model_patch.model.additional_in_dim > 0 + + skip_encoding = False + if self.image is not None and self.inpaint_image is not None: + if self.image.shape != self.inpaint_image.shape: + skip_encoding = True + + if skip_encoding: + self.encoded_image = None + else: + self.encoded_image = self.encode_latent_cond(self.image, self.inpaint_image) + if self.image is None: + self.encoded_image_size = (self.inpaint_image.shape[1], self.inpaint_image.shape[2]) + else: + self.encoded_image_size = (self.image.shape[1], self.image.shape[2]) + self.temp_data = None + + def encode_latent_cond(self, control_image=None, inpaint_image=None): + latent_image = None + if control_image is not None: + latent_image = comfy.latent_formats.Flux().process_in(self.vae.encode(control_image)) + + if self.is_inpaint: + if inpaint_image is None: + inpaint_image = torch.ones_like(control_image) * 0.5 + + if self.mask is not None: + mask_inpaint = comfy.utils.common_upscale(self.mask.view(self.mask.shape[0], -1, self.mask.shape[-2], self.mask.shape[-1]).mean(dim=1, keepdim=True), inpaint_image.shape[-2], inpaint_image.shape[-3], "bilinear", "center") + inpaint_image = ((inpaint_image - 0.5) * mask_inpaint.movedim(1, -1).round()) + 0.5 + + inpaint_image_latent = comfy.latent_formats.Flux().process_in(self.vae.encode(inpaint_image)) + + if self.mask is None: + mask_ = torch.zeros_like(inpaint_image_latent)[:, :1] + else: + mask_ = comfy.utils.common_upscale(self.mask.view(self.mask.shape[0], -1, self.mask.shape[-2], self.mask.shape[-1]).mean(dim=1, keepdim=True).to(device=inpaint_image_latent.device), inpaint_image_latent.shape[-1], inpaint_image_latent.shape[-2], "nearest", "center") + + if latent_image is None: + latent_image = comfy.latent_formats.Flux().process_in(self.vae.encode(torch.ones_like(inpaint_image) * 0.5)) + + return torch.cat([latent_image, mask_, inpaint_image_latent], dim=1) + else: + return latent_image + + def __call__(self, kwargs): + x = kwargs.get("x") + img = kwargs.get("img") + img_input = kwargs.get("img_input") + txt = kwargs.get("txt") + pe = kwargs.get("pe") + vec = kwargs.get("vec") + block_index = kwargs.get("block_index") + block_type = kwargs.get("block_type", "") + spacial_compression = self.vae.spacial_compression_encode() + if self.encoded_image is None or self.encoded_image_size != (x.shape[-2] * spacial_compression, x.shape[-1] * spacial_compression): + image_scaled = None + if self.image is not None: + image_scaled = comfy.utils.common_upscale(self.image.movedim(-1, 1), x.shape[-1] * spacial_compression, x.shape[-2] * spacial_compression, "area", "center").movedim(1, -1) + self.encoded_image_size = (image_scaled.shape[-3], image_scaled.shape[-2]) + + inpaint_scaled = None + if self.inpaint_image is not None: + inpaint_scaled = comfy.utils.common_upscale(self.inpaint_image.movedim(-1, 1), x.shape[-1] * spacial_compression, x.shape[-2] * spacial_compression, "area", "center").movedim(1, -1) + self.encoded_image_size = (inpaint_scaled.shape[-3], inpaint_scaled.shape[-2]) + + loaded_models = comfy.model_management.loaded_models(only_currently_used=True) + self.encoded_image = self.encode_latent_cond(image_scaled, inpaint_scaled) + comfy.model_management.load_models_gpu(loaded_models) + + cnet_blocks = self.model_patch.model.n_control_layers + div = round(30 / cnet_blocks) + + cnet_index = (block_index // div) + cnet_index_float = (block_index / div) + + kwargs.pop("img") # we do ops in place + kwargs.pop("txt") + + if cnet_index_float > (cnet_blocks - 1): + self.temp_data = None + return kwargs + + if self.temp_data is None or self.temp_data[0] > cnet_index: + if block_type == "noise_refiner": + self.temp_data = (-3, (None, self.model_patch.model(txt, self.encoded_image.to(img.dtype), pe, vec))) + else: + self.temp_data = (-1, (None, self.model_patch.model(txt, self.encoded_image.to(img.dtype), pe, vec))) + + if block_type == "noise_refiner": + next_layer = self.temp_data[0] + 1 + self.temp_data = (next_layer, self.model_patch.model.forward_noise_refiner_block(block_index, self.temp_data[1][1], img_input[:, :self.temp_data[1][1].shape[1]], None, pe, vec)) + if self.temp_data[1][0] is not None: + img[:, :self.temp_data[1][0].shape[1]] += (self.temp_data[1][0] * self.strength) + else: + while self.temp_data[0] < cnet_index and (self.temp_data[0] + 1) < cnet_blocks: + next_layer = self.temp_data[0] + 1 + self.temp_data = (next_layer, self.model_patch.model.forward_control_block(next_layer, self.temp_data[1][1], img_input[:, :self.temp_data[1][1].shape[1]], None, pe, vec)) + + if cnet_index_float == self.temp_data[0]: + img[:, :self.temp_data[1][0].shape[1]] += (self.temp_data[1][0] * self.strength) + if cnet_blocks == self.temp_data[0] + 1: + self.temp_data = None + + return kwargs + + def to(self, device_or_dtype): + if isinstance(device_or_dtype, torch.device): + if self.encoded_image is not None: + self.encoded_image = self.encoded_image.to(device_or_dtype) + self.temp_data = None + return self + + def models(self): + return [self.model_patch] + class QwenImageDiffsynthControlnet: @classmethod def INPUT_TYPES(s): @@ -279,9 +444,12 @@ class QwenImageDiffsynthControlnet: CATEGORY = "advanced/loaders/qwen" - def diffsynth_controlnet(self, model, model_patch, vae, image, strength, mask=None): + def diffsynth_controlnet(self, model, model_patch, vae, image=None, strength=1.0, inpaint_image=None, mask=None): model_patched = model.clone() - image = image[:, :, :, :3] + if image is not None: + image = image[:, :, :, :3] + if inpaint_image is not None: + inpaint_image = inpaint_image[:, :, :, :3] if mask is not None: if mask.ndim == 3: mask = mask.unsqueeze(1) @@ -289,9 +457,25 @@ class QwenImageDiffsynthControlnet: mask = mask.unsqueeze(2) mask = 1.0 - mask - model_patched.set_model_double_block_patch(DiffSynthCnetPatch(model_patch, vae, image, strength, mask)) + if isinstance(model_patch.model, comfy.ldm.lumina.controlnet.ZImage_Control): + patch = ZImageControlPatch(model_patch, vae, image, strength, inpaint_image=inpaint_image, mask=mask) + model_patched.set_model_noise_refiner_patch(patch) + model_patched.set_model_double_block_patch(patch) + else: + model_patched.set_model_double_block_patch(DiffSynthCnetPatch(model_patch, vae, image, strength, mask)) return (model_patched,) +class ZImageFunControlnet(QwenImageDiffsynthControlnet): + @classmethod + def INPUT_TYPES(s): + return {"required": { "model": ("MODEL",), + "model_patch": ("MODEL_PATCH",), + "vae": ("VAE",), + "strength": ("FLOAT", {"default": 1.0, "min": -10.0, "max": 10.0, "step": 0.01}), + }, + "optional": {"image": ("IMAGE",), "inpaint_image": ("IMAGE",), "mask": ("MASK",)}} + + CATEGORY = "advanced/loaders/zimage" class UsoStyleProjectorPatch: def __init__(self, model_patch, encoded_image): @@ -339,5 +523,6 @@ class USOStyleReference: NODE_CLASS_MAPPINGS = { "ModelPatchLoader": ModelPatchLoader, "QwenImageDiffsynthControlnet": QwenImageDiffsynthControlnet, + "ZImageFunControlnet": ZImageFunControlnet, "USOStyleReference": USOStyleReference, } diff --git a/comfy_extras/nodes_post_processing.py b/comfy_extras/nodes_post_processing.py index 34c388a5a..01afa13a1 100644 --- a/comfy_extras/nodes_post_processing.py +++ b/comfy_extras/nodes_post_processing.py @@ -4,11 +4,15 @@ import torch import torch.nn.functional as F from PIL import Image import math +from enum import Enum +from typing import TypedDict, Literal import comfy.utils import comfy.model_management +from comfy_extras.nodes_latent import reshape_latent_to import node_helpers from comfy_api.latest import ComfyExtension, io +from nodes import MAX_RESOLUTION class Blend(io.ComfyNode): @classmethod @@ -221,6 +225,7 @@ class ImageScaleToTotalPixels(io.ComfyNode): io.Image.Input("image"), io.Combo.Input("upscale_method", options=cls.upscale_methods), io.Float.Input("megapixels", default=1.0, min=0.01, max=16.0, step=0.01), + io.Int.Input("resolution_steps", default=1, min=1, max=256), ], outputs=[ io.Image.Output(), @@ -228,18 +233,365 @@ class ImageScaleToTotalPixels(io.ComfyNode): ) @classmethod - def execute(cls, image, upscale_method, megapixels) -> io.NodeOutput: + def execute(cls, image, upscale_method, megapixels, resolution_steps) -> io.NodeOutput: samples = image.movedim(-1,1) - total = int(megapixels * 1024 * 1024) + total = megapixels * 1024 * 1024 scale_by = math.sqrt(total / (samples.shape[3] * samples.shape[2])) - width = round(samples.shape[3] * scale_by) - height = round(samples.shape[2] * scale_by) + width = round(samples.shape[3] * scale_by / resolution_steps) * resolution_steps + height = round(samples.shape[2] * scale_by / resolution_steps) * resolution_steps - s = comfy.utils.common_upscale(samples, width, height, upscale_method, "disabled") + s = comfy.utils.common_upscale(samples, int(width), int(height), upscale_method, "disabled") s = s.movedim(1,-1) return io.NodeOutput(s) +class ResizeType(str, Enum): + SCALE_BY = "scale by multiplier" + SCALE_DIMENSIONS = "scale dimensions" + SCALE_LONGER_DIMENSION = "scale longer dimension" + SCALE_SHORTER_DIMENSION = "scale shorter dimension" + SCALE_WIDTH = "scale width" + SCALE_HEIGHT = "scale height" + SCALE_TOTAL_PIXELS = "scale total pixels" + MATCH_SIZE = "match size" + +def is_image(input: torch.Tensor) -> bool: + # images have 4 dimensions: [batch, height, width, channels] + # masks have 3 dimensions: [batch, height, width] + return len(input.shape) == 4 + +def init_image_mask_input(input: torch.Tensor, is_type_image: bool) -> torch.Tensor: + if is_type_image: + input = input.movedim(-1, 1) + else: + input = input.unsqueeze(1) + return input + +def finalize_image_mask_input(input: torch.Tensor, is_type_image: bool) -> torch.Tensor: + if is_type_image: + input = input.movedim(1, -1) + else: + input = input.squeeze(1) + return input + +def scale_by(input: torch.Tensor, multiplier: float, scale_method: str) -> torch.Tensor: + is_type_image = is_image(input) + input = init_image_mask_input(input, is_type_image) + width = round(input.shape[-1] * multiplier) + height = round(input.shape[-2] * multiplier) + + input = comfy.utils.common_upscale(input, width, height, scale_method, "disabled") + input = finalize_image_mask_input(input, is_type_image) + return input + +def scale_dimensions(input: torch.Tensor, width: int, height: int, scale_method: str, crop: str="disabled") -> torch.Tensor: + if width == 0 and height == 0: + return input + is_type_image = is_image(input) + input = init_image_mask_input(input, is_type_image) + + if width == 0: + width = max(1, round(input.shape[-1] * height / input.shape[-2])) + elif height == 0: + height = max(1, round(input.shape[-2] * width / input.shape[-1])) + + input = comfy.utils.common_upscale(input, width, height, scale_method, crop) + input = finalize_image_mask_input(input, is_type_image) + return input + +def scale_longer_dimension(input: torch.Tensor, longer_size: int, scale_method: str) -> torch.Tensor: + is_type_image = is_image(input) + input = init_image_mask_input(input, is_type_image) + width = input.shape[-1] + height = input.shape[-2] + + if height > width: + width = round((width / height) * longer_size) + height = longer_size + elif width > height: + height = round((height / width) * longer_size) + width = longer_size + else: + height = longer_size + width = longer_size + + input = comfy.utils.common_upscale(input, width, height, scale_method, "disabled") + input = finalize_image_mask_input(input, is_type_image) + return input + +def scale_shorter_dimension(input: torch.Tensor, shorter_size: int, scale_method: str) -> torch.Tensor: + is_type_image = is_image(input) + input = init_image_mask_input(input, is_type_image) + width = input.shape[-1] + height = input.shape[-2] + + if height < width: + width = round((width / height) * shorter_size) + height = shorter_size + elif width > height: + height = round((height / width) * shorter_size) + width = shorter_size + else: + height = shorter_size + width = shorter_size + + input = comfy.utils.common_upscale(input, width, height, scale_method, "disabled") + input = finalize_image_mask_input(input, is_type_image) + return input + +def scale_total_pixels(input: torch.Tensor, megapixels: float, scale_method: str) -> torch.Tensor: + is_type_image = is_image(input) + input = init_image_mask_input(input, is_type_image) + total = int(megapixels * 1024 * 1024) + + scale_by = math.sqrt(total / (input.shape[-1] * input.shape[-2])) + width = round(input.shape[-1] * scale_by) + height = round(input.shape[-2] * scale_by) + + input = comfy.utils.common_upscale(input, width, height, scale_method, "disabled") + input = finalize_image_mask_input(input, is_type_image) + return input + +def scale_match_size(input: torch.Tensor, match: torch.Tensor, scale_method: str, crop: str) -> torch.Tensor: + is_type_image = is_image(input) + input = init_image_mask_input(input, is_type_image) + match = init_image_mask_input(match, is_image(match)) + + width = match.shape[-1] + height = match.shape[-2] + input = comfy.utils.common_upscale(input, width, height, scale_method, crop) + input = finalize_image_mask_input(input, is_type_image) + return input + +class ResizeImageMaskNode(io.ComfyNode): + + scale_methods = ["nearest-exact", "bilinear", "area", "bicubic", "lanczos"] + crop_methods = ["disabled", "center"] + + class ResizeTypedDict(TypedDict): + resize_type: ResizeType + scale_method: Literal["nearest-exact", "bilinear", "area", "bicubic", "lanczos"] + crop: Literal["disabled", "center"] + multiplier: float + width: int + height: int + longer_size: int + shorter_size: int + megapixels: float + + @classmethod + def define_schema(cls): + template = io.MatchType.Template("input_type", [io.Image, io.Mask]) + crop_combo = io.Combo.Input("crop", options=cls.crop_methods, default="center") + return io.Schema( + node_id="ResizeImageMaskNode", + display_name="Resize Image/Mask", + category="transform", + inputs=[ + io.MatchType.Input("input", template=template), + io.DynamicCombo.Input("resize_type", options=[ + io.DynamicCombo.Option(ResizeType.SCALE_BY, [ + io.Float.Input("multiplier", default=1.00, min=0.01, max=8.0, step=0.01), + ]), + io.DynamicCombo.Option(ResizeType.SCALE_DIMENSIONS, [ + io.Int.Input("width", default=512, min=0, max=MAX_RESOLUTION, step=1), + io.Int.Input("height", default=512, min=0, max=MAX_RESOLUTION, step=1), + crop_combo, + ]), + io.DynamicCombo.Option(ResizeType.SCALE_LONGER_DIMENSION, [ + io.Int.Input("longer_size", default=512, min=0, max=MAX_RESOLUTION, step=1), + ]), + io.DynamicCombo.Option(ResizeType.SCALE_SHORTER_DIMENSION, [ + io.Int.Input("shorter_size", default=512, min=0, max=MAX_RESOLUTION, step=1), + ]), + io.DynamicCombo.Option(ResizeType.SCALE_WIDTH, [ + io.Int.Input("width", default=512, min=0, max=MAX_RESOLUTION, step=1), + ]), + io.DynamicCombo.Option(ResizeType.SCALE_HEIGHT, [ + io.Int.Input("height", default=512, min=0, max=MAX_RESOLUTION, step=1), + ]), + io.DynamicCombo.Option(ResizeType.SCALE_TOTAL_PIXELS, [ + io.Float.Input("megapixels", default=1.0, min=0.01, max=16.0, step=0.01), + ]), + io.DynamicCombo.Option(ResizeType.MATCH_SIZE, [ + io.MultiType.Input("match", [io.Image, io.Mask]), + crop_combo, + ]), + ]), + io.Combo.Input("scale_method", options=cls.scale_methods, default="area"), + ], + outputs=[io.MatchType.Output(template=template, display_name="resized")] + ) + + @classmethod + def execute(cls, input: io.Image.Type | io.Mask.Type, scale_method: io.Combo.Type, resize_type: ResizeTypedDict) -> io.NodeOutput: + selected_type = resize_type["resize_type"] + if selected_type == ResizeType.SCALE_BY: + return io.NodeOutput(scale_by(input, resize_type["multiplier"], scale_method)) + elif selected_type == ResizeType.SCALE_DIMENSIONS: + return io.NodeOutput(scale_dimensions(input, resize_type["width"], resize_type["height"], scale_method, resize_type["crop"])) + elif selected_type == ResizeType.SCALE_LONGER_DIMENSION: + return io.NodeOutput(scale_longer_dimension(input, resize_type["longer_size"], scale_method)) + elif selected_type == ResizeType.SCALE_SHORTER_DIMENSION: + return io.NodeOutput(scale_shorter_dimension(input, resize_type["shorter_size"], scale_method)) + elif selected_type == ResizeType.SCALE_WIDTH: + return io.NodeOutput(scale_dimensions(input, resize_type["width"], 0, scale_method)) + elif selected_type == ResizeType.SCALE_HEIGHT: + return io.NodeOutput(scale_dimensions(input, 0, resize_type["height"], scale_method)) + elif selected_type == ResizeType.SCALE_TOTAL_PIXELS: + return io.NodeOutput(scale_total_pixels(input, resize_type["megapixels"], scale_method)) + elif selected_type == ResizeType.MATCH_SIZE: + return io.NodeOutput(scale_match_size(input, resize_type["match"], scale_method, resize_type["crop"])) + raise ValueError(f"Unsupported resize type: {selected_type}") + +def batch_images(images: list[torch.Tensor]) -> torch.Tensor | None: + if len(images) == 0: + return None + # first, get the max channels count + max_channels = max(image.shape[-1] for image in images) + # then, pad all images to have the same channels count + padded_images: list[torch.Tensor] = [] + for image in images: + if image.shape[-1] < max_channels: + padded_images.append(torch.nn.functional.pad(image, (0,1), mode='constant', value=1.0)) + else: + padded_images.append(image) + # resize all images to be the same size as the first image + resized_images: list[torch.Tensor] = [] + first_image_shape = padded_images[0].shape + for image in padded_images: + if image.shape[1:] != first_image_shape[1:]: + resized_images.append(comfy.utils.common_upscale(image.movedim(-1,1), first_image_shape[2], first_image_shape[1], "bilinear", "center").movedim(1,-1)) + else: + resized_images.append(image) + # batch the images in the format [b, h, w, c] + return torch.cat(resized_images, dim=0) + +def batch_masks(masks: list[torch.Tensor]) -> torch.Tensor | None: + if len(masks) == 0: + return None + # resize all masks to be the same size as the first mask + resized_masks: list[torch.Tensor] = [] + first_mask_shape = masks[0].shape + for mask in masks: + if mask.shape[1:] != first_mask_shape[1:]: + mask = init_image_mask_input(mask, is_type_image=False) + mask = comfy.utils.common_upscale(mask, first_mask_shape[2], first_mask_shape[1], "bilinear", "center") + resized_masks.append(finalize_image_mask_input(mask, is_type_image=False)) + else: + resized_masks.append(mask) + # batch the masks in the format [b, h, w] + return torch.cat(resized_masks, dim=0) + +def batch_latents(latents: list[dict[str, torch.Tensor]]) -> dict[str, torch.Tensor] | None: + if len(latents) == 0: + return None + samples_out = latents[0].copy() + samples_out["batch_index"] = [] + first_samples = latents[0]["samples"] + tensors: list[torch.Tensor] = [] + for latent in latents: + # first, deal with latent tensors + tensors.append(reshape_latent_to(first_samples.shape, latent["samples"], repeat_batch=False)) + # next, deal with batch_index + samples_out["batch_index"].extend(latent.get("batch_index", [x for x in range(0, latent["samples"].shape[0])])) + samples_out["samples"] = torch.cat(tensors, dim=0) + return samples_out + +class BatchImagesNode(io.ComfyNode): + @classmethod + def define_schema(cls): + autogrow_template = io.Autogrow.TemplatePrefix(io.Image.Input("image"), prefix="image", min=2, max=50) + return io.Schema( + node_id="BatchImagesNode", + display_name="Batch Images", + category="image", + inputs=[ + io.Autogrow.Input("images", template=autogrow_template) + ], + outputs=[ + io.Image.Output() + ] + ) + + @classmethod + def execute(cls, images: io.Autogrow.Type) -> io.NodeOutput: + return io.NodeOutput(batch_images(list(images.values()))) + +class BatchMasksNode(io.ComfyNode): + @classmethod + def define_schema(cls): + autogrow_template = io.Autogrow.TemplatePrefix(io.Mask.Input("mask"), prefix="mask", min=2, max=50) + return io.Schema( + node_id="BatchMasksNode", + display_name="Batch Masks", + category="mask", + inputs=[ + io.Autogrow.Input("masks", template=autogrow_template) + ], + outputs=[ + io.Mask.Output() + ] + ) + + @classmethod + def execute(cls, masks: io.Autogrow.Type) -> io.NodeOutput: + return io.NodeOutput(batch_masks(list(masks.values()))) + +class BatchLatentsNode(io.ComfyNode): + @classmethod + def define_schema(cls): + autogrow_template = io.Autogrow.TemplatePrefix(io.Latent.Input("latent"), prefix="latent", min=2, max=50) + return io.Schema( + node_id="BatchLatentsNode", + display_name="Batch Latents", + category="latent", + inputs=[ + io.Autogrow.Input("latents", template=autogrow_template) + ], + outputs=[ + io.Latent.Output() + ] + ) + + @classmethod + def execute(cls, latents: io.Autogrow.Type) -> io.NodeOutput: + return io.NodeOutput(batch_latents(list(latents.values()))) + +class BatchImagesMasksLatentsNode(io.ComfyNode): + @classmethod + def define_schema(cls): + matchtype_template = io.MatchType.Template("input", allowed_types=[io.Image, io.Mask, io.Latent]) + autogrow_template = io.Autogrow.TemplatePrefix( + io.MatchType.Input("input", matchtype_template), + prefix="input", min=1, max=50) + return io.Schema( + node_id="BatchImagesMasksLatentsNode", + display_name="Batch Images/Masks/Latents", + category="util", + inputs=[ + io.Autogrow.Input("inputs", template=autogrow_template) + ], + outputs=[ + io.MatchType.Output(id=None, template=matchtype_template) + ] + ) + + @classmethod + def execute(cls, inputs: io.Autogrow.Type) -> io.NodeOutput: + batched = None + values = list(inputs.values()) + # latents + if isinstance(values[0], dict): + batched = batch_latents(values) + # images + elif is_image(values[0]): + batched = batch_images(values) + # masks + else: + batched = batch_masks(values) + return io.NodeOutput(batched) + class PostProcessingExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[io.ComfyNode]]: @@ -249,6 +601,11 @@ class PostProcessingExtension(ComfyExtension): Quantize, Sharpen, ImageScaleToTotalPixels, + ResizeImageMaskNode, + BatchImagesNode, + BatchMasksNode, + BatchLatentsNode, + # BatchImagesMasksLatentsNode, ] async def comfy_entrypoint() -> PostProcessingExtension: diff --git a/comfy_extras/nodes_preview_any.py b/comfy_extras/nodes_preview_any.py index e749fa6ae..139b07c93 100644 --- a/comfy_extras/nodes_preview_any.py +++ b/comfy_extras/nodes_preview_any.py @@ -39,5 +39,5 @@ NODE_CLASS_MAPPINGS = { } NODE_DISPLAY_NAME_MAPPINGS = { - "PreviewAny": "Preview Any", + "PreviewAny": "Preview as Text", } diff --git a/comfy_extras/nodes_primitive.py b/comfy_extras/nodes_primitive.py index 5a1aeba80..937321800 100644 --- a/comfy_extras/nodes_primitive.py +++ b/comfy_extras/nodes_primitive.py @@ -66,7 +66,7 @@ class Float(io.ComfyNode): display_name="Float", category="utils/primitive", inputs=[ - io.Float.Input("value", min=-sys.maxsize, max=sys.maxsize), + io.Float.Input("value", min=-sys.maxsize, max=sys.maxsize, step=0.1), ], outputs=[io.Float.Output()], ) diff --git a/comfy_extras/nodes_qwen.py b/comfy_extras/nodes_qwen.py index 525239ae5..fde8fac9a 100644 --- a/comfy_extras/nodes_qwen.py +++ b/comfy_extras/nodes_qwen.py @@ -3,7 +3,9 @@ import comfy.utils import math from typing_extensions import override from comfy_api.latest import ComfyExtension, io - +import comfy.model_management +import torch +import nodes class TextEncodeQwenImageEdit(io.ComfyNode): @classmethod @@ -104,12 +106,37 @@ class TextEncodeQwenImageEditPlus(io.ComfyNode): return io.NodeOutput(conditioning) +class EmptyQwenImageLayeredLatentImage(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="EmptyQwenImageLayeredLatentImage", + display_name="Empty Qwen Image Layered Latent", + category="latent/qwen", + inputs=[ + io.Int.Input("width", default=640, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("height", default=640, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("layers", default=3, min=0, max=nodes.MAX_RESOLUTION, step=1), + io.Int.Input("batch_size", default=1, min=1, max=4096), + ], + outputs=[ + io.Latent.Output(), + ], + ) + + @classmethod + def execute(cls, width, height, layers, batch_size=1) -> io.NodeOutput: + latent = torch.zeros([batch_size, 16, layers + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device()) + return io.NodeOutput({"samples": latent}) + + class QwenExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[io.ComfyNode]]: return [ TextEncodeQwenImageEdit, TextEncodeQwenImageEditPlus, + EmptyQwenImageLayeredLatentImage, ] diff --git a/comfy_extras/nodes_torch_compile.py b/comfy_extras/nodes_torch_compile.py index adbeece2f..c43e8ad63 100644 --- a/comfy_extras/nodes_torch_compile.py +++ b/comfy_extras/nodes_torch_compile.py @@ -2,6 +2,8 @@ from typing_extensions import override from comfy_api.latest import ComfyExtension, io from comfy_api.torch_helpers import set_torch_compile_wrapper +def skip_torch_compile_dict(guard_entries): + return [("transformer_options" not in entry.name) for entry in guard_entries] class TorchCompileModel(io.ComfyNode): @classmethod @@ -23,7 +25,7 @@ class TorchCompileModel(io.ComfyNode): @classmethod def execute(cls, model, backend) -> io.NodeOutput: m = model.clone() - set_torch_compile_wrapper(model=m, backend=backend) + set_torch_compile_wrapper(model=m, backend=backend, options={"guard_filter_fn": skip_torch_compile_dict}) return io.NodeOutput(m) diff --git a/comfy_extras/nodes_train.py b/comfy_extras/nodes_train.py index 9e6ec6780..364804205 100644 --- a/comfy_extras/nodes_train.py +++ b/comfy_extras/nodes_train.py @@ -1,26 +1,87 @@ -import datetime -import json import logging import os import numpy as np import safetensors import torch -from PIL import Image, ImageDraw, ImageFont -from PIL.PngImagePlugin import PngInfo import torch.utils.checkpoint -import tqdm +from tqdm.auto import trange +from PIL import Image, ImageDraw, ImageFont +from typing_extensions import override import comfy.samplers +import comfy.sampler_helpers import comfy.sd import comfy.utils import comfy.model_management import comfy_extras.nodes_custom_sampler import folder_paths import node_helpers -from comfy.cli_args import args -from comfy.comfy_types.node_typing import IO from comfy.weight_adapter import adapters, adapter_maps +from comfy_api.latest import ComfyExtension, io, ui +from comfy.utils import ProgressBar + + +class TrainGuider(comfy_extras.nodes_custom_sampler.Guider_Basic): + """ + CFGGuider with modifications for training specific logic + """ + def outer_sample( + self, + noise, + latent_image, + sampler, + sigmas, + denoise_mask=None, + callback=None, + disable_pbar=False, + seed=None, + latent_shapes=None, + ): + self.inner_model, self.conds, self.loaded_models = ( + comfy.sampler_helpers.prepare_sampling( + self.model_patcher, + noise.shape, + self.conds, + self.model_options, + force_full_load=True, # mirror behavior in TrainLoraNode.execute() to keep model loaded + ) + ) + device = self.model_patcher.load_device + + if denoise_mask is not None: + denoise_mask = comfy.sampler_helpers.prepare_mask( + denoise_mask, noise.shape, device + ) + + noise = noise.to(device) + latent_image = latent_image.to(device) + sigmas = sigmas.to(device) + comfy.samplers.cast_to_load_options( + self.model_options, device=device, dtype=self.model_patcher.model_dtype() + ) + + try: + self.model_patcher.pre_run() + output = self.inner_sample( + noise, + latent_image, + device, + sampler, + sigmas, + denoise_mask, + callback, + disable_pbar, + seed, + latent_shapes=latent_shapes, + ) + finally: + self.model_patcher.cleanup() + + comfy.sampler_helpers.cleanup_models(self.conds, self.loaded_models) + del self.inner_model + del self.loaded_models + return output def make_batch_extra_option_dict(d, indicies, full_size=None): @@ -56,7 +117,19 @@ def process_cond_list(d, prefix=""): class TrainSampler(comfy.samplers.Sampler): - def __init__(self, loss_fn, optimizer, loss_callback=None, batch_size=1, grad_acc=1, total_steps=1, seed=0, training_dtype=torch.bfloat16): + def __init__( + self, + loss_fn, + optimizer, + loss_callback=None, + batch_size=1, + grad_acc=1, + total_steps=1, + seed=0, + training_dtype=torch.bfloat16, + real_dataset=None, + bucket_latents=None, + ): self.loss_fn = loss_fn self.optimizer = optimizer self.loss_callback = loss_callback @@ -65,54 +138,210 @@ class TrainSampler(comfy.samplers.Sampler): self.grad_acc = grad_acc self.seed = seed self.training_dtype = training_dtype + self.real_dataset: list[torch.Tensor] | None = real_dataset + # Bucket mode data + self.bucket_latents: list[torch.Tensor] | None = ( + bucket_latents # list of (Bi, C, Hi, Wi) + ) + # Precompute bucket offsets and weights for sampling + if bucket_latents is not None: + self._init_bucket_data(bucket_latents) + else: + self.bucket_offsets = None + self.bucket_weights = None + self.num_images = None - def sample(self, model_wrap, sigmas, extra_args, callback, noise, latent_image=None, denoise_mask=None, disable_pbar=False): + def _init_bucket_data(self, bucket_latents): + """Initialize bucket offsets and weights for sampling.""" + self.bucket_offsets = [0] + bucket_sizes = [] + for lat in bucket_latents: + bucket_sizes.append(lat.shape[0]) + self.bucket_offsets.append(self.bucket_offsets[-1] + lat.shape[0]) + self.num_images = self.bucket_offsets[-1] + # Weights for sampling buckets proportional to their size + self.bucket_weights = torch.tensor(bucket_sizes, dtype=torch.float32) + + def fwd_bwd( + self, + model_wrap, + batch_sigmas, + batch_noise, + batch_latent, + cond, + indicies, + extra_args, + dataset_size, + bwd=True, + ): + xt = model_wrap.inner_model.model_sampling.noise_scaling( + batch_sigmas, batch_noise, batch_latent, False + ) + x0 = model_wrap.inner_model.model_sampling.noise_scaling( + torch.zeros_like(batch_sigmas), + torch.zeros_like(batch_noise), + batch_latent, + False, + ) + + model_wrap.conds["positive"] = [cond[i] for i in indicies] + batch_extra_args = make_batch_extra_option_dict( + extra_args, indicies, full_size=dataset_size + ) + + with torch.autocast(xt.device.type, dtype=self.training_dtype): + x0_pred = model_wrap( + xt.requires_grad_(True), + batch_sigmas.requires_grad_(True), + **batch_extra_args, + ) + loss = self.loss_fn(x0_pred, x0) + if bwd: + bwd_loss = loss / self.grad_acc + bwd_loss.backward() + return loss + + def _generate_batch_sigmas(self, model_wrap, batch_size, device): + """Generate random sigma values for a batch.""" + batch_sigmas = [ + model_wrap.inner_model.model_sampling.percent_to_sigma( + torch.rand((1,)).item() + ) + for _ in range(batch_size) + ] + return torch.tensor(batch_sigmas).to(device) + + def _train_step_bucket_mode(self, model_wrap, cond, extra_args, noisegen, latent_image, pbar): + """Execute one training step in bucket mode.""" + # Sample bucket (weighted by size), then sample batch from bucket + bucket_idx = torch.multinomial(self.bucket_weights, 1).item() + bucket_latent = self.bucket_latents[bucket_idx] # (Bi, C, Hi, Wi) + bucket_size = bucket_latent.shape[0] + bucket_offset = self.bucket_offsets[bucket_idx] + + # Sample indices from this bucket (use all if bucket_size < batch_size) + actual_batch_size = min(self.batch_size, bucket_size) + relative_indices = torch.randperm(bucket_size)[:actual_batch_size].tolist() + # Convert to absolute indices for fwd_bwd (cond is flattened, use absolute index) + absolute_indices = [bucket_offset + idx for idx in relative_indices] + + batch_latent = bucket_latent[relative_indices].to(latent_image) # (actual_batch_size, C, H, W) + batch_noise = noisegen.generate_noise({"samples": batch_latent}).to( + batch_latent.device + ) + batch_sigmas = self._generate_batch_sigmas(model_wrap, actual_batch_size, batch_latent.device) + + loss = self.fwd_bwd( + model_wrap, + batch_sigmas, + batch_noise, + batch_latent, + cond, # Use flattened cond with absolute indices + absolute_indices, + extra_args, + self.num_images, + bwd=True, + ) + if self.loss_callback: + self.loss_callback(loss.item()) + pbar.set_postfix({"loss": f"{loss.item():.4f}", "bucket": bucket_idx}) + + def _train_step_standard_mode(self, model_wrap, cond, extra_args, noisegen, latent_image, dataset_size, pbar): + """Execute one training step in standard (non-bucket, non-multi-res) mode.""" + indicies = torch.randperm(dataset_size)[: self.batch_size].tolist() + batch_latent = torch.stack([latent_image[i] for i in indicies]) + batch_noise = noisegen.generate_noise({"samples": batch_latent}).to( + batch_latent.device + ) + batch_sigmas = self._generate_batch_sigmas(model_wrap, min(self.batch_size, dataset_size), batch_latent.device) + + loss = self.fwd_bwd( + model_wrap, + batch_sigmas, + batch_noise, + batch_latent, + cond, + indicies, + extra_args, + dataset_size, + bwd=True, + ) + if self.loss_callback: + self.loss_callback(loss.item()) + pbar.set_postfix({"loss": f"{loss.item():.4f}"}) + + def _train_step_multires_mode(self, model_wrap, cond, extra_args, noisegen, latent_image, dataset_size, pbar): + """Execute one training step in multi-resolution mode (real_dataset is set).""" + indicies = torch.randperm(dataset_size)[: self.batch_size].tolist() + total_loss = 0 + for index in indicies: + single_latent = self.real_dataset[index].to(latent_image) + batch_noise = noisegen.generate_noise( + {"samples": single_latent} + ).to(single_latent.device) + batch_sigmas = ( + model_wrap.inner_model.model_sampling.percent_to_sigma( + torch.rand((1,)).item() + ) + ) + batch_sigmas = torch.tensor([batch_sigmas]).to(single_latent.device) + loss = self.fwd_bwd( + model_wrap, + batch_sigmas, + batch_noise, + single_latent, + cond, + [index], + extra_args, + dataset_size, + bwd=False, + ) + total_loss += loss + total_loss = total_loss / self.grad_acc / len(indicies) + total_loss.backward() + if self.loss_callback: + self.loss_callback(total_loss.item()) + pbar.set_postfix({"loss": f"{total_loss.item():.4f}"}) + + def sample( + self, + model_wrap, + sigmas, + extra_args, + callback, + noise, + latent_image=None, + denoise_mask=None, + disable_pbar=False, + ): model_wrap.conds = process_cond_list(model_wrap.conds) cond = model_wrap.conds["positive"] dataset_size = sigmas.size(0) torch.cuda.empty_cache() - for i in (pbar:=tqdm.trange(self.total_steps, desc="Training LoRA", smoothing=0.01, disable=not comfy.utils.PROGRESS_BAR_ENABLED)): - noisegen = comfy_extras.nodes_custom_sampler.Noise_RandomNoise(self.seed + i * 1000) - indicies = torch.randperm(dataset_size)[:self.batch_size].tolist() - - batch_latent = torch.stack([latent_image[i] for i in indicies]) - batch_noise = noisegen.generate_noise({"samples": batch_latent}).to(batch_latent.device) - batch_sigmas = [ - model_wrap.inner_model.model_sampling.percent_to_sigma( - torch.rand((1,)).item() - ) for _ in range(min(self.batch_size, dataset_size)) - ] - batch_sigmas = torch.tensor(batch_sigmas).to(batch_latent.device) - - xt = model_wrap.inner_model.model_sampling.noise_scaling( - batch_sigmas, - batch_noise, - batch_latent, - False + ui_pbar = ProgressBar(self.total_steps) + for i in ( + pbar := trange( + self.total_steps, + desc="Training LoRA", + smoothing=0.01, + disable=not comfy.utils.PROGRESS_BAR_ENABLED, ) - x0 = model_wrap.inner_model.model_sampling.noise_scaling( - torch.zeros_like(batch_sigmas), - torch.zeros_like(batch_noise), - batch_latent, - False + ): + noisegen = comfy_extras.nodes_custom_sampler.Noise_RandomNoise( + self.seed + i * 1000 ) - model_wrap.conds["positive"] = [ - cond[i] for i in indicies - ] - batch_extra_args = make_batch_extra_option_dict(extra_args, indicies, full_size=dataset_size) + if self.bucket_latents is not None: + self._train_step_bucket_mode(model_wrap, cond, extra_args, noisegen, latent_image, pbar) + elif self.real_dataset is None: + self._train_step_standard_mode(model_wrap, cond, extra_args, noisegen, latent_image, dataset_size, pbar) + else: + self._train_step_multires_mode(model_wrap, cond, extra_args, noisegen, latent_image, dataset_size, pbar) - with torch.autocast(xt.device.type, dtype=self.training_dtype): - x0_pred = model_wrap(xt, batch_sigmas, **batch_extra_args) - loss = self.loss_fn(x0_pred, x0) - loss.backward() - if self.loss_callback: - self.loss_callback(loss.item()) - pbar.set_postfix({"loss": f"{loss.item():.4f}"}) - - if (i+1) % self.grad_acc == 0: + if (i + 1) % self.grad_acc == 0: self.optimizer.step() self.optimizer.zero_grad() + ui_pbar.update(1) torch.cuda.empty_cache() return torch.zeros_like(latent_image) @@ -134,233 +363,6 @@ class BiasDiff(torch.nn.Module): return self.passive_memory_usage() -def load_and_process_images(image_files, input_dir, resize_method="None", w=None, h=None): - """Utility function to load and process a list of images. - - Args: - image_files: List of image filenames - input_dir: Base directory containing the images - resize_method: How to handle images of different sizes ("None", "Stretch", "Crop", "Pad") - - Returns: - torch.Tensor: Batch of processed images - """ - if not image_files: - raise ValueError("No valid images found in input") - - output_images = [] - - for file in image_files: - image_path = os.path.join(input_dir, file) - img = node_helpers.pillow(Image.open, image_path) - - if img.mode == "I": - img = img.point(lambda i: i * (1 / 255)) - img = img.convert("RGB") - - if w is None and h is None: - w, h = img.size[0], img.size[1] - - # Resize image to first image - if img.size[0] != w or img.size[1] != h: - if resize_method == "Stretch": - img = img.resize((w, h), Image.Resampling.LANCZOS) - elif resize_method == "Crop": - img = img.crop((0, 0, w, h)) - elif resize_method == "Pad": - img = img.resize((w, h), Image.Resampling.LANCZOS) - elif resize_method == "None": - raise ValueError( - "Your input image size does not match the first image in the dataset. Either select a valid resize method or use the same size for all images." - ) - - img_array = np.array(img).astype(np.float32) / 255.0 - img_tensor = torch.from_numpy(img_array)[None,] - output_images.append(img_tensor) - - return torch.cat(output_images, dim=0) - - -class LoadImageSetNode: - @classmethod - def INPUT_TYPES(s): - return { - "required": { - "images": ( - [ - f - for f in os.listdir(folder_paths.get_input_directory()) - if f.endswith((".png", ".jpg", ".jpeg", ".webp", ".bmp", ".gif", ".jpe", ".apng", ".tif", ".tiff")) - ], - {"image_upload": True, "allow_batch": True}, - ) - }, - "optional": { - "resize_method": ( - ["None", "Stretch", "Crop", "Pad"], - {"default": "None"}, - ), - }, - } - - INPUT_IS_LIST = True - RETURN_TYPES = ("IMAGE",) - FUNCTION = "load_images" - CATEGORY = "loaders" - EXPERIMENTAL = True - DESCRIPTION = "Loads a batch of images from a directory for training." - - @classmethod - def VALIDATE_INPUTS(s, images, resize_method): - filenames = images[0] if isinstance(images[0], list) else images - - for image in filenames: - if not folder_paths.exists_annotated_filepath(image): - return "Invalid image file: {}".format(image) - return True - - def load_images(self, input_files, resize_method): - input_dir = folder_paths.get_input_directory() - valid_extensions = [".png", ".jpg", ".jpeg", ".webp", ".bmp", ".gif", ".jpe", ".apng", ".tif", ".tiff"] - image_files = [ - f - for f in input_files - if any(f.lower().endswith(ext) for ext in valid_extensions) - ] - output_tensor = load_and_process_images(image_files, input_dir, resize_method) - return (output_tensor,) - - -class LoadImageSetFromFolderNode: - @classmethod - def INPUT_TYPES(s): - return { - "required": { - "folder": (folder_paths.get_input_subfolders(), {"tooltip": "The folder to load images from."}) - }, - "optional": { - "resize_method": ( - ["None", "Stretch", "Crop", "Pad"], - {"default": "None"}, - ), - }, - } - - RETURN_TYPES = ("IMAGE",) - FUNCTION = "load_images" - CATEGORY = "loaders" - EXPERIMENTAL = True - DESCRIPTION = "Loads a batch of images from a directory for training." - - def load_images(self, folder, resize_method): - sub_input_dir = os.path.join(folder_paths.get_input_directory(), folder) - valid_extensions = [".png", ".jpg", ".jpeg", ".webp"] - image_files = [ - f - for f in os.listdir(sub_input_dir) - if any(f.lower().endswith(ext) for ext in valid_extensions) - ] - output_tensor = load_and_process_images(image_files, sub_input_dir, resize_method) - return (output_tensor,) - - -class LoadImageTextSetFromFolderNode: - @classmethod - def INPUT_TYPES(s): - return { - "required": { - "folder": (folder_paths.get_input_subfolders(), {"tooltip": "The folder to load images from."}), - "clip": (IO.CLIP, {"tooltip": "The CLIP model used for encoding the text."}), - }, - "optional": { - "resize_method": ( - ["None", "Stretch", "Crop", "Pad"], - {"default": "None"}, - ), - "width": ( - IO.INT, - { - "default": -1, - "min": -1, - "max": 10000, - "step": 1, - "tooltip": "The width to resize the images to. -1 means use the original width.", - }, - ), - "height": ( - IO.INT, - { - "default": -1, - "min": -1, - "max": 10000, - "step": 1, - "tooltip": "The height to resize the images to. -1 means use the original height.", - }, - ) - }, - } - - RETURN_TYPES = ("IMAGE", IO.CONDITIONING,) - FUNCTION = "load_images" - CATEGORY = "loaders" - EXPERIMENTAL = True - DESCRIPTION = "Loads a batch of images and caption from a directory for training." - - def load_images(self, folder, clip, resize_method, width=None, height=None): - if clip is None: - raise RuntimeError("ERROR: clip input is invalid: None\n\nIf the clip is from a checkpoint loader node your checkpoint does not contain a valid clip or text encoder model.") - - logging.info(f"Loading images from folder: {folder}") - - sub_input_dir = os.path.join(folder_paths.get_input_directory(), folder) - valid_extensions = [".png", ".jpg", ".jpeg", ".webp"] - - image_files = [] - for item in os.listdir(sub_input_dir): - path = os.path.join(sub_input_dir, item) - if any(item.lower().endswith(ext) for ext in valid_extensions): - image_files.append(path) - elif os.path.isdir(path): - # Support kohya-ss/sd-scripts folder structure - repeat = 1 - if item.split("_")[0].isdigit(): - repeat = int(item.split("_")[0]) - image_files.extend([ - os.path.join(path, f) for f in os.listdir(path) if any(f.lower().endswith(ext) for ext in valid_extensions) - ] * repeat) - - caption_file_path = [ - f.replace(os.path.splitext(f)[1], ".txt") - for f in image_files - ] - captions = [] - for caption_file in caption_file_path: - caption_path = os.path.join(sub_input_dir, caption_file) - if os.path.exists(caption_path): - with open(caption_path, "r", encoding="utf-8") as f: - caption = f.read().strip() - captions.append(caption) - else: - captions.append("") - - width = width if width != -1 else None - height = height if height != -1 else None - output_tensor = load_and_process_images(image_files, sub_input_dir, resize_method, width, height) - - logging.info(f"Loaded {len(output_tensor)} images from {sub_input_dir}.") - - logging.info(f"Encoding captions from {sub_input_dir}.") - conditions = [] - empty_cond = clip.encode_from_tokens_scheduled(clip.tokenize("")) - for text in captions: - if text == "": - conditions.append(empty_cond) - tokens = clip.tokenize(text) - conditions.extend(clip.encode_from_tokens_scheduled(tokens)) - logging.info(f"Encoded {len(conditions)} captions from {sub_input_dir}.") - return (output_tensor, conditions) - - def draw_loss_graph(loss_map, steps): width, height = 500, 300 img = Image.new("RGB", (width, height), "white") @@ -379,10 +381,14 @@ def draw_loss_graph(loss_map, steps): return img -def find_all_highest_child_module_with_forward(model: torch.nn.Module, result = None, name = None): +def find_all_highest_child_module_with_forward( + model: torch.nn.Module, result=None, name=None +): if result is None: result = [] - elif hasattr(model, "forward") and not isinstance(model, (torch.nn.ModuleList, torch.nn.Sequential, torch.nn.ModuleDict)): + elif hasattr(model, "forward") and not isinstance( + model, (torch.nn.ModuleList, torch.nn.Sequential, torch.nn.ModuleDict) + ): result.append(model) logging.debug(f"Found module with forward: {name} ({model.__class__.__name__})") return result @@ -396,12 +402,13 @@ def patch(m): if not hasattr(m, "forward"): return org_forward = m.forward + def fwd(args, kwargs): return org_forward(*args, **kwargs) + def checkpointing_fwd(*args, **kwargs): - return torch.utils.checkpoint.checkpoint( - fwd, args, kwargs, use_reentrant=False - ) + return torch.utils.checkpoint.checkpoint(fwd, args, kwargs, use_reentrant=False) + m.org_forward = org_forward m.forward = checkpointing_fwd @@ -412,130 +419,489 @@ def unpatch(m): del m.org_forward -class TrainLoraNode: +def _process_latents_bucket_mode(latents): + """Process latents for bucket mode training. + + Args: + latents: list[{"samples": tensor}] where each tensor is (Bi, C, Hi, Wi) + + Returns: + list of latent tensors + """ + bucket_latents = [] + for latent_dict in latents: + bucket_latents.append(latent_dict["samples"]) # (Bi, C, Hi, Wi) + return bucket_latents + + +def _process_latents_standard_mode(latents): + """Process latents for standard (non-bucket) mode training. + + Args: + latents: list of latent dicts or single latent dict + + Returns: + Processed latents (tensor or list of tensors) + """ + if len(latents) == 1: + return latents[0]["samples"] # Single latent dict + + latent_list = [] + for latent in latents: + latent = latent["samples"] + bs = latent.shape[0] + if bs != 1: + for sub_latent in latent: + latent_list.append(sub_latent[None]) + else: + latent_list.append(latent) + return latent_list + + +def _process_conditioning(positive): + """Process conditioning - either single list or list of lists. + + Args: + positive: list of conditioning + + Returns: + Flattened conditioning list + """ + if len(positive) == 1: + return positive[0] # Single conditioning list + + # Multiple conditioning lists - flatten + flat_positive = [] + for cond in positive: + if isinstance(cond, list): + flat_positive.extend(cond) + else: + flat_positive.append(cond) + return flat_positive + + +def _prepare_latents_and_count(latents, dtype, bucket_mode): + """Convert latents to dtype and compute image counts. + + Args: + latents: Latents (tensor, list of tensors, or bucket list) + dtype: Target dtype + bucket_mode: Whether bucket mode is enabled + + Returns: + tuple: (processed_latents, num_images, multi_res) + """ + if bucket_mode: + # In bucket mode, latents is list of tensors (Bi, C, Hi, Wi) + latents = [t.to(dtype) for t in latents] + num_buckets = len(latents) + num_images = sum(t.shape[0] for t in latents) + multi_res = False # Not using multi_res path in bucket mode + + logging.info(f"Bucket mode: {num_buckets} buckets, {num_images} total samples") + for i, lat in enumerate(latents): + logging.info(f" Bucket {i}: shape {lat.shape}") + return latents, num_images, multi_res + + # Non-bucket mode + if isinstance(latents, list): + all_shapes = set() + latents = [t.to(dtype) for t in latents] + for latent in latents: + all_shapes.add(latent.shape) + logging.info(f"Latent shapes: {all_shapes}") + if len(all_shapes) > 1: + multi_res = True + else: + multi_res = False + latents = torch.cat(latents, dim=0) + num_images = len(latents) + elif isinstance(latents, torch.Tensor): + latents = latents.to(dtype) + num_images = latents.shape[0] + multi_res = False + else: + logging.error(f"Invalid latents type: {type(latents)}") + num_images = 0 + multi_res = False + + return latents, num_images, multi_res + + +def _validate_and_expand_conditioning(positive, num_images, bucket_mode): + """Validate conditioning count matches image count, expand if needed. + + Args: + positive: Conditioning list + num_images: Number of images + bucket_mode: Whether bucket mode is enabled + + Returns: + Validated/expanded conditioning list + + Raises: + ValueError: If conditioning count doesn't match image count + """ + if bucket_mode: + return positive # Skip validation in bucket mode + + logging.info(f"Total Images: {num_images}, Total Captions: {len(positive)}") + if len(positive) == 1 and num_images > 1: + return positive * num_images + elif len(positive) != num_images: + raise ValueError( + f"Number of positive conditions ({len(positive)}) does not match number of images ({num_images})." + ) + return positive + + +def _load_existing_lora(existing_lora): + """Load existing LoRA weights if provided. + + Args: + existing_lora: LoRA filename or "[None]" + + Returns: + tuple: (existing_weights dict, existing_steps int) + """ + if existing_lora == "[None]": + return {}, 0 + + lora_path = folder_paths.get_full_path_or_raise("loras", existing_lora) + # Extract steps from filename like "trained_lora_10_steps_20250225_203716" + existing_steps = int(existing_lora.split("_steps_")[0].split("_")[-1]) + existing_weights = {} + if lora_path: + existing_weights = comfy.utils.load_torch_file(lora_path) + return existing_weights, existing_steps + + +def _create_weight_adapter( + module, module_name, existing_weights, algorithm, lora_dtype, rank +): + """Create a weight adapter for a module with weight. + + Args: + module: The module to create adapter for + module_name: Name of the module + existing_weights: Dict of existing LoRA weights + algorithm: Algorithm name for new adapters + lora_dtype: dtype for LoRA weights + rank: Rank for new LoRA adapters + + Returns: + tuple: (train_adapter, lora_params dict) + """ + key = f"{module_name}.weight" + shape = module.weight.shape + lora_params = {} + + if len(shape) >= 2: + alpha = float(existing_weights.get(f"{key}.alpha", 1.0)) + dora_scale = existing_weights.get(f"{key}.dora_scale", None) + + # Try to load existing adapter + existing_adapter = None + for adapter_cls in adapters: + existing_adapter = adapter_cls.load( + module_name, existing_weights, alpha, dora_scale + ) + if existing_adapter is not None: + break + + if existing_adapter is None: + adapter_cls = adapter_maps[algorithm] + + if existing_adapter is not None: + train_adapter = existing_adapter.to_train().to(lora_dtype) + else: + # Use LoRA with alpha=1.0 by default + train_adapter = adapter_cls.create_train( + module.weight, rank=rank, alpha=1.0 + ).to(lora_dtype) + + for name, parameter in train_adapter.named_parameters(): + lora_params[f"{module_name}.{name}"] = parameter + + return train_adapter.train().requires_grad_(True), lora_params + else: + # 1D weight - use BiasDiff + diff = torch.nn.Parameter( + torch.zeros(module.weight.shape, dtype=lora_dtype, requires_grad=True) + ) + diff_module = BiasDiff(diff).train().requires_grad_(True) + lora_params[f"{module_name}.diff"] = diff + return diff_module, lora_params + + +def _create_bias_adapter(module, module_name, lora_dtype): + """Create a bias adapter for a module with bias. + + Args: + module: The module with bias + module_name: Name of the module + lora_dtype: dtype for LoRA weights + + Returns: + tuple: (bias_module, lora_params dict) + """ + bias = torch.nn.Parameter( + torch.zeros(module.bias.shape, dtype=lora_dtype, requires_grad=True) + ) + bias_module = BiasDiff(bias).train().requires_grad_(True) + lora_params = {f"{module_name}.diff_b": bias} + return bias_module, lora_params + + +def _setup_lora_adapters(mp, existing_weights, algorithm, lora_dtype, rank): + """Setup all LoRA adapters on the model. + + Args: + mp: Model patcher + existing_weights: Dict of existing LoRA weights + algorithm: Algorithm name for new adapters + lora_dtype: dtype for LoRA weights + rank: Rank for new LoRA adapters + + Returns: + tuple: (lora_sd dict, all_weight_adapters list) + """ + lora_sd = {} + all_weight_adapters = [] + + for n, m in mp.model.named_modules(): + if hasattr(m, "weight_function"): + if m.weight is not None: + adapter, params = _create_weight_adapter( + m, n, existing_weights, algorithm, lora_dtype, rank + ) + lora_sd.update(params) + key = f"{n}.weight" + mp.add_weight_wrapper(key, adapter) + all_weight_adapters.append(adapter) + + if hasattr(m, "bias") and m.bias is not None: + bias_adapter, bias_params = _create_bias_adapter(m, n, lora_dtype) + lora_sd.update(bias_params) + key = f"{n}.bias" + mp.add_weight_wrapper(key, bias_adapter) + all_weight_adapters.append(bias_adapter) + + return lora_sd, all_weight_adapters + + +def _create_optimizer(optimizer_name, parameters, learning_rate): + """Create optimizer based on name. + + Args: + optimizer_name: Name of optimizer ("Adam", "AdamW", "SGD", "RMSprop") + parameters: Parameters to optimize + learning_rate: Learning rate + + Returns: + Optimizer instance + """ + if optimizer_name == "Adam": + return torch.optim.Adam(parameters, lr=learning_rate) + elif optimizer_name == "AdamW": + return torch.optim.AdamW(parameters, lr=learning_rate) + elif optimizer_name == "SGD": + return torch.optim.SGD(parameters, lr=learning_rate) + elif optimizer_name == "RMSprop": + return torch.optim.RMSprop(parameters, lr=learning_rate) + + +def _create_loss_function(loss_function_name): + """Create loss function based on name. + + Args: + loss_function_name: Name of loss function ("MSE", "L1", "Huber", "SmoothL1") + + Returns: + Loss function instance + """ + if loss_function_name == "MSE": + return torch.nn.MSELoss() + elif loss_function_name == "L1": + return torch.nn.L1Loss() + elif loss_function_name == "Huber": + return torch.nn.HuberLoss() + elif loss_function_name == "SmoothL1": + return torch.nn.SmoothL1Loss() + + +def _run_training_loop( + guider, train_sampler, latents, num_images, seed, bucket_mode, multi_res +): + """Execute the training loop. + + Args: + guider: The guider object + train_sampler: The training sampler + latents: Latent tensors + num_images: Number of images + seed: Random seed + bucket_mode: Whether bucket mode is enabled + multi_res: Whether multi-resolution mode is enabled + """ + sigmas = torch.tensor(range(num_images)) + noise = comfy_extras.nodes_custom_sampler.Noise_RandomNoise(seed) + + if bucket_mode: + # Use first bucket's first latent as dummy for guider + dummy_latent = latents[0][:1].repeat(num_images, 1, 1, 1) + guider.sample( + noise.generate_noise({"samples": dummy_latent}), + dummy_latent, + train_sampler, + sigmas, + seed=noise.seed, + ) + elif multi_res: + # use first latent as dummy latent if multi_res + latents = latents[0].repeat(num_images, 1, 1, 1) + guider.sample( + noise.generate_noise({"samples": latents}), + latents, + train_sampler, + sigmas, + seed=noise.seed, + ) + else: + guider.sample( + noise.generate_noise({"samples": latents}), + latents, + train_sampler, + sigmas, + seed=noise.seed, + ) + + +class TrainLoraNode(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "model": (IO.MODEL, {"tooltip": "The model to train the LoRA on."}), - "latents": ( - "LATENT", - { - "tooltip": "The Latents to use for training, serve as dataset/input of the model." - }, + def define_schema(cls): + return io.Schema( + node_id="TrainLoraNode", + display_name="Train LoRA", + category="training", + is_experimental=True, + is_input_list=True, # All inputs become lists + inputs=[ + io.Model.Input("model", tooltip="The model to train the LoRA on."), + io.Latent.Input( + "latents", + tooltip="The Latents to use for training, serve as dataset/input of the model.", ), - "positive": ( - IO.CONDITIONING, - {"tooltip": "The positive conditioning to use for training."}, + io.Conditioning.Input( + "positive", tooltip="The positive conditioning to use for training." ), - "batch_size": ( - IO.INT, - { - "default": 1, - "min": 1, - "max": 10000, - "step": 1, - "tooltip": "The batch size to use for training.", - }, + io.Int.Input( + "batch_size", + default=1, + min=1, + max=10000, + tooltip="The batch size to use for training.", ), - "grad_accumulation_steps": ( - IO.INT, - { - "default": 1, - "min": 1, - "max": 1024, - "step": 1, - "tooltip": "The number of gradient accumulation steps to use for training.", - } + io.Int.Input( + "grad_accumulation_steps", + default=1, + min=1, + max=1024, + tooltip="The number of gradient accumulation steps to use for training.", ), - "steps": ( - IO.INT, - { - "default": 16, - "min": 1, - "max": 100000, - "tooltip": "The number of steps to train the LoRA for.", - }, + io.Int.Input( + "steps", + default=16, + min=1, + max=100000, + tooltip="The number of steps to train the LoRA for.", ), - "learning_rate": ( - IO.FLOAT, - { - "default": 0.0005, - "min": 0.0000001, - "max": 1.0, - "step": 0.000001, - "tooltip": "The learning rate to use for training.", - }, + io.Float.Input( + "learning_rate", + default=0.0005, + min=0.0000001, + max=1.0, + step=0.0000001, + tooltip="The learning rate to use for training.", ), - "rank": ( - IO.INT, - { - "default": 8, - "min": 1, - "max": 128, - "tooltip": "The rank of the LoRA layers.", - }, + io.Int.Input( + "rank", + default=8, + min=1, + max=128, + tooltip="The rank of the LoRA layers.", ), - "optimizer": ( - ["AdamW", "Adam", "SGD", "RMSprop"], - { - "default": "AdamW", - "tooltip": "The optimizer to use for training.", - }, + io.Combo.Input( + "optimizer", + options=["AdamW", "Adam", "SGD", "RMSprop"], + default="AdamW", + tooltip="The optimizer to use for training.", ), - "loss_function": ( - ["MSE", "L1", "Huber", "SmoothL1"], - { - "default": "MSE", - "tooltip": "The loss function to use for training.", - }, + io.Combo.Input( + "loss_function", + options=["MSE", "L1", "Huber", "SmoothL1"], + default="MSE", + tooltip="The loss function to use for training.", ), - "seed": ( - IO.INT, - { - "default": 0, - "min": 0, - "max": 0xFFFFFFFFFFFFFFFF, - "tooltip": "The seed to use for training (used in generator for LoRA weight initialization and noise sampling)", - }, + io.Int.Input( + "seed", + default=0, + min=0, + max=0xFFFFFFFFFFFFFFFF, + tooltip="The seed to use for training (used in generator for LoRA weight initialization and noise sampling)", ), - "training_dtype": ( - ["bf16", "fp32"], - {"default": "bf16", "tooltip": "The dtype to use for training."}, + io.Combo.Input( + "training_dtype", + options=["bf16", "fp32"], + default="bf16", + tooltip="The dtype to use for training.", ), - "lora_dtype": ( - ["bf16", "fp32"], - {"default": "bf16", "tooltip": "The dtype to use for lora."}, + io.Combo.Input( + "lora_dtype", + options=["bf16", "fp32"], + default="bf16", + tooltip="The dtype to use for lora.", ), - "algorithm": ( - list(adapter_maps.keys()), - {"default": list(adapter_maps.keys())[0], "tooltip": "The algorithm to use for training."}, + io.Combo.Input( + "algorithm", + options=list(adapter_maps.keys()), + default=list(adapter_maps.keys())[0], + tooltip="The algorithm to use for training.", ), - "gradient_checkpointing": ( - IO.BOOLEAN, - { - "default": True, - "tooltip": "Use gradient checkpointing for training.", - } + io.Boolean.Input( + "gradient_checkpointing", + default=True, + tooltip="Use gradient checkpointing for training.", ), - "existing_lora": ( - folder_paths.get_filename_list("loras") + ["[None]"], - { - "default": "[None]", - "tooltip": "The existing LoRA to append to. Set to None for new LoRA.", - }, + io.Combo.Input( + "existing_lora", + options=folder_paths.get_filename_list("loras") + ["[None]"], + default="[None]", + tooltip="The existing LoRA to append to. Set to None for new LoRA.", ), - }, - } + io.Boolean.Input( + "bucket_mode", + default=False, + tooltip="Enable resolution bucket mode. When enabled, expects pre-bucketed latents from ResolutionBucket node.", + ), + ], + outputs=[ + io.Model.Output( + display_name="model", tooltip="Model with LoRA applied" + ), + io.Custom("LORA_MODEL").Output( + display_name="lora", tooltip="LoRA weights" + ), + io.Custom("LOSS_MAP").Output( + display_name="loss_map", tooltip="Loss history" + ), + io.Int.Output(display_name="steps", tooltip="Total training steps"), + ], + ) - RETURN_TYPES = (IO.MODEL, IO.LORA_MODEL, IO.LOSS_MAP, IO.INT) - RETURN_NAMES = ("model_with_lora", "lora", "loss", "steps") - FUNCTION = "train" - CATEGORY = "training" - EXPERIMENTAL = True - - def train( - self, + @classmethod + def execute( + cls, model, latents, positive, @@ -552,261 +918,252 @@ class TrainLoraNode: algorithm, gradient_checkpointing, existing_lora, + bucket_mode, ): + # Extract scalars from lists (due to is_input_list=True) + model = model[0] + batch_size = batch_size[0] + steps = steps[0] + grad_accumulation_steps = grad_accumulation_steps[0] + learning_rate = learning_rate[0] + rank = rank[0] + optimizer_name = optimizer[0] + loss_function_name = loss_function[0] + seed = seed[0] + training_dtype = training_dtype[0] + lora_dtype = lora_dtype[0] + algorithm = algorithm[0] + gradient_checkpointing = gradient_checkpointing[0] + existing_lora = existing_lora[0] + bucket_mode = bucket_mode[0] + + # Process latents based on mode + if bucket_mode: + latents = _process_latents_bucket_mode(latents) + else: + latents = _process_latents_standard_mode(latents) + + # Process conditioning + positive = _process_conditioning(positive) + + # Setup model and dtype mp = model.clone() dtype = node_helpers.string_to_torch_dtype(training_dtype) lora_dtype = node_helpers.string_to_torch_dtype(lora_dtype) mp.set_model_compute_dtype(dtype) - latents = latents["samples"].to(dtype) - num_images = latents.shape[0] - logging.info(f"Total Images: {num_images}, Total Captions: {len(positive)}") - if len(positive) == 1 and num_images > 1: - positive = positive * num_images - elif len(positive) != num_images: - raise ValueError( - f"Number of positive conditions ({len(positive)}) does not match number of images ({num_images})." - ) + # Prepare latents and compute counts + latents, num_images, multi_res = _prepare_latents_and_count( + latents, dtype, bucket_mode + ) + + # Validate and expand conditioning + positive = _validate_and_expand_conditioning(positive, num_images, bucket_mode) with torch.inference_mode(False): - lora_sd = {} - generator = torch.Generator() - generator.manual_seed(seed) + # Setup models for training + mp.model.requires_grad_(False) # Load existing LoRA weights if provided - existing_weights = {} - existing_steps = 0 - if existing_lora != "[None]": - lora_path = folder_paths.get_full_path_or_raise("loras", existing_lora) - # Extract steps from filename like "trained_lora_10_steps_20250225_203716" - existing_steps = int(existing_lora.split("_steps_")[0].split("_")[-1]) - if lora_path: - existing_weights = comfy.utils.load_torch_file(lora_path) + existing_weights, existing_steps = _load_existing_lora(existing_lora) - all_weight_adapters = [] - for n, m in mp.model.named_modules(): - if hasattr(m, "weight_function"): - if m.weight is not None: - key = "{}.weight".format(n) - shape = m.weight.shape - if len(shape) >= 2: - alpha = float(existing_weights.get(f"{key}.alpha", 1.0)) - dora_scale = existing_weights.get( - f"{key}.dora_scale", None - ) - for adapter_cls in adapters: - existing_adapter = adapter_cls.load( - n, existing_weights, alpha, dora_scale - ) - if existing_adapter is not None: - break - else: - existing_adapter = None - adapter_cls = adapter_maps[algorithm] + # Setup LoRA adapters + lora_sd, all_weight_adapters = _setup_lora_adapters( + mp, existing_weights, algorithm, lora_dtype, rank + ) - if existing_adapter is not None: - train_adapter = existing_adapter.to_train().to(lora_dtype) - else: - # Use LoRA with alpha=1.0 by default - train_adapter = adapter_cls.create_train( - m.weight, rank=rank, alpha=1.0 - ).to(lora_dtype) - for name, parameter in train_adapter.named_parameters(): - lora_sd[f"{n}.{name}"] = parameter + # Create optimizer and loss function + optimizer = _create_optimizer( + optimizer_name, lora_sd.values(), learning_rate + ) + criterion = _create_loss_function(loss_function_name) - mp.add_weight_wrapper(key, train_adapter) - all_weight_adapters.append(train_adapter) - else: - diff = torch.nn.Parameter( - torch.zeros( - m.weight.shape, dtype=lora_dtype, requires_grad=True - ) - ) - diff_module = BiasDiff(diff) - mp.add_weight_wrapper(key, BiasDiff(diff)) - all_weight_adapters.append(diff_module) - lora_sd["{}.diff".format(n)] = diff - if hasattr(m, "bias") and m.bias is not None: - key = "{}.bias".format(n) - bias = torch.nn.Parameter( - torch.zeros(m.bias.shape, dtype=lora_dtype, requires_grad=True) - ) - bias_module = BiasDiff(bias) - lora_sd["{}.diff_b".format(n)] = bias - mp.add_weight_wrapper(key, BiasDiff(bias)) - all_weight_adapters.append(bias_module) - - if optimizer == "Adam": - optimizer = torch.optim.Adam(lora_sd.values(), lr=learning_rate) - elif optimizer == "AdamW": - optimizer = torch.optim.AdamW(lora_sd.values(), lr=learning_rate) - elif optimizer == "SGD": - optimizer = torch.optim.SGD(lora_sd.values(), lr=learning_rate) - elif optimizer == "RMSprop": - optimizer = torch.optim.RMSprop(lora_sd.values(), lr=learning_rate) - - # Setup loss function based on selection - if loss_function == "MSE": - criterion = torch.nn.MSELoss() - elif loss_function == "L1": - criterion = torch.nn.L1Loss() - elif loss_function == "Huber": - criterion = torch.nn.HuberLoss() - elif loss_function == "SmoothL1": - criterion = torch.nn.SmoothL1Loss() - - # setup models + # Setup gradient checkpointing if gradient_checkpointing: - for m in find_all_highest_child_module_with_forward(mp.model.diffusion_model): + for m in find_all_highest_child_module_with_forward( + mp.model.diffusion_model + ): patch(m) - mp.model.requires_grad_(False) - comfy.model_management.load_models_gpu([mp], memory_required=1e20, force_full_load=True) - # Setup sampler and guider like in test script + torch.cuda.empty_cache() + # With force_full_load=False we should be able to have offloading + # But for offloading in training we need custom AutoGrad hooks for fwd/bwd + comfy.model_management.load_models_gpu( + [mp], memory_required=1e20, force_full_load=True + ) + torch.cuda.empty_cache() + + # Setup loss tracking loss_map = {"loss": []} + def loss_callback(loss): loss_map["loss"].append(loss) - train_sampler = TrainSampler( - criterion, - optimizer, - loss_callback=loss_callback, - batch_size=batch_size, - grad_acc=grad_accumulation_steps, - total_steps=steps*grad_accumulation_steps, - seed=seed, - training_dtype=dtype - ) - guider = comfy_extras.nodes_custom_sampler.Guider_Basic(mp) - guider.set_conds(positive) # Set conditioning from input - # Training loop + # Create sampler + if bucket_mode: + train_sampler = TrainSampler( + criterion, + optimizer, + loss_callback=loss_callback, + batch_size=batch_size, + grad_acc=grad_accumulation_steps, + total_steps=steps * grad_accumulation_steps, + seed=seed, + training_dtype=dtype, + bucket_latents=latents, + ) + else: + train_sampler = TrainSampler( + criterion, + optimizer, + loss_callback=loss_callback, + batch_size=batch_size, + grad_acc=grad_accumulation_steps, + total_steps=steps * grad_accumulation_steps, + seed=seed, + training_dtype=dtype, + real_dataset=latents if multi_res else None, + ) + + # Setup guider + guider = TrainGuider(mp) + guider.set_conds(positive) + + # Run training loop try: - # Generate dummy sigmas and noise - sigmas = torch.tensor(range(num_images)) - noise = comfy_extras.nodes_custom_sampler.Noise_RandomNoise(seed) - guider.sample( - noise.generate_noise({"samples": latents}), - latents, + _run_training_loop( + guider, train_sampler, - sigmas, - seed=noise.seed + latents, + num_images, + seed, + bucket_mode, + multi_res, ) finally: for m in mp.model.modules(): unpatch(m) del train_sampler, optimizer + # Finalize adapters for adapter in all_weight_adapters: adapter.requires_grad_(False) for param in lora_sd: lora_sd[param] = lora_sd[param].to(lora_dtype) - return (mp, lora_sd, loss_map, steps + existing_steps) + return io.NodeOutput(mp, lora_sd, loss_map, steps + existing_steps) -class LoraModelLoader: - def __init__(self): - self.loaded_lora = None +class LoraModelLoader(io.ComfyNode):# + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LoraModelLoader", + display_name="Load LoRA Model", + category="loaders", + is_experimental=True, + inputs=[ + io.Model.Input( + "model", tooltip="The diffusion model the LoRA will be applied to." + ), + io.Custom("LORA_MODEL").Input( + "lora", tooltip="The LoRA model to apply to the diffusion model." + ), + io.Float.Input( + "strength_model", + default=1.0, + min=-100.0, + max=100.0, + tooltip="How strongly to modify the diffusion model. This value can be negative.", + ), + ], + outputs=[ + io.Model.Output( + display_name="model", tooltip="The modified diffusion model." + ), + ], + ) @classmethod - def INPUT_TYPES(s): - return { - "required": { - "model": ("MODEL", {"tooltip": "The diffusion model the LoRA will be applied to."}), - "lora": (IO.LORA_MODEL, {"tooltip": "The LoRA model to apply to the diffusion model."}), - "strength_model": ("FLOAT", {"default": 1.0, "min": -100.0, "max": 100.0, "step": 0.01, "tooltip": "How strongly to modify the diffusion model. This value can be negative."}), - } - } - - RETURN_TYPES = ("MODEL",) - OUTPUT_TOOLTIPS = ("The modified diffusion model.",) - FUNCTION = "load_lora_model" - - CATEGORY = "loaders" - DESCRIPTION = "Load Trained LoRA weights from Train LoRA node." - EXPERIMENTAL = True - - def load_lora_model(self, model, lora, strength_model): + def execute(cls, model, lora, strength_model): if strength_model == 0: - return (model, ) + return io.NodeOutput(model) - model_lora, _ = comfy.sd.load_lora_for_models(model, None, lora, strength_model, 0) - return (model_lora, ) + model_lora, _ = comfy.sd.load_lora_for_models( + model, None, lora, strength_model, 0 + ) + return io.NodeOutput(model_lora) -class SaveLoRA: - def __init__(self): - self.output_dir = folder_paths.get_output_directory() +class SaveLoRA(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="SaveLoRA", + display_name="Save LoRA Weights", + category="loaders", + is_experimental=True, + is_output_node=True, + inputs=[ + io.Custom("LORA_MODEL").Input( + "lora", + tooltip="The LoRA model to save. Do not use the model with LoRA layers.", + ), + io.String.Input( + "prefix", + default="loras/ComfyUI_trained_lora", + tooltip="The prefix to use for the saved LoRA file.", + ), + io.Int.Input( + "steps", + optional=True, + tooltip="Optional: The number of steps to LoRA has been trained for, used to name the saved file.", + ), + ], + outputs=[], + ) @classmethod - def INPUT_TYPES(s): - return { - "required": { - "lora": ( - IO.LORA_MODEL, - { - "tooltip": "The LoRA model to save. Do not use the model with LoRA layers." - }, - ), - "prefix": ( - "STRING", - { - "default": "loras/ComfyUI_trained_lora", - "tooltip": "The prefix to use for the saved LoRA file.", - }, - ), - }, - "optional": { - "steps": ( - IO.INT, - { - "forceInput": True, - "tooltip": "Optional: The number of steps to LoRA has been trained for, used to name the saved file.", - }, - ), - }, - } - - RETURN_TYPES = () - FUNCTION = "save" - CATEGORY = "loaders" - EXPERIMENTAL = True - OUTPUT_NODE = True - - def save(self, lora, prefix, steps=None): - full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(prefix, self.output_dir) + def execute(cls, lora, prefix, steps=None): + output_dir = folder_paths.get_output_directory() + full_output_folder, filename, counter, subfolder, filename_prefix = ( + folder_paths.get_save_image_path(prefix, output_dir) + ) if steps is None: output_checkpoint = f"{filename}_{counter:05}_.safetensors" else: output_checkpoint = f"{filename}_{steps}_steps_{counter:05}_.safetensors" output_checkpoint = os.path.join(full_output_folder, output_checkpoint) safetensors.torch.save_file(lora, output_checkpoint) - return {} + return io.NodeOutput() -class LossGraphNode: - def __init__(self): - self.output_dir = folder_paths.get_temp_directory() +class LossGraphNode(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LossGraphNode", + display_name="Plot Loss Graph", + category="training", + is_experimental=True, + is_output_node=True, + inputs=[ + io.Custom("LOSS_MAP").Input( + "loss", tooltip="Loss map from training node." + ), + io.String.Input( + "filename_prefix", + default="loss_graph", + tooltip="Prefix for the saved loss graph image.", + ), + ], + outputs=[], + hidden=[io.Hidden.prompt, io.Hidden.extra_pnginfo], + ) @classmethod - def INPUT_TYPES(s): - return { - "required": { - "loss": (IO.LOSS_MAP, {"default": {}}), - "filename_prefix": (IO.STRING, {"default": "loss_graph"}), - }, - "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"}, - } - - RETURN_TYPES = () - FUNCTION = "plot_loss" - OUTPUT_NODE = True - CATEGORY = "training" - EXPERIMENTAL = True - DESCRIPTION = "Plots the loss graph and saves it to the output directory." - - def plot_loss(self, loss, filename_prefix, prompt=None, extra_pnginfo=None): + def execute(cls, loss, filename_prefix, prompt=None, extra_pnginfo=None): loss_values = loss["loss"] width, height = 800, 480 margin = 40 @@ -849,47 +1206,27 @@ class LossGraphNode: (margin - 30, height - 10), f"{min_loss:.2f}", font=font, fill="black" ) - metadata = None - if not args.disable_metadata: - metadata = PngInfo() - if prompt is not None: - metadata.add_text("prompt", json.dumps(prompt)) - if extra_pnginfo is not None: - for x in extra_pnginfo: - metadata.add_text(x, json.dumps(extra_pnginfo[x])) + # Convert PIL image to tensor for PreviewImage + img_array = np.array(img).astype(np.float32) / 255.0 + img_tensor = torch.from_numpy(img_array)[None,] # [1, H, W, 3] - date = datetime.datetime.now().strftime("%Y%m%d_%H%M%S") - img.save( - os.path.join(self.output_dir, f"{filename_prefix}_{date}.png"), - pnginfo=metadata, - ) - return { - "ui": { - "images": [ - { - "filename": f"{filename_prefix}_{date}.png", - "subfolder": "", - "type": "temp", - } - ] - } - } + # Return preview UI + return io.NodeOutput(ui=ui.PreviewImage(img_tensor, cls=cls)) -NODE_CLASS_MAPPINGS = { - "TrainLoraNode": TrainLoraNode, - "SaveLoRANode": SaveLoRA, - "LoraModelLoader": LoraModelLoader, - "LoadImageSetFromFolderNode": LoadImageSetFromFolderNode, - "LoadImageTextSetFromFolderNode": LoadImageTextSetFromFolderNode, - "LossGraphNode": LossGraphNode, -} +# ========== Extension Setup ========== -NODE_DISPLAY_NAME_MAPPINGS = { - "TrainLoraNode": "Train LoRA", - "SaveLoRANode": "Save LoRA Weights", - "LoraModelLoader": "Load LoRA Model", - "LoadImageSetFromFolderNode": "Load Image Dataset from Folder", - "LoadImageTextSetFromFolderNode": "Load Image and Text Dataset from Folder", - "LossGraphNode": "Plot Loss Graph", -} + +class TrainingExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + TrainLoraNode, + LoraModelLoader, + SaveLoRA, + LossGraphNode, + ] + + +async def comfy_entrypoint() -> TrainingExtension: + return TrainingExtension() diff --git a/comfy_extras/nodes_upscale_model.py b/comfy_extras/nodes_upscale_model.py index 4d62b87be..ed587851c 100644 --- a/comfy_extras/nodes_upscale_model.py +++ b/comfy_extras/nodes_upscale_model.py @@ -78,18 +78,20 @@ class ImageUpscaleWithModel(io.ComfyNode): overlap = 32 oom = True - while oom: - try: - steps = in_img.shape[0] * comfy.utils.get_tiled_scale_steps(in_img.shape[3], in_img.shape[2], tile_x=tile, tile_y=tile, overlap=overlap) - pbar = comfy.utils.ProgressBar(steps) - s = comfy.utils.tiled_scale(in_img, lambda a: upscale_model(a), tile_x=tile, tile_y=tile, overlap=overlap, upscale_amount=upscale_model.scale, pbar=pbar) - oom = False - except model_management.OOM_EXCEPTION as e: - tile //= 2 - if tile < 128: - raise e + try: + while oom: + try: + steps = in_img.shape[0] * comfy.utils.get_tiled_scale_steps(in_img.shape[3], in_img.shape[2], tile_x=tile, tile_y=tile, overlap=overlap) + pbar = comfy.utils.ProgressBar(steps) + s = comfy.utils.tiled_scale(in_img, lambda a: upscale_model(a), tile_x=tile, tile_y=tile, overlap=overlap, upscale_amount=upscale_model.scale, pbar=pbar) + oom = False + except model_management.OOM_EXCEPTION as e: + tile //= 2 + if tile < 128: + raise e + finally: + upscale_model.to("cpu") - upscale_model.to("cpu") s = torch.clamp(s.movedim(-3,-1), min=0, max=1.0) return io.NodeOutput(s) diff --git a/comfy_extras/nodes_video.py b/comfy_extras/nodes_video.py index 69fabb12e..c609e03da 100644 --- a/comfy_extras/nodes_video.py +++ b/comfy_extras/nodes_video.py @@ -8,10 +8,7 @@ import json from typing import Optional from typing_extensions import override from fractions import Fraction -from comfy_api.input import AudioInput, ImageInput, VideoInput -from comfy_api.input_impl import VideoFromComponents, VideoFromFile -from comfy_api.util import VideoCodec, VideoComponents, VideoContainer -from comfy_api.latest import ComfyExtension, io, ui +from comfy_api.latest import ComfyExtension, io, ui, Input, InputImpl, Types from comfy.cli_args import args class SaveWEBM(io.ComfyNode): @@ -28,7 +25,6 @@ class SaveWEBM(io.ComfyNode): io.Float.Input("fps", default=24.0, min=0.01, max=1000.0, step=0.01), io.Float.Input("crf", default=32.0, min=0, max=63.0, step=1, tooltip="Higher crf means lower quality with a smaller file size, lower crf means higher quality higher filesize."), ], - outputs=[], hidden=[io.Hidden.prompt, io.Hidden.extra_pnginfo], is_output_node=True, ) @@ -79,16 +75,15 @@ class SaveVideo(io.ComfyNode): inputs=[ io.Video.Input("video", tooltip="The video to save."), io.String.Input("filename_prefix", default="video/ComfyUI", tooltip="The prefix for the file to save. This may include formatting information such as %date:yyyy-MM-dd% or %Empty Latent Image.width% to include values from nodes."), - io.Combo.Input("format", options=VideoContainer.as_input(), default="auto", tooltip="The format to save the video as."), - io.Combo.Input("codec", options=VideoCodec.as_input(), default="auto", tooltip="The codec to use for the video."), + io.Combo.Input("format", options=Types.VideoContainer.as_input(), default="auto", tooltip="The format to save the video as."), + io.Combo.Input("codec", options=Types.VideoCodec.as_input(), default="auto", tooltip="The codec to use for the video."), ], - outputs=[], hidden=[io.Hidden.prompt, io.Hidden.extra_pnginfo], is_output_node=True, ) @classmethod - def execute(cls, video: VideoInput, filename_prefix, format, codec) -> io.NodeOutput: + def execute(cls, video: Input.Video, filename_prefix, format: str, codec) -> io.NodeOutput: width, height = video.get_dimensions() full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path( filename_prefix, @@ -105,10 +100,10 @@ class SaveVideo(io.ComfyNode): metadata["prompt"] = cls.hidden.prompt if len(metadata) > 0: saved_metadata = metadata - file = f"{filename}_{counter:05}_.{VideoContainer.get_extension(format)}" + file = f"{filename}_{counter:05}_.{Types.VideoContainer.get_extension(format)}" video.save_to( os.path.join(full_output_folder, file), - format=format, + format=Types.VideoContainer(format), codec=codec, metadata=saved_metadata ) @@ -135,9 +130,9 @@ class CreateVideo(io.ComfyNode): ) @classmethod - def execute(cls, images: ImageInput, fps: float, audio: Optional[AudioInput] = None) -> io.NodeOutput: + def execute(cls, images: Input.Image, fps: float, audio: Optional[Input.Audio] = None) -> io.NodeOutput: return io.NodeOutput( - VideoFromComponents(VideoComponents(images=images, audio=audio, frame_rate=Fraction(fps))) + InputImpl.VideoFromComponents(Types.VideoComponents(images=images, audio=audio, frame_rate=Fraction(fps))) ) class GetVideoComponents(io.ComfyNode): @@ -159,11 +154,11 @@ class GetVideoComponents(io.ComfyNode): ) @classmethod - def execute(cls, video: VideoInput) -> io.NodeOutput: + def execute(cls, video: Input.Video) -> io.NodeOutput: components = video.get_components() - return io.NodeOutput(components.images, components.audio, float(components.frame_rate)) + class LoadVideo(io.ComfyNode): @classmethod def define_schema(cls): @@ -185,7 +180,7 @@ class LoadVideo(io.ComfyNode): @classmethod def execute(cls, file) -> io.NodeOutput: video_path = folder_paths.get_annotated_filepath(file) - return io.NodeOutput(VideoFromFile(video_path)) + return io.NodeOutput(InputImpl.VideoFromFile(video_path)) @classmethod def fingerprint_inputs(s, file): diff --git a/comfy_extras/nodes_wan.py b/comfy_extras/nodes_wan.py index b0bd471bf..d32aad98e 100644 --- a/comfy_extras/nodes_wan.py +++ b/comfy_extras/nodes_wan.py @@ -817,7 +817,7 @@ def get_sample_indices(original_fps, if required_duration > total_frames / original_fps: raise ValueError("required_duration must be less than video length") - if not fixed_start is None and fixed_start >= 0: + if fixed_start is not None and fixed_start >= 0: start_frame = fixed_start else: max_start = total_frames - required_origin_frames diff --git a/comfy_extras/nodes_wanmove.py b/comfy_extras/nodes_wanmove.py new file mode 100644 index 000000000..5f39afa46 --- /dev/null +++ b/comfy_extras/nodes_wanmove.py @@ -0,0 +1,535 @@ +import nodes +import node_helpers +import torch +import torchvision.transforms.functional as TF +import comfy.model_management +import comfy.utils +import numpy as np +from typing_extensions import override +from comfy_api.latest import ComfyExtension, io +from comfy_extras.nodes_wan import parse_json_tracks + +# https://github.com/ali-vilab/Wan-Move/blob/main/wan/modules/trajectory.py +from PIL import Image, ImageDraw + +SKIP_ZERO = False + +def get_pos_emb( + pos_k: torch.Tensor, # A 1D tensor containing positions for which to generate embeddings. + pos_emb_dim: int, + theta_func: callable = lambda i, d: torch.pow(10000, torch.mul(2, torch.div(i.to(torch.float32), d))), #Function to compute thetas based on position and embedding dimensions. + device: torch.device = torch.device("cpu"), + dtype: torch.dtype = torch.float32, +) -> torch.Tensor: # The position embeddings (batch_size, pos_emb_dim) + + assert pos_emb_dim % 2 == 0, "The dimension of position embeddings must be even." + pos_k = pos_k.to(device, dtype) + if SKIP_ZERO: + pos_k = pos_k + 1 + batch_size = pos_k.size(0) + + denominator = torch.arange(0, pos_emb_dim // 2, device=device, dtype=dtype) + # Expand denominator to match the shape needed for broadcasting + denominator_expanded = denominator.view(1, -1).expand(batch_size, -1) + + thetas = theta_func(denominator_expanded, pos_emb_dim) + + # Ensure pos_k is in the correct shape for broadcasting + pos_k_expanded = pos_k.view(-1, 1).to(dtype) + sin_thetas = torch.sin(torch.div(pos_k_expanded, thetas)) + cos_thetas = torch.cos(torch.div(pos_k_expanded, thetas)) + + # Concatenate sine and cosine embeddings along the last dimension + pos_emb = torch.cat([sin_thetas, cos_thetas], dim=-1) + + return pos_emb + +def create_pos_embeddings( + pred_tracks: torch.Tensor, # the predicted tracks, [T, N, 2] + pred_visibility: torch.Tensor, # the predicted visibility [T, N] + downsample_ratios: list[int], # the ratios for downsampling time, height, and width + height: int, # the height of the feature map + width: int, # the width of the feature map + track_num: int = -1, # the number of tracks to use + t_down_strategy: str = "sample", # the strategy for downsampling time dimension +): + assert t_down_strategy in ["sample", "average"], "Invalid strategy for downsampling time dimension." + + t, n, _ = pred_tracks.shape + t_down, h_down, w_down = downsample_ratios + track_pos = - torch.ones(n, (t-1) // t_down + 1, 2, dtype=torch.long) + + if track_num == -1: + track_num = n + + tracks_idx = torch.randperm(n)[:track_num] + tracks = pred_tracks[:, tracks_idx] + visibility = pred_visibility[:, tracks_idx] + + for t_idx in range(0, t, t_down): + if t_down_strategy == "sample" or t_idx == 0: + cur_tracks = tracks[t_idx] # [N, 2] + cur_visibility = visibility[t_idx] # [N] + else: + cur_tracks = tracks[t_idx:t_idx+t_down].mean(dim=0) + cur_visibility = torch.any(visibility[t_idx:t_idx+t_down], dim=0) + + for i in range(track_num): + if not cur_visibility[i] or cur_tracks[i][0] < 0 or cur_tracks[i][1] < 0 or cur_tracks[i][0] >= width or cur_tracks[i][1] >= height: + continue + x, y = cur_tracks[i] + x, y = int(x // w_down), int(y // h_down) + track_pos[i, t_idx // t_down, 0], track_pos[i, t_idx // t_down, 1] = y, x + + return track_pos # the position embeddings, [N, T', 2], 2 = height, width + +def replace_feature( + vae_feature: torch.Tensor, # [B, C', T', H', W'] + track_pos: torch.Tensor, # [B, N, T', 2] + strength: float = 1.0 +) -> torch.Tensor: + b, _, t, h, w = vae_feature.shape + assert b == track_pos.shape[0], "Batch size mismatch." + n = track_pos.shape[1] + + # Shuffle the trajectory order + track_pos = track_pos[:, torch.randperm(n), :, :] + + # Extract coordinates at time steps ≥ 1 and generate a valid mask + current_pos = track_pos[:, :, 1:, :] # [B, N, T-1, 2] + mask = (current_pos[..., 0] >= 0) & (current_pos[..., 1] >= 0) # [B, N, T-1] + + # Get all valid indices + valid_indices = mask.nonzero(as_tuple=False) # [num_valid, 3] + num_valid = valid_indices.shape[0] + + if num_valid == 0: + return vae_feature + + # Decompose valid indices into each dimension + batch_idx = valid_indices[:, 0] + track_idx = valid_indices[:, 1] + t_rel = valid_indices[:, 2] + t_target = t_rel + 1 # Convert to original time step indices + + # Extract target position coordinates + h_target = current_pos[batch_idx, track_idx, t_rel, 0].long() # Ensure integer indices + w_target = current_pos[batch_idx, track_idx, t_rel, 1].long() + + # Extract source position coordinates (t=0) + h_source = track_pos[batch_idx, track_idx, 0, 0].long() + w_source = track_pos[batch_idx, track_idx, 0, 1].long() + + # Get source features and assign to target positions + src_features = vae_feature[batch_idx, :, 0, h_source, w_source] + dst_features = vae_feature[batch_idx, :, t_target, h_target, w_target] + + vae_feature[batch_idx, :, t_target, h_target, w_target] = dst_features + (src_features - dst_features) * strength + + + return vae_feature + +# Visualize functions + +def _draw_gradient_polyline_on_overlay(overlay, line_width, points, start_color, opacity=1.0): + draw = ImageDraw.Draw(overlay, 'RGBA') + points = points[::-1] + + # Compute total length + total_length = 0 + segment_lengths = [] + for i in range(len(points) - 1): + dx = points[i + 1][0] - points[i][0] + dy = points[i + 1][1] - points[i][1] + length = (dx * dx + dy * dy) ** 0.5 + segment_lengths.append(length) + total_length += length + + if total_length == 0: + return + + accumulated_length = 0 + + # Draw the gradient polyline + for idx, (start_point, end_point) in enumerate(zip(points[:-1], points[1:])): + segment_length = segment_lengths[idx] + steps = max(int(segment_length), 1) + + for i in range(steps): + current_length = accumulated_length + (i / steps) * segment_length + ratio = current_length / total_length + + alpha = int(255 * (1 - ratio) * opacity) + color = (*start_color, alpha) + + x = int(start_point[0] + (end_point[0] - start_point[0]) * i / steps) + y = int(start_point[1] + (end_point[1] - start_point[1]) * i / steps) + + dynamic_line_width = max(int(line_width * (1 - ratio)), 1) + draw.line([(x, y), (x + 1, y)], fill=color, width=dynamic_line_width) + + accumulated_length += segment_length + + +def add_weighted(rgb, track): + rgb = np.array(rgb) # [H, W, C] "RGB" + track = np.array(track) # [H, W, C] "RGBA" + + alpha = track[:, :, 3] / 255.0 + alpha = np.stack([alpha] * 3, axis=-1) + blend_img = track[:, :, :3] * alpha + rgb * (1 - alpha) + + return Image.fromarray(blend_img.astype(np.uint8)) + +def draw_tracks_on_video(video, tracks, visibility=None, track_frame=24, circle_size=12, opacity=0.5, line_width=16): + color_map = [(102, 153, 255), (0, 255, 255), (255, 255, 0), (255, 102, 204), (0, 255, 0)] + + video = video.byte().cpu().numpy() # (81, 480, 832, 3) + tracks = tracks[0].long().detach().cpu().numpy() + if visibility is not None: + visibility = visibility[0].detach().cpu().numpy() + + num_frames, height, width = video.shape[:3] + num_tracks = tracks.shape[1] + alpha_opacity = int(255 * opacity) + + output_frames = [] + for t in range(num_frames): + frame_rgb = video[t].astype(np.float32) + + # Create a single RGBA overlay for all tracks in this frame + overlay = Image.new("RGBA", (width, height), (0, 0, 0, 0)) + draw_overlay = ImageDraw.Draw(overlay) + + polyline_data = [] + + # Draw all circles on a single overlay + for n in range(num_tracks): + if visibility is not None and visibility[t, n] == 0: + continue + + track_coord = tracks[t, n] + color = color_map[n % len(color_map)] + circle_color = color + (alpha_opacity,) + + draw_overlay.ellipse((track_coord[0] - circle_size, track_coord[1] - circle_size, track_coord[0] + circle_size, track_coord[1] + circle_size), + fill=circle_color + ) + + # Store polyline data for batch processing + tracks_coord = tracks[max(t - track_frame, 0):t + 1, n] + if len(tracks_coord) > 1: + polyline_data.append((tracks_coord, color)) + + # Blend circles overlay once + overlay_np = np.array(overlay) + alpha = overlay_np[:, :, 3:4] / 255.0 + frame_rgb = overlay_np[:, :, :3] * alpha + frame_rgb * (1 - alpha) + + # Draw all polylines on a single overlay + if polyline_data: + polyline_overlay = Image.new("RGBA", (width, height), (0, 0, 0, 0)) + for tracks_coord, color in polyline_data: + _draw_gradient_polyline_on_overlay(polyline_overlay, line_width, tracks_coord, color, opacity) + + # Blend polylines overlay once + polyline_np = np.array(polyline_overlay) + alpha = polyline_np[:, :, 3:4] / 255.0 + frame_rgb = polyline_np[:, :, :3] * alpha + frame_rgb * (1 - alpha) + + output_frames.append(Image.fromarray(frame_rgb.astype(np.uint8))) + + return output_frames + + +class WanMoveVisualizeTracks(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="WanMoveVisualizeTracks", + category="conditioning/video_models", + inputs=[ + io.Image.Input("images"), + io.Tracks.Input("tracks", optional=True), + io.Int.Input("line_resolution", default=24, min=1, max=1024), + io.Int.Input("circle_size", default=12, min=1, max=128), + io.Float.Input("opacity", default=0.75, min=0.0, max=1.0, step=0.01), + io.Int.Input("line_width", default=16, min=1, max=128), + ], + outputs=[ + io.Image.Output(), + ], + ) + + @classmethod + def execute(cls, images, line_resolution, circle_size, opacity, line_width, tracks=None) -> io.NodeOutput: + if tracks is None: + return io.NodeOutput(images) + + track_path = tracks["track_path"].unsqueeze(0) + track_visibility = tracks["track_visibility"].unsqueeze(0) + images_in = images * 255.0 + if images_in.shape[0] != track_path.shape[1]: + repeat_count = track_path.shape[1] // images.shape[0] + images_in = images_in.repeat(repeat_count, 1, 1, 1) + track_video = draw_tracks_on_video(images_in, track_path, track_visibility, track_frame=line_resolution, circle_size=circle_size, opacity=opacity, line_width=line_width) + track_video = torch.stack([TF.to_tensor(frame) for frame in track_video], dim=0).movedim(1, -1).float() + + return io.NodeOutput(track_video.to(comfy.model_management.intermediate_device())) + + +class WanMoveTracksFromCoords(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="WanMoveTracksFromCoords", + category="conditioning/video_models", + inputs=[ + io.String.Input("track_coords", force_input=True, default="[]", optional=True), + io.Mask.Input("track_mask", optional=True), + ], + outputs=[ + io.Tracks.Output(), + io.Int.Output(display_name="track_length"), + ], + ) + + @classmethod + def execute(cls, track_coords, track_mask=None) -> io.NodeOutput: + device=comfy.model_management.intermediate_device() + + tracks_data = parse_json_tracks(track_coords) + track_length = len(tracks_data[0]) + + track_list = [ + [[track[frame]['x'], track[frame]['y']] for track in tracks_data] + for frame in range(len(tracks_data[0])) + ] + tracks = torch.tensor(track_list, dtype=torch.float32, device=device) # [frames, num_tracks, 2] + + num_tracks = tracks.shape[-2] + if track_mask is None: + track_visibility = torch.ones((track_length, num_tracks), dtype=torch.bool, device=device) + else: + track_visibility = (track_mask > 0).any(dim=(1, 2)).unsqueeze(-1) + + out_track_info = {} + out_track_info["track_path"] = tracks + out_track_info["track_visibility"] = track_visibility + return io.NodeOutput(out_track_info, track_length) + + +class GenerateTracks(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="GenerateTracks", + category="conditioning/video_models", + inputs=[ + io.Int.Input("width", default=832, min=16, max=4096, step=16), + io.Int.Input("height", default=480, min=16, max=4096, step=16), + io.Float.Input("start_x", default=0.0, min=0.0, max=1.0, step=0.01, tooltip="Normalized X coordinate (0-1) for start position."), + io.Float.Input("start_y", default=0.0, min=0.0, max=1.0, step=0.01, tooltip="Normalized Y coordinate (0-1) for start position."), + io.Float.Input("end_x", default=1.0, min=0.0, max=1.0, step=0.01, tooltip="Normalized X coordinate (0-1) for end position."), + io.Float.Input("end_y", default=1.0, min=0.0, max=1.0, step=0.01, tooltip="Normalized Y coordinate (0-1) for end position."), + io.Int.Input("num_frames", default=81, min=1, max=1024), + io.Int.Input("num_tracks", default=5, min=1, max=100), + io.Float.Input("track_spread", default=0.025, min=0.0, max=1.0, step=0.001, tooltip="Normalized distance between tracks. Tracks are spread perpendicular to the motion direction."), + io.Boolean.Input("bezier", default=False, tooltip="Enable Bezier curve path using the mid point as control point."), + io.Float.Input("mid_x", default=0.5, min=0.0, max=1.0, step=0.01, tooltip="Normalized X control point for Bezier curve. Only used when 'bezier' is enabled."), + io.Float.Input("mid_y", default=0.5, min=0.0, max=1.0, step=0.01, tooltip="Normalized Y control point for Bezier curve. Only used when 'bezier' is enabled."), + io.Combo.Input( + "interpolation", + options=["linear", "ease_in", "ease_out", "ease_in_out", "constant"], + tooltip="Controls the timing/speed of movement along the path.", + ), + io.Mask.Input("track_mask", optional=True, tooltip="Optional mask to indicate visible frames."), + ], + outputs=[ + io.Tracks.Output(), + io.Int.Output(display_name="track_length"), + ], + ) + + @classmethod + def execute(cls, width, height, start_x, start_y, mid_x, mid_y, end_x, end_y, num_frames, num_tracks, + track_spread, bezier=False, interpolation="linear", track_mask=None) -> io.NodeOutput: + device = comfy.model_management.intermediate_device() + track_length = num_frames + + # normalized coordinates to pixel coordinates + start_x_px = start_x * width + start_y_px = start_y * height + mid_x_px = mid_x * width + mid_y_px = mid_y * height + end_x_px = end_x * width + end_y_px = end_y * height + + track_spread_px = track_spread * (width + height) / 2 # Use average of width/height for spread to keep it proportional + + t = torch.linspace(0, 1, num_frames, device=device) + if interpolation == "constant": # All points stay at start position + interp_values = torch.zeros_like(t) + elif interpolation == "linear": + interp_values = t + elif interpolation == "ease_in": + interp_values = t ** 2 + elif interpolation == "ease_out": + interp_values = 1 - (1 - t) ** 2 + elif interpolation == "ease_in_out": + interp_values = t * t * (3 - 2 * t) + + if bezier: # apply interpolation to t for timing control along the bezier path + t_interp = interp_values + one_minus_t = 1 - t_interp + x_positions = one_minus_t ** 2 * start_x_px + 2 * one_minus_t * t_interp * mid_x_px + t_interp ** 2 * end_x_px + y_positions = one_minus_t ** 2 * start_y_px + 2 * one_minus_t * t_interp * mid_y_px + t_interp ** 2 * end_y_px + tangent_x = 2 * one_minus_t * (mid_x_px - start_x_px) + 2 * t_interp * (end_x_px - mid_x_px) + tangent_y = 2 * one_minus_t * (mid_y_px - start_y_px) + 2 * t_interp * (end_y_px - mid_y_px) + else: # calculate base x and y positions for each frame (center track) + x_positions = start_x_px + (end_x_px - start_x_px) * interp_values + y_positions = start_y_px + (end_y_px - start_y_px) * interp_values + # For non-bezier, tangent is constant (direction from start to end) + tangent_x = torch.full_like(t, end_x_px - start_x_px) + tangent_y = torch.full_like(t, end_y_px - start_y_px) + + track_list = [] + for frame_idx in range(num_frames): + # Calculate perpendicular direction at this frame + tx = tangent_x[frame_idx].item() + ty = tangent_y[frame_idx].item() + length = (tx ** 2 + ty ** 2) ** 0.5 + + if length > 0: # Perpendicular unit vector (rotate 90 degrees) + perp_x = -ty / length + perp_y = tx / length + else: # If tangent is zero, spread horizontally + perp_x = 1.0 + perp_y = 0.0 + + frame_tracks = [] + for track_idx in range(num_tracks): # center tracks around the main path offset ranges from -(num_tracks-1)/2 to +(num_tracks-1)/2 + offset = (track_idx - (num_tracks - 1) / 2) * track_spread_px + track_x = x_positions[frame_idx].item() + perp_x * offset + track_y = y_positions[frame_idx].item() + perp_y * offset + frame_tracks.append([track_x, track_y]) + track_list.append(frame_tracks) + + tracks = torch.tensor(track_list, dtype=torch.float32, device=device) # [frames, num_tracks, 2] + + if track_mask is None: + track_visibility = torch.ones((track_length, num_tracks), dtype=torch.bool, device=device) + else: + track_visibility = (track_mask > 0).any(dim=(1, 2)).unsqueeze(-1) + + out_track_info = {} + out_track_info["track_path"] = tracks + out_track_info["track_visibility"] = track_visibility + return io.NodeOutput(out_track_info, track_length) + + +class WanMoveConcatTrack(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="WanMoveConcatTrack", + category="conditioning/video_models", + inputs=[ + io.Tracks.Input("tracks_1"), + io.Tracks.Input("tracks_2", optional=True), + ], + outputs=[ + io.Tracks.Output(), + ], + ) + + @classmethod + def execute(cls, tracks_1=None, tracks_2=None) -> io.NodeOutput: + if tracks_2 is None: + return io.NodeOutput(tracks_1) + + tracks_out = torch.cat([tracks_1["track_path"], tracks_2["track_path"]], dim=1) # Concatenate along the track dimension + mask_out = torch.cat([tracks_1["track_visibility"], tracks_2["track_visibility"]], dim=-1) + + out_track_info = {} + out_track_info["track_path"] = tracks_out + out_track_info["track_visibility"] = mask_out + return io.NodeOutput(out_track_info) + + +class WanMoveTrackToVideo(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="WanMoveTrackToVideo", + category="conditioning/video_models", + inputs=[ + io.Conditioning.Input("positive"), + io.Conditioning.Input("negative"), + io.Vae.Input("vae"), + io.Tracks.Input("tracks", optional=True), + io.Float.Input("strength", default=1.0, min=0.0, max=100.0, step=0.01, tooltip="Strength of the track conditioning."), + io.Int.Input("width", default=832, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4), + io.Int.Input("batch_size", default=1, min=1, max=4096), + io.Image.Input("start_image"), + io.ClipVisionOutput.Input("clip_vision_output", optional=True), + ], + outputs=[ + io.Conditioning.Output(display_name="positive"), + io.Conditioning.Output(display_name="negative"), + io.Latent.Output(display_name="latent"), + ], + ) + + @classmethod + def execute(cls, positive, negative, vae, width, height, length, batch_size, strength, tracks=None, start_image=None, clip_vision_output=None) -> io.NodeOutput: + device=comfy.model_management.intermediate_device() + latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=device) + if start_image is not None: + start_image = comfy.utils.common_upscale(start_image[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1) + image = torch.ones((length, height, width, start_image.shape[-1]), device=start_image.device, dtype=start_image.dtype) * 0.5 + image[:start_image.shape[0]] = start_image + + concat_latent_image = vae.encode(image[:, :, :, :3]) + mask = torch.ones((1, 1, latent.shape[2], concat_latent_image.shape[-2], concat_latent_image.shape[-1]), device=start_image.device, dtype=start_image.dtype) + mask[:, :, :((start_image.shape[0] - 1) // 4) + 1] = 0.0 + + if tracks is not None and strength > 0.0: + tracks_path = tracks["track_path"][:length] # [T, N, 2] + num_tracks = tracks_path.shape[-2] + + track_visibility = tracks.get("track_visibility", torch.ones((length, num_tracks), dtype=torch.bool, device=device)) + + track_pos = create_pos_embeddings(tracks_path, track_visibility, [4, 8, 8], height, width, track_num=num_tracks) + track_pos = comfy.utils.resize_to_batch_size(track_pos.unsqueeze(0), batch_size) + concat_latent_image_pos = replace_feature(concat_latent_image, track_pos, strength) + else: + concat_latent_image_pos = concat_latent_image + + positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": concat_latent_image_pos, "concat_mask": mask}) + negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": concat_latent_image, "concat_mask": mask}) + + if clip_vision_output is not None: + positive = node_helpers.conditioning_set_values(positive, {"clip_vision_output": clip_vision_output}) + negative = node_helpers.conditioning_set_values(negative, {"clip_vision_output": clip_vision_output}) + + out_latent = {} + out_latent["samples"] = latent + return io.NodeOutput(positive, negative, out_latent) + + +class WanMoveExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + WanMoveTrackToVideo, + WanMoveTracksFromCoords, + WanMoveConcatTrack, + WanMoveVisualizeTracks, + GenerateTracks, + ] + +async def comfy_entrypoint() -> WanMoveExtension: + return WanMoveExtension() diff --git a/comfyui_version.py b/comfyui_version.py index 1e554eb9f..750673f08 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.3.69" +__version__ = "0.8.0" diff --git a/cuda_malloc.py b/cuda_malloc.py index 6520d5123..ee2bc4b69 100644 --- a/cuda_malloc.py +++ b/cuda_malloc.py @@ -63,18 +63,22 @@ def cuda_malloc_supported(): return True +version = "" + +try: + torch_spec = importlib.util.find_spec("torch") + for folder in torch_spec.submodule_search_locations: + ver_file = os.path.join(folder, "version.py") + if os.path.isfile(ver_file): + spec = importlib.util.spec_from_file_location("torch_version_import", ver_file) + module = importlib.util.module_from_spec(spec) + spec.loader.exec_module(module) + version = module.__version__ +except: + pass + if not args.cuda_malloc: try: - version = "" - torch_spec = importlib.util.find_spec("torch") - for folder in torch_spec.submodule_search_locations: - ver_file = os.path.join(folder, "version.py") - if os.path.isfile(ver_file): - spec = importlib.util.spec_from_file_location("torch_version_import", ver_file) - module = importlib.util.module_from_spec(spec) - spec.loader.exec_module(module) - version = module.__version__ - if int(version[0]) >= 2 and "+cu" in version: # enable by default for torch version 2.0 and up only on cuda torch if PerformanceFeature.AutoTune not in args.fast: # Autotune has issues with cuda malloc args.cuda_malloc = cuda_malloc_supported() @@ -90,3 +94,6 @@ if args.cuda_malloc and not args.disable_cuda_malloc: env_var += ",backend:cudaMallocAsync" os.environ['PYTORCH_CUDA_ALLOC_CONF'] = env_var + +def get_torch_version_noimport(): + return str(version) diff --git a/execution.py b/execution.py index 17c77beab..648f204ec 100644 --- a/execution.py +++ b/execution.py @@ -13,6 +13,7 @@ import asyncio import torch import comfy.model_management +from latent_preview import set_preview_method import nodes from comfy_execution.caching import ( BasicCache, @@ -34,7 +35,7 @@ from comfy_execution.validation import validate_node_input from comfy_execution.progress import get_progress_state, reset_progress_state, add_progress_handler, WebUIProgressHandler from comfy_execution.utils import CurrentNodeContext from comfy_api.internal import _ComfyNodeInternal, _NodeOutputInternal, first_real_override, is_class, make_locked_method_func -from comfy_api.latest import io +from comfy_api.latest import io, _io class ExecutionResult(Enum): @@ -76,9 +77,9 @@ class IsChangedCache: return self.is_changed[node_id] # Intentionally do not use cached outputs here. We only want constants in IS_CHANGED - input_data_all, _, hidden_inputs = get_input_data(node["inputs"], class_def, node_id, None) + input_data_all, _, v3_data = get_input_data(node["inputs"], class_def, node_id, None) try: - is_changed = await _async_map_node_over_list(self.prompt_id, node_id, class_def, input_data_all, is_changed_name) + is_changed = await _async_map_node_over_list(self.prompt_id, node_id, class_def, input_data_all, is_changed_name, v3_data=v3_data) is_changed = await resolve_map_node_over_list_results(is_changed) node["is_changed"] = [None if isinstance(x, ExecutionBlocker) else x for x in is_changed] except Exception as e: @@ -146,13 +147,13 @@ SENSITIVE_EXTRA_DATA_KEYS = ("auth_token_comfy_org", "api_key_comfy_org") def get_input_data(inputs, class_def, unique_id, execution_list=None, dynprompt=None, extra_data={}): is_v3 = issubclass(class_def, _ComfyNodeInternal) + v3_data: io.V3Data = {} + hidden_inputs_v3 = {} + valid_inputs = class_def.INPUT_TYPES() if is_v3: - valid_inputs, schema = class_def.INPUT_TYPES(include_hidden=False, return_schema=True) - else: - valid_inputs = class_def.INPUT_TYPES() + valid_inputs, hidden, v3_data = _io.get_finalized_class_inputs(valid_inputs, inputs) input_data_all = {} missing_keys = {} - hidden_inputs_v3 = {} for x in inputs: input_data = inputs[x] _, input_category, input_info = get_input_info(class_def, x, valid_inputs) @@ -178,18 +179,18 @@ def get_input_data(inputs, class_def, unique_id, execution_list=None, dynprompt= input_data_all[x] = [input_data] if is_v3: - if schema.hidden: - if io.Hidden.prompt in schema.hidden: + if hidden is not None: + if io.Hidden.prompt.name in hidden: hidden_inputs_v3[io.Hidden.prompt] = dynprompt.get_original_prompt() if dynprompt is not None else {} - if io.Hidden.dynprompt in schema.hidden: + if io.Hidden.dynprompt.name in hidden: hidden_inputs_v3[io.Hidden.dynprompt] = dynprompt - if io.Hidden.extra_pnginfo in schema.hidden: + if io.Hidden.extra_pnginfo.name in hidden: hidden_inputs_v3[io.Hidden.extra_pnginfo] = extra_data.get('extra_pnginfo', None) - if io.Hidden.unique_id in schema.hidden: + if io.Hidden.unique_id.name in hidden: hidden_inputs_v3[io.Hidden.unique_id] = unique_id - if io.Hidden.auth_token_comfy_org in schema.hidden: + if io.Hidden.auth_token_comfy_org.name in hidden: hidden_inputs_v3[io.Hidden.auth_token_comfy_org] = extra_data.get("auth_token_comfy_org", None) - if io.Hidden.api_key_comfy_org in schema.hidden: + if io.Hidden.api_key_comfy_org.name in hidden: hidden_inputs_v3[io.Hidden.api_key_comfy_org] = extra_data.get("api_key_comfy_org", None) else: if "hidden" in valid_inputs: @@ -207,7 +208,8 @@ def get_input_data(inputs, class_def, unique_id, execution_list=None, dynprompt= input_data_all[x] = [extra_data.get("auth_token_comfy_org", None)] if h[x] == "API_KEY_COMFY_ORG": input_data_all[x] = [extra_data.get("api_key_comfy_org", None)] - return input_data_all, missing_keys, hidden_inputs_v3 + v3_data["hidden_inputs"] = hidden_inputs_v3 + return input_data_all, missing_keys, v3_data map_node_over_list = None #Don't hook this please @@ -223,7 +225,7 @@ async def resolve_map_node_over_list_results(results): raise exc return [x.result() if isinstance(x, asyncio.Task) else x for x in results] -async def _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, func, allow_interrupt=False, execution_block_cb=None, pre_execute_cb=None, hidden_inputs=None): +async def _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, func, allow_interrupt=False, execution_block_cb=None, pre_execute_cb=None, v3_data=None): # check if node wants the lists input_is_list = getattr(obj, "INPUT_IS_LIST", False) @@ -255,17 +257,20 @@ async def _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, f pre_execute_cb(index) # V3 if isinstance(obj, _ComfyNodeInternal) or (is_class(obj) and issubclass(obj, _ComfyNodeInternal)): - # if is just a class, then assign no resources or state, just create clone + # if is just a class, then assign no state, just create clone if is_class(obj): type_obj = obj obj.VALIDATE_CLASS() - class_clone = obj.PREPARE_CLASS_CLONE(hidden_inputs) + class_clone = obj.PREPARE_CLASS_CLONE(v3_data) # otherwise, use class instance to populate/reuse some fields else: type_obj = type(obj) type_obj.VALIDATE_CLASS() - class_clone = type_obj.PREPARE_CLASS_CLONE(hidden_inputs) + class_clone = type_obj.PREPARE_CLASS_CLONE(v3_data) f = make_locked_method_func(type_obj, func, class_clone) + # in case of dynamic inputs, restructure inputs to expected nested dict + if v3_data is not None: + inputs = _io.build_nested_inputs(inputs, v3_data) # V1 else: f = getattr(obj, func) @@ -320,8 +325,8 @@ def merge_result_data(results, obj): output.append([o[i] for o in results]) return output -async def get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=None, pre_execute_cb=None, hidden_inputs=None): - return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, hidden_inputs=hidden_inputs) +async def get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=None, pre_execute_cb=None, v3_data=None): + return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data) has_pending_task = any(isinstance(r, asyncio.Task) and not r.done() for r in return_values) if has_pending_task: return return_values, {}, False, has_pending_task @@ -460,7 +465,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, has_subgraph = False else: get_progress_state().start_progress(unique_id) - input_data_all, missing_keys, hidden_inputs = get_input_data(inputs, class_def, unique_id, execution_list, dynprompt, extra_data) + input_data_all, missing_keys, v3_data = get_input_data(inputs, class_def, unique_id, execution_list, dynprompt, extra_data) if server.client_id is not None: server.last_node_id = display_node_id server.send_sync("executing", { "node": unique_id, "display_node": display_node_id, "prompt_id": prompt_id }, server.client_id) @@ -475,7 +480,10 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, else: lazy_status_present = getattr(obj, "check_lazy_status", None) is not None if lazy_status_present: - required_inputs = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, "check_lazy_status", allow_interrupt=True, hidden_inputs=hidden_inputs) + # for check_lazy_status, the returned data should include the original key of the input + v3_data_lazy = v3_data.copy() + v3_data_lazy["create_dynamic_tuple"] = True + required_inputs = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, "check_lazy_status", allow_interrupt=True, v3_data=v3_data_lazy) required_inputs = await resolve_map_node_over_list_results(required_inputs) required_inputs = set(sum([r for r in required_inputs if isinstance(r,list)], [])) required_inputs = [x for x in required_inputs if isinstance(x,str) and ( @@ -507,7 +515,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, def pre_execute_cb(call_index): # TODO - How to handle this with async functions without contextvars (which requires Python 3.12)? GraphBuilder.set_default_prefix(unique_id, call_index, 0) - output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, hidden_inputs=hidden_inputs) + output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data) if has_pending_tasks: pending_async_nodes[unique_id] = output_data unblock = execution_list.add_external_block(unique_id) @@ -593,6 +601,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, if isinstance(ex, comfy.model_management.OOM_EXCEPTION): tips = "This error means you ran out of memory on your GPU.\n\nTIPS: If the workflow worked before you might have accidentally set the batch_size to a large number." + logging.info("Memory summary: {}".format(comfy.model_management.debug_memory_summary())) logging.error("Got an OOM, unloading all loaded models.") comfy.model_management.unload_all_models() @@ -664,6 +673,8 @@ class PromptExecutor: asyncio.run(self.execute_async(prompt, prompt_id, extra_data, execute_outputs)) async def execute_async(self, prompt, prompt_id, extra_data={}, execute_outputs=[]): + set_preview_method(extra_data.get("preview_method")) + nodes.interrupt_processing(False) if "client_id" in extra_data: @@ -745,18 +756,20 @@ async def validate_inputs(prompt_id, prompt, item, validated): class_type = prompt[unique_id]['class_type'] obj_class = nodes.NODE_CLASS_MAPPINGS[class_type] - class_inputs = obj_class.INPUT_TYPES() - valid_inputs = set(class_inputs.get('required',{})).union(set(class_inputs.get('optional',{}))) - errors = [] valid = True + v3_data = None validate_function_inputs = [] validate_has_kwargs = False if issubclass(obj_class, _ComfyNodeInternal): + obj_class: _io._ComfyNodeBaseInternal + class_inputs = obj_class.INPUT_TYPES() + class_inputs, _, v3_data = _io.get_finalized_class_inputs(class_inputs, inputs) validate_function_name = "validate_inputs" validate_function = first_real_override(obj_class, validate_function_name) else: + class_inputs = obj_class.INPUT_TYPES() validate_function_name = "VALIDATE_INPUTS" validate_function = getattr(obj_class, validate_function_name, None) if validate_function is not None: @@ -765,15 +778,18 @@ async def validate_inputs(prompt_id, prompt, item, validated): validate_has_kwargs = argspec.varkw is not None received_types = {} + valid_inputs = set(class_inputs.get('required',{})).union(set(class_inputs.get('optional',{}))) + for x in valid_inputs: input_type, input_category, extra_info = get_input_info(obj_class, x, class_inputs) assert extra_info is not None if x not in inputs: if input_category == "required": + details = f"{x}" if not v3_data else x.split(".")[-1] error = { "type": "required_input_missing", "message": "Required input is missing", - "details": f"{x}", + "details": details, "extra_info": { "input_name": x } @@ -907,8 +923,11 @@ async def validate_inputs(prompt_id, prompt, item, validated): errors.append(error) continue - if isinstance(input_type, list): - combo_options = input_type + if isinstance(input_type, list) or input_type == io.Combo.io_type: + if input_type == io.Combo.io_type: + combo_options = extra_info.get("options", []) + else: + combo_options = input_type if val not in combo_options: input_config = info list_info = "" @@ -935,7 +954,7 @@ async def validate_inputs(prompt_id, prompt, item, validated): continue if len(validate_function_inputs) > 0 or validate_has_kwargs: - input_data_all, _, hidden_inputs = get_input_data(inputs, obj_class, unique_id) + input_data_all, _, v3_data = get_input_data(inputs, obj_class, unique_id) input_filtered = {} for x in input_data_all: if x in validate_function_inputs or validate_has_kwargs: @@ -943,7 +962,7 @@ async def validate_inputs(prompt_id, prompt, item, validated): if 'input_types' in validate_function_inputs: input_filtered['input_types'] = [received_types] - ret = await _async_map_node_over_list(prompt_id, unique_id, obj_class, input_filtered, validate_function_name, hidden_inputs=hidden_inputs) + ret = await _async_map_node_over_list(prompt_id, unique_id, obj_class, input_filtered, validate_function_name, v3_data=v3_data) ret = await resolve_map_node_over_list_results(ret) for x in input_filtered: for i, r in enumerate(ret): diff --git a/folder_paths.py b/folder_paths.py index f110d832b..9c96540e3 100644 --- a/folder_paths.py +++ b/folder_paths.py @@ -38,6 +38,8 @@ folder_names_and_paths["gligen"] = ([os.path.join(models_dir, "gligen")], suppor folder_names_and_paths["upscale_models"] = ([os.path.join(models_dir, "upscale_models")], supported_pt_extensions) +folder_names_and_paths["latent_upscale_models"] = ([os.path.join(models_dir, "latent_upscale_models")], supported_pt_extensions) + folder_names_and_paths["custom_nodes"] = ([os.path.join(base_path, "custom_nodes")], set()) folder_names_and_paths["hypernetworks"] = ([os.path.join(models_dir, "hypernetworks")], supported_pt_extensions) @@ -135,6 +137,71 @@ def set_user_directory(user_dir: str) -> None: user_directory = user_dir +# System User Protection - Protects system directories from HTTP endpoint access +# System Users are internal-only users that cannot be accessed via HTTP endpoints. +# They use the '__' prefix convention (similar to Python's private member convention). +SYSTEM_USER_PREFIX = "__" + + +def get_system_user_directory(name: str = "system") -> str: + """ + Get the path to a System User directory. + + System User directories (prefixed with '__') are only accessible via internal API, + not through HTTP endpoints. Use this for storing system-internal data that + should not be exposed to users. + + Args: + name: System user name (e.g., "system", "cache"). Must be alphanumeric + with underscores allowed, but cannot start with underscore. + + Returns: + Absolute path to the system user directory. + + Raises: + ValueError: If name is empty, invalid, or starts with underscore. + + Example: + >>> get_system_user_directory("cache") + '/path/to/user/__cache' + """ + if not name or not isinstance(name, str): + raise ValueError("System user name cannot be empty") + if not name.replace("_", "").isalnum(): + raise ValueError(f"Invalid system user name: '{name}'") + if name.startswith("_"): + raise ValueError("System user name should not start with underscore") + return os.path.join(get_user_directory(), f"{SYSTEM_USER_PREFIX}{name}") + + +def get_public_user_directory(user_id: str) -> str | None: + """ + Get the path to a Public User directory for HTTP endpoint access. + + This function provides structural security by returning None for any + System User (prefixed with '__'). All HTTP endpoints should use this + function instead of directly constructing user paths. + + Args: + user_id: User identifier from HTTP request. + + Returns: + Absolute path to the user directory, or None if user_id is invalid + or refers to a System User. + + Example: + >>> get_public_user_directory("default") + '/path/to/user/default' + >>> get_public_user_directory("__system") + None + """ + if not user_id or not isinstance(user_id, str): + return None + if user_id.startswith(SYSTEM_USER_PREFIX): + return None + return os.path.join(get_user_directory(), user_id) + + #NOTE: used in http server so don't put folders that should not be accessed remotely def get_directory_by_type(type_name: str) -> str | None: if type_name == "output": diff --git a/latent_preview.py b/latent_preview.py index 95d3cb733..d52e3f7a1 100644 --- a/latent_preview.py +++ b/latent_preview.py @@ -2,17 +2,26 @@ import torch from PIL import Image from comfy.cli_args import args, LatentPreviewMethod from comfy.taesd.taesd import TAESD +from comfy.sd import VAE import comfy.model_management import folder_paths import comfy.utils import logging -MAX_PREVIEW_RESOLUTION = args.preview_size +default_preview_method = args.preview_method -def preview_to_image(latent_image): - latents_ubyte = (((latent_image + 1.0) / 2.0).clamp(0, 1) # change scale from -1..1 to 0..1 - .mul(0xFF) # to 0..255 - ) +MAX_PREVIEW_RESOLUTION = args.preview_size +VIDEO_TAES = ["taehv", "lighttaew2_2", "lighttaew2_1", "lighttaehy1_5"] + +def preview_to_image(latent_image, do_scale=True): + if do_scale: + latents_ubyte = (((latent_image + 1.0) / 2.0).clamp(0, 1) # change scale from -1..1 to 0..1 + .mul(0xFF) # to 0..255 + ) + else: + latents_ubyte = (latent_image.clamp(0, 1) + .mul(0xFF) # to 0..255 + ) if comfy.model_management.directml_enabled: latents_ubyte = latents_ubyte.to(dtype=torch.uint8) latents_ubyte = latents_ubyte.to(device="cpu", dtype=torch.uint8, non_blocking=comfy.model_management.device_supports_non_blocking(latent_image.device)) @@ -35,15 +44,22 @@ class TAESDPreviewerImpl(LatentPreviewer): x_sample = self.taesd.decode(x0[:1])[0].movedim(0, 2) return preview_to_image(x_sample) +class TAEHVPreviewerImpl(TAESDPreviewerImpl): + def decode_latent_to_preview(self, x0): + x_sample = self.taesd.decode(x0[:1, :, :1])[0][0] + return preview_to_image(x_sample, do_scale=False) class Latent2RGBPreviewer(LatentPreviewer): - def __init__(self, latent_rgb_factors, latent_rgb_factors_bias=None): + def __init__(self, latent_rgb_factors, latent_rgb_factors_bias=None, latent_rgb_factors_reshape=None): self.latent_rgb_factors = torch.tensor(latent_rgb_factors, device="cpu").transpose(0, 1) self.latent_rgb_factors_bias = None if latent_rgb_factors_bias is not None: self.latent_rgb_factors_bias = torch.tensor(latent_rgb_factors_bias, device="cpu") + self.latent_rgb_factors_reshape = latent_rgb_factors_reshape def decode_latent_to_preview(self, x0): + if self.latent_rgb_factors_reshape is not None: + x0 = self.latent_rgb_factors_reshape(x0) self.latent_rgb_factors = self.latent_rgb_factors.to(dtype=x0.dtype, device=x0.device) if self.latent_rgb_factors_bias is not None: self.latent_rgb_factors_bias = self.latent_rgb_factors_bias.to(dtype=x0.dtype, device=x0.device) @@ -78,14 +94,19 @@ def get_previewer(device, latent_format): if method == LatentPreviewMethod.TAESD: if taesd_decoder_path: - taesd = TAESD(None, taesd_decoder_path, latent_channels=latent_format.latent_channels).to(device) - previewer = TAESDPreviewerImpl(taesd) + if latent_format.taesd_decoder_name in VIDEO_TAES: + taesd = VAE(comfy.utils.load_torch_file(taesd_decoder_path)) + taesd.first_stage_model.show_progress_bar = False + previewer = TAEHVPreviewerImpl(taesd) + else: + taesd = TAESD(None, taesd_decoder_path, latent_channels=latent_format.latent_channels).to(device) + previewer = TAESDPreviewerImpl(taesd) else: logging.warning("Warning: TAESD previews enabled, but could not find models/vae_approx/{}".format(latent_format.taesd_decoder_name)) if previewer is None: if latent_format.latent_rgb_factors is not None: - previewer = Latent2RGBPreviewer(latent_format.latent_rgb_factors, latent_format.latent_rgb_factors_bias) + previewer = Latent2RGBPreviewer(latent_format.latent_rgb_factors, latent_format.latent_rgb_factors_bias, latent_format.latent_rgb_factors_reshape) return previewer def prepare_callback(model, steps, x0_output_dict=None): @@ -106,3 +127,11 @@ def prepare_callback(model, steps, x0_output_dict=None): pbar.update_absolute(step + 1, total_steps, preview_bytes) return callback +def set_preview_method(override: str = None): + if override and override != "default": + method = LatentPreviewMethod.from_string(override) + if method is not None: + args.preview_method = method + return + args.preview_method = default_preview_method + diff --git a/main.py b/main.py index e1b0f1620..0e07a95da 100644 --- a/main.py +++ b/main.py @@ -15,6 +15,7 @@ from comfy_execution.progress import get_progress_state from comfy_execution.utils import get_executing_context from comfy_api import feature_flags + if __name__ == "__main__": #NOTE: These do not do anything on core ComfyUI, they are for custom nodes. os.environ['HF_HUB_DISABLE_TELEMETRY'] = '1' @@ -22,6 +23,55 @@ if __name__ == "__main__": setup_logger(log_level=args.verbose, use_stdout=args.log_stdout) +if os.name == "nt": + os.environ['MIMALLOC_PURGE_DELAY'] = '0' + +if __name__ == "__main__": + os.environ['TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL'] = '1' + if args.default_device is not None: + default_dev = args.default_device + devices = list(range(32)) + devices.remove(default_dev) + devices.insert(0, default_dev) + devices = ','.join(map(str, devices)) + os.environ['CUDA_VISIBLE_DEVICES'] = str(devices) + os.environ['HIP_VISIBLE_DEVICES'] = str(devices) + + if args.cuda_device is not None: + os.environ['CUDA_VISIBLE_DEVICES'] = str(args.cuda_device) + os.environ['HIP_VISIBLE_DEVICES'] = str(args.cuda_device) + os.environ["ASCEND_RT_VISIBLE_DEVICES"] = str(args.cuda_device) + logging.info("Set cuda device to: {}".format(args.cuda_device)) + + if args.oneapi_device_selector is not None: + os.environ['ONEAPI_DEVICE_SELECTOR'] = args.oneapi_device_selector + logging.info("Set oneapi device selector to: {}".format(args.oneapi_device_selector)) + + if args.deterministic: + if 'CUBLAS_WORKSPACE_CONFIG' not in os.environ: + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ":4096:8" + + import cuda_malloc + if "rocm" in cuda_malloc.get_torch_version_noimport(): + os.environ['OCL_SET_SVM_SIZE'] = '262144' # set at the request of AMD + + +def handle_comfyui_manager_unavailable(): + if not args.windows_standalone_build: + logging.warning(f"\n\nYou appear to be running comfyui-manager from source, this is not recommended. Please install comfyui-manager using the following command:\ncommand:\n\t{sys.executable} -m pip install --pre comfyui_manager\n") + args.enable_manager = False + + +if args.enable_manager: + if importlib.util.find_spec("comfyui_manager"): + import comfyui_manager + + if not comfyui_manager.__file__ or not comfyui_manager.__file__.endswith('__init__.py'): + handle_comfyui_manager_unavailable() + else: + handle_comfyui_manager_unavailable() + + def apply_custom_paths(): # extra model paths extra_model_paths_config_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "extra_model_paths.yaml") @@ -79,6 +129,11 @@ def execute_prestartup_script(): for possible_module in possible_modules: module_path = os.path.join(custom_node_path, possible_module) + + if args.enable_manager: + if comfyui_manager.should_be_disabled(module_path): + continue + if os.path.isfile(module_path) or module_path.endswith(".disabled") or module_path == "__pycache__": continue @@ -101,6 +156,10 @@ def execute_prestartup_script(): logging.info("") apply_custom_paths() + +if args.enable_manager: + comfyui_manager.prestartup() + execute_prestartup_script() @@ -110,37 +169,6 @@ import shutil import threading import gc - -if os.name == "nt": - os.environ['MIMALLOC_PURGE_DELAY'] = '0' - -if __name__ == "__main__": - os.environ['TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL'] = '1' - if args.default_device is not None: - default_dev = args.default_device - devices = list(range(32)) - devices.remove(default_dev) - devices.insert(0, default_dev) - devices = ','.join(map(str, devices)) - os.environ['CUDA_VISIBLE_DEVICES'] = str(devices) - os.environ['HIP_VISIBLE_DEVICES'] = str(devices) - - if args.cuda_device is not None: - os.environ['CUDA_VISIBLE_DEVICES'] = str(args.cuda_device) - os.environ['HIP_VISIBLE_DEVICES'] = str(args.cuda_device) - os.environ["ASCEND_RT_VISIBLE_DEVICES"] = str(args.cuda_device) - logging.info("Set cuda device to: {}".format(args.cuda_device)) - - if args.oneapi_device_selector is not None: - os.environ['ONEAPI_DEVICE_SELECTOR'] = args.oneapi_device_selector - logging.info("Set oneapi device selector to: {}".format(args.oneapi_device_selector)) - - if args.deterministic: - if 'CUBLAS_WORKSPACE_CONFIG' not in os.environ: - os.environ['CUBLAS_WORKSPACE_CONFIG'] = ":4096:8" - - import cuda_malloc - if 'torch' in sys.modules: logging.warning("WARNING: Potential Error in code: Torch already imported, torch should never be imported before this point.") @@ -323,6 +351,9 @@ def start_comfyui(asyncio_loop=None): asyncio.set_event_loop(asyncio_loop) prompt_server = server.PromptServer(asyncio_loop) + if args.enable_manager and not args.disable_manager_ui: + comfyui_manager.start() + hook_breaker_ac10a0.save_functions() asyncio_loop.run_until_complete(nodes.init_extra_nodes( init_custom_nodes=(not args.disable_all_custom_nodes) or len(args.whitelist_custom_nodes) > 0, diff --git a/manager_requirements.txt b/manager_requirements.txt new file mode 100644 index 000000000..6585b0c19 --- /dev/null +++ b/manager_requirements.txt @@ -0,0 +1 @@ +comfyui_manager==4.0.4 diff --git a/models/latent_upscale_models/put_latent_upscale_models_here b/models/latent_upscale_models/put_latent_upscale_models_here new file mode 100644 index 000000000..e69de29bb diff --git a/nodes.py b/nodes.py index f6aeedc78..56b74ebe3 100644 --- a/nodes.py +++ b/nodes.py @@ -43,6 +43,9 @@ import folder_paths import latent_preview import node_helpers +if args.enable_manager: + import comfyui_manager + def before_node_execution(): comfy.model_management.throw_exception_if_processing_interrupted() @@ -292,7 +295,11 @@ class VAEDecode: DESCRIPTION = "Decodes latent images back into pixel space images." def decode(self, vae, samples): - images = vae.decode(samples["samples"]) + latent = samples["samples"] + if latent.is_nested: + latent = latent.unbind()[0] + + images = vae.decode(latent) if len(images.shape) == 5: #Combine batches images = images.reshape(-1, images.shape[-3], images.shape[-2], images.shape[-1]) return (images, ) @@ -340,7 +347,7 @@ class VAEEncode: CATEGORY = "latent" def encode(self, vae, pixels): - t = vae.encode(pixels[:,:,:,:3]) + t = vae.encode(pixels) return ({"samples":t}, ) class VAEEncodeTiled: @@ -358,7 +365,7 @@ class VAEEncodeTiled: CATEGORY = "_for_testing" def encode(self, vae, pixels, tile_size, overlap, temporal_size=64, temporal_overlap=8): - t = vae.encode_tiled(pixels[:,:,:,:3], tile_x=tile_size, tile_y=tile_size, overlap=overlap, tile_t=temporal_size, overlap_t=temporal_overlap) + t = vae.encode_tiled(pixels, tile_x=tile_size, tile_y=tile_size, overlap=overlap, tile_t=temporal_size, overlap_t=temporal_overlap) return ({"samples": t}, ) class VAEEncodeForInpaint: @@ -692,8 +699,10 @@ class LoraLoaderModelOnly(LoraLoader): return (self.load_lora(model, None, lora_name, strength_model, 0)[0],) class VAELoader: + video_taes = ["taehv", "lighttaew2_2", "lighttaew2_1", "lighttaehy1_5"] + image_taes = ["taesd", "taesdxl", "taesd3", "taef1"] @staticmethod - def vae_list(): + def vae_list(s): vaes = folder_paths.get_filename_list("vae") approx_vaes = folder_paths.get_filename_list("vae_approx") sdxl_taesd_enc = False @@ -722,6 +731,11 @@ class VAELoader: f1_taesd_dec = True elif v.startswith("taef1_decoder."): f1_taesd_enc = True + else: + for tae in s.video_taes: + if v.startswith(tae): + vaes.append(v) + if sd1_taesd_dec and sd1_taesd_enc: vaes.append("taesd") if sdxl_taesd_dec and sdxl_taesd_enc: @@ -765,7 +779,7 @@ class VAELoader: @classmethod def INPUT_TYPES(s): - return {"required": { "vae_name": (s.vae_list(), )}} + return {"required": { "vae_name": (s.vae_list(s), )}} RETURN_TYPES = ("VAE",) FUNCTION = "load_vae" @@ -776,10 +790,13 @@ class VAELoader: if vae_name == "pixel_space": sd = {} sd["pixel_space_vae"] = torch.tensor(1.0) - elif vae_name in ["taesd", "taesdxl", "taesd3", "taef1"]: + elif vae_name in self.image_taes: sd = self.load_taesd(vae_name) else: - vae_path = folder_paths.get_full_path_or_raise("vae", vae_name) + if os.path.splitext(vae_name)[0] in self.video_taes: + vae_path = folder_paths.get_full_path_or_raise("vae_approx", vae_name) + else: + vae_path = folder_paths.get_full_path_or_raise("vae", vae_name) sd = comfy.utils.load_torch_file(vae_path) vae = comfy.sd.VAE(sd=sd) vae.throw_exception_if_invalid() @@ -929,7 +946,7 @@ class CLIPLoader: @classmethod def INPUT_TYPES(s): return {"required": { "clip_name": (folder_paths.get_filename_list("text_encoders"), ), - "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream", "chroma", "ace", "omnigen2", "qwen_image", "hunyuan_image"], ), + "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream", "chroma", "ace", "omnigen2", "qwen_image", "hunyuan_image", "flux2", "ovis"], ), }, "optional": { "device": (["default", "cpu"], {"advanced": True}), @@ -957,7 +974,7 @@ class DualCLIPLoader: def INPUT_TYPES(s): return {"required": { "clip_name1": (folder_paths.get_filename_list("text_encoders"), ), "clip_name2": (folder_paths.get_filename_list("text_encoders"), ), - "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image"], ), + "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image", "hunyuan_video_15", "kandinsky5", "kandinsky5_image", "ltxv", "newbie"], ), }, "optional": { "device": (["default", "cpu"], {"advanced": True}), @@ -967,7 +984,7 @@ class DualCLIPLoader: CATEGORY = "advanced/loaders" - DESCRIPTION = "[Recipes]\n\nsdxl: clip-l, clip-g\nsd3: clip-l, clip-g / clip-l, t5 / clip-g, t5\nflux: clip-l, t5\nhidream: at least one of t5 or llama, recommended t5 and llama\nhunyuan_image: qwen2.5vl 7b and byt5 small" + DESCRIPTION = "[Recipes]\n\nsdxl: clip-l, clip-g\nsd3: clip-l, clip-g / clip-l, t5 / clip-g, t5\nflux: clip-l, t5\nhidream: at least one of t5 or llama, recommended t5 and llama\nhunyuan_image: qwen2.5vl 7b and byt5 small\nnewbie: gemma-3-4b-it, jina clip v2" def load_clip(self, clip_name1, clip_name2, type, device="default"): clip_type = getattr(comfy.sd.CLIPType, type.upper(), comfy.sd.CLIPType.STABLE_DIFFUSION) @@ -1650,8 +1667,6 @@ class LoadImage: output_masks = [] w, h = None, None - excluded_formats = ['MPO'] - for i in ImageSequence.Iterator(img): i = node_helpers.pillow(ImageOps.exif_transpose, i) @@ -1679,7 +1694,10 @@ class LoadImage: output_images.append(image) output_masks.append(mask.unsqueeze(0)) - if len(output_images) > 1 and img.format not in excluded_formats: + if img.format == "MPO": + break # ignore all frames except the first one for MPO format + + if len(output_images) > 1: output_image = torch.cat(output_images, dim=0) output_mask = torch.cat(output_masks, dim=0) else: @@ -1850,8 +1868,14 @@ class ImageBatch: FUNCTION = "batch" CATEGORY = "image" + DEPRECATED = True def batch(self, image1, image2): + if image1.shape[-1] != image2.shape[-1]: + if image1.shape[-1] > image2.shape[-1]: + image2 = torch.nn.functional.pad(image2, (0,1), mode='constant', value=1.0) + else: + image1 = torch.nn.functional.pad(image1, (0,1), mode='constant', value=1.0) if image1.shape[1:] != image2.shape[1:]: image2 = comfy.utils.common_upscale(image2.movedim(-1,1), image1.shape[2], image1.shape[1], "bilinear", "center").movedim(1,-1) s = torch.cat((image1, image2), dim=0) @@ -2223,11 +2247,19 @@ async def init_external_custom_nodes(): for possible_module in possible_modules: module_path = os.path.join(custom_node_path, possible_module) - if os.path.isfile(module_path) and os.path.splitext(module_path)[1] != ".py": continue - if module_path.endswith(".disabled"): continue + if os.path.isfile(module_path) and os.path.splitext(module_path)[1] != ".py": + continue + if module_path.endswith(".disabled"): + continue if args.disable_all_custom_nodes and possible_module not in args.whitelist_custom_nodes: logging.info(f"Skipping {possible_module} due to disable_all_custom_nodes and whitelist_custom_nodes") continue + + if args.enable_manager: + if comfyui_manager.should_be_disabled(module_path): + logging.info(f"Blocked by policy: {module_path}") + continue + time_before = time.perf_counter() success = await load_custom_node(module_path, base_node_names, module_parent="custom_nodes") node_import_times.append((time.perf_counter() - time_before, module_path, success)) @@ -2273,6 +2305,7 @@ async def init_builtin_extra_nodes(): "nodes_images.py", "nodes_video_model.py", "nodes_train.py", + "nodes_dataset.py", "nodes_sag.py", "nodes_perpneg.py", "nodes_stable3d.py", @@ -2302,6 +2335,8 @@ async def init_builtin_extra_nodes(): "nodes_mochi.py", "nodes_slg.py", "nodes_mahiro.py", + "nodes_lt_upsampler.py", + "nodes_lt_audio.py", "nodes_lt.py", "nodes_hooks.py", "nodes_load_3d.py", @@ -2330,7 +2365,10 @@ async def init_builtin_extra_nodes(): "nodes_easycache.py", "nodes_audio_encoder.py", "nodes_rope.py", + "nodes_logic.py", "nodes_nop.py", + "nodes_kandinsky5.py", + "nodes_wanmove.py", ] import_failed = [] @@ -2356,9 +2394,9 @@ async def init_builtin_api_nodes(): "nodes_recraft.py", "nodes_pixverse.py", "nodes_stability.py", - "nodes_pika.py", "nodes_runway.py", "nodes_sora.py", + "nodes_topaz.py", "nodes_tripo.py", "nodes_moonvalley.py", "nodes_rodin.py", diff --git a/pyproject.toml b/pyproject.toml index a14b383b3..951c2c978 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,9 +1,9 @@ [project] name = "ComfyUI" -version = "0.3.69" +version = "0.8.0" readme = "README.md" license = { file = "LICENSE" } -requires-python = ">=3.9" +requires-python = ">=3.10" [project.urls] homepage = "https://www.comfy.org/" @@ -15,12 +15,16 @@ lint.select = [ "N805", # invalid-first-argument-name-for-method "S307", # suspicious-eval-usage "S102", # exec + "E", "T", # print-usage "W", # The "F" series in Ruff stands for "Pyflakes" rules, which catch various Python syntax errors and undefined names. # See all rules here: https://docs.astral.sh/ruff/rules/#pyflakes-f "F", ] + +lint.ignore = ["E501", "E722", "E731", "E712", "E402", "E741"] + exclude = ["*.ipynb", "**/generated/*.pyi"] [tool.pylint] diff --git a/requirements.txt b/requirements.txt index 249c36dee..bc8346bcf 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ -comfyui-frontend-package==1.28.8 -comfyui-workflow-templates==0.2.11 +comfyui-frontend-package==1.35.9 +comfyui-workflow-templates==0.7.67 comfyui-embedded-docs==0.3.1 torch torchsde @@ -7,7 +7,7 @@ torchvision torchaudio numpy>=1.25.0 einops -transformers>=4.37.2 +transformers>=4.50.3 tokenizers>=0.13.3 sentencepiece safetensors>=0.4.2 @@ -21,6 +21,7 @@ psutil alembic SQLAlchemy av>=14.2.0 +comfy-kitchen>=0.2.3 #non essential dependencies: kornia>=0.7.1 diff --git a/server.py b/server.py index d059d3dc9..70c8b5e3b 100644 --- a/server.py +++ b/server.py @@ -7,6 +7,7 @@ import time import nodes import folder_paths import execution +from comfy_execution.jobs import JobStatus, get_job, get_all_jobs import uuid import urllib import json @@ -30,7 +31,7 @@ import comfy.model_management from comfy_api import feature_flags import node_helpers from comfyui_version import __version__ -from app.frontend_management import FrontendManager +from app.frontend_management import FrontendManager, parse_version from comfy_api.internal import _ComfyNodeInternal from app.user_manager import UserManager @@ -44,6 +45,15 @@ from protocol import BinaryEventTypes # Import cache control middleware from middleware.cache_middleware import cache_control +if args.enable_manager: + import comfyui_manager + + +def _remove_sensitive_from_queue(queue: list) -> list: + """Remove sensitive data (index 5) from queue item tuples.""" + return [item[:5] for item in queue] + + async def send_socket_catch_exception(function, message): try: await function(message) @@ -95,7 +105,7 @@ def create_cors_middleware(allowed_origin: str): response = await handler(request) response.headers['Access-Control-Allow-Origin'] = allowed_origin - response.headers['Access-Control-Allow-Methods'] = 'POST, GET, DELETE, PUT, OPTIONS' + response.headers['Access-Control-Allow-Methods'] = 'POST, GET, DELETE, PUT, OPTIONS, PATCH' response.headers['Access-Control-Allow-Headers'] = 'Content-Type, Authorization' response.headers['Access-Control-Allow-Credentials'] = 'true' return response @@ -164,6 +174,22 @@ def create_origin_only_middleware(): return origin_only_middleware + +def create_block_external_middleware(): + @web.middleware + async def block_external_middleware(request: web.Request, handler): + if request.method == "OPTIONS": + # Pre-flight request. Reply successfully: + response = web.Response() + else: + response = await handler(request) + + response.headers['Content-Security-Policy'] = "default-src 'self'; script-src 'self' 'unsafe-inline' 'unsafe-eval' blob:; style-src 'self' 'unsafe-inline'; img-src 'self' data: blob:; font-src 'self'; connect-src 'self'; frame-src 'self'; object-src 'self';" + return response + + return block_external_middleware + + class PromptServer(): def __init__(self, loop): PromptServer.instance = self @@ -193,6 +219,12 @@ class PromptServer(): else: middlewares.append(create_origin_only_middleware()) + if args.disable_api_nodes: + middlewares.append(create_block_external_middleware()) + + if args.enable_manager: + middlewares.append(comfyui_manager.create_middleware()) + max_upload_size = round(args.max_upload_size * 1024 * 1024) self.app = web.Application(client_max_size=max_upload_size, middlewares=middlewares) self.sockets = dict() @@ -292,7 +324,7 @@ class PromptServer(): @routes.get("/models/{folder}") async def get_models(request): folder = request.match_info.get("folder", None) - if not folder in folder_paths.folder_names_and_paths: + if folder not in folder_paths.folder_names_and_paths: return web.Response(status=404) files = folder_paths.get_filename_list(folder) return web.json_response(files) @@ -547,7 +579,7 @@ class PromptServer(): folder_name = request.match_info.get("folder_name", None) if folder_name is None: return web.Response(status=404) - if not "filename" in request.rel_url.query: + if "filename" not in request.rel_url.query: return web.Response(status=404) filename = request.rel_url.query["filename"] @@ -561,7 +593,7 @@ class PromptServer(): if out is None: return web.Response(status=404) dt = json.loads(out) - if not "__metadata__" in dt: + if "__metadata__" not in dt: return web.Response(status=404) return web.json_response(dt["__metadata__"]) @@ -580,7 +612,7 @@ class PromptServer(): system_stats = { "system": { - "os": os.name, + "os": sys.platform, "ram_total": ram_total, "ram_free": ram_free, "comfyui_version": __version__, @@ -669,6 +701,129 @@ class PromptServer(): out[node_class] = node_info(node_class) return web.json_response(out) + @routes.get("/api/jobs") + async def get_jobs(request): + """List all jobs with filtering, sorting, and pagination. + + Query parameters: + status: Filter by status (comma-separated): pending, in_progress, completed, failed + workflow_id: Filter by workflow ID + sort_by: Sort field: created_at (default), execution_duration + sort_order: Sort direction: asc, desc (default) + limit: Max items to return (positive integer) + offset: Items to skip (non-negative integer, default 0) + """ + query = request.rel_url.query + + status_param = query.get('status') + workflow_id = query.get('workflow_id') + sort_by = query.get('sort_by', 'created_at').lower() + sort_order = query.get('sort_order', 'desc').lower() + + status_filter = None + if status_param: + status_filter = [s.strip().lower() for s in status_param.split(',') if s.strip()] + invalid_statuses = [s for s in status_filter if s not in JobStatus.ALL] + if invalid_statuses: + return web.json_response( + {"error": f"Invalid status value(s): {', '.join(invalid_statuses)}. Valid values: {', '.join(JobStatus.ALL)}"}, + status=400 + ) + + if sort_by not in {'created_at', 'execution_duration'}: + return web.json_response( + {"error": "sort_by must be 'created_at' or 'execution_duration'"}, + status=400 + ) + + if sort_order not in {'asc', 'desc'}: + return web.json_response( + {"error": "sort_order must be 'asc' or 'desc'"}, + status=400 + ) + + limit = None + + # If limit is provided, validate that it is a positive integer, else continue without a limit + if 'limit' in query: + try: + limit = int(query.get('limit')) + if limit <= 0: + return web.json_response( + {"error": "limit must be a positive integer"}, + status=400 + ) + except (ValueError, TypeError): + return web.json_response( + {"error": "limit must be an integer"}, + status=400 + ) + + offset = 0 + if 'offset' in query: + try: + offset = int(query.get('offset')) + if offset < 0: + offset = 0 + except (ValueError, TypeError): + return web.json_response( + {"error": "offset must be an integer"}, + status=400 + ) + + running, queued = self.prompt_queue.get_current_queue_volatile() + history = self.prompt_queue.get_history() + + running = _remove_sensitive_from_queue(running) + queued = _remove_sensitive_from_queue(queued) + + jobs, total = get_all_jobs( + running, queued, history, + status_filter=status_filter, + workflow_id=workflow_id, + sort_by=sort_by, + sort_order=sort_order, + limit=limit, + offset=offset + ) + + has_more = (offset + len(jobs)) < total + + return web.json_response({ + 'jobs': jobs, + 'pagination': { + 'offset': offset, + 'limit': limit, + 'total': total, + 'has_more': has_more + } + }) + + @routes.get("/api/jobs/{job_id}") + async def get_job_by_id(request): + """Get a single job by ID.""" + job_id = request.match_info.get("job_id", None) + if not job_id: + return web.json_response( + {"error": "job_id is required"}, + status=400 + ) + + running, queued = self.prompt_queue.get_current_queue_volatile() + history = self.prompt_queue.get_history(prompt_id=job_id) + + running = _remove_sensitive_from_queue(running) + queued = _remove_sensitive_from_queue(queued) + + job = get_job(job_id, running, queued, history) + if job is None: + return web.json_response( + {"error": "Job not found"}, + status=404 + ) + + return web.json_response(job) + @routes.get("/history") async def get_history(request): max_items = request.rel_url.query.get("max_items", None) @@ -692,9 +847,8 @@ class PromptServer(): async def get_queue(request): queue_info = {} current_queue = self.prompt_queue.get_current_queue_volatile() - remove_sensitive = lambda queue: [x[:5] for x in queue] - queue_info['queue_running'] = remove_sensitive(current_queue[0]) - queue_info['queue_pending'] = remove_sensitive(current_queue[1]) + queue_info['queue_running'] = _remove_sensitive_from_queue(current_queue[0]) + queue_info['queue_pending'] = _remove_sensitive_from_queue(current_queue[1]) return web.json_response(queue_info) @routes.post("/prompt") @@ -849,11 +1003,31 @@ class PromptServer(): for name, dir in nodes.EXTENSION_WEB_DIRS.items(): self.app.add_routes([web.static('/extensions/' + name, dir)]) - workflow_templates_path = FrontendManager.templates_path() - if workflow_templates_path: - self.app.add_routes([ - web.static('/templates', workflow_templates_path) - ]) + installed_templates_version = FrontendManager.get_installed_templates_version() + use_legacy_templates = True + if installed_templates_version: + try: + use_legacy_templates = ( + parse_version(installed_templates_version) + < parse_version("0.3.0") + ) + except Exception as exc: + logging.warning( + "Unable to parse templates version '%s': %s", + installed_templates_version, + exc, + ) + + if use_legacy_templates: + workflow_templates_path = FrontendManager.legacy_templates_path() + if workflow_templates_path: + self.app.add_routes([ + web.static('/templates', workflow_templates_path) + ]) + else: + handler = FrontendManager.template_asset_handler() + if handler: + self.app.router.add_get("/templates/{path:.*}", handler) # Serve embedded documentation from the package embedded_docs_path = FrontendManager.embedded_docs_path() diff --git a/tests-unit/app_test/user_manager_system_user_test.py b/tests-unit/app_test/user_manager_system_user_test.py new file mode 100644 index 000000000..63b1ac5e5 --- /dev/null +++ b/tests-unit/app_test/user_manager_system_user_test.py @@ -0,0 +1,193 @@ +"""Tests for System User Protection in user_manager.py + +Tests cover: +- get_request_user_id(): 1st defense layer - blocks System Users from HTTP headers +- get_request_user_filepath(): 2nd defense layer - structural blocking via get_public_user_directory() +- add_user(): 3rd defense layer - prevents creation of System User names +- Defense layers integration tests +""" + +import pytest +from unittest.mock import MagicMock, patch +import tempfile + +import folder_paths +from app.user_manager import UserManager + + +@pytest.fixture +def mock_user_directory(): + """Create a temporary user directory.""" + with tempfile.TemporaryDirectory() as temp_dir: + original_dir = folder_paths.get_user_directory() + folder_paths.set_user_directory(temp_dir) + yield temp_dir + folder_paths.set_user_directory(original_dir) + + +@pytest.fixture +def user_manager(mock_user_directory): + """Create a UserManager instance for testing.""" + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + manager = UserManager() + # Add a default user for testing + manager.users = {"default": "default", "test_user_123": "Test User"} + yield manager + + +@pytest.fixture +def mock_request(): + """Create a mock request object.""" + request = MagicMock() + request.headers = {} + return request + + +class TestGetRequestUserId: + """Tests for get_request_user_id() - 1st defense layer. + + Verifies: + - System Users (__ prefix) in HTTP header are rejected with KeyError + - Public Users pass through successfully + """ + + def test_system_user_raises_error(self, user_manager, mock_request): + """Test System User in header raises KeyError.""" + mock_request.headers = {"comfy-user": "__system"} + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + with pytest.raises(KeyError, match="Unknown user"): + user_manager.get_request_user_id(mock_request) + + def test_system_user_cache_raises_error(self, user_manager, mock_request): + """Test System User cache raises KeyError.""" + mock_request.headers = {"comfy-user": "__cache"} + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + with pytest.raises(KeyError, match="Unknown user"): + user_manager.get_request_user_id(mock_request) + + def test_normal_user_works(self, user_manager, mock_request): + """Test normal user access works.""" + mock_request.headers = {"comfy-user": "default"} + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + user_id = user_manager.get_request_user_id(mock_request) + assert user_id == "default" + + def test_unknown_user_raises_error(self, user_manager, mock_request): + """Test unknown user raises KeyError.""" + mock_request.headers = {"comfy-user": "unknown_user"} + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + with pytest.raises(KeyError, match="Unknown user"): + user_manager.get_request_user_id(mock_request) + + +class TestGetRequestUserFilepath: + """Tests for get_request_user_filepath() - 2nd defense layer. + + Verifies: + - Returns None when get_public_user_directory() returns None (System User) + - Acts as backup defense if 1st layer is bypassed + """ + + def test_system_user_returns_none(self, user_manager, mock_request, mock_user_directory): + """Test System User returns None (structural blocking).""" + # First, we need to mock get_request_user_id to return System User + # But actually, get_request_user_id will raise KeyError first + # So we test via get_public_user_directory returning None + mock_request.headers = {"comfy-user": "default"} + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + # Patch get_public_user_directory to return None for testing + with patch.object(folder_paths, 'get_public_user_directory', return_value=None): + result = user_manager.get_request_user_filepath(mock_request, "test.txt") + assert result is None + + def test_normal_user_gets_path(self, user_manager, mock_request, mock_user_directory): + """Test normal user gets valid filepath.""" + mock_request.headers = {"comfy-user": "default"} + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + path = user_manager.get_request_user_filepath(mock_request, "test.txt") + assert path is not None + assert "default" in path + assert path.endswith("test.txt") + + +class TestAddUser: + """Tests for add_user() - 3rd defense layer (creation-time blocking). + + Verifies: + - System User name (__ prefix) creation is rejected with ValueError + - Sanitized usernames that become System User are also rejected + """ + + def test_system_user_prefix_name_raises(self, user_manager): + """Test System User prefix in name raises ValueError.""" + with pytest.raises(ValueError, match="System User prefix not allowed"): + user_manager.add_user("__system") + + def test_system_user_prefix_cache_raises(self, user_manager): + """Test System User cache prefix raises ValueError.""" + with pytest.raises(ValueError, match="System User prefix not allowed"): + user_manager.add_user("__cache") + + def test_sanitized_system_user_prefix_raises(self, user_manager): + """Test sanitized name becoming System User prefix raises ValueError (bypass prevention).""" + # "__test" directly starts with System User prefix + with pytest.raises(ValueError, match="System User prefix not allowed"): + user_manager.add_user("__test") + + def test_normal_user_creation(self, user_manager, mock_user_directory): + """Test normal user creation works.""" + user_id = user_manager.add_user("Normal User") + assert user_id is not None + assert not user_id.startswith("__") + assert "Normal-User" in user_id or "Normal_User" in user_id + + def test_empty_name_raises(self, user_manager): + """Test empty name raises ValueError.""" + with pytest.raises(ValueError, match="username not provided"): + user_manager.add_user("") + + def test_whitespace_only_raises(self, user_manager): + """Test whitespace-only name raises ValueError.""" + with pytest.raises(ValueError, match="username not provided"): + user_manager.add_user(" ") + + +class TestDefenseLayers: + """Integration tests for all three defense layers. + + Verifies: + - Each defense layer blocks System Users independently + - System User bypass is impossible through any layer + """ + + def test_layer1_get_request_user_id(self, user_manager, mock_request): + """Test 1st defense layer blocks System Users.""" + mock_request.headers = {"comfy-user": "__system"} + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + with pytest.raises(KeyError): + user_manager.get_request_user_id(mock_request) + + def test_layer2_get_public_user_directory(self): + """Test 2nd defense layer blocks System Users.""" + result = folder_paths.get_public_user_directory("__system") + assert result is None + + def test_layer3_add_user(self, user_manager): + """Test 3rd defense layer blocks System User creation.""" + with pytest.raises(ValueError): + user_manager.add_user("__system") diff --git a/tests-unit/comfy_extras_test/image_stitch_test.py b/tests-unit/comfy_extras_test/image_stitch_test.py index b5a0f022c..5c6a15ac4 100644 --- a/tests-unit/comfy_extras_test/image_stitch_test.py +++ b/tests-unit/comfy_extras_test/image_stitch_test.py @@ -25,7 +25,7 @@ class TestImageStitch: result = node.stitch(image1, "right", True, 0, "white", image2=None) - assert len(result) == 1 + assert len(result.result) == 1 assert torch.equal(result[0], image1) def test_basic_horizontal_stitch_right(self): diff --git a/tests-unit/comfy_quant/test_mixed_precision.py b/tests-unit/comfy_quant/test_mixed_precision.py index f8d1fd04e..7b2eac940 100644 --- a/tests-unit/comfy_quant/test_mixed_precision.py +++ b/tests-unit/comfy_quant/test_mixed_precision.py @@ -2,6 +2,7 @@ import unittest import torch import sys import os +import json # Add comfy to path sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "..")) @@ -15,6 +16,7 @@ if not has_gpu(): from comfy import ops from comfy.quant_ops import QuantizedTensor +import comfy.utils class SimpleModel(torch.nn.Module): @@ -37,11 +39,8 @@ class TestMixedPrecisionOps(unittest.TestCase): def test_all_layers_standard(self): """Test that model with no quantization works normally""" - # Configure no quantization - ops.MixedPrecisionOps._layer_quant_config = {} - # Create model - model = SimpleModel(operations=ops.MixedPrecisionOps) + model = SimpleModel(operations=ops.mixed_precision_ops({})) # Initialize weights manually model.layer1.weight = torch.nn.Parameter(torch.randn(20, 10, dtype=torch.bfloat16)) @@ -76,7 +75,6 @@ class TestMixedPrecisionOps(unittest.TestCase): "params": {} } } - ops.MixedPrecisionOps._layer_quant_config = layer_quant_config # Create state dict with mixed precision fp8_weight1 = torch.randn(20, 10, dtype=torch.float32).to(torch.float8_e4m3fn) @@ -98,28 +96,30 @@ class TestMixedPrecisionOps(unittest.TestCase): "layer3.weight_scale": torch.tensor(1.5, dtype=torch.float32), } + state_dict, _ = comfy.utils.convert_old_quants(state_dict, metadata={"_quantization_metadata": json.dumps({"layers": layer_quant_config})}) # Create model and load state dict (strict=False because custom loading pops keys) - model = SimpleModel(operations=ops.MixedPrecisionOps) + model = SimpleModel(operations=ops.mixed_precision_ops({})) model.load_state_dict(state_dict, strict=False) # Verify weights are wrapped in QuantizedTensor self.assertIsInstance(model.layer1.weight, QuantizedTensor) - self.assertEqual(model.layer1.weight._layout_type, "TensorCoreFP8Layout") + self.assertEqual(model.layer1.weight._layout_cls, "TensorCoreFP8E4M3Layout") # Layer 2 should NOT be quantized self.assertNotIsInstance(model.layer2.weight, QuantizedTensor) # Layer 3 should be quantized self.assertIsInstance(model.layer3.weight, QuantizedTensor) - self.assertEqual(model.layer3.weight._layout_type, "TensorCoreFP8Layout") + self.assertEqual(model.layer3.weight._layout_cls, "TensorCoreFP8E4M3Layout") # Verify scales were loaded - self.assertEqual(model.layer1.weight._layout_params['scale'].item(), 2.0) - self.assertEqual(model.layer3.weight._layout_params['scale'].item(), 1.5) + self.assertEqual(model.layer1.weight._params.scale.item(), 2.0) + self.assertEqual(model.layer3.weight._params.scale.item(), 1.5) # Forward pass input_tensor = torch.randn(5, 10, dtype=torch.bfloat16) - output = model(input_tensor) + with torch.inference_mode(): + output = model(input_tensor) self.assertEqual(output.shape, (5, 40)) @@ -132,7 +132,6 @@ class TestMixedPrecisionOps(unittest.TestCase): "params": {} } } - ops.MixedPrecisionOps._layer_quant_config = layer_quant_config # Create and load model fp8_weight = torch.randn(20, 10, dtype=torch.float32).to(torch.float8_e4m3fn) @@ -146,7 +145,8 @@ class TestMixedPrecisionOps(unittest.TestCase): "layer3.bias": torch.randn(40, dtype=torch.bfloat16), } - model = SimpleModel(operations=ops.MixedPrecisionOps) + state_dict1, _ = comfy.utils.convert_old_quants(state_dict1, metadata={"_quantization_metadata": json.dumps({"layers": layer_quant_config})}) + model = SimpleModel(operations=ops.mixed_precision_ops({})) model.load_state_dict(state_dict1, strict=False) # Save state dict @@ -154,8 +154,8 @@ class TestMixedPrecisionOps(unittest.TestCase): # Verify layer1.weight is a QuantizedTensor with scale preserved self.assertIsInstance(state_dict2["layer1.weight"], QuantizedTensor) - self.assertEqual(state_dict2["layer1.weight"]._layout_params['scale'].item(), 3.0) - self.assertEqual(state_dict2["layer1.weight"]._layout_type, "TensorCoreFP8Layout") + self.assertEqual(state_dict2["layer1.weight"]._params.scale.item(), 3.0) + self.assertEqual(state_dict2["layer1.weight"]._layout_cls, "TensorCoreFP8E4M3Layout") # Verify non-quantized layers are standard tensors self.assertNotIsInstance(state_dict2["layer2.weight"], QuantizedTensor) @@ -170,7 +170,6 @@ class TestMixedPrecisionOps(unittest.TestCase): "params": {} } } - ops.MixedPrecisionOps._layer_quant_config = layer_quant_config # Create and load model fp8_weight = torch.randn(20, 10, dtype=torch.float32).to(torch.float8_e4m3fn) @@ -184,7 +183,8 @@ class TestMixedPrecisionOps(unittest.TestCase): "layer3.bias": torch.randn(40, dtype=torch.bfloat16), } - model = SimpleModel(operations=ops.MixedPrecisionOps) + state_dict, _ = comfy.utils.convert_old_quants(state_dict, metadata={"_quantization_metadata": json.dumps({"layers": layer_quant_config})}) + model = SimpleModel(operations=ops.mixed_precision_ops({})) model.load_state_dict(state_dict, strict=False) # Add a weight function (simulating LoRA) @@ -210,7 +210,6 @@ class TestMixedPrecisionOps(unittest.TestCase): "params": {} } } - ops.MixedPrecisionOps._layer_quant_config = layer_quant_config # Create state dict state_dict = { @@ -222,8 +221,10 @@ class TestMixedPrecisionOps(unittest.TestCase): "layer3.bias": torch.randn(40, dtype=torch.bfloat16), } + state_dict, _ = comfy.utils.convert_old_quants(state_dict, metadata={"_quantization_metadata": json.dumps({"layers": layer_quant_config})}) + # Load should raise KeyError for unknown format in QUANT_FORMAT_MIXINS - model = SimpleModel(operations=ops.MixedPrecisionOps) + model = SimpleModel(operations=ops.mixed_precision_ops({})) with self.assertRaises(KeyError): model.load_state_dict(state_dict, strict=False) diff --git a/tests-unit/comfy_quant/test_quant_registry.py b/tests-unit/comfy_quant/test_quant_registry.py deleted file mode 100644 index 9cb54ede8..000000000 --- a/tests-unit/comfy_quant/test_quant_registry.py +++ /dev/null @@ -1,190 +0,0 @@ -import unittest -import torch -import sys -import os - -# Add comfy to path -sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "..")) - -def has_gpu(): - return torch.cuda.is_available() - -from comfy.cli_args import args -if not has_gpu(): - args.cpu = True - -from comfy.quant_ops import QuantizedTensor, TensorCoreFP8Layout - - -class TestQuantizedTensor(unittest.TestCase): - """Test the QuantizedTensor subclass with FP8 layout""" - - def test_creation(self): - """Test creating a QuantizedTensor with TensorCoreFP8Layout""" - fp8_data = torch.randn(256, 128, dtype=torch.float32).to(torch.float8_e4m3fn) - scale = torch.tensor(2.0) - layout_params = {'scale': scale, 'orig_dtype': torch.bfloat16} - - qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params) - - self.assertIsInstance(qt, QuantizedTensor) - self.assertEqual(qt.shape, (256, 128)) - self.assertEqual(qt.dtype, torch.float8_e4m3fn) - self.assertEqual(qt._layout_params['scale'], scale) - self.assertEqual(qt._layout_params['orig_dtype'], torch.bfloat16) - self.assertEqual(qt._layout_type, "TensorCoreFP8Layout") - - def test_dequantize(self): - """Test explicit dequantization""" - - fp8_data = torch.ones(10, 20, dtype=torch.float32).to(torch.float8_e4m3fn) - scale = torch.tensor(3.0) - layout_params = {'scale': scale, 'orig_dtype': torch.float32} - - qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params) - dequantized = qt.dequantize() - - self.assertEqual(dequantized.dtype, torch.float32) - self.assertTrue(torch.allclose(dequantized, torch.ones(10, 20) * 3.0, rtol=0.1)) - - def test_from_float(self): - """Test creating QuantizedTensor from float tensor""" - float_tensor = torch.randn(64, 32, dtype=torch.float32) - scale = torch.tensor(1.5) - - qt = QuantizedTensor.from_float( - float_tensor, - "TensorCoreFP8Layout", - scale=scale, - dtype=torch.float8_e4m3fn - ) - - self.assertIsInstance(qt, QuantizedTensor) - self.assertEqual(qt.dtype, torch.float8_e4m3fn) - self.assertEqual(qt.shape, (64, 32)) - - # Verify dequantization gives approximately original values - dequantized = qt.dequantize() - mean_rel_error = ((dequantized - float_tensor).abs() / (float_tensor.abs() + 1e-6)).mean() - self.assertLess(mean_rel_error, 0.1) - - -class TestGenericUtilities(unittest.TestCase): - """Test generic utility operations""" - - def test_detach(self): - """Test detach operation on quantized tensor""" - fp8_data = torch.randn(10, 20, dtype=torch.float32).to(torch.float8_e4m3fn) - scale = torch.tensor(1.5) - layout_params = {'scale': scale, 'orig_dtype': torch.float32} - qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params) - - # Detach should return a new QuantizedTensor - qt_detached = qt.detach() - - self.assertIsInstance(qt_detached, QuantizedTensor) - self.assertEqual(qt_detached.shape, qt.shape) - self.assertEqual(qt_detached._layout_type, "TensorCoreFP8Layout") - - def test_clone(self): - """Test clone operation on quantized tensor""" - fp8_data = torch.randn(10, 20, dtype=torch.float32).to(torch.float8_e4m3fn) - scale = torch.tensor(1.5) - layout_params = {'scale': scale, 'orig_dtype': torch.float32} - qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params) - - # Clone should return a new QuantizedTensor - qt_cloned = qt.clone() - - self.assertIsInstance(qt_cloned, QuantizedTensor) - self.assertEqual(qt_cloned.shape, qt.shape) - self.assertEqual(qt_cloned._layout_type, "TensorCoreFP8Layout") - - # Verify it's a deep copy - self.assertIsNot(qt_cloned._qdata, qt._qdata) - - @unittest.skipUnless(has_gpu(), "GPU not available") - def test_to_device(self): - """Test device transfer""" - fp8_data = torch.randn(10, 20, dtype=torch.float32).to(torch.float8_e4m3fn) - scale = torch.tensor(1.5) - layout_params = {'scale': scale, 'orig_dtype': torch.float32} - qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params) - - # Moving to same device should work (CPU to CPU) - qt_cpu = qt.to('cpu') - - self.assertIsInstance(qt_cpu, QuantizedTensor) - self.assertEqual(qt_cpu.device.type, 'cpu') - self.assertEqual(qt_cpu._layout_params['scale'].device.type, 'cpu') - - -class TestTensorCoreFP8Layout(unittest.TestCase): - """Test the TensorCoreFP8Layout implementation""" - - def test_quantize(self): - """Test quantization method""" - float_tensor = torch.randn(32, 64, dtype=torch.float32) - scale = torch.tensor(1.5) - - qdata, layout_params = TensorCoreFP8Layout.quantize( - float_tensor, - scale=scale, - dtype=torch.float8_e4m3fn - ) - - self.assertEqual(qdata.dtype, torch.float8_e4m3fn) - self.assertEqual(qdata.shape, float_tensor.shape) - self.assertIn('scale', layout_params) - self.assertIn('orig_dtype', layout_params) - self.assertEqual(layout_params['orig_dtype'], torch.float32) - - def test_dequantize(self): - """Test dequantization method""" - float_tensor = torch.ones(10, 20, dtype=torch.float32) * 3.0 - scale = torch.tensor(1.0) - - qdata, layout_params = TensorCoreFP8Layout.quantize( - float_tensor, - scale=scale, - dtype=torch.float8_e4m3fn - ) - - dequantized = TensorCoreFP8Layout.dequantize(qdata, **layout_params) - - # Should approximately match original - self.assertTrue(torch.allclose(dequantized, float_tensor, rtol=0.1, atol=0.1)) - - -class TestFallbackMechanism(unittest.TestCase): - """Test fallback for unsupported operations""" - - def test_unsupported_op_dequantizes(self): - """Test that unsupported operations fall back to dequantization""" - # Set seed for reproducibility - torch.manual_seed(42) - - # Create quantized tensor - a_fp32 = torch.randn(10, 20, dtype=torch.float32) - scale = torch.tensor(1.0) - a_q = QuantizedTensor.from_float( - a_fp32, - "TensorCoreFP8Layout", - scale=scale, - dtype=torch.float8_e4m3fn - ) - - # Call an operation that doesn't have a registered handler - # For example, torch.abs - result = torch.abs(a_q) - - # Should work via fallback (dequantize → abs → return) - self.assertNotIsInstance(result, QuantizedTensor) - expected = torch.abs(a_fp32) - # FP8 introduces quantization error, so use loose tolerance - mean_error = (result - expected).abs().mean() - self.assertLess(mean_error, 0.05, f"Mean error {mean_error:.4f} is too large") - - -if __name__ == "__main__": - unittest.main() diff --git a/tests-unit/execution_test/preview_method_override_test.py b/tests-unit/execution_test/preview_method_override_test.py new file mode 100644 index 000000000..79432d610 --- /dev/null +++ b/tests-unit/execution_test/preview_method_override_test.py @@ -0,0 +1,352 @@ +""" +Unit tests for Queue-specific Preview Method Override feature. + +Tests the preview method override functionality: +- LatentPreviewMethod.from_string() method +- set_preview_method() function in latent_preview.py +- default_preview_method variable +- Integration with args.preview_method +""" +import pytest +from comfy.cli_args import args, LatentPreviewMethod +from latent_preview import set_preview_method, default_preview_method + + +class TestLatentPreviewMethodFromString: + """Test LatentPreviewMethod.from_string() classmethod.""" + + @pytest.mark.parametrize("value,expected", [ + ("auto", LatentPreviewMethod.Auto), + ("latent2rgb", LatentPreviewMethod.Latent2RGB), + ("taesd", LatentPreviewMethod.TAESD), + ("none", LatentPreviewMethod.NoPreviews), + ]) + def test_valid_values_return_enum(self, value, expected): + """Valid string values should return corresponding enum.""" + assert LatentPreviewMethod.from_string(value) == expected + + @pytest.mark.parametrize("invalid", [ + "invalid", + "TAESD", # Case sensitive + "AUTO", # Case sensitive + "Latent2RGB", # Case sensitive + "latent", + "", + "default", # default is special, not a method + ]) + def test_invalid_values_return_none(self, invalid): + """Invalid string values should return None.""" + assert LatentPreviewMethod.from_string(invalid) is None + + +class TestLatentPreviewMethodEnumValues: + """Test LatentPreviewMethod enum has expected values.""" + + def test_enum_values(self): + """Verify enum values match expected strings.""" + assert LatentPreviewMethod.NoPreviews.value == "none" + assert LatentPreviewMethod.Auto.value == "auto" + assert LatentPreviewMethod.Latent2RGB.value == "latent2rgb" + assert LatentPreviewMethod.TAESD.value == "taesd" + + def test_enum_count(self): + """Verify exactly 4 preview methods exist.""" + assert len(LatentPreviewMethod) == 4 + + +class TestSetPreviewMethod: + """Test set_preview_method() function from latent_preview.py.""" + + def setup_method(self): + """Store original value before each test.""" + self.original = args.preview_method + + def teardown_method(self): + """Restore original value after each test.""" + args.preview_method = self.original + + def test_override_with_taesd(self): + """'taesd' should set args.preview_method to TAESD.""" + set_preview_method("taesd") + assert args.preview_method == LatentPreviewMethod.TAESD + + def test_override_with_latent2rgb(self): + """'latent2rgb' should set args.preview_method to Latent2RGB.""" + set_preview_method("latent2rgb") + assert args.preview_method == LatentPreviewMethod.Latent2RGB + + def test_override_with_auto(self): + """'auto' should set args.preview_method to Auto.""" + set_preview_method("auto") + assert args.preview_method == LatentPreviewMethod.Auto + + def test_override_with_none_value(self): + """'none' should set args.preview_method to NoPreviews.""" + set_preview_method("none") + assert args.preview_method == LatentPreviewMethod.NoPreviews + + def test_default_restores_original(self): + """'default' should restore to default_preview_method.""" + # First override to something else + set_preview_method("taesd") + assert args.preview_method == LatentPreviewMethod.TAESD + + # Then use 'default' to restore + set_preview_method("default") + assert args.preview_method == default_preview_method + + def test_none_param_restores_original(self): + """None parameter should restore to default_preview_method.""" + # First override to something else + set_preview_method("taesd") + assert args.preview_method == LatentPreviewMethod.TAESD + + # Then use None to restore + set_preview_method(None) + assert args.preview_method == default_preview_method + + def test_empty_string_restores_original(self): + """Empty string should restore to default_preview_method.""" + set_preview_method("taesd") + set_preview_method("") + assert args.preview_method == default_preview_method + + def test_invalid_value_restores_original(self): + """Invalid value should restore to default_preview_method.""" + set_preview_method("taesd") + set_preview_method("invalid_method") + assert args.preview_method == default_preview_method + + def test_case_sensitive_invalid_restores(self): + """Case-mismatched values should restore to default.""" + set_preview_method("taesd") + set_preview_method("TAESD") # Wrong case + assert args.preview_method == default_preview_method + + +class TestDefaultPreviewMethod: + """Test default_preview_method module variable.""" + + def test_default_is_not_none(self): + """default_preview_method should not be None.""" + assert default_preview_method is not None + + def test_default_is_enum_member(self): + """default_preview_method should be a LatentPreviewMethod enum.""" + assert isinstance(default_preview_method, LatentPreviewMethod) + + def test_default_matches_args_initial(self): + """default_preview_method should match CLI default or user setting.""" + # This tests that default_preview_method was captured at module load + # After set_preview_method(None), args should equal default + original = args.preview_method + set_preview_method("taesd") + set_preview_method(None) + assert args.preview_method == default_preview_method + args.preview_method = original + + +class TestArgsPreviewMethodModification: + """Test args.preview_method can be modified correctly.""" + + def setup_method(self): + """Store original value before each test.""" + self.original = args.preview_method + + def teardown_method(self): + """Restore original value after each test.""" + args.preview_method = self.original + + def test_args_accepts_all_enum_values(self): + """args.preview_method should accept all LatentPreviewMethod values.""" + for method in LatentPreviewMethod: + args.preview_method = method + assert args.preview_method == method + + def test_args_modification_and_restoration(self): + """args.preview_method should be modifiable and restorable.""" + original = args.preview_method + + args.preview_method = LatentPreviewMethod.TAESD + assert args.preview_method == LatentPreviewMethod.TAESD + + args.preview_method = original + assert args.preview_method == original + + +class TestExecutionFlow: + """Test the execution flow pattern used in execution.py.""" + + def setup_method(self): + """Store original value before each test.""" + self.original = args.preview_method + + def teardown_method(self): + """Restore original value after each test.""" + args.preview_method = self.original + + def test_sequential_executions_with_different_methods(self): + """Simulate multiple queue executions with different preview methods.""" + # Execution 1: taesd + set_preview_method("taesd") + assert args.preview_method == LatentPreviewMethod.TAESD + + # Execution 2: none + set_preview_method("none") + assert args.preview_method == LatentPreviewMethod.NoPreviews + + # Execution 3: default (restore) + set_preview_method("default") + assert args.preview_method == default_preview_method + + # Execution 4: auto + set_preview_method("auto") + assert args.preview_method == LatentPreviewMethod.Auto + + # Execution 5: no override (None) + set_preview_method(None) + assert args.preview_method == default_preview_method + + def test_override_then_default_pattern(self): + """Test the pattern: override -> execute -> next call restores.""" + # First execution with override + set_preview_method("latent2rgb") + assert args.preview_method == LatentPreviewMethod.Latent2RGB + + # Second execution without override restores default + set_preview_method(None) + assert args.preview_method == default_preview_method + + def test_extra_data_simulation(self): + """Simulate extra_data.get('preview_method') patterns.""" + # Simulate: extra_data = {"preview_method": "taesd"} + extra_data = {"preview_method": "taesd"} + set_preview_method(extra_data.get("preview_method")) + assert args.preview_method == LatentPreviewMethod.TAESD + + # Simulate: extra_data = {} + extra_data = {} + set_preview_method(extra_data.get("preview_method")) + assert args.preview_method == default_preview_method + + # Simulate: extra_data = {"preview_method": "default"} + extra_data = {"preview_method": "default"} + set_preview_method(extra_data.get("preview_method")) + assert args.preview_method == default_preview_method + + +class TestRealWorldScenarios: + """Tests using real-world prompt data patterns.""" + + def setup_method(self): + """Store original value before each test.""" + self.original = args.preview_method + + def teardown_method(self): + """Restore original value after each test.""" + args.preview_method = self.original + + def test_captured_prompt_without_preview_method(self): + """ + Test with captured prompt that has no preview_method. + Based on: tests-unit/execution_test/fixtures/default_prompt.json + """ + # Real captured extra_data structure (preview_method absent) + extra_data = { + "extra_pnginfo": {"workflow": {}}, + "client_id": "271314f0dabd48e5aaa488ed7a4ceb0d", + "create_time": 1765416558179 + } + + set_preview_method(extra_data.get("preview_method")) + assert args.preview_method == default_preview_method + + def test_captured_prompt_with_preview_method_taesd(self): + """Test captured prompt with preview_method: taesd.""" + extra_data = { + "extra_pnginfo": {"workflow": {}}, + "client_id": "271314f0dabd48e5aaa488ed7a4ceb0d", + "preview_method": "taesd" + } + + set_preview_method(extra_data.get("preview_method")) + assert args.preview_method == LatentPreviewMethod.TAESD + + def test_captured_prompt_with_preview_method_none(self): + """Test captured prompt with preview_method: none (disable preview).""" + extra_data = { + "extra_pnginfo": {"workflow": {}}, + "client_id": "test-client", + "preview_method": "none" + } + + set_preview_method(extra_data.get("preview_method")) + assert args.preview_method == LatentPreviewMethod.NoPreviews + + def test_captured_prompt_with_preview_method_latent2rgb(self): + """Test captured prompt with preview_method: latent2rgb.""" + extra_data = { + "extra_pnginfo": {"workflow": {}}, + "client_id": "test-client", + "preview_method": "latent2rgb" + } + + set_preview_method(extra_data.get("preview_method")) + assert args.preview_method == LatentPreviewMethod.Latent2RGB + + def test_captured_prompt_with_preview_method_auto(self): + """Test captured prompt with preview_method: auto.""" + extra_data = { + "extra_pnginfo": {"workflow": {}}, + "client_id": "test-client", + "preview_method": "auto" + } + + set_preview_method(extra_data.get("preview_method")) + assert args.preview_method == LatentPreviewMethod.Auto + + def test_captured_prompt_with_preview_method_default(self): + """Test captured prompt with preview_method: default (use CLI setting).""" + # First set to something else + set_preview_method("taesd") + assert args.preview_method == LatentPreviewMethod.TAESD + + # Then simulate a prompt with "default" + extra_data = { + "extra_pnginfo": {"workflow": {}}, + "client_id": "test-client", + "preview_method": "default" + } + + set_preview_method(extra_data.get("preview_method")) + assert args.preview_method == default_preview_method + + def test_sequential_queue_with_different_preview_methods(self): + """ + Simulate real queue scenario: multiple prompts with different settings. + This tests the actual usage pattern in ComfyUI. + """ + # Queue 1: User wants TAESD preview + extra_data_1 = {"client_id": "client-1", "preview_method": "taesd"} + set_preview_method(extra_data_1.get("preview_method")) + assert args.preview_method == LatentPreviewMethod.TAESD + + # Queue 2: User wants no preview (faster execution) + extra_data_2 = {"client_id": "client-2", "preview_method": "none"} + set_preview_method(extra_data_2.get("preview_method")) + assert args.preview_method == LatentPreviewMethod.NoPreviews + + # Queue 3: User doesn't specify (use server default) + extra_data_3 = {"client_id": "client-3"} + set_preview_method(extra_data_3.get("preview_method")) + assert args.preview_method == default_preview_method + + # Queue 4: User explicitly wants default + extra_data_4 = {"client_id": "client-4", "preview_method": "default"} + set_preview_method(extra_data_4.get("preview_method")) + assert args.preview_method == default_preview_method + + # Queue 5: User wants latent2rgb + extra_data_5 = {"client_id": "client-5", "preview_method": "latent2rgb"} + set_preview_method(extra_data_5.get("preview_method")) + assert args.preview_method == LatentPreviewMethod.Latent2RGB diff --git a/tests-unit/folder_paths_test/system_user_test.py b/tests-unit/folder_paths_test/system_user_test.py new file mode 100644 index 000000000..cd46459f1 --- /dev/null +++ b/tests-unit/folder_paths_test/system_user_test.py @@ -0,0 +1,206 @@ +"""Tests for System User Protection in folder_paths.py + +Tests cover: +- get_system_user_directory(): Internal API for custom nodes to access System User directories +- get_public_user_directory(): HTTP endpoint access with System User blocking +- Backward compatibility: Existing APIs unchanged +- Security: Path traversal and injection prevention +""" + +import pytest +import os +import tempfile + +from folder_paths import ( + get_system_user_directory, + get_public_user_directory, + get_user_directory, + set_user_directory, +) + + +@pytest.fixture(scope="module") +def mock_user_directory(): + """Create a temporary user directory for testing.""" + with tempfile.TemporaryDirectory() as temp_dir: + original_dir = get_user_directory() + set_user_directory(temp_dir) + yield temp_dir + set_user_directory(original_dir) + + +class TestGetSystemUserDirectory: + """Tests for get_system_user_directory() - internal API for System User directories. + + Verifies: + - Custom nodes can access System User directories via internal API + - Input validation prevents path traversal attacks + """ + + def test_default_name(self, mock_user_directory): + """Test default 'system' name.""" + path = get_system_user_directory() + assert path.endswith("__system") + assert mock_user_directory in path + + def test_custom_name(self, mock_user_directory): + """Test custom system user name.""" + path = get_system_user_directory("cache") + assert path.endswith("__cache") + assert "__cache" in path + + def test_name_with_underscore(self, mock_user_directory): + """Test name with underscore in middle.""" + path = get_system_user_directory("my_cache") + assert "__my_cache" in path + + def test_empty_name_raises(self): + """Test empty name raises ValueError.""" + with pytest.raises(ValueError, match="cannot be empty"): + get_system_user_directory("") + + def test_none_name_raises(self): + """Test None name raises ValueError.""" + with pytest.raises(ValueError, match="cannot be empty"): + get_system_user_directory(None) + + def test_name_starting_with_underscore_raises(self): + """Test name starting with underscore raises ValueError.""" + with pytest.raises(ValueError, match="should not start with underscore"): + get_system_user_directory("_system") + + def test_path_traversal_raises(self): + """Test path traversal attempt raises ValueError (security).""" + with pytest.raises(ValueError, match="Invalid system user name"): + get_system_user_directory("../escape") + + def test_path_traversal_middle_raises(self): + """Test path traversal in middle raises ValueError (security).""" + with pytest.raises(ValueError, match="Invalid system user name"): + get_system_user_directory("system/../other") + + def test_special_chars_raise(self): + """Test special characters raise ValueError (security).""" + with pytest.raises(ValueError, match="Invalid system user name"): + get_system_user_directory("system!") + + def test_returns_absolute_path(self, mock_user_directory): + """Test returned path is absolute.""" + path = get_system_user_directory("test") + assert os.path.isabs(path) + + +class TestGetPublicUserDirectory: + """Tests for get_public_user_directory() - HTTP endpoint access with System User blocking. + + Verifies: + - System Users (__ prefix) return None, blocking HTTP access + - Public Users get valid paths + - New endpoints using this function are automatically protected + """ + + def test_normal_user(self, mock_user_directory): + """Test normal user returns valid path.""" + path = get_public_user_directory("default") + assert path is not None + assert "default" in path + assert mock_user_directory in path + + def test_system_user_returns_none(self): + """Test System User (__ prefix) returns None - blocks HTTP access.""" + assert get_public_user_directory("__system") is None + + def test_system_user_cache_returns_none(self): + """Test System User cache returns None.""" + assert get_public_user_directory("__cache") is None + + def test_empty_user_returns_none(self): + """Test empty user returns None.""" + assert get_public_user_directory("") is None + + def test_none_user_returns_none(self): + """Test None user returns None.""" + assert get_public_user_directory(None) is None + + def test_header_injection_returns_none(self): + """Test header injection attempt returns None (security).""" + assert get_public_user_directory("__system\r\nX-Injected: true") is None + + def test_null_byte_injection_returns_none(self): + """Test null byte injection handling (security).""" + # Note: startswith check happens before any path operations + result = get_public_user_directory("user\x00__system") + # This should return a path since it doesn't start with __ + # The actual security comes from the path not being __* + assert result is not None or result is None # Depends on validation + + def test_path_traversal_attempt(self, mock_user_directory): + """Test path traversal attempt handling.""" + # This function doesn't validate paths, only reserved prefix + # Path traversal should be handled by the caller + path = get_public_user_directory("../../../etc/passwd") + # Returns path but doesn't start with __, so not None + # Actual path validation happens in user_manager + assert path is not None or "__" not in "../../../etc/passwd" + + def test_returns_absolute_path(self, mock_user_directory): + """Test returned path is absolute.""" + path = get_public_user_directory("testuser") + assert path is not None + assert os.path.isabs(path) + + +class TestBackwardCompatibility: + """Tests for backward compatibility with existing APIs. + + Verifies: + - get_user_directory() API unchanged + - Existing user data remains accessible + """ + + def test_get_user_directory_unchanged(self, mock_user_directory): + """Test get_user_directory() still works as before.""" + user_dir = get_user_directory() + assert user_dir is not None + assert os.path.isabs(user_dir) + assert user_dir == mock_user_directory + + def test_existing_user_accessible(self, mock_user_directory): + """Test existing users can access their directories.""" + path = get_public_user_directory("default") + assert path is not None + assert "default" in path + + +class TestEdgeCases: + """Tests for edge cases in System User detection. + + Verifies: + - Only __ prefix is blocked (not _, not middle __) + - Bypass attempts are prevented + """ + + def test_prefix_only(self): + """Test prefix-only string is blocked.""" + assert get_public_user_directory("__") is None + + def test_single_underscore_allowed(self): + """Test single underscore prefix is allowed (not System User).""" + path = get_public_user_directory("_system") + assert path is not None + assert "_system" in path + + def test_triple_underscore_blocked(self): + """Test triple underscore is blocked (starts with __).""" + assert get_public_user_directory("___system") is None + + def test_underscore_in_middle_allowed(self): + """Test underscore in middle is allowed.""" + path = get_public_user_directory("my__system") + assert path is not None + assert "my__system" in path + + def test_leading_space_allowed(self): + """Test leading space + prefix is allowed (doesn't start with __).""" + path = get_public_user_directory(" __system") + assert path is not None diff --git a/tests-unit/prompt_server_test/system_user_endpoint_test.py b/tests-unit/prompt_server_test/system_user_endpoint_test.py new file mode 100644 index 000000000..22ac00af9 --- /dev/null +++ b/tests-unit/prompt_server_test/system_user_endpoint_test.py @@ -0,0 +1,375 @@ +"""E2E Tests for System User Protection HTTP Endpoints + +Tests cover: +- HTTP endpoint blocking: System Users cannot access /userdata (GET, POST, DELETE, move) +- User creation blocking: System User names cannot be created via POST /users +- Backward compatibility: Public Users work as before +- Custom node scenario: Internal API works while HTTP is blocked +- Structural security: get_public_user_directory() provides automatic protection +""" + +import pytest +import os +from aiohttp import web +from app.user_manager import UserManager +from unittest.mock import patch +import folder_paths + + +@pytest.fixture +def mock_user_directory(tmp_path): + """Create a temporary user directory.""" + original_dir = folder_paths.get_user_directory() + folder_paths.set_user_directory(str(tmp_path)) + yield tmp_path + folder_paths.set_user_directory(original_dir) + + +@pytest.fixture +def user_manager_multi_user(mock_user_directory): + """Create UserManager in multi-user mode.""" + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + um = UserManager() + # Add test users + um.users = {"default": "default", "test_user_123": "Test User"} + yield um + + +@pytest.fixture +def app_multi_user(user_manager_multi_user): + """Create app with multi-user mode enabled.""" + app = web.Application() + routes = web.RouteTableDef() + user_manager_multi_user.add_routes(routes) + app.add_routes(routes) + return app + + +class TestSystemUserEndpointBlocking: + """E2E tests for System User blocking on all HTTP endpoints. + + Verifies: + - GET /userdata blocked for System Users + - POST /userdata blocked for System Users + - DELETE /userdata blocked for System Users + - POST /userdata/.../move/... blocked for System Users + """ + + @pytest.mark.asyncio + async def test_userdata_get_blocks_system_user( + self, aiohttp_client, app_multi_user, mock_user_directory + ): + """ + GET /userdata with System User header should be blocked. + """ + # Create test directory for System User (simulating internal creation) + system_user_dir = mock_user_directory / "__system" + system_user_dir.mkdir() + (system_user_dir / "secret.txt").write_text("sensitive data") + + client = await aiohttp_client(app_multi_user) + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + # Attempt to access System User's data via HTTP + resp = await client.get( + "/userdata?dir=.", + headers={"comfy-user": "__system"} + ) + + # Should be blocked (403 Forbidden or similar error) + assert resp.status in [400, 403, 500], \ + f"System User access should be blocked, got {resp.status}" + + @pytest.mark.asyncio + async def test_userdata_post_blocks_system_user( + self, aiohttp_client, app_multi_user, mock_user_directory + ): + """ + POST /userdata with System User header should be blocked. + """ + client = await aiohttp_client(app_multi_user) + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + resp = await client.post( + "/userdata/test.txt", + headers={"comfy-user": "__system"}, + data=b"malicious content" + ) + + assert resp.status in [400, 403, 500], \ + f"System User write should be blocked, got {resp.status}" + + # Verify no file was created + assert not (mock_user_directory / "__system" / "test.txt").exists() + + @pytest.mark.asyncio + async def test_userdata_delete_blocks_system_user( + self, aiohttp_client, app_multi_user, mock_user_directory + ): + """ + DELETE /userdata with System User header should be blocked. + """ + # Create a file in System User directory + system_user_dir = mock_user_directory / "__system" + system_user_dir.mkdir() + secret_file = system_user_dir / "secret.txt" + secret_file.write_text("do not delete") + + client = await aiohttp_client(app_multi_user) + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + resp = await client.delete( + "/userdata/secret.txt", + headers={"comfy-user": "__system"} + ) + + assert resp.status in [400, 403, 500], \ + f"System User delete should be blocked, got {resp.status}" + + # Verify file still exists + assert secret_file.exists() + + @pytest.mark.asyncio + async def test_v2_userdata_blocks_system_user( + self, aiohttp_client, app_multi_user, mock_user_directory + ): + """ + GET /v2/userdata with System User header should be blocked. + """ + client = await aiohttp_client(app_multi_user) + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + resp = await client.get( + "/v2/userdata", + headers={"comfy-user": "__system"} + ) + + assert resp.status in [400, 403, 500], \ + f"System User v2 access should be blocked, got {resp.status}" + + @pytest.mark.asyncio + async def test_move_userdata_blocks_system_user( + self, aiohttp_client, app_multi_user, mock_user_directory + ): + """ + POST /userdata/{file}/move/{dest} with System User header should be blocked. + """ + system_user_dir = mock_user_directory / "__system" + system_user_dir.mkdir() + (system_user_dir / "source.txt").write_text("sensitive data") + + client = await aiohttp_client(app_multi_user) + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + resp = await client.post( + "/userdata/source.txt/move/dest.txt", + headers={"comfy-user": "__system"} + ) + + assert resp.status in [400, 403, 500], \ + f"System User move should be blocked, got {resp.status}" + + # Verify source file still exists (move was blocked) + assert (system_user_dir / "source.txt").exists() + + +class TestSystemUserCreationBlocking: + """E2E tests for blocking System User name creation via POST /users. + + Verifies: + - POST /users returns 400 for System User name (not 500) + """ + + @pytest.mark.asyncio + async def test_post_users_blocks_system_user_name( + self, aiohttp_client, app_multi_user + ): + """POST /users with System User name should return 400 Bad Request.""" + client = await aiohttp_client(app_multi_user) + + resp = await client.post( + "/users", + json={"username": "__system"} + ) + + assert resp.status == 400, \ + f"System User creation should return 400, got {resp.status}" + + @pytest.mark.asyncio + async def test_post_users_blocks_system_user_prefix_variations( + self, aiohttp_client, app_multi_user + ): + """POST /users with any System User prefix variation should return 400 Bad Request.""" + client = await aiohttp_client(app_multi_user) + + system_user_names = ["__system", "__cache", "__config", "__anything"] + + for name in system_user_names: + resp = await client.post("/users", json={"username": name}) + assert resp.status == 400, \ + f"System User name '{name}' should return 400, got {resp.status}" + + +class TestPublicUserStillWorks: + """E2E tests for backward compatibility - Public Users should work as before. + + Verifies: + - Public Users can access their data via HTTP + - Public Users can create files via HTTP + """ + + @pytest.mark.asyncio + async def test_public_user_can_access_userdata( + self, aiohttp_client, app_multi_user, mock_user_directory + ): + """ + Public Users should still be able to access their data. + """ + # Create test directory for Public User + user_dir = mock_user_directory / "default" + user_dir.mkdir() + test_dir = user_dir / "workflows" + test_dir.mkdir() + (test_dir / "test.json").write_text('{"test": true}') + + client = await aiohttp_client(app_multi_user) + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + resp = await client.get( + "/userdata?dir=workflows", + headers={"comfy-user": "default"} + ) + + assert resp.status == 200 + data = await resp.json() + assert "test.json" in data + + @pytest.mark.asyncio + async def test_public_user_can_create_files( + self, aiohttp_client, app_multi_user, mock_user_directory + ): + """ + Public Users should still be able to create files. + """ + # Create user directory + user_dir = mock_user_directory / "default" + user_dir.mkdir() + + client = await aiohttp_client(app_multi_user) + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + resp = await client.post( + "/userdata/newfile.txt", + headers={"comfy-user": "default"}, + data=b"user content" + ) + + assert resp.status == 200 + assert (user_dir / "newfile.txt").exists() + + +class TestCustomNodeScenario: + """Tests for custom node use case: internal API access vs HTTP blocking. + + Verifies: + - Internal API (get_system_user_directory) works for custom nodes + - HTTP endpoint cannot access data created via internal API + """ + + def test_internal_api_can_access_system_user(self, mock_user_directory): + """ + Internal API (get_system_user_directory) should work for custom nodes. + """ + # Custom node uses internal API + system_path = folder_paths.get_system_user_directory("mynode_config") + + assert system_path is not None + assert "__mynode_config" in system_path + + # Can create and write to System User directory + os.makedirs(system_path, exist_ok=True) + config_file = os.path.join(system_path, "settings.json") + with open(config_file, "w") as f: + f.write('{"api_key": "secret"}') + + assert os.path.exists(config_file) + + @pytest.mark.asyncio + async def test_http_cannot_access_internal_data( + self, aiohttp_client, app_multi_user, mock_user_directory + ): + """ + HTTP endpoint cannot access data created via internal API. + """ + # Custom node creates data via internal API + system_path = folder_paths.get_system_user_directory("mynode_config") + os.makedirs(system_path, exist_ok=True) + with open(os.path.join(system_path, "secret.json"), "w") as f: + f.write('{"api_key": "secret"}') + + client = await aiohttp_client(app_multi_user) + + # Attacker tries to access via HTTP + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + resp = await client.get( + "/userdata/secret.json", + headers={"comfy-user": "__mynode_config"} + ) + + # Should be blocked + assert resp.status in [400, 403, 500] + + +class TestStructuralSecurity: + """Tests for structural security pattern. + + Verifies: + - get_public_user_directory() automatically blocks System Users + - New endpoints using this function are automatically protected + """ + + def test_get_public_user_directory_blocks_system_user(self): + """ + Any code using get_public_user_directory() is automatically protected. + """ + # This is the structural security - any new endpoint using this function + # will automatically block System Users + assert folder_paths.get_public_user_directory("__system") is None + assert folder_paths.get_public_user_directory("__cache") is None + assert folder_paths.get_public_user_directory("__anything") is None + + # Public Users work + assert folder_paths.get_public_user_directory("default") is not None + assert folder_paths.get_public_user_directory("user123") is not None + + def test_structural_security_pattern(self, mock_user_directory): + """ + Demonstrate the structural security pattern for new endpoints. + + Any new endpoint should follow this pattern: + 1. Get user from request + 2. Use get_public_user_directory() - automatically blocks System Users + 3. If None, return error + """ + def new_endpoint_handler(user_id: str) -> str | None: + """Example of how new endpoints should be implemented.""" + user_path = folder_paths.get_public_user_directory(user_id) + if user_path is None: + return None # Blocked + return user_path + + # System Users are automatically blocked + assert new_endpoint_handler("__system") is None + assert new_endpoint_handler("__secret") is None + + # Public Users work + assert new_endpoint_handler("default") is not None diff --git a/tests/execution/test_execution.py b/tests/execution/test_execution.py index ace0d2279..f73ca7e3c 100644 --- a/tests/execution/test_execution.py +++ b/tests/execution/test_execution.py @@ -99,6 +99,37 @@ class ComfyClient: with urllib.request.urlopen(url) as response: return json.loads(response.read()) + def get_jobs(self, status=None, limit=None, offset=None, sort_by=None, sort_order=None): + url = "http://{}/api/jobs".format(self.server_address) + params = {} + if status is not None: + params["status"] = status + if limit is not None: + params["limit"] = limit + if offset is not None: + params["offset"] = offset + if sort_by is not None: + params["sort_by"] = sort_by + if sort_order is not None: + params["sort_order"] = sort_order + + if params: + url_values = urllib.parse.urlencode(params) + url = "{}?{}".format(url, url_values) + + with urllib.request.urlopen(url) as response: + return json.loads(response.read()) + + def get_job(self, job_id): + url = "http://{}/api/jobs/{}".format(self.server_address, job_id) + try: + with urllib.request.urlopen(url) as response: + return json.loads(response.read()) + except urllib.error.HTTPError as e: + if e.code == 404: + return None + raise + def set_test_name(self, name): self.test_name = name @@ -877,3 +908,106 @@ class TestExecution: result = client.get_all_history(max_items=5, offset=len(all_history) - 1) assert len(result) <= 1, "Should return at most 1 item when offset is near end" + + # Jobs API tests + def test_jobs_api_job_structure( + self, client: ComfyClient, builder: GraphBuilder + ): + """Test that job objects have required fields""" + self._create_history_item(client, builder) + + jobs_response = client.get_jobs(status="completed", limit=1) + assert len(jobs_response["jobs"]) > 0, "Should have at least one job" + + job = jobs_response["jobs"][0] + assert "id" in job, "Job should have id" + assert "status" in job, "Job should have status" + assert "create_time" in job, "Job should have create_time" + assert "outputs_count" in job, "Job should have outputs_count" + assert "preview_output" in job, "Job should have preview_output" + + def test_jobs_api_preview_output_structure( + self, client: ComfyClient, builder: GraphBuilder + ): + """Test that preview_output has correct structure""" + self._create_history_item(client, builder) + + jobs_response = client.get_jobs(status="completed", limit=1) + job = jobs_response["jobs"][0] + + if job["preview_output"] is not None: + preview = job["preview_output"] + assert "filename" in preview, "Preview should have filename" + assert "nodeId" in preview, "Preview should have nodeId" + assert "mediaType" in preview, "Preview should have mediaType" + + def test_jobs_api_pagination( + self, client: ComfyClient, builder: GraphBuilder + ): + """Test jobs API pagination""" + for _ in range(5): + self._create_history_item(client, builder) + + first_page = client.get_jobs(limit=2, offset=0) + second_page = client.get_jobs(limit=2, offset=2) + + assert len(first_page["jobs"]) <= 2, "First page should have at most 2 jobs" + assert len(second_page["jobs"]) <= 2, "Second page should have at most 2 jobs" + + first_ids = {j["id"] for j in first_page["jobs"]} + second_ids = {j["id"] for j in second_page["jobs"]} + assert first_ids.isdisjoint(second_ids), "Pages should have different jobs" + + def test_jobs_api_sorting( + self, client: ComfyClient, builder: GraphBuilder + ): + """Test jobs API sorting""" + for _ in range(3): + self._create_history_item(client, builder) + + desc_jobs = client.get_jobs(sort_order="desc") + asc_jobs = client.get_jobs(sort_order="asc") + + if len(desc_jobs["jobs"]) >= 2: + desc_times = [j["create_time"] for j in desc_jobs["jobs"] if j["create_time"]] + asc_times = [j["create_time"] for j in asc_jobs["jobs"] if j["create_time"]] + if len(desc_times) >= 2: + assert desc_times == sorted(desc_times, reverse=True), "Desc should be newest first" + if len(asc_times) >= 2: + assert asc_times == sorted(asc_times), "Asc should be oldest first" + + def test_jobs_api_status_filter( + self, client: ComfyClient, builder: GraphBuilder + ): + """Test jobs API status filtering""" + self._create_history_item(client, builder) + + completed_jobs = client.get_jobs(status="completed") + assert len(completed_jobs["jobs"]) > 0, "Should have completed jobs from history" + + for job in completed_jobs["jobs"]: + assert job["status"] == "completed", "Should only return completed jobs" + + # Pending jobs are transient - just verify filter doesn't error + pending_jobs = client.get_jobs(status="pending") + for job in pending_jobs["jobs"]: + assert job["status"] == "pending", "Should only return pending jobs" + + def test_get_job_by_id( + self, client: ComfyClient, builder: GraphBuilder + ): + """Test getting a single job by ID""" + result = self._create_history_item(client, builder) + prompt_id = result.get_prompt_id() + + job = client.get_job(prompt_id) + assert job is not None, "Should find the job" + assert job["id"] == prompt_id, "Job ID should match" + assert "outputs" in job, "Single job should include outputs" + + def test_get_job_not_found( + self, client: ComfyClient, builder: GraphBuilder + ): + """Test getting a non-existent job returns 404""" + job = client.get_job("nonexistent-job-id") + assert job is None, "Non-existent job should return None" diff --git a/tests/execution/test_jobs.py b/tests/execution/test_jobs.py new file mode 100644 index 000000000..918c8080a --- /dev/null +++ b/tests/execution/test_jobs.py @@ -0,0 +1,361 @@ +"""Unit tests for comfy_execution/jobs.py""" + +from comfy_execution.jobs import ( + JobStatus, + is_previewable, + normalize_queue_item, + normalize_history_item, + get_outputs_summary, + apply_sorting, +) + + +class TestJobStatus: + """Test JobStatus constants.""" + + def test_status_values(self): + """Status constants should have expected string values.""" + assert JobStatus.PENDING == 'pending' + assert JobStatus.IN_PROGRESS == 'in_progress' + assert JobStatus.COMPLETED == 'completed' + assert JobStatus.FAILED == 'failed' + + def test_all_contains_all_statuses(self): + """ALL should contain all status values.""" + assert JobStatus.PENDING in JobStatus.ALL + assert JobStatus.IN_PROGRESS in JobStatus.ALL + assert JobStatus.COMPLETED in JobStatus.ALL + assert JobStatus.FAILED in JobStatus.ALL + assert len(JobStatus.ALL) == 4 + + +class TestIsPreviewable: + """Unit tests for is_previewable()""" + + def test_previewable_media_types(self): + """Images, video, audio media types should be previewable.""" + for media_type in ['images', 'video', 'audio']: + assert is_previewable(media_type, {}) is True + + def test_non_previewable_media_types(self): + """Other media types should not be previewable.""" + for media_type in ['latents', 'text', 'metadata', 'files']: + assert is_previewable(media_type, {}) is False + + def test_3d_extensions_previewable(self): + """3D file extensions should be previewable regardless of media_type.""" + for ext in ['.obj', '.fbx', '.gltf', '.glb']: + item = {'filename': f'model{ext}'} + assert is_previewable('files', item) is True + + def test_3d_extensions_case_insensitive(self): + """3D extension check should be case insensitive.""" + item = {'filename': 'MODEL.GLB'} + assert is_previewable('files', item) is True + + def test_video_format_previewable(self): + """Items with video/ format should be previewable.""" + item = {'format': 'video/mp4'} + assert is_previewable('files', item) is True + + def test_audio_format_previewable(self): + """Items with audio/ format should be previewable.""" + item = {'format': 'audio/wav'} + assert is_previewable('files', item) is True + + def test_other_format_not_previewable(self): + """Items with other format should not be previewable.""" + item = {'format': 'application/json'} + assert is_previewable('files', item) is False + + +class TestGetOutputsSummary: + """Unit tests for get_outputs_summary()""" + + def test_empty_outputs(self): + """Empty outputs should return 0 count and None preview.""" + count, preview = get_outputs_summary({}) + assert count == 0 + assert preview is None + + def test_counts_across_multiple_nodes(self): + """Outputs from multiple nodes should all be counted.""" + outputs = { + 'node1': {'images': [{'filename': 'a.png', 'type': 'output'}]}, + 'node2': {'images': [{'filename': 'b.png', 'type': 'output'}]}, + 'node3': {'images': [ + {'filename': 'c.png', 'type': 'output'}, + {'filename': 'd.png', 'type': 'output'} + ]} + } + count, preview = get_outputs_summary(outputs) + assert count == 4 + + def test_skips_animated_key_and_non_list_values(self): + """The 'animated' key and non-list values should be skipped.""" + outputs = { + 'node1': { + 'images': [{'filename': 'test.png', 'type': 'output'}], + 'animated': [True], # Should skip due to key name + 'metadata': 'string', # Should skip due to non-list + 'count': 42 # Should skip due to non-list + } + } + count, preview = get_outputs_summary(outputs) + assert count == 1 + + def test_preview_prefers_type_output(self): + """Items with type='output' should be preferred for preview.""" + outputs = { + 'node1': { + 'images': [ + {'filename': 'temp.png', 'type': 'temp'}, + {'filename': 'output.png', 'type': 'output'} + ] + } + } + count, preview = get_outputs_summary(outputs) + assert count == 2 + assert preview['filename'] == 'output.png' + + def test_preview_fallback_when_no_output_type(self): + """If no type='output', should use first previewable.""" + outputs = { + 'node1': { + 'images': [ + {'filename': 'temp1.png', 'type': 'temp'}, + {'filename': 'temp2.png', 'type': 'temp'} + ] + } + } + count, preview = get_outputs_summary(outputs) + assert preview['filename'] == 'temp1.png' + + def test_non_previewable_media_types_counted_but_no_preview(self): + """Non-previewable media types should be counted but not used as preview.""" + outputs = { + 'node1': { + 'latents': [ + {'filename': 'latent1.safetensors'}, + {'filename': 'latent2.safetensors'} + ] + } + } + count, preview = get_outputs_summary(outputs) + assert count == 2 + assert preview is None + + def test_previewable_media_types(self): + """Images, video, and audio media types should be previewable.""" + for media_type in ['images', 'video', 'audio']: + outputs = { + 'node1': { + media_type: [{'filename': 'test.file', 'type': 'output'}] + } + } + count, preview = get_outputs_summary(outputs) + assert preview is not None, f"{media_type} should be previewable" + + def test_3d_files_previewable(self): + """3D file extensions should be previewable.""" + for ext in ['.obj', '.fbx', '.gltf', '.glb']: + outputs = { + 'node1': { + 'files': [{'filename': f'model{ext}', 'type': 'output'}] + } + } + count, preview = get_outputs_summary(outputs) + assert preview is not None, f"3D file {ext} should be previewable" + + def test_format_mime_type_previewable(self): + """Files with video/ or audio/ format should be previewable.""" + for fmt in ['video/x-custom', 'audio/x-custom']: + outputs = { + 'node1': { + 'files': [{'filename': 'file.custom', 'format': fmt, 'type': 'output'}] + } + } + count, preview = get_outputs_summary(outputs) + assert preview is not None, f"Format {fmt} should be previewable" + + def test_preview_enriched_with_node_metadata(self): + """Preview should include nodeId, mediaType, and original fields.""" + outputs = { + 'node123': { + 'images': [{'filename': 'test.png', 'type': 'output', 'subfolder': 'outputs'}] + } + } + count, preview = get_outputs_summary(outputs) + assert preview['nodeId'] == 'node123' + assert preview['mediaType'] == 'images' + assert preview['subfolder'] == 'outputs' + + +class TestApplySorting: + """Unit tests for apply_sorting()""" + + def test_sort_by_create_time_desc(self): + """Default sort by create_time descending.""" + jobs = [ + {'id': 'a', 'create_time': 100}, + {'id': 'b', 'create_time': 300}, + {'id': 'c', 'create_time': 200}, + ] + result = apply_sorting(jobs, 'created_at', 'desc') + assert [j['id'] for j in result] == ['b', 'c', 'a'] + + def test_sort_by_create_time_asc(self): + """Sort by create_time ascending.""" + jobs = [ + {'id': 'a', 'create_time': 100}, + {'id': 'b', 'create_time': 300}, + {'id': 'c', 'create_time': 200}, + ] + result = apply_sorting(jobs, 'created_at', 'asc') + assert [j['id'] for j in result] == ['a', 'c', 'b'] + + def test_sort_by_execution_duration(self): + """Sort by execution_duration should order by duration.""" + jobs = [ + {'id': 'a', 'create_time': 100, 'execution_start_time': 100, 'execution_end_time': 5100}, # 5s + {'id': 'b', 'create_time': 300, 'execution_start_time': 300, 'execution_end_time': 1300}, # 1s + {'id': 'c', 'create_time': 200, 'execution_start_time': 200, 'execution_end_time': 3200}, # 3s + ] + result = apply_sorting(jobs, 'execution_duration', 'desc') + assert [j['id'] for j in result] == ['a', 'c', 'b'] + + def test_sort_with_none_values(self): + """Jobs with None values should sort as 0.""" + jobs = [ + {'id': 'a', 'create_time': 100, 'execution_start_time': 100, 'execution_end_time': 5100}, + {'id': 'b', 'create_time': 300, 'execution_start_time': None, 'execution_end_time': None}, + {'id': 'c', 'create_time': 200, 'execution_start_time': 200, 'execution_end_time': 3200}, + ] + result = apply_sorting(jobs, 'execution_duration', 'asc') + assert result[0]['id'] == 'b' # None treated as 0, comes first + + +class TestNormalizeQueueItem: + """Unit tests for normalize_queue_item()""" + + def test_basic_normalization(self): + """Queue item should be normalized to job dict.""" + item = ( + 10, # priority/number + 'prompt-123', # prompt_id + {'nodes': {}}, # prompt + { + 'create_time': 1234567890, + 'extra_pnginfo': {'workflow': {'id': 'workflow-abc'}} + }, # extra_data + ['node1'], # outputs_to_execute + ) + job = normalize_queue_item(item, JobStatus.PENDING) + + assert job['id'] == 'prompt-123' + assert job['status'] == 'pending' + assert job['priority'] == 10 + assert job['create_time'] == 1234567890 + assert 'execution_start_time' not in job + assert 'execution_end_time' not in job + assert 'execution_error' not in job + assert 'preview_output' not in job + assert job['outputs_count'] == 0 + assert job['workflow_id'] == 'workflow-abc' + + +class TestNormalizeHistoryItem: + """Unit tests for normalize_history_item()""" + + def test_completed_job(self): + """Completed history item should have correct status and times from messages.""" + history_item = { + 'prompt': ( + 5, # priority + 'prompt-456', + {'nodes': {}}, + { + 'create_time': 1234567890000, + 'extra_pnginfo': {'workflow': {'id': 'workflow-xyz'}} + }, + ['node1'], + ), + 'status': { + 'status_str': 'success', + 'completed': True, + 'messages': [ + ('execution_start', {'prompt_id': 'prompt-456', 'timestamp': 1234567890500}), + ('execution_success', {'prompt_id': 'prompt-456', 'timestamp': 1234567893000}), + ] + }, + 'outputs': {}, + } + job = normalize_history_item('prompt-456', history_item) + + assert job['id'] == 'prompt-456' + assert job['status'] == 'completed' + assert job['priority'] == 5 + assert job['execution_start_time'] == 1234567890500 + assert job['execution_end_time'] == 1234567893000 + assert job['workflow_id'] == 'workflow-xyz' + + def test_failed_job(self): + """Failed history item should have failed status and error from messages.""" + history_item = { + 'prompt': ( + 5, + 'prompt-789', + {'nodes': {}}, + {'create_time': 1234567890000}, + ['node1'], + ), + 'status': { + 'status_str': 'error', + 'completed': False, + 'messages': [ + ('execution_start', {'prompt_id': 'prompt-789', 'timestamp': 1234567890500}), + ('execution_error', { + 'prompt_id': 'prompt-789', + 'node_id': '5', + 'node_type': 'KSampler', + 'exception_message': 'CUDA out of memory', + 'exception_type': 'RuntimeError', + 'traceback': ['Traceback...', 'RuntimeError: CUDA out of memory'], + 'timestamp': 1234567891000, + }) + ] + }, + 'outputs': {}, + } + + job = normalize_history_item('prompt-789', history_item) + assert job['status'] == 'failed' + assert job['execution_start_time'] == 1234567890500 + assert job['execution_end_time'] == 1234567891000 + assert job['execution_error']['node_id'] == '5' + assert job['execution_error']['node_type'] == 'KSampler' + assert job['execution_error']['exception_message'] == 'CUDA out of memory' + + def test_include_outputs(self): + """When include_outputs=True, should include full output data.""" + history_item = { + 'prompt': ( + 5, + 'prompt-123', + {'nodes': {'1': {}}}, + {'create_time': 1234567890, 'client_id': 'abc'}, + ['node1'], + ), + 'status': {'status_str': 'success', 'completed': True, 'messages': []}, + 'outputs': {'node1': {'images': [{'filename': 'test.png'}]}}, + } + job = normalize_history_item('prompt-123', history_item, include_outputs=True) + + assert 'outputs' in job + assert 'workflow' in job + assert 'execution_status' in job + assert job['outputs'] == {'node1': {'images': [{'filename': 'test.png'}]}} + assert job['workflow'] == { + 'prompt': {'nodes': {'1': {}}}, + 'extra_data': {'create_time': 1234567890, 'client_id': 'abc'}, + } diff --git a/tests/execution/test_preview_method.py b/tests/execution/test_preview_method.py new file mode 100644 index 000000000..c3037553b --- /dev/null +++ b/tests/execution/test_preview_method.py @@ -0,0 +1,358 @@ +""" +E2E tests for Queue-specific Preview Method Override feature. + +Tests actual execution with different preview_method values. +Requires a running ComfyUI server with models. + +Usage: + COMFYUI_SERVER=http://localhost:8988 pytest test_preview_method_e2e.py -v -m preview_method + +Note: + These tests execute actual image generation and wait for completion. + Tests verify preview image transmission based on preview_method setting. +""" +import os +import json +import pytest +import uuid +import time +import random +import websocket +import urllib.request +from pathlib import Path + + +# Server configuration +SERVER_URL = os.environ.get("COMFYUI_SERVER", "http://localhost:8988") +SERVER_HOST = SERVER_URL.replace("http://", "").replace("https://", "") + +# Use existing inference graph fixture +GRAPH_FILE = Path(__file__).parent.parent / "inference" / "graphs" / "default_graph_sdxl1_0.json" + + +def is_server_running() -> bool: + """Check if ComfyUI server is running.""" + try: + request = urllib.request.Request(f"{SERVER_URL}/system_stats") + with urllib.request.urlopen(request, timeout=2.0): + return True + except Exception: + return False + + +def prepare_graph_for_test(graph: dict, steps: int = 5) -> dict: + """Prepare graph for testing: randomize seeds and reduce steps.""" + adapted = json.loads(json.dumps(graph)) # Deep copy + for node_id, node in adapted.items(): + inputs = node.get("inputs", {}) + # Handle both "seed" and "noise_seed" (used by KSamplerAdvanced) + if "seed" in inputs: + inputs["seed"] = random.randint(0, 2**32 - 1) + if "noise_seed" in inputs: + inputs["noise_seed"] = random.randint(0, 2**32 - 1) + # Reduce steps for faster testing (default 20 -> 5) + if "steps" in inputs: + inputs["steps"] = steps + return adapted + + +# Alias for backward compatibility +randomize_seed = prepare_graph_for_test + + +class PreviewMethodClient: + """Client for testing preview_method with WebSocket execution tracking.""" + + def __init__(self, server_address: str): + self.server_address = server_address + self.client_id = str(uuid.uuid4()) + self.ws = None + + def connect(self): + """Connect to WebSocket.""" + self.ws = websocket.WebSocket() + self.ws.settimeout(120) # 2 minute timeout for sampling + self.ws.connect(f"ws://{self.server_address}/ws?clientId={self.client_id}") + + def close(self): + """Close WebSocket connection.""" + if self.ws: + self.ws.close() + + def queue_prompt(self, prompt: dict, extra_data: dict = None) -> dict: + """Queue a prompt and return response with prompt_id.""" + data = { + "prompt": prompt, + "client_id": self.client_id, + "extra_data": extra_data or {} + } + req = urllib.request.Request( + f"http://{self.server_address}/prompt", + data=json.dumps(data).encode("utf-8"), + headers={"Content-Type": "application/json"} + ) + return json.loads(urllib.request.urlopen(req).read()) + + def wait_for_execution(self, prompt_id: str, timeout: float = 120.0) -> dict: + """ + Wait for execution to complete via WebSocket. + + Returns: + dict with keys: completed, error, preview_count, execution_time + """ + result = { + "completed": False, + "error": None, + "preview_count": 0, + "execution_time": 0.0 + } + + start_time = time.time() + self.ws.settimeout(timeout) + + try: + while True: + out = self.ws.recv() + elapsed = time.time() - start_time + + if isinstance(out, str): + message = json.loads(out) + msg_type = message.get("type") + data = message.get("data", {}) + + if data.get("prompt_id") != prompt_id: + continue + + if msg_type == "executing": + if data.get("node") is None: + # Execution complete + result["completed"] = True + result["execution_time"] = elapsed + break + + elif msg_type == "execution_error": + result["error"] = data + result["execution_time"] = elapsed + break + + elif msg_type == "progress": + # Progress update during sampling + pass + + elif isinstance(out, bytes): + # Binary data = preview image + result["preview_count"] += 1 + + except websocket.WebSocketTimeoutException: + result["error"] = "Timeout waiting for execution" + result["execution_time"] = time.time() - start_time + + return result + + +def load_graph() -> dict: + """Load the SDXL graph fixture with randomized seed.""" + with open(GRAPH_FILE) as f: + graph = json.load(f) + return randomize_seed(graph) # Avoid caching + + +# Skip all tests if server is not running +pytestmark = [ + pytest.mark.skipif( + not is_server_running(), + reason=f"ComfyUI server not running at {SERVER_URL}" + ), + pytest.mark.preview_method, + pytest.mark.execution, +] + + +@pytest.fixture +def client(): + """Create and connect a test client.""" + c = PreviewMethodClient(SERVER_HOST) + c.connect() + yield c + c.close() + + +@pytest.fixture +def graph(): + """Load the test graph.""" + return load_graph() + + +class TestPreviewMethodExecution: + """Test actual execution with different preview methods.""" + + def test_execution_with_latent2rgb(self, client, graph): + """ + Execute with preview_method=latent2rgb. + Should complete and potentially receive preview images. + """ + extra_data = {"preview_method": "latent2rgb"} + + response = client.queue_prompt(graph, extra_data) + assert "prompt_id" in response + + result = client.wait_for_execution(response["prompt_id"]) + + # Should complete (may error if model missing, but that's separate) + assert result["completed"] or result["error"] is not None + # Execution should take some time (sampling) + if result["completed"]: + assert result["execution_time"] > 0.5, "Execution too fast - likely didn't run" + # latent2rgb should produce previews + print(f"latent2rgb: {result['preview_count']} previews in {result['execution_time']:.2f}s") # noqa: T201 + + def test_execution_with_taesd(self, client, graph): + """ + Execute with preview_method=taesd. + TAESD provides higher quality previews. + """ + extra_data = {"preview_method": "taesd"} + + response = client.queue_prompt(graph, extra_data) + assert "prompt_id" in response + + result = client.wait_for_execution(response["prompt_id"]) + + assert result["completed"] or result["error"] is not None + if result["completed"]: + assert result["execution_time"] > 0.5 + # taesd should also produce previews + print(f"taesd: {result['preview_count']} previews in {result['execution_time']:.2f}s") # noqa: T201 + + def test_execution_with_none_preview(self, client, graph): + """ + Execute with preview_method=none. + No preview images should be generated. + """ + extra_data = {"preview_method": "none"} + + response = client.queue_prompt(graph, extra_data) + assert "prompt_id" in response + + result = client.wait_for_execution(response["prompt_id"]) + + assert result["completed"] or result["error"] is not None + if result["completed"]: + # With "none", should receive no preview images + assert result["preview_count"] == 0, \ + f"Expected no previews with 'none', got {result['preview_count']}" + print(f"none: {result['preview_count']} previews in {result['execution_time']:.2f}s") # noqa: T201 + + def test_execution_with_default(self, client, graph): + """ + Execute with preview_method=default. + Should use server's CLI default setting. + """ + extra_data = {"preview_method": "default"} + + response = client.queue_prompt(graph, extra_data) + assert "prompt_id" in response + + result = client.wait_for_execution(response["prompt_id"]) + + assert result["completed"] or result["error"] is not None + if result["completed"]: + print(f"default: {result['preview_count']} previews in {result['execution_time']:.2f}s") # noqa: T201 + + def test_execution_without_preview_method(self, client, graph): + """ + Execute without preview_method in extra_data. + Should use server's default preview method. + """ + extra_data = {} # No preview_method + + response = client.queue_prompt(graph, extra_data) + assert "prompt_id" in response + + result = client.wait_for_execution(response["prompt_id"]) + + assert result["completed"] or result["error"] is not None + if result["completed"]: + print(f"(no override): {result['preview_count']} previews in {result['execution_time']:.2f}s") # noqa: T201 + + +class TestPreviewMethodComparison: + """Compare preview behavior between different methods.""" + + def test_none_vs_latent2rgb_preview_count(self, client, graph): + """ + Compare preview counts: 'none' should have 0, others should have >0. + This is the key verification that preview_method actually works. + """ + results = {} + + # Run with none (randomize seed to avoid caching) + graph_none = randomize_seed(graph) + extra_data_none = {"preview_method": "none"} + response = client.queue_prompt(graph_none, extra_data_none) + results["none"] = client.wait_for_execution(response["prompt_id"]) + + # Run with latent2rgb (randomize seed again) + graph_rgb = randomize_seed(graph) + extra_data_rgb = {"preview_method": "latent2rgb"} + response = client.queue_prompt(graph_rgb, extra_data_rgb) + results["latent2rgb"] = client.wait_for_execution(response["prompt_id"]) + + # Verify both completed + assert results["none"]["completed"], f"'none' execution failed: {results['none']['error']}" + assert results["latent2rgb"]["completed"], f"'latent2rgb' execution failed: {results['latent2rgb']['error']}" + + # Key assertion: 'none' should have 0 previews + assert results["none"]["preview_count"] == 0, \ + f"'none' should have 0 previews, got {results['none']['preview_count']}" + + # 'latent2rgb' should have at least 1 preview (depends on steps) + assert results["latent2rgb"]["preview_count"] > 0, \ + f"'latent2rgb' should have >0 previews, got {results['latent2rgb']['preview_count']}" + + print("\nPreview count comparison:") # noqa: T201 + print(f" none: {results['none']['preview_count']} previews") # noqa: T201 + print(f" latent2rgb: {results['latent2rgb']['preview_count']} previews") # noqa: T201 + + +class TestPreviewMethodSequential: + """Test sequential execution with different preview methods.""" + + def test_sequential_different_methods(self, client, graph): + """ + Execute multiple prompts sequentially with different preview methods. + Each should complete independently with correct preview behavior. + """ + methods = ["latent2rgb", "none", "default"] + results = [] + + for method in methods: + # Randomize seed for each execution to avoid caching + graph_run = randomize_seed(graph) + extra_data = {"preview_method": method} + response = client.queue_prompt(graph_run, extra_data) + + result = client.wait_for_execution(response["prompt_id"]) + results.append({ + "method": method, + "completed": result["completed"], + "preview_count": result["preview_count"], + "execution_time": result["execution_time"], + "error": result["error"] + }) + + # All should complete or have clear errors + for r in results: + assert r["completed"] or r["error"] is not None, \ + f"Method {r['method']} neither completed nor errored" + + # "none" should have zero previews if completed + none_result = next(r for r in results if r["method"] == "none") + if none_result["completed"]: + assert none_result["preview_count"] == 0, \ + f"'none' should have 0 previews, got {none_result['preview_count']}" + + print("\nSequential execution results:") # noqa: T201 + for r in results: + status = "✓" if r["completed"] else f"✗ ({r['error']})" + print(f" {r['method']}: {status}, {r['preview_count']} previews, {r['execution_time']:.2f}s") # noqa: T201 diff --git a/tests/execution/test_public_api.py b/tests/execution/test_public_api.py new file mode 100644 index 000000000..52bc2fcd8 --- /dev/null +++ b/tests/execution/test_public_api.py @@ -0,0 +1,153 @@ +""" +Tests for public ComfyAPI and ComfyAPISync functions. + +These tests verify that the public API methods work correctly in both sync and async contexts, +ensuring that the sync wrapper generation (via get_type_hints() in async_to_sync.py) correctly +handles string annotations from 'from __future__ import annotations'. +""" + +import pytest +import time +import subprocess +import torch +from pytest import fixture +from comfy_execution.graph_utils import GraphBuilder +from tests.execution.test_execution import ComfyClient + + +@pytest.mark.execution +class TestPublicAPI: + """Test suite for public ComfyAPI and ComfyAPISync methods.""" + + @fixture(scope="class", autouse=True) + def _server(self, args_pytest): + """Start ComfyUI server for testing.""" + pargs = [ + 'python', 'main.py', + '--output-directory', args_pytest["output_dir"], + '--listen', args_pytest["listen"], + '--port', str(args_pytest["port"]), + '--extra-model-paths-config', 'tests/execution/extra_model_paths.yaml', + '--cpu', + ] + p = subprocess.Popen(pargs) + yield + p.kill() + torch.cuda.empty_cache() + + @fixture(scope="class", autouse=True) + def shared_client(self, args_pytest, _server): + """Create shared client with connection retry.""" + client = ComfyClient() + n_tries = 5 + for i in range(n_tries): + time.sleep(4) + try: + client.connect(listen=args_pytest["listen"], port=args_pytest["port"]) + break + except ConnectionRefusedError: + if i == n_tries - 1: + raise + yield client + del client + torch.cuda.empty_cache() + + @fixture + def client(self, shared_client, request): + """Set test name for each test.""" + shared_client.set_test_name(f"public_api[{request.node.name}]") + yield shared_client + + @fixture + def builder(self, request): + """Create GraphBuilder for each test.""" + yield GraphBuilder(prefix=request.node.name) + + def test_sync_progress_update_executes(self, client: ComfyClient, builder: GraphBuilder): + """Test that TestSyncProgressUpdate executes without errors. + + This test validates that api_sync.execution.set_progress() works correctly, + which is the primary code path fixed by adding get_type_hints() to async_to_sync.py. + """ + g = builder + image = g.node("StubImage", content="BLACK", height=256, width=256, batch_size=1) + + # Use TestSyncProgressUpdate with short sleep + progress_node = g.node("TestSyncProgressUpdate", + value=image.out(0), + sleep_seconds=0.5) + output = g.node("SaveImage", images=progress_node.out(0)) + + # Execute workflow + result = client.run(g) + + # Verify execution + assert result.did_run(progress_node), "Progress node should have executed" + assert result.did_run(output), "Output node should have executed" + + # Verify output + images = result.get_images(output) + assert len(images) == 1, "Should have produced 1 image" + + def test_async_progress_update_executes(self, client: ComfyClient, builder: GraphBuilder): + """Test that TestAsyncProgressUpdate executes without errors. + + This test validates that await api.execution.set_progress() works correctly + in async contexts. + """ + g = builder + image = g.node("StubImage", content="WHITE", height=256, width=256, batch_size=1) + + # Use TestAsyncProgressUpdate with short sleep + progress_node = g.node("TestAsyncProgressUpdate", + value=image.out(0), + sleep_seconds=0.5) + output = g.node("SaveImage", images=progress_node.out(0)) + + # Execute workflow + result = client.run(g) + + # Verify execution + assert result.did_run(progress_node), "Async progress node should have executed" + assert result.did_run(output), "Output node should have executed" + + # Verify output + images = result.get_images(output) + assert len(images) == 1, "Should have produced 1 image" + + def test_sync_and_async_progress_together(self, client: ComfyClient, builder: GraphBuilder): + """Test both sync and async progress updates in same workflow. + + This test ensures that both ComfyAPISync and ComfyAPI can coexist and work + correctly in the same workflow execution. + """ + g = builder + image1 = g.node("StubImage", content="BLACK", height=256, width=256, batch_size=1) + image2 = g.node("StubImage", content="WHITE", height=256, width=256, batch_size=1) + + # Use both types of progress nodes + sync_progress = g.node("TestSyncProgressUpdate", + value=image1.out(0), + sleep_seconds=0.3) + async_progress = g.node("TestAsyncProgressUpdate", + value=image2.out(0), + sleep_seconds=0.3) + + # Create outputs + output1 = g.node("SaveImage", images=sync_progress.out(0)) + output2 = g.node("SaveImage", images=async_progress.out(0)) + + # Execute workflow + result = client.run(g) + + # Both should execute successfully + assert result.did_run(sync_progress), "Sync progress node should have executed" + assert result.did_run(async_progress), "Async progress node should have executed" + assert result.did_run(output1), "First output node should have executed" + assert result.did_run(output2), "Second output node should have executed" + + # Verify outputs + images1 = result.get_images(output1) + images2 = result.get_images(output2) + assert len(images1) == 1, "Should have produced 1 image from sync node" + assert len(images2) == 1, "Should have produced 1 image from async node"