diff --git a/.ci/windows_nightly_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat b/.ci/windows_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat similarity index 100% rename from .ci/windows_nightly_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat rename to .ci/windows_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat diff --git a/.github/ISSUE_TEMPLATE/bug-report.yml b/.github/ISSUE_TEMPLATE/bug-report.yml index 39d1992d7..69ce998eb 100644 --- a/.github/ISSUE_TEMPLATE/bug-report.yml +++ b/.github/ISSUE_TEMPLATE/bug-report.yml @@ -15,6 +15,14 @@ body: steps to replicate what went wrong and others will be able to repeat your steps and see the same issue happen. If unsure, ask on the [ComfyUI Matrix Space](https://app.element.io/#/room/%23comfyui_space%3Amatrix.org) or the [Comfy Org Discord](https://discord.gg/comfyorg) first. + - type: checkboxes + id: custom-nodes-test + attributes: + label: Custom Node Testing + description: Please confirm you have tried to reproduce the issue with all custom nodes disabled. + options: + - label: I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help) + required: true - type: textarea attributes: label: Expected Behavior diff --git a/.github/ISSUE_TEMPLATE/user-support.yml b/.github/ISSUE_TEMPLATE/user-support.yml index df28804c6..50657d493 100644 --- a/.github/ISSUE_TEMPLATE/user-support.yml +++ b/.github/ISSUE_TEMPLATE/user-support.yml @@ -11,6 +11,14 @@ body: **2:** You have made an effort to find public answers to your question before asking here. In other words, you googled it first, and scrolled through recent help topics. If unsure, ask on the [ComfyUI Matrix Space](https://app.element.io/#/room/%23comfyui_space%3Amatrix.org) or the [Comfy Org Discord](https://discord.gg/comfyorg) first. + - type: checkboxes + id: custom-nodes-test + attributes: + label: Custom Node Testing + description: Please confirm you have tried to reproduce the issue with all custom nodes disabled. + options: + - label: I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help) + required: true - type: textarea attributes: label: Your question diff --git a/.github/workflows/update-api-stubs.yml b/.github/workflows/update-api-stubs.yml new file mode 100644 index 000000000..c99ec9fc1 --- /dev/null +++ b/.github/workflows/update-api-stubs.yml @@ -0,0 +1,56 @@ +name: Generate Pydantic Stubs from api.comfy.org + +on: + schedule: + - cron: '0 0 * * 1' + workflow_dispatch: + +jobs: + generate-models: + runs-on: ubuntu-latest + + steps: + - name: Checkout repository + uses: actions/checkout@v4 + + - name: Set up Python + uses: actions/setup-python@v4 + with: + python-version: '3.10' + + - name: Install dependencies + run: | + python -m pip install --upgrade pip + pip install 'datamodel-code-generator[http]' + npm install @redocly/cli + + - name: Download OpenAPI spec + run: | + curl -o openapi.yaml https://api.comfy.org/openapi + + - name: Filter OpenAPI spec with Redocly + run: | + npx @redocly/cli bundle openapi.yaml --output filtered-openapi.yaml --config comfy_api_nodes/redocly.yaml --remove-unused-components + + - name: Generate API models + run: | + datamodel-codegen --use-subclass-enum --input filtered-openapi.yaml --output comfy_api_nodes/apis --output-model-type pydantic_v2.BaseModel + + - name: Check for changes + id: git-check + run: | + git diff --exit-code comfy_api_nodes/apis || echo "changes=true" >> $GITHUB_OUTPUT + + - name: Create Pull Request + if: steps.git-check.outputs.changes == 'true' + uses: peter-evans/create-pull-request@v5 + with: + commit-message: 'chore: update API models from OpenAPI spec' + title: 'Update API models from api.comfy.org' + body: | + This PR updates the API models based on the latest api.comfy.org OpenAPI specification. + + Generated automatically by the a Github workflow. + branch: update-api-stubs + delete-branch: true + base: master diff --git a/.gitignore b/.gitignore index ddce17bd1..362aaab9a 100644 --- a/.gitignore +++ b/.gitignore @@ -183,3 +183,6 @@ cython_debug/ *.log web_custom_versions/ .DS_Store +openapi.yaml +filtered-openapi.yaml +uv.lock diff --git a/CODEOWNERS b/CODEOWNERS index e25edc385..a7ccbcb29 100644 --- a/CODEOWNERS +++ b/CODEOWNERS @@ -5,20 +5,20 @@ # Inlined the team members for now. # Maintainers -*.md @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne -/tests/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne -/tests-unit/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne -/notebooks/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne -/script_examples/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne -/.github/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne -/requirements.txt @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne -/pyproject.toml @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne +*.md @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne +/tests/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne +/tests-unit/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne +/notebooks/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne +/script_examples/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne +/.github/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne +/requirements.txt @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne +/pyproject.toml @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne # Python web server -/api_server/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @christian-byrne -/app/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @christian-byrne -/utils/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @christian-byrne +/api_server/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne +/app/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne +/utils/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne # Node developers -/comfy_extras/ @yoland68 @robinjhuang @huchenlei @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne -/comfy/comfy_types/ @yoland68 @robinjhuang @huchenlei @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne +/comfy_extras/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne +/comfy/comfy_types/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne diff --git a/README.md b/README.md index 076432443..2cc678bad 100644 --- a/README.md +++ b/README.md @@ -42,7 +42,6 @@ Supports all operating systems and GPU types (NVIDIA, AMD, Intel, Apple Silicon, See what ComfyUI can do with the [example workflows](https://comfyanonymous.github.io/ComfyUI_examples/). ## Upstream Features - - Nodes/graph/flowchart interface to experiment and create complex Stable Diffusion workflows without needing to code anything. - Image Models - SD1.x, SD2.x, @@ -62,9 +61,11 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith - [Hunyuan Video](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_video/) - [Nvidia Cosmos](https://comfyanonymous.github.io/ComfyUI_examples/cosmos/) - [Wan 2.1](https://comfyanonymous.github.io/ComfyUI_examples/wan/) +- Audio Models + - [Stable Audio](https://comfyanonymous.github.io/ComfyUI_examples/audio/) + - [ACE Step](https://comfyanonymous.github.io/ComfyUI_examples/audio/) - 3D Models - [Hunyuan3D 2.0](https://docs.comfy.org/tutorials/3d/hunyuan3D-2) -- [Stable Audio](https://comfyanonymous.github.io/ComfyUI_examples/audio/) - Asynchronous Queue system - Many optimizations: Only re-executes the parts of the workflow that changes between executions. - Smart memory management: can automatically run models on GPUs with as low as 1GB vram. @@ -86,7 +87,8 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith - [LCM models and Loras](https://comfyanonymous.github.io/ComfyUI_examples/lcm/) - Latent previews with [TAESD](#how-to-show-high-quality-previews) - Starts up very fast. -- Works fully offline: will never download anything. +- Works fully offline: core will never download anything unless you want to. +- Optional API nodes to use paid models from external providers through the online [Comfy API](https://docs.comfy.org/tutorials/api-nodes/overview). - [Config file](extra_model_paths.yaml.example) to set the search paths for models. ### Table of Contents diff --git a/comfy/__init__.py b/comfy/__init__.py index d91d0940c..4aa0289ec 100644 --- a/comfy/__init__.py +++ b/comfy/__init__.py @@ -1 +1 @@ -__version__ = "0.3.29" +__version__ = "0.3.40" \ No newline at end of file diff --git a/comfy/app/custom_node_manager.py b/comfy/app/custom_node_manager.py index ba488b2a3..8d43b7a58 100644 --- a/comfy/app/custom_node_manager.py +++ b/comfy/app/custom_node_manager.py @@ -103,6 +103,8 @@ class CustomNodeManager: def add_routes(self, routes, webapp, loadedModules): + example_workflow_folder_names = ["example_workflows", "example", "examples", "workflow", "workflows"] + @routes.get("/workflow_templates") async def get_workflow_templates(request): """Returns a web response that contains the map of custom_nodes names and their associated workflow templates. The ones without templates are omitted.""" @@ -110,7 +112,8 @@ class CustomNodeManager: files = [ file for folder in folder_paths.get_folder_paths("custom_nodes") - for file in glob.glob(os.path.join(folder, '*/example_workflows/*.json')) + for folder_name in example_workflow_folder_names + for file in glob.glob(os.path.join(folder, f'*/{folder_name}/*.json')) ] workflow_templates_dict = ( {} @@ -127,15 +130,22 @@ class CustomNodeManager: # Serve workflow templates from custom nodes. for module_name, module_dir in loadedModules: - workflows_dir = os.path.join(module_dir, "example_workflows") - if os.path.exists(workflows_dir): - webapp.add_routes( - [ - web.static( - "/api/workflow_templates/" + module_name, workflows_dir - ) - ] - ) + for folder_name in example_workflow_folder_names: + workflows_dir = os.path.join(module_dir, folder_name) + + if os.path.exists(workflows_dir): + if folder_name != "example_workflows": + logging.debug( + "Found example workflow folder '%s' for custom node '%s', consider renaming it to 'example_workflows'", + folder_name, module_name) + + webapp.add_routes( + [ + web.static( + "/api/workflow_templates/" + module_name, workflows_dir + ) + ] + ) @routes.get("/i18n") async def get_i18n(request): diff --git a/comfy/app/frontend_management.py b/comfy/app/frontend_management.py index 717b92dff..17e3c7ed0 100644 --- a/comfy/app/frontend_management.py +++ b/comfy/app/frontend_management.py @@ -151,6 +151,19 @@ comfyui-workflow-templates is not installed. """.strip() ) + @classmethod + def embedded_docs_path(cls) -> str: + """Get the path to embedded documentation""" + try: + import comfyui_embedded_docs + + return str( + importlib.resources.files(comfyui_embedded_docs) / "docs" + ) + except ImportError: + logging.info("comfyui-embedded-docs package not found") + return None + @classmethod def parse_version_string(cls, value: str) -> tuple[str, str, str]: """ diff --git a/comfy/app/user_manager.py b/comfy/app/user_manager.py index cf3880b77..8fdd388b6 100644 --- a/comfy/app/user_manager.py +++ b/comfy/app/user_manager.py @@ -200,6 +200,112 @@ class UserManager(): return web.json_response(results) + @routes.get("/v2/userdata") + async def list_userdata_v2(request): + """ + List files and directories in a user's data directory. + + This endpoint provides a structured listing of contents within a specified + subdirectory of the user's data storage. + + Query Parameters: + - path (optional): The relative path within the user's data directory + to list. Defaults to the root (''). + + Returns: + - 400: If the requested path is invalid, outside the user's data directory, or is not a directory. + - 404: If the requested path does not exist. + - 403: If the user is invalid. + - 500: If there is an error reading the directory contents. + - 200: JSON response containing a list of file and directory objects. + Each object includes: + - name: The name of the file or directory. + - type: 'file' or 'directory'. + - path: The relative path from the user's data root. + - size (for files): The size in bytes. + - modified (for files): The last modified timestamp (Unix epoch). + """ + requested_rel_path = request.rel_url.query.get('path', '') + + # URL-decode the path parameter + try: + requested_rel_path = parse.unquote(requested_rel_path) + except Exception as e: + logging.warning(f"Failed to decode path parameter: {requested_rel_path}, Error: {e}") + return web.Response(status=400, text="Invalid characters in path parameter") + + + # Check user validity and get the absolute path for the requested directory + try: + base_user_path = self.get_request_user_filepath(request, None, create_dir=False) + + if requested_rel_path: + target_abs_path = self.get_request_user_filepath(request, requested_rel_path, create_dir=False) + else: + target_abs_path = base_user_path + + except KeyError as e: + # Invalid user detected by get_request_user_id inside get_request_user_filepath + logging.warning(f"Access denied for user: {e}") + return web.Response(status=403, text="Invalid user specified in request") + + + if not target_abs_path: + # Path traversal or other issue detected by get_request_user_filepath + return web.Response(status=400, text="Invalid path requested") + + # Handle cases where the user directory or target path doesn't exist + if not os.path.exists(target_abs_path): + # Check if it's the base user directory that's missing (new user case) + if target_abs_path == base_user_path: + # It's okay if the base user directory doesn't exist yet, return empty list + return web.json_response([]) + else: + # A specific subdirectory was requested but doesn't exist + return web.Response(status=404, text="Requested path not found") + + if not os.path.isdir(target_abs_path): + return web.Response(status=400, text="Requested path is not a directory") + + results = [] + try: + for root, dirs, files in os.walk(target_abs_path, topdown=True): + # Process directories + for dir_name in dirs: + dir_path = os.path.join(root, dir_name) + rel_path = os.path.relpath(dir_path, base_user_path).replace(os.sep, '/') + results.append({ + "name": dir_name, + "path": rel_path, + "type": "directory" + }) + + # Process files + for file_name in files: + file_path = os.path.join(root, file_name) + rel_path = os.path.relpath(file_path, base_user_path).replace(os.sep, '/') + entry_info = { + "name": file_name, + "path": rel_path, + "type": "file" + } + try: + stats = os.stat(file_path) # Use os.stat for potentially better performance with os.walk + entry_info["size"] = stats.st_size + entry_info["modified"] = stats.st_mtime + except OSError as stat_error: + logging.warning(f"Could not stat file {file_path}: {stat_error}") + pass # Include file with available info + results.append(entry_info) + except OSError as e: + logging.error(f"Error listing directory {target_abs_path}: {e}") + return web.Response(status=500, text="Error reading directory contents") + + # Sort results alphabetically, directories first then files + results.sort(key=lambda x: (x['type'] != 'directory', x['name'].lower())) + + return web.json_response(results) + def get_user_data_path(request, check_exists=False, param="file"): file = request.match_info.get(param, None) if not file: diff --git a/comfy/cli_args.py b/comfy/cli_args.py index ae37538cc..39c1db6c9 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -64,6 +64,7 @@ def _create_parser() -> EnhancedConfigArgParser: fpunet_group.add_argument("--fp16-unet", action="store_true", help="Run the diffusion model in fp16") fpunet_group.add_argument("--fp8_e4m3fn-unet", action="store_true", help="Store unet weights in fp8_e4m3fn.") fpunet_group.add_argument("--fp8_e5m2-unet", action="store_true", help="Store unet weights in fp8_e5m2.") + fpunet_group.add_argument("--fp8_e8m0fnu-unet", action="store_true", help="Store unet weights in fp8_e8m0fnu.") fpvae_group = parser.add_mutually_exclusive_group() fpvae_group.add_argument("--fp16-vae", action="store_true", help="Run the VAE in fp16, might cause black images.") @@ -87,6 +88,7 @@ def _create_parser() -> EnhancedConfigArgParser: parser.add_argument("--oneapi-device-selector", type=str, default=None, metavar="SELECTOR_STRING", help="Sets the oneAPI device(s) this instance will use.") parser.add_argument("--disable-ipex-optimize", action="store_true", help="Disables ipex.optimize default when loading models with Intel's Extension for Pytorch.") + parser.add_argument("--supports-fp8-compute", action="store_true", help="ComfyUI will act like if the device supports fp8 compute.") parser.add_argument("--preview-method", type=LatentPreviewMethod, default=LatentPreviewMethod.Auto, help="Default preview method for sampler nodes.", action=EnumAction) @@ -123,6 +125,7 @@ def _create_parser() -> EnhancedConfigArgParser: vram_group.add_argument("--cpu", action="store_true", help="To use the CPU for everything (slow).") 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("--default-hashing-function", type=str, choices=['md5', 'sha1', 'sha256', 'sha512'], default='sha256', help="Allows you to choose the hash function to use for duplicate filename / contents comparison. Default is sha256.") parser.add_argument("--disable-smart-memory", action="store_true", help="Force ComfyUI to agressively offload to regular ram instead of keeping models in vram when it can.") @@ -131,6 +134,7 @@ def _create_parser() -> EnhancedConfigArgParser: parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. Pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: fp16_accumulation fp8_matrix_mult cublas_ops", default=set()) + parser.add_argument("--mmap-torch-files", action="store_true", help="Use mmap when loading ckpt/pt files.") parser.add_argument("--dont-print-server", action="store_true", help="Don't print server output.") parser.add_argument("--quick-test-for-ci", action="store_true", help="Quick test for CI. Raises an error if nodes cannot be imported,") parser.add_argument("--windows-standalone-build", default=hasattr(sys, 'frozen') and getattr(sys, 'frozen'), @@ -139,6 +143,7 @@ def _create_parser() -> EnhancedConfigArgParser: 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("--disable-api-nodes", action="store_true", help="Disable loading all api nodes.") parser.add_argument("--multi-user", action="store_true", help="Enables per-user storage.") parser.add_argument("--create-directories", action="store_true", @@ -248,6 +253,13 @@ def _create_parser() -> EnhancedConfigArgParser: parser.add_argument("--enable-compress-response-body", action="store_true", help="Enable compressing response body.") + parser.add_argument( + "--comfy-api-base", + type=str, + default="https://api.comfy.org", + help="Set the base URL for the ComfyUI API. (default: https://api.comfy.org)", + ) + parser.add_argument("--workflows", type=str, nargs='+', default=[], help="Execute the API workflow(s) specified in the provided files. For each workflow, its outputs will be printed to a line to standard out. Application logging will be redirected to standard error. Use `-` to signify standard in.") # now give plugins a chance to add configuration diff --git a/comfy/cli_args_types.py b/comfy/cli_args_types.py index c68827254..8bbb91b64 100644 --- a/comfy/cli_args_types.py +++ b/comfy/cli_args_types.py @@ -78,7 +78,7 @@ class Configuration(dict): fp8_e5m2_text_enc (bool): Use FP8 precision for the text encoder (e5m2 variant). fp16_text_enc (bool): Use FP16 precision for the text encoder. fp32_text_enc (bool): Use FP32 precision for the text encoder. - openapi_device_selector (Optional[str]): Sets the oneAPI device(s) this instance will use. + oneapi_device_selector (Optional[str]): Sets the oneAPI device(s) this instance will use. directml (Optional[int]): Use DirectML. -1 for auto-selection. disable_ipex_optimize (bool): Disable IPEX optimization for Intel GPUs. preview_method (LatentPreviewMethod): Method for generating previews. Defaults to "auto". @@ -87,7 +87,7 @@ class Configuration(dict): use_quad_cross_attention (bool): Use sub-quadratic cross-attention optimization. use_pytorch_cross_attention (bool): Use PyTorch's cross-attention function. use_sage_attention (bool): Use Sage Attention - use_flas_attention (bool): Use FlashAttention + use_flash_attention (bool): Use FlashAttention disable_xformers (bool): Disable xformers. gpu_only (bool): Run everything on the GPU. highvram (bool): Keep models in GPU memory. @@ -132,6 +132,19 @@ class Configuration(dict): panic_when (list[str]): List of fully qualified exception class names to panic (sys.exit(1)) when a workflow raises it. enable_compress_response_body (bool): Enable compressing response body. workflows (list[str]): Execute the API workflow(s) specified in the provided files. For each workflow, its outputs will be printed to a line to standard out. Application logging will be redirected to standard error. Use `-` to signify standard in. + fp8_e8m0fnu_unet (bool): Store unet weights in fp8_e8m0fnu. + bf16_text_enc (bool): Store text encoder weights in bf16. + supports_fp8_compute (bool): ComfyUI will act like if the device supports fp8 compute. + cache_classic (bool): WARNING: Unused. Use the old style (aggressive) caching. + cache_none (bool): Reduced RAM/VRAM usage at the expense of executing every node for each run. + async_offload (bool): Use async weight offloading. + default_hashing_function (str): Allows you to choose the hash function to use for duplicate filename / contents comparison. Default is sha256. + mmap_torch_files (bool): Use mmap when loading ckpt/pt files. + dont_print_server (bool): Don't print server output. + disable_api_nodes (bool): Disable loading all api nodes. + front_end_version (str): Specifies the version of the frontend to be used. + front_end_root (Optional[str]): The local filesystem path to the directory where the frontend is located. Overrides --front-end-version. + comfy_api_base (str): Set the base URL for the ComfyUI API. (default: https://api.comfy.org) """ def __init__(self, **kwargs): @@ -215,7 +228,7 @@ class Configuration(dict): self.force_hf_local_dir_mode = False self.preview_size: int = 512 self.logging_level: str = "INFO" - self.openapi_device_selector: Optional[str] = None + self.oneapi_device_selector: Optional[str] = None self.log_stdout: bool = False # from guill @@ -232,6 +245,19 @@ class Configuration(dict): self.user_directory: Optional[str] = None self.panic_when: list[str] = [] self.workflows: list[str] = [] + + self.fp8_e8m0fnu_unet: bool = False + self.bf16_text_enc: bool = False + self.supports_fp8_compute: bool = False + self.cache_classic: bool = False + self.cache_none: bool = False + self.async_offload: bool = False + self.default_hashing_function: str = 'sha256' + self.mmap_torch_files: bool = False + self.disable_api_nodes: bool = False + self.front_end_version: str = "comfyanonymous/ComfyUI@latest" + self.front_end_root: Optional[str] = None + self.comfy_api_base: str = "https://api.comfy.org" for key, value in kwargs.items(): self[key] = value # this must always be last diff --git a/comfy/clip_vision.py b/comfy/clip_vision.py index bb79485e2..d188306a5 100644 --- a/comfy/clip_vision.py +++ b/comfy/clip_vision.py @@ -23,6 +23,7 @@ class Output: 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) diff --git a/comfy/cmd/execution.py b/comfy/cmd/execution.py index 61e7ffc7e..7f3867dee 100644 --- a/comfy/cmd/execution.py +++ b/comfy/cmd/execution.py @@ -175,6 +175,8 @@ def get_input_data(inputs, class_def, unique_id, outputs=None, dynprompt=None, e input_data_all[x] = [unique_id] if h[x] == "AUTH_TOKEN_COMFY_ORG": 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 @@ -1122,6 +1124,7 @@ class PromptQueue(AbstractPromptQueue): if queue_item.completed: queue_item.completed.set_result(outputs_) + # Note: slow def get_current_queue(self) -> Tuple[typing.List[QueueTuple], typing.List[QueueTuple]]: with self.mutex: out: typing.List[QueueTuple] = [] @@ -1129,6 +1132,13 @@ class PromptQueue(AbstractPromptQueue): out += [x.queue_tuple] return out, copy.deepcopy([item.queue_tuple for item in self.queue]) + # read-safe as long as queue items are immutable + def get_current_queue_volatile(self): + with self.mutex: + running = [x for x in self.currently_running.values()] + queued = copy.copy(self.queue) + return (running, queued) + def get_tasks_remaining(self): with self.mutex: return len(self.queue) + len(self.currently_running) diff --git a/comfy/cmd/main_pre.py b/comfy/cmd/main_pre.py index 0d8507983..7e6656600 100644 --- a/comfy/cmd/main_pre.py +++ b/comfy/cmd/main_pre.py @@ -15,6 +15,8 @@ os.environ["TORCHINDUCTOR_FX_GRAPH_CACHE"] = "1" os.environ["TORCHINDUCTOR_AUTOGRAD_CACHE"] = "1" os.environ["BITSANDBYTES_NOWELCOME"] = "1" os.environ["NO_ALBUMENTATIONS_UPDATE"] = "1" +os.environ['HF_HUB_DISABLE_TELEMETRY'] = '1' +os.environ['DO_NOT_TRACK'] = '1' import ctypes import importlib.util diff --git a/comfy/cmd/server.py b/comfy/cmd/server.py index 9fcb90cf0..a62dbe5be 100644 --- a/comfy/cmd/server.py +++ b/comfy/cmd/server.py @@ -18,7 +18,7 @@ from asyncio import Future, AbstractEventLoop, Task from enum import Enum from io import BytesIO from posixpath import join as urljoin -from typing import List, Optional +from typing import List, Optional, Union from urllib.parse import quote, urlencode import aiofiles @@ -192,7 +192,7 @@ class PromptServer(ExecutorToClientProgress): self.internal_routes = InternalRoutes(self) # todo: this is probably read by custom nodes elsewhere self.supports: List[str] = ["custom_nodes_from_web"] - self.prompt_queue: AbstractPromptQueue | AsyncAbstractPromptQueue | None = None + self.prompt_queue: AbstractPromptQueue | AsyncAbstractPromptQueue | None = execution.PromptQueue(self) self.loop: AbstractEventLoop = loop self.messages: asyncio.Queue = asyncio.Queue() self.client_session: Optional[aiohttp.ClientSession] = None @@ -318,7 +318,6 @@ class PromptServer(ExecutorToClientProgress): a.update(f.read()) b.update(image.file.read()) image.file.seek(0) - f.close() return a.hexdigest() == b.hexdigest() return False @@ -430,6 +429,7 @@ class PromptServer(ExecutorToClientProgress): async def view_image(request): if "filename" in request.rel_url.query: filename = request.rel_url.query["filename"] + # todo: do we ever need annotated filenames support on this? if not filename: return web.Response(status=400) @@ -510,9 +510,8 @@ class PromptServer(ExecutorToClientProgress): # Get content type from mimetype, defaulting to 'application/octet-stream' content_type = mimetypes.guess_type(filename)[0] or 'application/octet-stream' - # For security, force certain extensions to download instead of display - file_extension = os.path.splitext(filename)[1].lower() - if file_extension in {'.html', '.htm', '.js', '.css'}: + # For security, force certain mimetypes to download instead of display + if content_type in {'text/html', 'text/html-sandboxed', 'application/xhtml+xml', 'text/javascript', 'text/css'}: content_type = 'application/octet-stream' # Forces download return web.FileResponse( @@ -615,6 +614,9 @@ class PromptServer(ExecutorToClientProgress): info['deprecated'] = True if getattr(obj_class, "EXPERIMENTAL", False): info['experimental'] = True + + if hasattr(obj_class, 'API_NODE'): + info['api_node'] = obj_class.API_NODE return info @routes.get("/object_info") @@ -651,7 +653,7 @@ class PromptServer(ExecutorToClientProgress): @routes.get("/queue") async def get_queue(request): queue_info = {} - current_queue = self.prompt_queue.get_current_queue() + current_queue = self.prompt_queue.get_current_queue_volatile() queue_info['queue_running'] = current_queue[0] queue_info['queue_pending'] = current_queue[1] return web.json_response(queue_info) @@ -990,6 +992,13 @@ class PromptServer(ExecutorToClientProgress): web.static('/templates', workflow_templates_path) ]) + # Serve embedded documentation from the package + embedded_docs_path = FrontendManager.embedded_docs_path() + if embedded_docs_path: + self.app.add_routes([ + web.static('/docs', embedded_docs_path) + ]) + self.app.add_routes([ web.static('/', self.web_root, follow_symlinks=True), ]) @@ -1126,3 +1135,15 @@ class PromptServer(ExecutorToClientProgress): @classmethod def get_too_busy_queue_size(cls): return args.max_queue_size + + def send_progress_text( + self, text: Union[bytes, bytearray, str], node_id: str, sid=None + ): + if isinstance(text, str): + text = text.encode("utf-8") + node_id_bytes = str(node_id).encode("utf-8") + + # Pack the node_id length as a 4-byte unsigned integer, followed by the node_id bytes + message = struct.pack(">I", len(node_id_bytes)) + node_id_bytes + text + + self.send_sync(BinaryEventTypes.TEXT, message, sid) diff --git a/comfy/comfy_types/node_typing.py b/comfy/comfy_types/node_typing.py index 8ed4c3d53..ce58e833b 100644 --- a/comfy/comfy_types/node_typing.py +++ b/comfy/comfy_types/node_typing.py @@ -1,7 +1,7 @@ """Comfy-specific type hinting""" from __future__ import annotations -from typing import Literal, TypedDict +from typing import Literal, TypedDict, Optional from typing_extensions import NotRequired from abc import ABC, abstractmethod from enum import Enum @@ -48,6 +48,7 @@ class IO(StrEnum): FACE_ANALYSIS = "FACE_ANALYSIS" BBOX = "BBOX" SEGS = "SEGS" + VIDEO = "VIDEO" ANY = "*" """Always matches any type, but at a price. @@ -115,6 +116,15 @@ class InputTypeOptions(TypedDict): """When a link exists, rather than receiving the evaluated value, you will receive the link (i.e. `["nodeId", ]`). Designed for node expansion.""" tooltip: NotRequired[str] """Tooltip for the input (or widget), shown on pointer hover""" + socketless: NotRequired[bool] + """All inputs (including widgets) have an input socket to connect links. When ``true``, if there is a widget for this input, no socket will be created. + Available from frontend v1.17.5 + Ref: https://github.com/Comfy-Org/ComfyUI_frontend/pull/3548 + """ + widgetType: NotRequired[str] + """Specifies a type to be used for widget initialization if different from the input type. + Available from frontend v1.18.0 + https://github.com/Comfy-Org/ComfyUI_frontend/pull/3550""" # class InputTypeNumber(InputTypeOptions): # default: float | int min: NotRequired[float] @@ -224,6 +234,8 @@ class ComfyNodeABC(ABC): """Flags a node as experimental, informing users that it may change or not work as expected.""" DEPRECATED: bool """Flags a node as deprecated, indicating to users that they should find alternatives to this node.""" + API_NODE: Optional[bool] + """Flags a node as an API node. See: https://docs.comfy.org/tutorials/api-nodes/overview.""" @classmethod @abstractmethod @@ -262,7 +274,7 @@ class ComfyNodeABC(ABC): Comfy Docs: https://docs.comfy.org/custom-nodes/backend/lists#list-processing """ - OUTPUT_IS_LIST: tuple[bool] + OUTPUT_IS_LIST: tuple[bool, ...] """A tuple indicating which node outputs are lists, but will be connected to nodes that expect individual items. Connected nodes that do not implement `INPUT_IS_LIST` will be executed once for every item in the list. @@ -281,7 +293,7 @@ class ComfyNodeABC(ABC): Comfy Docs: https://docs.comfy.org/custom-nodes/backend/lists#list-processing """ - RETURN_TYPES: tuple[IO] + RETURN_TYPES: tuple[IO, ...] """A tuple representing the outputs of this node. Usage:: @@ -290,12 +302,12 @@ class ComfyNodeABC(ABC): Comfy Docs: https://docs.comfy.org/custom-nodes/backend/server_overview#return-types """ - RETURN_NAMES: tuple[str] + RETURN_NAMES: tuple[str, ...] """The output slot names for each item in `RETURN_TYPES`, e.g. ``RETURN_NAMES = ("count", "filter_string")`` Comfy Docs: https://docs.comfy.org/custom-nodes/backend/server_overview#return-names """ - OUTPUT_TOOLTIPS: tuple[str] + OUTPUT_TOOLTIPS: tuple[str, ...] """A tuple of strings to use as tooltips for node outputs, one for each item in `RETURN_TYPES`.""" FUNCTION: str """The name of the function to execute as a literal string, e.g. `FUNCTION = "execute"` diff --git a/comfy/component_model/executor_types.py b/comfy/component_model/executor_types.py index dd73cabf0..efc5cdce9 100644 --- a/comfy/component_model/executor_types.py +++ b/comfy/component_model/executor_types.py @@ -1,11 +1,10 @@ from __future__ import annotations # for Python 3.7-3.9 +import PIL.Image import concurrent.futures import typing from enum import Enum from typing import Optional, Literal, Protocol, Union, NamedTuple, List - -import PIL.Image from typing_extensions import NotRequired, TypedDict from .outputs_types import OutputsDict @@ -121,6 +120,9 @@ class ExecutorToClientProgress(Protocol): """ pass + def send_progress_text(self, text: Union[bytes, bytearray, str], node_id: str, sid=None): + pass + def queue_updated(self, queue_remaining: Optional[int] = None): """ Indicates that the local client's queue has been updated diff --git a/comfy/component_model/queue_types.py b/comfy/component_model/queue_types.py index 05a0432d3..217df738c 100644 --- a/comfy/component_model/queue_types.py +++ b/comfy/component_model/queue_types.py @@ -127,6 +127,7 @@ class QueueItem(NamedQueueTuple): class BinaryEventTypes(Enum): PREVIEW_IMAGE = 1 UNENCODED_PREVIEW_IMAGE = 2 + TEXT = 3 class ExecutorToClientMessage(TypedDict, total=False): diff --git a/comfy/conds.py b/comfy/conds.py index 5a86754bd..cfa3ab173 100644 --- a/comfy/conds.py +++ b/comfy/conds.py @@ -26,6 +26,9 @@ class CONDRegular: conds.append(x.cond) return torch.cat(conds) + def size(self): + return list(self.cond.size()) + class CONDNoiseShape(CONDRegular): def process_cond(self, batch_size, device, area, **kwargs): @@ -82,3 +85,48 @@ class CONDConstant(CONDRegular): def concat(self, others): return self.cond + + def size(self): + return [1] + + +class CONDList(CONDRegular): + def __init__(self, cond): + self.cond = cond + + def process_cond(self, batch_size, device, **kwargs): + out = [] + for c in self.cond: + out.append(comfy.utils.repeat_to_batch_size(c, batch_size).to(device)) + + return self._copy_with(out) + + def can_concat(self, other): + if len(self.cond) != len(other.cond): + return False + for i in range(len(self.cond)): + if self.cond[i].shape != other.cond[i].shape: + return False + + return True + + def concat(self, others): + out = [] + for i in range(len(self.cond)): + o = [self.cond[i]] + for x in others: + o.append(x.cond[i]) + out.append(torch.cat(o)) + + return out + + def size(self): # hackish implementation to make the mem estimation work + o = 0 + c = 1 + for c in self.cond: + size = c.size() + o += math.prod(size) + if len(size) > 1: + c = size[1] + + return [1, c, o // c] diff --git a/comfy/controlnet.py b/comfy/controlnet.py index d946146c1..784ec047e 100644 --- a/comfy/controlnet.py +++ b/comfy/controlnet.py @@ -394,8 +394,9 @@ class ControlLora(ControlNet): pass for k in self.control_weights: - if k not in {"lora_controlnet"}: - utils.set_attr_param(self.control_model, k, self.control_weights[k].to(dtype).to(model_management.get_torch_device())) + if (k not in {"lora_controlnet"}): + if (k.endswith(".up") or k.endswith(".down") or k.endswith(".weight") or k.endswith(".bias")) and ("__" not in k): + utils.set_attr_param(self.control_model, k, self.control_weights[k].to(dtype).to(model_management.get_torch_device())) def copy(self): c = ControlLora(self.control_weights, global_average_pooling=self.global_average_pooling) diff --git a/comfy/distributed/distributed_progress.py b/comfy/distributed/distributed_progress.py index 1a122b28d..95de2b9ea 100644 --- a/comfy/distributed/distributed_progress.py +++ b/comfy/distributed/distributed_progress.py @@ -68,6 +68,10 @@ class DistributedExecutorToClientProgress(ExecutorToClientProgress): # these can gracefully expire pass + def send_progress_text(self, text: str, node_id: str = None): + # todo: we'll fill this out later + pass + def send_sync(self, event: SendSyncEvent, data: SendSyncData, diff --git a/comfy/distributed/server_stub.py b/comfy/distributed/server_stub.py index c21d3a6cc..3a24a1333 100644 --- a/comfy/distributed/server_stub.py +++ b/comfy/distributed/server_stub.py @@ -25,6 +25,9 @@ class ServerStub(ExecutorToClientProgress): def queue_updated(self, queue_remaining: Optional[int] = None): pass + def send_progress_text(self, text: str, node_id: str = None): + pass + @property def receive_all_progress_notifications(self) -> bool: return False diff --git a/comfy/image_encoders/dino2.py b/comfy/image_encoders/dino2.py index 130ed6fd7..976f98c65 100644 --- a/comfy/image_encoders/dino2.py +++ b/comfy/image_encoders/dino2.py @@ -116,7 +116,7 @@ class Dino2Embeddings(torch.nn.Module): def forward(self, pixel_values): x = self.patch_embeddings(pixel_values) # TODO: mask_token? - x = torch.cat((self.cls_token.expand(x.shape[0], -1, -1), x), dim=1) + x = torch.cat((self.cls_token.to(device=x.device, dtype=x.dtype).expand(x.shape[0], -1, -1), x), dim=1) x = x + comfy.model_management.cast_to_device(self.position_embeddings, x.device, x.dtype) return x diff --git a/comfy/k_diffusion/sampling.py b/comfy/k_diffusion/sampling.py index c6149ca11..c2128bd5b 100644 --- a/comfy/k_diffusion/sampling.py +++ b/comfy/k_diffusion/sampling.py @@ -1321,6 +1321,7 @@ def res_multistep(model, x, sigmas, extra_args=None, callback=None, disable=None phi1_fn = lambda t: torch.expm1(t) / t phi2_fn = lambda t: (phi1_fn(t) - 1.0) / t + old_sigma_down = None old_denoised = None uncond_denoised = None def post_cfg_function(args): @@ -1348,9 +1349,9 @@ def res_multistep(model, x, sigmas, extra_args=None, callback=None, disable=None x = x + d * dt else: # Second order multistep method in https://arxiv.org/pdf/2308.02157 - t, t_next, t_prev = t_fn(sigmas[i]), t_fn(sigma_down), t_fn(sigmas[i - 1]) + t, t_old, t_next, t_prev = t_fn(sigmas[i]), t_fn(old_sigma_down), t_fn(sigma_down), t_fn(sigmas[i - 1]) h = t_next - t - c2 = (t_prev - t) / h + c2 = (t_prev - t_old) / h phi1_val, phi2_val = phi1_fn(-h), phi2_fn(-h) b1 = torch.nan_to_num(phi1_val - phi2_val / c2, nan=0.0) @@ -1370,6 +1371,7 @@ def res_multistep(model, x, sigmas, extra_args=None, callback=None, disable=None old_denoised = uncond_denoised else: old_denoised = denoised + old_sigma_down = sigma_down return x @torch.no_grad() @@ -1389,28 +1391,52 @@ def sample_res_multistep_ancestral_cfg_pp(model, x, sigmas, extra_args=None, cal return res_multistep(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, s_noise=s_noise, noise_sampler=noise_sampler, eta=eta, cfg_pp=True) @torch.no_grad() -def sample_gradient_estimation(model, x, sigmas, extra_args=None, callback=None, disable=None, ge_gamma=2.): +def sample_gradient_estimation(model, x, sigmas, extra_args=None, callback=None, disable=None, ge_gamma=2., cfg_pp=False): """Gradient-estimation sampler. Paper: https://openreview.net/pdf?id=o2ND9v0CeK""" extra_args = {} if extra_args is None else extra_args s_in = x.new_ones([x.shape[0]]) old_d = None + uncond_denoised = None + def post_cfg_function(args): + nonlocal uncond_denoised + uncond_denoised = args["uncond_denoised"] + return args["denoised"] + + if cfg_pp: + model_options = extra_args.get("model_options", {}).copy() + extra_args["model_options"] = comfy.model_patcher.set_model_options_post_cfg_function(model_options, post_cfg_function, disable_cfg1_optimization=True) + for i in trange(len(sigmas) - 1, disable=disable): denoised = model(x, sigmas[i] * s_in, **extra_args) - d = to_d(x, sigmas[i], denoised) + if cfg_pp: + d = to_d(x, sigmas[i], uncond_denoised) + else: + d = to_d(x, sigmas[i], denoised) if callback is not None: callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised}) dt = sigmas[i + 1] - sigmas[i] if i == 0: # Euler method - x = x + d * dt + if cfg_pp: + x = denoised + d * sigmas[i + 1] + else: + x = x + d * dt else: # Gradient estimation - d_bar = ge_gamma * d + (1 - ge_gamma) * old_d - x = x + d_bar * dt + if cfg_pp: + d_bar = (ge_gamma - 1) * (d - old_d) + x = denoised + d * sigmas[i + 1] + d_bar * dt + else: + d_bar = ge_gamma * d + (1 - ge_gamma) * old_d + x = x + d_bar * dt old_d = d return x +@torch.no_grad() +def sample_gradient_estimation_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, ge_gamma=2.): + return sample_gradient_estimation(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, ge_gamma=ge_gamma, cfg_pp=True) + @torch.no_grad() def sample_er_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, s_noise=1., noise_sampler=None, noise_scaler=None, max_stage=3): """ diff --git a/comfy/latent_formats.py b/comfy/latent_formats.py index d334c4b0e..71a876498 100644 --- a/comfy/latent_formats.py +++ b/comfy/latent_formats.py @@ -480,3 +480,7 @@ class Hunyuan3Dv2mini(LatentFormat): latent_channels = 64 latent_dimensions = 1 scale_factor = 1.0188137142395404 + +class ACEAudio(LatentFormat): + latent_channels = 8 + latent_dimensions = 2 diff --git a/comfy/ldm/ace/attention.py b/comfy/ldm/ace/attention.py new file mode 100644 index 000000000..f20a01669 --- /dev/null +++ b/comfy/ldm/ace/attention.py @@ -0,0 +1,761 @@ +# Original from: https://github.com/ace-step/ACE-Step/blob/main/models/attention.py +# Copyright 2024 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Tuple, Union, Optional + +import torch +import torch.nn.functional as F +from torch import nn + +import comfy.model_management +from comfy.ldm.modules.attention import optimized_attention + +class Attention(nn.Module): + def __init__( + self, + query_dim: int, + cross_attention_dim: Optional[int] = None, + heads: int = 8, + kv_heads: Optional[int] = None, + dim_head: int = 64, + dropout: float = 0.0, + bias: bool = False, + qk_norm: Optional[str] = None, + added_kv_proj_dim: Optional[int] = None, + added_proj_bias: Optional[bool] = True, + out_bias: bool = True, + scale_qk: bool = True, + only_cross_attention: bool = False, + eps: float = 1e-5, + rescale_output_factor: float = 1.0, + residual_connection: bool = False, + processor=None, + out_dim: int = None, + out_context_dim: int = None, + context_pre_only=None, + pre_only=False, + elementwise_affine: bool = True, + is_causal: bool = False, + dtype=None, device=None, operations=None + ): + super().__init__() + + self.inner_dim = out_dim if out_dim is not None else dim_head * heads + self.inner_kv_dim = self.inner_dim if kv_heads is None else dim_head * kv_heads + self.query_dim = query_dim + self.use_bias = bias + self.is_cross_attention = cross_attention_dim is not None + self.cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim + self.rescale_output_factor = rescale_output_factor + self.residual_connection = residual_connection + self.dropout = dropout + self.fused_projections = False + self.out_dim = out_dim if out_dim is not None else query_dim + self.out_context_dim = out_context_dim if out_context_dim is not None else query_dim + self.context_pre_only = context_pre_only + self.pre_only = pre_only + self.is_causal = is_causal + + self.scale_qk = scale_qk + self.scale = dim_head**-0.5 if self.scale_qk else 1.0 + + self.heads = out_dim // dim_head if out_dim is not None else heads + # for slice_size > 0 the attention score computation + # is split across the batch axis to save memory + # You can set slice_size with `set_attention_slice` + self.sliceable_head_dim = heads + + self.added_kv_proj_dim = added_kv_proj_dim + self.only_cross_attention = only_cross_attention + + if self.added_kv_proj_dim is None and self.only_cross_attention: + raise ValueError( + "`only_cross_attention` can only be set to True if `added_kv_proj_dim` is not None. Make sure to set either `only_cross_attention=False` or define `added_kv_proj_dim`." + ) + + self.group_norm = None + self.spatial_norm = None + + self.norm_q = None + self.norm_k = None + + self.norm_cross = None + self.to_q = operations.Linear(query_dim, self.inner_dim, bias=bias, dtype=dtype, device=device) + + if not self.only_cross_attention: + # only relevant for the `AddedKVProcessor` classes + self.to_k = operations.Linear(self.cross_attention_dim, self.inner_kv_dim, bias=bias, dtype=dtype, device=device) + self.to_v = operations.Linear(self.cross_attention_dim, self.inner_kv_dim, bias=bias, dtype=dtype, device=device) + else: + self.to_k = None + self.to_v = None + + self.added_proj_bias = added_proj_bias + if self.added_kv_proj_dim is not None: + self.add_k_proj = operations.Linear(added_kv_proj_dim, self.inner_kv_dim, bias=added_proj_bias, dtype=dtype, device=device) + self.add_v_proj = operations.Linear(added_kv_proj_dim, self.inner_kv_dim, bias=added_proj_bias, dtype=dtype, device=device) + if self.context_pre_only is not None: + self.add_q_proj = operations.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias, dtype=dtype, device=device) + else: + self.add_q_proj = None + self.add_k_proj = None + self.add_v_proj = None + + if not self.pre_only: + self.to_out = nn.ModuleList([]) + self.to_out.append(operations.Linear(self.inner_dim, self.out_dim, bias=out_bias, dtype=dtype, device=device)) + self.to_out.append(nn.Dropout(dropout)) + else: + self.to_out = None + + if self.context_pre_only is not None and not self.context_pre_only: + self.to_add_out = operations.Linear(self.inner_dim, self.out_context_dim, bias=out_bias, dtype=dtype, device=device) + else: + self.to_add_out = None + + self.norm_added_q = None + self.norm_added_k = None + self.processor = processor + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + **cross_attention_kwargs, + ) -> torch.Tensor: + return self.processor( + self, + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + + +class CustomLiteLAProcessor2_0: + """Attention processor used typically in processing the SD3-like self-attention projections. add rms norm for query and key and apply RoPE""" + + def __init__(self): + self.kernel_func = nn.ReLU(inplace=False) + self.eps = 1e-15 + self.pad_val = 1.0 + + def apply_rotary_emb( + self, + x: torch.Tensor, + freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]], + ) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Apply rotary embeddings to input tensors using the given frequency tensor. This function applies rotary embeddings + to the given query or key 'x' 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 (`torch.Tensor`): + Query or key tensor to apply rotary embeddings. [B, H, S, D] xk (torch.Tensor): Key tensor to apply + freqs_cis (`Tuple[torch.Tensor]`): Precomputed frequency tensor for complex exponentials. ([S, D], [S, D],) + + Returns: + Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor and key tensor with rotary embeddings. + """ + cos, sin = freqs_cis # [S, D] + cos = cos[None, None] + sin = sin[None, None] + cos, sin = cos.to(x.device), sin.to(x.device) + + x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1) # [B, S, H, D//2] + x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3) + out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype) + + return out + + def __call__( + self, + attn: Attention, + hidden_states: torch.FloatTensor, + encoder_hidden_states: torch.FloatTensor = None, + attention_mask: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + rotary_freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]] = None, + rotary_freqs_cis_cross: Union[torch.Tensor, Tuple[torch.Tensor]] = None, + *args, + **kwargs, + ) -> torch.FloatTensor: + hidden_states_len = hidden_states.shape[1] + + input_ndim = hidden_states.ndim + if input_ndim == 4: + batch_size, channel, height, width = hidden_states.shape + hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) + if encoder_hidden_states is not None: + context_input_ndim = encoder_hidden_states.ndim + if context_input_ndim == 4: + batch_size, channel, height, width = encoder_hidden_states.shape + encoder_hidden_states = encoder_hidden_states.view(batch_size, channel, height * width).transpose(1, 2) + + batch_size = hidden_states.shape[0] + + # `sample` projections. + dtype = hidden_states.dtype + query = attn.to_q(hidden_states) + key = attn.to_k(hidden_states) + value = attn.to_v(hidden_states) + + # `context` projections. + has_encoder_hidden_state_proj = hasattr(attn, "add_q_proj") and hasattr(attn, "add_k_proj") and hasattr(attn, "add_v_proj") + if encoder_hidden_states is not None and has_encoder_hidden_state_proj: + encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states) + encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states) + encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states) + + # attention + if not attn.is_cross_attention: + query = torch.cat([query, encoder_hidden_states_query_proj], dim=1) + key = torch.cat([key, encoder_hidden_states_key_proj], dim=1) + value = torch.cat([value, encoder_hidden_states_value_proj], dim=1) + else: + query = hidden_states + key = encoder_hidden_states + value = encoder_hidden_states + + inner_dim = key.shape[-1] + head_dim = inner_dim // attn.heads + + query = query.transpose(-1, -2).reshape(batch_size, attn.heads, head_dim, -1) + key = key.transpose(-1, -2).reshape(batch_size, attn.heads, head_dim, -1).transpose(-1, -2) + value = value.transpose(-1, -2).reshape(batch_size, attn.heads, head_dim, -1) + + # RoPE需要 [B, H, S, D] 输入 + # 此时 query是 [B, H, D, S], 需要转成 [B, H, S, D] 才能应用RoPE + query = query.permute(0, 1, 3, 2) # [B, H, S, D] (从 [B, H, D, S]) + + # Apply query and key normalization if needed + if attn.norm_q is not None: + query = attn.norm_q(query) + if attn.norm_k is not None: + key = attn.norm_k(key) + + # Apply RoPE if needed + if rotary_freqs_cis is not None: + query = self.apply_rotary_emb(query, rotary_freqs_cis) + if not attn.is_cross_attention: + key = self.apply_rotary_emb(key, rotary_freqs_cis) + elif rotary_freqs_cis_cross is not None and has_encoder_hidden_state_proj: + key = self.apply_rotary_emb(key, rotary_freqs_cis_cross) + + # 此时 query是 [B, H, S, D]，需要还原成 [B, H, D, S] + query = query.permute(0, 1, 3, 2) # [B, H, D, S] + + if attention_mask is not None: + # attention_mask: [B, S] -> [B, 1, S, 1] + attention_mask = attention_mask[:, None, :, None].to(key.dtype) # [B, 1, S, 1] + query = query * attention_mask.permute(0, 1, 3, 2) # [B, H, S, D] * [B, 1, S, 1] + if not attn.is_cross_attention: + key = key * attention_mask # key: [B, h, S, D] 与 mask [B, 1, S, 1] 相乘 + value = value * attention_mask.permute(0, 1, 3, 2) # 如果 value 是 [B, h, D, S]，那么需调整mask以匹配S维度 + + if attn.is_cross_attention and encoder_attention_mask is not None and has_encoder_hidden_state_proj: + encoder_attention_mask = encoder_attention_mask[:, None, :, None].to(key.dtype) # [B, 1, S_enc, 1] + # 此时 key: [B, h, S_enc, D], value: [B, h, D, S_enc] + key = key * encoder_attention_mask # [B, h, S_enc, D] * [B, 1, S_enc, 1] + value = value * encoder_attention_mask.permute(0, 1, 3, 2) # [B, h, D, S_enc] * [B, 1, 1, S_enc] + + query = self.kernel_func(query) + key = self.kernel_func(key) + + query, key, value = query.float(), key.float(), value.float() + + value = F.pad(value, (0, 0, 0, 1), mode="constant", value=self.pad_val) + + vk = torch.matmul(value, key) + + hidden_states = torch.matmul(vk, query) + + if hidden_states.dtype in [torch.float16, torch.bfloat16]: + hidden_states = hidden_states.float() + + hidden_states = hidden_states[:, :, :-1] / (hidden_states[:, :, -1:] + self.eps) + + hidden_states = hidden_states.view(batch_size, attn.heads * head_dim, -1).permute(0, 2, 1) + + hidden_states = hidden_states.to(dtype) + if encoder_hidden_states is not None: + encoder_hidden_states = encoder_hidden_states.to(dtype) + + # Split the attention outputs. + if encoder_hidden_states is not None and not attn.is_cross_attention and has_encoder_hidden_state_proj: + hidden_states, encoder_hidden_states = ( + hidden_states[:, : hidden_states_len], + hidden_states[:, hidden_states_len:], + ) + + # linear proj + hidden_states = attn.to_out[0](hidden_states) + # dropout + hidden_states = attn.to_out[1](hidden_states) + if encoder_hidden_states is not None and not attn.context_pre_only and not attn.is_cross_attention and hasattr(attn, "to_add_out"): + encoder_hidden_states = attn.to_add_out(encoder_hidden_states) + + if input_ndim == 4: + hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width) + if encoder_hidden_states is not None and context_input_ndim == 4: + encoder_hidden_states = encoder_hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width) + + if torch.get_autocast_gpu_dtype() == torch.float16: + hidden_states = hidden_states.clip(-65504, 65504) + if encoder_hidden_states is not None: + encoder_hidden_states = encoder_hidden_states.clip(-65504, 65504) + + return hidden_states, encoder_hidden_states + + +class CustomerAttnProcessor2_0: + r""" + Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). + """ + + def apply_rotary_emb( + self, + x: torch.Tensor, + freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]], + ) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Apply rotary embeddings to input tensors using the given frequency tensor. This function applies rotary embeddings + to the given query or key 'x' 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 (`torch.Tensor`): + Query or key tensor to apply rotary embeddings. [B, H, S, D] xk (torch.Tensor): Key tensor to apply + freqs_cis (`Tuple[torch.Tensor]`): Precomputed frequency tensor for complex exponentials. ([S, D], [S, D],) + + Returns: + Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor and key tensor with rotary embeddings. + """ + cos, sin = freqs_cis # [S, D] + cos = cos[None, None] + sin = sin[None, None] + cos, sin = cos.to(x.device), sin.to(x.device) + + x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1) # [B, S, H, D//2] + x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3) + out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype) + + return out + + def __call__( + self, + attn: Attention, + hidden_states: torch.FloatTensor, + encoder_hidden_states: torch.FloatTensor = None, + attention_mask: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + rotary_freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]] = None, + rotary_freqs_cis_cross: Union[torch.Tensor, Tuple[torch.Tensor]] = None, + *args, + **kwargs, + ) -> torch.Tensor: + + residual = hidden_states + input_ndim = hidden_states.ndim + + if input_ndim == 4: + batch_size, channel, height, width = hidden_states.shape + hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) + + batch_size, sequence_length, _ = ( + hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape + ) + + has_encoder_hidden_state_proj = hasattr(attn, "add_q_proj") and hasattr(attn, "add_k_proj") and hasattr(attn, "add_v_proj") + + if attn.group_norm is not None: + hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) + + query = attn.to_q(hidden_states) + + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + key = attn.to_k(encoder_hidden_states) + value = attn.to_v(encoder_hidden_states) + + inner_dim = key.shape[-1] + head_dim = inner_dim // attn.heads + + query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + + key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + + if attn.norm_q is not None: + query = attn.norm_q(query) + if attn.norm_k is not None: + key = attn.norm_k(key) + + # Apply RoPE if needed + if rotary_freqs_cis is not None: + query = self.apply_rotary_emb(query, rotary_freqs_cis) + if not attn.is_cross_attention: + key = self.apply_rotary_emb(key, rotary_freqs_cis) + elif rotary_freqs_cis_cross is not None and has_encoder_hidden_state_proj: + key = self.apply_rotary_emb(key, rotary_freqs_cis_cross) + + if attn.is_cross_attention and encoder_attention_mask is not None and has_encoder_hidden_state_proj: + # attention_mask: N x S1 + # encoder_attention_mask: N x S2 + # cross attention 整合attention_mask和encoder_attention_mask + combined_mask = attention_mask[:, :, None] * encoder_attention_mask[:, None, :] + attention_mask = torch.where(combined_mask == 1, 0.0, -torch.inf) + attention_mask = attention_mask[:, None, :, :].expand(-1, attn.heads, -1, -1).to(query.dtype) + + elif not attn.is_cross_attention and attention_mask is not None: + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + # scaled_dot_product_attention expects attention_mask shape to be + # (batch, heads, source_length, target_length) + attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) + + # the output of sdp = (batch, num_heads, seq_len, head_dim) + hidden_states = optimized_attention( + query, key, value, heads=query.shape[1], mask=attention_mask, skip_reshape=True, + ).to(query.dtype) + + # linear proj + hidden_states = attn.to_out[0](hidden_states) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + if input_ndim == 4: + hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width) + + if attn.residual_connection: + hidden_states = hidden_states + residual + + hidden_states = hidden_states / attn.rescale_output_factor + + return hidden_states + +def val2list(x: list or tuple or any, repeat_time=1) -> list: # type: ignore + """Repeat `val` for `repeat_time` times and return the list or val if list/tuple.""" + if isinstance(x, (list, tuple)): + return list(x) + return [x for _ in range(repeat_time)] + + +def val2tuple(x: list or tuple or any, min_len: int = 1, idx_repeat: int = -1) -> tuple: # type: ignore + """Return tuple with min_len by repeating element at idx_repeat.""" + # convert to list first + x = val2list(x) + + # repeat elements if necessary + if len(x) > 0: + x[idx_repeat:idx_repeat] = [x[idx_repeat] for _ in range(min_len - len(x))] + + return tuple(x) + + +def t2i_modulate(x, shift, scale): + return x * (1 + scale) + shift + + +def get_same_padding(kernel_size: Union[int, Tuple[int, ...]]) -> Union[int, Tuple[int, ...]]: + if isinstance(kernel_size, tuple): + return tuple([get_same_padding(ks) for ks in kernel_size]) + else: + assert kernel_size % 2 > 0, f"kernel size {kernel_size} should be odd number" + return kernel_size // 2 + +class ConvLayer(nn.Module): + def __init__( + self, + in_dim: int, + out_dim: int, + kernel_size=3, + stride=1, + dilation=1, + groups=1, + padding: Union[int, None] = None, + use_bias=False, + norm=None, + act=None, + dtype=None, device=None, operations=None + ): + super().__init__() + if padding is None: + padding = get_same_padding(kernel_size) + padding *= dilation + + self.in_dim = in_dim + self.out_dim = out_dim + self.kernel_size = kernel_size + self.stride = stride + self.dilation = dilation + self.groups = groups + self.padding = padding + self.use_bias = use_bias + + self.conv = operations.Conv1d( + in_dim, + out_dim, + kernel_size=kernel_size, + stride=stride, + padding=padding, + dilation=dilation, + groups=groups, + bias=use_bias, + device=device, + dtype=dtype + ) + if norm is not None: + self.norm = operations.RMSNorm(out_dim, elementwise_affine=False, dtype=dtype, device=device) + else: + self.norm = None + if act is not None: + self.act = nn.SiLU(inplace=True) + else: + self.act = None + + def forward(self, x: torch.Tensor) -> torch.Tensor: + x = self.conv(x) + if self.norm: + x = self.norm(x) + if self.act: + x = self.act(x) + return x + + +class GLUMBConv(nn.Module): + def __init__( + self, + in_features: int, + hidden_features: int, + out_feature=None, + kernel_size=3, + stride=1, + padding: Union[int, None] = None, + use_bias=False, + norm=(None, None, None), + act=("silu", "silu", None), + dilation=1, + dtype=None, device=None, operations=None + ): + out_feature = out_feature or in_features + super().__init__() + use_bias = val2tuple(use_bias, 3) + norm = val2tuple(norm, 3) + act = val2tuple(act, 3) + + self.glu_act = nn.SiLU(inplace=False) + self.inverted_conv = ConvLayer( + in_features, + hidden_features * 2, + 1, + use_bias=use_bias[0], + norm=norm[0], + act=act[0], + dtype=dtype, + device=device, + operations=operations, + ) + self.depth_conv = ConvLayer( + hidden_features * 2, + hidden_features * 2, + kernel_size, + stride=stride, + groups=hidden_features * 2, + padding=padding, + use_bias=use_bias[1], + norm=norm[1], + act=None, + dilation=dilation, + dtype=dtype, + device=device, + operations=operations, + ) + self.point_conv = ConvLayer( + hidden_features, + out_feature, + 1, + use_bias=use_bias[2], + norm=norm[2], + act=act[2], + dtype=dtype, + device=device, + operations=operations, + ) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + x = x.transpose(1, 2) + x = self.inverted_conv(x) + x = self.depth_conv(x) + + x, gate = torch.chunk(x, 2, dim=1) + gate = self.glu_act(gate) + x = x * gate + + x = self.point_conv(x) + x = x.transpose(1, 2) + + return x + + +class LinearTransformerBlock(nn.Module): + """ + A Sana block with global shared adaptive layer norm (adaLN-single) conditioning. + """ + def __init__( + self, + dim, + num_attention_heads, + attention_head_dim, + use_adaln_single=True, + cross_attention_dim=None, + added_kv_proj_dim=None, + context_pre_only=False, + mlp_ratio=4.0, + add_cross_attention=False, + add_cross_attention_dim=None, + qk_norm=None, + dtype=None, device=None, operations=None + ): + super().__init__() + + self.norm1 = operations.RMSNorm(dim, elementwise_affine=False, eps=1e-6) + self.attn = Attention( + query_dim=dim, + cross_attention_dim=cross_attention_dim, + added_kv_proj_dim=added_kv_proj_dim, + dim_head=attention_head_dim, + heads=num_attention_heads, + out_dim=dim, + bias=True, + qk_norm=qk_norm, + processor=CustomLiteLAProcessor2_0(), + dtype=dtype, + device=device, + operations=operations, + ) + + self.add_cross_attention = add_cross_attention + self.context_pre_only = context_pre_only + + if add_cross_attention and add_cross_attention_dim is not None: + self.cross_attn = Attention( + query_dim=dim, + cross_attention_dim=add_cross_attention_dim, + added_kv_proj_dim=add_cross_attention_dim, + dim_head=attention_head_dim, + heads=num_attention_heads, + out_dim=dim, + context_pre_only=context_pre_only, + bias=True, + qk_norm=qk_norm, + processor=CustomerAttnProcessor2_0(), + dtype=dtype, + device=device, + operations=operations, + ) + + self.norm2 = operations.RMSNorm(dim, 1e-06, elementwise_affine=False) + + self.ff = GLUMBConv( + in_features=dim, + hidden_features=int(dim * mlp_ratio), + use_bias=(True, True, False), + norm=(None, None, None), + act=("silu", "silu", None), + dtype=dtype, + device=device, + operations=operations, + ) + self.use_adaln_single = use_adaln_single + if use_adaln_single: + self.scale_shift_table = nn.Parameter(torch.empty(6, dim, dtype=dtype, device=device)) + + def forward( + self, + hidden_states: torch.FloatTensor, + encoder_hidden_states: torch.FloatTensor = None, + attention_mask: torch.FloatTensor = None, + encoder_attention_mask: torch.FloatTensor = None, + rotary_freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]] = None, + rotary_freqs_cis_cross: Union[torch.Tensor, Tuple[torch.Tensor]] = None, + temb: torch.FloatTensor = None, + ): + + N = hidden_states.shape[0] + + # step 1: AdaLN single + if self.use_adaln_single: + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ( + comfy.model_management.cast_to(self.scale_shift_table[None], dtype=temb.dtype, device=temb.device) + temb.reshape(N, 6, -1) + ).chunk(6, dim=1) + + norm_hidden_states = self.norm1(hidden_states) + if self.use_adaln_single: + norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa + + # step 2: attention + if not self.add_cross_attention: + attn_output, encoder_hidden_states = self.attn( + hidden_states=norm_hidden_states, + attention_mask=attention_mask, + encoder_hidden_states=encoder_hidden_states, + encoder_attention_mask=encoder_attention_mask, + rotary_freqs_cis=rotary_freqs_cis, + rotary_freqs_cis_cross=rotary_freqs_cis_cross, + ) + else: + attn_output, _ = self.attn( + hidden_states=norm_hidden_states, + attention_mask=attention_mask, + encoder_hidden_states=None, + encoder_attention_mask=None, + rotary_freqs_cis=rotary_freqs_cis, + rotary_freqs_cis_cross=None, + ) + + if self.use_adaln_single: + attn_output = gate_msa * attn_output + hidden_states = attn_output + hidden_states + + if self.add_cross_attention: + attn_output = self.cross_attn( + hidden_states=hidden_states, + attention_mask=attention_mask, + encoder_hidden_states=encoder_hidden_states, + encoder_attention_mask=encoder_attention_mask, + rotary_freqs_cis=rotary_freqs_cis, + rotary_freqs_cis_cross=rotary_freqs_cis_cross, + ) + hidden_states = attn_output + hidden_states + + # step 3: add norm + norm_hidden_states = self.norm2(hidden_states) + if self.use_adaln_single: + norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp + + # step 4: feed forward + ff_output = self.ff(norm_hidden_states) + if self.use_adaln_single: + ff_output = gate_mlp * ff_output + + hidden_states = hidden_states + ff_output + + return hidden_states diff --git a/comfy/ldm/ace/lyric_encoder.py b/comfy/ldm/ace/lyric_encoder.py new file mode 100644 index 000000000..ff4359b26 --- /dev/null +++ b/comfy/ldm/ace/lyric_encoder.py @@ -0,0 +1,1067 @@ +# Original from: https://github.com/ace-step/ACE-Step/blob/main/models/lyrics_utils/lyric_encoder.py +from typing import Optional, Tuple, Union +import math +import torch +from torch import nn + +import comfy.model_management + +class ConvolutionModule(nn.Module): + """ConvolutionModule in Conformer model.""" + + def __init__(self, + channels: int, + kernel_size: int = 15, + activation: nn.Module = nn.ReLU(), + norm: str = "batch_norm", + causal: bool = False, + bias: bool = True, + dtype=None, device=None, operations=None): + """Construct an ConvolutionModule object. + Args: + channels (int): The number of channels of conv layers. + kernel_size (int): Kernel size of conv layers. + causal (int): Whether use causal convolution or not + """ + super().__init__() + + self.pointwise_conv1 = operations.Conv1d( + channels, + 2 * channels, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + dtype=dtype, device=device + ) + # self.lorder is used to distinguish if it's a causal convolution, + # if self.lorder > 0: it's a causal convolution, the input will be + # padded with self.lorder frames on the left in forward. + # else: it's a symmetrical convolution + if causal: + padding = 0 + self.lorder = kernel_size - 1 + else: + # kernel_size should be an odd number for none causal convolution + assert (kernel_size - 1) % 2 == 0 + padding = (kernel_size - 1) // 2 + self.lorder = 0 + self.depthwise_conv = operations.Conv1d( + channels, + channels, + kernel_size, + stride=1, + padding=padding, + groups=channels, + bias=bias, + dtype=dtype, device=device + ) + + assert norm in ['batch_norm', 'layer_norm'] + if norm == "batch_norm": + self.use_layer_norm = False + self.norm = nn.BatchNorm1d(channels) + else: + self.use_layer_norm = True + self.norm = operations.LayerNorm(channels, dtype=dtype, device=device) + + self.pointwise_conv2 = operations.Conv1d( + channels, + channels, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + dtype=dtype, device=device + ) + self.activation = activation + + def forward( + self, + x: torch.Tensor, + mask_pad: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool), + cache: torch.Tensor = torch.zeros((0, 0, 0)), + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Compute convolution module. + Args: + x (torch.Tensor): Input tensor (#batch, time, channels). + mask_pad (torch.Tensor): used for batch padding (#batch, 1, time), + (0, 0, 0) means fake mask. + cache (torch.Tensor): left context cache, it is only + used in causal convolution (#batch, channels, cache_t), + (0, 0, 0) meas fake cache. + Returns: + torch.Tensor: Output tensor (#batch, time, channels). + """ + # exchange the temporal dimension and the feature dimension + x = x.transpose(1, 2) # (#batch, channels, time) + + # mask batch padding + if mask_pad.size(2) > 0: # time > 0 + x.masked_fill_(~mask_pad, 0.0) + + if self.lorder > 0: + if cache.size(2) == 0: # cache_t == 0 + x = nn.functional.pad(x, (self.lorder, 0), 'constant', 0.0) + else: + assert cache.size(0) == x.size(0) # equal batch + assert cache.size(1) == x.size(1) # equal channel + x = torch.cat((cache, x), dim=2) + assert (x.size(2) > self.lorder) + new_cache = x[:, :, -self.lorder:] + else: + # It's better we just return None if no cache is required, + # However, for JIT export, here we just fake one tensor instead of + # None. + new_cache = torch.zeros((0, 0, 0), dtype=x.dtype, device=x.device) + + # GLU mechanism + x = self.pointwise_conv1(x) # (batch, 2*channel, dim) + x = nn.functional.glu(x, dim=1) # (batch, channel, dim) + + # 1D Depthwise Conv + x = self.depthwise_conv(x) + if self.use_layer_norm: + x = x.transpose(1, 2) + x = self.activation(self.norm(x)) + if self.use_layer_norm: + x = x.transpose(1, 2) + x = self.pointwise_conv2(x) + # mask batch padding + if mask_pad.size(2) > 0: # time > 0 + x.masked_fill_(~mask_pad, 0.0) + + return x.transpose(1, 2), new_cache + +class PositionwiseFeedForward(torch.nn.Module): + """Positionwise feed forward layer. + + FeedForward are appied on each position of the sequence. + The output dim is same with the input dim. + + Args: + idim (int): Input dimenstion. + hidden_units (int): The number of hidden units. + dropout_rate (float): Dropout rate. + activation (torch.nn.Module): Activation function + """ + + def __init__( + self, + idim: int, + hidden_units: int, + dropout_rate: float, + activation: torch.nn.Module = torch.nn.ReLU(), + dtype=None, device=None, operations=None + ): + """Construct a PositionwiseFeedForward object.""" + super(PositionwiseFeedForward, self).__init__() + self.w_1 = operations.Linear(idim, hidden_units, dtype=dtype, device=device) + self.activation = activation + self.dropout = torch.nn.Dropout(dropout_rate) + self.w_2 = operations.Linear(hidden_units, idim, dtype=dtype, device=device) + + def forward(self, xs: torch.Tensor) -> torch.Tensor: + """Forward function. + + Args: + xs: input tensor (B, L, D) + Returns: + output tensor, (B, L, D) + """ + return self.w_2(self.dropout(self.activation(self.w_1(xs)))) + +class Swish(torch.nn.Module): + """Construct an Swish object.""" + + def forward(self, x: torch.Tensor) -> torch.Tensor: + """Return Swish activation function.""" + return x * torch.sigmoid(x) + +class MultiHeadedAttention(nn.Module): + """Multi-Head Attention layer. + + Args: + n_head (int): The number of heads. + n_feat (int): The number of features. + dropout_rate (float): Dropout rate. + + """ + + def __init__(self, + n_head: int, + n_feat: int, + dropout_rate: float, + key_bias: bool = True, + dtype=None, device=None, operations=None): + """Construct an MultiHeadedAttention object.""" + super().__init__() + assert n_feat % n_head == 0 + # We assume d_v always equals d_k + self.d_k = n_feat // n_head + self.h = n_head + self.linear_q = operations.Linear(n_feat, n_feat, dtype=dtype, device=device) + self.linear_k = operations.Linear(n_feat, n_feat, bias=key_bias, dtype=dtype, device=device) + self.linear_v = operations.Linear(n_feat, n_feat, dtype=dtype, device=device) + self.linear_out = operations.Linear(n_feat, n_feat, dtype=dtype, device=device) + self.dropout = nn.Dropout(p=dropout_rate) + + def forward_qkv( + self, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """Transform query, key and value. + + Args: + query (torch.Tensor): Query tensor (#batch, time1, size). + key (torch.Tensor): Key tensor (#batch, time2, size). + value (torch.Tensor): Value tensor (#batch, time2, size). + + Returns: + torch.Tensor: Transformed query tensor, size + (#batch, n_head, time1, d_k). + torch.Tensor: Transformed key tensor, size + (#batch, n_head, time2, d_k). + torch.Tensor: Transformed value tensor, size + (#batch, n_head, time2, d_k). + + """ + n_batch = query.size(0) + q = self.linear_q(query).view(n_batch, -1, self.h, self.d_k) + k = self.linear_k(key).view(n_batch, -1, self.h, self.d_k) + v = self.linear_v(value).view(n_batch, -1, self.h, self.d_k) + q = q.transpose(1, 2) # (batch, head, time1, d_k) + k = k.transpose(1, 2) # (batch, head, time2, d_k) + v = v.transpose(1, 2) # (batch, head, time2, d_k) + return q, k, v + + def forward_attention( + self, + value: torch.Tensor, + scores: torch.Tensor, + mask: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool) + ) -> torch.Tensor: + """Compute attention context vector. + + Args: + value (torch.Tensor): Transformed value, size + (#batch, n_head, time2, d_k). + scores (torch.Tensor): Attention score, size + (#batch, n_head, time1, time2). + mask (torch.Tensor): Mask, size (#batch, 1, time2) or + (#batch, time1, time2), (0, 0, 0) means fake mask. + + Returns: + torch.Tensor: Transformed value (#batch, time1, d_model) + weighted by the attention score (#batch, time1, time2). + + """ + n_batch = value.size(0) + + if mask is not None and mask.size(2) > 0: # time2 > 0 + mask = mask.unsqueeze(1).eq(0) # (batch, 1, *, time2) + # For last chunk, time2 might be larger than scores.size(-1) + mask = mask[:, :, :, :scores.size(-1)] # (batch, 1, *, time2) + scores = scores.masked_fill(mask, -float('inf')) + attn = torch.softmax(scores, dim=-1).masked_fill( + mask, 0.0) # (batch, head, time1, time2) + + else: + attn = torch.softmax(scores, dim=-1) # (batch, head, time1, time2) + + p_attn = self.dropout(attn) + x = torch.matmul(p_attn, value) # (batch, head, time1, d_k) + x = (x.transpose(1, 2).contiguous().view(n_batch, -1, + self.h * self.d_k) + ) # (batch, time1, d_model) + + return self.linear_out(x) # (batch, time1, d_model) + + def forward( + self, + query: torch.Tensor, + key: torch.Tensor, + value: torch.Tensor, + mask: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool), + pos_emb: torch.Tensor = torch.empty(0), + cache: torch.Tensor = torch.zeros((0, 0, 0, 0)) + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Compute scaled dot product attention. + + Args: + query (torch.Tensor): Query tensor (#batch, time1, size). + key (torch.Tensor): Key tensor (#batch, time2, size). + value (torch.Tensor): Value tensor (#batch, time2, size). + mask (torch.Tensor): Mask tensor (#batch, 1, time2) or + (#batch, time1, time2). + 1.When applying cross attention between decoder and encoder, + the batch padding mask for input is in (#batch, 1, T) shape. + 2.When applying self attention of encoder, + the mask is in (#batch, T, T) shape. + 3.When applying self attention of decoder, + the mask is in (#batch, L, L) shape. + 4.If the different position in decoder see different block + of the encoder, such as Mocha, the passed in mask could be + in (#batch, L, T) shape. But there is no such case in current + CosyVoice. + cache (torch.Tensor): Cache tensor (1, head, cache_t, d_k * 2), + where `cache_t == chunk_size * num_decoding_left_chunks` + and `head * d_k == size` + + + Returns: + torch.Tensor: Output tensor (#batch, time1, d_model). + torch.Tensor: Cache tensor (1, head, cache_t + time1, d_k * 2) + where `cache_t == chunk_size * num_decoding_left_chunks` + and `head * d_k == size` + + """ + q, k, v = self.forward_qkv(query, key, value) + if cache.size(0) > 0: + key_cache, value_cache = torch.split(cache, + cache.size(-1) // 2, + dim=-1) + k = torch.cat([key_cache, k], dim=2) + v = torch.cat([value_cache, v], dim=2) + new_cache = torch.cat((k, v), dim=-1) + + scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.d_k) + return self.forward_attention(v, scores, mask), new_cache + + +class RelPositionMultiHeadedAttention(MultiHeadedAttention): + """Multi-Head Attention layer with relative position encoding. + Paper: https://arxiv.org/abs/1901.02860 + Args: + n_head (int): The number of heads. + n_feat (int): The number of features. + dropout_rate (float): Dropout rate. + """ + + def __init__(self, + n_head: int, + n_feat: int, + dropout_rate: float, + key_bias: bool = True, + dtype=None, device=None, operations=None): + """Construct an RelPositionMultiHeadedAttention object.""" + super().__init__(n_head, n_feat, dropout_rate, key_bias, dtype=dtype, device=device, operations=operations) + # linear transformation for positional encoding + self.linear_pos = operations.Linear(n_feat, n_feat, bias=False, dtype=dtype, device=device) + # these two learnable bias are used in matrix c and matrix d + # as described in https://arxiv.org/abs/1901.02860 Section 3.3 + self.pos_bias_u = nn.Parameter(torch.empty(self.h, self.d_k, dtype=dtype, device=device)) + self.pos_bias_v = nn.Parameter(torch.empty(self.h, self.d_k, dtype=dtype, device=device)) + # torch.nn.init.xavier_uniform_(self.pos_bias_u) + # torch.nn.init.xavier_uniform_(self.pos_bias_v) + + def rel_shift(self, x: torch.Tensor) -> torch.Tensor: + """Compute relative positional encoding. + + Args: + x (torch.Tensor): Input tensor (batch, head, time1, 2*time1-1). + time1 means the length of query vector. + + Returns: + torch.Tensor: Output tensor. + + """ + zero_pad = torch.zeros((x.size()[0], x.size()[1], x.size()[2], 1), + device=x.device, + dtype=x.dtype) + x_padded = torch.cat([zero_pad, x], dim=-1) + + x_padded = x_padded.view(x.size()[0], + x.size()[1], + x.size(3) + 1, x.size(2)) + x = x_padded[:, :, 1:].view_as(x)[ + :, :, :, : x.size(-1) // 2 + 1 + ] # only keep the positions from 0 to time2 + return x + + def forward( + self, + query: torch.Tensor, + key: torch.Tensor, + value: torch.Tensor, + mask: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool), + pos_emb: torch.Tensor = torch.empty(0), + cache: torch.Tensor = torch.zeros((0, 0, 0, 0)) + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Compute 'Scaled Dot Product Attention' with rel. positional encoding. + Args: + query (torch.Tensor): Query tensor (#batch, time1, size). + key (torch.Tensor): Key tensor (#batch, time2, size). + value (torch.Tensor): Value tensor (#batch, time2, size). + mask (torch.Tensor): Mask tensor (#batch, 1, time2) or + (#batch, time1, time2), (0, 0, 0) means fake mask. + pos_emb (torch.Tensor): Positional embedding tensor + (#batch, time2, size). + cache (torch.Tensor): Cache tensor (1, head, cache_t, d_k * 2), + where `cache_t == chunk_size * num_decoding_left_chunks` + and `head * d_k == size` + Returns: + torch.Tensor: Output tensor (#batch, time1, d_model). + torch.Tensor: Cache tensor (1, head, cache_t + time1, d_k * 2) + where `cache_t == chunk_size * num_decoding_left_chunks` + and `head * d_k == size` + """ + q, k, v = self.forward_qkv(query, key, value) + q = q.transpose(1, 2) # (batch, time1, head, d_k) + + if cache.size(0) > 0: + key_cache, value_cache = torch.split(cache, + cache.size(-1) // 2, + dim=-1) + k = torch.cat([key_cache, k], dim=2) + v = torch.cat([value_cache, v], dim=2) + # NOTE(xcsong): We do cache slicing in encoder.forward_chunk, since it's + # non-trivial to calculate `next_cache_start` here. + new_cache = torch.cat((k, v), dim=-1) + + n_batch_pos = pos_emb.size(0) + p = self.linear_pos(pos_emb).view(n_batch_pos, -1, self.h, self.d_k) + p = p.transpose(1, 2) # (batch, head, time1, d_k) + + # (batch, head, time1, d_k) + q_with_bias_u = (q + comfy.model_management.cast_to(self.pos_bias_u, dtype=q.dtype, device=q.device)).transpose(1, 2) + # (batch, head, time1, d_k) + q_with_bias_v = (q + comfy.model_management.cast_to(self.pos_bias_v, dtype=q.dtype, device=q.device)).transpose(1, 2) + + # compute attention score + # first compute matrix a and matrix c + # as described in https://arxiv.org/abs/1901.02860 Section 3.3 + # (batch, head, time1, time2) + matrix_ac = torch.matmul(q_with_bias_u, k.transpose(-2, -1)) + + # compute matrix b and matrix d + # (batch, head, time1, time2) + matrix_bd = torch.matmul(q_with_bias_v, p.transpose(-2, -1)) + # NOTE(Xiang Lyu): Keep rel_shift since espnet rel_pos_emb is used + if matrix_ac.shape != matrix_bd.shape: + matrix_bd = self.rel_shift(matrix_bd) + + scores = (matrix_ac + matrix_bd) / math.sqrt( + self.d_k) # (batch, head, time1, time2) + + return self.forward_attention(v, scores, mask), new_cache + + + +def subsequent_mask( + size: int, + device: torch.device = torch.device("cpu"), +) -> torch.Tensor: + """Create mask for subsequent steps (size, size). + + This mask is used only in decoder which works in an auto-regressive mode. + This means the current step could only do attention with its left steps. + + In encoder, fully attention is used when streaming is not necessary and + the sequence is not long. In this case, no attention mask is needed. + + When streaming is need, chunk-based attention is used in encoder. See + subsequent_chunk_mask for the chunk-based attention mask. + + Args: + size (int): size of mask + str device (str): "cpu" or "cuda" or torch.Tensor.device + dtype (torch.device): result dtype + + Returns: + torch.Tensor: mask + + Examples: + >>> subsequent_mask(3) + [[1, 0, 0], + [1, 1, 0], + [1, 1, 1]] + """ + arange = torch.arange(size, device=device) + mask = arange.expand(size, size) + arange = arange.unsqueeze(-1) + mask = mask <= arange + return mask + + +def subsequent_chunk_mask( + size: int, + chunk_size: int, + num_left_chunks: int = -1, + device: torch.device = torch.device("cpu"), + ) -> torch.Tensor: + """Create mask for subsequent steps (size, size) with chunk size, + this is for streaming encoder + + Args: + size (int): size of mask + chunk_size (int): size of chunk + num_left_chunks (int): number of left chunks + <0: use full chunk + >=0: use num_left_chunks + device (torch.device): "cpu" or "cuda" or torch.Tensor.device + + Returns: + torch.Tensor: mask + + Examples: + >>> subsequent_chunk_mask(4, 2) + [[1, 1, 0, 0], + [1, 1, 0, 0], + [1, 1, 1, 1], + [1, 1, 1, 1]] + """ + ret = torch.zeros(size, size, device=device, dtype=torch.bool) + for i in range(size): + if num_left_chunks < 0: + start = 0 + else: + start = max((i // chunk_size - num_left_chunks) * chunk_size, 0) + ending = min((i // chunk_size + 1) * chunk_size, size) + ret[i, start:ending] = True + return ret + +def add_optional_chunk_mask(xs: torch.Tensor, + masks: torch.Tensor, + use_dynamic_chunk: bool, + use_dynamic_left_chunk: bool, + decoding_chunk_size: int, + static_chunk_size: int, + num_decoding_left_chunks: int, + enable_full_context: bool = True): + """ Apply optional mask for encoder. + + Args: + xs (torch.Tensor): padded input, (B, L, D), L for max length + mask (torch.Tensor): mask for xs, (B, 1, L) + use_dynamic_chunk (bool): whether to use dynamic chunk or not + use_dynamic_left_chunk (bool): whether to use dynamic left chunk for + training. + decoding_chunk_size (int): decoding chunk size for dynamic chunk, it's + 0: default for training, use random dynamic chunk. + <0: for decoding, use full chunk. + >0: for decoding, use fixed chunk size as set. + static_chunk_size (int): chunk size for static chunk training/decoding + if it's greater than 0, if use_dynamic_chunk is true, + this parameter will be ignored + num_decoding_left_chunks: number of left chunks, this is for decoding, + the chunk size is decoding_chunk_size. + >=0: use num_decoding_left_chunks + <0: use all left chunks + enable_full_context (bool): + True: chunk size is either [1, 25] or full context(max_len) + False: chunk size ~ U[1, 25] + + Returns: + torch.Tensor: chunk mask of the input xs. + """ + # Whether to use chunk mask or not + if use_dynamic_chunk: + max_len = xs.size(1) + if decoding_chunk_size < 0: + chunk_size = max_len + num_left_chunks = -1 + elif decoding_chunk_size > 0: + chunk_size = decoding_chunk_size + num_left_chunks = num_decoding_left_chunks + else: + # chunk size is either [1, 25] or full context(max_len). + # Since we use 4 times subsampling and allow up to 1s(100 frames) + # delay, the maximum frame is 100 / 4 = 25. + chunk_size = torch.randint(1, max_len, (1, )).item() + num_left_chunks = -1 + if chunk_size > max_len // 2 and enable_full_context: + chunk_size = max_len + else: + chunk_size = chunk_size % 25 + 1 + if use_dynamic_left_chunk: + max_left_chunks = (max_len - 1) // chunk_size + num_left_chunks = torch.randint(0, max_left_chunks, + (1, )).item() + chunk_masks = subsequent_chunk_mask(xs.size(1), chunk_size, + num_left_chunks, + xs.device) # (L, L) + chunk_masks = chunk_masks.unsqueeze(0) # (1, L, L) + chunk_masks = masks & chunk_masks # (B, L, L) + elif static_chunk_size > 0: + num_left_chunks = num_decoding_left_chunks + chunk_masks = subsequent_chunk_mask(xs.size(1), static_chunk_size, + num_left_chunks, + xs.device) # (L, L) + chunk_masks = chunk_masks.unsqueeze(0) # (1, L, L) + chunk_masks = masks & chunk_masks # (B, L, L) + else: + chunk_masks = masks + return chunk_masks + + +class ConformerEncoderLayer(nn.Module): + """Encoder layer module. + Args: + size (int): Input dimension. + self_attn (torch.nn.Module): Self-attention module instance. + `MultiHeadedAttention` or `RelPositionMultiHeadedAttention` + instance can be used as the argument. + feed_forward (torch.nn.Module): Feed-forward module instance. + `PositionwiseFeedForward` instance can be used as the argument. + feed_forward_macaron (torch.nn.Module): Additional feed-forward module + instance. + `PositionwiseFeedForward` instance can be used as the argument. + conv_module (torch.nn.Module): Convolution module instance. + `ConvlutionModule` instance can be used as the argument. + dropout_rate (float): Dropout rate. + normalize_before (bool): + True: use layer_norm before each sub-block. + False: use layer_norm after each sub-block. + """ + + def __init__( + self, + size: int, + self_attn: torch.nn.Module, + feed_forward: Optional[nn.Module] = None, + feed_forward_macaron: Optional[nn.Module] = None, + conv_module: Optional[nn.Module] = None, + dropout_rate: float = 0.1, + normalize_before: bool = True, + dtype=None, device=None, operations=None + ): + """Construct an EncoderLayer object.""" + super().__init__() + self.self_attn = self_attn + self.feed_forward = feed_forward + self.feed_forward_macaron = feed_forward_macaron + self.conv_module = conv_module + self.norm_ff = operations.LayerNorm(size, eps=1e-5, dtype=dtype, device=device) # for the FNN module + self.norm_mha = operations.LayerNorm(size, eps=1e-5, dtype=dtype, device=device) # for the MHA module + if feed_forward_macaron is not None: + self.norm_ff_macaron = operations.LayerNorm(size, eps=1e-5, dtype=dtype, device=device) + self.ff_scale = 0.5 + else: + self.ff_scale = 1.0 + if self.conv_module is not None: + self.norm_conv = operations.LayerNorm(size, eps=1e-5, dtype=dtype, device=device) # for the CNN module + self.norm_final = operations.LayerNorm( + size, eps=1e-5, dtype=dtype, device=device) # for the final output of the block + self.dropout = nn.Dropout(dropout_rate) + self.size = size + self.normalize_before = normalize_before + + def forward( + self, + x: torch.Tensor, + mask: torch.Tensor, + pos_emb: torch.Tensor, + mask_pad: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool), + att_cache: torch.Tensor = torch.zeros((0, 0, 0, 0)), + cnn_cache: torch.Tensor = torch.zeros((0, 0, 0, 0)), + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + """Compute encoded features. + + Args: + x (torch.Tensor): (#batch, time, size) + mask (torch.Tensor): Mask tensor for the input (#batch, time，time), + (0, 0, 0) means fake mask. + pos_emb (torch.Tensor): positional encoding, must not be None + for ConformerEncoderLayer. + mask_pad (torch.Tensor): batch padding mask used for conv module. + (#batch, 1，time), (0, 0, 0) means fake mask. + att_cache (torch.Tensor): Cache tensor of the KEY & VALUE + (#batch=1, head, cache_t1, d_k * 2), head * d_k == size. + cnn_cache (torch.Tensor): Convolution cache in conformer layer + (#batch=1, size, cache_t2) + Returns: + torch.Tensor: Output tensor (#batch, time, size). + torch.Tensor: Mask tensor (#batch, time, time). + torch.Tensor: att_cache tensor, + (#batch=1, head, cache_t1 + time, d_k * 2). + torch.Tensor: cnn_cahce tensor (#batch, size, cache_t2). + """ + + # whether to use macaron style + if self.feed_forward_macaron is not None: + residual = x + if self.normalize_before: + x = self.norm_ff_macaron(x) + x = residual + self.ff_scale * self.dropout( + self.feed_forward_macaron(x)) + if not self.normalize_before: + x = self.norm_ff_macaron(x) + + # multi-headed self-attention module + residual = x + if self.normalize_before: + x = self.norm_mha(x) + x_att, new_att_cache = self.self_attn(x, x, x, mask, pos_emb, + att_cache) + x = residual + self.dropout(x_att) + if not self.normalize_before: + x = self.norm_mha(x) + + # convolution module + # Fake new cnn cache here, and then change it in conv_module + new_cnn_cache = torch.zeros((0, 0, 0), dtype=x.dtype, device=x.device) + if self.conv_module is not None: + residual = x + if self.normalize_before: + x = self.norm_conv(x) + x, new_cnn_cache = self.conv_module(x, mask_pad, cnn_cache) + x = residual + self.dropout(x) + + if not self.normalize_before: + x = self.norm_conv(x) + + # feed forward module + residual = x + if self.normalize_before: + x = self.norm_ff(x) + + x = residual + self.ff_scale * self.dropout(self.feed_forward(x)) + if not self.normalize_before: + x = self.norm_ff(x) + + if self.conv_module is not None: + x = self.norm_final(x) + + return x, mask, new_att_cache, new_cnn_cache + + + +class EspnetRelPositionalEncoding(torch.nn.Module): + """Relative positional encoding module (new implementation). + + Details can be found in https://github.com/espnet/espnet/pull/2816. + + See : Appendix B in https://arxiv.org/abs/1901.02860 + + Args: + d_model (int): Embedding dimension. + dropout_rate (float): Dropout rate. + max_len (int): Maximum input length. + + """ + + def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000): + """Construct an PositionalEncoding object.""" + super(EspnetRelPositionalEncoding, self).__init__() + self.d_model = d_model + self.xscale = math.sqrt(self.d_model) + self.dropout = torch.nn.Dropout(p=dropout_rate) + self.pe = None + self.extend_pe(torch.tensor(0.0).expand(1, max_len)) + + def extend_pe(self, x: torch.Tensor): + """Reset the positional encodings.""" + if self.pe is not None: + # self.pe contains both positive and negative parts + # the length of self.pe is 2 * input_len - 1 + if self.pe.size(1) >= x.size(1) * 2 - 1: + if self.pe.dtype != x.dtype or self.pe.device != x.device: + self.pe = self.pe.to(dtype=x.dtype, device=x.device) + return + # Suppose `i` means to the position of query vecotr and `j` means the + # position of key vector. We use position relative positions when keys + # are to the left (i>j) and negative relative positions otherwise (i Tuple[torch.Tensor, torch.Tensor]: + """Add positional encoding. + + Args: + x (torch.Tensor): Input tensor (batch, time, `*`). + + Returns: + torch.Tensor: Encoded tensor (batch, time, `*`). + + """ + self.extend_pe(x) + x = x * self.xscale + pos_emb = self.position_encoding(size=x.size(1), offset=offset) + return self.dropout(x), self.dropout(pos_emb) + + def position_encoding(self, + offset: Union[int, torch.Tensor], + size: int) -> torch.Tensor: + """ For getting encoding in a streaming fashion + + Attention!!!!! + we apply dropout only once at the whole utterance level in a none + streaming way, but will call this function several times with + increasing input size in a streaming scenario, so the dropout will + be applied several times. + + Args: + offset (int or torch.tensor): start offset + size (int): required size of position encoding + + Returns: + torch.Tensor: Corresponding encoding + """ + pos_emb = self.pe[ + :, + self.pe.size(1) // 2 - size + 1: self.pe.size(1) // 2 + size, + ] + return pos_emb + + + +class LinearEmbed(torch.nn.Module): + """Linear transform the input without subsampling + + Args: + idim (int): Input dimension. + odim (int): Output dimension. + dropout_rate (float): Dropout rate. + + """ + + def __init__(self, idim: int, odim: int, dropout_rate: float, + pos_enc_class: torch.nn.Module, dtype=None, device=None, operations=None): + """Construct an linear object.""" + super().__init__() + self.out = torch.nn.Sequential( + operations.Linear(idim, odim, dtype=dtype, device=device), + operations.LayerNorm(odim, eps=1e-5, dtype=dtype, device=device), + torch.nn.Dropout(dropout_rate), + ) + self.pos_enc = pos_enc_class #rel_pos_espnet + + def position_encoding(self, offset: Union[int, torch.Tensor], + size: int) -> torch.Tensor: + return self.pos_enc.position_encoding(offset, size) + + def forward( + self, + x: torch.Tensor, + offset: Union[int, torch.Tensor] = 0 + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """Input x. + + Args: + x (torch.Tensor): Input tensor (#batch, time, idim). + x_mask (torch.Tensor): Input mask (#batch, 1, time). + + Returns: + torch.Tensor: linear input tensor (#batch, time', odim), + where time' = time . + torch.Tensor: linear input mask (#batch, 1, time'), + where time' = time . + + """ + x = self.out(x) + x, pos_emb = self.pos_enc(x, offset) + return x, pos_emb + + +ATTENTION_CLASSES = { + "selfattn": MultiHeadedAttention, + "rel_selfattn": RelPositionMultiHeadedAttention, +} + +ACTIVATION_CLASSES = { + "hardtanh": torch.nn.Hardtanh, + "tanh": torch.nn.Tanh, + "relu": torch.nn.ReLU, + "selu": torch.nn.SELU, + "swish": getattr(torch.nn, "SiLU", Swish), + "gelu": torch.nn.GELU, +} + + +def make_pad_mask(lengths: torch.Tensor, max_len: int = 0) -> torch.Tensor: + """Make mask tensor containing indices of padded part. + + See description of make_non_pad_mask. + + Args: + lengths (torch.Tensor): Batch of lengths (B,). + Returns: + torch.Tensor: Mask tensor containing indices of padded part. + + Examples: + >>> lengths = [5, 3, 2] + >>> make_pad_mask(lengths) + masks = [[0, 0, 0, 0 ,0], + [0, 0, 0, 1, 1], + [0, 0, 1, 1, 1]] + """ + batch_size = lengths.size(0) + max_len = max_len if max_len > 0 else lengths.max().item() + seq_range = torch.arange(0, + max_len, + dtype=torch.int64, + device=lengths.device) + seq_range_expand = seq_range.unsqueeze(0).expand(batch_size, max_len) + seq_length_expand = lengths.unsqueeze(-1) + mask = seq_range_expand >= seq_length_expand + return mask + +#https://github.com/FunAudioLLM/CosyVoice/blob/main/examples/magicdata-read/cosyvoice/conf/cosyvoice.yaml +class ConformerEncoder(torch.nn.Module): + """Conformer encoder module.""" + + def __init__( + self, + input_size: int, + output_size: int = 1024, + attention_heads: int = 16, + linear_units: int = 4096, + num_blocks: int = 6, + dropout_rate: float = 0.1, + positional_dropout_rate: float = 0.1, + attention_dropout_rate: float = 0.0, + input_layer: str = 'linear', + pos_enc_layer_type: str = 'rel_pos_espnet', + normalize_before: bool = True, + static_chunk_size: int = 1, # 1: causal_mask; 0: full_mask + use_dynamic_chunk: bool = False, + use_dynamic_left_chunk: bool = False, + positionwise_conv_kernel_size: int = 1, + macaron_style: bool =False, + selfattention_layer_type: str = "rel_selfattn", + activation_type: str = "swish", + use_cnn_module: bool = False, + cnn_module_kernel: int = 15, + causal: bool = False, + cnn_module_norm: str = "batch_norm", + key_bias: bool = True, + dtype=None, device=None, operations=None + ): + """Construct ConformerEncoder + + Args: + input_size to use_dynamic_chunk, see in BaseEncoder + positionwise_conv_kernel_size (int): Kernel size of positionwise + conv1d layer. + macaron_style (bool): Whether to use macaron style for + positionwise layer. + selfattention_layer_type (str): Encoder attention layer type, + the parameter has no effect now, it's just for configure + compatibility. #'rel_selfattn' + activation_type (str): Encoder activation function type. + use_cnn_module (bool): Whether to use convolution module. + cnn_module_kernel (int): Kernel size of convolution module. + causal (bool): whether to use causal convolution or not. + key_bias: whether use bias in attention.linear_k, False for whisper models. + """ + super().__init__() + self.output_size = output_size + self.embed = LinearEmbed(input_size, output_size, dropout_rate, + EspnetRelPositionalEncoding(output_size, positional_dropout_rate), dtype=dtype, device=device, operations=operations) + self.normalize_before = normalize_before + self.after_norm = operations.LayerNorm(output_size, eps=1e-5, dtype=dtype, device=device) + self.use_dynamic_chunk = use_dynamic_chunk + + self.static_chunk_size = static_chunk_size + self.use_dynamic_chunk = use_dynamic_chunk + self.use_dynamic_left_chunk = use_dynamic_left_chunk + activation = ACTIVATION_CLASSES[activation_type]() + + # self-attention module definition + encoder_selfattn_layer_args = ( + attention_heads, + output_size, + attention_dropout_rate, + key_bias, + ) + # feed-forward module definition + positionwise_layer_args = ( + output_size, + linear_units, + dropout_rate, + activation, + ) + # convolution module definition + convolution_layer_args = (output_size, cnn_module_kernel, activation, + cnn_module_norm, causal) + + self.encoders = torch.nn.ModuleList([ + ConformerEncoderLayer( + output_size, + RelPositionMultiHeadedAttention( + *encoder_selfattn_layer_args, dtype=dtype, device=device, operations=operations), + PositionwiseFeedForward(*positionwise_layer_args, dtype=dtype, device=device, operations=operations), + PositionwiseFeedForward( + *positionwise_layer_args, dtype=dtype, device=device, operations=operations) if macaron_style else None, + ConvolutionModule( + *convolution_layer_args, dtype=dtype, device=device, operations=operations) if use_cnn_module else None, + dropout_rate, + normalize_before, dtype=dtype, device=device, operations=operations + ) for _ in range(num_blocks) + ]) + + def forward_layers(self, xs: torch.Tensor, chunk_masks: torch.Tensor, + pos_emb: torch.Tensor, + mask_pad: torch.Tensor) -> torch.Tensor: + for layer in self.encoders: + xs, chunk_masks, _, _ = layer(xs, chunk_masks, pos_emb, mask_pad) + return xs + + def forward( + self, + xs: torch.Tensor, + pad_mask: torch.Tensor, + decoding_chunk_size: int = 0, + num_decoding_left_chunks: int = -1, + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Embed positions in tensor. + + Args: + xs: padded input tensor (B, T, D) + xs_lens: input length (B) + decoding_chunk_size: decoding chunk size for dynamic chunk + 0: default for training, use random dynamic chunk. + <0: for decoding, use full chunk. + >0: for decoding, use fixed chunk size as set. + num_decoding_left_chunks: number of left chunks, this is for decoding, + the chunk size is decoding_chunk_size. + >=0: use num_decoding_left_chunks + <0: use all left chunks + Returns: + encoder output tensor xs, and subsampled masks + xs: padded output tensor (B, T' ~= T/subsample_rate, D) + masks: torch.Tensor batch padding mask after subsample + (B, 1, T' ~= T/subsample_rate) + NOTE(xcsong): + We pass the `__call__` method of the modules instead of `forward` to the + checkpointing API because `__call__` attaches all the hooks of the module. + https://discuss.pytorch.org/t/any-different-between-model-input-and-model-forward-input/3690/2 + """ + masks = None + if pad_mask is not None: + masks = pad_mask.to(torch.bool).unsqueeze(1) # (B, 1, T) + xs, pos_emb = self.embed(xs) + mask_pad = masks # (B, 1, T/subsample_rate) + chunk_masks = add_optional_chunk_mask(xs, masks, + self.use_dynamic_chunk, + self.use_dynamic_left_chunk, + decoding_chunk_size, + self.static_chunk_size, + num_decoding_left_chunks) + + xs = self.forward_layers(xs, chunk_masks, pos_emb, mask_pad) + if self.normalize_before: + xs = self.after_norm(xs) + # Here we assume the mask is not changed in encoder layers, so just + # return the masks before encoder layers, and the masks will be used + # for cross attention with decoder later + return xs, masks + diff --git a/comfy/ldm/ace/model.py b/comfy/ldm/ace/model.py new file mode 100644 index 000000000..12c524701 --- /dev/null +++ b/comfy/ldm/ace/model.py @@ -0,0 +1,385 @@ +# Original from: https://github.com/ace-step/ACE-Step/blob/main/models/ace_step_transformer.py + +# Copyright 2024 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Optional, List, Union + +import torch +from torch import nn + +import comfy.model_management + +from comfy.ldm.lightricks.model import TimestepEmbedding, Timesteps +from .attention import LinearTransformerBlock, t2i_modulate +from .lyric_encoder import ConformerEncoder as LyricEncoder + + +def cross_norm(hidden_states, controlnet_input): + # input N x T x c + mean_hidden_states, std_hidden_states = hidden_states.mean(dim=(1,2), keepdim=True), hidden_states.std(dim=(1,2), keepdim=True) + mean_controlnet_input, std_controlnet_input = controlnet_input.mean(dim=(1,2), keepdim=True), controlnet_input.std(dim=(1,2), keepdim=True) + controlnet_input = (controlnet_input - mean_controlnet_input) * (std_hidden_states / (std_controlnet_input + 1e-12)) + mean_hidden_states + return controlnet_input + + +# Copied from transformers.models.mixtral.modeling_mixtral.MixtralRotaryEmbedding with Mixtral->Qwen2 +class Qwen2RotaryEmbedding(nn.Module): + def __init__(self, dim, max_position_embeddings=2048, base=10000, dtype=None, device=None): + super().__init__() + + self.dim = dim + self.max_position_embeddings = max_position_embeddings + self.base = base + inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device=device).float() / self.dim)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + + # Build here to make `torch.jit.trace` work. + self._set_cos_sin_cache( + seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.float32 + ) + + def _set_cos_sin_cache(self, seq_len, device, dtype): + self.max_seq_len_cached = seq_len + t = torch.arange(self.max_seq_len_cached, device=device, dtype=torch.int64).type_as(self.inv_freq) + + freqs = torch.outer(t, self.inv_freq) + # Different from paper, but it uses a different permutation in order to obtain the same calculation + emb = torch.cat((freqs, freqs), dim=-1) + self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False) + self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False) + + def forward(self, x, seq_len=None): + # x: [bs, num_attention_heads, seq_len, head_size] + if seq_len > self.max_seq_len_cached: + self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype) + + return ( + self.cos_cached[:seq_len].to(dtype=x.dtype), + self.sin_cached[:seq_len].to(dtype=x.dtype), + ) + + +class T2IFinalLayer(nn.Module): + """ + The final layer of Sana. + """ + + def __init__(self, hidden_size, patch_size=[16, 1], out_channels=256, dtype=None, device=None, operations=None): + super().__init__() + self.norm_final = operations.RMSNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) + self.linear = operations.Linear(hidden_size, patch_size[0] * patch_size[1] * out_channels, bias=True, dtype=dtype, device=device) + self.scale_shift_table = nn.Parameter(torch.empty(2, hidden_size, dtype=dtype, device=device)) + self.out_channels = out_channels + self.patch_size = patch_size + + def unpatchfy( + self, + hidden_states: torch.Tensor, + width: int, + ): + # 4 unpatchify + new_height, new_width = 1, hidden_states.size(1) + hidden_states = hidden_states.reshape( + shape=(hidden_states.shape[0], new_height, new_width, self.patch_size[0], self.patch_size[1], self.out_channels) + ).contiguous() + hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states) + output = hidden_states.reshape( + shape=(hidden_states.shape[0], self.out_channels, new_height * self.patch_size[0], new_width * self.patch_size[1]) + ).contiguous() + if width > new_width: + output = torch.nn.functional.pad(output, (0, width - new_width, 0, 0), 'constant', 0) + elif width < new_width: + output = output[:, :, :, :width] + return output + + def forward(self, x, t, output_length): + shift, scale = (comfy.model_management.cast_to(self.scale_shift_table[None], device=t.device, dtype=t.dtype) + t[:, None]).chunk(2, dim=1) + x = t2i_modulate(self.norm_final(x), shift, scale) + x = self.linear(x) + # unpatchify + output = self.unpatchfy(x, output_length) + return output + + +class PatchEmbed(nn.Module): + """2D Image to Patch Embedding""" + + def __init__( + self, + height=16, + width=4096, + patch_size=(16, 1), + in_channels=8, + embed_dim=1152, + bias=True, + dtype=None, device=None, operations=None + ): + super().__init__() + patch_size_h, patch_size_w = patch_size + self.early_conv_layers = nn.Sequential( + operations.Conv2d(in_channels, in_channels*256, kernel_size=patch_size, stride=patch_size, padding=0, bias=bias, dtype=dtype, device=device), + operations.GroupNorm(num_groups=32, num_channels=in_channels*256, eps=1e-6, affine=True, dtype=dtype, device=device), + operations.Conv2d(in_channels*256, embed_dim, kernel_size=1, stride=1, padding=0, bias=bias, dtype=dtype, device=device) + ) + self.patch_size = patch_size + self.height, self.width = height // patch_size_h, width // patch_size_w + self.base_size = self.width + + def forward(self, latent): + # early convolutions, N x C x H x W -> N x 256 * sqrt(patch_size) x H/patch_size x W/patch_size + latent = self.early_conv_layers(latent) + latent = latent.flatten(2).transpose(1, 2) # BCHW -> BNC + return latent + + +class ACEStepTransformer2DModel(nn.Module): + # _supports_gradient_checkpointing = True + + def __init__( + self, + in_channels: Optional[int] = 8, + num_layers: int = 28, + inner_dim: int = 1536, + attention_head_dim: int = 64, + num_attention_heads: int = 24, + mlp_ratio: float = 4.0, + out_channels: int = 8, + max_position: int = 32768, + rope_theta: float = 1000000.0, + speaker_embedding_dim: int = 512, + text_embedding_dim: int = 768, + ssl_encoder_depths: List[int] = [9, 9], + ssl_names: List[str] = ["mert", "m-hubert"], + ssl_latent_dims: List[int] = [1024, 768], + lyric_encoder_vocab_size: int = 6681, + lyric_hidden_size: int = 1024, + patch_size: List[int] = [16, 1], + max_height: int = 16, + max_width: int = 4096, + audio_model=None, + dtype=None, device=None, operations=None + + ): + super().__init__() + + self.dtype = dtype + self.num_attention_heads = num_attention_heads + self.attention_head_dim = attention_head_dim + inner_dim = num_attention_heads * attention_head_dim + self.inner_dim = inner_dim + self.out_channels = out_channels + self.max_position = max_position + self.patch_size = patch_size + + self.rope_theta = rope_theta + + self.rotary_emb = Qwen2RotaryEmbedding( + dim=self.attention_head_dim, + max_position_embeddings=self.max_position, + base=self.rope_theta, + dtype=dtype, + device=device, + ) + + # 2. Define input layers + self.in_channels = in_channels + + self.num_layers = num_layers + # 3. Define transformers blocks + self.transformer_blocks = nn.ModuleList( + [ + LinearTransformerBlock( + dim=self.inner_dim, + num_attention_heads=self.num_attention_heads, + attention_head_dim=attention_head_dim, + mlp_ratio=mlp_ratio, + add_cross_attention=True, + add_cross_attention_dim=self.inner_dim, + dtype=dtype, + device=device, + operations=operations, + ) + for i in range(self.num_layers) + ] + ) + + 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=self.inner_dim, dtype=dtype, device=device, operations=operations) + self.t_block = nn.Sequential(nn.SiLU(), operations.Linear(self.inner_dim, 6 * self.inner_dim, bias=True, dtype=dtype, device=device)) + + # speaker + self.speaker_embedder = operations.Linear(speaker_embedding_dim, self.inner_dim, dtype=dtype, device=device) + + # genre + self.genre_embedder = operations.Linear(text_embedding_dim, self.inner_dim, dtype=dtype, device=device) + + # lyric + self.lyric_embs = operations.Embedding(lyric_encoder_vocab_size, lyric_hidden_size, dtype=dtype, device=device) + self.lyric_encoder = LyricEncoder(input_size=lyric_hidden_size, static_chunk_size=0, dtype=dtype, device=device, operations=operations) + self.lyric_proj = operations.Linear(lyric_hidden_size, self.inner_dim, dtype=dtype, device=device) + + projector_dim = 2 * self.inner_dim + + self.projectors = nn.ModuleList([ + nn.Sequential( + operations.Linear(self.inner_dim, projector_dim, dtype=dtype, device=device), + nn.SiLU(), + operations.Linear(projector_dim, projector_dim, dtype=dtype, device=device), + nn.SiLU(), + operations.Linear(projector_dim, ssl_dim, dtype=dtype, device=device), + ) for ssl_dim in ssl_latent_dims + ]) + + self.proj_in = PatchEmbed( + height=max_height, + width=max_width, + patch_size=patch_size, + embed_dim=self.inner_dim, + bias=True, + dtype=dtype, + device=device, + operations=operations, + ) + + self.final_layer = T2IFinalLayer(self.inner_dim, patch_size=patch_size, out_channels=out_channels, dtype=dtype, device=device, operations=operations) + + def forward_lyric_encoder( + self, + lyric_token_idx: Optional[torch.LongTensor] = None, + lyric_mask: Optional[torch.LongTensor] = None, + out_dtype=None, + ): + # N x T x D + lyric_embs = self.lyric_embs(lyric_token_idx, out_dtype=out_dtype) + prompt_prenet_out, _mask = self.lyric_encoder(lyric_embs, lyric_mask, decoding_chunk_size=1, num_decoding_left_chunks=-1) + prompt_prenet_out = self.lyric_proj(prompt_prenet_out) + return prompt_prenet_out + + def encode( + self, + encoder_text_hidden_states: Optional[torch.Tensor] = None, + text_attention_mask: Optional[torch.LongTensor] = None, + speaker_embeds: Optional[torch.FloatTensor] = None, + lyric_token_idx: Optional[torch.LongTensor] = None, + lyric_mask: Optional[torch.LongTensor] = None, + lyrics_strength=1.0, + ): + + bs = encoder_text_hidden_states.shape[0] + device = encoder_text_hidden_states.device + + # speaker embedding + encoder_spk_hidden_states = self.speaker_embedder(speaker_embeds).unsqueeze(1) + + # genre embedding + encoder_text_hidden_states = self.genre_embedder(encoder_text_hidden_states) + + # lyric + encoder_lyric_hidden_states = self.forward_lyric_encoder( + lyric_token_idx=lyric_token_idx, + lyric_mask=lyric_mask, + out_dtype=encoder_text_hidden_states.dtype, + ) + + encoder_lyric_hidden_states *= lyrics_strength + + encoder_hidden_states = torch.cat([encoder_spk_hidden_states, encoder_text_hidden_states, encoder_lyric_hidden_states], dim=1) + + encoder_hidden_mask = None + if text_attention_mask is not None: + speaker_mask = torch.ones(bs, 1, device=device) + encoder_hidden_mask = torch.cat([speaker_mask, text_attention_mask, lyric_mask], dim=1) + + return encoder_hidden_states, encoder_hidden_mask + + def decode( + self, + hidden_states: torch.Tensor, + attention_mask: torch.Tensor, + encoder_hidden_states: torch.Tensor, + encoder_hidden_mask: torch.Tensor, + timestep: Optional[torch.Tensor], + output_length: int = 0, + block_controlnet_hidden_states: Optional[Union[List[torch.Tensor], torch.Tensor]] = None, + controlnet_scale: Union[float, torch.Tensor] = 1.0, + ): + embedded_timestep = self.timestep_embedder(self.time_proj(timestep).to(dtype=hidden_states.dtype)) + temb = self.t_block(embedded_timestep) + + hidden_states = self.proj_in(hidden_states) + + # controlnet logic + if block_controlnet_hidden_states is not None: + control_condi = cross_norm(hidden_states, block_controlnet_hidden_states) + hidden_states = hidden_states + control_condi * controlnet_scale + + # inner_hidden_states = [] + + rotary_freqs_cis = self.rotary_emb(hidden_states, seq_len=hidden_states.shape[1]) + encoder_rotary_freqs_cis = self.rotary_emb(encoder_hidden_states, seq_len=encoder_hidden_states.shape[1]) + + for index_block, block in enumerate(self.transformer_blocks): + hidden_states = block( + hidden_states=hidden_states, + attention_mask=attention_mask, + encoder_hidden_states=encoder_hidden_states, + encoder_attention_mask=encoder_hidden_mask, + rotary_freqs_cis=rotary_freqs_cis, + rotary_freqs_cis_cross=encoder_rotary_freqs_cis, + temb=temb, + ) + + output = self.final_layer(hidden_states, embedded_timestep, output_length) + return output + + def forward( + self, + x, + timestep, + attention_mask=None, + context: Optional[torch.Tensor] = None, + text_attention_mask: Optional[torch.LongTensor] = None, + speaker_embeds: Optional[torch.FloatTensor] = None, + lyric_token_idx: Optional[torch.LongTensor] = None, + lyric_mask: Optional[torch.LongTensor] = None, + block_controlnet_hidden_states: Optional[Union[List[torch.Tensor], torch.Tensor]] = None, + controlnet_scale: Union[float, torch.Tensor] = 1.0, + lyrics_strength=1.0, + **kwargs + ): + hidden_states = x + encoder_text_hidden_states = context + encoder_hidden_states, encoder_hidden_mask = self.encode( + encoder_text_hidden_states=encoder_text_hidden_states, + text_attention_mask=text_attention_mask, + speaker_embeds=speaker_embeds, + lyric_token_idx=lyric_token_idx, + lyric_mask=lyric_mask, + lyrics_strength=lyrics_strength, + ) + + output_length = hidden_states.shape[-1] + + output = self.decode( + hidden_states=hidden_states, + attention_mask=attention_mask, + encoder_hidden_states=encoder_hidden_states, + encoder_hidden_mask=encoder_hidden_mask, + timestep=timestep, + output_length=output_length, + block_controlnet_hidden_states=block_controlnet_hidden_states, + controlnet_scale=controlnet_scale, + ) + + return output diff --git a/comfy/ldm/ace/vae/autoencoder_dc.py b/comfy/ldm/ace/vae/autoencoder_dc.py new file mode 100644 index 000000000..e7b1d4801 --- /dev/null +++ b/comfy/ldm/ace/vae/autoencoder_dc.py @@ -0,0 +1,644 @@ +# Rewritten from diffusers +import torch +import torch.nn as nn +import torch.nn.functional as F +from typing import Tuple, Union + +import comfy.model_management +import comfy.ops +ops = comfy.ops.disable_weight_init + + +class RMSNorm(ops.RMSNorm): + def __init__(self, dim, eps=1e-5, elementwise_affine=True, bias=False): + super().__init__(dim, eps=eps, elementwise_affine=elementwise_affine) + if elementwise_affine: + self.bias = nn.Parameter(torch.empty(dim)) if bias else None + + def forward(self, x): + x = super().forward(x) + if self.elementwise_affine: + if self.bias is not None: + x = x + comfy.model_management.cast_to(self.bias, dtype=x.dtype, device=x.device) + return x + + +def get_normalization(norm_type, num_features, num_groups=32, eps=1e-5): + if norm_type == "batch_norm": + return nn.BatchNorm2d(num_features) + elif norm_type == "group_norm": + return ops.GroupNorm(num_groups, num_features) + elif norm_type == "layer_norm": + return ops.LayerNorm(num_features) + elif norm_type == "rms_norm": + return RMSNorm(num_features, eps=eps, elementwise_affine=True, bias=True) + else: + raise ValueError(f"Unknown normalization type: {norm_type}") + + +def get_activation(activation_type): + if activation_type == "relu": + return nn.ReLU() + elif activation_type == "relu6": + return nn.ReLU6() + elif activation_type == "silu": + return nn.SiLU() + elif activation_type == "leaky_relu": + return nn.LeakyReLU(0.2) + else: + raise ValueError(f"Unknown activation type: {activation_type}") + + +class ResBlock(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + norm_type: str = "batch_norm", + act_fn: str = "relu6", + ) -> None: + super().__init__() + + self.norm_type = norm_type + self.nonlinearity = get_activation(act_fn) if act_fn is not None else nn.Identity() + self.conv1 = ops.Conv2d(in_channels, in_channels, 3, 1, 1) + self.conv2 = ops.Conv2d(in_channels, out_channels, 3, 1, 1, bias=False) + self.norm = get_normalization(norm_type, out_channels) + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + residual = hidden_states + hidden_states = self.conv1(hidden_states) + hidden_states = self.nonlinearity(hidden_states) + hidden_states = self.conv2(hidden_states) + + if self.norm_type == "rms_norm": + # move channel to the last dimension so we apply RMSnorm across channel dimension + hidden_states = self.norm(hidden_states.movedim(1, -1)).movedim(-1, 1) + else: + hidden_states = self.norm(hidden_states) + + return hidden_states + residual + +class SanaMultiscaleAttentionProjection(nn.Module): + def __init__( + self, + in_channels: int, + num_attention_heads: int, + kernel_size: int, + ) -> None: + super().__init__() + + channels = 3 * in_channels + self.proj_in = ops.Conv2d( + channels, + channels, + kernel_size, + padding=kernel_size // 2, + groups=channels, + bias=False, + ) + self.proj_out = ops.Conv2d(channels, channels, 1, 1, 0, groups=3 * num_attention_heads, bias=False) + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + hidden_states = self.proj_in(hidden_states) + hidden_states = self.proj_out(hidden_states) + return hidden_states + +class SanaMultiscaleLinearAttention(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + num_attention_heads: int = None, + attention_head_dim: int = 8, + mult: float = 1.0, + norm_type: str = "batch_norm", + kernel_sizes: tuple = (5,), + eps: float = 1e-15, + residual_connection: bool = False, + ): + super().__init__() + + self.eps = eps + self.attention_head_dim = attention_head_dim + self.norm_type = norm_type + self.residual_connection = residual_connection + + num_attention_heads = ( + int(in_channels // attention_head_dim * mult) + if num_attention_heads is None + else num_attention_heads + ) + inner_dim = num_attention_heads * attention_head_dim + + self.to_q = ops.Linear(in_channels, inner_dim, bias=False) + self.to_k = ops.Linear(in_channels, inner_dim, bias=False) + self.to_v = ops.Linear(in_channels, inner_dim, bias=False) + + self.to_qkv_multiscale = nn.ModuleList() + for kernel_size in kernel_sizes: + self.to_qkv_multiscale.append( + SanaMultiscaleAttentionProjection(inner_dim, num_attention_heads, kernel_size) + ) + + self.nonlinearity = nn.ReLU() + self.to_out = ops.Linear(inner_dim * (1 + len(kernel_sizes)), out_channels, bias=False) + self.norm_out = get_normalization(norm_type, out_channels) + + def apply_linear_attention(self, query, key, value): + value = F.pad(value, (0, 0, 0, 1), mode="constant", value=1) + scores = torch.matmul(value, key.transpose(-1, -2)) + hidden_states = torch.matmul(scores, query) + + hidden_states = hidden_states.to(dtype=torch.float32) + hidden_states = hidden_states[:, :, :-1] / (hidden_states[:, :, -1:] + self.eps) + return hidden_states + + def apply_quadratic_attention(self, query, key, value): + scores = torch.matmul(key.transpose(-1, -2), query) + scores = scores.to(dtype=torch.float32) + scores = scores / (torch.sum(scores, dim=2, keepdim=True) + self.eps) + hidden_states = torch.matmul(value, scores.to(value.dtype)) + return hidden_states + + def forward(self, hidden_states): + height, width = hidden_states.shape[-2:] + if height * width > self.attention_head_dim: + use_linear_attention = True + else: + use_linear_attention = False + + residual = hidden_states + + batch_size, _, height, width = list(hidden_states.size()) + original_dtype = hidden_states.dtype + + hidden_states = hidden_states.movedim(1, -1) + query = self.to_q(hidden_states) + key = self.to_k(hidden_states) + value = self.to_v(hidden_states) + hidden_states = torch.cat([query, key, value], dim=3) + hidden_states = hidden_states.movedim(-1, 1) + + multi_scale_qkv = [hidden_states] + for block in self.to_qkv_multiscale: + multi_scale_qkv.append(block(hidden_states)) + + hidden_states = torch.cat(multi_scale_qkv, dim=1) + + if use_linear_attention: + # for linear attention upcast hidden_states to float32 + hidden_states = hidden_states.to(dtype=torch.float32) + + hidden_states = hidden_states.reshape(batch_size, -1, 3 * self.attention_head_dim, height * width) + + query, key, value = hidden_states.chunk(3, dim=2) + query = self.nonlinearity(query) + key = self.nonlinearity(key) + + if use_linear_attention: + hidden_states = self.apply_linear_attention(query, key, value) + hidden_states = hidden_states.to(dtype=original_dtype) + else: + hidden_states = self.apply_quadratic_attention(query, key, value) + + hidden_states = torch.reshape(hidden_states, (batch_size, -1, height, width)) + hidden_states = self.to_out(hidden_states.movedim(1, -1)).movedim(-1, 1) + + if self.norm_type == "rms_norm": + hidden_states = self.norm_out(hidden_states.movedim(1, -1)).movedim(-1, 1) + else: + hidden_states = self.norm_out(hidden_states) + + if self.residual_connection: + hidden_states = hidden_states + residual + + return hidden_states + + +class EfficientViTBlock(nn.Module): + def __init__( + self, + in_channels: int, + mult: float = 1.0, + attention_head_dim: int = 32, + qkv_multiscales: tuple = (5,), + norm_type: str = "batch_norm", + ) -> None: + super().__init__() + + self.attn = SanaMultiscaleLinearAttention( + in_channels=in_channels, + out_channels=in_channels, + mult=mult, + attention_head_dim=attention_head_dim, + norm_type=norm_type, + kernel_sizes=qkv_multiscales, + residual_connection=True, + ) + + self.conv_out = GLUMBConv( + in_channels=in_channels, + out_channels=in_channels, + norm_type="rms_norm", + ) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + x = self.attn(x) + x = self.conv_out(x) + return x + + +class GLUMBConv(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + expand_ratio: float = 4, + norm_type: str = None, + residual_connection: bool = True, + ) -> None: + super().__init__() + + hidden_channels = int(expand_ratio * in_channels) + self.norm_type = norm_type + self.residual_connection = residual_connection + + self.nonlinearity = nn.SiLU() + self.conv_inverted = ops.Conv2d(in_channels, hidden_channels * 2, 1, 1, 0) + self.conv_depth = ops.Conv2d(hidden_channels * 2, hidden_channels * 2, 3, 1, 1, groups=hidden_channels * 2) + self.conv_point = ops.Conv2d(hidden_channels, out_channels, 1, 1, 0, bias=False) + + self.norm = None + if norm_type == "rms_norm": + self.norm = RMSNorm(out_channels, eps=1e-5, elementwise_affine=True, bias=True) + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + if self.residual_connection: + residual = hidden_states + + hidden_states = self.conv_inverted(hidden_states) + hidden_states = self.nonlinearity(hidden_states) + + hidden_states = self.conv_depth(hidden_states) + hidden_states, gate = torch.chunk(hidden_states, 2, dim=1) + hidden_states = hidden_states * self.nonlinearity(gate) + + hidden_states = self.conv_point(hidden_states) + + if self.norm_type == "rms_norm": + # move channel to the last dimension so we apply RMSnorm across channel dimension + hidden_states = self.norm(hidden_states.movedim(1, -1)).movedim(-1, 1) + + if self.residual_connection: + hidden_states = hidden_states + residual + + return hidden_states + + +def get_block( + block_type: str, + in_channels: int, + out_channels: int, + attention_head_dim: int, + norm_type: str, + act_fn: str, + qkv_mutliscales: tuple = (), +): + if block_type == "ResBlock": + block = ResBlock(in_channels, out_channels, norm_type, act_fn) + elif block_type == "EfficientViTBlock": + block = EfficientViTBlock( + in_channels, + attention_head_dim=attention_head_dim, + norm_type=norm_type, + qkv_multiscales=qkv_mutliscales + ) + else: + raise ValueError(f"Block with {block_type=} is not supported.") + + return block + + +class DCDownBlock2d(nn.Module): + def __init__(self, in_channels: int, out_channels: int, downsample: bool = False, shortcut: bool = True) -> None: + super().__init__() + + self.downsample = downsample + self.factor = 2 + self.stride = 1 if downsample else 2 + self.group_size = in_channels * self.factor**2 // out_channels + self.shortcut = shortcut + + out_ratio = self.factor**2 + if downsample: + assert out_channels % out_ratio == 0 + out_channels = out_channels // out_ratio + + self.conv = ops.Conv2d( + in_channels, + out_channels, + kernel_size=3, + stride=self.stride, + padding=1, + ) + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + x = self.conv(hidden_states) + if self.downsample: + x = F.pixel_unshuffle(x, self.factor) + + if self.shortcut: + y = F.pixel_unshuffle(hidden_states, self.factor) + y = y.unflatten(1, (-1, self.group_size)) + y = y.mean(dim=2) + hidden_states = x + y + else: + hidden_states = x + + return hidden_states + + +class DCUpBlock2d(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + interpolate: bool = False, + shortcut: bool = True, + interpolation_mode: str = "nearest", + ) -> None: + super().__init__() + + self.interpolate = interpolate + self.interpolation_mode = interpolation_mode + self.shortcut = shortcut + self.factor = 2 + self.repeats = out_channels * self.factor**2 // in_channels + + out_ratio = self.factor**2 + if not interpolate: + out_channels = out_channels * out_ratio + + self.conv = ops.Conv2d(in_channels, out_channels, 3, 1, 1) + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + if self.interpolate: + x = F.interpolate(hidden_states, scale_factor=self.factor, mode=self.interpolation_mode) + x = self.conv(x) + else: + x = self.conv(hidden_states) + x = F.pixel_shuffle(x, self.factor) + + if self.shortcut: + y = hidden_states.repeat_interleave(self.repeats, dim=1, output_size=hidden_states.shape[1] * self.repeats) + y = F.pixel_shuffle(y, self.factor) + hidden_states = x + y + else: + hidden_states = x + + return hidden_states + + +class Encoder(nn.Module): + def __init__( + self, + in_channels: int, + latent_channels: int, + attention_head_dim: int = 32, + block_type: str or tuple = "ResBlock", + block_out_channels: tuple = (128, 256, 512, 512, 1024, 1024), + layers_per_block: tuple = (2, 2, 2, 2, 2, 2), + qkv_multiscales: tuple = ((), (), (), (5,), (5,), (5,)), + downsample_block_type: str = "pixel_unshuffle", + out_shortcut: bool = True, + ): + super().__init__() + + num_blocks = len(block_out_channels) + + if isinstance(block_type, str): + block_type = (block_type,) * num_blocks + + if layers_per_block[0] > 0: + self.conv_in = ops.Conv2d( + in_channels, + block_out_channels[0] if layers_per_block[0] > 0 else block_out_channels[1], + kernel_size=3, + stride=1, + padding=1, + ) + else: + self.conv_in = DCDownBlock2d( + in_channels=in_channels, + out_channels=block_out_channels[0] if layers_per_block[0] > 0 else block_out_channels[1], + downsample=downsample_block_type == "pixel_unshuffle", + shortcut=False, + ) + + down_blocks = [] + for i, (out_channel, num_layers) in enumerate(zip(block_out_channels, layers_per_block)): + down_block_list = [] + + for _ in range(num_layers): + block = get_block( + block_type[i], + out_channel, + out_channel, + attention_head_dim=attention_head_dim, + norm_type="rms_norm", + act_fn="silu", + qkv_mutliscales=qkv_multiscales[i], + ) + down_block_list.append(block) + + if i < num_blocks - 1 and num_layers > 0: + downsample_block = DCDownBlock2d( + in_channels=out_channel, + out_channels=block_out_channels[i + 1], + downsample=downsample_block_type == "pixel_unshuffle", + shortcut=True, + ) + down_block_list.append(downsample_block) + + down_blocks.append(nn.Sequential(*down_block_list)) + + self.down_blocks = nn.ModuleList(down_blocks) + + self.conv_out = ops.Conv2d(block_out_channels[-1], latent_channels, 3, 1, 1) + + self.out_shortcut = out_shortcut + if out_shortcut: + self.out_shortcut_average_group_size = block_out_channels[-1] // latent_channels + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + hidden_states = self.conv_in(hidden_states) + for down_block in self.down_blocks: + hidden_states = down_block(hidden_states) + + if self.out_shortcut: + x = hidden_states.unflatten(1, (-1, self.out_shortcut_average_group_size)) + x = x.mean(dim=2) + hidden_states = self.conv_out(hidden_states) + x + else: + hidden_states = self.conv_out(hidden_states) + + return hidden_states + + +class Decoder(nn.Module): + def __init__( + self, + in_channels: int, + latent_channels: int, + attention_head_dim: int = 32, + block_type: str or tuple = "ResBlock", + block_out_channels: tuple = (128, 256, 512, 512, 1024, 1024), + layers_per_block: tuple = (2, 2, 2, 2, 2, 2), + qkv_multiscales: tuple = ((), (), (), (5,), (5,), (5,)), + norm_type: str or tuple = "rms_norm", + act_fn: str or tuple = "silu", + upsample_block_type: str = "pixel_shuffle", + in_shortcut: bool = True, + ): + super().__init__() + + num_blocks = len(block_out_channels) + + if isinstance(block_type, str): + block_type = (block_type,) * num_blocks + if isinstance(norm_type, str): + norm_type = (norm_type,) * num_blocks + if isinstance(act_fn, str): + act_fn = (act_fn,) * num_blocks + + self.conv_in = ops.Conv2d(latent_channels, block_out_channels[-1], 3, 1, 1) + + self.in_shortcut = in_shortcut + if in_shortcut: + self.in_shortcut_repeats = block_out_channels[-1] // latent_channels + + up_blocks = [] + for i, (out_channel, num_layers) in reversed(list(enumerate(zip(block_out_channels, layers_per_block)))): + up_block_list = [] + + if i < num_blocks - 1 and num_layers > 0: + upsample_block = DCUpBlock2d( + block_out_channels[i + 1], + out_channel, + interpolate=upsample_block_type == "interpolate", + shortcut=True, + ) + up_block_list.append(upsample_block) + + for _ in range(num_layers): + block = get_block( + block_type[i], + out_channel, + out_channel, + attention_head_dim=attention_head_dim, + norm_type=norm_type[i], + act_fn=act_fn[i], + qkv_mutliscales=qkv_multiscales[i], + ) + up_block_list.append(block) + + up_blocks.insert(0, nn.Sequential(*up_block_list)) + + self.up_blocks = nn.ModuleList(up_blocks) + + channels = block_out_channels[0] if layers_per_block[0] > 0 else block_out_channels[1] + + self.norm_out = RMSNorm(channels, 1e-5, elementwise_affine=True, bias=True) + self.conv_act = nn.ReLU() + self.conv_out = None + + if layers_per_block[0] > 0: + self.conv_out = ops.Conv2d(channels, in_channels, 3, 1, 1) + else: + self.conv_out = DCUpBlock2d( + channels, in_channels, interpolate=upsample_block_type == "interpolate", shortcut=False + ) + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + if self.in_shortcut: + x = hidden_states.repeat_interleave( + self.in_shortcut_repeats, dim=1, output_size=hidden_states.shape[1] * self.in_shortcut_repeats + ) + hidden_states = self.conv_in(hidden_states) + x + else: + hidden_states = self.conv_in(hidden_states) + + for up_block in reversed(self.up_blocks): + hidden_states = up_block(hidden_states) + + hidden_states = self.norm_out(hidden_states.movedim(1, -1)).movedim(-1, 1) + hidden_states = self.conv_act(hidden_states) + hidden_states = self.conv_out(hidden_states) + return hidden_states + + +class AutoencoderDC(nn.Module): + def __init__( + self, + in_channels: int = 2, + latent_channels: int = 8, + attention_head_dim: int = 32, + encoder_block_types: Union[str, Tuple[str]] = ["ResBlock", "ResBlock", "ResBlock", "EfficientViTBlock"], + decoder_block_types: Union[str, Tuple[str]] = ["ResBlock", "ResBlock", "ResBlock", "EfficientViTBlock"], + encoder_block_out_channels: Tuple[int, ...] = (128, 256, 512, 1024), + decoder_block_out_channels: Tuple[int, ...] = (128, 256, 512, 1024), + encoder_layers_per_block: Tuple[int] = (2, 2, 3, 3), + decoder_layers_per_block: Tuple[int] = (3, 3, 3, 3), + encoder_qkv_multiscales: Tuple[Tuple[int, ...], ...] = ((), (), (5,), (5,)), + decoder_qkv_multiscales: Tuple[Tuple[int, ...], ...] = ((), (), (5,), (5,)), + upsample_block_type: str = "interpolate", + downsample_block_type: str = "Conv", + decoder_norm_types: Union[str, Tuple[str]] = "rms_norm", + decoder_act_fns: Union[str, Tuple[str]] = "silu", + scaling_factor: float = 0.41407, + ) -> None: + super().__init__() + + self.encoder = Encoder( + in_channels=in_channels, + latent_channels=latent_channels, + attention_head_dim=attention_head_dim, + block_type=encoder_block_types, + block_out_channels=encoder_block_out_channels, + layers_per_block=encoder_layers_per_block, + qkv_multiscales=encoder_qkv_multiscales, + downsample_block_type=downsample_block_type, + ) + + self.decoder = Decoder( + in_channels=in_channels, + latent_channels=latent_channels, + attention_head_dim=attention_head_dim, + block_type=decoder_block_types, + block_out_channels=decoder_block_out_channels, + layers_per_block=decoder_layers_per_block, + qkv_multiscales=decoder_qkv_multiscales, + norm_type=decoder_norm_types, + act_fn=decoder_act_fns, + upsample_block_type=upsample_block_type, + ) + + self.scaling_factor = scaling_factor + self.spatial_compression_ratio = 2 ** (len(encoder_block_out_channels) - 1) + + def encode(self, x: torch.Tensor) -> torch.Tensor: + """Internal encoding function.""" + encoded = self.encoder(x) + return encoded * self.scaling_factor + + def decode(self, z: torch.Tensor) -> torch.Tensor: + # Scale the latents back + z = z / self.scaling_factor + decoded = self.decoder(z) + return decoded + + def forward(self, x: torch.Tensor) -> torch.Tensor: + z = self.encode(x) + return self.decode(z) + diff --git a/comfy/ldm/ace/vae/music_dcae_pipeline.py b/comfy/ldm/ace/vae/music_dcae_pipeline.py new file mode 100644 index 000000000..af81280eb --- /dev/null +++ b/comfy/ldm/ace/vae/music_dcae_pipeline.py @@ -0,0 +1,109 @@ +# Original from: https://github.com/ace-step/ACE-Step/blob/main/music_dcae/music_dcae_pipeline.py +import torch +from .autoencoder_dc import AutoencoderDC +import logging +try: + import torchaudio +except: + logging.warning("torchaudio missing, ACE model will be broken") + +import torchvision.transforms as transforms +from .music_vocoder import ADaMoSHiFiGANV1 + + +class MusicDCAE(torch.nn.Module): + def __init__(self, source_sample_rate=None, dcae_config={}, vocoder_config={}): + super(MusicDCAE, self).__init__() + + self.dcae = AutoencoderDC(**dcae_config) + self.vocoder = ADaMoSHiFiGANV1(**vocoder_config) + + if source_sample_rate is None: + self.source_sample_rate = 48000 + else: + self.source_sample_rate = source_sample_rate + + # self.resampler = torchaudio.transforms.Resample(source_sample_rate, 44100) + + self.transform = transforms.Compose([ + transforms.Normalize(0.5, 0.5), + ]) + self.min_mel_value = -11.0 + self.max_mel_value = 3.0 + self.audio_chunk_size = int(round((1024 * 512 / 44100 * 48000))) + self.mel_chunk_size = 1024 + self.time_dimention_multiple = 8 + self.latent_chunk_size = self.mel_chunk_size // self.time_dimention_multiple + self.scale_factor = 0.1786 + self.shift_factor = -1.9091 + + def load_audio(self, audio_path): + audio, sr = torchaudio.load(audio_path) + return audio, sr + + def forward_mel(self, audios): + mels = [] + for i in range(len(audios)): + image = self.vocoder.mel_transform(audios[i]) + mels.append(image) + mels = torch.stack(mels) + return mels + + @torch.no_grad() + def encode(self, audios, audio_lengths=None, sr=None): + if audio_lengths is None: + audio_lengths = torch.tensor([audios.shape[2]] * audios.shape[0]) + audio_lengths = audio_lengths.to(audios.device) + + if sr is None: + sr = self.source_sample_rate + + if sr != 44100: + audios = torchaudio.functional.resample(audios, sr, 44100) + + max_audio_len = audios.shape[-1] + if max_audio_len % (8 * 512) != 0: + audios = torch.nn.functional.pad(audios, (0, 8 * 512 - max_audio_len % (8 * 512))) + + mels = self.forward_mel(audios) + mels = (mels - self.min_mel_value) / (self.max_mel_value - self.min_mel_value) + mels = self.transform(mels) + latents = [] + for mel in mels: + latent = self.dcae.encoder(mel.unsqueeze(0)) + latents.append(latent) + latents = torch.cat(latents, dim=0) + # latent_lengths = (audio_lengths / sr * 44100 / 512 / self.time_dimention_multiple).long() + latents = (latents - self.shift_factor) * self.scale_factor + return latents + # return latents, latent_lengths + + @torch.no_grad() + def decode(self, latents, audio_lengths=None, sr=None): + latents = latents / self.scale_factor + self.shift_factor + + pred_wavs = [] + + for latent in latents: + mels = self.dcae.decoder(latent.unsqueeze(0)) + mels = mels * 0.5 + 0.5 + mels = mels * (self.max_mel_value - self.min_mel_value) + self.min_mel_value + wav = self.vocoder.decode(mels[0]).squeeze(1) + + if sr is not None: + # resampler = torchaudio.transforms.Resample(44100, sr).to(latents.device).to(latents.dtype) + wav = torchaudio.functional.resample(wav, 44100, sr) + # wav = resampler(wav) + else: + sr = 44100 + pred_wavs.append(wav) + + if audio_lengths is not None: + pred_wavs = [wav[:, :length].cpu() for wav, length in zip(pred_wavs, audio_lengths)] + return torch.stack(pred_wavs) + # return sr, pred_wavs + + def forward(self, audios, audio_lengths=None, sr=None): + latents, latent_lengths = self.encode(audios=audios, audio_lengths=audio_lengths, sr=sr) + sr, pred_wavs = self.decode(latents=latents, audio_lengths=audio_lengths, sr=sr) + return sr, pred_wavs, latents, latent_lengths diff --git a/comfy/ldm/ace/vae/music_log_mel.py b/comfy/ldm/ace/vae/music_log_mel.py new file mode 100755 index 000000000..9c584eb7f --- /dev/null +++ b/comfy/ldm/ace/vae/music_log_mel.py @@ -0,0 +1,113 @@ +# Original from: https://github.com/ace-step/ACE-Step/blob/main/music_dcae/music_log_mel.py +import torch +import torch.nn as nn +from torch import Tensor +import logging +try: + from torchaudio.transforms import MelScale +except: + logging.warning("torchaudio missing, ACE model will be broken") + +import comfy.model_management + +class LinearSpectrogram(nn.Module): + def __init__( + self, + n_fft=2048, + win_length=2048, + hop_length=512, + center=False, + mode="pow2_sqrt", + ): + super().__init__() + + self.n_fft = n_fft + self.win_length = win_length + self.hop_length = hop_length + self.center = center + self.mode = mode + + self.register_buffer("window", torch.hann_window(win_length)) + + def forward(self, y: Tensor) -> Tensor: + if y.ndim == 3: + y = y.squeeze(1) + + y = torch.nn.functional.pad( + y.unsqueeze(1), + ( + (self.win_length - self.hop_length) // 2, + (self.win_length - self.hop_length + 1) // 2, + ), + mode="reflect", + ).squeeze(1) + dtype = y.dtype + spec = torch.stft( + y.float(), + self.n_fft, + hop_length=self.hop_length, + win_length=self.win_length, + window=comfy.model_management.cast_to(self.window, dtype=torch.float32, device=y.device), + center=self.center, + pad_mode="reflect", + normalized=False, + onesided=True, + return_complex=True, + ) + spec = torch.view_as_real(spec) + + if self.mode == "pow2_sqrt": + spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6) + spec = spec.to(dtype) + return spec + + +class LogMelSpectrogram(nn.Module): + def __init__( + self, + sample_rate=44100, + n_fft=2048, + win_length=2048, + hop_length=512, + n_mels=128, + center=False, + f_min=0.0, + f_max=None, + ): + super().__init__() + + self.sample_rate = sample_rate + self.n_fft = n_fft + self.win_length = win_length + self.hop_length = hop_length + self.center = center + self.n_mels = n_mels + self.f_min = f_min + self.f_max = f_max or sample_rate // 2 + + self.spectrogram = LinearSpectrogram(n_fft, win_length, hop_length, center) + self.mel_scale = MelScale( + self.n_mels, + self.sample_rate, + self.f_min, + self.f_max, + self.n_fft // 2 + 1, + "slaney", + "slaney", + ) + + def compress(self, x: Tensor) -> Tensor: + return torch.log(torch.clamp(x, min=1e-5)) + + def decompress(self, x: Tensor) -> Tensor: + return torch.exp(x) + + def forward(self, x: Tensor, return_linear: bool = False) -> Tensor: + linear = self.spectrogram(x) + x = self.mel_scale(linear) + x = self.compress(x) + # print(x.shape) + if return_linear: + return x, self.compress(linear) + + return x diff --git a/comfy/ldm/ace/vae/music_vocoder.py b/comfy/ldm/ace/vae/music_vocoder.py new file mode 100755 index 000000000..2f989fa86 --- /dev/null +++ b/comfy/ldm/ace/vae/music_vocoder.py @@ -0,0 +1,538 @@ +# Original from: https://github.com/ace-step/ACE-Step/blob/main/music_dcae/music_vocoder.py +import torch +from torch import nn + +from functools import partial +from math import prod +from typing import Callable, Tuple, List + +import numpy as np +import torch.nn.functional as F +from torch.nn.utils.parametrize import remove_parametrizations as remove_weight_norm + +from .music_log_mel import LogMelSpectrogram + +import comfy.model_management +import comfy.ops +ops = comfy.ops.disable_weight_init + + +def drop_path( + x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True +): + """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks). + + This is the same as the DropConnect impl I created for EfficientNet, etc networks, however, + the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper... + See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for + changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use + 'survival rate' as the argument. + + """ # noqa: E501 + + if drop_prob == 0.0 or not training: + return x + keep_prob = 1 - drop_prob + shape = (x.shape[0],) + (1,) * ( + x.ndim - 1 + ) # work with diff dim tensors, not just 2D ConvNets + random_tensor = x.new_empty(shape).bernoulli_(keep_prob) + if keep_prob > 0.0 and scale_by_keep: + random_tensor.div_(keep_prob) + return x * random_tensor + + +class DropPath(nn.Module): + """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).""" # noqa: E501 + + def __init__(self, drop_prob: float = 0.0, scale_by_keep: bool = True): + super(DropPath, self).__init__() + self.drop_prob = drop_prob + self.scale_by_keep = scale_by_keep + + def forward(self, x): + return drop_path(x, self.drop_prob, self.training, self.scale_by_keep) + + def extra_repr(self): + return f"drop_prob={round(self.drop_prob,3):0.3f}" + + +class LayerNorm(nn.Module): + r"""LayerNorm that supports two data formats: channels_last (default) or channels_first. + The ordering of the dimensions in the inputs. channels_last corresponds to inputs with + shape (batch_size, height, width, channels) while channels_first corresponds to inputs + with shape (batch_size, channels, height, width). + """ # noqa: E501 + + def __init__(self, normalized_shape, eps=1e-6, data_format="channels_last"): + super().__init__() + self.weight = nn.Parameter(torch.ones(normalized_shape)) + self.bias = nn.Parameter(torch.zeros(normalized_shape)) + self.eps = eps + self.data_format = data_format + if self.data_format not in ["channels_last", "channels_first"]: + raise NotImplementedError + self.normalized_shape = (normalized_shape,) + + def forward(self, x): + if self.data_format == "channels_last": + return F.layer_norm( + x, self.normalized_shape, comfy.model_management.cast_to(self.weight, dtype=x.dtype, device=x.device), comfy.model_management.cast_to(self.bias, dtype=x.dtype, device=x.device), self.eps + ) + elif self.data_format == "channels_first": + u = x.mean(1, keepdim=True) + s = (x - u).pow(2).mean(1, keepdim=True) + x = (x - u) / torch.sqrt(s + self.eps) + x = comfy.model_management.cast_to(self.weight[:, None], dtype=x.dtype, device=x.device) * x + comfy.model_management.cast_to(self.bias[:, None], dtype=x.dtype, device=x.device) + return x + + +class ConvNeXtBlock(nn.Module): + r"""ConvNeXt Block. There are two equivalent implementations: + (1) DwConv -> LayerNorm (channels_first) -> 1x1 Conv -> GELU -> 1x1 Conv; all in (N, C, H, W) + (2) DwConv -> Permute to (N, H, W, C); LayerNorm (channels_last) -> Linear -> GELU -> Linear; Permute back + We use (2) as we find it slightly faster in PyTorch + + Args: + dim (int): Number of input channels. + drop_path (float): Stochastic depth rate. Default: 0.0 + layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6. + mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.0. + kernel_size (int): Kernel size for depthwise conv. Default: 7. + dilation (int): Dilation for depthwise conv. Default: 1. + """ # noqa: E501 + + def __init__( + self, + dim: int, + drop_path: float = 0.0, + layer_scale_init_value: float = 1e-6, + mlp_ratio: float = 4.0, + kernel_size: int = 7, + dilation: int = 1, + ): + super().__init__() + + self.dwconv = ops.Conv1d( + dim, + dim, + kernel_size=kernel_size, + padding=int(dilation * (kernel_size - 1) / 2), + groups=dim, + ) # depthwise conv + self.norm = LayerNorm(dim, eps=1e-6) + self.pwconv1 = ops.Linear( + dim, int(mlp_ratio * dim) + ) # pointwise/1x1 convs, implemented with linear layers + self.act = nn.GELU() + self.pwconv2 = ops.Linear(int(mlp_ratio * dim), dim) + self.gamma = ( + nn.Parameter(torch.empty((dim)), requires_grad=False) + if layer_scale_init_value > 0 + else None + ) + self.drop_path = DropPath( + drop_path) if drop_path > 0.0 else nn.Identity() + + def forward(self, x, apply_residual: bool = True): + input = x + + x = self.dwconv(x) + x = x.permute(0, 2, 1) # (N, C, L) -> (N, L, C) + x = self.norm(x) + x = self.pwconv1(x) + x = self.act(x) + x = self.pwconv2(x) + + if self.gamma is not None: + x = comfy.model_management.cast_to(self.gamma, dtype=x.dtype, device=x.device) * x + + x = x.permute(0, 2, 1) # (N, L, C) -> (N, C, L) + x = self.drop_path(x) + + if apply_residual: + x = input + x + + return x + + +class ParallelConvNeXtBlock(nn.Module): + def __init__(self, kernel_sizes: List[int], *args, **kwargs): + super().__init__() + self.blocks = nn.ModuleList( + [ + ConvNeXtBlock(kernel_size=kernel_size, *args, **kwargs) + for kernel_size in kernel_sizes + ] + ) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + return torch.stack( + [block(x, apply_residual=False) for block in self.blocks] + [x], + dim=1, + ).sum(dim=1) + + +class ConvNeXtEncoder(nn.Module): + def __init__( + self, + input_channels=3, + depths=[3, 3, 9, 3], + dims=[96, 192, 384, 768], + drop_path_rate=0.0, + layer_scale_init_value=1e-6, + kernel_sizes: Tuple[int] = (7,), + ): + super().__init__() + assert len(depths) == len(dims) + + self.channel_layers = nn.ModuleList() + stem = nn.Sequential( + ops.Conv1d( + input_channels, + dims[0], + kernel_size=7, + padding=3, + padding_mode="replicate", + ), + LayerNorm(dims[0], eps=1e-6, data_format="channels_first"), + ) + self.channel_layers.append(stem) + + for i in range(len(depths) - 1): + mid_layer = nn.Sequential( + LayerNorm(dims[i], eps=1e-6, data_format="channels_first"), + ops.Conv1d(dims[i], dims[i + 1], kernel_size=1), + ) + self.channel_layers.append(mid_layer) + + block_fn = ( + partial(ConvNeXtBlock, kernel_size=kernel_sizes[0]) + if len(kernel_sizes) == 1 + else partial(ParallelConvNeXtBlock, kernel_sizes=kernel_sizes) + ) + + self.stages = nn.ModuleList() + drop_path_rates = [ + x.item() for x in torch.linspace(0, drop_path_rate, sum(depths)) + ] + + cur = 0 + for i in range(len(depths)): + stage = nn.Sequential( + *[ + block_fn( + dim=dims[i], + drop_path=drop_path_rates[cur + j], + layer_scale_init_value=layer_scale_init_value, + ) + for j in range(depths[i]) + ] + ) + self.stages.append(stage) + cur += depths[i] + + self.norm = LayerNorm(dims[-1], eps=1e-6, data_format="channels_first") + + def forward( + self, + x: torch.Tensor, + ) -> torch.Tensor: + for channel_layer, stage in zip(self.channel_layers, self.stages): + x = channel_layer(x) + x = stage(x) + + return self.norm(x) + + +def get_padding(kernel_size, dilation=1): + return (kernel_size * dilation - dilation) // 2 + + +class ResBlock1(torch.nn.Module): + def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)): + super().__init__() + + self.convs1 = nn.ModuleList( + [ + torch.nn.utils.parametrizations.weight_norm( + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=dilation[0], + padding=get_padding(kernel_size, dilation[0]), + ) + ), + torch.nn.utils.parametrizations.weight_norm( + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=dilation[1], + padding=get_padding(kernel_size, dilation[1]), + ) + ), + torch.nn.utils.parametrizations.weight_norm( + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=dilation[2], + padding=get_padding(kernel_size, dilation[2]), + ) + ), + ] + ) + + self.convs2 = nn.ModuleList( + [ + torch.nn.utils.parametrizations.weight_norm( + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=1, + padding=get_padding(kernel_size, 1), + ) + ), + torch.nn.utils.parametrizations.weight_norm( + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=1, + padding=get_padding(kernel_size, 1), + ) + ), + torch.nn.utils.parametrizations.weight_norm( + 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.silu(x) + xt = c1(xt) + xt = F.silu(xt) + xt = c2(xt) + x = xt + x + return x + + def remove_weight_norm(self): + for conv in self.convs1: + remove_weight_norm(conv) + for conv in self.convs2: + remove_weight_norm(conv) + + +class HiFiGANGenerator(nn.Module): + def __init__( + self, + *, + hop_length: int = 512, + upsample_rates: Tuple[int] = (8, 8, 2, 2, 2), + upsample_kernel_sizes: Tuple[int] = (16, 16, 8, 2, 2), + resblock_kernel_sizes: Tuple[int] = (3, 7, 11), + resblock_dilation_sizes: Tuple[Tuple[int]] = ( + (1, 3, 5), (1, 3, 5), (1, 3, 5)), + num_mels: int = 128, + upsample_initial_channel: int = 512, + use_template: bool = True, + pre_conv_kernel_size: int = 7, + post_conv_kernel_size: int = 7, + post_activation: Callable = partial(nn.SiLU, inplace=True), + ): + super().__init__() + + assert ( + prod(upsample_rates) == hop_length + ), f"hop_length must be {prod(upsample_rates)}" + + self.conv_pre = torch.nn.utils.parametrizations.weight_norm( + ops.Conv1d( + num_mels, + upsample_initial_channel, + pre_conv_kernel_size, + 1, + padding=get_padding(pre_conv_kernel_size), + ) + ) + + self.num_upsamples = len(upsample_rates) + self.num_kernels = len(resblock_kernel_sizes) + + self.noise_convs = nn.ModuleList() + self.use_template = use_template + self.ups = nn.ModuleList() + + for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)): + c_cur = upsample_initial_channel // (2 ** (i + 1)) + self.ups.append( + torch.nn.utils.parametrizations.weight_norm( + ops.ConvTranspose1d( + upsample_initial_channel // (2**i), + upsample_initial_channel // (2 ** (i + 1)), + k, + u, + padding=(k - u) // 2, + ) + ) + ) + + if not use_template: + continue + + if i + 1 < len(upsample_rates): + stride_f0 = np.prod(upsample_rates[i + 1:]) + self.noise_convs.append( + ops.Conv1d( + 1, + c_cur, + kernel_size=stride_f0 * 2, + stride=stride_f0, + padding=stride_f0 // 2, + ) + ) + else: + self.noise_convs.append(ops.Conv1d(1, c_cur, kernel_size=1)) + + self.resblocks = nn.ModuleList() + for i in range(len(self.ups)): + ch = upsample_initial_channel // (2 ** (i + 1)) + for k, d in zip(resblock_kernel_sizes, resblock_dilation_sizes): + self.resblocks.append(ResBlock1(ch, k, d)) + + self.activation_post = post_activation() + self.conv_post = torch.nn.utils.parametrizations.weight_norm( + ops.Conv1d( + ch, + 1, + post_conv_kernel_size, + 1, + padding=get_padding(post_conv_kernel_size), + ) + ) + + def forward(self, x, template=None): + x = self.conv_pre(x) + + for i in range(self.num_upsamples): + x = F.silu(x, inplace=True) + x = self.ups[i](x) + + if self.use_template: + x = x + self.noise_convs[i](template) + + 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 = self.activation_post(x) + x = self.conv_post(x) + x = torch.tanh(x) + + return x + + def remove_weight_norm(self): + for up in self.ups: + remove_weight_norm(up) + for block in self.resblocks: + block.remove_weight_norm() + remove_weight_norm(self.conv_pre) + remove_weight_norm(self.conv_post) + + +class ADaMoSHiFiGANV1(nn.Module): + def __init__( + self, + input_channels: int = 128, + depths: List[int] = [3, 3, 9, 3], + dims: List[int] = [128, 256, 384, 512], + drop_path_rate: float = 0.0, + kernel_sizes: Tuple[int] = (7,), + upsample_rates: Tuple[int] = (4, 4, 2, 2, 2, 2, 2), + upsample_kernel_sizes: Tuple[int] = (8, 8, 4, 4, 4, 4, 4), + resblock_kernel_sizes: Tuple[int] = (3, 7, 11, 13), + resblock_dilation_sizes: Tuple[Tuple[int]] = ( + (1, 3, 5), (1, 3, 5), (1, 3, 5), (1, 3, 5)), + num_mels: int = 512, + upsample_initial_channel: int = 1024, + use_template: bool = False, + pre_conv_kernel_size: int = 13, + post_conv_kernel_size: int = 13, + sampling_rate: int = 44100, + n_fft: int = 2048, + win_length: int = 2048, + hop_length: int = 512, + f_min: int = 40, + f_max: int = 16000, + n_mels: int = 128, + ): + super().__init__() + + self.backbone = ConvNeXtEncoder( + input_channels=input_channels, + depths=depths, + dims=dims, + drop_path_rate=drop_path_rate, + kernel_sizes=kernel_sizes, + ) + + self.head = HiFiGANGenerator( + hop_length=hop_length, + upsample_rates=upsample_rates, + upsample_kernel_sizes=upsample_kernel_sizes, + resblock_kernel_sizes=resblock_kernel_sizes, + resblock_dilation_sizes=resblock_dilation_sizes, + num_mels=num_mels, + upsample_initial_channel=upsample_initial_channel, + use_template=use_template, + pre_conv_kernel_size=pre_conv_kernel_size, + post_conv_kernel_size=post_conv_kernel_size, + ) + self.sampling_rate = sampling_rate + self.mel_transform = LogMelSpectrogram( + sample_rate=sampling_rate, + n_fft=n_fft, + win_length=win_length, + hop_length=hop_length, + f_min=f_min, + f_max=f_max, + n_mels=n_mels, + ) + self.eval() + + @torch.no_grad() + def decode(self, mel): + y = self.backbone(mel) + y = self.head(y) + return y + + @torch.no_grad() + def encode(self, x): + return self.mel_transform(x) + + def forward(self, mel): + y = self.backbone(mel) + y = self.head(y) + return y diff --git a/comfy/ldm/audio/autoencoder.py b/comfy/ldm/audio/autoencoder.py index 255db6308..3a554909a 100644 --- a/comfy/ldm/audio/autoencoder.py +++ b/comfy/ldm/audio/autoencoder.py @@ -75,16 +75,10 @@ class SnakeBeta(nn.Module): return x def WNConv1d(*args, **kwargs): - try: - return torch.nn.utils.parametrizations.weight_norm(ops.Conv1d(*args, **kwargs)) - except: - return torch.nn.utils.weight_norm(ops.Conv1d(*args, **kwargs)) #support pytorch 2.1 and older + return torch.nn.utils.parametrizations.weight_norm(ops.Conv1d(*args, **kwargs)) def WNConvTranspose1d(*args, **kwargs): - try: - return torch.nn.utils.parametrizations.weight_norm(ops.ConvTranspose1d(*args, **kwargs)) - except: - return torch.nn.utils.weight_norm(ops.ConvTranspose1d(*args, **kwargs)) #support pytorch 2.1 and older + return torch.nn.utils.parametrizations.weight_norm(ops.ConvTranspose1d(*args, **kwargs)) def get_activation(activation: Literal["elu", "snake", "none"], antialias=False, channels=None) -> nn.Module: if activation == "elu": diff --git a/comfy/ldm/chroma/layers.py b/comfy/ldm/chroma/layers.py new file mode 100644 index 000000000..2a0dec606 --- /dev/null +++ b/comfy/ldm/chroma/layers.py @@ -0,0 +1,181 @@ +import torch +from torch import Tensor, nn + +from comfy.ldm.flux.math import attention +from comfy.ldm.flux.layers import ( + MLPEmbedder, + RMSNorm, + QKNorm, + SelfAttention, + ModulationOut, +) + + + +class ChromaModulationOut(ModulationOut): + @classmethod + def from_offset(cls, tensor: torch.Tensor, offset: int = 0) -> ModulationOut: + return cls( + shift=tensor[:, offset : offset + 1, :], + scale=tensor[:, offset + 1 : offset + 2, :], + gate=tensor[:, offset + 2 : offset + 3, :], + ) + + + + +class Approximator(nn.Module): + def __init__(self, in_dim: int, out_dim: int, hidden_dim: int, n_layers = 5, dtype=None, device=None, operations=None): + super().__init__() + self.in_proj = operations.Linear(in_dim, hidden_dim, bias=True, dtype=dtype, device=device) + self.layers = nn.ModuleList([MLPEmbedder(hidden_dim, hidden_dim, dtype=dtype, device=device, operations=operations) for x in range( n_layers)]) + self.norms = nn.ModuleList([RMSNorm(hidden_dim, dtype=dtype, device=device, operations=operations) for x in range( n_layers)]) + self.out_proj = operations.Linear(hidden_dim, out_dim, dtype=dtype, device=device) + + @property + def device(self): + # Get the device of the module (assumes all parameters are on the same device) + return next(self.parameters()).device + + def forward(self, x: Tensor) -> Tensor: + x = self.in_proj(x) + + for layer, norms in zip(self.layers, self.norms): + x = x + layer(norms(x)) + + x = self.out_proj(x) + + return x + + +class DoubleStreamBlock(nn.Module): + def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, dtype=None, device=None, operations=None): + super().__init__() + + mlp_hidden_dim = int(hidden_size * mlp_ratio) + self.num_heads = num_heads + self.hidden_size = hidden_size + 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_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.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_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.flipped_img_txt = flipped_img_txt + + def forward(self, img: Tensor, txt: Tensor, pe: Tensor, vec: Tensor, attn_mask=None): + (img_mod1, img_mod2), (txt_mod1, txt_mod2) = vec + + # prepare image for attention + img_modulated = torch.addcmul(img_mod1.shift, 1 + img_mod1.scale, self.img_norm1(img)) + img_qkv = self.img_attn.qkv(img_modulated) + img_q, img_k, img_v = img_qkv.view(img_qkv.shape[0], img_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) + img_q, img_k = self.img_attn.norm(img_q, img_k, img_v) + + # prepare txt for attention + txt_modulated = torch.addcmul(txt_mod1.shift, 1 + txt_mod1.scale, self.txt_norm1(txt)) + txt_qkv = self.txt_attn.qkv(txt_modulated) + txt_q, txt_k, txt_v = txt_qkv.view(txt_qkv.shape[0], txt_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) + txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v) + + # run actual attention + attn = attention(torch.cat((txt_q, img_q), dim=2), + torch.cat((txt_k, img_k), dim=2), + torch.cat((txt_v, img_v), dim=2), + pe=pe, mask=attn_mask) + + txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1] :] + + # calculate the img bloks + img.addcmul_(img_mod1.gate, self.img_attn.proj(img_attn)) + img.addcmul_(img_mod2.gate, self.img_mlp(torch.addcmul(img_mod2.shift, 1 + img_mod2.scale, self.img_norm2(img)))) + + # calculate the txt bloks + txt.addcmul_(txt_mod1.gate, self.txt_attn.proj(txt_attn)) + txt.addcmul_(txt_mod2.gate, self.txt_mlp(torch.addcmul(txt_mod2.shift, 1 + txt_mod2.scale, self.txt_norm2(txt)))) + + if txt.dtype == torch.float16: + txt = torch.nan_to_num(txt, nan=0.0, posinf=65504, neginf=-65504) + + return img, txt + + +class SingleStreamBlock(nn.Module): + """ + A DiT block with parallel linear layers as described in + https://arxiv.org/abs/2302.05442 and adapted modulation interface. + """ + + def __init__( + self, + hidden_size: int, + num_heads: int, + mlp_ratio: float = 4.0, + qk_scale: float = None, + dtype=None, + device=None, + operations=None + ): + super().__init__() + self.hidden_dim = hidden_size + self.num_heads = num_heads + head_dim = hidden_size // num_heads + self.scale = qk_scale or head_dim**-0.5 + + self.mlp_hidden_dim = int(hidden_size * mlp_ratio) + # qkv and mlp_in + self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim, 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.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") + + def forward(self, x: Tensor, pe: Tensor, vec: Tensor, attn_mask=None) -> Tensor: + mod = vec + x_mod = torch.addcmul(mod.shift, 1 + mod.scale, self.pre_norm(x)) + qkv, mlp = torch.split(self.linear1(x_mod), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1) + + q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) + q, k = self.norm(q, k, v) + + # compute attention + attn = attention(q, k, v, pe=pe, mask=attn_mask) + # compute activation in mlp stream, cat again and run second linear layer + output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2)) + x.addcmul_(mod.gate, output) + if x.dtype == torch.float16: + x = torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504) + return x + + +class LastLayer(nn.Module): + def __init__(self, hidden_size: int, patch_size: int, out_channels: int, 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, out_channels, bias=True, dtype=dtype, device=device) + + def forward(self, x: Tensor, vec: Tensor) -> Tensor: + shift, scale = vec + shift = shift.squeeze(1) + scale = scale.squeeze(1) + x = torch.addcmul(shift[:, None, :], 1 + scale[:, None, :], self.norm_final(x)) + x = self.linear(x) + return x diff --git a/comfy/ldm/chroma/model.py b/comfy/ldm/chroma/model.py new file mode 100644 index 000000000..c75023a31 --- /dev/null +++ b/comfy/ldm/chroma/model.py @@ -0,0 +1,271 @@ +#Original code can be found on: https://github.com/black-forest-labs/flux + +from dataclasses import dataclass + +import torch +from torch import Tensor, nn +from einops import rearrange, repeat +import comfy.ldm.common_dit + +from comfy.ldm.flux.layers import ( + EmbedND, + timestep_embedding, +) + +from .layers import ( + DoubleStreamBlock, + LastLayer, + SingleStreamBlock, + Approximator, + ChromaModulationOut, +) + + +@dataclass +class ChromaParams: + in_channels: int + out_channels: int + context_in_dim: int + hidden_size: int + mlp_ratio: float + num_heads: int + depth: int + depth_single_blocks: int + axes_dim: list + theta: int + patch_size: int + qkv_bias: bool + in_dim: int + out_dim: int + hidden_dim: int + n_layers: int + + + + +class Chroma(nn.Module): + """ + Transformer model for flow matching on sequences. + """ + + def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, operations=None, **kwargs): + super().__init__() + self.dtype = dtype + params = ChromaParams(**kwargs) + self.params = params + self.patch_size = params.patch_size + self.in_channels = params.in_channels + self.out_channels = params.out_channels + 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}" + ) + pe_dim = params.hidden_size // params.num_heads + if sum(params.axes_dim) != pe_dim: + raise ValueError(f"Got {params.axes_dim} but expected positional dim {pe_dim}") + self.hidden_size = params.hidden_size + self.num_heads = params.num_heads + self.in_dim = params.in_dim + self.out_dim = params.out_dim + self.hidden_dim = params.hidden_dim + self.n_layers = params.n_layers + 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.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, dtype=dtype, device=device) + # set as nn identity for now, will overwrite it later. + self.distilled_guidance_layer = Approximator( + in_dim=self.in_dim, + hidden_dim=self.hidden_dim, + out_dim=self.out_dim, + n_layers=self.n_layers, + dtype=dtype, device=device, operations=operations + ) + + + self.double_blocks = nn.ModuleList( + [ + DoubleStreamBlock( + self.hidden_size, + self.num_heads, + mlp_ratio=params.mlp_ratio, + qkv_bias=params.qkv_bias, + dtype=dtype, device=device, operations=operations + ) + for _ in range(params.depth) + ] + ) + + self.single_blocks = nn.ModuleList( + [ + SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, 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.skip_mmdit = [] + self.skip_dit = [] + self.lite = False + + def get_modulations(self, tensor: torch.Tensor, block_type: str, *, idx: int = 0): + # This function slices up the modulations tensor which has the following layout: + # single : num_single_blocks * 3 elements + # double_img : num_double_blocks * 6 elements + # double_txt : num_double_blocks * 6 elements + # final : 2 elements + if block_type == "final": + return (tensor[:, -2:-1, :], tensor[:, -1:, :]) + single_block_count = self.params.depth_single_blocks + double_block_count = self.params.depth + offset = 3 * idx + if block_type == "single": + return ChromaModulationOut.from_offset(tensor, offset) + # Double block modulations are 6 elements so we double 3 * idx. + offset *= 2 + if block_type in {"double_img", "double_txt"}: + # Advance past the single block modulations. + offset += 3 * single_block_count + if block_type == "double_txt": + # Advance past the double block img modulations. + offset += 6 * double_block_count + return ( + ChromaModulationOut.from_offset(tensor, offset), + ChromaModulationOut.from_offset(tensor, offset + 3), + ) + raise ValueError("Bad block_type") + + + def forward_orig( + self, + img: Tensor, + img_ids: Tensor, + txt: Tensor, + txt_ids: Tensor, + timesteps: Tensor, + guidance: Tensor = None, + control = None, + transformer_options={}, + attn_mask: Tensor = None, + ) -> Tensor: + patches_replace = transformer_options.get("patches_replace", {}) + if img.ndim != 3 or txt.ndim != 3: + raise ValueError("Input img and txt tensors must have 3 dimensions.") + + # running on sequences img + img = self.img_in(img) + + # distilled vector guidance + mod_index_length = 344 + distill_timestep = timestep_embedding(timesteps.detach().clone(), 16).to(img.device, img.dtype) + # guidance = guidance * + distil_guidance = timestep_embedding(guidance.detach().clone(), 16).to(img.device, img.dtype) + + # get all modulation index + modulation_index = timestep_embedding(torch.arange(mod_index_length, device=img.device), 32).to(img.device, img.dtype) + # we need to broadcast the modulation index here so each batch has all of the index + modulation_index = modulation_index.unsqueeze(0).repeat(img.shape[0], 1, 1).to(img.device, img.dtype) + # and we need to broadcast timestep and guidance along too + timestep_guidance = torch.cat([distill_timestep, distil_guidance], dim=1).unsqueeze(1).repeat(1, mod_index_length, 1).to(img.dtype).to(img.device, img.dtype) + # then and only then we could concatenate it together + input_vec = torch.cat([timestep_guidance, modulation_index], dim=-1).to(img.device, img.dtype) + + mod_vectors = self.distilled_guidance_layer(input_vec) + + txt = self.txt_in(txt) + + ids = torch.cat((txt_ids, img_ids), dim=1) + pe = self.pe_embedder(ids) + + blocks_replace = patches_replace.get("dit", {}) + for i, block in enumerate(self.double_blocks): + if i not in self.skip_mmdit: + double_mod = ( + self.get_modulations(mod_vectors, "double_img", idx=i), + self.get_modulations(mod_vectors, "double_txt", idx=i), + ) + if ("double_block", i) in blocks_replace: + def block_wrap(args): + out = {} + out["img"], out["txt"] = block(img=args["img"], + txt=args["txt"], + vec=args["vec"], + pe=args["pe"], + attn_mask=args.get("attn_mask")) + return out + + out = blocks_replace[("double_block", i)]({"img": img, + "txt": txt, + "vec": double_mod, + "pe": pe, + "attn_mask": attn_mask}, + {"original_block": block_wrap}) + txt = out["txt"] + img = out["img"] + else: + img, txt = block(img=img, + txt=txt, + vec=double_mod, + pe=pe, + attn_mask=attn_mask) + + if control is not None: # Controlnet + control_i = control.get("input") + if i < len(control_i): + add = control_i[i] + if add is not None: + img += add + + img = torch.cat((txt, img), 1) + + for i, block in enumerate(self.single_blocks): + if i not in self.skip_dit: + single_mod = self.get_modulations(mod_vectors, "single", idx=i) + if ("single_block", i) in blocks_replace: + def block_wrap(args): + out = {} + out["img"] = block(args["img"], + vec=args["vec"], + pe=args["pe"], + attn_mask=args.get("attn_mask")) + return out + + out = blocks_replace[("single_block", i)]({"img": img, + "vec": single_mod, + "pe": pe, + "attn_mask": attn_mask}, + {"original_block": block_wrap}) + img = out["img"] + else: + img = block(img, vec=single_mod, pe=pe, attn_mask=attn_mask) + + if control is not None: # Controlnet + control_o = control.get("output") + if i < len(control_o): + add = control_o[i] + if add is not None: + img[:, txt.shape[1] :, ...] += add + + img = img[:, txt.shape[1] :, ...] + final_mod = self.get_modulations(mod_vectors, "final") + img = self.final_layer(img, vec=final_mod) # (N, T, patch_size ** 2 * out_channels) + return img + + def forward(self, x, timestep, context, guidance, control=None, transformer_options={}, **kwargs): + bs, c, h, w = x.shape + patch_size = 2 + x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size)) + + img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size) + + h_len = ((h + (patch_size // 2)) // patch_size) + w_len = ((w + (patch_size // 2)) // patch_size) + img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype) + img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1) + img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0) + img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs) + + txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype) + out = self.forward_orig(img, img_ids, context, txt_ids, timestep, guidance, control, transformer_options, attn_mask=kwargs.get("attention_mask", None)) + 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,:w] diff --git a/comfy/ldm/cosmos/blocks.py b/comfy/ldm/cosmos/blocks.py index 84fd6d839..a12f892d2 100644 --- a/comfy/ldm/cosmos/blocks.py +++ b/comfy/ldm/cosmos/blocks.py @@ -23,7 +23,6 @@ from einops import rearrange, repeat from einops.layers.torch import Rearrange from torch import nn -from comfy.ldm.modules.diffusionmodules.mmdit import RMSNorm from comfy.ldm.modules.attention import optimized_attention @@ -37,11 +36,11 @@ def apply_rotary_pos_emb( return t_out -def get_normalization(name: str, channels: int, weight_args={}): +def get_normalization(name: str, channels: int, weight_args={}, operations=None): if name == "I": return nn.Identity() elif name == "R": - return RMSNorm(channels, elementwise_affine=True, eps=1e-6, **weight_args) + return operations.RMSNorm(channels, elementwise_affine=True, eps=1e-6, **weight_args) else: raise ValueError(f"Normalization {name} not found") @@ -120,15 +119,15 @@ class Attention(nn.Module): self.to_q = nn.Sequential( operations.Linear(query_dim, inner_dim, bias=qkv_bias, **weight_args), - get_normalization(qkv_norm[0], norm_dim), + get_normalization(qkv_norm[0], norm_dim, weight_args=weight_args, operations=operations), ) self.to_k = nn.Sequential( operations.Linear(context_dim, inner_dim, bias=qkv_bias, **weight_args), - get_normalization(qkv_norm[1], norm_dim), + get_normalization(qkv_norm[1], norm_dim, weight_args=weight_args, operations=operations), ) self.to_v = nn.Sequential( operations.Linear(context_dim, inner_dim, bias=qkv_bias, **weight_args), - get_normalization(qkv_norm[2], norm_dim), + get_normalization(qkv_norm[2], norm_dim, weight_args=weight_args, operations=operations), ) self.to_out = nn.Sequential( diff --git a/comfy/ldm/cosmos/model.py b/comfy/ldm/cosmos/model.py index 7cf0dc2b6..4836e0b69 100644 --- a/comfy/ldm/cosmos/model.py +++ b/comfy/ldm/cosmos/model.py @@ -27,8 +27,6 @@ from torchvision import transforms from enum import Enum import logging -from ..modules.diffusionmodules.mmdit import RMSNorm - from .blocks import ( FinalLayer, GeneralDITTransformerBlock, @@ -195,7 +193,7 @@ class GeneralDIT(nn.Module): if self.affline_emb_norm: logging.debug("Building affine embedding normalization layer") - self.affline_norm = RMSNorm(model_channels, elementwise_affine=True, eps=1e-6) + self.affline_norm = operations.RMSNorm(model_channels, elementwise_affine=True, eps=1e-6, device=device, dtype=dtype) else: self.affline_norm = nn.Identity() diff --git a/comfy/ldm/flux/controlnet.py b/comfy/ldm/flux/controlnet.py index 588533339..2395398e5 100644 --- a/comfy/ldm/flux/controlnet.py +++ b/comfy/ldm/flux/controlnet.py @@ -124,6 +124,9 @@ class ControlNetFlux(Flux): if img.ndim != 3 or txt.ndim != 3: raise ValueError("Input img and txt tensors must have 3 dimensions.") + if y is None: + y = torch.zeros((img.shape[0], self.params.vec_in_dim), device=img.device, dtype=img.dtype) + # running on sequences img img = self.img_in(img) @@ -177,7 +180,7 @@ class ControlNetFlux(Flux): out["output"] = out_output[:self.main_model_single] return out - def forward(self, x, timesteps, context, y, guidance=None, hint=None, **kwargs): + def forward(self, x, timesteps, context, y=None, guidance=None, hint=None, **kwargs): patch_size = 2 if self.latent_input: hint = common_dit.pad_to_patch_size(hint, (patch_size, patch_size)) diff --git a/comfy/ldm/flux/model.py b/comfy/ldm/flux/model.py index 5e320fcd7..9c28022b5 100644 --- a/comfy/ldm/flux/model.py +++ b/comfy/ldm/flux/model.py @@ -103,6 +103,10 @@ class Flux(nn.Module): transformer_options={}, 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_replace = transformer_options.get("patches_replace", {}) if img.ndim != 3 or txt.ndim != 3: raise ValueError("Input img and txt tensors must have 3 dimensions.") @@ -190,7 +194,7 @@ class Flux(nn.Module): img = self.final_layer(img, vec) # (N, T, patch_size ** 2 * out_channels) return img - def forward(self, x, timestep, context, y, guidance=None, control=None, transformer_options={}, **kwargs): + def forward(self, x, timestep, context, y=None, guidance=None, control=None, transformer_options={}, **kwargs): bs, c, h, w = x.shape patch_size = self.patch_size x = common_dit.pad_to_patch_size(x, (patch_size, patch_size)) diff --git a/comfy/ldm/genmo/joint_model/asymm_models_joint.py b/comfy/ldm/genmo/joint_model/asymm_models_joint.py index 32de0ee85..9c3346c46 100644 --- a/comfy/ldm/genmo/joint_model/asymm_models_joint.py +++ b/comfy/ldm/genmo/joint_model/asymm_models_joint.py @@ -10,7 +10,6 @@ from einops import rearrange from .layers import ( FeedForward, PatchEmbed, - RMSNorm, TimestepEmbedder, ) from .rope_mixed import ( @@ -88,10 +87,10 @@ class AsymmetricAttention(nn.Module): # Query and key normalization for stability. assert qk_norm - self.q_norm_x = RMSNorm(self.head_dim, device=device, dtype=dtype) - self.k_norm_x = RMSNorm(self.head_dim, device=device, dtype=dtype) - self.q_norm_y = RMSNorm(self.head_dim, device=device, dtype=dtype) - self.k_norm_y = RMSNorm(self.head_dim, device=device, dtype=dtype) + self.q_norm_x = operations.RMSNorm(self.head_dim, eps=1e-5, device=device, dtype=dtype) + self.k_norm_x = operations.RMSNorm(self.head_dim, eps=1e-5, device=device, dtype=dtype) + self.q_norm_y = operations.RMSNorm(self.head_dim, eps=1e-5, device=device, dtype=dtype) + self.k_norm_y = operations.RMSNorm(self.head_dim, eps=1e-5, device=device, dtype=dtype) # Output layers. y features go back down from dim_x -> dim_y. self.proj_x = operations.Linear(dim_x, dim_x, bias=out_bias, device=device, dtype=dtype) diff --git a/comfy/ldm/genmo/joint_model/layers.py b/comfy/ldm/genmo/joint_model/layers.py index 3232a0df7..caedc6bad 100644 --- a/comfy/ldm/genmo/joint_model/layers.py +++ b/comfy/ldm/genmo/joint_model/layers.py @@ -152,14 +152,3 @@ class PatchEmbed(nn.Module): x = self.norm(x) return x - - -class RMSNorm(torch.nn.Module): - def __init__(self, hidden_size, eps=1e-5, device=None, dtype=None): - super().__init__() - self.eps = eps - self.weight = torch.nn.Parameter(torch.empty(hidden_size, device=device, dtype=dtype)) - self.register_parameter("bias", None) - - def forward(self, x): - return rms_norm(x, self.weight, self.eps) diff --git a/comfy/ldm/hidream/model.py b/comfy/ldm/hidream/model.py index fcb5a9c51..0305747bf 100644 --- a/comfy/ldm/hidream/model.py +++ b/comfy/ldm/hidream/model.py @@ -699,10 +699,13 @@ class HiDreamImageTransformer2DModel(nn.Module): y: Optional[torch.Tensor] = None, context: Optional[torch.Tensor] = None, encoder_hidden_states_llama3=None, + image_cond=None, control = None, transformer_options = {}, ) -> torch.Tensor: bs, c, h, w = x.shape + if image_cond is not None: + x = torch.cat([x, image_cond], dim=-1) hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size)) timesteps = t pooled_embeds = y diff --git a/comfy/ldm/hunyuan_video/model.py b/comfy/ldm/hunyuan_video/model.py index f878b57f9..a2d62d220 100644 --- a/comfy/ldm/hunyuan_video/model.py +++ b/comfy/ldm/hunyuan_video/model.py @@ -219,6 +219,7 @@ class HunyuanVideo(nn.Module): y: Tensor, guidance: Tensor = None, guiding_frame_index=None, + ref_latent=None, control=None, transformer_options={}, ) -> Tensor: @@ -229,6 +230,14 @@ class HunyuanVideo(nn.Module): img = self.img_in(img) vec = self.time_in(timestep_embedding(timesteps, 256, time_factor=1.0).to(img.dtype)) + if ref_latent is not None: + ref_latent_ids = self.img_ids(ref_latent) + ref_latent = self.img_in(ref_latent) + img = torch.cat([ref_latent, img], dim=-2) + ref_latent_ids[..., 0] = -1 + ref_latent_ids[..., 2] += (initial_shape[-1] // self.patch_size[-1]) + img_ids = torch.cat([ref_latent_ids, img_ids], dim=-2) + if guiding_frame_index is not None: token_replace_vec = self.time_in(timestep_embedding(guiding_frame_index, 256, time_factor=1.0)) vec_ = self.vector_in(y[:, :self.params.vec_in_dim]) @@ -304,6 +313,8 @@ class HunyuanVideo(nn.Module): img[:, : img_len] += add img = img[:, : img_len] + if ref_latent is not None: + img = img[:, ref_latent.shape[1]:] img = self.final_layer(img, vec, modulation_dims=modulation_dims) # (N, T, patch_size ** 2 * out_channels) @@ -315,7 +326,7 @@ class HunyuanVideo(nn.Module): img = img.reshape(initial_shape[0], self.out_channels, initial_shape[2], initial_shape[3], initial_shape[4]) return img - def forward(self, x, timestep, context, y, guidance=None, attention_mask=None, guiding_frame_index=None, control=None, transformer_options={}, **kwargs): + def img_ids(self, x): bs, c, t, h, w = x.shape patch_size = self.patch_size t_len = ((t + (patch_size[0] // 2)) // patch_size[0]) @@ -325,7 +336,11 @@ class HunyuanVideo(nn.Module): img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(0, t_len - 1, steps=t_len, device=x.device, dtype=x.dtype).reshape(-1, 1, 1) img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).reshape(1, -1, 1) img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).reshape(1, 1, -1) - img_ids = repeat(img_ids, "t h w c -> b (t h w) c", b=bs) + return repeat(img_ids, "t h w c -> b (t h w) c", b=bs) + + def forward(self, x, timestep, context, y, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, control=None, transformer_options={}, **kwargs): + bs, c, t, h, w = x.shape + img_ids = self.img_ids(x) txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype) - out = self.forward_orig(x, img_ids, context, txt_ids, attention_mask, timestep, y, guidance, guiding_frame_index, control, transformer_options) + out = self.forward_orig(x, img_ids, context, txt_ids, attention_mask, timestep, y, guidance, guiding_frame_index, ref_latent, control=control, transformer_options=transformer_options) return out diff --git a/comfy/ldm/hydit/models.py b/comfy/ldm/hydit/models.py index 320318cef..993b61e74 100644 --- a/comfy/ldm/hydit/models.py +++ b/comfy/ldm/hydit/models.py @@ -3,7 +3,7 @@ import torch import torch.nn as nn from ... import ops -from ..modules.diffusionmodules.mmdit import Mlp, TimestepEmbedder, PatchEmbed, RMSNorm +from ..modules.diffusionmodules.mmdit import Mlp, TimestepEmbedder, PatchEmbed from ..modules.diffusionmodules.util import timestep_embedding from torch.utils import checkpoint @@ -51,7 +51,7 @@ class HunYuanDiTBlock(nn.Module): if norm_type == "layer": norm_layer = operations.LayerNorm elif norm_type == "rms": - norm_layer = RMSNorm + norm_layer = operations.RMSNorm else: raise ValueError(f"Unknown norm_type: {norm_type}") diff --git a/comfy/ldm/lightricks/model.py b/comfy/ldm/lightricks/model.py index 4863ea70c..23ed454d1 100644 --- a/comfy/ldm/lightricks/model.py +++ b/comfy/ldm/lightricks/model.py @@ -2,7 +2,6 @@ import torch from torch import nn from ..common_dit import rms_norm -from ..genmo.joint_model.layers import RMSNorm from einops import rearrange import math from typing import Dict, Optional, Tuple @@ -263,8 +262,8 @@ class CrossAttention(nn.Module): self.heads = heads self.dim_head = dim_head - self.q_norm = RMSNorm(inner_dim, dtype=dtype, device=device) - self.k_norm = RMSNorm(inner_dim, dtype=dtype, device=device) + self.q_norm = operations.RMSNorm(inner_dim, dtype=dtype, device=device) + self.k_norm = operations.RMSNorm(inner_dim, dtype=dtype, device=device) self.to_q = operations.Linear(query_dim, inner_dim, bias=True, dtype=dtype, device=device) self.to_k = operations.Linear(context_dim, inner_dim, bias=True, dtype=dtype, device=device) diff --git a/comfy/ldm/lumina/model.py b/comfy/ldm/lumina/model.py index 45e74d69a..2cda78da2 100644 --- a/comfy/ldm/lumina/model.py +++ b/comfy/ldm/lumina/model.py @@ -8,7 +8,7 @@ import torch.nn as nn import torch.nn.functional as F from ..common_dit import pad_to_patch_size -from ..modules.diffusionmodules.mmdit import TimestepEmbedder, RMSNorm +from ..modules.diffusionmodules.mmdit import TimestepEmbedder from ..modules.attention import optimized_attention_masked from ..flux.layers import EmbedND @@ -64,8 +64,8 @@ class JointAttention(nn.Module): ) if qk_norm: - self.q_norm = RMSNorm(self.head_dim, elementwise_affine=True, **operation_settings) - self.k_norm = RMSNorm(self.head_dim, elementwise_affine=True, **operation_settings) + self.q_norm = operation_settings.get("operations").RMSNorm(self.head_dim, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.k_norm = operation_settings.get("operations").RMSNorm(self.head_dim, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) else: self.q_norm = self.k_norm = nn.Identity() @@ -242,11 +242,11 @@ class JointTransformerBlock(nn.Module): operation_settings=operation_settings, ) self.layer_id = layer_id - self.attention_norm1 = RMSNorm(dim, eps=norm_eps, elementwise_affine=True, **operation_settings) - self.ffn_norm1 = RMSNorm(dim, eps=norm_eps, elementwise_affine=True, **operation_settings) + self.attention_norm1 = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.ffn_norm1 = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) - self.attention_norm2 = RMSNorm(dim, eps=norm_eps, elementwise_affine=True, **operation_settings) - self.ffn_norm2 = RMSNorm(dim, eps=norm_eps, elementwise_affine=True, **operation_settings) + self.attention_norm2 = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.ffn_norm2 = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) self.modulation = modulation if modulation: @@ -431,7 +431,7 @@ class NextDiT(nn.Module): self.t_embedder = TimestepEmbedder(min(dim, 1024), **operation_settings) self.cap_embedder = nn.Sequential( - RMSNorm(cap_feat_dim, eps=norm_eps, elementwise_affine=True, **operation_settings), + 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( cap_feat_dim, dim, @@ -457,7 +457,7 @@ class NextDiT(nn.Module): for layer_id in range(n_layers) ] ) - self.norm_final = RMSNorm(dim, eps=norm_eps, elementwise_affine=True, **operation_settings) + 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) assert (dim // n_heads) == sum(axes_dims) diff --git a/comfy/ldm/modules/attention.py b/comfy/ldm/modules/attention.py index de0dea230..aa9faf723 100644 --- a/comfy/ldm/modules/attention.py +++ b/comfy/ldm/modules/attention.py @@ -17,9 +17,15 @@ if model_management.xformers_enabled(): import xformers.ops # pylint: disable=import-error if model_management.sage_attention_enabled(): - from sageattention import sageattn # pylint: disable=import-error -else: - sageattn = torch.nn.functional.scaled_dot_product_attention + try: + from sageattention import sageattn # pylint: disable=import-error + except ModuleNotFoundError as e: + if e.name == "sageattention": + logging.error(f"\n\nTo use the `--use-sage-attention` feature, the `sageattention` package must be installed first.\ncommand:\n\t{sys.executable} -m pip install sageattention") + else: + raise e + sageattn = torch.nn.functional.scaled_dot_product_attention + if model_management.flash_attention_enabled(): from flash_attn import flash_attn_func # pylint: disable=import-error diff --git a/comfy/ldm/wan/model.py b/comfy/ldm/wan/model.py index 5e7848bd5..1d6edb354 100644 --- a/comfy/ldm/wan/model.py +++ b/comfy/ldm/wan/model.py @@ -9,7 +9,6 @@ from einops import repeat from comfy.ldm.modules.attention import optimized_attention from comfy.ldm.flux.layers import EmbedND from comfy.ldm.flux.math import apply_rope -from comfy.ldm.modules.diffusionmodules.mmdit import RMSNorm import comfy.ldm.common_dit import comfy.model_management @@ -49,8 +48,8 @@ class WanSelfAttention(nn.Module): self.k = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) self.v = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) self.o = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) - self.norm_q = RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity() - self.norm_k = RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity() + self.norm_q = operation_settings.get("operations").RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity() + self.norm_k = operation_settings.get("operations").RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity() def forward(self, x, freqs): r""" @@ -114,7 +113,7 @@ class WanI2VCrossAttention(WanSelfAttention): self.k_img = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) self.v_img = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) # self.alpha = nn.Parameter(torch.zeros((1, ))) - self.norm_k_img = RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity() + self.norm_k_img = operation_settings.get("operations").RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity() def forward(self, x, context, context_img_len): r""" @@ -248,6 +247,60 @@ class VaceWanAttentionBlock(WanAttentionBlock): return c_skip, c +class WanCamAdapter(nn.Module): + def __init__(self, in_dim, out_dim, kernel_size, stride, num_residual_blocks=1, operation_settings={}): + super(WanCamAdapter, self).__init__() + + # Pixel Unshuffle: reduce spatial dimensions by a factor of 8 + self.pixel_unshuffle = nn.PixelUnshuffle(downscale_factor=8) + + # Convolution: reduce spatial dimensions by a factor + # of 2 (without overlap) + self.conv = operation_settings.get("operations").Conv2d(in_dim * 64, out_dim, kernel_size=kernel_size, stride=stride, padding=0, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + + # Residual blocks for feature extraction + self.residual_blocks = nn.Sequential( + *[WanCamResidualBlock(out_dim, operation_settings = operation_settings) for _ in range(num_residual_blocks)] + ) + + def forward(self, x): + # Reshape to merge the frame dimension into batch + bs, c, f, h, w = x.size() + x = x.permute(0, 2, 1, 3, 4).contiguous().view(bs * f, c, h, w) + + # Pixel Unshuffle operation + x_unshuffled = self.pixel_unshuffle(x) + + # Convolution operation + x_conv = self.conv(x_unshuffled) + + # Feature extraction with residual blocks + out = self.residual_blocks(x_conv) + + # Reshape to restore original bf dimension + out = out.view(bs, f, out.size(1), out.size(2), out.size(3)) + + # Permute dimensions to reorder (if needed), e.g., swap channels and feature frames + out = out.permute(0, 2, 1, 3, 4) + + return out + + +class WanCamResidualBlock(nn.Module): + def __init__(self, dim, operation_settings={}): + super(WanCamResidualBlock, self).__init__() + self.conv1 = operation_settings.get("operations").Conv2d(dim, dim, kernel_size=3, padding=1, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.relu = nn.ReLU(inplace=True) + self.conv2 = operation_settings.get("operations").Conv2d(dim, dim, kernel_size=3, padding=1, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + + def forward(self, x): + residual = x + out = self.relu(self.conv1(x)) + out = self.conv2(out) + out += residual + return out + + class Head(nn.Module): def __init__(self, dim, out_dim, patch_size, eps=1e-6, operation_settings={}): @@ -486,13 +539,20 @@ class WanModel(torch.nn.Module): x = self.unpatchify(x, grid_sizes) return x - def forward(self, x, timestep, context, clip_fea=None, transformer_options={}, **kwargs): + def forward(self, x, timestep, context, clip_fea=None, time_dim_concat=None, transformer_options={}, **kwargs): bs, c, t, h, w = x.shape x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size) + 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 time_dim_concat is not None: + time_dim_concat = comfy.ldm.common_dit.pad_to_patch_size(time_dim_concat, self.patch_size) + x = torch.cat([x, time_dim_concat], dim=2) + t_len = ((x.shape[2] + (patch_size[0] // 2)) // patch_size[0]) + img_ids = torch.zeros((t_len, h_len, w_len, 3), device=x.device, dtype=x.dtype) img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(0, t_len - 1, steps=t_len, device=x.device, dtype=x.dtype).reshape(-1, 1, 1) img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).reshape(1, -1, 1) @@ -582,6 +642,7 @@ class VaceWanModel(WanModel): t, context, vace_context, + vace_strength, clip_fea=None, freqs=None, transformer_options={}, @@ -607,8 +668,11 @@ class VaceWanModel(WanModel): context = torch.concat([context_clip, context], dim=1) context_img_len = clip_fea.shape[-2] + orig_shape = list(vace_context.shape) + vace_context = vace_context.movedim(0, 1).reshape([-1] + orig_shape[2:]) c = self.vace_patch_embedding(vace_context.float()).to(vace_context.dtype) c = c.flatten(2).transpose(1, 2) + c = list(c.split(orig_shape[0], dim=0)) # arguments x_orig = x @@ -628,8 +692,99 @@ class VaceWanModel(WanModel): ii = self.vace_layers_mapping.get(i, None) if ii is not None: - c_skip, c = self.vace_blocks[ii](c, x=x_orig, e=e0, freqs=freqs, context=context, context_img_len=context_img_len) - x += c_skip + for iii in range(len(c)): + c_skip, c[iii] = self.vace_blocks[ii](c[iii], x=x_orig, e=e0, freqs=freqs, context=context, context_img_len=context_img_len) + x += c_skip * vace_strength[iii] + del c_skip + # head + x = self.head(x, e) + + # unpatchify + x = self.unpatchify(x, grid_sizes) + return x + +class CameraWanModel(WanModel): + r""" + Wan diffusion backbone supporting both text-to-video and image-to-video. + """ + + def __init__(self, + model_type='camera', + patch_size=(1, 2, 2), + text_len=512, + in_dim=16, + dim=2048, + ffn_dim=8192, + freq_dim=256, + text_dim=4096, + out_dim=16, + num_heads=16, + num_layers=32, + window_size=(-1, -1), + qk_norm=True, + cross_attn_norm=True, + eps=1e-6, + flf_pos_embed_token_number=None, + image_model=None, + in_dim_control_adapter=24, + device=None, + dtype=None, + operations=None, + ): + + super().__init__(model_type='i2v', patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, flf_pos_embed_token_number=flf_pos_embed_token_number, image_model=image_model, device=device, dtype=dtype, operations=operations) + operation_settings = {"operations": operations, "device": device, "dtype": dtype} + + self.control_adapter = WanCamAdapter(in_dim_control_adapter, dim, kernel_size=patch_size[1:], stride=patch_size[1:], operation_settings=operation_settings) + + + def forward_orig( + self, + x, + t, + context, + clip_fea=None, + freqs=None, + camera_conditions = None, + transformer_options={}, + **kwargs, + ): + # embeddings + x = self.patch_embedding(x.float()).to(x.dtype) + if self.control_adapter is not None and camera_conditions is not None: + x_camera = self.control_adapter(camera_conditions).to(x.dtype) + x = x + x_camera + grid_sizes = x.shape[2:] + x = x.flatten(2).transpose(1, 2) + + # time embeddings + e = self.time_embedding( + sinusoidal_embedding_1d(self.freq_dim, t).to(dtype=x[0].dtype)) + e0 = self.time_projection(e).unflatten(1, (6, self.dim)) + + # context + context = self.text_embedding(context) + + context_img_len = None + if clip_fea is not None: + if self.img_emb is not None: + context_clip = self.img_emb(clip_fea) # bs x 257 x dim + context = torch.concat([context_clip, context], dim=1) + context_img_len = clip_fea.shape[-2] + + patches_replace = transformer_options.get("patches_replace", {}) + blocks_replace = patches_replace.get("dit", {}) + for i, block in enumerate(self.blocks): + if ("double_block", i) in blocks_replace: + def block_wrap(args): + out = {} + out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], context_img_len=context_img_len) + return out + out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap}) + x = out["img"] + else: + x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len) + # head x = self.head(x, e) diff --git a/comfy/lora.py b/comfy/lora.py index 43a726623..659bb6cff 100644 --- a/comfy/lora.py +++ b/comfy/lora.py @@ -25,6 +25,7 @@ from . import model_base from . import model_management from . import utils from .lora_types import PatchDict, PatchOffset, PatchConversionFunction, PatchType, ModelPatchesDictValue +from . import weight_adapter LORA_CLIP_MAP = { "mlp.fc1": "mlp_fc1", @@ -52,133 +53,12 @@ def load_lora(lora, to_load, log_missing=True) -> PatchDict: dora_scale = lora[dora_scale_name] loaded_keys.add(dora_scale_name) - reshape_name = "{}.reshape_weight".format(x) - reshape = None - if reshape_name in lora.keys(): - try: - reshape = lora[reshape_name].tolist() - loaded_keys.add(reshape_name) - except: - pass - - regular_lora = "{}.lora_up.weight".format(x) - diffusers_lora = "{}_lora.up.weight".format(x) - diffusers2_lora = "{}.lora_B.weight".format(x) - diffusers3_lora = "{}.lora.up.weight".format(x) - mochi_lora = "{}.lora_B".format(x) - transformers_lora = "{}.lora_linear_layer.up.weight".format(x) - A_name = B_name = None - - mid_name = None - if regular_lora in lora.keys(): - A_name = regular_lora - B_name = "{}.lora_down.weight".format(x) - mid_name = "{}.lora_mid.weight".format(x) - elif diffusers_lora in lora.keys(): - A_name = diffusers_lora - B_name = "{}_lora.down.weight".format(x) - elif diffusers2_lora in lora.keys(): - A_name = diffusers2_lora - B_name = "{}.lora_A.weight".format(x) - elif diffusers3_lora in lora.keys(): - A_name = diffusers3_lora - B_name = "{}.lora.down.weight".format(x) - elif mochi_lora in lora.keys(): - A_name = mochi_lora - B_name = "{}.lora_A".format(x) - elif transformers_lora in lora.keys(): - A_name = transformers_lora - B_name = "{}.lora_linear_layer.down.weight".format(x) - - if A_name is not None: - mid = None - if mid_name is not None and mid_name in lora.keys(): - mid = lora[mid_name] - loaded_keys.add(mid_name) - patch_dict[to_load[x]] = ("lora", (lora[A_name], lora[B_name], alpha, mid, dora_scale, reshape)) - loaded_keys.add(A_name) - loaded_keys.add(B_name) - - ######## loha - hada_w1_a_name = "{}.hada_w1_a".format(x) - hada_w1_b_name = "{}.hada_w1_b".format(x) - hada_w2_a_name = "{}.hada_w2_a".format(x) - hada_w2_b_name = "{}.hada_w2_b".format(x) - hada_t1_name = "{}.hada_t1".format(x) - hada_t2_name = "{}.hada_t2".format(x) - if hada_w1_a_name in lora.keys(): - hada_t1 = None - hada_t2 = None - if hada_t1_name in lora.keys(): - hada_t1 = lora[hada_t1_name] - hada_t2 = lora[hada_t2_name] - loaded_keys.add(hada_t1_name) - loaded_keys.add(hada_t2_name) - - patch_dict[to_load[x]] = ("loha", (lora[hada_w1_a_name], lora[hada_w1_b_name], alpha, lora[hada_w2_a_name], lora[hada_w2_b_name], hada_t1, hada_t2, dora_scale)) - loaded_keys.add(hada_w1_a_name) - loaded_keys.add(hada_w1_b_name) - loaded_keys.add(hada_w2_a_name) - loaded_keys.add(hada_w2_b_name) - - ######## lokr - lokr_w1_name = "{}.lokr_w1".format(x) - lokr_w2_name = "{}.lokr_w2".format(x) - lokr_w1_a_name = "{}.lokr_w1_a".format(x) - lokr_w1_b_name = "{}.lokr_w1_b".format(x) - lokr_t2_name = "{}.lokr_t2".format(x) - lokr_w2_a_name = "{}.lokr_w2_a".format(x) - lokr_w2_b_name = "{}.lokr_w2_b".format(x) - - lokr_w1 = None - if lokr_w1_name in lora.keys(): - lokr_w1 = lora[lokr_w1_name] - loaded_keys.add(lokr_w1_name) - - lokr_w2 = None - if lokr_w2_name in lora.keys(): - lokr_w2 = lora[lokr_w2_name] - loaded_keys.add(lokr_w2_name) - - lokr_w1_a = None - if lokr_w1_a_name in lora.keys(): - lokr_w1_a = lora[lokr_w1_a_name] - loaded_keys.add(lokr_w1_a_name) - - lokr_w1_b = None - if lokr_w1_b_name in lora.keys(): - lokr_w1_b = lora[lokr_w1_b_name] - loaded_keys.add(lokr_w1_b_name) - - lokr_w2_a = None - if lokr_w2_a_name in lora.keys(): - lokr_w2_a = lora[lokr_w2_a_name] - loaded_keys.add(lokr_w2_a_name) - - lokr_w2_b = None - if lokr_w2_b_name in lora.keys(): - lokr_w2_b = lora[lokr_w2_b_name] - loaded_keys.add(lokr_w2_b_name) - - lokr_t2 = None - if lokr_t2_name in lora.keys(): - lokr_t2 = lora[lokr_t2_name] - loaded_keys.add(lokr_t2_name) - - if (lokr_w1 is not None) or (lokr_w2 is not None) or (lokr_w1_a is not None) or (lokr_w2_a is not None): - patch_dict[to_load[x]] = ("lokr", (lokr_w1, lokr_w2, alpha, lokr_w1_a, lokr_w1_b, lokr_w2_a, lokr_w2_b, lokr_t2, dora_scale)) - - # glora - a1_name = "{}.a1.weight".format(x) - a2_name = "{}.a2.weight".format(x) - b1_name = "{}.b1.weight".format(x) - b2_name = "{}.b2.weight".format(x) - if a1_name in lora: - patch_dict[to_load[x]] = ("glora", (lora[a1_name], lora[a2_name], lora[b1_name], lora[b2_name], alpha, dora_scale)) - loaded_keys.add(a1_name) - loaded_keys.add(a2_name) - loaded_keys.add(b1_name) - loaded_keys.add(b2_name) + for adapter_cls in weight_adapter.adapters: + adapter = adapter_cls.load(x, lora, alpha, dora_scale, loaded_keys) + if adapter is not None: + patch_dict[to_load[x]] = adapter + loaded_keys.update(adapter.loaded_keys) + continue w_norm_name = "{}.w_norm".format(x) b_norm_name = "{}.b_norm".format(x) @@ -408,30 +288,23 @@ def model_lora_keys_unet(model, key_map=None): key_map["transformer.{}".format(key_lora)] = k key_map["diffusion_model.{}".format(key_lora)] = k # Old loras + if isinstance(model, model_base.HiDream): + for k in sdk: + if k.startswith("diffusion_model."): + if k.endswith(".weight"): + key_lora = k[len("diffusion_model."):-len(".weight")] + key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = k #SimpleTuner lycoris format + key_map["transformer.{}".format(key_lora)] = k #SimpleTuner regular format + + if isinstance(model, model_base.ACEStep): + for k in sdk: + if k.startswith("diffusion_model.") and k.endswith(".weight"): #Official ACE step lora format + key_lora = k[len("diffusion_model."):-len(".weight")] + key_map["{}".format(key_lora)] = k + return key_map -def weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function): - dora_scale = model_management.cast_to_device(dora_scale, weight.device, intermediate_dtype) - lora_diff *= alpha - weight_calc = weight + function(lora_diff).type(weight.dtype) - weight_norm = ( - weight_calc.transpose(0, 1) - .reshape(weight_calc.shape[1], -1) - .norm(dim=1, keepdim=True) - .reshape(weight_calc.shape[1], *[1] * (weight_calc.dim() - 1)) - .transpose(0, 1) - ) - - weight_calc *= (dora_scale / weight_norm).type(weight.dtype) - if strength != 1.0: - weight_calc -= weight - weight += strength * (weight_calc) - else: - weight[:] = weight_calc - return weight - - def pad_tensor_to_shape(tensor: torch.Tensor, new_shape: list[int]) -> torch.Tensor: """ Pad a tensor to a new shape with zeros. @@ -488,6 +361,16 @@ def calculate_weight(patches: ModelPatchesDictValue, weight, key, intermediate_d v = (calculate_weight(v[1:], v[0][1](model_management.cast_to_device(v[0][0], weight.device, intermediate_dtype, copy=True), inplace=True), key, intermediate_dtype=intermediate_dtype),) patch_type: PatchType = "" + if isinstance(v, weight_adapter.WeightAdapterBase): + output = v.calculate_weight(weight, key, strength, strength_model, offset, function, intermediate_dtype, original_weights) + if output is None: + logging.warning("Calculate Weight Failed: {} {}".format(v.name, key)) + else: + weight = output + if old_weight is not None: + weight = old_weight + continue + if len(v) == 1: patch_type = "diff" elif len(v) == 2: @@ -514,158 +397,6 @@ def calculate_weight(patches: ModelPatchesDictValue, weight, key, intermediate_d diff_weight = model_management.cast_to_device(target_weight, weight.device, intermediate_dtype) - \ model_management.cast_to_device(original_weights[key][0][0], weight.device, intermediate_dtype) weight += function(strength * model_management.cast_to_device(diff_weight, weight.device, weight.dtype)) - elif patch_type == "lora": # lora/locon - mat1 = model_management.cast_to_device(v[0], weight.device, intermediate_dtype) - mat2 = model_management.cast_to_device(v[1], weight.device, intermediate_dtype) - dora_scale = v[4] - reshape = v[5] - - if reshape is not None: - weight = pad_tensor_to_shape(weight, reshape) - - if v[2] is not None: - alpha = v[2] / mat2.shape[0] - else: - alpha = 1.0 - - if v[3] is not None: - # locon mid weights, hopefully the math is fine because I didn't properly test it - mat3 = model_management.cast_to_device(v[3], weight.device, intermediate_dtype) - final_shape = [mat2.shape[1], mat2.shape[0], mat3.shape[2], mat3.shape[3]] - mat2 = torch.mm(mat2.transpose(0, 1).flatten(start_dim=1), mat3.transpose(0, 1).flatten(start_dim=1)).reshape(final_shape).transpose(0, 1) - try: - lora_diff = torch.mm(mat1.flatten(start_dim=1), mat2.flatten(start_dim=1)).reshape(weight.shape) - if dora_scale is not None: - weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function) - else: - weight += function(((strength * alpha) * lora_diff).type(weight.dtype)) - except Exception as e: - logging.error("ERROR {} {} {}".format(patch_type, key, e)) - elif patch_type == "lokr": - w1 = v[0] - w2 = v[1] - w1_a = v[3] - w1_b = v[4] - w2_a = v[5] - w2_b = v[6] - t2 = v[7] - dora_scale = v[8] - dim = None - - if w1 is None: - dim = w1_b.shape[0] - w1 = torch.mm(model_management.cast_to_device(w1_a, weight.device, intermediate_dtype), - model_management.cast_to_device(w1_b, weight.device, intermediate_dtype)) - else: - w1 = model_management.cast_to_device(w1, weight.device, intermediate_dtype) - - if w2 is None: - dim = w2_b.shape[0] - if t2 is None: - w2 = torch.mm(model_management.cast_to_device(w2_a, weight.device, intermediate_dtype), - model_management.cast_to_device(w2_b, weight.device, intermediate_dtype)) - else: - w2 = torch.einsum('i j k l, j r, i p -> p r k l', - model_management.cast_to_device(t2, weight.device, intermediate_dtype), - model_management.cast_to_device(w2_b, weight.device, intermediate_dtype), - model_management.cast_to_device(w2_a, weight.device, intermediate_dtype)) - else: - w2 = model_management.cast_to_device(w2, weight.device, intermediate_dtype) - - if len(w2.shape) == 4: - w1 = w1.unsqueeze(2).unsqueeze(2) - if v[2] is not None and dim is not None: - alpha = v[2] / dim - else: - alpha = 1.0 - - try: - lora_diff = torch.kron(w1, w2).reshape(weight.shape) - if dora_scale is not None: - weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function) - else: - weight += function(((strength * alpha) * lora_diff).type(weight.dtype)) - except Exception as e: - logging.error("ERROR {} {} {}".format(patch_type, key, e)) - elif patch_type == "loha": - w1a = v[0] - w1b = v[1] - if v[2] is not None: - alpha = v[2] / w1b.shape[0] - else: - alpha = 1.0 - - w2a = v[3] - w2b = v[4] - dora_scale = v[7] - if v[5] is not None: # cp decomposition - t1 = v[5] - t2 = v[6] - m1 = torch.einsum('i j k l, j r, i p -> p r k l', - model_management.cast_to_device(t1, weight.device, intermediate_dtype), - model_management.cast_to_device(w1b, weight.device, intermediate_dtype), - model_management.cast_to_device(w1a, weight.device, intermediate_dtype)) - - m2 = torch.einsum('i j k l, j r, i p -> p r k l', - model_management.cast_to_device(t2, weight.device, intermediate_dtype), - model_management.cast_to_device(w2b, weight.device, intermediate_dtype), - model_management.cast_to_device(w2a, weight.device, intermediate_dtype)) - else: - m1 = torch.mm(model_management.cast_to_device(w1a, weight.device, intermediate_dtype), - model_management.cast_to_device(w1b, weight.device, intermediate_dtype)) - m2 = torch.mm(model_management.cast_to_device(w2a, weight.device, intermediate_dtype), - model_management.cast_to_device(w2b, weight.device, intermediate_dtype)) - - try: - lora_diff = (m1 * m2).reshape(weight.shape) - if dora_scale is not None: - weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function) - else: - weight += function(((strength * alpha) * lora_diff).type(weight.dtype)) - except Exception as e: - logging.error("ERROR {} {} {}".format(patch_type, key, e)) - elif patch_type == "glora": - rank = 0 - dora_scale = v[5] - - old_glora = False - if v[3].shape[1] == v[2].shape[0] == v[0].shape[0] == v[1].shape[1]: - rank = v[0].shape[0] - old_glora = True - - if v[3].shape[0] == v[2].shape[1] == v[0].shape[1] == v[1].shape[0]: - if old_glora and v[1].shape[0] == weight.shape[0] and weight.shape[0] == weight.shape[1]: - pass - else: - old_glora = False - rank = v[1].shape[0] - - a1 = model_management.cast_to_device(v[0].flatten(start_dim=1), weight.device, intermediate_dtype) - a2 = model_management.cast_to_device(v[1].flatten(start_dim=1), weight.device, intermediate_dtype) - b1 = model_management.cast_to_device(v[2].flatten(start_dim=1), weight.device, intermediate_dtype) - b2 = model_management.cast_to_device(v[3].flatten(start_dim=1), weight.device, intermediate_dtype) - - if v[4] is not None: - alpha = v[4] / rank - else: - alpha = 1.0 - - try: - if old_glora: - lora_diff = (torch.mm(b2, b1) + torch.mm(torch.mm(weight.flatten(start_dim=1).to(dtype=intermediate_dtype), a2), a1)).reshape(weight.shape) # old lycoris glora - else: - if weight.dim() > 2: - lora_diff = torch.einsum("o i ..., i j -> o j ...", torch.einsum("o i ..., i j -> o j ...", weight.to(dtype=intermediate_dtype), a1), a2).reshape(weight.shape) - else: - lora_diff = torch.mm(torch.mm(weight.to(dtype=intermediate_dtype), a1), a2).reshape(weight.shape) - lora_diff += torch.mm(b1, b2).reshape(weight.shape) - - if dora_scale is not None: - weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function) - else: - weight += function(((strength * alpha) * lora_diff).type(weight.dtype)) - except Exception as e: - logging.error("ERROR {} {} {}".format(patch_type, key, e)) else: logging.warning("patch type not recognized {} {}".format(patch_type, key)) diff --git a/comfy/model_base.py b/comfy/model_base.py index a442cb318..cd3b07b85 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -28,6 +28,7 @@ from . import latent_formats from . import model_management from . import ops from . import utils +from .ldm.ace.model import ACEStepTransformer2DModel from .ldm.audio.dit import AudioDiffusionTransformer from .ldm.audio.embedders import NumberConditioner from .ldm.aura.mmdit import MMDiT as AuraMMDiT @@ -47,7 +48,8 @@ from .ldm.modules.diffusionmodules.openaimodel import UNetModel, Timestep from .ldm.modules.diffusionmodules.upscaling import ImageConcatWithNoiseAugmentation from .ldm.modules.encoders.noise_aug_modules import CLIPEmbeddingNoiseAugmentation from .ldm.pixart.pixartms import PixArtMS -from .ldm.wan.model import WanModel, VaceWanModel +from .ldm.wan.model import WanModel, VaceWanModel, CameraWanModel +from .ldm.chroma import model as chroma_model from .model_management_types import ModelManageable from .ops import Operations from .patcher_extension import WrapperExecutor, WrappersMP, get_all_wrappers @@ -111,6 +113,13 @@ class ComfyUIModel(Protocol): ... +def convert_tensor(extra, dtype): + if hasattr(extra, "dtype"): + if extra.dtype != torch.int and extra.dtype != torch.long: + extra = extra.to(dtype) + return extra + + class BaseModel(torch.nn.Module): def __init__(self, model_config, model_type=ModelType.EPS, device: torch.device = None, unet_model: Type[TModule] = UNetModel): super().__init__() @@ -148,6 +157,7 @@ class BaseModel(torch.nn.Module): logging.debug("model_type {}".format(model_type.name)) logging.debug("adm {}".format(self.adm_channels)) self.memory_usage_factor = model_config.memory_usage_factor + self.memory_usage_factor_conds = () self.training = False def apply_model(self, x, t, c_concat=None, c_crossattn=None, control=None, transformer_options={}, **kwargs): @@ -178,9 +188,14 @@ class BaseModel(torch.nn.Module): extra_conds = {} for o in kwargs: extra = kwargs[o] + if hasattr(extra, "dtype"): - if extra.dtype != torch.int and extra.dtype != torch.long: - extra = extra.to(dtype) + extra = convert_tensor(extra, dtype) + elif isinstance(extra, list): + ex = [] + for ext in extra: + ex.append(convert_tensor(ext, dtype)) + extra = ex extra_conds[o] = extra t = self.process_timestep(t, x=x, **extra_conds) @@ -342,19 +357,28 @@ class BaseModel(torch.nn.Module): def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs): return self.model_sampling.noise_scaling(sigma.reshape([sigma.shape[0]] + [1] * (len(noise.shape) - 1)), noise, latent_image) - def memory_required(self, input_shape): + def memory_required(self, input_shape, cond_shapes={}): + input_shapes = [input_shape] + for c in self.memory_usage_factor_conds: + shape = cond_shapes.get(c, None) + if shape is not None and len(shape) > 0: + input_shapes += shape + if model_management.xformers_enabled() or model_management.pytorch_attention_flash_attention(): dtype = self.get_dtype() if self.manual_cast_dtype is not None: dtype = self.manual_cast_dtype # TODO: this needs to be tweaked - area = input_shape[0] * math.prod(input_shape[2:]) + area = sum(map(lambda input_shape: input_shape[0] * math.prod(input_shape[2:]), input_shapes)) return (area * model_management.dtype_size(dtype) * 0.01 * self.memory_usage_factor) * (1024 * 1024) else: # TODO: this formula might be too aggressive since I tweaked the sub-quad and split algorithms to use less memory. - area = input_shape[0] * math.prod(input_shape[2:]) + area = sum(map(lambda input_shape: input_shape[0] * math.prod(input_shape[2:]), input_shapes)) return (area * 0.15 * self.memory_usage_factor) * (1024 * 1024) + def extra_conds_shapes(self, **kwargs): + return {} + def unclip_adm(unclip_conditioning, device, noise_augmentor, noise_augment_merge=0.0, seed=None): adm_inputs = [] @@ -820,8 +844,8 @@ class PixArt(BaseModel): class Flux(BaseModel): - def __init__(self, model_config, model_type=ModelType.FLUX, device=None): - super().__init__(model_config, model_type, device=device, unet_model=flux_model.Flux) + def __init__(self, model_config, model_type=ModelType.FLUX, device=None, unet_model=flux_model.Flux): + super().__init__(model_config, model_type, device=device, unet_model=unet_model) def concat_cond(self, **kwargs): try: @@ -958,6 +982,10 @@ class HunyuanVideo(BaseModel): guiding_frame_index = kwargs.get("guiding_frame_index", None) if guiding_frame_index is not None: out['guiding_frame_index'] = conds.CONDRegular(torch.FloatTensor([guiding_frame_index])) + + ref_latent = kwargs.get("ref_latent", None) + if ref_latent is not None: + out['ref_latent'] = conds.CONDRegular(self.process_latent_in(ref_latent)) return out def scale_latent_inpaint(self, latent_image, **kwargs): @@ -1082,6 +1110,10 @@ class WAN21(BaseModel): clip_vision_output = kwargs.get("clip_vision_output", None) if clip_vision_output is not None: out['clip_fea'] = conds.CONDRegular(clip_vision_output.penultimate_hidden_states) + + time_dim_concat = kwargs.get("time_dim_concat", None) + if time_dim_concat is not None: + out['time_dim_concat'] = conds.CONDRegular(self.process_latent_in(time_dim_concat)) return out @@ -1097,20 +1129,39 @@ class WAN21_Vace(WAN21): vace_frames = kwargs.get("vace_frames", None) if vace_frames is None: noise_shape[1] = 32 - vace_frames = torch.zeros(noise_shape, device=noise.device, dtype=noise.dtype) - - for i in range(0, vace_frames.shape[1], 16): - vace_frames = vace_frames.clone() - vace_frames[:, i:i + 16] = self.process_latent_in(vace_frames[:, i:i + 16]) + vace_frames = [torch.zeros(noise_shape, device=noise.device, dtype=noise.dtype)] mask = kwargs.get("vace_mask", None) if mask is None: noise_shape[1] = 64 - mask = torch.ones(noise_shape, device=noise.device, dtype=noise.dtype) + mask = [torch.ones(noise_shape, device=noise.device, dtype=noise.dtype)] * len(vace_frames) - out['vace_context'] = conds.CONDRegular(torch.cat([vace_frames.to(noise), mask.to(noise)], dim=1)) + vace_frames_out = [] + for j in range(len(vace_frames)): + vf = vace_frames[j].clone() + for i in range(0, vf.shape[1], 16): + vf[:, i:i + 16] = self.process_latent_in(vf[:, i:i + 16]) + vf = torch.cat([vf, mask[j]], dim=1) + vace_frames_out.append(vf) + + vace_frames = torch.stack(vace_frames_out, dim=1) + out['vace_context'] = conds.CONDRegular(vace_frames) + + vace_strength = kwargs.get("vace_strength", [1.0] * len(vace_frames_out)) + out['vace_strength'] = conds.CONDConstant(vace_strength) return out +class WAN21_Camera(WAN21): + def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None): + super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=CameraWanModel) + self.image_to_video = image_to_video + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + camera_conditions = kwargs.get("camera_conditions", None) + if camera_conditions is not None: + out['camera_conditions'] = conds.CONDRegular(camera_conditions) + return out class Hunyuan3Dv2(BaseModel): def __init__(self, model_config, model_type=ModelType.FLOW, device=None): @@ -1143,4 +1194,38 @@ class HiDream(BaseModel): conditioning_llama3 = kwargs.get("conditioning_llama3", None) if conditioning_llama3 is not None: out['encoder_hidden_states_llama3'] = conds.CONDRegular(conditioning_llama3) + image_cond = kwargs.get("concat_latent_image", None) + if image_cond is not None: + out['image_cond'] = conds.CONDNoiseShape(self.process_latent_in(image_cond)) + return out + +class Chroma(Flux): + def __init__(self, model_config, model_type=ModelType.FLUX, device=None): + super().__init__(model_config, model_type, device=device, unet_model=chroma_model.Chroma) + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + + guidance = kwargs.get("guidance", 0) + if guidance is not None: + out['guidance'] = conds.CONDRegular(torch.FloatTensor([guidance])) + return out + +class ACEStep(BaseModel): + def __init__(self, model_config, model_type=ModelType.FLOW, device=None): + super().__init__(model_config, model_type, device=device, unet_model=ACEStepTransformer2DModel) + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + noise = kwargs.get("noise", None) + + cross_attn = kwargs.get("cross_attn", None) + if cross_attn is not None: + out['c_crossattn'] = conds.CONDRegular(cross_attn) + + conditioning_lyrics = kwargs.get("conditioning_lyrics", None) + if cross_attn is not None: + out['lyric_token_idx'] = conds.CONDRegular(conditioning_lyrics) + out['speaker_embeds'] = conds.CONDRegular(torch.zeros(noise.shape[0], 512, device=noise.device, dtype=noise.dtype)) + out['lyrics_strength'] = conds.CONDConstant(kwargs.get("lyrics_strength", 1.0)) return out diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 610b08b8d..57772a356 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -168,7 +168,9 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): 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 - dit_config["vec_in_dim"] = 768 + 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 @@ -178,7 +180,16 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["axes_dim"] = [16, 56, 56] dit_config["theta"] = 10000 dit_config["qkv_bias"] = True - dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys + 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 + dit_config["out_channels"] = 64 + dit_config["in_dim"] = 64 + dit_config["out_dim"] = 3072 + dit_config["hidden_dim"] = 5120 + dit_config["n_layers"] = 5 + else: + dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys return dit_config if '{}t5_yproj.weight'.format(key_prefix) in state_dict_keys: # Genmo mochi preview @@ -215,10 +226,39 @@ 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["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 + dit_config["cross_attention_dim"] = shape[1] if metadata is not None and "config" in metadata: dit_config.update(json.loads(metadata["config"]).get("transformer", {})) return dit_config + if '{}genre_embedder.weight'.format(key_prefix) in state_dict_keys: #ACE-Step model + dit_config = {} + dit_config["audio_model"] = "ace" + dit_config["attention_head_dim"] = 128 + dit_config["in_channels"] = 8 + dit_config["inner_dim"] = 2560 + dit_config["max_height"] = 16 + dit_config["max_position"] = 32768 + dit_config["max_width"] = 32768 + dit_config["mlp_ratio"] = 2.5 + dit_config["num_attention_heads"] = 20 + dit_config["num_layers"] = 24 + dit_config["out_channels"] = 8 + dit_config["patch_size"] = [16, 1] + dit_config["rope_theta"] = 1000000.0 + dit_config["speaker_embedding_dim"] = 512 + dit_config["text_embedding_dim"] = 768 + + dit_config["ssl_encoder_depths"] = [8, 8] + dit_config["ssl_latent_dims"] = [1024, 768] + dit_config["ssl_names"] = ["mert", "m-hubert"] + dit_config["lyric_encoder_vocab_size"] = 6693 + dit_config["lyric_hidden_size"] = 1024 + return dit_config + if '{}t_block.1.weight'.format(key_prefix) in state_dict_keys: # PixArt patch_size = 2 dit_config = {} @@ -325,6 +365,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["model_type"] = "vace" dit_config["vace_in_dim"] = state_dict['{}vace_patch_embedding.weight'.format(key_prefix)].shape[1] dit_config["vace_layers"] = count_blocks(state_dict_keys, '{}vace_blocks.'.format(key_prefix) + '{}.') + elif '{}control_adapter.conv.weight'.format(key_prefix) in state_dict_keys: + dit_config["model_type"] = "camera" else: if '{}img_emb.proj.0.bias'.format(key_prefix) in state_dict_keys: dit_config["model_type"] = "i2v" @@ -584,6 +626,9 @@ def convert_config(unet_config): def unet_config_from_diffusers_unet(state_dict, dtype=None): + if "conv_in.weight" not in state_dict: + return None + match = {} transformer_depth = [] diff --git a/comfy/model_management.py b/comfy/model_management.py index 12db8b038..9c3cf68bb 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -340,14 +340,25 @@ try: except: pass + +SUPPORT_FP8_OPS = args.supports_fp8_compute try: if is_amd(): + try: + rocm_version = tuple(map(int, str(torch.version.hip).split(".")[:2])) + except: + rocm_version = (6, -1) arch = torch.cuda.get_device_properties(get_torch_device()).gcnArchName logger.info("AMD arch: {}".format(arch)) + logging.info("ROCm version: {}".format(rocm_version)) if args.use_split_cross_attention == False and args.use_quad_cross_attention == False: - if torch_version_numeric[0] >= 2 and torch_version_numeric[1] >= 7: # works on 2.6 but doesn't actually seem to improve much - if any((a in arch) for a in ["gfx1100", "gfx1101"]): # TODO: more arches + if torch_version_numeric >= (2, 7): # works on 2.6 but doesn't actually seem to improve much + if any((a in arch) for a in ["gfx90a", "gfx942", "gfx1100", "gfx1101", "gfx1151"]): # TODO: more arches, TODO: gfx1201 and gfx950 ENABLE_PYTORCH_ATTENTION = True + if torch_version_numeric >= (2, 7) and rocm_version >= (6, 4): + if any((a in arch) for a in ["gfx1201", "gfx942", "gfx950"]): # TODO: more arches + SUPPORT_FP8_OPS = True + except: pass @@ -366,7 +377,7 @@ except: pass try: - if torch_version_numeric[0] == 2 and torch_version_numeric[1] >= 5: + if torch_version_numeric >= (2, 5): torch.backends.cuda.allow_fp16_bf16_reduction_math_sdp(True) # pylint: disable=no-member except: logger.warning("Warning, could not set allow_fp16_bf16_reduction_math_sdp") @@ -793,7 +804,7 @@ def unet_initial_load_device(parameters, dtype): return torch_dev cpu_dev = torch.device("cpu") - if DISABLE_SMART_MEMORY: + if DISABLE_SMART_MEMORY or vram_state == VRAMState.NO_VRAM: return cpu_dev model_size = dtype_size(dtype) * parameters @@ -825,6 +836,8 @@ def unet_dtype(device=None, model_params=0, supported_dtypes=(torch.float16, tor return torch.float8_e4m3fn if args.fp8_e5m2_unet: return torch.float8_e5m2 + if args.fp8_e8m0fnu_unet: + return torch.float8_e8m0fnu fp8_dtype = None if weight_dtype in FLOAT8_TYPES: @@ -1055,15 +1068,60 @@ def force_channels_last(): return False -def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False): +STREAMS = {} +NUM_STREAMS = 1 +if args.async_offload: + NUM_STREAMS = 2 + logging.info("Using async weight offloading with {} streams".format(NUM_STREAMS)) + +stream_counters = {} +def get_offload_stream(device): + stream_counter = stream_counters.get(device, 0) + if NUM_STREAMS <= 1: + return None + + if device in STREAMS: + ss = STREAMS[device] + s = ss[stream_counter] + stream_counter = (stream_counter + 1) % len(ss) + if is_device_cuda(device): + ss[stream_counter].wait_stream(torch.cuda.current_stream()) + stream_counters[device] = stream_counter + return s + elif is_device_cuda(device): + ss = [] + for k in range(NUM_STREAMS): + ss.append(torch.cuda.Stream(device=device, priority=0)) + STREAMS[device] = ss + s = ss[stream_counter] + stream_counter = (stream_counter + 1) % len(ss) + stream_counters[device] = stream_counter + return s + return None + +def sync_stream(device, stream): + if stream is None: + return + if is_device_cuda(device): + torch.cuda.current_stream().wait_stream(stream) + +def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, stream=None): if device is None or weight.device == device: if not copy: if dtype is None or weight.dtype == dtype: return weight + if stream is not None: + with stream: + return weight.to(dtype=dtype, copy=copy) return weight.to(dtype=dtype, copy=copy) - r = torch.empty_like(weight, dtype=dtype, device=device) - r.copy_(weight, non_blocking=non_blocking) + if stream is not None: + with stream: + r = torch.empty_like(weight, dtype=dtype, device=device) + r.copy_(weight, non_blocking=non_blocking) + else: + r = torch.empty_like(weight, dtype=dtype, device=device) + r.copy_(weight, non_blocking=non_blocking) return r @@ -1380,6 +1438,9 @@ def bfloat16_support_mps(device): def supports_fp8_compute(device=None): + if SUPPORT_FP8_OPS: + return True + if not is_nvidia(): return False @@ -1395,11 +1456,11 @@ def supports_fp8_compute(device=None): if props.minor < 9: return False - if torch_version_numeric[0] < 2 or (torch_version_numeric[0] == 2 and torch_version_numeric[1] < 3): + if torch_version_numeric < (2, 3): return False if WINDOWS: - if (torch_version_numeric[0] == 2 and torch_version_numeric[1] < 4): + if torch_version_numeric < (2, 4): return False return True diff --git a/comfy/model_sampling.py b/comfy/model_sampling.py index 0c06466eb..e6c23cc7a 100644 --- a/comfy/model_sampling.py +++ b/comfy/model_sampling.py @@ -131,13 +131,14 @@ class ModelSamplingDiscrete(torch.nn.Module): self.num_timesteps = int(timesteps) self.linear_start = linear_start self.linear_end = linear_end + self.zsnr = zsnr # self.register_buffer('betas', torch.tensor(betas, dtype=torch.float32)) # self.register_buffer('alphas_cumprod', torch.tensor(alphas_cumprod, dtype=torch.float32)) # self.register_buffer('alphas_cumprod_prev', torch.tensor(alphas_cumprod_prev, dtype=torch.float32)) sigmas = ((1 - alphas_cumprod) / alphas_cumprod) ** 0.5 - if zsnr: + if self.zsnr: sigmas = rescale_zero_terminal_snr_sigmas(sigmas) self.set_sigmas(sigmas) diff --git a/comfy/node_helpers.py b/comfy/node_helpers.py index 7ed32913a..3acf1c503 100644 --- a/comfy/node_helpers.py +++ b/comfy/node_helpers.py @@ -1,19 +1,24 @@ import hashlib - from PIL import ImageFile, UnidentifiedImageError from .cli_args import args from .component_model.files import get_package_as_path -def conditioning_set_values(conditioning, values: dict = None): +def conditioning_set_values(conditioning, values: dict = None, append=False): if values is None: values = {} c = [] for t in conditioning: n = [t[0], t[1].copy()] for k in values: - n[1][k] = values[k] + val = values[k] + if append: + old_val = n[1].get(k, None) + if old_val is not None: + val = old_val + val + + n[1][k] = val c.append(n) return c diff --git a/comfy/nodes/base_nodes.py b/comfy/nodes/base_nodes.py index 11a1df5f6..dbabb9d18 100644 --- a/comfy/nodes/base_nodes.py +++ b/comfy/nodes/base_nodes.py @@ -247,6 +247,9 @@ class ConditioningZeroOut: pooled_output = d.get("pooled_output", None) if pooled_output is not None: d["pooled_output"] = torch.zeros_like(pooled_output) + conditioning_lyrics = d.get("conditioning_lyrics", None) + if conditioning_lyrics is not None: + d["conditioning_lyrics"] = torch.zeros_like(conditioning_lyrics) n = [torch.zeros_like(t[0]), d] c.append(n) return (c, ) @@ -949,7 +952,7 @@ class CLIPLoader: @classmethod def INPUT_TYPES(s): return {"required": { "clip_name": (get_filename_list_with_downloadable("text_encoders", KNOWN_CLIP_MODELS),), - "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream"], ), + "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream", "chroma", "ace"], ), }, "optional": { "device": (["default", "cpu"], {"advanced": True}), @@ -1135,16 +1138,7 @@ class unCLIPConditioning: if strength == 0: return (conditioning, ) - c = [] - for t in conditioning: - o = t[1].copy() - x = {"clip_vision_output": clip_vision_output, "strength": strength, "noise_augmentation": noise_augmentation} - if "unclip_conditioning" in o: - o["unclip_conditioning"] = o["unclip_conditioning"][:] + [x] - else: - o["unclip_conditioning"] = [x] - n = [t[0], o] - c.append(n) + c = node_helpers.conditioning_set_values(conditioning, {"unclip_conditioning": [{"clip_vision_output": clip_vision_output, "strength": strength, "noise_augmentation": noise_augmentation}]}, append=True) return (c, ) class GLIGENLoader: @@ -1966,6 +1960,7 @@ class ImagePadForOutpaint: mask[:, top:top + height, left:left + width] = t + # mask is already in batch shape return new_image, mask @@ -2098,11 +2093,13 @@ NODE_DISPLAY_NAME_MAPPINGS = { "ImagePadForOutpaint": "Pad Image for Outpainting", "ImageBatch": "Batch Images", "ImageCrop": "Image Crop", + "ImageStitch": "Image Stitch", "ImageBlend": "Image Blend", "ImageBlur": "Image Blur", "ImageQuantize": "Image Quantize", "ImageSharpen": "Image Sharpen", "ImageScaleToTotalPixels": "Scale Image to Total Pixels", + "GetImageSize": "Get Image Size", # _for_testing "VAEDecodeTiled": "VAE Decode (Tiled)", "VAEEncodeTiled": "VAE Encode (Tiled)", diff --git a/comfy/ops.py b/comfy/ops.py index ef9677dab..48aeea317 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -15,11 +15,11 @@ You should have received a copy of the GNU General Public License along with this program. If not, see . """ +import contextlib import logging -from typing import Optional, Type, Union - import torch from torch import Tensor +from typing import Optional, Type, Union from . import model_management, rmsnorm from .cli_args import args, PerformanceFeature @@ -44,20 +44,30 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None): if device is None: device = input.device + offload_stream = model_management.get_offload_stream(device) + if offload_stream is not None: + wf_context = offload_stream + else: + wf_context = contextlib.nullcontext() + bias = None non_blocking = True if torch.jit.is_tracing() or torch.jit.is_scripting() else model_management.device_supports_non_blocking(device) if s.bias is not None: has_function = len(s.bias_function) > 0 - bias = model_management.cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=has_function) + bias = model_management.cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=has_function, stream=offload_stream) if has_function: - for f in s.bias_function: - bias = f(bias) + with wf_context: + for f in s.bias_function: + bias = f(bias) has_function = len(s.weight_function) > 0 - weight = model_management.cast_to(s.weight, dtype, device, non_blocking=non_blocking, copy=has_function) + weight = model_management.cast_to(s.weight, dtype, device, non_blocking=non_blocking, copy=has_function, stream=offload_stream) if has_function: - for f in s.weight_function: - weight = f(weight) + with wf_context: + for f in s.weight_function: + weight = f(weight) + + model_management.sync_stream(device, offload_stream) return weight, bias @@ -351,10 +361,10 @@ def fp8_linear(self, input): 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) - input = input.reshape(-1, input_shape[2]).to(dtype) + input = input.reshape(-1, input_shape[2]).to(dtype).contiguous() else: scale_input = scale_input.to(input.device) - input = (input * (1.0 / scale_input).to(input_dtype)).reshape(-1, input_shape[2]).to(dtype) + input = (input * (1.0 / scale_input).to(input_dtype)).reshape(-1, input_shape[2]).to(dtype).contiguous() if bias is not None: o = torch._scaled_mm(input, w, out_dtype=input_dtype, bias=bias, scale_a=scale_input, scale_b=scale_weight) diff --git a/comfy/rmsnorm.py b/comfy/rmsnorm.py index accb91c01..3a40268b1 100644 --- a/comfy/rmsnorm.py +++ b/comfy/rmsnorm.py @@ -30,7 +30,7 @@ if RMSNorm is None: def __init__( self, normalized_shape, - eps=None, + eps=1e-6, elementwise_affine=True, device=None, dtype=None, diff --git a/comfy/sampler_helpers.py b/comfy/sampler_helpers.py index c0c4baf5f..918821249 100644 --- a/comfy/sampler_helpers.py +++ b/comfy/sampler_helpers.py @@ -1,3 +1,8 @@ +from __future__ import annotations + +import math + +import collections import uuid from . import model_management @@ -111,6 +116,23 @@ def cleanup_additional_models(models): m.cleanup() +def estimate_memory(model, noise_shape, conds): + cond_shapes = collections.defaultdict(list) + cond_shapes_min = {} + for _, cs in conds.items(): + for cond in cs: + for k, v in model.model.extra_conds_shapes(**cond).items(): + cond_shapes[k].append(v) + if cond_shapes_min.get(k, None) is None: + cond_shapes_min[k] = [v] + elif math.prod(v) > math.prod(cond_shapes_min[k][0]): + cond_shapes_min[k] = [v] + + memory_required = model.model.memory_required([noise_shape[0] * 2] + list(noise_shape[1:]), cond_shapes=cond_shapes) + 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): executor = patcher_extension.WrapperExecutor.new_executor( _prepare_sampling, @@ -118,14 +140,14 @@ def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None ) return executor.execute(model, noise_shape, conds, model_options=model_options) + def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None): 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 = model.memory_required([noise_shape[0] * 2] + list(noise_shape[1:])) + inference_memory - minimum_memory_required = model.memory_required([noise_shape[0]] + list(noise_shape[1:])) + inference_memory - model_management.load_models_gpu([model] + models, memory_required=memory_required, minimum_memory_required=minimum_memory_required) + memory_required, minimum_memory_required = estimate_memory(model, noise_shape, conds) + model_management.load_models_gpu([model] + models, memory_required=memory_required + inference_memory, minimum_memory_required=minimum_memory_required + inference_memory) real_model = model.model return real_model, conds, models diff --git a/comfy/sampler_names.py b/comfy/sampler_names.py index ec3b0a649..302bae799 100644 --- a/comfy/sampler_names.py +++ b/comfy/sampler_names.py @@ -2,7 +2,7 @@ KSAMPLER_NAMES = ["euler", "euler_cfg_pp", "euler_ancestral", "euler_ancestral_c "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_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", - "gradient_estimation", "er_sde", "seeds_2", "seeds_3"] + "gradient_estimation", "gradient_estimation_cfg_pp", "er_sde", "seeds_2", "seeds_3"] SCHEDULER_NAMES = ["normal", "karras", "exponential", "sgm_uniform", "simple", "ddim_uniform", "beta", "linear_quadratic", "kl_optimal"] SAMPLER_NAMES = KSAMPLER_NAMES + ["ddim", "uni_pc", "uni_pc_bh2"] diff --git a/comfy/samplers.py b/comfy/samplers.py index 98a211f69..2ce059857 100644 --- a/comfy/samplers.py +++ b/comfy/samplers.py @@ -269,7 +269,13 @@ def _calc_cond_batch(model: BaseModel, conds, x_in: torch.Tensor, timestep: torc for i in range(1, len(to_batch_temp) + 1): batch_amount = to_batch_temp[:len(to_batch_temp) // i] input_shape = [len(batch_amount) * first_shape[0]] + list(first_shape)[1:] - if model.memory_required(input_shape) * 1.5 < free_memory: + cond_shapes = collections.defaultdict(list) + for tt in batch_amount: + cond = {k: v.size() for k, v in to_run[tt][0].conditioning.items()} + for k, v in to_run[tt][0].conditioning.items(): + cond_shapes[k].append(v.size()) + + if model.memory_required(input_shape, cond_shapes=cond_shapes) * 1.5 < free_memory: to_batch = batch_amount break diff --git a/comfy/sd.py b/comfy/sd.py index e9d6dcfa0..0f4f5c8a4 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -23,6 +23,7 @@ from . import sd1_clip from . import sdxl_clip from . import utils from .hooks import EnumHookMode +from .ldm.ace.vae.music_dcae_pipeline import MusicDCAE from .ldm.audio.autoencoder import AudioOobleckVAE from .ldm.cascade.stage_a import StageA from .ldm.cascade.stage_c_coder import StageC_coder @@ -52,6 +53,7 @@ from .text_encoders import sa_t5 from .text_encoders import sd2_clip from .text_encoders import sd3_clip from .text_encoders import wan +from .text_encoders import ace from .utils import ProgressBar logger = logging.getLogger(__name__) @@ -130,6 +132,7 @@ class CLIP: self.layer_idx = None self.use_clip_schedule = False logger.debug("CLIP/text encoder model load device: {}, offload device: {}, current: {}, dtype: {}".format(load_device, offload_device, params['device'], dtype)) + self.tokenizer_options = {} def clone(self): n = CLIP(no_init=True) @@ -138,6 +141,7 @@ class CLIP: # cloning the tokenizer allows the vocab updates to work more idiomatically n.tokenizer = self.tokenizer.clone() n.layer_idx = self.layer_idx + n.tokenizer_options = self.tokenizer_options.copy() n.use_clip_schedule = self.use_clip_schedule n.apply_hooks_to_conds = self.apply_hooks_to_conds return n @@ -145,10 +149,18 @@ class CLIP: def add_patches(self, patches, strength_patch=1.0, strength_model=1.0): return self.patcher.add_patches(patches, strength_patch, strength_model) + def set_tokenizer_option(self, option_name, value): + self.tokenizer_options[option_name] = value + def clip_layer(self, layer_idx): self.layer_idx = layer_idx def tokenize(self, text, return_word_ids=False, **kwargs): + tokenizer_options = kwargs.get("tokenizer_options", {}) + if len(self.tokenizer_options) > 0: + tokenizer_options = {**self.tokenizer_options, **tokenizer_options} + if len(tokenizer_options) > 0: + kwargs["tokenizer_options"] = tokenizer_options return self.tokenizer.tokenize_with_weights(text, return_word_ids, **kwargs) def add_hooks_to_dict(self, pooled_dict: dict[str]): @@ -282,6 +294,7 @@ class VAE: self.downscale_index_formula = None self.upscale_index_formula = None + self.extra_1d_channel = None if config is None: if "decoder.mid.block_1.mix_factor" in sd: @@ -439,6 +452,20 @@ class VAE: ddconfig = {"embed_dim": 64, "num_freqs": 8, "include_pi": False, "heads": 16, "width": 1024, "num_decoder_layers": 16, "qkv_bias": False, "qk_norm": True, "geo_decoder_mlp_expand_ratio": mlp_expand, "geo_decoder_downsample_ratio": downsample_ratio, "geo_decoder_ln_post": ln_post} self.first_stage_model = ShapeVAE(**ddconfig) self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32] + elif "vocoder.backbone.channel_layers.0.0.bias" in sd: #Ace Step Audio + self.first_stage_model = MusicDCAE(source_sample_rate=44100) + self.memory_used_encode = lambda shape, dtype: (shape[2] * 330) * model_management.dtype_size(dtype) + 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.upscale_ratio = 4096 + self.downscale_ratio = 4096 + self.latent_dim = 2 + self.process_output = lambda audio: audio + self.process_input = lambda audio: audio + self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32] + self.disable_offload = True + self.extra_1d_channel = 16 else: logger.warning("WARNING: No VAE weights detected, VAE not initalized.") self.first_stage_model = None @@ -497,7 +524,13 @@ class VAE: return output def decode_tiled_1d(self, samples, tile_x=128, overlap=32): - decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).float() + if samples.ndim == 3: + decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).float() + else: + og_shape = samples.shape + samples = samples.reshape((og_shape[0], og_shape[1] * og_shape[2], -1)) + decode_fn = lambda a: self.first_stage_model.decode(a.reshape((-1, og_shape[1], og_shape[2], a.shape[-1])).to(self.vae_dtype).to(self.device)).float() + return self.process_output(utils.tiled_scale_multidim(samples, decode_fn, tile=(tile_x,), overlap=overlap, upscale_amount=self.upscale_ratio, out_channels=self.output_channels, output_device=self.output_device)) def decode_tiled_3d(self, samples, tile_t=999, tile_x=32, tile_y=32, overlap=(1, 8, 8)): @@ -517,9 +550,24 @@ class VAE: samples /= 3.0 return samples - def encode_tiled_1d(self, samples, tile_x=128 * 2048, overlap=32 * 2048): - encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).float() - return utils.tiled_scale_multidim(samples, encode_fn, tile=(tile_x,), overlap=overlap, upscale_amount=(1 / self.downscale_ratio), out_channels=self.latent_channels, output_device=self.output_device) + def encode_tiled_1d(self, samples, tile_x=256 * 2048, overlap=64 * 2048): + if self.latent_dim == 1: + encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).float() + out_channels = self.latent_channels + upscale_amount = 1 / self.downscale_ratio + else: + extra_channel_size = self.extra_1d_channel + out_channels = self.latent_channels * extra_channel_size + tile_x = tile_x // extra_channel_size + overlap = overlap // extra_channel_size + upscale_amount = 1 / self.downscale_ratio + encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).reshape(1, out_channels, -1).float() + + out = utils.tiled_scale_multidim(samples, encode_fn, tile=(tile_x,), overlap=overlap, upscale_amount=upscale_amount, out_channels=out_channels, output_device=self.output_device) + if self.latent_dim == 1: + return out + else: + return out.reshape(samples.shape[0], self.latent_channels, extra_channel_size, -1) def encode_tiled_3d(self, samples, tile_t=9999, tile_x=512, tile_y=512, overlap=(1, 64, 64)): encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).float() @@ -544,7 +592,7 @@ class VAE: except model_management.OOM_EXCEPTION: logger.warning("Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.") dims = samples_in.ndim - 2 - if dims == 1: + if dims == 1 or self.extra_1d_channel is not None: pixel_samples = self.decode_tiled_1d(samples_in) elif dims == 2: pixel_samples = self.decode_tiled_(samples_in) @@ -613,7 +661,7 @@ class VAE: tile = 256 overlap = tile // 4 samples = self.encode_tiled_3d(pixel_samples, tile_x=tile, tile_y=tile, overlap=(1, overlap, overlap)) - elif self.latent_dim == 1: + elif self.latent_dim == 1 or self.extra_1d_channel is not None: samples = self.encode_tiled_1d(pixel_samples) else: samples = self.encode_tiled_(pixel_samples) @@ -723,6 +771,8 @@ class CLIPType(Enum): LUMINA2 = 12 WAN = 13 HIDREAM = 14 + CHROMA = 15 + ACE = 16 @dataclasses.dataclass @@ -840,7 +890,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip elif clip_type == CLIPType.LTXV: clip_target.clip = lt.ltxv_te(**t5xxl_detect(clip_data)) clip_target.tokenizer = lt.LTXVT5Tokenizer - elif clip_type == CLIPType.PIXART: + elif clip_type == CLIPType.PIXART or clip_type == CLIPType.CHROMA: clip_target.clip = pixart_t5.pixart_te(**t5xxl_detect(clip_data)) clip_target.tokenizer = pixart_t5.PixArtTokenizer elif clip_type == CLIPType.WAN: @@ -861,8 +911,13 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip clip_target.clip = aura_t5.AuraT5Model clip_target.tokenizer = aura_t5.AuraT5Tokenizer elif te_model == TEModel.T5_BASE: - clip_target.clip = sa_t5.SAT5Model - clip_target.tokenizer = sa_t5.SAT5Tokenizer + if clip_type == CLIPType.ACE or "spiece_model" in clip_data[0]: + clip_target.clip = ace.AceT5Model + clip_target.tokenizer = ace.AceT5Tokenizer + tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None) + else: + clip_target.clip = sa_t5.SAT5Model + clip_target.tokenizer = sa_t5.SAT5Tokenizer elif te_model == TEModel.GEMMA_2_2B: clip_target.clip = lumina2.te(**llama_detect(clip_data)) clip_target.tokenizer = lumina2.LuminaTokenizer diff --git a/comfy/sd1_clip.py b/comfy/sd1_clip.py index c88c36ba8..c6430a7fd 100644 --- a/comfy/sd1_clip.py +++ b/comfy/sd1_clip.py @@ -537,7 +537,7 @@ SDTokenizerT = TypeVar('SDTokenizerT', bound='SDTokenizer') class SDTokenizer: - def __init__(self, tokenizer_path: torch.Tensor | bytes | bytearray | memoryview | str | Path | Traversable = 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, tokenizer_data=None, tokenizer_args=None): + def __init__(self, tokenizer_path: torch.Tensor | bytes | bytearray | memoryview | str | Path | Traversable = 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, tokenizer_data=None, tokenizer_args=None): if tokenizer_data is None: tokenizer_data = dict() if tokenizer_args is None: @@ -558,6 +558,7 @@ class SDTokenizer: self.max_length = tokenizer_data.get("{}_max_length".format(embedding_key), max_length) self.min_length = min_length self.end_token = None + self.min_padding = min_padding empty = self.tokenizer('')["input_ids"] self.tokenizer_adds_end_token = has_end_token @@ -624,13 +625,15 @@ class SDTokenizer: return (embed, "{} {}".format(embedding_name[len(stripped):], leftover)) return (embed, leftover) - def tokenize_with_weights(self, text: str, return_word_ids=False, **kwargs): + def tokenize_with_weights(self, text: str, return_word_ids=False, tokenizer_options={}, **kwargs): ''' Takes a prompt and converts it to a list of (token, weight, word id) elements. Tokens can both be integer tokens and pre computed CLIP tensors. Word id values are unique per word and embedding, where the id 0 is reserved for non word tokens. Returned list has the dimensions NxM where M is the input size of CLIP ''' + min_length = tokenizer_options.get("{}_min_length".format(self.embedding_key), self.min_length) + min_padding = tokenizer_options.get("{}_min_padding".format(self.embedding_key), self.min_padding) text = escape_important(text) parsed_weights = token_weights(text, 1.0) @@ -717,10 +720,12 @@ class SDTokenizer: # fill last batch if self.end_token is not None: batch.append((self.end_token, 1.0, 0)) - if self.pad_to_max_length: + if min_padding is not None: + batch.extend([(self.pad_token, 1.0, 0)] * min_padding) + if self.pad_to_max_length and len(batch) < self.max_length: batch.extend([(self.pad_token, 1.0, 0)] * (self.max_length - len(batch))) - if self.min_length is not None and len(batch) < self.min_length: - batch.extend([(self.pad_token, 1.0, 0)] * (self.min_length - len(batch))) + if min_length is not None and len(batch) < min_length: + batch.extend([(self.pad_token, 1.0, 0)] * (min_length - len(batch))) if not return_word_ids: batched_tokens = [[(t, w) for t, w, _ in x] for x in batched_tokens] @@ -752,7 +757,7 @@ class SD1Tokenizer: def tokenize_with_weights(self, text: str, return_word_ids=False, **kwargs): out = {} - out[self.clip_name] = self.sd_tokenizer.tokenize_with_weights(text, return_word_ids) + out[self.clip_name] = self.sd_tokenizer.tokenize_with_weights(text, return_word_ids, **kwargs) return out def untokenize(self, token_weight_pair): diff --git a/comfy/sdxl_clip.py b/comfy/sdxl_clip.py index b5bfe4baa..950405a2a 100644 --- a/comfy/sdxl_clip.py +++ b/comfy/sdxl_clip.py @@ -36,8 +36,8 @@ class SDXLTokenizer: def tokenize_with_weights(self, text: str, return_word_ids=False, **kwargs): out = {} - out["g"] = self.clip_g.tokenize_with_weights(text, return_word_ids) - out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids) + out["g"] = self.clip_g.tokenize_with_weights(text, return_word_ids, **kwargs) + out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids, **kwargs) return out def untokenize(self, token_weight_pair): diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 9f7ebc9e6..a845912db 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -7,6 +7,7 @@ from . import sd1_clip from . import sdxl_clip from . import supported_models_base from . import utils +from .text_encoders import ace from .text_encoders import aura_t5 from .text_encoders import cosmos from .text_encoders import flux @@ -535,6 +536,7 @@ class SDXL_instructpix2pix(SDXL): def get_model(self, state_dict, prefix="", device=None): return model_base.SDXL_instructpix2pix(self, model_type=self.model_type(state_dict, prefix), device=device) + class LotusD(SD20): unet_config = { "model_channels": 320, @@ -551,6 +553,7 @@ class LotusD(SD20): def get_model(self, state_dict, prefix="", device=None): return model_base.Lotus(self, device=device) + class SD3(supported_models_base.BASE): unet_config = { "in_channels": 16, @@ -839,6 +842,10 @@ class LTXV(supported_models_base.BASE): vae_key_prefix = ["vae."] text_encoder_key_prefix = ["text_encoders."] + def __init__(self, unet_config): + super().__init__(unet_config) + self.memory_usage_factor = (unet_config.get("cross_attention_dim", 2048) / 2048) * 5.5 + def get_model(self, state_dict, prefix="", device=None): out = model_base.LTXV(self, device=device) return out @@ -912,6 +919,7 @@ class HunyuanVideoI2V(HunyuanVideo): out = model_base.HunyuanVideoI2V(self, device=device) return out + class HunyuanVideoSkyreelsI2V(HunyuanVideo): unet_config = { "image_model": "hunyuan_video", @@ -1059,16 +1067,34 @@ class WAN21_FunControl2V(WAN21_T2V): out = model_base.WAN21(self, image_to_video=False, device=device) return out + +class WAN21_Camera(WAN21_T2V): + unet_config = { + "image_model": "wan2.1", + "model_type": "camera", + "in_dim": 32, + } + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.WAN21_Camera(self, image_to_video=False, device=device) + return out + + class WAN21_Vace(WAN21_T2V): unet_config = { "image_model": "wan2.1", "model_type": "vace", } + def __init__(self, unet_config): + super().__init__(unet_config) + self.memory_usage_factor = 1.2 * self.memory_usage_factor + def get_model(self, state_dict, prefix="", device=None): out = model_base.WAN21_Vace(self, image_to_video=False, device=device) return out + class Hunyuan3Dv2(supported_models_base.BASE): unet_config = { "image_model": "hunyuan3d2", @@ -1096,9 +1122,12 @@ class Hunyuan3Dv2(supported_models_base.BASE): out = model_base.Hunyuan3Dv2(self, device=device) return out - def clip_target(self, state_dict={}): + def clip_target(self, state_dict=None): + if state_dict is None: + state_dict = {} return None + class Hunyuan3Dv2mini(Hunyuan3Dv2): unet_config = { "image_model": "hunyuan3d2", @@ -1107,6 +1136,7 @@ class Hunyuan3Dv2mini(Hunyuan3Dv2): latent_format = latent_formats.Hunyuan3Dv2mini + class HiDream(supported_models_base.BASE): unet_config = { "image_model": "hidream", @@ -1133,10 +1163,73 @@ class HiDream(supported_models_base.BASE): out = model_base.HiDream(self, device=device) return out - def clip_target(self, state_dict={}): - return None # TODO + def clip_target(self, state_dict=None): + if state_dict is None: + state_dict = {} + return None # TODO -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, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, Lumina2, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream] +class Chroma(supported_models_base.BASE): + unet_config = { + "image_model": "chroma", + } + + unet_extra_config = { + } + + sampling_settings = { + "multiplier": 1.0, + } + + latent_format = latent_formats.Flux + + memory_usage_factor = 3.2 + + supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32] + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.Chroma(self, device=device) + return out + + def clip_target(self, state_dict=None): + if state_dict is None: + state_dict = {} + pref = self.text_encoder_key_prefix[0] + t5_detect = sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref)) + return supported_models_base.ClipTarget(pixart_t5.PixArtTokenizer, pixart_t5.pixart_te(**t5_detect)) + + +class ACEStep(supported_models_base.BASE): + unet_config = { + "audio_model": "ace", + } + + unet_extra_config = { + } + + sampling_settings = { + "shift": 3.0, + } + + latent_format = latent_formats.ACEAudio + + memory_usage_factor = 0.5 + + 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.ACEStep(self, device=device) + return out + + def clip_target(self, state_dict=None): + if state_dict is None: + state_dict = {} + return supported_models_base.ClipTarget(ace.AceT5Tokenizer, ace.AceT5Model) + + +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, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, Lumina2, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream, Chroma, ACEStep] models += [SVD_img2vid] diff --git a/comfy/text_encoders/ace.py b/comfy/text_encoders/ace.py new file mode 100644 index 000000000..d650bb10d --- /dev/null +++ b/comfy/text_encoders/ace.py @@ -0,0 +1,153 @@ +from comfy import sd1_clip +from .spiece_tokenizer import SPieceTokenizer +import comfy.text_encoders.t5 +import os +import re +import torch +import logging + +from tokenizers import Tokenizer +from .ace_text_cleaners import multilingual_cleaners, japanese_to_romaji + +SUPPORT_LANGUAGES = { + "en": 259, "de": 260, "fr": 262, "es": 284, "it": 285, + "pt": 286, "pl": 294, "tr": 295, "ru": 267, "cs": 293, + "nl": 297, "ar": 5022, "zh": 5023, "ja": 5412, "hu": 5753, + "ko": 6152, "hi": 6680 +} + +structure_pattern = re.compile(r"\[.*?\]") + +DEFAULT_VOCAB_FILE = os.path.join(os.path.join(os.path.dirname(os.path.realpath(__file__)), "ace_lyrics_tokenizer"), "vocab.json") + + +class VoiceBpeTokenizer: + def __init__(self, vocab_file=DEFAULT_VOCAB_FILE): + self.tokenizer = None + if vocab_file is not None: + self.tokenizer = Tokenizer.from_file(vocab_file) + + def preprocess_text(self, txt, lang): + txt = multilingual_cleaners(txt, lang) + return txt + + def encode(self, txt, lang='en'): + # lang = lang.split("-")[0] # remove the region + # self.check_input_length(txt, lang) + txt = self.preprocess_text(txt, lang) + lang = "zh-cn" if lang == "zh" else lang + txt = f"[{lang}]{txt}" + txt = txt.replace(" ", "[SPACE]") + return self.tokenizer.encode(txt).ids + + def get_lang(self, line): + if line.startswith("[") and line[3:4] == ']': + lang = line[1:3].lower() + if lang in SUPPORT_LANGUAGES: + return lang, line[4:] + return "en", line + + def __call__(self, string): + lines = string.split("\n") + lyric_token_idx = [261] + for line in lines: + line = line.strip() + if not line: + lyric_token_idx += [2] + continue + + lang, line = self.get_lang(line) + + if lang not in SUPPORT_LANGUAGES: + lang = "en" + if "zh" in lang: + lang = "zh" + if "spa" in lang: + lang = "es" + + try: + line_out = japanese_to_romaji(line) + if line_out != line: + lang = "ja" + line = line_out + except: + pass + + try: + if structure_pattern.match(line): + token_idx = self.encode(line, "en") + else: + token_idx = self.encode(line, lang) + lyric_token_idx = lyric_token_idx + token_idx + [2] + except Exception as e: + logging.warning("tokenize error {} for line {} major_language {}".format(e, line, lang)) + return {"input_ids": lyric_token_idx} + + @staticmethod + def from_pretrained(path, **kwargs): + return VoiceBpeTokenizer(path, **kwargs) + + def get_vocab(self): + return {} + + +class UMT5BaseModel(sd1_clip.SDClipModel): + def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}): + textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "umt5_config_base.json") + 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=True, zero_out_masked=False, model_options=model_options) + +class UMT5BaseTokenizer(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=768, embedding_key='umt5base', tokenizer_class=SPieceTokenizer, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=0, tokenizer_data=tokenizer_data) + + def state_dict(self): + return {"spiece_model": self.tokenizer.serialize_model()} + +class LyricsTokenizer(sd1_clip.SDTokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + tokenizer = os.path.join(os.path.join(os.path.dirname(os.path.realpath(__file__)), "ace_lyrics_tokenizer"), "vocab.json") + super().__init__(tokenizer, pad_with_end=False, embedding_size=1024, embedding_key='lyrics', tokenizer_class=VoiceBpeTokenizer, has_start_token=True, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=2, has_end_token=False, tokenizer_data=tokenizer_data) + +class AceT5Tokenizer: + def __init__(self, embedding_directory=None, tokenizer_data={}): + self.voicebpe = LyricsTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data) + self.umt5base = UMT5BaseTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data) + + def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs): + out = {} + out["lyrics"] = self.voicebpe.tokenize_with_weights(kwargs.get("lyrics", ""), return_word_ids, **kwargs) + out["umt5base"] = self.umt5base.tokenize_with_weights(text, return_word_ids, **kwargs) + return out + + def untokenize(self, token_weight_pair): + return self.umt5base.untokenize(token_weight_pair) + + def state_dict(self): + return self.umt5base.state_dict() + +class AceT5Model(torch.nn.Module): + def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs): + super().__init__() + self.umt5base = UMT5BaseModel(device=device, dtype=dtype, model_options=model_options) + self.dtypes = set() + if dtype is not None: + self.dtypes.add(dtype) + + def set_clip_options(self, options): + self.umt5base.set_clip_options(options) + + def reset_clip_options(self): + self.umt5base.reset_clip_options() + + def encode_token_weights(self, token_weight_pairs): + token_weight_pairs_umt5base = token_weight_pairs["umt5base"] + token_weight_pairs_lyrics = token_weight_pairs["lyrics"] + + t5_out, t5_pooled = self.umt5base.encode_token_weights(token_weight_pairs_umt5base) + + lyrics_embeds = torch.tensor(list(map(lambda a: a[0], token_weight_pairs_lyrics[0]))).unsqueeze(0) + return t5_out, None, {"conditioning_lyrics": lyrics_embeds} + + def load_sd(self, sd): + return self.umt5base.load_sd(sd) diff --git a/comfy/text_encoders/ace_lyrics_tokenizer/vocab.json b/comfy/text_encoders/ace_lyrics_tokenizer/vocab.json new file mode 100644 index 000000000..519ed340c --- /dev/null +++ b/comfy/text_encoders/ace_lyrics_tokenizer/vocab.json @@ -0,0 +1,15535 @@ +{ + "version": "1.0", + "truncation": null, + "padding": null, + "added_tokens": [ + { + "id": 0, + "special": true, + "content": "[STOP]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 1, + "special": true, + "content": "[UNK]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 2, + "special": true, + "content": "[SPACE]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 259, + "special": true, + "content": "[en]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 260, + "special": true, + "content": "[de]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 261, + "special": true, + "content": "[START]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 262, + "special": true, + "content": "[fr]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 284, + "special": true, + "content": "[es]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 285, + "special": true, + "content": "[it]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 286, + "special": true, + "content": "[pt]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 294, + "special": true, + "content": "[pl]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 295, + "special": true, + "content": "[tr]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 267, + "special": true, + "content": "[ru]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 293, + "special": true, + "content": "[cs]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 297, + "special": true, + "content": "[nl]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 5022, + "special": true, + "content": "[ar]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 5023, + "special": true, + "content": "[zh-cn]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 5412, + "special": true, + "content": "[ja]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 5753, + "special": true, + "content": "[hu]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 6152, + "special": true, + "content": "[ko]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 6680, + "special": true, + "content": "[hi]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 6681, + "special": true, + "content": "[start]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 6682, + "special": true, + "content": "[intro]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 6683, + "special": true, + "content": "[verse]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 6684, + "special": true, + "content": "[chorus]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 6685, + "special": true, + "content": "[bridge]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 6686, + "special": true, + "content": "[outro]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 6687, + "special": true, + "content": "[end]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 6688, + "special": true, + "content": "[inst]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 6689, + "special": true, + "content": "[solo]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 6690, + "special": true, + "content": "[hook]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 6691, + "special": true, + "content": "[pre-chorus]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + }, + { + "id": 6692, + "special": true, + "content": "[break]", + "single_word": false, + "lstrip": false, + "rstrip": false, + "normalized": false + } + ], + "normalizer": null, + "pre_tokenizer": { + "type": "Whitespace" + }, + "post_processor": null, + "decoder": null, + "model": { + "type": "BPE", + "dropout": null, + "unk_token": "[UNK]", + "continuing_subword_prefix": null, + "end_of_word_suffix": null, + "fuse_unk": false, + "vocab": { + "[STOP]": 0, + "[UNK]": 1, + "[SPACE]": 2, + "!": 3, + "'": 4, + "(": 5, + ")": 6, + ",": 7, + "-": 8, + ".": 9, + "/": 10, + ":": 11, + ";": 12, + "?": 13, + "a": 14, + "b": 15, + "c": 16, + "d": 17, + "e": 18, + "f": 19, + "g": 20, + "h": 21, + "i": 22, + "j": 23, + "k": 24, + "l": 25, + "m": 26, + "n": 27, + "o": 28, + "p": 29, + "q": 30, + "r": 31, + "s": 32, + "t": 33, + "u": 34, + "v": 35, + "w": 36, + "x": 37, + "y": 38, + "z": 39, + "th": 40, + "in": 41, + "the": 42, + "an": 43, + "er": 44, + "ou": 45, + "re": 46, + "on": 47, + "at": 48, + "ed": 49, + "en": 50, + "to": 51, + "ing": 52, + "and": 53, + "is": 54, + "as": 55, + "al": 56, + "or": 57, + "of": 58, + "ar": 59, + "it": 60, + "es": 61, + "he": 62, + "st": 63, + "le": 64, + "om": 65, + "se": 66, + "be": 67, + "ad": 68, + "ow": 69, + "ly": 70, + "ch": 71, + "wh": 72, + "that": 73, + "you": 74, + "li": 75, + "ve": 76, + "ac": 77, + "ti": 78, + "ld": 79, + "me": 80, + "was": 81, + "gh": 82, + "id": 83, + "ll": 84, + "wi": 85, + "ent": 86, + "for": 87, + "ay": 88, + "ro": 89, + "ver": 90, + "ic": 91, + "her": 92, + "ke": 93, + "his": 94, + "no": 95, + "ut": 96, + "un": 97, + "ir": 98, + "lo": 99, + "we": 100, + "ri": 101, + "ha": 102, + "with": 103, + "ght": 104, + "out": 105, + "im": 106, + "ion": 107, + "all": 108, + "ab": 109, + "one": 110, + "ne": 111, + "ge": 112, + "ould": 113, + "ter": 114, + "mo": 115, + "had": 116, + "ce": 117, + "she": 118, + "go": 119, + "sh": 120, + "ur": 121, + "am": 122, + "so": 123, + "pe": 124, + "my": 125, + "de": 126, + "are": 127, + "but": 128, + "ome": 129, + "fr": 130, + "ther": 131, + "fe": 132, + "su": 133, + "do": 134, + "con": 135, + "te": 136, + "ain": 137, + "ere": 138, + "po": 139, + "if": 140, + "they": 141, + "us": 142, + "ag": 143, + "tr": 144, + "now": 145, + "oun": 146, + "this": 147, + "have": 148, + "not": 149, + "sa": 150, + "il": 151, + "up": 152, + "thing": 153, + "from": 154, + "ap": 155, + "him": 156, + "ack": 157, + "ation": 158, + "ant": 159, + "our": 160, + "op": 161, + "like": 162, + "ust": 163, + "ess": 164, + "bo": 165, + "ok": 166, + "ul": 167, + "ind": 168, + "ex": 169, + "com": 170, + "some": 171, + "there": 172, + "ers": 173, + "co": 174, + "res": 175, + "man": 176, + "ard": 177, + "pl": 178, + "wor": 179, + "way": 180, + "tion": 181, + "fo": 182, + "ca": 183, + "were": 184, + "by": 185, + "ate": 186, + "pro": 187, + "ted": 188, + "ound": 189, + "own": 190, + "would": 191, + "ts": 192, + "what": 193, + "qu": 194, + "ally": 195, + "ight": 196, + "ck": 197, + "gr": 198, + "when": 199, + "ven": 200, + "can": 201, + "ough": 202, + "ine": 203, + "end": 204, + "per": 205, + "ous": 206, + "od": 207, + "ide": 208, + "know": 209, + "ty": 210, + "very": 211, + "si": 212, + "ak": 213, + "who": 214, + "about": 215, + "ill": 216, + "them": 217, + "est": 218, + "red": 219, + "ye": 220, + "could": 221, + "ong": 222, + "your": 223, + "their": 224, + "em": 225, + "just": 226, + "other": 227, + "into": 228, + "any": 229, + "whi": 230, + "um": 231, + "tw": 232, + "ast": 233, + "der": 234, + "did": 235, + "ie": 236, + "been": 237, + "ace": 238, + "ink": 239, + "ity": 240, + "back": 241, + "ting": 242, + "br": 243, + "more": 244, + "ake": 245, + "pp": 246, + "then": 247, + "sp": 248, + "el": 249, + "use": 250, + "bl": 251, + "said": 252, + "over": 253, + "get": 254, + "ß": 255, + "ä": 256, + "ö": 257, + "ü": 258, + "[en]": 259, + "[de]": 260, + "[START]": 261, + "[fr]": 262, + "œ": 263, + "ï": 264, + "ê": 265, + "â": 266, + "[ru]": 267, + "ÿ": 268, + "è": 269, + "à": 270, + "ë": 271, + "ù": 272, + "î": 273, + "ç": 274, + "æ": 275, + "ô": 276, + "û": 277, + "á": 278, + "é": 279, + "í": 280, + "ó": 281, + "ú": 282, + "ñ": 283, + "[es]": 284, + "[it]": 285, + "[pt]": 286, + "ń": 287, + "ś": 288, + "ę": 289, + "ą": 290, + "ż": 291, + "ć": 292, + "[cs]": 293, + "[pl]": 294, + "[tr]": 295, + "ã": 296, + "[nl]": 297, + "ş": 298, + "ğ": 299, + "ı": 300, + "ò": 301, + "ì": 302, + "¿": 303, + "…": 304, + "i̇": 305, + "õ": 306, + "\"": 307, + "´": 308, + "ø": 309, + "č": 310, + "ō": 311, + "š": 312, + "ž": 313, + "̇": 314, + "ei": 315, + "ich": 316, + "ein": 317, + "au": 318, + "sch": 319, + "und": 320, + "die": 321, + "da": 322, + "den": 323, + "gen": 324, + "zu": 325, + "hr": 326, + "ten": 327, + "mi": 328, + "sie": 329, + "das": 330, + "eine": 331, + "icht": 332, + "ber": 333, + "ach": 334, + "auf": 335, + "lich": 336, + "nicht": 337, + "mm": 338, + "ben": 339, + "war": 340, + "mit": 341, + "sich": 342, + "ig": 343, + "aus": 344, + "ist": 345, + "wie": 346, + "och": 347, + "ung": 348, + "ann": 349, + "ür": 350, + "hn": 351, + "ihr": 352, + "sen": 353, + "tz": 354, + "dem": 355, + "eit": 356, + "hat": 357, + "wir": 358, + "von": 359, + "wei": 360, + "ier": 361, + "ra": 362, + "einen": 363, + "vor": 364, + "als": 365, + "wo": 366, + "rei": 367, + "ste": 368, + "lie": 369, + "auch": 370, + "du": 371, + "des": 372, + "ko": 373, + "über": 374, + "bei": 375, + "hen": 376, + "hm": 377, + "lei": 378, + "aber": 379, + "wen": 380, + "hl": 381, + "ger": 382, + "nach": 383, + "ft": 384, + "imm": 385, + "je": 386, + "schen": 387, + "wer": 388, + "ser": 389, + "än": 390, + "sein": 391, + "ol": 392, + "cht": 393, + "für": 394, + "kl": 395, + "ff": 396, + "einem": 397, + "nen": 398, + "ja": 399, + "noch": 400, + "hatte": 401, + "pf": 402, + "hin": 403, + "di": 404, + "chen": 405, + "rü": 406, + "iel": 407, + "sel": 408, + "dass": 409, + "ihn": 410, + "mir": 411, + "schl": 412, + "ön": 413, + "gan": 414, + "gt": 415, + "einer": 416, + "sten": 417, + "mich": 418, + "wenn": 419, + "ell": 420, + "gte": 421, + "mal": 422, + "gel": 423, + "ken": 424, + "nur": 425, + "mmen": 426, + "fü": 427, + "ern": 428, + "ör": 429, + "unter": 430, + "ander": 431, + "dur": 432, + "uch": 433, + "ta": 434, + "men": 435, + "mach": 436, + "doch": 437, + "durch": 438, + "os": 439, + "gl": 440, + "hal": 441, + "ihre": 442, + "wä": 443, + "immer": 444, + "ihm": 445, + "kann": 446, + "ort": 447, + "dann": 448, + "lan": 449, + "tzt": 450, + "oder": 451, + "hren": 452, + "et": 453, + "kön": 454, + "ick": 455, + "fa": 456, + "wieder": 457, + "daß": 458, + "mein": 459, + "fen": 460, + "ganz": 461, + "diese": 462, + "ster": 463, + "dar": 464, + "wa": 465, + "ges": 466, + "na": 467, + "fl": 468, + "igen": 469, + "sche": 470, + "ungen": 471, + "mehr": 472, + "ßen": 473, + "ot": 474, + "kon": 475, + "gew": 476, + "haben": 477, + "geh": 478, + "ät": 479, + "sind": 480, + "dr": 481, + "wel": 482, + "uns": 483, + "vo": 484, + "ma": 485, + "ute": 486, + "schon": 487, + "bes": 488, + "gesch": 489, + "bt": 490, + "che": 491, + "son": 492, + "ob": 493, + "la": 494, + "rück": 495, + "seine": 496, + "kr": 497, + "fre": 498, + "eil": 499, + "zum": 500, + "hier": 501, + "kt": 502, + "ige": 503, + "spr": 504, + "leben": 505, + "bst": 506, + "zeit": 507, + "gro": 508, + "denn": 509, + "ho": 510, + "scha": 511, + "bar": 512, + "alle": 513, + "gegen": 514, + "wür": 515, + "mü": 516, + "ze": 517, + "werden": 518, + "jetzt": 519, + "kommen": 520, + "nie": 521, + "sei": 522, + "heit": 523, + "soll": 524, + "glei": 525, + "meine": 526, + "woll": 527, + "ner": 528, + "habe": 529, + "wur": 530, + "lichen": 531, + "assen": 532, + "nte": 533, + "sehen": 534, + "wird": 535, + "bis": 536, + "gar": 537, + "ien": 538, + "mus": 539, + "uß": 540, + "är": 541, + "stell": 542, + "keit": 543, + "zwei": 544, + "selbst": 545, + "sta": 546, + "pa": 547, + "sagte": 548, + "tet": 549, + "kam": 550, + "ssen": 551, + "viel": 552, + "ug": 553, + "zen": 554, + "hei": 555, + "mann": 556, + "will": 557, + "geb": 558, + "waren": 559, + "ück": 560, + "äch": 561, + "mer": 562, + "ru": 563, + "hau": 564, + "eigen": 565, + "ang": 566, + "weg": 567, + "blick": 568, + "fra": 569, + "alles": 570, + "ka": 571, + "augen": 572, + "fin": 573, + "liche": 574, + "unser": 575, + "dern": 576, + "herr": 577, + "nun": 578, + "vie": 579, + "chte": 580, + "wohl": 581, + "fall": 582, + "ht": 583, + "ün": 584, + "etwas": 585, + "stand": 586, + "äu": 587, + "mö": 588, + "tel": 589, + "rie": 590, + "dich": 591, + "dies": 592, + "hand": 593, + "bin": 594, + "ffen": 595, + "nichts": 596, + "dan": 597, + "hne": 598, + "ihnen": 599, + "esen": 600, + "dieser": 601, + "frau": 602, + "art": 603, + "dir": 604, + "isch": 605, + "erst": 606, + "gleich": 607, + "komm": 608, + "hör": 609, + "ße": 610, + "dig": 611, + "sehr": 612, + "zei": 613, + "sam": 614, + "aum": 615, + "hät": 616, + "ingen": 617, + "gut": 618, + "mut": 619, + "cken": 620, + "konnte": 621, + "stimm": 622, + "zur": 623, + "itz": 624, + "weil": 625, + "würde": 626, + "fä": 627, + "können": 628, + "keine": 629, + "fer": 630, + "ischen": 631, + "voll": 632, + "eines": 633, + "setz": 634, + "zie": 635, + "del": 636, + "tete": 637, + "seiner": 638, + "ieren": 639, + "gest": 640, + "zurück": 641, + "wurde": 642, + "schn": 643, + "pr": 644, + "ließ": 645, + "tra": 646, + "mä": 647, + "gend": 648, + "fol": 649, + "ik": 650, + "schla": 651, + "schaft": 652, + "ater": 653, + "weiß": 654, + "seinen": 655, + "lassen": 656, + "lu": 657, + "unden": 658, + "teil": 659, + "neu": 660, + "iert": 661, + "menschen": 662, + "hmen": 663, + "str": 664, + "gi": 665, + "sah": 666, + "ihren": 667, + "eln": 668, + "weiter": 669, + "gehen": 670, + "iger": 671, + "macht": 672, + "tag": 673, + "also": 674, + "halten": 675, + "nis": 676, + "acht": 677, + "geben": 678, + "og": 679, + "nat": 680, + "mar": 681, + "det": 682, + "ohne": 683, + "haus": 684, + "tro": 685, + "ange": 686, + "lau": 687, + "spiel": 688, + "tre": 689, + "schr": 690, + "inn": 691, + "los": 692, + "machen": 693, + "hätte": 694, + "beg": 695, + "wirk": 696, + "alt": 697, + "glich": 698, + "tes": 699, + "richt": 700, + "freund": 701, + "ihrer": 702, + "fel": 703, + "bel": 704, + "sol": 705, + "einmal": 706, + "eben": 707, + "hol": 708, + "hän": 709, + "tern": 710, + "hö": 711, + "schw": 712, + "recht": 713, + "wahr": 714, + "seinem": 715, + "stehen": 716, + "hlen": 717, + "ins": 718, + "ging": 719, + "wollte": 720, + "wissen": 721, + "ungs": 722, + "ald": 723, + "ass": 724, + "jahr": 725, + "mor": 726, + "welt": 727, + "under": 728, + "zusa": 729, + "kopf": 730, + "lang": 731, + "hinter": 732, + "atz": 733, + "stra": 734, + "angen": 735, + "ank": 736, + "ade": 737, + "glau": 738, + "fach": 739, + "hatten": 740, + "fort": 741, + "eicht": 742, + "iff": 743, + "ler": 744, + "mei": 745, + "diesem": 746, + "kein": 747, + "frei": 748, + "führ": 749, + "vom": 750, + "β": 751, + "ai": 752, + "ait": 753, + "que": 754, + "les": 755, + "av": 756, + "ais": 757, + "oi": 758, + "eu": 759, + "lle": 760, + "par": 761, + "ans": 762, + "ment": 763, + "ét": 764, + "une": 765, + "pas": 766, + "qui": 767, + "elle": 768, + "dé": 769, + "pour": 770, + "dans": 771, + "ré": 772, + "tou": 773, + "vous": 774, + "vi": 775, + "ouv": 776, + "mon": 777, + "sur": 778, + "ci": 779, + "plu": 780, + "ère": 781, + "mais": 782, + "ois": 783, + "plus": 784, + "ée": 785, + "aient": 786, + "mp": 787, + "lui": 788, + "ave": 789, + "était": 790, + "ses": 791, + "tout": 792, + "oir": 793, + "avait": 794, + "és": 795, + "mes": 796, + "nous": 797, + "eux": 798, + "bi": 799, + "ons": 800, + "pu": 801, + "ces": 802, + "tu": 803, + "leur": 804, + "don": 805, + "eur": 806, + "ette": 807, + "aire": 808, + "avec": 809, + "dit": 810, + "té": 811, + "ille": 812, + "comme": 813, + "cr": 814, + "ux": 815, + "ès": 816, + "aux": 817, + "jour": 818, + "ils": 819, + "bien": 820, + "cou": 821, + "quel": 822, + "peu": 823, + "cette": 824, + "cu": 825, + "mê": 826, + "fait": 827, + "gu": 828, + "être": 829, + "ité": 830, + "ens": 831, + "ni": 832, + "lé": 833, + "dis": 834, + "ble": 835, + "né": 836, + "puis": 837, + "même": 838, + "ques": 839, + "fi": 840, + "age": 841, + "moi": 842, + "ence": 843, + "ont": 844, + "main": 845, + "ors": 846, + "aut": 847, + "ance": 848, + "mé": 849, + "sans": 850, + "sé": 851, + "lon": 852, + "hom": 853, + "car": 854, + "able": 855, + "cher": 856, + "deux": 857, + "enf": 858, + "où": 859, + "ph": 860, + "ure": 861, + "temp": 862, + "pos": 863, + "rent": 864, + "pé": 865, + "faire": 866, + "pi": 867, + "tres": 868, + "ça": 869, + "endre": 870, + "bon": 871, + "sou": 872, + "int": 873, + "pré": 874, + "sent": 875, + "tant": 876, + "cer": 877, + "là": 878, + "lais": 879, + "près": 880, + "bre": 881, + "cour": 882, + "pet": 883, + "comp": 884, + "lait": 885, + "trouv": 886, + "entre": 887, + "sont": 888, + "dev": 889, + "nu": 890, + "temps": 891, + "dou": 892, + "rait": 893, + "bou": 894, + "quand": 895, + "jours": 896, + "avoir": 897, + "été": 898, + "ale": 899, + "pre": 900, + "fois": 901, + "orte": 902, + "vé": 903, + "non": 904, + "tous": 905, + "jus": 906, + "coup": 907, + "homme": 908, + "ête": 909, + "aussi": 910, + "urs": 911, + "seu": 912, + "ord": 913, + "min": 914, + "gé": 915, + "core": 916, + "va": 917, + "vre": 918, + "encore": 919, + "sem": 920, + "ite": 921, + "autre": 922, + "pris": 923, + "peut": 924, + "ue": 925, + "ante": 926, + "gn": 927, + "rép": 928, + "hu": 929, + "sion": 930, + "votre": 931, + "dire": 932, + "ez": 933, + "fem": 934, + "leurs": 935, + "met": 936, + "cri": 937, + "mis": 938, + "tour": 939, + "rai": 940, + "jam": 941, + "regar": 942, + "rien": 943, + "vers": 944, + "suis": 945, + "pouv": 946, + "vis": 947, + "grand": 948, + "ants": 949, + "cor": 950, + "rer": 951, + "cé": 952, + "tent": 953, + "pres": 954, + "vou": 955, + "alors": 956, + "sieur": 957, + "aine": 958, + "quoi": 959, + "fon": 960, + "endant": 961, + "arri": 962, + "eure": 963, + "après": 964, + "donc": 965, + "itu": 966, + "lè": 967, + "sait": 968, + "toi": 969, + "cha": 970, + "ail": 971, + "asse": 972, + "imp": 973, + "voy": 974, + "conn": 975, + "pla": 976, + "petit": 977, + "avant": 978, + "nom": 979, + "tin": 980, + "dont": 981, + "sous": 982, + "emp": 983, + "person": 984, + "elles": 985, + "beau": 986, + "parti": 987, + "cho": 988, + "prit": 989, + "toujours": 990, + "rais": 991, + "jamais": 992, + "trav": 993, + "tions": 994, + "très": 995, + "voi": 996, + "ren": 997, + "yeux": 998, + "voir": 999, + "premi": 1000, + "gne": 1001, + "heure": 1002, + "rou": 1003, + "eff": 1004, + "notre": 1005, + "ments": 1006, + "ton": 1007, + "fais": 1008, + "cela": 1009, + "répon": 1010, + "cons": 1011, + "air": 1012, + "ôt": 1013, + "pendant": 1014, + "ici": 1015, + "toute": 1016, + "jet": 1017, + "port": 1018, + "étaient": 1019, + "pen": 1020, + "hé": 1021, + "autres": 1022, + "père": 1023, + "oc": 1024, + "quelques": 1025, + "ique": 1026, + "lis": 1027, + "femme": 1028, + "jou": 1029, + "teur": 1030, + "monde": 1031, + "nes": 1032, + "dre": 1033, + "aff": 1034, + "rap": 1035, + "part": 1036, + "lement": 1037, + "cla": 1038, + "fut": 1039, + "quelque": 1040, + "prendre": 1041, + "rê": 1042, + "aille": 1043, + "sais": 1044, + "ches": 1045, + "let": 1046, + "char": 1047, + "ères": 1048, + "ents": 1049, + "moins": 1050, + "eau": 1051, + "aî": 1052, + "jeu": 1053, + "heur": 1054, + "ées": 1055, + "tri": 1056, + "point": 1057, + "mom": 1058, + "vent": 1059, + "nouv": 1060, + "gran": 1061, + "trois": 1062, + "sant": 1063, + "toutes": 1064, + "contre": 1065, + "èrent": 1066, + "chez": 1067, + "avez": 1068, + "ût": 1069, + "att": 1070, + "pau": 1071, + "porte": 1072, + "ouver": 1073, + "lit": 1074, + "prés": 1075, + "chose": 1076, + "vit": 1077, + "monsieur": 1078, + "hab": 1079, + "tête": 1080, + "ju": 1081, + "tement": 1082, + "ction": 1083, + "vrai": 1084, + "lar": 1085, + "cet": 1086, + "regard": 1087, + "lant": 1088, + "som": 1089, + "moment": 1090, + "illes": 1091, + "ple": 1092, + "ps": 1093, + "mère": 1094, + "cl": 1095, + "sour": 1096, + "ys": 1097, + "trop": 1098, + "enne": 1099, + "jusqu": 1100, + "avaient": 1101, + "avais": 1102, + "jeune": 1103, + "depuis": 1104, + "personne": 1105, + "fit": 1106, + "cert": 1107, + "jo": 1108, + "oui": 1109, + "rest": 1110, + "semb": 1111, + "cap": 1112, + "mat": 1113, + "mu": 1114, + "long": 1115, + "fran": 1116, + "faut": 1117, + "iti": 1118, + "bli": 1119, + "chev": 1120, + "pri": 1121, + "ente": 1122, + "ainsi": 1123, + "cham": 1124, + "lors": 1125, + "cas": 1126, + "ili": 1127, + "bé": 1128, + "nos": 1129, + "sui": 1130, + "rit": 1131, + "cro": 1132, + "gue": 1133, + "ía": 1134, + "por": 1135, + "las": 1136, + "ón": 1137, + "una": 1138, + "aba": 1139, + "dos": 1140, + "era": 1141, + "mb": 1142, + "para": 1143, + "ás": 1144, + "mos": 1145, + "ando": 1146, + "como": 1147, + "más": 1148, + "ción": 1149, + "tan": 1150, + "dad": 1151, + "ado": 1152, + "fu": 1153, + "cia": 1154, + "mente": 1155, + "sus": 1156, + "tar": 1157, + "za": 1158, + "ba": 1159, + "pero": 1160, + "sin": 1161, + "lla": 1162, + "án": 1163, + "ia": 1164, + "ran": 1165, + "ga": 1166, + "yo": 1167, + "tos": 1168, + "cos": 1169, + "ya": 1170, + "ones": 1171, + "había": 1172, + "hi": 1173, + "esta": 1174, + "mas": 1175, + "tor": 1176, + "aban": 1177, + "dor": 1178, + "ían": 1179, + "tas": 1180, + "én": 1181, + "endo": 1182, + "aque": 1183, + "ero": 1184, + "io": 1185, + "qué": 1186, + "cab": 1187, + "tal": 1188, + "señ": 1189, + "ora": 1190, + "todo": 1191, + "sal": 1192, + "cuando": 1193, + "gun": 1194, + "bu": 1195, + "ras": 1196, + "esto": 1197, + "pare": 1198, + "él": 1199, + "tras": 1200, + "jos": 1201, + "mien": 1202, + "pue": 1203, + "cre": 1204, + "pon": 1205, + "día": 1206, + "tros": 1207, + "sab": 1208, + "sobre": 1209, + "ese": 1210, + "mbre": 1211, + "eron": 1212, + "añ": 1213, + "ido": 1214, + "porque": 1215, + "ella": 1216, + "cen": 1217, + "muy": 1218, + "cal": 1219, + "este": 1220, + "has": 1221, + "có": 1222, + "gra": 1223, + "ros": 1224, + "aquel": 1225, + "dijo": 1226, + "cía": 1227, + "zo": 1228, + "ciones": 1229, + "mbi": 1230, + "elo": 1231, + "tó": 1232, + "ina": 1233, + "todos": 1234, + "tien": 1235, + "estaba": 1236, + "deci": 1237, + "cio": 1238, + "ño": 1239, + "lor": 1240, + "nues": 1241, + "medi": 1242, + "len": 1243, + "vida": 1244, + "ali": 1245, + "pues": 1246, + "ales": 1247, + "vol": 1248, + "mí": 1249, + "rar": 1250, + "cion": 1251, + "hasta": 1252, + "señor": 1253, + "cono": 1254, + "ah": 1255, + "dios": 1256, + "esa": 1257, + "ún": 1258, + "var": 1259, + "san": 1260, + "gui": 1261, + "otros": 1262, + "tado": 1263, + "buen": 1264, + "ña": 1265, + "tiemp": 1266, + "hacer": 1267, + "jer": 1268, + "vu": 1269, + "ana": 1270, + "así": 1271, + "antes": 1272, + "vez": 1273, + "miento": 1274, + "jar": 1275, + "lab": 1276, + "casa": 1277, + "eso": 1278, + "ego": 1279, + "dió": 1280, + "está": 1281, + "encia": 1282, + "eli": 1283, + "ías": 1284, + "tiempo": 1285, + "zar": 1286, + "van": 1287, + "mun": 1288, + "erta": 1289, + "tambi": 1290, + "sí": 1291, + "aun": 1292, + "mismo": 1293, + "entes": 1294, + "mano": 1295, + "ele": 1296, + "nada": 1297, + "segu": 1298, + "mej": 1299, + "erra": 1300, + "tir": 1301, + "uno": 1302, + "donde": 1303, + "toda": 1304, + "desde": 1305, + "también": 1306, + "cuer": 1307, + "hombre": 1308, + "otro": 1309, + "lib": 1310, + "trar": 1311, + "cual": 1312, + "hay": 1313, + "cada": 1314, + "taba": 1315, + "mento": 1316, + "tenía": 1317, + "quer": 1318, + "eran": 1319, + "siemp": 1320, + "siempre": 1321, + "erto": 1322, + "quí": 1323, + "gos": 1324, + "pués": 1325, + "ellos": 1326, + "después": 1327, + "nue": 1328, + "llo": 1329, + "inter": 1330, + "cómo": 1331, + "ahora": 1332, + "uste": 1333, + "traba": 1334, + "lado": 1335, + "ino": 1336, + "poco": 1337, + "erte": 1338, + "mujer": 1339, + "quier": 1340, + "algun": 1341, + "fue": 1342, + "ojos": 1343, + "enton": 1344, + "vos": 1345, + "esper": 1346, + "much": 1347, + "otra": 1348, + "az": 1349, + "eza": 1350, + "aquí": 1351, + "cias": 1352, + "gua": 1353, + "mucho": 1354, + "decir": 1355, + "esti": 1356, + "idad": 1357, + "algo": 1358, + "ocu": 1359, + "entonces": 1360, + "dido": 1361, + "entos": 1362, + "gri": 1363, + "dado": 1364, + "ios": 1365, + "dose": 1366, + "usted": 1367, + "quien": 1368, + "ami": 1369, + "unto": 1370, + "mejor": 1371, + "bas": 1372, + "solo": 1373, + "pregun": 1374, + "tur": 1375, + "alg": 1376, + "todas": 1377, + "parte": 1378, + "emb": 1379, + "cto": 1380, + "mundo": 1381, + "tiene": 1382, + "tante": 1383, + "palab": 1384, + "tran": 1385, + "aquella": 1386, + "cios": 1387, + "aunque": 1388, + "cuen": 1389, + "tener": 1390, + "fun": 1391, + "respon": 1392, + "allí": 1393, + "xi": 1394, + "han": 1395, + "pens": 1396, + "contra": 1397, + "tura": 1398, + "val": 1399, + "dio": 1400, + "tanto": 1401, + "camin": 1402, + "mó": 1403, + "esp": 1404, + "ada": 1405, + "ío": 1406, + "hacia": 1407, + "dej": 1408, + "estar": 1409, + "ión": 1410, + "gas": 1411, + "vas": 1412, + "noche": 1413, + "ér": 1414, + "años": 1415, + "padre": 1416, + "gus": 1417, + "ár": 1418, + "sino": 1419, + "manos": 1420, + "cido": 1421, + "estu": 1422, + "hubi": 1423, + "vir": 1424, + "bri": 1425, + "raz": 1426, + "chi": 1427, + "puede": 1428, + "menos": 1429, + "habi": 1430, + "homb": 1431, + "neces": 1432, + "may": 1433, + "eros": 1434, + "ría": 1435, + "hecho": 1436, + "escu": 1437, + "lti": 1438, + "ándo": 1439, + "bus": 1440, + "cosas": 1441, + "tú": 1442, + "espa": 1443, + "reci": 1444, + "ctor": 1445, + "prim": 1446, + "dia": 1447, + "dese": 1448, + "mientras": 1449, + "hor": 1450, + "fuer": 1451, + "ida": 1452, + "posi": 1453, + "lante": 1454, + "ano": 1455, + "estas": 1456, + "pli": 1457, + "luego": 1458, + "sión": 1459, + "cin": 1460, + "tierra": 1461, + "guar": 1462, + "cado": 1463, + "encon": 1464, + "pren": 1465, + "mayor": 1466, + "fal": 1467, + "ð": 1468, + "ħ": 1469, + "ň": 1470, + "ə": 1471, + "θ": 1472, + "’": 1473, + "“": 1474, + "”": 1475, + "zi": 1476, + "gli": 1477, + "tto": 1478, + "ono": 1479, + "nel": 1480, + "tti": 1481, + "della": 1482, + "zione": 1483, + "tta": 1484, + "tà": 1485, + "uo": 1486, + "come": 1487, + "alla": 1488, + "oni": 1489, + "ggi": 1490, + "ssi": 1491, + "più": 1492, + "ini": 1493, + "bb": 1494, + "sto": 1495, + "sono": 1496, + "eri": 1497, + "sse": 1498, + "sc": 1499, + "sul": 1500, + "vano": 1501, + "sti": 1502, + "suo": 1503, + "cchi": 1504, + "zza": 1505, + "anche": 1506, + "tte": 1507, + "sci": 1508, + "col": 1509, + "sso": 1510, + "ssa": 1511, + "dei": 1512, + "aveva": 1513, + "zz": 1514, + "amo": 1515, + "gno": 1516, + "sua": 1517, + "ria": 1518, + "sì": 1519, + "ché": 1520, + "dal": 1521, + "ona": 1522, + "spe": 1523, + "gni": 1524, + "tt": 1525, + "delle": 1526, + "questo": 1527, + "nella": 1528, + "dere": 1529, + "anno": 1530, + "dell": 1531, + "uni": 1532, + "bbe": 1533, + "anti": 1534, + "ene": 1535, + "gio": 1536, + "uto": 1537, + "qual": 1538, + "glia": 1539, + "quando": 1540, + "tutto": 1541, + "glio": 1542, + "zioni": 1543, + "cam": 1544, + "esso": 1545, + "ss": 1546, + "mol": 1547, + "loro": 1548, + "perché": 1549, + "cosa": 1550, + "due": 1551, + "poi": 1552, + "sco": 1553, + "cco": 1554, + "gna": 1555, + "tem": 1556, + "prima": 1557, + "così": 1558, + "essere": 1559, + "ani": 1560, + "bra": 1561, + "rio": 1562, + "anco": 1563, + "cui": 1564, + "spi": 1565, + "via": 1566, + "gior": 1567, + "bile": 1568, + "ggio": 1569, + "mai": 1570, + "tare": 1571, + "indi": 1572, + "rebbe": 1573, + "senza": 1574, + "zio": 1575, + "tutti": 1576, + "stato": 1577, + "zia": 1578, + "dalla": 1579, + "mia": 1580, + "vita": 1581, + "quella": 1582, + "qua": 1583, + "dove": 1584, + "allo": 1585, + "sempre": 1586, + "zzo": 1587, + "sia": 1588, + "dopo": 1589, + "porta": 1590, + "ccia": 1591, + "erano": 1592, + "anni": 1593, + "chia": 1594, + "enza": 1595, + "propri": 1596, + "anda": 1597, + "cca": 1598, + "occhi": 1599, + "questa": 1600, + "ffi": 1601, + "ron": 1602, + "mio": 1603, + "ris": 1604, + "ogni": 1605, + "rin": 1606, + "far": 1607, + "menti": 1608, + "ancora": 1609, + "fatto": 1610, + "mani": 1611, + "senti": 1612, + "pra": 1613, + "tempo": 1614, + "essi": 1615, + "bbi": 1616, + "lare": 1617, + "pers": 1618, + "sor": 1619, + "anza": 1620, + "pie": 1621, + "verso": 1622, + "altro": 1623, + "tato": 1624, + "cato": 1625, + "ato": 1626, + "volta": 1627, + "cc": 1628, + "fare": 1629, + "ciò": 1630, + "bili": 1631, + "nuo": 1632, + "quello": 1633, + "colo": 1634, + "ppo": 1635, + "trova": 1636, + "ore": 1637, + "rono": 1638, + "molto": 1639, + "almente": 1640, + "sca": 1641, + "vole": 1642, + "tali": 1643, + "sulla": 1644, + "sce": 1645, + "meno": 1646, + "anto": 1647, + "pun": 1648, + "stu": 1649, + "capi": 1650, + "giu": 1651, + "mini": 1652, + "pia": 1653, + "lavo": 1654, + "vero": 1655, + "rsi": 1656, + "altri": 1657, + "scia": 1658, + "suoi": 1659, + "glie": 1660, + "sotto": 1661, + "bene": 1662, + "scri": 1663, + "tale": 1664, + "degli": 1665, + "alc": 1666, + "uomo": 1667, + "pel": 1668, + "pote": 1669, + "essa": 1670, + "scu": 1671, + "signo": 1672, + "stro": 1673, + "uti": 1674, + "sione": 1675, + "gre": 1676, + "fini": 1677, + "lun": 1678, + "esi": 1679, + "passa": 1680, + "rà": 1681, + "mentre": 1682, + "hanno": 1683, + "usci": 1684, + "gia": 1685, + "già": 1686, + "mina": 1687, + "tica": 1688, + "giorno": 1689, + "esse": 1690, + "modo": 1691, + "spa": 1692, + "proprio": 1693, + "ori": 1694, + "contro": 1695, + "stru": 1696, + "diven": 1697, + "disse": 1698, + "rato": 1699, + "noi": 1700, + "vere": 1701, + "può": 1702, + "dice": 1703, + "cci": 1704, + "secon": 1705, + "ccio": 1706, + "qualche": 1707, + "tutta": 1708, + "gg": 1709, + "mondo": 1710, + "forma": 1711, + "mma": 1712, + "pensa": 1713, + "deva": 1714, + "fosse": 1715, + "sopra": 1716, + "tamente": 1717, + "ness": 1718, + "quanto": 1719, + "raga": 1720, + "unque": 1721, + "care": 1722, + "stre": 1723, + "grande": 1724, + "picco": 1725, + "guarda": 1726, + "nell": 1727, + "possi": 1728, + "presen": 1729, + "rò": 1730, + "paro": 1731, + "tua": 1732, + "vin": 1733, + "ane": 1734, + "stesso": 1735, + "dav": 1736, + "nei": 1737, + "nelle": 1738, + "ghi": 1739, + "pio": 1740, + "lato": 1741, + "sid": 1742, + "fine": 1743, + "fuo": 1744, + "quasi": 1745, + "ulti": 1746, + "ito": 1747, + "sue": 1748, + "fil": 1749, + "allora": 1750, + "veni": 1751, + "tano": 1752, + "ello": 1753, + "ão": 1754, + "não": 1755, + "uma": 1756, + "ela": 1757, + "lh": 1758, + "ção": 1759, + "cê": 1760, + "inha": 1761, + "você": 1762, + "ec": 1763, + "dade": 1764, + "ao": 1765, + "ram": 1766, + "vel": 1767, + "ém": 1768, + "pode": 1769, + "estava": 1770, + "isso": 1771, + "mui": 1772, + "faz": 1773, + "ões": 1774, + "pes": 1775, + "ix": 1776, + "sim": 1777, + "olh": 1778, + "isa": 1779, + "ên": 1780, + "tinha": 1781, + "meu": 1782, + "são": 1783, + "minha": 1784, + "muito": 1785, + "foi": 1786, + "bem": 1787, + "diz": 1788, + "parec": 1789, + "ço": 1790, + "pesso": 1791, + "pois": 1792, + "mesmo": 1793, + "ções": 1794, + "seus": 1795, + "até": 1796, + "ência": 1797, + "lhe": 1798, + "tiv": 1799, + "mã": 1800, + "só": 1801, + "tão": 1802, + "tudo": 1803, + "então": 1804, + "inda": 1805, + "bal": 1806, + "indo": 1807, + "ndo": 1808, + "já": 1809, + "vam": 1810, + "eito": 1811, + "depois": 1812, + "mel": 1813, + "lha": 1814, + "ainda": 1815, + "fazer": 1816, + "pou": 1817, + "pergun": 1818, + "deix": 1819, + "tamb": 1820, + "ala": 1821, + "pelo": 1822, + "também": 1823, + "fica": 1824, + "prec": 1825, + "eles": 1826, + "havia": 1827, + "lá": 1828, + "nas": 1829, + "gem": 1830, + "mem": 1831, + "ós": 1832, + "deu": 1833, + "eiro": 1834, + "..": 1835, + "assim": 1836, + "ior": 1837, + "har": 1838, + "aqui": 1839, + "cul": 1840, + "sar": 1841, + "outra": 1842, + "olhos": 1843, + "ima": 1844, + "mim": 1845, + "ago": 1846, + "pessoas": 1847, + "eram": 1848, + "eira": 1849, + "pela": 1850, + "coisa": 1851, + "mão": 1852, + "conh": 1853, + "agora": 1854, + "iam": 1855, + "há": 1856, + "suas": 1857, + "guém": 1858, + "cabe": 1859, + "nem": 1860, + "ível": 1861, + "consegu": 1862, + "trabal": 1863, + "lev": 1864, + "lem": 1865, + "vai": 1866, + "tei": 1867, + "pró": 1868, + "quem": 1869, + "onde": 1870, + "cabeça": 1871, + "nunca": 1872, + "mentos": 1873, + "hum": 1874, + "dele": 1875, + "verdade": 1876, + "tá": 1877, + "hos": 1878, + "algum": 1879, + "dizer": 1880, + "penas": 1881, + "nós": 1882, + "enquanto": 1883, + "outro": 1884, + "lho": 1885, + "melhor": 1886, + "primei": 1887, + "iu": 1888, + "apenas": 1889, + "estou": 1890, + "conte": 1891, + "homem": 1892, + "dois": 1893, + "ças": 1894, + "pouco": 1895, + "senhor": 1896, + "tando": 1897, + "espera": 1898, + "pai": 1899, + "rios": 1900, + "baix": 1901, + "ase": 1902, + "isas": 1903, + "hora": 1904, + "ficar": 1905, + "seja": 1906, + "ân": 1907, + "clar": 1908, + "inc": 1909, + "fos": 1910, + "ouvi": 1911, + "vem": 1912, + "tava": 1913, + "ário": 1914, + "sos": 1915, + "inho": 1916, + "rando": 1917, + "ês": 1918, + "coisas": 1919, + "aconte": 1920, + "lher": 1921, + "anos": 1922, + "talvez": 1923, + "estão": 1924, + "liv": 1925, + "outros": 1926, + "qualquer": 1927, + "gou": 1928, + "lí": 1929, + "tivesse": 1930, + "rado": 1931, + "precisa": 1932, + "mãe": 1933, + "dela": 1934, + "entra": 1935, + "maior": 1936, + "noite": 1937, + "tiva": 1938, + "pala": 1939, + "ração": 1940, + "deus": 1941, + "sas": 1942, + "inte": 1943, + "fei": 1944, + "palav": 1945, + "trás": 1946, + "cidade": 1947, + "lugar": 1948, + "vezes": 1949, + "encontra": 1950, + "tru": 1951, + "eci": 1952, + "ın": 1953, + "bir": 1954, + "yor": 1955, + "ek": 1956, + "dı": 1957, + "ey": 1958, + "tı": 1959, + "mı": 1960, + "iz": 1961, + "ır": 1962, + "gö": 1963, + "sı": 1964, + "bil": 1965, + "lı": 1966, + "üz": 1967, + "iç": 1968, + "iy": 1969, + "ım": 1970, + "uz": 1971, + "cak": 1972, + "iş": 1973, + "ını": 1974, + "iyor": 1975, + "baş": 1976, + "dü": 1977, + "değ": 1978, + "kar": 1979, + "ev": 1980, + "öy": 1981, + "bun": 1982, + "yap": 1983, + "sun": 1984, + "gör": 1985, + "yı": 1986, + "ki": 1987, + "ara": 1988, + "alı": 1989, + "onu": 1990, + "çı": 1991, + "şey": 1992, + "sın": 1993, + "kı": 1994, + "kad": 1995, + "ağ": 1996, + "değil": 1997, + "ük": 1998, + "çok": 1999, + "şı": 2000, + "ül": 2001, + "için": 2002, + "eye": 2003, + "oldu": 2004, + "mış": 2005, + "kal": 2006, + "mek": 2007, + "öyle": 2008, + "yordu": 2009, + "yüz": 2010, + "miş": 2011, + "mak": 2012, + "ola": 2013, + "yan": 2014, + "cek": 2015, + "yorum": 2016, + "bak": 2017, + "üm": 2018, + "ları": 2019, + "oğ": 2020, + "kadar": 2021, + "arı": 2022, + "ında": 2023, + "gün": 2024, + "yok": 2025, + "yer": 2026, + "dım": 2027, + "daha": 2028, + "ına": 2029, + "dim": 2030, + "bilir": 2031, + "iki": 2032, + "siz": 2033, + "diğ": 2034, + "bü": 2035, + "düş": 2036, + "üç": 2037, + "unu": 2038, + "aman": 2039, + "fak": 2040, + "ede": 2041, + "sonra": 2042, + "hiç": 2043, + "aki": 2044, + "ğı": 2045, + "bul": 2046, + "maz": 2047, + "anla": 2048, + "bura": 2049, + "geç": 2050, + "maya": 2051, + "konu": 2052, + "din": 2053, + "tek": 2054, + "zaman": 2055, + "eler": 2056, + "öz": 2057, + "dır": 2058, + "gibi": 2059, + "şa": 2060, + "leri": 2061, + "kim": 2062, + "ku": 2063, + "fakat": 2064, + "yar": 2065, + "göz": 2066, + "cı": 2067, + "yorsun": 2068, + "bek": 2069, + "inde": 2070, + "pek": 2071, + "bunu": 2072, + "lik": 2073, + "iler": 2074, + "edi": 2075, + "öl": 2076, + "sür": 2077, + "sır": 2078, + "çık": 2079, + "sıl": 2080, + "alar": 2081, + "kes": 2082, + "yak": 2083, + "çek": 2084, + "yıl": 2085, + "ecek": 2086, + "ız": 2087, + "git": 2088, + "kap": 2089, + "ama": 2090, + "ıl": 2091, + "ların": 2092, + "biz": 2093, + "tır": 2094, + "oy": 2095, + "ancak": 2096, + "doğ": 2097, + "bana": 2098, + "şim": 2099, + "başla": 2100, + "lü": 2101, + "madı": 2102, + "beni": 2103, + "yük": 2104, + "lık": 2105, + "beş": 2106, + "nasıl": 2107, + "tık": 2108, + "tür": 2109, + "daki": 2110, + "ceğ": 2111, + "zı": 2112, + "iyi": 2113, + "dok": 2114, + "benim": 2115, + "cağ": 2116, + "yen": 2117, + "şu": 2118, + "mez": 2119, + "düşün": 2120, + "kendi": 2121, + "şimdi": 2122, + "yol": 2123, + "yu": 2124, + "iste": 2125, + "sek": 2126, + "mam": 2127, + "söyle": 2128, + "dik": 2129, + "kur": 2130, + "olduğ": 2131, + "sını": 2132, + "biliyor": 2133, + "kan": 2134, + "yal": 2135, + "meye": 2136, + "muş": 2137, + "kaç": 2138, + "iye": 2139, + "tü": 2140, + "ef": 2141, + "tım": 2142, + "evet": 2143, + "yet": 2144, + "burada": 2145, + "tim": 2146, + "biraz": 2147, + "kor": 2148, + "doğru": 2149, + "inin": 2150, + "kız": 2151, + "diye": 2152, + "dör": 2153, + "etti": 2154, + "onun": 2155, + "isti": 2156, + "ği": 2157, + "sana": 2158, + "üş": 2159, + "arka": 2160, + "hayır": 2161, + "karşı": 2162, + "ile": 2163, + "hak": 2164, + "ıyor": 2165, + "neden": 2166, + "sev": 2167, + "sız": 2168, + "çocu": 2169, + "çalı": 2170, + "olur": 2171, + "bır": 2172, + "gir": 2173, + "ise": 2174, + "ih": 2175, + "kır": 2176, + "dön": 2177, + "böyle": 2178, + "seni": 2179, + "!\"": 2180, + "dört": 2181, + "söy": 2182, + "oş": 2183, + "musun": 2184, + "laş": 2185, + "ip": 2186, + "kay": 2187, + "hem": 2188, + "büyük": 2189, + "aç": 2190, + "bırak": 2191, + "misin": 2192, + "söz": 2193, + "değiş": 2194, + "ünü": 2195, + "gül": 2196, + "kö": 2197, + "karı": 2198, + "tamam": 2199, + "olu": 2200, + "yeni": 2201, + "lam": 2202, + "mıştı": 2203, + "yaş": 2204, + "iniz": 2205, + "kadın": 2206, + "bunun": 2207, + "mey": 2208, + "altı": 2209, + "yi": 2210, + "inden": 2211, + "senin": 2212, + "yat": 2213, + "top": 2214, + "isi": 2215, + "dün": 2216, + "hiçbir": 2217, + "yon": 2218, + "dın": 2219, + "tün": 2220, + "başka": 2221, + "hep": 2222, + "irmi": 2223, + "devam": 2224, + "olacak": 2225, + "artık": 2226, + "durum": 2227, + "imiz": 2228, + "üzel": 2229, + "lerini": 2230, + "sağ": 2231, + "gerek": 2232, + "yirmi": 2233, + "şek": 2234, + "bağ": 2235, + "lara": 2236, + "yür": 2237, + "ması": 2238, + "katı": 2239, + "dedi": 2240, + "gü": 2241, + "sorun": 2242, + "üne": 2243, + "mız": 2244, + "yapı": 2245, + "mil": 2246, + "ğını": 2247, + "tara": 2248, + "vardı": 2249, + "konuş": 2250, + "arak": 2251, + "larak": 2252, + "çocuk": 2253, + "bütün": 2254, + "ley": 2255, + "dür": 2256, + "güzel": 2257, + "ayı": 2258, + "yapa": 2259, + "nı": 2260, + "ayr": 2261, + "öne": 2262, + "yordum": 2263, + "ban": 2264, + "i̇ş": 2265, + "dum": 2266, + "yorlar": 2267, + "larını": 2268, + "çıkar": 2269, + "zan": 2270, + "seç": 2271, + "liyor": 2272, + "tak": 2273, + "şık": 2274, + "tekrar": 2275, + "aş": 2276, + "eş": 2277, + "mişti": 2278, + "kin": 2279, + "imi": 2280, + "eğ": 2281, + "gidi": 2282, + "leş": 2283, + "başladı": 2284, + "gide": 2285, + "otur": 2286, + "dde": 2287, + "ından": 2288, + "üzer": 2289, + "ının": 2290, + "nız": 2291, + "uy": 2292, + "yedi": 2293, + "kat": 2294, + "olarak": 2295, + "ladı": 2296, + "yalnız": 2297, + "bah": 2298, + "iyet": 2299, + "sak": 2300, + "açık": 2301, + "sında": 2302, + "...": 2303, + "insan": 2304, + "aynı": 2305, + "eder": 2306, + "istan": 2307, + "uzun": 2308, + "geri": 2309, + "erek": 2310, + "olan": 2311, + "gerçek": 2312, + "alan": 2313, + "dış": 2314, + "alık": 2315, + "fark": 2316, + "üst": 2317, + "sade": 2318, + "kiş": 2319, + "ldı": 2320, + "zor": 2321, + "etir": 2322, + "herkes": 2323, + "ömer": 2324, + "unda": 2325, + "haf": 2326, + "buna": 2327, + "ydı": 2328, + "peki": 2329, + "adam": 2330, + "haz": 2331, + "sına": 2332, + "kapı": 2333, + "görüş": 2334, + "sadece": 2335, + "aldı": 2336, + "geldi": 2337, + "rz": 2338, + "sz": 2339, + "cz": 2340, + "ię": 2341, + "dz": 2342, + "ał": 2343, + "się": 2344, + "rze": 2345, + "że": 2346, + "wy": 2347, + "rzy": 2348, + "ła": 2349, + "ło": 2350, + "ny": 2351, + "dzie": 2352, + "dzi": 2353, + "czy": 2354, + "cie": 2355, + "prze": 2356, + "dy": 2357, + "kie": 2358, + "ry": 2359, + "ją": 2360, + "ów": 2361, + "przy": 2362, + "mie": 2363, + "szy": 2364, + "cze": 2365, + "bie": 2366, + "cy": 2367, + "nia": 2368, + "ści": 2369, + "sze": 2370, + "jest": 2371, + "ży": 2372, + "ną": 2373, + "któ": 2374, + "ała": 2375, + "mnie": 2376, + "ły": 2377, + "cza": 2378, + "jak": 2379, + "roz": 2380, + "ró": 2381, + "zna": 2382, + "łu": 2383, + "ść": 2384, + "wia": 2385, + "wszy": 2386, + "spo": 2387, + "gdy": 2388, + "wał": 2389, + "wię": 2390, + "łem": 2391, + "ję": 2392, + "sk": 2393, + "rę": 2394, + "dob": 2395, + "już": 2396, + "bę": 2397, + "ałem": 2398, + "sza": 2399, + "pod": 2400, + "dla": 2401, + "pan": 2402, + "nę": 2403, + "może": 2404, + "śli": 2405, + "ało": 2406, + "lko": 2407, + "nych": 2408, + "powie": 2409, + "cię": 2410, + "tylko": 2411, + "naj": 2412, + "tego": 2413, + "ski": 2414, + "nego": 2415, + "wszyst": 2416, + "szcze": 2417, + "jed": 2418, + "jej": 2419, + "two": 2420, + "ąd": 2421, + "śmy": 2422, + "czę": 2423, + "wać": 2424, + "jego": 2425, + "ża": 2426, + "sy": 2427, + "praw": 2428, + "tym": 2429, + "który": 2430, + "ały": 2431, + "trze": 2432, + "niej": 2433, + "nym": 2434, + "gło": 2435, + "jąc": 2436, + "mówi": 2437, + "ska": 2438, + "nej": 2439, + "słu": 2440, + "wła": 2441, + "będzie": 2442, + "dę": 2443, + "pó": 2444, + "bez": 2445, + "nic": 2446, + "pła": 2447, + "ście": 2448, + "są": 2449, + "trzy": 2450, + "kiem": 2451, + "był": 2452, + "mog": 2453, + "robi": 2454, + "tam": 2455, + "mię": 2456, + "zy": 2457, + "pew": 2458, + "myś": 2459, + "przed": 2460, + "sko": 2461, + "które": 2462, + "lę": 2463, + "wsze": 2464, + "ąc": 2465, + "było": 2466, + "sobie": 2467, + "py": 2468, + "cią": 2469, + "jeszcze": 2470, + "tę": 2471, + "czas": 2472, + "szę": 2473, + "gł": 2474, + "kę": 2475, + "czu": 2476, + "przez": 2477, + "sło": 2478, + "wz": 2479, + "kto": 2480, + "ków": 2481, + "czo": 2482, + "liśmy": 2483, + "więc": 2484, + "rą": 2485, + "wó": 2486, + "rza": 2487, + "ności": 2488, + "wet": 2489, + "nął": 2490, + "śmie": 2491, + "nawet": 2492, + "musi": 2493, + "swo": 2494, + "tej": 2495, + "wą": 2496, + "wu": 2497, + "wią": 2498, + "niu": 2499, + "czą": 2500, + "dzo": 2501, + "skie": 2502, + "jeśli": 2503, + "czego": 2504, + "chy": 2505, + "dł": 2506, + "tych": 2507, + "bym": 2508, + "żo": 2509, + "eś": 2510, + "sią": 2511, + "kiedy": 2512, + "wró": 2513, + "dze": 2514, + "dro": 2515, + "rów": 2516, + "pani": 2517, + "kul": 2518, + "nad": 2519, + "chwi": 2520, + "nim": 2521, + "być": 2522, + "chodzi": 2523, + "nio": 2524, + "dobrze": 2525, + "teraz": 2526, + "wokul": 2527, + "coś": 2528, + "kł": 2529, + "pier": 2530, + "gdzie": 2531, + "dzy": 2532, + "pię": 2533, + "dź": 2534, + "ką": 2535, + "gó": 2536, + "zda": 2537, + "chce": 2538, + "stę": 2539, + "świa": 2540, + "wszystko": 2541, + "peł": 2542, + "wiem": 2543, + "wiel": 2544, + "każ": 2545, + "rzu": 2546, + "sły": 2547, + "jedna": 2548, + "myśl": 2549, + "mój": 2550, + "jestem": 2551, + "óż": 2552, + "miej": 2553, + "moż": 2554, + "kła": 2555, + "resz": 2556, + "dłu": 2557, + "stwo": 2558, + "nię": 2559, + "masz": 2560, + "żeby": 2561, + "niem": 2562, + "jakie": 2563, + "sty": 2564, + "nią": 2565, + "wej": 2566, + "oj": 2567, + "sła": 2568, + "ność": 2569, + "zło": 2570, + "szczę": 2571, + "lej": 2572, + "wego": 2573, + "cał": 2574, + "dział": 2575, + "kich": 2576, + "dza": 2577, + "dzię": 2578, + "oczy": 2579, + "zosta": 2580, + "czło": 2581, + "nam": 2582, + "kil": 2583, + "szu": 2584, + "wę": 2585, + "miał": 2586, + "strze": 2587, + "cej": 2588, + "ej": 2589, + "znaj": 2590, + "dać": 2591, + "miejs": 2592, + "kró": 2593, + "kry": 2594, + "bardzo": 2595, + "śnie": 2596, + "lą": 2597, + "gie": 2598, + "ciebie": 2599, + "dni": 2600, + "potrze": 2601, + "wokulski": 2602, + "uwa": 2603, + "umie": 2604, + "jednak": 2605, + "kra": 2606, + "wróci": 2607, + "człowie": 2608, + "czyć": 2609, + "była": 2610, + "żeli": 2611, + "mę": 2612, + "cę": 2613, + "zrobi": 2614, + "mogę": 2615, + "prowa": 2616, + "rem": 2617, + "niech": 2618, + "cznie": 2619, + "kro": 2620, + "tą": 2621, + "chci": 2622, + "bro": 2623, + "dzieć": 2624, + "szą": 2625, + "pad": 2626, + "trz": 2627, + "jem": 2628, + "tów": 2629, + "dru": 2630, + "taj": 2631, + "rzekł": 2632, + "niego": 2633, + "takie": 2634, + "wała": 2635, + "towa": 2636, + "kapła": 2637, + "widzi": 2638, + "podob": 2639, + "dzę": 2640, + "tał": 2641, + "stęp": 2642, + "bą": 2643, + "poko": 2644, + "wem": 2645, + "gę": 2646, + "aby": 2647, + "albo": 2648, + "spra": 2649, + "zno": 2650, + "smo": 2651, + "jesz": 2652, + "księ": 2653, + "jesteś": 2654, + "poz": 2655, + "nigdy": 2656, + "ksią": 2657, + "cóż": 2658, + "ws": 2659, + "pow": 2660, + "tka": 2661, + "świe": 2662, + "szka": 2663, + "samo": 2664, + "sł": 2665, + "rzę": 2666, + "nale": 2667, + "chcesz": 2668, + "nik": 2669, + "pę": 2670, + "chyba": 2671, + "ciąg": 2672, + "jący": 2673, + "woj": 2674, + "nasze": 2675, + "mniej": 2676, + "więcej": 2677, + "zwy": 2678, + "osta": 2679, + "waż": 2680, + "śmier": 2681, + "wier": 2682, + "dzą": 2683, + "zaś": 2684, + "gdyby": 2685, + "jaki": 2686, + "wol": 2687, + "win": 2688, + "dą": 2689, + "ścia": 2690, + "rozma": 2691, + "wal": 2692, + "panie": 2693, + "star": 2694, + "kaz": 2695, + "jeżeli": 2696, + "wra": 2697, + "koń": 2698, + "siebie": 2699, + "znowu": 2700, + "czem": 2701, + "stwa": 2702, + "isto": 2703, + "pół": 2704, + "dał": 2705, + "kobie": 2706, + "ałam": 2707, + "wych": 2708, + "cesa": 2709, + "nich": 2710, + "zawsze": 2711, + "dzić": 2712, + "też": 2713, + "lepie": 2714, + "proszę": 2715, + "kre": 2716, + "twa": 2717, + "łą": 2718, + "chu": 2719, + "cą": 2720, + "prz": 2721, + "łe": 2722, + "szedł": 2723, + "odpowie": 2724, + "myśli": 2725, + "świą": 2726, + "ź": 2727, + "ł": 2728, + "&": 2729, + "=": 2730, + "ă": 2731, + "đ": 2732, + "ţ": 2733, + "–": 2734, + "‘": 2735, + "ij": 2736, + "aa": 2737, + "een": 2738, + "het": 2739, + "aar": 2740, + "oor": 2741, + "ijn": 2742, + "dat": 2743, + "oe": 2744, + "ijk": 2745, + "aan": 2746, + "voor": 2747, + "iet": 2748, + "zijn": 2749, + "niet": 2750, + "oo": 2751, + "moet": 2752, + "heb": 2753, + "uit": 2754, + "wij": 2755, + "aat": 2756, + "lijk": 2757, + "sl": 2758, + "daar": 2759, + "deze": 2760, + "worden": 2761, + "moeten": 2762, + "onder": 2763, + "hebben": 2764, + "ook": 2765, + "ct": 2766, + "nog": 2767, + "aal": 2768, + "eer": 2769, + "bij": 2770, + "mijn": 2771, + "kom": 2772, + "atie": 2773, + "eft": 2774, + "kel": 2775, + "rij": 2776, + "heid": 2777, + "af": 2778, + "stel": 2779, + "maar": 2780, + "wee": 2781, + "heeft": 2782, + "waar": 2783, + "eren": 2784, + "wat": 2785, + "wil": 2786, + "aag": 2787, + "bet": 2788, + "hij": 2789, + "kun": 2790, + "uw": 2791, + "dt": 2792, + "door": 2793, + "tij": 2794, + "ond": 2795, + "geen": 2796, + "gev": 2797, + "veel": 2798, + "naar": 2799, + "aten": 2800, + "kunnen": 2801, + "echt": 2802, + "goe": 2803, + "twee": 2804, + "delijk": 2805, + "uur": 2806, + "toe": 2807, + "meer": 2808, + "onze": 2809, + "tijd": 2810, + "hoe": 2811, + "tot": 2812, + "zou": 2813, + "aak": 2814, + "amen": 2815, + "woor": 2816, + "wordt": 2817, + "gelijk": 2818, + "gaan": 2819, + "ker": 2820, + "eld": 2821, + "hou": 2822, + "zel": 2823, + "tegen": 2824, + "komen": 2825, + "werk": 2826, + "goed": 2827, + "zal": 2828, + "zij": 2829, + "slag": 2830, + "zien": 2831, + "echter": 2832, + "itie": 2833, + "tie": 2834, + "elijk": 2835, + "ische": 2836, + "belan": 2837, + "haar": 2838, + "vr": 2839, + "grijk": 2840, + "doen": 2841, + "land": 2842, + "belangrijk": 2843, + "open": 2844, + "ctie": 2845, + "zelf": 2846, + "mij": 2847, + "iteit": 2848, + "stem": 2849, + "mee": 2850, + "aren": 2851, + "dien": 2852, + "gaat": 2853, + "prob": 2854, + "moe": 2855, + "ullen": 2856, + "zich": 2857, + "daarom": 2858, + "orm": 2859, + "staat": 2860, + "zit": 2861, + "dui": 2862, + "dus": 2863, + "ds": 2864, + "verslag": 2865, + "kelijk": 2866, + "proble": 2867, + "schap": 2868, + "gd": 2869, + "hun": 2870, + "erd": 2871, + "zet": 2872, + "staan": 2873, + "maal": 2874, + "inder": 2875, + "eid": 2876, + "kken": 2877, + "ged": 2878, + "zullen": 2879, + "mensen": 2880, + "jaar": 2881, + "regel": 2882, + "ieder": 2883, + "volgen": 2884, + "geven": 2885, + "even": 2886, + "blij": 2887, + "ië": 2888, + "uwe": 2889, + "maken": 2890, + "oek": 2891, + "nieuwe": 2892, + "baar": 2893, + "andere": 2894, + "ruik": 2895, + "agen": 2896, + "ouw": 2897, + "willen": 2898, + "aakt": 2899, + "hoo": 2900, + "anden": 2901, + "lig": 2902, + "samen": 2903, + "zeer": 2904, + "duidelijk": 2905, + "antwoor": 2906, + "heel": 2907, + "punt": 2908, + "houden": 2909, + "vraag": 2910, + "gele": 2911, + "eens": 2912, + "besch": 2913, + "omen": 2914, + "erg": 2915, + "doel": 2916, + "dag": 2917, + "uren": 2918, + "ings": 2919, + "oren": 2920, + "delen": 2921, + "steun": 2922, + "innen": 2923, + "pol": 2924, + "oon": 2925, + "sn": 2926, + "zonder": 2927, + "nodig": 2928, + "alleen": 2929, + "mid": 2930, + "ragen": 2931, + "iets": 2932, + "versch": 2933, + "gebruik": 2934, + "rouw": 2935, + "stellen": 2936, + "menten": 2937, + "eerste": 2938, + "laat": 2939, + "groot": 2940, + "ood": 2941, + "toch": 2942, + "laten": 2943, + "aard": 2944, + "sle": 2945, + "deel": 2946, + "plaat": 2947, + "ree": 2948, + "betre": 2949, + "lid": 2950, + "uiten": 2951, + "racht": 2952, + "beleid": 2953, + "stie": 2954, + "staten": 2955, + "ggen": 2956, + "reken": 2957, + "alen": 2958, + "ming": 2959, + "mogelijk": 2960, + "grote": 2961, + "altijd": 2962, + "enkel": 2963, + "wik": 2964, + "politie": 2965, + "elk": 2966, + "handel": 2967, + "kwe": 2968, + "maat": 2969, + "elen": 2970, + "vrij": 2971, + "jes": 2972, + "aam": 2973, + "huis": 2974, + "weer": 2975, + "lidstaten": 2976, + "king": 2977, + "kle": 2978, + "bed": 2979, + "geval": 2980, + "wikkel": 2981, + "kwestie": 2982, + "stee": 2983, + "hel": 2984, + "komst": 2985, + "iden": 2986, + "eerd": 2987, + "tweede": 2988, + "probleem": 2989, + "ussen": 2990, + "snel": 2991, + "tig": 2992, + "ult": 2993, + "nemen": 2994, + "commis": 2995, + "verschil": 2996, + "zoek": 2997, + "krij": 2998, + "graag": 2999, + "denk": 3000, + "landen": 3001, + "reden": 3002, + "besl": 3003, + "oeg": 3004, + "beter": 3005, + "heden": 3006, + "mag": 3007, + "boven": 3008, + "cont": 3009, + "fd": 3010, + "hele": 3011, + "vier": 3012, + "gez": 3013, + "kw": 3014, + "aas": 3015, + "ontwikkel": 3016, + "drie": 3017, + "vaak": 3018, + "plaats": 3019, + "gang": 3020, + "ijf": 3021, + "natuur": 3022, + "tussen": 3023, + "bat": 3024, + "komt": 3025, + "wacht": 3026, + "aad": 3027, + "achter": 3028, + "gebie": 3029, + "verk": 3030, + "ligt": 3031, + "nieuw": 3032, + "vand": 3033, + "ý": 3034, + "ď": 3035, + "ě": 3036, + "ř": 3037, + "ť": 3038, + "ů": 3039, + "„": 3040, + "ní": 3041, + "ně": 3042, + "ře": 3043, + "ná": 3044, + "vě": 3045, + "vá": 3046, + "rá": 3047, + "vy": 3048, + "mě": 3049, + "ři": 3050, + "ří": 3051, + "že": 3052, + "jí": 3053, + "vý": 3054, + "ji": 3055, + "dě": 3056, + "če": 3057, + "tě": 3058, + "ky": 3059, + "še": 3060, + "ké": 3061, + "ší": 3062, + "pře": 3063, + "ví": 3064, + "ný": 3065, + "ži": 3066, + "má": 3067, + "cí": 3068, + "zá": 3069, + "ské": 3070, + "dá": 3071, + "byl": 3072, + "tí": 3073, + "pří": 3074, + "při": 3075, + "či": 3076, + "vní": 3077, + "ča": 3078, + "dí": 3079, + "dní": 3080, + "ká": 3081, + "nou": 3082, + "vět": 3083, + "pě": 3084, + "kou": 3085, + "ých": 3086, + "bě": 3087, + "prá": 3088, + "jako": 3089, + "ží": 3090, + "zí": 3091, + "jsou": 3092, + "jsem": 3093, + "lní": 3094, + "cké": 3095, + "vat": 3096, + "před": 3097, + "hla": 3098, + "stá": 3099, + "čí": 3100, + "ši": 3101, + "kla": 3102, + "ště": 3103, + "lou": 3104, + "mů": 3105, + "chá": 3106, + "pů": 3107, + "také": 3108, + "dů": 3109, + "nost": 3110, + "tře": 3111, + "sku": 3112, + "vše": 3113, + "tní": 3114, + "byla": 3115, + "ční": 3116, + "jeho": 3117, + "bý": 3118, + "vání": 3119, + "ných": 3120, + "tři": 3121, + "vz": 3122, + "stře": 3123, + "dva": 3124, + "hle": 3125, + "čá": 3126, + "nosti": 3127, + "vš": 3128, + "hra": 3129, + "jen": 3130, + "slo": 3131, + "však": 3132, + "kdy": 3133, + "bylo": 3134, + "bude": 3135, + "jší": 3136, + "vých": 3137, + "ním": 3138, + "sm": 3139, + "koli": 3140, + "rů": 3141, + "může": 3142, + "není": 3143, + "hod": 3144, + "bí": 3145, + "tý": 3146, + "stě": 3147, + "uje": 3148, + "sá": 3149, + "pět": 3150, + "krá": 3151, + "tom": 3152, + "ství": 3153, + "vně": 3154, + "sed": 3155, + "své": 3156, + "pí": 3157, + "musí": 3158, + "už": 3159, + "tím": 3160, + "jící": 3161, + "jedno": 3162, + "čas": 3163, + "čty": 3164, + "ský": 3165, + "evro": 3166, + "toho": 3167, + "hy": 3168, + "kter": 3169, + "rní": 3170, + "stí": 3171, + "svě": 3172, + "pak": 3173, + "všech": 3174, + "ků": 3175, + "ng": 3176, + "ád": 3177, + "chází": 3178, + "být": 3179, + "první": 3180, + "mno": 3181, + "ského": 3182, + "pá": 3183, + "nebo": 3184, + "kem": 3185, + "sla": 3186, + "ného": 3187, + "zde": 3188, + "další": 3189, + "řa": 3190, + "čtyři": 3191, + "hrá": 3192, + "druh": 3193, + "lně": 3194, + "vla": 3195, + "ských": 3196, + "ško": 3197, + "půso": 3198, + "proto": 3199, + "vů": 3200, + "ská": 3201, + "šest": 3202, + "dně": 3203, + "ještě": 3204, + "mezi": 3205, + "několi": 3206, + "již": 3207, + "čně": 3208, + "slu": 3209, + "zná": 3210, + "sedm": 3211, + "vlá": 3212, + "osm": 3213, + "byly": 3214, + "vám": 3215, + "cký": 3216, + "tech": 3217, + "ději": 3218, + "velmi": 3219, + "leži": 3220, + "vala": 3221, + "lý": 3222, + "tvo": 3223, + "spole": 3224, + "stup": 3225, + "mož": 3226, + "evrop": 3227, + "stal": 3228, + "jde": 3229, + "rodi": 3230, + "její": 3231, + "poli": 3232, + "devět": 3233, + "sme": 3234, + "až": 3235, + "této": 3236, + "tento": 3237, + "kaž": 3238, + "nula": 3239, + "bych": 3240, + "moc": 3241, + "stou": 3242, + "kdo": 3243, + "zd": 3244, + "praco": 3245, + "tomu": 3246, + "ným": 3247, + "živo": 3248, + "zem": 3249, + "násle": 3250, + "sky": 3251, + "jich": 3252, + "měl": 3253, + "děla": 3254, + "jsme": 3255, + "nice": 3256, + "stej": 3257, + "stní": 3258, + "náro": 3259, + "nit": 3260, + "později": 3261, + "tako": 3262, + "nce": 3263, + "čer": 3264, + "ším": 3265, + "něco": 3266, + "vál": 3267, + "řej": 3268, + "krát": 3269, + "ální": 3270, + "asi": 3271, + "které": 3272, + "stav": 3273, + "mají": 3274, + "mys": 3275, + "době": 3276, + "sně": 3277, + "zku": 3278, + "tů": 3279, + "chod": 3280, + "spě": 3281, + "jejich": 3282, + "součas": 3283, + "vali": 3284, + "kte": 3285, + "prů": 3286, + "zení": 3287, + "pat": 3288, + "potře": 3289, + "dnes": 3290, + "zemí": 3291, + "znam": 3292, + "mám": 3293, + "tedy": 3294, + "hlavní": 3295, + "použí": 3296, + "bní": 3297, + "vede": 3298, + "lep": 3299, + "jek": 3300, + "prav": 3301, + "politi": 3302, + "dne": 3303, + "čení": 3304, + "než": 3305, + "děl": 3306, + "čo": 3307, + "cích": 3308, + "sté": 3309, + "dlou": 3310, + "několik": 3311, + "vyu": 3312, + "ckých": 3313, + "nové": 3314, + "čin": 3315, + "dělá": 3316, + "ký": 3317, + "obla": 3318, + "podle": 3319, + "důleži": 3320, + "poku": 3321, + "kone": 3322, + "dý": 3323, + "dvě": 3324, + "žád": 3325, + "nout": 3326, + "tku": 3327, + "tvr": 3328, + "ckého": 3329, + "rov": 3330, + "tele": 3331, + "psa": 3332, + "svět": 3333, + "tivní": 3334, + "dosta": 3335, + "šel": 3336, + "druhé": 3337, + "skou": 3338, + "žo": 3339, + "jedná": 3340, + "význam": 3341, + "problé": 3342, + "publi": 3343, + "ván": 3344, + "odpo": 3345, + "podpo": 3346, + "dle": 3347, + "jaké": 3348, + "šení": 3349, + "vím": 3350, + "během": 3351, + "nachází": 3352, + "slou": 3353, + "pouze": 3354, + "otá": 3355, + "plo": 3356, + "tové": 3357, + "větši": 3358, + "komi": 3359, + "vají": 3360, + "tyto": 3361, + "zápa": 3362, + "změ": 3363, + "moh": 3364, + "více": 3365, + "společ": 3366, + "auto": 3367, + "proti": 3368, + "dět": 3369, + "cháze": 3370, + "žel": 3371, + "«": 3372, + "»": 3373, + "а": 3374, + "б": 3375, + "в": 3376, + "г": 3377, + "д": 3378, + "е": 3379, + "ж": 3380, + "з": 3381, + "и": 3382, + "й": 3383, + "к": 3384, + "л": 3385, + "м": 3386, + "н": 3387, + "о": 3388, + "п": 3389, + "р": 3390, + "с": 3391, + "т": 3392, + "у": 3393, + "ф": 3394, + "х": 3395, + "ц": 3396, + "ч": 3397, + "ш": 3398, + "щ": 3399, + "ъ": 3400, + "ы": 3401, + "ь": 3402, + "э": 3403, + "ю": 3404, + "я": 3405, + "ё": 3406, + "‑": 3407, + "−": 3408, + "ст": 3409, + "ен": 3410, + "но": 3411, + "на": 3412, + "пр": 3413, + "то": 3414, + "по": 3415, + "ра": 3416, + "го": 3417, + "ко": 3418, + "не": 3419, + "во": 3420, + "ва": 3421, + "ет": 3422, + "ер": 3423, + "ни": 3424, + "ел": 3425, + "ит": 3426, + "ны": 3427, + "за": 3428, + "ро": 3429, + "ени": 3430, + "ка": 3431, + "ли": 3432, + "ем": 3433, + "да": 3434, + "об": 3435, + "ла": 3436, + "до": 3437, + "ся": 3438, + "ть": 3439, + "от": 3440, + "ло": 3441, + "ль": 3442, + "ед": 3443, + "со": 3444, + "ми": 3445, + "ре": 3446, + "мо": 3447, + "ци": 3448, + "про": 3449, + "та": 3450, + "это": 3451, + "ки": 3452, + "ру": 3453, + "при": 3454, + "ти": 3455, + "се": 3456, + "ста": 3457, + "вы": 3458, + "мы": 3459, + "ви": 3460, + "бы": 3461, + "ма": 3462, + "ес": 3463, + "ля": 3464, + "сти": 3465, + "ле": 3466, + "что": 3467, + "ме": 3468, + "ри": 3469, + "ча": 3470, + "од": 3471, + "ей": 3472, + "ель": 3473, + "ения": 3474, + "га": 3475, + "ну": 3476, + "си": 3477, + "па": 3478, + "раз": 3479, + "бо": 3480, + "сто": 3481, + "су": 3482, + "са": 3483, + "ду": 3484, + "его": 3485, + "ест": 3486, + "ин": 3487, + "ить": 3488, + "из": 3489, + "же": 3490, + "му": 3491, + "пер": 3492, + "под": 3493, + "ение": 3494, + "сь": 3495, + "ку": 3496, + "пред": 3497, + "ного": 3498, + "ных": 3499, + "вер": 3500, + "те": 3501, + "ной": 3502, + "ции": 3503, + "де": 3504, + "ры": 3505, + "дел": 3506, + "лю": 3507, + "ве": 3508, + "он": 3509, + "мен": 3510, + "ги": 3511, + "ня": 3512, + "бу": 3513, + "пра": 3514, + "все": 3515, + "ется": 3516, + "сть": 3517, + "жа": 3518, + "дол": 3519, + "жи": 3520, + "бе": 3521, + "кон": 3522, + "сл": 3523, + "ши": 3524, + "ди": 3525, + "ств": 3526, + "ско": 3527, + "ные": 3528, + "чи": 3529, + "ют": 3530, + "дер": 3531, + "стра": 3532, + "ты": 3533, + "ход": 3534, + "щи": 3535, + "зо": 3536, + "зна": 3537, + "ности": 3538, + "чес": 3539, + "вля": 3540, + "вать": 3541, + "ор": 3542, + "пол": 3543, + "вет": 3544, + "так": 3545, + "ша": 3546, + "ту": 3547, + "сво": 3548, + "пре": 3549, + "она": 3550, + "итель": 3551, + "ный": 3552, + "сло": 3553, + "как": 3554, + "вл": 3555, + "ность": 3556, + "хо": 3557, + "мож": 3558, + "пе": 3559, + "для": 3560, + "ния": 3561, + "ное": 3562, + "рас": 3563, + "долж": 3564, + "дар": 3565, + "тель": 3566, + "ска": 3567, + "пу": 3568, + "ство": 3569, + "кото": 3570, + "раб": 3571, + "ее": 3572, + "род": 3573, + "эти": 3574, + "соб": 3575, + "ору": 3576, + "жен": 3577, + "ным": 3578, + "ити": 3579, + "ние": 3580, + "ком": 3581, + "дет": 3582, + "сту": 3583, + "гу": 3584, + "пи": 3585, + "меж": 3586, + "ению": 3587, + "тер": 3588, + "работ": 3589, + "воз": 3590, + "ция": 3591, + "кой": 3592, + "щест": 3593, + "гра": 3594, + "зи": 3595, + "ря": 3596, + "между": 3597, + "ства": 3598, + "вс": 3599, + "ело": 3600, + "ше": 3601, + "мер": 3602, + "ба": 3603, + "зы": 3604, + "лу": 3605, + "аль": 3606, + "дей": 3607, + "гла": 3608, + "народ": 3609, + "кти": 3610, + "предста": 3611, + "лся": 3612, + "явля": 3613, + "ски": 3614, + "нов": 3615, + "един": 3616, + "ров": 3617, + "ис": 3618, + "нима": 3619, + "рем": 3620, + "ходи": 3621, + "также": 3622, + "дру": 3623, + "ать": 3624, + "след": 3625, + "гово": 3626, + "ная": 3627, + "ющи": 3628, + "ень": 3629, + "которы": 3630, + "хот": 3631, + "ву": 3632, + "их": 3633, + "ему": 3634, + "чит": 3635, + "важ": 3636, + "орга": 3637, + "чески": 3638, + "ще": 3639, + "ке": 3640, + "ха": 3641, + "пос": 3642, + "том": 3643, + "боль": 3644, + "мне": 3645, + "пас": 3646, + "объ": 3647, + "прав": 3648, + "конф": 3649, + "слу": 3650, + "поддер": 3651, + "стви": 3652, + "наш": 3653, + "лько": 3654, + "стоя": 3655, + "ную": 3656, + "лем": 3657, + "енных": 3658, + "кра": 3659, + "ды": 3660, + "международ": 3661, + "гда": 3662, + "необ": 3663, + "госу": 3664, + "ству": 3665, + "ении": 3666, + "государ": 3667, + "кто": 3668, + "им": 3669, + "чест": 3670, + "рет": 3671, + "вопро": 3672, + "лен": 3673, + "ели": 3674, + "рова": 3675, + "ций": 3676, + "нам": 3677, + "этой": 3678, + "жения": 3679, + "необходи": 3680, + "меня": 3681, + "было": 3682, + "сили": 3683, + "фи": 3684, + "вя": 3685, + "шь": 3686, + "этого": 3687, + "они": 3688, + "органи": 3689, + "безо": 3690, + "проб": 3691, + "име": 3692, + "реш": 3693, + "би": 3694, + "безопас": 3695, + "ются": 3696, + "оста": 3697, + "енно": 3698, + "год": 3699, + "ела": 3700, + "представ": 3701, + "ться": 3702, + "слово": 3703, + "организа": 3704, + "должны": 3705, + "этом": 3706, + "бла": 3707, + "че": 3708, + "чу": 3709, + "благо": 3710, + "этому": 3711, + "врем": 3712, + "спе": 3713, + "ном": 3714, + "ений": 3715, + "спо": 3716, + "нас": 3717, + "нет": 3718, + "зу": 3719, + "вед": 3720, + "еще": 3721, + "сказа": 3722, + "сей": 3723, + "ерен": 3724, + "дан": 3725, + "сам": 3726, + "еля": 3727, + "ран": 3728, + "зыва": 3729, + "является": 3730, + "будет": 3731, + "ктив": 3732, + "тре": 3733, + "деле": 3734, + "мот": 3735, + "конферен": 3736, + "лась": 3737, + "час": 3738, + "сторо": 3739, + "кого": 3740, + "ез": 3741, + "ней": 3742, + "ос": 3743, + "лись": 3744, + "разору": 3745, + "пере": 3746, + "сси": 3747, + "ными": 3748, + "проц": 3749, + "голо": 3750, + "чело": 3751, + "боле": 3752, + "челове": 3753, + "сер": 3754, + "пл": 3755, + "чет": 3756, + "стран": 3757, + "пя": 3758, + "был": 3759, + "кла": 3760, + "тов": 3761, + "жд": 3762, + "дела": 3763, + "ера": 3764, + "уже": 3765, + "совет": 3766, + "ген": 3767, + "безопасности": 3768, + "ца": 3769, + "седа": 3770, + "поз": 3771, + "ответ": 3772, + "проблем": 3773, + "нако": 3774, + "тем": 3775, + "доста": 3776, + "пы": 3777, + "ща": 3778, + "вой": 3779, + "сущест": 3780, + "необходимо": 3781, + "быть": 3782, + "может": 3783, + "дем": 3784, + "чтобы": 3785, + "ек": 3786, + "чер": 3787, + "усили": 3788, + "рес": 3789, + "руд": 3790, + "единенных": 3791, + "доб": 3792, + "дости": 3793, + "ствен": 3794, + "ядер": 3795, + "годня": 3796, + "каза": 3797, + "сегодня": 3798, + "сейчас": 3799, + "только": 3800, + "вод": 3801, + "есь": 3802, + "много": 3803, + "буду": 3804, + "ев": 3805, + "есть": 3806, + "три": 3807, + "общест": 3808, + "явл": 3809, + "высту": 3810, + "ред": 3811, + "счит": 3812, + "сит": 3813, + "делега": 3814, + "лож": 3815, + "этот": 3816, + "фор": 3817, + "клю": 3818, + "возмож": 3819, + "вания": 3820, + "бли": 3821, + "или": 3822, + "вз": 3823, + "наций": 3824, + "ского": 3825, + "приня": 3826, + "пла": 3827, + "оч": 3828, + "иться": 3829, + "сте": 3830, + "наши": 3831, + "которые": 3832, + "ар": 3833, + "имеет": 3834, + "сот": 3835, + "знач": 3836, + "перь": 3837, + "следу": 3838, + "ены": 3839, + "таки": 3840, + "объединенных": 3841, + "стро": 3842, + "теперь": 3843, + "бле": 3844, + "благодар": 3845, + "разв": 3846, + "ан": 3847, + "жива": 3848, + "очень": 3849, + "ят": 3850, + "без": 3851, + "обес": 3852, + "гро": 3853, + "лось": 3854, + "сы": 3855, + "организации": 3856, + "член": 3857, + "того": 3858, + "ональ": 3859, + "жда": 3860, + "всех": 3861, + "свя": 3862, + "более": 3863, + "сов": 3864, + "когда": 3865, + "вот": 3866, + "кре": 3867, + "кры": 3868, + "поэтому": 3869, + "воль": 3870, + "ой": 3871, + "генера": 3872, + "чем": 3873, + "лы": 3874, + "полити": 3875, + "вен": 3876, + "конференции": 3877, + "процес": 3878, + "бя": 3879, + "ите": 3880, + "отно": 3881, + "развити": 3882, + "аф": 3883, + "ющ": 3884, + "вно": 3885, + "мир": 3886, + "нии": 3887, + "кая": 3888, + "ас": 3889, + "ительно": 3890, + "вто": 3891, + "ением": 3892, + "генераль": 3893, + "прот": 3894, + "всем": 3895, + "самбле": 3896, + "ассамбле": 3897, + "ом": 3898, + "зд": 3899, + "смот": 3900, + "реги": 3901, + "чего": 3902, + "однако": 3903, + "усилия": 3904, + "действи": 3905, + "чно": 3906, + "уча": 3907, + "образ": 3908, + "вос": 3909, + "эта": 3910, + "перего": 3911, + "говор": 3912, + "вам": 3913, + "моло": 3914, + "время": 3915, + "дь": 3916, + "хотел": 3917, + "гру": 3918, + "заявл": 3919, + "предоста": 3920, + "поль": 3921, + "нее": 3922, + "резо": 3923, + "перегово": 3924, + "резолю": 3925, + "крет": 3926, + "поддерж": 3927, + "обеспе": 3928, + "него": 3929, + "представит": 3930, + "наде": 3931, + "кри": 3932, + "чь": 3933, + "проек": 3934, + "лет": 3935, + "други": 3936, + "_": 3937, + "،": 3938, + "؛": 3939, + "؟": 3940, + "ء": 3941, + "آ": 3942, + "أ": 3943, + "ؤ": 3944, + "إ": 3945, + "ئ": 3946, + "ا": 3947, + "ب": 3948, + "ة": 3949, + "ت": 3950, + "ث": 3951, + "ج": 3952, + "ح": 3953, + "خ": 3954, + "د": 3955, + "ذ": 3956, + "ر": 3957, + "ز": 3958, + "س": 3959, + "ش": 3960, + "ص": 3961, + "ض": 3962, + "ط": 3963, + "ظ": 3964, + "ع": 3965, + "غ": 3966, + "ـ": 3967, + "ف": 3968, + "ق": 3969, + "ك": 3970, + "ل": 3971, + "م": 3972, + "ن": 3973, + "ه": 3974, + "و": 3975, + "ى": 3976, + "ي": 3977, + "ً": 3978, + "ٌ": 3979, + "ٍ": 3980, + "َ": 3981, + "ُ": 3982, + "ِ": 3983, + "ّ": 3984, + "ْ": 3985, + "ٰ": 3986, + "چ": 3987, + "ڨ": 3988, + "ک": 3989, + "ھ": 3990, + "ی": 3991, + "ۖ": 3992, + "ۗ": 3993, + "ۘ": 3994, + "ۚ": 3995, + "ۛ": 3996, + "—": 3997, + "☭": 3998, + "ﺃ": 3999, + "ﻻ": 4000, + "ال": 4001, + "َا": 4002, + "وَ": 4003, + "َّ": 4004, + "ِي": 4005, + "أَ": 4006, + "لَ": 4007, + "نَ": 4008, + "الْ": 4009, + "هُ": 4010, + "ُو": 4011, + "ما": 4012, + "نْ": 4013, + "من": 4014, + "عَ": 4015, + "نا": 4016, + "لا": 4017, + "مَ": 4018, + "تَ": 4019, + "فَ": 4020, + "أن": 4021, + "لي": 4022, + "مِ": 4023, + "ان": 4024, + "في": 4025, + "رَ": 4026, + "يَ": 4027, + "هِ": 4028, + "مْ": 4029, + "قَ": 4030, + "بِ": 4031, + "لى": 4032, + "ين": 4033, + "إِ": 4034, + "لِ": 4035, + "وا": 4036, + "كَ": 4037, + "ها": 4038, + "ًا": 4039, + "مُ": 4040, + "ون": 4041, + "الم": 4042, + "بَ": 4043, + "يا": 4044, + "ذا": 4045, + "سا": 4046, + "الل": 4047, + "مي": 4048, + "يْ": 4049, + "را": 4050, + "ري": 4051, + "لك": 4052, + "مَا": 4053, + "نَّ": 4054, + "لم": 4055, + "إن": 4056, + "ست": 4057, + "وم": 4058, + "َّا": 4059, + "لَا": 4060, + "هم": 4061, + "ِّ": 4062, + "كُ": 4063, + "كان": 4064, + "سَ": 4065, + "با": 4066, + "دي": 4067, + "حَ": 4068, + "عْ": 4069, + "بي": 4070, + "الأ": 4071, + "ول": 4072, + "فِي": 4073, + "رِ": 4074, + "دا": 4075, + "مِنْ": 4076, + "ُونَ": 4077, + "وْ": 4078, + "هَا": 4079, + "ُّ": 4080, + "الس": 4081, + "الَ": 4082, + "ني": 4083, + "لْ": 4084, + "تُ": 4085, + "هل": 4086, + "رة": 4087, + "دَ": 4088, + "سْ": 4089, + "تِ": 4090, + "نَا": 4091, + "رْ": 4092, + "اللَّ": 4093, + "سامي": 4094, + "كن": 4095, + "كل": 4096, + "هَ": 4097, + "عَلَ": 4098, + "على": 4099, + "مع": 4100, + "إلى": 4101, + "قد": 4102, + "الر": 4103, + "ُوا": 4104, + "ير": 4105, + "عن": 4106, + "يُ": 4107, + "نِ": 4108, + "بْ": 4109, + "الح": 4110, + "هُمْ": 4111, + "قا": 4112, + "ذه": 4113, + "الت": 4114, + "ِينَ": 4115, + "جَ": 4116, + "هذا": 4117, + "عد": 4118, + "الع": 4119, + "دْ": 4120, + "قَالَ": 4121, + "رُ": 4122, + "يم": 4123, + "ية": 4124, + "نُ": 4125, + "خَ": 4126, + "رب": 4127, + "الك": 4128, + "وَا": 4129, + "أنا": 4130, + "ةِ": 4131, + "الن": 4132, + "حد": 4133, + "عِ": 4134, + "تا": 4135, + "هو": 4136, + "فا": 4137, + "عا": 4138, + "الش": 4139, + "لُ": 4140, + "يت": 4141, + "ذَا": 4142, + "يع": 4143, + "الذ": 4144, + "حْ": 4145, + "الص": 4146, + "إِنَّ": 4147, + "جا": 4148, + "علي": 4149, + "كَا": 4150, + "بُ": 4151, + "تع": 4152, + "وق": 4153, + "مل": 4154, + "لَّ": 4155, + "يد": 4156, + "أخ": 4157, + "رف": 4158, + "تي": 4159, + "الِ": 4160, + "ّا": 4161, + "ذلك": 4162, + "أَنْ": 4163, + "سِ": 4164, + "توم": 4165, + "مر": 4166, + "مَنْ": 4167, + "بل": 4168, + "الق": 4169, + "الله": 4170, + "ِيَ": 4171, + "كم": 4172, + "ذَ": 4173, + "عل": 4174, + "حب": 4175, + "سي": 4176, + "عُ": 4177, + "الج": 4178, + "الد": 4179, + "شَ": 4180, + "تك": 4181, + "فْ": 4182, + "صَ": 4183, + "لل": 4184, + "دِ": 4185, + "بر": 4186, + "فِ": 4187, + "ته": 4188, + "أع": 4189, + "تْ": 4190, + "قْ": 4191, + "الْأَ": 4192, + "ئِ": 4193, + "عَنْ": 4194, + "ور": 4195, + "حا": 4196, + "الَّ": 4197, + "مت": 4198, + "فر": 4199, + "دُ": 4200, + "هنا": 4201, + "وَأَ": 4202, + "تب": 4203, + "ةُ": 4204, + "أي": 4205, + "سب": 4206, + "ريد": 4207, + "وج": 4208, + "كُمْ": 4209, + "حِ": 4210, + "كْ": 4211, + "در": 4212, + "َاء": 4213, + "هذه": 4214, + "الط": 4215, + "الْمُ": 4216, + "دة": 4217, + "قل": 4218, + "غَ": 4219, + "يوم": 4220, + "الَّذ": 4221, + "كر": 4222, + "تر": 4223, + "كِ": 4224, + "كي": 4225, + "عَلَى": 4226, + "رَب": 4227, + "عة": 4228, + "قُ": 4229, + "جْ": 4230, + "فض": 4231, + "لة": 4232, + "هْ": 4233, + "رَا": 4234, + "وَلَ": 4235, + "الْمَ": 4236, + "أَنَّ": 4237, + "يَا": 4238, + "أُ": 4239, + "شي": 4240, + "اللَّهُ": 4241, + "لَى": 4242, + "قِ": 4243, + "أت": 4244, + "عَلَيْ": 4245, + "اللَّهِ": 4246, + "الب": 4247, + "ضَ": 4248, + "ةً": 4249, + "قي": 4250, + "ار": 4251, + "بد": 4252, + "خْ": 4253, + "سْتَ": 4254, + "طَ": 4255, + "قَدْ": 4256, + "ذهب": 4257, + "أم": 4258, + "ماذا": 4259, + "وَإِ": 4260, + "ةٌ": 4261, + "ونَ": 4262, + "ليلى": 4263, + "ولا": 4264, + "حُ": 4265, + "هي": 4266, + "صل": 4267, + "الخ": 4268, + "ود": 4269, + "ليس": 4270, + "لدي": 4271, + "قال": 4272, + "كَانَ": 4273, + "مَّ": 4274, + "حي": 4275, + "تم": 4276, + "لن": 4277, + "وَلَا": 4278, + "بع": 4279, + "يمكن": 4280, + "سُ": 4281, + "ةَ": 4282, + "حت": 4283, + "رًا": 4284, + "كا": 4285, + "شا": 4286, + "هِمْ": 4287, + "لَهُ": 4288, + "زَ": 4289, + "داً": 4290, + "مس": 4291, + "كث": 4292, + "الْعَ": 4293, + "جِ": 4294, + "صْ": 4295, + "فَا": 4296, + "له": 4297, + "وي": 4298, + "عَا": 4299, + "هُوَ": 4300, + "بِي": 4301, + "بَا": 4302, + "أس": 4303, + "ثَ": 4304, + "لِي": 4305, + "رض": 4306, + "الرَّ": 4307, + "لِكَ": 4308, + "تَّ": 4309, + "فُ": 4310, + "قة": 4311, + "فعل": 4312, + "مِن": 4313, + "الآ": 4314, + "ثُ": 4315, + "سم": 4316, + "مَّا": 4317, + "بِهِ": 4318, + "تق": 4319, + "خر": 4320, + "لقد": 4321, + "خل": 4322, + "شر": 4323, + "أنت": 4324, + "لَّا": 4325, + "سن": 4326, + "السَّ": 4327, + "الذي": 4328, + "سَا": 4329, + "وما": 4330, + "زل": 4331, + "وب": 4332, + "أْ": 4333, + "إذا": 4334, + "رِي": 4335, + "حة": 4336, + "نِي": 4337, + "الْحَ": 4338, + "وَقَالَ": 4339, + "به": 4340, + "ةٍ": 4341, + "سأ": 4342, + "رٌ": 4343, + "بال": 4344, + "مة": 4345, + "شْ": 4346, + "وت": 4347, + "عند": 4348, + "فس": 4349, + "بَعْ": 4350, + "هر": 4351, + "قط": 4352, + "أح": 4353, + "إنه": 4354, + "وع": 4355, + "فت": 4356, + "غا": 4357, + "هناك": 4358, + "بت": 4359, + "مِنَ": 4360, + "سر": 4361, + "ذَلِكَ": 4362, + "رس": 4363, + "حدث": 4364, + "غْ": 4365, + "ِّي": 4366, + "الإ": 4367, + "وَيَ": 4368, + "جل": 4369, + "است": 4370, + "قِي": 4371, + "عب": 4372, + "وس": 4373, + "يش": 4374, + "الَّذِينَ": 4375, + "تاب": 4376, + "دِي": 4377, + "جب": 4378, + "كون": 4379, + "بن": 4380, + "الث": 4381, + "لَيْ": 4382, + "بعد": 4383, + "وَالْ": 4384, + "فَأَ": 4385, + "عم": 4386, + "هُم": 4387, + "تن": 4388, + "ذْ": 4389, + "أص": 4390, + "أين": 4391, + "رَبِّ": 4392, + "الذين": 4393, + "إِن": 4394, + "بين": 4395, + "جُ": 4396, + "عَلَيْهِ": 4397, + "حَا": 4398, + "لو": 4399, + "ستط": 4400, + "ظر": 4401, + "لَمْ": 4402, + "ءِ": 4403, + "كُل": 4404, + "طل": 4405, + "تَا": 4406, + "ضُ": 4407, + "كنت": 4408, + "لًا": 4409, + "مٌ": 4410, + "قبل": 4411, + "ــ": 4412, + "ذِ": 4413, + "قَوْ": 4414, + "صِ": 4415, + "مًا": 4416, + "كانت": 4417, + "صا": 4418, + "يق": 4419, + "الف": 4420, + "النا": 4421, + "مٍ": 4422, + "إِنْ": 4423, + "النَّ": 4424, + "جد": 4425, + "وَمَا": 4426, + "تت": 4427, + "بح": 4428, + "مكان": 4429, + "كيف": 4430, + "ّة": 4431, + "الا": 4432, + "جَا": 4433, + "أو": 4434, + "ساعد": 4435, + "ضِ": 4436, + "إلا": 4437, + "راً": 4438, + "قَا": 4439, + "رأ": 4440, + "عت": 4441, + "أحد": 4442, + "هد": 4443, + "ضا": 4444, + "طر": 4445, + "أق": 4446, + "ماء": 4447, + "دَّ": 4448, + "البا": 4449, + "مُو": 4450, + "أَوْ": 4451, + "طا": 4452, + "قُو": 4453, + "خِ": 4454, + "تل": 4455, + "ستطيع": 4456, + "دَا": 4457, + "النَّا": 4458, + "إلَى": 4459, + "وَتَ": 4460, + "هَذَا": 4461, + "بة": 4462, + "عليك": 4463, + "جر": 4464, + "المن": 4465, + "زا": 4466, + "رٍ": 4467, + "دع": 4468, + "ًّا": 4469, + "سة": 4470, + "ثُمَّ": 4471, + "شيء": 4472, + "الغ": 4473, + "تح": 4474, + "رُونَ": 4475, + "اليوم": 4476, + "مِي": 4477, + "نُوا": 4478, + "أر": 4479, + "تُمْ": 4480, + "عر": 4481, + "يف": 4482, + "أب": 4483, + "دًا": 4484, + "صَا": 4485, + "التَّ": 4486, + "أريد": 4487, + "الز": 4488, + "يَوْ": 4489, + "إلي": 4490, + "جي": 4491, + "يَعْ": 4492, + "فضل": 4493, + "الإن": 4494, + "أنه": 4495, + "1": 4496, + "2": 4497, + "3": 4498, + "4": 4499, + "5": 4500, + "·": 4501, + "×": 4502, + "̃": 4503, + "̌": 4504, + "ε": 4505, + "λ": 4506, + "μ": 4507, + "•": 4508, + "‧": 4509, + "─": 4510, + "□": 4511, + "、": 4512, + "。": 4513, + "〈": 4514, + "〉": 4515, + "《": 4516, + "》": 4517, + "「": 4518, + "」": 4519, + "『": 4520, + "』": 4521, + "ア": 4522, + "オ": 4523, + "カ": 4524, + "チ": 4525, + "ド": 4526, + "ベ": 4527, + "ャ": 4528, + "ヤ": 4529, + "ン": 4530, + "・": 4531, + "ー": 4532, + "ㄟ": 4533, + "！": 4534, + "（": 4535, + "）": 4536, + "，": 4537, + "－": 4538, + "／": 4539, + "：": 4540, + "；": 4541, + "？": 4542, + "ｐ": 4543, + "i4": 4544, + "zh": 4545, + "i2": 4546, + "ng1": 4547, + "u4": 4548, + "i1": 4549, + "ng2": 4550, + "u3": 4551, + "de5": 4552, + "e4": 4553, + "i3": 4554, + "ng4": 4555, + "an4": 4556, + "shi4": 4557, + "an2": 4558, + "u2": 4559, + "u1": 4560, + "ng3": 4561, + "a1": 4562, + "an1": 4563, + "e2": 4564, + "a4": 4565, + "ei4": 4566, + "ong1": 4567, + "ai4": 4568, + "ao4": 4569, + "ang1": 4570, + "an3": 4571, + "wei4": 4572, + "uo2": 4573, + "n1": 4574, + "en2": 4575, + "ao3": 4576, + "e1": 4577, + "qi": 4578, + "eng2": 4579, + "zho": 4580, + "ang3": 4581, + "ang4": 4582, + "ang2": 4583, + "uo4": 4584, + "ge4": 4585, + "yi1": 4586, + "guo2": 4587, + "a3": 4588, + "he2": 4589, + "e3": 4590, + "yi2": 4591, + "di4": 4592, + "zhong1": 4593, + "bu4": 4594, + "ai2": 4595, + "n2": 4596, + "zai4": 4597, + "shi2": 4598, + "eng1": 4599, + "ren2": 4600, + "ong2": 4601, + "xian4": 4602, + "xu": 4603, + "n4": 4604, + "li4": 4605, + "en4": 4606, + "yu2": 4607, + "ei2": 4608, + "yi2ge4": 4609, + "ou4": 4610, + "ei3": 4611, + "ui4": 4612, + "a2": 4613, + "you3": 4614, + "ao1": 4615, + "da4": 4616, + "cheng2": 4617, + "en1": 4618, + "eng4": 4619, + "yi4": 4620, + "si1": 4621, + "zhi4": 4622, + "jia1": 4623, + "yuan2": 4624, + "ta1": 4625, + "de5yi2ge4": 4626, + "ke1": 4627, + "shu3": 4628, + "xi1": 4629, + "ji2": 4630, + "ao2": 4631, + "ou3": 4632, + "ong4": 4633, + "xia4": 4634, + "ai1": 4635, + "gong1": 4636, + "zhi1": 4637, + "en3": 4638, + "wei2": 4639, + "xue2": 4640, + "qu1": 4641, + "zhou1": 4642, + "er3": 4643, + "ming2": 4644, + "zhong3": 4645, + "li3": 4646, + "wu4": 4647, + "yi3": 4648, + "uo1": 4649, + "e5": 4650, + "ji4": 4651, + "xing2": 4652, + "jian4": 4653, + "hua4": 4654, + "yu3": 4655, + "uo3": 4656, + "ji1": 4657, + "ai3": 4658, + "zuo4": 4659, + "hou4": 4660, + "hui4": 4661, + "ei1": 4662, + "nian2": 4663, + "qi2": 4664, + "dao4": 4665, + "sheng1": 4666, + "de2": 4667, + "dai4": 4668, + "uan2": 4669, + "zhe4": 4670, + "zheng4": 4671, + "ben3": 4672, + "shang4": 4673, + "zhu3": 4674, + "bei4": 4675, + "ye4": 4676, + "chu1": 4677, + "zhan4": 4678, + "le5": 4679, + "lai2": 4680, + "shi3": 4681, + "nan2": 4682, + "ren4": 4683, + "you2": 4684, + "ke4": 4685, + "ba1": 4686, + "fu4": 4687, + "dui4": 4688, + "ya4": 4689, + "mei3": 4690, + "zi4": 4691, + "xin1": 4692, + "jing1": 4693, + "zhu": 4694, + "n3": 4695, + "yong4": 4696, + "mu4": 4697, + "jiao4": 4698, + "ye3": 4699, + "jin4": 4700, + "bian4": 4701, + "lu4": 4702, + "qi1": 4703, + "she4": 4704, + "xiang1": 4705, + "ong3": 4706, + "shu4": 4707, + "dong4": 4708, + "suo3": 4709, + "guan1": 4710, + "san1": 4711, + "te4": 4712, + "duo1": 4713, + "fu2": 4714, + "min2": 4715, + "la1": 4716, + "zhi2": 4717, + "zhen4": 4718, + "ou1": 4719, + "wu3": 4720, + "ma3": 4721, + "i5": 4722, + "zi5": 4723, + "ju4": 4724, + "er4": 4725, + "yao4": 4726, + "xia4de5yi2ge4": 4727, + "si4": 4728, + "tu2": 4729, + "shan1": 4730, + "zui4": 4731, + "yin1": 4732, + "er2": 4733, + "tong2": 4734, + "dong1": 4735, + "yu4": 4736, + "yan2": 4737, + "qian2": 4738, + "shu3xia4de5yi2ge4": 4739, + "jun1": 4740, + "ke3": 4741, + "wen2": 4742, + "fa3": 4743, + "luo2": 4744, + "zhu4": 4745, + "xi4": 4746, + "kou3": 4747, + "bei3": 4748, + "jian1": 4749, + "fa1": 4750, + "dian4": 4751, + "jiang1": 4752, + "wei4yu2": 4753, + "xiang4": 4754, + "zhi3": 4755, + "eng3": 4756, + "fang1": 4757, + "lan2": 4758, + "shu": 4759, + "ri4": 4760, + "lian2": 4761, + "shou3": 4762, + "qiu2": 4763, + "jin1": 4764, + "huo4": 4765, + "shu3xia4de5yi2ge4zhong3": 4766, + "fen1": 4767, + "nei4": 4768, + "gai1": 4769, + "mei3guo2": 4770, + "un2": 4771, + "ge2": 4772, + "bao3": 4773, + "qing1": 4774, + "gao1": 4775, + "tai2": 4776, + "xiao3": 4777, + "jie2": 4778, + "tian1": 4779, + "chang2": 4780, + "quan2": 4781, + "lie4": 4782, + "hai3": 4783, + "fei1": 4784, + "ti3": 4785, + "jue2": 4786, + "ou2": 4787, + "ci3": 4788, + "zu2": 4789, + "ni2": 4790, + "biao3": 4791, + "zhong1guo2": 4792, + "du4": 4793, + "yue4": 4794, + "xing4": 4795, + "sheng4": 4796, + "che1": 4797, + "dan1": 4798, + "jie1": 4799, + "lin2": 4800, + "ping2": 4801, + "fu3": 4802, + "gu3": 4803, + "jie4": 4804, + "v3": 4805, + "sheng3": 4806, + "na4": 4807, + "yuan4": 4808, + "zhang3": 4809, + "guan3": 4810, + "dao3": 4811, + "zu3": 4812, + "ding4": 4813, + "dian3": 4814, + "ceng2": 4815, + "ren2kou3": 4816, + "tai4": 4817, + "tong1": 4818, + "guo4": 4819, + "neng2": 4820, + "chang3": 4821, + "hua2": 4822, + "liu2": 4823, + "ying1": 4824, + "xiao4": 4825, + "ci4": 4826, + "bian4hua4": 4827, + "liang3": 4828, + "gong4": 4829, + "zhong4": 4830, + "de5yi1": 4831, + "se4": 4832, + "kai1": 4833, + "wang2": 4834, + "jiu4": 4835, + "shi1": 4836, + "shou4": 4837, + "mei2": 4838, + "feng1": 4839, + "ze2": 4840, + "tu2shi4": 4841, + "ti2": 4842, + "qi4": 4843, + "jiu3": 4844, + "shen1": 4845, + "zhe3": 4846, + "ren2kou3bian4hua4": 4847, + "ren2kou3bian4hua4tu2shi4": 4848, + "di4qu1": 4849, + "yang2": 4850, + "men5": 4851, + "long2": 4852, + "bing4": 4853, + "chan3": 4854, + "zhu1": 4855, + "wei3": 4856, + "wai4": 4857, + "xing1": 4858, + "bo1": 4859, + "bi3": 4860, + "tang2": 4861, + "hua1": 4862, + "bo2": 4863, + "shui3": 4864, + "shu1": 4865, + "dou1": 4866, + "sai4": 4867, + "chao2": 4868, + "bi4": 4869, + "ling2": 4870, + "lei4": 4871, + "da4xue2": 4872, + "fen4": 4873, + "shu3de5": 4874, + "mu3": 4875, + "jiao1": 4876, + "dang1": 4877, + "cheng1": 4878, + "tong3": 4879, + "nv3": 4880, + "qi3": 4881, + "yan3": 4882, + "mian4": 4883, + "luo4": 4884, + "jing4": 4885, + "ge1": 4886, + "ru4": 4887, + "dan4": 4888, + "ri4ben3": 4889, + "pu3": 4890, + "yun4": 4891, + "huang2": 4892, + "wo3": 4893, + "lv": 4894, + "hai2": 4895, + "shi4yi1": 4896, + "xie1": 4897, + "ying3": 4898, + "wu2": 4899, + "shen2": 4900, + "wang3": 4901, + "guang3": 4902, + "liu4": 4903, + "su4": 4904, + "shi4zhen4": 4905, + "can1": 4906, + "cao3": 4907, + "xia2": 4908, + "ka3": 4909, + "da2": 4910, + "hu4": 4911, + "ban4": 4912, + "dang3": 4913, + "hu2": 4914, + "zong3": 4915, + "deng3": 4916, + "de5yi2ge4shi4zhen4": 4917, + "chuan2": 4918, + "mo4": 4919, + "zhang1": 4920, + "ban1": 4921, + "mo2": 4922, + "cha2": 4923, + "ce4": 4924, + "zhu3yao4": 4925, + "tou2": 4926, + "ju2": 4927, + "shi4wei4yu2": 4928, + "sa4": 4929, + "un1": 4930, + "ke3yi3": 4931, + "du1": 4932, + "han4": 4933, + "liang4": 4934, + "sha1": 4935, + "jia3": 4936, + "zi1": 4937, + "lv4": 4938, + "fu1": 4939, + "xian1": 4940, + "xu4": 4941, + "guang1": 4942, + "meng2": 4943, + "bao4": 4944, + "you4": 4945, + "rong2": 4946, + "zhi1yi1": 4947, + "wei1": 4948, + "mao2": 4949, + "guo2jia1": 4950, + "cong2": 4951, + "gou4": 4952, + "tie3": 4953, + "zhen1": 4954, + "du2": 4955, + "bian1": 4956, + "ci2": 4957, + "qu3": 4958, + "fan4": 4959, + "xiang3": 4960, + "men2": 4961, + "ju1": 4962, + "hong2": 4963, + "zi3": 4964, + "ta1men5": 4965, + "ji3": 4966, + "zong1": 4967, + "zhou1de5yi2ge4shi4zhen4": 4968, + "tuan2": 4969, + "jing3": 4970, + "gong1si1": 4971, + "xie4": 4972, + "li2": 4973, + "li4shi3": 4974, + "bao1": 4975, + "gang3": 4976, + "gui1": 4977, + "zheng1": 4978, + "zhi2wu4": 4979, + "ta1de5": 4980, + "pin3": 4981, + "zhuan1": 4982, + "chong2": 4983, + "shi3yong4": 4984, + "wa3": 4985, + "shuo1": 4986, + "chuan1": 4987, + "lei2": 4988, + "wan1": 4989, + "huo2": 4990, + "su1": 4991, + "zao3": 4992, + "gai3": 4993, + "qu4": 4994, + "gu4": 4995, + "xi2": 4996, + "hang2": 4997, + "ying4": 4998, + "cun1": 4999, + "gen1": 5000, + "ying2": 5001, + "ting2": 5002, + "cheng2shi4": 5003, + "jiang3": 5004, + "ling3": 5005, + "lun2": 5006, + "bu4fen4": 5007, + "deng1": 5008, + "xuan3": 5009, + "dong4wu4": 5010, + "de2guo2": 5011, + "xian3": 5012, + "fan3": 5013, + "zhe5": 5014, + "han2": 5015, + "hao4": 5016, + "mi4": 5017, + "ran2": 5018, + "qin1": 5019, + "tiao2": 5020, + "zhan3": 5021, + "[ar]": 5022, + "[zh-cn]": 5023, + "¡": 5024, + "é": 5025, + "shi": 5026, + "tsu": 5027, + "teki": 5028, + "nai": 5029, + "aru": 5030, + "uu": 5031, + "kai": 5032, + "shite": 5033, + "mono": 5034, + "koto": 5035, + "kara": 5036, + "shita": 5037, + "suru": 5038, + "masu": 5039, + "tai": 5040, + "ware": 5041, + "shin": 5042, + "oku": 5043, + "yuu": 5044, + "iru": 5045, + "jiko": 5046, + "desu": 5047, + "rare": 5048, + "shou": 5049, + "sha": 5050, + "sekai": 5051, + "kyou": 5052, + "mashita": 5053, + "nara": 5054, + "kei": 5055, + "ita": 5056, + "ari": 5057, + "itsu": 5058, + "kono": 5059, + "naka": 5060, + "chou": 5061, + "sore": 5062, + "naru": 5063, + "gaku": 5064, + "reba": 5065, + "hito": 5066, + "sai": 5067, + "nan": 5068, + "dai": 5069, + "tsuku": 5070, + "shiki": 5071, + "sare": 5072, + "naku": 5073, + "jun": 5074, + "kaku": 5075, + "zai": 5076, + "wata": 5077, + "shuu": 5078, + "ii": 5079, + "kare": 5080, + "shii": 5081, + "made": 5082, + "sho": 5083, + "kereba": 5084, + "shika": 5085, + "ichi": 5086, + "deki": 5087, + "nin": 5088, + "wareware": 5089, + "nakereba": 5090, + "oite": 5091, + "yaku": 5092, + "mujun": 5093, + "yoku": 5094, + "butsu": 5095, + "omo": 5096, + "gae": 5097, + "naranai": 5098, + "tachi": 5099, + "chuu": 5100, + "kangae": 5101, + "toki": 5102, + "koro": 5103, + "mujunteki": 5104, + "naga": 5105, + "jin": 5106, + "shima": 5107, + "iku": 5108, + "imasu": 5109, + "hon": 5110, + "kae": 5111, + "kore": 5112, + "kita": 5113, + "datta": 5114, + "jitsu": 5115, + "mae": 5116, + "toku": 5117, + "douitsu": 5118, + "ritsu": 5119, + "kyuu": 5120, + "hyou": 5121, + "rareta": 5122, + "keisei": 5123, + "kkan": 5124, + "rareru": 5125, + "mou": 5126, + "doko": 5127, + "ryou": 5128, + "dake": 5129, + "nakatta": 5130, + "soko": 5131, + "tabe": 5132, + "hana": 5133, + "fuku": 5134, + "yasu": 5135, + "wataku": 5136, + "yama": 5137, + "kyo": 5138, + "genzai": 5139, + "boku": 5140, + "ata": 5141, + "kawa": 5142, + "masen": 5143, + "juu": 5144, + "natte": 5145, + "watakushi": 5146, + "yotte": 5147, + "hai": 5148, + "jishin": 5149, + "rete": 5150, + "oka": 5151, + "kagaku": 5152, + "natta": 5153, + "karu": 5154, + "nari": 5155, + "mata": 5156, + "kuru": 5157, + "gai": 5158, + "kari": 5159, + "shakai": 5160, + "koui": 5161, + "yori": 5162, + "setsu": 5163, + "reru": 5164, + "tokoro": 5165, + "jutsu": 5166, + "saku": 5167, + "ttai": 5168, + "ningen": 5169, + "tame": 5170, + "kankyou": 5171, + "ooku": 5172, + "watashi": 5173, + "tsukuru": 5174, + "sugi": 5175, + "jibun": 5176, + "shitsu": 5177, + "keru": 5178, + "kishi": 5179, + "shikashi": 5180, + "moto": 5181, + "mari": 5182, + "itte": 5183, + "deshita": 5184, + "nde": 5185, + "arimasu": 5186, + "koe": 5187, + "zettai": 5188, + "kkanteki": 5189, + "rekishi": 5190, + "dekiru": 5191, + "tsuka": 5192, + "itta": 5193, + "kobutsu": 5194, + "miru": 5195, + "shoku": 5196, + "shimasu": 5197, + "gijutsu": 5198, + "gyou": 5199, + "joushiki": 5200, + "atta": 5201, + "hodo": 5202, + "koko": 5203, + "tsukurareta": 5204, + "zoku": 5205, + "hitei": 5206, + "koku": 5207, + "rekishiteki": 5208, + "kete": 5209, + "kako": 5210, + "nagara": 5211, + "kakaru": 5212, + "shutai": 5213, + "haji": 5214, + "taku": 5215, + "douitsuteki": 5216, + "mete": 5217, + "tsuu": 5218, + "sarete": 5219, + "genjitsu": 5220, + "bai": 5221, + "nawa": 5222, + "jikan": 5223, + "waru": 5224, + "rt": 5225, + "atsu": 5226, + "soku": 5227, + "kouiteki": 5228, + "kata": 5229, + "tetsu": 5230, + "gawa": 5231, + "kedo": 5232, + "reta": 5233, + "sayou": 5234, + "tteru": 5235, + "tori": 5236, + "kimi": 5237, + "mura": 5238, + "sareru": 5239, + "machi": 5240, + "kya": 5241, + "osa": 5242, + "konna": 5243, + "aku": 5244, + "sareta": 5245, + "ipp": 5246, + "shiku": 5247, + "uchi": 5248, + "hitotsu": 5249, + "hatara": 5250, + "tachiba": 5251, + "shiro": 5252, + "katachi": 5253, + "tomo": 5254, + "ete": 5255, + "meru": 5256, + "nichi": 5257, + "dare": 5258, + "katta": 5259, + "eru": 5260, + "suki": 5261, + "ooki": 5262, + "maru": 5263, + "moku": 5264, + "oko": 5265, + "kangaerareru": 5266, + "oto": 5267, + "tanni": 5268, + "tada": 5269, + "taiteki": 5270, + "motte": 5271, + "kinou": 5272, + "shinai": 5273, + "kki": 5274, + "tari": 5275, + "ranai": 5276, + "kkou": 5277, + "mirai": 5278, + "ppon": 5279, + "goto": 5280, + "hitsu": 5281, + "teru": 5282, + "mochi": 5283, + "katsu": 5284, + "nyuu": 5285, + "zuka": 5286, + "tsuite": 5287, + "nomi": 5288, + "sugu": 5289, + "kuda": 5290, + "tetsugaku": 5291, + "ika": 5292, + "ronri": 5293, + "oki": 5294, + "nippon": 5295, + "shimashita": 5296, + "chishiki": 5297, + "chokkanteki": 5298, + "suko": 5299, + "kuu": 5300, + "arou": 5301, + "katte": 5302, + "kuri": 5303, + "inai": 5304, + "hyougen": 5305, + "ishiki": 5306, + "doku": 5307, + "atte": 5308, + "atara": 5309, + "wari": 5310, + "kao": 5311, + "seisan": 5312, + "hanashi": 5313, + "kake": 5314, + "naji": 5315, + "sunawa": 5316, + "sunawachi": 5317, + "ugo": 5318, + "suu": 5319, + "bara": 5320, + "hiro": 5321, + "iwa": 5322, + "betsu": 5323, + "yoi": 5324, + "seru": 5325, + "shiteru": 5326, + "rarete": 5327, + "toshi": 5328, + "seki": 5329, + "tairitsu": 5330, + "wakara": 5331, + "tokyo": 5332, + "kka": 5333, + "kyoku": 5334, + "iro": 5335, + "mite": 5336, + "saki": 5337, + "kanji": 5338, + "mita": 5339, + "sube": 5340, + "ryoku": 5341, + "matta": 5342, + "kudasai": 5343, + "omoi": 5344, + "wareru": 5345, + "hitsuyou": 5346, + "kashi": 5347, + "renai": 5348, + "kankei": 5349, + "gatte": 5350, + "ochi": 5351, + "motsu": 5352, + "sonzai": 5353, + "taishite": 5354, + "ame": 5355, + "seimei": 5356, + "kano": 5357, + "giri": 5358, + "kangaeru": 5359, + "yue": 5360, + "asa": 5361, + "onaji": 5362, + "yoru": 5363, + "niku": 5364, + "osaka": 5365, + "sukoshi": 5366, + "tama": 5367, + "kanojo": 5368, + "kite": 5369, + "mondai": 5370, + "amari": 5371, + "eki": 5372, + "kojin": 5373, + "haya": 5374, + "dete": 5375, + "atarashii": 5376, + "awa": 5377, + "gakkou": 5378, + "tsuzu": 5379, + "shukan": 5380, + "imashita": 5381, + "atae": 5382, + "darou": 5383, + "hataraku": 5384, + "gata": 5385, + "dachi": 5386, + "matsu": 5387, + "arimasen": 5388, + "seibutsu": 5389, + "mitsu": 5390, + "heya": 5391, + "yasui": 5392, + "deni": 5393, + "noko": 5394, + "haha": 5395, + "domo": 5396, + "kami": 5397, + "sudeni": 5398, + "nao": 5399, + "raku": 5400, + "ike": 5401, + "meta": 5402, + "kodomo": 5403, + "soshite": 5404, + "game": 5405, + "bakari": 5406, + "tote": 5407, + "hatsu": 5408, + "mise": 5409, + "mokuteki": 5410, + "dakara": 5411, + "[ja]": 5412, + "ő": 5413, + "ű": 5414, + "そ": 5415, + "な": 5416, + "ん": 5417, + "포": 5418, + "�": 5419, + "gy": 5420, + "eg": 5421, + "cs": 5422, + "ál": 5423, + "egy": 5424, + "át": 5425, + "ott": 5426, + "ett": 5427, + "meg": 5428, + "hogy": 5429, + "ég": 5430, + "ól": 5431, + "nek": 5432, + "volt": 5433, + "ág": 5434, + "nk": 5435, + "ék": 5436, + "ít": 5437, + "ák": 5438, + "ud": 5439, + "szer": 5440, + "mind": 5441, + "oz": 5442, + "ép": 5443, + "ért": 5444, + "mond": 5445, + "szt": 5446, + "nak": 5447, + "ől": 5448, + "csak": 5449, + "oly": 5450, + "áll": 5451, + "ány": 5452, + "mint": 5453, + "már": 5454, + "ött": 5455, + "nagy": 5456, + "ész": 5457, + "azt": 5458, + "elő": 5459, + "tud": 5460, + "ény": 5461, + "áz": 5462, + "még": 5463, + "köz": 5464, + "ely": 5465, + "ség": 5466, + "hoz": 5467, + "uk": 5468, + "kez": 5469, + "ám": 5470, + "aj": 5471, + "unk": 5472, + "vagy": 5473, + "szem": 5474, + "ember": 5475, + "fog": 5476, + "mert": 5477, + "ös": 5478, + "ság": 5479, + "leg": 5480, + "ünk": 5481, + "hát": 5482, + "ony": 5483, + "ezt": 5484, + "minden": 5485, + "ült": 5486, + "jó": 5487, + "kis": 5488, + "áj": 5489, + "úgy": 5490, + "most": 5491, + "ír": 5492, + "itt": 5493, + "elt": 5494, + "mondta": 5495, + "kell": 5496, + "ált": 5497, + "érd": 5498, + "tö": 5499, + "vár": 5500, + "lát": 5501, + "ők": 5502, + "vet": 5503, + "után": 5504, + "két": 5505, + "nap": 5506, + "ív": 5507, + "ály": 5508, + "vég": 5509, + "ök": 5510, + "dul": 5511, + "néz": 5512, + "ában": 5513, + "kül": 5514, + "akkor": 5515, + "szél": 5516, + "új": 5517, + "olyan": 5518, + "ked": 5519, + "hely": 5520, + "tör": 5521, + "ból": 5522, + "elm": 5523, + "ára": 5524, + "ló": 5525, + "volna": 5526, + "lehet": 5527, + "ebb": 5528, + "sok": 5529, + "olt": 5530, + "eket": 5531, + "bor": 5532, + "fej": 5533, + "gond": 5534, + "akar": 5535, + "fél": 5536, + "úl": 5537, + "otta": 5538, + "valami": 5539, + "jel": 5540, + "éd": 5541, + "arc": 5542, + "hall": 5543, + "föl": 5544, + "ába": 5545, + "olg": 5546, + "kir": 5547, + "old": 5548, + "kérd": 5549, + "jár": 5550, + "úr": 5551, + "zs": 5552, + "élet": 5553, + "ját": 5554, + "ov": 5555, + "éz": 5556, + "vil": 5557, + "őr": 5558, + "ög": 5559, + "lesz": 5560, + "koz": 5561, + "ább": 5562, + "király": 5563, + "eng": 5564, + "igaz": 5565, + "haj": 5566, + "kod": 5567, + "ról": 5568, + "több": 5569, + "szó": 5570, + "ében": 5571, + "öt": 5572, + "nyi": 5573, + "szól": 5574, + "gondol": 5575, + "egész": 5576, + "így": 5577, + "ős": 5578, + "obb": 5579, + "osan": 5580, + "ből": 5581, + "abb": 5582, + "őt": 5583, + "nál": 5584, + "kép": 5585, + "aztán": 5586, + "tart": 5587, + "beszél": 5588, + "előtt": 5589, + "aszt": 5590, + "maj": 5591, + "kör": 5592, + "hang": 5593, + "íz": 5594, + "incs": 5595, + "év": 5596, + "ód": 5597, + "ók": 5598, + "hozz": 5599, + "okat": 5600, + "nagyon": 5601, + "ház": 5602, + "ped": 5603, + "ezte": 5604, + "etlen": 5605, + "neki": 5606, + "majd": 5607, + "szony": 5608, + "ának": 5609, + "felé": 5610, + "egyszer": 5611, + "adt": 5612, + "gyer": 5613, + "amikor": 5614, + "foly": 5615, + "szak": 5616, + "őd": 5617, + "hú": 5618, + "ász": 5619, + "amely": 5620, + "ére": 5621, + "ilyen": 5622, + "oda": 5623, + "ják": 5624, + "tár": 5625, + "ával": 5626, + "lak": 5627, + "gyan": 5628, + "ély": 5629, + "út": 5630, + "kezd": 5631, + "mell": 5632, + "mikor": 5633, + "hez": 5634, + "való": 5635, + "szeret": 5636, + "rend": 5637, + "vissza": 5638, + "fő": 5639, + "asszony": 5640, + "ről": 5641, + "pedig": 5642, + "szép": 5643, + "ták": 5644, + "öv": 5645, + "világ": 5646, + "maga": 5647, + "szik": 5648, + "éj": 5649, + "ént": 5650, + "jött": 5651, + "szí": 5652, + "gat": 5653, + "ettem": 5654, + "hány": 5655, + "ást": 5656, + "ahol": 5657, + "őket": 5658, + "hár": 5659, + "nő": 5660, + "csi": 5661, + "talál": 5662, + "elte": 5663, + "látt": 5664, + "tört": 5665, + "hagy": 5666, + "esz": 5667, + "nél": 5668, + "kut": 5669, + "lány": 5670, + "amit": 5671, + "ső": 5672, + "ellen": 5673, + "magát": 5674, + "ugyan": 5675, + "külön": 5676, + "asz": 5677, + "mindig": 5678, + "lép": 5679, + "talán": 5680, + "szor": 5681, + "illan": 5682, + "nincs": 5683, + "vagyok": 5684, + "telen": 5685, + "ismer": 5686, + "isten": 5687, + "ított": 5688, + "jobb": 5689, + "ves": 5690, + "dult": 5691, + "juk": 5692, + "szen": 5693, + "öm": 5694, + "lett": 5695, + "egyik": 5696, + "bár": 5697, + "szi": 5698, + "szív": 5699, + "azon": 5700, + "eszt": 5701, + "föld": 5702, + "kuty": 5703, + "pillan": 5704, + "fér": 5705, + "től": 5706, + "tű": 5707, + "ébe": 5708, + "tött": 5709, + "barát": 5710, + "íg": 5711, + "ahogy": 5712, + "eh": 5713, + "ep": 5714, + "jelent": 5715, + "tat": 5716, + "szeg": 5717, + "mintha": 5718, + "egyen": 5719, + "szab": 5720, + "bizony": 5721, + "jon": 5722, + "öreg": 5723, + "dolg": 5724, + "csap": 5725, + "tiszt": 5726, + "állt": 5727, + "ancs": 5728, + "idő": 5729, + "ügy": 5730, + "miért": 5731, + "ót": 5732, + "csin": 5733, + "ének": 5734, + "vér": 5735, + "jól": 5736, + "alatt": 5737, + "mely": 5738, + "semmi": 5739, + "nyug": 5740, + "vág": 5741, + "követ": 5742, + "össze": 5743, + "mad": 5744, + "acs": 5745, + "fiú": 5746, + "másik": 5747, + "jön": 5748, + "szám": 5749, + "rész": 5750, + "kér": 5751, + "ével": 5752, + "[hu]": 5753, + "%": 5754, + "0": 5755, + "6": 5756, + "7": 5757, + "8": 5758, + "9": 5759, + "A": 5760, + "B": 5761, + "C": 5762, + "D": 5763, + "E": 5764, + "F": 5765, + "G": 5766, + "H": 5767, + "I": 5768, + "J": 5769, + "K": 5770, + "L": 5771, + "M": 5772, + "N": 5773, + "O": 5774, + "P": 5775, + "Q": 5776, + "R": 5777, + "S": 5778, + "T": 5779, + "U": 5780, + "V": 5781, + "W": 5782, + "X": 5783, + "Y": 5784, + "Z": 5785, + "Ł": 5786, + "α": 5787, + "ς": 5788, + "♥": 5789, + "か": 5790, + "ズ": 5791, + "因": 5792, + "国": 5793, + "怎": 5794, + "抱": 5795, + "推": 5796, + "有": 5797, + "樣": 5798, + "為": 5799, + "群": 5800, + "麼": 5801, + "eo": 5802, + "eul": 5803, + "eun": 5804, + "eon": 5805, + "ae": 5806, + "yeon": 5807, + "yeo": 5808, + "ui": 5809, + "hae": 5810, + "geo": 5811, + "neun": 5812, + "ssda": 5813, + "seo": 5814, + "eong": 5815, + "kk": 5816, + "jeo": 5817, + "deul": 5818, + "eum": 5819, + "yeong": 5820, + "geos": 5821, + "hag": 5822, + "aneun": 5823, + "iss": 5824, + "dae": 5825, + "eob": 5826, + "eol": 5827, + "geu": 5828, + "jeong": 5829, + "sae": 5830, + "doe": 5831, + "geul": 5832, + "eulo": 5833, + "bn": 5834, + "sang": 5835, + "bnida": 5836, + "haneun": 5837, + "jeog": 5838, + "saeng": 5839, + "ineun": 5840, + "anh": 5841, + "salam": 5842, + "eom": 5843, + "nae": 5844, + "gwa": 5845, + "yeol": 5846, + "eseo": 5847, + "myeon": 5848, + "ttae": 5849, + "hw": 5850, + "eobs": 5851, + "jang": 5852, + "gw": 5853, + "ileul": 5854, + "yeog": 5855, + "jeon": 5856, + "sig": 5857, + "jag": 5858, + "hago": 5859, + "deun": 5860, + "seong": 5861, + "gag": 5862, + "ham": 5863, + "dang": 5864, + "leul": 5865, + "sil": 5866, + "dong": 5867, + "handa": 5868, + "eossda": 5869, + "aeg": 5870, + "seon": 5871, + "haessda": 5872, + "issda": 5873, + "ege": 5874, + "mul": 5875, + "jung": 5876, + "jig": 5877, + "issneun": 5878, + "geun": 5879, + "seubnida": 5880, + "won": 5881, + "daneun": 5882, + "eoh": 5883, + "deo": 5884, + "gam": 5885, + "jal": 5886, + "haeng": 5887, + "yang": 5888, + "bang": 5889, + "jae": 5890, + "saenggag": 5891, + "hage": 5892, + "sog": 5893, + "eoss": 5894, + "jasin": 5895, + "jil": 5896, + "eog": 5897, + "gyeong": 5898, + "gong": 5899, + "deon": 5900, + "haess": 5901, + "eung": 5902, + "joh": 5903, + "nal": 5904, + "myeong": 5905, + "eona": 5906, + "igo": 5907, + "gyeol": 5908, + "yag": 5909, + "gwan": 5910, + "uli": 5911, + "yong": 5912, + "lyeo": 5913, + "jog": 5914, + "eohge": 5915, + "bog": 5916, + "tong": 5917, + "manh": 5918, + "jeol": 5919, + "geol": 5920, + "aga": 5921, + "naneun": 5922, + "uneun": 5923, + "cheol": 5924, + "dol": 5925, + "bad": 5926, + "hamyeon": 5927, + "yeossda": 5928, + "ibnida": 5929, + "gye": 5930, + "eos": 5931, + "hwal": 5932, + "salamdeul": 5933, + "jiman": 5934, + "dangsin": 5935, + "jib": 5936, + "ttaemun": 5937, + "ib": 5938, + "eneun": 5939, + "eug": 5940, + "jeom": 5941, + "geuleon": 5942, + "hwa": 5943, + "assda": 5944, + "beob": 5945, + "bae": 5946, + "yeoss": 5947, + "chin": 5948, + "chaeg": 5949, + "geon": 5950, + "naega": 5951, + "iga": 5952, + "sigan": 5953, + "gil": 5954, + "hyeon": 5955, + "lyeog": 5956, + "gug": 5957, + "pyeon": 5958, + "wae": 5959, + "jul": 5960, + "seul": 5961, + "deung": 5962, + "hajiman": 5963, + "eumyeon": 5964, + "pil": 5965, + "nyeon": 5966, + "tae": 5967, + "pyo": 5968, + "jineun": 5969, + "beon": 5970, + "hada": 5971, + "seol": 5972, + "sip": 5973, + "daleun": 5974, + "salm": 5975, + "gyo": 5976, + "cheon": 5977, + "hagi": 5978, + "cheoleom": 5979, + "gal": 5980, + "ila": 5981, + "kkaji": 5982, + "anhneun": 5983, + "habnida": 5984, + "tteon": 5985, + "haeseo": 5986, + "doenda": 5987, + "ttal": 5988, + "ilo": 5989, + "seub": 5990, + "byeon": 5991, + "myeo": 5992, + "beol": 5993, + "jeung": 5994, + "chim": 5995, + "hwang": 5996, + "euneun": 5997, + "jong": 5998, + "boda": 5999, + "nol": 6000, + "neom": 6001, + "buteo": 6002, + "jigeum": 6003, + "eobsda": 6004, + "daelo": 6005, + "yul": 6006, + "pyeong": 6007, + "seoneun": 6008, + "salang": 6009, + "seut": 6010, + "heom": 6011, + "hyang": 6012, + "gwang": 6013, + "eobsneun": 6014, + "hwag": 6015, + "gess": 6016, + "jagi": 6017, + "ileon": 6018, + "wihae": 6019, + "daehan": 6020, + "gaji": 6021, + "meog": 6022, + "jyeo": 6023, + "chaj": 6024, + "byeong": 6025, + "eod": 6026, + "gyeo": 6027, + "eoji": 6028, + "gul": 6029, + "modeun": 6030, + "insaeng": 6031, + "geulae": 6032, + "sasil": 6033, + "sib": 6034, + "chal": 6035, + "ilago": 6036, + "geum": 6037, + "doeneun": 6038, + "bol": 6039, + "gajang": 6040, + "geuligo": 6041, + "hyeong": 6042, + "haengbog": 6043, + "chul": 6044, + "chae": 6045, + "mang": 6046, + "dam": 6047, + "choe": 6048, + "sijag": 6049, + "cheong": 6050, + "ilaneun": 6051, + "ulineun": 6052, + "aen": 6053, + "kke": 6054, + "munje": 6055, + "teu": 6056, + "geuneun": 6057, + "bge": 6058, + "cheo": 6059, + "baeg": 6060, + "jug": 6061, + "sangdae": 6062, + "geugeos": 6063, + "dog": 6064, + "eus": 6065, + "jab": 6066, + "hyeo": 6067, + "tteohge": 6068, + "chil": 6069, + "swi": 6070, + "jileul": 6071, + "chang": 6072, + "ganeun": 6073, + "iji": 6074, + "dago": 6075, + "yohan": 6076, + "teug": 6077, + "ppun": 6078, + "aleul": 6079, + "haengdong": 6080, + "sesang": 6081, + "edo": 6082, + "mandeul": 6083, + "amyeon": 6084, + "kkae": 6085, + "bag": 6086, + "ideul": 6087, + "pum": 6088, + "meol": 6089, + "neul": 6090, + "hamkke": 6091, + "chung": 6092, + "dab": 6093, + "yug": 6094, + "sag": 6095, + "gwangye": 6096, + "ileohge": 6097, + "balo": 6098, + "neunde": 6099, + "hamyeo": 6100, + "geuleoh": 6101, + "anila": 6102, + "bangbeob": 6103, + "dasi": 6104, + "byeol": 6105, + "gyeon": 6106, + "gamjeong": 6107, + "oneul": 6108, + "janeun": 6109, + "yeom": 6110, + "lago": 6111, + "igi": 6112, + "hwan": 6113, + "teul": 6114, + "eoseo": 6115, + "sik": 6116, + "jaga": 6117, + "geuleom": 6118, + "geuleona": 6119, + "jeongdo": 6120, + "gyeog": 6121, + "geuleohge": 6122, + "geudeul": 6123, + "eut": 6124, + "imyeon": 6125, + "jjae": 6126, + "keun": 6127, + "isang": 6128, + "malhaessda": 6129, + "euge": 6130, + "nop": 6131, + "ingan": 6132, + "bomyeon": 6133, + "taeg": 6134, + "dwi": 6135, + "saneun": 6136, + "wan": 6137, + "anhgo": 6138, + "nugu": 6139, + "sung": 6140, + "damyeon": 6141, + "adeul": 6142, + "peul": 6143, + "ttala": 6144, + "geosdo": 6145, + "aji": 6146, + "meon": 6147, + "eumyeo": 6148, + "dolog": 6149, + "neung": 6150, + "modu": 6151, + "[ko]": 6152, + "\u0014": 6153, + "\u0016": 6154, + "$": 6155, + "*": 6156, + "|": 6157, + "°": 6158, + "º": 6159, + "ँ": 6160, + "ं": 6161, + "ः": 6162, + "अ": 6163, + "आ": 6164, + "इ": 6165, + "ई": 6166, + "उ": 6167, + "ऊ": 6168, + "ऋ": 6169, + "ऎ": 6170, + "ए": 6171, + "ऐ": 6172, + "ऑ": 6173, + "ऒ": 6174, + "ओ": 6175, + "औ": 6176, + "क": 6177, + "ख": 6178, + "ग": 6179, + "घ": 6180, + "ङ": 6181, + "च": 6182, + "छ": 6183, + "ज": 6184, + "झ": 6185, + "ञ": 6186, + "ट": 6187, + "ठ": 6188, + "ड": 6189, + "ढ": 6190, + "ण": 6191, + "त": 6192, + "थ": 6193, + "द": 6194, + "ध": 6195, + "न": 6196, + "ऩ": 6197, + "प": 6198, + "फ": 6199, + "ब": 6200, + "भ": 6201, + "म": 6202, + "य": 6203, + "र": 6204, + "ऱ": 6205, + "ल": 6206, + "ळ": 6207, + "व": 6208, + "श": 6209, + "ष": 6210, + "स": 6211, + "ह": 6212, + "़": 6213, + "ा": 6214, + "ि": 6215, + "ी": 6216, + "ु": 6217, + "ू": 6218, + "ृ": 6219, + "ॄ": 6220, + "ॅ": 6221, + "ॆ": 6222, + "े": 6223, + "ै": 6224, + "ॉ": 6225, + "ॊ": 6226, + "ो": 6227, + "ौ": 6228, + "्": 6229, + "ॐ": 6230, + "ॖ": 6231, + "क़": 6232, + "ख़": 6233, + "ग़": 6234, + "ज़": 6235, + "ड़": 6236, + "ढ़": 6237, + "फ़": 6238, + "य़": 6239, + "ॠ": 6240, + "।": 6241, + "॥": 6242, + "०": 6243, + "१": 6244, + "२": 6245, + "३": 6246, + "४": 6247, + "५": 6248, + "६": 6249, + "७": 6250, + "८": 6251, + "९": 6252, + "॰": 6253, + "ॲ": 6254, + "​": 6255, + "‌": 6256, + "‍": 6257, + "‎": 6258, + "₹": 6259, + "के": 6260, + "है": 6261, + "ें": 6262, + "्र": 6263, + "ार": 6264, + "ने": 6265, + "या": 6266, + "में": 6267, + "से": 6268, + "की": 6269, + "का": 6270, + "ों": 6271, + "ता": 6272, + "कर": 6273, + "स्": 6274, + "कि": 6275, + "को": 6276, + "र्": 6277, + "ना": 6278, + "क्": 6279, + "ही": 6280, + "और": 6281, + "पर": 6282, + "ते": 6283, + "हो": 6284, + "प्र": 6285, + "ान": 6286, + "्य": 6287, + "ला": 6288, + "वा": 6289, + "ले": 6290, + "सा": 6291, + "हैं": 6292, + "लि": 6293, + "जा": 6294, + "हा": 6295, + "भी": 6296, + "वि": 6297, + "इस": 6298, + "ती": 6299, + "न्": 6300, + "रा": 6301, + "मा": 6302, + "दे": 6303, + "दि": 6304, + "बा": 6305, + "ति": 6306, + "था": 6307, + "नि": 6308, + "कार": 6309, + "एक": 6310, + "हीं": 6311, + "हु": 6312, + "ंग": 6313, + "ैं": 6314, + "नी": 6315, + "सी": 6316, + "अप": 6317, + "त्": 6318, + "नहीं": 6319, + "री": 6320, + "मे": 6321, + "मु": 6322, + "ित": 6323, + "तो": 6324, + "पा": 6325, + "ली": 6326, + "लिए": 6327, + "गा": 6328, + "ल्": 6329, + "रह": 6330, + "रे": 6331, + "क्ष": 6332, + "मैं": 6333, + "सम": 6334, + "उस": 6335, + "जि": 6336, + "त्र": 6337, + "मि": 6338, + "चा": 6339, + "ोग": 6340, + "सं": 6341, + "द्": 6342, + "सि": 6343, + "आप": 6344, + "तु": 6345, + "दा": 6346, + "कु": 6347, + "यों": 6348, + "वे": 6349, + "जी": 6350, + "्या": 6351, + "उन": 6352, + "िक": 6353, + "ये": 6354, + "भा": 6355, + "्ट": 6356, + "हम": 6357, + "स्ट": 6358, + "शा": 6359, + "ड़": 6360, + "ंद": 6361, + "खा": 6362, + "म्": 6363, + "श्": 6364, + "यह": 6365, + "सक": 6366, + "पू": 6367, + "किया": 6368, + "अपने": 6369, + "रू": 6370, + "सु": 6371, + "मी": 6372, + "हि": 6373, + "जो": 6374, + "थे": 6375, + "रि": 6376, + "दी": 6377, + "थी": 6378, + "गी": 6379, + "लोग": 6380, + "गया": 6381, + "तर": 6382, + "न्ह": 6383, + "च्": 6384, + "वार": 6385, + "बी": 6386, + "प्": 6387, + "दो": 6388, + "टी": 6389, + "शि": 6390, + "करने": 6391, + "गे": 6392, + "ैसे": 6393, + "इन": 6394, + "ंड": 6395, + "साथ": 6396, + "पु": 6397, + "बे": 6398, + "बार": 6399, + "वी": 6400, + "अन": 6401, + "हर": 6402, + "उन्ह": 6403, + "होता": 6404, + "जब": 6405, + "कुछ": 6406, + "मान": 6407, + "क्र": 6408, + "बि": 6409, + "पह": 6410, + "फि": 6411, + "सर": 6412, + "ारी": 6413, + "रो": 6414, + "दू": 6415, + "कहा": 6416, + "तक": 6417, + "शन": 6418, + "ब्": 6419, + "स्थ": 6420, + "वह": 6421, + "बाद": 6422, + "ओं": 6423, + "गु": 6424, + "ज्": 6425, + "्रे": 6426, + "गर": 6427, + "रहे": 6428, + "वर्": 6429, + "हू": 6430, + "ार्": 6431, + "पी": 6432, + "बहु": 6433, + "मुझ": 6434, + "्रा": 6435, + "दिया": 6436, + "सब": 6437, + "करते": 6438, + "अपनी": 6439, + "बहुत": 6440, + "कह": 6441, + "टे": 6442, + "हुए": 6443, + "किसी": 6444, + "रहा": 6445, + "ष्ट": 6446, + "ज़": 6447, + "बना": 6448, + "सो": 6449, + "डि": 6450, + "कोई": 6451, + "व्य": 6452, + "बात": 6453, + "रु": 6454, + "वो": 6455, + "मुझे": 6456, + "द्ध": 6457, + "चार": 6458, + "मेरे": 6459, + "वर": 6460, + "्री": 6461, + "जाता": 6462, + "नों": 6463, + "प्रा": 6464, + "देख": 6465, + "टा": 6466, + "क्या": 6467, + "अध": 6468, + "लग": 6469, + "लो": 6470, + "पि": 6471, + "यु": 6472, + "चे": 6473, + "जिस": 6474, + "ंत": 6475, + "ानी": 6476, + "पै": 6477, + "जन": 6478, + "ारे": 6479, + "ची": 6480, + "मिल": 6481, + "दु": 6482, + "देश": 6483, + "च्छ": 6484, + "ष्": 6485, + "सू": 6486, + "खे": 6487, + "चु": 6488, + "िया": 6489, + "लगा": 6490, + "बु": 6491, + "उनके": 6492, + "ज्ञ": 6493, + "क्षा": 6494, + "तरह": 6495, + "्यादा": 6496, + "वाले": 6497, + "पूर्": 6498, + "मैंने": 6499, + "काम": 6500, + "रूप": 6501, + "होती": 6502, + "उप": 6503, + "जान": 6504, + "प्रकार": 6505, + "भार": 6506, + "मन": 6507, + "हुआ": 6508, + "टर": 6509, + "हूँ": 6510, + "परि": 6511, + "पास": 6512, + "अनु": 6513, + "राज": 6514, + "लोगों": 6515, + "अब": 6516, + "समझ": 6517, + "डी": 6518, + "मौ": 6519, + "शु": 6520, + "चि": 6521, + "पे": 6522, + "कृ": 6523, + "सकते": 6524, + "मह": 6525, + "योग": 6526, + "दर्": 6527, + "उसे": 6528, + "ंध": 6529, + "डा": 6530, + "जाए": 6531, + "बो": 6532, + "ूल": 6533, + "मो": 6534, + "ोंने": 6535, + "ंस": 6536, + "तुम": 6537, + "पहले": 6538, + "बता": 6539, + "तथा": 6540, + "यो": 6541, + "गई": 6542, + "उत्": 6543, + "सकता": 6544, + "कम": 6545, + "ज्यादा": 6546, + "रख": 6547, + "समय": 6548, + "ारा": 6549, + "अगर": 6550, + "स्त": 6551, + "चल": 6552, + "फिर": 6553, + "वारा": 6554, + "करना": 6555, + "शी": 6556, + "गए": 6557, + "बन": 6558, + "ौर": 6559, + "होने": 6560, + "चाह": 6561, + "खु": 6562, + "हाँ": 6563, + "उन्हें": 6564, + "उन्होंने": 6565, + "छो": 6566, + "म्ह": 6567, + "प्रति": 6568, + "निक": 6569, + "वन": 6570, + "्यू": 6571, + "रही": 6572, + "तुम्ह": 6573, + "जैसे": 6574, + "ियों": 6575, + "क्यों": 6576, + "लों": 6577, + "फ़": 6578, + "ंत्र": 6579, + "होते": 6580, + "क्ति": 6581, + "त्य": 6582, + "कर्": 6583, + "कई": 6584, + "वं": 6585, + "किन": 6586, + "पो": 6587, + "कारण": 6588, + "ड़ी": 6589, + "भि": 6590, + "इसके": 6591, + "बर": 6592, + "उसके": 6593, + "द्वारा": 6594, + "शे": 6595, + "कॉ": 6596, + "दिन": 6597, + "न्न": 6598, + "ड़ा": 6599, + "स्व": 6600, + "निर्": 6601, + "मुख": 6602, + "लिया": 6603, + "टि": 6604, + "ज्ञान": 6605, + "क्त": 6606, + "द्र": 6607, + "ग्": 6608, + "क्स": 6609, + "मै": 6610, + "गो": 6611, + "जे": 6612, + "ट्र": 6613, + "मार": 6614, + "त्व": 6615, + "धार": 6616, + "भाव": 6617, + "करता": 6618, + "खि": 6619, + "कं": 6620, + "चाहि": 6621, + "यर": 6622, + "प्त": 6623, + "कों": 6624, + "ंच": 6625, + "जु": 6626, + "मत": 6627, + "अच्छ": 6628, + "हुई": 6629, + "कभी": 6630, + "लेकिन": 6631, + "भू": 6632, + "अपना": 6633, + "दूस": 6634, + "चाहिए": 6635, + "यू": 6636, + "घर": 6637, + "सबसे": 6638, + "मेरी": 6639, + "नाम": 6640, + "ढ़": 6641, + "ंट": 6642, + "ेंगे": 6643, + "बै": 6644, + "फा": 6645, + "एवं": 6646, + "यी": 6647, + "ग्र": 6648, + "क्षे": 6649, + "आज": 6650, + "आपको": 6651, + "भाग": 6652, + "ठा": 6653, + "कै": 6654, + "भारत": 6655, + "उनकी": 6656, + "पहु": 6657, + "सभी": 6658, + "धा": 6659, + "णा": 6660, + "सान": 6661, + "होगा": 6662, + "तब": 6663, + "संग": 6664, + "पर्": 6665, + "अव": 6666, + "तना": 6667, + "गि": 6668, + "यन": 6669, + "स्था": 6670, + "चित": 6671, + "ट्": 6672, + "छा": 6673, + "जाने": 6674, + "क्षेत्र": 6675, + "वाली": 6676, + "पूर्ण": 6677, + "समा": 6678, + "कारी": 6679, + "[hi]": 6680 + }, + "merges": [ + "t h", + "i n", + "th e", + "a n", + "e r", + "o u", + "r e", + "o n", + "a t", + "e d", + "e n", + "t o", + "in g", + "an d", + "i s", + "a s", + "a l", + "o r", + "o f", + "a r", + "i t", + "e s", + "h e", + "s t", + "l e", + "o m", + "s e", + "b e", + "a d", + "o w", + "l y", + "c h", + "w h", + "th at", + "y ou", + "l i", + "v e", + "a c", + "t i", + "l d", + "m e", + "w as", + "g h", + "i d", + "l l", + "w i", + "en t", + "f or", + "a y", + "r o", + "v er", + "i c", + "h er", + "k e", + "h is", + "n o", + "u t", + "u n", + "i r", + "l o", + "w e", + "r i", + "h a", + "wi th", + "gh t", + "ou t", + "i m", + "i on", + "al l", + "a b", + "on e", + "n e", + "g e", + "ou ld", + "t er", + "m o", + "h ad", + "c e", + "s he", + "g o", + "s h", + "u r", + "a m", + "s o", + "p e", + "m y", + "d e", + "a re", + "b ut", + "om e", + "f r", + "the r", + "f e", + "s u", + "d o", + "c on", + "t e", + "a in", + "er e", + "p o", + "i f", + "the y", + "u s", + "a g", + "t r", + "n ow", + "ou n", + "th is", + "ha ve", + "no t", + "s a", + "i l", + "u p", + "th ing", + "fr om", + "a p", + "h im", + "ac k", + "at ion", + "an t", + "ou r", + "o p", + "li ke", + "u st", + "es s", + "b o", + "o k", + "u l", + "in d", + "e x", + "c om", + "s ome", + "the re", + "er s", + "c o", + "re s", + "m an", + "ar d", + "p l", + "w or", + "w ay", + "ti on", + "f o", + "c a", + "w ere", + "b y", + "at e", + "p ro", + "t ed", + "oun d", + "ow n", + "w ould", + "t s", + "wh at", + "q u", + "al ly", + "i ght", + "c k", + "g r", + "wh en", + "v en", + "c an", + "ou gh", + "in e", + "en d", + "p er", + "ou s", + "o d", + "id e", + "k now", + "t y", + "ver y", + "s i", + "a k", + "wh o", + "ab out", + "i ll", + "the m", + "es t", + "re d", + "y e", + "c ould", + "on g", + "you r", + "the ir", + "e m", + "j ust", + "o ther", + "in to", + "an y", + "wh i", + "u m", + "t w", + "as t", + "d er", + "d id", + "i e", + "be en", + "ac e", + "in k", + "it y", + "b ack", + "t ing", + "b r", + "mo re", + "a ke", + "p p", + "the n", + "s p", + "e l", + "u se", + "b l", + "sa id", + "o ver", + "ge t", + "e n", + "e r", + "c h", + "e i", + "i e", + "u n", + "i ch", + "ei n", + "s t", + "a n", + "t e", + "g e", + "a u", + "i n", + "s ch", + "d er", + "un d", + "d ie", + "d a", + "e s", + "a l", + "d en", + "a r", + "g en", + "z u", + "d e", + "h r", + "o n", + "t en", + "e l", + "o r", + "m i", + "s ie", + "da s", + "a t", + "b e", + "ein e", + "ich t", + "b er", + "l e", + "a ch", + "v er", + "s e", + "au f", + "w i", + "s o", + "t er", + "l ich", + "c k", + "u r", + "n icht", + "m m", + "b en", + "a s", + "w ar", + "r e", + "mi t", + "s ich", + "i g", + "l l", + "au s", + "i st", + "w ie", + "o ch", + "un g", + "an n", + "ü r", + "h n", + "i hr", + "s a", + "s en", + "t z", + "de m", + "ei t", + "u m", + "h at", + "wi r", + "v on", + "h a", + "s p", + "w ei", + "i er", + "r o", + "h er", + "r a", + "ein en", + "n e", + "v or", + "al s", + "an d", + "al l", + "w as", + "w o", + "r ei", + "st e", + "l ie", + "au ch", + "d u", + "d es", + "k o", + "ü ber", + "a m", + "b ei", + "h en", + "h m", + "l ei", + "a ber", + "w en", + "h l", + "g er", + "i m", + "u t", + "n ach", + "h e", + "i s", + "b r", + "f t", + "en t", + "i mm", + "j e", + "sch en", + "w er", + "s er", + "a b", + "ä n", + "m e", + "s ein", + "i t", + "o l", + "ch t", + "f ür", + "k l", + "f f", + "eine m", + "n en", + "w e", + "j a", + "u s", + "n och", + "hat te", + "t r", + "p f", + "h in", + "d i", + "ch en", + "b l", + "m an", + "r ü", + "ie l", + "s el", + "das s", + "i hn", + "mi r", + "sch l", + "ö n", + "g an", + "g t", + "ein er", + "st en", + "m ich", + "wen n", + "el l", + "g te", + "in d", + "m al", + "ge l", + "k en", + "n ur", + "mm en", + "f ü", + "er n", + "ö r", + "un ter", + "f r", + "an der", + "g r", + "i l", + "d ur", + "u ch", + "f e", + "t a", + "m en", + "m ach", + "d och", + "t i", + "dur ch", + "o s", + "g l", + "h al", + "ihr e", + "w ä", + "imm er", + "i hm", + "k ann", + "or t", + "d ann", + "l an", + "tz t", + "o der", + "hr en", + "e t", + "k ön", + "i ck", + "f a", + "in g", + "i r", + "wie der", + "da ß", + "m ein", + "f en", + "gan z", + "die se", + "st er", + "da r", + "w a", + "ge s", + "n a", + "f l", + "i gen", + "sch e", + "un gen", + "me hr", + "ß en", + "o t", + "k on", + "ge w", + "ha ben", + "ge h", + "ä t", + "s ind", + "d r", + "w el", + "un s", + "v o", + "m a", + "u te", + "sch on", + "b es", + "ge sch", + "b t", + "ch e", + "s on", + "o b", + "l a", + "p p", + "rü ck", + "s eine", + "k r", + "f re", + "ei l", + "zu m", + "u l", + "h ier", + "k t", + "i ge", + "sp r", + "k e", + "le ben", + "b st", + "z eit", + "i on", + "g ro", + "den n", + "h o", + "sch a", + "b ar", + "al le", + "ge gen", + "w ür", + "m ü", + "z e", + "wer den", + "je tzt", + "ko mmen", + "n ie", + "s ei", + "h eit", + "so ll", + "g lei", + "m eine", + "wo ll", + "n er", + "ha be", + "w ur", + "lich en", + "p er", + "as sen", + "n te", + "se hen", + "wir d", + "b is", + "g ar", + "i en", + "m us", + "u ß", + "ä r", + "st ell", + "k eit", + "z wei", + "sel bst", + "st a", + "p a", + "sa gte", + "te t", + "k am", + "s sen", + "v iel", + "u g", + "z en", + "h ei", + "m ann", + "wi ll", + "ge b", + "war en", + "ü ck", + "ä ch", + "m er", + "r u", + "w or", + "h au", + "ei gen", + "an g", + "we g", + "bl ick", + "f ra", + "all es", + "k a", + "au gen", + "f in", + "lich e", + "t o", + "un ser", + "der n", + "her r", + "n un", + "v ie", + "ch te", + "wo hl", + "f all", + "h t", + "ü n", + "et was", + "st and", + "en d", + "ä u", + "e m", + "m ö", + "te l", + "r ie", + "d ich", + "die s", + "h and", + "b in", + "ff en", + "nicht s", + "d an", + "p l", + "hn e", + "ihn en", + "es en", + "die ser", + "fr au", + "an t", + "ar t", + "di r", + "i sch", + "er st", + "glei ch", + "ko mm", + "h ör", + "ß e", + "d ig", + "se hr", + "z ei", + "sa m", + "au m", + "h ät", + "in gen", + "g ut", + "b o", + "m ut", + "ck en", + "kon nte", + "st imm", + "p ro", + "zu r", + "i tz", + "wei l", + "wür de", + "f ä", + "kön nen", + "k eine", + "f er", + "i schen", + "vo ll", + "ein es", + "se tz", + "z ie", + "de l", + "te te", + "sein er", + "ier en", + "ge st", + "zu rück", + "wur de", + "sch n", + "p r", + "lie ß", + "t ra", + "m ä", + "gen d", + "f ol", + "i k", + "schl a", + "scha ft", + "at er", + "wei ß", + "s einen", + "l assen", + "l u", + "und en", + "t eil", + "ne u", + "ier t", + "men schen", + "hm en", + "st r", + "g i", + "sa h", + "ihr en", + "el n", + "wei ter", + "ge hen", + "ig er", + "mach t", + "ta g", + "al so", + "hal ten", + "n is", + "ach t", + "ge ben", + "f or", + "o g", + "n at", + "m ar", + "de t", + "o hne", + "h aus", + "t ro", + "an ge", + "l au", + "sp iel", + "t re", + "sch r", + "in n", + "s u", + "l os", + "mach en", + "hät te", + "be g", + "wir k", + "al t", + "g lich", + "te s", + "r icht", + "fre und", + "m o", + "ihr er", + "f el", + "b el", + "so l", + "ein mal", + "e ben", + "h ol", + "h än", + "q u", + "ter n", + "h ö", + "sch w", + "re cht", + "wa hr", + "s einem", + "ste hen", + "hl en", + "in s", + "g ing", + "woll te", + "wi ssen", + "ung s", + "al d", + "as s", + "ja hr", + "m or", + "wel t", + "un der", + "zu sa", + "at ion", + "ko pf", + "lan g", + "hin ter", + "at z", + "st ra", + "an gen", + "an k", + "a de", + "gl au", + "f ach", + "hat ten", + "l o", + "f ort", + "ei cht", + "i ff", + "l er", + "m ei", + "diese m", + "k ein", + "f rei", + "fü hr", + "vo m", + "e s", + "e n", + "a i", + "o u", + "o n", + "l e", + "d e", + "r e", + "q u", + "a n", + "e r", + "en t", + "e t", + "l a", + "n e", + "i l", + "a r", + "i s", + "ai t", + "t e", + "a u", + "i n", + "qu e", + "i t", + "u r", + "s e", + "l es", + "c h", + "c e", + "m e", + "o r", + "ou r", + "a s", + "p r", + "a v", + "o m", + "ai s", + "u n", + "an t", + "ou s", + "t r", + "t i", + "l u", + "o i", + "e u", + "l le", + "s i", + "p ar", + "d es", + "an s", + "m ent", + "é t", + "es t", + "j e", + "u ne", + "a l", + "p as", + "t re", + "qu i", + "d u", + "r i", + "c on", + "s on", + "c om", + "e lle", + "d é", + "p our", + "d ans", + "l i", + "s a", + "r é", + "t ou", + "v ous", + "d i", + "v i", + "a g", + "a m", + "a t", + "ou v", + "a p", + "ti on", + "m on", + "s ur", + "c i", + "o s", + "p lu", + "s u", + "en d", + "a b", + "è re", + "ai n", + "m ais", + "o is", + "r es", + "plu s", + "é e", + "ai ent", + "m p", + "ch e", + "lu i", + "av e", + "ét ait", + "m a", + "s es", + "tou t", + "i r", + "v o", + "a c", + "s er", + "an d", + "f f", + "oi r", + "g r", + "av ait", + "é s", + "m es", + "n ous", + "eu x", + "b i", + "t er", + "c o", + "on s", + "p u", + "c es", + "g e", + "t u", + "le ur", + "pr o", + "d on", + "e ur", + "et te", + "ai re", + "ave c", + "d it", + "t é", + "i e", + "u s", + "il le", + "p er", + "com me", + "c r", + "or t", + "m i", + "e x", + "u x", + "v er", + "m o", + "è s", + "v e", + "au x", + "r a", + "j our", + "il s", + "bi en", + "c ou", + "p e", + "que l", + "p eu", + "c ette", + "t es", + "p o", + "in s", + "c u", + "m ê", + "s o", + "f ait", + "g u", + "m ar", + "ê tre", + "l o", + "it é", + "f r", + "a tion", + "en s", + "b r", + "n i", + "l é", + "d is", + "b le", + "m an", + "n é", + "pu is", + "mê me", + "qu es", + "f i", + "e l", + "ag e", + "g ar", + "m oi", + "en ce", + "on t", + "m ain", + "or s", + "au t", + "an ce", + "v en", + "m é", + "s ans", + "e m", + "s é", + "l on", + "h om", + "r o", + "u t", + "c ar", + "ab le", + "i m", + "de r", + "ch er", + "n o", + "vi e", + "au s", + "b e", + "de ux", + "en f", + "o ù", + "t en", + "p h", + "u re", + "te mp", + "p os", + "r ent", + "p é", + "f aire", + "p i", + "tr es", + "ç a", + "an g", + "end re", + "f or", + "p a", + "b on", + "s ou", + "in t", + "pr é", + "s ent", + "t ant", + "n er", + "c er", + "l à", + "l ais", + "pr ès", + "b re", + "c our", + "p et", + "i on", + "i ne", + "com p", + "l ait", + "tr ouv", + "t a", + "ent re", + "son t", + "de v", + "n u", + "temp s", + "d ou", + "r ait", + "b ou", + "qu and", + "jour s", + "l an", + "er s", + "av oir", + "ét é", + "a le", + "p re", + "f ois", + "or te", + "v é", + "m er", + "n on", + "t ous", + "j us", + "cou p", + "t s", + "hom me", + "ê te", + "a d", + "aus si", + "ur s", + "se u", + "or d", + "o b", + "m in", + "g é", + "co re", + "v a", + "v re", + "en core", + "se m", + "i te", + "au tre", + "pr is", + "peu t", + "u e", + "an te", + "m al", + "g n", + "ré p", + "h u", + "si on", + "vo tre", + "di re", + "e z", + "f em", + "leur s", + "m et", + "f in", + "c ri", + "m is", + "t our", + "r ai", + "j am", + "re gar", + "ri en", + "ver s", + "su is", + "p ouv", + "o p", + "v is", + "gr and", + "ant s", + "c or", + "re r", + "ar d", + "c é", + "t ent", + "pr es", + "v ou", + "f a", + "al ors", + "si eur", + "ai ne", + "le r", + "qu oi", + "f on", + "end ant", + "ar ri", + "eu re", + "a près", + "don c", + "it u", + "l è", + "s ait", + "t oi", + "ch a", + "ai l", + "as se", + "i mp", + "vo y", + "con n", + "p la", + "pet it", + "av ant", + "n om", + "t in", + "don t", + "d a", + "s ous", + "e mp", + "per son", + "el les", + "be au", + "par ti", + "ch o", + "pr it", + "tou jours", + "m en", + "r ais", + "jam ais", + "tr av", + "tion s", + "tr ès", + "v oi", + "r en", + "y eux", + "f er", + "v oir", + "pre mi", + "c a", + "g ne", + "h eure", + "r ou", + "e ff", + "no tre", + "ment s", + "t on", + "f ais", + "ce la", + "i er", + "rép on", + "con s", + "ai r", + "ô t", + "p endant", + "i ci", + "tou te", + "j et", + "p ort", + "ét aient", + "p en", + "h é", + "au tres", + "p ère", + "o c", + "quel ques", + "i que", + "l is", + "fem me", + "j ou", + "te ur", + "mon de", + "u se", + "n es", + "d re", + "a ff", + "r ap", + "par t", + "le ment", + "c la", + "f ut", + "quel que", + "pr endre", + "r ê", + "ai lle", + "s ais", + "ch es", + "le t", + "ch ar", + "è res", + "ent s", + "b er", + "g er", + "mo ins", + "e au", + "a î", + "j eu", + "h eur", + "é es", + "tr i", + "po int", + "m om", + "v ent", + "n ouv", + "gr an", + "tr ois", + "s ant", + "tout es", + "con tre", + "è rent", + "che z", + "ave z", + "û t", + "a lle", + "at t", + "p au", + "p orte", + "ouv er", + "b ar", + "l it", + "f ort", + "o t", + "as s", + "pr és", + "cho se", + "v it", + "mon sieur", + "h ab", + "t ête", + "j u", + "te ment", + "c tion", + "v rai", + "la r", + "c et", + "regar d", + "l ant", + "de m", + "s om", + "mom ent", + "il les", + "p le", + "p s", + "b es", + "m ère", + "c l", + "s our", + "y s", + "tr op", + "en ne", + "jus qu", + "av aient", + "av ais", + "jeu ne", + "de puis", + "person ne", + "f it", + "cer t", + "j o", + "g es", + "ou i", + "r est", + "sem b", + "c ap", + "m at", + "m u", + "lon g", + "fr an", + "f aut", + "it i", + "b li", + "che v", + "pr i", + "ent e", + "ain si", + "ch am", + "l ors", + "c as", + "d o", + "il i", + "b é", + "n os", + "an ge", + "su i", + "r it", + "cr o", + "gu e", + "d e", + "e n", + "e s", + "o s", + "l a", + "e r", + "q u", + "a r", + "a n", + "o n", + "qu e", + "a s", + "o r", + "e l", + "d o", + "a l", + "c i", + "u n", + "r e", + "a b", + "i n", + "t e", + "t o", + "s e", + "d i", + "t r", + "d a", + "c on", + "t a", + "s u", + "m i", + "c o", + "t i", + "l e", + "l os", + "n o", + "l o", + "í a", + "c u", + "c a", + "s i", + "v i", + "m e", + "p or", + "m o", + "p ar", + "r a", + "r i", + "la s", + "c h", + "r o", + "m a", + "p er", + "ó n", + "m en", + "de s", + "un a", + "m p", + "s o", + "ab a", + "p u", + "d os", + "t u", + "g u", + "er a", + "de l", + "h a", + "m u", + "l i", + "en t", + "m b", + "h ab", + "es t", + "g o", + "p a", + "r es", + "par a", + "p o", + "á s", + "m os", + "tr a", + "t en", + "an do", + "p i", + "qu i", + "b i", + "m an", + "co mo", + "v e", + "m ás", + "j o", + "ci ón", + "i s", + "t an", + "v o", + "da d", + "c e", + "a do", + "v er", + "f u", + "ci a", + "c er", + "p e", + "c as", + "c ar", + "men te", + "n i", + "su s", + "t ar", + "n a", + "f i", + "t er", + "z a", + "p ro", + "tr o", + "s a", + "l u", + "b a", + "per o", + "s er", + "c es", + "d as", + "d u", + "s in", + "e mp", + "m ar", + "l la", + "e x", + "á n", + "c or", + "i a", + "v a", + "r an", + "ch o", + "g a", + "y o", + "t os", + "c os", + "mi s", + "l es", + "t es", + "v en", + "h o", + "y a", + "en te", + "on es", + "hab ía", + "n u", + "u s", + "p as", + "h i", + "n os", + "es ta", + "la n", + "m as", + "t or", + "l le", + "h e", + "s on", + "b re", + "p re", + "ab an", + "d or", + "í an", + "i r", + "t as", + "é n", + "r u", + "en do", + "a que", + "er o", + "i o", + "qu é", + "m in", + "c ab", + "j a", + "de r", + "t al", + "é s", + "se ñ", + "or a", + "to do", + "la r", + "d on", + "g ar", + "s al", + "p r", + "cu ando", + "j e", + "h u", + "g un", + "b u", + "g i", + "d ar", + "n e", + "r as", + "de n", + "es to", + "par e", + "p en", + "é l", + "tr as", + "c an", + "b o", + "j os", + "mi en", + "pu e", + "c re", + "co mp", + "p on", + "d ía", + "tr os", + "s ab", + "so bre", + "es e", + "mb re", + "er on", + "a ñ", + "m or", + "f or", + "i do", + "por que", + "el la", + "p ri", + "g ran", + "f a", + "c en", + "di s", + "c ri", + "mu y", + "ch a", + "c al", + "es te", + "h as", + "c ó", + "g ra", + "r os", + "p os", + "o b", + "al l", + "aque l", + "j u", + "p res", + "m er", + "di jo", + "c ía", + "ent re", + "z o", + "ci ones", + "bi en", + "mb i", + "el o", + "t ó", + "in a", + "to dos", + "g en", + "ti en", + "est aba", + "de ci", + "ci o", + "h er", + "ñ o", + "l or", + "nu es", + "me di", + "l en", + "vi da", + "f e", + "al i", + "m on", + "c la", + "d re", + "pu es", + "al es", + "vo l", + "m í", + "r ar", + "b le", + "ci on", + "has ta", + "señ or", + "con o", + "a h", + "di os", + "s en", + "es a", + "ú n", + "v ar", + "s an", + "gu i", + "a c", + "o tros", + "ta do", + "bu en", + "ñ a", + "ti emp", + "ha cer", + "j er", + "f er", + "v u", + "f in", + "an a", + "as í", + "an tes", + "t in", + "ve z", + "mien to", + "j ar", + "la b", + "ch e", + "cas a", + "d r", + "es o", + "e go", + "di ó", + "an te", + "est á", + "m al", + "en cia", + "el i", + "í as", + "tiemp o", + "z ar", + "v an", + "m un", + "er ta", + "ta mbi", + "s í", + "b ar", + "a un", + "al e", + "mis mo", + "ent es", + "vi s", + "man o", + "el e", + "na da", + "se gu", + "me j", + "er ra", + "ab le", + "b e", + "ti r", + "un o", + "don de", + "to da", + "des de", + "r en", + "tambi én", + "cu er", + "per son", + "ho mbre", + "o tro", + "li b", + "tr ar", + "cu al", + "ha y", + "a u", + "ca da", + "t aba", + "i mp", + "men to", + "ten ía", + "qu er", + "er an", + "si emp", + "siemp re", + "er to", + "qu í", + "g os", + "pu és", + "el los", + "des pués", + "nu e", + "g an", + "l lo", + "in ter", + "có mo", + "tr i", + "ah ora", + "us te", + "tr aba", + "la do", + "in o", + "po co", + "er te", + "mu jer", + "i m", + "qui er", + "al gun", + "fu e", + "o jos", + "ent on", + "v os", + "es per", + "mu ch", + "o tra", + "a z", + "a d", + "in g", + "e za", + "a quí", + "ci as", + "gu a", + "mu cho", + "deci r", + "es ti", + "i dad", + "al go", + "e z", + "o cu", + "enton ces", + "di do", + "ent os", + "g ri", + "da do", + "i os", + "so l", + "dos e", + "uste d", + "qui en", + "a mi", + "un to", + "f r", + "mi r", + "mej or", + "b as", + "so lo", + "pre gun", + "tu r", + "al g", + "p la", + "to das", + "par te", + "e mb", + "c to", + "mun do", + "tien e", + "tan te", + "pa lab", + "tr an", + "aque lla", + "ci os", + "aun que", + "a y", + "cu en", + "ten er", + "f un", + "res pon", + "all í", + "x i", + "h an", + "pen s", + "con tra", + "tu ra", + "v al", + "di o", + "tr es", + "t re", + "tan to", + "ca min", + "m ó", + "es p", + "a da", + "í o", + "in s", + "ha cia", + "de j", + "est ar", + "i ón", + "g as", + "b er", + "v as", + "no che", + "é r", + "añ os", + "pa dre", + "gu s", + "á r", + "sin o", + "man os", + "ci do", + "es tu", + "a de", + "hu bi", + "vi r", + "b ri", + "ra z", + "ch i", + "pue de", + "men os", + "hab i", + "ho mb", + "ne ces", + "ma y", + "er os", + "r ía", + "he cho", + "es cu", + "l ti", + "án do", + "b us", + "cos as", + "t ú", + "es pa", + "re ci", + "c tor", + "pri m", + "di a", + "de se", + "mien tras", + "h or", + "fu er", + "i da", + "pos i", + "lan te", + "t on", + "an o", + "est as", + "p li", + "ch ar", + "lu ego", + "si ón", + "ci n", + "ti erra", + "m es", + "gu ar", + "ca do", + "en con", + "pr en", + "may or", + "f al", + "e r", + "o n", + "a n", + "t o", + "d i", + "r e", + "l a", + "i n", + "e n", + "a l", + "t a", + "c h", + "e l", + "r i", + "c o", + "t i", + "t e", + "s i", + "r a", + "u n", + "l e", + "l i", + "ch e", + "r o", + "c i", + "c a", + "s e", + "q u", + "m a", + "p o", + "s o", + "i l", + "d o", + "e s", + "v a", + "p er", + "l o", + "c on", + "d el", + "p a", + "m o", + "s a", + "p i", + "d a", + "m i", + "g i", + "s u", + "d e", + "v i", + "z i", + "m e", + "g li", + "n o", + "m en", + "v o", + "t u", + "n on", + "v e", + "t to", + "s t", + "on e", + "an o", + "ch i", + "er a", + "er e", + "f a", + "c e", + "z a", + "un a", + "b i", + "p re", + "s ta", + "o r", + "a r", + "f i", + "on o", + "t ra", + "n a", + "n el", + "n e", + "p ro", + "t ro", + "al e", + "v er", + "n i", + "c u", + "t ti", + "men te", + "del la", + "t er", + "zi one", + "g u", + "p e", + "t ta", + "an do", + "t à", + "al i", + "u o", + "qu el", + "co m", + "s en", + "co me", + "b a", + "al la", + "p ri", + "d u", + "qu es", + "l u", + "on i", + "g gi", + "pa r", + "s si", + "v en", + "in a", + "g a", + "pi ù", + "ci a", + "i m", + "co r", + "m an", + "in o", + "in i", + "t en", + "r an", + "b b", + "g o", + "s to", + "t re", + "a ve", + "a v", + "s ono", + "er i", + "a c", + "s se", + "er o", + "h a", + "s c", + "su l", + "f or", + "v ano", + "po r", + "s ti", + "su o", + "c chi", + "t an", + "z za", + "an che", + "p u", + "i o", + "t te", + "vo l", + "es s", + "s ci", + "co l", + "r u", + "p en", + "f u", + "al l", + "s so", + "s te", + "se m", + "s sa", + "d en", + "a d", + "t ri", + "de i", + "in e", + "ave va", + "men to", + "z z", + "a mo", + "g no", + "f o", + "un o", + "su a", + "g en", + "ri a", + "g e", + "st ra", + "s ì", + "c er", + "ch é", + "b u", + "a p", + "c en", + "d al", + "on a", + "s pe", + "g ni", + "b o", + "t t", + "del le", + "ques to", + "nel la", + "f f", + "d ere", + "an no", + "del l", + "un i", + "bb e", + "an ti", + "g ra", + "s p", + "en e", + "gi o", + "u to", + "qu al", + "gli a", + "qu ando", + "tu tto", + "c an", + "gli o", + "zi oni", + "ca m", + "h o", + "es so", + "s s", + "mo l", + "a t", + "lo ro", + "per ché", + "co sa", + "du e", + "po i", + "ca r", + "s co", + "ci o", + "to r", + "c co", + "c re", + "a m", + "g na", + "te m", + "pri ma", + "lu i", + "co sì", + "qu e", + "gu ar", + "ess ere", + "an i", + "con o", + "b ra", + "al le", + "m on", + "ri o", + "an co", + "cu i", + "s pi", + "vi a", + "g ran", + "gi or", + "a i", + "bi le", + "u l", + "ggi o", + "f e", + "an te", + "ma i", + "ta re", + "in ter", + "in di", + "re bbe", + "sen za", + "so lo", + "zi o", + "e d", + "en te", + "tu tti", + "sta to", + "zi a", + "d alla", + "tu ra", + "mi a", + "vi ta", + "quel la", + "qu a", + "ma r", + "do ve", + "g h", + "al lo", + "sem pre", + "zz o", + "si a", + "mo r", + "do po", + "por ta", + "d re", + "c cia", + "er ano", + "an ni", + "di o", + "chi a", + "en za", + "pro pri", + "qu i", + "m u", + "m b", + "an da", + "c ca", + "o cchi", + "ques ta", + "f fi", + "le i", + "par te", + "d on", + "r on", + "mi o", + "tan to", + "ri s", + "o gni", + "di s", + "r in", + "fa r", + "men ti", + "t el", + "anco ra", + "f ra", + "fa tto", + "man i", + "sen ti", + "p ra", + "tem po", + "es si", + "b bi", + "f in", + "a re", + "la re", + "per s", + "f on", + "b el", + "so r", + "d er", + "pre n", + "an za", + "di re", + "pi e", + "o ra", + "ver so", + "se gu", + "al tro", + "ta to", + "ca to", + "a to", + "vol ta", + "c c", + "fa re", + "pa re", + "ci ò", + "li b", + "bi li", + "n uo", + "s er", + "quel lo", + "co lo", + "p po", + "ca sa", + "tro va", + "o re", + "f er", + "r ono", + "d es", + "mol to", + "al mente", + "s ca", + "vo le", + "t ali", + "sul la", + "s ce", + "men o", + "an to", + "p un", + "s tu", + "ca pi", + "so l", + "gi u", + "m ini", + "m ano", + "z e", + "pi a", + "par ti", + "s al", + "la vo", + "ver o", + "r si", + "al tri", + "es ti", + "s cia", + "suo i", + "gli e", + "so tto", + "b ene", + "sc ri", + "t ale", + "de gli", + "n u", + "al c", + "uo mo", + "p el", + "f re", + "po te", + "es sa", + "s cu", + "si gno", + "el e", + "st ro", + "u ti", + "di a", + "si one", + "g re", + "f ini", + "ar ri", + "l un", + "c ri", + "e si", + "pa ssa", + "r à", + "men tre", + "an d", + "h anno", + "el o", + "u sci", + "gi a", + "gi à", + "di e", + "m ina", + "b e", + "ti ca", + "gior no", + "t in", + "es se", + "mo do", + "c al", + "s pa", + "propri o", + "l en", + "o ri", + "con tro", + "st ru", + "di ven", + "di sse", + "ra to", + "no i", + "v ere", + "pu ò", + "di ce", + "s an", + "es a", + "c ci", + "se con", + "re n", + "c cio", + "qual che", + "tu tta", + "g g", + "mon do", + "for ma", + "p li", + "m ma", + "pen sa", + "de va", + "tu r", + "fo sse", + "so pra", + "ta mente", + "n ess", + "qu anto", + "ra ga", + "un que", + "ca re", + "st re", + "gran de", + "pi cco", + "guar da", + "b en", + "nel l", + "a ff", + "po ssi", + "pre sen", + "r ò", + "pa ro", + "tu a", + "v in", + "an e", + "a s", + "ste sso", + "da v", + "ne i", + "nel le", + "gh i", + "pi o", + "ta r", + "an a", + "la to", + "si d", + "f ine", + "f uo", + "m er", + "z o", + "qua si", + "ul ti", + "i to", + "su e", + "si e", + "f il", + "allo ra", + "m in", + "ven i", + "t ano", + "el lo", + "d e", + "r a", + "e s", + "d o", + "e n", + "q u", + "c o", + "a s", + "o s", + "e r", + "a r", + "s e", + "qu e", + "a n", + "i n", + "i s", + "t o", + "ã o", + "t e", + "d a", + "m a", + "e l", + "t a", + "o r", + "i a", + "r e", + "e m", + "a l", + "co m", + "p a", + "o u", + "c a", + "u m", + "r o", + "v a", + "t i", + "s o", + "m en", + "n ão", + "h a", + "co n", + "m e", + "r i", + "pa ra", + "p o", + "d i", + "s a", + "v o", + "u ma", + "c i", + "n a", + "p or", + "n o", + "g u", + "s u", + "h o", + "an do", + "t ra", + "e i", + "v i", + "e u", + "i m", + "do s", + "el e", + "r es", + "m o", + "en t", + "f i", + "l a", + "e ra", + "l e", + "de s", + "el a", + "men te", + "l h", + "p er", + "l i", + "ç ão", + "m as", + "t er", + "m u", + "es t", + "v e", + "g o", + "l o", + "u s", + "ma is", + "v er", + "c ê", + "in ha", + "vo cê", + "f a", + "t u", + "c u", + "p ar", + "com o", + "p ro", + "s i", + "m os", + "e c", + "p re", + "d as", + "ç a", + "es ta", + "s er", + "u n", + "da de", + "d is", + "f o", + "e x", + "c h", + "i r", + "ra n", + "t ar", + "en te", + "g a", + "t r", + "p e", + "t os", + "b o", + "c ia", + "p en", + "c ar", + "s en", + "su a", + "se m", + "c as", + "f or", + "to u", + "n os", + "te m", + "r ia", + "m es", + "se u", + "co r", + "o n", + "a o", + "p os", + "ra m", + "v el", + "é m", + "t en", + "po de", + "t es", + "esta va", + "c e", + "b a", + "qu ando", + "m i", + "qu er", + "men to", + "se gu", + "t as", + "is so", + "mu i", + "g ar", + "t ro", + "d u", + "fa z", + "õ es", + "p es", + "an to", + "l u", + "p i", + "i x", + "ve z", + "s im", + "j a", + "p r", + "m in", + "b e", + "ra s", + "m an", + "p res", + "est á", + "c er", + "b re", + "p as", + "d ia", + "m b", + "dis se", + "n i", + "r os", + "es se", + "v ia", + "o lh", + "is a", + "an te", + "ê n", + "z a", + "qu i", + "b i", + "t inha", + "me u", + "s ão", + "m inha", + "a c", + "ri o", + "m ar", + "a t", + "p el", + "mui to", + "ta l", + "to r", + "fo i", + "h or", + "j o", + "b em", + "g i", + "f al", + "vo l", + "po n", + "di z", + "l ar", + "gu n", + "m or", + "r u", + "par ec", + "ç o", + "do r", + "pes so", + "n e", + "f er", + "b er", + "p u", + "po is", + "in a", + "es p", + "d ar", + "en do", + "de n", + "so bre", + "co s", + "p ri", + "al i", + "mes mo", + "ç ões", + "g ra", + "se us", + "me i", + "b ra", + "vi da", + "an tes", + "b ri", + "at é", + "ên cia", + "lh e", + "ti v", + "m ã", + "al g", + "qu anto", + "s ó", + "g os", + "de r", + "t ão", + "tu do", + "ent ão", + "r ou", + "es s", + "in da", + "b al", + "in do", + "ci o", + "n do", + "j á", + "va m", + "re i", + "l es", + "ei to", + "v is", + "tem po", + "de pois", + "c ha", + "m el", + "ch e", + "l ha", + "a inda", + "faz er", + "con tra", + "p ou", + "per gun", + "de ix", + "ta mb", + "ra r", + "al a", + "v en", + "t in", + "pel o", + "tamb ém", + "fi ca", + "pre c", + "el es", + "tra n", + "ha via", + "l á", + "to dos", + "j u", + "qu al", + "c an", + "ta do", + "cas a", + "es sa", + "n as", + "g em", + "m em", + "se i", + "na da", + "sen ti", + "c ri", + "ó s", + "de u", + "ei ro", + ". .", + "f un", + "as sim", + "s ou", + "ent re", + "com e", + "i or", + "h ar", + "f e", + "por que", + "s or", + "f in", + "ta mente", + "a qui", + "cu l", + "t ó", + "for ma", + "s ar", + "ou tra", + "olh os", + "i ma", + "m im", + "a go", + "in s", + "co u", + "g ran", + "v al", + "pesso as", + "era m", + "ei ra", + "a que", + "com p", + "de i", + "p ela", + "co isa", + "m ão", + "con h", + "ca da", + "ago ra", + "ia m", + "h á", + "con s", + "su as", + "gu ém", + "o b", + "l an", + "es ti", + "á s", + "la do", + "in ter", + "ca be", + "por ta", + "n em", + "í vel", + "r is", + "j e", + "n un", + "sem pre", + "con segu", + "h as", + "tra bal", + "f u", + "le v", + "l em", + "l as", + "va i", + "tr os", + "t ante", + "te i", + "pr ó", + "que m", + "tu ra", + "on de", + "cabe ça", + "nun ca", + "men tos", + "h um", + "de le", + "ver dade", + "t á", + "h os", + "el i", + "ent es", + "m er", + "alg um", + "diz er", + "s in", + "pen as", + "n ós", + "en quanto", + "ou tro", + "l ho", + "es te", + "mel hor", + "est ar", + "g an", + "b ar", + "pri mei", + "a u", + "i u", + "pen sa", + "a penas", + "p ra", + "es tou", + "con te", + "res pon", + "ho mem", + "do is", + "a do", + "c al", + "a b", + "l os", + "ç as", + "pou co", + "sen hor", + "t ando", + "esp era", + "pa i", + "ri os", + "no i", + "i da", + "ba ix", + "as e", + "is as", + "f r", + "ho ra", + "mu ndo", + "pas sa", + "fi car", + "to do", + "se ja", + "al mente", + "â n", + "c lar", + "a d", + "in c", + "f os", + "lo n", + "g ri", + "ou vi", + "v em", + "g e", + "ta va", + "á rio", + "mo n", + "s os", + "in ho", + "ma l", + "t an", + "t re", + "gran de", + "ran do", + "b u", + "v ou", + "ê s", + "co isas", + "a conte", + "lh er", + "g en", + "ci on", + "an os", + "i do", + "tal vez", + "est ão", + "li v", + "sa b", + "su r", + "ou tros", + "c re", + "qual quer", + "g ou", + "t ri", + "l í", + "tiv esse", + "ra do", + "prec isa", + "mã e", + "su s", + "t anto", + "de la", + "men os", + "s al", + "en tra", + "p é", + "ma ior", + "noi te", + "ti va", + "p ala", + "so n", + "ra ção", + "de us", + "s as", + "un i", + "l or", + "u l", + "in te", + "f ei", + "an o", + "par ti", + "pala v", + "tr ás", + "par te", + "b el", + "ci dade", + "lu gar", + "v os", + "vez es", + "do u", + "en contra", + "tr u", + "e ci", + "a r", + "e r", + "a n", + "e n", + "i n", + "i r", + "o r", + "d e", + "a k", + "ı n", + "a l", + "d i", + "d a", + "b u", + "b ir", + "y or", + "i l", + "e k", + "y a", + "m a", + "l a", + "e l", + "u n", + "k a", + "l ar", + "i m", + "d ı", + "e t", + "o n", + "d u", + "o l", + "e y", + "t ı", + "m i", + "h a", + "b a", + "l er", + "ü n", + "m ı", + "i z", + "l e", + "ı r", + "m e", + "i s", + "n e", + "o k", + "t a", + "s a", + "u m", + "r a", + "g ö", + "i k", + "s ı", + "d en", + "e s", + "b il", + "t i", + "l ı", + "ü z", + "i ç", + "ü r", + "g i", + "u r", + "t e", + "b en", + "d an", + "i y", + "ı m", + "u z", + "v e", + "c ak", + "a y", + "c e", + "i ş", + "ın ı", + "i yor", + "ba ş", + "d ü", + "a t", + "a m", + "g el", + "de ğ", + "k ar", + "i ̇", + "m u", + "e v", + "ö y", + "bu n", + "v ar", + "ya p", + "s en", + "an a", + "s un", + "in i", + "gö r", + "y ı", + "k i", + "l i", + "ar a", + "al ı", + "on u", + "ç ı", + "ş ey", + "s ın", + "k ı", + "ka d", + "s e", + "t an", + "a ğ", + "değ il", + "s in", + "ü k", + "a z", + "ç ok", + "s on", + "ş ı", + "b i", + "ü l", + "t u", + "v er", + "iç in", + "g e", + "k en", + "ey e", + "ol du", + "mı ş", + "y e", + "k al", + "m ek", + "l an", + "öy le", + "yor du", + "er i", + "y üz", + "mi ş", + "b e", + "m ak", + "o la", + "in e", + "y an", + "h er", + "c ek", + "yor um", + "b ak", + "ü m", + "ö n", + "lar ı", + "o ğ", + "d er", + "kad ar", + "h al", + "ar ı", + "s t", + "s an", + "ın da", + "du r", + "g ün", + "v a", + "y ok", + "y er", + "dı m", + "k o", + "da ha", + "l u", + "ın a", + "di m", + "e m", + "bil ir", + "ik i", + "s iz", + "s i", + "n a", + "di ğ", + "s u", + "b ü", + "ha y", + "s or", + "dü ş", + "ü ç", + "un u", + "ö r", + "d ir", + "m ü", + "c a", + "am an", + "f ak", + "a da", + "e de", + "son ra", + "h iç", + "ak i", + "ğ ı", + "bu l", + "r u", + "ma z", + "an la", + "bu ra", + "ge ç", + "ma ya", + "l en", + "k onu", + "c i", + "c u", + "d in", + "t ek", + "z aman", + "el er", + "ö z", + "dı r", + "gi bi", + "o t", + "ş a", + "g er", + "ler i", + "k im", + "k u", + "fak at", + "y ar", + "gö z", + "c ı", + "yor sun", + "b ek", + "in de", + "r o", + "p ek", + "bun u", + "l ik", + "m an", + "il er", + "e di", + "ö l", + "s ür", + "b in", + "s ır", + "çı k", + "sı l", + "al ar", + "k es", + "y ak", + "ç ek", + "yı l", + "e cek", + "ı z", + "gi t", + "ka p", + "a ma", + "ı l", + "lar ın", + "b iz", + "tı r", + "o y", + "an cak", + "d oğ", + "ç a", + "b ana", + "ş im", + "baş la", + "l ü", + "ma dı", + "ben i", + "t ir", + "y ük", + "lı k", + "be ş", + "b el", + "b er", + "m er", + "na sıl", + "tı k", + "k e", + "t ür", + "a v", + ". .", + "d aki", + "p ar", + "t er", + "ce ğ", + "t en", + "z ı", + "iy i", + "d ok", + "ben im", + "c ağ", + "n er", + "y en", + "ş u", + "me z", + "düş ün", + "ken di", + "şim di", + "y ol", + "y u", + "de v", + "is te", + "s ek", + "ma m", + "s öyle", + "di k", + "t o", + "k ur", + "oldu ğ", + "s ını", + "t ar", + "bil iyor", + "k an", + "y al", + "m eye", + "mu ş", + "f a", + "ka ç", + "bil e", + "iy e", + "t ü", + "e f", + "tı m", + "ev et", + "ç o", + "y et", + "g en", + "bura da", + "t im", + "bir az", + "es i", + "k or", + "doğ ru", + "in in", + "kı z", + "di ye", + "d ör", + "et ti", + "on un", + "is ti", + "ğ i", + "h e", + "s ana", + "ü ş", + "ar ka", + "hay ır", + "kar şı", + "h ar", + "il e", + "h ak", + "ı yor", + "ne den", + "s ev", + "sı z", + "ço cu", + "me m", + "ç alı", + "ol ur", + "b ır", + "g ir", + "is e", + "i h", + "c an", + "k ır", + "d ön", + "b öyle", + "sen i", + "! \"", + "al t", + "dör t", + "s öy", + "o ş", + "mu sun", + "la ş", + "h an", + "i p", + "ka y", + "h em", + "bü yük", + "a ç", + "bır ak", + "mi sin", + "s öz", + "u l", + "değ iş", + "ün ü", + "g ül", + "k ö", + "kar ı", + "ta mam", + "ol u", + "r ar", + "yen i", + "la m", + "mış tı", + "ya ş", + "al a", + "in iz", + "kad ın", + "bun un", + "m ey", + "al tı", + "y i", + "s o", + "in den", + "sen in", + "ya t", + "to p", + "s er", + "is i", + "d ün", + "s es", + "hiç bir", + "y on", + "d ın", + "t ün", + "baş ka", + "a s", + "he p", + "i t", + "ir mi", + "dev am", + "ola cak", + "ar tık", + "r e", + "dur um", + "im iz", + "üz el", + "ler ini", + "sa ğ", + "p ro", + "ger ek", + "y irmi", + "ş ek", + "ba ğ", + "me di", + "lar a", + "a h", + "t ur", + "y ür", + "ma sı", + "ka tı", + "de di", + "g ü", + "sor un", + "el i", + "ün e", + "mı z", + "yap ı", + "m il", + "ğ ını", + "t ara", + "m en", + "ha t", + "var dı", + "m et", + "konu ş", + "ar ak", + "lar ak", + "çocu k", + "bü tün", + "l ey", + "d ür", + "g üzel", + "ay ı", + "yap a", + "n ı", + "ay r", + "ö ne", + "yordu m", + "b an", + "i̇ ş", + "du m", + "un a", + "on a", + "yor lar", + "lar ını", + "çı kar", + "z an", + "se ç", + "l iyor", + "t ak", + "şı k", + "tek rar", + "a ş", + "e ş", + "miş ti", + "f ar", + "k in", + "im i", + "i f", + "e ğ", + "gi di", + "le ş", + "başla dı", + "gi de", + "ot ur", + "d de", + "ın dan", + "üz er", + "ın ın", + "n ız", + "u y", + "ye di", + "ka t", + "o larak", + "la dı", + "yal nız", + "ba h", + "iy et", + "m al", + "s ak", + "a çık", + "sın da", + ".. .", + "in san", + "ay nı", + "e der", + "is tan", + "uz un", + "sa h", + "d o", + "g eri", + "er ek", + "ol an", + "ger çek", + "f en", + "al an", + "dı ş", + "alı k", + "far k", + "ü st", + "sa de", + "r i", + "k iş", + "l dı", + "z or", + "et ir", + "her kes", + "s al", + "ö mer", + "s el", + "un da", + "ha f", + "bun a", + "y dı", + "pek i", + "ada m", + "ha z", + "sın a", + "kap ı", + "gör üş", + "sade ce", + "al dı", + "gel di", + "i e", + "n ie", + "n a", + "r z", + "s z", + "c z", + "p o", + "s t", + "c h", + "i ę", + "d z", + "n i", + "a ł", + "r a", + "j e", + "r o", + "d o", + "s ię", + "z a", + "g o", + "e m", + "w i", + "c i", + "rz e", + "k o", + "l e", + "l i", + "w a", + "t o", + "k a", + "m i", + "ż e", + "t a", + "w ie", + "b y", + "m o", + "w y", + "rz y", + "ł a", + "j a", + "n o", + "ł o", + "w o", + "p a", + "m a", + "t e", + "t y", + "n y", + "k i", + "d a", + "n e", + "dz ie", + "dz i", + "cz y", + "c ie", + "m y", + "p rze", + "d y", + "o d", + "l a", + "k ie", + "r y", + "st a", + "j ą", + "ó w", + "c e", + "p rzy", + "c o", + "k u", + "m ie", + "sz y", + "cz e", + "r e", + "b a", + "s i", + "b ie", + "m u", + "w e", + "c y", + "ni a", + "ś ci", + "sz e", + "je st", + "k t", + "s a", + "b o", + "t u", + "ż y", + "n ą", + "b i", + "r u", + "a le", + "kt ó", + "p ra", + "ał a", + "m nie", + "p ie", + "ł y", + "cz a", + "ja k", + "ro z", + "r ó", + "l u", + "z na", + "g a", + "ra z", + "ł u", + "ta k", + "j u", + "p i", + "ś ć", + "s o", + "wi a", + "m ó", + "ch o", + "w szy", + "p e", + "s po", + "c a", + "g dy", + "w ał", + "w ię", + "d e", + "b e", + "p ro", + "ł em", + "j ę", + "s k", + "z e", + "l o", + "g i", + "r ę", + "do b", + "d u", + "ju ż", + "st o", + "b ę", + "ał em", + "sz a", + "m e", + "po d", + "d la", + "pa n", + "n ę", + "z o", + "mo że", + "ś li", + "s ie", + "ał o", + "t em", + "l ko", + "ny ch", + "po wie", + "c ię", + "s u", + "ty lko", + "i n", + "b u", + "na j", + "ch a", + "te go", + "p u", + "s ki", + "ne go", + "wszy st", + "sz cze", + "je d", + "je j", + "t wo", + "ą d", + "ś my", + "cz ę", + "wa ć", + "je go", + "ż a", + "i m", + "s y", + "pra w", + "ty m", + "któ ry", + "ał y", + "t rze", + "nie j", + "s e", + "ny m", + "i ch", + "o b", + ". .", + "g ło", + "ją c", + "mó wi", + "s ka", + "o n", + "ne j", + "s łu", + "w ła", + "bę dzie", + "d ę", + "p ó", + "be z", + "ni c", + "p ła", + "ś cie", + "mi a", + "s ą", + "t rzy", + "kie m", + "by ł", + "mo g", + "ro bi", + "ta m", + "c u", + "te n", + "m ię", + "z y", + "pe w", + "ci a", + "my ś", + "prze d", + "s ko", + "n u", + "któ re", + "a l", + "l ę", + "w sze", + "ą c", + "by ło", + "so bie", + "p y", + "ci ą", + "ba r", + "je szcze", + "h a", + "t ę", + "b ra", + "cza s", + "sz ę", + "g ł", + "k ę", + "ma r", + "cz u", + "prze z", + "f i", + "s ło", + "w z", + "k to", + "k ów", + "cz o", + "li śmy", + "st ra", + "wię c", + "r ą", + "ma m", + "w ó", + "rz a", + "g ro", + "no ści", + "f a", + "we t", + "ną ł", + "ś mie", + "na wet", + "mu si", + "s wo", + "te j", + "w ą", + "w u", + "wi ą", + "ni u", + "cz ą", + "b li", + "dz o", + "s kie", + "n em", + "je śli", + "cze go", + "ch y", + "d ł", + "ty ch", + "by m", + "ż o", + "e ś", + "si ą", + "kie dy", + "na s", + "w ró", + "dz e", + "d ro", + "t ra", + "r ów", + "pa ni", + "z ie", + "ku l", + "na d", + "ch wi", + "ni m", + "t ro", + "by ć", + "cho dzi", + "ni o", + "dob rze", + "te raz", + "wo kul", + "co ś", + "k ł", + "pie r", + "h e", + "g dzie", + "dz y", + "p ię", + "d ź", + "k ą", + "g ó", + "z da", + "ch ce", + "st ę", + "o r", + "ś wia", + "wszyst ko", + "st ro", + "pe ł", + "wie m", + "wie l", + "ka ż", + "ki m", + "rz u", + "s ły", + "jed na", + "z u", + "myś l", + "mó j", + "g u", + "wa r", + "jest em", + "ó ż", + "mie j", + "mo ż", + "k ła", + "re sz", + "d łu", + "st wo", + "n ię", + "ma sz", + "że by", + "nie m", + "ja kie", + "st y", + "ni ą", + "we j", + "o j", + "g ra", + "s ła", + "no ść", + "z ło", + "sz czę", + ".. .", + "r i", + "le j", + "we go", + "c ał", + "dzi ał", + "ki ch", + "dz a", + "dz ię", + "o czy", + "zo sta", + "cz ło", + "na m", + "ki l", + "o na", + "sz u", + "w ę", + "pa r", + "mi ał", + "st rze", + "ce j", + "e j", + "zna j", + "da ć", + "miej s", + "k ró", + "k ry", + "bar dzo", + "si a", + "z i", + "ś nie", + "l ą", + "g ie", + "cie bie", + "d ni", + "st u", + "po trze", + "wokul ski", + "u wa", + "u mie", + "jedna k", + "k ra", + "wró ci", + "czło wie", + "czy ć", + "by ła", + "że li", + "m ę", + "c ę", + "z robi", + "mog ę", + "pro wa", + "r em", + "nie ch", + "cz nie", + "k ro", + "t ą", + "ch ci", + "b ro", + "dzie ć", + "sz ą", + "pa d", + "t rz", + "t ru", + "je m", + "a ni", + "t ów", + "a r", + "d ru", + "ta j", + "rze kł", + "sa m", + "st e", + "nie go", + "ta kie", + "w ała", + "to wa", + "ka pła", + "wi dzi", + "po dob", + "dz ę", + "t ał", + "stę p", + "b ą", + "po ko", + "w em", + "g ę", + "a by", + "g e", + "al bo", + "s pra", + "z no", + "de n", + "s mo", + "je sz", + "k się", + "jest eś", + "po z", + "ni gdy", + "k sią", + "c óż", + "w s", + "po w", + "t ka", + "ś wie", + "sz ka", + "sa mo", + "s ł", + "rz ę", + "na le", + "chce sz", + "ni k", + "p ę", + "chy ba", + "cią g", + "ją cy", + "wo j", + "na sze", + "mnie j", + "wię cej", + "z wy", + "o sta", + "f e", + "wa ż", + "h o", + "se r", + "śmie r", + "wie r", + "dz ą", + "za ś", + "gdy by", + "ja ki", + "wo l", + "wi n", + "d ą", + "ści a", + "roz ma", + "wa l", + "pa nie", + "sta r", + "ka z", + "je żeli", + "d em", + "w ra", + "ko ń", + "sie bie", + "zno wu", + "p ró", + "cz em", + "st wa", + "i sto", + "pó ł", + "d ał", + "ko bie", + "ała m", + "wy ch", + "ce sa", + "ni ch", + "za wsze", + "dzi ć", + "te ż", + "le pie", + "pro szę", + "k re", + "t wa", + "o t", + "ł ą", + "ch u", + "c ą", + "p rz", + "ł e", + "sze dł", + "od powie", + "my śli", + "ś wią", + "e n", + "e r", + "d e", + "a n", + "e t", + "i j", + "i n", + "e l", + "a a", + "s t", + "o r", + "g e", + "i s", + "a t", + "i e", + "c h", + "o n", + "e en", + "h et", + "i t", + "v er", + "aa r", + "a l", + "o or", + "g en", + "v an", + "o p", + "d en", + "h e", + "o m", + "t e", + "w e", + "i k", + "r e", + "z e", + "ij n", + "d at", + "b e", + "d er", + "in g", + "o e", + "ij k", + "a an", + "ch t", + "v oor", + "l e", + "i et", + "r o", + "m o", + "k en", + "z ijn", + "m en", + "i g", + "j e", + "n iet", + "a r", + "o o", + "i d", + "u n", + "i l", + "s ch", + "mo et", + "st e", + "u r", + "o l", + "he b", + "u it", + "g el", + "w ij", + "a s", + "m e", + "t en", + "w or", + "o u", + "v en", + "l en", + "aa t", + "d it", + "m et", + "r a", + "b en", + "s p", + "o ver", + "d ie", + "n o", + "w er", + "l ijk", + "f t", + "s l", + "an d", + "v e", + "t er", + "i er", + "i en", + "t o", + "d aar", + "g r", + "b el", + "de ze", + "d u", + "a g", + "k an", + "wor den", + "in gen", + "moet en", + "n en", + "on der", + "heb ben", + "r u", + "oo k", + "s en", + "c t", + "k t", + "no g", + "aa l", + "w as", + "u l", + "e er", + "b ij", + "m ijn", + "p ro", + "v ol", + "d o", + "k om", + "at ie", + "e ft", + "k el", + "al s", + "r ij", + "he id", + "a f", + "st el", + "m aar", + "a p", + "we e", + "a d", + "he eft", + "w aar", + "i cht", + "d an", + "er en", + "n e", + "w el", + "w at", + "w il", + "a cht", + "aa g", + "ge b", + "c on", + "z o", + "k e", + "b et", + "h ij", + "d ig", + "k un", + "u w", + "d t", + "d oor", + "t ij", + "a m", + "an g", + "on d", + "er s", + "is ch", + "ge en", + "i ge", + "ge v", + "ve el", + "n u", + "m a", + "on s", + "o f", + "b l", + "n aar", + "g ro", + "p l", + "an der", + "at en", + "kun nen", + "e cht", + "h ier", + "g oe", + "an t", + "u s", + "t wee", + "on t", + "de lijk", + "el e", + "u ur", + "al le", + "t oe", + "me er", + "i st", + "n a", + "n ie", + "on ze", + "l o", + "i m", + "p en", + "h ad", + "tij d", + "h oe", + "to t", + "z ou", + "a k", + "aa k", + "a men", + "d r", + "w oor", + "s e", + "wor dt", + "o t", + "gel ijk", + "g aan", + "i c", + "g er", + "k er", + "el d", + "e m", + "h ou", + "de l", + "z en", + "z el", + "te gen", + "b o", + "kom en", + "c om", + "i gen", + "e it", + "wer k", + "goe d", + "z al", + "z ij", + "sl ag", + "e s", + "z ien", + "a st", + "echt er", + "it ie", + "t ie", + "el ijk", + "m is", + "isch e", + "bel an", + "h aar", + "i ch", + "b er", + "h an", + "v r", + "al e", + "c i", + "gr ijk", + "in d", + "do en", + "l and", + "belan grijk", + "p un", + "op en", + "ct ie", + "zel f", + "m ij", + "it eit", + "ste m", + "me e", + "ar en", + "al l", + "b r", + "re cht", + "d ien", + "h u", + "g aat", + "pro b", + "m oe", + "p er", + "a u", + "ul len", + "z ich", + "daar om", + "or m", + "k l", + "v o", + "en t", + "st aat", + "z it", + "du i", + "n at", + "du s", + "d s", + "ver slag", + "kel ijk", + "prob le", + "w et", + "ge m", + "c r", + "i on", + "p r", + "sch ap", + "g d", + "h un", + "z a", + "er d", + "z et", + "st aan", + "st r", + "m aal", + "in der", + "e id", + "st en", + "p ar", + "k ken", + "ge d", + "z ullen", + "re s", + "men sen", + "j aar", + "re gel", + "ie der", + "vol gen", + "ge ven", + "e ven", + "l u", + "bl ij", + "i ë", + "k o", + "u we", + "m an", + "ma ken", + "l ie", + "g a", + "oe k", + "nie uwe", + "b aar", + "h o", + "h er", + "in ter", + "ander e", + "ru ik", + "s u", + "a gen", + "or t", + "m er", + "ou w", + "st er", + "wil len", + "aa kt", + "h oo", + "an den", + "f f", + "l ig", + "t re", + "s amen", + "ze er", + "dui delijk", + "ant woor", + "he el", + "men t", + "pun t", + "hou den", + "we g", + "vr aag", + "gel e", + "een s", + "be sch", + "om en", + "er g", + "do el", + "d ag", + "sp e", + "ur en", + "ing s", + "or en", + "l ang", + "de len", + "m ar", + "ste un", + "in nen", + "p ol", + "o on", + "i de", + "s n", + "s ie", + "r icht", + "z onder", + "no dig", + "all een", + "m id", + "ra gen", + "iet s", + "ver sch", + "geb ruik", + "st u", + "ro uw", + "stel len", + "be g", + "men ten", + "v in", + "eer ste", + "l aat", + "gro ot", + "oo d", + "to ch", + "l aten", + "aar d", + "s le", + "de el", + "st and", + "pl aat", + "re e", + "bet re", + "d i", + "l id", + "uit en", + "ra cht", + "bel eid", + "g et", + "ar t", + "st ie", + "st aten", + "g gen", + "re ken", + "e in", + "al en", + "m ing", + "mo gelijk", + "gro te", + "al tijd", + "z or", + "en kel", + "w ik", + "pol itie", + "e igen", + "el k", + "han del", + "g t", + "k we", + "m aat", + "el en", + "i p", + "v rij", + "s om", + "je s", + "aa m", + "hu is", + "v al", + "we er", + "lid staten", + "k ing", + "k le", + "be d", + "gev al", + "stel l", + "a i", + "wik kel", + "kwe stie", + "t al", + "ste e", + "a b", + "h el", + "kom st", + "p as", + "s s", + "it u", + "i den", + "eer d", + "m in", + "c e", + "p o", + "twee de", + "proble em", + "w aren", + "us sen", + "sn el", + "t ig", + "ge w", + "j u", + "ul t", + "ne men", + "com mis", + "versch il", + "k on", + "z oek", + "k rij", + "gr aag", + "den k", + "l anden", + "re den", + "be sl", + "oe g", + "bet er", + "he den", + "m ag", + "p e", + "bo ven", + "a c", + "con t", + "f d", + "h ele", + "k r", + "v ier", + "w in", + "ge z", + "k w", + "m il", + "v or", + "he m", + "ra m", + "aa s", + "ont wikkel", + "dr ie", + "v aak", + "plaat s", + "l a", + "g ang", + "ij f", + "f in", + "nat uur", + "t ussen", + "u g", + "in e", + "d a", + "b at", + "kom t", + "w acht", + "aa d", + "u t", + "é n", + "acht er", + "geb ie", + "ver k", + "lig t", + "c es", + "nie uw", + "van d", + "s t", + "n í", + "j e", + "p o", + "c h", + "r o", + "n a", + "s e", + "t o", + "n e", + "l e", + "k o", + "l a", + "d o", + "r a", + "n o", + "t e", + "h o", + "n ě", + "v a", + "l i", + "l o", + "ř e", + "c e", + "d e", + "v e", + "b y", + "n i", + "s k", + "t a", + "n á", + "z a", + "p ro", + "v o", + "v ě", + "m e", + "v á", + "s o", + "k a", + "r á", + "v y", + "z e", + "m i", + "p a", + "t i", + "st a", + "m ě", + "n é", + "ř i", + "ř í", + "m o", + "ž e", + "m a", + "j í", + "v ý", + "j i", + "d ě", + "r e", + "d a", + "k u", + "j a", + "c i", + "r u", + "č e", + "o b", + "t ě", + "m u", + "k y", + "d i", + "š e", + "k é", + "š í", + "t u", + "v i", + "p ře", + "v í", + "s i", + "n ý", + "o d", + "so u", + "v é", + "n y", + "r i", + "d y", + "b u", + "b o", + "t y", + "l á", + "l u", + "n u", + "ž i", + "m á", + "st i", + "c í", + "z á", + "p ra", + "sk é", + "m í", + "c o", + "d u", + "d á", + "by l", + "st o", + "s a", + "t í", + "je d", + "p ří", + "p ři", + "t é", + "s í", + "č i", + "v ní", + "č a", + "d í", + "z i", + "st u", + "p e", + "b a", + "d ní", + "ro z", + "va l", + "l í", + "s po", + "k á", + "b e", + "p i", + "no u", + "ta k", + "st e", + "r y", + "l é", + "vě t", + "se m", + "p ě", + "ko n", + "ne j", + "l y", + "ko u", + "ý ch", + "b ě", + "p r", + "f i", + "p rá", + "a le", + "ja ko", + "po d", + "ž í", + "z í", + "j sou", + "j sem", + "ch o", + "l ní", + "c ké", + "t á", + "m y", + "a k", + "h u", + "va t", + "pře d", + "h la", + "k e", + "st á", + "č í", + "š i", + "s le", + "k la", + "š tě", + "lo u", + "m ů", + "z na", + "ch á", + "o r", + "p ů", + "h a", + "b i", + "ta ké", + "d ů", + "no st", + "t ře", + "te r", + "p u", + "i n", + "v r", + "ve l", + "sk u", + "v še", + "t ní", + "do b", + "by la", + "č ní", + "ja k", + "v u", + "je ho", + "b ý", + "vá ní", + "ný ch", + "po u", + "te n", + "t ři", + "v z", + "st ře", + "d va", + "h le", + "č á", + "no sti", + "c k", + "v š", + "vo u", + "s u", + "h e", + "h ra", + "je n", + "s y", + "da l", + "po z", + "s lo", + "te l", + "d ru", + "de n", + "vš ak", + "g i", + "k dy", + "by lo", + "bu de", + "st ra", + "j ší", + "m é", + "me n", + "vý ch", + "ní m", + "s m", + "ko li", + "r ů", + "t ra", + "mů že", + "ne ní", + "ho d", + "b í", + "do u", + "sk a", + "t ý", + "st ě", + "u je", + "s á", + "pě t", + "ne s", + "k rá", + "to m", + "st ví", + "v ně", + "se d", + "s vé", + "p í", + "z o", + "mu sí", + "u ž", + "tí m", + "jí cí", + "jed no", + "t r", + "ča s", + "e v", + "č ty", + "sk ý", + "ni c", + "ev ro", + "to ho", + "h y", + "k ter", + "r ní", + "st í", + "s vě", + "pa k", + "vše ch", + "k ů", + "n g", + "á d", + "chá zí", + "a ni", + "a r", + "jed na", + "bý t", + "t ro", + "k ra", + "pr vní", + "m no", + "ské ho", + "p á", + "p la", + "le m", + "ne bo", + "ke m", + "st ro", + "s la", + "né ho", + "z de", + "dal ší", + "ř a", + "čty ři", + "h rá", + "dru h", + "l ně", + "v la", + "sk ých", + "š ko", + "pů so", + "pro to", + "v ů", + "sk á", + "ve n", + "še st", + "d ně", + "je ště", + "me zi", + "te k", + "s ko", + "ch a", + "ně koli", + "be z", + "g ra", + "ji ž", + "č ně", + "j á", + "s lu", + "z ná", + "ve r", + "sed m", + "k ro", + "ta m", + "a no", + "v lá", + "o sm", + "byl y", + "vá m", + "ck ý", + "te ch", + "dě ji", + "vel mi", + "le ži", + "va la", + "l ý", + "t vo", + "spo le", + "ch u", + "stu p", + "mo ž", + "evro p", + "g e", + "sta l", + "j de", + "ch y", + "ro di", + "je jí", + "po li", + "de vět", + "s me", + "a ž", + "té to", + "re m", + "d é", + "f or", + "u ni", + "f o", + "ten to", + "a u", + "ka ž", + "nu la", + "na d", + "by ch", + "mo c", + "sto u", + "e x", + "le n", + "k do", + "z d", + "pra co", + "to mu", + "ný m", + "ži vo", + "ze m", + "f e", + "f u", + "ná sle", + "j o", + "sk y", + "ji ch", + "h á", + "mě l", + "dě la", + "j sme", + "p re", + "ni ce", + "ste j", + "ne m", + "st ní", + "he m", + "ná ro", + "z u", + "b li", + "ni t", + "pa r", + "a l", + "poz ději", + "ta ko", + "n ce", + "če r", + "ší m", + "ně co", + "vá l", + "ře j", + "krá t", + "á lní", + "u r", + ". .", + "a si", + "kter é", + "sta v", + "ma jí", + "my s", + "do bě", + "s ně", + "ce n", + "z y", + "z ku", + "t ů", + "ch od", + "s pě", + "je jich", + "sou čas", + "d r", + "va li", + "ri e", + "k te", + "pr ů", + "ze ní", + "pa t", + "a n", + "po tře", + "de m", + "d nes", + "ze mí", + "sa mo", + "zna m", + "b ra", + "má m", + "te dy", + "g o", + "hla vní", + "pou ží", + "b ní", + "ve de", + "le p", + "je k", + "pra v", + "poli ti", + "d ne", + "je m", + "le t", + "če ní", + "pro b", + "ne ž", + "dě l", + "fi l", + "č o", + "cí ch", + "st é", + "d lou", + "h i", + "a by", + "to u", + "několi k", + "d la", + "vy u", + "vi t", + "ho u", + "ck ých", + "no vé", + "či n", + "st y", + "dě lá", + "k ý", + "ob la", + "pod le", + "ra n", + "dů leži", + "ta to", + "po ku", + "ko ne", + "d ý", + "d vě", + "ž ád", + "nou t", + "t ku", + "t vr", + "cké ho", + "ro v", + "r é", + "te le", + "p sa", + "s vět", + "ti vní", + "do sta", + "te m", + "še l", + "druh é", + "s kou", + "ž o", + "jed ná", + "vý znam", + "prob lé", + "pu bli", + "vá n", + "od po", + "pod po", + "d le", + "ja ké", + "še ní", + "ví m", + "bě hem", + "na chází", + "s lou", + "pou ze", + "o tá", + "p lo", + "to vé", + "vět ši", + "ko mi", + "va jí", + "ty to", + "zá pa", + "z mě", + "mo h", + "ví ce", + "spole č", + "au to", + "pro ti", + "st ru", + "dě t", + "chá ze", + "že l", + "с т", + "е н", + "н о", + "н а", + "п р", + "т о", + "п о", + "р а", + "г о", + "к о", + "н е", + "в о", + "в а", + "е т", + "е р", + "н и", + "е л", + "и т", + "н ы", + "з а", + "р о", + "ен и", + "к а", + "л и", + "е м", + "д а", + "о б", + "л а", + "д о", + "с я", + "т ь", + "о т", + "л о", + "л ь", + "е д", + "с о", + "м и", + "р е", + "м о", + "ц и", + "пр о", + "т а", + "э то", + "к и", + "р у", + "пр и", + "т и", + "с е", + "ст а", + "в ы", + "м ы", + "в и", + "б ы", + "м а", + "е с", + "л я", + "ст и", + "л е", + "ч то", + "м е", + "р и", + "ч а", + "о д", + "е й", + "ел ь", + "ени я", + "г а", + "н у", + "с и", + "п а", + "ра з", + "б о", + "ст о", + "с у", + "с а", + "д у", + "е го", + "е ст", + "и н", + "ит ь", + "и з", + "ж е", + "м у", + "п ер", + "по д", + "ени е", + "с ь", + "к у", + "пр ед", + "но го", + "ны х", + "в ер", + "т е", + "но й", + "ци и", + "д е", + "р ы", + "д ел", + "л ю", + "в е", + "о н", + "м ен", + "г и", + "н я", + "б у", + "пр а", + "в се", + "ет ся", + "ст ь", + "ж а", + "до л", + "ж и", + "б е", + "ко н", + "с л", + "ш и", + "д и", + "ст в", + "с ко", + "ны е", + "ч и", + "ю т", + "д ер", + "ст ра", + "т ы", + "х од", + "щ и", + "з о", + "з на", + "но сти", + "ч ес", + "в ля", + "ва ть", + "о р", + "по л", + "в ет", + "та к", + "ш а", + "т у", + "с во", + "пр е", + "о на", + "ит ель", + "ны й", + "с ло", + "ка к", + "в л", + "но сть", + "х о", + "мо ж", + "п е", + "д ля", + "ни я", + "но е", + "ра с", + "дол ж", + "да р", + "т ель", + "с ка", + "п у", + "ст во", + "ко то", + "ра б", + "е е", + "ро д", + "э ти", + "с об", + "о ру", + "ж ен", + "ны м", + "ит и", + "ни е", + "ко м", + "д ет", + "ст у", + "г у", + "п и", + "ме ж", + "ени ю", + "т ер", + "раб от", + "во з", + "ци я", + "ко й", + "щ ест", + "г ра", + "з и", + "р я", + "меж ду", + "ст ва", + "в с", + "ел о", + "ш е", + "м ер", + "б а", + "з ы", + "л у", + "а ль", + "д ей", + "г ла", + "на род", + "к ти", + "пред ста", + "л ся", + "я вля", + "с ки", + "но в", + "ед ин", + "ро в", + "и с", + "ни ма", + "р ем", + "ход и", + "так же", + "д ру", + "а ть", + "сл ед", + "го во", + "на я", + "ю щи", + "ен ь", + "кото ры", + "х от", + "в у", + "и х", + "ем у", + "ч ит", + "ва ж", + "ор га", + "чес ки", + "щ е", + "к е", + "х а", + "по с", + "то м", + "бо ль", + "м не", + "па с", + "об ъ", + "пра в", + "кон ф", + "сл у", + "под дер", + "ст ви", + "на ш", + "ль ко", + "сто я", + "ну ю", + "л ем", + "ен ных", + "к ра", + "д ы", + "между народ", + "г да", + "не об", + "го су", + "ств у", + "ени и", + "госу дар", + "к то", + "и м", + "ч ест", + "р ет", + "во про", + "л ен", + "ел и", + "ро ва", + "ци й", + "на м", + "это й", + "ж ения", + "необ ходи", + "мен я", + "бы ло", + "си ли", + "ф и", + "в я", + "ш ь", + "это го", + "о ни", + "орга ни", + "бе зо", + "пр об", + "и ме", + "ре ш", + "б и", + "безо пас", + "ют ся", + "о ста", + "ен но", + "го д", + "ел а", + "предста в", + "ть ся", + "сло во", + "органи за", + "долж ны", + "это м", + "б ла", + "ч е", + "ч у", + "бла го", + "это му", + "в рем", + "с пе", + "но м", + "ени й", + "с по", + "на с", + "не т", + "з у", + "в ед", + "е ще", + "ска за", + "се й", + "ер ен", + "да н", + "са м", + "ел я", + "ра н", + "зы ва", + "явля ется", + "бу дет", + "кти в", + "т ре", + "дел е", + "м от", + "конф ерен", + "ла сь", + "ча с", + "сто ро", + "ко го", + "е з", + "не й", + "о с", + "ли сь", + "раз ору", + "пер е", + "с си", + "ны ми", + "про ц", + "го ло", + "ч ело", + "бо ле", + "чело ве", + "с ер", + "п л", + "ч ет", + "стра н", + "п я", + "бы л", + "к ла", + "то в", + "ж д", + "дел а", + "е ра", + "у же", + "со вет", + "г ен", + "безопас ности", + "ц а", + "се да", + "по з", + "от вет", + "проб лем", + "на ко", + "т ем", + "до ста", + "п ы", + "щ а", + "во й", + "су щест", + "необходи мо", + "бы ть", + "мож ет", + "д ем", + "что бы", + "е к", + "ч ер", + "у сили", + "ре с", + "ру д", + "един енных", + "д об", + "до сти", + "ств ен", + "я дер", + "год ня", + "ка за", + "се годня", + "сей час", + "то лько", + "во д", + "ес ь", + "м ного", + "бу ду", + "е в", + "ест ь", + "т ри", + "об щест", + ". .", + "я вл", + "вы сту", + "р ед", + "с чит", + "с ит", + "деле га", + "ло ж", + "это т", + "ф ор", + "к лю", + "воз мож", + "ва ния", + "б ли", + "и ли", + "в з", + "на ций", + "ско го", + "при ня", + "п ла", + "о ч", + "ить ся", + "ст е", + "на ши", + "которы е", + "а р", + "име ет", + "с от", + "зна ч", + "пер ь", + "след у", + "ен ы", + "та ки", + "объ единенных", + "ст ро", + "те перь", + "б ле", + "благо дар", + "раз в", + "а н", + "жи ва", + "оч ень", + "я т", + "бе з", + "об ес", + "г ро", + "ло сь", + "с ы", + "организа ции", + "ч лен", + "то го", + "она ль", + "ж да", + "все х", + "с вя", + "боле е", + "со в", + "ко гда", + "во т", + "к ре", + "к ры", + "по этому", + "во ль", + "о й", + "ген ера", + "ч ем", + "л ы", + "пол ити", + "в ен", + "конферен ции", + "проц ес", + "б я", + "ит е", + "от но", + "разв ити", + "а ф", + "ю щ", + "в но", + "ми р", + "ни и", + "ка я", + "а с", + "итель но", + "в то", + "ени ем", + "генера ль", + "пр от", + "вс ем", + "сам бле", + "ас самбле", + "о м", + "з д", + "с мот", + "ре ги", + "ч его", + "од нако", + "усили я", + "дей стви", + "ч но", + "у ча", + "об раз", + "во с", + "э та", + "пер его", + "гово р", + "ва м", + "мо ло", + "врем я", + "д ь", + "хот ел", + "г ру", + "за явл", + "пре доста", + "по ль", + "не е", + "ре зо", + "перего во", + "резо лю", + "к рет", + "поддер ж", + "обес пе", + "не го", + "представ ит", + "на де", + "к ри", + "ч ь", + "про ек", + "л ет", + "дру ги", + "ا ل", + "َ ا", + "و َ", + "ّ َ", + "ِ ي", + "أ َ", + "ل َ", + "ن َ", + "ال ْ", + "ه ُ", + "ُ و", + "م ا", + "ن ْ", + "م ن", + "ع َ", + "ن ا", + "ل ا", + "م َ", + "ت َ", + "ف َ", + "أ ن", + "ل ي", + "م ِ", + "ا ن", + "ف ي", + "ر َ", + "ي َ", + "ه ِ", + "م ْ", + "ق َ", + "ب ِ", + "ل ى", + "ي ن", + "إ ِ", + "ل ِ", + "و ا", + "ك َ", + "ه ا", + "ً ا", + "م ُ", + "و ن", + "ال م", + "ب َ", + "ي ا", + "ذ ا", + "س ا", + "ال ل", + "م ي", + "ي ْ", + "ر ا", + "ر ي", + "ل ك", + "م َا", + "ن َّ", + "ل م", + "إ ن", + "س ت", + "و م", + "ّ َا", + "ل َا", + "ه م", + "ّ ِ", + "ك ُ", + "ك ان", + "س َ", + "ب ا", + "د ي", + "ح َ", + "ع ْ", + "ب ي", + "ال أ", + "و ل", + "ف ِي", + "ر ِ", + "د ا", + "مِ نْ", + "ُو نَ", + "و ْ", + "ه َا", + "ّ ُ", + "ال س", + "ال َ", + "ن ي", + "ل ْ", + "ت ُ", + "ه ل", + "ر ة", + "د َ", + "س ْ", + "ت ِ", + "ن َا", + "ر ْ", + "الل َّ", + "سا مي", + "ك ن", + "ك ل", + "ه َ", + "عَ لَ", + "ع لى", + "م ع", + "إ لى", + "ق د", + "ال ر", + "ُو ا", + "ي ر", + "ع ن", + "ي ُ", + "ن ِ", + "ب ْ", + "ال ح", + "هُ مْ", + "ق ا", + "ذ ه", + "ال ت", + "ِي نَ", + "ج َ", + "ه ذا", + "ع د", + "ال ع", + "د ْ", + "قَ الَ", + "ر ُ", + "ي م", + "ي ة", + "ن ُ", + "خ َ", + "ر ب", + "ال ك", + "و َا", + "أ نا", + "ة ِ", + "ال ن", + "ح د", + "ع ِ", + "ت ا", + "ه و", + "ف ا", + "ع ا", + "ال ش", + "ل ُ", + "ي ت", + "ذ َا", + "ي ع", + "ال ذ", + "ح ْ", + "ال ص", + "إِ نَّ", + "ج ا", + "ع لي", + "ك َا", + "ب ُ", + "ت ع", + "و ق", + "م ل", + "ل َّ", + "ي د", + "أ خ", + "ر ف", + "ت ي", + "ال ِ", + "ّ ا", + "ذ لك", + "أَ نْ", + "س ِ", + "ت وم", + "م ر", + "مَ نْ", + "ب ل", + "ال ق", + "الل ه", + "ِي َ", + "ك م", + "ذ َ", + "ع ل", + "ح ب", + "س ي", + "ع ُ", + "ال ج", + "ال د", + "ش َ", + "ت ك", + "ف ْ", + "ص َ", + "ل ل", + "د ِ", + "ب ر", + "ف ِ", + "ت ه", + "أ ع", + "ت ْ", + "ق ْ", + "الْ أَ", + "ئ ِ", + "عَ نْ", + "و ر", + "ح ا", + "ال َّ", + "م ت", + "ف ر", + "د ُ", + "ه نا", + "وَ أَ", + "ت ب", + "ة ُ", + "أ ي", + "س ب", + "ري د", + "و ج", + "كُ مْ", + "ح ِ", + "ك ْ", + "د ر", + "َا ء", + "ه ذه", + "ال ط", + "الْ مُ", + "د ة", + "ق ل", + "غ َ", + "ي وم", + "الَّ ذ", + "ك ر", + "ت ر", + "ك ِ", + "ك ي", + "عَلَ ى", + "رَ ب", + "ع ة", + "ق ُ", + "ج ْ", + "ف ض", + "ل ة", + "ه ْ", + "ر َا", + "وَ لَ", + "الْ مَ", + "أَ نَّ", + "ي َا", + "أ ُ", + "ش ي", + "اللَّ هُ", + "لَ ى", + "ق ِ", + "أ ت", + "عَلَ يْ", + "اللَّ هِ", + "ال ب", + "ض َ", + "ة ً", + "ق ي", + "ا ر", + "ب د", + "خ ْ", + "سْ تَ", + "ط َ", + "قَ دْ", + "ذه ب", + "أ م", + "ما ذا", + "وَ إِ", + "ة ٌ", + "و نَ", + "لي لى", + "و لا", + "ح ُ", + "ه ي", + "ص ل", + "ال خ", + "و د", + "لي س", + "ل دي", + "ق ال", + "كَا نَ", + "م َّ", + "ح ي", + "ت م", + "ل ن", + "وَ لَا", + "ب ع", + "يم كن", + "س ُ", + "ة َ", + "ح ت", + "ر ًا", + "ك ا", + "ش ا", + "هِ مْ", + "لَ هُ", + "ز َ", + "دا ً", + "م س", + "ك ث", + "الْ عَ", + "ج ِ", + "ص ْ", + "ف َا", + "ل ه", + "و ي", + "ع َا", + "هُ وَ", + "ب ِي", + "ب َا", + "أ س", + "ث َ", + "ل ِي", + "ر ض", + "الر َّ", + "لِ كَ", + "ت َّ", + "ف ُ", + "ق ة", + "ف عل", + "مِ ن", + "ال آ", + "ث ُ", + "س م", + "م َّا", + "بِ هِ", + "ت ق", + "خ ر", + "ل قد", + "خ ل", + "ش ر", + "أن ت", + "ل َّا", + "س ن", + "الس َّ", + "الذ ي", + "س َا", + "و ما", + "ز ل", + "و ب", + "أ ْ", + "إ ذا", + "ر ِي", + "ح ة", + "ن ِي", + "الْ حَ", + "وَ قَالَ", + "ب ه", + "ة ٍ", + "س أ", + "ر ٌ", + "ب ال", + "م ة", + "ش ْ", + "و ت", + "عن د", + "ف س", + "بَ عْ", + "ه ر", + "ق ط", + "أ ح", + "إن ه", + "و ع", + "ف ت", + "غ ا", + "هنا ك", + "ب ت", + "مِ نَ", + "س ر", + "ذَ لِكَ", + "ر س", + "حد ث", + "غ ْ", + "ّ ِي", + "ال إ", + "وَ يَ", + "ج ل", + "ا ست", + "ق ِي", + "ع ب", + "و س", + "ي ش", + "الَّذ ِينَ", + "تا ب", + "د ِي", + "ج ب", + "ك ون", + "ب ن", + "ال ث", + "لَ يْ", + "ب عد", + "وَ الْ", + "فَ أَ", + "ع م", + "هُ م", + "ت ن", + "ذ ْ", + "أ ص", + "أ ين", + "رَب ِّ", + "الذ ين", + "إِ ن", + "ب ين", + "ج ُ", + "عَلَيْ هِ", + "ح َا", + "ل و", + "ست ط", + "ظ ر", + "لَ مْ", + "ء ِ", + "كُ ل", + "ط ل", + "ت َا", + "ض ُ", + "كن ت", + "ل ًا", + "م ٌ", + "ق بل", + "ـ ـ", + "ذ ِ", + "قَ وْ", + "ص ِ", + "م ًا", + "كان ت", + "ص ا", + "ي ق", + "ال ف", + "ال نا", + "م ٍ", + "إِ نْ", + "ال نَّ", + "ج د", + "وَ مَا", + "ت ت", + "ب ح", + "م كان", + "كي ف", + "ّ ة", + "ال ا", + "ج َا", + "أ و", + "سا عد", + "ض ِ", + "إ لا", + "را ً", + "ق َا", + "ر أ", + "ع ت", + "أ حد", + "ه د", + "ض ا", + "ط ر", + "أ ق", + "ما ء", + "د َّ", + "ال با", + "م ُو", + "أَ وْ", + "ط ا", + "ق ُو", + "خ ِ", + "ت ل", + "ستط يع", + "د َا", + "الن َّا", + "إ لَى", + "وَ تَ", + "هَ ذَا", + "ب ة", + "علي ك", + "ج ر", + "ال من", + "ز ا", + "ر ٍ", + "د ع", + "ّ ًا", + "س ة", + "ثُ مَّ", + "شي ء", + "ال غ", + "ت ح", + "ر ُونَ", + "ال يوم", + "م ِي", + "ن ُوا", + "أ ر", + "تُ مْ", + "ع ر", + "ي ف", + "أ ب", + "د ًا", + "ص َا", + "الت َّ", + "أ ريد", + "ال ز", + "يَ وْ", + "إ لي", + "ج ي", + "يَ عْ", + "فض ل", + "ال إن", + "أن ه", + "n g", + "i 4", + "a n", + "s h", + "z h", + "i 2", + "ng 1", + "u 4", + "i 1", + "ng 2", + "d e", + "j i", + "a o", + "x i", + "u 3", + "de 5", + "e 4", + "i 3", + "ng 4", + "an 4", + "e n", + "u o", + "sh i4", + "an 2", + "u 2", + "c h", + "u 1", + "ng 3", + "a 1", + "an 1", + "e 2", + "a 4", + "e i4", + "o ng1", + "a i4", + "ao 4", + "h u", + "a ng1", + "l i", + "y o", + "an 3", + "w ei4", + "uo 2", + "n 1", + "en 2", + "ao 3", + "e 1", + "y u", + "q i", + "e ng2", + "zh o", + "a ng3", + "a ng4", + "a ng2", + "uo 4", + "m i", + "g e4", + "y i1", + "g uo2", + "e r", + "b i", + "a 3", + "h e2", + "e 3", + "y i2", + "d i4", + "zh ong1", + "b u4", + "g u", + "a i2", + "n 2", + "z ai4", + "sh i2", + "e ng1", + "r en2", + "o ng2", + "xi an4", + "y i", + "x u", + "n 4", + "l i4", + "en 4", + "y u2", + "e i2", + "yi2 ge4", + "o u4", + "e i3", + "d i", + "u i4", + "a 2", + "yo u3", + "ao 1", + "d a4", + "ch eng2", + "en 1", + "e ng4", + "y i4", + "s i1", + "zh i4", + "ji a1", + "yu an2", + "n i", + "t a1", + "de5 yi2ge4", + "k e1", + "sh u3", + "x i1", + "j i2", + "ao 2", + "t i", + "o u3", + "o ng4", + "xi a4", + "a i1", + "g ong1", + "zh i1", + "en 3", + "w ei2", + "j u", + "xu e2", + "q u1", + "zho u1", + "er 3", + "mi ng2", + "zho ng3", + "l i3", + "w u4", + "y i3", + "uo 1", + "e 5", + "j i4", + "xi ng2", + "ji an4", + "hu a4", + "y u3", + "uo 3", + "j i1", + "a i3", + "z uo4", + "h ou4", + "hu i4", + "e i1", + "ni an2", + "q i2", + "p i", + "d ao4", + "sh eng1", + "de 2", + "d ai4", + "u an2", + "zh e4", + "zh eng4", + "b en3", + "sh ang4", + "zh u3", + "b ei4", + "y e4", + "ch u1", + "zh an4", + "l e5", + "l ai2", + "sh i3", + "n an2", + "r en4", + "yo u2", + "k e4", + "b a1", + "f u4", + "d ui4", + "y a4", + "m ei3", + "z i4", + "xi n1", + "ji ng1", + "zh u", + "n 3", + "yo ng4", + "m u4", + "ji ao4", + "y e3", + "ji n4", + "bi an4", + "l u4", + "q i1", + "sh e4", + "xi ang1", + "o ng3", + "sh u4", + "d ong4", + "s uo3", + "gu an1", + "s an1", + "b o", + "t e4", + "d uo1", + "f u2", + "mi n2", + "l a1", + "zh i2", + "zh en4", + "o u1", + "w u3", + "m a3", + "i 5", + "z i5", + "j u4", + "er 4", + "y ao4", + "xia4 de5yi2ge4", + "s i4", + "t u2", + "sh an1", + "z ui4", + "ch u", + "yi n1", + "er 2", + "t ong2", + "d ong1", + "y u4", + "y an2", + "qi an2", + "shu3 xia4de5yi2ge4", + "ju n1", + "k e3", + "w en2", + "f a3", + "l uo2", + "zh u4", + "x i4", + "k ou3", + "b ei3", + "ji an1", + "f a1", + "di an4", + "ji ang1", + "wei4 yu2", + "xi ang4", + "zh i3", + "e ng3", + "f ang1", + "l an2", + "sh u", + "r i4", + "li an2", + "sh ou3", + "m o", + "qi u2", + "ji n1", + "h uo4", + "shu3xia4de5yi2ge4 zhong3", + "f en1", + "n ei4", + "g ai1", + "mei3 guo2", + "u n2", + "g e2", + "b ao3", + "qi ng1", + "g ao1", + "t ai2", + "d u", + "xi ao3", + "ji e2", + "ti an1", + "ch ang2", + "q uan2", + "li e4", + "h ai3", + "f ei1", + "t i3", + "ju e2", + "o u2", + "c i3", + "z u2", + "n i2", + "bi ao3", + "zhong1 guo2", + "d u4", + "yu e4", + "xi ng4", + "sh eng4", + "ch e1", + "d an1", + "ji e1", + "li n2", + "pi ng2", + "f u3", + "g u3", + "ji e4", + "w o", + "v 3", + "sh eng3", + "n a4", + "yu an4", + "zh ang3", + "gu an3", + "d ao3", + "z u3", + "di ng4", + "di an3", + "c eng2", + "ren2 kou3", + "t ai4", + "t ong1", + "g uo4", + "n eng2", + "ch ang3", + "hu a2", + "li u2", + "yi ng1", + "xi ao4", + "c i4", + "bian4 hua4", + "li ang3", + "g ong4", + "zho ng4", + "de5 yi1", + "s e4", + "k ai1", + "w ang2", + "ji u4", + "sh i1", + "sh ou4", + "m ei2", + "k u", + "s u", + "f eng1", + "z e2", + "tu2 shi4", + "t i2", + "q i4", + "ji u3", + "sh en1", + "zh e3", + "ren2kou3 bian4hua4", + "ren2kou3bian4hua4 tu2shi4", + "di4 qu1", + "y ang2", + "m en", + "men 5", + "l ong2", + "bi ng4", + "ch an3", + "zh u1", + "w ei3", + "w ai4", + "xi ng1", + "bo 1", + "b i3", + "t ang2", + "hu a1", + "bo 2", + "shu i3", + "sh u1", + "d ou1", + "s ai4", + "ch ao2", + "b i4", + "li ng2", + "l ei4", + "da4 xue2", + "f en4", + "shu3 de5", + "m u3", + "ji ao1", + "d ang1", + "ch eng1", + "t ong3", + "n v3", + "q i3", + "y an3", + "mi an4", + "l uo4", + "ji ng4", + "g e1", + "r u4", + "d an4", + "ri4 ben3", + "p u3", + "yu n4", + "hu ang2", + "wo 3", + "l v", + "h ai2", + "shi4 yi1", + "xi e1", + "yi ng3", + "w u2", + "sh en2", + "w ang3", + "gu ang3", + "li u4", + "s u4", + "shi4 zhen4", + "c an1", + "c ao3", + "xi a2", + "k a3", + "d a2", + "h u4", + "b an4", + "d ang3", + "h u2", + "z ong3", + "de ng3", + "de5yi2ge4 shi4zhen4", + "ch uan2", + "mo 4", + "zh ang1", + "b an1", + "mo 2", + "ch a2", + "c e4", + "zhu3 yao4", + "t ou2", + "j u2", + "shi4 wei4yu2", + "s a4", + "u n1", + "ke3 yi3", + "d u1", + "h an4", + "li ang4", + "sh a1", + "ji a3", + "z i1", + "lv 4", + "f u1", + "xi an1", + "x u4", + "gu ang1", + "m eng2", + "b ao4", + "yo u4", + "r ong2", + "zhi1 yi1", + "w ei1", + "m ao2", + "guo2 jia1", + "c ong2", + "g ou4", + "ti e3", + "zh en1", + "d u2", + "bi an1", + "c i2", + "q u3", + "f an4", + "xi ang3", + "m en2", + "j u1", + "h ong2", + "z i3", + "ta1 men5", + "ji 3", + "z ong1", + "zhou1 de5yi2ge4shi4zhen4", + "t uan2", + "ji ng3", + "gong1 si1", + "xi e4", + "l i2", + "li4 shi3", + "b ao1", + "g ang3", + "gu i1", + "zh eng1", + "zhi2 wu4", + "ta1 de5", + "pi n3", + "zhu an1", + "ch ong2", + "shi3 yong4", + "w a3", + "sh uo1", + "chu an1", + "l ei2", + "w an1", + "h uo2", + "q u", + "s u1", + "z ao3", + "g ai3", + "q u4", + "g u4", + "l u", + "x i2", + "h ang2", + "yi ng4", + "c un1", + "g en1", + "yi ng2", + "ti ng2", + "cheng2 shi4", + "ji ang3", + "li ng3", + "l un2", + "bu4 fen4", + "de ng1", + "xu an3", + "dong4 wu4", + "de2 guo2", + "xi an3", + "f an3", + "zh e5", + "h an2", + "h ao4", + "m i4", + "r an2", + "qi n1", + "ti ao2", + "zh an3", + "h i", + "k a", + "n o", + "t e", + "s u", + "s hi", + "t a", + "t o", + "n a", + "w a", + "o u", + "r u", + "n i", + "k u", + "k i", + "g a", + "d e", + "k o", + "m a", + "r e", + "r a", + "m o", + "t su", + "w o", + "e n", + "r i", + "s a", + "d a", + "s e", + "j i", + "h a", + "c hi", + "k e", + "te ki", + "m i", + "y ou", + "s h", + "s o", + "y o", + "y a", + "na i", + "t te", + "a ru", + "b a", + "u u", + "t ta", + "ka i", + "ka n", + "shi te", + "m e", + "d o", + "mo no", + "se i", + "r o", + "ko to", + "ka ra", + "shi ta", + "b u", + "m u", + "c h", + "su ru", + "k ou", + "g o", + "ma su", + "ta i", + "f u", + "k en", + "i u", + "g en", + "wa re", + "shi n", + "z u", + "a i", + "o n", + "o ku", + "g i", + "d ou", + "n e", + "y uu", + "i ru", + "i te", + "ji ko", + "de su", + "j u", + "ra re", + "sh u", + "b e", + "sh ou", + "s ha", + "se kai", + "s ou", + "k you", + "ma shita", + "s en", + "na ra", + "sa n", + "ke i", + "i ta", + "a ri", + "i tsu", + "ko no", + "j ou", + "na ka", + "ch ou", + "so re", + "g u", + "na ru", + "ga ku", + "re ba", + "g e", + "h o", + "i n", + "hi to", + "sa i", + "na n", + "da i", + "tsu ku", + "shi ki", + "sa re", + "na ku", + "p p", + "bu n", + "ju n", + "so no", + "ka ku", + "z ai", + "b i", + "to u", + "wa ta", + "sh uu", + "i i", + "te i", + "ka re", + "y u", + "shi i", + "ma de", + "sh o", + "a n", + "ke reba", + "shi ka", + "i chi", + "ha n", + "de ki", + "ni n", + "ware ware", + "na kereba", + "o ite", + "h ou", + "ya ku", + "ra i", + "mu jun", + "l e", + "yo ku", + "bu tsu", + "o o", + "ko n", + "o mo", + "ga e", + "nara nai", + "ta chi", + "z en", + "ch uu", + "kan gae", + "ta ra", + "to ki", + "ko ro", + "mujun teki", + "z e", + "na ga", + "ji n", + "shi ma", + "te n", + "i ki", + "i ku", + "no u", + "i masu", + "r ou", + "h on", + "ka e", + "t to", + "ko re", + "ta n", + "ki ta", + "i s", + "da tta", + "ji tsu", + "ma e", + "i e", + "me i", + "da n", + "h e", + "to ku", + "dou itsu", + "ri tsu", + "k yuu", + "h you", + "rare ta", + "kei sei", + "k kan", + "rare ru", + "m ou", + "do ko", + "r you", + "da ke", + "naka tta", + "so ko", + "ta be", + "e r", + "ha na", + "c o", + "fu ku", + "p a", + "so n", + "ya su", + "ch o", + "wata ku", + "ya ma", + "z a", + "k yo", + "gen zai", + "b oku", + "a ta", + "j a", + "ka wa", + "ma sen", + "j uu", + "ro n", + "b o", + "na tte", + "wataku shi", + "yo tte", + "ma i", + "g ou", + "ha i", + "mo n", + "ba n", + "ji shin", + "c a", + "re te", + "n en", + "o ka", + "ka gaku", + "na tta", + "p o", + "ka ru", + "na ri", + "m en", + "ma ta", + "e i", + "ku ru", + "ga i", + "ka ri", + "sha kai", + "kou i", + "yo ri", + "se tsu", + "j o", + "re ru", + "to koro", + "ju tsu", + "i on", + "sa ku", + "tta i", + "c ha", + "nin gen", + "n u", + "c e", + "ta me", + "kan kyou", + "de n", + "o oku", + "i ma", + "wata shi", + "tsuku ru", + "su gi", + "b en", + "ji bun", + "shi tsu", + "ke ru", + "ki n", + "ki shi", + "shika shi", + "mo to", + "ma ri", + "i tte", + "de shita", + "n de", + "ari masu", + "te r", + "z ou", + "ko e", + "ze ttai", + "kkan teki", + "h en", + "re kishi", + "deki ru", + "tsu ka", + "l a", + "i tta", + "o i", + "ko butsu", + "mi ru", + "sh oku", + "shi masu", + "gi jutsu", + "g you", + "jou shiki", + "a tta", + "ho do", + "ko ko", + "tsuku rareta", + "z oku", + "hi tei", + "ko ku", + "rekishi teki", + "ke te", + "o ri", + "i mi", + "ka ko", + "naga ra", + "ka karu", + "shu tai", + "ha ji", + "ma n", + "ta ku", + "ra n", + "douitsu teki", + "z o", + "me te", + "re i", + "tsu u", + "sare te", + "gen jitsu", + "p e", + "s t", + "ba i", + "na wa", + "ji kan", + "wa ru", + "r t", + "a tsu", + "so ku", + "koui teki", + "a ra", + "u ma", + "a no", + "i de", + "ka ta", + "te tsu", + "ga wa", + "ke do", + "re ta", + "mi n", + "sa you", + "tte ru", + "to ri", + "p u", + "ki mi", + "b ou", + "mu ra", + "sare ru", + "ma chi", + "k ya", + "o sa", + "kon na", + "a ku", + "a l", + "sare ta", + "i pp", + "shi ku", + "u chi", + "hito tsu", + "ha tara", + "tachi ba", + "shi ro", + "ka tachi", + "to mo", + "e te", + "me ru", + "ni chi", + "da re", + "ka tta", + "e ru", + "su ki", + "a ge", + "oo ki", + "ma ru", + "mo ku", + "o ko", + "kangae rareru", + "o to", + "tan ni", + "ta da", + "tai teki", + "mo tte", + "ki nou", + "shi nai", + "k ki", + "u e", + "ta ri", + "l i", + "ra nai", + "k kou", + "mi rai", + "pp on", + "go to", + "hi n", + "hi tsu", + "te ru", + "mo chi", + "ka tsu", + "re n", + "n yuu", + "su i", + "zu ka", + "tsu ite", + "no mi", + "su gu", + "ku da", + "tetsu gaku", + "i ka", + "ron ri", + "o ki", + "ni ppon", + "p er", + "shi mashita", + "chi shiki", + "cho kkanteki", + "su ko", + "t ion", + "ku u", + "a na", + "a rou", + "ka tte", + "ku ri", + "i nai", + "hyou gen", + "i shiki", + "do ku", + "a tte", + "a tara", + "to n", + "wa ri", + "ka o", + "sei san", + "hana shi", + "s i", + "ka ke", + "na ji", + "su nawa", + "sunawa chi", + "u go", + "su u", + "ba ra", + "le v", + "hi ro", + "i wa", + "be tsu", + "yo i", + "se ru", + "shite ru", + "rare te", + "to shi", + "se ki", + "tai ritsu", + "wa kara", + "to kyo", + "k ka", + "k yoku", + "u n", + "i ro", + "mi te", + "sa ki", + "kan ji", + "mi ta", + "su be", + "r yoku", + "ma tta", + "kuda sai", + "omo i", + "ta no", + "ware ru", + "co m", + "hitsu you", + "ka shi", + "re nai", + "kan kei", + "a to", + "ga tte", + "o chi", + "mo tsu", + "in g", + "son zai", + "l l", + "o re", + "tai shite", + "a me", + "sei mei", + "ka no", + "gi ri", + "kangae ru", + "yu e", + "a sa", + "o naji", + "yo ru", + "ni ku", + "osa ka", + "suko shi", + "c k", + "ta ma", + "kano jo", + "ki te", + "mon dai", + "a mari", + "e ki", + "ko jin", + "ha ya", + "i t", + "de te", + "atara shii", + "a wa", + "ga kkou", + "tsu zu", + "shu kan", + "i mashita", + "mi na", + "ata e", + "da rou", + "hatara ku", + "ga ta", + "da chi", + "ma tsu", + "ari masen", + "sei butsu", + "mi tsu", + "he ya", + "yasu i", + "d i", + "de ni", + "no ko", + "ha ha", + "do mo", + "ka mi", + "su deni", + "na o", + "ra ku", + "i ke", + "a ki", + "me ta", + "l o", + "ko domo", + "so shite", + "ga me", + "ba kari", + "to te", + "ha tsu", + "mi se", + "moku teki", + "da kara", + "s z", + "e l", + "g y", + "e n", + "t t", + "e m", + "a n", + "a k", + "e r", + "a z", + "a l", + "e t", + "o l", + "e g", + "e k", + "m i", + "o n", + "é s", + "c s", + "a t", + "á r", + "h o", + "e z", + "á l", + "i s", + "á n", + "o r", + "a r", + "e gy", + "e s", + "é r", + "á t", + "o tt", + "e tt", + "m eg", + "t a", + "o k", + "o s", + "ho gy", + "n em", + "é g", + "n y", + "k i", + "é l", + "h a", + "á s", + "ü l", + "i n", + "mi n", + "n a", + "e d", + "o m", + "i k", + "k ö", + "m a", + "n i", + "v a", + "v ol", + "é t", + "b b", + "f el", + "i g", + "l e", + "r a", + "é n", + "t e", + "d e", + "a d", + "ó l", + "b e", + "on d", + "j a", + "r e", + "u l", + "b en", + "n ek", + "u t", + "vol t", + "b an", + "ö r", + "o g", + "a p", + "o d", + "á g", + "n k", + "é k", + "v al", + "k or", + "a m", + "i l", + "í t", + "á k", + "b a", + "u d", + "sz er", + "min d", + "o z", + "é p", + "el l", + "ér t", + "m ond", + "i t", + "sz t", + "n ak", + "a mi", + "n e", + "ő l", + "cs ak", + "n é", + "ma g", + "ol y", + "m er", + "ál l", + "án y", + "ö n", + "ö l", + "min t", + "m ár", + "ö tt", + "na gy", + "é sz", + "az t", + "el ő", + "t ud", + "o t", + "é ny", + "á z", + "m ég", + "kö z", + "el y", + "s ég", + "en t", + "s em", + "ta m", + "h et", + "h al", + "f i", + "a s", + "v an", + "ho z", + "v e", + "u k", + "k ez", + "á m", + "v el", + "b er", + "a j", + "u nk", + "i z", + "va gy", + "m os", + "sz em", + "em ber", + "f og", + "mer t", + "ü k", + "l en", + "ö s", + "e j", + "t al", + "h at", + "t ak", + "h i", + "m ás", + "s ág", + "ett e", + "l eg", + "ü nk", + "h át", + "sz a", + "on y", + "ez t", + "mind en", + "en d", + "ül t", + "h an", + "j ó", + "k is", + "á j", + "in t", + "ú gy", + "i d", + "mos t", + "ar t", + "í r", + "k er", + "i tt", + "a tt", + "el t", + "mond ta", + "k ell", + "l á", + "ak i", + "ál t", + "ér d", + "t ö", + "l an", + "v ár", + "h ol", + "t el", + "l át", + "ő k", + "v et", + "s e", + "ut án", + "k ét", + "na p", + "í v", + "ál y", + "v ég", + "ö k", + "i r", + "d ul", + "v is", + "né z", + "t er", + "á ban", + "k ül", + "ak kor", + "k ap", + "sz él", + "y en", + "ú j", + "i m", + "oly an", + "es en", + "k ed", + "h ely", + "t ör", + "b ól", + "el m", + "r á", + "ár a", + "r ó", + "l ó", + "vol na", + "t an", + "le het", + "e bb", + "t en", + "t ek", + "s ok", + "k al", + "f or", + "u g", + "ol t", + "k a", + "ek et", + "b or", + "f ej", + "g ond", + "a g", + "ak ar", + "f él", + "ú l", + "b el", + "ott a", + "mi t", + "val ami", + "j el", + "é d", + "ar c", + "u r", + "hal l", + "t i", + "f öl", + "á ba", + "ol g", + "ki r", + "ol d", + "m ar", + "k érd", + "j ár", + "ú r", + "sz e", + "z s", + "él et", + "j át", + "o v", + "u s", + "é z", + "v il", + "v er", + "ő r", + "á d", + "ö g", + "le sz", + "on t", + "b iz", + "k oz", + "á bb", + "kir ály", + "es t", + "a b", + "en g", + "ig az", + "b ar", + "ha j", + "d i", + "o b", + "k od", + "r ól", + "v ez", + "tö bb", + "sz ó", + "é ben", + "ö t", + "ny i", + "t á", + "sz ól", + "gond ol", + "eg ész", + "í gy", + "ő s", + "o bb", + "os an", + "b ől", + "a bb", + "c i", + "ő t", + "n ál", + "k ép", + "azt án", + "v i", + "t art", + "be szél", + "m en", + "elő tt", + "a szt", + "ma j", + "kö r", + "han g", + "í z", + "in cs", + "a i", + "é v", + "ó d", + "ó k", + "hoz z", + "t em", + "ok at", + "an y", + "nagy on", + "h áz", + "p er", + "p ed", + "ez te", + "et len", + "nek i", + "maj d", + "sz ony", + "án ak", + "fel é", + "egy szer", + "j e", + "ad t", + "gy er", + "ami kor", + "f oly", + "sz ak", + "ő d", + "h ú", + "á sz", + "am ely", + "h ar", + "ér e", + "il yen", + "od a", + "j ák", + "t ár", + "á val", + "l ak", + "t ó", + "m ent", + "gy an", + "él y", + "ú t", + "v ar", + "kez d", + "m ell", + "mi kor", + "h ez", + "val ó", + "k o", + "m es", + "szer et", + "r end", + "l et", + "vis sza", + "ig en", + "f ő", + "va s", + "as szony", + "r ől", + "ped ig", + "p i", + "sz ép", + "t ák", + "ö v", + "an i", + "vil ág", + "p en", + "mag a", + "t et", + "sz ik", + "é j", + "én t", + "j ött", + "s an", + "sz í", + "i de", + "g at", + "ett em", + "ul t", + "h ány", + "ás t", + "a hol", + "ők et", + "h ár", + "k el", + "n ő", + "cs i", + "tal ál", + "el te", + "lá tt", + "tör t", + "ha gy", + "e sz", + "s en", + "n él", + "p ar", + "v ál", + "k ut", + "l ány", + "ami t", + "s ő", + "ell en", + "mag át", + "in k", + "u gyan", + "kül ön", + "a sz", + "mind ig", + "l ép", + "tal án", + "u n", + "sz or", + "k e", + "il lan", + "n incs", + "z et", + "vagy ok", + "tel en", + "is mer", + "s or", + "is ten", + "ít ott", + "j obb", + "v es", + "dul t", + "j uk", + "sz en", + "r o", + "ö m", + "l ett", + "k ar", + "egy ik", + "b ár", + "sz i", + "sz ív", + "az on", + "e szt", + "föl d", + "kut y", + "p illan", + "f ér", + "k om", + "t ől", + "t ű", + "é be", + "t ött", + "bar át", + "í g", + "a hogy", + "e h", + "e p", + "s o", + "v en", + "jel ent", + "t at", + "sz eg", + "mint ha", + "f al", + "egy en", + "mi l", + "sza b", + "r i", + "é m", + "biz ony", + "j on", + "ör eg", + "d olg", + "cs ap", + "ti szt", + "áll t", + "an cs", + "id ő", + "k at", + "ü gy", + "mi ért", + "ó t", + "ü r", + "cs in", + "h az", + "b et", + "én ek", + "v ér", + "j ól", + "al att", + "m ely", + "l o", + "sem mi", + "ny ug", + "v ág", + "kö vet", + "ös sze", + "ma d", + "l i", + "a cs", + "fi ú", + "kö n", + "más ik", + "j ön", + "sz ám", + "g er", + "s ó", + "r ész", + "k ér", + "z el", + "é vel", + "e o", + "e u", + "a n", + "eu l", + "eu n", + "eo n", + "a e", + "d a", + "a l", + "s s", + "i n", + "i l", + "a g", + "an g", + "y eon", + "y eo", + "d o", + "c h", + "n g", + "j i", + "h an", + "g a", + "g o", + "u i", + "h ae", + "a m", + "u l", + "u n", + "g eo", + "s i", + "n eun", + "ss da", + "s eo", + "eon g", + "y o", + "i da", + "t t", + "k k", + "j eo", + "d eul", + "w a", + "eu m", + "g e", + "o n", + "o g", + "s al", + "m an", + "yeon g", + "geo s", + "h ag", + "an eun", + "j a", + "g i", + "s u", + "i ss", + "o l", + "d ae", + "eo b", + "h a", + "j u", + "eo l", + "g eu", + "j eong", + "s ae", + "do e", + "g eul", + "s eu", + "s in", + "eul o", + "b n", + "s ang", + "bn ida", + "h al", + "b o", + "han eun", + "m al", + "i m", + "m o", + "b u", + "jeo g", + "sae ng", + "in eun", + "an h", + "m a", + "sal am", + "j o", + "s a", + "eo m", + "n ae", + "w i", + "l o", + "g wa", + "yeo l", + "n a", + "e seo", + "y e", + "m yeon", + "tt ae", + "h w", + "j e", + "eob s", + "j ang", + "g u", + "g w", + "il eul", + "yeo g", + "j eon", + "si g", + "j ag", + "j in", + "y u", + "o e", + "s e", + "hag o", + "d eun", + "y a", + "m un", + "s eong", + "g ag", + "h am", + "d ang", + "b a", + "l eul", + "s il", + "do ng", + "kk a", + "b al", + "da l", + "han da", + "eo ssda", + "ae g", + "l i", + "ha ji", + "s eon", + "o ng", + "hae ssda", + "d e", + "i ssda", + "e ge", + "b un", + "m ul", + "ju ng", + "ji g", + "m u", + "iss neun", + "b i", + "g eun", + "seu bnida", + "w on", + "p p", + "d aneun", + "eo h", + "d eo", + "ga m", + "j al", + "hae ng", + "ag o", + "y ang", + "b ul", + "b ang", + "u m", + "s o", + "h i", + "j ae", + "si m", + "saeng gag", + "hag e", + "s og", + "eo ss", + "d an", + "ja sin", + "j il", + "eo g", + "g yeong", + "doe n", + "go ng", + "m i", + "ch i", + "d eu", + "d eon", + "hae ss", + "d u", + "n am", + "eun g", + "jo h", + "n al", + "m yeong", + "w o", + "eon a", + "i go", + "g yeol", + "y ag", + "gw an", + "ul i", + "yo ng", + "n o", + "l yeo", + "j og", + "eoh ge", + "ga t", + "b og", + "mo s", + "t ong", + "ch a", + "man h", + "jeo l", + "geo l", + "h oe", + "ag a", + "n aneun", + "g an", + "un eun", + "ch eol", + "ch e", + "do l", + "b on", + "b an", + "ba d", + "ch u", + "ham yeon", + "yeo ssda", + "i bnida", + "g ye", + "eo s", + "hw al", + "salam deul", + "ji man", + "dang sin", + "ji b", + "ttae mun", + "m ae", + "i b", + "e neun", + "eu g", + "jeo m", + "geul eon", + "h wa", + "a ssda", + "b eob", + "bu t", + "b ae", + "yeo ss", + "ch in", + "ch aeg", + "g eon", + "g ae", + "nae ga", + "i ga", + "m og", + "sig an", + "g il", + "h yeon", + "l yeog", + "gu g", + "p yeon", + "s an", + "w ae", + "j ul", + "s eul", + "deun g", + "haji man", + "eum yeon", + "p il", + "m ol", + "n eu", + "a ss", + "n yeon", + "t ae", + "h u", + "p yo", + "s ul", + "g ang", + "j ineun", + "b eon", + "ha da", + "seo l", + "si p", + "dal eun", + "a p", + "sal m", + "g yo", + "ch eon", + "hag i", + "in a", + "cheol eom", + "g al", + "il a", + "kka ji", + "anh neun", + "ha bnida", + "tt eon", + "n u", + "hae seo", + "doen da", + "s ol", + "tt al", + "l a", + "il o", + "seu b", + "b yeon", + "m yeo", + "b eol", + "s on", + "n un", + "j un", + "j am", + "j eung", + "tt o", + "e n", + "mo m", + "h o", + "ch im", + "hw ang", + "eun eun", + "jo ng", + "bo da", + "n ol", + "n eom", + "but eo", + "jig eum", + "eobs da", + "dae lo", + "i g", + "y ul", + "p yeong", + "seon eun", + "sal ang", + "seu t", + "h im", + "n an", + "h eom", + "h yang", + "p i", + "gw ang", + "eobs neun", + "hw ag", + "ge ss", + "jag i", + "il eon", + "wi hae", + "dae han", + "ga ji", + "m eog", + "j yeo", + "cha j", + "b yeong", + "eo d", + "g yeo", + "do n", + "eo ji", + "g ul", + "mo deun", + "j on", + "in saeng", + "geul ae", + "h ang", + "sa sil", + "si b", + "ch al", + "il ago", + "doe l", + "g eum", + "doe neun", + "b ol", + "ga jang", + "geul igo", + "e l", + "h yeong", + "haeng bog", + "ch ul", + "h on", + "ch ae", + "s am", + "m ang", + "in da", + "da m", + "w ol", + "ch oe", + "d ul", + "si jag", + "ch eong", + "il aneun", + "ul ineun", + "ae n", + "kk e", + "mun je", + "a do", + "t eu", + "g un", + "geun eun", + "b ge", + "ch eo", + "b aeg", + "ju g", + "t a", + "sang dae", + "geu geos", + "do g", + "eu s", + "deu s", + "ja b", + "h yeo", + "tt eohge", + "u g", + "ma j", + "ch il", + "s wi", + "j ileul", + "ch ang", + "g aneun", + "m ag", + "i ji", + "da go", + "m in", + "yo han", + "t eug", + "pp un", + "al eul", + "haeng dong", + "p o", + "m il", + "ch am", + "se sang", + "e do", + "p an", + "man deul", + "am yeon", + "a b", + "kk ae", + "b ag", + "i deul", + "p um", + "m eol", + "s un", + "n eul", + "ham kke", + "chu ng", + "da b", + "yu g", + "s ag", + "gwang ye", + "il eohge", + "bal o", + "neun de", + "ham yeo", + "go s", + "geul eoh", + "an ila", + "bang beob", + "da si", + "b yeol", + "g yeon", + "gam jeong", + "on eul", + "j aneun", + "yeo m", + "l ago", + "i gi", + "hw an", + "t eul", + "eo seo", + "si k", + "ch o", + "jag a", + "geul eom", + "geul eona", + "jeong do", + "g yeog", + "geul eohge", + "geu deul", + "eu t", + "im yeon", + "j jae", + "k eun", + "i sang", + "mal haessda", + "eu ge", + "no p", + "in gan", + "bo myeon", + "t aeg", + "seu s", + "d wi", + "s aneun", + "w an", + "anh go", + "t an", + "nu gu", + "su ng", + "da myeon", + "a deul", + "p eul", + "ttal a", + "d i", + "geos do", + "a ji", + "m eon", + "eum yeo", + "dol og", + "neun g", + "mo du", + "क े", + "ह ै", + "े ं", + "् र", + "ा र", + "न े", + "य ा", + "म ें", + "स े", + "क ी", + "क ा", + "ो ं", + "त ा", + "क र", + "स ्", + "क ि", + "क ो", + "र ्", + "न ा", + "क ्", + "ह ी", + "औ र", + "प र", + "त े", + "ह ो", + "प ्र", + "ा न", + "् य", + "ल ा", + "व ा", + "ल े", + "स ा", + "है ं", + "ल ि", + "ज ा", + "ह ा", + "भ ी", + "व ि", + "इ स", + "त ी", + "न ्", + "र ा", + "म ा", + "द े", + "द ि", + "ब ा", + "त ि", + "थ ा", + "न ि", + "क ार", + "ए क", + "ही ं", + "ह ु", + "ं ग", + "ै ं", + "न ी", + "स ी", + "अ प", + "त ्", + "न हीं", + "र ी", + "म े", + "म ु", + "ि त", + "त ो", + "प ा", + "ल ी", + "लि ए", + "ग ा", + "ल ्", + "र ह", + "र े", + "क् ष", + "म ैं", + "स म", + "उ स", + "ज ि", + "त ्र", + "म ि", + "च ा", + "ो ग", + "स ं", + "द ्", + "स ि", + "आ प", + "त ु", + "द ा", + "क ु", + "य ों", + "व े", + "ज ी", + "् या", + "उ न", + "ि क", + "य े", + "भ ा", + "् ट", + "ह म", + "स् ट", + "श ा", + "ड ़", + "ं द", + "ख ा", + "म ्", + "श ्", + "य ह", + "स क", + "प ू", + "कि या", + "अप ने", + "र ू", + "स ु", + "म ी", + "ह ि", + "ज ो", + "थ े", + "र ि", + "द ी", + "थ ी", + "ग ी", + "ल ोग", + "ग या", + "त र", + "न् ह", + "च ्", + "व ार", + "ब ी", + "प ्", + "द ो", + "ट ी", + "श ि", + "कर ने", + "ग े", + "ै से", + "इ न", + "ं ड", + "सा थ", + "प ु", + "ब े", + "ब ार", + "व ी", + "अ न", + "ह र", + "उ न्ह", + "हो ता", + "ज ब", + "कु छ", + "म ान", + "क ्र", + "ब ि", + "प ह", + "फ ि", + "स र", + "ार ी", + "र ो", + "द ू", + "क हा", + "त क", + "श न", + "ब ्", + "स् थ", + "व ह", + "बा द", + "ओ ं", + "ग ु", + "ज ्", + "्र े", + "ग र", + "रह े", + "व र्", + "ह ू", + "ार ्", + "प ी", + "ब हु", + "मु झ", + "्र ा", + "दि या", + "स ब", + "कर ते", + "अप नी", + "बहु त", + "क ह", + "ट े", + "हु ए", + "कि सी", + "र हा", + "ष ्ट", + "ज ़", + "ब ना", + "स ो", + "ड ि", + "को ई", + "व ्य", + "बा त", + "र ु", + "व ो", + "मुझ े", + "द् ध", + "च ार", + "मे रे", + "व र", + "्र ी", + "जा ता", + "न ों", + "प्र ा", + "दे ख", + "ट ा", + "क् या", + "अ ध", + "ल ग", + "ल ो", + "प ि", + "य ु", + "च े", + "जि स", + "ं त", + "ान ी", + "प ै", + "ज न", + "ार े", + "च ी", + "मि ल", + "द ु", + "दे श", + "च् छ", + "ष ्", + "स ू", + "ख े", + "च ु", + "ि या", + "ल गा", + "ब ु", + "उन के", + "ज् ञ", + "क्ष ा", + "त रह", + "्या दा", + "वा ले", + "पू र्", + "मैं ने", + "का म", + "रू प", + "हो ती", + "उ प", + "ज ान", + "प्र कार", + "भ ार", + "म न", + "हु आ", + "ट र", + "हू ँ", + "पर ि", + "पा स", + "अन ु", + "रा ज", + "लोग ों", + "अ ब", + "सम झ", + "ड ी", + "म ौ", + "श ु", + "च ि", + "प े", + "क ृ", + "सक ते", + "म ह", + "य ोग", + "द र्", + "उ से", + "ं ध", + "ड ा", + "जा ए", + "ब ो", + "ू ल", + "म ो", + "ों ने", + "ं स", + "तु म", + "पह ले", + "ब ता", + "त था", + "य ो", + "ग ई", + "उ त्", + "सक ता", + "क म", + "ज ्यादा", + "र ख", + "सम य", + "ार ा", + "अ गर", + "स् त", + "च ल", + "फि र", + "वार ा", + "कर ना", + "श ी", + "ग ए", + "ब न", + "ौ र", + "हो ने", + "चा ह", + "ख ु", + "हा ँ", + "उन्ह ें", + "उन्ह ोंने", + "छ ो", + "म् ह", + "प्र ति", + "नि क", + "व न", + "्य ू", + "र ही", + "तु म्ह", + "ज ैसे", + "ि यों", + "क् यों", + "ल ों", + "फ ़", + "ं त्र", + "हो ते", + "क् ति", + "त ्य", + "कर ्", + "क ई", + "व ं", + "कि न", + "प ो", + "कार ण", + "ड़ ी", + "भ ि", + "इस के", + "ब र", + "उस के", + "द् वारा", + "श े", + "क ॉ", + "दि न", + "न् न", + "ड़ ा", + "स् व", + "नि र्", + "मु ख", + "लि या", + "ट ि", + "ज्ञ ान", + "क् त", + "द ्र", + "ग ्", + "क् स", + "म ै", + "ग ो", + "ज े", + "ट ्र", + "म ार", + "त् व", + "ध ार", + "भा व", + "कर ता", + "ख ि", + "क ं", + "चा हि", + "य र", + "प् त", + "क ों", + "ं च", + "ज ु", + "म त", + "अ च्छ", + "हु ई", + "क भी", + "ले किन", + "भ ू", + "अप ना", + "दू स", + "चाहि ए", + "य ू", + "घ र", + "सब से", + "मे री", + "ना म", + "ढ ़", + "ं ट", + "ें गे", + "ब ै", + "फ ा", + "ए वं", + "य ी", + "ग ्र", + "क्ष े", + "आ ज", + "आप को", + "भा ग", + "ठ ा", + "क ै", + "भार त", + "उन की", + "प हु", + "स भी", + "ध ा", + "ण ा", + "स ान", + "हो गा", + "त ब", + "स ंग", + "प र्", + "अ व", + "त ना", + "ग ि", + "य न", + "स् था", + "च ित", + "ट ्", + "छ ा", + "जा ने", + "क्षे त्र", + "वा ली", + "पूर् ण", + "स मा", + "कार ी" + ] + } +} \ No newline at end of file diff --git a/comfy/text_encoders/ace_text_cleaners.py b/comfy/text_encoders/ace_text_cleaners.py new file mode 100644 index 000000000..cd31d8d8c --- /dev/null +++ b/comfy/text_encoders/ace_text_cleaners.py @@ -0,0 +1,395 @@ +# basic text cleaners for the ACE step model +# I didn't copy the ones from the reference code because I didn't want to deal with the dependencies +# TODO: more languages than english? + +import re + +def japanese_to_romaji(japanese_text): + """ + Convert Japanese hiragana and katakana to romaji (Latin alphabet representation). + + Args: + japanese_text (str): Text containing hiragana and/or katakana characters + + Returns: + str: The romaji (Latin alphabet) equivalent + """ + # Dictionary mapping kana characters to their romaji equivalents + kana_map = { + # Katakana characters + 'ア': 'a', 'イ': 'i', 'ウ': 'u', 'エ': 'e', 'オ': 'o', + 'カ': 'ka', 'キ': 'ki', 'ク': 'ku', 'ケ': 'ke', 'コ': 'ko', + 'サ': 'sa', 'シ': 'shi', 'ス': 'su', 'セ': 'se', 'ソ': 'so', + 'タ': 'ta', 'チ': 'chi', 'ツ': 'tsu', 'テ': 'te', 'ト': 'to', + 'ナ': 'na', 'ニ': 'ni', 'ヌ': 'nu', 'ネ': 'ne', 'ノ': 'no', + 'ハ': 'ha', 'ヒ': 'hi', 'フ': 'fu', 'ヘ': 'he', 'ホ': 'ho', + 'マ': 'ma', 'ミ': 'mi', 'ム': 'mu', 'メ': 'me', 'モ': 'mo', + 'ヤ': 'ya', 'ユ': 'yu', 'ヨ': 'yo', + 'ラ': 'ra', 'リ': 'ri', 'ル': 'ru', 'レ': 're', 'ロ': 'ro', + 'ワ': 'wa', 'ヲ': 'wo', 'ン': 'n', + + # Katakana voiced consonants + 'ガ': 'ga', 'ギ': 'gi', 'グ': 'gu', 'ゲ': 'ge', 'ゴ': 'go', + 'ザ': 'za', 'ジ': 'ji', 'ズ': 'zu', 'ゼ': 'ze', 'ゾ': 'zo', + 'ダ': 'da', 'ヂ': 'ji', 'ヅ': 'zu', 'デ': 'de', 'ド': 'do', + 'バ': 'ba', 'ビ': 'bi', 'ブ': 'bu', 'ベ': 'be', 'ボ': 'bo', + 'パ': 'pa', 'ピ': 'pi', 'プ': 'pu', 'ペ': 'pe', 'ポ': 'po', + + # Katakana combinations + 'キャ': 'kya', 'キュ': 'kyu', 'キョ': 'kyo', + 'シャ': 'sha', 'シュ': 'shu', 'ショ': 'sho', + 'チャ': 'cha', 'チュ': 'chu', 'チョ': 'cho', + 'ニャ': 'nya', 'ニュ': 'nyu', 'ニョ': 'nyo', + 'ヒャ': 'hya', 'ヒュ': 'hyu', 'ヒョ': 'hyo', + 'ミャ': 'mya', 'ミュ': 'myu', 'ミョ': 'myo', + 'リャ': 'rya', 'リュ': 'ryu', 'リョ': 'ryo', + 'ギャ': 'gya', 'ギュ': 'gyu', 'ギョ': 'gyo', + 'ジャ': 'ja', 'ジュ': 'ju', 'ジョ': 'jo', + 'ビャ': 'bya', 'ビュ': 'byu', 'ビョ': 'byo', + 'ピャ': 'pya', 'ピュ': 'pyu', 'ピョ': 'pyo', + + # Katakana small characters and special cases + 'ッ': '', # Small tsu (doubles the following consonant) + 'ャ': 'ya', 'ュ': 'yu', 'ョ': 'yo', + + # Katakana extras + 'ヴ': 'vu', 'ファ': 'fa', 'フィ': 'fi', 'フェ': 'fe', 'フォ': 'fo', + 'ウィ': 'wi', 'ウェ': 'we', 'ウォ': 'wo', + + # Hiragana characters + 'あ': 'a', 'い': 'i', 'う': 'u', 'え': 'e', 'お': 'o', + 'か': 'ka', 'き': 'ki', 'く': 'ku', 'け': 'ke', 'こ': 'ko', + 'さ': 'sa', 'し': 'shi', 'す': 'su', 'せ': 'se', 'そ': 'so', + 'た': 'ta', 'ち': 'chi', 'つ': 'tsu', 'て': 'te', 'と': 'to', + 'な': 'na', 'に': 'ni', 'ぬ': 'nu', 'ね': 'ne', 'の': 'no', + 'は': 'ha', 'ひ': 'hi', 'ふ': 'fu', 'へ': 'he', 'ほ': 'ho', + 'ま': 'ma', 'み': 'mi', 'む': 'mu', 'め': 'me', 'も': 'mo', + 'や': 'ya', 'ゆ': 'yu', 'よ': 'yo', + 'ら': 'ra', 'り': 'ri', 'る': 'ru', 'れ': 're', 'ろ': 'ro', + 'わ': 'wa', 'を': 'wo', 'ん': 'n', + + # Hiragana voiced consonants + 'が': 'ga', 'ぎ': 'gi', 'ぐ': 'gu', 'げ': 'ge', 'ご': 'go', + 'ざ': 'za', 'じ': 'ji', 'ず': 'zu', 'ぜ': 'ze', 'ぞ': 'zo', + 'だ': 'da', 'ぢ': 'ji', 'づ': 'zu', 'で': 'de', 'ど': 'do', + 'ば': 'ba', 'び': 'bi', 'ぶ': 'bu', 'べ': 'be', 'ぼ': 'bo', + 'ぱ': 'pa', 'ぴ': 'pi', 'ぷ': 'pu', 'ぺ': 'pe', 'ぽ': 'po', + + # Hiragana combinations + 'きゃ': 'kya', 'きゅ': 'kyu', 'きょ': 'kyo', + 'しゃ': 'sha', 'しゅ': 'shu', 'しょ': 'sho', + 'ちゃ': 'cha', 'ちゅ': 'chu', 'ちょ': 'cho', + 'にゃ': 'nya', 'にゅ': 'nyu', 'にょ': 'nyo', + 'ひゃ': 'hya', 'ひゅ': 'hyu', 'ひょ': 'hyo', + 'みゃ': 'mya', 'みゅ': 'myu', 'みょ': 'myo', + 'りゃ': 'rya', 'りゅ': 'ryu', 'りょ': 'ryo', + 'ぎゃ': 'gya', 'ぎゅ': 'gyu', 'ぎょ': 'gyo', + 'じゃ': 'ja', 'じゅ': 'ju', 'じょ': 'jo', + 'びゃ': 'bya', 'びゅ': 'byu', 'びょ': 'byo', + 'ぴゃ': 'pya', 'ぴゅ': 'pyu', 'ぴょ': 'pyo', + + # Hiragana small characters and special cases + 'っ': '', # Small tsu (doubles the following consonant) + 'ゃ': 'ya', 'ゅ': 'yu', 'ょ': 'yo', + + # Common punctuation and spaces + '　': ' ', # Japanese space + '、': ', ', '。': '. ', + } + + result = [] + i = 0 + + while i < len(japanese_text): + # Check for small tsu (doubling the following consonant) + if i < len(japanese_text) - 1 and (japanese_text[i] == 'っ' or japanese_text[i] == 'ッ'): + if i < len(japanese_text) - 1 and japanese_text[i+1] in kana_map: + next_romaji = kana_map[japanese_text[i+1]] + if next_romaji and next_romaji[0] not in 'aiueon': + result.append(next_romaji[0]) # Double the consonant + i += 1 + continue + + # Check for combinations with small ya, yu, yo + if i < len(japanese_text) - 1 and japanese_text[i+1] in ('ゃ', 'ゅ', 'ょ', 'ャ', 'ュ', 'ョ'): + combo = japanese_text[i:i+2] + if combo in kana_map: + result.append(kana_map[combo]) + i += 2 + continue + + # Regular character + if japanese_text[i] in kana_map: + result.append(kana_map[japanese_text[i]]) + else: + # If it's not in our map, keep it as is (might be kanji, romaji, etc.) + result.append(japanese_text[i]) + + i += 1 + + return ''.join(result) + +def number_to_text(num, ordinal=False): + """ + Convert a number (int or float) to its text representation. + + Args: + num: The number to convert + + Returns: + str: Text representation of the number + """ + + if not isinstance(num, (int, float)): + return "Input must be a number" + + # Handle special case of zero + if num == 0: + return "zero" + + # Handle negative numbers + negative = num < 0 + num = abs(num) + + # Handle floats + if isinstance(num, float): + # Split into integer and decimal parts + int_part = int(num) + + # Convert both parts + int_text = _int_to_text(int_part) + + # Handle decimal part (convert to string and remove '0.') + decimal_str = str(num).split('.')[1] + decimal_text = " point " + " ".join(_digit_to_text(int(digit)) for digit in decimal_str) + + result = int_text + decimal_text + else: + # Handle integers + result = _int_to_text(num) + + # Add 'negative' prefix for negative numbers + if negative: + result = "negative " + result + + return result + + +def _int_to_text(num): + """Helper function to convert an integer to text""" + + ones = ["", "one", "two", "three", "four", "five", "six", "seven", "eight", "nine", + "ten", "eleven", "twelve", "thirteen", "fourteen", "fifteen", "sixteen", + "seventeen", "eighteen", "nineteen"] + + tens = ["", "", "twenty", "thirty", "forty", "fifty", "sixty", "seventy", "eighty", "ninety"] + + if num < 20: + return ones[num] + + if num < 100: + return tens[num // 10] + (" " + ones[num % 10] if num % 10 != 0 else "") + + if num < 1000: + return ones[num // 100] + " hundred" + (" " + _int_to_text(num % 100) if num % 100 != 0 else "") + + if num < 1000000: + return _int_to_text(num // 1000) + " thousand" + (" " + _int_to_text(num % 1000) if num % 1000 != 0 else "") + + if num < 1000000000: + return _int_to_text(num // 1000000) + " million" + (" " + _int_to_text(num % 1000000) if num % 1000000 != 0 else "") + + return _int_to_text(num // 1000000000) + " billion" + (" " + _int_to_text(num % 1000000000) if num % 1000000000 != 0 else "") + + +def _digit_to_text(digit): + """Convert a single digit to text""" + digits = ["zero", "one", "two", "three", "four", "five", "six", "seven", "eight", "nine"] + return digits[digit] + + +_whitespace_re = re.compile(r"\s+") + + +# List of (regular expression, replacement) pairs for abbreviations: +_abbreviations = { + "en": [ + (re.compile("\\b%s\\." % x[0], re.IGNORECASE), x[1]) + for x in [ + ("mrs", "misess"), + ("mr", "mister"), + ("dr", "doctor"), + ("st", "saint"), + ("co", "company"), + ("jr", "junior"), + ("maj", "major"), + ("gen", "general"), + ("drs", "doctors"), + ("rev", "reverend"), + ("lt", "lieutenant"), + ("hon", "honorable"), + ("sgt", "sergeant"), + ("capt", "captain"), + ("esq", "esquire"), + ("ltd", "limited"), + ("col", "colonel"), + ("ft", "fort"), + ] + ], +} + + +def expand_abbreviations_multilingual(text, lang="en"): + for regex, replacement in _abbreviations[lang]: + text = re.sub(regex, replacement, text) + return text + + +_symbols_multilingual = { + "en": [ + (re.compile(r"%s" % re.escape(x[0]), re.IGNORECASE), x[1]) + for x in [ + ("&", " and "), + ("@", " at "), + ("%", " percent "), + ("#", " hash "), + ("$", " dollar "), + ("£", " pound "), + ("°", " degree "), + ] + ], +} + + +def expand_symbols_multilingual(text, lang="en"): + for regex, replacement in _symbols_multilingual[lang]: + text = re.sub(regex, replacement, text) + text = text.replace(" ", " ") # Ensure there are no double spaces + return text.strip() + + +_ordinal_re = { + "en": re.compile(r"([0-9]+)(st|nd|rd|th)"), +} +_number_re = re.compile(r"[0-9]+") +_currency_re = { + "USD": re.compile(r"((\$[0-9\.\,]*[0-9]+)|([0-9\.\,]*[0-9]+\$))"), + "GBP": re.compile(r"((£[0-9\.\,]*[0-9]+)|([0-9\.\,]*[0-9]+£))"), + "EUR": re.compile(r"(([0-9\.\,]*[0-9]+€)|((€[0-9\.\,]*[0-9]+)))"), +} + +_comma_number_re = re.compile(r"\b\d{1,3}(,\d{3})*(\.\d+)?\b") +_dot_number_re = re.compile(r"\b\d{1,3}(.\d{3})*(\,\d+)?\b") +_decimal_number_re = re.compile(r"([0-9]+[.,][0-9]+)") + + +def _remove_commas(m): + text = m.group(0) + if "," in text: + text = text.replace(",", "") + return text + + +def _remove_dots(m): + text = m.group(0) + if "." in text: + text = text.replace(".", "") + return text + + +def _expand_decimal_point(m, lang="en"): + amount = m.group(1).replace(",", ".") + return number_to_text(float(amount)) + + +def _expand_currency(m, lang="en", currency="USD"): + amount = float((re.sub(r"[^\d.]", "", m.group(0).replace(",", ".")))) + full_amount = number_to_text(amount) + + and_equivalents = { + "en": ", ", + "es": " con ", + "fr": " et ", + "de": " und ", + "pt": " e ", + "it": " e ", + "pl": ", ", + "cs": ", ", + "ru": ", ", + "nl": ", ", + "ar": ", ", + "tr": ", ", + "hu": ", ", + "ko": ", ", + } + + if amount.is_integer(): + last_and = full_amount.rfind(and_equivalents[lang]) + if last_and != -1: + full_amount = full_amount[:last_and] + + return full_amount + + +def _expand_ordinal(m, lang="en"): + return number_to_text(int(m.group(1)), ordinal=True) + + +def _expand_number(m, lang="en"): + return number_to_text(int(m.group(0))) + + +def expand_numbers_multilingual(text, lang="en"): + if lang in ["en", "ru"]: + text = re.sub(_comma_number_re, _remove_commas, text) + else: + text = re.sub(_dot_number_re, _remove_dots, text) + try: + text = re.sub(_currency_re["GBP"], lambda m: _expand_currency(m, lang, "GBP"), text) + text = re.sub(_currency_re["USD"], lambda m: _expand_currency(m, lang, "USD"), text) + text = re.sub(_currency_re["EUR"], lambda m: _expand_currency(m, lang, "EUR"), text) + except: + pass + + text = re.sub(_decimal_number_re, lambda m: _expand_decimal_point(m, lang), text) + text = re.sub(_ordinal_re[lang], lambda m: _expand_ordinal(m, lang), text) + text = re.sub(_number_re, lambda m: _expand_number(m, lang), text) + return text + + +def lowercase(text): + return text.lower() + + +def collapse_whitespace(text): + return re.sub(_whitespace_re, " ", text) + + +def multilingual_cleaners(text, lang): + text = text.replace('"', "") + if lang == "tr": + text = text.replace("İ", "i") + text = text.replace("Ö", "ö") + text = text.replace("Ü", "ü") + text = lowercase(text) + try: + text = expand_numbers_multilingual(text, lang) + except: + pass + try: + text = expand_abbreviations_multilingual(text, lang) + except: + pass + try: + text = expand_symbols_multilingual(text, lang=lang) + except: + pass + text = collapse_whitespace(text) + return text + + +def basic_cleaners(text): + """Basic pipeline that lowercases and collapses whitespace without transliteration.""" + text = lowercase(text) + text = collapse_whitespace(text) + return text diff --git a/comfy/text_encoders/flux.py b/comfy/text_encoders/flux.py index b16f97d78..d202c1396 100644 --- a/comfy/text_encoders/flux.py +++ b/comfy/text_encoders/flux.py @@ -25,8 +25,8 @@ class FluxTokenizer: def tokenize_with_weights(self, text: str, return_word_ids=False, **kwargs): out = { - "l": self.clip_l.tokenize_with_weights(text, return_word_ids), - "t5xxl": self.t5xxl.tokenize_with_weights(text, return_word_ids) + "l": self.clip_l.tokenize_with_weights(text, return_word_ids, **kwargs), + "t5xxl": self.t5xxl.tokenize_with_weights(text, return_word_ids, **kwargs) } return out diff --git a/comfy/text_encoders/hidream.py b/comfy/text_encoders/hidream.py index 8e1abcfc1..dbcf52784 100644 --- a/comfy/text_encoders/hidream.py +++ b/comfy/text_encoders/hidream.py @@ -16,11 +16,11 @@ class HiDreamTokenizer: def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs): out = {} - out["g"] = self.clip_g.tokenize_with_weights(text, return_word_ids) - out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids) - t5xxl = self.t5xxl.tokenize_with_weights(text, return_word_ids) + out["g"] = self.clip_g.tokenize_with_weights(text, return_word_ids, **kwargs) + out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids, **kwargs) + t5xxl = self.t5xxl.tokenize_with_weights(text, return_word_ids, **kwargs) out["t5xxl"] = [t5xxl[0]] # Use only first 128 tokens - out["llama"] = self.llama.tokenize_with_weights(text, return_word_ids) + out["llama"] = self.llama.tokenize_with_weights(text, return_word_ids, **kwargs) return out def untokenize(self, token_weight_pair): diff --git a/comfy/text_encoders/hunyuan_video.py b/comfy/text_encoders/hunyuan_video.py index 0adb3e27d..52ac19e74 100644 --- a/comfy/text_encoders/hunyuan_video.py +++ b/comfy/text_encoders/hunyuan_video.py @@ -59,13 +59,13 @@ class HunyuanVideoTokenizer: def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, image_embeds=None, image_interleave=1, **kwargs): out = {} - out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids) + out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids, **kwargs) if llama_template is None: llama_text = self.llama_template.format(text) else: llama_text = llama_template.format(text) - llama_text_tokens = self.llama.tokenize_with_weights(llama_text, return_word_ids) + llama_text_tokens = self.llama.tokenize_with_weights(llama_text, return_word_ids, **kwargs) embed_count = 0 for r in llama_text_tokens: for i in range(len(r)): diff --git a/comfy/text_encoders/hydit.py b/comfy/text_encoders/hydit.py index 7b3db6f29..436b1c6f8 100644 --- a/comfy/text_encoders/hydit.py +++ b/comfy/text_encoders/hydit.py @@ -56,8 +56,8 @@ class HyditTokenizer: def tokenize_with_weights(self, text: str, return_word_ids=False, **kwargs): out = {} - out["hydit_clip"] = self.hydit_clip.tokenize_with_weights(text, return_word_ids) - out["mt5xl"] = self.mt5xl.tokenize_with_weights(text, return_word_ids) + out["hydit_clip"] = self.hydit_clip.tokenize_with_weights(text, return_word_ids, **kwargs) + out["mt5xl"] = self.mt5xl.tokenize_with_weights(text, return_word_ids, **kwargs) return out def untokenize(self, token_weight_pair): diff --git a/comfy/text_encoders/long_clipl.json b/comfy/text_encoders/long_clipl.json deleted file mode 100644 index 5e2056ff3..000000000 --- a/comfy/text_encoders/long_clipl.json +++ /dev/null @@ -1,25 +0,0 @@ -{ - "_name_or_path": "openai/clip-vit-large-patch14", - "architectures": [ - "CLIPTextModel" - ], - "attention_dropout": 0.0, - "bos_token_id": 0, - "dropout": 0.0, - "eos_token_id": 49407, - "hidden_act": "quick_gelu", - "hidden_size": 768, - "initializer_factor": 1.0, - "initializer_range": 0.02, - "intermediate_size": 3072, - "layer_norm_eps": 1e-05, - "max_position_embeddings": 248, - "model_type": "clip_text_model", - "num_attention_heads": 12, - "num_hidden_layers": 12, - "pad_token_id": 1, - "projection_dim": 768, - "torch_dtype": "float32", - "transformers_version": "4.24.0", - "vocab_size": 49408 -} diff --git a/comfy/text_encoders/sd3_clip.py b/comfy/text_encoders/sd3_clip.py index 48a432b5c..e08de16ab 100644 --- a/comfy/text_encoders/sd3_clip.py +++ b/comfy/text_encoders/sd3_clip.py @@ -55,9 +55,9 @@ class SD3Tokenizer: def tokenize_with_weights(self, text: str, return_word_ids=False, **kwargs): out = {} - out["g"] = self.clip_g.tokenize_with_weights(text, return_word_ids) - out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids) - out["t5xxl"] = self.t5xxl.tokenize_with_weights(text, return_word_ids) + out["g"] = self.clip_g.tokenize_with_weights(text, return_word_ids, **kwargs) + out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids, **kwargs) + out["t5xxl"] = self.t5xxl.tokenize_with_weights(text, return_word_ids, **kwargs) return out def untokenize(self, token_weight_pair): diff --git a/comfy/text_encoders/umt5_config_base.json b/comfy/text_encoders/umt5_config_base.json new file mode 100644 index 000000000..6b3618f07 --- /dev/null +++ b/comfy/text_encoders/umt5_config_base.json @@ -0,0 +1,22 @@ +{ + "d_ff": 2048, + "d_kv": 64, + "d_model": 768, + "decoder_start_token_id": 0, + "dropout_rate": 0.1, + "eos_token_id": 1, + "dense_act_fn": "gelu_pytorch_tanh", + "initializer_factor": 1.0, + "is_encoder_decoder": true, + "is_gated_act": true, + "layer_norm_epsilon": 1e-06, + "model_type": "umt5", + "num_decoder_layers": 12, + "num_heads": 12, + "num_layers": 12, + "output_past": true, + "pad_token_id": 0, + "relative_attention_num_buckets": 32, + "tie_word_embeddings": false, + "vocab_size": 256384 +} diff --git a/comfy/utils.py b/comfy/utils.py index d8b18eb49..6f03e7669 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -17,37 +17,39 @@ """ from __future__ import annotations +import itertools +import math +import sys + import contextlib import contextvars -import itertools import json import logging -import math +import numpy as np import os import random +import safetensors.torch import struct -import sys +import torch import warnings +from PIL import Image from contextlib import contextmanager +from einops import rearrange from pathlib import Path from pickle import UnpicklingError -from typing import Optional, Any - -import numpy as np -import safetensors.torch -import torch -from PIL import Image -from einops import rearrange from torch.nn.functional import interpolate from tqdm import tqdm +from typing import Optional, Any from . import interruption, checkpoint_pickle +from .cli_args import args from .component_model import files from .component_model.deprecation import _deprecate_method from .component_model.executor_types import ExecutorToClientProgress, ProgressMessage from .component_model.queue_types import BinaryEventTypes from .execution_context import current_execution_context +MMAP_TORCH_FILES = args.mmap_torch_files logger = logging.getLogger(__name__) ALWAYS_SAFE_LOAD = False @@ -121,16 +123,14 @@ def load_torch_file(ckpt: str, safe_load=False, device=None, return_metadata=Fal sd.update(safetensors.torch.load_file(str(checkpoint_file), device=device.type)) else: try: + torch_args = {} + if MMAP_TORCH_FILES: + torch_args["mmap"] = True + if safe_load or ALWAYS_SAFE_LOAD: - if not 'weights_only' in torch.load.__code__.co_varnames: - logger.warning("Warning torch.load doesn't support weights_only on this pytorch version, loading unsafely.") - safe_load = False - if safe_load: - pl_sd = torch.load(ckpt, map_location=device, weights_only=True) + pl_sd = torch.load(ckpt, map_location=device, weights_only=True, **torch_args) else: pl_sd = torch.load(ckpt, map_location=device, pickle_module=checkpoint_pickle) - if "global_step" in pl_sd: - logger.debug(f"Global Step: {pl_sd['global_step']}") if "state_dict" in pl_sd: sd = pl_sd["state_dict"] else: diff --git a/comfy/weight_adapter/__init__.py b/comfy/weight_adapter/__init__.py new file mode 100644 index 000000000..d2a1d0151 --- /dev/null +++ b/comfy/weight_adapter/__init__.py @@ -0,0 +1,17 @@ +from .base import WeightAdapterBase +from .lora import LoRAAdapter +from .loha import LoHaAdapter +from .lokr import LoKrAdapter +from .glora import GLoRAAdapter +from .oft import OFTAdapter +from .boft import BOFTAdapter + + +adapters: list[type[WeightAdapterBase]] = [ + LoRAAdapter, + LoHaAdapter, + LoKrAdapter, + GLoRAAdapter, + OFTAdapter, + BOFTAdapter, +] diff --git a/comfy/weight_adapter/base.py b/comfy/weight_adapter/base.py new file mode 100644 index 000000000..29873519d --- /dev/null +++ b/comfy/weight_adapter/base.py @@ -0,0 +1,104 @@ +from typing import Optional + +import torch +import torch.nn as nn + +import comfy.model_management + + +class WeightAdapterBase: + name: str + loaded_keys: set[str] + weights: list[torch.Tensor] + + @classmethod + def load(cls, x: str, lora: dict[str, torch.Tensor]) -> Optional["WeightAdapterBase"]: + raise NotImplementedError + + def to_train(self) -> "WeightAdapterTrainBase": + raise NotImplementedError + + def calculate_weight( + self, + weight, + key, + strength, + strength_model, + offset, + function, + intermediate_dtype=torch.float32, + original_weight=None, + ): + raise NotImplementedError + + +class WeightAdapterTrainBase(nn.Module): + def __init__(self): + super().__init__() + + # [TODO] Collaborate with LoRA training PR #7032 + + +def weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function): + dora_scale = comfy.model_management.cast_to_device(dora_scale, weight.device, intermediate_dtype) + lora_diff *= alpha + weight_calc = weight + function(lora_diff).type(weight.dtype) + + wd_on_output_axis = dora_scale.shape[0] == weight_calc.shape[0] + if wd_on_output_axis: + weight_norm = ( + weight.reshape(weight.shape[0], -1) + .norm(dim=1, keepdim=True) + .reshape(weight.shape[0], *[1] * (weight.dim() - 1)) + ) + else: + weight_norm = ( + weight_calc.transpose(0, 1) + .reshape(weight_calc.shape[1], -1) + .norm(dim=1, keepdim=True) + .reshape(weight_calc.shape[1], *[1] * (weight_calc.dim() - 1)) + .transpose(0, 1) + ) + weight_norm = weight_norm + torch.finfo(weight.dtype).eps + + weight_calc *= (dora_scale / weight_norm).type(weight.dtype) + if strength != 1.0: + weight_calc -= weight + weight += strength * (weight_calc) + else: + weight[:] = weight_calc + return weight + + +def pad_tensor_to_shape(tensor: torch.Tensor, new_shape: list[int]) -> torch.Tensor: + """ + Pad a tensor to a new shape with zeros. + + Args: + tensor (torch.Tensor): The original tensor to be padded. + new_shape (List[int]): The desired shape of the padded tensor. + + Returns: + torch.Tensor: A new tensor padded with zeros to the specified shape. + + Note: + If the new shape is smaller than the original tensor in any dimension, + the original tensor will be truncated in that dimension. + """ + if any([new_shape[i] < tensor.shape[i] for i in range(len(new_shape))]): + raise ValueError("The new shape must be larger than the original tensor in all dimensions") + + if len(new_shape) != len(tensor.shape): + raise ValueError("The new shape must have the same number of dimensions as the original tensor") + + # Create a new tensor filled with zeros + padded_tensor = torch.zeros(new_shape, dtype=tensor.dtype, device=tensor.device) + + # Create slicing tuples for both tensors + orig_slices = tuple(slice(0, dim) for dim in tensor.shape) + new_slices = tuple(slice(0, dim) for dim in tensor.shape) + + # Copy the original tensor into the new tensor + padded_tensor[new_slices] = tensor[orig_slices] + + return padded_tensor diff --git a/comfy/weight_adapter/boft.py b/comfy/weight_adapter/boft.py new file mode 100644 index 000000000..b2a2f1bd4 --- /dev/null +++ b/comfy/weight_adapter/boft.py @@ -0,0 +1,115 @@ +import logging +from typing import Optional + +import torch +import comfy.model_management +from .base import WeightAdapterBase, weight_decompose + + +class BOFTAdapter(WeightAdapterBase): + name = "boft" + + def __init__(self, loaded_keys, weights): + self.loaded_keys = loaded_keys + self.weights = weights + + @classmethod + def load( + cls, + x: str, + lora: dict[str, torch.Tensor], + alpha: float, + dora_scale: torch.Tensor, + loaded_keys: set[str] = None, + ) -> Optional["BOFTAdapter"]: + if loaded_keys is None: + loaded_keys = set() + blocks_name = "{}.oft_blocks".format(x) + rescale_name = "{}.rescale".format(x) + + blocks = None + if blocks_name in lora.keys(): + blocks = lora[blocks_name] + if blocks.ndim == 4: + loaded_keys.add(blocks_name) + else: + blocks = None + if blocks is None: + return None + + rescale = None + if rescale_name in lora.keys(): + rescale = lora[rescale_name] + loaded_keys.add(rescale_name) + + weights = (blocks, rescale, alpha, dora_scale) + return cls(loaded_keys, weights) + + def calculate_weight( + self, + weight, + key, + strength, + strength_model, + offset, + function, + intermediate_dtype=torch.float32, + original_weight=None, + ): + v = self.weights + blocks = v[0] + rescale = v[1] + alpha = v[2] + dora_scale = v[3] + + blocks = comfy.model_management.cast_to_device(blocks, weight.device, intermediate_dtype) + if rescale is not None: + rescale = comfy.model_management.cast_to_device(rescale, weight.device, intermediate_dtype) + + boft_m, block_num, boft_b, *_ = blocks.shape + + try: + # Get r + I = torch.eye(boft_b, device=blocks.device, dtype=blocks.dtype) + # for Q = -Q^T + q = blocks - blocks.transpose(-1, -2) + normed_q = q + if alpha > 0: # alpha in boft/bboft is for constraint + q_norm = torch.norm(q) + 1e-8 + if q_norm > alpha: + normed_q = q * alpha / q_norm + # use float() to prevent unsupported type in .inverse() + r = (I + normed_q) @ (I - normed_q).float().inverse() + r = r.to(weight) + inp = org = weight + + r_b = boft_b//2 + for i in range(boft_m): + bi = r[i] + g = 2 + k = 2**i * r_b + if strength != 1: + bi = bi * strength + (1-strength) * I + inp = ( + inp.unflatten(0, (-1, g, k)) + .transpose(1, 2) + .flatten(0, 2) + .unflatten(0, (-1, boft_b)) + ) + inp = torch.einsum("b i j, b j ...-> b i ...", bi, inp) + inp = ( + inp.flatten(0, 1).unflatten(0, (-1, k, g)).transpose(1, 2).flatten(0, 2) + ) + + if rescale is not None: + inp = inp * rescale + + lora_diff = inp - org + lora_diff = comfy.model_management.cast_to_device(lora_diff, weight.device, intermediate_dtype) + if dora_scale is not None: + weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function) + else: + weight += function((strength * lora_diff).type(weight.dtype)) + except Exception as e: + logging.error("ERROR {} {} {}".format(self.name, key, e)) + return weight diff --git a/comfy/weight_adapter/glora.py b/comfy/weight_adapter/glora.py new file mode 100644 index 000000000..939abbba5 --- /dev/null +++ b/comfy/weight_adapter/glora.py @@ -0,0 +1,93 @@ +import logging +from typing import Optional + +import torch +import comfy.model_management +from .base import WeightAdapterBase, weight_decompose + + +class GLoRAAdapter(WeightAdapterBase): + name = "glora" + + def __init__(self, loaded_keys, weights): + self.loaded_keys = loaded_keys + self.weights = weights + + @classmethod + def load( + cls, + x: str, + lora: dict[str, torch.Tensor], + alpha: float, + dora_scale: torch.Tensor, + loaded_keys: set[str] = None, + ) -> Optional["GLoRAAdapter"]: + if loaded_keys is None: + loaded_keys = set() + a1_name = "{}.a1.weight".format(x) + a2_name = "{}.a2.weight".format(x) + b1_name = "{}.b1.weight".format(x) + b2_name = "{}.b2.weight".format(x) + if a1_name in lora: + weights = (lora[a1_name], lora[a2_name], lora[b1_name], lora[b2_name], alpha, dora_scale) + loaded_keys.add(a1_name) + loaded_keys.add(a2_name) + loaded_keys.add(b1_name) + loaded_keys.add(b2_name) + return cls(loaded_keys, weights) + else: + return None + + def calculate_weight( + self, + weight, + key, + strength, + strength_model, + offset, + function, + intermediate_dtype=torch.float32, + original_weight=None, + ): + v = self.weights + dora_scale = v[5] + + old_glora = False + if v[3].shape[1] == v[2].shape[0] == v[0].shape[0] == v[1].shape[1]: + rank = v[0].shape[0] + old_glora = True + + if v[3].shape[0] == v[2].shape[1] == v[0].shape[1] == v[1].shape[0]: + if old_glora and v[1].shape[0] == weight.shape[0] and weight.shape[0] == weight.shape[1]: + pass + else: + old_glora = False + rank = v[1].shape[0] + + a1 = comfy.model_management.cast_to_device(v[0].flatten(start_dim=1), weight.device, intermediate_dtype) + a2 = comfy.model_management.cast_to_device(v[1].flatten(start_dim=1), weight.device, intermediate_dtype) + b1 = comfy.model_management.cast_to_device(v[2].flatten(start_dim=1), weight.device, intermediate_dtype) + b2 = comfy.model_management.cast_to_device(v[3].flatten(start_dim=1), weight.device, intermediate_dtype) + + if v[4] is not None: + alpha = v[4] / rank + else: + alpha = 1.0 + + try: + if old_glora: + lora_diff = (torch.mm(b2, b1) + torch.mm(torch.mm(weight.flatten(start_dim=1).to(dtype=intermediate_dtype), a2), a1)).reshape(weight.shape) #old lycoris glora + else: + if weight.dim() > 2: + lora_diff = torch.einsum("o i ..., i j -> o j ...", torch.einsum("o i ..., i j -> o j ...", weight.to(dtype=intermediate_dtype), a1), a2).reshape(weight.shape) + else: + lora_diff = torch.mm(torch.mm(weight.to(dtype=intermediate_dtype), a1), a2).reshape(weight.shape) + lora_diff += torch.mm(b1, b2).reshape(weight.shape) + + if dora_scale is not None: + weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function) + else: + weight += function(((strength * alpha) * lora_diff).type(weight.dtype)) + except Exception as e: + logging.error("ERROR {} {} {}".format(self.name, key, e)) + return weight diff --git a/comfy/weight_adapter/loha.py b/comfy/weight_adapter/loha.py new file mode 100644 index 000000000..ce79abad5 --- /dev/null +++ b/comfy/weight_adapter/loha.py @@ -0,0 +1,100 @@ +import logging +from typing import Optional + +import torch +import comfy.model_management +from .base import WeightAdapterBase, weight_decompose + + +class LoHaAdapter(WeightAdapterBase): + name = "loha" + + def __init__(self, loaded_keys, weights): + self.loaded_keys = loaded_keys + self.weights = weights + + @classmethod + def load( + cls, + x: str, + lora: dict[str, torch.Tensor], + alpha: float, + dora_scale: torch.Tensor, + loaded_keys: set[str] = None, + ) -> Optional["LoHaAdapter"]: + if loaded_keys is None: + loaded_keys = set() + + hada_w1_a_name = "{}.hada_w1_a".format(x) + hada_w1_b_name = "{}.hada_w1_b".format(x) + hada_w2_a_name = "{}.hada_w2_a".format(x) + hada_w2_b_name = "{}.hada_w2_b".format(x) + hada_t1_name = "{}.hada_t1".format(x) + hada_t2_name = "{}.hada_t2".format(x) + if hada_w1_a_name in lora.keys(): + hada_t1 = None + hada_t2 = None + if hada_t1_name in lora.keys(): + hada_t1 = lora[hada_t1_name] + hada_t2 = lora[hada_t2_name] + loaded_keys.add(hada_t1_name) + loaded_keys.add(hada_t2_name) + + weights = (lora[hada_w1_a_name], lora[hada_w1_b_name], alpha, lora[hada_w2_a_name], lora[hada_w2_b_name], hada_t1, hada_t2, dora_scale) + loaded_keys.add(hada_w1_a_name) + loaded_keys.add(hada_w1_b_name) + loaded_keys.add(hada_w2_a_name) + loaded_keys.add(hada_w2_b_name) + return cls(loaded_keys, weights) + else: + return None + + def calculate_weight( + self, + weight, + key, + strength, + strength_model, + offset, + function, + intermediate_dtype=torch.float32, + original_weight=None, + ): + v = self.weights + w1a = v[0] + w1b = v[1] + if v[2] is not None: + alpha = v[2] / w1b.shape[0] + else: + alpha = 1.0 + + w2a = v[3] + w2b = v[4] + dora_scale = v[7] + if v[5] is not None: #cp decomposition + t1 = v[5] + t2 = v[6] + m1 = torch.einsum('i j k l, j r, i p -> p r k l', + comfy.model_management.cast_to_device(t1, weight.device, intermediate_dtype), + comfy.model_management.cast_to_device(w1b, weight.device, intermediate_dtype), + comfy.model_management.cast_to_device(w1a, weight.device, intermediate_dtype)) + + m2 = torch.einsum('i j k l, j r, i p -> p r k l', + comfy.model_management.cast_to_device(t2, weight.device, intermediate_dtype), + comfy.model_management.cast_to_device(w2b, weight.device, intermediate_dtype), + comfy.model_management.cast_to_device(w2a, weight.device, intermediate_dtype)) + else: + m1 = torch.mm(comfy.model_management.cast_to_device(w1a, weight.device, intermediate_dtype), + comfy.model_management.cast_to_device(w1b, weight.device, intermediate_dtype)) + m2 = torch.mm(comfy.model_management.cast_to_device(w2a, weight.device, intermediate_dtype), + comfy.model_management.cast_to_device(w2b, weight.device, intermediate_dtype)) + + try: + lora_diff = (m1 * m2).reshape(weight.shape) + if dora_scale is not None: + weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function) + else: + weight += function(((strength * alpha) * lora_diff).type(weight.dtype)) + except Exception as e: + logging.error("ERROR {} {} {}".format(self.name, key, e)) + return weight diff --git a/comfy/weight_adapter/lokr.py b/comfy/weight_adapter/lokr.py new file mode 100644 index 000000000..51233db2d --- /dev/null +++ b/comfy/weight_adapter/lokr.py @@ -0,0 +1,133 @@ +import logging +from typing import Optional + +import torch +import comfy.model_management +from .base import WeightAdapterBase, weight_decompose + + +class LoKrAdapter(WeightAdapterBase): + name = "lokr" + + def __init__(self, loaded_keys, weights): + self.loaded_keys = loaded_keys + self.weights = weights + + @classmethod + def load( + cls, + x: str, + lora: dict[str, torch.Tensor], + alpha: float, + dora_scale: torch.Tensor, + loaded_keys: set[str] = None, + ) -> Optional["LoKrAdapter"]: + if loaded_keys is None: + loaded_keys = set() + lokr_w1_name = "{}.lokr_w1".format(x) + lokr_w2_name = "{}.lokr_w2".format(x) + lokr_w1_a_name = "{}.lokr_w1_a".format(x) + lokr_w1_b_name = "{}.lokr_w1_b".format(x) + lokr_t2_name = "{}.lokr_t2".format(x) + lokr_w2_a_name = "{}.lokr_w2_a".format(x) + lokr_w2_b_name = "{}.lokr_w2_b".format(x) + + lokr_w1 = None + if lokr_w1_name in lora.keys(): + lokr_w1 = lora[lokr_w1_name] + loaded_keys.add(lokr_w1_name) + + lokr_w2 = None + if lokr_w2_name in lora.keys(): + lokr_w2 = lora[lokr_w2_name] + loaded_keys.add(lokr_w2_name) + + lokr_w1_a = None + if lokr_w1_a_name in lora.keys(): + lokr_w1_a = lora[lokr_w1_a_name] + loaded_keys.add(lokr_w1_a_name) + + lokr_w1_b = None + if lokr_w1_b_name in lora.keys(): + lokr_w1_b = lora[lokr_w1_b_name] + loaded_keys.add(lokr_w1_b_name) + + lokr_w2_a = None + if lokr_w2_a_name in lora.keys(): + lokr_w2_a = lora[lokr_w2_a_name] + loaded_keys.add(lokr_w2_a_name) + + lokr_w2_b = None + if lokr_w2_b_name in lora.keys(): + lokr_w2_b = lora[lokr_w2_b_name] + loaded_keys.add(lokr_w2_b_name) + + lokr_t2 = None + if lokr_t2_name in lora.keys(): + lokr_t2 = lora[lokr_t2_name] + loaded_keys.add(lokr_t2_name) + + if (lokr_w1 is not None) or (lokr_w2 is not None) or (lokr_w1_a is not None) or (lokr_w2_a is not None): + weights = (lokr_w1, lokr_w2, alpha, lokr_w1_a, lokr_w1_b, lokr_w2_a, lokr_w2_b, lokr_t2, dora_scale) + return cls(loaded_keys, weights) + else: + return None + + def calculate_weight( + self, + weight, + key, + strength, + strength_model, + offset, + function, + intermediate_dtype=torch.float32, + original_weight=None, + ): + v = self.weights + w1 = v[0] + w2 = v[1] + w1_a = v[3] + w1_b = v[4] + w2_a = v[5] + w2_b = v[6] + t2 = v[7] + dora_scale = v[8] + dim = None + + if w1 is None: + dim = w1_b.shape[0] + w1 = torch.mm(comfy.model_management.cast_to_device(w1_a, weight.device, intermediate_dtype), + comfy.model_management.cast_to_device(w1_b, weight.device, intermediate_dtype)) + else: + w1 = comfy.model_management.cast_to_device(w1, weight.device, intermediate_dtype) + + if w2 is None: + dim = w2_b.shape[0] + if t2 is None: + w2 = torch.mm(comfy.model_management.cast_to_device(w2_a, weight.device, intermediate_dtype), + comfy.model_management.cast_to_device(w2_b, weight.device, intermediate_dtype)) + else: + w2 = torch.einsum('i j k l, j r, i p -> p r k l', + comfy.model_management.cast_to_device(t2, weight.device, intermediate_dtype), + comfy.model_management.cast_to_device(w2_b, weight.device, intermediate_dtype), + comfy.model_management.cast_to_device(w2_a, weight.device, intermediate_dtype)) + else: + w2 = comfy.model_management.cast_to_device(w2, weight.device, intermediate_dtype) + + if len(w2.shape) == 4: + w1 = w1.unsqueeze(2).unsqueeze(2) + if v[2] is not None and dim is not None: + alpha = v[2] / dim + else: + alpha = 1.0 + + try: + lora_diff = torch.kron(w1, w2).reshape(weight.shape) + if dora_scale is not None: + weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function) + else: + weight += function(((strength * alpha) * lora_diff).type(weight.dtype)) + except Exception as e: + logging.error("ERROR {} {} {}".format(self.name, key, e)) + return weight diff --git a/comfy/weight_adapter/lora.py b/comfy/weight_adapter/lora.py new file mode 100644 index 000000000..b2e623924 --- /dev/null +++ b/comfy/weight_adapter/lora.py @@ -0,0 +1,142 @@ +import logging +from typing import Optional + +import torch +import comfy.model_management +from .base import WeightAdapterBase, weight_decompose, pad_tensor_to_shape + + +class LoRAAdapter(WeightAdapterBase): + name = "lora" + + def __init__(self, loaded_keys, weights): + self.loaded_keys = loaded_keys + self.weights = weights + + @classmethod + def load( + cls, + x: str, + lora: dict[str, torch.Tensor], + alpha: float, + dora_scale: torch.Tensor, + loaded_keys: set[str] = None, + ) -> Optional["LoRAAdapter"]: + if loaded_keys is None: + loaded_keys = set() + + reshape_name = "{}.reshape_weight".format(x) + regular_lora = "{}.lora_up.weight".format(x) + diffusers_lora = "{}_lora.up.weight".format(x) + diffusers2_lora = "{}.lora_B.weight".format(x) + diffusers3_lora = "{}.lora.up.weight".format(x) + mochi_lora = "{}.lora_B".format(x) + transformers_lora = "{}.lora_linear_layer.up.weight".format(x) + A_name = None + + if regular_lora in lora.keys(): + A_name = regular_lora + B_name = "{}.lora_down.weight".format(x) + mid_name = "{}.lora_mid.weight".format(x) + elif diffusers_lora in lora.keys(): + A_name = diffusers_lora + B_name = "{}_lora.down.weight".format(x) + mid_name = None + elif diffusers2_lora in lora.keys(): + A_name = diffusers2_lora + B_name = "{}.lora_A.weight".format(x) + mid_name = None + elif diffusers3_lora in lora.keys(): + A_name = diffusers3_lora + B_name = "{}.lora.down.weight".format(x) + mid_name = None + elif mochi_lora in lora.keys(): + A_name = mochi_lora + B_name = "{}.lora_A".format(x) + mid_name = None + elif transformers_lora in lora.keys(): + A_name = transformers_lora + B_name = "{}.lora_linear_layer.down.weight".format(x) + mid_name = None + + if A_name is not None: + mid = None + if mid_name is not None and mid_name in lora.keys(): + mid = lora[mid_name] + loaded_keys.add(mid_name) + reshape = None + if reshape_name in lora.keys(): + try: + reshape = lora[reshape_name].tolist() + loaded_keys.add(reshape_name) + except: + pass + weights = (lora[A_name], lora[B_name], alpha, mid, dora_scale, reshape) + loaded_keys.add(A_name) + loaded_keys.add(B_name) + return cls(loaded_keys, weights) + else: + return None + + def calculate_weight( + self, + weight, + key, + strength, + strength_model, + offset, + function, + intermediate_dtype=torch.float32, + original_weight=None, + ): + v = self.weights + mat1 = comfy.model_management.cast_to_device( + v[0], weight.device, intermediate_dtype + ) + mat2 = comfy.model_management.cast_to_device( + v[1], weight.device, intermediate_dtype + ) + dora_scale = v[4] + reshape = v[5] + + if reshape is not None: + weight = pad_tensor_to_shape(weight, reshape) + + if v[2] is not None: + alpha = v[2] / mat2.shape[0] + else: + alpha = 1.0 + + if v[3] is not None: + # locon mid weights, hopefully the math is fine because I didn't properly test it + mat3 = comfy.model_management.cast_to_device( + v[3], weight.device, intermediate_dtype + ) + final_shape = [mat2.shape[1], mat2.shape[0], mat3.shape[2], mat3.shape[3]] + mat2 = ( + torch.mm( + mat2.transpose(0, 1).flatten(start_dim=1), + mat3.transpose(0, 1).flatten(start_dim=1), + ) + .reshape(final_shape) + .transpose(0, 1) + ) + try: + lora_diff = torch.mm( + mat1.flatten(start_dim=1), mat2.flatten(start_dim=1) + ).reshape(weight.shape) + if dora_scale is not None: + weight = weight_decompose( + dora_scale, + weight, + lora_diff, + alpha, + strength, + intermediate_dtype, + function, + ) + else: + weight += function(((strength * alpha) * lora_diff).type(weight.dtype)) + except Exception as e: + logging.error("ERROR {} {} {}".format(self.name, key, e)) + return weight diff --git a/comfy/weight_adapter/oft.py b/comfy/weight_adapter/oft.py new file mode 100644 index 000000000..25009eca3 --- /dev/null +++ b/comfy/weight_adapter/oft.py @@ -0,0 +1,96 @@ +import logging +from typing import Optional + +import torch +import comfy.model_management +from .base import WeightAdapterBase, weight_decompose + + +class OFTAdapter(WeightAdapterBase): + name = "oft" + + def __init__(self, loaded_keys, weights): + self.loaded_keys = loaded_keys + self.weights = weights + + @classmethod + def load( + cls, + x: str, + lora: dict[str, torch.Tensor], + alpha: float, + dora_scale: torch.Tensor, + loaded_keys: set[str] = None, + ) -> Optional["OFTAdapter"]: + if loaded_keys is None: + loaded_keys = set() + blocks_name = "{}.oft_blocks".format(x) + rescale_name = "{}.rescale".format(x) + + blocks = None + if blocks_name in lora.keys(): + blocks = lora[blocks_name] + if blocks.ndim == 3: + loaded_keys.add(blocks_name) + else: + blocks = None + if blocks is None: + return None + + rescale = None + if rescale_name in lora.keys(): + rescale = lora[rescale_name] + loaded_keys.add(rescale_name) + + weights = (blocks, rescale, alpha, dora_scale) + return cls(loaded_keys, weights) + + def calculate_weight( + self, + weight, + key, + strength, + strength_model, + offset, + function, + intermediate_dtype=torch.float32, + original_weight=None, + ): + v = self.weights + blocks = v[0] + rescale = v[1] + alpha = v[2] + dora_scale = v[3] + + blocks = comfy.model_management.cast_to_device(blocks, weight.device, intermediate_dtype) + if rescale is not None: + rescale = comfy.model_management.cast_to_device(rescale, weight.device, intermediate_dtype) + + block_num, block_size, *_ = blocks.shape + + try: + # Get r + I = torch.eye(block_size, device=blocks.device, dtype=blocks.dtype) + # for Q = -Q^T + q = blocks - blocks.transpose(1, 2) + normed_q = q + if alpha > 0: # alpha in oft/boft is for constraint + q_norm = torch.norm(q) + 1e-8 + if q_norm > alpha: + normed_q = q * alpha / q_norm + # use float() to prevent unsupported type in .inverse() + r = (I + normed_q) @ (I - normed_q).float().inverse() + r = r.to(weight) + _, *shape = weight.shape + lora_diff = torch.einsum( + "k n m, k n ... -> k m ...", + (r * strength) - strength * I, + weight.view(block_num, block_size, *shape), + ).view(-1, *shape) + if dora_scale is not None: + weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function) + else: + weight += function((strength * lora_diff).type(weight.dtype)) + except Exception as e: + logging.error("ERROR {} {} {}".format(self.name, key, e)) + return weight diff --git a/comfy_api/input/__init__.py b/comfy_api/input/__init__.py new file mode 100644 index 000000000..66667946f --- /dev/null +++ b/comfy_api/input/__init__.py @@ -0,0 +1,8 @@ +from .basic_types import ImageInput, AudioInput +from .video_types import VideoInput + +__all__ = [ + "ImageInput", + "AudioInput", + "VideoInput", +] diff --git a/comfy_api/input/basic_types.py b/comfy_api/input/basic_types.py new file mode 100644 index 000000000..033fb7e27 --- /dev/null +++ b/comfy_api/input/basic_types.py @@ -0,0 +1,20 @@ +import torch +from typing import TypedDict + +ImageInput = torch.Tensor +""" +An image in format [B, H, W, C] where B is the batch size, C is the number of channels, +""" + +class AudioInput(TypedDict): + """ + TypedDict representing audio input. + """ + + waveform: torch.Tensor + """ + Tensor in the format [B, C, T] where B is the batch size, C is the number of channels, + """ + + sample_rate: int + diff --git a/comfy_api/input/video_types.py b/comfy_api/input/video_types.py new file mode 100644 index 000000000..dc22d34ff --- /dev/null +++ b/comfy_api/input/video_types.py @@ -0,0 +1,55 @@ +from __future__ import annotations +from abc import ABC, abstractmethod +from typing import Optional +from comfy_api.util import VideoContainer, VideoCodec, VideoComponents + +class VideoInput(ABC): + """ + Abstract base class for video input types. + """ + + @abstractmethod + def get_components(self) -> VideoComponents: + """ + Abstract method to get the video components (images, audio, and frame rate). + + Returns: + VideoComponents containing images, audio, and frame rate + """ + pass + + @abstractmethod + def save_to( + self, + path: str, + format: VideoContainer = VideoContainer.AUTO, + codec: VideoCodec = VideoCodec.AUTO, + metadata: Optional[dict] = None + ): + """ + Abstract method to save the video input to a file. + """ + pass + + # Provide a default implementation, but subclasses can provide optimized versions + # if possible. + def get_dimensions(self) -> tuple[int, int]: + """ + Returns the dimensions of the video input. + + Returns: + Tuple of (width, height) + """ + components = self.get_components() + return components.images.shape[2], components.images.shape[1] + + def get_duration(self) -> float: + """ + Returns the duration of the video in seconds. + + Returns: + Duration in seconds + """ + components = self.get_components() + frame_count = components.images.shape[0] + return float(frame_count / components.frame_rate) diff --git a/comfy_api/input_impl/__init__.py b/comfy_api/input_impl/__init__.py new file mode 100644 index 000000000..02901b8b9 --- /dev/null +++ b/comfy_api/input_impl/__init__.py @@ -0,0 +1,7 @@ +from .video_types import VideoFromFile, VideoFromComponents + +__all__ = [ + # Implementations + "VideoFromFile", + "VideoFromComponents", +] diff --git a/comfy_api/input_impl/video_types.py b/comfy_api/input_impl/video_types.py new file mode 100644 index 000000000..197f6558c --- /dev/null +++ b/comfy_api/input_impl/video_types.py @@ -0,0 +1,303 @@ +from __future__ import annotations +from av.container import InputContainer +from av.subtitles.stream import SubtitleStream +from fractions import Fraction +from typing import Optional +from comfy_api.input import AudioInput +import av +import io +import json +import numpy as np +import torch +from comfy_api.input import VideoInput +from comfy_api.util import VideoContainer, VideoCodec, VideoComponents + + +def container_to_output_format(container_format: str | None) -> str | None: + """ + A container's `format` may be a comma-separated list of formats. + E.g., iso container's `format` may be `mov,mp4,m4a,3gp,3g2,mj2`. + However, writing to a file/stream with `av.open` requires a single format, + or `None` to auto-detect. + """ + if not container_format: + return None # Auto-detect + + if "," not in container_format: + return container_format + + formats = container_format.split(",") + return formats[0] + + +def get_open_write_kwargs( + dest: str | io.BytesIO, container_format: str, to_format: str | None +) -> dict: + """Get kwargs for writing a `VideoFromFile` to a file/stream with `av.open`""" + open_kwargs = { + "mode": "w", + # If isobmff, preserve custom metadata tags (workflow, prompt, extra_pnginfo) + "options": {"movflags": "use_metadata_tags"}, + } + + is_write_to_buffer = isinstance(dest, io.BytesIO) + if is_write_to_buffer: + # Set output format explicitly, since it cannot be inferred from file extension + if to_format == VideoContainer.AUTO: + to_format = container_format.lower() + elif isinstance(to_format, str): + to_format = to_format.lower() + open_kwargs["format"] = container_to_output_format(to_format) + + return open_kwargs + + +class VideoFromFile(VideoInput): + """ + Class representing video input from a file. + """ + + def __init__(self, file: str | io.BytesIO): + """ + Initialize the VideoFromFile object based off of either a path on disk or a BytesIO object + containing the file contents. + """ + self.__file = file + + def get_dimensions(self) -> tuple[int, int]: + """ + Returns the dimensions of the video input. + + Returns: + Tuple of (width, height) + """ + if isinstance(self.__file, io.BytesIO): + self.__file.seek(0) # Reset the BytesIO object to the beginning + with av.open(self.__file, mode='r') as container: + for stream in container.streams: + if stream.type == 'video': + assert isinstance(stream, av.VideoStream) + return stream.width, stream.height + raise ValueError(f"No video stream found in file '{self.__file}'") + + def get_duration(self) -> float: + """ + Returns the duration of the video in seconds. + + Returns: + Duration in seconds + """ + if isinstance(self.__file, io.BytesIO): + self.__file.seek(0) + with av.open(self.__file, mode="r") as container: + if container.duration is not None: + return float(container.duration / av.time_base) + + # Fallback: calculate from frame count and frame rate + video_stream = next( + (s for s in container.streams if s.type == "video"), None + ) + if video_stream and video_stream.frames and video_stream.average_rate: + return float(video_stream.frames / video_stream.average_rate) + + # Last resort: decode frames to count them + if video_stream and video_stream.average_rate: + frame_count = 0 + container.seek(0) + for packet in container.demux(video_stream): + for _ in packet.decode(): + frame_count += 1 + if frame_count > 0: + return float(frame_count / video_stream.average_rate) + + raise ValueError(f"Could not determine duration for file '{self.__file}'") + + def get_components_internal(self, container: InputContainer) -> VideoComponents: + # Get video frames + frames = [] + for frame in container.decode(video=0): + img = frame.to_ndarray(format='rgb24') # shape: (H, W, 3) + img = torch.from_numpy(img) / 255.0 # shape: (H, W, 3) + frames.append(img) + + images = torch.stack(frames) if len(frames) > 0 else torch.zeros(0, 3, 0, 0) + + # Get frame rate + video_stream = next(s for s in container.streams if s.type == 'video') + frame_rate = Fraction(video_stream.average_rate) if video_stream and video_stream.average_rate else Fraction(1) + + # Get audio if available + audio = None + try: + container.seek(0) # Reset the container to the beginning + for stream in container.streams: + if stream.type != 'audio': + continue + assert isinstance(stream, av.AudioStream) + audio_frames = [] + for packet in container.demux(stream): + for frame in packet.decode(): + assert isinstance(frame, av.AudioFrame) + audio_frames.append(frame.to_ndarray()) # shape: (channels, samples) + if len(audio_frames) > 0: + audio_data = np.concatenate(audio_frames, axis=1) # shape: (channels, total_samples) + audio_tensor = torch.from_numpy(audio_data).unsqueeze(0) # shape: (1, channels, total_samples) + audio = AudioInput({ + "waveform": audio_tensor, + "sample_rate": int(stream.sample_rate) if stream.sample_rate else 1, + }) + except StopIteration: + pass # No audio stream + + metadata = container.metadata + return VideoComponents(images=images, audio=audio, frame_rate=frame_rate, metadata=metadata) + + def get_components(self) -> VideoComponents: + if isinstance(self.__file, io.BytesIO): + self.__file.seek(0) # Reset the BytesIO object to the beginning + with av.open(self.__file, mode='r') as container: + return self.get_components_internal(container) + raise ValueError(f"No video stream found in file '{self.__file}'") + + def save_to( + self, + path: str | io.BytesIO, + format: VideoContainer = VideoContainer.AUTO, + codec: VideoCodec = VideoCodec.AUTO, + metadata: Optional[dict] = None + ): + if isinstance(self.__file, io.BytesIO): + self.__file.seek(0) # Reset the BytesIO object to the beginning + with av.open(self.__file, mode='r') as container: + container_format = container.format.name + video_encoding = container.streams.video[0].codec.name if len(container.streams.video) > 0 else None + reuse_streams = True + if format != VideoContainer.AUTO and format not in container_format.split(","): + reuse_streams = False + if codec != VideoCodec.AUTO and codec != video_encoding and video_encoding is not None: + reuse_streams = False + + if not reuse_streams: + components = self.get_components_internal(container) + video = VideoFromComponents(components) + return video.save_to( + path, + format=format, + codec=codec, + metadata=metadata + ) + + streams = container.streams + + open_kwargs = get_open_write_kwargs(path, container_format, format) + with av.open(path, **open_kwargs) as output_container: + # Copy over the original metadata + for key, value in container.metadata.items(): + if metadata is None or key not in metadata: + output_container.metadata[key] = value + + # Add our new metadata + if metadata is not None: + for key, value in metadata.items(): + if isinstance(value, str): + output_container.metadata[key] = value + else: + output_container.metadata[key] = json.dumps(value) + + # Add streams to the new container + stream_map = {} + for stream in streams: + if isinstance(stream, (av.VideoStream, av.AudioStream, SubtitleStream)): + out_stream = output_container.add_stream_from_template(template=stream, opaque=True) + stream_map[stream] = out_stream + + # Write packets to the new container + for packet in container.demux(): + if packet.stream in stream_map and packet.dts is not None: + packet.stream = stream_map[packet.stream] + output_container.mux(packet) + +class VideoFromComponents(VideoInput): + """ + Class representing video input from tensors. + """ + + def __init__(self, components: VideoComponents): + self.__components = components + + def get_components(self) -> VideoComponents: + return VideoComponents( + images=self.__components.images, + audio=self.__components.audio, + frame_rate=self.__components.frame_rate + ) + + def save_to( + self, + path: str, + format: VideoContainer = VideoContainer.AUTO, + codec: VideoCodec = VideoCodec.AUTO, + metadata: Optional[dict] = None + ): + if format != VideoContainer.AUTO and format != VideoContainer.MP4: + 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: + # Add metadata before writing any streams + if metadata is not None: + for key, value in metadata.items(): + output.metadata[key] = json.dumps(value) + + frame_rate = Fraction(round(self.__components.frame_rate * 1000), 1000) + # Create a video stream + video_stream = output.add_stream('h264', rate=frame_rate) + video_stream.width = self.__components.images.shape[2] + video_stream.height = self.__components.images.shape[1] + video_stream.pix_fmt = 'yuv420p' + + # Create an audio stream + audio_sample_rate = 1 + audio_stream: Optional[av.AudioStream] = None + if self.__components.audio: + audio_sample_rate = int(self.__components.audio['sample_rate']) + audio_stream = output.add_stream('aac', rate=audio_sample_rate) + audio_stream.sample_rate = audio_sample_rate + audio_stream.format = 'fltp' + + # Encode video + for i, frame in enumerate(self.__components.images): + img = (frame * 255).clamp(0, 255).byte().cpu().numpy() # shape: (H, W, 3) + frame = av.VideoFrame.from_ndarray(img, format='rgb24') + frame = frame.reformat(format='yuv420p') # Convert to YUV420P as required by h264 + packet = video_stream.encode(frame) + output.mux(packet) + + # Flush video + packet = video_stream.encode(None) + output.mux(packet) + + if audio_stream and self.__components.audio: + # Encode audio + samples_per_frame = int(audio_sample_rate / frame_rate) + num_frames = self.__components.audio['waveform'].shape[2] // samples_per_frame + for i in range(num_frames): + start = i * samples_per_frame + end = start + samples_per_frame + # TODO(Feature) - Add support for stereo audio + chunk = ( + self.__components.audio["waveform"][0, 0, start:end] + .unsqueeze(0) + .contiguous() + .numpy() + ) + audio_frame = av.AudioFrame.from_ndarray(chunk, format='fltp', layout='mono') + audio_frame.sample_rate = audio_sample_rate + audio_frame.pts = i * samples_per_frame + for packet in audio_stream.encode(audio_frame): + output.mux(packet) + + # Flush audio + for packet in audio_stream.encode(None): + output.mux(packet) + diff --git a/comfy_api/torch_helpers/__init__.py b/comfy_api/torch_helpers/__init__.py new file mode 100644 index 000000000..be7ae7a61 --- /dev/null +++ b/comfy_api/torch_helpers/__init__.py @@ -0,0 +1,5 @@ +from .torch_compile import set_torch_compile_wrapper + +__all__ = [ + "set_torch_compile_wrapper", +] diff --git a/comfy_api/torch_helpers/torch_compile.py b/comfy_api/torch_helpers/torch_compile.py new file mode 100644 index 000000000..9223f58db --- /dev/null +++ b/comfy_api/torch_helpers/torch_compile.py @@ -0,0 +1,69 @@ +from __future__ import annotations +import torch + +import comfy.utils +from comfy.patcher_extension import WrappersMP +from typing import TYPE_CHECKING, Callable, Optional +if TYPE_CHECKING: + from comfy.model_patcher import ModelPatcher + from comfy.patcher_extension import WrapperExecutor + + +COMPILE_KEY = "torch.compile" +TORCH_COMPILE_KWARGS = "torch_compile_kwargs" + + +def apply_torch_compile_factory(compiled_module_dict: dict[str, Callable]) -> Callable: + ''' + Create a wrapper that will refer to the compiled_diffusion_model. + ''' + def apply_torch_compile_wrapper(executor: WrapperExecutor, *args, **kwargs): + try: + orig_modules = {} + for key, value in compiled_module_dict.items(): + orig_modules[key] = comfy.utils.get_attr(executor.class_obj, key) + comfy.utils.set_attr(executor.class_obj, key, value) + return executor(*args, **kwargs) + finally: + for key, value in orig_modules.items(): + comfy.utils.set_attr(executor.class_obj, key, value) + return apply_torch_compile_wrapper + + +def set_torch_compile_wrapper(model: ModelPatcher, backend: str, options: Optional[dict[str,str]]=None, + mode: Optional[str]=None, fullgraph=False, dynamic: Optional[bool]=None, + keys: list[str]=["diffusion_model"], *args, **kwargs): + ''' + Perform torch.compile that will be applied at sample time for either the whole model or specific params of the BaseModel instance. + + When keys is None, it will default to using ["diffusion_model"], compiling the whole diffusion_model. + When a list of keys is provided, it will perform torch.compile on only the selected modules. + ''' + # clear out any other torch.compile wrappers + model.remove_wrappers_with_key(WrappersMP.APPLY_MODEL, COMPILE_KEY) + # if no keys, default to 'diffusion_model' + if not keys: + keys = ["diffusion_model"] + # create kwargs dict that can be referenced later + compile_kwargs = { + "backend": backend, + "options": options, + "mode": mode, + "fullgraph": fullgraph, + "dynamic": dynamic, + } + # get a dict of compiled keys + compiled_modules = {} + for key in keys: + compiled_modules[key] = torch.compile( + model=model.get_model_object(key), + **compile_kwargs, + ) + # add torch.compile wrapper + wrapper_func = apply_torch_compile_factory( + compiled_module_dict=compiled_modules, + ) + # store wrapper to run on BaseModel's apply_model function + model.add_wrapper_with_key(WrappersMP.APPLY_MODEL, COMPILE_KEY, wrapper_func) + # keep compile kwargs for reference + model.model_options[TORCH_COMPILE_KWARGS] = compile_kwargs diff --git a/comfy_api/util/__init__.py b/comfy_api/util/__init__.py new file mode 100644 index 000000000..9019c46db --- /dev/null +++ b/comfy_api/util/__init__.py @@ -0,0 +1,8 @@ +from .video_types import VideoContainer, VideoCodec, VideoComponents + +__all__ = [ + # Utility Types + "VideoContainer", + "VideoCodec", + "VideoComponents", +] diff --git a/comfy_api/util/video_types.py b/comfy_api/util/video_types.py new file mode 100644 index 000000000..d09663db9 --- /dev/null +++ b/comfy_api/util/video_types.py @@ -0,0 +1,51 @@ +from __future__ import annotations +from dataclasses import dataclass +from enum import Enum +from fractions import Fraction +from typing import Optional +from comfy_api.input import ImageInput, AudioInput + +class VideoCodec(str, Enum): + AUTO = "auto" + H264 = "h264" + + @classmethod + def as_input(cls) -> list[str]: + """ + Returns a list of codec names that can be used as node input. + """ + return [member.value for member in cls] + +class VideoContainer(str, Enum): + AUTO = "auto" + MP4 = "mp4" + + @classmethod + def as_input(cls) -> list[str]: + """ + Returns a list of container names that can be used as node input. + """ + return [member.value for member in cls] + + @classmethod + def get_extension(cls, value) -> str: + """ + Returns the file extension for the container. + """ + if isinstance(value, str): + value = cls(value) + if value == VideoContainer.MP4 or value == VideoContainer.AUTO: + return "mp4" + return "" + +@dataclass +class VideoComponents: + """ + Dataclass representing the components of a video. + """ + + images: ImageInput + frame_rate: Fraction + audio: Optional[AudioInput] = None + metadata: Optional[dict] = None + diff --git a/comfy_api_nodes/README.md b/comfy_api_nodes/README.md new file mode 100644 index 000000000..64a389cc1 --- /dev/null +++ b/comfy_api_nodes/README.md @@ -0,0 +1,65 @@ +# ComfyUI API Nodes + +## Introduction + +Below are a collection of nodes that work by calling external APIs. More information available in our [docs](https://docs.comfy.org/tutorials/api-nodes/overview#api-nodes). + +## Development + +While developing, you should be testing against the Staging environment. To test against staging: + +**Install ComfyUI_frontend** + +Follow the instructions [here](https://github.com/Comfy-Org/ComfyUI_frontend) to start the frontend server. By default, it will connect to Staging authentication. + +> **Hint:** If you use --front-end-version argument for ComfyUI, it will use production authentication. + +```bash +python run main.py --comfy-api-base https://stagingapi.comfy.org +``` + +To authenticate to staging, please login and then ask one of Comfy Org team to whitelist you for access to staging. + +API stubs are generated through automatic codegen tools from OpenAPI definitions. Since the Comfy Org OpenAPI definition contains many things from the Comfy Registry as well, we use redocly/cli to filter out only the paths relevant for API nodes. + +### Redocly Instructions + +**Tip** +When developing locally, use the `redocly-dev.yaml` file to generate pydantic models. This lets you use stubs for APIs that are not marked `Released` yet. + +Before your API node PR merges, make sure to add the `Released` tag to the `openapi.yaml` file and test in staging. + +```bash +# Download the OpenAPI file from staging server. +curl -o openapi.yaml https://stagingapi.comfy.org/openapi + +# Filter out unneeded API definitions. +npm install -g @redocly/cli +redocly bundle openapi.yaml --output filtered-openapi.yaml --config comfy_api_nodes/redocly-dev.yaml --remove-unused-components + +# Generate the pydantic datamodels for validation. +datamodel-codegen --use-subclass-enum --field-constraints --strict-types bytes --input filtered-openapi.yaml --output comfy_api_nodes/apis/__init__.py --output-model-type pydantic_v2.BaseModel + +``` + + +# Merging to Master + +Before merging to comfyanonymous/ComfyUI master, follow these steps: + +1. Add the "Released" tag to the ComfyUI OpenAPI yaml file for each endpoint you are using in the nodes. +1. Make sure the ComfyUI API is deployed to prod with your changes. +1. Run the code generation again with `redocly.yaml` and the production OpenAPI yaml file. + +```bash +# Download the OpenAPI file from prod server. +curl -o openapi.yaml https://api.comfy.org/openapi + +# Filter out unneeded API definitions. +npm install -g @redocly/cli +redocly bundle openapi.yaml --output filtered-openapi.yaml --config comfy_api_nodes/redocly.yaml --remove-unused-components + +# Generate the pydantic datamodels for validation. +datamodel-codegen --use-subclass-enum --field-constraints --strict-types bytes --input filtered-openapi.yaml --output comfy_api_nodes/apis/__init__.py --output-model-type pydantic_v2.BaseModel + +``` diff --git a/comfy_api_nodes/__init__.py b/comfy_api_nodes/__init__.py new file mode 100644 index 000000000..e69de29bb diff --git a/comfy_api_nodes/apinode_utils.py b/comfy_api_nodes/apinode_utils.py new file mode 100644 index 000000000..788e2803f --- /dev/null +++ b/comfy_api_nodes/apinode_utils.py @@ -0,0 +1,678 @@ +from __future__ import annotations +import io +import logging +import mimetypes +from typing import Optional, Union +from comfy.utils import common_upscale +from comfy_api.input_impl import VideoFromFile +from comfy_api.util import VideoContainer, VideoCodec +from comfy_api.input.video_types import VideoInput +from comfy_api.input.basic_types import AudioInput +from comfy_api_nodes.apis.client import ( + ApiClient, + ApiEndpoint, + HttpMethod, + SynchronousOperation, + UploadRequest, + UploadResponse, +) +from server import PromptServer + + +import numpy as np +from PIL import Image +import requests +import torch +import math +import base64 +import uuid +from io import BytesIO +import av + + +def download_url_to_video_output(video_url: str, timeout: int = None) -> VideoFromFile: + """Downloads a video from a URL and returns a `VIDEO` output. + + Args: + video_url: The URL of the video to download. + + Returns: + A Comfy node `VIDEO` output. + """ + video_io = download_url_to_bytesio(video_url, timeout) + if video_io is None: + error_msg = f"Failed to download video from {video_url}" + logging.error(error_msg) + raise ValueError(error_msg) + return VideoFromFile(video_io) + + +def downscale_image_tensor(image, total_pixels=1536 * 1024) -> torch.Tensor: + """Downscale input image tensor to roughly the specified total pixels.""" + samples = image.movedim(-1, 1) + total = int(total_pixels) + scale_by = math.sqrt(total / (samples.shape[3] * samples.shape[2])) + if scale_by >= 1: + return image + width = round(samples.shape[3] * scale_by) + height = round(samples.shape[2] * scale_by) + + s = common_upscale(samples, width, height, "lanczos", "disabled") + s = s.movedim(1, -1) + return s + + +def validate_and_cast_response( + response, timeout: int = None, node_id: Union[str, None] = None +) -> torch.Tensor: + """Validates and casts a response to a torch.Tensor. + + Args: + response: The response to validate and cast. + timeout: Request timeout in seconds. Defaults to None (no timeout). + + Returns: + A torch.Tensor representing the image (1, H, W, C). + + Raises: + ValueError: If the response is not valid. + """ + # validate raw JSON response + data = response.data + if not data or len(data) == 0: + raise ValueError("No images returned from API endpoint") + + # Initialize list to store image tensors + image_tensors: list[torch.Tensor] = [] + + # Process each image in the data array + for image_data in data: + image_url = image_data.url + b64_data = image_data.b64_json + + if not image_url and not b64_data: + raise ValueError("No image was generated in the response") + + if b64_data: + img_data = base64.b64decode(b64_data) + img = Image.open(io.BytesIO(img_data)) + + elif image_url: + if node_id: + PromptServer.instance.send_progress_text( + f"Result URL: {image_url}", node_id + ) + img_response = requests.get(image_url, timeout=timeout) + if img_response.status_code != 200: + raise ValueError("Failed to download the image") + img = Image.open(io.BytesIO(img_response.content)) + + img = img.convert("RGBA") + + # Convert to numpy array, normalize to float32 between 0 and 1 + img_array = np.array(img).astype(np.float32) / 255.0 + img_tensor = torch.from_numpy(img_array) + + # Add to list of tensors + image_tensors.append(img_tensor) + + return torch.stack(image_tensors, dim=0) + + +def validate_aspect_ratio( + aspect_ratio: str, + minimum_ratio: float, + maximum_ratio: float, + minimum_ratio_str: str, + maximum_ratio_str: str, +) -> float: + """Validates and casts an aspect ratio string to a float. + + Args: + aspect_ratio: The aspect ratio string to validate. + minimum_ratio: The minimum aspect ratio. + maximum_ratio: The maximum aspect ratio. + minimum_ratio_str: The minimum aspect ratio string. + maximum_ratio_str: The maximum aspect ratio string. + + Returns: + The validated and cast aspect ratio. + + Raises: + Exception: If the aspect ratio is not valid. + """ + # get ratio values + numbers = aspect_ratio.split(":") + if len(numbers) != 2: + raise TypeError( + f"Aspect ratio must be in the format X:Y, such as 16:9, but was {aspect_ratio}." + ) + try: + numerator = int(numbers[0]) + denominator = int(numbers[1]) + except ValueError as exc: + raise TypeError( + f"Aspect ratio must contain numbers separated by ':', such as 16:9, but was {aspect_ratio}." + ) from exc + calculated_ratio = numerator / denominator + # if not close to minimum and maximum, check bounds + if not math.isclose(calculated_ratio, minimum_ratio) or not math.isclose( + calculated_ratio, maximum_ratio + ): + if calculated_ratio < minimum_ratio: + raise TypeError( + f"Aspect ratio cannot reduce to any less than {minimum_ratio_str} ({minimum_ratio}), but was {aspect_ratio} ({calculated_ratio})." + ) + elif calculated_ratio > maximum_ratio: + raise TypeError( + f"Aspect ratio cannot reduce to any greater than {maximum_ratio_str} ({maximum_ratio}), but was {aspect_ratio} ({calculated_ratio})." + ) + return aspect_ratio + + +def mimetype_to_extension(mime_type: str) -> str: + """Converts a MIME type to a file extension.""" + return mime_type.split("/")[-1].lower() + + +def download_url_to_bytesio(url: str, timeout: int = None) -> BytesIO: + """Downloads content from a URL using requests and returns it as BytesIO. + + Args: + url: The URL to download. + timeout: Request timeout in seconds. Defaults to None (no timeout). + + Returns: + BytesIO object containing the downloaded content. + """ + response = requests.get(url, stream=True, timeout=timeout) + response.raise_for_status() # Raises HTTPError for bad responses (4XX or 5XX) + return BytesIO(response.content) + + +def bytesio_to_image_tensor(image_bytesio: BytesIO, mode: str = "RGBA") -> torch.Tensor: + """Converts image data from BytesIO to a torch.Tensor. + + Args: + image_bytesio: BytesIO object containing the image data. + mode: The PIL mode to convert the image to (e.g., "RGB", "RGBA"). + + Returns: + A torch.Tensor representing the image (1, H, W, C). + + Raises: + PIL.UnidentifiedImageError: If the image data cannot be identified. + ValueError: If the specified mode is invalid. + """ + image = Image.open(image_bytesio) + image = image.convert(mode) + image_array = np.array(image).astype(np.float32) / 255.0 + return torch.from_numpy(image_array).unsqueeze(0) + + +def download_url_to_image_tensor(url: str, timeout: int = None) -> torch.Tensor: + """Downloads an image from a URL and returns a [B, H, W, C] tensor.""" + image_bytesio = download_url_to_bytesio(url, timeout) + return bytesio_to_image_tensor(image_bytesio) + + +def process_image_response(response: requests.Response) -> torch.Tensor: + """Uses content from a Response object and converts it to a torch.Tensor""" + return bytesio_to_image_tensor(BytesIO(response.content)) + + +def _tensor_to_pil(image: torch.Tensor, total_pixels: int = 2048 * 2048) -> Image.Image: + """Converts a single torch.Tensor image [H, W, C] to a PIL Image, optionally downscaling.""" + if len(image.shape) > 3: + image = image[0] + # TODO: remove alpha if not allowed and present + input_tensor = image.cpu() + input_tensor = downscale_image_tensor( + input_tensor.unsqueeze(0), total_pixels=total_pixels + ).squeeze() + image_np = (input_tensor.numpy() * 255).astype(np.uint8) + img = Image.fromarray(image_np) + return img + + +def _pil_to_bytesio(img: Image.Image, mime_type: str = "image/png") -> BytesIO: + """Converts a PIL Image to a BytesIO object.""" + if not mime_type: + mime_type = "image/png" + + img_byte_arr = io.BytesIO() + # Derive PIL format from MIME type (e.g., 'image/png' -> 'PNG') + pil_format = mime_type.split("/")[-1].upper() + if pil_format == "JPG": + pil_format = "JPEG" + img.save(img_byte_arr, format=pil_format) + img_byte_arr.seek(0) + return img_byte_arr + + +def tensor_to_bytesio( + image: torch.Tensor, + name: Optional[str] = None, + total_pixels: int = 2048 * 2048, + mime_type: str = "image/png", +) -> BytesIO: + """Converts a torch.Tensor image to a named BytesIO object. + + Args: + image: Input torch.Tensor image. + name: Optional filename for the BytesIO object. + total_pixels: Maximum total pixels for potential downscaling. + mime_type: Target image MIME type (e.g., 'image/png', 'image/jpeg', 'image/webp', 'video/mp4'). + + Returns: + Named BytesIO object containing the image data. + """ + if not mime_type: + mime_type = "image/png" + + pil_image = _tensor_to_pil(image, total_pixels=total_pixels) + img_binary = _pil_to_bytesio(pil_image, mime_type=mime_type) + img_binary.name = ( + f"{name if name else uuid.uuid4()}.{mimetype_to_extension(mime_type)}" + ) + return img_binary + + +def tensor_to_base64_string( + image_tensor: torch.Tensor, + total_pixels: int = 2048 * 2048, + mime_type: str = "image/png", +) -> str: + """Convert [B, H, W, C] or [H, W, C] tensor to a base64 string. + + Args: + image_tensor: Input torch.Tensor image. + total_pixels: Maximum total pixels for potential downscaling. + mime_type: Target image MIME type (e.g., 'image/png', 'image/jpeg', 'image/webp', 'video/mp4'). + + Returns: + Base64 encoded string of the image. + """ + pil_image = _tensor_to_pil(image_tensor, total_pixels=total_pixels) + img_byte_arr = _pil_to_bytesio(pil_image, mime_type=mime_type) + img_bytes = img_byte_arr.getvalue() + # Encode bytes to base64 string + base64_encoded_string = base64.b64encode(img_bytes).decode("utf-8") + return base64_encoded_string + + +def tensor_to_data_uri( + image_tensor: torch.Tensor, + total_pixels: int = 2048 * 2048, + mime_type: str = "image/png", +) -> str: + """Converts a tensor image to a Data URI string. + + Args: + image_tensor: Input torch.Tensor image. + total_pixels: Maximum total pixels for potential downscaling. + mime_type: Target image MIME type (e.g., 'image/png', 'image/jpeg', 'image/webp'). + + Returns: + Data URI string (e.g., 'data:image/png;base64,...'). + """ + base64_string = tensor_to_base64_string(image_tensor, total_pixels, mime_type) + return f"data:{mime_type};base64,{base64_string}" + + +def text_filepath_to_base64_string(filepath: str) -> str: + """Converts a text file to a base64 string.""" + with open(filepath, "rb") as f: + file_content = f.read() + return base64.b64encode(file_content).decode("utf-8") + + +def text_filepath_to_data_uri(filepath: str) -> str: + """Converts a text file to a data URI.""" + base64_string = text_filepath_to_base64_string(filepath) + mime_type, _ = mimetypes.guess_type(filepath) + if mime_type is None: + mime_type = "application/octet-stream" + return f"data:{mime_type};base64,{base64_string}" + + +def upload_file_to_comfyapi( + file_bytes_io: BytesIO, + filename: str, + upload_mime_type: str, + auth_kwargs: Optional[dict[str, str]] = None, +) -> str: + """ + Uploads a single file to ComfyUI API and returns its download URL. + + Args: + file_bytes_io: BytesIO object containing the file data. + filename: The filename of the file. + upload_mime_type: MIME type of the file. + auth_kwargs: Optional authentication token(s). + + Returns: + The download URL for the uploaded file. + """ + request_object = UploadRequest(file_name=filename, content_type=upload_mime_type) + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/customers/storage", + method=HttpMethod.POST, + request_model=UploadRequest, + response_model=UploadResponse, + ), + request=request_object, + auth_kwargs=auth_kwargs, + ) + + response: UploadResponse = operation.execute() + upload_response = ApiClient.upload_file( + response.upload_url, file_bytes_io, content_type=upload_mime_type + ) + upload_response.raise_for_status() + + return response.download_url + + +def video_to_base64_string( + video: VideoInput, + container_format: VideoContainer = None, + codec: VideoCodec = None +) -> str: + """ + Converts a video input to a base64 string. + + Args: + video: The video input to convert + container_format: Optional container format to use (defaults to video.container if available) + codec: Optional codec to use (defaults to video.codec if available) + """ + video_bytes_io = 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_bytes_io.seek(0) + return base64.b64encode(video_bytes_io.getvalue()).decode("utf-8") + + +def upload_video_to_comfyapi( + video: VideoInput, + auth_kwargs: Optional[dict[str, str]] = None, + container: VideoContainer = VideoContainer.MP4, + codec: VideoCodec = VideoCodec.H264, + max_duration: Optional[int] = None, +) -> str: + """ + Uploads a single video to ComfyUI API and returns its download URL. + Uses the specified container and codec for saving the video before upload. + + Args: + video: VideoInput object (Comfy VIDEO type). + auth_kwargs: Optional authentication token(s). + container: The video container format to use (default: MP4). + codec: The video codec to use (default: H264). + max_duration: Optional maximum duration of the video in seconds. If the video is longer than this, an error will be raised. + + Returns: + The download URL for the uploaded video file. + """ + if max_duration is not None: + try: + actual_duration = video.duration_seconds + if actual_duration is not None and actual_duration > max_duration: + raise ValueError( + f"Video duration ({actual_duration:.2f}s) exceeds the maximum allowed ({max_duration}s)." + ) + except Exception as e: + logging.error(f"Error getting video duration: {e}") + 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()}" + + # Convert VideoInput to BytesIO using specified container/codec + video_bytes_io = io.BytesIO() + video.save_to(video_bytes_io, format=container, codec=codec) + video_bytes_io.seek(0) + + return upload_file_to_comfyapi( + video_bytes_io, filename, upload_mime_type, auth_kwargs + ) + + +def audio_tensor_to_contiguous_ndarray(waveform: torch.Tensor) -> np.ndarray: + """ + Prepares audio waveform for av library by converting to a contiguous numpy array. + + Args: + waveform: a tensor of shape (1, channels, samples) derived from a Comfy `AUDIO` type. + + Returns: + Contiguous numpy array of the audio waveform. If the audio was batched, + the first item is taken. + """ + if waveform.ndim != 3 or waveform.shape[0] != 1: + raise ValueError("Expected waveform tensor shape (1, channels, samples)") + + # If batch is > 1, take first item + if waveform.shape[0] > 1: + waveform = waveform[0] + + # Prepare for av: remove batch dim, move to CPU, make contiguous, convert to numpy array + audio_data_np = waveform.squeeze(0).cpu().contiguous().numpy() + if audio_data_np.dtype != np.float32: + audio_data_np = audio_data_np.astype(np.float32) + + return audio_data_np + + +def audio_ndarray_to_bytesio( + audio_data_np: np.ndarray, + sample_rate: int, + container_format: str = "mp4", + codec_name: str = "aac", +) -> BytesIO: + """ + Encodes a numpy array of audio data into a BytesIO object. + """ + audio_bytes_io = io.BytesIO() + with av.open(audio_bytes_io, mode="w", format=container_format) as output_container: + audio_stream = output_container.add_stream(codec_name, rate=sample_rate) + frame = av.AudioFrame.from_ndarray( + audio_data_np, + format="fltp", + layout="stereo" if audio_data_np.shape[0] > 1 else "mono", + ) + frame.sample_rate = sample_rate + frame.pts = 0 + + for packet in audio_stream.encode(frame): + output_container.mux(packet) + + # Flush stream + for packet in audio_stream.encode(None): + output_container.mux(packet) + + audio_bytes_io.seek(0) + return audio_bytes_io + + +def upload_audio_to_comfyapi( + audio: AudioInput, + auth_kwargs: Optional[dict[str, str]] = None, + container_format: str = "mp4", + codec_name: str = "aac", + mime_type: str = "audio/mp4", + filename: str = "uploaded_audio.mp4", +) -> str: + """ + Uploads a single audio input to ComfyUI API and returns its download URL. + Encodes the raw waveform into the specified format before uploading. + + Args: + audio: a Comfy `AUDIO` type (contains waveform tensor and sample_rate) + auth_kwargs: Optional authentication token(s). + + Returns: + The download URL for the uploaded audio file. + """ + sample_rate: int = audio["sample_rate"] + waveform: torch.Tensor = audio["waveform"] + audio_data_np = audio_tensor_to_contiguous_ndarray(waveform) + audio_bytes_io = audio_ndarray_to_bytesio( + audio_data_np, sample_rate, container_format, codec_name + ) + + return upload_file_to_comfyapi(audio_bytes_io, filename, mime_type, auth_kwargs) + + +def audio_to_base64_string( + audio: AudioInput, container_format: str = "mp4", codec_name: str = "aac" +) -> str: + """Converts an audio input to a base64 string.""" + sample_rate: int = audio["sample_rate"] + waveform: torch.Tensor = audio["waveform"] + audio_data_np = audio_tensor_to_contiguous_ndarray(waveform) + audio_bytes_io = audio_ndarray_to_bytesio( + audio_data_np, sample_rate, container_format, codec_name + ) + audio_bytes = audio_bytes_io.getvalue() + return base64.b64encode(audio_bytes).decode("utf-8") + + +def upload_images_to_comfyapi( + image: torch.Tensor, + max_images=8, + auth_kwargs: Optional[dict[str, str]] = None, + mime_type: Optional[str] = None, +) -> list[str]: + """ + Uploads images to ComfyUI API and returns download URLs. + To upload multiple images, stack them in the batch dimension first. + + Args: + image: Input torch.Tensor image. + max_images: Maximum number of images to upload. + auth_kwargs: Optional authentication token(s). + mime_type: Optional MIME type for the image. + """ + # if batch, try to upload each file if max_images is greater than 0 + idx_image = 0 + download_urls: list[str] = [] + is_batch = len(image.shape) > 3 + batch_length = 1 + if is_batch: + batch_length = image.shape[0] + while True: + curr_image = image + if len(image.shape) > 3: + curr_image = image[idx_image] + # get BytesIO version of image + img_binary = tensor_to_bytesio(curr_image, mime_type=mime_type) + # first, request upload/download urls from comfy API + if not mime_type: + request_object = UploadRequest(file_name=img_binary.name) + else: + request_object = UploadRequest( + file_name=img_binary.name, content_type=mime_type + ) + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/customers/storage", + method=HttpMethod.POST, + request_model=UploadRequest, + response_model=UploadResponse, + ), + request=request_object, + auth_kwargs=auth_kwargs, + ) + response = operation.execute() + + upload_response = ApiClient.upload_file( + response.upload_url, img_binary, content_type=mime_type + ) + # verify success + try: + upload_response.raise_for_status() + except requests.exceptions.HTTPError as e: + raise ValueError(f"Could not upload one or more images: {e}") from e + # add download_url to list + download_urls.append(response.download_url) + + idx_image += 1 + # stop uploading additional files if done + if is_batch and max_images > 0: + if idx_image >= max_images: + break + if idx_image >= batch_length: + break + return download_urls + + +def resize_mask_to_image( + mask: torch.Tensor, + image: torch.Tensor, + upscale_method="nearest-exact", + crop="disabled", + allow_gradient=True, + add_channel_dim=False, +): + """ + Resize mask to be the same dimensions as an image, while maintaining proper format for API calls. + """ + _, H, W, _ = image.shape + mask = mask.unsqueeze(-1) + mask = mask.movedim(-1, 1) + mask = common_upscale( + mask, width=W, height=H, upscale_method=upscale_method, crop=crop + ) + mask = mask.movedim(1, -1) + if not add_channel_dim: + mask = mask.squeeze(-1) + if not allow_gradient: + mask = (mask > 0.5).float() + return mask + + +def validate_string( + string: str, + strip_whitespace=True, + field_name="prompt", + min_length=None, + max_length=None, +): + if string is None: + raise Exception(f"Field '{field_name}' cannot be empty.") + if strip_whitespace: + string = string.strip() + if min_length and len(string) < min_length: + raise Exception( + f"Field '{field_name}' cannot be shorter than {min_length} characters; was {len(string)} characters long." + ) + if max_length and len(string) > max_length: + raise Exception( + f" Field '{field_name} cannot be longer than {max_length} characters; was {len(string)} characters long." + ) + + +def image_tensor_pair_to_batch( + image1: torch.Tensor, image2: torch.Tensor +) -> torch.Tensor: + """ + Converts a pair of image tensors to a batch tensor. + If the images are not the same size, the smaller image is resized to + match the larger image. + """ + if image1.shape[1:] != image2.shape[1:]: + image2 = common_upscale( + image2.movedim(-1, 1), + image1.shape[2], + image1.shape[1], + "bilinear", + "center", + ).movedim(1, -1) + return torch.cat((image1, image2), dim=0) diff --git a/comfy_api_nodes/apis/PixverseController.py b/comfy_api_nodes/apis/PixverseController.py new file mode 100644 index 000000000..310c0f546 --- /dev/null +++ b/comfy_api_nodes/apis/PixverseController.py @@ -0,0 +1,17 @@ +# generated by datamodel-codegen: +# filename: filtered-openapi.yaml +# timestamp: 2025-04-29T23:44:54+00:00 + +from __future__ import annotations + +from typing import Optional + +from pydantic import BaseModel + +from . import PixverseDto + + +class ResponseData(BaseModel): + ErrCode: Optional[int] = None + ErrMsg: Optional[str] = None + Resp: Optional[PixverseDto.V2OpenAPII2VResp] = None diff --git a/comfy_api_nodes/apis/PixverseDto.py b/comfy_api_nodes/apis/PixverseDto.py new file mode 100644 index 000000000..323c38e96 --- /dev/null +++ b/comfy_api_nodes/apis/PixverseDto.py @@ -0,0 +1,57 @@ +# generated by datamodel-codegen: +# filename: filtered-openapi.yaml +# timestamp: 2025-04-29T23:44:54+00:00 + +from __future__ import annotations + +from typing import Optional + +from pydantic import BaseModel, Field + + +class V2OpenAPII2VResp(BaseModel): + video_id: Optional[int] = Field(None, description='Video_id') + + +class V2OpenAPIT2VReq(BaseModel): + aspect_ratio: str = Field( + ..., description='Aspect ratio (16:9, 4:3, 1:1, 3:4, 9:16)', examples=['16:9'] + ) + duration: int = Field( + ..., + description='Video duration (5, 8 seconds, --model=v3.5 only allows 5,8; --quality=1080p does not support 8s)', + examples=[5], + ) + model: str = Field( + ..., description='Model version (only supports v3.5)', examples=['v3.5'] + ) + motion_mode: Optional[str] = Field( + 'normal', + description='Motion mode (normal, fast, --fast only available when duration=5; --quality=1080p does not support fast)', + examples=['normal'], + ) + negative_prompt: Optional[str] = Field( + None, description='Negative prompt\n', max_length=2048 + ) + prompt: str = Field(..., description='Prompt', max_length=2048) + quality: str = Field( + ..., + description='Video quality ("360p"(Turbo model), "540p", "720p", "1080p")', + examples=['540p'], + ) + seed: Optional[int] = Field(None, description='Random seed, range: 0 - 2147483647') + style: Optional[str] = Field( + None, + description='Style (effective when model=v3.5, "anime", "3d_animation", "clay", "comic", "cyberpunk") Do not include style parameter unless needed', + examples=['anime'], + ) + template_id: Optional[int] = Field( + None, + description='Template ID (template_id must be activated before use)', + examples=[302325299692608], + ) + water_mark: Optional[bool] = Field( + False, + description='Watermark (true: add watermark, false: no watermark)', + examples=[False], + ) diff --git a/comfy_api_nodes/apis/__init__.py b/comfy_api_nodes/apis/__init__.py new file mode 100644 index 000000000..e38d38cc9 --- /dev/null +++ b/comfy_api_nodes/apis/__init__.py @@ -0,0 +1,3453 @@ +# generated by datamodel-codegen: +# filename: filtered-openapi.yaml +# timestamp: 2025-05-19T21:38:55+00:00 + +from __future__ import annotations + +from datetime import date, datetime +from enum import Enum +from typing import Any, Dict, List, Literal, Optional, Union +from uuid import UUID + +from pydantic import AnyUrl, BaseModel, ConfigDict, Field, RootModel, StrictBytes + + +class APIKey(BaseModel): + created_at: Optional[datetime] = None + description: Optional[str] = None + id: Optional[str] = None + key_prefix: Optional[str] = None + name: Optional[str] = None + + +class APIKeyWithPlaintext(APIKey): + plaintext_key: Optional[str] = Field( + None, description='The full API key (only returned at creation)' + ) + + +class AuditLog(BaseModel): + createdAt: Optional[datetime] = Field( + None, description='The date and time the event was created' + ) + event_id: Optional[str] = Field(None, description='the id of the event') + event_type: Optional[str] = Field(None, description='the type of the event') + params: Optional[Dict[str, Any]] = Field( + None, description='data related to the event' + ) + + +class OutputFormat(str, Enum): + jpeg = 'jpeg' + png = 'png' + + +class BFLFluxPro11GenerateRequest(BaseModel): + height: int = Field(..., description='Height of the generated image') + image_prompt: Optional[str] = Field(None, description='Optional image prompt') + output_format: Optional[OutputFormat] = Field( + None, description='Output image format' + ) + prompt: str = Field(..., description='The main text prompt for image generation') + prompt_upsampling: Optional[bool] = Field( + None, description='Whether to use prompt upsampling' + ) + safety_tolerance: Optional[int] = Field(None, description='Safety tolerance level') + seed: Optional[int] = Field(None, description='Random seed for reproducibility') + webhook_secret: Optional[str] = Field( + None, description='Optional webhook secret for async processing' + ) + webhook_url: Optional[str] = Field( + None, description='Optional webhook URL for async processing' + ) + width: int = Field(..., description='Width of the generated image') + + +class BFLFluxPro11GenerateResponse(BaseModel): + id: str = Field(..., description='Job ID for tracking') + polling_url: str = Field(..., description='URL to poll for results') + + +class BFLFluxProGenerateRequest(BaseModel): + guidance_scale: Optional[float] = Field( + None, description='The guidance scale for generation.', ge=1.0, le=20.0 + ) + height: int = Field( + ..., description='The height of the image to generate.', ge=64, le=2048 + ) + negative_prompt: Optional[str] = Field( + None, description='The negative prompt for image generation.' + ) + num_images: Optional[int] = Field( + None, description='The number of images to generate.', ge=1, le=4 + ) + num_inference_steps: Optional[int] = Field( + None, description='The number of inference steps.', ge=1, le=100 + ) + prompt: str = Field(..., description='The text prompt for image generation.') + seed: Optional[int] = Field(None, description='The seed value for reproducibility.') + width: int = Field( + ..., description='The width of the image to generate.', ge=64, le=2048 + ) + + +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.') + + +class Status(str, Enum): + in_progress = 'in_progress' + completed = 'completed' + incomplete = 'incomplete' + + +class Type(str, Enum): + computer_call = 'computer_call' + + +class ComputerToolCall(BaseModel): + action: Dict[str, Any] + call_id: str = Field( + ..., + description='An identifier used when responding to the tool call with output.\n', + ) + id: str = Field(..., description='The unique ID of the computer call.') + status: Status = Field( + ..., + description='The status of the item. One of `in_progress`, `completed`, or\n`incomplete`. Populated when items are returned via API.\n', + ) + type: Type = Field( + ..., description='The type of the computer call. Always `computer_call`.' + ) + + +class Environment(str, Enum): + windows = 'windows' + mac = 'mac' + linux = 'linux' + ubuntu = 'ubuntu' + browser = 'browser' + + +class Type1(str, Enum): + computer_use_preview = 'computer_use_preview' + + +class ComputerUsePreviewTool(BaseModel): + display_height: int = Field(..., description='The height of the computer display.') + display_width: int = Field(..., description='The width of the computer display.') + environment: Environment = Field( + ..., description='The type of computer environment to control.' + ) + type: Literal['ComputerUsePreviewTool'] = Field( + ..., + description='The type of the computer use tool. Always `computer_use_preview`.', + ) + + +class CreateAPIKeyRequest(BaseModel): + description: Optional[str] = None + name: str + + +class Customer(BaseModel): + createdAt: Optional[datetime] = Field( + None, description='The date and time the user was created' + ) + email: Optional[str] = Field(None, description='The email address for this user') + id: str = Field(..., description='The firebase UID of the user') + is_admin: Optional[bool] = Field(None, description='Whether the user is an admin') + metronome_id: Optional[str] = Field(None, description='The Metronome customer ID') + name: Optional[str] = Field(None, description='The name for this user') + stripe_id: Optional[str] = Field(None, description='The Stripe customer ID') + updatedAt: Optional[datetime] = Field( + None, description='The date and time the user was last updated' + ) + + +class CustomerStorageResourceResponse(BaseModel): + download_url: Optional[str] = Field( + None, + description='The signed URL to use for downloading the file from the specified path', + ) + existing_file: Optional[bool] = Field( + None, description='Whether an existing file with the same hash was found' + ) + expires_at: Optional[datetime] = Field( + None, description='When the signed URL will expire' + ) + upload_url: Optional[str] = Field( + None, + description='The signed URL to use for uploading the file to the specified path', + ) + + +class Role(str, Enum): + user = 'user' + assistant = 'assistant' + system = 'system' + developer = 'developer' + + +class Type2(str, Enum): + message = 'message' + + +class ErrorResponse(BaseModel): + error: str + message: str + + +class Type3(str, Enum): + file_search = 'file_search' + + +class FileSearchTool(BaseModel): + type: Literal['FileSearchTool'] = Field(..., description='The type of tool') + vector_store_ids: List[str] = Field( + ..., description='IDs of vector stores to search in' + ) + + +class Result(BaseModel): + file_id: Optional[str] = Field(None, description='The unique ID of the file.\n') + filename: Optional[str] = Field(None, description='The name of the file.\n') + score: Optional[float] = Field( + None, description='The relevance score of the file - a value between 0 and 1.\n' + ) + text: Optional[str] = Field( + None, description='The text that was retrieved from the file.\n' + ) + + +class Status1(str, Enum): + in_progress = 'in_progress' + searching = 'searching' + completed = 'completed' + incomplete = 'incomplete' + failed = 'failed' + + +class Type4(str, Enum): + file_search_call = 'file_search_call' + + +class FileSearchToolCall(BaseModel): + id: str = Field(..., description='The unique ID of the file search tool call.\n') + queries: List[str] = Field( + ..., description='The queries used to search for files.\n' + ) + results: Optional[List[Result]] = Field( + None, description='The results of the file search tool call.\n' + ) + status: Status1 = Field( + ..., + description='The status of the file search tool call. One of `in_progress`, \n`searching`, `incomplete` or `failed`,\n', + ) + type: Type4 = Field( + ..., + description='The type of the file search tool call. Always `file_search_call`.\n', + ) + + +class Type5(str, Enum): + function = 'function' + + +class FunctionTool(BaseModel): + description: Optional[str] = Field( + None, description='Description of what the function does' + ) + name: str = Field(..., description='Name of the function') + parameters: Dict[str, Any] = Field( + ..., description='JSON Schema object describing the function parameters' + ) + type: Literal['FunctionTool'] = Field(..., description='The type of tool') + + +class Status2(str, Enum): + in_progress = 'in_progress' + completed = 'completed' + incomplete = 'incomplete' + + +class Type6(str, Enum): + function_call = 'function_call' + + +class FunctionToolCall(BaseModel): + arguments: str = Field( + ..., description='A JSON string of the arguments to pass to the function.\n' + ) + call_id: str = Field( + ..., + description='The unique ID of the function tool call generated by the model.\n', + ) + id: Optional[str] = Field( + None, description='The unique ID of the function tool call.\n' + ) + name: str = Field(..., description='The name of the function to run.\n') + status: Optional[Status2] = Field( + None, + description='The status of the item. One of `in_progress`, `completed`, or\n`incomplete`. Populated when items are returned via API.\n', + ) + type: Type6 = Field( + ..., description='The type of the function tool call. Always `function_call`.\n' + ) + + +class GeminiCitation(BaseModel): + authors: Optional[List[str]] = None + endIndex: Optional[int] = None + license: Optional[str] = None + publicationDate: Optional[date] = None + startIndex: Optional[int] = None + title: Optional[str] = None + uri: Optional[str] = None + + +class GeminiCitationMetadata(BaseModel): + citations: Optional[List[GeminiCitation]] = None + + +class Role1(str, Enum): + user = 'user' + model = 'model' + + +class GeminiFunctionDeclaration(BaseModel): + description: Optional[str] = None + name: str + parameters: Dict[str, Any] = Field( + ..., description='JSON schema for the function parameters' + ) + + +class GeminiGenerationConfig(BaseModel): + maxOutputTokens: Optional[int] = Field( + None, + description='Maximum number of tokens that can be generated in the response. A token is approximately 4 characters. 100 tokens correspond to roughly 60-80 words.\n', + examples=[2048], + ge=16, + le=8192, + ) + seed: Optional[int] = Field( + None, + description="When 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. Available for the following models:, gemini-2.5-flash-preview-04-1, gemini-2.5-pro-preview-05-0, gemini-2.0-flash-lite-00, gemini-2.0-flash-001\n", + examples=[343940597], + ) + stopSequences: Optional[List[str]] = None + temperature: Optional[float] = Field( + 1, + description="The temperature is used for sampling during response generation, which occurs when topP and topK are applied. Temperature controls the degree of randomness in token selection. Lower temperatures are good for prompts that require a less open-ended or creative response, while higher temperatures can lead to more diverse or creative results. A temperature of 0 means that the highest probability tokens are always selected. In this case, responses for a given prompt are mostly deterministic, but a small amount of variation is still possible. If the model returns a response that's too generic, too short, or the model gives a fallback response, try increasing the temperature\n", + ge=0.0, + le=2.0, + ) + topK: Optional[int] = Field( + 40, + description="Top-K changes how the model selects tokens for output. A top-K of 1 means the next selected token is the most probable among all tokens in the model's vocabulary. A top-K of 3 means that the next token is selected from among the 3 most probable tokens by using temperature.\n", + examples=[40], + ge=1, + ) + topP: Optional[float] = Field( + 0.95, + description='If specified, nucleus sampling is used.\nTop-P changes how the model selects tokens for output. Tokens are selected from the most (see top-K) to least probable until the sum of their probabilities equals the top-P value. For example, if tokens A, B, and C have a probability of 0.3, 0.2, and 0.1 and the top-P value is 0.5, then the model will select either A or B as the next token by using temperature and excludes C as a candidate.\nSpecify a lower value for less random responses and a higher value for more random responses.\n', + ge=0.0, + le=1.0, + ) + + +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 GeminiOffset(BaseModel): + nanos: Optional[int] = Field( + None, + description='Signed fractions of a second at nanosecond resolution. Negative second values with fractions must still have non-negative nanos values.\n', + examples=[0], + ge=0, + le=999999999, + ) + seconds: Optional[int] = Field( + None, + description='Signed seconds of the span of time. Must be from -315,576,000,000 to +315,576,000,000 inclusive.\n', + examples=[60], + ge=-315576000000, + le=315576000000, + ) + + +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 Probability(str, Enum): + NEGLIGIBLE = 'NEGLIGIBLE' + LOW = 'LOW' + MEDIUM = 'MEDIUM' + HIGH = 'HIGH' + UNKNOWN = 'UNKNOWN' + + +class GeminiSafetyRating(BaseModel): + category: Optional[GeminiSafetyCategory] = None + probability: Optional[Probability] = Field( + None, + description='The probability that the content violates the specified safety category', + ) + + +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 GeminiTextPart(BaseModel): + text: Optional[str] = Field( + None, + description='A text prompt or code snippet.', + examples=['Answer as concisely as possible'], + ) + + +class GeminiTool(BaseModel): + functionDeclarations: Optional[List[GeminiFunctionDeclaration]] = None + + +class GeminiVideoMetadata(BaseModel): + endOffset: Optional[GeminiOffset] = None + startOffset: Optional[GeminiOffset] = None + + +class IdeogramColorPalette1(BaseModel): + name: str = Field(..., description='Name of the preset color palette') + + +class Member(BaseModel): + color: Optional[str] = Field( + None, description='Hexadecimal color code', pattern='^#[0-9A-Fa-f]{6}$' + ) + weight: Optional[float] = Field( + None, description='Optional weight for the color (0-1)', ge=0.0, le=1.0 + ) + + +class IdeogramColorPalette2(BaseModel): + members: List[Member] = Field( + ..., description='Array of color definitions with optional weights' + ) + + +class IdeogramColorPalette( + RootModel[Union[IdeogramColorPalette1, IdeogramColorPalette2]] +): + root: Union[IdeogramColorPalette1, IdeogramColorPalette2] = Field( + ..., + description='A color palette specification that can either use a preset name or explicit color definitions with weights', + ) + + +class ImageRequest(BaseModel): + aspect_ratio: Optional[str] = Field( + None, + description="Optional. The aspect ratio (e.g., 'ASPECT_16_9', 'ASPECT_1_1'). Cannot be used with resolution. Defaults to 'ASPECT_1_1' if unspecified.", + ) + color_palette: Optional[Dict[str, Any]] = Field( + None, description='Optional. Color palette object. Only for V_2, V_2_TURBO.' + ) + magic_prompt_option: Optional[str] = Field( + None, description="Optional. MagicPrompt usage ('AUTO', 'ON', 'OFF')." + ) + model: str = Field(..., description="The model used (e.g., 'V_2', 'V_2A_TURBO')") + negative_prompt: Optional[str] = Field( + None, + description='Optional. Description of what to exclude. Only for V_1, V_1_TURBO, V_2, V_2_TURBO.', + ) + num_images: Optional[int] = Field( + 1, + description='Optional. Number of images to generate (1-8). Defaults to 1.', + ge=1, + le=8, + ) + prompt: str = Field( + ..., description='Required. The prompt to use to generate the image.' + ) + resolution: Optional[str] = Field( + None, + description="Optional. Resolution (e.g., 'RESOLUTION_1024_1024'). Only for model V_2. Cannot be used with aspect_ratio.", + ) + seed: Optional[int] = Field( + None, + description='Optional. A number between 0 and 2147483647.', + ge=0, + le=2147483647, + ) + style_type: Optional[str] = Field( + None, + description="Optional. Style type ('AUTO', 'GENERAL', 'REALISTIC', 'DESIGN', 'RENDER_3D', 'ANIME'). Only for models V_2 and above.", + ) + + +class IdeogramGenerateRequest(BaseModel): + image_request: ImageRequest = Field( + ..., description='The image generation request parameters.' + ) + + +class Datum(BaseModel): + is_image_safe: Optional[bool] = Field( + None, description='Indicates whether the image is considered safe.' + ) + prompt: Optional[str] = Field( + None, description='The prompt used to generate this image.' + ) + resolution: Optional[str] = Field( + None, description="The resolution of the generated image (e.g., '1024x1024')." + ) + seed: Optional[int] = Field( + None, description='The seed value used for this generation.' + ) + style_type: Optional[str] = Field( + None, + description="The style type used for generation (e.g., 'REALISTIC', 'ANIME').", + ) + url: Optional[str] = Field(None, description='URL to the generated image.') + + +class IdeogramGenerateResponse(BaseModel): + created: Optional[datetime] = Field( + None, description='Timestamp when the generation was created.' + ) + data: Optional[List[Datum]] = Field( + None, description='Array of generated image information.' + ) + + +class StyleCode(RootModel[str]): + root: str = Field(..., pattern='^[0-9A-Fa-f]{8}$') + + +class Datum1(BaseModel): + is_image_safe: Optional[bool] = None + prompt: Optional[str] = None + resolution: Optional[str] = None + seed: Optional[int] = None + style_type: Optional[str] = None + url: Optional[str] = None + + +class IdeogramV3IdeogramResponse(BaseModel): + created: Optional[datetime] = None + data: Optional[List[Datum1]] = None + + +class RenderingSpeed1(str, Enum): + TURBO = 'TURBO' + DEFAULT = 'DEFAULT' + QUALITY = 'QUALITY' + + +class IdeogramV3ReframeRequest(BaseModel): + color_palette: Optional[Dict[str, Any]] = None + image: Optional[StrictBytes] = None + num_images: Optional[int] = Field(None, ge=1, le=8) + rendering_speed: Optional[RenderingSpeed1] = None + resolution: str + seed: Optional[int] = Field(None, ge=0, le=2147483647) + style_codes: Optional[List[str]] = None + style_reference_images: Optional[List[StrictBytes]] = None + + +class MagicPrompt(str, Enum): + AUTO = 'AUTO' + ON = 'ON' + OFF = 'OFF' + + +class StyleType(str, Enum): + AUTO = 'AUTO' + GENERAL = 'GENERAL' + REALISTIC = 'REALISTIC' + DESIGN = 'DESIGN' + + +class IdeogramV3RemixRequest(BaseModel): + aspect_ratio: Optional[str] = None + color_palette: Optional[Dict[str, Any]] = None + image: Optional[StrictBytes] = None + image_weight: Optional[int] = Field(50, ge=1, le=100) + magic_prompt: Optional[MagicPrompt] = None + negative_prompt: Optional[str] = None + num_images: Optional[int] = Field(None, ge=1, le=8) + prompt: str + rendering_speed: Optional[RenderingSpeed1] = None + resolution: Optional[str] = None + seed: Optional[int] = Field(None, ge=0, le=2147483647) + style_codes: Optional[List[str]] = None + style_reference_images: Optional[List[StrictBytes]] = None + style_type: Optional[StyleType] = None + + +class IdeogramV3ReplaceBackgroundRequest(BaseModel): + color_palette: Optional[Dict[str, Any]] = None + image: Optional[StrictBytes] = None + magic_prompt: Optional[MagicPrompt] = None + num_images: Optional[int] = Field(None, ge=1, le=8) + prompt: str + rendering_speed: Optional[RenderingSpeed1] = None + seed: Optional[int] = Field(None, ge=0, le=2147483647) + style_codes: Optional[List[str]] = None + style_reference_images: Optional[List[StrictBytes]] = None + + +class ColorPalette(BaseModel): + name: str = Field(..., description='Name of the color palette', examples=['PASTEL']) + + +class MagicPrompt2(str, Enum): + ON = 'ON' + OFF = 'OFF' + + +class StyleType1(str, Enum): + GENERAL = 'GENERAL' + + +class ImagenImageGenerationInstance(BaseModel): + prompt: str = Field(..., description='Text prompt for image generation') + + +class AspectRatio(str, Enum): + field_1_1 = '1:1' + field_9_16 = '9:16' + field_16_9 = '16:9' + field_3_4 = '3:4' + field_4_3 = '4:3' + + +class PersonGeneration(str, Enum): + dont_allow = 'dont_allow' + allow_adult = 'allow_adult' + allow_all = 'allow_all' + + +class SafetySetting(str, Enum): + block_most = 'block_most' + block_some = 'block_some' + block_few = 'block_few' + block_fewest = 'block_fewest' + + +class ImagenImagePrediction(BaseModel): + bytesBase64Encoded: Optional[str] = Field( + None, description='Base64-encoded image content' + ) + mimeType: Optional[str] = Field( + None, description='MIME type of the generated image' + ) + prompt: Optional[str] = Field( + None, description='Enhanced or rewritten prompt used to generate this image' + ) + + +class MimeType(str, Enum): + image_png = 'image/png' + image_jpeg = 'image/jpeg' + + +class ImagenOutputOptions(BaseModel): + compressionQuality: Optional[int] = Field(None, ge=0, le=100) + mimeType: Optional[MimeType] = None + + +class Includable(str, Enum): + file_search_call_results = 'file_search_call.results' + message_input_image_image_url = 'message.input_image.image_url' + computer_call_output_output_image_url = 'computer_call_output.output.image_url' + + +class Type7(str, Enum): + input_file = 'input_file' + + +class InputFileContent(BaseModel): + file_data: Optional[str] = Field( + None, description='The content of the file to be sent to the model.\n' + ) + file_id: Optional[str] = Field( + None, description='The ID of the file to be sent to the model.' + ) + filename: Optional[str] = Field( + None, description='The name of the file to be sent to the model.' + ) + type: Type7 = Field( + ..., description='The type of the input item. Always `input_file`.' + ) + + +class Detail(str, Enum): + low = 'low' + high = 'high' + auto = 'auto' + + +class Type8(str, Enum): + input_image = 'input_image' + + +class InputImageContent(BaseModel): + detail: Detail = Field( + ..., + description='The detail level of the image to be sent to the model. One of `high`, `low`, or `auto`. Defaults to `auto`.', + ) + file_id: Optional[str] = Field( + None, description='The ID of the file to be sent to the model.' + ) + image_url: Optional[str] = Field( + None, + description='The URL of the image to be sent to the model. A fully qualified URL or base64 encoded image in a data URL.', + ) + type: Type8 = Field( + ..., description='The type of the input item. Always `input_image`.' + ) + + +class Role3(str, Enum): + user = 'user' + system = 'system' + developer = 'developer' + + +class Type9(str, Enum): + message = 'message' + + +class Type10(str, Enum): + input_text = 'input_text' + + +class InputTextContent(BaseModel): + text: str = Field(..., description='The text input to the model.') + type: Type10 = Field( + ..., description='The type of the input item. Always `input_text`.' + ) + + +class KlingAudioUploadType(str, Enum): + file = 'file' + url = 'url' + + +class KlingCameraConfig(BaseModel): + horizontal: Optional[float] = Field( + None, + description="Controls camera's movement along horizontal axis (x-axis). Negative indicates left, positive indicates right.", + ge=-10.0, + le=10.0, + ) + pan: Optional[float] = Field( + None, + description="Controls camera's rotation in vertical plane (x-axis). Negative indicates downward rotation, positive indicates upward rotation.", + ge=-10.0, + le=10.0, + ) + roll: Optional[float] = Field( + None, + description="Controls camera's rolling amount (z-axis). Negative indicates counterclockwise, positive indicates clockwise.", + ge=-10.0, + le=10.0, + ) + tilt: Optional[float] = Field( + None, + description="Controls camera's rotation in horizontal plane (y-axis). Negative indicates left rotation, positive indicates right rotation.", + ge=-10.0, + le=10.0, + ) + vertical: Optional[float] = Field( + None, + description="Controls camera's movement along vertical axis (y-axis). Negative indicates downward, positive indicates upward.", + ge=-10.0, + le=10.0, + ) + zoom: Optional[float] = Field( + None, + description="Controls change in camera's focal length. Negative indicates narrower field of view, positive indicates wider field of view.", + ge=-10.0, + le=10.0, + ) + + +class KlingCameraControlType(str, Enum): + simple = 'simple' + down_back = 'down_back' + forward_up = 'forward_up' + right_turn_forward = 'right_turn_forward' + left_turn_forward = 'left_turn_forward' + + +class KlingCharacterEffectModelName(str, Enum): + kling_v1 = 'kling-v1' + kling_v1_5 = 'kling-v1-5' + kling_v1_6 = 'kling-v1-6' + + +class KlingDualCharacterEffectsScene(str, Enum): + hug = 'hug' + kiss = 'kiss' + heart_gesture = 'heart_gesture' + + +class KlingDualCharacterImages(RootModel[List[str]]): + root: List[str] = Field(..., max_length=2, min_length=2) + + +class KlingErrorResponse(BaseModel): + code: int = Field( + ..., + description='- 1000: Authentication failed\n- 1001: Authorization is empty\n- 1002: Authorization is invalid\n- 1003: Authorization is not yet valid\n- 1004: Authorization has expired\n- 1100: Account exception\n- 1101: Account in arrears (postpaid scenario)\n- 1102: Resource pack depleted or expired (prepaid scenario)\n- 1103: Unauthorized access to requested resource\n- 1200: Invalid request parameters\n- 1201: Invalid parameters\n- 1202: Invalid request method\n- 1203: Requested resource does not exist\n- 1300: Trigger platform strategy\n- 1301: Trigger content security policy\n- 1302: API request too frequent\n- 1303: Concurrency/QPS exceeds limit\n- 1304: Trigger IP whitelist policy\n- 5000: Internal server error\n- 5001: Service temporarily unavailable\n- 5002: Server internal timeout\n', + ) + message: str = Field(..., description='Human-readable error message') + request_id: str = Field( + ..., description='Request ID for tracking and troubleshooting' + ) + + +class Trajectory(BaseModel): + x: Optional[int] = Field( + None, + description='The horizontal coordinate of trajectory point. Based on bottom-left corner of image as origin (0,0).', + ) + y: Optional[int] = Field( + None, + description='The vertical coordinate of trajectory point. Based on bottom-left corner of image as origin (0,0).', + ) + + +class DynamicMask(BaseModel): + mask: Optional[AnyUrl] = Field( + None, + description='Dynamic Brush Application Area (Mask image created by users using the motion brush). The aspect ratio must match the input image.', + ) + trajectories: Optional[List[Trajectory]] = None + + +class TaskInfo(BaseModel): + external_task_id: Optional[str] = None + + +class KlingImageGenAspectRatio(str, Enum): + field_16_9 = '16:9' + field_9_16 = '9:16' + field_1_1 = '1:1' + field_4_3 = '4:3' + field_3_4 = '3:4' + field_3_2 = '3:2' + field_2_3 = '2:3' + field_21_9 = '21:9' + + +class KlingImageGenImageReferenceType(str, Enum): + subject = 'subject' + face = 'face' + + +class KlingImageGenModelName(str, Enum): + kling_v1 = 'kling-v1' + kling_v1_5 = 'kling-v1-5' + kling_v2 = 'kling-v2' + + +class KlingImageGenerationsRequest(BaseModel): + aspect_ratio: Optional[KlingImageGenAspectRatio] = '16:9' + callback_url: Optional[AnyUrl] = Field( + None, description='The callback notification address' + ) + human_fidelity: Optional[float] = Field( + 0.45, description='Subject reference similarity', ge=0.0, le=1.0 + ) + image: Optional[str] = Field( + None, description='Reference Image - Base64 encoded string or image URL' + ) + image_fidelity: Optional[float] = Field( + 0.5, description='Reference intensity for user-uploaded images', ge=0.0, le=1.0 + ) + image_reference: Optional[KlingImageGenImageReferenceType] = None + model_name: Optional[KlingImageGenModelName] = 'kling-v1' + n: Optional[int] = Field(1, description='Number of generated images', ge=1, le=9) + negative_prompt: Optional[str] = Field( + None, description='Negative text prompt', max_length=200 + ) + prompt: str = Field(..., description='Positive text prompt', max_length=500) + + +class KlingImageResult(BaseModel): + index: Optional[int] = Field(None, description='Image Number (0-9)') + url: Optional[AnyUrl] = Field(None, description='URL for generated image') + + +class KlingLipSyncMode(str, Enum): + text2video = 'text2video' + audio2video = 'audio2video' + + +class KlingLipSyncVoiceLanguage(str, Enum): + zh = 'zh' + en = 'en' + + +class ResourcePackType(str, Enum): + decreasing_total = 'decreasing_total' + constant_period = 'constant_period' + + +class Status4(str, Enum): + toBeOnline = 'toBeOnline' + online = 'online' + expired = 'expired' + runOut = 'runOut' + + +class ResourcePackSubscribeInfo(BaseModel): + effective_time: Optional[int] = Field( + None, description='Effective time, Unix timestamp in ms' + ) + invalid_time: Optional[int] = Field( + None, description='Expiration time, Unix timestamp in ms' + ) + purchase_time: Optional[int] = Field( + None, description='Purchase time, Unix timestamp in ms' + ) + remaining_quantity: Optional[float] = Field( + None, description='Remaining quantity (updated with a 12-hour delay)' + ) + resource_pack_id: Optional[str] = Field(None, description='Resource package ID') + resource_pack_name: Optional[str] = Field(None, description='Resource package name') + resource_pack_type: Optional[ResourcePackType] = Field( + None, + description='Resource package type (decreasing_total=decreasing total, constant_period=constant periodicity)', + ) + status: Optional[Status4] = Field(None, description='Resource Package Status') + total_quantity: Optional[float] = Field(None, description='Total quantity') + + +class Data3(BaseModel): + code: Optional[int] = Field(None, description='Error code; 0 indicates success') + msg: Optional[str] = Field(None, description='Error information') + resource_pack_subscribe_infos: Optional[List[ResourcePackSubscribeInfo]] = Field( + None, description='Resource package list' + ) + + +class KlingResourcePackageResponse(BaseModel): + code: Optional[int] = Field(None, description='Error code; 0 indicates success') + data: Optional[Data3] = None + message: Optional[str] = Field(None, description='Error information') + request_id: Optional[str] = Field( + None, + description='Request ID, generated by the system, used to track requests and troubleshoot problems', + ) + + +class KlingSingleImageEffectDuration(str, Enum): + field_5 = '5' + + +class KlingSingleImageEffectModelName(str, Enum): + kling_v1_6 = 'kling-v1-6' + + +class KlingSingleImageEffectsScene(str, Enum): + bloombloom = 'bloombloom' + dizzydizzy = 'dizzydizzy' + fuzzyfuzzy = 'fuzzyfuzzy' + squish = 'squish' + expansion = 'expansion' + + +class KlingTaskStatus(str, Enum): + submitted = 'submitted' + processing = 'processing' + succeed = 'succeed' + failed = 'failed' + + +class KlingTextToVideoModelName(str, Enum): + kling_v1 = 'kling-v1' + kling_v1_6 = 'kling-v1-6' + + +class KlingVideoGenAspectRatio(str, Enum): + field_16_9 = '16:9' + field_9_16 = '9:16' + field_1_1 = '1:1' + + +class KlingVideoGenCfgScale(RootModel[float]): + root: float = Field( + ..., + description="Flexibility in video generation. The higher the value, the lower the model's degree of flexibility, and the stronger the relevance to the user's prompt.", + ge=0.0, + le=1.0, + ) + + +class KlingVideoGenDuration(str, Enum): + field_5 = '5' + field_10 = '10' + + +class KlingVideoGenMode(str, Enum): + std = 'std' + pro = 'pro' + + +class KlingVideoGenModelName(str, Enum): + kling_v1 = 'kling-v1' + kling_v1_5 = 'kling-v1-5' + kling_v1_6 = 'kling-v1-6' + kling_v2_master = 'kling-v2-master' + + +class KlingVideoResult(BaseModel): + duration: Optional[str] = Field(None, description='Total video duration') + id: Optional[str] = Field(None, description='Generated video ID') + url: Optional[AnyUrl] = Field(None, description='URL for generated video') + + +class KlingVirtualTryOnModelName(str, Enum): + kolors_virtual_try_on_v1 = 'kolors-virtual-try-on-v1' + kolors_virtual_try_on_v1_5 = 'kolors-virtual-try-on-v1-5' + + +class KlingVirtualTryOnRequest(BaseModel): + callback_url: Optional[AnyUrl] = Field( + None, description='The callback notification address' + ) + cloth_image: Optional[str] = Field( + None, + description='Reference clothing image - Base64 encoded string or image URL', + ) + human_image: str = Field( + ..., description='Reference human image - Base64 encoded string or image URL' + ) + model_name: Optional[KlingVirtualTryOnModelName] = 'kolors-virtual-try-on-v1' + + +class TaskResult6(BaseModel): + images: Optional[List[KlingImageResult]] = None + + +class Data7(BaseModel): + created_at: Optional[int] = Field(None, description='Task creation time') + task_id: Optional[str] = Field(None, description='Task ID') + task_result: Optional[TaskResult6] = None + task_status: Optional[KlingTaskStatus] = None + task_status_msg: Optional[str] = Field(None, description='Task status information') + updated_at: Optional[int] = Field(None, description='Task update time') + + +class KlingVirtualTryOnResponse(BaseModel): + code: Optional[int] = Field(None, description='Error code') + data: Optional[Data7] = None + message: Optional[str] = Field(None, description='Error message') + request_id: Optional[str] = Field(None, description='Request ID') + + +class LumaAspectRatio(str, Enum): + field_1_1 = '1:1' + field_16_9 = '16:9' + field_9_16 = '9:16' + field_4_3 = '4:3' + field_3_4 = '3:4' + field_21_9 = '21:9' + field_9_21 = '9:21' + + +class LumaAssets(BaseModel): + image: Optional[AnyUrl] = Field(None, description='The URL of the image') + progress_video: Optional[AnyUrl] = Field( + None, description='The URL of the progress video' + ) + video: Optional[AnyUrl] = Field(None, description='The URL of the video') + + +class GenerationType(str, Enum): + add_audio = 'add_audio' + + +class LumaAudioGenerationRequest(BaseModel): + callback_url: Optional[AnyUrl] = Field( + None, description='The callback URL for the audio' + ) + generation_type: Optional[GenerationType] = 'add_audio' + negative_prompt: Optional[str] = Field( + None, description='The negative prompt of the audio' + ) + prompt: Optional[str] = Field(None, description='The prompt of the audio') + + +class LumaError(BaseModel): + detail: Optional[str] = Field(None, description='The error message') + + +class Type11(str, Enum): + generation = 'generation' + + +class LumaGenerationReference(BaseModel): + id: UUID = Field(..., description='The ID of the generation') + type: Literal['generation'] + + +class GenerationType1(str, Enum): + video = 'video' + + +class LumaGenerationType(str, Enum): + video = 'video' + image = 'image' + + +class GenerationType2(str, Enum): + image = 'image' + + +class LumaImageIdentity(BaseModel): + images: Optional[List[AnyUrl]] = Field( + None, description='The URLs of the image identity' + ) + + +class LumaImageModel(str, Enum): + photon_1 = 'photon-1' + photon_flash_1 = 'photon-flash-1' + + +class LumaImageRef(BaseModel): + url: Optional[AnyUrl] = Field(None, description='The URL of the image reference') + weight: Optional[float] = Field( + None, description='The weight of the image reference' + ) + + +class Type12(str, Enum): + image = 'image' + + +class LumaImageReference(BaseModel): + type: Literal['image'] + url: AnyUrl = Field(..., description='The URL of the image') + + +class LumaKeyframe(RootModel[Union[LumaGenerationReference, LumaImageReference]]): + root: Union[LumaGenerationReference, LumaImageReference] = Field( + ..., + description='A keyframe can be either a Generation reference, an Image, or a Video', + discriminator='type', + ) + + +class LumaKeyframes(BaseModel): + frame0: Optional[LumaKeyframe] = None + frame1: Optional[LumaKeyframe] = None + + +class LumaModifyImageRef(BaseModel): + url: Optional[AnyUrl] = Field(None, description='The URL of the image reference') + weight: Optional[float] = Field( + None, description='The weight of the modify image reference' + ) + + +class LumaState(str, Enum): + queued = 'queued' + dreaming = 'dreaming' + completed = 'completed' + failed = 'failed' + + +class GenerationType3(str, Enum): + upscale_video = 'upscale_video' + + +class LumaVideoModel(str, Enum): + ray_2 = 'ray-2' + ray_flash_2 = 'ray-flash-2' + ray_1_6 = 'ray-1-6' + + +class LumaVideoModelOutputDuration1(str, Enum): + field_5s = '5s' + field_9s = '9s' + + +class LumaVideoModelOutputDuration( + RootModel[Union[LumaVideoModelOutputDuration1, str]] +): + root: Union[LumaVideoModelOutputDuration1, str] + + +class LumaVideoModelOutputResolution1(str, Enum): + field_540p = '540p' + field_720p = '720p' + field_1080p = '1080p' + field_4k = '4k' + + +class LumaVideoModelOutputResolution( + RootModel[Union[LumaVideoModelOutputResolution1, str]] +): + root: Union[LumaVideoModelOutputResolution1, str] + + +class MinimaxBaseResponse(BaseModel): + status_code: int = Field( + ..., + description='Status code. 0 indicates success, other values indicate errors.', + ) + status_msg: str = Field( + ..., description='Specific error details or success message.' + ) + + +class File(BaseModel): + bytes: Optional[int] = Field(None, description='File size in bytes') + created_at: Optional[int] = Field( + None, description='Unix timestamp when the file was created, in seconds' + ) + download_url: Optional[str] = Field( + None, description='The URL to download the video' + ) + file_id: Optional[int] = Field(None, description='Unique identifier for the file') + filename: Optional[str] = Field(None, description='The name of the file') + purpose: Optional[str] = Field(None, description='The purpose of using the file') + + +class MinimaxFileRetrieveResponse(BaseModel): + base_resp: MinimaxBaseResponse + file: File + + +class Status5(str, Enum): + Queueing = 'Queueing' + Preparing = 'Preparing' + Processing = 'Processing' + Success = 'Success' + Fail = 'Fail' + + +class MinimaxTaskResultResponse(BaseModel): + base_resp: MinimaxBaseResponse + file_id: Optional[str] = Field( + None, + description='After the task status changes to Success, this field returns the file ID corresponding to the generated video.', + ) + status: Status5 = Field( + ..., + description="Task status: 'Queueing' (in queue), 'Preparing' (task is preparing), 'Processing' (generating), 'Success' (task completed successfully), or 'Fail' (task failed).", + ) + task_id: str = Field(..., description='The task ID being queried.') + + +class Model(str, Enum): + T2V_01_Director = 'T2V-01-Director' + I2V_01_Director = 'I2V-01-Director' + S2V_01 = 'S2V-01' + I2V_01 = 'I2V-01' + I2V_01_live = 'I2V-01-live' + T2V_01 = 'T2V-01' + + +class SubjectReferenceItem(BaseModel): + image: Optional[str] = Field( + None, description='URL or base64 encoding of the subject reference image.' + ) + mask: Optional[str] = Field( + None, + description='URL or base64 encoding of the mask for the subject reference image.', + ) + + +class MinimaxVideoGenerationRequest(BaseModel): + callback_url: Optional[str] = Field( + None, + description='Optional. URL to receive real-time status updates about the video generation task.', + ) + first_frame_image: Optional[str] = Field( + None, + description='URL or base64 encoding of the first frame image. Required when model is I2V-01, I2V-01-Director, or I2V-01-live.', + ) + model: Model = Field( + ..., + description='Required. ID of model. Options: T2V-01-Director, I2V-01-Director, S2V-01, I2V-01, I2V-01-live, T2V-01', + ) + prompt: Optional[str] = Field( + None, + description='Description of the video. Should be less than 2000 characters. Supports camera movement instructions in [brackets].', + max_length=2000, + ) + prompt_optimizer: Optional[bool] = Field( + True, + description='If true (default), the model will automatically optimize the prompt. Set to false for more precise control.', + ) + subject_reference: Optional[List[SubjectReferenceItem]] = Field( + None, + description='Only available when model is S2V-01. The model will generate a video based on the subject uploaded through this parameter.', + ) + + +class MinimaxVideoGenerationResponse(BaseModel): + base_resp: MinimaxBaseResponse + task_id: str = Field( + ..., description='The task ID for the asynchronous video generation task.' + ) + + +class Truncation(str, Enum): + disabled = 'disabled' + auto = 'auto' + + +class ModelResponseProperties(BaseModel): + instructions: Optional[str] = Field( + None, description='Instructions for the model on how to generate the response' + ) + max_output_tokens: Optional[int] = Field( + None, description='Maximum number of tokens to generate' + ) + model: Optional[str] = Field( + None, description='The model used to generate the response' + ) + temperature: Optional[float] = Field( + 1, description='Controls randomness in the response', ge=0.0, le=2.0 + ) + top_p: Optional[float] = Field( + 1, + description='Controls diversity of the response via nucleus sampling', + ge=0.0, + le=1.0, + ) + truncation: Optional[Truncation] = Field( + 'disabled', description='How to handle truncation of the response' + ) + + +class Moderation(str, Enum): + low = 'low' + auto = 'auto' + + +class OutputFormat1(str, Enum): + png = 'png' + webp = 'webp' + jpeg = 'jpeg' + + +class OpenAIImageEditRequest(BaseModel): + background: Optional[str] = Field( + None, description='Background transparency', examples=['opaque'] + ) + model: str = Field( + ..., description='The model to use for image editing', examples=['gpt-image-1'] + ) + moderation: Optional[Moderation] = Field( + None, description='Content moderation setting', examples=['auto'] + ) + n: Optional[int] = Field( + None, description='The number of images to generate', examples=[1] + ) + output_compression: Optional[int] = Field( + None, description='Compression level for JPEG or WebP (0-100)', examples=[100] + ) + output_format: Optional[OutputFormat1] = Field( + None, description='Format of the output image', examples=['png'] + ) + prompt: str = Field( + ..., + description='A text description of the desired edit', + examples=['Give the rocketship rainbow coloring'], + ) + quality: Optional[str] = Field( + None, description='The quality of the edited image', examples=['low'] + ) + size: Optional[str] = Field( + None, description='Size of the output image', examples=['1024x1024'] + ) + user: Optional[str] = Field( + None, + description='A unique identifier for end-user monitoring', + examples=['user-1234'], + ) + + +class Background(str, Enum): + transparent = 'transparent' + opaque = 'opaque' + + +class Quality(str, Enum): + low = 'low' + medium = 'medium' + high = 'high' + standard = 'standard' + hd = 'hd' + + +class ResponseFormat(str, Enum): + url = 'url' + b64_json = 'b64_json' + + +class Style(str, Enum): + vivid = 'vivid' + natural = 'natural' + + +class OpenAIImageGenerationRequest(BaseModel): + background: Optional[Background] = Field( + None, description='Background transparency', examples=['opaque'] + ) + model: Optional[str] = Field( + None, description='The model to use for image generation', examples=['dall-e-3'] + ) + moderation: Optional[Moderation] = Field( + None, description='Content moderation setting', examples=['auto'] + ) + n: Optional[int] = Field( + None, + description='The number of images to generate (1-10). Only 1 supported for dall-e-3.', + examples=[1], + ) + output_compression: Optional[int] = Field( + None, description='Compression level for JPEG or WebP (0-100)', examples=[100] + ) + output_format: Optional[OutputFormat1] = Field( + None, description='Format of the output image', examples=['png'] + ) + prompt: str = Field( + ..., + description='A text description of the desired image', + examples=['Draw a rocket in front of a blackhole in deep space'], + ) + quality: Optional[Quality] = Field( + None, description='The quality of the generated image', examples=['high'] + ) + response_format: Optional[ResponseFormat] = Field( + None, description='Response format of image data', examples=['b64_json'] + ) + size: Optional[str] = Field( + None, + description='Size of the image (e.g., 1024x1024, 1536x1024, auto)', + examples=['1024x1536'], + ) + style: Optional[Style] = Field( + None, description='Style of the image (only for dall-e-3)', examples=['vivid'] + ) + user: Optional[str] = Field( + None, + description='A unique identifier for end-user monitoring', + examples=['user-1234'], + ) + + +class Datum2(BaseModel): + b64_json: Optional[str] = Field(None, description='Base64 encoded image data') + revised_prompt: Optional[str] = Field(None, description='Revised prompt') + url: Optional[str] = Field(None, description='URL of the image') + + +class InputTokensDetails(BaseModel): + image_tokens: Optional[int] = None + text_tokens: Optional[int] = None + + +class Usage(BaseModel): + input_tokens: Optional[int] = None + input_tokens_details: Optional[InputTokensDetails] = None + output_tokens: Optional[int] = None + total_tokens: Optional[int] = None + + +class OpenAIImageGenerationResponse(BaseModel): + data: Optional[List[Datum2]] = None + usage: Optional[Usage] = None + + +class OpenAIModels(str, Enum): + gpt_4 = 'gpt-4' + gpt_4_0314 = 'gpt-4-0314' + gpt_4_0613 = 'gpt-4-0613' + gpt_4_32k = 'gpt-4-32k' + gpt_4_32k_0314 = 'gpt-4-32k-0314' + gpt_4_32k_0613 = 'gpt-4-32k-0613' + gpt_4_0125_preview = 'gpt-4-0125-preview' + gpt_4_turbo = 'gpt-4-turbo' + gpt_4_turbo_2024_04_09 = 'gpt-4-turbo-2024-04-09' + gpt_4_turbo_preview = 'gpt-4-turbo-preview' + gpt_4_1106_preview = 'gpt-4-1106-preview' + gpt_4_vision_preview = 'gpt-4-vision-preview' + gpt_3_5_turbo = 'gpt-3.5-turbo' + gpt_3_5_turbo_16k = 'gpt-3.5-turbo-16k' + gpt_3_5_turbo_0301 = 'gpt-3.5-turbo-0301' + gpt_3_5_turbo_0613 = 'gpt-3.5-turbo-0613' + gpt_3_5_turbo_1106 = 'gpt-3.5-turbo-1106' + gpt_3_5_turbo_0125 = 'gpt-3.5-turbo-0125' + gpt_3_5_turbo_16k_0613 = 'gpt-3.5-turbo-16k-0613' + gpt_4_1 = 'gpt-4.1' + gpt_4_1_mini = 'gpt-4.1-mini' + gpt_4_1_nano = 'gpt-4.1-nano' + gpt_4_1_2025_04_14 = 'gpt-4.1-2025-04-14' + gpt_4_1_mini_2025_04_14 = 'gpt-4.1-mini-2025-04-14' + gpt_4_1_nano_2025_04_14 = 'gpt-4.1-nano-2025-04-14' + o1 = 'o1' + o1_mini = 'o1-mini' + o1_preview = 'o1-preview' + o1_pro = 'o1-pro' + o1_2024_12_17 = 'o1-2024-12-17' + o1_preview_2024_09_12 = 'o1-preview-2024-09-12' + o1_mini_2024_09_12 = 'o1-mini-2024-09-12' + o1_pro_2025_03_19 = 'o1-pro-2025-03-19' + o3 = 'o3' + o3_mini = 'o3-mini' + o3_2025_04_16 = 'o3-2025-04-16' + o3_mini_2025_01_31 = 'o3-mini-2025-01-31' + o4_mini = 'o4-mini' + o4_mini_2025_04_16 = 'o4-mini-2025-04-16' + gpt_4o = 'gpt-4o' + gpt_4o_mini = 'gpt-4o-mini' + gpt_4o_2024_11_20 = 'gpt-4o-2024-11-20' + gpt_4o_2024_08_06 = 'gpt-4o-2024-08-06' + gpt_4o_2024_05_13 = 'gpt-4o-2024-05-13' + gpt_4o_mini_2024_07_18 = 'gpt-4o-mini-2024-07-18' + gpt_4o_audio_preview = 'gpt-4o-audio-preview' + gpt_4o_audio_preview_2024_10_01 = 'gpt-4o-audio-preview-2024-10-01' + gpt_4o_audio_preview_2024_12_17 = 'gpt-4o-audio-preview-2024-12-17' + gpt_4o_mini_audio_preview = 'gpt-4o-mini-audio-preview' + gpt_4o_mini_audio_preview_2024_12_17 = 'gpt-4o-mini-audio-preview-2024-12-17' + gpt_4o_search_preview = 'gpt-4o-search-preview' + gpt_4o_mini_search_preview = 'gpt-4o-mini-search-preview' + gpt_4o_search_preview_2025_03_11 = 'gpt-4o-search-preview-2025-03-11' + gpt_4o_mini_search_preview_2025_03_11 = 'gpt-4o-mini-search-preview-2025-03-11' + computer_use_preview = 'computer-use-preview' + computer_use_preview_2025_03_11 = 'computer-use-preview-2025-03-11' + chatgpt_4o_latest = 'chatgpt-4o-latest' + + +class Reason(str, Enum): + max_output_tokens = 'max_output_tokens' + content_filter = 'content_filter' + + +class IncompleteDetails(BaseModel): + reason: Optional[Reason] = Field( + None, description='The reason why the response is incomplete.' + ) + + +class Object(str, Enum): + response = 'response' + + +class Status6(str, Enum): + completed = 'completed' + failed = 'failed' + in_progress = 'in_progress' + incomplete = 'incomplete' + + +class Type13(str, Enum): + output_audio = 'output_audio' + + +class OutputAudioContent(BaseModel): + data: str = Field(..., description='Base64-encoded audio data') + transcript: str = Field(..., description='Transcript of the audio') + type: Type13 = Field(..., description='The type of output content') + + +class Role4(str, Enum): + assistant = 'assistant' + + +class Type14(str, Enum): + message = 'message' + + +class Type15(str, Enum): + output_text = 'output_text' + + +class OutputTextContent(BaseModel): + text: str = Field(..., description='The text content') + type: Type15 = Field(..., description='The type of output content') + + +class AspectRatio1(RootModel[float]): + root: float = Field( + ..., + description='Aspect ratio (width / height)', + ge=0.4, + le=2.5, + title='Aspectratio', + ) + + +class IngredientsMode(str, Enum): + creative = 'creative' + precise = 'precise' + + +class PikaBodyGenerate22C2vGenerate22PikascenesPost(BaseModel): + aspectRatio: Optional[AspectRatio1] = Field( + None, description='Aspect ratio (width / height)', title='Aspectratio' + ) + duration: Optional[int] = Field(5, title='Duration') + images: Optional[List[StrictBytes]] = Field(None, title='Images') + ingredientsMode: IngredientsMode = Field(..., title='Ingredientsmode') + negativePrompt: Optional[str] = Field(None, title='Negativeprompt') + promptText: Optional[str] = Field(None, title='Prompttext') + resolution: Optional[str] = Field('1080p', title='Resolution') + seed: Optional[int] = Field(None, title='Seed') + + +class PikaBodyGeneratePikadditionsGeneratePikadditionsPost(BaseModel): + image: Optional[StrictBytes] = Field(None, title='Image') + negativePrompt: Optional[str] = Field(None, title='Negativeprompt') + promptText: Optional[str] = Field(None, title='Prompttext') + seed: Optional[int] = Field(None, title='Seed') + video: Optional[StrictBytes] = Field(None, title='Video') + + +class PikaBodyGeneratePikaswapsGeneratePikaswapsPost(BaseModel): + image: Optional[StrictBytes] = Field(None, title='Image') + modifyRegionMask: Optional[StrictBytes] = Field( + None, + description='A mask image that specifies the region to modify, where the mask is white and the background is black', + title='Modifyregionmask', + ) + modifyRegionRoi: Optional[str] = Field( + None, + description='Plaintext description of the object / region to modify', + title='Modifyregionroi', + ) + negativePrompt: Optional[str] = Field(None, title='Negativeprompt') + promptText: Optional[str] = Field(None, title='Prompttext') + seed: Optional[int] = Field(None, title='Seed') + video: Optional[StrictBytes] = Field(None, title='Video') + + +class PikaDurationEnum(int, Enum): + integer_5 = 5 + integer_10 = 10 + + +class PikaGenerateResponse(BaseModel): + video_id: str = Field(..., title='Video Id') + + +class PikaResolutionEnum(str, Enum): + field_1080p = '1080p' + field_720p = '720p' + + +class PikaStatusEnum(str, Enum): + queued = 'queued' + started = 'started' + finished = 'finished' + + +class PikaValidationError(BaseModel): + loc: List[Union[str, int]] = Field(..., title='Location') + msg: str = Field(..., title='Message') + type: str = Field(..., title='Error Type') + + +class PikaVideoResponse(BaseModel): + id: str = Field(..., title='Id') + progress: Optional[int] = Field(None, title='Progress') + status: PikaStatusEnum + url: Optional[str] = Field(None, title='Url') + + +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 Resp(BaseModel): + img_id: Optional[int] = None + + +class PixverseImageUploadResponse(BaseModel): + ErrCode: Optional[int] = None + ErrMsg: Optional[str] = None + Resp_1: Optional[Resp] = Field(None, alias='Resp') + + +class Duration(int, Enum): + integer_5 = 5 + integer_8 = 8 + + +class Model1(str, Enum): + v3_5 = 'v3.5' + + +class MotionMode(str, Enum): + normal = 'normal' + fast = 'fast' + + +class Quality1(str, Enum): + field_360p = '360p' + field_540p = '540p' + field_720p = '720p' + field_1080p = '1080p' + + +class Style1(str, Enum): + anime = 'anime' + field_3d_animation = '3d_animation' + clay = 'clay' + comic = 'comic' + cyberpunk = 'cyberpunk' + + +class PixverseImageVideoRequest(BaseModel): + duration: Duration + img_id: int + model: Model1 + motion_mode: Optional[MotionMode] = None + prompt: str + quality: Quality1 + seed: Optional[int] = None + style: Optional[Style1] = None + template_id: Optional[int] = None + water_mark: Optional[bool] = None + + +class AspectRatio2(str, Enum): + field_16_9 = '16:9' + field_4_3 = '4:3' + field_1_1 = '1:1' + field_3_4 = '3:4' + field_9_16 = '9:16' + + +class PixverseTextVideoRequest(BaseModel): + aspect_ratio: AspectRatio2 + duration: Duration + model: Model1 + motion_mode: Optional[MotionMode] = None + negative_prompt: Optional[str] = None + prompt: str + quality: Quality1 + seed: Optional[int] = None + style: Optional[Style1] = None + template_id: Optional[int] = None + water_mark: Optional[bool] = None + + +class PixverseTransitionVideoRequest(BaseModel): + duration: Duration + first_frame_img: int + last_frame_img: int + model: Model1 + motion_mode: MotionMode + prompt: str + quality: Quality1 + seed: int + style: Optional[Style1] = None + template_id: Optional[int] = None + water_mark: Optional[bool] = None + + +class Resp1(BaseModel): + video_id: Optional[int] = None + + +class PixverseVideoResponse(BaseModel): + ErrCode: Optional[int] = None + ErrMsg: Optional[str] = None + Resp: Optional[Resp1] = None + + +class Status7(int, Enum): + integer_1 = 1 + integer_5 = 5 + integer_6 = 6 + integer_7 = 7 + integer_8 = 8 + + +class Resp2(BaseModel): + create_time: Optional[str] = None + id: Optional[int] = None + modify_time: Optional[str] = None + negative_prompt: Optional[str] = None + outputHeight: Optional[int] = None + outputWidth: Optional[int] = None + prompt: Optional[str] = None + resolution_ratio: Optional[int] = None + seed: Optional[int] = None + size: Optional[int] = None + status: Optional[Status7] = Field( + None, + description='Video generation status codes:\n* 1 - Generation successful\n* 5 - Generating\n* 6 - Deleted\n* 7 - Contents moderation failed\n* 8 - Generation failed\n', + ) + style: Optional[str] = None + url: Optional[str] = None + + +class PixverseVideoResultResponse(BaseModel): + ErrCode: Optional[int] = None + ErrMsg: Optional[str] = None + Resp: Optional[Resp2] = None + + +class RgbItem(RootModel[int]): + root: int = Field(..., ge=0, le=255) + + +class RGBColor(BaseModel): + rgb: List[RgbItem] = Field(..., max_length=3, min_length=3) + + +class GenerateSummary(str, Enum): + auto = 'auto' + concise = 'concise' + detailed = 'detailed' + + +class Summary(str, Enum): + auto = 'auto' + concise = 'concise' + detailed = 'detailed' + + +class ReasoningEffort(str, Enum): + low = 'low' + medium = 'medium' + high = 'high' + + +class Status8(str, Enum): + in_progress = 'in_progress' + completed = 'completed' + incomplete = 'incomplete' + + +class Type16(str, Enum): + summary_text = 'summary_text' + + +class SummaryItem(BaseModel): + text: str = Field( + ..., + description='A short summary of the reasoning used by the model when generating\nthe response.\n', + ) + type: Type16 = Field( + ..., description='The type of the object. Always `summary_text`.\n' + ) + + +class Type17(str, Enum): + reasoning = 'reasoning' + + +class ReasoningItem(BaseModel): + id: str = Field( + ..., description='The unique identifier of the reasoning content.\n' + ) + status: Optional[Status8] = Field( + None, + description='The status of the item. One of `in_progress`, `completed`, or\n`incomplete`. Populated when items are returned via API.\n', + ) + summary: List[SummaryItem] = Field(..., description='Reasoning text contents.\n') + type: Type17 = Field( + ..., description='The type of the object. Always `reasoning`.\n' + ) + + +class Controls(BaseModel): + artistic_level: Optional[int] = Field( + None, + description='Defines artistic tone of your image. At a simple level, the person looks straight at the camera in a static and clean style. Dynamic and eccentric levels introduce movement and creativity.', + ge=0, + le=5, + ) + background_color: Optional[RGBColor] = None + colors: Optional[List[RGBColor]] = Field( + None, description='An array of preferable colors' + ) + no_text: Optional[bool] = Field(None, description='Do not embed text layouts') + + +class RecraftImageGenerationRequest(BaseModel): + controls: Optional[Controls] = Field( + None, description='The controls for the generated image' + ) + model: str = Field( + ..., description='The model to use for generation (e.g., "recraftv3")' + ) + n: int = Field(..., description='The number of images to generate', ge=1, le=4) + prompt: str = Field( + ..., description='The text prompt describing the image to generate' + ) + size: str = Field( + ..., description='The size of the generated image (e.g., "1024x1024")' + ) + style: Optional[str] = Field( + None, + description='The style to apply to the generated image (e.g., "digital_illustration")', + ) + style_id: Optional[str] = Field( + None, + description='The style ID to apply to the generated image (e.g., "123e4567-e89b-12d3-a456-426614174000"). If style_id is provided, style should not be provided.', + ) + + +class Datum3(BaseModel): + image_id: Optional[str] = Field( + None, description='Unique identifier for the generated image' + ) + url: Optional[str] = Field(None, description='URL to access the generated image') + + +class RecraftImageGenerationResponse(BaseModel): + created: int = Field( + ..., description='Unix timestamp when the generation was created' + ) + credits: int = Field(..., description='Number of credits used for the generation') + data: List[Datum3] = Field(..., description='Array of generated image information') + + +class RenderingSpeed(str, Enum): + BALANCED = 'BALANCED' + TURBO = 'TURBO' + QUALITY = 'QUALITY' + + +class ResponseErrorCode(str, Enum): + server_error = 'server_error' + rate_limit_exceeded = 'rate_limit_exceeded' + invalid_prompt = 'invalid_prompt' + vector_store_timeout = 'vector_store_timeout' + invalid_image = 'invalid_image' + invalid_image_format = 'invalid_image_format' + invalid_base64_image = 'invalid_base64_image' + invalid_image_url = 'invalid_image_url' + image_too_large = 'image_too_large' + image_too_small = 'image_too_small' + image_parse_error = 'image_parse_error' + image_content_policy_violation = 'image_content_policy_violation' + invalid_image_mode = 'invalid_image_mode' + image_file_too_large = 'image_file_too_large' + unsupported_image_media_type = 'unsupported_image_media_type' + empty_image_file = 'empty_image_file' + failed_to_download_image = 'failed_to_download_image' + image_file_not_found = 'image_file_not_found' + + +class Type18(str, Enum): + json_object = 'json_object' + + +class ResponseFormatJsonObject(BaseModel): + type: Type18 = Field( + ..., + description='The type of response format being defined. Always `json_object`.', + ) + + +class ResponseFormatJsonSchemaSchema(BaseModel): + pass + model_config = ConfigDict( + extra='allow', + ) + + +class Type19(str, Enum): + text = 'text' + + +class ResponseFormatText(BaseModel): + type: Type19 = Field( + ..., description='The type of response format being defined. Always `text`.' + ) + + +class Truncation1(str, Enum): + auto = 'auto' + disabled = 'disabled' + + +class InputTokensDetails1(BaseModel): + cached_tokens: int = Field( + ..., + description='The number of tokens that were retrieved from the cache. \n[More on prompt caching](/docs/guides/prompt-caching).\n', + ) + + +class OutputTokensDetails(BaseModel): + reasoning_tokens: int = Field(..., description='The number of reasoning tokens.') + + +class ResponseUsage(BaseModel): + input_tokens: int = Field(..., description='The number of input tokens.') + input_tokens_details: InputTokensDetails1 = Field( + ..., description='A detailed breakdown of the input tokens.' + ) + output_tokens: int = Field(..., description='The number of output tokens.') + output_tokens_details: OutputTokensDetails = Field( + ..., description='A detailed breakdown of the output tokens.' + ) + total_tokens: int = Field(..., description='The total number of tokens used.') + + +class Rodin3DCheckStatusRequest(BaseModel): + subscription_key: str = Field( + ..., description='subscription from generate endpoint' + ) + + +class Rodin3DCheckStatusResponse(BaseModel): + pass + + +class Rodin3DDownloadRequest(BaseModel): + task_uuid: str = Field(..., description='Task UUID') + + +class RodinGenerateJobsData(BaseModel): + subscription_key: Optional[str] = Field(None, description='Subscription Key.') + uuids: Optional[List[str]] = Field(None, description='subjobs uuid.') + + +class RodinMaterialType(str, Enum): + PBR = 'PBR' + Shaded = 'Shaded' + + +class RodinMeshModeType(str, Enum): + Quad = 'Quad' + Raw = 'Raw' + + +class RodinQualityType(str, Enum): + extra_low = 'extra-low' + low = 'low' + medium = 'medium' + high = 'high' + + +class RodinResourceItem(BaseModel): + name: Optional[str] = Field(None, description='File name') + url: Optional[str] = Field(None, description='Download url') + + +class RodinTierType(str, Enum): + Regular = 'Regular' + Sketch = 'Sketch' + Detail = 'Detail' + Smooth = 'Smooth' + + +class RunwayAspectRatioEnum(str, Enum): + field_1280_720 = '1280:720' + field_720_1280 = '720:1280' + field_1104_832 = '1104:832' + field_832_1104 = '832:1104' + field_960_960 = '960:960' + field_1584_672 = '1584:672' + field_1280_768 = '1280:768' + field_768_1280 = '768:1280' + + +class RunwayDurationEnum(int, Enum): + integer_5 = 5 + integer_10 = 10 + + +class RunwayImageToVideoResponse(BaseModel): + id: Optional[str] = Field(None, description='Task ID') + + +class RunwayModelEnum(str, Enum): + gen4_turbo = 'gen4_turbo' + gen3a_turbo = 'gen3a_turbo' + + +class Position(str, Enum): + first = 'first' + last = 'last' + + +class RunwayPromptImageDetailedObject(BaseModel): + position: Position = Field( + ..., + description="The position of the image in the output video. 'last' is currently supported for gen3a_turbo only.", + ) + uri: str = Field( + ..., description='A HTTPS URL or data URI containing an encoded image.' + ) + + +class RunwayPromptImageObject( + RootModel[Union[str, List[RunwayPromptImageDetailedObject]]] +): + root: Union[str, List[RunwayPromptImageDetailedObject]] = Field( + ..., + description='Image(s) to use for the video generation. Can be a single URI or an array of image objects with positions.', + ) + + +class RunwayTaskStatusEnum(str, Enum): + SUCCEEDED = 'SUCCEEDED' + RUNNING = 'RUNNING' + FAILED = 'FAILED' + PENDING = 'PENDING' + CANCELLED = 'CANCELLED' + THROTTLED = 'THROTTLED' + + +class RunwayTaskStatusResponse(BaseModel): + createdAt: datetime = Field(..., description='Task creation timestamp') + id: str = Field(..., description='Task ID') + output: Optional[List[str]] = Field(None, description='Array of output video URLs') + progress: Optional[float] = Field( + None, + description='Float value between 0 and 1 representing the progress of the task. Only available if status is RUNNING.', + ge=0.0, + le=1.0, + ) + status: RunwayTaskStatusEnum + + +class RunwayTextToImageAspectRatioEnum(str, Enum): + field_1920_1080 = '1920:1080' + field_1080_1920 = '1080:1920' + field_1024_1024 = '1024:1024' + field_1360_768 = '1360:768' + field_1080_1080 = '1080:1080' + field_1168_880 = '1168:880' + field_1440_1080 = '1440:1080' + field_1080_1440 = '1080:1440' + field_1808_768 = '1808:768' + field_2112_912 = '2112:912' + +class Model4(str, Enum): + gen4_image = 'gen4_image' + + +class ReferenceImage(BaseModel): + uri: Optional[str] = Field( + None, description='A HTTPS URL or data URI containing an encoded image' + ) + + +class RunwayTextToImageRequest(BaseModel): + model: Model4 = Field(..., description='Model to use for generation') + promptText: str = Field( + ..., description='Text prompt for the image generation', max_length=1000 + ) + ratio: RunwayTextToImageAspectRatioEnum + referenceImages: Optional[List[ReferenceImage]] = Field( + None, description='Array of reference images to guide the generation' + ) + + +class RunwayTextToImageResponse(BaseModel): + id: Optional[str] = Field(None, description='Task ID') + + +class StabilityError(BaseModel): + errors: List[str] = Field( + ..., + description='One or more error messages indicating what went wrong.', + examples=[[{'some-field': 'is required'}]], + min_length=1, + ) + id: str = Field( + ..., + description='A unique identifier associated with this error. Please include this in any [support tickets](https://kb.stability.ai/knowledge-base/kb-tickets/new) you file, as it will greatly assist us in diagnosing the root cause of the problem.\n', + examples=['a1b2c3d4e5f6a1b2c3d4e5f6a1b2c3d4'], + min_length=1, + ) + name: str = Field( + ..., + description='Short-hand name for an error, useful for discriminating between errors with the same status code.', + examples=['bad_request'], + min_length=1, + ) + + +class Status9(str, Enum): + in_progress = 'in-progress' + + +class StabilityGetResultResponse202(BaseModel): + id: Optional[str] = Field( + None, description='The ID of the generation result.', examples=[1234567890] + ) + status: Optional[Status9] = None + + +class Type20(str, Enum): + json_schema = 'json_schema' + + +class TextResponseFormatJsonSchema(BaseModel): + description: Optional[str] = Field( + None, + description='A description of what the response format is for, used by the model to\ndetermine how to respond in the format.\n', + ) + name: str = Field( + ..., + description='The name of the response format. Must be a-z, A-Z, 0-9, or contain\nunderscores and dashes, with a maximum length of 64.\n', + ) + schema_: ResponseFormatJsonSchemaSchema = Field(..., alias='schema') + strict: Optional[bool] = Field( + False, + description='Whether to enable strict schema adherence when generating the output.\nIf set to true, the model will always follow the exact schema defined\nin the `schema` field. Only a subset of JSON Schema is supported when\n`strict` is `true`. To learn more, read the [Structured Outputs\nguide](/docs/guides/structured-outputs).\n', + ) + type: Type20 = Field( + ..., + description='The type of response format being defined. Always `json_schema`.', + ) + + +class Type21(str, Enum): + function = 'function' + + +class ToolChoiceFunction(BaseModel): + name: str = Field(..., description='The name of the function to call.') + type: Type21 = Field( + ..., description='For function calling, the type is always `function`.' + ) + + +class ToolChoiceOptions(str, Enum): + none = 'none' + auto = 'auto' + required = 'required' + + +class Type22(str, Enum): + file_search = 'file_search' + web_search_preview = 'web_search_preview' + computer_use_preview = 'computer_use_preview' + web_search_preview_2025_03_11 = 'web_search_preview_2025_03_11' + + +class ToolChoiceTypes(BaseModel): + type: Type22 = Field( + ..., + description='The type of hosted tool the model should to use. Learn more about\n[built-in tools](/docs/guides/tools).\n\nAllowed values are:\n- `file_search`\n- `web_search_preview`\n- `computer_use_preview`\n', + ) + + +class TripoAnimation(str, Enum): + preset_idle = 'preset:idle' + preset_walk = 'preset:walk' + preset_climb = 'preset:climb' + preset_jump = 'preset:jump' + preset_run = 'preset:run' + preset_slash = 'preset:slash' + preset_shoot = 'preset:shoot' + preset_hurt = 'preset:hurt' + preset_fall = 'preset:fall' + preset_turn = 'preset:turn' + + +class TripoBalance(BaseModel): + balance: float + frozen: float + + +class TripoConvertFormat(str, Enum): + GLTF = 'GLTF' + USDZ = 'USDZ' + FBX = 'FBX' + OBJ = 'OBJ' + STL = 'STL' + field_3MF = '3MF' + + +class Code(int, Enum): + integer_1001 = 1001 + integer_2000 = 2000 + integer_2001 = 2001 + integer_2002 = 2002 + integer_2003 = 2003 + integer_2004 = 2004 + integer_2006 = 2006 + integer_2007 = 2007 + integer_2008 = 2008 + integer_2010 = 2010 + + +class TripoErrorResponse(BaseModel): + code: Code + message: str + suggestion: str + + +class TripoImageToModel(str, Enum): + image_to_model = 'image_to_model' + + +class TripoModelStyle(str, Enum): + person_person2cartoon = 'person:person2cartoon' + animal_venom = 'animal:venom' + object_clay = 'object:clay' + object_steampunk = 'object:steampunk' + object_christmas = 'object:christmas' + object_barbie = 'object:barbie' + gold = 'gold' + ancient_bronze = 'ancient_bronze' + + +class TripoModelVersion(str, Enum): + V2_5 = 'v2.5-20250123' + V2_0 = 'v2.0-20240919' + V1_4 = 'v1.4-20240625' + + +class TripoMultiviewMode(str, Enum): + LEFT = 'LEFT' + RIGHT = 'RIGHT' + + +class TripoMultiviewToModel(str, Enum): + multiview_to_model = 'multiview_to_model' + + +class TripoOrientation(str, Enum): + align_image = 'align_image' + default = 'default' + + +class TripoResponseSuccessCode(RootModel[int]): + root: int = Field( + ..., + description='Standard success code for Tripo API responses. Typically 0 for success.', + examples=[0], + ) + + +class TripoSpec(str, Enum): + mixamo = 'mixamo' + tripo = 'tripo' + + +class TripoStandardFormat(str, Enum): + glb = 'glb' + fbx = 'fbx' + + +class TripoStylizeOptions(str, Enum): + lego = 'lego' + voxel = 'voxel' + voronoi = 'voronoi' + minecraft = 'minecraft' + + +class Code1(int, Enum): + integer_0 = 0 + + +class Data8(BaseModel): + task_id: str = Field(..., description='used for getTask') + + +class TripoSuccessTask(BaseModel): + code: Code1 + data: Data8 + + +class Topology(str, Enum): + bip = 'bip' + quad = 'quad' + + +class Output(BaseModel): + base_model: Optional[str] = None + model: Optional[str] = None + pbr_model: Optional[str] = None + rendered_image: Optional[str] = None + riggable: Optional[bool] = None + topology: Optional[Topology] = None + + +class Status10(str, Enum): + queued = 'queued' + running = 'running' + success = 'success' + failed = 'failed' + cancelled = 'cancelled' + unknown = 'unknown' + banned = 'banned' + expired = 'expired' + + +class TripoTask(BaseModel): + create_time: int + input: Dict[str, Any] + output: Output + progress: int = Field(..., ge=0, le=100) + status: Status10 + task_id: str + type: str + + +class TripoTextToModel(str, Enum): + text_to_model = 'text_to_model' + + +class TripoTextureAlignment(str, Enum): + original_image = 'original_image' + geometry = 'geometry' + + +class TripoTextureFormat(str, Enum): + BMP = 'BMP' + DPX = 'DPX' + HDR = 'HDR' + JPEG = 'JPEG' + OPEN_EXR = 'OPEN_EXR' + PNG = 'PNG' + TARGA = 'TARGA' + TIFF = 'TIFF' + WEBP = 'WEBP' + + +class TripoTextureQuality(str, Enum): + standard = 'standard' + detailed = 'detailed' + + +class TripoTopology(str, Enum): + bip = 'bip' + quad = 'quad' + + +class TripoTypeAnimatePrerigcheck(str, Enum): + animate_prerigcheck = 'animate_prerigcheck' + + +class TripoTypeAnimateRetarget(str, Enum): + animate_retarget = 'animate_retarget' + + +class TripoTypeAnimateRig(str, Enum): + animate_rig = 'animate_rig' + + +class TripoTypeConvertModel(str, Enum): + convert_model = 'convert_model' + + +class TripoTypeRefineModel(str, Enum): + refine_model = 'refine_model' + + +class TripoTypeStylizeModel(str, Enum): + stylize_model = 'stylize_model' + + +class TripoTypeTextureModel(str, Enum): + texture_model = 'texture_model' + + +class Veo2GenVidPollRequest(BaseModel): + operationName: str = Field( + ..., + description='Full operation name (from predict response)', + examples=[ + 'projects/PROJECT_ID/locations/us-central1/publishers/google/models/MODEL_ID/operations/OPERATION_ID' + ], + ) + + +class Error(BaseModel): + code: Optional[int] = Field(None, description='Error code') + message: Optional[str] = Field(None, description='Error message') + + +class Video(BaseModel): + bytesBase64Encoded: Optional[str] = Field( + None, description='Base64-encoded video content' + ) + gcsUri: Optional[str] = Field(None, description='Cloud Storage URI of the video') + mimeType: Optional[str] = Field(None, description='Video MIME type') + + +class Response(BaseModel): + field_type: Optional[str] = Field( + None, + alias='@type', + examples=[ + 'type.googleapis.com/cloud.ai.large_models.vision.GenerateVideoResponse' + ], + ) + raiMediaFilteredCount: Optional[int] = Field( + None, description='Count of media filtered by responsible AI policies' + ) + raiMediaFilteredReasons: Optional[List[str]] = Field( + None, description='Reasons why media was filtered by responsible AI policies' + ) + videos: Optional[List[Video]] = None + + +class Veo2GenVidPollResponse(BaseModel): + done: Optional[bool] = None + error: Optional[Error] = Field( + None, description='Error details if operation failed' + ) + name: Optional[str] = None + response: Optional[Response] = Field( + None, description='The actual prediction response if done is true' + ) + + +class Image(BaseModel): + bytesBase64Encoded: str + gcsUri: Optional[str] = None + mimeType: Optional[str] = None + + +class Image1(BaseModel): + bytesBase64Encoded: Optional[str] = None + gcsUri: str + mimeType: Optional[str] = None + + +class Instance(BaseModel): + image: Optional[Union[Image, Image1]] = Field( + None, description='Optional image to guide video generation' + ) + prompt: str = Field(..., description='Text description of the video') + + +class PersonGeneration1(str, Enum): + ALLOW = 'ALLOW' + BLOCK = 'BLOCK' + + +class Parameters(BaseModel): + aspectRatio: Optional[str] = Field(None, examples=['16:9']) + durationSeconds: Optional[int] = None + enhancePrompt: Optional[bool] = None + negativePrompt: Optional[str] = None + personGeneration: Optional[PersonGeneration1] = None + sampleCount: Optional[int] = None + seed: Optional[int] = None + storageUri: Optional[str] = Field( + None, description='Optional Cloud Storage URI to upload the video' + ) + + +class Veo2GenVidRequest(BaseModel): + instances: Optional[List[Instance]] = None + parameters: Optional[Parameters] = None + + +class Veo2GenVidResponse(BaseModel): + name: str = Field( + ..., + description='Operation resource name', + examples=[ + 'projects/PROJECT_ID/locations/us-central1/publishers/google/models/MODEL_ID/operations/a1b07c8e-7b5a-4aba-bb34-3e1ccb8afcc8' + ], + ) + + +class SearchContextSize(str, Enum): + low = 'low' + medium = 'medium' + high = 'high' + + +class Type23(str, Enum): + web_search_preview = 'web_search_preview' + web_search_preview_2025_03_11 = 'web_search_preview_2025_03_11' + + +class WebSearchPreviewTool(BaseModel): + search_context_size: Optional[SearchContextSize] = Field( + None, + description='High level guidance for the amount of context window space to use for the search. One of `low`, `medium`, or `high`. `medium` is the default.', + ) + type: Literal['WebSearchPreviewTool'] = Field( + ..., + description='The type of the web search tool. One of `web_search_preview` or `web_search_preview_2025_03_11`.', + ) + + +class Status11(str, Enum): + in_progress = 'in_progress' + searching = 'searching' + completed = 'completed' + failed = 'failed' + + +class Type24(str, Enum): + web_search_call = 'web_search_call' + + +class WebSearchToolCall(BaseModel): + id: str = Field(..., description='The unique ID of the web search tool call.\n') + status: Status11 = Field( + ..., description='The status of the web search tool call.\n' + ) + type: Type24 = Field( + ..., + description='The type of the web search tool call. Always `web_search_call`.\n', + ) + + +class CreateModelResponseProperties(ModelResponseProperties): + pass + + +class GeminiInlineData(BaseModel): + data: Optional[str] = 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\n', + ) + mimeType: Optional[GeminiMimeType] = None + + +class GeminiPart(BaseModel): + inlineData: Optional[GeminiInlineData] = None + text: Optional[str] = Field( + None, + description='A text prompt or code snippet.', + examples=['Write a story about a robot learning to paint'], + ) + + +class GeminiPromptFeedback(BaseModel): + blockReason: Optional[str] = None + blockReasonMessage: Optional[str] = None + safetyRatings: Optional[List[GeminiSafetyRating]] = None + + +class GeminiSafetySetting(BaseModel): + category: GeminiSafetyCategory + threshold: GeminiSafetyThreshold + + +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. For limits on the inputs, such as the maximum number of tokens or the number of images, see the model specifications on the Google models page.\n', + ) + role: Role1 = Field( + ..., + description='The identity of the entity that creates the message. The following values are supported: user: This indicates that the message is sent by a real person, typically a user-generated message. model: This indicates that the message is generated by the model. The model value is used to insert messages from the model into the conversation during multi-turn conversations. For non-multi-turn conversations, this field can be left blank or unset.\n', + examples=['user'], + ) + + +class IdeogramV3EditRequest(BaseModel): + color_palette: Optional[IdeogramColorPalette] = None + image: Optional[StrictBytes] = Field( + None, + description='The image being edited (max size 10MB); only JPEG, WebP and PNG formats are supported at this time.', + ) + magic_prompt: Optional[str] = Field( + None, + description='Determine if MagicPrompt should be used in generating the request or not.', + ) + mask: Optional[StrictBytes] = Field( + None, + description='A black and white image of the same size as the image being edited (max size 10MB). Black regions in the mask should match up with the regions of the image that you would like to edit; only JPEG, WebP and PNG formats are supported at this time.', + ) + num_images: Optional[int] = Field( + None, description='The number of images to generate.' + ) + prompt: str = Field( + ..., description='The prompt used to describe the edited result.' + ) + rendering_speed: RenderingSpeed + seed: Optional[int] = Field( + None, description='Random seed. Set for reproducible generation.' + ) + style_codes: Optional[List[StyleCode]] = Field( + None, + description='A list of 8 character hexadecimal codes representing the style of the image. Cannot be used in conjunction with style_reference_images or style_type.', + ) + style_reference_images: Optional[List[StrictBytes]] = Field( + None, + description='A set of images to use as style references (maximum total size 10MB across all style references). The images should be in JPEG, PNG or WebP format.', + ) + + +class IdeogramV3Request(BaseModel): + aspect_ratio: Optional[str] = Field( + None, description='Aspect ratio in format WxH', examples=['1x3'] + ) + color_palette: Optional[ColorPalette] = None + magic_prompt: Optional[MagicPrompt2] = Field( + None, description='Whether to enable magic prompt enhancement' + ) + negative_prompt: Optional[str] = Field( + None, description='Text prompt specifying what to avoid in the generation' + ) + num_images: Optional[int] = Field( + None, description='Number of images to generate', ge=1 + ) + prompt: str = Field(..., description='The text prompt for image generation') + rendering_speed: RenderingSpeed + resolution: Optional[str] = Field( + None, description='Image resolution in format WxH', examples=['1280x800'] + ) + seed: Optional[int] = Field( + None, description='Seed value for reproducible generation' + ) + style_codes: Optional[List[StyleCode]] = Field( + None, description='Array of style codes in hexadecimal format' + ) + style_reference_images: Optional[List[str]] = Field( + None, description='Array of reference image URLs or identifiers' + ) + style_type: Optional[StyleType1] = Field( + None, description='The type of style to apply' + ) + + +class ImagenGenerateImageResponse(BaseModel): + predictions: Optional[List[ImagenImagePrediction]] = None + + +class ImagenImageGenerationParameters(BaseModel): + addWatermark: Optional[bool] = None + aspectRatio: Optional[AspectRatio] = None + enhancePrompt: Optional[bool] = None + includeRaiReason: Optional[bool] = None + includeSafetyAttributes: Optional[bool] = None + outputOptions: Optional[ImagenOutputOptions] = None + personGeneration: Optional[PersonGeneration] = None + safetySetting: Optional[SafetySetting] = None + sampleCount: Optional[int] = Field(None, ge=1, le=4) + seed: Optional[int] = None + storageUri: Optional[AnyUrl] = None + + +class InputContent( + RootModel[Union[InputTextContent, InputImageContent, InputFileContent]] +): + root: Union[InputTextContent, InputImageContent, InputFileContent] + + +class InputMessageContentList(RootModel[List[InputContent]]): + root: List[InputContent] = Field( + ..., + description='A list of one or many input items to the model, containing different content \ntypes.\n', + title='Input item content list', + ) + + +class KlingCameraControl(BaseModel): + config: Optional[KlingCameraConfig] = None + type: Optional[KlingCameraControlType] = None + + +class KlingDualCharacterEffectInput(BaseModel): + duration: KlingVideoGenDuration + images: KlingDualCharacterImages + mode: Optional[KlingVideoGenMode] = 'std' + model_name: Optional[KlingCharacterEffectModelName] = 'kling-v1' + + +class KlingImage2VideoRequest(BaseModel): + aspect_ratio: Optional[KlingVideoGenAspectRatio] = '16:9' + callback_url: Optional[AnyUrl] = Field( + None, + description='The callback notification address. Server will notify when the task status changes.', + ) + camera_control: Optional[KlingCameraControl] = None + cfg_scale: Optional[KlingVideoGenCfgScale] = Field( + default_factory=lambda: KlingVideoGenCfgScale.model_validate(0.5) + ) + duration: Optional[KlingVideoGenDuration] = '5' + dynamic_masks: Optional[List[DynamicMask]] = Field( + None, + description='Dynamic Brush Configuration List (up to 6 groups). For 5-second videos, trajectory length must not exceed 77 coordinates.', + ) + external_task_id: Optional[str] = Field( + None, + description='Customized Task ID. Must be unique within a single user account.', + ) + image: Optional[str] = Field( + None, + description='Reference Image - URL or Base64 encoded string, cannot exceed 10MB, resolution not less than 300*300px, aspect ratio between 1:2.5 ~ 2.5:1. Base64 should not include data:image prefix.', + ) + image_tail: Optional[str] = Field( + None, + description='Reference Image - End frame control. URL or Base64 encoded string, cannot exceed 10MB, resolution not less than 300*300px. Base64 should not include data:image prefix.', + ) + mode: Optional[KlingVideoGenMode] = 'std' + model_name: Optional[KlingVideoGenModelName] = 'kling-v2-master' + negative_prompt: Optional[str] = Field( + None, description='Negative text prompt', max_length=2500 + ) + prompt: Optional[str] = Field( + None, description='Positive text prompt', max_length=2500 + ) + static_mask: Optional[str] = Field( + None, + description='Static Brush Application Area (Mask image created by users using the motion brush). The aspect ratio must match the input image.', + ) + + +class TaskResult(BaseModel): + videos: Optional[List[KlingVideoResult]] = None + + +class Data(BaseModel): + created_at: Optional[int] = Field(None, description='Task creation time') + task_id: Optional[str] = Field(None, description='Task ID') + task_info: Optional[TaskInfo] = None + task_result: Optional[TaskResult] = None + task_status: Optional[KlingTaskStatus] = None + updated_at: Optional[int] = Field(None, description='Task update time') + + +class KlingImage2VideoResponse(BaseModel): + code: Optional[int] = Field(None, description='Error code') + data: Optional[Data] = None + message: Optional[str] = Field(None, description='Error message') + request_id: Optional[str] = Field(None, description='Request ID') + + +class TaskResult1(BaseModel): + images: Optional[List[KlingImageResult]] = None + + +class Data1(BaseModel): + created_at: Optional[int] = Field(None, description='Task creation time') + task_id: Optional[str] = Field(None, description='Task ID') + task_result: Optional[TaskResult1] = None + task_status: Optional[KlingTaskStatus] = None + task_status_msg: Optional[str] = Field(None, description='Task status information') + updated_at: Optional[int] = Field(None, description='Task update time') + + +class KlingImageGenerationsResponse(BaseModel): + code: Optional[int] = Field(None, description='Error code') + data: Optional[Data1] = None + message: Optional[str] = Field(None, description='Error message') + request_id: Optional[str] = Field(None, description='Request ID') + + +class KlingLipSyncInputObject(BaseModel): + audio_file: Optional[str] = Field( + None, + description='Local Path of Audio File. Supported formats: .mp3/.wav/.m4a/.aac, maximum file size of 5MB. Base64 code.', + ) + audio_type: Optional[KlingAudioUploadType] = None + audio_url: Optional[str] = Field( + None, + description='Audio File Download URL. Supported formats: .mp3/.wav/.m4a/.aac, maximum file size of 5MB.', + ) + mode: KlingLipSyncMode + text: Optional[str] = Field( + None, + description='Text Content for Lip-Sync Video Generation. Required when mode is text2video. Maximum length is 120 characters.', + ) + video_id: Optional[str] = Field( + None, + description='The ID of the video generated by Kling AI. Only supports 5-second and 10-second videos generated within the last 30 days.', + ) + video_url: Optional[str] = Field( + None, + description='Get link for uploaded video. Video files support .mp4/.mov, file size does not exceed 100MB, video length between 2-10s.', + ) + voice_id: Optional[str] = Field( + None, + description='Voice ID. Required when mode is text2video. The system offers a variety of voice options to choose from.', + ) + voice_language: Optional[KlingLipSyncVoiceLanguage] = 'en' + voice_speed: Optional[float] = Field( + 1, + description='Speech Rate. Valid range: 0.8~2.0, accurate to one decimal place.', + ge=0.8, + le=2.0, + ) + + +class KlingLipSyncRequest(BaseModel): + callback_url: Optional[AnyUrl] = Field( + None, + description='The callback notification address. Server will notify when the task status changes.', + ) + input: KlingLipSyncInputObject + + +class TaskResult2(BaseModel): + videos: Optional[List[KlingVideoResult]] = None + + +class Data2(BaseModel): + created_at: Optional[int] = Field(None, description='Task creation time') + task_id: Optional[str] = Field(None, description='Task ID') + task_info: Optional[TaskInfo] = None + task_result: Optional[TaskResult2] = None + task_status: Optional[KlingTaskStatus] = None + updated_at: Optional[int] = Field(None, description='Task update time') + + +class KlingLipSyncResponse(BaseModel): + code: Optional[int] = Field(None, description='Error code') + data: Optional[Data2] = None + message: Optional[str] = Field(None, description='Error message') + request_id: Optional[str] = Field(None, description='Request ID') + + +class KlingSingleImageEffectInput(BaseModel): + duration: KlingSingleImageEffectDuration + image: str = Field( + ..., + description='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.', + ) + model_name: KlingSingleImageEffectModelName + + +class KlingText2VideoRequest(BaseModel): + aspect_ratio: Optional[KlingVideoGenAspectRatio] = '16:9' + callback_url: Optional[AnyUrl] = Field( + None, description='The callback notification address' + ) + camera_control: Optional[KlingCameraControl] = None + cfg_scale: Optional[KlingVideoGenCfgScale] = Field( + default_factory=lambda: KlingVideoGenCfgScale.model_validate(0.5) + ) + duration: Optional[KlingVideoGenDuration] = '5' + external_task_id: Optional[str] = Field(None, description='Customized Task ID') + mode: Optional[KlingVideoGenMode] = 'std' + model_name: Optional[KlingTextToVideoModelName] = 'kling-v1' + negative_prompt: Optional[str] = Field( + None, description='Negative text prompt', max_length=2500 + ) + prompt: Optional[str] = Field( + None, description='Positive text prompt', max_length=2500 + ) + + +class Data4(BaseModel): + created_at: Optional[int] = Field(None, description='Task creation time') + task_id: Optional[str] = Field(None, description='Task ID') + task_info: Optional[TaskInfo] = None + task_result: Optional[TaskResult2] = None + task_status: Optional[KlingTaskStatus] = None + updated_at: Optional[int] = Field(None, description='Task update time') + + +class KlingText2VideoResponse(BaseModel): + code: Optional[int] = Field(None, description='Error code') + data: Optional[Data4] = None + message: Optional[str] = Field(None, description='Error message') + request_id: Optional[str] = Field(None, description='Request ID') + + +class KlingVideoEffectsInput( + RootModel[Union[KlingSingleImageEffectInput, KlingDualCharacterEffectInput]] +): + root: Union[KlingSingleImageEffectInput, KlingDualCharacterEffectInput] + + +class KlingVideoEffectsRequest(BaseModel): + callback_url: Optional[AnyUrl] = Field( + None, + description='The callback notification address for the result of this task.', + ) + effect_scene: Union[KlingDualCharacterEffectsScene, KlingSingleImageEffectsScene] + external_task_id: Optional[str] = Field( + None, + description='Customized Task ID. Must be unique within a single user account.', + ) + input: KlingVideoEffectsInput + + +class Data5(BaseModel): + created_at: Optional[int] = Field(None, description='Task creation time') + task_id: Optional[str] = Field(None, description='Task ID') + task_info: Optional[TaskInfo] = None + task_result: Optional[TaskResult2] = None + task_status: Optional[KlingTaskStatus] = None + updated_at: Optional[int] = Field(None, description='Task update time') + + +class KlingVideoEffectsResponse(BaseModel): + code: Optional[int] = Field(None, description='Error code') + data: Optional[Data5] = None + message: Optional[str] = Field(None, description='Error message') + request_id: Optional[str] = Field(None, description='Request ID') + + +class KlingVideoExtendRequest(BaseModel): + callback_url: Optional[AnyUrl] = Field( + None, + description='The callback notification address. Server will notify when the task status changes.', + ) + cfg_scale: Optional[KlingVideoGenCfgScale] = Field( + default_factory=lambda: KlingVideoGenCfgScale.model_validate(0.5) + ) + negative_prompt: Optional[str] = Field( + None, + description='Negative text prompt for elements to avoid in the extended video', + max_length=2500, + ) + prompt: Optional[str] = Field( + None, + description='Positive text prompt for guiding the video extension', + max_length=2500, + ) + video_id: Optional[str] = Field( + None, + description='The ID of the video to be extended. Supports videos generated by text-to-video, image-to-video, and previous video extension operations. Cannot exceed 3 minutes total duration after extension.', + ) + + +class Data6(BaseModel): + created_at: Optional[int] = Field(None, description='Task creation time') + task_id: Optional[str] = Field(None, description='Task ID') + task_info: Optional[TaskInfo] = None + task_result: Optional[TaskResult2] = None + task_status: Optional[KlingTaskStatus] = None + updated_at: Optional[int] = Field(None, description='Task update time') + + +class KlingVideoExtendResponse(BaseModel): + code: Optional[int] = Field(None, description='Error code') + data: Optional[Data6] = None + message: Optional[str] = Field(None, description='Error message') + request_id: Optional[str] = Field(None, description='Request ID') + + +class LumaGenerationRequest(BaseModel): + aspect_ratio: LumaAspectRatio + callback_url: Optional[AnyUrl] = Field( + None, + description='The callback URL of the generation, a POST request with Generation object will be sent to the callback URL when the generation is dreaming, completed, or failed', + ) + duration: LumaVideoModelOutputDuration + generation_type: Optional[GenerationType1] = 'video' + keyframes: Optional[LumaKeyframes] = None + loop: Optional[bool] = Field(None, description='Whether to loop the video') + model: LumaVideoModel + prompt: str = Field(..., description='The prompt of the generation') + resolution: LumaVideoModelOutputResolution + + +class CharacterRef(BaseModel): + identity0: Optional[LumaImageIdentity] = None + + +class LumaImageGenerationRequest(BaseModel): + aspect_ratio: Optional[LumaAspectRatio] = '16:9' + callback_url: Optional[AnyUrl] = Field( + None, description='The callback URL for the generation' + ) + character_ref: Optional[CharacterRef] = None + generation_type: Optional[GenerationType2] = 'image' + image_ref: Optional[List[LumaImageRef]] = None + model: Optional[LumaImageModel] = 'photon-1' + modify_image_ref: Optional[LumaModifyImageRef] = None + prompt: Optional[str] = Field(None, description='The prompt of the generation') + style_ref: Optional[List[LumaImageRef]] = None + + +class LumaUpscaleVideoGenerationRequest(BaseModel): + callback_url: Optional[AnyUrl] = Field( + None, description='The callback URL for the upscale' + ) + generation_type: Optional[GenerationType3] = 'upscale_video' + resolution: Optional[LumaVideoModelOutputResolution] = None + + +class OutputContent(RootModel[Union[OutputTextContent, OutputAudioContent]]): + root: Union[OutputTextContent, OutputAudioContent] + + +class OutputMessage(BaseModel): + content: List[OutputContent] = Field(..., description='The content of the message') + role: Role4 = Field(..., description='The role of the message') + type: Type14 = Field(..., description='The type of output item') + + +class PikaBodyGenerate22I2vGenerate22I2vPost(BaseModel): + duration: Optional[PikaDurationEnum] = 5 + image: Optional[StrictBytes] = Field(None, title='Image') + negativePrompt: Optional[str] = Field(None, title='Negativeprompt') + promptText: Optional[str] = Field(None, title='Prompttext') + resolution: Optional[PikaResolutionEnum] = '1080p' + seed: Optional[int] = Field(None, title='Seed') + + +class PikaBodyGenerate22KeyframeGenerate22PikaframesPost(BaseModel): + duration: Optional[int] = Field(None, ge=5, le=10, title='Duration') + keyFrames: Optional[List[StrictBytes]] = Field( + None, description='Array of keyframe images', title='Keyframes' + ) + negativePrompt: Optional[str] = Field(None, title='Negativeprompt') + promptText: str = Field(..., title='Prompttext') + resolution: Optional[PikaResolutionEnum] = '1080p' + seed: Optional[int] = Field(None, title='Seed') + + +class PikaBodyGenerate22T2vGenerate22T2vPost(BaseModel): + aspectRatio: Optional[float] = Field( + 1.7777777777777777, + description='Aspect ratio (width / height)', + ge=0.4, + le=2.5, + title='Aspectratio', + ) + duration: Optional[PikaDurationEnum] = 5 + negativePrompt: Optional[str] = Field(None, title='Negativeprompt') + promptText: str = Field(..., title='Prompttext') + resolution: Optional[PikaResolutionEnum] = '1080p' + seed: Optional[int] = Field(None, title='Seed') + + +class PikaBodyGeneratePikaffectsGeneratePikaffectsPost(BaseModel): + image: Optional[StrictBytes] = Field(None, title='Image') + negativePrompt: Optional[str] = Field(None, title='Negativeprompt') + pikaffect: Optional[Pikaffect] = None + promptText: Optional[str] = Field(None, title='Prompttext') + seed: Optional[int] = Field(None, title='Seed') + + +class PikaHTTPValidationError(BaseModel): + detail: Optional[List[PikaValidationError]] = Field(None, title='Detail') + + +class Reasoning(BaseModel): + effort: Optional[ReasoningEffort] = 'medium' + generate_summary: Optional[GenerateSummary] = Field( + None, + description="**Deprecated:** use `summary` instead.\n\nA summary of the reasoning performed by the model. This can be\nuseful for debugging and understanding the model's reasoning process.\nOne of `auto`, `concise`, or `detailed`.\n", + ) + summary: Optional[Summary] = Field( + None, + description="A summary of the reasoning performed by the model. This can be\nuseful for debugging and understanding the model's reasoning process.\nOne of `auto`, `concise`, or `detailed`.\n", + ) + + +class ResponseError(BaseModel): + code: ResponseErrorCode + message: str = Field(..., description='A human-readable description of the error.') + + +class Rodin3DDownloadResponse(BaseModel): + list: Optional[RodinResourceItem] = None + + +class Rodin3DGenerateRequest(BaseModel): + images: str = Field(..., description='The reference images to generate 3D Assets.') + material: Optional[RodinMaterialType] = None + mesh_mode: Optional[RodinMeshModeType] = None + quality: Optional[RodinQualityType] = None + seed: Optional[int] = Field(None, description='Seed.') + tier: Optional[RodinTierType] = None + + +class Rodin3DGenerateResponse(BaseModel): + jobs: Optional[RodinGenerateJobsData] = None + message: Optional[str] = Field(None, description='message') + prompt: Optional[str] = Field(None, description='prompt') + submit_time: Optional[str] = Field(None, description='Time') + uuid: Optional[str] = Field(None, description='Task UUID') + + +class RunwayImageToVideoRequest(BaseModel): + duration: RunwayDurationEnum + model: RunwayModelEnum + promptImage: RunwayPromptImageObject + promptText: Optional[str] = Field( + None, description='Text prompt for the generation', max_length=1000 + ) + ratio: RunwayAspectRatioEnum + seed: int = Field( + ..., description='Random seed for generation', ge=0, le=4294967295 + ) + + +class TextResponseFormatConfiguration( + RootModel[ + Union[ + ResponseFormatText, TextResponseFormatJsonSchema, ResponseFormatJsonObject + ] + ] +): + root: Union[ + ResponseFormatText, TextResponseFormatJsonSchema, ResponseFormatJsonObject + ] = Field( + ..., + description='An object specifying the format that the model must output.\n\nConfiguring `{ "type": "json_schema" }` enables Structured Outputs, \nwhich ensures the model will match your supplied JSON schema. Learn more in the \n[Structured Outputs guide](/docs/guides/structured-outputs).\n\nThe default format is `{ "type": "text" }` with no additional options.\n\n**Not recommended for gpt-4o and newer models:**\n\nSetting to `{ "type": "json_object" }` enables the older JSON mode, which\nensures the message the model generates is valid JSON. Using `json_schema`\nis preferred for models that support it.\n', + ) + + +class Tool( + RootModel[ + Union[ + FileSearchTool, FunctionTool, WebSearchPreviewTool, ComputerUsePreviewTool + ] + ] +): + root: Union[ + FileSearchTool, FunctionTool, WebSearchPreviewTool, ComputerUsePreviewTool + ] = Field(..., discriminator='type') + + +class EasyInputMessage(BaseModel): + content: Union[str, InputMessageContentList] = Field( + ..., + description='Text, image, or audio input to the model, used to generate a response.\nCan also contain previous assistant responses.\n', + ) + role: Role = Field( + ..., + description='The role of the message input. One of `user`, `assistant`, `system`, or\n`developer`.\n', + ) + type: Optional[Type2] = Field( + None, description='The type of the message input. Always `message`.\n' + ) + + +class GeminiContent(BaseModel): + parts: List[GeminiPart] + role: Role1 = Field(..., examples=['user']) + + +class GeminiGenerateContentRequest(BaseModel): + contents: List[GeminiContent] + generationConfig: Optional[GeminiGenerationConfig] = None + safetySettings: Optional[List[GeminiSafetySetting]] = None + systemInstruction: Optional[GeminiSystemInstructionContent] = None + tools: Optional[List[GeminiTool]] = None + videoMetadata: Optional[GeminiVideoMetadata] = None + + +class ImagenGenerateImageRequest(BaseModel): + instances: List[ImagenImageGenerationInstance] + parameters: ImagenImageGenerationParameters + + +class InputMessage(BaseModel): + content: Optional[InputMessageContentList] = None + role: Optional[Role3] = None + status: Optional[Status2] = None + type: Optional[Type9] = None + + +class Item( + RootModel[ + Union[ + InputMessage, + OutputMessage, + FileSearchToolCall, + ComputerToolCall, + WebSearchToolCall, + FunctionToolCall, + ReasoningItem, + ] + ] +): + root: Union[ + InputMessage, + OutputMessage, + FileSearchToolCall, + ComputerToolCall, + WebSearchToolCall, + FunctionToolCall, + ReasoningItem, + ] = Field(..., description='Content item used to generate a response.\n') + + +class LumaGeneration(BaseModel): + assets: Optional[LumaAssets] = None + created_at: Optional[datetime] = Field( + None, description='The date and time when the generation was created' + ) + failure_reason: Optional[str] = Field( + None, description='The reason for the state of the generation' + ) + generation_type: Optional[LumaGenerationType] = None + id: Optional[UUID] = Field(None, description='The ID of the generation') + model: Optional[str] = Field(None, description='The model used for the generation') + request: Optional[ + Union[ + LumaGenerationRequest, + LumaImageGenerationRequest, + LumaUpscaleVideoGenerationRequest, + LumaAudioGenerationRequest, + ] + ] = Field(None, description='The request of the generation') + state: Optional[LumaState] = None + + +class OutputItem( + RootModel[ + Union[ + OutputMessage, + FileSearchToolCall, + FunctionToolCall, + WebSearchToolCall, + ComputerToolCall, + ReasoningItem, + ] + ] +): + root: Union[ + OutputMessage, + FileSearchToolCall, + FunctionToolCall, + WebSearchToolCall, + ComputerToolCall, + ReasoningItem, + ] + + +class Text(BaseModel): + format: Optional[TextResponseFormatConfiguration] = None + + +class ResponseProperties(BaseModel): + instructions: Optional[str] = Field( + None, + description="Inserts a system (or developer) message as the first item in the model's context.\n\nWhen using along with `previous_response_id`, the instructions from a previous\nresponse will not be carried over to the next response. This makes it simple\nto swap out system (or developer) messages in new responses.\n", + ) + max_output_tokens: Optional[int] = Field( + None, + description='An upper bound for the number of tokens that can be generated for a response, including visible output tokens and [reasoning tokens](/docs/guides/reasoning).\n', + ) + model: Optional[OpenAIModels] = None + previous_response_id: Optional[str] = Field( + None, + description='The unique ID of the previous response to the model. Use this to\ncreate multi-turn conversations. Learn more about \n[conversation state](/docs/guides/conversation-state).\n', + ) + reasoning: Optional[Reasoning] = None + text: Optional[Text] = None + tool_choice: Optional[ + Union[ToolChoiceOptions, ToolChoiceTypes, ToolChoiceFunction] + ] = Field( + None, + description='How the model should select which tool (or tools) to use when generating\na response. See the `tools` parameter to see how to specify which tools\nthe model can call.\n', + ) + tools: Optional[List[Tool]] = None + truncation: Optional[Truncation1] = Field( + 'disabled', + description="The truncation strategy to use for the model response.\n- `auto`: If the context of this response and previous ones exceeds\n the model's context window size, the model will truncate the \n response to fit the context window by dropping input items in the\n middle of the conversation. \n- `disabled` (default): If a model response will exceed the context window \n size for a model, the request will fail with a 400 error.\n", + ) + + +class GeminiCandidate(BaseModel): + citationMetadata: Optional[GeminiCitationMetadata] = None + content: Optional[GeminiContent] = None + finishReason: Optional[str] = None + safetyRatings: Optional[List[GeminiSafetyRating]] = None + + +class GeminiGenerateContentResponse(BaseModel): + candidates: Optional[List[GeminiCandidate]] = None + promptFeedback: Optional[GeminiPromptFeedback] = None + + +class InputItem(RootModel[Union[EasyInputMessage, Item]]): + root: Union[EasyInputMessage, Item] + + +class OpenAICreateResponse(CreateModelResponseProperties, ResponseProperties): + include: Optional[List[Includable]] = Field( + None, + description='Specify additional output data to include in the model response. Currently\nsupported values are:\n- `file_search_call.results`: Include the search results of\n the file search tool call.\n- `message.input_image.image_url`: Include image urls from the input message.\n- `computer_call_output.output.image_url`: Include image urls from the computer call output.\n', + ) + input: Union[str, List[InputItem]] = Field( + ..., + description='Text, image, or file inputs to the model, used to generate a response.\n\nLearn more:\n- [Text inputs and outputs](/docs/guides/text)\n- [Image inputs](/docs/guides/images)\n- [File inputs](/docs/guides/pdf-files)\n- [Conversation state](/docs/guides/conversation-state)\n- [Function calling](/docs/guides/function-calling)\n', + ) + parallel_tool_calls: Optional[bool] = Field( + True, description='Whether to allow the model to run tool calls in parallel.\n' + ) + store: Optional[bool] = Field( + True, + description='Whether to store the generated model response for later retrieval via\nAPI.\n', + ) + stream: Optional[bool] = Field( + False, + description='If set to true, the model response data will be streamed to the client\nas it is generated using [server-sent events](https://developer.mozilla.org/en-US/docs/Web/API/Server-sent_events/Using_server-sent_events#Event_stream_format).\nSee the [Streaming section below](/docs/api-reference/responses-streaming)\nfor more information.\n', + ) + usage: Optional[ResponseUsage] = None + + +class OpenAIResponse(ModelResponseProperties, ResponseProperties): + created_at: Optional[float] = Field( + None, + description='Unix timestamp (in seconds) of when this Response was created.', + ) + error: Optional[ResponseError] = None + id: Optional[str] = Field(None, description='Unique identifier for this Response.') + incomplete_details: Optional[IncompleteDetails] = Field( + None, description='Details about why the response is incomplete.\n' + ) + object: Optional[Object] = Field( + None, description='The object type of this resource - always set to `response`.' + ) + output: Optional[List[OutputItem]] = Field( + None, + description="An array of content items generated by the model.\n\n- The length and order of items in the `output` array is dependent\n on the model's response.\n- Rather than accessing the first item in the `output` array and \n assuming it's an `assistant` message with the content generated by\n the model, you might consider using the `output_text` property where\n supported in SDKs.\n", + ) + output_text: Optional[str] = Field( + None, + description='SDK-only convenience property that contains the aggregated text output \nfrom all `output_text` items in the `output` array, if any are present. \nSupported in the Python and JavaScript SDKs.\n', + ) + parallel_tool_calls: Optional[bool] = Field( + True, description='Whether to allow the model to run tool calls in parallel.\n' + ) + status: Optional[Status6] = Field( + None, + description='The status of the response generation. One of `completed`, `failed`, `in_progress`, or `incomplete`.', + ) + usage: Optional[ResponseUsage] = None diff --git a/comfy_api_nodes/apis/bfl_api.py b/comfy_api_nodes/apis/bfl_api.py new file mode 100644 index 000000000..0e90aef7c --- /dev/null +++ b/comfy_api_nodes/apis/bfl_api.py @@ -0,0 +1,174 @@ +from __future__ import annotations + +from enum import Enum +from typing import Any, Dict, Optional + +from pydantic import BaseModel, Field, confloat, conint + + +class BFLOutputFormat(str, Enum): + png = 'png' + jpeg = 'jpeg' + + +class BFLFluxExpandImageRequest(BaseModel): + prompt: str = Field(..., description='The description of the changes you want to make. This text guides the expansion process, allowing you to specify features, styles, or modifications for the expanded areas.') + prompt_upsampling: Optional[bool] = Field( + None, description='Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation.' + ) + seed: Optional[int] = Field(None, description='The seed value for reproducibility.') + top: conint(ge=0, le=2048) = Field(..., description='Number of pixels to expand at the top of the image') + bottom: conint(ge=0, le=2048) = Field(..., description='Number of pixels to expand at the bottom of the image') + left: conint(ge=0, le=2048) = Field(..., description='Number of pixels to expand at the left side of the image') + right: conint(ge=0, le=2048) = Field(..., description='Number of pixels to expand at the right side of the image') + steps: conint(ge=15, le=50) = Field(..., description='Number of steps for the image generation process') + guidance: confloat(ge=1.5, le=100) = Field(..., description='Guidance strength for the image generation process') + safety_tolerance: Optional[conint(ge=0, le=6)] = Field( + 6, description='Tolerance level for input and output moderation. Between 0 and 6, 0 being most strict, 6 being least strict. Defaults to 2.' + ) + output_format: Optional[BFLOutputFormat] = Field( + BFLOutputFormat.png, description="Output format for the generated image. Can be 'jpeg' or 'png'.", examples=['png'] + ) + image: str = Field(None, description='A Base64-encoded string representing the image you wish to expand') + + +class BFLFluxFillImageRequest(BaseModel): + prompt: str = Field(..., description='The description of the changes you want to make. This text guides the expansion process, allowing you to specify features, styles, or modifications for the expanded areas.') + prompt_upsampling: Optional[bool] = Field( + None, description='Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation.' + ) + seed: Optional[int] = Field(None, description='The seed value for reproducibility.') + steps: conint(ge=15, le=50) = Field(..., description='Number of steps for the image generation process') + guidance: confloat(ge=1.5, le=100) = Field(..., description='Guidance strength for the image generation process') + safety_tolerance: Optional[conint(ge=0, le=6)] = Field( + 6, description='Tolerance level for input and output moderation. Between 0 and 6, 0 being most strict, 6 being least strict. Defaults to 2.' + ) + output_format: Optional[BFLOutputFormat] = Field( + BFLOutputFormat.png, description="Output format for the generated image. Can be 'jpeg' or 'png'.", examples=['png'] + ) + image: str = Field(None, description='A Base64-encoded string representing the image you wish to modify. Can contain alpha mask if desired.') + mask: str = Field(None, description='A Base64-encoded string representing the mask of the areas you with to modify.') + + +class BFLFluxCannyImageRequest(BaseModel): + prompt: str = Field(..., description='Text prompt for image generation') + prompt_upsampling: Optional[bool] = Field( + None, description='Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation.' + ) + canny_low_threshold: Optional[int] = Field(None, description='Low threshold for Canny edge detection') + canny_high_threshold: Optional[int] = Field(None, description='High threshold for Canny edge detection') + seed: Optional[int] = Field(None, description='The seed value for reproducibility.') + steps: conint(ge=15, le=50) = Field(..., description='Number of steps for the image generation process') + guidance: confloat(ge=1, le=100) = Field(..., description='Guidance strength for the image generation process') + safety_tolerance: Optional[conint(ge=0, le=6)] = Field( + 6, description='Tolerance level for input and output moderation. Between 0 and 6, 0 being most strict, 6 being least strict. Defaults to 2.' + ) + output_format: Optional[BFLOutputFormat] = Field( + BFLOutputFormat.png, description="Output format for the generated image. Can be 'jpeg' or 'png'.", examples=['png'] + ) + control_image: Optional[str] = Field(None, description='Base64 encoded image to use as control input if no preprocessed image is provided') + preprocessed_image: Optional[str] = Field(None, description='Optional pre-processed image that will bypass the control preprocessing step') + + +class BFLFluxDepthImageRequest(BaseModel): + prompt: str = Field(..., description='Text prompt for image generation') + prompt_upsampling: Optional[bool] = Field( + None, description='Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation.' + ) + seed: Optional[int] = Field(None, description='The seed value for reproducibility.') + steps: conint(ge=15, le=50) = Field(..., description='Number of steps for the image generation process') + guidance: confloat(ge=1, le=100) = Field(..., description='Guidance strength for the image generation process') + safety_tolerance: Optional[conint(ge=0, le=6)] = Field( + 6, description='Tolerance level for input and output moderation. Between 0 and 6, 0 being most strict, 6 being least strict. Defaults to 2.' + ) + output_format: Optional[BFLOutputFormat] = Field( + BFLOutputFormat.png, description="Output format for the generated image. Can be 'jpeg' or 'png'.", examples=['png'] + ) + control_image: Optional[str] = Field(None, description='Base64 encoded image to use as control input if no preprocessed image is provided') + preprocessed_image: Optional[str] = Field(None, description='Optional pre-processed image that will bypass the control preprocessing step') + + +class BFLFluxProGenerateRequest(BaseModel): + prompt: str = Field(..., description='The text prompt for image generation.') + prompt_upsampling: Optional[bool] = Field( + None, description='Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation.' + ) + seed: Optional[int] = Field(None, description='The seed value for reproducibility.') + width: conint(ge=256, le=1440) = Field(1024, description='Width of the generated image in pixels. Must be a multiple of 32.') + height: conint(ge=256, le=1440) = Field(768, description='Height of the generated image in pixels. Must be a multiple of 32.') + safety_tolerance: Optional[conint(ge=0, le=6)] = Field( + 6, description='Tolerance level for input and output moderation. Between 0 and 6, 0 being most strict, 6 being least strict. Defaults to 2.' + ) + output_format: Optional[BFLOutputFormat] = Field( + BFLOutputFormat.png, description="Output format for the generated image. Can be 'jpeg' or 'png'.", examples=['png'] + ) + image_prompt: Optional[str] = Field(None, description='Optional image to remix in base64 format') + # image_prompt_strength: Optional[confloat(ge=0.0, le=1.0)] = Field( + # None, description='Blend between the prompt and the image prompt.' + # ) + + +class BFLFluxKontextProGenerateRequest(BaseModel): + prompt: str = Field(..., description='The text prompt for what you wannt to edit.') + input_image: Optional[str] = Field(None, description='Image to edit in base64 format') + seed: Optional[int] = Field(None, description='The seed value for reproducibility.') + guidance: confloat(ge=0.1, le=99.0) = Field(..., description='Guidance strength for the image generation process') + steps: conint(ge=1, le=150) = Field(..., description='Number of steps for the image generation process') + safety_tolerance: Optional[conint(ge=0, le=2)] = Field( + 2, description='Tolerance level for input and output moderation. Between 0 and 2, 0 being most strict, 6 being least strict. Defaults to 2.' + ) + output_format: Optional[BFLOutputFormat] = Field( + BFLOutputFormat.png, description="Output format for the generated image. Can be 'jpeg' or 'png'.", examples=['png'] + ) + aspect_ratio: Optional[str] = Field(None, description='Aspect ratio of the image between 21:9 and 9:21.') + prompt_upsampling: Optional[bool] = Field( + None, description='Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation.' + ) + + +class BFLFluxProUltraGenerateRequest(BaseModel): + prompt: str = Field(..., description='The text prompt for image generation.') + prompt_upsampling: Optional[bool] = Field( + None, description='Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation.' + ) + seed: Optional[int] = Field(None, description='The seed value for reproducibility.') + aspect_ratio: Optional[str] = Field(None, description='Aspect ratio of the image between 21:9 and 9:21.') + safety_tolerance: Optional[conint(ge=0, le=6)] = Field( + 6, description='Tolerance level for input and output moderation. Between 0 and 6, 0 being most strict, 6 being least strict. Defaults to 2.' + ) + output_format: Optional[BFLOutputFormat] = Field( + BFLOutputFormat.png, description="Output format for the generated image. Can be 'jpeg' or 'png'.", examples=['png'] + ) + raw: Optional[bool] = Field(None, description='Generate less processed, more natural-looking images.') + image_prompt: Optional[str] = Field(None, description='Optional image to remix in base64 format') + image_prompt_strength: Optional[confloat(ge=0.0, le=1.0)] = Field( + None, description='Blend between the prompt and the image prompt.' + ) + + +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.') + + +class BFLStatus(str, Enum): + task_not_found = "Task not found" + pending = "Pending" + request_moderated = "Request Moderated" + content_moderated = "Content Moderated" + ready = "Ready" + error = "Error" + + +class BFLFluxProStatusResponse(BaseModel): + id: str = Field(..., description="The unique identifier for the generation task.") + status: BFLStatus = Field(..., description="The status of the task.") + result: Optional[Dict[str, Any]] = Field( + None, description="The result of the task (null if not completed)." + ) + progress: confloat(ge=0.0, le=1.0) = Field( + ..., description="The progress of the task (0.0 to 1.0)." + ) + details: Optional[Dict[str, Any]] = Field( + None, description="Additional details about the task (null if not available)." + ) diff --git a/comfy_api_nodes/apis/client.py b/comfy_api_nodes/apis/client.py new file mode 100644 index 000000000..2a4bac88b --- /dev/null +++ b/comfy_api_nodes/apis/client.py @@ -0,0 +1,1126 @@ +""" +API Client Framework for api.comfy.org. + +This module provides a flexible framework for making API requests from ComfyUI nodes. +It supports both synchronous and asynchronous API operations with proper type validation. + +Key Components: +-------------- +1. ApiClient - Handles HTTP requests with authentication and error handling +2. ApiEndpoint - Defines a single HTTP endpoint with its request/response models +3. ApiOperation - Executes a single synchronous API operation + +Usage Examples: +-------------- + +# Example 1: Synchronous API Operation +# ------------------------------------ +# For a simple API call that returns the result immediately: + +# 1. Create the API client +api_client = ApiClient( + base_url="https://api.example.com", + auth_token="your_auth_token_here", + comfy_api_key="your_comfy_api_key_here", + timeout=30.0, + verify_ssl=True +) + +# 2. Define the endpoint +user_info_endpoint = ApiEndpoint( + path="/v1/users/me", + method=HttpMethod.GET, + request_model=EmptyRequest, # No request body needed + response_model=UserProfile, # Pydantic model for the response + query_params=None +) + +# 3. Create the request object +request = EmptyRequest() + +# 4. Create and execute the operation +operation = ApiOperation( + endpoint=user_info_endpoint, + request=request +) +user_profile = operation.execute(client=api_client) # Returns immediately with the result + + +# Example 2: Asynchronous API Operation with Polling +# ------------------------------------------------- +# For an API that starts a task and requires polling for completion: + +# 1. Define the endpoints (initial request and polling) +generate_image_endpoint = ApiEndpoint( + path="/v1/images/generate", + method=HttpMethod.POST, + request_model=ImageGenerationRequest, + response_model=TaskCreatedResponse, + query_params=None +) + +check_task_endpoint = ApiEndpoint( + path="/v1/tasks/{task_id}", + method=HttpMethod.GET, + request_model=EmptyRequest, + response_model=ImageGenerationResult, + query_params=None +) + +# 2. Create the request object +request = ImageGenerationRequest( + prompt="a beautiful sunset over mountains", + width=1024, + height=1024, + num_images=1 +) + +# 3. Create and execute the polling operation +operation = PollingOperation( + initial_endpoint=generate_image_endpoint, + initial_request=request, + poll_endpoint=check_task_endpoint, + task_id_field="task_id", + status_field="status", + completed_statuses=["completed"], + failed_statuses=["failed", "error"] +) + +# This will make the initial request and then poll until completion +result = operation.execute(client=api_client) # Returns the final ImageGenerationResult when done +""" + +from __future__ import annotations +import logging +import time +import io +import socket +from typing import Dict, Type, Optional, Any, TypeVar, Generic, Callable, Tuple +from enum import Enum +import json +import requests +from urllib.parse import urljoin, urlparse +from pydantic import BaseModel, Field +import uuid # For generating unique operation IDs + +from server import PromptServer +from comfy.cli_args import args +from comfy import utils +from . import request_logger + +T = TypeVar("T", bound=BaseModel) +R = TypeVar("R", bound=BaseModel) +P = TypeVar("P", bound=BaseModel) # For poll response + +PROGRESS_BAR_MAX = 100 + + +class NetworkError(Exception): + """Base exception for network-related errors with diagnostic information.""" + pass + + +class LocalNetworkError(NetworkError): + """Exception raised when local network connectivity issues are detected.""" + pass + + +class ApiServerError(NetworkError): + """Exception raised when the API server is unreachable but internet is working.""" + pass + + +class EmptyRequest(BaseModel): + """Base class for empty request bodies. + For GET requests, fields will be sent as query parameters.""" + + pass + + +class UploadRequest(BaseModel): + file_name: str = Field(..., description="Filename to upload") + content_type: Optional[str] = Field( + None, + description="Mime type of the file. For example: image/png, image/jpeg, video/mp4, etc.", + ) + + +class UploadResponse(BaseModel): + download_url: str = Field(..., description="URL to GET uploaded file") + upload_url: str = Field(..., description="URL to PUT file to upload") + + +class HttpMethod(str, Enum): + GET = "GET" + POST = "POST" + PUT = "PUT" + DELETE = "DELETE" + PATCH = "PATCH" + + +class ApiClient: + """ + Client for making HTTP requests to an API with authentication, error handling, and retry logic. + """ + + def __init__( + self, + base_url: str, + auth_token: Optional[str] = None, + comfy_api_key: Optional[str] = None, + timeout: float = 3600.0, + verify_ssl: bool = True, + max_retries: int = 3, + retry_delay: float = 1.0, + retry_backoff_factor: float = 2.0, + retry_status_codes: Optional[Tuple[int, ...]] = None, + ): + self.base_url = base_url + self.auth_token = auth_token + self.comfy_api_key = comfy_api_key + self.timeout = timeout + self.verify_ssl = verify_ssl + self.max_retries = max_retries + self.retry_delay = retry_delay + self.retry_backoff_factor = retry_backoff_factor + # Default retry status codes: 408 (Request Timeout), 429 (Too Many Requests), + # 500, 502, 503, 504 (Server Errors) + self.retry_status_codes = retry_status_codes or (408, 429, 500, 502, 503, 504) + + def _generate_operation_id(self, path: str) -> str: + """Generates a unique operation ID for logging.""" + return f"{path.strip('/').replace('/', '_')}_{uuid.uuid4().hex[:8]}" + + def _create_json_payload_args( + self, + data: Optional[Dict[str, Any]] = None, + headers: Optional[Dict[str, str]] = None, + ) -> Dict[str, Any]: + return { + "json": data, + "headers": headers, + } + + def _create_form_data_args( + self, + data: Dict[str, Any], + files: Dict[str, Any], + headers: Optional[Dict[str, str]] = None, + multipart_parser = None, + ) -> Dict[str, Any]: + if headers and "Content-Type" in headers: + del headers["Content-Type"] + + if multipart_parser: + data = multipart_parser(data) + + return { + "data": data, + "files": files, + "headers": headers, + } + + def _create_urlencoded_form_data_args( + self, + data: Dict[str, Any], + headers: Optional[Dict[str, str]] = None, + ) -> Dict[str, Any]: + headers = headers or {} + headers["Content-Type"] = "application/x-www-form-urlencoded" + + return { + "data": data, + "headers": headers, + } + + def get_headers(self) -> Dict[str, str]: + """Get headers for API requests, including authentication if available""" + headers = {"Content-Type": "application/json", "Accept": "application/json"} + + if self.auth_token: + headers["Authorization"] = f"Bearer {self.auth_token}" + elif self.comfy_api_key: + headers["X-API-KEY"] = self.comfy_api_key + + return headers + + def _check_connectivity(self, target_url: str) -> Dict[str, bool]: + """ + Check connectivity to determine if network issues are local or server-related. + + Args: + target_url: URL to check connectivity to + + Returns: + Dictionary with connectivity status details + """ + results = { + "internet_accessible": False, + "api_accessible": False, + "is_local_issue": False, + "is_api_issue": False + } + + # First check basic internet connectivity using a reliable external site + try: + # Use a reliable external domain for checking basic connectivity + check_response = requests.get("https://www.google.com", + timeout=5.0, + verify=self.verify_ssl) + if check_response.status_code < 500: + results["internet_accessible"] = True + except (requests.RequestException, socket.error): + results["internet_accessible"] = False + results["is_local_issue"] = True + return results + + # Now check API server connectivity + try: + # Extract domain from the target URL to do a simpler health check + parsed_url = urlparse(target_url) + api_base = f"{parsed_url.scheme}://{parsed_url.netloc}" + + # Try to reach the API domain + api_response = requests.get(f"{api_base}/health", timeout=5.0, verify=self.verify_ssl) + if api_response.status_code < 500: + results["api_accessible"] = True + else: + results["api_accessible"] = False + results["is_api_issue"] = True + except requests.RequestException: + results["api_accessible"] = False + # If we can reach the internet but not the API, it's an API issue + results["is_api_issue"] = True + + return results + + def request( + self, + method: str, + path: str, + params: Optional[Dict[str, Any]] = None, + data: Optional[Dict[str, Any]] = None, + files: Optional[Dict[str, Any]] = None, + headers: Optional[Dict[str, str]] = None, + content_type: str = "application/json", + multipart_parser: Callable = None, + retry_count: int = 0, # Used internally for tracking retries + ) -> Dict[str, Any]: + """ + Make an HTTP request to the API with automatic retries for transient errors. + + Args: + method: HTTP method (GET, POST, etc.) + path: API endpoint path (will be joined with base_url) + params: Query parameters + data: body data + files: Files to upload + headers: Additional headers + content_type: Content type of the request. Defaults to application/json. + retry_count: Internal parameter for tracking retries, do not set manually + + Returns: + Parsed JSON response + + Raises: + LocalNetworkError: If local network connectivity issues are detected + ApiServerError: If the API server is unreachable but internet is working + Exception: For other request failures + """ + # Use urljoin but ensure path is relative to avoid absolute path behavior + relative_path = path.lstrip('/') + url = urljoin(self.base_url, relative_path) + self.check_auth(self.auth_token, self.comfy_api_key) + # Combine default headers with any provided headers + request_headers = self.get_headers() + if headers: + request_headers.update(headers) + + # Let requests handle the content type when files are present. + if files: + del request_headers["Content-Type"] + + logging.debug(f"[DEBUG] Request Headers: {request_headers}") + logging.debug(f"[DEBUG] Files: {files}") + logging.debug(f"[DEBUG] Params: {params}") + logging.debug(f"[DEBUG] Data: {data}") + + if content_type == "application/x-www-form-urlencoded": + payload_args = self._create_urlencoded_form_data_args(data, request_headers) + elif content_type == "multipart/form-data": + payload_args = self._create_form_data_args( + data, files, request_headers, multipart_parser + ) + else: + payload_args = self._create_json_payload_args(data, request_headers) + + operation_id = self._generate_operation_id(path) + request_logger.log_request_response( + operation_id=operation_id, + request_method=method, + request_url=url, + request_headers=request_headers, + request_params=params, + request_data=data if content_type == "application/json" else "[form-data or other]" + ) + + try: + response = requests.request( + method=method, + url=url, + params=params, + timeout=self.timeout, + verify=self.verify_ssl, + **payload_args, + ) + + # Check if we should retry based on status code + if (response.status_code in self.retry_status_codes and + retry_count < self.max_retries): + + # Calculate delay with exponential backoff + delay = self.retry_delay * (self.retry_backoff_factor ** retry_count) + + logging.warning( + f"Request failed with status {response.status_code}. " + f"Retrying in {delay:.2f}s ({retry_count + 1}/{self.max_retries})" + ) + + time.sleep(delay) + return self.request( + method=method, + path=path, + params=params, + data=data, + files=files, + headers=headers, + content_type=content_type, + multipart_parser=multipart_parser, + retry_count=retry_count + 1, + ) + + # Raise exception for error status codes + response.raise_for_status() + + # Log successful response + response_content_to_log = response.content + try: + # Attempt to parse JSON for prettier logging, fallback to raw content + response_content_to_log = response.json() + except json.JSONDecodeError: + pass # Keep as bytes/str if not JSON + + request_logger.log_request_response( + operation_id=operation_id, + request_method=method, # Pass request details again for context in log + request_url=url, + response_status_code=response.status_code, + response_headers=dict(response.headers), + response_content=response_content_to_log + ) + + except requests.ConnectionError as e: + error_message = f"ConnectionError: {str(e)}" + request_logger.log_request_response( + operation_id=operation_id, + request_method=method, + request_url=url, + error_message=error_message + ) + # Only perform connectivity check if we've exhausted all retries + if retry_count >= self.max_retries: + # Check connectivity to determine if it's a local or API issue + connectivity = self._check_connectivity(self.base_url) + + if connectivity["is_local_issue"]: + raise LocalNetworkError( + "Unable to connect to the API server due to local network issues. " + "Please check your internet connection and try again." + ) from e + elif connectivity["is_api_issue"]: + raise ApiServerError( + f"The API server at {self.base_url} is currently unreachable. " + f"The service may be experiencing issues. Please try again later." + ) from e + + # If we haven't exhausted retries yet, retry the request + if retry_count < self.max_retries: + delay = self.retry_delay * (self.retry_backoff_factor ** retry_count) + logging.warning( + f"Connection error: {str(e)}. " + f"Retrying in {delay:.2f}s ({retry_count + 1}/{self.max_retries})" + ) + time.sleep(delay) + return self.request( + method=method, + path=path, + params=params, + data=data, + files=files, + headers=headers, + content_type=content_type, + multipart_parser=multipart_parser, + retry_count=retry_count + 1, + ) + + # If we've exhausted retries and didn't identify the specific issue, + # raise a generic exception + final_error_message = ( + f"Unable to connect to the API server after {self.max_retries} attempts. " + f"Please check your internet connection or try again later." + ) + request_logger.log_request_response( # Log final failure + operation_id=operation_id, + request_method=method, request_url=url, + error_message=final_error_message + ) + raise Exception(final_error_message) from e + + except requests.Timeout as e: + error_message = f"Timeout: {str(e)}" + request_logger.log_request_response( + operation_id=operation_id, + request_method=method, request_url=url, + error_message=error_message + ) + # Retry timeouts if we haven't exhausted retries + if retry_count < self.max_retries: + delay = self.retry_delay * (self.retry_backoff_factor ** retry_count) + logging.warning( + f"Request timed out. " + f"Retrying in {delay:.2f}s ({retry_count + 1}/{self.max_retries})" + ) + time.sleep(delay) + return self.request( + method=method, + path=path, + params=params, + data=data, + files=files, + headers=headers, + content_type=content_type, + multipart_parser=multipart_parser, + retry_count=retry_count + 1, + ) + final_error_message = ( + f"Request timed out after {self.timeout} seconds and {self.max_retries} retry attempts. " + f"The server might be experiencing high load or the operation is taking longer than expected." + ) + request_logger.log_request_response( # Log final failure + operation_id=operation_id, + request_method=method, request_url=url, + error_message=final_error_message + ) + raise Exception(final_error_message) from e + + except requests.HTTPError as e: + status_code = e.response.status_code if hasattr(e, "response") else None + original_error_message = f"HTTP Error: {str(e)}" + error_content_for_log = None + if hasattr(e, "response") and e.response is not None: + error_content_for_log = e.response.content + try: + error_content_for_log = e.response.json() + except json.JSONDecodeError: + pass + + + # Try to extract detailed error message from JSON response for user display + # but log the full error content. + user_display_error_message = original_error_message + + try: + if hasattr(e, "response") and e.response is not None and e.response.content: + error_json = e.response.json() + if "error" in error_json and "message" in error_json["error"]: + user_display_error_message = f"API Error: {error_json['error']['message']}" + if "type" in error_json["error"]: + user_display_error_message += f" (Type: {error_json['error']['type']})" + elif isinstance(error_json, dict): # Handle cases where error is just a JSON dict + user_display_error_message = f"API Error: {json.dumps(error_json)}" + else: # Non-dict JSON error + user_display_error_message = f"API Error: {str(error_json)}" + except json.JSONDecodeError: + # If not JSON, use the raw content if it's not too long, or a summary + if hasattr(e, "response") and e.response is not None and e.response.content: + raw_content = e.response.content.decode(errors='ignore') + if len(raw_content) < 200: # Arbitrary limit for display + user_display_error_message = f"API Error (raw): {raw_content}" + else: + user_display_error_message = f"API Error (raw, status {status_code})" + + request_logger.log_request_response( + operation_id=operation_id, + request_method=method, request_url=url, + response_status_code=status_code, + response_headers=dict(e.response.headers) if hasattr(e, "response") and e.response is not None else None, + response_content=error_content_for_log, + error_message=original_error_message # Log the original exception string as error + ) + + logging.debug(f"[DEBUG] API Error: {user_display_error_message} (Status: {status_code})") + if hasattr(e, "response") and e.response is not None and e.response.content: + logging.debug(f"[DEBUG] Response content: {e.response.content}") + + # Retry if the status code is in our retry list and we haven't exhausted retries + if (status_code in self.retry_status_codes and + retry_count < self.max_retries): + + delay = self.retry_delay * (self.retry_backoff_factor ** retry_count) + logging.warning( + f"HTTP error {status_code}. " + f"Retrying in {delay:.2f}s ({retry_count + 1}/{self.max_retries})" + ) + time.sleep(delay) + return self.request( + method=method, + path=path, + params=params, + data=data, + files=files, + headers=headers, + content_type=content_type, + multipart_parser=multipart_parser, + retry_count=retry_count + 1, + ) + + # Specific error messages for common status codes for user display + if status_code == 401: + user_display_error_message = "Unauthorized: Please login first to use this node." + elif status_code == 402: + user_display_error_message = "Payment Required: Please add credits to your account to use this node." + elif status_code == 409: + user_display_error_message = "There is a problem with your account. Please contact support@comfy.org." + elif status_code == 429: + user_display_error_message = "Rate Limit Exceeded: Please try again later." + # else, user_display_error_message remains as parsed from response or original HTTPError string + + raise Exception(user_display_error_message) # Raise with the user-friendly message + + # Parse and return JSON response + if response.content: + return response.json() + return {} + + def check_auth(self, auth_token, comfy_api_key): + """Verify that an auth token is present or comfy_api_key is present""" + if auth_token is None and comfy_api_key is None: + raise Exception("Unauthorized: Please login first to use this node.") + return auth_token or comfy_api_key + + @staticmethod + def upload_file( + upload_url: str, + file: io.BytesIO | str, + content_type: str | None = None, + max_retries: int = 3, + retry_delay: float = 1.0, + retry_backoff_factor: float = 2.0, + ): + """Upload a file to the API with retry logic. + + Args: + upload_url: The URL to upload to + file: Either a file path string, BytesIO object, or tuple of (file_path, filename) + content_type: Optional mime type to set for the upload + max_retries: Maximum number of retry attempts + retry_delay: Initial delay between retries in seconds + retry_backoff_factor: Multiplier for the delay after each retry + """ + headers = {} + if content_type: + headers["Content-Type"] = content_type + + # Prepare the file data + if isinstance(file, io.BytesIO): + file.seek(0) # Ensure we're at the start of the file + data = file.read() + elif isinstance(file, str): + with open(file, "rb") as f: + data = f.read() + else: + raise ValueError("File must be either a BytesIO object or a file path string") + + # Try the upload with retries + last_exception = None + operation_id = f"upload_{upload_url.split('/')[-1]}_{uuid.uuid4().hex[:8]}" # Simplified ID for uploads + + # Log initial attempt (without full file data for brevity) + request_logger.log_request_response( + operation_id=operation_id, + request_method="PUT", + request_url=upload_url, + request_headers=headers, + request_data=f"[File data of type {content_type or 'unknown'}, size {len(data)} bytes]" + ) + + for retry_attempt in range(max_retries + 1): + try: + response = requests.put(upload_url, data=data, headers=headers) + response.raise_for_status() + request_logger.log_request_response( + operation_id=operation_id, + request_method="PUT", request_url=upload_url, # For context + response_status_code=response.status_code, + response_headers=dict(response.headers), + response_content="File uploaded successfully." # Or response.text if available + ) + return response + + except (requests.ConnectionError, requests.Timeout, requests.HTTPError) as e: + last_exception = e + error_message_for_log = f"{type(e).__name__}: {str(e)}" + response_content_for_log = None + status_code_for_log = None + headers_for_log = None + + if hasattr(e, 'response') and e.response is not None: + status_code_for_log = e.response.status_code + headers_for_log = dict(e.response.headers) + try: + response_content_for_log = e.response.json() + except json.JSONDecodeError: + response_content_for_log = e.response.content + + + request_logger.log_request_response( + operation_id=operation_id, + request_method="PUT", request_url=upload_url, + response_status_code=status_code_for_log, + response_headers=headers_for_log, + response_content=response_content_for_log, + error_message=error_message_for_log + ) + + if retry_attempt < max_retries: + delay = retry_delay * (retry_backoff_factor ** retry_attempt) + logging.warning( + f"File upload failed: {str(e)}. " + f"Retrying in {delay:.2f}s ({retry_attempt + 1}/{max_retries})" + ) + time.sleep(delay) + else: + break # Max retries reached + + # If we've exhausted all retries, determine the final error type and raise + final_error_message = f"Failed to upload file after {max_retries + 1} attempts. Error: {str(last_exception)}" + try: + # Check basic internet connectivity + check_response = requests.get("https://www.google.com", timeout=5.0, verify=True) # Assuming verify=True is desired + if check_response.status_code >= 500: # Google itself has an issue (rare) + final_error_message = (f"Failed to upload file. Internet connectivity check to Google failed " + f"(status {check_response.status_code}). Original error: {str(last_exception)}") + # Not raising LocalNetworkError here as Google itself might be down. + # If Google is reachable, the issue is likely with the upload server or a more specific local problem + # not caught by a simple Google ping (e.g., DNS for the specific upload URL, firewall). + # The original last_exception is probably most relevant. + + except (requests.RequestException, socket.error) as conn_check_exc: + # Could not reach Google, likely a local network issue + final_error_message = (f"Failed to upload file due to network connectivity issues " + f"(cannot reach Google: {str(conn_check_exc)}). " + f"Original upload error: {str(last_exception)}") + request_logger.log_request_response( # Log final failure reason + operation_id=operation_id, + request_method="PUT", request_url=upload_url, + error_message=final_error_message + ) + raise LocalNetworkError(final_error_message) from last_exception + + request_logger.log_request_response( # Log final failure reason if not LocalNetworkError + operation_id=operation_id, + request_method="PUT", request_url=upload_url, + error_message=final_error_message + ) + raise Exception(final_error_message) from last_exception + + +class ApiEndpoint(Generic[T, R]): + """Defines an API endpoint with its request and response types""" + + def __init__( + self, + path: str, + method: HttpMethod, + request_model: Type[T], + response_model: Type[R], + query_params: Optional[Dict[str, Any]] = None, + ): + """Initialize an API endpoint definition. + + Args: + path: The URL path for this endpoint, can include placeholders like {id} + method: The HTTP method to use (GET, POST, etc.) + request_model: Pydantic model class that defines the structure and validation rules for API requests to this endpoint + response_model: Pydantic model class that defines the structure and validation rules for API responses from this endpoint + query_params: Optional dictionary of query parameters to include in the request + """ + self.path = path + self.method = method + self.request_model = request_model + self.response_model = response_model + self.query_params = query_params or {} + + +class SynchronousOperation(Generic[T, R]): + """ + Represents a single synchronous API operation. + """ + + def __init__( + self, + endpoint: ApiEndpoint[T, R], + request: T, + files: Optional[Dict[str, Any]] = None, + api_base: str | None = None, + auth_token: Optional[str] = None, + comfy_api_key: Optional[str] = None, + auth_kwargs: Optional[Dict[str,str]] = None, + timeout: float = 604800.0, + verify_ssl: bool = True, + content_type: str = "application/json", + multipart_parser: Callable = None, + max_retries: int = 3, + retry_delay: float = 1.0, + retry_backoff_factor: float = 2.0, + ): + self.endpoint = endpoint + self.request = request + self.response = None + self.error = None + self.api_base: str = api_base or args.comfy_api_base + self.auth_token = auth_token + self.comfy_api_key = comfy_api_key + if auth_kwargs is not None: + self.auth_token = auth_kwargs.get("auth_token", self.auth_token) + self.comfy_api_key = auth_kwargs.get("comfy_api_key", self.comfy_api_key) + self.timeout = timeout + self.verify_ssl = verify_ssl + self.files = files + self.content_type = content_type + self.multipart_parser = multipart_parser + self.max_retries = max_retries + self.retry_delay = retry_delay + self.retry_backoff_factor = retry_backoff_factor + + def execute(self, client: Optional[ApiClient] = None) -> R: + """Execute the API operation using the provided client or create one with retry support""" + try: + # Create client if not provided + if client is None: + client = ApiClient( + base_url=self.api_base, + auth_token=self.auth_token, + comfy_api_key=self.comfy_api_key, + timeout=self.timeout, + verify_ssl=self.verify_ssl, + max_retries=self.max_retries, + retry_delay=self.retry_delay, + retry_backoff_factor=self.retry_backoff_factor, + ) + + # Convert request model to dict, but use None for EmptyRequest + request_dict = ( + None + if isinstance(self.request, EmptyRequest) + else self.request.model_dump(exclude_none=True) + ) + if request_dict: + for key, value in request_dict.items(): + if isinstance(value, Enum): + request_dict[key] = value.value + + # Debug log for request + logging.debug( + f"[DEBUG] API Request: {self.endpoint.method.value} {self.endpoint.path}" + ) + logging.debug(f"[DEBUG] Request Data: {json.dumps(request_dict, indent=2)}") + logging.debug(f"[DEBUG] Query Params: {self.endpoint.query_params}") + + # Make the request with built-in retry + resp = client.request( + method=self.endpoint.method.value, + path=self.endpoint.path, + data=request_dict, + params=self.endpoint.query_params, + files=self.files, + content_type=self.content_type, + multipart_parser=self.multipart_parser + ) + + # Debug log for response + logging.debug("=" * 50) + logging.debug("[DEBUG] RESPONSE DETAILS:") + logging.debug("[DEBUG] Status Code: 200 (Success)") + logging.debug(f"[DEBUG] Response Body: {json.dumps(resp, indent=2)}") + logging.debug("=" * 50) + + # Parse and return the response + return self._parse_response(resp) + + except LocalNetworkError as e: + # Propagate specific network error types + logging.error(f"[ERROR] Local network error: {str(e)}") + raise + + except ApiServerError as e: + # Propagate API server errors + logging.error(f"[ERROR] API server error: {str(e)}") + raise + + except Exception as e: + logging.error(f"[ERROR] API Exception: {str(e)}") + raise Exception(str(e)) + + def _parse_response(self, resp): + """Parse response data - can be overridden by subclasses""" + # The response is already the complete object, don't extract just the "data" field + # as that would lose the outer structure (created timestamp, etc.) + + # Parse response using the provided model + self.response = self.endpoint.response_model.model_validate(resp) + logging.debug(f"[DEBUG] Parsed Response: {self.response}") + return self.response + + +class TaskStatus(str, Enum): + """Enum for task status values""" + + COMPLETED = "completed" + FAILED = "failed" + PENDING = "pending" + + +class PollingOperation(Generic[T, R]): + """ + Represents an asynchronous API operation that requires polling for completion. + """ + + def __init__( + self, + poll_endpoint: ApiEndpoint[EmptyRequest, R], + completed_statuses: list, + failed_statuses: list, + status_extractor: Callable[[R], str], + progress_extractor: Callable[[R], float] = None, + result_url_extractor: Callable[[R], str] = None, + request: Optional[T] = None, + api_base: str | None = None, + auth_token: Optional[str] = None, + comfy_api_key: Optional[str] = None, + auth_kwargs: Optional[Dict[str,str]] = None, + poll_interval: float = 5.0, + max_poll_attempts: int = 120, # Default max polling attempts (10 minutes with 5s interval) + max_retries: int = 3, # Max retries per individual API call + retry_delay: float = 1.0, + retry_backoff_factor: float = 2.0, + estimated_duration: Optional[float] = None, + node_id: Optional[str] = None, + ): + self.poll_endpoint = poll_endpoint + self.request = request + self.api_base: str = api_base or args.comfy_api_base + self.auth_token = auth_token + self.comfy_api_key = comfy_api_key + if auth_kwargs is not None: + self.auth_token = auth_kwargs.get("auth_token", self.auth_token) + self.comfy_api_key = auth_kwargs.get("comfy_api_key", self.comfy_api_key) + self.poll_interval = poll_interval + self.max_poll_attempts = max_poll_attempts + self.max_retries = max_retries + self.retry_delay = retry_delay + self.retry_backoff_factor = retry_backoff_factor + self.estimated_duration = estimated_duration + + # Polling configuration + self.status_extractor = status_extractor or ( + lambda x: getattr(x, "status", None) + ) + self.progress_extractor = progress_extractor + self.result_url_extractor = result_url_extractor + self.node_id = node_id + self.completed_statuses = completed_statuses + self.failed_statuses = failed_statuses + + # For storing response data + self.final_response = None + self.error = None + + def execute(self, client: Optional[ApiClient] = None) -> R: + """Execute the polling operation using the provided client. If failed, raise an exception.""" + try: + if client is None: + client = ApiClient( + base_url=self.api_base, + auth_token=self.auth_token, + comfy_api_key=self.comfy_api_key, + max_retries=self.max_retries, + retry_delay=self.retry_delay, + retry_backoff_factor=self.retry_backoff_factor, + ) + return self._poll_until_complete(client) + except LocalNetworkError as e: + # Provide clear message for local network issues + raise Exception( + f"Polling failed due to local network issues. Please check your internet connection. " + f"Details: {str(e)}" + ) from e + except ApiServerError as e: + # Provide clear message for API server issues + raise Exception( + f"Polling failed due to API server issues. The service may be experiencing problems. " + f"Please try again later. Details: {str(e)}" + ) from e + except Exception as e: + raise Exception(f"Error during polling: {str(e)}") + + def _display_text_on_node(self, text: str): + """Sends text to the client which will be displayed on the node in the UI""" + if not self.node_id: + return + + PromptServer.instance.send_progress_text(text, self.node_id) + + def _display_time_progress_on_node(self, time_completed: int): + if not self.node_id: + return + + if self.estimated_duration is not None: + estimated_time_remaining = max( + 0, int(self.estimated_duration) - int(time_completed) + ) + message = f"Task in progress: {time_completed:.0f}s (~{estimated_time_remaining:.0f}s remaining)" + else: + message = f"Task in progress: {time_completed:.0f}s" + self._display_text_on_node(message) + + def _check_task_status(self, response: R) -> TaskStatus: + """Check task status using the status extractor function""" + try: + status = self.status_extractor(response) + if status in self.completed_statuses: + return TaskStatus.COMPLETED + elif status in self.failed_statuses: + return TaskStatus.FAILED + return TaskStatus.PENDING + except Exception as e: + logging.error(f"Error extracting status: {e}") + return TaskStatus.PENDING + + def _poll_until_complete(self, client: ApiClient) -> R: + """Poll until the task is complete""" + poll_count = 0 + consecutive_errors = 0 + max_consecutive_errors = min(5, self.max_retries * 2) # Limit consecutive errors + + if self.progress_extractor: + progress = utils.ProgressBar(PROGRESS_BAR_MAX) + + while poll_count < self.max_poll_attempts: + try: + poll_count += 1 + logging.debug(f"[DEBUG] Polling attempt #{poll_count}") + + request_dict = ( + self.request.model_dump(exclude_none=True) + if self.request is not None + else None + ) + + if poll_count == 1: + logging.debug( + f"[DEBUG] Poll Request: {self.poll_endpoint.method.value} {self.poll_endpoint.path}" + ) + logging.debug( + f"[DEBUG] Poll Request Data: {json.dumps(request_dict, indent=2) if request_dict else 'None'}" + ) + + # Query task status + resp = client.request( + method=self.poll_endpoint.method.value, + path=self.poll_endpoint.path, + params=self.poll_endpoint.query_params, + data=request_dict, + ) + + # Successfully got a response, reset consecutive error count + consecutive_errors = 0 + + # Parse response + response_obj = self.poll_endpoint.response_model.model_validate(resp) + + # Check if task is complete + status = self._check_task_status(response_obj) + logging.debug(f"[DEBUG] Task Status: {status}") + + # If progress extractor is provided, extract progress + if self.progress_extractor: + new_progress = self.progress_extractor(response_obj) + if new_progress is not None: + progress.update_absolute(new_progress, total=PROGRESS_BAR_MAX) + + if status == TaskStatus.COMPLETED: + message = "Task completed successfully" + if self.result_url_extractor: + result_url = self.result_url_extractor(response_obj) + if result_url: + message = f"Result URL: {result_url}" + else: + message = "Task completed successfully!" + logging.debug(f"[DEBUG] {message}") + self._display_text_on_node(message) + self.final_response = response_obj + if self.progress_extractor: + progress.update(100) + return self.final_response + elif status == TaskStatus.FAILED: + message = f"Task failed: {json.dumps(resp)}" + logging.error(f"[DEBUG] {message}") + raise Exception(message) + else: + logging.debug("[DEBUG] Task still pending, continuing to poll...") + + # Wait before polling again + logging.debug( + f"[DEBUG] Waiting {self.poll_interval} seconds before next poll" + ) + for i in range(int(self.poll_interval)): + time_completed = (poll_count * self.poll_interval) + i + self._display_time_progress_on_node(time_completed) + time.sleep(1) + + except (LocalNetworkError, ApiServerError) as e: + # For network-related errors, increment error count and potentially abort + consecutive_errors += 1 + if consecutive_errors >= max_consecutive_errors: + raise Exception( + f"Polling aborted after {consecutive_errors} consecutive network errors: {str(e)}" + ) from e + + # Log the error but continue polling + logging.warning( + f"Network error during polling (attempt {poll_count}/{self.max_poll_attempts}): {str(e)}. " + f"Will retry in {self.poll_interval} seconds." + ) + time.sleep(self.poll_interval) + + except Exception as e: + # For other errors, increment count and potentially abort + consecutive_errors += 1 + if consecutive_errors >= max_consecutive_errors or status == TaskStatus.FAILED: + raise Exception( + f"Polling aborted after {consecutive_errors} consecutive errors: {str(e)}" + ) from e + + logging.error(f"[DEBUG] Polling error: {str(e)}") + logging.warning( + f"Error during polling (attempt {poll_count}/{self.max_poll_attempts}): {str(e)}. " + f"Will retry in {self.poll_interval} seconds." + ) + time.sleep(self.poll_interval) + + # If we've exhausted all polling attempts + raise Exception( + f"Polling timed out after {poll_count} attempts ({poll_count * self.poll_interval} seconds). " + f"The operation may still be running on the server but is taking longer than expected." + ) diff --git a/comfy_api_nodes/apis/luma_api.py b/comfy_api_nodes/apis/luma_api.py new file mode 100644 index 000000000..632c4ab96 --- /dev/null +++ b/comfy_api_nodes/apis/luma_api.py @@ -0,0 +1,253 @@ +from __future__ import annotations + + +import torch + +from enum import Enum +from typing import Optional, Union + +from pydantic import BaseModel, Field, confloat + + + +class LumaIO: + LUMA_REF = "LUMA_REF" + LUMA_CONCEPTS = "LUMA_CONCEPTS" + + +class LumaReference: + def __init__(self, image: torch.Tensor, weight: float): + self.image = image + self.weight = weight + + def create_api_model(self, download_url: str): + return LumaImageRef(url=download_url, weight=self.weight) + +class LumaReferenceChain: + def __init__(self, first_ref: LumaReference=None): + self.refs: list[LumaReference] = [] + if first_ref: + self.refs.append(first_ref) + + def add(self, luma_ref: LumaReference=None): + self.refs.append(luma_ref) + + def create_api_model(self, download_urls: list[str], max_refs=4): + if len(self.refs) == 0: + return None + api_refs: list[LumaImageRef] = [] + for ref, url in zip(self.refs, download_urls): + api_ref = LumaImageRef(url=url, weight=ref.weight) + api_refs.append(api_ref) + return api_refs + + def clone(self): + c = LumaReferenceChain() + for ref in self.refs: + c.add(ref) + return c + + +class LumaConcept: + def __init__(self, key: str): + self.key = key + + +class LumaConceptChain: + def __init__(self, str_list: list[str] = None): + self.concepts: list[LumaConcept] = [] + if str_list is not None: + for c in str_list: + if c != "None": + self.add(LumaConcept(key=c)) + + def add(self, concept: LumaConcept): + self.concepts.append(concept) + + def create_api_model(self): + if len(self.concepts) == 0: + return None + api_concepts: list[LumaConceptObject] = [] + for concept in self.concepts: + if concept.key == "None": + continue + api_concepts.append(LumaConceptObject(key=concept.key)) + if len(api_concepts) == 0: + return None + return api_concepts + + def clone(self): + c = LumaConceptChain() + for concept in self.concepts: + c.add(concept) + return c + + def clone_and_merge(self, other: LumaConceptChain): + c = self.clone() + for concept in other.concepts: + c.add(concept) + return c + + +def get_luma_concepts(include_none=False): + concepts = [] + if include_none: + concepts.append("None") + return concepts + [ + "truck_left", + "pan_right", + "pedestal_down", + "low_angle", + "pedestal_up", + "selfie", + "pan_left", + "roll_right", + "zoom_in", + "over_the_shoulder", + "orbit_right", + "orbit_left", + "static", + "tiny_planet", + "high_angle", + "bolt_cam", + "dolly_zoom", + "overhead", + "zoom_out", + "handheld", + "roll_left", + "pov", + "aerial_drone", + "push_in", + "crane_down", + "truck_right", + "tilt_down", + "elevator_doors", + "tilt_up", + "ground_level", + "pull_out", + "aerial", + "crane_up", + "eye_level" + ] + + +class LumaImageModel(str, Enum): + photon_1 = "photon-1" + photon_flash_1 = "photon-flash-1" + + +class LumaVideoModel(str, Enum): + ray_2 = "ray-2" + ray_flash_2 = "ray-flash-2" + ray_1_6 = "ray-1-6" + + +class LumaAspectRatio(str, Enum): + ratio_1_1 = "1:1" + ratio_16_9 = "16:9" + ratio_9_16 = "9:16" + ratio_4_3 = "4:3" + ratio_3_4 = "3:4" + ratio_21_9 = "21:9" + ratio_9_21 = "9:21" + + +class LumaVideoOutputResolution(str, Enum): + res_540p = "540p" + res_720p = "720p" + res_1080p = "1080p" + res_4k = "4k" + + +class LumaVideoModelOutputDuration(str, Enum): + dur_5s = "5s" + dur_9s = "9s" + + +class LumaGenerationType(str, Enum): + video = 'video' + image = 'image' + + +class LumaState(str, Enum): + queued = "queued" + dreaming = "dreaming" + completed = "completed" + failed = "failed" + + +class LumaAssets(BaseModel): + video: Optional[str] = Field(None, description='The URL of the video') + image: Optional[str] = Field(None, description='The URL of the image') + progress_video: Optional[str] = Field(None, description='The URL of the progress video') + + +class LumaImageRef(BaseModel): + '''Used for image gen''' + url: str = Field(..., description='The URL of the image reference') + weight: confloat(ge=0.0, le=1.0) = Field(..., description='The weight of the image reference') + + +class LumaImageReference(BaseModel): + '''Used for video gen''' + type: Optional[str] = Field('image', description='Input type, defaults to image') + url: str = Field(..., description='The URL of the image') + + +class LumaModifyImageRef(BaseModel): + url: str = Field(..., description='The URL of the image reference') + weight: confloat(ge=0.0, le=1.0) = Field(..., description='The weight of the image reference') + + +class LumaCharacterRef(BaseModel): + identity0: LumaImageIdentity = Field(..., description='The image identity object') + + +class LumaImageIdentity(BaseModel): + images: list[str] = Field(..., description='The URLs of the image identity') + + +class LumaGenerationReference(BaseModel): + type: str = Field('generation', description='Input type, defaults to generation') + id: str = Field(..., description='The ID of the generation') + + +class LumaKeyframes(BaseModel): + frame0: Optional[Union[LumaImageReference, LumaGenerationReference]] = Field(None, description='') + frame1: Optional[Union[LumaImageReference, LumaGenerationReference]] = Field(None, description='') + + +class LumaConceptObject(BaseModel): + key: str = Field(..., description='Camera Concept name') + + +class LumaImageGenerationRequest(BaseModel): + prompt: str = Field(..., description='The prompt of the generation') + model: LumaImageModel = Field(LumaImageModel.photon_1, description='The image model used for the generation') + aspect_ratio: Optional[LumaAspectRatio] = Field(LumaAspectRatio.ratio_16_9, description='The aspect ratio of the generation') + image_ref: Optional[list[LumaImageRef]] = Field(None, description='List of image reference objects') + style_ref: Optional[list[LumaImageRef]] = Field(None, description='List of style reference objects') + character_ref: Optional[LumaCharacterRef] = Field(None, description='The image identity object') + modify_image_ref: Optional[LumaModifyImageRef] = Field(None, description='The modify image reference object') + + +class LumaGenerationRequest(BaseModel): + prompt: str = Field(..., description='The prompt of the generation') + model: LumaVideoModel = Field(LumaVideoModel.ray_2, description='The video model used for the generation') + duration: Optional[LumaVideoModelOutputDuration] = Field(None, description='The duration of the generation') + aspect_ratio: Optional[LumaAspectRatio] = Field(None, description='The aspect ratio of the generation') + resolution: Optional[LumaVideoOutputResolution] = Field(None, description='The resolution of the generation') + loop: Optional[bool] = Field(None, description='Whether to loop the video') + keyframes: Optional[LumaKeyframes] = Field(None, description='The keyframes of the generation') + concepts: Optional[list[LumaConceptObject]] = Field(None, description='Camera Concepts to apply to generation') + + +class LumaGeneration(BaseModel): + id: str = Field(..., description='The ID of the generation') + generation_type: LumaGenerationType = Field(..., description='Generation type, image or video') + state: LumaState = Field(..., description='The state of the generation') + failure_reason: Optional[str] = Field(None, description='The reason for the state of the generation') + created_at: str = Field(..., description='The date and time when the generation was created') + assets: Optional[LumaAssets] = Field(None, description='The assets of the generation') + model: str = Field(..., description='The model used for the generation') + request: Union[LumaGenerationRequest, LumaImageGenerationRequest] = Field(..., description="The request used for the generation") diff --git a/comfy_api_nodes/apis/pixverse_api.py b/comfy_api_nodes/apis/pixverse_api.py new file mode 100644 index 000000000..9bb29c383 --- /dev/null +++ b/comfy_api_nodes/apis/pixverse_api.py @@ -0,0 +1,146 @@ +from __future__ import annotations + +from enum import Enum +from typing import Optional + +from pydantic import BaseModel, Field + + +pixverse_templates = { + "Microwave": 324641385496960, + "Suit Swagger": 328545151283968, + "Anything, Robot": 313358700761536, + "Subject 3 Fever": 327828816843648, + "kiss kiss": 315446315336768, +} + + +class PixverseIO: + TEMPLATE = "PIXVERSE_TEMPLATE" + + +class PixverseStatus(int, Enum): + successful = 1 + generating = 5 + deleted = 6 + contents_moderation = 7 + failed = 8 + + +class PixverseAspectRatio(str, Enum): + ratio_16_9 = "16:9" + ratio_4_3 = "4:3" + ratio_1_1 = "1:1" + ratio_3_4 = "3:4" + ratio_9_16 = "9:16" + + +class PixverseQuality(str, Enum): + res_360p = "360p" + res_540p = "540p" + res_720p = "720p" + res_1080p = "1080p" + + +class PixverseDuration(int, Enum): + dur_5 = 5 + dur_8 = 8 + + +class PixverseMotionMode(str, Enum): + normal = "normal" + fast = "fast" + + +class PixverseStyle(str, Enum): + anime = "anime" + animation_3d = "3d_animation" + clay = "clay" + comic = "comic" + cyberpunk = "cyberpunk" + + +# NOTE: forgoing descriptions for now in return for dev speed +class PixverseTextVideoRequest(BaseModel): + aspect_ratio: PixverseAspectRatio = Field(...) + quality: PixverseQuality = Field(...) + duration: PixverseDuration = Field(...) + model: Optional[str] = Field("v3.5") + motion_mode: Optional[PixverseMotionMode] = Field(PixverseMotionMode.normal) + prompt: str = Field(...) + negative_prompt: Optional[str] = Field(None) + seed: Optional[int] = Field(None) + style: Optional[str] = Field(None) + template_id: Optional[int] = Field(None) + water_mark: Optional[bool] = Field(None) + + +class PixverseImageVideoRequest(BaseModel): + quality: PixverseQuality = Field(...) + duration: PixverseDuration = Field(...) + img_id: int = Field(...) + model: Optional[str] = Field("v3.5") + motion_mode: Optional[PixverseMotionMode] = Field(PixverseMotionMode.normal) + prompt: str = Field(...) + negative_prompt: Optional[str] = Field(None) + seed: Optional[int] = Field(None) + style: Optional[str] = Field(None) + template_id: Optional[int] = Field(None) + water_mark: Optional[bool] = Field(None) + + +class PixverseTransitionVideoRequest(BaseModel): + quality: PixverseQuality = Field(...) + duration: PixverseDuration = Field(...) + first_frame_img: int = Field(...) + last_frame_img: int = Field(...) + model: Optional[str] = Field("v3.5") + motion_mode: Optional[PixverseMotionMode] = Field(PixverseMotionMode.normal) + prompt: str = Field(...) + # negative_prompt: Optional[str] = Field(None) + seed: Optional[int] = Field(None) + # style: Optional[str] = Field(None) + # template_id: Optional[int] = Field(None) + # water_mark: Optional[bool] = Field(None) + + +class PixverseImageUploadResponse(BaseModel): + ErrCode: Optional[int] = None + ErrMsg: Optional[str] = None + Resp: Optional[PixverseImgIdResponseObject] = Field(None, alias='Resp') + + +class PixverseImgIdResponseObject(BaseModel): + img_id: Optional[int] = None + + +class PixverseVideoResponse(BaseModel): + ErrCode: Optional[int] = Field(None) + ErrMsg: Optional[str] = Field(None) + Resp: Optional[PixverseVideoIdResponseObject] = Field(None) + + +class PixverseVideoIdResponseObject(BaseModel): + video_id: int = Field(..., description='Video_id') + + +class PixverseGenerationStatusResponse(BaseModel): + ErrCode: Optional[int] = Field(None) + ErrMsg: Optional[str] = Field(None) + Resp: Optional[PixverseGenerationStatusResponseObject] = Field(None) + + +class PixverseGenerationStatusResponseObject(BaseModel): + create_time: Optional[str] = Field(None) + id: Optional[int] = Field(None) + modify_time: Optional[str] = Field(None) + negative_prompt: Optional[str] = Field(None) + outputHeight: Optional[int] = Field(None) + outputWidth: Optional[int] = Field(None) + prompt: Optional[str] = Field(None) + resolution_ratio: Optional[int] = Field(None) + seed: Optional[int] = Field(None) + size: Optional[int] = Field(None) + status: Optional[int] = Field(None) + style: Optional[str] = Field(None) + url: Optional[str] = Field(None) diff --git a/comfy_api_nodes/apis/recraft_api.py b/comfy_api_nodes/apis/recraft_api.py new file mode 100644 index 000000000..c36d95f24 --- /dev/null +++ b/comfy_api_nodes/apis/recraft_api.py @@ -0,0 +1,262 @@ +from __future__ import annotations + + + +from enum import Enum +from typing import Optional + +from pydantic import BaseModel, Field, conint, confloat + + +class RecraftColor: + def __init__(self, r: int, g: int, b: int): + self.color = [r, g, b] + + def create_api_model(self): + return RecraftColorObject(rgb=self.color) + + +class RecraftColorChain: + def __init__(self): + self.colors: list[RecraftColor] = [] + + def get_first(self): + if len(self.colors) > 0: + return self.colors[0] + return None + + def add(self, color: RecraftColor): + self.colors.append(color) + + def create_api_model(self): + if not self.colors: + return None + colors_api = [x.create_api_model() for x in self.colors] + return colors_api + + def clone(self): + c = RecraftColorChain() + for color in self.colors: + c.add(color) + return c + + def clone_and_merge(self, other: RecraftColorChain): + c = self.clone() + for color in other.colors: + c.add(color) + return c + + +class RecraftControls: + def __init__(self, colors: RecraftColorChain=None, background_color: RecraftColorChain=None, + artistic_level: int=None, no_text: bool=None): + self.colors = colors + self.background_color = background_color + self.artistic_level = artistic_level + self.no_text = no_text + + def create_api_model(self): + if self.colors is None and self.background_color is None and self.artistic_level is None and self.no_text is None: + return None + colors_api = None + background_color_api = None + if self.colors: + colors_api = self.colors.create_api_model() + if self.background_color: + first_background = self.background_color.get_first() + background_color_api = first_background.create_api_model() if first_background else None + + return RecraftControlsObject(colors=colors_api, background_color=background_color_api, + artistic_level=self.artistic_level, no_text=self.no_text) + + +class RecraftStyle: + def __init__(self, style: str=None, substyle: str=None, style_id: str=None): + self.style = style + if substyle == "None": + substyle = None + self.substyle = substyle + self.style_id = style_id + + +class RecraftIO: + STYLEV3 = "RECRAFT_V3_STYLE" + COLOR = "RECRAFT_COLOR" + CONTROLS = "RECRAFT_CONTROLS" + + +class RecraftStyleV3(str, Enum): + #any = 'any' NOTE: this does not work for some reason... why? + realistic_image = 'realistic_image' + digital_illustration = 'digital_illustration' + vector_illustration = 'vector_illustration' + logo_raster = 'logo_raster' + + +def get_v3_substyles(style_v3: str, include_none=True) -> list[str]: + substyles: list[str] = [] + if include_none: + substyles.append("None") + return substyles + dict_recraft_substyles_v3.get(style_v3, []) + + +dict_recraft_substyles_v3 = { + RecraftStyleV3.realistic_image: [ + "b_and_w", + "enterprise", + "evening_light", + "faded_nostalgia", + "forest_life", + "hard_flash", + "hdr", + "motion_blur", + "mystic_naturalism", + "natural_light", + "natural_tones", + "organic_calm", + "real_life_glow", + "retro_realism", + "retro_snapshot", + "studio_portrait", + "urban_drama", + "village_realism", + "warm_folk" + ], + RecraftStyleV3.digital_illustration: [ + "2d_art_poster", + "2d_art_poster_2", + "antiquarian", + "bold_fantasy", + "child_book", + "child_books", + "cover", + "crosshatch", + "digital_engraving", + "engraving_color", + "expressionism", + "freehand_details", + "grain", + "grain_20", + "graphic_intensity", + "hand_drawn", + "hand_drawn_outline", + "handmade_3d", + "hard_comics", + "infantile_sketch", + "long_shadow", + "modern_folk", + "multicolor", + "neon_calm", + "noir", + "nostalgic_pastel", + "outline_details", + "pastel_gradient", + "pastel_sketch", + "pixel_art", + "plastic", + "pop_art", + "pop_renaissance", + "seamless", + "street_art", + "tablet_sketch", + "urban_glow", + "urban_sketching", + "vanilla_dreams", + "young_adult_book", + "young_adult_book_2" + ], + RecraftStyleV3.vector_illustration: [ + "bold_stroke", + "chemistry", + "colored_stencil", + "contour_pop_art", + "cosmics", + "cutout", + "depressive", + "editorial", + "emotional_flat", + "engraving", + "infographical", + "line_art", + "line_circuit", + "linocut", + "marker_outline", + "mosaic", + "naivector", + "roundish_flat", + "seamless", + "segmented_colors", + "sharp_contrast", + "thin", + "vector_photo", + "vivid_shapes" + ], + RecraftStyleV3.logo_raster: [ + "emblem_graffiti", + "emblem_pop_art", + "emblem_punk", + "emblem_stamp", + "emblem_vintage" + ], +} + + +class RecraftModel(str, Enum): + recraftv3 = 'recraftv3' + recraftv2 = 'recraftv2' + + +class RecraftImageSize(str, Enum): + res_1024x1024 = '1024x1024' + res_1365x1024 = '1365x1024' + res_1024x1365 = '1024x1365' + res_1536x1024 = '1536x1024' + res_1024x1536 = '1024x1536' + res_1820x1024 = '1820x1024' + res_1024x1820 = '1024x1820' + res_1024x2048 = '1024x2048' + res_2048x1024 = '2048x1024' + res_1434x1024 = '1434x1024' + res_1024x1434 = '1024x1434' + res_1024x1280 = '1024x1280' + res_1280x1024 = '1280x1024' + res_1024x1707 = '1024x1707' + res_1707x1024 = '1707x1024' + + +class RecraftColorObject(BaseModel): + rgb: list[int] = Field(..., description='An array of 3 integer values in range of 0...255 defining RGB Color Model') + + +class RecraftControlsObject(BaseModel): + colors: Optional[list[RecraftColorObject]] = Field(None, description='An array of preferable colors') + background_color: Optional[RecraftColorObject] = Field(None, description='Use given color as a desired background color') + no_text: Optional[bool] = Field(None, description='Do not embed text layouts') + artistic_level: Optional[conint(ge=0, le=5)] = Field(None, description='Defines artistic tone of your image. At a simple level, the person looks straight at the camera in a static and clean style. Dynamic and eccentric levels introduce movement and creativity. The value should be in range [0..5].') + + +class RecraftImageGenerationRequest(BaseModel): + prompt: str = Field(..., description='The text prompt describing the image to generate') + size: Optional[RecraftImageSize] = Field(None, description='The size of the generated image (e.g., "1024x1024")') + n: conint(ge=1, le=6) = Field(..., description='The number of images to generate') + negative_prompt: Optional[str] = Field(None, description='A text description of undesired elements on an image') + model: Optional[RecraftModel] = Field(RecraftModel.recraftv3, description='The model to use for generation (e.g., "recraftv3")') + style: Optional[str] = Field(None, description='The style to apply to the generated image (e.g., "digital_illustration")') + substyle: Optional[str] = Field(None, description='The substyle to apply to the generated image, depending on the style input') + controls: Optional[RecraftControlsObject] = Field(None, description='A set of custom parameters to tweak generation process') + style_id: Optional[str] = Field(None, description='Use a previously uploaded style as a reference; UUID') + strength: Optional[confloat(ge=0.0, le=1.0)] = Field(None, description='Defines the difference with the original image, should lie in [0, 1], where 0 means almost identical, and 1 means miserable similarity') + random_seed: Optional[int] = Field(None, description="Seed for video generation") + # text_layout + + +class RecraftReturnedObject(BaseModel): + image_id: str = Field(..., description='Unique identifier for the generated image') + url: str = Field(..., description='URL to access the generated image') + + +class RecraftImageGenerationResponse(BaseModel): + created: int = Field(..., description='Unix timestamp when the generation was created') + credits: int = Field(..., description='Number of credits used for the generation') + data: Optional[list[RecraftReturnedObject]] = Field(None, description='Array of generated image information') + image: Optional[RecraftReturnedObject] = Field(None, description='Single generated image') diff --git a/comfy_api_nodes/apis/request_logger.py b/comfy_api_nodes/apis/request_logger.py new file mode 100644 index 000000000..93517ede9 --- /dev/null +++ b/comfy_api_nodes/apis/request_logger.py @@ -0,0 +1,125 @@ +import os +import datetime +import json +import logging +import folder_paths + +# Get the logger instance +logger = logging.getLogger(__name__) + +def get_log_directory(): + """ + Ensures the API log directory exists within ComfyUI's temp directory + and returns its path. + """ + base_temp_dir = folder_paths.get_temp_directory() + log_dir = os.path.join(base_temp_dir, "api_logs") + try: + os.makedirs(log_dir, exist_ok=True) + except Exception as e: + logger.error(f"Error creating API log directory {log_dir}: {e}") + # Fallback to base temp directory if sub-directory creation fails + return base_temp_dir + return log_dir + +def _format_data_for_logging(data): + """Helper to format data (dict, str, bytes) for logging.""" + if isinstance(data, bytes): + try: + return data.decode('utf-8') # Try to decode as text + except UnicodeDecodeError: + return f"[Binary data of length {len(data)} bytes]" + elif isinstance(data, (dict, list)): + try: + return json.dumps(data, indent=2, ensure_ascii=False) + except TypeError: + return str(data) # Fallback for non-serializable objects + return str(data) + +def log_request_response( + operation_id: str, + request_method: str, + request_url: str, + request_headers: dict | None = None, + request_params: dict | None = None, + request_data: any = None, + response_status_code: int | None = None, + response_headers: dict | None = None, + response_content: any = None, + error_message: str | None = None +): + """ + Logs API request and response details to a file in the temp/api_logs directory. + """ + log_dir = get_log_directory() + timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S_%f") + filename = f"{timestamp}_{operation_id.replace('/', '_').replace(':', '_')}.log" + filepath = os.path.join(log_dir, filename) + + log_content = [] + + log_content.append(f"Timestamp: {datetime.datetime.now().isoformat()}") + log_content.append(f"Operation ID: {operation_id}") + log_content.append("-" * 30 + " REQUEST " + "-" * 30) + log_content.append(f"Method: {request_method}") + log_content.append(f"URL: {request_url}") + if request_headers: + log_content.append(f"Headers:\n{_format_data_for_logging(request_headers)}") + if request_params: + log_content.append(f"Params:\n{_format_data_for_logging(request_params)}") + if request_data: + log_content.append(f"Data/Body:\n{_format_data_for_logging(request_data)}") + + log_content.append("\n" + "-" * 30 + " RESPONSE " + "-" * 30) + if response_status_code is not None: + log_content.append(f"Status Code: {response_status_code}") + if response_headers: + log_content.append(f"Headers:\n{_format_data_for_logging(response_headers)}") + if response_content: + log_content.append(f"Content:\n{_format_data_for_logging(response_content)}") + if error_message: + log_content.append(f"Error:\n{error_message}") + + try: + with open(filepath, "w", encoding="utf-8") as f: + f.write("\n".join(log_content)) + logger.debug(f"API log saved to: {filepath}") + except Exception as e: + logger.error(f"Error writing API log to {filepath}: {e}") + +if __name__ == '__main__': + # Example usage (for testing the logger directly) + logger.setLevel(logging.DEBUG) + # Mock folder_paths for direct execution if not running within ComfyUI full context + if not hasattr(folder_paths, 'get_temp_directory'): + class MockFolderPaths: + def get_temp_directory(self): + # Create a local temp dir for testing if needed + p = os.path.join(os.path.dirname(__file__), 'temp_test_logs') + os.makedirs(p, exist_ok=True) + return p + folder_paths = MockFolderPaths() + + log_request_response( + operation_id="test_operation_get", + request_method="GET", + request_url="https://api.example.com/test", + request_headers={"Authorization": "Bearer testtoken"}, + request_params={"param1": "value1"}, + response_status_code=200, + response_content={"message": "Success!"} + ) + log_request_response( + operation_id="test_operation_post_error", + request_method="POST", + request_url="https://api.example.com/submit", + request_data={"key": "value", "nested": {"num": 123}}, + error_message="Connection timed out" + ) + log_request_response( + operation_id="test_binary_response", + request_method="GET", + request_url="https://api.example.com/image.png", + response_status_code=200, + response_content=b'\x89PNG\r\n\x1a\n\x00\x00\x00\rIHDR...' # Sample binary data + ) diff --git a/comfy_api_nodes/apis/rodin_api.py b/comfy_api_nodes/apis/rodin_api.py new file mode 100644 index 000000000..b0cf171fa --- /dev/null +++ b/comfy_api_nodes/apis/rodin_api.py @@ -0,0 +1,57 @@ +from __future__ import annotations + +from enum import Enum +from typing import Optional, List +from pydantic import BaseModel, Field + + +class Rodin3DGenerateRequest(BaseModel): + seed: int = Field(..., description="seed_") + tier: str = Field(..., description="Tier of generation.") + material: str = Field(..., description="The material type.") + quality: str = Field(..., description="The generation quality of the mesh.") + mesh_mode: str = Field(..., description="It controls the type of faces of generated models.") + +class GenerateJobsData(BaseModel): + uuids: List[str] = Field(..., description="str LIST") + subscription_key: str = Field(..., description="subscription key") + +class Rodin3DGenerateResponse(BaseModel): + message: Optional[str] = Field(None, description="Return message.") + prompt: Optional[str] = Field(None, description="Generated Prompt from image.") + submit_time: Optional[str] = Field(None, description="Submit Time") + uuid: Optional[str] = Field(None, description="Task str") + jobs: Optional[GenerateJobsData] = Field(None, description="Details of jobs") + +class JobStatus(str, Enum): + """ + Status for jobs + """ + Done = "Done" + Failed = "Failed" + Generating = "Generating" + Waiting = "Waiting" + +class Rodin3DCheckStatusRequest(BaseModel): + subscription_key: str = Field(..., description="subscription from generate endpoint") + +class JobItem(BaseModel): + uuid: str = Field(..., description="uuid") + status: JobStatus = Field(...,description="Status Currently") + +class Rodin3DCheckStatusResponse(BaseModel): + jobs: List[JobItem] = Field(..., description="Job status List") + +class Rodin3DDownloadRequest(BaseModel): + task_uuid: str = Field(..., description="Task str") + +class RodinResourceItem(BaseModel): + url: str = Field(..., description="Download Url") + name: str = Field(..., description="File name with ext") + +class Rodin3DDownloadResponse(BaseModel): + list: List[RodinResourceItem] = Field(..., description="Source List") + + + + diff --git a/comfy_api_nodes/apis/stability_api.py b/comfy_api_nodes/apis/stability_api.py new file mode 100644 index 000000000..47c87daec --- /dev/null +++ b/comfy_api_nodes/apis/stability_api.py @@ -0,0 +1,127 @@ +from __future__ import annotations + +from enum import Enum +from typing import Optional + +from pydantic import BaseModel, Field, confloat + + +class StabilityFormat(str, Enum): + png = 'png' + jpeg = 'jpeg' + webp = 'webp' + + +class StabilityAspectRatio(str, Enum): + ratio_1_1 = "1:1" + ratio_16_9 = "16:9" + ratio_9_16 = "9:16" + ratio_3_2 = "3:2" + ratio_2_3 = "2:3" + ratio_5_4 = "5:4" + ratio_4_5 = "4:5" + ratio_21_9 = "21:9" + ratio_9_21 = "9:21" + + +def get_stability_style_presets(include_none=True): + presets = [] + if include_none: + presets.append("None") + return presets + [x.value for x in StabilityStylePreset] + + +class StabilityStylePreset(str, Enum): + _3d_model = "3d-model" + analog_film = "analog-film" + anime = "anime" + cinematic = "cinematic" + comic_book = "comic-book" + digital_art = "digital-art" + enhance = "enhance" + fantasy_art = "fantasy-art" + isometric = "isometric" + line_art = "line-art" + low_poly = "low-poly" + modeling_compound = "modeling-compound" + neon_punk = "neon-punk" + origami = "origami" + photographic = "photographic" + pixel_art = "pixel-art" + tile_texture = "tile-texture" + + +class Stability_SD3_5_Model(str, Enum): + sd3_5_large = "sd3.5-large" + # sd3_5_large_turbo = "sd3.5-large-turbo" + sd3_5_medium = "sd3.5-medium" + + +class Stability_SD3_5_GenerationMode(str, Enum): + text_to_image = "text-to-image" + image_to_image = "image-to-image" + + +class StabilityStable3_5Request(BaseModel): + model: str = Field(...) + mode: str = Field(...) + prompt: str = Field(...) + negative_prompt: Optional[str] = Field(None) + aspect_ratio: Optional[str] = Field(None) + seed: Optional[int] = Field(None) + output_format: Optional[str] = Field(StabilityFormat.png.value) + image: Optional[str] = Field(None) + style_preset: Optional[str] = Field(None) + cfg_scale: float = Field(...) + strength: Optional[confloat(ge=0.0, le=1.0)] = Field(None) + + +class StabilityUpscaleConservativeRequest(BaseModel): + prompt: str = Field(...) + negative_prompt: Optional[str] = Field(None) + seed: Optional[int] = Field(None) + output_format: Optional[str] = Field(StabilityFormat.png.value) + image: Optional[str] = Field(None) + creativity: Optional[confloat(ge=0.2, le=0.5)] = Field(None) + + +class StabilityUpscaleCreativeRequest(BaseModel): + prompt: str = Field(...) + negative_prompt: Optional[str] = Field(None) + seed: Optional[int] = Field(None) + output_format: Optional[str] = Field(StabilityFormat.png.value) + image: Optional[str] = Field(None) + creativity: Optional[confloat(ge=0.1, le=0.5)] = Field(None) + style_preset: Optional[str] = Field(None) + + +class StabilityStableUltraRequest(BaseModel): + prompt: str = Field(...) + negative_prompt: Optional[str] = Field(None) + aspect_ratio: Optional[str] = Field(None) + seed: Optional[int] = Field(None) + output_format: Optional[str] = Field(StabilityFormat.png.value) + image: Optional[str] = Field(None) + style_preset: Optional[str] = Field(None) + strength: Optional[confloat(ge=0.0, le=1.0)] = Field(None) + + +class StabilityStableUltraResponse(BaseModel): + image: Optional[str] = Field(None) + finish_reason: Optional[str] = Field(None) + seed: Optional[int] = Field(None) + + +class StabilityResultsGetResponse(BaseModel): + image: Optional[str] = Field(None) + finish_reason: Optional[str] = Field(None) + seed: Optional[int] = Field(None) + id: Optional[str] = Field(None) + name: Optional[str] = Field(None) + errors: Optional[list[str]] = Field(None) + status: Optional[str] = Field(None) + result: Optional[str] = Field(None) + + +class StabilityAsyncResponse(BaseModel): + id: Optional[str] = Field(None) diff --git a/comfy_api_nodes/apis/tripo_api.py b/comfy_api_nodes/apis/tripo_api.py new file mode 100644 index 000000000..626e8d277 --- /dev/null +++ b/comfy_api_nodes/apis/tripo_api.py @@ -0,0 +1,275 @@ +from __future__ import annotations +from comfy_api_nodes.apis import ( + TripoModelVersion, + TripoTextureQuality, +) +from enum import Enum +from typing import Optional, List, Dict, Any, Union + +from pydantic import BaseModel, Field, RootModel + +class TripoStyle(str, Enum): + PERSON_TO_CARTOON = "person:person2cartoon" + ANIMAL_VENOM = "animal:venom" + OBJECT_CLAY = "object:clay" + OBJECT_STEAMPUNK = "object:steampunk" + OBJECT_CHRISTMAS = "object:christmas" + OBJECT_BARBIE = "object:barbie" + GOLD = "gold" + ANCIENT_BRONZE = "ancient_bronze" + NONE = "None" + +class TripoTaskType(str, Enum): + TEXT_TO_MODEL = "text_to_model" + IMAGE_TO_MODEL = "image_to_model" + MULTIVIEW_TO_MODEL = "multiview_to_model" + TEXTURE_MODEL = "texture_model" + REFINE_MODEL = "refine_model" + ANIMATE_PRERIGCHECK = "animate_prerigcheck" + ANIMATE_RIG = "animate_rig" + ANIMATE_RETARGET = "animate_retarget" + STYLIZE_MODEL = "stylize_model" + CONVERT_MODEL = "convert_model" + +class TripoTextureAlignment(str, Enum): + ORIGINAL_IMAGE = "original_image" + GEOMETRY = "geometry" + +class TripoOrientation(str, Enum): + ALIGN_IMAGE = "align_image" + DEFAULT = "default" + +class TripoOutFormat(str, Enum): + GLB = "glb" + FBX = "fbx" + +class TripoTopology(str, Enum): + BIP = "bip" + QUAD = "quad" + +class TripoSpec(str, Enum): + MIXAMO = "mixamo" + TRIPO = "tripo" + +class TripoAnimation(str, Enum): + IDLE = "preset:idle" + WALK = "preset:walk" + CLIMB = "preset:climb" + JUMP = "preset:jump" + RUN = "preset:run" + SLASH = "preset:slash" + SHOOT = "preset:shoot" + HURT = "preset:hurt" + FALL = "preset:fall" + TURN = "preset:turn" + +class TripoStylizeStyle(str, Enum): + LEGO = "lego" + VOXEL = "voxel" + VORONOI = "voronoi" + MINECRAFT = "minecraft" + +class TripoConvertFormat(str, Enum): + GLTF = "GLTF" + USDZ = "USDZ" + FBX = "FBX" + OBJ = "OBJ" + STL = "STL" + _3MF = "3MF" + +class TripoTextureFormat(str, Enum): + BMP = "BMP" + DPX = "DPX" + HDR = "HDR" + JPEG = "JPEG" + OPEN_EXR = "OPEN_EXR" + PNG = "PNG" + TARGA = "TARGA" + TIFF = "TIFF" + WEBP = "WEBP" + +class TripoTaskStatus(str, Enum): + QUEUED = "queued" + RUNNING = "running" + SUCCESS = "success" + FAILED = "failed" + CANCELLED = "cancelled" + UNKNOWN = "unknown" + BANNED = "banned" + EXPIRED = "expired" + +class TripoFileTokenReference(BaseModel): + type: Optional[str] = Field(None, description='The type of the reference') + file_token: str + +class TripoUrlReference(BaseModel): + type: Optional[str] = Field(None, description='The type of the reference') + url: str + +class TripoObjectStorage(BaseModel): + bucket: str + key: str + +class TripoObjectReference(BaseModel): + type: str + object: TripoObjectStorage + +class TripoFileEmptyReference(BaseModel): + pass + +class TripoFileReference(RootModel): + root: Union[TripoFileTokenReference, TripoUrlReference, TripoObjectReference, TripoFileEmptyReference] + +class TripoGetStsTokenRequest(BaseModel): + format: str = Field(..., description='The format of the image') + +class TripoTextToModelRequest(BaseModel): + type: TripoTaskType = Field(TripoTaskType.TEXT_TO_MODEL, description='Type of task') + prompt: str = Field(..., description='The text prompt describing the model to generate', max_length=1024) + negative_prompt: Optional[str] = Field(None, description='The negative text prompt', max_length=1024) + model_version: Optional[TripoModelVersion] = TripoModelVersion.V2_5 + face_limit: Optional[int] = Field(None, description='The number of faces to limit the generation to') + texture: Optional[bool] = Field(True, description='Whether to apply texture to the generated model') + pbr: Optional[bool] = Field(True, description='Whether to apply PBR to the generated model') + image_seed: Optional[int] = Field(None, description='The seed for the text') + model_seed: Optional[int] = Field(None, description='The seed for the model') + texture_seed: Optional[int] = Field(None, description='The seed for the texture') + texture_quality: Optional[TripoTextureQuality] = TripoTextureQuality.standard + 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') + +class TripoImageToModelRequest(BaseModel): + type: TripoTaskType = Field(TripoTaskType.IMAGE_TO_MODEL, description='Type of task') + file: TripoFileReference = Field(..., description='The file reference to convert to a model') + model_version: Optional[TripoModelVersion] = Field(None, description='The model version to use for generation') + face_limit: Optional[int] = Field(None, description='The number of faces to limit the generation to') + texture: Optional[bool] = Field(True, description='Whether to apply texture to the generated model') + pbr: Optional[bool] = Field(True, description='Whether to apply PBR to the generated model') + model_seed: Optional[int] = Field(None, description='The seed for the model') + texture_seed: Optional[int] = Field(None, description='The seed for the texture') + texture_quality: Optional[TripoTextureQuality] = TripoTextureQuality.standard + texture_alignment: Optional[TripoTextureAlignment] = Field(TripoTextureAlignment.ORIGINAL_IMAGE, description='The texture alignment method') + style: Optional[TripoStyle] = Field(None, description='The style to apply to the generated model') + auto_size: Optional[bool] = Field(False, description='Whether to auto-size the model') + orientation: Optional[TripoOrientation] = TripoOrientation.DEFAULT + quad: Optional[bool] = Field(False, description='Whether to apply quad to the generated model') + +class TripoMultiviewToModelRequest(BaseModel): + type: TripoTaskType = TripoTaskType.MULTIVIEW_TO_MODEL + files: List[TripoFileReference] = Field(..., description='The file references to convert to a model') + model_version: Optional[TripoModelVersion] = Field(None, description='The model version to use for generation') + orthographic_projection: Optional[bool] = Field(False, description='Whether to use orthographic projection') + face_limit: Optional[int] = Field(None, description='The number of faces to limit the generation to') + texture: Optional[bool] = Field(True, description='Whether to apply texture to the generated model') + pbr: Optional[bool] = Field(True, description='Whether to apply PBR to the generated model') + model_seed: Optional[int] = Field(None, description='The seed for the model') + texture_seed: Optional[int] = Field(None, description='The seed for the texture') + texture_quality: Optional[TripoTextureQuality] = TripoTextureQuality.standard + texture_alignment: Optional[TripoTextureAlignment] = TripoTextureAlignment.ORIGINAL_IMAGE + auto_size: Optional[bool] = Field(False, description='Whether to auto-size the model') + orientation: Optional[TripoOrientation] = Field(TripoOrientation.DEFAULT, description='The orientation for the model') + quad: Optional[bool] = Field(False, description='Whether to apply quad to the generated model') + +class TripoTextureModelRequest(BaseModel): + type: TripoTaskType = Field(TripoTaskType.TEXTURE_MODEL, description='Type of task') + original_model_task_id: str = Field(..., description='The task ID of the original model') + texture: Optional[bool] = Field(True, description='Whether to apply texture to the model') + pbr: Optional[bool] = Field(True, description='Whether to apply PBR to the model') + model_seed: Optional[int] = Field(None, description='The seed for the model') + texture_seed: Optional[int] = Field(None, description='The seed for the texture') + texture_quality: Optional[TripoTextureQuality] = Field(None, description='The quality of the texture') + texture_alignment: Optional[TripoTextureAlignment] = Field(TripoTextureAlignment.ORIGINAL_IMAGE, description='The texture alignment method') + +class TripoRefineModelRequest(BaseModel): + type: TripoTaskType = Field(TripoTaskType.REFINE_MODEL, description='Type of task') + draft_model_task_id: str = Field(..., description='The task ID of the draft model') + +class TripoAnimatePrerigcheckRequest(BaseModel): + type: TripoTaskType = Field(TripoTaskType.ANIMATE_PRERIGCHECK, description='Type of task') + original_model_task_id: str = Field(..., description='The task ID of the original model') + +class TripoAnimateRigRequest(BaseModel): + type: TripoTaskType = Field(TripoTaskType.ANIMATE_RIG, description='Type of task') + original_model_task_id: str = Field(..., description='The task ID of the original model') + out_format: Optional[TripoOutFormat] = Field(TripoOutFormat.GLB, description='The output format') + spec: Optional[TripoSpec] = Field(TripoSpec.TRIPO, description='The specification for rigging') + +class TripoAnimateRetargetRequest(BaseModel): + type: TripoTaskType = Field(TripoTaskType.ANIMATE_RETARGET, description='Type of task') + original_model_task_id: str = Field(..., description='The task ID of the original model') + animation: TripoAnimation = Field(..., description='The animation to apply') + out_format: Optional[TripoOutFormat] = Field(TripoOutFormat.GLB, description='The output format') + bake_animation: Optional[bool] = Field(True, description='Whether to bake the animation') + +class TripoStylizeModelRequest(BaseModel): + type: TripoTaskType = Field(TripoTaskType.STYLIZE_MODEL, description='Type of task') + style: TripoStylizeStyle = Field(..., description='The style to apply to the model') + original_model_task_id: str = Field(..., description='The task ID of the original model') + block_size: Optional[int] = Field(80, description='The block size for stylization') + +class TripoConvertModelRequest(BaseModel): + type: TripoTaskType = Field(TripoTaskType.CONVERT_MODEL, description='Type of task') + format: TripoConvertFormat = Field(..., description='The format to convert to') + original_model_task_id: str = Field(..., description='The task ID of the original model') + quad: Optional[bool] = Field(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') + 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') + +class TripoTaskRequest(RootModel): + root: Union[ + TripoTextToModelRequest, + TripoImageToModelRequest, + TripoMultiviewToModelRequest, + TripoTextureModelRequest, + TripoRefineModelRequest, + TripoAnimatePrerigcheckRequest, + TripoAnimateRigRequest, + TripoAnimateRetargetRequest, + TripoStylizeModelRequest, + TripoConvertModelRequest + ] + +class TripoTaskOutput(BaseModel): + model: Optional[str] = Field(None, description='URL to the model') + base_model: Optional[str] = Field(None, description='URL to the base model') + pbr_model: Optional[str] = Field(None, description='URL to the PBR model') + rendered_image: Optional[str] = Field(None, description='URL to the rendered image') + riggable: Optional[bool] = Field(None, description='Whether the model is riggable') + +class TripoTask(BaseModel): + task_id: str = Field(..., description='The task ID') + type: Optional[str] = Field(None, description='The type of task') + status: Optional[TripoTaskStatus] = Field(None, description='The status of the task') + input: Optional[Dict[str, Any]] = Field(None, description='The input parameters for the task') + output: Optional[TripoTaskOutput] = Field(None, description='The output of the task') + progress: Optional[int] = Field(None, description='The progress of the task', ge=0, le=100) + create_time: Optional[int] = Field(None, description='The creation time of the task') + running_left_time: Optional[int] = Field(None, description='The estimated time left for the task') + queue_position: Optional[int] = Field(None, description='The position in the queue') + +class TripoTaskResponse(BaseModel): + code: int = Field(0, description='The response code') + data: TripoTask = Field(..., description='The task data') + +class TripoGeneralResponse(BaseModel): + code: int = Field(0, description='The response code') + data: Dict[str, str] = Field(..., description='The task ID data') + +class TripoBalanceData(BaseModel): + balance: float = Field(..., description='The account balance') + frozen: float = Field(..., description='The frozen balance') + +class TripoBalanceResponse(BaseModel): + code: int = Field(0, description='The response code') + data: TripoBalanceData = Field(..., description='The balance data') + +class TripoErrorResponse(BaseModel): + code: int = Field(..., description='The error code') + message: str = Field(..., description='The error message') + suggestion: str = Field(..., description='The suggestion for fixing the error') diff --git a/comfy_api_nodes/canary.py b/comfy_api_nodes/canary.py new file mode 100644 index 000000000..4df7590b6 --- /dev/null +++ b/comfy_api_nodes/canary.py @@ -0,0 +1,10 @@ +import av + +ver = av.__version__.split(".") +if int(ver[0]) < 14: + raise Exception("INSTALL NEW VERSION OF PYAV TO USE API NODES.") + +if int(ver[0]) == 14 and int(ver[1]) < 2: + raise Exception("INSTALL NEW VERSION OF PYAV TO USE API NODES.") + +NODE_CLASS_MAPPINGS = {} diff --git a/comfy_api_nodes/mapper_utils.py b/comfy_api_nodes/mapper_utils.py new file mode 100644 index 000000000..6fab8f4bb --- /dev/null +++ b/comfy_api_nodes/mapper_utils.py @@ -0,0 +1,116 @@ +from enum import Enum + +from pydantic.fields import FieldInfo +from pydantic import BaseModel +from pydantic_core import PydanticUndefined + +from comfy.comfy_types.node_typing import IO, InputTypeOptions + +NodeInput = tuple[IO, InputTypeOptions] + + +def _create_base_config(field_info: FieldInfo) -> InputTypeOptions: + config = {} + if hasattr(field_info, "default") and field_info.default is not PydanticUndefined: + config["default"] = field_info.default + if hasattr(field_info, "description") and field_info.description is not None: + config["tooltip"] = field_info.description + return config + + +def _get_number_constraints_config(field_info: FieldInfo) -> dict: + config = {} + if hasattr(field_info, "metadata"): + metadata = field_info.metadata + for constraint in metadata: + if hasattr(constraint, "ge"): + config["min"] = constraint.ge + if hasattr(constraint, "le"): + config["max"] = constraint.le + if hasattr(constraint, "multiple_of"): + config["step"] = constraint.multiple_of + return config + + +def _model_field_to_image_input(field_info: FieldInfo, **kwargs) -> NodeInput: + return IO.IMAGE, { + **_create_base_config(field_info), + **kwargs, + } + + +def _model_field_to_string_input(field_info: FieldInfo, **kwargs) -> NodeInput: + return IO.STRING, { + **_create_base_config(field_info), + **kwargs, + } + + +def _model_field_to_float_input(field_info: FieldInfo, **kwargs) -> NodeInput: + return IO.FLOAT, { + **_create_base_config(field_info), + **_get_number_constraints_config(field_info), + **kwargs, + } + + +def _model_field_to_int_input(field_info: FieldInfo, **kwargs) -> NodeInput: + return IO.INT, { + **_create_base_config(field_info), + **_get_number_constraints_config(field_info), + **kwargs, + } + + +def _model_field_to_combo_input( + field_info: FieldInfo, enum_type: type[Enum] = None, **kwargs +) -> NodeInput: + combo_config = {} + if enum_type is not None: + combo_config["options"] = [option.value for option in enum_type] + combo_config = { + **combo_config, + **_create_base_config(field_info), + **kwargs, + } + return IO.COMBO, combo_config + + +def model_field_to_node_input( + input_type: IO, base_model: type[BaseModel], field_name: str, **kwargs +) -> NodeInput: + """ + Maps a field from a Pydantic model to a Comfy node input. + + Args: + input_type: The type of the input. + base_model: The Pydantic model to map the field from. + field_name: The name of the field to map. + **kwargs: Additional key/values to include in the input options. + + Note: + For combo inputs, pass an `Enum` to the `enum_type` keyword argument to populate the options automatically. + + Example: + >>> model_field_to_node_input(IO.STRING, MyModel, "my_field", multiline=True) + >>> model_field_to_node_input(IO.COMBO, MyModel, "my_field", enum_type=MyEnum) + >>> model_field_to_node_input(IO.FLOAT, MyModel, "my_field", slider=True) + """ + field_info: FieldInfo = base_model.model_fields[field_name] + result: NodeInput + + if input_type == IO.IMAGE: + result = _model_field_to_image_input(field_info, **kwargs) + elif input_type == IO.STRING: + result = _model_field_to_string_input(field_info, **kwargs) + elif input_type == IO.FLOAT: + result = _model_field_to_float_input(field_info, **kwargs) + elif input_type == IO.INT: + result = _model_field_to_int_input(field_info, **kwargs) + elif input_type == IO.COMBO: + result = _model_field_to_combo_input(field_info, **kwargs) + else: + message = f"Invalid input type: {input_type}" + raise ValueError(message) + + return result diff --git a/comfy_api_nodes/nodes_bfl.py b/comfy_api_nodes/nodes_bfl.py new file mode 100644 index 000000000..010564704 --- /dev/null +++ b/comfy_api_nodes/nodes_bfl.py @@ -0,0 +1,1088 @@ +import io +from inspect import cleandoc +from typing import Union, Optional +from comfy.comfy_types.node_typing import IO, ComfyNodeABC +from comfy_api_nodes.apis.bfl_api import ( + BFLStatus, + BFLFluxExpandImageRequest, + BFLFluxFillImageRequest, + BFLFluxCannyImageRequest, + BFLFluxDepthImageRequest, + BFLFluxProGenerateRequest, + BFLFluxKontextProGenerateRequest, + BFLFluxProUltraGenerateRequest, + BFLFluxProGenerateResponse, +) +from comfy_api_nodes.apis.client import ( + ApiEndpoint, + HttpMethod, + SynchronousOperation, +) +from comfy_api_nodes.apinode_utils import ( + downscale_image_tensor, + validate_aspect_ratio, + process_image_response, + resize_mask_to_image, + validate_string, +) + +import numpy as np +from PIL import Image +import requests +import torch +import base64 +import time +from server import PromptServer + + +def convert_mask_to_image(mask: torch.Tensor): + """ + Make mask have the expected amount of dims (4) and channels (3) to be recognized as an image. + """ + mask = mask.unsqueeze(-1) + mask = torch.cat([mask]*3, dim=-1) + return mask + + +def handle_bfl_synchronous_operation( + operation: SynchronousOperation, + timeout_bfl_calls=360, + node_id: Union[str, None] = None, +): + response_api: BFLFluxProGenerateResponse = operation.execute() + return _poll_until_generated( + response_api.polling_url, timeout=timeout_bfl_calls, node_id=node_id + ) + + +def _poll_until_generated( + polling_url: str, timeout=360, node_id: Union[str, None] = None +): + # used bfl-comfy-nodes to verify code implementation: + # https://github.com/black-forest-labs/bfl-comfy-nodes/tree/main + start_time = time.time() + retries_404 = 0 + max_retries_404 = 5 + retry_404_seconds = 2 + retry_202_seconds = 2 + retry_pending_seconds = 1 + request = requests.Request(method=HttpMethod.GET, url=polling_url) + # NOTE: should True loop be replaced with checking if workflow has been interrupted? + while True: + if node_id: + time_elapsed = time.time() - start_time + PromptServer.instance.send_progress_text( + f"Generating ({time_elapsed:.0f}s)", node_id + ) + + response = requests.Session().send(request.prepare()) + if response.status_code == 200: + result = response.json() + if result["status"] == BFLStatus.ready: + img_url = result["result"]["sample"] + if node_id: + PromptServer.instance.send_progress_text( + f"Result URL: {img_url}", node_id + ) + img_response = requests.get(img_url) + return process_image_response(img_response) + elif result["status"] in [ + BFLStatus.request_moderated, + BFLStatus.content_moderated, + ]: + status = result["status"] + raise Exception( + f"BFL API did not return an image due to: {status}." + ) + elif result["status"] == BFLStatus.error: + raise Exception(f"BFL API encountered an error: {result}.") + elif result["status"] == BFLStatus.pending: + time.sleep(retry_pending_seconds) + continue + elif response.status_code == 404: + if retries_404 < max_retries_404: + retries_404 += 1 + time.sleep(retry_404_seconds) + continue + raise Exception( + f"BFL API could not find task after {max_retries_404} tries." + ) + elif response.status_code == 202: + time.sleep(retry_202_seconds) + elif time.time() - start_time > timeout: + raise Exception( + f"BFL API experienced a timeout; could not return request under {timeout} seconds." + ) + else: + raise Exception(f"BFL API encountered an error: {response.json()}") + +def convert_image_to_base64(image: torch.Tensor): + scaled_image = downscale_image_tensor(image, total_pixels=2048 * 2048) + # remove batch dimension if present + if len(scaled_image.shape) > 3: + scaled_image = scaled_image[0] + image_np = (scaled_image.numpy() * 255).astype(np.uint8) + img = Image.fromarray(image_np) + img_byte_arr = io.BytesIO() + img.save(img_byte_arr, format="PNG") + return base64.b64encode(img_byte_arr.getvalue()).decode() + + +class FluxProUltraImageNode(ComfyNodeABC): + """ + Generates images using Flux Pro 1.1 Ultra via api based on prompt and resolution. + """ + + MINIMUM_RATIO = 1 / 4 + MAXIMUM_RATIO = 4 / 1 + MINIMUM_RATIO_STR = "1:4" + MAXIMUM_RATIO_STR = "4:1" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Prompt for the image generation", + }, + ), + "prompt_upsampling": ( + IO.BOOLEAN, + { + "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).", + }, + ), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 0xFFFFFFFFFFFFFFFF, + "control_after_generate": True, + "tooltip": "The random seed used for creating the noise.", + }, + ), + "aspect_ratio": ( + IO.STRING, + { + "default": "16:9", + "tooltip": "Aspect ratio of image; must be between 1:4 and 4:1.", + }, + ), + "raw": ( + IO.BOOLEAN, + { + "default": False, + "tooltip": "When True, generate less processed, more natural-looking images.", + }, + ), + }, + "optional": { + "image_prompt": (IO.IMAGE,), + "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.", + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + @classmethod + def VALIDATE_INPUTS(cls, aspect_ratio: str): + try: + validate_aspect_ratio( + aspect_ratio, + minimum_ratio=cls.MINIMUM_RATIO, + maximum_ratio=cls.MAXIMUM_RATIO, + minimum_ratio_str=cls.MINIMUM_RATIO_STR, + maximum_ratio_str=cls.MAXIMUM_RATIO_STR, + ) + except Exception as e: + return str(e) + return True + + RETURN_TYPES = (IO.IMAGE,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/image/BFL" + + def api_call( + self, + prompt: str, + aspect_ratio: str, + prompt_upsampling=False, + raw=False, + seed=0, + image_prompt=None, + image_prompt_strength=0.1, + unique_id: Union[str, None] = None, + **kwargs, + ): + if image_prompt is None: + validate_string(prompt, strip_whitespace=False) + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/bfl/flux-pro-1.1-ultra/generate", + method=HttpMethod.POST, + request_model=BFLFluxProUltraGenerateRequest, + response_model=BFLFluxProGenerateResponse, + ), + request=BFLFluxProUltraGenerateRequest( + prompt=prompt, + prompt_upsampling=prompt_upsampling, + seed=seed, + aspect_ratio=validate_aspect_ratio( + aspect_ratio, + minimum_ratio=self.MINIMUM_RATIO, + maximum_ratio=self.MAXIMUM_RATIO, + minimum_ratio_str=self.MINIMUM_RATIO_STR, + maximum_ratio_str=self.MAXIMUM_RATIO_STR, + ), + raw=raw, + image_prompt=( + image_prompt + if image_prompt is None + else convert_image_to_base64(image_prompt) + ), + image_prompt_strength=( + None if image_prompt is None else round(image_prompt_strength, 2) + ), + ), + auth_kwargs=kwargs, + ) + output_image = handle_bfl_synchronous_operation(operation, node_id=unique_id) + return (output_image,) + + +class FluxKontextProImageNode(ComfyNodeABC): + """ + Edits images using Flux.1 Kontext [pro] via api based on prompt and aspect ratio. + """ + + MINIMUM_RATIO = 1 / 4 + MAXIMUM_RATIO = 4 / 1 + MINIMUM_RATIO_STR = "1:4" + MAXIMUM_RATIO_STR = "4:1" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Prompt for the image generation - specify what and how to edit.", + }, + ), + "aspect_ratio": ( + IO.STRING, + { + "default": "16:9", + "tooltip": "Aspect ratio of image; must be between 1:4 and 4:1.", + }, + ), + "guidance": ( + IO.FLOAT, + { + "default": 3.0, + "min": 0.1, + "max": 99.0, + "step": 0.1, + "tooltip": "Guidance strength for the image generation process" + }, + ), + "steps": ( + IO.INT, + { + "default": 50, + "min": 1, + "max": 150, + "tooltip": "Number of steps for the image generation process" + }, + ), + "seed": ( + IO.INT, + { + "default": 1234, + "min": 0, + "max": 0xFFFFFFFFFFFFFFFF, + "control_after_generate": True, + "tooltip": "The random seed used for creating the noise.", + }, + ), + "prompt_upsampling": ( + IO.BOOLEAN, + { + "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).", + }, + ), + }, + "optional": { + "input_image": (IO.IMAGE,), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + @classmethod + def VALIDATE_INPUTS(cls, aspect_ratio: str): + try: + validate_aspect_ratio( + aspect_ratio, + minimum_ratio=cls.MINIMUM_RATIO, + maximum_ratio=cls.MAXIMUM_RATIO, + minimum_ratio_str=cls.MINIMUM_RATIO_STR, + maximum_ratio_str=cls.MAXIMUM_RATIO_STR, + ) + except Exception as e: + return str(e) + return True + + RETURN_TYPES = (IO.IMAGE,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/image/BFL" + + BFL_PATH = "/proxy/bfl/flux-kontext-pro/generate" + + def api_call( + self, + prompt: str, + aspect_ratio: str, + guidance: float, + steps: int, + input_image: Optional[torch.Tensor]=None, + seed=0, + prompt_upsampling=False, + unique_id: Union[str, None] = None, + **kwargs, + ): + if input_image is None: + validate_string(prompt, strip_whitespace=False) + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path=self.BFL_PATH, + method=HttpMethod.POST, + request_model=BFLFluxKontextProGenerateRequest, + response_model=BFLFluxProGenerateResponse, + ), + request=BFLFluxKontextProGenerateRequest( + prompt=prompt, + prompt_upsampling=prompt_upsampling, + guidance=round(guidance, 1), + steps=steps, + seed=seed, + aspect_ratio=validate_aspect_ratio( + aspect_ratio, + minimum_ratio=self.MINIMUM_RATIO, + maximum_ratio=self.MAXIMUM_RATIO, + minimum_ratio_str=self.MINIMUM_RATIO_STR, + maximum_ratio_str=self.MAXIMUM_RATIO_STR, + ), + input_image=( + input_image + if input_image is None + else convert_image_to_base64(input_image) + ) + ), + auth_kwargs=kwargs, + ) + output_image = handle_bfl_synchronous_operation(operation, node_id=unique_id) + return (output_image,) + + +class FluxKontextMaxImageNode(FluxKontextProImageNode): + """ + Edits images using Flux.1 Kontext [max] via api based on prompt and aspect ratio. + """ + + DESCRIPTION = cleandoc(__doc__ or "") + BFL_PATH = "/proxy/bfl/flux-kontext-max/generate" + + +class FluxProImageNode(ComfyNodeABC): + """ + Generates images synchronously based on prompt and resolution. + """ + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Prompt for the image generation", + }, + ), + "prompt_upsampling": ( + IO.BOOLEAN, + { + "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).", + }, + ), + "width": ( + IO.INT, + { + "default": 1024, + "min": 256, + "max": 1440, + "step": 32, + }, + ), + "height": ( + IO.INT, + { + "default": 768, + "min": 256, + "max": 1440, + "step": 32, + }, + ), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 0xFFFFFFFFFFFFFFFF, + "control_after_generate": True, + "tooltip": "The random seed used for creating the noise.", + }, + ), + }, + "optional": { + "image_prompt": (IO.IMAGE,), + # "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.", + # }, + # ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = (IO.IMAGE,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/image/BFL" + + def api_call( + self, + prompt: str, + prompt_upsampling, + width: int, + height: int, + seed=0, + image_prompt=None, + # image_prompt_strength=0.1, + unique_id: Union[str, None] = None, + **kwargs, + ): + image_prompt = ( + image_prompt + if image_prompt is None + else convert_image_to_base64(image_prompt) + ) + + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/bfl/flux-pro-1.1/generate", + method=HttpMethod.POST, + request_model=BFLFluxProGenerateRequest, + response_model=BFLFluxProGenerateResponse, + ), + request=BFLFluxProGenerateRequest( + prompt=prompt, + prompt_upsampling=prompt_upsampling, + width=width, + height=height, + seed=seed, + image_prompt=image_prompt, + ), + auth_kwargs=kwargs, + ) + output_image = handle_bfl_synchronous_operation(operation, node_id=unique_id) + return (output_image,) + + +class FluxProExpandNode(ComfyNodeABC): + """ + Outpaints image based on prompt. + """ + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "image": (IO.IMAGE,), + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Prompt for the image generation", + }, + ), + "prompt_upsampling": ( + IO.BOOLEAN, + { + "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).", + }, + ), + "top": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 2048, + "tooltip": "Number of pixels to expand at the top of the image" + }, + ), + "bottom": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 2048, + "tooltip": "Number of pixels to expand at the bottom of the image" + }, + ), + "left": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 2048, + "tooltip": "Number of pixels to expand at the left side of the image" + }, + ), + "right": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 2048, + "tooltip": "Number of pixels to expand at the right side of the image" + }, + ), + "guidance": ( + IO.FLOAT, + { + "default": 60, + "min": 1.5, + "max": 100, + "tooltip": "Guidance strength for the image generation process" + }, + ), + "steps": ( + IO.INT, + { + "default": 50, + "min": 15, + "max": 50, + "tooltip": "Number of steps for the image generation process" + }, + ), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 0xFFFFFFFFFFFFFFFF, + "control_after_generate": True, + "tooltip": "The random seed used for creating the noise.", + }, + ), + }, + "optional": {}, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = (IO.IMAGE,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/image/BFL" + + def api_call( + self, + image: torch.Tensor, + prompt: str, + prompt_upsampling: bool, + top: int, + bottom: int, + left: int, + right: int, + steps: int, + guidance: float, + seed=0, + unique_id: Union[str, None] = None, + **kwargs, + ): + image = convert_image_to_base64(image) + + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/bfl/flux-pro-1.0-expand/generate", + method=HttpMethod.POST, + request_model=BFLFluxExpandImageRequest, + response_model=BFLFluxProGenerateResponse, + ), + request=BFLFluxExpandImageRequest( + prompt=prompt, + prompt_upsampling=prompt_upsampling, + top=top, + bottom=bottom, + left=left, + right=right, + steps=steps, + guidance=guidance, + seed=seed, + image=image, + ), + auth_kwargs=kwargs, + ) + output_image = handle_bfl_synchronous_operation(operation, node_id=unique_id) + return (output_image,) + + + +class FluxProFillNode(ComfyNodeABC): + """ + Inpaints image based on mask and prompt. + """ + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "image": (IO.IMAGE,), + "mask": (IO.MASK,), + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Prompt for the image generation", + }, + ), + "prompt_upsampling": ( + IO.BOOLEAN, + { + "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).", + }, + ), + "guidance": ( + IO.FLOAT, + { + "default": 60, + "min": 1.5, + "max": 100, + "tooltip": "Guidance strength for the image generation process" + }, + ), + "steps": ( + IO.INT, + { + "default": 50, + "min": 15, + "max": 50, + "tooltip": "Number of steps for the image generation process" + }, + ), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 0xFFFFFFFFFFFFFFFF, + "control_after_generate": True, + "tooltip": "The random seed used for creating the noise.", + }, + ), + }, + "optional": {}, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = (IO.IMAGE,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/image/BFL" + + def api_call( + self, + image: torch.Tensor, + mask: torch.Tensor, + prompt: str, + prompt_upsampling: bool, + steps: int, + guidance: float, + seed=0, + unique_id: Union[str, None] = None, + **kwargs, + ): + # prepare mask + mask = resize_mask_to_image(mask, image) + mask = convert_image_to_base64(convert_mask_to_image(mask)) + # make sure image will have alpha channel removed + image = convert_image_to_base64(image[:, :, :, :3]) + + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/bfl/flux-pro-1.0-fill/generate", + method=HttpMethod.POST, + request_model=BFLFluxFillImageRequest, + response_model=BFLFluxProGenerateResponse, + ), + request=BFLFluxFillImageRequest( + prompt=prompt, + prompt_upsampling=prompt_upsampling, + steps=steps, + guidance=guidance, + seed=seed, + image=image, + mask=mask, + ), + auth_kwargs=kwargs, + ) + output_image = handle_bfl_synchronous_operation(operation, node_id=unique_id) + return (output_image,) + + +class FluxProCannyNode(ComfyNodeABC): + """ + Generate image using a control image (canny). + """ + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "control_image": (IO.IMAGE,), + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Prompt for the image generation", + }, + ), + "prompt_upsampling": ( + IO.BOOLEAN, + { + "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).", + }, + ), + "canny_low_threshold": ( + IO.FLOAT, + { + "default": 0.1, + "min": 0.01, + "max": 0.99, + "step": 0.01, + "tooltip": "Low threshold for Canny edge detection; ignored if skip_processing is True" + }, + ), + "canny_high_threshold": ( + IO.FLOAT, + { + "default": 0.4, + "min": 0.01, + "max": 0.99, + "step": 0.01, + "tooltip": "High threshold for Canny edge detection; ignored if skip_processing is True" + }, + ), + "skip_preprocessing": ( + IO.BOOLEAN, + { + "default": False, + "tooltip": "Whether to skip preprocessing; set to True if control_image already is canny-fied, False if it is a raw image.", + }, + ), + "guidance": ( + IO.FLOAT, + { + "default": 30, + "min": 1, + "max": 100, + "tooltip": "Guidance strength for the image generation process" + }, + ), + "steps": ( + IO.INT, + { + "default": 50, + "min": 15, + "max": 50, + "tooltip": "Number of steps for the image generation process" + }, + ), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 0xFFFFFFFFFFFFFFFF, + "control_after_generate": True, + "tooltip": "The random seed used for creating the noise.", + }, + ), + }, + "optional": {}, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = (IO.IMAGE,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/image/BFL" + + def api_call( + self, + control_image: torch.Tensor, + prompt: str, + prompt_upsampling: bool, + canny_low_threshold: float, + canny_high_threshold: float, + skip_preprocessing: bool, + steps: int, + guidance: float, + seed=0, + unique_id: Union[str, None] = None, + **kwargs, + ): + control_image = convert_image_to_base64(control_image[:, :, :, :3]) + preprocessed_image = None + + # scale canny threshold between 0-500, to match BFL's API + def scale_value(value: float, min_val=0, max_val=500): + return min_val + value * (max_val - min_val) + canny_low_threshold = int(round(scale_value(canny_low_threshold))) + canny_high_threshold = int(round(scale_value(canny_high_threshold))) + + + if skip_preprocessing: + preprocessed_image = control_image + control_image = None + canny_low_threshold = None + canny_high_threshold = None + + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/bfl/flux-pro-1.0-canny/generate", + method=HttpMethod.POST, + request_model=BFLFluxCannyImageRequest, + response_model=BFLFluxProGenerateResponse, + ), + request=BFLFluxCannyImageRequest( + prompt=prompt, + prompt_upsampling=prompt_upsampling, + steps=steps, + guidance=guidance, + seed=seed, + control_image=control_image, + canny_low_threshold=canny_low_threshold, + canny_high_threshold=canny_high_threshold, + preprocessed_image=preprocessed_image, + ), + auth_kwargs=kwargs, + ) + output_image = handle_bfl_synchronous_operation(operation, node_id=unique_id) + return (output_image,) + + +class FluxProDepthNode(ComfyNodeABC): + """ + Generate image using a control image (depth). + """ + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "control_image": (IO.IMAGE,), + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Prompt for the image generation", + }, + ), + "prompt_upsampling": ( + IO.BOOLEAN, + { + "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).", + }, + ), + "skip_preprocessing": ( + IO.BOOLEAN, + { + "default": False, + "tooltip": "Whether to skip preprocessing; set to True if control_image already is depth-ified, False if it is a raw image.", + }, + ), + "guidance": ( + IO.FLOAT, + { + "default": 15, + "min": 1, + "max": 100, + "tooltip": "Guidance strength for the image generation process" + }, + ), + "steps": ( + IO.INT, + { + "default": 50, + "min": 15, + "max": 50, + "tooltip": "Number of steps for the image generation process" + }, + ), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 0xFFFFFFFFFFFFFFFF, + "control_after_generate": True, + "tooltip": "The random seed used for creating the noise.", + }, + ), + }, + "optional": {}, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = (IO.IMAGE,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/image/BFL" + + def api_call( + self, + control_image: torch.Tensor, + prompt: str, + prompt_upsampling: bool, + skip_preprocessing: bool, + steps: int, + guidance: float, + seed=0, + unique_id: Union[str, None] = None, + **kwargs, + ): + control_image = convert_image_to_base64(control_image[:,:,:,:3]) + preprocessed_image = None + + if skip_preprocessing: + preprocessed_image = control_image + control_image = None + + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/bfl/flux-pro-1.0-depth/generate", + method=HttpMethod.POST, + request_model=BFLFluxDepthImageRequest, + response_model=BFLFluxProGenerateResponse, + ), + request=BFLFluxDepthImageRequest( + prompt=prompt, + prompt_upsampling=prompt_upsampling, + steps=steps, + guidance=guidance, + seed=seed, + control_image=control_image, + preprocessed_image=preprocessed_image, + ), + auth_kwargs=kwargs, + ) + output_image = handle_bfl_synchronous_operation(operation, node_id=unique_id) + return (output_image,) + + +# A dictionary that contains all nodes you want to export with their names +# NOTE: names should be globally unique +NODE_CLASS_MAPPINGS = { + "FluxProUltraImageNode": FluxProUltraImageNode, + # "FluxProImageNode": FluxProImageNode, + "FluxKontextProImageNode": FluxKontextProImageNode, + "FluxKontextMaxImageNode": FluxKontextMaxImageNode, + "FluxProExpandNode": FluxProExpandNode, + "FluxProFillNode": FluxProFillNode, + "FluxProCannyNode": FluxProCannyNode, + "FluxProDepthNode": FluxProDepthNode, +} + +# A dictionary that contains the friendly/humanly readable titles for the nodes +NODE_DISPLAY_NAME_MAPPINGS = { + "FluxProUltraImageNode": "Flux 1.1 [pro] Ultra Image", + # "FluxProImageNode": "Flux 1.1 [pro] Image", + "FluxKontextProImageNode": "Flux.1 Kontext [pro] Image", + "FluxKontextMaxImageNode": "Flux.1 Kontext [max] Image", + "FluxProExpandNode": "Flux.1 Expand Image", + "FluxProFillNode": "Flux.1 Fill Image", + "FluxProCannyNode": "Flux.1 Canny Control Image", + "FluxProDepthNode": "Flux.1 Depth Control Image", +} diff --git a/comfy_api_nodes/nodes_gemini.py b/comfy_api_nodes/nodes_gemini.py new file mode 100644 index 000000000..ae7b04846 --- /dev/null +++ b/comfy_api_nodes/nodes_gemini.py @@ -0,0 +1,446 @@ +""" +API Nodes for Gemini Multimodal LLM Usage via Remote API +See: https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/inference +""" + +import os +from enum import Enum +from typing import Optional, Literal + +import torch + +import folder_paths +from comfy.comfy_types.node_typing import IO, ComfyNodeABC, InputTypeDict +from server import PromptServer +from comfy_api_nodes.apis import ( + GeminiContent, + GeminiGenerateContentRequest, + GeminiGenerateContentResponse, + GeminiInlineData, + GeminiPart, + GeminiMimeType, +) +from comfy_api_nodes.apis.client import ( + ApiEndpoint, + HttpMethod, + SynchronousOperation, +) +from comfy_api_nodes.apinode_utils import ( + validate_string, + audio_to_base64_string, + video_to_base64_string, + tensor_to_base64_string, +) + + +GEMINI_BASE_ENDPOINT = "/proxy/vertexai/gemini" +GEMINI_MAX_INPUT_FILE_SIZE = 20 * 1024 * 1024 # 20 MB + + +class GeminiModel(str, Enum): + """ + Gemini Model Names allowed by comfy-api + """ + + gemini_2_5_pro_preview_05_06 = "gemini-2.5-pro-preview-05-06" + gemini_2_5_flash_preview_04_17 = "gemini-2.5-flash-preview-04-17" + + +def get_gemini_endpoint( + model: GeminiModel, +) -> ApiEndpoint[GeminiGenerateContentRequest, GeminiGenerateContentResponse]: + """ + Get the API endpoint for a given Gemini model. + + Args: + model: The Gemini model to use, either as enum or string value. + + Returns: + ApiEndpoint configured for the specific Gemini model. + """ + if isinstance(model, str): + model = GeminiModel(model) + return ApiEndpoint( + path=f"{GEMINI_BASE_ENDPOINT}/{model.value}", + method=HttpMethod.POST, + request_model=GeminiGenerateContentRequest, + response_model=GeminiGenerateContentResponse, + ) + + +class GeminiNode(ComfyNodeABC): + """ + Node to generate text responses from a Gemini model. + + This node allows users to interact with Google's Gemini AI models, providing + multimodal inputs (text, images, audio, video, files) to generate coherent + text responses. The node works with the latest Gemini models, handling the + API communication and response parsing. + """ + + @classmethod + def INPUT_TYPES(cls) -> InputTypeDict: + return { + "required": { + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Text inputs to the model, used to generate a response. You can include detailed instructions, questions, or context for the model.", + }, + ), + "model": ( + IO.COMBO, + { + "tooltip": "The Gemini model to use for generating responses.", + "options": [model.value for model in GeminiModel], + "default": GeminiModel.gemini_2_5_pro_preview_05_06.value, + }, + ), + "seed": ( + IO.INT, + { + "default": 42, + "min": 0, + "max": 0xFFFFFFFFFFFFFFFF, + "control_after_generate": True, + "tooltip": "When 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.", + }, + ), + }, + "optional": { + "images": ( + IO.IMAGE, + { + "default": None, + "tooltip": "Optional image(s) to use as context for the model. To include multiple images, you can use the Batch Images node.", + }, + ), + "audio": ( + IO.AUDIO, + { + "tooltip": "Optional audio to use as context for the model.", + "default": None, + }, + ), + "video": ( + IO.VIDEO, + { + "tooltip": "Optional video to use as context for the model.", + "default": None, + }, + ), + "files": ( + "GEMINI_INPUT_FILES", + { + "default": None, + "tooltip": "Optional file(s) to use as context for the model. Accepts inputs from the Gemini Generate Content Input Files node.", + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + DESCRIPTION = "Generate text responses with Google's Gemini AI model. You can provide multiple types of inputs (text, images, audio, video) as context for generating more relevant and meaningful responses." + RETURN_TYPES = ("STRING",) + FUNCTION = "api_call" + CATEGORY = "api node/text/Gemini" + API_NODE = True + + def get_parts_from_response( + self, response: GeminiGenerateContentResponse + ) -> list[GeminiPart]: + """ + Extract all parts from the Gemini API response. + + Args: + response: The API response from Gemini. + + Returns: + List of response parts from the first candidate. + """ + return response.candidates[0].content.parts + + def get_parts_by_type( + self, response: GeminiGenerateContentResponse, part_type: Literal["text"] | str + ) -> list[GeminiPart]: + """ + Filter response parts by their type. + + Args: + response: The API response from Gemini. + part_type: Type of parts to extract ("text" or a MIME type). + + Returns: + List of response parts matching the requested type. + """ + parts = [] + for part in self.get_parts_from_response(response): + if part_type == "text" and hasattr(part, "text") and part.text: + parts.append(part) + elif ( + hasattr(part, "inlineData") + and part.inlineData + and part.inlineData.mimeType == part_type + ): + parts.append(part) + # Skip parts that don't match the requested type + return parts + + def get_text_from_response(self, response: GeminiGenerateContentResponse) -> str: + """ + Extract and concatenate all text parts from the response. + + Args: + response: The API response from Gemini. + + Returns: + Combined text from all text parts in the response. + """ + parts = self.get_parts_by_type(response, "text") + return "\n".join([part.text for part in parts]) + + def create_video_parts(self, video_input: IO.VIDEO, **kwargs) -> list[GeminiPart]: + """ + Convert video input to Gemini API compatible parts. + + Args: + video_input: Video tensor from ComfyUI. + **kwargs: Additional arguments to pass to the conversion function. + + Returns: + List of GeminiPart objects containing the encoded video. + """ + from comfy_api.util import VideoContainer, VideoCodec + base_64_string = video_to_base64_string( + video_input, + container_format=VideoContainer.MP4, + codec=VideoCodec.H264 + ) + return [ + GeminiPart( + inlineData=GeminiInlineData( + mimeType=GeminiMimeType.video_mp4, + data=base_64_string, + ) + ) + ] + + def create_audio_parts(self, audio_input: IO.AUDIO) -> list[GeminiPart]: + """ + Convert audio input to Gemini API compatible parts. + + Args: + audio_input: Audio input from ComfyUI, containing waveform tensor and sample rate. + + Returns: + List of GeminiPart objects containing the encoded audio. + """ + audio_parts: list[GeminiPart] = [] + for batch_index in range(audio_input["waveform"].shape[0]): + # Recreate an IO.AUDIO object for the given batch dimension index + audio_at_index = { + "waveform": audio_input["waveform"][batch_index].unsqueeze(0), + "sample_rate": audio_input["sample_rate"], + } + # Convert to MP3 format for compatibility with Gemini API + audio_bytes = audio_to_base64_string( + audio_at_index, + container_format="mp3", + codec_name="libmp3lame", + ) + audio_parts.append( + GeminiPart( + inlineData=GeminiInlineData( + mimeType=GeminiMimeType.audio_mp3, + data=audio_bytes, + ) + ) + ) + return audio_parts + + def create_image_parts(self, 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. + """ + 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) + ) + image_parts.append( + GeminiPart( + inlineData=GeminiInlineData( + mimeType=GeminiMimeType.image_png, + data=image_as_b64, + ) + ) + ) + return image_parts + + def create_text_part(self, text: str) -> GeminiPart: + """ + Create a text part for the Gemini API request. + + Args: + text: The text content to include in the request. + + Returns: + A GeminiPart object with the text content. + """ + return GeminiPart(text=text) + + def api_call( + self, + prompt: str, + model: GeminiModel, + images: Optional[IO.IMAGE] = None, + audio: Optional[IO.AUDIO] = None, + video: Optional[IO.VIDEO] = None, + files: Optional[list[GeminiPart]] = None, + unique_id: Optional[str] = None, + **kwargs, + ) -> tuple[str]: + # Validate inputs + validate_string(prompt, strip_whitespace=False) + + # Create parts list with text prompt as the first part + parts: list[GeminiPart] = [self.create_text_part(prompt)] + + # Add other modal parts + if images is not None: + image_parts = self.create_image_parts(images) + parts.extend(image_parts) + if audio is not None: + parts.extend(self.create_audio_parts(audio)) + if video is not None: + parts.extend(self.create_video_parts(video)) + if files is not None: + parts.extend(files) + + # Create response + response = SynchronousOperation( + endpoint=get_gemini_endpoint(model), + request=GeminiGenerateContentRequest( + contents=[ + GeminiContent( + role="user", + parts=parts, + ) + ] + ), + auth_kwargs=kwargs, + ).execute() + + # Get result output + output_text = self.get_text_from_response(response) + if unique_id and output_text: + PromptServer.instance.send_progress_text(output_text, node_id=unique_id) + + return (output_text or "Empty response from Gemini model...",) + + +class GeminiInputFiles(ComfyNodeABC): + """ + Loads and formats input files for use with the Gemini API. + + This node allows users to include text (.txt) and PDF (.pdf) files as input + context for the Gemini model. Files are converted to the appropriate format + required by the API and can be chained together to include multiple files + in a single request. + """ + + @classmethod + def INPUT_TYPES(cls) -> InputTypeDict: + """ + For details about the supported file input types, see: + https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/inference + """ + input_dir = folder_paths.get_input_directory() + input_files = [ + f + for f in os.scandir(input_dir) + if f.is_file() + and (f.name.endswith(".txt") or f.name.endswith(".pdf")) + and f.stat().st_size < GEMINI_MAX_INPUT_FILE_SIZE + ] + input_files = sorted(input_files, key=lambda x: x.name) + input_files = [f.name for f in input_files] + return { + "required": { + "file": ( + IO.COMBO, + { + "tooltip": "Input files to include as context for the model. Only accepts text (.txt) and PDF (.pdf) files for now.", + "options": input_files, + "default": input_files[0] if input_files else None, + }, + ), + }, + "optional": { + "GEMINI_INPUT_FILES": ( + "GEMINI_INPUT_FILES", + { + "tooltip": "An optional additional file(s) to batch together with the file loaded from this node. Allows chaining of input files so that a single message can include multiple input files.", + "default": None, + }, + ), + }, + } + + DESCRIPTION = "Loads and prepares input files to include as inputs for Gemini LLM nodes. The files will be read by the Gemini model when generating a response. The contents of the text file count toward the token limit. 🛈 TIP: Can be chained together with other Gemini Input File nodes." + RETURN_TYPES = ("GEMINI_INPUT_FILES",) + FUNCTION = "prepare_files" + CATEGORY = "api node/text/Gemini" + + def create_file_part(self, file_path: str) -> GeminiPart: + mime_type = ( + GeminiMimeType.pdf + if file_path.endswith(".pdf") + else GeminiMimeType.text_plain + ) + # Use base64 string directly, not the data URI + with open(file_path, "rb") as f: + file_content = f.read() + import base64 + base64_str = base64.b64encode(file_content).decode("utf-8") + + return GeminiPart( + inlineData=GeminiInlineData( + mimeType=mime_type, + data=base64_str, + ) + ) + + def prepare_files( + self, file: str, GEMINI_INPUT_FILES: list[GeminiPart] = [] + ) -> tuple[list[GeminiPart]]: + """ + Loads and formats input files for Gemini API. + """ + file_path = folder_paths.get_annotated_filepath(file) + input_file_content = self.create_file_part(file_path) + files = [input_file_content] + GEMINI_INPUT_FILES + return (files,) + + +NODE_CLASS_MAPPINGS = { + "GeminiNode": GeminiNode, + "GeminiInputFiles": GeminiInputFiles, +} + +NODE_DISPLAY_NAME_MAPPINGS = { + "GeminiNode": "Google Gemini", + "GeminiInputFiles": "Gemini Input Files", +} diff --git a/comfy_api_nodes/nodes_ideogram.py b/comfy_api_nodes/nodes_ideogram.py new file mode 100644 index 000000000..b8487355f --- /dev/null +++ b/comfy_api_nodes/nodes_ideogram.py @@ -0,0 +1,801 @@ +from comfy.comfy_types.node_typing import IO, ComfyNodeABC, InputTypeDict +from inspect import cleandoc +from PIL import Image +import numpy as np +import io +import torch +from comfy_api_nodes.apis import ( + IdeogramGenerateRequest, + IdeogramGenerateResponse, + ImageRequest, + IdeogramV3Request, + IdeogramV3EditRequest, +) + +from comfy_api_nodes.apis.client import ( + ApiEndpoint, + HttpMethod, + SynchronousOperation, +) + +from comfy_api_nodes.apinode_utils import ( + download_url_to_bytesio, + bytesio_to_image_tensor, + resize_mask_to_image, +) +from server import PromptServer + +V1_V1_RES_MAP = { + "Auto":"AUTO", + "512 x 1536":"RESOLUTION_512_1536", + "576 x 1408":"RESOLUTION_576_1408", + "576 x 1472":"RESOLUTION_576_1472", + "576 x 1536":"RESOLUTION_576_1536", + "640 x 1024":"RESOLUTION_640_1024", + "640 x 1344":"RESOLUTION_640_1344", + "640 x 1408":"RESOLUTION_640_1408", + "640 x 1472":"RESOLUTION_640_1472", + "640 x 1536":"RESOLUTION_640_1536", + "704 x 1152":"RESOLUTION_704_1152", + "704 x 1216":"RESOLUTION_704_1216", + "704 x 1280":"RESOLUTION_704_1280", + "704 x 1344":"RESOLUTION_704_1344", + "704 x 1408":"RESOLUTION_704_1408", + "704 x 1472":"RESOLUTION_704_1472", + "720 x 1280":"RESOLUTION_720_1280", + "736 x 1312":"RESOLUTION_736_1312", + "768 x 1024":"RESOLUTION_768_1024", + "768 x 1088":"RESOLUTION_768_1088", + "768 x 1152":"RESOLUTION_768_1152", + "768 x 1216":"RESOLUTION_768_1216", + "768 x 1232":"RESOLUTION_768_1232", + "768 x 1280":"RESOLUTION_768_1280", + "768 x 1344":"RESOLUTION_768_1344", + "832 x 960":"RESOLUTION_832_960", + "832 x 1024":"RESOLUTION_832_1024", + "832 x 1088":"RESOLUTION_832_1088", + "832 x 1152":"RESOLUTION_832_1152", + "832 x 1216":"RESOLUTION_832_1216", + "832 x 1248":"RESOLUTION_832_1248", + "864 x 1152":"RESOLUTION_864_1152", + "896 x 960":"RESOLUTION_896_960", + "896 x 1024":"RESOLUTION_896_1024", + "896 x 1088":"RESOLUTION_896_1088", + "896 x 1120":"RESOLUTION_896_1120", + "896 x 1152":"RESOLUTION_896_1152", + "960 x 832":"RESOLUTION_960_832", + "960 x 896":"RESOLUTION_960_896", + "960 x 1024":"RESOLUTION_960_1024", + "960 x 1088":"RESOLUTION_960_1088", + "1024 x 640":"RESOLUTION_1024_640", + "1024 x 768":"RESOLUTION_1024_768", + "1024 x 832":"RESOLUTION_1024_832", + "1024 x 896":"RESOLUTION_1024_896", + "1024 x 960":"RESOLUTION_1024_960", + "1024 x 1024":"RESOLUTION_1024_1024", + "1088 x 768":"RESOLUTION_1088_768", + "1088 x 832":"RESOLUTION_1088_832", + "1088 x 896":"RESOLUTION_1088_896", + "1088 x 960":"RESOLUTION_1088_960", + "1120 x 896":"RESOLUTION_1120_896", + "1152 x 704":"RESOLUTION_1152_704", + "1152 x 768":"RESOLUTION_1152_768", + "1152 x 832":"RESOLUTION_1152_832", + "1152 x 864":"RESOLUTION_1152_864", + "1152 x 896":"RESOLUTION_1152_896", + "1216 x 704":"RESOLUTION_1216_704", + "1216 x 768":"RESOLUTION_1216_768", + "1216 x 832":"RESOLUTION_1216_832", + "1232 x 768":"RESOLUTION_1232_768", + "1248 x 832":"RESOLUTION_1248_832", + "1280 x 704":"RESOLUTION_1280_704", + "1280 x 720":"RESOLUTION_1280_720", + "1280 x 768":"RESOLUTION_1280_768", + "1280 x 800":"RESOLUTION_1280_800", + "1312 x 736":"RESOLUTION_1312_736", + "1344 x 640":"RESOLUTION_1344_640", + "1344 x 704":"RESOLUTION_1344_704", + "1344 x 768":"RESOLUTION_1344_768", + "1408 x 576":"RESOLUTION_1408_576", + "1408 x 640":"RESOLUTION_1408_640", + "1408 x 704":"RESOLUTION_1408_704", + "1472 x 576":"RESOLUTION_1472_576", + "1472 x 640":"RESOLUTION_1472_640", + "1472 x 704":"RESOLUTION_1472_704", + "1536 x 512":"RESOLUTION_1536_512", + "1536 x 576":"RESOLUTION_1536_576", + "1536 x 640":"RESOLUTION_1536_640", +} + +V1_V2_RATIO_MAP = { + "1:1":"ASPECT_1_1", + "4:3":"ASPECT_4_3", + "3:4":"ASPECT_3_4", + "16:9":"ASPECT_16_9", + "9:16":"ASPECT_9_16", + "2:1":"ASPECT_2_1", + "1:2":"ASPECT_1_2", + "3:2":"ASPECT_3_2", + "2:3":"ASPECT_2_3", + "4:5":"ASPECT_4_5", + "5:4":"ASPECT_5_4", +} + +V3_RATIO_MAP = { + "1:3":"1x3", + "3:1":"3x1", + "1:2":"1x2", + "2:1":"2x1", + "9:16":"9x16", + "16:9":"16x9", + "10:16":"10x16", + "16:10":"16x10", + "2:3":"2x3", + "3:2":"3x2", + "3:4":"3x4", + "4:3":"4x3", + "4:5":"4x5", + "5:4":"5x4", + "1:1":"1x1", +} + +V3_RESOLUTIONS= [ + "Auto", + "512x1536", + "576x1408", + "576x1472", + "576x1536", + "640x1344", + "640x1408", + "640x1472", + "640x1536", + "704x1152", + "704x1216", + "704x1280", + "704x1344", + "704x1408", + "704x1472", + "736x1312", + "768x1088", + "768x1216", + "768x1280", + "768x1344", + "800x1280", + "832x960", + "832x1024", + "832x1088", + "832x1152", + "832x1216", + "832x1248", + "864x1152", + "896x960", + "896x1024", + "896x1088", + "896x1120", + "896x1152", + "960x832", + "960x896", + "960x1024", + "960x1088", + "1024x832", + "1024x896", + "1024x960", + "1024x1024", + "1088x768", + "1088x832", + "1088x896", + "1088x960", + "1120x896", + "1152x704", + "1152x832", + "1152x864", + "1152x896", + "1216x704", + "1216x768", + "1216x832", + "1248x832", + "1280x704", + "1280x768", + "1280x800", + "1312x736", + "1344x640", + "1344x704", + "1344x768", + "1408x576", + "1408x640", + "1408x704", + "1472x576", + "1472x640", + "1472x704", + "1536x512", + "1536x576", + "1536x640" +] + +def download_and_process_images(image_urls): + """Helper function to download and process multiple images from URLs""" + + # Initialize list to store image tensors + image_tensors = [] + + for image_url in image_urls: + # Using functions from apinode_utils.py to handle downloading and processing + image_bytesio = download_url_to_bytesio(image_url) # Download image content to BytesIO + img_tensor = bytesio_to_image_tensor(image_bytesio, mode="RGB") # Convert to torch.Tensor with RGB mode + image_tensors.append(img_tensor) + + # Stack tensors to match (N, width, height, channels) + if image_tensors: + stacked_tensors = torch.cat(image_tensors, dim=0) + else: + raise Exception("No valid images were processed") + + return stacked_tensors + + +def display_image_urls_on_node(image_urls, node_id): + if node_id and image_urls: + if len(image_urls) == 1: + PromptServer.instance.send_progress_text( + f"Generated Image URL:\n{image_urls[0]}", node_id + ) + else: + urls_text = "Generated Image URLs:\n" + "\n".join( + f"{i+1}. {url}" for i, url in enumerate(image_urls) + ) + PromptServer.instance.send_progress_text(urls_text, node_id) + + +class IdeogramV1(ComfyNodeABC): + """ + Generates images using the Ideogram V1 model. + """ + + def __init__(self): + pass + + @classmethod + def INPUT_TYPES(cls) -> InputTypeDict: + return { + "required": { + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Prompt for the image generation", + }, + ), + "turbo": ( + IO.BOOLEAN, + { + "default": False, + "tooltip": "Whether to use turbo mode (faster generation, potentially lower quality)", + } + ), + }, + "optional": { + "aspect_ratio": ( + IO.COMBO, + { + "options": list(V1_V2_RATIO_MAP.keys()), + "default": "1:1", + "tooltip": "The aspect ratio for image generation.", + }, + ), + "magic_prompt_option": ( + IO.COMBO, + { + "options": ["AUTO", "ON", "OFF"], + "default": "AUTO", + "tooltip": "Determine if MagicPrompt should be used in generation", + }, + ), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 2147483647, + "step": 1, + "control_after_generate": True, + "display": "number", + }, + ), + "negative_prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Description of what to exclude from the image", + }, + ), + "num_images": ( + IO.INT, + {"default": 1, "min": 1, "max": 8, "step": 1, "display": "number"}, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = (IO.IMAGE,) + FUNCTION = "api_call" + CATEGORY = "api node/image/Ideogram" + DESCRIPTION = cleandoc(__doc__ or "") + API_NODE = True + + def api_call( + self, + prompt, + turbo=False, + aspect_ratio="1:1", + magic_prompt_option="AUTO", + seed=0, + negative_prompt="", + num_images=1, + unique_id=None, + **kwargs, + ): + # Determine the model based on turbo setting + aspect_ratio = V1_V2_RATIO_MAP.get(aspect_ratio, None) + model = "V_1_TURBO" if turbo else "V_1" + + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/ideogram/generate", + method=HttpMethod.POST, + request_model=IdeogramGenerateRequest, + response_model=IdeogramGenerateResponse, + ), + request=IdeogramGenerateRequest( + image_request=ImageRequest( + prompt=prompt, + model=model, + num_images=num_images, + seed=seed, + aspect_ratio=aspect_ratio if aspect_ratio != "ASPECT_1_1" else None, + magic_prompt_option=( + magic_prompt_option if magic_prompt_option != "AUTO" else None + ), + negative_prompt=negative_prompt if negative_prompt else None, + ) + ), + auth_kwargs=kwargs, + ) + + response = operation.execute() + + if not response.data or len(response.data) == 0: + raise Exception("No images were generated in the response") + + image_urls = [image_data.url for image_data in response.data if image_data.url] + + if not image_urls: + raise Exception("No image URLs were generated in the response") + + display_image_urls_on_node(image_urls, unique_id) + return (download_and_process_images(image_urls),) + + +class IdeogramV2(ComfyNodeABC): + """ + Generates images using the Ideogram V2 model. + """ + + def __init__(self): + pass + + @classmethod + def INPUT_TYPES(cls) -> InputTypeDict: + return { + "required": { + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Prompt for the image generation", + }, + ), + "turbo": ( + IO.BOOLEAN, + { + "default": False, + "tooltip": "Whether to use turbo mode (faster generation, potentially lower quality)", + } + ), + }, + "optional": { + "aspect_ratio": ( + IO.COMBO, + { + "options": list(V1_V2_RATIO_MAP.keys()), + "default": "1:1", + "tooltip": "The aspect ratio for image generation. Ignored if resolution is not set to AUTO.", + }, + ), + "resolution": ( + IO.COMBO, + { + "options": list(V1_V1_RES_MAP.keys()), + "default": "Auto", + "tooltip": "The resolution for image generation. If not set to AUTO, this overrides the aspect_ratio setting.", + }, + ), + "magic_prompt_option": ( + IO.COMBO, + { + "options": ["AUTO", "ON", "OFF"], + "default": "AUTO", + "tooltip": "Determine if MagicPrompt should be used in generation", + }, + ), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 2147483647, + "step": 1, + "control_after_generate": True, + "display": "number", + }, + ), + "style_type": ( + IO.COMBO, + { + "options": ["AUTO", "GENERAL", "REALISTIC", "DESIGN", "RENDER_3D", "ANIME"], + "default": "NONE", + "tooltip": "Style type for generation (V2 only)", + }, + ), + "negative_prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Description of what to exclude from the image", + }, + ), + "num_images": ( + IO.INT, + {"default": 1, "min": 1, "max": 8, "step": 1, "display": "number"}, + ), + #"color_palette": ( + # IO.STRING, + # { + # "multiline": False, + # "default": "", + # "tooltip": "Color palette preset name or hex colors with weights", + # }, + #), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = (IO.IMAGE,) + FUNCTION = "api_call" + CATEGORY = "api node/image/Ideogram" + DESCRIPTION = cleandoc(__doc__ or "") + API_NODE = True + + def api_call( + self, + prompt, + turbo=False, + aspect_ratio="1:1", + resolution="Auto", + magic_prompt_option="AUTO", + seed=0, + style_type="NONE", + negative_prompt="", + num_images=1, + color_palette="", + unique_id=None, + **kwargs, + ): + aspect_ratio = V1_V2_RATIO_MAP.get(aspect_ratio, None) + resolution = V1_V1_RES_MAP.get(resolution, None) + # Determine the model based on turbo setting + model = "V_2_TURBO" if turbo else "V_2" + + # Handle resolution vs aspect_ratio logic + # If resolution is not AUTO, it overrides aspect_ratio + final_resolution = None + final_aspect_ratio = None + + if resolution != "AUTO": + final_resolution = resolution + else: + final_aspect_ratio = aspect_ratio if aspect_ratio != "ASPECT_1_1" else None + + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/ideogram/generate", + method=HttpMethod.POST, + request_model=IdeogramGenerateRequest, + response_model=IdeogramGenerateResponse, + ), + request=IdeogramGenerateRequest( + image_request=ImageRequest( + prompt=prompt, + model=model, + num_images=num_images, + seed=seed, + aspect_ratio=final_aspect_ratio, + resolution=final_resolution, + magic_prompt_option=( + magic_prompt_option if magic_prompt_option != "AUTO" else None + ), + style_type=style_type if style_type != "NONE" else None, + negative_prompt=negative_prompt if negative_prompt else None, + color_palette=color_palette if color_palette else None, + ) + ), + auth_kwargs=kwargs, + ) + + response = operation.execute() + + if not response.data or len(response.data) == 0: + raise Exception("No images were generated in the response") + + image_urls = [image_data.url for image_data in response.data if image_data.url] + + if not image_urls: + raise Exception("No image URLs were generated in the response") + + display_image_urls_on_node(image_urls, unique_id) + return (download_and_process_images(image_urls),) + +class IdeogramV3(ComfyNodeABC): + """ + Generates images using the Ideogram V3 model. Supports both regular image generation from text prompts and image editing with mask. + """ + + def __init__(self): + pass + + @classmethod + def INPUT_TYPES(cls) -> InputTypeDict: + return { + "required": { + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Prompt for the image generation or editing", + }, + ), + }, + "optional": { + "image": ( + IO.IMAGE, + { + "default": None, + "tooltip": "Optional reference image for image editing.", + }, + ), + "mask": ( + IO.MASK, + { + "default": None, + "tooltip": "Optional mask for inpainting (white areas will be replaced)", + }, + ), + "aspect_ratio": ( + IO.COMBO, + { + "options": list(V3_RATIO_MAP.keys()), + "default": "1:1", + "tooltip": "The aspect ratio for image generation. Ignored if resolution is not set to Auto.", + }, + ), + "resolution": ( + IO.COMBO, + { + "options": V3_RESOLUTIONS, + "default": "Auto", + "tooltip": "The resolution for image generation. If not set to Auto, this overrides the aspect_ratio setting.", + }, + ), + "magic_prompt_option": ( + IO.COMBO, + { + "options": ["AUTO", "ON", "OFF"], + "default": "AUTO", + "tooltip": "Determine if MagicPrompt should be used in generation", + }, + ), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 2147483647, + "step": 1, + "control_after_generate": True, + "display": "number", + }, + ), + "num_images": ( + IO.INT, + {"default": 1, "min": 1, "max": 8, "step": 1, "display": "number"}, + ), + "rendering_speed": ( + IO.COMBO, + { + "options": ["BALANCED", "TURBO", "QUALITY"], + "default": "BALANCED", + "tooltip": "Controls the trade-off between generation speed and quality", + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = (IO.IMAGE,) + FUNCTION = "api_call" + CATEGORY = "api node/image/Ideogram" + DESCRIPTION = cleandoc(__doc__ or "") + API_NODE = True + + def api_call( + self, + prompt, + image=None, + mask=None, + resolution="Auto", + aspect_ratio="1:1", + magic_prompt_option="AUTO", + seed=0, + num_images=1, + rendering_speed="BALANCED", + unique_id=None, + **kwargs, + ): + # Check if both image and mask are provided for editing mode + if image is not None and mask is not None: + # Edit mode + path = "/proxy/ideogram/ideogram-v3/edit" + + # Process image and mask + input_tensor = image.squeeze().cpu() + # Resize mask to match image dimension + mask = resize_mask_to_image(mask, image, allow_gradient=False) + # Invert mask, as Ideogram API will edit black areas instead of white areas (opposite of convention). + mask = 1.0 - mask + + # Validate mask dimensions match image + if mask.shape[1:] != image.shape[1:-1]: + raise Exception("Mask and Image must be the same size") + + # Process image + img_np = (input_tensor.numpy() * 255).astype(np.uint8) + img = Image.fromarray(img_np) + img_byte_arr = io.BytesIO() + img.save(img_byte_arr, format="PNG") + img_byte_arr.seek(0) + img_binary = img_byte_arr + img_binary.name = "image.png" + + # Process mask - white areas will be replaced + mask_np = (mask.squeeze().cpu().numpy() * 255).astype(np.uint8) + mask_img = Image.fromarray(mask_np) + mask_byte_arr = io.BytesIO() + mask_img.save(mask_byte_arr, format="PNG") + mask_byte_arr.seek(0) + mask_binary = mask_byte_arr + mask_binary.name = "mask.png" + + # Create edit request + edit_request = IdeogramV3EditRequest( + prompt=prompt, + rendering_speed=rendering_speed, + ) + + # Add optional parameters + if magic_prompt_option != "AUTO": + edit_request.magic_prompt = magic_prompt_option + if seed != 0: + edit_request.seed = seed + if num_images > 1: + edit_request.num_images = num_images + + # Execute the operation for edit mode + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path=path, + method=HttpMethod.POST, + request_model=IdeogramV3EditRequest, + response_model=IdeogramGenerateResponse, + ), + request=edit_request, + files={ + "image": img_binary, + "mask": mask_binary, + }, + content_type="multipart/form-data", + auth_kwargs=kwargs, + ) + + elif image is not None or mask is not None: + # If only one of image or mask is provided, raise an error + raise Exception("Ideogram V3 image editing requires both an image AND a mask") + else: + # Generation mode + path = "/proxy/ideogram/ideogram-v3/generate" + + # Create generation request + gen_request = IdeogramV3Request( + prompt=prompt, + rendering_speed=rendering_speed, + ) + + # Handle resolution vs aspect ratio + if resolution != "Auto": + gen_request.resolution = resolution + elif aspect_ratio != "1:1": + v3_aspect = V3_RATIO_MAP.get(aspect_ratio) + if v3_aspect: + gen_request.aspect_ratio = v3_aspect + + # Add optional parameters + if magic_prompt_option != "AUTO": + gen_request.magic_prompt = magic_prompt_option + if seed != 0: + gen_request.seed = seed + if num_images > 1: + gen_request.num_images = num_images + + # Execute the operation for generation mode + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path=path, + method=HttpMethod.POST, + request_model=IdeogramV3Request, + response_model=IdeogramGenerateResponse, + ), + request=gen_request, + auth_kwargs=kwargs, + ) + + # Execute the operation and process response + response = operation.execute() + + if not response.data or len(response.data) == 0: + raise Exception("No images were generated in the response") + + image_urls = [image_data.url for image_data in response.data if image_data.url] + + if not image_urls: + raise Exception("No image URLs were generated in the response") + + display_image_urls_on_node(image_urls, unique_id) + return (download_and_process_images(image_urls),) + + +NODE_CLASS_MAPPINGS = { + "IdeogramV1": IdeogramV1, + "IdeogramV2": IdeogramV2, + "IdeogramV3": IdeogramV3, +} + +NODE_DISPLAY_NAME_MAPPINGS = { + "IdeogramV1": "Ideogram V1", + "IdeogramV2": "Ideogram V2", + "IdeogramV3": "Ideogram V3", +} diff --git a/comfy_api_nodes/nodes_kling.py b/comfy_api_nodes/nodes_kling.py new file mode 100644 index 000000000..641cd6353 --- /dev/null +++ b/comfy_api_nodes/nodes_kling.py @@ -0,0 +1,1758 @@ +"""Kling API Nodes + +For source of truth on the allowed permutations of request fields, please reference: +- [Compatibility Table](https://app.klingai.com/global/dev/document-api/apiReference/model/skillsMap) +""" + +from __future__ import annotations +from typing import Optional, TypeVar, Any +from collections.abc import Callable +import math +import logging + +import torch + +from comfy_api_nodes.apis import ( + KlingTaskStatus, + KlingCameraControl, + KlingCameraConfig, + KlingCameraControlType, + KlingVideoGenDuration, + KlingVideoGenMode, + KlingVideoGenAspectRatio, + KlingVideoGenModelName, + KlingText2VideoRequest, + KlingText2VideoResponse, + KlingImage2VideoRequest, + KlingImage2VideoResponse, + KlingVideoExtendRequest, + KlingVideoExtendResponse, + KlingLipSyncVoiceLanguage, + KlingLipSyncInputObject, + KlingLipSyncRequest, + KlingLipSyncResponse, + KlingVirtualTryOnModelName, + KlingVirtualTryOnRequest, + KlingVirtualTryOnResponse, + KlingVideoResult, + KlingImageResult, + KlingImageGenerationsRequest, + KlingImageGenerationsResponse, + KlingImageGenImageReferenceType, + KlingImageGenModelName, + KlingImageGenAspectRatio, + KlingVideoEffectsRequest, + KlingVideoEffectsResponse, + KlingDualCharacterEffectsScene, + KlingSingleImageEffectsScene, + KlingDualCharacterEffectInput, + KlingSingleImageEffectInput, + KlingCharacterEffectModelName, + KlingSingleImageEffectModelName, +) +from comfy_api_nodes.apis.client import ( + ApiEndpoint, + HttpMethod, + SynchronousOperation, + PollingOperation, + EmptyRequest, +) +from comfy_api_nodes.apinode_utils import ( + tensor_to_base64_string, + download_url_to_video_output, + upload_video_to_comfyapi, + upload_audio_to_comfyapi, + download_url_to_image_tensor, +) +from comfy_api_nodes.mapper_utils import model_field_to_node_input +from comfy_api_nodes.util.validation_utils import ( + validate_image_dimensions, + validate_image_aspect_ratio, + validate_video_dimensions, + validate_video_duration, +) +from comfy_api.input.basic_types import AudioInput +from comfy_api.input.video_types import VideoInput +from comfy_api.input_impl import VideoFromFile +from comfy.comfy_types.node_typing import IO, InputTypeOptions, ComfyNodeABC + +KLING_API_VERSION = "v1" +PATH_TEXT_TO_VIDEO = f"/proxy/kling/{KLING_API_VERSION}/videos/text2video" +PATH_IMAGE_TO_VIDEO = f"/proxy/kling/{KLING_API_VERSION}/videos/image2video" +PATH_VIDEO_EXTEND = f"/proxy/kling/{KLING_API_VERSION}/videos/video-extend" +PATH_LIP_SYNC = f"/proxy/kling/{KLING_API_VERSION}/videos/lip-sync" +PATH_VIDEO_EFFECTS = f"/proxy/kling/{KLING_API_VERSION}/videos/effects" +PATH_CHARACTER_IMAGE = f"/proxy/kling/{KLING_API_VERSION}/images/generations" +PATH_VIRTUAL_TRY_ON = f"/proxy/kling/{KLING_API_VERSION}/images/kolors-virtual-try-on" +PATH_IMAGE_GENERATIONS = f"/proxy/kling/{KLING_API_VERSION}/images/generations" + +MAX_PROMPT_LENGTH_T2V = 2500 +MAX_PROMPT_LENGTH_I2V = 500 +MAX_PROMPT_LENGTH_IMAGE_GEN = 500 +MAX_NEGATIVE_PROMPT_LENGTH_IMAGE_GEN = 200 +MAX_PROMPT_LENGTH_LIP_SYNC = 120 + +AVERAGE_DURATION_T2V = 319 +AVERAGE_DURATION_I2V = 164 +AVERAGE_DURATION_LIP_SYNC = 455 +AVERAGE_DURATION_VIRTUAL_TRY_ON = 19 +AVERAGE_DURATION_IMAGE_GEN = 32 +AVERAGE_DURATION_VIDEO_EFFECTS = 320 +AVERAGE_DURATION_VIDEO_EXTEND = 320 + +R = TypeVar("R") + + +class KlingApiError(Exception): + """Base exception for Kling API errors.""" + + pass + + +def poll_until_finished( + auth_kwargs: dict[str, str], + api_endpoint: ApiEndpoint[Any, R], + result_url_extractor: Optional[Callable[[R], str]] = None, + estimated_duration: Optional[int] = None, + node_id: Optional[str] = None, +) -> R: + """Polls the Kling API endpoint until the task reaches a terminal state, then returns the response.""" + return PollingOperation( + poll_endpoint=api_endpoint, + completed_statuses=[ + KlingTaskStatus.succeed.value, + ], + failed_statuses=[KlingTaskStatus.failed.value], + status_extractor=lambda response: ( + response.data.task_status.value + if response.data and response.data.task_status + else None + ), + auth_kwargs=auth_kwargs, + result_url_extractor=result_url_extractor, + estimated_duration=estimated_duration, + node_id=node_id, + ).execute() + + +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) + + +def is_valid_prompt(prompt: str) -> bool: + """Verifies that the prompt is not empty.""" + return bool(prompt) + + +def is_valid_task_creation_response(response: KlingText2VideoResponse) -> bool: + """Verifies that the initial response contains a task ID.""" + return bool(response.data.task_id) + + +def is_valid_video_response(response: KlingText2VideoResponse) -> bool: + """Verifies that the response contains a task result with at least one video.""" + return ( + response.data is not None + and response.data.task_result is not None + and response.data.task_result.videos is not None + and len(response.data.task_result.videos) > 0 + ) + + +def is_valid_image_response(response: KlingVirtualTryOnResponse) -> bool: + """Verifies that the response contains a task result with at least one image.""" + return ( + response.data is not None + and response.data.task_result is not None + and response.data.task_result.images is not None + and len(response.data.task_result.images) > 0 + ) + + +def validate_prompts(prompt: str, negative_prompt: str, max_length: int) -> bool: + """Verifies that the positive prompt is not empty and that neither promt is too long.""" + if not prompt: + raise ValueError("Positive prompt is empty") + if len(prompt) > max_length: + raise ValueError(f"Positive prompt is too long: {len(prompt)} characters") + if negative_prompt and len(negative_prompt) > max_length: + raise ValueError( + f"Negative prompt is too long: {len(negative_prompt)} characters" + ) + return True + + +def validate_task_creation_response(response) -> None: + """Validates that the Kling task creation request was successful.""" + if not is_valid_task_creation_response(response): + error_msg = f"Kling initial request failed. Code: {response.code}, Message: {response.message}, Data: {response.data}" + logging.error(error_msg) + raise KlingApiError(error_msg) + + +def validate_video_result_response(response) -> None: + """Validates that the Kling task result contains a video.""" + if not is_valid_video_response(response): + error_msg = f"Kling task {response.data.task_id} succeeded but no video data found in response." + logging.error(f"Error: {error_msg}.\nResponse: {response}") + raise KlingApiError(error_msg) + + +def validate_image_result_response(response) -> None: + """Validates that the Kling task result contains an image.""" + if not is_valid_image_response(response): + error_msg = f"Kling task {response.data.task_id} succeeded but no image data found in response." + logging.error(f"Error: {error_msg}.\nResponse: {response}") + raise KlingApiError(error_msg) + + +def validate_input_image(image: torch.Tensor) -> None: + """ + Validates the input image adheres to the expectations of the Kling API: + - The image resolution should not be less than 300*300px + - The aspect ratio of the image should be between 1:2.5 ~ 2.5:1 + + See: https://app.klingai.com/global/dev/document-api/apiReference/model/imageToVideo + """ + validate_image_dimensions(image, min_width=300, min_height=300) + validate_image_aspect_ratio(image, min_aspect_ratio=1 / 2.5, max_aspect_ratio=2.5) + + +def get_camera_control_input_config( + tooltip: str, default: float = 0.0 +) -> tuple[IO, InputTypeOptions]: + """Returns common InputTypeOptions for Kling camera control configurations.""" + input_config = { + "default": default, + "min": -10.0, + "max": 10.0, + "step": 0.25, + "display": "slider", + "tooltip": tooltip, + } + return IO.FLOAT, input_config + + +def get_video_from_response(response) -> KlingVideoResult: + """Returns the first video object from the Kling video generation task result. + Will raise an error if the response is not valid. + """ + video = response.data.task_result.videos[0] + logging.info( + "Kling task %s succeeded. Video URL: %s", response.data.task_id, video.url + ) + return video + + +def get_video_url_from_response(response) -> Optional[str]: + """Returns the first video url from the Kling video generation task result. + Will not raise an error if the response is not valid. + """ + if response and is_valid_video_response(response): + return str(get_video_from_response(response).url) + else: + return None + + +def get_images_from_response(response) -> list[KlingImageResult]: + """Returns the list of image objects from the Kling image generation task result. + Will raise an error if the response is not valid. + """ + images = response.data.task_result.images + logging.info("Kling task %s succeeded. Images: %s", response.data.task_id, images) + return images + + +def get_images_urls_from_response(response) -> Optional[str]: + """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. + """ + if response and is_valid_image_response(response): + images = get_images_from_response(response) + image_urls = [str(image.url) for image in images] + return "\n".join(image_urls) + else: + return None + + +def video_result_to_node_output( + video: KlingVideoResult, +) -> tuple[VideoFromFile, str, str]: + """Converts a KlingVideoResult to a tuple of (VideoFromFile, str, str) to be used as a ComfyUI node output.""" + return ( + download_url_to_video_output(video.url), + str(video.id), + str(video.duration), + ) + + +def image_result_to_node_output( + images: list[KlingImageResult], +) -> torch.Tensor: + """ + Converts a KlingImageResult to a tuple containing a [B, H, W, C] tensor. + If multiple images are returned, they will be stacked along the batch dimension. + """ + if len(images) == 1: + return download_url_to_image_tensor(images[0].url) + else: + return torch.cat([download_url_to_image_tensor(image.url) for image in images]) + + +class KlingNodeBase(ComfyNodeABC): + """Base class for Kling nodes.""" + + FUNCTION = "api_call" + CATEGORY = "api node/video/Kling" + API_NODE = True + + +class KlingCameraControls(KlingNodeBase): + """Kling Camera Controls Node""" + + @classmethod + def INPUT_TYPES(cls): + return { + "required": { + "camera_control_type": model_field_to_node_input( + IO.COMBO, + KlingCameraControl, + "type", + enum_type=KlingCameraControlType, + ), + "horizontal_movement": get_camera_control_input_config( + "Controls camera's movement along horizontal axis (x-axis). Negative indicates left, positive indicates right" + ), + "vertical_movement": get_camera_control_input_config( + "Controls camera's movement along vertical axis (y-axis). Negative indicates downward, positive indicates upward." + ), + "pan": get_camera_control_input_config( + "Controls camera's rotation in vertical plane (x-axis). Negative indicates downward rotation, positive indicates upward rotation.", + default=0.5, + ), + "tilt": get_camera_control_input_config( + "Controls camera's rotation in horizontal plane (y-axis). Negative indicates left rotation, positive indicates right rotation.", + ), + "roll": get_camera_control_input_config( + "Controls camera's rolling amount (z-axis). Negative indicates counterclockwise, positive indicates clockwise.", + ), + "zoom": get_camera_control_input_config( + "Controls change in camera's focal length. Negative indicates narrower field of view, positive indicates wider field of view.", + ), + } + } + + DESCRIPTION = "Allows specifying configuration options for Kling Camera Controls and motion control effects." + RETURN_TYPES = ("CAMERA_CONTROL",) + RETURN_NAMES = ("camera_control",) + FUNCTION = "main" + API_NODE = False # This is just a helper node, it doesn't make an API call + + @classmethod + def VALIDATE_INPUTS( + cls, + horizontal_movement: float, + vertical_movement: float, + pan: float, + tilt: float, + roll: float, + zoom: float, + ) -> bool | str: + if not is_valid_camera_control_configs( + [ + horizontal_movement, + vertical_movement, + pan, + tilt, + roll, + zoom, + ] + ): + return "Invalid camera control configs: at least one of the values must be non-zero" + return True + + def main( + self, + camera_control_type: str, + horizontal_movement: float, + vertical_movement: float, + pan: float, + tilt: float, + roll: float, + zoom: float, + ) -> tuple[KlingCameraControl]: + return ( + KlingCameraControl( + type=KlingCameraControlType(camera_control_type), + config=KlingCameraConfig( + horizontal=horizontal_movement, + vertical=vertical_movement, + pan=pan, + roll=roll, + tilt=tilt, + zoom=zoom, + ), + ), + ) + + +class KlingTextToVideoNode(KlingNodeBase): + """Kling Text to Video Node""" + + @staticmethod + def get_mode_string_mapping() -> dict[str, tuple[str, str, str]]: + """ + Returns a mapping of mode strings to their corresponding (mode, duration, model_name) tuples. + Only includes config combos that support the `image_tail` request field. + + See: [Kling API Docs Capability Map](https://app.klingai.com/global/dev/document-api/apiReference/model/skillsMap) + """ + return { + "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"), + "pro mode / 10s duration / kling-v2-master": ("pro", "10", "kling-v2-master"), + "standard mode / 5s duration / kling-v2-master": ("std", "5", "kling-v2-master"), + "standard mode / 10s duration / kling-v2-master": ("std", "10", "kling-v2-master"), + } + + @classmethod + def INPUT_TYPES(s): + modes = list(KlingTextToVideoNode.get_mode_string_mapping().keys()) + return { + "required": { + "prompt": model_field_to_node_input( + IO.STRING, KlingText2VideoRequest, "prompt", multiline=True + ), + "negative_prompt": model_field_to_node_input( + IO.STRING, KlingText2VideoRequest, "negative_prompt", multiline=True + ), + "cfg_scale": model_field_to_node_input( + IO.FLOAT, + KlingText2VideoRequest, + "cfg_scale", + default=1.0, + min=0.0, + max=1.0, + ), + "aspect_ratio": model_field_to_node_input( + IO.COMBO, + KlingText2VideoRequest, + "aspect_ratio", + enum_type=KlingVideoGenAspectRatio, + ), + "mode": ( + modes, + { + "default": modes[4], + "tooltip": "The configuration to use for the video generation following the format: mode / duration / model_name.", + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = ("VIDEO", "STRING", "STRING") + RETURN_NAMES = ("VIDEO", "video_id", "duration") + DESCRIPTION = "Kling Text to Video Node" + + def get_response( + self, task_id: str, auth_kwargs: dict[str, str], node_id: Optional[str] = None + ) -> KlingText2VideoResponse: + return poll_until_finished( + auth_kwargs, + ApiEndpoint( + path=f"{PATH_TEXT_TO_VIDEO}/{task_id}", + method=HttpMethod.GET, + request_model=EmptyRequest, + response_model=KlingText2VideoResponse, + ), + result_url_extractor=get_video_url_from_response, + estimated_duration=AVERAGE_DURATION_T2V, + node_id=node_id, + ) + + def api_call( + self, + prompt: str, + negative_prompt: str, + cfg_scale: float, + mode: str, + aspect_ratio: str, + camera_control: Optional[KlingCameraControl] = None, + model_name: Optional[str] = None, + duration: Optional[str] = None, + unique_id: Optional[str] = None, + **kwargs, + ) -> tuple[VideoFromFile, str, str]: + validate_prompts(prompt, negative_prompt, MAX_PROMPT_LENGTH_T2V) + if model_name is None: + mode, duration, model_name = self.get_mode_string_mapping()[mode] + initial_operation = SynchronousOperation( + endpoint=ApiEndpoint( + path=PATH_TEXT_TO_VIDEO, + method=HttpMethod.POST, + request_model=KlingText2VideoRequest, + response_model=KlingText2VideoResponse, + ), + request=KlingText2VideoRequest( + prompt=prompt if prompt else None, + negative_prompt=negative_prompt if negative_prompt else None, + duration=KlingVideoGenDuration(duration), + mode=KlingVideoGenMode(mode), + model_name=KlingVideoGenModelName(model_name), + cfg_scale=cfg_scale, + aspect_ratio=KlingVideoGenAspectRatio(aspect_ratio), + camera_control=camera_control, + ), + auth_kwargs=kwargs, + ) + + task_creation_response = initial_operation.execute() + validate_task_creation_response(task_creation_response) + + task_id = task_creation_response.data.task_id + final_response = self.get_response( + task_id, auth_kwargs=kwargs, node_id=unique_id + ) + validate_video_result_response(final_response) + + video = get_video_from_response(final_response) + return video_result_to_node_output(video) + + +class KlingCameraControlT2VNode(KlingTextToVideoNode): + """ + Kling Text to Video Camera Control Node. This node is a text to video node, but it supports controlling the camera. + Duration, mode, and model_name request fields are hard-coded because camera control is only supported in pro mode with the kling-v1-5 model at 5s duration as of 2025-05-02. + """ + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "prompt": model_field_to_node_input( + IO.STRING, KlingText2VideoRequest, "prompt", multiline=True + ), + "negative_prompt": model_field_to_node_input( + IO.STRING, + KlingText2VideoRequest, + "negative_prompt", + multiline=True, + ), + "cfg_scale": model_field_to_node_input( + IO.FLOAT, + KlingText2VideoRequest, + "cfg_scale", + default=0.75, + min=0.0, + max=1.0, + ), + "aspect_ratio": model_field_to_node_input( + IO.COMBO, + KlingText2VideoRequest, + "aspect_ratio", + enum_type=KlingVideoGenAspectRatio, + ), + "camera_control": ( + "CAMERA_CONTROL", + { + "tooltip": "Can be created using the Kling Camera Controls node. Controls the camera movement and motion during the video generation.", + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + DESCRIPTION = "Transform text into cinematic videos with professional camera movements that simulate real-world cinematography. Control virtual camera actions including zoom, rotation, pan, tilt, and first-person view, while maintaining focus on your original text." + + def api_call( + self, + prompt: str, + negative_prompt: str, + cfg_scale: float, + aspect_ratio: str, + camera_control: Optional[KlingCameraControl] = None, + unique_id: Optional[str] = None, + **kwargs, + ): + return super().api_call( + model_name=KlingVideoGenModelName.kling_v1, + cfg_scale=cfg_scale, + mode=KlingVideoGenMode.std, + aspect_ratio=KlingVideoGenAspectRatio(aspect_ratio), + duration=KlingVideoGenDuration.field_5, + prompt=prompt, + negative_prompt=negative_prompt, + camera_control=camera_control, + **kwargs, + ) + + +class KlingImage2VideoNode(KlingNodeBase): + """Kling Image to Video Node""" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "start_frame": model_field_to_node_input( + IO.IMAGE, + KlingImage2VideoRequest, + "image", + tooltip="The reference image used to generate the video.", + ), + "prompt": model_field_to_node_input( + IO.STRING, KlingImage2VideoRequest, "prompt", multiline=True + ), + "negative_prompt": model_field_to_node_input( + IO.STRING, + KlingImage2VideoRequest, + "negative_prompt", + multiline=True, + ), + "model_name": model_field_to_node_input( + IO.COMBO, + KlingImage2VideoRequest, + "model_name", + enum_type=KlingVideoGenModelName, + ), + "cfg_scale": model_field_to_node_input( + IO.FLOAT, + KlingImage2VideoRequest, + "cfg_scale", + default=0.8, + min=0.0, + max=1.0, + ), + "mode": model_field_to_node_input( + IO.COMBO, + KlingImage2VideoRequest, + "mode", + enum_type=KlingVideoGenMode, + ), + "aspect_ratio": model_field_to_node_input( + IO.COMBO, + KlingImage2VideoRequest, + "aspect_ratio", + enum_type=KlingVideoGenAspectRatio, + ), + "duration": model_field_to_node_input( + IO.COMBO, + KlingImage2VideoRequest, + "duration", + enum_type=KlingVideoGenDuration, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = ("VIDEO", "STRING", "STRING") + RETURN_NAMES = ("VIDEO", "video_id", "duration") + DESCRIPTION = "Kling Image to Video Node" + + def get_response( + self, task_id: str, auth_kwargs: dict[str, str], node_id: Optional[str] = None + ) -> KlingImage2VideoResponse: + return poll_until_finished( + auth_kwargs, + ApiEndpoint( + path=f"{PATH_IMAGE_TO_VIDEO}/{task_id}", + method=HttpMethod.GET, + request_model=KlingImage2VideoRequest, + response_model=KlingImage2VideoResponse, + ), + result_url_extractor=get_video_url_from_response, + estimated_duration=AVERAGE_DURATION_I2V, + node_id=node_id, + ) + + def api_call( + self, + start_frame: torch.Tensor, + prompt: str, + negative_prompt: str, + model_name: str, + cfg_scale: float, + mode: str, + aspect_ratio: str, + duration: str, + camera_control: Optional[KlingCameraControl] = None, + end_frame: Optional[torch.Tensor] = None, + unique_id: Optional[str] = None, + **kwargs, + ) -> tuple[VideoFromFile]: + validate_prompts(prompt, negative_prompt, MAX_PROMPT_LENGTH_I2V) + validate_input_image(start_frame) + + if camera_control is not None: + # Camera control type for image 2 video is always `simple` + camera_control.type = KlingCameraControlType.simple + + initial_operation = SynchronousOperation( + endpoint=ApiEndpoint( + path=PATH_IMAGE_TO_VIDEO, + method=HttpMethod.POST, + request_model=KlingImage2VideoRequest, + response_model=KlingImage2VideoResponse, + ), + request=KlingImage2VideoRequest( + model_name=KlingVideoGenModelName(model_name), + image=tensor_to_base64_string(start_frame), + image_tail=( + tensor_to_base64_string(end_frame) + if end_frame is not None + else None + ), + prompt=prompt, + negative_prompt=negative_prompt if negative_prompt else None, + cfg_scale=cfg_scale, + mode=KlingVideoGenMode(mode), + duration=KlingVideoGenDuration(duration), + camera_control=camera_control, + ), + auth_kwargs=kwargs, + ) + + task_creation_response = initial_operation.execute() + validate_task_creation_response(task_creation_response) + task_id = task_creation_response.data.task_id + + final_response = self.get_response( + task_id, auth_kwargs=kwargs, node_id=unique_id + ) + validate_video_result_response(final_response) + + video = get_video_from_response(final_response) + return video_result_to_node_output(video) + + +class KlingCameraControlI2VNode(KlingImage2VideoNode): + """ + Kling Image to Video Camera Control Node. This node is a image to video node, but it supports controlling the camera. + Duration, mode, and model_name request fields are hard-coded because camera control is only supported in pro mode with the kling-v1-5 model at 5s duration as of 2025-05-02. + """ + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "start_frame": model_field_to_node_input( + IO.IMAGE, KlingImage2VideoRequest, "image" + ), + "prompt": model_field_to_node_input( + IO.STRING, KlingImage2VideoRequest, "prompt", multiline=True + ), + "negative_prompt": model_field_to_node_input( + IO.STRING, + KlingImage2VideoRequest, + "negative_prompt", + multiline=True, + ), + "cfg_scale": model_field_to_node_input( + IO.FLOAT, + KlingImage2VideoRequest, + "cfg_scale", + default=0.75, + min=0.0, + max=1.0, + ), + "aspect_ratio": model_field_to_node_input( + IO.COMBO, + KlingImage2VideoRequest, + "aspect_ratio", + enum_type=KlingVideoGenAspectRatio, + ), + "camera_control": ( + "CAMERA_CONTROL", + { + "tooltip": "Can be created using the Kling Camera Controls node. Controls the camera movement and motion during the video generation.", + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + DESCRIPTION = "Transform still images into cinematic videos with professional camera movements that simulate real-world cinematography. Control virtual camera actions including zoom, rotation, pan, tilt, and first-person view, while maintaining focus on your original image." + + def api_call( + self, + start_frame: torch.Tensor, + prompt: str, + negative_prompt: str, + cfg_scale: float, + aspect_ratio: str, + camera_control: KlingCameraControl, + unique_id: Optional[str] = None, + **kwargs, + ): + return super().api_call( + model_name=KlingVideoGenModelName.kling_v1_5, + start_frame=start_frame, + cfg_scale=cfg_scale, + mode=KlingVideoGenMode.pro, + aspect_ratio=KlingVideoGenAspectRatio(aspect_ratio), + duration=KlingVideoGenDuration.field_5, + prompt=prompt, + negative_prompt=negative_prompt, + camera_control=camera_control, + unique_id=unique_id, + **kwargs, + ) + + +class KlingStartEndFrameNode(KlingImage2VideoNode): + """ + Kling First Last Frame Node. This node allows creation of a video from a first and last frame. It calls the normal image to video endpoint, but only allows the subset of input options that support the `image_tail` request field. + """ + + @staticmethod + def get_mode_string_mapping() -> dict[str, tuple[str, str, str]]: + """ + Returns a mapping of mode strings to their corresponding (mode, duration, model_name) tuples. + Only includes config combos that support the `image_tail` request field. + + See: [Kling API Docs Capability Map](https://app.klingai.com/global/dev/document-api/apiReference/model/skillsMap) + """ + return { + "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"), + } + + @classmethod + def INPUT_TYPES(s): + modes = list(KlingStartEndFrameNode.get_mode_string_mapping().keys()) + return { + "required": { + "start_frame": model_field_to_node_input( + IO.IMAGE, KlingImage2VideoRequest, "image" + ), + "end_frame": model_field_to_node_input( + IO.IMAGE, KlingImage2VideoRequest, "image_tail" + ), + "prompt": model_field_to_node_input( + IO.STRING, KlingImage2VideoRequest, "prompt", multiline=True + ), + "negative_prompt": model_field_to_node_input( + IO.STRING, + KlingImage2VideoRequest, + "negative_prompt", + multiline=True, + ), + "cfg_scale": model_field_to_node_input( + IO.FLOAT, + KlingImage2VideoRequest, + "cfg_scale", + default=0.5, + min=0.0, + max=1.0, + ), + "aspect_ratio": model_field_to_node_input( + IO.COMBO, + KlingImage2VideoRequest, + "aspect_ratio", + enum_type=KlingVideoGenAspectRatio, + ), + "mode": ( + modes, + { + "default": modes[2], + "tooltip": "The configuration to use for the video generation following the format: mode / duration / model_name.", + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + DESCRIPTION = "Generate a video sequence that transitions between your provided start and end images. The node creates all frames in between, producing a smooth transformation from the first frame to the last." + + def api_call( + self, + start_frame: torch.Tensor, + end_frame: torch.Tensor, + prompt: str, + negative_prompt: str, + cfg_scale: float, + aspect_ratio: str, + mode: str, + unique_id: Optional[str] = None, + **kwargs, + ): + mode, duration, model_name = KlingStartEndFrameNode.get_mode_string_mapping()[ + mode + ] + return super().api_call( + prompt=prompt, + negative_prompt=negative_prompt, + model_name=model_name, + start_frame=start_frame, + cfg_scale=cfg_scale, + mode=mode, + aspect_ratio=aspect_ratio, + duration=duration, + end_frame=end_frame, + unique_id=unique_id, + **kwargs, + ) + + +class KlingVideoExtendNode(KlingNodeBase): + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "prompt": model_field_to_node_input( + IO.STRING, KlingVideoExtendRequest, "prompt", multiline=True + ), + "negative_prompt": model_field_to_node_input( + IO.STRING, + KlingVideoExtendRequest, + "negative_prompt", + multiline=True, + ), + "cfg_scale": model_field_to_node_input( + IO.FLOAT, + KlingVideoExtendRequest, + "cfg_scale", + default=0.5, + min=0.0, + max=1.0, + ), + "video_id": model_field_to_node_input( + IO.STRING, KlingVideoExtendRequest, "video_id", forceInput=True + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = ("VIDEO", "STRING", "STRING") + RETURN_NAMES = ("VIDEO", "video_id", "duration") + DESCRIPTION = "Kling Video Extend Node. Extend videos made by other Kling nodes. The video_id is created by using other Kling Nodes." + + def get_response( + self, task_id: str, auth_kwargs: dict[str, str], node_id: Optional[str] = None + ) -> KlingVideoExtendResponse: + return poll_until_finished( + auth_kwargs, + ApiEndpoint( + path=f"{PATH_VIDEO_EXTEND}/{task_id}", + method=HttpMethod.GET, + request_model=EmptyRequest, + response_model=KlingVideoExtendResponse, + ), + result_url_extractor=get_video_url_from_response, + estimated_duration=AVERAGE_DURATION_VIDEO_EXTEND, + node_id=node_id, + ) + + def api_call( + self, + prompt: str, + negative_prompt: str, + cfg_scale: float, + video_id: str, + unique_id: Optional[str] = None, + **kwargs, + ) -> tuple[VideoFromFile, str, str]: + validate_prompts(prompt, negative_prompt, MAX_PROMPT_LENGTH_T2V) + initial_operation = SynchronousOperation( + endpoint=ApiEndpoint( + path=PATH_VIDEO_EXTEND, + method=HttpMethod.POST, + request_model=KlingVideoExtendRequest, + response_model=KlingVideoExtendResponse, + ), + request=KlingVideoExtendRequest( + prompt=prompt if prompt else None, + negative_prompt=negative_prompt if negative_prompt else None, + cfg_scale=cfg_scale, + video_id=video_id, + ), + auth_kwargs=kwargs, + ) + + task_creation_response = initial_operation.execute() + validate_task_creation_response(task_creation_response) + task_id = task_creation_response.data.task_id + + final_response = self.get_response( + task_id, auth_kwargs=kwargs, node_id=unique_id + ) + validate_video_result_response(final_response) + + video = get_video_from_response(final_response) + return video_result_to_node_output(video) + + +class KlingVideoEffectsBase(KlingNodeBase): + """Kling Video Effects Base""" + + RETURN_TYPES = ("VIDEO", "STRING", "STRING") + RETURN_NAMES = ("VIDEO", "video_id", "duration") + + def get_response( + self, task_id: str, auth_kwargs: dict[str, str], node_id: Optional[str] = None + ) -> KlingVideoEffectsResponse: + return poll_until_finished( + auth_kwargs, + ApiEndpoint( + path=f"{PATH_VIDEO_EFFECTS}/{task_id}", + method=HttpMethod.GET, + request_model=EmptyRequest, + response_model=KlingVideoEffectsResponse, + ), + result_url_extractor=get_video_url_from_response, + estimated_duration=AVERAGE_DURATION_VIDEO_EFFECTS, + node_id=node_id, + ) + + def api_call( + self, + dual_character: bool, + effect_scene: KlingDualCharacterEffectsScene | KlingSingleImageEffectsScene, + model_name: str, + duration: KlingVideoGenDuration, + image_1: torch.Tensor, + image_2: Optional[torch.Tensor] = None, + mode: Optional[KlingVideoGenMode] = None, + unique_id: Optional[str] = None, + **kwargs, + ): + if dual_character: + request_input_field = KlingDualCharacterEffectInput( + model_name=model_name, + mode=mode, + images=[ + tensor_to_base64_string(image_1), + tensor_to_base64_string(image_2), + ], + duration=duration, + ) + else: + request_input_field = KlingSingleImageEffectInput( + model_name=model_name, + image=tensor_to_base64_string(image_1), + duration=duration, + ) + + initial_operation = SynchronousOperation( + endpoint=ApiEndpoint( + path=PATH_VIDEO_EFFECTS, + method=HttpMethod.POST, + request_model=KlingVideoEffectsRequest, + response_model=KlingVideoEffectsResponse, + ), + request=KlingVideoEffectsRequest( + effect_scene=effect_scene, + input=request_input_field, + ), + auth_kwargs=kwargs, + ) + + task_creation_response = initial_operation.execute() + validate_task_creation_response(task_creation_response) + task_id = task_creation_response.data.task_id + + final_response = self.get_response( + task_id, auth_kwargs=kwargs, node_id=unique_id + ) + validate_video_result_response(final_response) + + video = get_video_from_response(final_response) + return video_result_to_node_output(video) + + +class KlingDualCharacterVideoEffectNode(KlingVideoEffectsBase): + """Kling Dual Character Video Effect Node""" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "image_left": (IO.IMAGE, {"tooltip": "Left side image"}), + "image_right": (IO.IMAGE, {"tooltip": "Right side image"}), + "effect_scene": model_field_to_node_input( + IO.COMBO, + KlingVideoEffectsRequest, + "effect_scene", + enum_type=KlingDualCharacterEffectsScene, + ), + "model_name": model_field_to_node_input( + IO.COMBO, + KlingDualCharacterEffectInput, + "model_name", + enum_type=KlingCharacterEffectModelName, + ), + "mode": model_field_to_node_input( + IO.COMBO, + KlingDualCharacterEffectInput, + "mode", + enum_type=KlingVideoGenMode, + ), + "duration": model_field_to_node_input( + IO.COMBO, + KlingDualCharacterEffectInput, + "duration", + enum_type=KlingVideoGenDuration, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + DESCRIPTION = "Achieve different special effects when generating a video based on the effect_scene. First image will be positioned on left side, second on right side of the composite." + RETURN_TYPES = ("VIDEO", "STRING") + RETURN_NAMES = ("VIDEO", "duration") + + def api_call( + self, + image_left: torch.Tensor, + image_right: torch.Tensor, + effect_scene: KlingDualCharacterEffectsScene, + model_name: KlingCharacterEffectModelName, + mode: KlingVideoGenMode, + duration: KlingVideoGenDuration, + unique_id: Optional[str] = None, + **kwargs, + ): + video, _, duration = super().api_call( + dual_character=True, + effect_scene=effect_scene, + model_name=model_name, + mode=mode, + duration=duration, + image_1=image_left, + image_2=image_right, + unique_id=unique_id, + **kwargs, + ) + return video, duration + + +class KlingSingleImageVideoEffectNode(KlingVideoEffectsBase): + """Kling Single Image Video Effect Node""" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "image": ( + IO.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" + }, + ), + "effect_scene": model_field_to_node_input( + IO.COMBO, + KlingVideoEffectsRequest, + "effect_scene", + enum_type=KlingSingleImageEffectsScene, + ), + "model_name": model_field_to_node_input( + IO.COMBO, + KlingSingleImageEffectInput, + "model_name", + enum_type=KlingSingleImageEffectModelName, + ), + "duration": model_field_to_node_input( + IO.COMBO, + KlingSingleImageEffectInput, + "duration", + enum_type=KlingVideoGenDuration, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + DESCRIPTION = "Achieve different special effects when generating a video based on the effect_scene." + + def api_call( + self, + image: torch.Tensor, + effect_scene: KlingSingleImageEffectsScene, + model_name: KlingSingleImageEffectModelName, + duration: KlingVideoGenDuration, + unique_id: Optional[str] = None, + **kwargs, + ): + return super().api_call( + dual_character=False, + effect_scene=effect_scene, + model_name=model_name, + duration=duration, + image_1=image, + unique_id=unique_id, + **kwargs, + ) + + +class KlingLipSyncBase(KlingNodeBase): + """Kling Lip Sync Base""" + + RETURN_TYPES = ("VIDEO", "STRING", "STRING") + RETURN_NAMES = ("VIDEO", "video_id", "duration") + + def validate_lip_sync_video(self, video: VideoInput): + """ + Validates the input video adheres to the expectations of the Kling Lip Sync API: + - Video length does not exceed 10s and is not shorter than 2s + - Length and width dimensions should both be between 720px and 1920px + + See: https://app.klingai.com/global/dev/document-api/apiReference/model/videoTolip + """ + validate_video_dimensions(video, 720, 1920) + validate_video_duration(video, 2, 10) + + def validate_text(self, text: str): + if not text: + raise ValueError("Text is required") + if len(text) > MAX_PROMPT_LENGTH_LIP_SYNC: + raise ValueError( + f"Text is too long. Maximum length is {MAX_PROMPT_LENGTH_LIP_SYNC} characters." + ) + + def get_response( + self, task_id: str, auth_kwargs: dict[str, str], node_id: Optional[str] = None + ) -> KlingLipSyncResponse: + """Polls the Kling API endpoint until the task reaches a terminal state.""" + return poll_until_finished( + auth_kwargs, + ApiEndpoint( + path=f"{PATH_LIP_SYNC}/{task_id}", + method=HttpMethod.GET, + request_model=EmptyRequest, + response_model=KlingLipSyncResponse, + ), + result_url_extractor=get_video_url_from_response, + estimated_duration=AVERAGE_DURATION_LIP_SYNC, + node_id=node_id, + ) + + def api_call( + self, + video: VideoInput, + audio: Optional[AudioInput] = None, + voice_language: Optional[str] = None, + mode: Optional[str] = None, + text: Optional[str] = None, + voice_speed: Optional[float] = None, + voice_id: Optional[str] = None, + unique_id: Optional[str] = None, + **kwargs, + ) -> tuple[VideoFromFile, str, str]: + if text: + self.validate_text(text) + self.validate_lip_sync_video(video) + + # Upload video to Comfy API and get download URL + video_url = upload_video_to_comfyapi(video, auth_kwargs=kwargs) + logging.info("Uploaded video to Comfy API. URL: %s", video_url) + + # Upload the audio file to Comfy API and get download URL + if audio: + audio_url = upload_audio_to_comfyapi(audio, auth_kwargs=kwargs) + logging.info("Uploaded audio to Comfy API. URL: %s", audio_url) + else: + audio_url = None + + initial_operation = SynchronousOperation( + endpoint=ApiEndpoint( + path=PATH_LIP_SYNC, + method=HttpMethod.POST, + request_model=KlingLipSyncRequest, + response_model=KlingLipSyncResponse, + ), + request=KlingLipSyncRequest( + input=KlingLipSyncInputObject( + video_url=video_url, + mode=mode, + text=text, + voice_language=voice_language, + voice_speed=voice_speed, + audio_type="url", + audio_url=audio_url, + voice_id=voice_id, + ), + ), + auth_kwargs=kwargs, + ) + + task_creation_response = initial_operation.execute() + validate_task_creation_response(task_creation_response) + task_id = task_creation_response.data.task_id + + final_response = self.get_response( + task_id, auth_kwargs=kwargs, node_id=unique_id + ) + validate_video_result_response(final_response) + + video = get_video_from_response(final_response) + return video_result_to_node_output(video) + + +class KlingLipSyncAudioToVideoNode(KlingLipSyncBase): + """Kling Lip Sync Audio to Video Node. Syncs mouth movements in a video file to the audio content of an audio file.""" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "video": (IO.VIDEO, {}), + "audio": (IO.AUDIO, {}), + "voice_language": model_field_to_node_input( + IO.COMBO, + KlingLipSyncInputObject, + "voice_language", + enum_type=KlingLipSyncVoiceLanguage, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + DESCRIPTION = "Kling Lip Sync Audio to Video Node. Syncs mouth movements in a video file to the audio content of an audio file. When using, ensure that the audio contains clearly distinguishable vocals and that the video contains a distinct face. The audio file should not be larger than 5MB. The video file should not be larger than 100MB, should have height/width between 720px and 1920px, and should be between 2s and 10s in length." + + def api_call( + self, + video: VideoInput, + audio: AudioInput, + voice_language: str, + unique_id: Optional[str] = None, + **kwargs, + ): + return super().api_call( + video=video, + audio=audio, + voice_language=voice_language, + mode="audio2video", + unique_id=unique_id, + **kwargs, + ) + + +class KlingLipSyncTextToVideoNode(KlingLipSyncBase): + """Kling Lip Sync Text to Video Node. Syncs mouth movements in a video file to a text prompt.""" + + @staticmethod + def get_voice_config() -> dict[str, tuple[str, str]]: + return { + # English voices + "Melody": ("girlfriend_4_speech02", "en"), + "Sunny": ("genshin_vindi2", "en"), + "Sage": ("zhinen_xuesheng", "en"), + "Ace": ("AOT", "en"), + "Blossom": ("ai_shatang", "en"), + "Peppy": ("genshin_klee2", "en"), + "Dove": ("genshin_kirara", "en"), + "Shine": ("ai_kaiya", "en"), + "Anchor": ("oversea_male1", "en"), + "Lyric": ("ai_chenjiahao_712", "en"), + "Tender": ("chat1_female_new-3", "en"), + "Siren": ("chat_0407_5-1", "en"), + "Zippy": ("cartoon-boy-07", "en"), + "Bud": ("uk_boy1", "en"), + "Sprite": ("cartoon-girl-01", "en"), + "Candy": ("PeppaPig_platform", "en"), + "Beacon": ("ai_huangzhong_712", "en"), + "Rock": ("ai_huangyaoshi_712", "en"), + "Titan": ("ai_laoguowang_712", "en"), + "Grace": ("chengshu_jiejie", "en"), + "Helen": ("you_pingjing", "en"), + "Lore": ("calm_story1", "en"), + "Crag": ("uk_man2", "en"), + "Prattle": ("laopopo_speech02", "en"), + "Hearth": ("heainainai_speech02", "en"), + "The Reader": ("reader_en_m-v1", "en"), + "Commercial Lady": ("commercial_lady_en_f-v1", "en"), + # Chinese voices + "阳光少年": ("genshin_vindi2", "zh"), + "懂事小弟": ("zhinen_xuesheng", "zh"), + "运动少年": ("tiyuxi_xuedi", "zh"), + "青春少女": ("ai_shatang", "zh"), + "温柔小妹": ("genshin_klee2", "zh"), + "元气少女": ("genshin_kirara", "zh"), + "阳光男生": ("ai_kaiya", "zh"), + "幽默小哥": ("tiexin_nanyou", "zh"), + "文艺小哥": ("ai_chenjiahao_712", "zh"), + "甜美邻家": ("girlfriend_1_speech02", "zh"), + "温柔姐姐": ("chat1_female_new-3", "zh"), + "职场女青": ("girlfriend_2_speech02", "zh"), + "活泼男童": ("cartoon-boy-07", "zh"), + "俏皮女童": ("cartoon-girl-01", "zh"), + "稳重老爸": ("ai_huangyaoshi_712", "zh"), + "温柔妈妈": ("you_pingjing", "zh"), + "严肃上司": ("ai_laoguowang_712", "zh"), + "优雅贵妇": ("chengshu_jiejie", "zh"), + "慈祥爷爷": ("zhuxi_speech02", "zh"), + "唠叨爷爷": ("uk_oldman3", "zh"), + "唠叨奶奶": ("laopopo_speech02", "zh"), + "和蔼奶奶": ("heainainai_speech02", "zh"), + "东北老铁": ("dongbeilaotie_speech02", "zh"), + "重庆小伙": ("chongqingxiaohuo_speech02", "zh"), + "四川妹子": ("chuanmeizi_speech02", "zh"), + "潮汕大叔": ("chaoshandashu_speech02", "zh"), + "台湾男生": ("ai_taiwan_man2_speech02", "zh"), + "西安掌柜": ("xianzhanggui_speech02", "zh"), + "天津姐姐": ("tianjinjiejie_speech02", "zh"), + "新闻播报男": ("diyinnansang_DB_CN_M_04-v2", "zh"), + "译制片男": ("yizhipiannan-v1", "zh"), + "撒娇女友": ("tianmeixuemei-v1", "zh"), + "刀片烟嗓": ("daopianyansang-v1", "zh"), + "乖巧正太": ("mengwa-v1", "zh"), + } + + @classmethod + def INPUT_TYPES(s): + voice_options = list(s.get_voice_config().keys()) + return { + "required": { + "video": (IO.VIDEO, {}), + "text": model_field_to_node_input( + IO.STRING, KlingLipSyncInputObject, "text", multiline=True + ), + "voice": (voice_options, {"default": voice_options[0]}), + "voice_speed": model_field_to_node_input( + IO.FLOAT, KlingLipSyncInputObject, "voice_speed", slider=True + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + DESCRIPTION = "Kling Lip Sync Text to Video Node. Syncs mouth movements in a video file to a text prompt. The video file should not be larger than 100MB, should have height/width between 720px and 1920px, and should be between 2s and 10s in length." + + def api_call( + self, + video: VideoInput, + text: str, + voice: str, + voice_speed: float, + unique_id: Optional[str] = None, + **kwargs, + ): + voice_id, voice_language = KlingLipSyncTextToVideoNode.get_voice_config()[voice] + return super().api_call( + video=video, + text=text, + voice_language=voice_language, + voice_id=voice_id, + voice_speed=voice_speed, + mode="text2video", + unique_id=unique_id, + **kwargs, + ) + + +class KlingImageGenerationBase(KlingNodeBase): + """Kling Image Generation Base Node.""" + + RETURN_TYPES = ("IMAGE",) + CATEGORY = "api node/image/Kling" + + def validate_prompt(self, prompt: str, negative_prompt: Optional[str] = None): + if not prompt or len(prompt) > MAX_PROMPT_LENGTH_IMAGE_GEN: + raise ValueError( + f"Prompt must be less than {MAX_PROMPT_LENGTH_IMAGE_GEN} characters" + ) + if negative_prompt and len(negative_prompt) > MAX_PROMPT_LENGTH_IMAGE_GEN: + raise ValueError( + f"Negative prompt must be less than {MAX_PROMPT_LENGTH_IMAGE_GEN} characters" + ) + + +class KlingVirtualTryOnNode(KlingImageGenerationBase): + """Kling Virtual Try On Node.""" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "human_image": (IO.IMAGE, {}), + "cloth_image": (IO.IMAGE, {}), + "model_name": model_field_to_node_input( + IO.COMBO, + KlingVirtualTryOnRequest, + "model_name", + enum_type=KlingVirtualTryOnModelName, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + DESCRIPTION = "Kling Virtual Try On Node. Input a human image and a cloth image to try on the cloth on the human. You can merge multiple clothing item pictures into one image with a white background." + + def get_response( + self, task_id: str, auth_kwargs: dict[str, str], node_id: Optional[str] = None + ) -> KlingVirtualTryOnResponse: + return poll_until_finished( + auth_kwargs, + ApiEndpoint( + path=f"{PATH_VIRTUAL_TRY_ON}/{task_id}", + method=HttpMethod.GET, + request_model=EmptyRequest, + response_model=KlingVirtualTryOnResponse, + ), + result_url_extractor=get_images_urls_from_response, + estimated_duration=AVERAGE_DURATION_VIRTUAL_TRY_ON, + node_id=node_id, + ) + + def api_call( + self, + human_image: torch.Tensor, + cloth_image: torch.Tensor, + model_name: KlingVirtualTryOnModelName, + unique_id: Optional[str] = None, + **kwargs, + ): + initial_operation = SynchronousOperation( + endpoint=ApiEndpoint( + path=PATH_VIRTUAL_TRY_ON, + method=HttpMethod.POST, + request_model=KlingVirtualTryOnRequest, + response_model=KlingVirtualTryOnResponse, + ), + request=KlingVirtualTryOnRequest( + human_image=tensor_to_base64_string(human_image), + cloth_image=tensor_to_base64_string(cloth_image), + model_name=model_name, + ), + auth_kwargs=kwargs, + ) + + task_creation_response = initial_operation.execute() + validate_task_creation_response(task_creation_response) + task_id = task_creation_response.data.task_id + + final_response = self.get_response( + task_id, auth_kwargs=kwargs, node_id=unique_id + ) + validate_image_result_response(final_response) + + images = get_images_from_response(final_response) + return (image_result_to_node_output(images),) + + +class KlingImageGenerationNode(KlingImageGenerationBase): + """Kling Image Generation Node. Generate an image from a text prompt with an optional reference image.""" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "prompt": model_field_to_node_input( + IO.STRING, + KlingImageGenerationsRequest, + "prompt", + multiline=True, + max_length=MAX_PROMPT_LENGTH_IMAGE_GEN, + ), + "negative_prompt": model_field_to_node_input( + IO.STRING, + KlingImageGenerationsRequest, + "negative_prompt", + multiline=True, + ), + "image_type": model_field_to_node_input( + IO.COMBO, + KlingImageGenerationsRequest, + "image_reference", + enum_type=KlingImageGenImageReferenceType, + ), + "image_fidelity": model_field_to_node_input( + IO.FLOAT, + KlingImageGenerationsRequest, + "image_fidelity", + slider=True, + step=0.01, + ), + "human_fidelity": model_field_to_node_input( + IO.FLOAT, + KlingImageGenerationsRequest, + "human_fidelity", + slider=True, + step=0.01, + ), + "model_name": model_field_to_node_input( + IO.COMBO, + KlingImageGenerationsRequest, + "model_name", + enum_type=KlingImageGenModelName, + ), + "aspect_ratio": model_field_to_node_input( + IO.COMBO, + KlingImageGenerationsRequest, + "aspect_ratio", + enum_type=KlingImageGenAspectRatio, + ), + "n": model_field_to_node_input( + IO.INT, + KlingImageGenerationsRequest, + "n", + ), + }, + "optional": { + "image": (IO.IMAGE, {}), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + DESCRIPTION = "Kling Image Generation Node. Generate an image from a text prompt with an optional reference image." + + def get_response( + self, + task_id: str, + auth_kwargs: Optional[dict[str, str]], + node_id: Optional[str] = None, + ) -> KlingImageGenerationsResponse: + return poll_until_finished( + auth_kwargs, + ApiEndpoint( + path=f"{PATH_IMAGE_GENERATIONS}/{task_id}", + method=HttpMethod.GET, + request_model=EmptyRequest, + response_model=KlingImageGenerationsResponse, + ), + result_url_extractor=get_images_urls_from_response, + estimated_duration=AVERAGE_DURATION_IMAGE_GEN, + node_id=node_id, + ) + + def api_call( + self, + model_name: KlingImageGenModelName, + prompt: str, + negative_prompt: str, + image_type: KlingImageGenImageReferenceType, + image_fidelity: float, + human_fidelity: float, + n: int, + aspect_ratio: KlingImageGenAspectRatio, + image: Optional[torch.Tensor] = None, + unique_id: Optional[str] = None, + **kwargs, + ): + self.validate_prompt(prompt, negative_prompt) + + if image is not None: + image = tensor_to_base64_string(image) + + initial_operation = SynchronousOperation( + endpoint=ApiEndpoint( + path=PATH_IMAGE_GENERATIONS, + method=HttpMethod.POST, + request_model=KlingImageGenerationsRequest, + response_model=KlingImageGenerationsResponse, + ), + request=KlingImageGenerationsRequest( + model_name=model_name, + prompt=prompt, + negative_prompt=negative_prompt, + image=image, + image_reference=image_type, + image_fidelity=image_fidelity, + human_fidelity=human_fidelity, + n=n, + aspect_ratio=aspect_ratio, + ), + auth_kwargs=kwargs, + ) + + task_creation_response = initial_operation.execute() + validate_task_creation_response(task_creation_response) + task_id = task_creation_response.data.task_id + + final_response = self.get_response( + task_id, auth_kwargs=kwargs, node_id=unique_id + ) + validate_image_result_response(final_response) + + images = get_images_from_response(final_response) + return (image_result_to_node_output(images),) + + +NODE_CLASS_MAPPINGS = { + "KlingCameraControls": KlingCameraControls, + "KlingTextToVideoNode": KlingTextToVideoNode, + "KlingImage2VideoNode": KlingImage2VideoNode, + "KlingCameraControlI2VNode": KlingCameraControlI2VNode, + "KlingCameraControlT2VNode": KlingCameraControlT2VNode, + "KlingStartEndFrameNode": KlingStartEndFrameNode, + "KlingVideoExtendNode": KlingVideoExtendNode, + "KlingLipSyncAudioToVideoNode": KlingLipSyncAudioToVideoNode, + "KlingLipSyncTextToVideoNode": KlingLipSyncTextToVideoNode, + "KlingVirtualTryOnNode": KlingVirtualTryOnNode, + "KlingImageGenerationNode": KlingImageGenerationNode, + "KlingSingleImageVideoEffectNode": KlingSingleImageVideoEffectNode, + "KlingDualCharacterVideoEffectNode": KlingDualCharacterVideoEffectNode, +} + +NODE_DISPLAY_NAME_MAPPINGS = { + "KlingCameraControls": "Kling Camera Controls", + "KlingTextToVideoNode": "Kling Text to Video", + "KlingImage2VideoNode": "Kling Image to Video", + "KlingCameraControlI2VNode": "Kling Image to Video (Camera Control)", + "KlingCameraControlT2VNode": "Kling Text to Video (Camera Control)", + "KlingStartEndFrameNode": "Kling Start-End Frame to Video", + "KlingVideoExtendNode": "Kling Video Extend", + "KlingLipSyncAudioToVideoNode": "Kling Lip Sync Video with Audio", + "KlingLipSyncTextToVideoNode": "Kling Lip Sync Video with Text", + "KlingVirtualTryOnNode": "Kling Virtual Try On", + "KlingImageGenerationNode": "Kling Image Generation", + "KlingSingleImageVideoEffectNode": "Kling Video Effects", + "KlingDualCharacterVideoEffectNode": "Kling Dual Character Video Effects", +} diff --git a/comfy_api_nodes/nodes_luma.py b/comfy_api_nodes/nodes_luma.py new file mode 100644 index 000000000..525dc38e6 --- /dev/null +++ b/comfy_api_nodes/nodes_luma.py @@ -0,0 +1,737 @@ +from __future__ import annotations +from inspect import cleandoc +from typing import Optional +from comfy.comfy_types.node_typing import IO, ComfyNodeABC +from comfy_api.input_impl.video_types import VideoFromFile +from comfy_api_nodes.apis.luma_api import ( + LumaImageModel, + LumaVideoModel, + LumaVideoOutputResolution, + LumaVideoModelOutputDuration, + LumaAspectRatio, + LumaState, + LumaImageGenerationRequest, + LumaGenerationRequest, + LumaGeneration, + LumaCharacterRef, + LumaModifyImageRef, + LumaImageIdentity, + LumaReference, + LumaReferenceChain, + LumaImageReference, + LumaKeyframes, + LumaConceptChain, + LumaIO, + get_luma_concepts, +) +from comfy_api_nodes.apis.client import ( + ApiEndpoint, + HttpMethod, + SynchronousOperation, + PollingOperation, + EmptyRequest, +) +from comfy_api_nodes.apinode_utils import ( + upload_images_to_comfyapi, + process_image_response, + validate_string, +) +from server import PromptServer + +import requests +import torch +from io import BytesIO + +LUMA_T2V_AVERAGE_DURATION = 105 +LUMA_I2V_AVERAGE_DURATION = 100 + +def image_result_url_extractor(response: LumaGeneration): + return response.assets.image if hasattr(response, "assets") and hasattr(response.assets, "image") else None + +def video_result_url_extractor(response: LumaGeneration): + return response.assets.video if hasattr(response, "assets") and hasattr(response.assets, "video") else None + +class LumaReferenceNode(ComfyNodeABC): + """ + Holds an image and weight for use with Luma Generate Image node. + """ + + RETURN_TYPES = (LumaIO.LUMA_REF,) + RETURN_NAMES = ("luma_ref",) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "create_luma_reference" + CATEGORY = "api node/image/Luma" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "image": ( + IO.IMAGE, + { + "tooltip": "Image to use as reference.", + }, + ), + "weight": ( + IO.FLOAT, + { + "default": 1.0, + "min": 0.0, + "max": 1.0, + "step": 0.01, + "tooltip": "Weight of image reference.", + }, + ), + }, + "optional": {"luma_ref": (LumaIO.LUMA_REF,)}, + } + + def create_luma_reference( + self, image: torch.Tensor, weight: float, luma_ref: LumaReferenceChain = None + ): + if luma_ref is not None: + luma_ref = luma_ref.clone() + else: + luma_ref = LumaReferenceChain() + luma_ref.add(LumaReference(image=image, weight=round(weight, 2))) + return (luma_ref,) + + +class LumaConceptsNode(ComfyNodeABC): + """ + Holds one or more Camera Concepts for use with Luma Text to Video and Luma Image to Video nodes. + """ + + RETURN_TYPES = (LumaIO.LUMA_CONCEPTS,) + RETURN_NAMES = ("luma_concepts",) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "create_concepts" + CATEGORY = "api node/video/Luma" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "concept1": (get_luma_concepts(include_none=True),), + "concept2": (get_luma_concepts(include_none=True),), + "concept3": (get_luma_concepts(include_none=True),), + "concept4": (get_luma_concepts(include_none=True),), + }, + "optional": { + "luma_concepts": ( + LumaIO.LUMA_CONCEPTS, + { + "tooltip": "Optional Camera Concepts to add to the ones chosen here." + }, + ), + }, + } + + def create_concepts( + self, + concept1: str, + concept2: str, + concept3: str, + concept4: str, + luma_concepts: LumaConceptChain = None, + ): + chain = LumaConceptChain(str_list=[concept1, concept2, concept3, concept4]) + if luma_concepts is not None: + chain = luma_concepts.clone_and_merge(chain) + return (chain,) + + +class LumaImageGenerationNode(ComfyNodeABC): + """ + Generates images synchronously based on prompt and aspect ratio. + """ + + RETURN_TYPES = (IO.IMAGE,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/image/Luma" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Prompt for the image generation", + }, + ), + "model": ([model.value for model in LumaImageModel],), + "aspect_ratio": ( + [ratio.value for ratio in LumaAspectRatio], + { + "default": LumaAspectRatio.ratio_16_9, + }, + ), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 0xFFFFFFFFFFFFFFFF, + "control_after_generate": True, + "tooltip": "Seed to determine if node should re-run; actual results are nondeterministic regardless of seed.", + }, + ), + "style_image_weight": ( + IO.FLOAT, + { + "default": 1.0, + "min": 0.0, + "max": 1.0, + "step": 0.01, + "tooltip": "Weight of style image. Ignored if no style_image provided.", + }, + ), + }, + "optional": { + "image_luma_ref": ( + LumaIO.LUMA_REF, + { + "tooltip": "Luma Reference node connection to influence generation with input images; up to 4 images can be considered." + }, + ), + "style_image": ( + IO.IMAGE, + {"tooltip": "Style reference image; only 1 image will be used."}, + ), + "character_image": ( + IO.IMAGE, + { + "tooltip": "Character reference images; can be a batch of multiple, up to 4 images can be considered." + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + def api_call( + self, + prompt: str, + model: str, + aspect_ratio: str, + seed, + style_image_weight: float, + image_luma_ref: LumaReferenceChain = None, + style_image: torch.Tensor = None, + character_image: torch.Tensor = None, + unique_id: str = None, + **kwargs, + ): + validate_string(prompt, strip_whitespace=True, min_length=3) + # handle image_luma_ref + api_image_ref = None + if image_luma_ref is not None: + api_image_ref = self._convert_luma_refs( + image_luma_ref, max_refs=4, auth_kwargs=kwargs, + ) + # handle style_luma_ref + api_style_ref = None + if style_image is not None: + api_style_ref = self._convert_style_image( + style_image, weight=style_image_weight, auth_kwargs=kwargs, + ) + # handle character_ref images + character_ref = None + if character_image is not None: + download_urls = upload_images_to_comfyapi( + character_image, max_images=4, auth_kwargs=kwargs, + ) + character_ref = LumaCharacterRef( + identity0=LumaImageIdentity(images=download_urls) + ) + + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/luma/generations/image", + method=HttpMethod.POST, + request_model=LumaImageGenerationRequest, + response_model=LumaGeneration, + ), + request=LumaImageGenerationRequest( + prompt=prompt, + model=model, + aspect_ratio=aspect_ratio, + image_ref=api_image_ref, + style_ref=api_style_ref, + character_ref=character_ref, + ), + auth_kwargs=kwargs, + ) + response_api: LumaGeneration = operation.execute() + + operation = PollingOperation( + poll_endpoint=ApiEndpoint( + path=f"/proxy/luma/generations/{response_api.id}", + method=HttpMethod.GET, + request_model=EmptyRequest, + response_model=LumaGeneration, + ), + completed_statuses=[LumaState.completed], + failed_statuses=[LumaState.failed], + status_extractor=lambda x: x.state, + result_url_extractor=image_result_url_extractor, + node_id=unique_id, + auth_kwargs=kwargs, + ) + response_poll = operation.execute() + + img_response = requests.get(response_poll.assets.image) + img = process_image_response(img_response) + return (img,) + + def _convert_luma_refs( + self, luma_ref: LumaReferenceChain, max_refs: int, auth_kwargs: Optional[dict[str,str]] = None + ): + luma_urls = [] + ref_count = 0 + for ref in luma_ref.refs: + download_urls = upload_images_to_comfyapi( + ref.image, max_images=1, auth_kwargs=auth_kwargs + ) + luma_urls.append(download_urls[0]) + ref_count += 1 + if ref_count >= max_refs: + break + return luma_ref.create_api_model(download_urls=luma_urls, max_refs=max_refs) + + def _convert_style_image( + self, style_image: torch.Tensor, weight: float, auth_kwargs: Optional[dict[str,str]] = None + ): + chain = LumaReferenceChain( + first_ref=LumaReference(image=style_image, weight=weight) + ) + return self._convert_luma_refs(chain, max_refs=1, auth_kwargs=auth_kwargs) + + +class LumaImageModifyNode(ComfyNodeABC): + """ + Modifies images synchronously based on prompt and aspect ratio. + """ + + RETURN_TYPES = (IO.IMAGE,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/image/Luma" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "image": (IO.IMAGE,), + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Prompt for the image generation", + }, + ), + "image_weight": ( + IO.FLOAT, + { + "default": 0.1, + "min": 0.0, + "max": 0.98, + "step": 0.01, + "tooltip": "Weight of the image; the closer to 1.0, the less the image will be modified.", + }, + ), + "model": ([model.value for model in LumaImageModel],), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 0xFFFFFFFFFFFFFFFF, + "control_after_generate": True, + "tooltip": "Seed to determine if node should re-run; actual results are nondeterministic regardless of seed.", + }, + ), + }, + "optional": {}, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + def api_call( + self, + prompt: str, + model: str, + image: torch.Tensor, + image_weight: float, + seed, + unique_id: str = None, + **kwargs, + ): + # first, upload image + download_urls = upload_images_to_comfyapi( + image, max_images=1, auth_kwargs=kwargs, + ) + image_url = download_urls[0] + # next, make Luma call with download url provided + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/luma/generations/image", + method=HttpMethod.POST, + request_model=LumaImageGenerationRequest, + response_model=LumaGeneration, + ), + request=LumaImageGenerationRequest( + prompt=prompt, + model=model, + modify_image_ref=LumaModifyImageRef( + url=image_url, weight=round(max(min(1.0-image_weight, 0.98), 0.0), 2) + ), + ), + auth_kwargs=kwargs, + ) + response_api: LumaGeneration = operation.execute() + + operation = PollingOperation( + poll_endpoint=ApiEndpoint( + path=f"/proxy/luma/generations/{response_api.id}", + method=HttpMethod.GET, + request_model=EmptyRequest, + response_model=LumaGeneration, + ), + completed_statuses=[LumaState.completed], + failed_statuses=[LumaState.failed], + status_extractor=lambda x: x.state, + result_url_extractor=image_result_url_extractor, + node_id=unique_id, + auth_kwargs=kwargs, + ) + response_poll = operation.execute() + + img_response = requests.get(response_poll.assets.image) + img = process_image_response(img_response) + return (img,) + + +class LumaTextToVideoGenerationNode(ComfyNodeABC): + """ + Generates videos synchronously based on prompt and output_size. + """ + + RETURN_TYPES = (IO.VIDEO,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/video/Luma" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Prompt for the video generation", + }, + ), + "model": ([model.value for model in LumaVideoModel],), + "aspect_ratio": ( + [ratio.value for ratio in LumaAspectRatio], + { + "default": LumaAspectRatio.ratio_16_9, + }, + ), + "resolution": ( + [resolution.value for resolution in LumaVideoOutputResolution], + { + "default": LumaVideoOutputResolution.res_540p, + }, + ), + "duration": ([dur.value for dur in LumaVideoModelOutputDuration],), + "loop": ( + IO.BOOLEAN, + { + "default": False, + }, + ), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 0xFFFFFFFFFFFFFFFF, + "control_after_generate": True, + "tooltip": "Seed to determine if node should re-run; actual results are nondeterministic regardless of seed.", + }, + ), + }, + "optional": { + "luma_concepts": ( + LumaIO.LUMA_CONCEPTS, + { + "tooltip": "Optional Camera Concepts to dictate camera motion via the Luma Concepts node." + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + def api_call( + self, + prompt: str, + model: str, + aspect_ratio: str, + resolution: str, + duration: str, + loop: bool, + seed, + luma_concepts: LumaConceptChain = None, + unique_id: str = None, + **kwargs, + ): + validate_string(prompt, strip_whitespace=False, min_length=3) + duration = duration if model != LumaVideoModel.ray_1_6 else None + resolution = resolution if model != LumaVideoModel.ray_1_6 else None + + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/luma/generations", + method=HttpMethod.POST, + request_model=LumaGenerationRequest, + response_model=LumaGeneration, + ), + request=LumaGenerationRequest( + prompt=prompt, + model=model, + resolution=resolution, + aspect_ratio=aspect_ratio, + duration=duration, + loop=loop, + concepts=luma_concepts.create_api_model() if luma_concepts else None, + ), + auth_kwargs=kwargs, + ) + response_api: LumaGeneration = operation.execute() + + if unique_id: + PromptServer.instance.send_progress_text(f"Luma video generation started: {response_api.id}", unique_id) + + operation = PollingOperation( + poll_endpoint=ApiEndpoint( + path=f"/proxy/luma/generations/{response_api.id}", + method=HttpMethod.GET, + request_model=EmptyRequest, + response_model=LumaGeneration, + ), + completed_statuses=[LumaState.completed], + failed_statuses=[LumaState.failed], + status_extractor=lambda x: x.state, + result_url_extractor=video_result_url_extractor, + node_id=unique_id, + estimated_duration=LUMA_T2V_AVERAGE_DURATION, + auth_kwargs=kwargs, + ) + response_poll = operation.execute() + + vid_response = requests.get(response_poll.assets.video) + return (VideoFromFile(BytesIO(vid_response.content)),) + + +class LumaImageToVideoGenerationNode(ComfyNodeABC): + """ + Generates videos synchronously based on prompt, input images, and output_size. + """ + + RETURN_TYPES = (IO.VIDEO,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/video/Luma" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Prompt for the video generation", + }, + ), + "model": ([model.value for model in LumaVideoModel],), + # "aspect_ratio": ([ratio.value for ratio in LumaAspectRatio], { + # "default": LumaAspectRatio.ratio_16_9, + # }), + "resolution": ( + [resolution.value for resolution in LumaVideoOutputResolution], + { + "default": LumaVideoOutputResolution.res_540p, + }, + ), + "duration": ([dur.value for dur in LumaVideoModelOutputDuration],), + "loop": ( + IO.BOOLEAN, + { + "default": False, + }, + ), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 0xFFFFFFFFFFFFFFFF, + "control_after_generate": True, + "tooltip": "Seed to determine if node should re-run; actual results are nondeterministic regardless of seed.", + }, + ), + }, + "optional": { + "first_image": ( + IO.IMAGE, + {"tooltip": "First frame of generated video."}, + ), + "last_image": (IO.IMAGE, {"tooltip": "Last frame of generated video."}), + "luma_concepts": ( + LumaIO.LUMA_CONCEPTS, + { + "tooltip": "Optional Camera Concepts to dictate camera motion via the Luma Concepts node." + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + def api_call( + self, + prompt: str, + model: str, + resolution: str, + duration: str, + loop: bool, + seed, + first_image: torch.Tensor = None, + last_image: torch.Tensor = None, + luma_concepts: LumaConceptChain = None, + unique_id: str = None, + **kwargs, + ): + if first_image is None and last_image is None: + raise Exception( + "At least one of first_image and last_image requires an input." + ) + keyframes = self._convert_to_keyframes(first_image, last_image, auth_kwargs=kwargs) + duration = duration if model != LumaVideoModel.ray_1_6 else None + resolution = resolution if model != LumaVideoModel.ray_1_6 else None + + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/luma/generations", + method=HttpMethod.POST, + request_model=LumaGenerationRequest, + response_model=LumaGeneration, + ), + request=LumaGenerationRequest( + prompt=prompt, + model=model, + aspect_ratio=LumaAspectRatio.ratio_16_9, # ignored, but still needed by the API for some reason + resolution=resolution, + duration=duration, + loop=loop, + keyframes=keyframes, + concepts=luma_concepts.create_api_model() if luma_concepts else None, + ), + auth_kwargs=kwargs, + ) + response_api: LumaGeneration = operation.execute() + + if unique_id: + PromptServer.instance.send_progress_text(f"Luma video generation started: {response_api.id}", unique_id) + + operation = PollingOperation( + poll_endpoint=ApiEndpoint( + path=f"/proxy/luma/generations/{response_api.id}", + method=HttpMethod.GET, + request_model=EmptyRequest, + response_model=LumaGeneration, + ), + completed_statuses=[LumaState.completed], + failed_statuses=[LumaState.failed], + status_extractor=lambda x: x.state, + result_url_extractor=video_result_url_extractor, + node_id=unique_id, + estimated_duration=LUMA_I2V_AVERAGE_DURATION, + auth_kwargs=kwargs, + ) + response_poll = operation.execute() + + vid_response = requests.get(response_poll.assets.video) + return (VideoFromFile(BytesIO(vid_response.content)),) + + def _convert_to_keyframes( + self, + first_image: torch.Tensor = None, + last_image: torch.Tensor = None, + auth_kwargs: Optional[dict[str,str]] = None, + ): + if first_image is None and last_image is None: + return None + frame0 = None + frame1 = None + if first_image is not None: + download_urls = upload_images_to_comfyapi( + first_image, max_images=1, auth_kwargs=auth_kwargs, + ) + frame0 = LumaImageReference(type="image", url=download_urls[0]) + if last_image is not None: + download_urls = upload_images_to_comfyapi( + last_image, max_images=1, auth_kwargs=auth_kwargs, + ) + frame1 = LumaImageReference(type="image", url=download_urls[0]) + return LumaKeyframes(frame0=frame0, frame1=frame1) + + +# A dictionary that contains all nodes you want to export with their names +# NOTE: names should be globally unique +NODE_CLASS_MAPPINGS = { + "LumaImageNode": LumaImageGenerationNode, + "LumaImageModifyNode": LumaImageModifyNode, + "LumaVideoNode": LumaTextToVideoGenerationNode, + "LumaImageToVideoNode": LumaImageToVideoGenerationNode, + "LumaReferenceNode": LumaReferenceNode, + "LumaConceptsNode": LumaConceptsNode, +} + +# A dictionary that contains the friendly/humanly readable titles for the nodes +NODE_DISPLAY_NAME_MAPPINGS = { + "LumaImageNode": "Luma Text to Image", + "LumaImageModifyNode": "Luma Image to Image", + "LumaVideoNode": "Luma Text to Video", + "LumaImageToVideoNode": "Luma Image to Video", + "LumaReferenceNode": "Luma Reference", + "LumaConceptsNode": "Luma Concepts", +} diff --git a/comfy_api_nodes/nodes_minimax.py b/comfy_api_nodes/nodes_minimax.py new file mode 100644 index 000000000..9b46636db --- /dev/null +++ b/comfy_api_nodes/nodes_minimax.py @@ -0,0 +1,332 @@ +from typing import Union +import logging +import torch + +from comfy.comfy_types.node_typing import IO +from comfy_api.input_impl.video_types import VideoFromFile +from comfy_api_nodes.apis import ( + MinimaxVideoGenerationRequest, + MinimaxVideoGenerationResponse, + MinimaxFileRetrieveResponse, + MinimaxTaskResultResponse, + SubjectReferenceItem, + Model +) +from comfy_api_nodes.apis.client import ( + ApiEndpoint, + HttpMethod, + SynchronousOperation, + PollingOperation, + EmptyRequest, +) +from comfy_api_nodes.apinode_utils import ( + download_url_to_bytesio, + upload_images_to_comfyapi, + validate_string, +) +from server import PromptServer + + +I2V_AVERAGE_DURATION = 114 +T2V_AVERAGE_DURATION = 234 + +class MinimaxTextToVideoNode: + """ + Generates videos synchronously based on a prompt, and optional parameters using MiniMax's API. + """ + + AVERAGE_DURATION = T2V_AVERAGE_DURATION + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "prompt_text": ( + "STRING", + { + "multiline": True, + "default": "", + "tooltip": "Text prompt to guide the video generation", + }, + ), + "model": ( + [ + "T2V-01", + "T2V-01-Director", + ], + { + "default": "T2V-01", + "tooltip": "Model to use for video generation", + }, + ), + }, + "optional": { + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 0xFFFFFFFFFFFFFFFF, + "control_after_generate": True, + "tooltip": "The random seed used for creating the noise.", + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = ("VIDEO",) + DESCRIPTION = "Generates videos from prompts using MiniMax's API" + FUNCTION = "generate_video" + CATEGORY = "api node/video/MiniMax" + API_NODE = True + OUTPUT_NODE = True + + def generate_video( + self, + prompt_text, + seed=0, + model="T2V-01", + image: torch.Tensor=None, # used for ImageToVideo + subject: torch.Tensor=None, # used for SubjectToVideo + unique_id: Union[str, None]=None, + **kwargs, + ): + ''' + Function used between MiniMax nodes - supports T2V, I2V, and S2V, based on provided arguments. + ''' + if image is None: + validate_string(prompt_text, field_name="prompt_text") + # upload image, if passed in + image_url = None + if image is not None: + image_url = upload_images_to_comfyapi(image, max_images=1, auth_kwargs=kwargs)[0] + + # TODO: figure out how to deal with subject properly, API returns invalid params when using S2V-01 model + subject_reference = None + if subject is not None: + subject_url = upload_images_to_comfyapi(subject, max_images=1, auth_kwargs=kwargs)[0] + subject_reference = [SubjectReferenceItem(image=subject_url)] + + + video_generate_operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/minimax/video_generation", + method=HttpMethod.POST, + request_model=MinimaxVideoGenerationRequest, + response_model=MinimaxVideoGenerationResponse, + ), + request=MinimaxVideoGenerationRequest( + model=Model(model), + prompt=prompt_text, + callback_url=None, + first_frame_image=image_url, + subject_reference=subject_reference, + prompt_optimizer=None, + ), + auth_kwargs=kwargs, + ) + response = video_generate_operation.execute() + + task_id = response.task_id + if not task_id: + raise Exception(f"MiniMax generation failed: {response.base_resp}") + + video_generate_operation = PollingOperation( + poll_endpoint=ApiEndpoint( + path="/proxy/minimax/query/video_generation", + method=HttpMethod.GET, + request_model=EmptyRequest, + response_model=MinimaxTaskResultResponse, + query_params={"task_id": task_id}, + ), + completed_statuses=["Success"], + failed_statuses=["Fail"], + status_extractor=lambda x: x.status.value, + estimated_duration=self.AVERAGE_DURATION, + node_id=unique_id, + auth_kwargs=kwargs, + ) + task_result = video_generate_operation.execute() + + file_id = task_result.file_id + if file_id is None: + raise Exception("Request was not successful. Missing file ID.") + file_retrieve_operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/minimax/files/retrieve", + method=HttpMethod.GET, + request_model=EmptyRequest, + response_model=MinimaxFileRetrieveResponse, + query_params={"file_id": int(file_id)}, + ), + request=EmptyRequest(), + auth_kwargs=kwargs, + ) + file_result = file_retrieve_operation.execute() + + file_url = file_result.file.download_url + if file_url is None: + raise Exception( + f"No video was found in the response. Full response: {file_result.model_dump()}" + ) + logging.info(f"Generated video URL: {file_url}") + if unique_id: + if hasattr(file_result.file, "backup_download_url"): + message = f"Result URL: {file_url}\nBackup URL: {file_result.file.backup_download_url}" + else: + message = f"Result URL: {file_url}" + PromptServer.instance.send_progress_text(message, unique_id) + + video_io = download_url_to_bytesio(file_url) + if video_io is None: + error_msg = f"Failed to download video from {file_url}" + logging.error(error_msg) + raise Exception(error_msg) + return (VideoFromFile(video_io),) + + +class MinimaxImageToVideoNode(MinimaxTextToVideoNode): + """ + Generates videos synchronously based on an image and prompt, and optional parameters using MiniMax's API. + """ + + AVERAGE_DURATION = I2V_AVERAGE_DURATION + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "image": ( + IO.IMAGE, + { + "tooltip": "Image to use as first frame of video generation" + }, + ), + "prompt_text": ( + "STRING", + { + "multiline": True, + "default": "", + "tooltip": "Text prompt to guide the video generation", + }, + ), + "model": ( + [ + "I2V-01-Director", + "I2V-01", + "I2V-01-live", + ], + { + "default": "I2V-01", + "tooltip": "Model to use for video generation", + }, + ), + }, + "optional": { + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 0xFFFFFFFFFFFFFFFF, + "control_after_generate": True, + "tooltip": "The random seed used for creating the noise.", + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = ("VIDEO",) + DESCRIPTION = "Generates videos from an image and prompts using MiniMax's API" + FUNCTION = "generate_video" + CATEGORY = "api node/video/MiniMax" + API_NODE = True + OUTPUT_NODE = True + + +class MinimaxSubjectToVideoNode(MinimaxTextToVideoNode): + """ + Generates videos synchronously based on an image and prompt, and optional parameters using MiniMax's API. + """ + + AVERAGE_DURATION = T2V_AVERAGE_DURATION + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "subject": ( + IO.IMAGE, + { + "tooltip": "Image of subject to reference video generation" + }, + ), + "prompt_text": ( + "STRING", + { + "multiline": True, + "default": "", + "tooltip": "Text prompt to guide the video generation", + }, + ), + "model": ( + [ + "S2V-01", + ], + { + "default": "S2V-01", + "tooltip": "Model to use for video generation", + }, + ), + }, + "optional": { + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 0xFFFFFFFFFFFFFFFF, + "control_after_generate": True, + "tooltip": "The random seed used for creating the noise.", + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = ("VIDEO",) + DESCRIPTION = "Generates videos from an image and prompts using MiniMax's API" + FUNCTION = "generate_video" + CATEGORY = "api node/video/MiniMax" + API_NODE = True + OUTPUT_NODE = True + + +# A dictionary that contains all nodes you want to export with their names +# NOTE: names should be globally unique +NODE_CLASS_MAPPINGS = { + "MinimaxTextToVideoNode": MinimaxTextToVideoNode, + "MinimaxImageToVideoNode": MinimaxImageToVideoNode, + # "MinimaxSubjectToVideoNode": MinimaxSubjectToVideoNode, +} + +# A dictionary that contains the friendly/humanly readable titles for the nodes +NODE_DISPLAY_NAME_MAPPINGS = { + "MinimaxTextToVideoNode": "MiniMax Text to Video", + "MinimaxImageToVideoNode": "MiniMax Image to Video", + "MinimaxSubjectToVideoNode": "MiniMax Subject to Video", +} diff --git a/comfy_api_nodes/nodes_openai.py b/comfy_api_nodes/nodes_openai.py new file mode 100644 index 000000000..be1d2de4a --- /dev/null +++ b/comfy_api_nodes/nodes_openai.py @@ -0,0 +1,1008 @@ +import io +from typing import TypedDict, Optional +import json +import os +import time +import re +import uuid +from enum import Enum +from inspect import cleandoc +import numpy as np +import torch +from PIL import Image +from comfy.comfy_types.node_typing import IO, ComfyNodeABC, InputTypeDict +from server import PromptServer +import folder_paths + + +from comfy_api_nodes.apis import ( + OpenAIImageGenerationRequest, + OpenAIImageEditRequest, + OpenAIImageGenerationResponse, + OpenAICreateResponse, + OpenAIResponse, + CreateModelResponseProperties, + Item, + Includable, + OutputContent, + InputImageContent, + Detail, + InputTextContent, + InputMessage, + InputMessageContentList, + InputContent, + InputFileContent, +) + +from comfy_api_nodes.apis.client import ( + ApiEndpoint, + HttpMethod, + SynchronousOperation, + PollingOperation, + EmptyRequest, +) + +from comfy_api_nodes.apinode_utils import ( + downscale_image_tensor, + validate_and_cast_response, + validate_string, + tensor_to_base64_string, + text_filepath_to_data_uri, +) +from comfy_api_nodes.mapper_utils import model_field_to_node_input + + +RESPONSES_ENDPOINT = "/proxy/openai/v1/responses" +STARTING_POINT_ID_PATTERN = r"" + + +class HistoryEntry(TypedDict): + """Type definition for a single history entry in the chat.""" + + prompt: str + response: str + response_id: str + timestamp: float + + +class ChatHistory(TypedDict): + """Type definition for the chat history dictionary.""" + + __annotations__: dict[str, list[HistoryEntry]] + + +class SupportedOpenAIModel(str, Enum): + o4_mini = "o4-mini" + o1 = "o1" + o3 = "o3" + o1_pro = "o1-pro" + gpt_4o = "gpt-4o" + gpt_4_1 = "gpt-4.1" + gpt_4_1_mini = "gpt-4.1-mini" + gpt_4_1_nano = "gpt-4.1-nano" + + +class OpenAIDalle2(ComfyNodeABC): + """ + Generates images synchronously via OpenAI's DALL·E 2 endpoint. + """ + + def __init__(self): + pass + + @classmethod + def INPUT_TYPES(cls) -> InputTypeDict: + return { + "required": { + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Text prompt for DALL·E", + }, + ), + }, + "optional": { + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 2**31 - 1, + "step": 1, + "display": "number", + "control_after_generate": True, + "tooltip": "not implemented yet in backend", + }, + ), + "size": ( + IO.COMBO, + { + "options": ["256x256", "512x512", "1024x1024"], + "default": "1024x1024", + "tooltip": "Image size", + }, + ), + "n": ( + IO.INT, + { + "default": 1, + "min": 1, + "max": 8, + "step": 1, + "display": "number", + "tooltip": "How many images to generate", + }, + ), + "image": ( + IO.IMAGE, + { + "default": None, + "tooltip": "Optional reference image for image editing.", + }, + ), + "mask": ( + IO.MASK, + { + "default": None, + "tooltip": "Optional mask for inpainting (white areas will be replaced)", + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = (IO.IMAGE,) + FUNCTION = "api_call" + CATEGORY = "api node/image/OpenAI" + DESCRIPTION = cleandoc(__doc__ or "") + API_NODE = True + + def api_call( + self, + prompt, + seed=0, + image=None, + mask=None, + n=1, + size="1024x1024", + unique_id=None, + **kwargs, + ): + validate_string(prompt, strip_whitespace=False) + model = "dall-e-2" + path = "/proxy/openai/images/generations" + content_type = "application/json" + request_class = OpenAIImageGenerationRequest + img_binary = None + + if image is not None and mask is not None: + path = "/proxy/openai/images/edits" + content_type = "multipart/form-data" + request_class = OpenAIImageEditRequest + + input_tensor = image.squeeze().cpu() + height, width, channels = input_tensor.shape + rgba_tensor = torch.ones(height, width, 4, device="cpu") + rgba_tensor[:, :, :channels] = input_tensor + + if mask.shape[1:] != image.shape[1:-1]: + raise Exception("Mask and Image must be the same size") + rgba_tensor[:, :, 3] = 1 - mask.squeeze().cpu() + + rgba_tensor = downscale_image_tensor(rgba_tensor.unsqueeze(0)).squeeze() + + image_np = (rgba_tensor.numpy() * 255).astype(np.uint8) + img = Image.fromarray(image_np) + img_byte_arr = io.BytesIO() + img.save(img_byte_arr, format="PNG") + img_byte_arr.seek(0) + img_binary = img_byte_arr # .getvalue() + img_binary.name = "image.png" + elif image is not None or mask is not None: + raise Exception("Dall-E 2 image editing requires an image AND a mask") + + # Build the operation + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path=path, + method=HttpMethod.POST, + request_model=request_class, + response_model=OpenAIImageGenerationResponse, + ), + request=request_class( + model=model, + prompt=prompt, + n=n, + size=size, + seed=seed, + ), + files=( + { + "image": img_binary, + } + if img_binary + else None + ), + content_type=content_type, + auth_kwargs=kwargs, + ) + + response = operation.execute() + + img_tensor = validate_and_cast_response(response, node_id=unique_id) + return (img_tensor,) + + +class OpenAIDalle3(ComfyNodeABC): + """ + Generates images synchronously via OpenAI's DALL·E 3 endpoint. + """ + + def __init__(self): + pass + + @classmethod + def INPUT_TYPES(cls) -> InputTypeDict: + return { + "required": { + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Text prompt for DALL·E", + }, + ), + }, + "optional": { + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 2**31 - 1, + "step": 1, + "display": "number", + "control_after_generate": True, + "tooltip": "not implemented yet in backend", + }, + ), + "quality": ( + IO.COMBO, + { + "options": ["standard", "hd"], + "default": "standard", + "tooltip": "Image quality", + }, + ), + "style": ( + IO.COMBO, + { + "options": ["natural", "vivid"], + "default": "natural", + "tooltip": "Vivid causes the model to lean towards generating hyper-real and dramatic images. Natural causes the model to produce more natural, less hyper-real looking images.", + }, + ), + "size": ( + IO.COMBO, + { + "options": ["1024x1024", "1024x1792", "1792x1024"], + "default": "1024x1024", + "tooltip": "Image size", + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = (IO.IMAGE,) + FUNCTION = "api_call" + CATEGORY = "api node/image/OpenAI" + DESCRIPTION = cleandoc(__doc__ or "") + API_NODE = True + + def api_call( + self, + prompt, + seed=0, + style="natural", + quality="standard", + size="1024x1024", + unique_id=None, + **kwargs, + ): + validate_string(prompt, strip_whitespace=False) + model = "dall-e-3" + + # build the operation + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/openai/images/generations", + method=HttpMethod.POST, + request_model=OpenAIImageGenerationRequest, + response_model=OpenAIImageGenerationResponse, + ), + request=OpenAIImageGenerationRequest( + model=model, + prompt=prompt, + quality=quality, + size=size, + style=style, + seed=seed, + ), + auth_kwargs=kwargs, + ) + + response = operation.execute() + + img_tensor = validate_and_cast_response(response, node_id=unique_id) + return (img_tensor,) + + +class OpenAIGPTImage1(ComfyNodeABC): + """ + Generates images synchronously via OpenAI's GPT Image 1 endpoint. + """ + + def __init__(self): + pass + + @classmethod + def INPUT_TYPES(cls) -> InputTypeDict: + return { + "required": { + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Text prompt for GPT Image 1", + }, + ), + }, + "optional": { + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 2**31 - 1, + "step": 1, + "display": "number", + "control_after_generate": True, + "tooltip": "not implemented yet in backend", + }, + ), + "quality": ( + IO.COMBO, + { + "options": ["low", "medium", "high"], + "default": "low", + "tooltip": "Image quality, affects cost and generation time.", + }, + ), + "background": ( + IO.COMBO, + { + "options": ["opaque", "transparent"], + "default": "opaque", + "tooltip": "Return image with or without background", + }, + ), + "size": ( + IO.COMBO, + { + "options": ["auto", "1024x1024", "1024x1536", "1536x1024"], + "default": "auto", + "tooltip": "Image size", + }, + ), + "n": ( + IO.INT, + { + "default": 1, + "min": 1, + "max": 8, + "step": 1, + "display": "number", + "tooltip": "How many images to generate", + }, + ), + "image": ( + IO.IMAGE, + { + "default": None, + "tooltip": "Optional reference image for image editing.", + }, + ), + "mask": ( + IO.MASK, + { + "default": None, + "tooltip": "Optional mask for inpainting (white areas will be replaced)", + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = (IO.IMAGE,) + FUNCTION = "api_call" + CATEGORY = "api node/image/OpenAI" + DESCRIPTION = cleandoc(__doc__ or "") + API_NODE = True + + def api_call( + self, + prompt, + seed=0, + quality="low", + background="opaque", + image=None, + mask=None, + n=1, + size="1024x1024", + unique_id=None, + **kwargs, + ): + validate_string(prompt, strip_whitespace=False) + model = "gpt-image-1" + path = "/proxy/openai/images/generations" + content_type = "application/json" + request_class = OpenAIImageGenerationRequest + img_binaries = [] + mask_binary = None + files = [] + + if image is not None: + path = "/proxy/openai/images/edits" + request_class = OpenAIImageEditRequest + content_type = "multipart/form-data" + + 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() + + image_np = (scaled_image.numpy() * 255).astype(np.uint8) + img = Image.fromarray(image_np) + img_byte_arr = io.BytesIO() + img.save(img_byte_arr, format="PNG") + img_byte_arr.seek(0) + img_binary = img_byte_arr + img_binary.name = f"image_{i}.png" + + img_binaries.append(img_binary) + if batch_size == 1: + files.append(("image", img_binary)) + else: + files.append(("image[]", img_binary)) + + 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() + + scaled_mask = downscale_image_tensor(rgba_mask.unsqueeze(0)).squeeze() + + mask_np = (scaled_mask.numpy() * 255).astype(np.uint8) + mask_img = Image.fromarray(mask_np) + mask_img_byte_arr = io.BytesIO() + mask_img.save(mask_img_byte_arr, format="PNG") + mask_img_byte_arr.seek(0) + mask_binary = mask_img_byte_arr + mask_binary.name = "mask.png" + files.append(("mask", mask_binary)) + + # Build the operation + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path=path, + method=HttpMethod.POST, + request_model=request_class, + response_model=OpenAIImageGenerationResponse, + ), + request=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, + auth_kwargs=kwargs, + ) + + response = operation.execute() + + img_tensor = validate_and_cast_response(response, node_id=unique_id) + return (img_tensor,) + + +class OpenAITextNode(ComfyNodeABC): + """ + Base class for OpenAI text generation nodes. + """ + + RETURN_TYPES = (IO.STRING,) + FUNCTION = "api_call" + CATEGORY = "api node/text/OpenAI" + API_NODE = True + + +class OpenAIChatNode(OpenAITextNode): + """ + Node to generate text responses from an OpenAI model. + """ + + def __init__(self) -> None: + """Initialize the chat node with a new session ID and empty history.""" + self.current_session_id: str = str(uuid.uuid4()) + self.history: dict[str, list[HistoryEntry]] = {} + self.previous_response_id: Optional[str] = None + + @classmethod + def INPUT_TYPES(cls) -> InputTypeDict: + return { + "required": { + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Text inputs to the model, used to generate a response.", + }, + ), + "persist_context": ( + IO.BOOLEAN, + { + "default": True, + "tooltip": "Persist chat context between calls (multi-turn conversation)", + }, + ), + "model": model_field_to_node_input( + IO.COMBO, + OpenAICreateResponse, + "model", + enum_type=SupportedOpenAIModel, + ), + }, + "optional": { + "images": ( + IO.IMAGE, + { + "default": None, + "tooltip": "Optional image(s) to use as context for the model. To include multiple images, you can use the Batch Images node.", + }, + ), + "files": ( + "OPENAI_INPUT_FILES", + { + "default": None, + "tooltip": "Optional file(s) to use as context for the model. Accepts inputs from the OpenAI Chat Input Files node.", + }, + ), + "advanced_options": ( + "OPENAI_CHAT_CONFIG", + { + "default": None, + "tooltip": "Optional configuration for the model. Accepts inputs from the OpenAI Chat Advanced Options node.", + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + DESCRIPTION = "Generate text responses from an OpenAI model." + + def get_result_response( + self, + response_id: str, + include: Optional[list[Includable]] = None, + auth_kwargs: Optional[dict[str, str]] = None, + ) -> OpenAIResponse: + """ + Retrieve a model response with the given ID from the OpenAI API. + + Args: + response_id (str): The ID of the response to retrieve. + include (Optional[List[Includable]]): Additional fields to include + in the response. See the `include` parameter for Response + creation above for more information. + + """ + return PollingOperation( + poll_endpoint=ApiEndpoint( + path=f"{RESPONSES_ENDPOINT}/{response_id}", + method=HttpMethod.GET, + request_model=EmptyRequest, + response_model=OpenAIResponse, + query_params={"include": include}, + ), + completed_statuses=["completed"], + failed_statuses=["failed"], + status_extractor=lambda response: response.status, + auth_kwargs=auth_kwargs, + ).execute() + + def get_message_content_from_response( + self, response: OpenAIResponse + ) -> list[OutputContent]: + """Extract message content from the API response.""" + for output in response.output: + if output.root.type == "message": + return output.root.content + raise TypeError("No output message found in response") + + def get_text_from_message_content( + self, message_content: list[OutputContent] + ) -> str: + """Extract text content from message content.""" + for content_item in message_content: + if content_item.root.type == "output_text": + return str(content_item.root.text) + return "No text output found in response" + + def get_history_text(self, session_id: str) -> str: + """Convert the entire history for a given session to JSON string.""" + return json.dumps(self.history[session_id]) + + def display_history_on_node(self, session_id: str, node_id: str) -> None: + """Display formatted chat history on the node UI.""" + render_spec = { + "node_id": node_id, + "component": "ChatHistoryWidget", + "props": { + "history": self.get_history_text(session_id), + }, + } + PromptServer.instance.send_sync( + "display_component", + render_spec, + ) + + def add_to_history( + self, session_id: str, prompt: str, output_text: str, response_id: str + ) -> None: + """Add a new entry to the chat history.""" + if session_id not in self.history: + self.history[session_id] = [] + self.history[session_id].append( + { + "prompt": prompt, + "response": output_text, + "response_id": response_id, + "timestamp": time.time(), + } + ) + + def parse_output_text_from_response(self, response: OpenAIResponse) -> str: + """Extract text output from the API response.""" + message_contents = self.get_message_content_from_response(response) + return self.get_text_from_message_content(message_contents) + + def generate_new_session_id(self) -> str: + """Generate a new unique session ID.""" + return str(uuid.uuid4()) + + def get_session_id(self, persist_context: bool) -> str: + """Get the current or generate a new session ID based on context persistence.""" + return ( + self.current_session_id + if persist_context + else self.generate_new_session_id() + ) + + def tensor_to_input_image_content( + self, image: torch.Tensor, detail_level: Detail = "auto" + ) -> InputImageContent: + """Convert a tensor to an input image content object.""" + return InputImageContent( + detail=detail_level, + image_url=f"data:image/png;base64,{tensor_to_base64_string(image)}", + type="input_image", + ) + + def create_input_message_contents( + self, + prompt: str, + image: Optional[torch.Tensor] = None, + files: Optional[list[InputFileContent]] = None, + ) -> InputMessageContentList: + """Create a list of input message contents from prompt and optional image.""" + content_list: list[InputContent] = [ + InputTextContent(text=prompt, type="input_text"), + ] + if image is not None: + for i in range(image.shape[0]): + content_list.append( + self.tensor_to_input_image_content(image[i].unsqueeze(0)) + ) + if files is not None: + content_list.extend(files) + + return InputMessageContentList( + root=content_list, + ) + + def parse_response_id_from_prompt(self, prompt: str) -> Optional[str]: + """Extract response ID from prompt if it exists.""" + parsed_id = re.search(STARTING_POINT_ID_PATTERN, prompt) + return parsed_id.group(1) if parsed_id else None + + def strip_response_tag_from_prompt(self, prompt: str) -> str: + """Remove the response ID tag from the prompt.""" + return re.sub(STARTING_POINT_ID_PATTERN, "", prompt.strip()) + + def delete_history_after_response_id( + self, new_start_id: str, session_id: str + ) -> None: + """Delete history entries after a specific response ID.""" + if session_id not in self.history: + return + + new_history = [] + i = 0 + while ( + i < len(self.history[session_id]) + and self.history[session_id][i]["response_id"] != new_start_id + ): + new_history.append(self.history[session_id][i]) + i += 1 + + # Since it's the new starting point (not the response being edited), we include it as well + if i < len(self.history[session_id]): + new_history.append(self.history[session_id][i]) + + self.history[session_id] = new_history + + def api_call( + self, + prompt: str, + persist_context: bool, + model: SupportedOpenAIModel, + unique_id: Optional[str] = None, + images: Optional[torch.Tensor] = None, + files: Optional[list[InputFileContent]] = None, + advanced_options: Optional[CreateModelResponseProperties] = None, + **kwargs, + ) -> tuple[str]: + # Validate inputs + validate_string(prompt, strip_whitespace=False) + + session_id = self.get_session_id(persist_context) + response_id_override = self.parse_response_id_from_prompt(prompt) + if response_id_override: + is_starting_from_beginning = response_id_override == "start" + if is_starting_from_beginning: + self.history[session_id] = [] + previous_response_id = None + else: + previous_response_id = response_id_override + self.delete_history_after_response_id(response_id_override, session_id) + prompt = self.strip_response_tag_from_prompt(prompt) + elif persist_context: + previous_response_id = self.previous_response_id + else: + previous_response_id = None + + # Create response + create_response = SynchronousOperation( + endpoint=ApiEndpoint( + path=RESPONSES_ENDPOINT, + method=HttpMethod.POST, + request_model=OpenAICreateResponse, + response_model=OpenAIResponse, + ), + request=OpenAICreateResponse( + input=[ + Item( + root=InputMessage( + content=self.create_input_message_contents( + prompt, images, files + ), + role="user", + ) + ), + ], + store=True, + stream=False, + model=model, + previous_response_id=previous_response_id, + **( + advanced_options.model_dump(exclude_none=True) + if advanced_options + else {} + ), + ), + auth_kwargs=kwargs, + ).execute() + response_id = create_response.id + + # Get result output + result_response = self.get_result_response(response_id, auth_kwargs=kwargs) + output_text = self.parse_output_text_from_response(result_response) + + # Update history + self.add_to_history(session_id, prompt, output_text, response_id) + self.display_history_on_node(session_id, unique_id) + self.previous_response_id = response_id + + return (output_text,) + + +class OpenAIInputFiles(ComfyNodeABC): + """ + Loads and formats input files for OpenAI API. + """ + + @classmethod + def INPUT_TYPES(cls) -> InputTypeDict: + """ + For details about the supported file input types, see: + https://platform.openai.com/docs/guides/pdf-files?api-mode=responses + """ + input_dir = folder_paths.get_input_directory() + input_files = [ + f + for f in os.scandir(input_dir) + if f.is_file() + and (f.name.endswith(".txt") or f.name.endswith(".pdf")) + and f.stat().st_size < 32 * 1024 * 1024 + ] + input_files = sorted(input_files, key=lambda x: x.name) + input_files = [f.name for f in input_files] + return { + "required": { + "file": ( + IO.COMBO, + { + "tooltip": "Input files to include as context for the model. Only accepts text (.txt) and PDF (.pdf) files for now.", + "options": input_files, + "default": input_files[0] if input_files else None, + }, + ), + }, + "optional": { + "OPENAI_INPUT_FILES": ( + "OPENAI_INPUT_FILES", + { + "tooltip": "An optional additional file(s) to batch together with the file loaded from this node. Allows chaining of input files so that a single message can include multiple input files.", + "default": None, + }, + ), + }, + } + + DESCRIPTION = "Loads and prepares input files (text, pdf, etc.) to include as inputs for the OpenAI Chat Node. The files will be read by the OpenAI model when generating a response. 🛈 TIP: Can be chained together with other OpenAI Input File nodes." + RETURN_TYPES = ("OPENAI_INPUT_FILES",) + FUNCTION = "prepare_files" + CATEGORY = "api node/text/OpenAI" + + def create_input_file_content(self, file_path: str) -> InputFileContent: + return InputFileContent( + file_data=text_filepath_to_data_uri(file_path), + filename=os.path.basename(file_path), + type="input_file", + ) + + def prepare_files( + self, file: str, OPENAI_INPUT_FILES: list[InputFileContent] = [] + ) -> tuple[list[InputFileContent]]: + """ + Loads and formats input files for OpenAI API. + """ + file_path = folder_paths.get_annotated_filepath(file) + input_file_content = self.create_input_file_content(file_path) + files = [input_file_content] + OPENAI_INPUT_FILES + return (files,) + + +class OpenAIChatConfig(ComfyNodeABC): + """Allows setting additional configuration for the OpenAI Chat Node.""" + + RETURN_TYPES = ("OPENAI_CHAT_CONFIG",) + FUNCTION = "configure" + DESCRIPTION = ( + "Allows specifying advanced configuration options for the OpenAI Chat Nodes." + ) + CATEGORY = "api node/text/OpenAI" + + @classmethod + def INPUT_TYPES(cls) -> InputTypeDict: + return { + "required": { + "truncation": ( + IO.COMBO, + { + "options": ["auto", "disabled"], + "default": "auto", + "tooltip": "The truncation strategy to use for the model response. auto: If the context of this response and previous ones exceeds the model's context window size, the model will truncate the response to fit the context window by dropping input items in the middle of the conversation.disabled: If a model response will exceed the context window size for a model, the request will fail with a 400 error", + }, + ), + }, + "optional": { + "max_output_tokens": model_field_to_node_input( + IO.INT, + OpenAICreateResponse, + "max_output_tokens", + min=16, + default=4096, + max=16384, + tooltip="An upper bound for the number of tokens that can be generated for a response, including visible output tokens", + ), + "instructions": model_field_to_node_input( + IO.STRING, OpenAICreateResponse, "instructions", multiline=True + ), + }, + } + + def configure( + self, + truncation: bool, + instructions: Optional[str] = None, + max_output_tokens: Optional[int] = None, + ) -> tuple[CreateModelResponseProperties]: + """ + Configure advanced options for the OpenAI Chat Node. + + Note: + While `top_p` and `temperature` are listed as properties in the + spec, they are not supported for all models (e.g., o4-mini). + They are not exposed as inputs at all to avoid having to manually + remove depending on model choice. + """ + return ( + CreateModelResponseProperties( + instructions=instructions, + truncation=truncation, + max_output_tokens=max_output_tokens, + ), + ) + + +NODE_CLASS_MAPPINGS = { + "OpenAIDalle2": OpenAIDalle2, + "OpenAIDalle3": OpenAIDalle3, + "OpenAIGPTImage1": OpenAIGPTImage1, + "OpenAIChatNode": OpenAIChatNode, + "OpenAIInputFiles": OpenAIInputFiles, + "OpenAIChatConfig": OpenAIChatConfig, +} + +NODE_DISPLAY_NAME_MAPPINGS = { + "OpenAIDalle2": "OpenAI DALL·E 2", + "OpenAIDalle3": "OpenAI DALL·E 3", + "OpenAIGPTImage1": "OpenAI GPT Image 1", + "OpenAIChatNode": "OpenAI Chat", + "OpenAIInputFiles": "OpenAI Chat Input Files", + "OpenAIChatConfig": "OpenAI Chat Advanced Options", +} diff --git a/comfy_api_nodes/nodes_pika.py b/comfy_api_nodes/nodes_pika.py new file mode 100644 index 000000000..1cc708564 --- /dev/null +++ b/comfy_api_nodes/nodes_pika.py @@ -0,0 +1,782 @@ +""" +Pika x ComfyUI API Nodes + +Pika API docs: https://pika-827374fb.mintlify.app/api-reference +""" +from __future__ import annotations + +import io +import logging +from typing import Optional, TypeVar + +import numpy as np +import torch + +from comfy.comfy_types.node_typing import IO, ComfyNodeABC, InputTypeOptions +from comfy_api.input_impl import VideoFromFile +from comfy_api.input_impl.video_types import VideoCodec, VideoContainer, VideoInput +from comfy_api_nodes.apinode_utils import ( + download_url_to_video_output, + tensor_to_bytesio, +) +from comfy_api_nodes.apis import ( + IngredientsMode, + PikaBodyGenerate22C2vGenerate22PikascenesPost, + PikaBodyGenerate22I2vGenerate22I2vPost, + PikaBodyGenerate22KeyframeGenerate22PikaframesPost, + PikaBodyGenerate22T2vGenerate22T2vPost, + PikaBodyGeneratePikadditionsGeneratePikadditionsPost, + PikaBodyGeneratePikaffectsGeneratePikaffectsPost, + PikaBodyGeneratePikaswapsGeneratePikaswapsPost, + PikaDurationEnum, + Pikaffect, + PikaGenerateResponse, + PikaResolutionEnum, + PikaVideoResponse, +) +from comfy_api_nodes.apis.client import ( + ApiEndpoint, + EmptyRequest, + HttpMethod, + PollingOperation, + SynchronousOperation, +) +from comfy_api_nodes.mapper_utils import model_field_to_node_input + +R = TypeVar("R") + +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" + + +class PikaApiError(Exception): + """Exception for Pika API errors.""" + + pass + + +def is_valid_video_response(response: PikaVideoResponse) -> bool: + """Check if the video response is valid.""" + return hasattr(response, "url") and response.url is not None + + +def is_valid_initial_response(response: PikaGenerateResponse) -> bool: + """Check if the initial response is valid.""" + return hasattr(response, "video_id") and response.video_id is not None + + +class PikaNodeBase(ComfyNodeABC): + """Base class for Pika nodes.""" + + @classmethod + def get_base_inputs_types( + cls, request_model + ) -> dict[str, tuple[IO, InputTypeOptions]]: + """Get the base required inputs types common to all Pika nodes.""" + return { + "prompt_text": model_field_to_node_input( + IO.STRING, + request_model, + "promptText", + multiline=True, + ), + "negative_prompt": model_field_to_node_input( + IO.STRING, + request_model, + "negativePrompt", + multiline=True, + ), + "seed": model_field_to_node_input( + IO.INT, + request_model, + "seed", + min=0, + max=0xFFFFFFFF, + control_after_generate=True, + ), + "resolution": model_field_to_node_input( + IO.COMBO, + request_model, + "resolution", + enum_type=PikaResolutionEnum, + ), + "duration": model_field_to_node_input( + IO.COMBO, + request_model, + "duration", + enum_type=PikaDurationEnum, + ), + } + + CATEGORY = "api node/video/Pika" + API_NODE = True + FUNCTION = "api_call" + RETURN_TYPES = ("VIDEO",) + + def poll_for_task_status( + self, + task_id: str, + auth_kwargs: Optional[dict[str, str]] = None, + node_id: Optional[str] = None, + ) -> PikaGenerateResponse: + polling_operation = PollingOperation( + poll_endpoint=ApiEndpoint( + path=f"{PATH_VIDEO_GET}/{task_id}", + method=HttpMethod.GET, + request_model=EmptyRequest, + response_model=PikaVideoResponse, + ), + completed_statuses=[ + "finished", + ], + failed_statuses=["failed", "cancelled"], + status_extractor=lambda response: ( + response.status.value if response.status else None + ), + progress_extractor=lambda response: ( + response.progress if hasattr(response, "progress") else None + ), + auth_kwargs=auth_kwargs, + result_url_extractor=lambda response: ( + response.url if hasattr(response, "url") else None + ), + node_id=node_id, + estimated_duration=60 + ) + return polling_operation.execute() + + def execute_task( + self, + initial_operation: SynchronousOperation[R, PikaGenerateResponse], + auth_kwargs: Optional[dict[str, str]] = None, + node_id: Optional[str] = None, + ) -> tuple[VideoFromFile]: + """Executes the initial operation then polls for the task status until it is completed. + + Args: + initial_operation: The initial operation to execute. + auth_kwargs: The authentication token(s) to use for the API call. + + Returns: + A tuple containing the video file as a VIDEO output. + """ + initial_response = initial_operation.execute() + if not is_valid_initial_response(initial_response): + error_msg = f"Pika initial request failed. Code: {initial_response.code}, Message: {initial_response.message}, Data: {initial_response.data}" + logging.error(error_msg) + raise PikaApiError(error_msg) + + task_id = initial_response.video_id + final_response = self.poll_for_task_status(task_id, auth_kwargs) + if not is_valid_video_response(final_response): + error_msg = ( + f"Pika task {task_id} succeeded but no video data found in response." + ) + logging.error(error_msg) + raise PikaApiError(error_msg) + + video_url = str(final_response.url) + logging.info("Pika task %s succeeded. Video URL: %s", task_id, video_url) + + return (download_url_to_video_output(video_url),) + + +class PikaImageToVideoV2_2(PikaNodeBase): + """Pika 2.2 Image to Video Node.""" + + @classmethod + def INPUT_TYPES(cls): + return { + "required": { + "image": ( + IO.IMAGE, + {"tooltip": "The image to convert to video"}, + ), + **cls.get_base_inputs_types(PikaBodyGenerate22I2vGenerate22I2vPost), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + DESCRIPTION = "Sends an image and prompt to the Pika API v2.2 to generate a video." + + def api_call( + self, + image: torch.Tensor, + prompt_text: str, + negative_prompt: str, + seed: int, + resolution: str, + duration: int, + unique_id: str, + **kwargs, + ) -> tuple[VideoFromFile]: + # Convert image to BytesIO + image_bytes_io = tensor_to_bytesio(image) + image_bytes_io.seek(0) + + pika_files = {"image": ("image.png", image_bytes_io, "image/png")} + + # Prepare non-file data + pika_request_data = PikaBodyGenerate22I2vGenerate22I2vPost( + promptText=prompt_text, + negativePrompt=negative_prompt, + seed=seed, + resolution=resolution, + duration=duration, + ) + + initial_operation = SynchronousOperation( + endpoint=ApiEndpoint( + path=PATH_IMAGE_TO_VIDEO, + method=HttpMethod.POST, + request_model=PikaBodyGenerate22I2vGenerate22I2vPost, + response_model=PikaGenerateResponse, + ), + request=pika_request_data, + files=pika_files, + content_type="multipart/form-data", + auth_kwargs=kwargs, + ) + + return self.execute_task(initial_operation, auth_kwargs=kwargs, node_id=unique_id) + + +class PikaTextToVideoNodeV2_2(PikaNodeBase): + """Pika Text2Video v2.2 Node.""" + + @classmethod + def INPUT_TYPES(cls): + return { + "required": { + **cls.get_base_inputs_types(PikaBodyGenerate22T2vGenerate22T2vPost), + "aspect_ratio": model_field_to_node_input( + IO.FLOAT, + PikaBodyGenerate22T2vGenerate22T2vPost, + "aspectRatio", + step=0.001, + min=0.4, + max=2.5, + default=1.7777777777777777, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + DESCRIPTION = "Sends a text prompt to the Pika API v2.2 to generate a video." + + def api_call( + self, + prompt_text: str, + negative_prompt: str, + seed: int, + resolution: str, + duration: int, + aspect_ratio: float, + unique_id: str, + **kwargs, + ) -> tuple[VideoFromFile]: + initial_operation = SynchronousOperation( + endpoint=ApiEndpoint( + path=PATH_TEXT_TO_VIDEO, + method=HttpMethod.POST, + request_model=PikaBodyGenerate22T2vGenerate22T2vPost, + response_model=PikaGenerateResponse, + ), + request=PikaBodyGenerate22T2vGenerate22T2vPost( + promptText=prompt_text, + negativePrompt=negative_prompt, + seed=seed, + resolution=resolution, + duration=duration, + aspectRatio=aspect_ratio, + ), + auth_kwargs=kwargs, + content_type="application/x-www-form-urlencoded", + ) + + return self.execute_task(initial_operation, auth_kwargs=kwargs, node_id=unique_id) + + +class PikaScenesV2_2(PikaNodeBase): + """PikaScenes v2.2 Node.""" + + @classmethod + def INPUT_TYPES(cls): + image_ingredient_input = ( + IO.IMAGE, + {"tooltip": "Image that will be used as ingredient to create a video."}, + ) + return { + "required": { + **cls.get_base_inputs_types( + PikaBodyGenerate22C2vGenerate22PikascenesPost, + ), + "ingredients_mode": model_field_to_node_input( + IO.COMBO, + PikaBodyGenerate22C2vGenerate22PikascenesPost, + "ingredientsMode", + enum_type=IngredientsMode, + default="creative", + ), + "aspect_ratio": model_field_to_node_input( + IO.FLOAT, + PikaBodyGenerate22C2vGenerate22PikascenesPost, + "aspectRatio", + step=0.001, + min=0.4, + max=2.5, + default=1.7777777777777777, + ), + }, + "optional": { + "image_ingredient_1": image_ingredient_input, + "image_ingredient_2": image_ingredient_input, + "image_ingredient_3": image_ingredient_input, + "image_ingredient_4": image_ingredient_input, + "image_ingredient_5": image_ingredient_input, + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + 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." + + def api_call( + self, + prompt_text: str, + negative_prompt: str, + seed: int, + resolution: str, + duration: int, + ingredients_mode: str, + aspect_ratio: float, + unique_id: str, + 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, + **kwargs, + ) -> tuple[VideoFromFile]: + # Convert all passed images to BytesIO + 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: + image_bytes_io = tensor_to_bytesio(image) + image_bytes_io.seek(0) + all_image_bytes_io.append(image_bytes_io) + + 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 = PikaBodyGenerate22C2vGenerate22PikascenesPost( + ingredientsMode=ingredients_mode, + promptText=prompt_text, + negativePrompt=negative_prompt, + seed=seed, + resolution=resolution, + duration=duration, + aspectRatio=aspect_ratio, + ) + + initial_operation = SynchronousOperation( + endpoint=ApiEndpoint( + path=PATH_PIKASCENES, + method=HttpMethod.POST, + request_model=PikaBodyGenerate22C2vGenerate22PikascenesPost, + response_model=PikaGenerateResponse, + ), + request=pika_request_data, + files=pika_files, + content_type="multipart/form-data", + auth_kwargs=kwargs, + ) + + return self.execute_task(initial_operation, auth_kwargs=kwargs, node_id=unique_id) + + +class PikAdditionsNode(PikaNodeBase): + """Pika Pikadditions Node. Add an image into a video.""" + + @classmethod + def INPUT_TYPES(cls): + return { + "required": { + "video": (IO.VIDEO, {"tooltip": "The video to add an image to."}), + "image": (IO.IMAGE, {"tooltip": "The image to add to the video."}), + "prompt_text": model_field_to_node_input( + IO.STRING, + PikaBodyGeneratePikadditionsGeneratePikadditionsPost, + "promptText", + multiline=True, + ), + "negative_prompt": model_field_to_node_input( + IO.STRING, + PikaBodyGeneratePikadditionsGeneratePikadditionsPost, + "negativePrompt", + multiline=True, + ), + "seed": model_field_to_node_input( + IO.INT, + PikaBodyGeneratePikadditionsGeneratePikadditionsPost, + "seed", + min=0, + max=0xFFFFFFFF, + control_after_generate=True, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + 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." + + def api_call( + self, + video: VideoInput, + image: torch.Tensor, + prompt_text: str, + negative_prompt: str, + seed: int, + unique_id: str, + **kwargs, + ) -> tuple[VideoFromFile]: + # Convert video to BytesIO + video_bytes_io = io.BytesIO() + video.save_to(video_bytes_io, format=VideoContainer.MP4, codec=VideoCodec.H264) + video_bytes_io.seek(0) + + # Convert image to BytesIO + image_bytes_io = tensor_to_bytesio(image) + image_bytes_io.seek(0) + + pika_files = [ + ("video", ("video.mp4", video_bytes_io, "video/mp4")), + ("image", ("image.png", image_bytes_io, "image/png")), + ] + + # Prepare non-file data + pika_request_data = PikaBodyGeneratePikadditionsGeneratePikadditionsPost( + promptText=prompt_text, + negativePrompt=negative_prompt, + seed=seed, + ) + + initial_operation = SynchronousOperation( + endpoint=ApiEndpoint( + path=PATH_PIKADDITIONS, + method=HttpMethod.POST, + request_model=PikaBodyGeneratePikadditionsGeneratePikadditionsPost, + response_model=PikaGenerateResponse, + ), + request=pika_request_data, + files=pika_files, + content_type="multipart/form-data", + auth_kwargs=kwargs, + ) + + return self.execute_task(initial_operation, auth_kwargs=kwargs, node_id=unique_id) + + +class PikaSwapsNode(PikaNodeBase): + """Pika Pikaswaps Node.""" + + @classmethod + def INPUT_TYPES(cls): + return { + "required": { + "video": (IO.VIDEO, {"tooltip": "The video to swap an object in."}), + "image": ( + IO.IMAGE, + { + "tooltip": "The image used to replace the masked object in the video." + }, + ), + "mask": ( + IO.MASK, + {"tooltip": "Use the mask to define areas in the video to replace"}, + ), + "prompt_text": model_field_to_node_input( + IO.STRING, + PikaBodyGeneratePikaswapsGeneratePikaswapsPost, + "promptText", + multiline=True, + ), + "negative_prompt": model_field_to_node_input( + IO.STRING, + PikaBodyGeneratePikaswapsGeneratePikaswapsPost, + "negativePrompt", + multiline=True, + ), + "seed": model_field_to_node_input( + IO.INT, + PikaBodyGeneratePikaswapsGeneratePikaswapsPost, + "seed", + min=0, + max=0xFFFFFFFF, + control_after_generate=True, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + 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." + RETURN_TYPES = ("VIDEO",) + + def api_call( + self, + video: VideoInput, + image: torch.Tensor, + mask: torch.Tensor, + prompt_text: str, + negative_prompt: str, + seed: int, + unique_id: str, + **kwargs, + ) -> tuple[VideoFromFile]: + # Convert video to BytesIO + video_bytes_io = io.BytesIO() + video.save_to(video_bytes_io, format=VideoContainer.MP4, codec=VideoCodec.H264) + video_bytes_io.seek(0) + + # Convert mask to binary mask with three channels + mask = torch.round(mask) + mask = mask.repeat(1, 3, 1, 1) + + # Convert 3-channel binary mask to BytesIO + mask_bytes_io = io.BytesIO() + mask_bytes_io.write(mask.numpy().astype(np.uint8)) + mask_bytes_io.seek(0) + + # Convert image to BytesIO + image_bytes_io = tensor_to_bytesio(image) + image_bytes_io.seek(0) + + pika_files = [ + ("video", ("video.mp4", video_bytes_io, "video/mp4")), + ("image", ("image.png", image_bytes_io, "image/png")), + ("modifyRegionMask", ("mask.png", mask_bytes_io, "image/png")), + ] + + # Prepare non-file data + pika_request_data = PikaBodyGeneratePikaswapsGeneratePikaswapsPost( + promptText=prompt_text, + negativePrompt=negative_prompt, + seed=seed, + ) + + initial_operation = SynchronousOperation( + endpoint=ApiEndpoint( + path=PATH_PIKADDITIONS, + method=HttpMethod.POST, + request_model=PikaBodyGeneratePikadditionsGeneratePikadditionsPost, + response_model=PikaGenerateResponse, + ), + request=pika_request_data, + files=pika_files, + content_type="multipart/form-data", + auth_kwargs=kwargs, + ) + + return self.execute_task(initial_operation, auth_kwargs=kwargs, node_id=unique_id) + + +class PikaffectsNode(PikaNodeBase): + """Pika Pikaffects Node.""" + + @classmethod + def INPUT_TYPES(cls): + return { + "required": { + "image": ( + IO.IMAGE, + {"tooltip": "The reference image to apply the Pikaffect to."}, + ), + "pikaffect": model_field_to_node_input( + IO.COMBO, + PikaBodyGeneratePikaffectsGeneratePikaffectsPost, + "pikaffect", + enum_type=Pikaffect, + default="Cake-ify", + ), + "prompt_text": model_field_to_node_input( + IO.STRING, + PikaBodyGeneratePikaffectsGeneratePikaffectsPost, + "promptText", + multiline=True, + ), + "negative_prompt": model_field_to_node_input( + IO.STRING, + PikaBodyGeneratePikaffectsGeneratePikaffectsPost, + "negativePrompt", + multiline=True, + ), + "seed": model_field_to_node_input( + IO.INT, + PikaBodyGeneratePikaffectsGeneratePikaffectsPost, + "seed", + min=0, + max=0xFFFFFFFF, + control_after_generate=True, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + 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" + + def api_call( + self, + image: torch.Tensor, + pikaffect: str, + prompt_text: str, + negative_prompt: str, + seed: int, + unique_id: str, + **kwargs, + ) -> tuple[VideoFromFile]: + + initial_operation = SynchronousOperation( + endpoint=ApiEndpoint( + path=PATH_PIKAFFECTS, + method=HttpMethod.POST, + request_model=PikaBodyGeneratePikaffectsGeneratePikaffectsPost, + response_model=PikaGenerateResponse, + ), + request=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", + auth_kwargs=kwargs, + ) + + return self.execute_task(initial_operation, auth_kwargs=kwargs, node_id=unique_id) + + +class PikaStartEndFrameNode2_2(PikaNodeBase): + """PikaFrames v2.2 Node.""" + + @classmethod + def INPUT_TYPES(cls): + return { + "required": { + "image_start": (IO.IMAGE, {"tooltip": "The first image to combine."}), + "image_end": (IO.IMAGE, {"tooltip": "The last image to combine."}), + **cls.get_base_inputs_types( + PikaBodyGenerate22KeyframeGenerate22PikaframesPost + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + 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." + + def api_call( + self, + image_start: torch.Tensor, + image_end: torch.Tensor, + prompt_text: str, + negative_prompt: str, + seed: int, + resolution: str, + duration: int, + unique_id: str, + **kwargs, + ) -> tuple[VideoFromFile]: + + 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 = SynchronousOperation( + endpoint=ApiEndpoint( + path=PATH_PIKAFRAMES, + method=HttpMethod.POST, + request_model=PikaBodyGenerate22KeyframeGenerate22PikaframesPost, + response_model=PikaGenerateResponse, + ), + request=PikaBodyGenerate22KeyframeGenerate22PikaframesPost( + promptText=prompt_text, + negativePrompt=negative_prompt, + seed=seed, + resolution=resolution, + duration=duration, + ), + files=pika_files, + content_type="multipart/form-data", + auth_kwargs=kwargs, + ) + + return self.execute_task(initial_operation, auth_kwargs=kwargs, node_id=unique_id) + + +NODE_CLASS_MAPPINGS = { + "PikaImageToVideoNode2_2": PikaImageToVideoV2_2, + "PikaTextToVideoNode2_2": PikaTextToVideoNodeV2_2, + "PikaScenesV2_2": PikaScenesV2_2, + "Pikadditions": PikAdditionsNode, + "Pikaswaps": PikaSwapsNode, + "Pikaffects": PikaffectsNode, + "PikaStartEndFrameNode2_2": PikaStartEndFrameNode2_2, +} + +NODE_DISPLAY_NAME_MAPPINGS = { + "PikaImageToVideoNode2_2": "Pika Image to Video", + "PikaTextToVideoNode2_2": "Pika Text to Video", + "PikaScenesV2_2": "Pika Scenes (Video Image Composition)", + "Pikadditions": "Pikadditions (Video Object Insertion)", + "Pikaswaps": "Pika Swaps (Video Object Replacement)", + "Pikaffects": "Pikaffects (Video Effects)", + "PikaStartEndFrameNode2_2": "Pika Start and End Frame to Video", +} diff --git a/comfy_api_nodes/nodes_pixverse.py b/comfy_api_nodes/nodes_pixverse.py new file mode 100644 index 000000000..ef4a9a802 --- /dev/null +++ b/comfy_api_nodes/nodes_pixverse.py @@ -0,0 +1,525 @@ +from inspect import cleandoc +from typing import Optional +from comfy_api_nodes.apis.pixverse_api import ( + PixverseTextVideoRequest, + PixverseImageVideoRequest, + PixverseTransitionVideoRequest, + PixverseImageUploadResponse, + PixverseVideoResponse, + PixverseGenerationStatusResponse, + PixverseAspectRatio, + PixverseQuality, + PixverseDuration, + PixverseMotionMode, + PixverseStatus, + PixverseIO, + pixverse_templates, +) +from comfy_api_nodes.apis.client import ( + ApiEndpoint, + HttpMethod, + SynchronousOperation, + PollingOperation, + EmptyRequest, +) +from comfy_api_nodes.apinode_utils import ( + tensor_to_bytesio, + validate_string, +) +from comfy.comfy_types.node_typing import IO, ComfyNodeABC +from comfy_api.input_impl import VideoFromFile + +import torch +import requests +from io import BytesIO + + +AVERAGE_DURATION_T2V = 32 +AVERAGE_DURATION_I2V = 30 +AVERAGE_DURATION_T2T = 52 + + +def get_video_url_from_response( + response: PixverseGenerationStatusResponse, +) -> Optional[str]: + if response.Resp is None or response.Resp.url is None: + return None + return str(response.Resp.url) + + +def upload_image_to_pixverse(image: torch.Tensor, auth_kwargs=None): + # first, upload image to Pixverse and get image id to use in actual generation call + files = {"image": tensor_to_bytesio(image)} + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/pixverse/image/upload", + method=HttpMethod.POST, + request_model=EmptyRequest, + response_model=PixverseImageUploadResponse, + ), + request=EmptyRequest(), + files=files, + content_type="multipart/form-data", + auth_kwargs=auth_kwargs, + ) + response_upload: PixverseImageUploadResponse = operation.execute() + + if response_upload.Resp is None: + raise Exception( + f"PixVerse image upload request failed: '{response_upload.ErrMsg}'" + ) + + return response_upload.Resp.img_id + + +class PixverseTemplateNode: + """ + Select template for PixVerse Video generation. + """ + + RETURN_TYPES = (PixverseIO.TEMPLATE,) + RETURN_NAMES = ("pixverse_template",) + FUNCTION = "create_template" + CATEGORY = "api node/video/PixVerse" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "template": (list(pixverse_templates.keys()),), + } + } + + def create_template(self, template: str): + template_id = pixverse_templates.get(template, None) + if template_id is None: + raise Exception(f"Template '{template}' is not recognized.") + # just return the integer + return (template_id,) + + +class PixverseTextToVideoNode(ComfyNodeABC): + """ + Generates videos based on prompt and output_size. + """ + + RETURN_TYPES = (IO.VIDEO,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/video/PixVerse" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Prompt for the video generation", + }, + ), + "aspect_ratio": ([ratio.value for ratio in PixverseAspectRatio],), + "quality": ( + [resolution.value for resolution in PixverseQuality], + { + "default": PixverseQuality.res_540p, + }, + ), + "duration_seconds": ([dur.value for dur in PixverseDuration],), + "motion_mode": ([mode.value for mode in PixverseMotionMode],), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 2147483647, + "control_after_generate": True, + "tooltip": "Seed for video generation.", + }, + ), + }, + "optional": { + "negative_prompt": ( + IO.STRING, + { + "default": "", + "forceInput": True, + "tooltip": "An optional text description of undesired elements on an image.", + }, + ), + "pixverse_template": ( + PixverseIO.TEMPLATE, + { + "tooltip": "An optional template to influence style of generation, created by the PixVerse Template node." + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + def api_call( + self, + prompt: str, + aspect_ratio: str, + quality: str, + duration_seconds: int, + motion_mode: str, + seed, + negative_prompt: str = None, + pixverse_template: int = None, + unique_id: Optional[str] = None, + **kwargs, + ): + validate_string(prompt, strip_whitespace=False) + # 1080p is limited to 5 seconds duration + # only normal motion_mode supported for 1080p or for non-5 second duration + if quality == PixverseQuality.res_1080p: + motion_mode = PixverseMotionMode.normal + duration_seconds = PixverseDuration.dur_5 + elif duration_seconds != PixverseDuration.dur_5: + motion_mode = PixverseMotionMode.normal + + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/pixverse/video/text/generate", + method=HttpMethod.POST, + request_model=PixverseTextVideoRequest, + response_model=PixverseVideoResponse, + ), + request=PixverseTextVideoRequest( + prompt=prompt, + aspect_ratio=aspect_ratio, + quality=quality, + duration=duration_seconds, + motion_mode=motion_mode, + negative_prompt=negative_prompt if negative_prompt else None, + template_id=pixverse_template, + seed=seed, + ), + auth_kwargs=kwargs, + ) + response_api = operation.execute() + + if response_api.Resp is None: + raise Exception(f"PixVerse request failed: '{response_api.ErrMsg}'") + + operation = PollingOperation( + poll_endpoint=ApiEndpoint( + path=f"/proxy/pixverse/video/result/{response_api.Resp.video_id}", + method=HttpMethod.GET, + request_model=EmptyRequest, + response_model=PixverseGenerationStatusResponse, + ), + completed_statuses=[PixverseStatus.successful], + failed_statuses=[ + PixverseStatus.contents_moderation, + PixverseStatus.failed, + PixverseStatus.deleted, + ], + status_extractor=lambda x: x.Resp.status, + auth_kwargs=kwargs, + node_id=unique_id, + result_url_extractor=get_video_url_from_response, + estimated_duration=AVERAGE_DURATION_T2V, + ) + response_poll = operation.execute() + + vid_response = requests.get(response_poll.Resp.url) + + return (VideoFromFile(BytesIO(vid_response.content)),) + + +class PixverseImageToVideoNode(ComfyNodeABC): + """ + Generates videos based on prompt and output_size. + """ + + RETURN_TYPES = (IO.VIDEO,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/video/PixVerse" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "image": (IO.IMAGE,), + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Prompt for the video generation", + }, + ), + "quality": ( + [resolution.value for resolution in PixverseQuality], + { + "default": PixverseQuality.res_540p, + }, + ), + "duration_seconds": ([dur.value for dur in PixverseDuration],), + "motion_mode": ([mode.value for mode in PixverseMotionMode],), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 2147483647, + "control_after_generate": True, + "tooltip": "Seed for video generation.", + }, + ), + }, + "optional": { + "negative_prompt": ( + IO.STRING, + { + "default": "", + "forceInput": True, + "tooltip": "An optional text description of undesired elements on an image.", + }, + ), + "pixverse_template": ( + PixverseIO.TEMPLATE, + { + "tooltip": "An optional template to influence style of generation, created by the PixVerse Template node." + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + def api_call( + self, + image: torch.Tensor, + prompt: str, + quality: str, + duration_seconds: int, + motion_mode: str, + seed, + negative_prompt: str = None, + pixverse_template: int = None, + unique_id: Optional[str] = None, + **kwargs, + ): + validate_string(prompt, strip_whitespace=False) + img_id = upload_image_to_pixverse(image, auth_kwargs=kwargs) + + # 1080p is limited to 5 seconds duration + # only normal motion_mode supported for 1080p or for non-5 second duration + if quality == PixverseQuality.res_1080p: + motion_mode = PixverseMotionMode.normal + duration_seconds = PixverseDuration.dur_5 + elif duration_seconds != PixverseDuration.dur_5: + motion_mode = PixverseMotionMode.normal + + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/pixverse/video/img/generate", + method=HttpMethod.POST, + request_model=PixverseImageVideoRequest, + response_model=PixverseVideoResponse, + ), + request=PixverseImageVideoRequest( + img_id=img_id, + prompt=prompt, + quality=quality, + duration=duration_seconds, + motion_mode=motion_mode, + negative_prompt=negative_prompt if negative_prompt else None, + template_id=pixverse_template, + seed=seed, + ), + auth_kwargs=kwargs, + ) + response_api = operation.execute() + + if response_api.Resp is None: + raise Exception(f"PixVerse request failed: '{response_api.ErrMsg}'") + + operation = PollingOperation( + poll_endpoint=ApiEndpoint( + path=f"/proxy/pixverse/video/result/{response_api.Resp.video_id}", + method=HttpMethod.GET, + request_model=EmptyRequest, + response_model=PixverseGenerationStatusResponse, + ), + completed_statuses=[PixverseStatus.successful], + failed_statuses=[ + PixverseStatus.contents_moderation, + PixverseStatus.failed, + PixverseStatus.deleted, + ], + status_extractor=lambda x: x.Resp.status, + auth_kwargs=kwargs, + node_id=unique_id, + result_url_extractor=get_video_url_from_response, + estimated_duration=AVERAGE_DURATION_I2V, + ) + response_poll = operation.execute() + + vid_response = requests.get(response_poll.Resp.url) + return (VideoFromFile(BytesIO(vid_response.content)),) + + +class PixverseTransitionVideoNode(ComfyNodeABC): + """ + Generates videos based on prompt and output_size. + """ + + RETURN_TYPES = (IO.VIDEO,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/video/PixVerse" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "first_frame": (IO.IMAGE,), + "last_frame": (IO.IMAGE,), + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Prompt for the video generation", + }, + ), + "quality": ( + [resolution.value for resolution in PixverseQuality], + { + "default": PixverseQuality.res_540p, + }, + ), + "duration_seconds": ([dur.value for dur in PixverseDuration],), + "motion_mode": ([mode.value for mode in PixverseMotionMode],), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 2147483647, + "control_after_generate": True, + "tooltip": "Seed for video generation.", + }, + ), + }, + "optional": { + "negative_prompt": ( + IO.STRING, + { + "default": "", + "forceInput": True, + "tooltip": "An optional text description of undesired elements on an image.", + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + def api_call( + self, + first_frame: torch.Tensor, + last_frame: torch.Tensor, + prompt: str, + quality: str, + duration_seconds: int, + motion_mode: str, + seed, + negative_prompt: str = None, + unique_id: Optional[str] = None, + **kwargs, + ): + validate_string(prompt, strip_whitespace=False) + first_frame_id = upload_image_to_pixverse(first_frame, auth_kwargs=kwargs) + last_frame_id = upload_image_to_pixverse(last_frame, auth_kwargs=kwargs) + + # 1080p is limited to 5 seconds duration + # only normal motion_mode supported for 1080p or for non-5 second duration + if quality == PixverseQuality.res_1080p: + motion_mode = PixverseMotionMode.normal + duration_seconds = PixverseDuration.dur_5 + elif duration_seconds != PixverseDuration.dur_5: + motion_mode = PixverseMotionMode.normal + + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/pixverse/video/transition/generate", + method=HttpMethod.POST, + request_model=PixverseTransitionVideoRequest, + response_model=PixverseVideoResponse, + ), + request=PixverseTransitionVideoRequest( + first_frame_img=first_frame_id, + last_frame_img=last_frame_id, + prompt=prompt, + quality=quality, + duration=duration_seconds, + motion_mode=motion_mode, + negative_prompt=negative_prompt if negative_prompt else None, + seed=seed, + ), + auth_kwargs=kwargs, + ) + response_api = operation.execute() + + if response_api.Resp is None: + raise Exception(f"PixVerse request failed: '{response_api.ErrMsg}'") + + operation = PollingOperation( + poll_endpoint=ApiEndpoint( + path=f"/proxy/pixverse/video/result/{response_api.Resp.video_id}", + method=HttpMethod.GET, + request_model=EmptyRequest, + response_model=PixverseGenerationStatusResponse, + ), + completed_statuses=[PixverseStatus.successful], + failed_statuses=[ + PixverseStatus.contents_moderation, + PixverseStatus.failed, + PixverseStatus.deleted, + ], + status_extractor=lambda x: x.Resp.status, + auth_kwargs=kwargs, + node_id=unique_id, + result_url_extractor=get_video_url_from_response, + estimated_duration=AVERAGE_DURATION_T2V, + ) + response_poll = operation.execute() + + vid_response = requests.get(response_poll.Resp.url) + return (VideoFromFile(BytesIO(vid_response.content)),) + + +NODE_CLASS_MAPPINGS = { + "PixverseTextToVideoNode": PixverseTextToVideoNode, + "PixverseImageToVideoNode": PixverseImageToVideoNode, + "PixverseTransitionVideoNode": PixverseTransitionVideoNode, + "PixverseTemplateNode": PixverseTemplateNode, +} + +NODE_DISPLAY_NAME_MAPPINGS = { + "PixverseTextToVideoNode": "PixVerse Text to Video", + "PixverseImageToVideoNode": "PixVerse Image to Video", + "PixverseTransitionVideoNode": "PixVerse Transition Video", + "PixverseTemplateNode": "PixVerse Template", +} diff --git a/comfy_api_nodes/nodes_recraft.py b/comfy_api_nodes/nodes_recraft.py new file mode 100644 index 000000000..e369c4b7e --- /dev/null +++ b/comfy_api_nodes/nodes_recraft.py @@ -0,0 +1,1138 @@ +from __future__ import annotations +from inspect import cleandoc +from typing import Optional +from comfy.utils import ProgressBar +from comfy_extras.nodes_images import SVG # Added +from comfy.comfy_types.node_typing import IO +from comfy_api_nodes.apis.recraft_api import ( + RecraftImageGenerationRequest, + RecraftImageGenerationResponse, + RecraftImageSize, + RecraftModel, + RecraftStyle, + RecraftStyleV3, + RecraftColor, + RecraftColorChain, + RecraftControls, + RecraftIO, + get_v3_substyles, +) +from comfy_api_nodes.apis.client import ( + ApiEndpoint, + HttpMethod, + SynchronousOperation, + EmptyRequest, +) +from comfy_api_nodes.apinode_utils import ( + bytesio_to_image_tensor, + download_url_to_bytesio, + tensor_to_bytesio, + resize_mask_to_image, + validate_string, +) +from server import PromptServer + +import torch +from io import BytesIO +from PIL import UnidentifiedImageError + + +def handle_recraft_file_request( + image: torch.Tensor, + path: str, + mask: torch.Tensor=None, + total_pixels=4096*4096, + timeout=1024, + request=None, + auth_kwargs: dict[str,str] = None, + ) -> list[BytesIO]: + """ + Handle sending common Recraft file-only request to get back file bytes. + """ + if request is None: + request = EmptyRequest() + + files = { + 'image': tensor_to_bytesio(image, total_pixels=total_pixels).read() + } + if mask is not None: + files['mask'] = tensor_to_bytesio(mask, total_pixels=total_pixels).read() + + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path=path, + method=HttpMethod.POST, + request_model=type(request), + response_model=RecraftImageGenerationResponse, + ), + request=request, + files=files, + content_type="multipart/form-data", + auth_kwargs=auth_kwargs, + multipart_parser=recraft_multipart_parser, + ) + response: RecraftImageGenerationResponse = operation.execute() + all_bytesio = [] + if response.image is not None: + all_bytesio.append(download_url_to_bytesio(response.image.url, timeout=timeout)) + else: + for data in response.data: + all_bytesio.append(download_url_to_bytesio(data.url, timeout=timeout)) + + return all_bytesio + + +def recraft_multipart_parser(data, parent_key=None, formatter: callable=None, converted_to_check: list[list]=None, is_list=False) -> dict: + """ + Formats data such that multipart/form-data will work with requests library + when both files and data are present. + + The OpenAI client that Recraft uses has a bizarre way of serializing lists: + + It does NOT keep track of indeces of each list, so for background_color, that must be serialized as: + 'background_color[rgb][]' = [0, 0, 255] + where the array is assigned to a key that has '[]' at the end, to signal it's an array. + + This has the consequence of nested lists having the exact same key, forcing arrays to merge; all colors inputs fall under the same key: + if 1 color -> 'controls[colors][][rgb][]' = [0, 0, 255] + if 2 colors -> 'controls[colors][][rgb][]' = [0, 0, 255, 255, 0, 0] + if 3 colors -> 'controls[colors][][rgb][]' = [0, 0, 255, 255, 0, 0, 0, 255, 0] + etc. + Whoever made this serialization up at OpenAI added the constraint that lists must be of uniform length on objects of same 'type'. + """ + # Modification of a function that handled a different type of multipart parsing, big ups: + # https://gist.github.com/kazqvaizer/4cebebe5db654a414132809f9f88067b + + def handle_converted_lists(data, parent_key, lists_to_check=tuple[list]): + # if list already exists exists, just extend list with data + for check_list in lists_to_check: + for conv_tuple in check_list: + if conv_tuple[0] == parent_key and type(conv_tuple[1]) is list: + conv_tuple[1].append(formatter(data)) + return True + return False + + if converted_to_check is None: + converted_to_check = [] + + + if formatter is None: + formatter = lambda v: v # Multipart representation of value + + if type(data) is not dict: + # if list already exists exists, just extend list with data + added = handle_converted_lists(data, parent_key, converted_to_check) + if added: + return {} + # otherwise if is_list, create new list with data + if is_list: + return {parent_key: [formatter(data)]} + # return new key with data + return {parent_key: formatter(data)} + + converted = [] + next_check = [converted] + next_check.extend(converted_to_check) + + for key, value in data.items(): + current_key = key if parent_key is None else f"{parent_key}[{key}]" + if type(value) is dict: + converted.extend(recraft_multipart_parser(value, current_key, formatter, next_check).items()) + elif type(value) is list: + for ind, list_value in enumerate(value): + iter_key = f"{current_key}[]" + converted.extend(recraft_multipart_parser(list_value, iter_key, formatter, next_check, is_list=True).items()) + else: + converted.append((current_key, formatter(value))) + + return dict(converted) + + +class handle_recraft_image_output: + """ + Catch an exception related to receiving SVG data instead of image, when Infinite Style Library style_id is in use. + """ + def __init__(self): + pass + + def __enter__(self): + pass + + def __exit__(self, exc_type, exc_val, exc_tb): + if exc_type is not None and exc_type is UnidentifiedImageError: + raise Exception("Received output data was not an image; likely an SVG. If you used style_id, make sure it is not a Vector art style.") + + +class RecraftColorRGBNode: + """ + Create Recraft Color by choosing specific RGB values. + """ + + RETURN_TYPES = (RecraftIO.COLOR,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + RETURN_NAMES = ("recraft_color",) + FUNCTION = "create_color" + CATEGORY = "api node/image/Recraft" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "r": (IO.INT, { + "default": 0, + "min": 0, + "max": 255, + "tooltip": "Red value of color." + }), + "g": (IO.INT, { + "default": 0, + "min": 0, + "max": 255, + "tooltip": "Green value of color." + }), + "b": (IO.INT, { + "default": 0, + "min": 0, + "max": 255, + "tooltip": "Blue value of color." + }), + }, + "optional": { + "recraft_color": (RecraftIO.COLOR,), + } + } + + def create_color(self, r: int, g: int, b: int, recraft_color: RecraftColorChain=None): + recraft_color = recraft_color.clone() if recraft_color else RecraftColorChain() + recraft_color.add(RecraftColor(r, g, b)) + return (recraft_color, ) + + +class RecraftControlsNode: + """ + Create Recraft Controls for customizing Recraft generation. + """ + + RETURN_TYPES = (RecraftIO.CONTROLS,) + RETURN_NAMES = ("recraft_controls",) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "create_controls" + CATEGORY = "api node/image/Recraft" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + }, + "optional": { + "colors": (RecraftIO.COLOR,), + "background_color": (RecraftIO.COLOR,), + } + } + + def create_controls(self, colors: RecraftColorChain=None, background_color: RecraftColorChain=None): + return (RecraftControls(colors=colors, background_color=background_color), ) + + +class RecraftStyleV3RealisticImageNode: + """ + Select realistic_image style and optional substyle. + """ + + RETURN_TYPES = (RecraftIO.STYLEV3,) + RETURN_NAMES = ("recraft_style",) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "create_style" + CATEGORY = "api node/image/Recraft" + + RECRAFT_STYLE = RecraftStyleV3.realistic_image + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "substyle": (get_v3_substyles(s.RECRAFT_STYLE),), + } + } + + def create_style(self, substyle: str): + if substyle == "None": + substyle = None + return (RecraftStyle(self.RECRAFT_STYLE, substyle),) + + +class RecraftStyleV3DigitalIllustrationNode(RecraftStyleV3RealisticImageNode): + """ + Select digital_illustration style and optional substyle. + """ + + RECRAFT_STYLE = RecraftStyleV3.digital_illustration + + +class RecraftStyleV3VectorIllustrationNode(RecraftStyleV3RealisticImageNode): + """ + Select vector_illustration style and optional substyle. + """ + + RECRAFT_STYLE = RecraftStyleV3.vector_illustration + + +class RecraftStyleV3LogoRasterNode(RecraftStyleV3RealisticImageNode): + """ + Select vector_illustration style and optional substyle. + """ + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "substyle": (get_v3_substyles(s.RECRAFT_STYLE, include_none=False),), + } + } + + RECRAFT_STYLE = RecraftStyleV3.logo_raster + + +class RecraftStyleInfiniteStyleLibrary: + """ + Select style based on preexisting UUID from Recraft's Infinite Style Library. + """ + + RETURN_TYPES = (RecraftIO.STYLEV3,) + RETURN_NAMES = ("recraft_style",) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "create_style" + CATEGORY = "api node/image/Recraft" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "style_id": (IO.STRING, { + "default": "", + "tooltip": "UUID of style from Infinite Style Library.", + }) + } + } + + def create_style(self, style_id: str): + if not style_id: + raise Exception("The style_id input cannot be empty.") + return (RecraftStyle(style_id=style_id),) + + +class RecraftTextToImageNode: + """ + Generates images synchronously based on prompt and resolution. + """ + + RETURN_TYPES = (IO.IMAGE,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/image/Recraft" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Prompt for the image generation.", + }, + ), + "size": ( + [res.value for res in RecraftImageSize], + { + "default": RecraftImageSize.res_1024x1024, + "tooltip": "The size of the generated image.", + }, + ), + "n": ( + IO.INT, + { + "default": 1, + "min": 1, + "max": 6, + "tooltip": "The number of images to generate.", + }, + ), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 0xFFFFFFFFFFFFFFFF, + "control_after_generate": True, + "tooltip": "Seed to determine if node should re-run; actual results are nondeterministic regardless of seed.", + }, + ), + }, + "optional": { + "recraft_style": (RecraftIO.STYLEV3,), + "negative_prompt": ( + IO.STRING, + { + "default": "", + "forceInput": True, + "tooltip": "An optional text description of undesired elements on an image.", + }, + ), + "recraft_controls": ( + RecraftIO.CONTROLS, + { + "tooltip": "Optional additional controls over the generation via the Recraft Controls node." + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + def api_call( + self, + prompt: str, + size: str, + n: int, + seed, + recraft_style: RecraftStyle = None, + negative_prompt: str = None, + recraft_controls: RecraftControls = None, + unique_id: Optional[str] = None, + **kwargs, + ): + validate_string(prompt, strip_whitespace=False, max_length=1000) + default_style = RecraftStyle(RecraftStyleV3.realistic_image) + if recraft_style is None: + recraft_style = default_style + + controls_api = None + if recraft_controls: + controls_api = recraft_controls.create_api_model() + + if not negative_prompt: + negative_prompt = None + + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/recraft/image_generation", + method=HttpMethod.POST, + request_model=RecraftImageGenerationRequest, + response_model=RecraftImageGenerationResponse, + ), + request=RecraftImageGenerationRequest( + prompt=prompt, + negative_prompt=negative_prompt, + model=RecraftModel.recraftv3, + size=size, + n=n, + style=recraft_style.style, + substyle=recraft_style.substyle, + style_id=recraft_style.style_id, + controls=controls_api, + ), + auth_kwargs=kwargs, + ) + response: RecraftImageGenerationResponse = operation.execute() + images = [] + urls = [] + for data in response.data: + with handle_recraft_image_output(): + if unique_id and data.url: + urls.append(data.url) + urls_string = '\n'.join(urls) + PromptServer.instance.send_progress_text( + f"Result URL: {urls_string}", unique_id + ) + image = bytesio_to_image_tensor( + download_url_to_bytesio(data.url, timeout=1024) + ) + if len(image.shape) < 4: + image = image.unsqueeze(0) + images.append(image) + output_image = torch.cat(images, dim=0) + + return (output_image,) + + +class RecraftImageToImageNode: + """ + Modify image based on prompt and strength. + """ + + RETURN_TYPES = (IO.IMAGE,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/image/Recraft" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "image": (IO.IMAGE, ), + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Prompt for the image generation.", + }, + ), + "n": ( + IO.INT, + { + "default": 1, + "min": 1, + "max": 6, + "tooltip": "The number of images to generate.", + }, + ), + "strength": ( + IO.FLOAT, + { + "default": 0.5, + "min": 0.0, + "max": 1.0, + "step": 0.01, + "tooltip": "Defines the difference with the original image, should lie in [0, 1], where 0 means almost identical, and 1 means miserable similarity." + } + ), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 0xFFFFFFFFFFFFFFFF, + "control_after_generate": True, + "tooltip": "Seed to determine if node should re-run; actual results are nondeterministic regardless of seed.", + }, + ), + }, + "optional": { + "recraft_style": (RecraftIO.STYLEV3,), + "negative_prompt": ( + IO.STRING, + { + "default": "", + "forceInput": True, + "tooltip": "An optional text description of undesired elements on an image.", + }, + ), + "recraft_controls": ( + RecraftIO.CONTROLS, + { + "tooltip": "Optional additional controls over the generation via the Recraft Controls node." + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + }, + } + + def api_call( + self, + image: torch.Tensor, + prompt: str, + n: int, + strength: float, + seed, + recraft_style: RecraftStyle = None, + negative_prompt: str = None, + recraft_controls: RecraftControls = None, + **kwargs, + ): + validate_string(prompt, strip_whitespace=False, max_length=1000) + default_style = RecraftStyle(RecraftStyleV3.realistic_image) + if recraft_style is None: + recraft_style = default_style + + controls_api = None + if recraft_controls: + controls_api = recraft_controls.create_api_model() + + if not negative_prompt: + negative_prompt = None + + request = RecraftImageGenerationRequest( + prompt=prompt, + negative_prompt=negative_prompt, + model=RecraftModel.recraftv3, + n=n, + strength=round(strength, 2), + style=recraft_style.style, + substyle=recraft_style.substyle, + style_id=recraft_style.style_id, + controls=controls_api, + ) + + images = [] + total = image.shape[0] + pbar = ProgressBar(total) + for i in range(total): + sub_bytes = handle_recraft_file_request( + image=image[i], + path="/proxy/recraft/images/imageToImage", + request=request, + auth_kwargs=kwargs, + ) + with handle_recraft_image_output(): + images.append(torch.cat([bytesio_to_image_tensor(x) for x in sub_bytes], dim=0)) + pbar.update(1) + + images_tensor = torch.cat(images, dim=0) + return (images_tensor, ) + + +class RecraftImageInpaintingNode: + """ + Modify image based on prompt and mask. + """ + + RETURN_TYPES = (IO.IMAGE,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/image/Recraft" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "image": (IO.IMAGE, ), + "mask": (IO.MASK, ), + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Prompt for the image generation.", + }, + ), + "n": ( + IO.INT, + { + "default": 1, + "min": 1, + "max": 6, + "tooltip": "The number of images to generate.", + }, + ), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 0xFFFFFFFFFFFFFFFF, + "control_after_generate": True, + "tooltip": "Seed to determine if node should re-run; actual results are nondeterministic regardless of seed.", + }, + ), + }, + "optional": { + "recraft_style": (RecraftIO.STYLEV3,), + "negative_prompt": ( + IO.STRING, + { + "default": "", + "forceInput": True, + "tooltip": "An optional text description of undesired elements on an image.", + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + }, + } + + def api_call( + self, + image: torch.Tensor, + mask: torch.Tensor, + prompt: str, + n: int, + seed, + recraft_style: RecraftStyle = None, + negative_prompt: str = None, + **kwargs, + ): + validate_string(prompt, strip_whitespace=False, max_length=1000) + default_style = RecraftStyle(RecraftStyleV3.realistic_image) + if recraft_style is None: + recraft_style = default_style + + if not negative_prompt: + negative_prompt = None + + request = RecraftImageGenerationRequest( + prompt=prompt, + negative_prompt=negative_prompt, + model=RecraftModel.recraftv3, + n=n, + style=recraft_style.style, + substyle=recraft_style.substyle, + style_id=recraft_style.style_id, + ) + + # prepare mask tensor + mask = resize_mask_to_image(mask, image, allow_gradient=False, add_channel_dim=True) + + images = [] + total = image.shape[0] + pbar = ProgressBar(total) + for i in range(total): + sub_bytes = handle_recraft_file_request( + image=image[i], + mask=mask[i:i+1], + path="/proxy/recraft/images/inpaint", + request=request, + auth_kwargs=kwargs, + ) + with handle_recraft_image_output(): + images.append(torch.cat([bytesio_to_image_tensor(x) for x in sub_bytes], dim=0)) + pbar.update(1) + + images_tensor = torch.cat(images, dim=0) + return (images_tensor, ) + + +class RecraftTextToVectorNode: + """ + Generates SVG synchronously based on prompt and resolution. + """ + + RETURN_TYPES = ("SVG",) # Changed + DESCRIPTION = cleandoc(__doc__ or "") if 'cleandoc' in globals() else __doc__ # Keep cleandoc if other nodes use it + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/image/Recraft" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Prompt for the image generation.", + }, + ), + "substyle": (get_v3_substyles(RecraftStyleV3.vector_illustration),), + "size": ( + [res.value for res in RecraftImageSize], + { + "default": RecraftImageSize.res_1024x1024, + "tooltip": "The size of the generated image.", + }, + ), + "n": ( + IO.INT, + { + "default": 1, + "min": 1, + "max": 6, + "tooltip": "The number of images to generate.", + }, + ), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 0xFFFFFFFFFFFFFFFF, + "control_after_generate": True, + "tooltip": "Seed to determine if node should re-run; actual results are nondeterministic regardless of seed.", + }, + ), + }, + "optional": { + "negative_prompt": ( + IO.STRING, + { + "default": "", + "forceInput": True, + "tooltip": "An optional text description of undesired elements on an image.", + }, + ), + "recraft_controls": ( + RecraftIO.CONTROLS, + { + "tooltip": "Optional additional controls over the generation via the Recraft Controls node." + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + def api_call( + self, + prompt: str, + substyle: str, + size: str, + n: int, + seed, + negative_prompt: str = None, + recraft_controls: RecraftControls = None, + unique_id: Optional[str] = None, + **kwargs, + ): + validate_string(prompt, strip_whitespace=False, max_length=1000) + # create RecraftStyle so strings will be formatted properly (i.e. "None" will become None) + recraft_style = RecraftStyle(RecraftStyleV3.vector_illustration, substyle=substyle) + + controls_api = None + if recraft_controls: + controls_api = recraft_controls.create_api_model() + + if not negative_prompt: + negative_prompt = None + + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/recraft/image_generation", + method=HttpMethod.POST, + request_model=RecraftImageGenerationRequest, + response_model=RecraftImageGenerationResponse, + ), + request=RecraftImageGenerationRequest( + prompt=prompt, + negative_prompt=negative_prompt, + model=RecraftModel.recraftv3, + size=size, + n=n, + style=recraft_style.style, + substyle=recraft_style.substyle, + controls=controls_api, + ), + auth_kwargs=kwargs, + ) + response: RecraftImageGenerationResponse = operation.execute() + svg_data = [] + urls = [] + for data in response.data: + if unique_id and data.url: + urls.append(data.url) + # Print result on each iteration in case of error + PromptServer.instance.send_progress_text( + f"Result URL: {' '.join(urls)}", unique_id + ) + svg_data.append(download_url_to_bytesio(data.url, timeout=1024)) + + return (SVG(svg_data),) + + +class RecraftVectorizeImageNode: + """ + Generates SVG synchronously from an input image. + """ + + RETURN_TYPES = ("SVG",) # Changed + DESCRIPTION = cleandoc(__doc__ or "") if 'cleandoc' in globals() else __doc__ # Keep cleandoc if other nodes use it + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/image/Recraft" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "image": (IO.IMAGE, ), + }, + "optional": { + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + }, + } + + def api_call( + self, + image: torch.Tensor, + **kwargs, + ): + svgs = [] + total = image.shape[0] + pbar = ProgressBar(total) + for i in range(total): + sub_bytes = handle_recraft_file_request( + image=image[i], + path="/proxy/recraft/images/vectorize", + auth_kwargs=kwargs, + ) + svgs.append(SVG(sub_bytes)) + pbar.update(1) + + return (SVG.combine_all(svgs), ) + + +class RecraftReplaceBackgroundNode: + """ + Replace background on image, based on provided prompt. + """ + + RETURN_TYPES = (IO.IMAGE,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/image/Recraft" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "image": (IO.IMAGE, ), + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Prompt for the image generation.", + }, + ), + "n": ( + IO.INT, + { + "default": 1, + "min": 1, + "max": 6, + "tooltip": "The number of images to generate.", + }, + ), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 0xFFFFFFFFFFFFFFFF, + "control_after_generate": True, + "tooltip": "Seed to determine if node should re-run; actual results are nondeterministic regardless of seed.", + }, + ), + }, + "optional": { + "recraft_style": (RecraftIO.STYLEV3,), + "negative_prompt": ( + IO.STRING, + { + "default": "", + "forceInput": True, + "tooltip": "An optional text description of undesired elements on an image.", + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + }, + } + + def api_call( + self, + image: torch.Tensor, + prompt: str, + n: int, + seed, + recraft_style: RecraftStyle = None, + negative_prompt: str = None, + **kwargs, + ): + default_style = RecraftStyle(RecraftStyleV3.realistic_image) + if recraft_style is None: + recraft_style = default_style + + if not negative_prompt: + negative_prompt = None + + request = RecraftImageGenerationRequest( + prompt=prompt, + negative_prompt=negative_prompt, + model=RecraftModel.recraftv3, + n=n, + style=recraft_style.style, + substyle=recraft_style.substyle, + style_id=recraft_style.style_id, + ) + + images = [] + total = image.shape[0] + pbar = ProgressBar(total) + for i in range(total): + sub_bytes = handle_recraft_file_request( + image=image[i], + path="/proxy/recraft/images/replaceBackground", + request=request, + auth_kwargs=kwargs, + ) + images.append(torch.cat([bytesio_to_image_tensor(x) for x in sub_bytes], dim=0)) + pbar.update(1) + + images_tensor = torch.cat(images, dim=0) + return (images_tensor, ) + + +class RecraftRemoveBackgroundNode: + """ + Remove background from image, and return processed image and mask. + """ + + RETURN_TYPES = (IO.IMAGE, IO.MASK) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/image/Recraft" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "image": (IO.IMAGE, ), + }, + "optional": { + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + }, + } + + def api_call( + self, + image: torch.Tensor, + **kwargs, + ): + images = [] + total = image.shape[0] + pbar = ProgressBar(total) + for i in range(total): + sub_bytes = handle_recraft_file_request( + image=image[i], + path="/proxy/recraft/images/removeBackground", + auth_kwargs=kwargs, + ) + images.append(torch.cat([bytesio_to_image_tensor(x) for x in sub_bytes], dim=0)) + pbar.update(1) + + images_tensor = torch.cat(images, dim=0) + # use alpha channel as masks, in B,H,W format + masks_tensor = images_tensor[:,:,:,-1:].squeeze(-1) + return (images_tensor, masks_tensor) + + +class RecraftCrispUpscaleNode: + """ + Upscale image synchronously. + Enhances a given raster image using ‘crisp upscale’ tool, increasing image resolution, making the image sharper and cleaner. + """ + + RETURN_TYPES = (IO.IMAGE,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/image/Recraft" + + RECRAFT_PATH = "/proxy/recraft/images/crispUpscale" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "image": (IO.IMAGE, ), + }, + "optional": { + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + }, + } + + def api_call( + self, + image: torch.Tensor, + **kwargs, + ): + images = [] + total = image.shape[0] + pbar = ProgressBar(total) + for i in range(total): + sub_bytes = handle_recraft_file_request( + image=image[i], + path=self.RECRAFT_PATH, + auth_kwargs=kwargs, + ) + images.append(torch.cat([bytesio_to_image_tensor(x) for x in sub_bytes], dim=0)) + pbar.update(1) + + images_tensor = torch.cat(images, dim=0) + return (images_tensor,) + + +class RecraftCreativeUpscaleNode(RecraftCrispUpscaleNode): + """ + Upscale image synchronously. + Enhances a given raster image using ‘creative upscale’ tool, boosting resolution with a focus on refining small details and faces. + """ + + RETURN_TYPES = (IO.IMAGE,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/image/Recraft" + + RECRAFT_PATH = "/proxy/recraft/images/creativeUpscale" + + +# A dictionary that contains all nodes you want to export with their names +# NOTE: names should be globally unique +NODE_CLASS_MAPPINGS = { + "RecraftTextToImageNode": RecraftTextToImageNode, + "RecraftImageToImageNode": RecraftImageToImageNode, + "RecraftImageInpaintingNode": RecraftImageInpaintingNode, + "RecraftTextToVectorNode": RecraftTextToVectorNode, + "RecraftVectorizeImageNode": RecraftVectorizeImageNode, + "RecraftRemoveBackgroundNode": RecraftRemoveBackgroundNode, + "RecraftReplaceBackgroundNode": RecraftReplaceBackgroundNode, + "RecraftCrispUpscaleNode": RecraftCrispUpscaleNode, + "RecraftCreativeUpscaleNode": RecraftCreativeUpscaleNode, + "RecraftStyleV3RealisticImage": RecraftStyleV3RealisticImageNode, + "RecraftStyleV3DigitalIllustration": RecraftStyleV3DigitalIllustrationNode, + "RecraftStyleV3LogoRaster": RecraftStyleV3LogoRasterNode, + "RecraftStyleV3InfiniteStyleLibrary": RecraftStyleInfiniteStyleLibrary, + "RecraftColorRGB": RecraftColorRGBNode, + "RecraftControls": RecraftControlsNode, +} + +# A dictionary that contains the friendly/humanly readable titles for the nodes +NODE_DISPLAY_NAME_MAPPINGS = { + "RecraftTextToImageNode": "Recraft Text to Image", + "RecraftImageToImageNode": "Recraft Image to Image", + "RecraftImageInpaintingNode": "Recraft Image Inpainting", + "RecraftTextToVectorNode": "Recraft Text to Vector", + "RecraftVectorizeImageNode": "Recraft Vectorize Image", + "RecraftRemoveBackgroundNode": "Recraft Remove Background", + "RecraftReplaceBackgroundNode": "Recraft Replace Background", + "RecraftCrispUpscaleNode": "Recraft Crisp Upscale Image", + "RecraftCreativeUpscaleNode": "Recraft Creative Upscale Image", + "RecraftStyleV3RealisticImage": "Recraft Style - Realistic Image", + "RecraftStyleV3DigitalIllustration": "Recraft Style - Digital Illustration", + "RecraftStyleV3LogoRaster": "Recraft Style - Logo Raster", + "RecraftStyleV3InfiniteStyleLibrary": "Recraft Style - Infinite Style Library", + "RecraftColorRGB": "Recraft Color RGB", + "RecraftControls": "Recraft Controls", +} diff --git a/comfy_api_nodes/nodes_rodin.py b/comfy_api_nodes/nodes_rodin.py new file mode 100644 index 000000000..67f90478c --- /dev/null +++ b/comfy_api_nodes/nodes_rodin.py @@ -0,0 +1,462 @@ +""" +ComfyUI X Rodin3D(Deemos) API Nodes + +Rodin API docs: https://developer.hyper3d.ai/ + +""" + +from __future__ import annotations +from inspect import cleandoc +from comfy.comfy_types.node_typing import IO +import folder_paths as comfy_paths +import requests +import os +import datetime +import shutil +import time +import io +import logging +import math +from PIL import Image +from comfy_api_nodes.apis.rodin_api import ( + Rodin3DGenerateRequest, + Rodin3DGenerateResponse, + Rodin3DCheckStatusRequest, + Rodin3DCheckStatusResponse, + Rodin3DDownloadRequest, + Rodin3DDownloadResponse, + JobStatus, +) +from comfy_api_nodes.apis.client import ( + ApiEndpoint, + HttpMethod, + SynchronousOperation, + PollingOperation, +) + + +COMMON_PARAMETERS = { + "Seed": ( + IO.INT, + { + "default":0, + "min":0, + "max":65535, + "display":"number" + } + ), + "Material_Type": ( + IO.COMBO, + { + "options": ["PBR", "Shaded"], + "default": "PBR" + } + ), + "Polygon_count": ( + IO.COMBO, + { + "options": ["4K-Quad", "8K-Quad", "18K-Quad", "50K-Quad", "200K-Triangle"], + "default": "18K-Quad" + } + ) +} + +def create_task_error(response: Rodin3DGenerateResponse): + """Check if the response has error""" + return hasattr(response, "error") + + + +class Rodin3DAPI: + """ + Generate 3D Assets using Rodin API + """ + RETURN_TYPES = (IO.STRING,) + RETURN_NAMES = ("3D Model Path",) + CATEGORY = "api node/3d/Rodin" + DESCRIPTION = cleandoc(__doc__ or "") + FUNCTION = "api_call" + API_NODE = True + + def tensor_to_filelike(self, tensor, max_pixels: int = 2048*2048): + """ + Converts a PyTorch tensor to a file-like object. + + Args: + - tensor (torch.Tensor): A tensor representing an image of shape (H, W, C) + where C is the number of channels (3 for RGB), H is height, and W is width. + + Returns: + - io.BytesIO: A file-like object containing the image data. + """ + array = tensor.cpu().numpy() + array = (array * 255).astype('uint8') + image = Image.fromarray(array, 'RGB') + + original_width, original_height = image.size + original_pixels = original_width * original_height + if original_pixels > max_pixels: + scale = math.sqrt(max_pixels / original_pixels) + new_width = int(original_width * scale) + new_height = int(original_height * scale) + else: + new_width, new_height = original_width, original_height + + if new_width != original_width or new_height != original_height: + image = image.resize((new_width, new_height), Image.Resampling.LANCZOS) + + img_byte_arr = io.BytesIO() + image.save(img_byte_arr, format='PNG') # PNG is used for lossless compression + img_byte_arr.seek(0) + return img_byte_arr + + def check_rodin_status(self, response: Rodin3DCheckStatusResponse) -> str: + has_failed = any(job.status == JobStatus.Failed for job in response.jobs) + all_done = all(job.status == JobStatus.Done for job in response.jobs) + status_list = [str(job.status) for job in response.jobs] + logging.info(f"[ Rodin3D API - CheckStatus ] Generate Status: {status_list}") + if has_failed: + logging.error(f"[ Rodin3D API - CheckStatus ] Generate Failed: {status_list}, Please try again.") + raise Exception("[ Rodin3D API ] Generate Failed, Please Try again.") + elif all_done: + return "DONE" + else: + return "Generating" + + def CreateGenerateTask(self, images=None, seed=1, material="PBR", quality="medium", tier="Regular", mesh_mode="Quad", **kwargs): + if images == None: + raise Exception("Rodin 3D generate requires at least 1 image.") + if len(images) >= 5: + raise Exception("Rodin 3D generate requires up to 5 image.") + + path = "/proxy/rodin/api/v2/rodin" + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path=path, + method=HttpMethod.POST, + request_model=Rodin3DGenerateRequest, + response_model=Rodin3DGenerateResponse, + ), + request=Rodin3DGenerateRequest( + seed=seed, + tier=tier, + material=material, + quality=quality, + mesh_mode=mesh_mode + ), + files=[ + ( + "images", + open(image, "rb") if isinstance(image, str) else self.tensor_to_filelike(image) + ) + for image in images if image is not None + ], + content_type = "multipart/form-data", + auth_kwargs=kwargs, + ) + + response = operation.execute() + + if create_task_error(response): + error_message = f"Rodin3D Create 3D generate Task Failed. Message: {response.message}, error: {response.error}" + logging.error(error_message) + raise Exception(error_message) + + logging.info("[ Rodin3D API - Submit Jobs ] Submit Generate Task Success!") + subscription_key = response.jobs.subscription_key + task_uuid = response.uuid + logging.info(f"[ Rodin3D API - Submit Jobs ] UUID: {task_uuid}") + return task_uuid, subscription_key + + def poll_for_task_status(self, subscription_key, **kwargs) -> Rodin3DCheckStatusResponse: + + path = "/proxy/rodin/api/v2/status" + + poll_operation = PollingOperation( + poll_endpoint=ApiEndpoint( + path = path, + method=HttpMethod.POST, + request_model=Rodin3DCheckStatusRequest, + response_model=Rodin3DCheckStatusResponse, + ), + request=Rodin3DCheckStatusRequest( + subscription_key = subscription_key + ), + completed_statuses=["DONE"], + failed_statuses=["FAILED"], + status_extractor=self.check_rodin_status, + poll_interval=3.0, + auth_kwargs=kwargs, + ) + + logging.info("[ Rodin3D API - CheckStatus ] Generate Start!") + + return poll_operation.execute() + + + + def GetRodinDownloadList(self, uuid, **kwargs) -> Rodin3DDownloadResponse: + logging.info("[ Rodin3D API - Downloading ] Generate Successfully!") + + path = "/proxy/rodin/api/v2/download" + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path=path, + method=HttpMethod.POST, + request_model=Rodin3DDownloadRequest, + response_model=Rodin3DDownloadResponse, + ), + request=Rodin3DDownloadRequest( + task_uuid=uuid + ), + auth_kwargs=kwargs + ) + + return operation.execute() + + def GetQualityAndMode(self, PolyCount): + if PolyCount == "200K-Triangle": + mesh_mode = "Raw" + quality = "medium" + else: + mesh_mode = "Quad" + if PolyCount == "4K-Quad": + quality = "extra-low" + elif PolyCount == "8K-Quad": + quality = "low" + elif PolyCount == "18K-Quad": + quality = "medium" + elif PolyCount == "50K-Quad": + quality = "high" + else: + quality = "medium" + + return mesh_mode, quality + + def DownLoadFiles(self, Url_List): + Save_path = os.path.join(comfy_paths.get_output_directory(), "Rodin3D", datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S")) + os.makedirs(Save_path, exist_ok=True) + model_file_path = None + for Item in Url_List.list: + url = Item.url + file_name = Item.name + file_path = os.path.join(Save_path, file_name) + if file_path.endswith(".glb"): + model_file_path = file_path + logging.info(f"[ Rodin3D API - download_files ] Downloading file: {file_path}") + max_retries = 5 + for attempt in range(max_retries): + try: + with requests.get(url, stream=True) as r: + r.raise_for_status() + with open(file_path, "wb") as f: + shutil.copyfileobj(r.raw, f) + break + except Exception as e: + logging.info(f"[ Rodin3D API - download_files ] Error downloading {file_path}:{e}") + if attempt < max_retries - 1: + logging.info("Retrying...") + time.sleep(2) + else: + logging.info(f"[ Rodin3D API - download_files ] Failed to download {file_path} after {max_retries} attempts.") + + return model_file_path + + +class Rodin3D_Regular(Rodin3DAPI): + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "Images": + ( + IO.IMAGE, + { + "forceInput":True, + } + ) + }, + "optional": { + **COMMON_PARAMETERS + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + }, + } + + def api_call( + self, + Images, + Seed, + Material_Type, + Polygon_count, + **kwargs + ): + tier = "Regular" + num_images = Images.shape[0] + m_images = [] + for i in range(num_images): + m_images.append(Images[i]) + mesh_mode, quality = self.GetQualityAndMode(Polygon_count) + task_uuid, subscription_key = self.CreateGenerateTask(images=m_images, seed=Seed, material=Material_Type, quality=quality, tier=tier, mesh_mode=mesh_mode, **kwargs) + self.poll_for_task_status(subscription_key, **kwargs) + Download_List = self.GetRodinDownloadList(task_uuid, **kwargs) + model = self.DownLoadFiles(Download_List) + + return (model,) + +class Rodin3D_Detail(Rodin3DAPI): + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "Images": + ( + IO.IMAGE, + { + "forceInput":True, + } + ) + }, + "optional": { + **COMMON_PARAMETERS + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + }, + } + + def api_call( + self, + Images, + Seed, + Material_Type, + Polygon_count, + **kwargs + ): + tier = "Detail" + num_images = Images.shape[0] + m_images = [] + for i in range(num_images): + m_images.append(Images[i]) + mesh_mode, quality = self.GetQualityAndMode(Polygon_count) + task_uuid, subscription_key = self.CreateGenerateTask(images=m_images, seed=Seed, material=Material_Type, quality=quality, tier=tier, mesh_mode=mesh_mode, **kwargs) + self.poll_for_task_status(subscription_key, **kwargs) + Download_List = self.GetRodinDownloadList(task_uuid, **kwargs) + model = self.DownLoadFiles(Download_List) + + return (model,) + +class Rodin3D_Smooth(Rodin3DAPI): + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "Images": + ( + IO.IMAGE, + { + "forceInput":True, + } + ) + }, + "optional": { + **COMMON_PARAMETERS + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + }, + } + + def api_call( + self, + Images, + Seed, + Material_Type, + Polygon_count, + **kwargs + ): + tier = "Smooth" + num_images = Images.shape[0] + m_images = [] + for i in range(num_images): + m_images.append(Images[i]) + mesh_mode, quality = self.GetQualityAndMode(Polygon_count) + task_uuid, subscription_key = self.CreateGenerateTask(images=m_images, seed=Seed, material=Material_Type, quality=quality, tier=tier, mesh_mode=mesh_mode, **kwargs) + self.poll_for_task_status(subscription_key, **kwargs) + Download_List = self.GetRodinDownloadList(task_uuid, **kwargs) + model = self.DownLoadFiles(Download_List) + + return (model,) + +class Rodin3D_Sketch(Rodin3DAPI): + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "Images": + ( + IO.IMAGE, + { + "forceInput":True, + } + ) + }, + "optional": { + "Seed": + ( + IO.INT, + { + "default":0, + "min":0, + "max":65535, + "display":"number" + } + ) + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + }, + } + + def api_call( + self, + Images, + Seed, + **kwargs + ): + tier = "Sketch" + num_images = Images.shape[0] + m_images = [] + for i in range(num_images): + m_images.append(Images[i]) + material_type = "PBR" + quality = "medium" + mesh_mode = "Quad" + task_uuid, subscription_key = self.CreateGenerateTask(images=m_images, seed=Seed, material=material_type, quality=quality, tier=tier, mesh_mode=mesh_mode, **kwargs) + self.poll_for_task_status(subscription_key, **kwargs) + Download_List = self.GetRodinDownloadList(task_uuid, **kwargs) + model = self.DownLoadFiles(Download_List) + + return (model,) + +# A dictionary that contains all nodes you want to export with their names +# NOTE: names should be globally unique +NODE_CLASS_MAPPINGS = { + "Rodin3D_Regular": Rodin3D_Regular, + "Rodin3D_Detail": Rodin3D_Detail, + "Rodin3D_Smooth": Rodin3D_Smooth, + "Rodin3D_Sketch": Rodin3D_Sketch, +} + +# A dictionary that contains the friendly/humanly readable titles for the nodes +NODE_DISPLAY_NAME_MAPPINGS = { + "Rodin3D_Regular": "Rodin 3D Generate - Regular Generate", + "Rodin3D_Detail": "Rodin 3D Generate - Detail Generate", + "Rodin3D_Smooth": "Rodin 3D Generate - Smooth Generate", + "Rodin3D_Sketch": "Rodin 3D Generate - Sketch Generate", +} diff --git a/comfy_api_nodes/nodes_runway.py b/comfy_api_nodes/nodes_runway.py new file mode 100644 index 000000000..af4b321f9 --- /dev/null +++ b/comfy_api_nodes/nodes_runway.py @@ -0,0 +1,635 @@ +"""Runway API Nodes + +API Docs: + - https://docs.dev.runwayml.com/api/#tag/Task-management/paths/~1v1~1tasks~1%7Bid%7D/delete + +User Guides: + - https://help.runwayml.com/hc/en-us/sections/30265301423635-Gen-3-Alpha + - https://help.runwayml.com/hc/en-us/articles/37327109429011-Creating-with-Gen-4-Video + - https://help.runwayml.com/hc/en-us/articles/33927968552339-Creating-with-Act-One-on-Gen-3-Alpha-and-Turbo + - https://help.runwayml.com/hc/en-us/articles/34170748696595-Creating-with-Keyframes-on-Gen-3 + +""" + +from typing import Union, Optional, Any +from enum import Enum + +import torch + +from comfy_api_nodes.apis import ( + RunwayImageToVideoRequest, + RunwayImageToVideoResponse, + RunwayTaskStatusResponse as TaskStatusResponse, + RunwayTaskStatusEnum as TaskStatus, + RunwayModelEnum as Model, + RunwayDurationEnum as Duration, + RunwayAspectRatioEnum as AspectRatio, + RunwayPromptImageObject, + RunwayPromptImageDetailedObject, + RunwayTextToImageRequest, + RunwayTextToImageResponse, + Model4, + ReferenceImage, + RunwayTextToImageAspectRatioEnum, +) +from comfy_api_nodes.apis.client import ( + ApiEndpoint, + HttpMethod, + SynchronousOperation, + PollingOperation, + EmptyRequest, +) +from comfy_api_nodes.apinode_utils import ( + upload_images_to_comfyapi, + download_url_to_video_output, + image_tensor_pair_to_batch, + validate_string, + download_url_to_image_tensor, +) +from comfy_api_nodes.mapper_utils import model_field_to_node_input +from comfy_api.input_impl import VideoFromFile +from comfy.comfy_types.node_typing import IO, ComfyNodeABC + +PATH_IMAGE_TO_VIDEO = "/proxy/runway/image_to_video" +PATH_TEXT_TO_IMAGE = "/proxy/runway/text_to_image" +PATH_GET_TASK_STATUS = "/proxy/runway/tasks" + +AVERAGE_DURATION_I2V_SECONDS = 64 +AVERAGE_DURATION_FLF_SECONDS = 256 +AVERAGE_DURATION_T2I_SECONDS = 41 + + +class RunwayApiError(Exception): + """Base exception for Runway API errors.""" + + pass + + +class RunwayGen4TurboAspectRatio(str, Enum): + """Aspect ratios supported for Image to Video API when using gen4_turbo model.""" + + field_1280_720 = "1280:720" + field_720_1280 = "720:1280" + field_1104_832 = "1104:832" + field_832_1104 = "832:1104" + field_960_960 = "960:960" + field_1584_672 = "1584:672" + + +class RunwayGen3aAspectRatio(str, Enum): + """Aspect ratios supported for Image to Video API when using gen3a_turbo model.""" + + field_768_1280 = "768:1280" + field_1280_768 = "1280:768" + + +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 response.output and len(response.output) > 0: + return response.output[0] + return None + + +# TODO: replace with updated image validation utils (upstream) +def validate_input_image(image: torch.Tensor) -> bool: + """ + Validate the input image is within the size limits for the Runway API. + See: https://docs.dev.runwayml.com/assets/inputs/#common-error-reasons + """ + return image.shape[2] < 8000 and image.shape[1] < 8000 + + +def poll_until_finished( + auth_kwargs: dict[str, str], + api_endpoint: ApiEndpoint[Any, TaskStatusResponse], + estimated_duration: Optional[int] = None, + node_id: Optional[str] = None, +) -> TaskStatusResponse: + """Polls the Runway API endpoint until the task reaches a terminal state, then returns the response.""" + return PollingOperation( + poll_endpoint=api_endpoint, + completed_statuses=[ + TaskStatus.SUCCEEDED.value, + ], + failed_statuses=[ + TaskStatus.FAILED.value, + TaskStatus.CANCELLED.value, + ], + status_extractor=lambda response: (response.status.value), + auth_kwargs=auth_kwargs, + result_url_extractor=get_video_url_from_task_status, + estimated_duration=estimated_duration, + node_id=node_id, + progress_extractor=extract_progress_from_task_status, + ).execute() + + +def extract_progress_from_task_status( + response: TaskStatusResponse, +) -> Union[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]: + """Returns the image URL from the task status response if it exists.""" + if response.output and len(response.output) > 0: + return response.output[0] + return None + + +class RunwayVideoGenNode(ComfyNodeABC): + """Runway Video Node Base.""" + + RETURN_TYPES = ("VIDEO",) + FUNCTION = "api_call" + CATEGORY = "api node/video/Runway" + API_NODE = True + + def validate_task_created(self, response: RunwayImageToVideoResponse) -> bool: + """ + Validate the task creation response from the Runway API matches + expected format. + """ + if not bool(response.id): + raise RunwayApiError("Invalid initial response from Runway API.") + return True + + def validate_response(self, response: RunwayImageToVideoResponse) -> bool: + """ + Validate the successful task status response from the Runway API + matches expected format. + """ + if not response.output or len(response.output) == 0: + raise RunwayApiError( + "Runway task succeeded but no video data found in response." + ) + return True + + def get_response( + self, task_id: str, auth_kwargs: dict[str, str], node_id: Optional[str] = None + ) -> RunwayImageToVideoResponse: + """Poll the task status until it is finished then get the response.""" + return poll_until_finished( + auth_kwargs, + ApiEndpoint( + path=f"{PATH_GET_TASK_STATUS}/{task_id}", + method=HttpMethod.GET, + request_model=EmptyRequest, + response_model=TaskStatusResponse, + ), + estimated_duration=AVERAGE_DURATION_FLF_SECONDS, + node_id=node_id, + ) + + def generate_video( + self, + request: RunwayImageToVideoRequest, + auth_kwargs: dict[str, str], + node_id: Optional[str] = None, + ) -> tuple[VideoFromFile]: + initial_operation = SynchronousOperation( + endpoint=ApiEndpoint( + path=PATH_IMAGE_TO_VIDEO, + method=HttpMethod.POST, + request_model=RunwayImageToVideoRequest, + response_model=RunwayImageToVideoResponse, + ), + request=request, + auth_kwargs=auth_kwargs, + ) + + initial_response = initial_operation.execute() + self.validate_task_created(initial_response) + task_id = initial_response.id + + final_response = self.get_response(task_id, auth_kwargs, node_id) + self.validate_response(final_response) + + video_url = get_video_url_from_task_status(final_response) + return (download_url_to_video_output(video_url),) + + +class RunwayImageToVideoNodeGen3a(RunwayVideoGenNode): + """Runway Image to Video Node using Gen3a Turbo model.""" + + DESCRIPTION = "Generate a video from a single starting frame using Gen3a Turbo model. Before diving in, review these best practices to ensure that your input selections will set your generation up for success: https://help.runwayml.com/hc/en-us/articles/33927968552339-Creating-with-Act-One-on-Gen-3-Alpha-and-Turbo." + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "prompt": model_field_to_node_input( + IO.STRING, RunwayImageToVideoRequest, "promptText", multiline=True + ), + "start_frame": ( + IO.IMAGE, + {"tooltip": "Start frame to be used for the video"}, + ), + "duration": model_field_to_node_input( + IO.COMBO, RunwayImageToVideoRequest, "duration", enum_type=Duration + ), + "ratio": model_field_to_node_input( + IO.COMBO, + RunwayImageToVideoRequest, + "ratio", + enum_type=RunwayGen3aAspectRatio, + ), + "seed": model_field_to_node_input( + IO.INT, + RunwayImageToVideoRequest, + "seed", + control_after_generate=True, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + def api_call( + self, + prompt: str, + start_frame: torch.Tensor, + duration: str, + ratio: str, + seed: int, + unique_id: Optional[str] = None, + **kwargs, + ) -> tuple[VideoFromFile]: + # Validate inputs + validate_string(prompt, min_length=1) + validate_input_image(start_frame) + + # Upload image + download_urls = upload_images_to_comfyapi( + start_frame, + max_images=1, + mime_type="image/png", + auth_kwargs=kwargs, + ) + if len(download_urls) != 1: + raise RunwayApiError("Failed to upload one or more images to comfy api.") + + return self.generate_video( + RunwayImageToVideoRequest( + promptText=prompt, + seed=seed, + model=Model("gen3a_turbo"), + duration=Duration(duration), + ratio=AspectRatio(ratio), + promptImage=RunwayPromptImageObject( + root=[ + RunwayPromptImageDetailedObject( + uri=str(download_urls[0]), position="first" + ) + ] + ), + ), + auth_kwargs=kwargs, + node_id=unique_id, + ) + + +class RunwayImageToVideoNodeGen4(RunwayVideoGenNode): + """Runway Image to Video Node using Gen4 Turbo model.""" + + DESCRIPTION = "Generate a video from a single starting frame using Gen4 Turbo model. Before diving in, review these best practices to ensure that your input selections will set your generation up for success: https://help.runwayml.com/hc/en-us/articles/37327109429011-Creating-with-Gen-4-Video." + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "prompt": model_field_to_node_input( + IO.STRING, RunwayImageToVideoRequest, "promptText", multiline=True + ), + "start_frame": ( + IO.IMAGE, + {"tooltip": "Start frame to be used for the video"}, + ), + "duration": model_field_to_node_input( + IO.COMBO, RunwayImageToVideoRequest, "duration", enum_type=Duration + ), + "ratio": model_field_to_node_input( + IO.COMBO, + RunwayImageToVideoRequest, + "ratio", + enum_type=RunwayGen4TurboAspectRatio, + ), + "seed": model_field_to_node_input( + IO.INT, + RunwayImageToVideoRequest, + "seed", + control_after_generate=True, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + def api_call( + self, + prompt: str, + start_frame: torch.Tensor, + duration: str, + ratio: str, + seed: int, + unique_id: Optional[str] = None, + **kwargs, + ) -> tuple[VideoFromFile]: + # Validate inputs + validate_string(prompt, min_length=1) + validate_input_image(start_frame) + + # Upload image + download_urls = upload_images_to_comfyapi( + start_frame, + max_images=1, + mime_type="image/png", + auth_kwargs=kwargs, + ) + if len(download_urls) != 1: + raise RunwayApiError("Failed to upload one or more images to comfy api.") + + return self.generate_video( + RunwayImageToVideoRequest( + promptText=prompt, + seed=seed, + model=Model("gen4_turbo"), + duration=Duration(duration), + ratio=AspectRatio(ratio), + promptImage=RunwayPromptImageObject( + root=[ + RunwayPromptImageDetailedObject( + uri=str(download_urls[0]), position="first" + ) + ] + ), + ), + auth_kwargs=kwargs, + node_id=unique_id, + ) + + +class RunwayFirstLastFrameNode(RunwayVideoGenNode): + """Runway First-Last Frame Node.""" + + DESCRIPTION = "Upload first and last keyframes, draft a prompt, and generate a video. More complex transitions, such as cases where the Last frame is completely different from the First frame, may benefit from the longer 10s duration. This would give the generation more time to smoothly transition between the two inputs. Before diving in, review these best practices to ensure that your input selections will set your generation up for success: https://help.runwayml.com/hc/en-us/articles/34170748696595-Creating-with-Keyframes-on-Gen-3." + + def get_response( + self, task_id: str, auth_kwargs: dict[str, str], node_id: Optional[str] = None + ) -> RunwayImageToVideoResponse: + return poll_until_finished( + auth_kwargs, + ApiEndpoint( + path=f"{PATH_GET_TASK_STATUS}/{task_id}", + method=HttpMethod.GET, + request_model=EmptyRequest, + response_model=TaskStatusResponse, + ), + estimated_duration=AVERAGE_DURATION_FLF_SECONDS, + node_id=node_id, + ) + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "prompt": model_field_to_node_input( + IO.STRING, RunwayImageToVideoRequest, "promptText", multiline=True + ), + "start_frame": ( + IO.IMAGE, + {"tooltip": "Start frame to be used for the video"}, + ), + "end_frame": ( + IO.IMAGE, + { + "tooltip": "End frame to be used for the video. Supported for gen3a_turbo only." + }, + ), + "duration": model_field_to_node_input( + IO.COMBO, RunwayImageToVideoRequest, "duration", enum_type=Duration + ), + "ratio": model_field_to_node_input( + IO.COMBO, + RunwayImageToVideoRequest, + "ratio", + enum_type=RunwayGen3aAspectRatio, + ), + "seed": model_field_to_node_input( + IO.INT, + RunwayImageToVideoRequest, + "seed", + control_after_generate=True, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "unique_id": "UNIQUE_ID", + "comfy_api_key": "API_KEY_COMFY_ORG", + }, + } + + def api_call( + self, + prompt: str, + start_frame: torch.Tensor, + end_frame: torch.Tensor, + duration: str, + ratio: str, + seed: int, + unique_id: Optional[str] = None, + **kwargs, + ) -> tuple[VideoFromFile]: + # Validate inputs + validate_string(prompt, min_length=1) + validate_input_image(start_frame) + validate_input_image(end_frame) + + # Upload images + stacked_input_images = image_tensor_pair_to_batch(start_frame, end_frame) + download_urls = upload_images_to_comfyapi( + stacked_input_images, + max_images=2, + mime_type="image/png", + auth_kwargs=kwargs, + ) + if len(download_urls) != 2: + raise RunwayApiError("Failed to upload one or more images to comfy api.") + + return self.generate_video( + RunwayImageToVideoRequest( + promptText=prompt, + seed=seed, + model=Model("gen3a_turbo"), + duration=Duration(duration), + ratio=AspectRatio(ratio), + promptImage=RunwayPromptImageObject( + root=[ + RunwayPromptImageDetailedObject( + uri=str(download_urls[0]), position="first" + ), + RunwayPromptImageDetailedObject( + uri=str(download_urls[1]), position="last" + ), + ] + ), + ), + auth_kwargs=kwargs, + node_id=unique_id, + ) + + +class RunwayTextToImageNode(ComfyNodeABC): + """Runway Text to Image Node.""" + + RETURN_TYPES = ("IMAGE",) + FUNCTION = "api_call" + CATEGORY = "api node/image/Runway" + API_NODE = True + DESCRIPTION = "Generate an image from a text prompt using Runway's Gen 4 model. You can also include reference images to guide the generation." + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "prompt": model_field_to_node_input( + IO.STRING, RunwayTextToImageRequest, "promptText", multiline=True + ), + "ratio": model_field_to_node_input( + IO.COMBO, + RunwayTextToImageRequest, + "ratio", + enum_type=RunwayTextToImageAspectRatioEnum, + ), + }, + "optional": { + "reference_image": ( + IO.IMAGE, + {"tooltip": "Optional reference image to guide the generation"}, + ) + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + def validate_task_created(self, response: RunwayTextToImageResponse) -> bool: + """ + Validate the task creation response from the Runway API matches + expected format. + """ + if not bool(response.id): + raise RunwayApiError("Invalid initial response from Runway API.") + return True + + def validate_response(self, response: TaskStatusResponse) -> bool: + """ + Validate the successful task status response from the Runway API + matches expected format. + """ + if not response.output or len(response.output) == 0: + raise RunwayApiError( + "Runway task succeeded but no image data found in response." + ) + return True + + def get_response( + self, task_id: str, auth_kwargs: dict[str, str], node_id: Optional[str] = None + ) -> TaskStatusResponse: + """Poll the task status until it is finished then get the response.""" + return poll_until_finished( + auth_kwargs, + ApiEndpoint( + path=f"{PATH_GET_TASK_STATUS}/{task_id}", + method=HttpMethod.GET, + request_model=EmptyRequest, + response_model=TaskStatusResponse, + ), + estimated_duration=AVERAGE_DURATION_T2I_SECONDS, + node_id=node_id, + ) + + def api_call( + self, + prompt: str, + ratio: str, + reference_image: Optional[torch.Tensor] = None, + unique_id: Optional[str] = None, + **kwargs, + ) -> tuple[torch.Tensor]: + # Validate inputs + validate_string(prompt, min_length=1) + + # Prepare reference images if provided + reference_images = None + if reference_image is not None: + validate_input_image(reference_image) + download_urls = upload_images_to_comfyapi( + reference_image, + max_images=1, + mime_type="image/png", + auth_kwargs=kwargs, + ) + if len(download_urls) != 1: + raise RunwayApiError("Failed to upload reference image to comfy api.") + + reference_images = [ReferenceImage(uri=str(download_urls[0]))] + + # Create request + request = RunwayTextToImageRequest( + promptText=prompt, + model=Model4.gen4_image, + ratio=ratio, + referenceImages=reference_images, + ) + + # Execute initial request + initial_operation = SynchronousOperation( + endpoint=ApiEndpoint( + path=PATH_TEXT_TO_IMAGE, + method=HttpMethod.POST, + request_model=RunwayTextToImageRequest, + response_model=RunwayTextToImageResponse, + ), + request=request, + auth_kwargs=kwargs, + ) + + initial_response = initial_operation.execute() + self.validate_task_created(initial_response) + task_id = initial_response.id + + # Poll for completion + final_response = self.get_response( + task_id, auth_kwargs=kwargs, node_id=unique_id + ) + self.validate_response(final_response) + + # Download and return image + image_url = get_image_url_from_task_status(final_response) + return (download_url_to_image_tensor(image_url),) + + +NODE_CLASS_MAPPINGS = { + "RunwayFirstLastFrameNode": RunwayFirstLastFrameNode, + "RunwayImageToVideoNodeGen3a": RunwayImageToVideoNodeGen3a, + "RunwayImageToVideoNodeGen4": RunwayImageToVideoNodeGen4, + "RunwayTextToImageNode": RunwayTextToImageNode, +} + +NODE_DISPLAY_NAME_MAPPINGS = { + "RunwayFirstLastFrameNode": "Runway First-Last-Frame to Video", + "RunwayImageToVideoNodeGen3a": "Runway Image to Video (Gen3a Turbo)", + "RunwayImageToVideoNodeGen4": "Runway Image to Video (Gen4 Turbo)", + "RunwayTextToImageNode": "Runway Text to Image", +} diff --git a/comfy_api_nodes/nodes_stability.py b/comfy_api_nodes/nodes_stability.py new file mode 100644 index 000000000..02e421678 --- /dev/null +++ b/comfy_api_nodes/nodes_stability.py @@ -0,0 +1,614 @@ +from inspect import cleandoc +from comfy.comfy_types.node_typing import IO +from comfy_api_nodes.apis.stability_api import ( + StabilityUpscaleConservativeRequest, + StabilityUpscaleCreativeRequest, + StabilityAsyncResponse, + StabilityResultsGetResponse, + StabilityStable3_5Request, + StabilityStableUltraRequest, + StabilityStableUltraResponse, + StabilityAspectRatio, + Stability_SD3_5_Model, + Stability_SD3_5_GenerationMode, + get_stability_style_presets, +) +from comfy_api_nodes.apis.client import ( + ApiEndpoint, + HttpMethod, + SynchronousOperation, + PollingOperation, + EmptyRequest, +) +from comfy_api_nodes.apinode_utils import ( + bytesio_to_image_tensor, + tensor_to_bytesio, + validate_string, +) + +import torch +import base64 +from io import BytesIO +from enum import Enum + + +class StabilityPollStatus(str, Enum): + finished = "finished" + in_progress = "in_progress" + failed = "failed" + + +def get_async_dummy_status(x: StabilityResultsGetResponse): + if x.name is not None or x.errors is not None: + return StabilityPollStatus.failed + elif x.finish_reason is not None: + return StabilityPollStatus.finished + return StabilityPollStatus.in_progress + + +class StabilityStableImageUltraNode: + """ + Generates images synchronously based on prompt and resolution. + """ + + RETURN_TYPES = (IO.IMAGE,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/image/Stability AI" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "What you wish to see in the output image. A strong, descriptive prompt that clearly defines" + + "What you wish to see in the output image. A strong, descriptive prompt that clearly defines" + + "elements, colors, and subjects will lead to better results. " + + "To control the weight of a given word use the format `(word:weight)`," + + "where `word` is the word you'd like to control the weight of and `weight`" + + "is a value between 0 and 1. For example: `The sky was a crisp (blue:0.3) and (green:0.8)`" + + "would convey a sky that was blue and green, but more green than blue." + }, + ), + "aspect_ratio": ([x.value for x in StabilityAspectRatio], + { + "default": StabilityAspectRatio.ratio_1_1, + "tooltip": "Aspect ratio of generated image.", + }, + ), + "style_preset": (get_stability_style_presets(), + { + "tooltip": "Optional desired style of generated image.", + }, + ), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 4294967294, + "control_after_generate": True, + "tooltip": "The random seed used for creating the noise.", + }, + ), + }, + "optional": { + "image": (IO.IMAGE,), + "negative_prompt": ( + IO.STRING, + { + "default": "", + "forceInput": True, + "tooltip": "A blurb of text describing what you do not wish to see in the output image. This is an advanced feature." + }, + ), + "image_denoise": ( + IO.FLOAT, + { + "default": 0.5, + "min": 0.0, + "max": 1.0, + "step": 0.01, + "tooltip": "Denoise of input image; 0.0 yields image identical to input, 1.0 is as if no image was provided at all.", + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + }, + } + + def api_call(self, prompt: str, aspect_ratio: str, style_preset: str, seed: int, + negative_prompt: str=None, image: torch.Tensor = None, image_denoise: float=None, + **kwargs): + validate_string(prompt, strip_whitespace=False) + # prepare image binary if image present + image_binary = None + if image is not None: + image_binary = tensor_to_bytesio(image, total_pixels=1504*1504).read() + else: + image_denoise = None + + if not negative_prompt: + negative_prompt = None + if style_preset == "None": + style_preset = None + + files = { + "image": image_binary + } + + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/stability/v2beta/stable-image/generate/ultra", + method=HttpMethod.POST, + request_model=StabilityStableUltraRequest, + response_model=StabilityStableUltraResponse, + ), + request=StabilityStableUltraRequest( + prompt=prompt, + negative_prompt=negative_prompt, + aspect_ratio=aspect_ratio, + seed=seed, + strength=image_denoise, + style_preset=style_preset, + ), + files=files, + content_type="multipart/form-data", + auth_kwargs=kwargs, + ) + response_api = operation.execute() + + if response_api.finish_reason != "SUCCESS": + raise Exception(f"Stable Image Ultra generation failed: {response_api.finish_reason}.") + + image_data = base64.b64decode(response_api.image) + returned_image = bytesio_to_image_tensor(BytesIO(image_data)) + + return (returned_image,) + + +class StabilityStableImageSD_3_5Node: + """ + Generates images synchronously based on prompt and resolution. + """ + + RETURN_TYPES = (IO.IMAGE,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/image/Stability AI" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "What you wish to see in the output image. A strong, descriptive prompt that clearly defines elements, colors, and subjects will lead to better results." + }, + ), + "model": ([x.value for x in Stability_SD3_5_Model],), + "aspect_ratio": ([x.value for x in StabilityAspectRatio], + { + "default": StabilityAspectRatio.ratio_1_1, + "tooltip": "Aspect ratio of generated image.", + }, + ), + "style_preset": (get_stability_style_presets(), + { + "tooltip": "Optional desired style of generated image.", + }, + ), + "cfg_scale": ( + IO.FLOAT, + { + "default": 4.0, + "min": 1.0, + "max": 10.0, + "step": 0.1, + "tooltip": "How strictly the diffusion process adheres to the prompt text (higher values keep your image closer to your prompt)", + }, + ), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 4294967294, + "control_after_generate": True, + "tooltip": "The random seed used for creating the noise.", + }, + ), + }, + "optional": { + "image": (IO.IMAGE,), + "negative_prompt": ( + IO.STRING, + { + "default": "", + "forceInput": True, + "tooltip": "Keywords of what you do not wish to see in the output image. This is an advanced feature." + }, + ), + "image_denoise": ( + IO.FLOAT, + { + "default": 0.5, + "min": 0.0, + "max": 1.0, + "step": 0.01, + "tooltip": "Denoise of input image; 0.0 yields image identical to input, 1.0 is as if no image was provided at all.", + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + }, + } + + def api_call(self, model: str, prompt: str, aspect_ratio: str, style_preset: str, seed: int, cfg_scale: float, + negative_prompt: str=None, image: torch.Tensor = None, image_denoise: float=None, + **kwargs): + validate_string(prompt, strip_whitespace=False) + # prepare image binary if image present + image_binary = None + mode = Stability_SD3_5_GenerationMode.text_to_image + if image is not None: + image_binary = tensor_to_bytesio(image, total_pixels=1504*1504).read() + mode = Stability_SD3_5_GenerationMode.image_to_image + aspect_ratio = None + else: + image_denoise = None + + if not negative_prompt: + negative_prompt = None + if style_preset == "None": + style_preset = None + + files = { + "image": image_binary + } + + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/stability/v2beta/stable-image/generate/sd3", + method=HttpMethod.POST, + request_model=StabilityStable3_5Request, + response_model=StabilityStableUltraResponse, + ), + request=StabilityStable3_5Request( + prompt=prompt, + negative_prompt=negative_prompt, + aspect_ratio=aspect_ratio, + seed=seed, + strength=image_denoise, + style_preset=style_preset, + cfg_scale=cfg_scale, + model=model, + mode=mode, + ), + files=files, + content_type="multipart/form-data", + auth_kwargs=kwargs, + ) + response_api = operation.execute() + + if response_api.finish_reason != "SUCCESS": + raise Exception(f"Stable Diffusion 3.5 Image generation failed: {response_api.finish_reason}.") + + image_data = base64.b64decode(response_api.image) + returned_image = bytesio_to_image_tensor(BytesIO(image_data)) + + return (returned_image,) + + +class StabilityUpscaleConservativeNode: + """ + Upscale image with minimal alterations to 4K resolution. + """ + + RETURN_TYPES = (IO.IMAGE,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/image/Stability AI" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "image": (IO.IMAGE,), + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "What you wish to see in the output image. A strong, descriptive prompt that clearly defines elements, colors, and subjects will lead to better results." + }, + ), + "creativity": ( + IO.FLOAT, + { + "default": 0.35, + "min": 0.2, + "max": 0.5, + "step": 0.01, + "tooltip": "Controls the likelihood of creating additional details not heavily conditioned by the init image.", + }, + ), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 4294967294, + "control_after_generate": True, + "tooltip": "The random seed used for creating the noise.", + }, + ), + }, + "optional": { + "negative_prompt": ( + IO.STRING, + { + "default": "", + "forceInput": True, + "tooltip": "Keywords of what you do not wish to see in the output image. This is an advanced feature." + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + }, + } + + def api_call(self, image: torch.Tensor, prompt: str, creativity: float, seed: int, negative_prompt: str=None, + **kwargs): + validate_string(prompt, strip_whitespace=False) + image_binary = tensor_to_bytesio(image, total_pixels=1024*1024).read() + + if not negative_prompt: + negative_prompt = None + + files = { + "image": image_binary + } + + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/stability/v2beta/stable-image/upscale/conservative", + method=HttpMethod.POST, + request_model=StabilityUpscaleConservativeRequest, + response_model=StabilityStableUltraResponse, + ), + request=StabilityUpscaleConservativeRequest( + prompt=prompt, + negative_prompt=negative_prompt, + creativity=round(creativity,2), + seed=seed, + ), + files=files, + content_type="multipart/form-data", + auth_kwargs=kwargs, + ) + response_api = operation.execute() + + if response_api.finish_reason != "SUCCESS": + raise Exception(f"Stability Upscale Conservative generation failed: {response_api.finish_reason}.") + + image_data = base64.b64decode(response_api.image) + returned_image = bytesio_to_image_tensor(BytesIO(image_data)) + + return (returned_image,) + + +class StabilityUpscaleCreativeNode: + """ + Upscale image with minimal alterations to 4K resolution. + """ + + RETURN_TYPES = (IO.IMAGE,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/image/Stability AI" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "image": (IO.IMAGE,), + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "What you wish to see in the output image. A strong, descriptive prompt that clearly defines elements, colors, and subjects will lead to better results." + }, + ), + "creativity": ( + IO.FLOAT, + { + "default": 0.3, + "min": 0.1, + "max": 0.5, + "step": 0.01, + "tooltip": "Controls the likelihood of creating additional details not heavily conditioned by the init image.", + }, + ), + "style_preset": (get_stability_style_presets(), + { + "tooltip": "Optional desired style of generated image.", + }, + ), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 4294967294, + "control_after_generate": True, + "tooltip": "The random seed used for creating the noise.", + }, + ), + }, + "optional": { + "negative_prompt": ( + IO.STRING, + { + "default": "", + "forceInput": True, + "tooltip": "Keywords of what you do not wish to see in the output image. This is an advanced feature." + }, + ), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + }, + } + + def api_call(self, image: torch.Tensor, prompt: str, creativity: float, style_preset: str, seed: int, negative_prompt: str=None, + **kwargs): + validate_string(prompt, strip_whitespace=False) + image_binary = tensor_to_bytesio(image, total_pixels=1024*1024).read() + + if not negative_prompt: + negative_prompt = None + if style_preset == "None": + style_preset = None + + files = { + "image": image_binary + } + + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/stability/v2beta/stable-image/upscale/creative", + method=HttpMethod.POST, + request_model=StabilityUpscaleCreativeRequest, + response_model=StabilityAsyncResponse, + ), + request=StabilityUpscaleCreativeRequest( + prompt=prompt, + negative_prompt=negative_prompt, + creativity=round(creativity,2), + style_preset=style_preset, + seed=seed, + ), + files=files, + content_type="multipart/form-data", + auth_kwargs=kwargs, + ) + response_api = operation.execute() + + operation = PollingOperation( + poll_endpoint=ApiEndpoint( + path=f"/proxy/stability/v2beta/results/{response_api.id}", + method=HttpMethod.GET, + request_model=EmptyRequest, + response_model=StabilityResultsGetResponse, + ), + poll_interval=3, + completed_statuses=[StabilityPollStatus.finished], + failed_statuses=[StabilityPollStatus.failed], + status_extractor=lambda x: get_async_dummy_status(x), + auth_kwargs=kwargs, + ) + response_poll: StabilityResultsGetResponse = operation.execute() + + if response_poll.finish_reason != "SUCCESS": + raise Exception(f"Stability Upscale Creative generation failed: {response_poll.finish_reason}.") + + image_data = base64.b64decode(response_poll.result) + returned_image = bytesio_to_image_tensor(BytesIO(image_data)) + + return (returned_image,) + + +class StabilityUpscaleFastNode: + """ + Quickly upscales an image via Stability API call to 4x its original size; intended for upscaling low-quality/compressed images. + """ + + RETURN_TYPES = (IO.IMAGE,) + DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value + FUNCTION = "api_call" + API_NODE = True + CATEGORY = "api node/image/Stability AI" + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "image": (IO.IMAGE,), + }, + "optional": { + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + }, + } + + def api_call(self, image: torch.Tensor, + **kwargs): + image_binary = tensor_to_bytesio(image, total_pixels=4096*4096).read() + + files = { + "image": image_binary + } + + operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/stability/v2beta/stable-image/upscale/fast", + method=HttpMethod.POST, + request_model=EmptyRequest, + response_model=StabilityStableUltraResponse, + ), + request=EmptyRequest(), + files=files, + content_type="multipart/form-data", + auth_kwargs=kwargs, + ) + response_api = operation.execute() + + if response_api.finish_reason != "SUCCESS": + raise Exception(f"Stability Upscale Fast failed: {response_api.finish_reason}.") + + image_data = base64.b64decode(response_api.image) + returned_image = bytesio_to_image_tensor(BytesIO(image_data)) + + return (returned_image,) + + +# A dictionary that contains all nodes you want to export with their names +# NOTE: names should be globally unique +NODE_CLASS_MAPPINGS = { + "StabilityStableImageUltraNode": StabilityStableImageUltraNode, + "StabilityStableImageSD_3_5Node": StabilityStableImageSD_3_5Node, + "StabilityUpscaleConservativeNode": StabilityUpscaleConservativeNode, + "StabilityUpscaleCreativeNode": StabilityUpscaleCreativeNode, + "StabilityUpscaleFastNode": StabilityUpscaleFastNode, +} + +# A dictionary that contains the friendly/humanly readable titles for the nodes +NODE_DISPLAY_NAME_MAPPINGS = { + "StabilityStableImageUltraNode": "Stability AI Stable Image Ultra", + "StabilityStableImageSD_3_5Node": "Stability AI Stable Diffusion 3.5 Image", + "StabilityUpscaleConservativeNode": "Stability AI Upscale Conservative", + "StabilityUpscaleCreativeNode": "Stability AI Upscale Creative", + "StabilityUpscaleFastNode": "Stability AI Upscale Fast", +} diff --git a/comfy_api_nodes/nodes_tripo.py b/comfy_api_nodes/nodes_tripo.py new file mode 100644 index 000000000..65f3b21f5 --- /dev/null +++ b/comfy_api_nodes/nodes_tripo.py @@ -0,0 +1,574 @@ +import os +from folder_paths import get_output_directory +from comfy_api_nodes.mapper_utils import model_field_to_node_input +from comfy.comfy_types.node_typing import IO +from comfy_api_nodes.apis import ( + TripoOrientation, + TripoModelVersion, +) +from comfy_api_nodes.apis.tripo_api import ( + TripoTaskType, + TripoStyle, + TripoFileReference, + TripoFileEmptyReference, + TripoUrlReference, + TripoTaskResponse, + TripoTaskStatus, + TripoTextToModelRequest, + TripoImageToModelRequest, + TripoMultiviewToModelRequest, + TripoTextureModelRequest, + TripoRefineModelRequest, + TripoAnimateRigRequest, + TripoAnimateRetargetRequest, + TripoConvertModelRequest, +) + +from comfy_api_nodes.apis.client import ( + ApiEndpoint, + HttpMethod, + SynchronousOperation, + PollingOperation, + EmptyRequest, +) +from comfy_api_nodes.apinode_utils import ( + upload_images_to_comfyapi, + download_url_to_bytesio, +) + + +def upload_image_to_tripo(image, **kwargs): + urls = upload_images_to_comfyapi(image, max_images=1, auth_kwargs=kwargs) + return TripoFileReference(TripoUrlReference(url=urls[0], type="jpeg")) + +def get_model_url_from_response(response: TripoTaskResponse) -> str: + if response.data is not None: + for key in ["pbr_model", "model", "base_model"]: + if getattr(response.data.output, key, None) is not None: + return getattr(response.data.output, key) + raise RuntimeError(f"Failed to get model url from response: {response}") + + +def poll_until_finished( + kwargs: dict[str, str], + response: TripoTaskResponse, +) -> tuple[str, str]: + """Polls the Tripo API endpoint until the task reaches a terminal state, then returns the response.""" + if response.code != 0: + raise RuntimeError(f"Failed to generate mesh: {response.error}") + task_id = response.data.task_id + response_poll = PollingOperation( + poll_endpoint=ApiEndpoint( + path=f"/proxy/tripo/v2/openapi/task/{task_id}", + method=HttpMethod.GET, + request_model=EmptyRequest, + response_model=TripoTaskResponse, + ), + completed_statuses=[TripoTaskStatus.SUCCESS], + failed_statuses=[ + TripoTaskStatus.FAILED, + TripoTaskStatus.CANCELLED, + TripoTaskStatus.UNKNOWN, + TripoTaskStatus.BANNED, + TripoTaskStatus.EXPIRED, + ], + status_extractor=lambda x: x.data.status, + auth_kwargs=kwargs, + node_id=kwargs["unique_id"], + result_url_extractor=get_model_url_from_response, + progress_extractor=lambda x: x.data.progress, + ).execute() + if response_poll.data.status == TripoTaskStatus.SUCCESS: + url = get_model_url_from_response(response_poll) + bytesio = download_url_to_bytesio(url) + # Save the downloaded model file + model_file = f"tripo_model_{task_id}.glb" + with open(os.path.join(get_output_directory(), model_file), "wb") as f: + f.write(bytesio.getvalue()) + return model_file, task_id + raise RuntimeError(f"Failed to generate mesh: {response_poll}") + +class TripoTextToModelNode: + """ + Generates 3D models synchronously based on a text prompt using Tripo's API. + """ + AVERAGE_DURATION = 80 + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "prompt": ("STRING", {"multiline": True}), + }, + "optional": { + "negative_prompt": ("STRING", {"multiline": True}), + "model_version": model_field_to_node_input(IO.COMBO, TripoTextToModelRequest, "model_version", enum_type=TripoModelVersion), + "style": model_field_to_node_input(IO.COMBO, TripoTextToModelRequest, "style", enum_type=TripoStyle, default="None"), + "texture": ("BOOLEAN", {"default": True}), + "pbr": ("BOOLEAN", {"default": True}), + "image_seed": ("INT", {"default": 42}), + "model_seed": ("INT", {"default": 42}), + "texture_seed": ("INT", {"default": 42}), + "texture_quality": (["standard", "detailed"], {"default": "standard"}), + "face_limit": ("INT", {"min": -1, "max": 500000, "default": -1}), + "quad": ("BOOLEAN", {"default": False}) + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = ("STRING", "MODEL_TASK_ID",) + RETURN_NAMES = ("model_file", "model task_id") + FUNCTION = "generate_mesh" + CATEGORY = "api node/3d/Tripo" + API_NODE = True + OUTPUT_NODE = True + + def generate_mesh(self, prompt, negative_prompt=None, model_version=None, style=None, texture=None, pbr=None, image_seed=None, model_seed=None, texture_seed=None, texture_quality=None, face_limit=None, quad=None, **kwargs): + style_enum = None if style == "None" else style + if not prompt: + raise RuntimeError("Prompt is required") + response = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/tripo/v2/openapi/task", + method=HttpMethod.POST, + request_model=TripoTextToModelRequest, + response_model=TripoTaskResponse, + ), + request=TripoTextToModelRequest( + type=TripoTaskType.TEXT_TO_MODEL, + prompt=prompt, + negative_prompt=negative_prompt if negative_prompt else None, + model_version=model_version, + style=style_enum, + texture=texture, + pbr=pbr, + image_seed=image_seed, + model_seed=model_seed, + texture_seed=texture_seed, + texture_quality=texture_quality, + face_limit=face_limit, + auto_size=True, + quad=quad + ), + auth_kwargs=kwargs, + ).execute() + return poll_until_finished(kwargs, response) + +class TripoImageToModelNode: + """ + Generates 3D models synchronously based on a single image using Tripo's API. + """ + AVERAGE_DURATION = 80 + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "image": ("IMAGE",), + }, + "optional": { + "model_version": model_field_to_node_input(IO.COMBO, TripoImageToModelRequest, "model_version", enum_type=TripoModelVersion), + "style": model_field_to_node_input(IO.COMBO, TripoTextToModelRequest, "style", enum_type=TripoStyle, default="None"), + "texture": ("BOOLEAN", {"default": True}), + "pbr": ("BOOLEAN", {"default": True}), + "model_seed": ("INT", {"default": 42}), + "orientation": model_field_to_node_input(IO.COMBO, TripoImageToModelRequest, "orientation", enum_type=TripoOrientation), + "texture_seed": ("INT", {"default": 42}), + "texture_quality": (["standard", "detailed"], {"default": "standard"}), + "texture_alignment": (["original_image", "geometry"], {"default": "original_image"}), + "face_limit": ("INT", {"min": -1, "max": 500000, "default": -1}), + "quad": ("BOOLEAN", {"default": False}) + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = ("STRING", "MODEL_TASK_ID",) + RETURN_NAMES = ("model_file", "model task_id") + FUNCTION = "generate_mesh" + CATEGORY = "api node/3d/Tripo" + API_NODE = True + OUTPUT_NODE = True + + def generate_mesh(self, image, model_version=None, style=None, texture=None, pbr=None, model_seed=None, orientation=None, texture_alignment=None, texture_seed=None, texture_quality=None, face_limit=None, quad=None, **kwargs): + style_enum = None if style == "None" else style + if image is None: + raise RuntimeError("Image is required") + tripo_file = upload_image_to_tripo(image, **kwargs) + response = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/tripo/v2/openapi/task", + method=HttpMethod.POST, + request_model=TripoImageToModelRequest, + response_model=TripoTaskResponse, + ), + request=TripoImageToModelRequest( + type=TripoTaskType.IMAGE_TO_MODEL, + file=tripo_file, + model_version=model_version, + style=style_enum, + texture=texture, + pbr=pbr, + model_seed=model_seed, + orientation=orientation, + texture_alignment=texture_alignment, + texture_seed=texture_seed, + texture_quality=texture_quality, + face_limit=face_limit, + auto_size=True, + quad=quad + ), + auth_kwargs=kwargs, + ).execute() + return poll_until_finished(kwargs, response) + +class TripoMultiviewToModelNode: + """ + Generates 3D models synchronously based on up to four images (front, left, back, right) using Tripo's API. + """ + AVERAGE_DURATION = 80 + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "image": ("IMAGE",), + }, + "optional": { + "image_left": ("IMAGE",), + "image_back": ("IMAGE",), + "image_right": ("IMAGE",), + "model_version": model_field_to_node_input(IO.COMBO, TripoMultiviewToModelRequest, "model_version", enum_type=TripoModelVersion), + "orientation": model_field_to_node_input(IO.COMBO, TripoImageToModelRequest, "orientation", enum_type=TripoOrientation), + "texture": ("BOOLEAN", {"default": True}), + "pbr": ("BOOLEAN", {"default": True}), + "model_seed": ("INT", {"default": 42}), + "texture_seed": ("INT", {"default": 42}), + "texture_quality": (["standard", "detailed"], {"default": "standard"}), + "texture_alignment": (["original_image", "geometry"], {"default": "original_image"}), + "face_limit": ("INT", {"min": -1, "max": 500000, "default": -1}), + "quad": ("BOOLEAN", {"default": False}) + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = ("STRING", "MODEL_TASK_ID",) + RETURN_NAMES = ("model_file", "model task_id") + FUNCTION = "generate_mesh" + CATEGORY = "api node/3d/Tripo" + API_NODE = True + OUTPUT_NODE = True + + def generate_mesh(self, image, image_left=None, image_back=None, image_right=None, model_version=None, orientation=None, texture=None, pbr=None, model_seed=None, texture_seed=None, texture_quality=None, texture_alignment=None, face_limit=None, quad=None, **kwargs): + if image is None: + raise RuntimeError("front image for multiview is required") + images = [] + image_dict = { + "image": image, + "image_left": image_left, + "image_back": image_back, + "image_right": image_right + } + if image_left is None and image_back is None and image_right is None: + raise RuntimeError("At least one of left, back, or right image must be provided for multiview") + for image_name in ["image", "image_left", "image_back", "image_right"]: + image_ = image_dict[image_name] + if image_ is not None: + tripo_file = upload_image_to_tripo(image_, **kwargs) + images.append(tripo_file) + else: + images.append(TripoFileEmptyReference()) + response = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/tripo/v2/openapi/task", + method=HttpMethod.POST, + request_model=TripoMultiviewToModelRequest, + response_model=TripoTaskResponse, + ), + request=TripoMultiviewToModelRequest( + type=TripoTaskType.MULTIVIEW_TO_MODEL, + files=images, + model_version=model_version, + orientation=orientation, + texture=texture, + pbr=pbr, + model_seed=model_seed, + texture_seed=texture_seed, + texture_quality=texture_quality, + texture_alignment=texture_alignment, + face_limit=face_limit, + quad=quad, + ), + auth_kwargs=kwargs, + ).execute() + return poll_until_finished(kwargs, response) + +class TripoTextureNode: + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "model_task_id": ("MODEL_TASK_ID",), + }, + "optional": { + "texture": ("BOOLEAN", {"default": True}), + "pbr": ("BOOLEAN", {"default": True}), + "texture_seed": ("INT", {"default": 42}), + "texture_quality": (["standard", "detailed"], {"default": "standard"}), + "texture_alignment": (["original_image", "geometry"], {"default": "original_image"}), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = ("STRING", "MODEL_TASK_ID",) + RETURN_NAMES = ("model_file", "model task_id") + FUNCTION = "generate_mesh" + CATEGORY = "api node/3d/Tripo" + API_NODE = True + OUTPUT_NODE = True + AVERAGE_DURATION = 80 + + def generate_mesh(self, model_task_id, texture=None, pbr=None, texture_seed=None, texture_quality=None, texture_alignment=None, **kwargs): + response = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/tripo/v2/openapi/task", + method=HttpMethod.POST, + request_model=TripoTextureModelRequest, + response_model=TripoTaskResponse, + ), + request=TripoTextureModelRequest( + original_model_task_id=model_task_id, + texture=texture, + pbr=pbr, + texture_seed=texture_seed, + texture_quality=texture_quality, + texture_alignment=texture_alignment + ), + auth_kwargs=kwargs, + ).execute() + return poll_until_finished(kwargs, response) + + +class TripoRefineNode: + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "model_task_id": ("MODEL_TASK_ID", { + "tooltip": "Must be a v1.4 Tripo model" + }), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + DESCRIPTION = "Refine a draft model created by v1.4 Tripo models only." + + RETURN_TYPES = ("STRING", "MODEL_TASK_ID",) + RETURN_NAMES = ("model_file", "model task_id") + FUNCTION = "generate_mesh" + CATEGORY = "api node/3d/Tripo" + API_NODE = True + OUTPUT_NODE = True + AVERAGE_DURATION = 240 + + def generate_mesh(self, model_task_id, **kwargs): + response = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/tripo/v2/openapi/task", + method=HttpMethod.POST, + request_model=TripoRefineModelRequest, + response_model=TripoTaskResponse, + ), + request=TripoRefineModelRequest( + draft_model_task_id=model_task_id + ), + auth_kwargs=kwargs, + ).execute() + return poll_until_finished(kwargs, response) + + +class TripoRigNode: + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "original_model_task_id": ("MODEL_TASK_ID",), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = ("STRING", "RIG_TASK_ID") + RETURN_NAMES = ("model_file", "rig task_id") + FUNCTION = "generate_mesh" + CATEGORY = "api node/3d/Tripo" + API_NODE = True + OUTPUT_NODE = True + AVERAGE_DURATION = 180 + + def generate_mesh(self, original_model_task_id, **kwargs): + response = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/tripo/v2/openapi/task", + method=HttpMethod.POST, + request_model=TripoAnimateRigRequest, + response_model=TripoTaskResponse, + ), + request=TripoAnimateRigRequest( + original_model_task_id=original_model_task_id, + out_format="glb", + spec="tripo" + ), + auth_kwargs=kwargs, + ).execute() + return poll_until_finished(kwargs, response) + +class TripoRetargetNode: + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "original_model_task_id": ("RIG_TASK_ID",), + "animation": ([ + "preset:idle", + "preset:walk", + "preset:climb", + "preset:jump", + "preset:slash", + "preset:shoot", + "preset:hurt", + "preset:fall", + "preset:turn", + ],), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = ("STRING", "RETARGET_TASK_ID") + RETURN_NAMES = ("model_file", "retarget task_id") + FUNCTION = "generate_mesh" + CATEGORY = "api node/3d/Tripo" + API_NODE = True + OUTPUT_NODE = True + AVERAGE_DURATION = 30 + + def generate_mesh(self, animation, original_model_task_id, **kwargs): + response = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/tripo/v2/openapi/task", + method=HttpMethod.POST, + request_model=TripoAnimateRetargetRequest, + response_model=TripoTaskResponse, + ), + request=TripoAnimateRetargetRequest( + original_model_task_id=original_model_task_id, + animation=animation, + out_format="glb", + bake_animation=True + ), + auth_kwargs=kwargs, + ).execute() + return poll_until_finished(kwargs, response) + +class TripoConversionNode: + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "original_model_task_id": ("MODEL_TASK_ID,RIG_TASK_ID,RETARGET_TASK_ID",), + "format": (["GLTF", "USDZ", "FBX", "OBJ", "STL", "3MF"],), + }, + "optional": { + "quad": ("BOOLEAN", {"default": False}), + "face_limit": ("INT", {"min": -1, "max": 500000, "default": -1}), + "texture_size": ("INT", {"min": 128, "max": 4096, "default": 4096}), + "texture_format": (["BMP", "DPX", "HDR", "JPEG", "OPEN_EXR", "PNG", "TARGA", "TIFF", "WEBP"], {"default": "JPEG"}) + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + @classmethod + def VALIDATE_INPUTS(cls, input_types): + # The min and max of input1 and input2 are still validated because + # we didn't take `input1` or `input2` as arguments + if input_types["original_model_task_id"] not in ("MODEL_TASK_ID", "RIG_TASK_ID", "RETARGET_TASK_ID"): + return "original_model_task_id must be MODEL_TASK_ID, RIG_TASK_ID or RETARGET_TASK_ID type" + return True + + RETURN_TYPES = () + FUNCTION = "generate_mesh" + CATEGORY = "api node/3d/Tripo" + API_NODE = True + OUTPUT_NODE = True + AVERAGE_DURATION = 30 + + def generate_mesh(self, original_model_task_id, format, quad, face_limit, texture_size, texture_format, **kwargs): + if not original_model_task_id: + raise RuntimeError("original_model_task_id is required") + response = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/tripo/v2/openapi/task", + method=HttpMethod.POST, + request_model=TripoConvertModelRequest, + response_model=TripoTaskResponse, + ), + request=TripoConvertModelRequest( + original_model_task_id=original_model_task_id, + format=format, + quad=quad if quad else None, + face_limit=face_limit if face_limit != -1 else None, + texture_size=texture_size if texture_size != 4096 else None, + texture_format=texture_format if texture_format != "JPEG" else None + ), + auth_kwargs=kwargs, + ).execute() + return poll_until_finished(kwargs, response) + +NODE_CLASS_MAPPINGS = { + "TripoTextToModelNode": TripoTextToModelNode, + "TripoImageToModelNode": TripoImageToModelNode, + "TripoMultiviewToModelNode": TripoMultiviewToModelNode, + "TripoTextureNode": TripoTextureNode, + "TripoRefineNode": TripoRefineNode, + "TripoRigNode": TripoRigNode, + "TripoRetargetNode": TripoRetargetNode, + "TripoConversionNode": TripoConversionNode, +} + +NODE_DISPLAY_NAME_MAPPINGS = { + "TripoTextToModelNode": "Tripo: Text to Model", + "TripoImageToModelNode": "Tripo: Image to Model", + "TripoMultiviewToModelNode": "Tripo: Multiview to Model", + "TripoTextureNode": "Tripo: Texture model", + "TripoRefineNode": "Tripo: Refine Draft model", + "TripoRigNode": "Tripo: Rig model", + "TripoRetargetNode": "Tripo: Retarget rigged model", + "TripoConversionNode": "Tripo: Convert model", +} diff --git a/comfy_api_nodes/nodes_veo2.py b/comfy_api_nodes/nodes_veo2.py new file mode 100644 index 000000000..df846d5dd --- /dev/null +++ b/comfy_api_nodes/nodes_veo2.py @@ -0,0 +1,308 @@ +import io +import logging +import base64 +import requests +import torch +from typing import Optional + +from comfy.comfy_types.node_typing import IO, ComfyNodeABC +from comfy_api.input_impl.video_types import VideoFromFile +from comfy_api_nodes.apis import ( + Veo2GenVidRequest, + Veo2GenVidResponse, + Veo2GenVidPollRequest, + Veo2GenVidPollResponse +) +from comfy_api_nodes.apis.client import ( + ApiEndpoint, + HttpMethod, + SynchronousOperation, + PollingOperation, +) + +from comfy_api_nodes.apinode_utils import ( + downscale_image_tensor, + tensor_to_base64_string +) + +AVERAGE_DURATION_VIDEO_GEN = 32 + +def convert_image_to_base64(image: torch.Tensor): + if image is None: + return None + + scaled_image = downscale_image_tensor(image, total_pixels=2048*2048) + return tensor_to_base64_string(scaled_image) + + +def get_video_url_from_response(poll_response: Veo2GenVidPollResponse) -> Optional[str]: + if ( + poll_response.response + and hasattr(poll_response.response, "videos") + and poll_response.response.videos + and len(poll_response.response.videos) > 0 + ): + video = poll_response.response.videos[0] + else: + return None + if hasattr(video, "gcsUri") and video.gcsUri: + return str(video.gcsUri) + return None + + +class VeoVideoGenerationNode(ComfyNodeABC): + """ + Generates videos from text prompts using Google's Veo API. + + This node can create videos from text descriptions and optional image inputs, + with control over parameters like aspect ratio, duration, and more. + """ + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Text description of the video", + }, + ), + "aspect_ratio": ( + IO.COMBO, + { + "options": ["16:9", "9:16"], + "default": "16:9", + "tooltip": "Aspect ratio of the output video", + }, + ), + }, + "optional": { + "negative_prompt": ( + IO.STRING, + { + "multiline": True, + "default": "", + "tooltip": "Negative text prompt to guide what to avoid in the video", + }, + ), + "duration_seconds": ( + IO.INT, + { + "default": 5, + "min": 5, + "max": 8, + "step": 1, + "display": "number", + "tooltip": "Duration of the output video in seconds", + }, + ), + "enhance_prompt": ( + IO.BOOLEAN, + { + "default": True, + "tooltip": "Whether to enhance the prompt with AI assistance", + } + ), + "person_generation": ( + IO.COMBO, + { + "options": ["ALLOW", "BLOCK"], + "default": "ALLOW", + "tooltip": "Whether to allow generating people in the video", + }, + ), + "seed": ( + IO.INT, + { + "default": 0, + "min": 0, + "max": 0xFFFFFFFF, + "step": 1, + "display": "number", + "control_after_generate": True, + "tooltip": "Seed for video generation (0 for random)", + }, + ), + "image": (IO.IMAGE, { + "default": None, + "tooltip": "Optional reference image to guide video generation", + }), + }, + "hidden": { + "auth_token": "AUTH_TOKEN_COMFY_ORG", + "comfy_api_key": "API_KEY_COMFY_ORG", + "unique_id": "UNIQUE_ID", + }, + } + + RETURN_TYPES = (IO.VIDEO,) + FUNCTION = "generate_video" + CATEGORY = "api node/video/Veo" + DESCRIPTION = "Generates videos from text prompts using Google's Veo API" + API_NODE = True + + def generate_video( + self, + prompt, + aspect_ratio="16:9", + negative_prompt="", + duration_seconds=5, + enhance_prompt=True, + person_generation="ALLOW", + seed=0, + image=None, + unique_id: Optional[str] = None, + **kwargs, + ): + # Prepare the instances for the request + instances = [] + + instance = { + "prompt": prompt + } + + # Add image if provided + if image is not None: + image_base64 = convert_image_to_base64(image) + if image_base64: + instance["image"] = { + "bytesBase64Encoded": image_base64, + "mimeType": "image/png" + } + + instances.append(instance) + + # Create parameters dictionary + parameters = { + "aspectRatio": aspect_ratio, + "personGeneration": person_generation, + "durationSeconds": duration_seconds, + "enhancePrompt": enhance_prompt, + } + + # Add optional parameters if provided + if negative_prompt: + parameters["negativePrompt"] = negative_prompt + if seed > 0: + parameters["seed"] = seed + + # Initial request to start video generation + initial_operation = SynchronousOperation( + endpoint=ApiEndpoint( + path="/proxy/veo/generate", + method=HttpMethod.POST, + request_model=Veo2GenVidRequest, + response_model=Veo2GenVidResponse + ), + request=Veo2GenVidRequest( + instances=instances, + parameters=parameters + ), + auth_kwargs=kwargs, + ) + + initial_response = initial_operation.execute() + operation_name = initial_response.name + + logging.info(f"Veo generation started with operation name: {operation_name}") + + # Define status extractor function + def status_extractor(response): + # Only return "completed" if the operation is done, regardless of success or failure + # We'll check for errors after polling completes + return "completed" if response.done else "pending" + + # Define progress extractor function + def progress_extractor(response): + # Could be enhanced if the API provides progress information + return None + + # Define the polling operation + poll_operation = PollingOperation( + poll_endpoint=ApiEndpoint( + path="/proxy/veo/poll", + method=HttpMethod.POST, + request_model=Veo2GenVidPollRequest, + response_model=Veo2GenVidPollResponse + ), + completed_statuses=["completed"], + failed_statuses=[], # No failed statuses, we'll handle errors after polling + status_extractor=status_extractor, + progress_extractor=progress_extractor, + request=Veo2GenVidPollRequest( + operationName=operation_name + ), + auth_kwargs=kwargs, + poll_interval=5.0, + result_url_extractor=get_video_url_from_response, + node_id=unique_id, + estimated_duration=AVERAGE_DURATION_VIDEO_GEN, + ) + + # Execute the polling operation + poll_response = poll_operation.execute() + + # Now check for errors in the final response + # Check for error in poll response + if hasattr(poll_response, 'error') and poll_response.error: + error_message = f"Veo API error: {poll_response.error.message} (code: {poll_response.error.code})" + logging.error(error_message) + raise Exception(error_message) + + # Check for RAI filtered content + if (hasattr(poll_response.response, 'raiMediaFilteredCount') and + poll_response.response.raiMediaFilteredCount > 0): + + # Extract reason message if available + if (hasattr(poll_response.response, 'raiMediaFilteredReasons') and + poll_response.response.raiMediaFilteredReasons): + reason = poll_response.response.raiMediaFilteredReasons[0] + error_message = f"Content filtered by Google's Responsible AI practices: {reason} ({poll_response.response.raiMediaFilteredCount} videos filtered.)" + else: + error_message = f"Content filtered by Google's Responsible AI practices ({poll_response.response.raiMediaFilteredCount} videos filtered.)" + + logging.error(error_message) + raise Exception(error_message) + + # Extract video data + video_data = None + if poll_response.response and hasattr(poll_response.response, 'videos') and poll_response.response.videos and len(poll_response.response.videos) > 0: + video = poll_response.response.videos[0] + + # Check if video is provided as base64 or URL + if hasattr(video, 'bytesBase64Encoded') and video.bytesBase64Encoded: + # Decode base64 string to bytes + video_data = base64.b64decode(video.bytesBase64Encoded) + elif hasattr(video, 'gcsUri') and video.gcsUri: + # Download from URL + video_url = video.gcsUri + video_response = requests.get(video_url) + video_data = video_response.content + else: + raise Exception("Video returned but no data or URL was provided") + else: + raise Exception("Video generation completed but no video was returned") + + if not video_data: + raise Exception("No video data was returned") + + logging.info("Video generation completed successfully") + + # Convert video data to BytesIO object + video_io = io.BytesIO(video_data) + + # Return VideoFromFile object + return (VideoFromFile(video_io),) + + +# Register the node +NODE_CLASS_MAPPINGS = { + "VeoVideoGenerationNode": VeoVideoGenerationNode, +} + +NODE_DISPLAY_NAME_MAPPINGS = { + "VeoVideoGenerationNode": "Google Veo2 Video Generation", +} diff --git a/comfy_api_nodes/redocly-dev.yaml b/comfy_api_nodes/redocly-dev.yaml new file mode 100644 index 000000000..d9e3cab70 --- /dev/null +++ b/comfy_api_nodes/redocly-dev.yaml @@ -0,0 +1,10 @@ +# This file is used to filter the Comfy Org OpenAPI spec for schemas related to API Nodes. +# This is used for development purposes to generate stubs for unreleased API endpoints. +apis: + filter: + root: openapi.yaml + decorators: + filter-in: + property: tags + value: ['API Nodes'] + matchStrategy: all diff --git a/comfy_api_nodes/redocly.yaml b/comfy_api_nodes/redocly.yaml new file mode 100644 index 000000000..d102345b1 --- /dev/null +++ b/comfy_api_nodes/redocly.yaml @@ -0,0 +1,10 @@ +# This file is used to filter the Comfy Org OpenAPI spec for schemas related to API Nodes. + +apis: + filter: + root: openapi.yaml + decorators: + filter-in: + property: tags + value: ['API Nodes', 'Released'] + matchStrategy: all diff --git a/comfy_api_nodes/util/__init__.py b/comfy_api_nodes/util/__init__.py new file mode 100644 index 000000000..e69de29bb diff --git a/comfy_api_nodes/util/validation_utils.py b/comfy_api_nodes/util/validation_utils.py new file mode 100644 index 000000000..031b9fbd3 --- /dev/null +++ b/comfy_api_nodes/util/validation_utils.py @@ -0,0 +1,100 @@ +import logging +from typing import Optional + +import torch +from comfy_api.input.video_types import VideoInput + + +def get_image_dimensions(image: torch.Tensor) -> tuple[int, int]: + if len(image.shape) == 4: + return image.shape[1], image.shape[2] + elif len(image.shape) == 3: + return image.shape[0], image.shape[1] + else: + raise ValueError("Invalid image tensor shape.") + + +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, +): + height, width = get_image_dimensions(image) + + if min_width is not None and width < min_width: + raise ValueError(f"Image width must be at least {min_width}px, got {width}px") + if max_width is not None and width > max_width: + raise ValueError(f"Image width must be at most {max_width}px, got {width}px") + if min_height is not None and height < min_height: + raise ValueError( + f"Image height must be at least {min_height}px, got {height}px" + ) + if max_height is not None and height > max_height: + raise ValueError(f"Image height must be at most {max_height}px, got {height}px") + + +def validate_image_aspect_ratio( + image: torch.Tensor, + min_aspect_ratio: Optional[float] = None, + max_aspect_ratio: Optional[float] = None, +): + width, height = get_image_dimensions(image) + aspect_ratio = width / height + + if min_aspect_ratio is not None and aspect_ratio < min_aspect_ratio: + raise ValueError( + f"Image aspect ratio must be at least {min_aspect_ratio}, got {aspect_ratio}" + ) + if max_aspect_ratio is not None and aspect_ratio > max_aspect_ratio: + raise ValueError( + f"Image aspect ratio must be at most {max_aspect_ratio}, got {aspect_ratio}" + ) + + +def validate_video_dimensions( + video: VideoInput, + min_width: Optional[int] = None, + max_width: Optional[int] = None, + min_height: Optional[int] = None, + max_height: Optional[int] = None, +): + try: + width, height = video.get_dimensions() + except Exception as e: + logging.error("Error getting dimensions of video: %s", e) + return + + if min_width is not None and width < min_width: + raise ValueError(f"Video width must be at least {min_width}px, got {width}px") + if max_width is not None and width > max_width: + raise ValueError(f"Video width must be at most {max_width}px, got {width}px") + if min_height is not None and height < min_height: + raise ValueError( + f"Video height must be at least {min_height}px, got {height}px" + ) + if max_height is not None and height > max_height: + raise ValueError(f"Video height must be at most {max_height}px, got {height}px") + + +def validate_video_duration( + video: VideoInput, + min_duration: Optional[float] = None, + max_duration: Optional[float] = None, +): + try: + duration = video.get_duration() + except Exception as e: + logging.error("Error getting duration of video: %s", e) + return + + epsilon = 0.0001 + if min_duration is not None and min_duration - epsilon > duration: + raise ValueError( + f"Video duration must be at least {min_duration}s, got {duration}s" + ) + if max_duration is not None and duration > max_duration + epsilon: + raise ValueError( + f"Video duration must be at most {max_duration}s, got {duration}s" + ) diff --git a/comfy_config/config_parser.py b/comfy_config/config_parser.py new file mode 100644 index 000000000..a9cbd94dd --- /dev/null +++ b/comfy_config/config_parser.py @@ -0,0 +1,97 @@ +import os +from pathlib import Path +from typing import Optional + +from pydantic_settings import PydanticBaseSettingsSource, TomlConfigSettingsSource + +from comfy_config.types import ( + ComfyConfig, + ProjectConfig, + PyProjectConfig, + PyProjectSettings +) + +""" +Extract configuration from a custom node directory's pyproject.toml file or a Python file. + +This function reads and parses the pyproject.toml file in the specified directory +to extract project and ComfyUI-specific configuration information. If no +pyproject.toml file is found, it creates a minimal configuration using the +folder name as the project name. If a Python file is provided, it uses the +file name (without extension) as the project name. + +Args: + path (str): Path to the directory containing the pyproject.toml file, or + path to a .py file. If pyproject.toml doesn't exist in a directory, + the folder name will be used as the default project name. If a .py + file is provided, the filename (without .py extension) will be used + as the project name. + +Returns: + Optional[PyProjectConfig]: A PyProjectConfig object containing: + - project: Basic project information (name, version, dependencies, etc.) + - tool_comfy: ComfyUI-specific configuration (publisher_id, models, etc.) + Returns None if configuration extraction fails or if the provided file + is not a Python file. + +Notes: + - If pyproject.toml is missing in a directory, creates a default config with folder name + - If a .py file is provided, creates a default config with filename (without extension) + - Returns None for non-Python files + +Example: + >>> from comfy_config import config_parser + >>> # For directory + >>> custom_node_dir = os.path.dirname(os.path.realpath(__file__)) + >>> project_config = config_parser.extract_node_configuration(custom_node_dir) + >>> print(project_config.project.name) # "my_custom_node" or name from pyproject.toml + >>> + >>> # For single-file Python node file + >>> py_file_path = os.path.realpath(__file__) # "/path/to/my_node.py" + >>> project_config = config_parser.extract_node_configuration(py_file_path) + >>> print(project_config.project.name) # "my_node" +""" +def extract_node_configuration(path) -> Optional[PyProjectConfig]: + if os.path.isfile(path): + file_path = Path(path) + + if file_path.suffix.lower() != '.py': + return None + + project_name = file_path.stem + project = ProjectConfig(name=project_name) + comfy = ComfyConfig() + return PyProjectConfig(project=project, tool_comfy=comfy) + + folder_name = os.path.basename(path) + toml_path = Path(path) / "pyproject.toml" + + if not toml_path.exists(): + project = ProjectConfig(name=folder_name) + comfy = ComfyConfig() + return PyProjectConfig(project=project, tool_comfy=comfy) + + raw_settings = load_pyproject_settings(toml_path) + + project_data = raw_settings.project + + tool_data = raw_settings.tool + comfy_data = tool_data.get("comfy", {}) if tool_data else {} + + return PyProjectConfig(project=project_data, tool_comfy=comfy_data) + + +def load_pyproject_settings(toml_path: Path) -> PyProjectSettings: + class PyProjectLoader(PyProjectSettings): + @classmethod + def settings_customise_sources( + cls, + settings_cls, + init_settings: PydanticBaseSettingsSource, + env_settings: PydanticBaseSettingsSource, + dotenv_settings: PydanticBaseSettingsSource, + file_secret_settings: PydanticBaseSettingsSource, + ): + return (TomlConfigSettingsSource(settings_cls, toml_path),) + + return PyProjectLoader() diff --git a/comfy_config/types.py b/comfy_config/types.py new file mode 100644 index 000000000..611982083 --- /dev/null +++ b/comfy_config/types.py @@ -0,0 +1,80 @@ +from pydantic import BaseModel, Field +from pydantic_settings import BaseSettings, SettingsConfigDict +from typing import List, Optional + +# IMPORTANT: The type definitions specified in pyproject.toml for custom nodes +# must remain synchronized with the corresponding files in the https://github.com/Comfy-Org/comfy-cli/blob/main/comfy_cli/registry/types.py. +# Any changes to one must be reflected in the other to maintain consistency. + +class NodeVersion(BaseModel): + changelog: str + dependencies: List[str] + deprecated: bool + id: str + version: str + download_url: str + + +class Node(BaseModel): + id: str + name: str + description: str + author: Optional[str] = None + license: Optional[str] = None + icon: Optional[str] = None + repository: Optional[str] = None + tags: List[str] = Field(default_factory=list) + latest_version: Optional[NodeVersion] = None + + +class PublishNodeVersionResponse(BaseModel): + node_version: NodeVersion + signedUrl: str + + +class URLs(BaseModel): + homepage: str = Field(default="", alias="Homepage") + documentation: str = Field(default="", alias="Documentation") + repository: str = Field(default="", alias="Repository") + issues: str = Field(default="", alias="Issues") + + +class Model(BaseModel): + location: str + model_url: str + + +class ComfyConfig(BaseModel): + publisher_id: str = Field(default="", alias="PublisherId") + display_name: str = Field(default="", alias="DisplayName") + icon: str = Field(default="", alias="Icon") + models: List[Model] = Field(default_factory=list, alias="Models") + includes: List[str] = Field(default_factory=list) + + +class License(BaseModel): + file: str = "" + text: str = "" + + +class ProjectConfig(BaseModel): + name: str = "" + description: str = "" + version: str = "1.0.0" + requires_python: str = Field(default=">= 3.9", alias="requires-python") + dependencies: List[str] = Field(default_factory=list) + license: License = Field(default_factory=License) + urls: URLs = Field(default_factory=URLs) + + +class PyProjectConfig(BaseModel): + project: ProjectConfig = Field(default_factory=ProjectConfig) + tool_comfy: ComfyConfig = Field(default_factory=ComfyConfig) + + +class PyProjectSettings(BaseSettings): + project: dict = Field(default_factory=dict) + + tool: dict = Field(default_factory=dict) + + model_config = SettingsConfigDict() diff --git a/comfy_extras/nodes/nodes_audio.py b/comfy_extras/nodes/nodes_audio.py index 6f506ef0d..3b34557db 100644 --- a/comfy_extras/nodes/nodes_audio.py +++ b/comfy_extras/nodes/nodes_audio.py @@ -1,12 +1,11 @@ from __future__ import annotations +import av import hashlib import io import json import os import random -import struct - import torch import comfy.model_management @@ -107,60 +106,122 @@ class VAEDecodeAudio: return ({"waveform": audio, "sample_rate": 44100},) -def create_vorbis_comment_block(comment_dict, last_block): - vendor_string = b'ComfyUI' - vendor_length = len(vendor_string) - - comments = [] - for key, value in comment_dict.items(): - comment = f"{key}={value}".encode('utf-8') - comments.append(struct.pack('I', len(comment_data))[1:] + comment_data - - return comment_block +def save_audio(self, audio, filename_prefix="ComfyUI", format="flac", prompt=None, extra_pnginfo=None, quality="128k"): + try: + import torchaudio # pylint: disable=import-error + except ImportError as exc_info: + raise TorchAudioNotFoundError() -def insert_or_replace_vorbis_comment(flac_io, comment_dict): - if len(comment_dict) == 0: - return flac_io + 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] = [] - flac_io.seek(4) + # 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]) - blocks = [] - last_block = False + # Opus supported sample rates + OPUS_RATES = [8000, 12000, 16000, 24000, 48000] - while not last_block: - header = flac_io.read(4) - last_block = (header[0] & 0x80) != 0 - block_type = header[0] & 0x7F - block_length = struct.unpack('>I', b'\x00' + header[1:])[0] - block_data = flac_io.read(block_length) + 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) - if block_type == 4 or block_type == 1: - pass - else: - header = bytes([(header[0] & (~0x80))]) + header[1:] - blocks.append(header + block_data) + # Use original sample rate initially + sample_rate = audio["sample_rate"] - blocks.append(create_vorbis_comment_block(comment_dict, last_block=True)) + # 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 - new_flac_io = io.BytesIO() - new_flac_io.write(b'fLaC') - for block in blocks: - new_flac_io.write(block) + # Resample if necessary + if sample_rate != audio["sample_rate"]: + waveform = torchaudio.functional.resample(waveform, audio["sample_rate"], sample_rate) - new_flac_io.write(flac_io.read()) - return new_flac_io + # Create in-memory WAV buffer + wav_buffer = io.BytesIO() + torchaudio.save(wav_buffer, waveform, sample_rate, format="WAV") + wav_buffer.seek(0) # Rewind for reading + # Use PyAV to convert and add metadata + input_container = av.open(wav_buffer) + + # 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 + + # Set up the output stream with appropriate properties + input_container.streams.audio[0] + if format == "opus": + out_stream = output_container.add_stream("libopus", rate=sample_rate) + 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) + 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) + + + # Copy frames from input to output + for frame in input_container.decode(audio=0): + frame.pts = None # Let PyAV handle timestamps + output_container.mux(out_stream.encode(frame)) + + # Flush encoder + output_container.mux(out_stream.encode(None)) + + # Close containers + output_container.close() + input_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): @@ -170,57 +231,71 @@ class SaveAudio: @classmethod def INPUT_TYPES(s): - return {"required": {"audio": ("AUDIO",), - "filename_prefix": ("STRING", {"default": "audio/ComfyUI"})}, + return {"required": { "audio": ("AUDIO", ), + "filename_prefix": ("STRING", {"default": "audio/ComfyUI"}), + }, "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"}, } RETURN_TYPES = () - FUNCTION = "save_audio" + FUNCTION = "save_flac" OUTPUT_NODE = True CATEGORY = "audio" - def save_audio(self, audio, filename_prefix="ComfyUI", prompt=None, extra_pnginfo=None): - try: - import torchaudio # pylint: disable=import-error - except ImportError as exc_info: - raise TorchAudioNotFoundError() + 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) - 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] = [] +class SaveAudioMP3: + def __init__(self): + self.output_dir = folder_paths.get_output_directory() + self.type = "output" + self.prefix_append = "" - 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]) + @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"}, + } - 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}_.flac" + RETURN_TYPES = () + FUNCTION = "save_mp3" - buff = io.BytesIO() - torchaudio.save(buff, waveform, audio["sample_rate"], format="FLAC") + OUTPUT_NODE = True - buff = insert_or_replace_vorbis_comment(buff, metadata) + CATEGORY = "audio" - with open(os.path.join(full_output_folder, file), 'wb') as f: - f.write(buff.getbuffer()) + 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) - results.append({ - "filename": file, - "subfolder": subfolder, - "type": self.type - }) - counter += 1 +class SaveAudioOpus: + def __init__(self): + self.output_dir = folder_paths.get_output_directory() + self.type = "output" + self.prefix_append = "" - return {"ui": {"audio": results}} + @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"}, + } + RETURN_TYPES = () + FUNCTION = "save_opus" + + 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): @@ -279,7 +354,20 @@ NODE_CLASS_MAPPINGS = { "VAEEncodeAudio": VAEEncodeAudio, "VAEDecodeAudio": VAEDecodeAudio, "SaveAudio": SaveAudio, + "SaveAudioMP3": SaveAudioMP3, + "SaveAudioOpus": SaveAudioOpus, "LoadAudio": LoadAudio, "PreviewAudio": PreviewAudio, "ConditioningStableAudio": ConditioningStableAudio, } + +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)", +} diff --git a/comfy_extras/nodes/nodes_cond.py b/comfy_extras/nodes/nodes_cond.py index 4c3a1d5bf..58c16f621 100644 --- a/comfy_extras/nodes/nodes_cond.py +++ b/comfy_extras/nodes/nodes_cond.py @@ -20,6 +20,30 @@ class CLIPTextEncodeControlnet: c.append(n) return (c, ) +class T5TokenizerOptions: + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "clip": ("CLIP", ), + "min_padding": ("INT", {"default": 0, "min": 0, "max": 10000, "step": 1}), + "min_length": ("INT", {"default": 0, "min": 0, "max": 10000, "step": 1}), + } + } + + CATEGORY = "_for_testing/conditioning" + RETURN_TYPES = ("CLIP",) + FUNCTION = "set_options" + + def set_options(self, clip, min_padding, min_length): + clip = clip.clone() + for t5_type in ["t5xxl", "pile_t5xl", "t5base", "mt5xl", "umt5xxl"]: + clip.set_tokenizer_option("{}_min_padding".format(t5_type), min_padding) + clip.set_tokenizer_option("{}_min_length".format(t5_type), min_length) + + return (clip, ) + NODE_CLASS_MAPPINGS = { - "CLIPTextEncodeControlnet": CLIPTextEncodeControlnet + "CLIPTextEncodeControlnet": CLIPTextEncodeControlnet, + "T5TokenizerOptions": T5TokenizerOptions, } diff --git a/comfy_extras/nodes/nodes_custom_sampler.py b/comfy_extras/nodes/nodes_custom_sampler.py index 1d1552f04..44c98b6ba 100644 --- a/comfy_extras/nodes/nodes_custom_sampler.py +++ b/comfy_extras/nodes/nodes_custom_sampler.py @@ -1,3 +1,4 @@ +import math import comfy.sampler_names from comfy import samplers from comfy import model_management @@ -254,6 +255,55 @@ class SetFirstSigma: sigmas[0] = sigma return (sigmas, ) +class ExtendIntermediateSigmas: + @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" + + FUNCTION = "extend" + + def extend(self, sigmas: torch.Tensor, steps: int, start_at_sigma: float, end_at_sigma: float, spacing: str): + if start_at_sigma < 0: + start_at_sigma = float("inf") + + interpolator = { + 'linear': lambda x: x, + 'cosine': lambda x: torch.sin(x*math.pi/2), + 'sine': lambda x: 1 - torch.cos(x*math.pi/2) + }[spacing] + + # linear space for our interpolation function + x = torch.linspace(0, 1, steps + 1, device=sigmas.device)[1:-1] + computed_spacing = interpolator(x) + + extended_sigmas = [] + for i in range(len(sigmas) - 1): + sigma_current = sigmas[i] + sigma_next = sigmas[i+1] + + extended_sigmas.append(sigma_current) + + if end_at_sigma <= sigma_current <= start_at_sigma: + interpolated_steps = computed_spacing * (sigma_next - sigma_current) + sigma_current + extended_sigmas.extend(interpolated_steps.tolist()) + + # Add the last sigma value + if len(sigmas) > 0: + extended_sigmas.append(sigmas[-1]) + + extended_sigmas = torch.FloatTensor(extended_sigmas) + + return (extended_sigmas,) + class KSamplerSelect: @classmethod def INPUT_TYPES(s): @@ -734,6 +784,7 @@ NODE_CLASS_MAPPINGS = { "SplitSigmasDenoise": SplitSigmasDenoise, "FlipSigmas": FlipSigmas, "SetFirstSigma": SetFirstSigma, + "ExtendIntermediateSigmas": ExtendIntermediateSigmas, "CFGGuider": CFGGuider, "DualCFGGuider": DualCFGGuider, diff --git a/comfy_extras/nodes/nodes_hunyuan.py b/comfy_extras/nodes/nodes_hunyuan.py index 84377f1ad..c2bf9539c 100644 --- a/comfy_extras/nodes/nodes_hunyuan.py +++ b/comfy_extras/nodes/nodes_hunyuan.py @@ -78,7 +78,7 @@ class HunyuanImageToVideo: "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}), "length": ("INT", {"default": 53, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}), "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}), - "guidance_type": (["v1 (concat)", "v2 (replace)"], ) + "guidance_type": (["v1 (concat)", "v2 (replace)", "custom"], ) }, "optional": {"start_image": ("IMAGE", ), }} @@ -102,10 +102,12 @@ class HunyuanImageToVideo: if guidance_type == "v1 (concat)": cond = {"concat_latent_image": concat_latent_image, "concat_mask": mask} - else: + elif guidance_type == "v2 (replace)": cond = {'guiding_frame_index': 0} latent[:, :, :concat_latent_image.shape[2]] = concat_latent_image out_latent["noise_mask"] = mask + elif guidance_type == "custom": + cond = {"ref_latent": concat_latent_image} positive = node_helpers.conditioning_set_values(positive, cond) diff --git a/comfy_extras/nodes/nodes_images.py b/comfy_extras/nodes/nodes_images.py index 386cbe28b..ff3f7a901 100644 --- a/comfy_extras/nodes/nodes_images.py +++ b/comfy_extras/nodes/nodes_images.py @@ -1,22 +1,24 @@ import json -import os -from typing import Literal, Tuple - import numpy as np +import os +import re import torch from PIL import Image from PIL.PngImagePlugin import PngInfo +from inspect import cleandoc +from io import BytesIO +from typing import Literal, Tuple from comfy import utils from comfy.cli_args import args from comfy.cmd import folder_paths -from comfy.component_model.tensor_types import ImageBatch, RGBImageBatch, RGBAImageBatch +from comfy.comfy_types import IO +from comfy.component_model.tensor_types import ImageBatch, RGBImageBatch +from comfy.execution_context import current_execution_context from comfy.nodes.base_nodes import ImageScale from comfy.nodes.common import MAX_RESOLUTION from comfy.nodes.package_typing import CustomNode -from comfy_extras.constants.resolutions import SDXL_SD3_FLUX_RESOLUTIONS, LTVX_RESOLUTIONS, SD_RESOLUTIONS, \ - IDEOGRAM_RESOLUTIONS, COSMOS_RESOLUTIONS, HUNYUAN_VIDEO_RESOLUTIONS, WAN_VIDEO_14B_RESOLUTIONS, \ - WAN_VIDEO_1_3B_RESOLUTIONS, WAN_VIDEO_14B_EXTENDED_RESOLUTIONS, RESOLUTION_NAMES, RESOLUTION_MAP +from comfy_extras.constants.resolutions import RESOLUTION_MAP, RESOLUTION_NAMES, SD_RESOLUTIONS def levels_adjustment(image: ImageBatch, black_level: float = 0.0, mid_level: float = 0.5, white_level: float = 1.0, clip: bool = True) -> ImageBatch: @@ -44,14 +46,10 @@ def levels_adjustment(image: ImageBatch, black_level: float = 0.0, mid_level: fl return torch.where(x <= 0.0031308, x * 12.92, 1.055 * x ** (1 / 2.4) - 0.055) linear = srgb_to_linear(image) - adjusted = (linear - black_level) / (white_level - black_level) - power_factor = torch.log2(torch.tensor(0.5, device=image.device)) / torch.log2(torch.tensor(mid_level, device=image.device)) - # apply power function to avoid nans adjusted = torch.where(adjusted > 0, torch.pow(adjusted.clamp(min=1e-8), power_factor), adjusted) - result = linear_to_srgb(adjusted) if clip: @@ -72,7 +70,6 @@ class ImageCrop: RETURN_TYPES = ("IMAGE",) FUNCTION = "crop" - CATEGORY = "image/transform" def crop(self, image, width, height, x, y): @@ -93,7 +90,6 @@ class RepeatImageBatch: RETURN_TYPES = ("IMAGE",) FUNCTION = "repeat" - CATEGORY = "image/batch" def repeat(self, image, amount): @@ -111,7 +107,6 @@ class ImageFromBatch: RETURN_TYPES = ("IMAGE",) FUNCTION = "frombatch" - CATEGORY = "image/batch" def frombatch(self, image, batch_index, length): @@ -122,6 +117,24 @@ class ImageFromBatch: return (s,) +class ImageAddNoise: + @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 = "add_noise" + CATEGORY = "image" + + def add_noise(self, image, seed, strength): + 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,) + + class SaveAnimatedWEBP: def __init__(self): self.output_dir = folder_paths.get_output_directory() @@ -139,16 +152,13 @@ class SaveAnimatedWEBP: "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): @@ -209,9 +219,7 @@ class SaveAnimatedPNG: RETURN_TYPES = () FUNCTION = "save_images" - OUTPUT_NODE = True - CATEGORY = "image/animation" def save_images(self, images, fps, compress_level, filename_prefix="ComfyUI", prompt=None, extra_pnginfo=None): @@ -245,28 +253,50 @@ class SaveAnimatedPNG: class ImageShape: - def __init__(self): - pass - @classmethod def INPUT_TYPES(cls): - return { - "required": { - "image": ("IMAGE",), - } - } + return {"required": {"image": ("IMAGE",), }} RETURN_TYPES = ("INT", "INT") RETURN_NAMES = ("width", "height") - FUNCTION = "image_width_height" - + FUNCTION = "get_shape" CATEGORY = "image/operations" - def image_width_height(self, image: ImageBatch): + def get_shape(self, image: ImageBatch): shape = image.shape return shape[2], shape[1] +class GetImageSize: + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "image": (IO.IMAGE,), + }, + "hidden": { + "unique_id": "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): + height = image.shape[1] + width = image.shape[2] + batch_size = image.shape[0] + + if unique_id: + server = current_execution_context().server + server.send_progress_text(f"width: {width}, height: {height}\n batch size: {batch_size}", unique_id) + + return width, height, batch_size + + class ImageResize: @classmethod def INPUT_TYPES(cls): @@ -286,7 +316,7 @@ class ImageResize: FUNCTION = "resize_image" CATEGORY = "image/transform" - def resize_image(self, image: ImageBatch, resize_mode: Literal["cover", "contain", "auto"], resolutions: Literal["SDXL/SD3/Flux", "SD1.5"], interpolation: str, aspect_ratio_tolerance=0.05) -> tuple[RGBImageBatch]: + def resize_image(self, image: ImageBatch, resize_mode: Literal["cover", "contain", "auto"], resolutions: str, interpolation: str, aspect_ratio_tolerance=0.05) -> tuple[RGBImageBatch]: supported_resolutions = RESOLUTION_MAP.get(resolutions, SD_RESOLUTIONS) return self.resize_image_with_supported_resolutions(image, resize_mode, supported_resolutions, interpolation, aspect_ratio_tolerance=aspect_ratio_tolerance) @@ -295,12 +325,9 @@ class ImageResize: for img in image: h, w = img.shape[:2] current_aspect_ratio = w / h - aspect_ratio_diffs = [(abs(res[0] / res[1] - current_aspect_ratio), res) for res in supported_resolutions] min_diff = min(aspect_ratio_diffs, key=lambda x: x[0])[0] close_enough_resolutions = [res for diff, res in aspect_ratio_diffs if diff <= min_diff + aspect_ratio_tolerance] - - # pick the highest resolution from the filtered set target_resolution = max(close_enough_resolutions, key=lambda res: res[0] * res[1]) if resize_mode == "cover": @@ -315,31 +342,25 @@ class ImageResize: else: new_w, new_h = int(w * target_resolution[1] / h), target_resolution[1] - # convert to b, c, h, w img_tensor = img.permute(2, 0, 1).unsqueeze(0) resized = utils.common_upscale(img_tensor, new_w, new_h, interpolation, "disabled") - # handle padding or cropping if resize_mode == "contain": canvas = torch.zeros((1, 3, target_resolution[1], target_resolution[0]), device=resized.device, dtype=resized.dtype) - y1 = (target_resolution[1] - new_h) // 2 - x1 = (target_resolution[0] - new_w) // 2 + y1, x1 = (target_resolution[1] - new_h) // 2, (target_resolution[0] - new_w) // 2 canvas[:, :, y1:y1 + new_h, x1:x1 + new_w] = resized resized = canvas elif resize_mode == "cover": - y1 = (new_h - target_resolution[1]) // 2 - x1 = (new_w - target_resolution[0]) // 2 + y1, x1 = (new_h - target_resolution[1]) // 2, (new_w - target_resolution[0]) // 2 resized = resized[:, :, y1:y1 + target_resolution[1], x1:x1 + target_resolution[0]] else: # auto if new_w != target_resolution[0] or new_h != target_resolution[1]: canvas = torch.zeros((1, 3, target_resolution[1], target_resolution[0]), device=resized.device, dtype=resized.dtype) - y1 = (target_resolution[1] - new_h) // 2 - x1 = (target_resolution[0] - new_w) // 2 + y1, x1 = (target_resolution[1] - new_h) // 2, (target_resolution[0] - new_w) // 2 canvas[:, :, y1:y1 + new_h, x1:x1 + new_w] = resized resized = canvas resized_images.append(resized.squeeze(0).permute(1, 2, 0).clamp(0.0, 1.0)) - return (torch.stack(resized_images),) @@ -385,14 +406,10 @@ class ImageLevels(CustomNode): return (adjusted_image,) -class ImageLuminance: +class ImageLuminance(CustomNode): @classmethod def INPUT_TYPES(cls): - return { - "required": { - "image": ("IMAGE",), - } - } + return {"required": {"image": ("IMAGE",), }} RETURN_TYPES = ("IMAGE",) FUNCTION = "compute_luminance" @@ -400,16 +417,173 @@ class ImageLuminance: def compute_luminance(self, image: ImageBatch) -> Tuple[ImageBatch]: assert image.dim() == 4 and image.shape[-1] == 3, "Input should be of shape (B, H, W, 3)" - - # define srgb luminance coefficients coeffs = torch.tensor([0.2126, 0.7152, 0.0722], device=image.device, dtype=image.dtype) - luminance = torch.sum(image * coeffs, dim=-1, keepdim=True) luminance = luminance.expand(-1, -1, -1, 3) - return (luminance,) +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) + + +class ImageStitch: + """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",), }, + } + + RETURN_TYPES = ("IMAGE",) + FUNCTION = "stitch" + CATEGORY = "image/transform" + DESCRIPTION = "Stitches image2 to image1 in the specified direction." + + def stitch(self, image1, direction, match_image_size, spacing_width, spacing_color, image2=None): + if image2 is None: + return (image1,) + + if image1.shape[0] != image2.shape[0]: + max_batch = max(image1.shape[0], image2.shape[0]) + if image1.shape[0] < max_batch: + image1 = torch.cat([image1, image1[-1:].repeat(max_batch - image1.shape[0], 1, 1, 1)]) + if image2.shape[0] < max_batch: + image2 = torch.cat([image2, image2[-1:].repeat(max_batch - image2.shape[0], 1, 1, 1)]) + + if match_image_size: + h1, w1 = image1.shape[1:3] + h2, w2 = image2.shape[1:3] + aspect_ratio = w2 / h2 + if direction in ["left", "right"]: + target_h, target_w = h1, int(h1 * aspect_ratio) + else: + target_w, target_h = w1, int(w1 / aspect_ratio) + image2 = utils.common_upscale(image2.movedim(-1, 1), target_w, target_h, "lanczos", "disabled").movedim(1, -1) + else: + h1, w1 = image1.shape[1:3] + h2, w2 = image2.shape[1:3] + if direction in ["left", "right"]: + if h1 != h2: + target_h = max(h1, h2) + if h1 < target_h: + pad_h = target_h - h1 + pad_top, pad_bottom = pad_h // 2, pad_h - pad_h // 2 + image1 = torch.nn.functional.pad(image1, (0, 0, 0, 0, pad_top, pad_bottom), mode='constant', value=0.0) + if h2 < target_h: + pad_h = target_h - h2 + pad_top, pad_bottom = pad_h // 2, pad_h - pad_h // 2 + image2 = torch.nn.functional.pad(image2, (0, 0, 0, 0, pad_top, pad_bottom), mode='constant', value=0.0) + else: + if w1 != w2: + target_w = max(w1, w2) + if w1 < target_w: + pad_w = target_w - w1 + pad_left, pad_right = pad_w // 2, pad_w - pad_w // 2 + image1 = torch.nn.functional.pad(image1, (0, 0, pad_left, pad_right), mode='constant', value=0.0) + if w2 < target_w: + pad_w = target_w - w2 + pad_left, pad_right = pad_w // 2, pad_w - pad_w // 2 + image2 = torch.nn.functional.pad(image2, (0, 0, pad_left, pad_right), mode='constant', value=0.0) + + images_to_stitch = [image2, image1] if direction in ["left", "up"] else [image1, image2] + if spacing_width > 0: + color_map = {"white": 1.0, "black": 0.0, "red": (1.0, 0.0, 0.0), "green": (0.0, 1.0, 0.0), "blue": (0.0, 0.0, 1.0), } + color_val = color_map[spacing_color] + if direction in ["left", "right"]: + spacing_shape = (image1.shape[0], max(image1.shape[1], image2.shape[1]), spacing_width, image1.shape[-1],) + else: + spacing_shape = (image1.shape[0], spacing_width, max(image1.shape[2], image2.shape[2]), image1.shape[-1],) + spacing = torch.full(spacing_shape, 0.0, device=image1.device) + if isinstance(color_val, tuple): + for i, c in enumerate(color_val): + if i < spacing.shape[-1]: + spacing[..., i] = c + if spacing.shape[-1] == 4: + spacing[..., 3] = 1.0 + else: + spacing[..., : min(3, spacing.shape[-1])] = color_val + if spacing.shape[-1] == 4: + spacing[..., 3] = 1.0 + images_to_stitch.insert(1, spacing) + + concat_dim = 2 if direction in ["left", "right"] else 1 + return (torch.cat(images_to_stitch, dim=concat_dim),) + + +class SaveSVGNode: + """Save SVG files on disk.""" + + def __init__(self): + self.output_dir = folder_paths.get_output_directory() + self.type = "output" + self.prefix_append = "" + + RETURN_TYPES = () + DESCRIPTION = cleandoc(__doc__ or "") + FUNCTION = "save_svg" + CATEGORY = "image/save" + OUTPUT_NODE = True + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "svg": ("SVG",), + "filename_prefix": ("STRING", {"default": "svg/ComfyUI", "tooltip": "The prefix for the file to save."}) + }, + "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"} + } + + 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() + metadata_dict = {} + if prompt is not None: + metadata_dict["prompt"] = prompt + if extra_pnginfo is not None: + metadata_dict.update(extra_pnginfo) + 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" + svg_bytes.seek(0) + svg_content = svg_bytes.read().decode('utf-8') + + if metadata_json: + metadata_element = f' \n' + svg_content = re.sub(r'(]*>)', lambda m: m.group(1) + '\n' + metadata_element, svg_content, flags=re.UNICODE) + + 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}) + counter += 1 + return {"ui": {"images": results}} + + NODE_CLASS_MAPPINGS = { "ImageResize": ImageResize, "ImageResize1": ImageResize1, @@ -421,9 +595,17 @@ NODE_CLASS_MAPPINGS = { "ImageFromBatch": ImageFromBatch, "SaveAnimatedWEBP": SaveAnimatedWEBP, "SaveAnimatedPNG": SaveAnimatedPNG, + "ImageAddNoise": ImageAddNoise, + "SaveSVGNode": SaveSVGNode, + "ImageStitch": ImageStitch, + "GetImageSize": GetImageSize, } NODE_DISPLAY_NAME_MAPPINGS = { "ImageResize": "Fit Image to Diffusion Size", - "ImageResize1": "Fit Image to Width Height" + "ImageResize1": "Fit Image to Width Height", + "ImageShape": "Get Image Shape", + "GetImageSize": "Get Image Size", + "ImageStitch": "Stitch Images", + "SaveSVGNode": "Save SVG", } diff --git a/comfy_extras/nodes/nodes_load_3d.py b/comfy_extras/nodes/nodes_load_3d.py index 6ee46b4df..676ba07ac 100644 --- a/comfy_extras/nodes/nodes_load_3d.py +++ b/comfy_extras/nodes/nodes_load_3d.py @@ -2,6 +2,8 @@ import os from comfy.cmd import folder_paths from comfy.nodes import base_nodes as nodes +from comfy.comfy_types import IO +from comfy_api.input_impl import VideoFromFile def normalize_path(path): @@ -15,7 +17,7 @@ class Load3D(): os.makedirs(input_dir, exist_ok=True) - files = [normalize_path(os.path.join("3d", f)) for f in os.listdir(input_dir) if f.endswith(('.gltf', '.glb', '.obj', '.mtl', '.fbx', '.stl'))] + files = [normalize_path(os.path.join("3d", f)) for f in os.listdir(input_dir) if f.endswith(('.gltf', '.glb', '.obj', '.fbx', '.stl'))] return {"required": { "model_file": (sorted(files), {"file_upload": True}), @@ -24,8 +26,8 @@ class Load3D(): "height": ("INT", {"default": 1024, "min": 1, "max": 4096, "step": 1}), }} - RETURN_TYPES = ("IMAGE", "MASK", "STRING", "IMAGE", "IMAGE", "LOAD3D_CAMERA") - RETURN_NAMES = ("image", "mask", "mesh_path", "normal", "lineart", "camera_info") + 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 @@ -44,7 +46,14 @@ class Load3D(): normal_image, ignore_mask2 = load_image_node.load_image(image=normal_path) lineart_image, ignore_mask3 = load_image_node.load_image(image=lineart_path) - return output_image, output_mask, model_file, normal_image, lineart_image, image['camera_info'] + 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, lineart_image, image['camera_info'], video class Load3DAnimation(): @@ -63,8 +72,8 @@ class Load3DAnimation(): "height": ("INT", {"default": 1024, "min": 1, "max": 4096, "step": 1}), }} - RETURN_TYPES = ("IMAGE", "MASK", "STRING", "IMAGE", "LOAD3D_CAMERA") - RETURN_NAMES = ("image", "mask", "mesh_path", "normal", "camera_info") + 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 @@ -81,7 +90,14 @@ class Load3DAnimation(): ignore_image, output_mask = load_image_node.load_image(image=mask_path) normal_image, ignore_mask2 = load_image_node.load_image(image=normal_path) - return output_image, output_mask, model_file, normal_image, image['camera_info'] + 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(): diff --git a/comfy_extras/nodes/nodes_lt.py b/comfy_extras/nodes/nodes_lt.py index 34e2514a4..07e8bb58f 100644 --- a/comfy_extras/nodes/nodes_lt.py +++ b/comfy_extras/nodes/nodes_lt.py @@ -42,6 +42,7 @@ class LTXVImgToVideo: "height": ("INT", {"default": 512, "min": 64, "max": nodes.MAX_RESOLUTION, "step": 32}), "length": ("INT", {"default": 97, "min": 9, "max": nodes.MAX_RESOLUTION, "step": 8}), "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}), + "strength": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0}), }} RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT") @@ -50,7 +51,7 @@ class LTXVImgToVideo: CATEGORY = "conditioning/video_models" FUNCTION = "generate" - def generate(self, positive, negative, image, vae, width, height, length, batch_size): + def generate(self, positive, negative, image, vae, width, height, length, batch_size, strength): pixels = comfy.utils.common_upscale(image.movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1) encode_pixels = pixels[:, :, :, :3] t = vae.encode(encode_pixels) @@ -63,7 +64,7 @@ class LTXVImgToVideo: dtype=torch.float32, device=latent.device, ) - conditioning_latent_frames_mask[:, :, :t.shape[2]] = 0 + conditioning_latent_frames_mask[:, :, :t.shape[2]] = 1.0 - strength return (positive, negative, {"samples": latent, "noise_mask": conditioning_latent_frames_mask},) @@ -158,6 +159,15 @@ class LTXVAddGuide: return node_helpers.conditioning_set_values(cond, {"keyframe_idxs": keyframe_idxs}) def append_keyframe(self, positive, negative, frame_idx, latent_image, noise_mask, guiding_latent, strength, scale_factors): + _, latent_idx = self.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 + positive = self.add_keyframe_index(positive, frame_idx, guiding_latent, scale_factors) negative = self.add_keyframe_index(negative, frame_idx, guiding_latent, scale_factors) @@ -393,7 +403,7 @@ def encode_single_frame(output_file, image_array: np.ndarray, crf): container = av.open(output_file, "w", format="mp4") try: stream = container.add_stream( - "h264", rate=1, options={"crf": str(crf), "preset": "veryfast"} + "libx264", rate=1, options={"crf": str(crf), "preset": "veryfast"} ) stream.height = image_array.shape[0] stream.width = image_array.shape[1] diff --git a/comfy_extras/nodes/nodes_mask.py b/comfy_extras/nodes/nodes_mask.py index 97c86d325..8818d8bb5 100644 --- a/comfy_extras/nodes/nodes_mask.py +++ b/comfy_extras/nodes/nodes_mask.py @@ -1,10 +1,13 @@ import numpy as np +import random import scipy.ndimage import torch from comfy import node_helpers from comfy import utils +from comfy.cmd import folder_paths from comfy.component_model.tensor_types import MaskBatch, RGBImageBatch +from comfy.nodes.base_nodes import SaveImage from comfy.nodes.common import MAX_RESOLUTION @@ -379,6 +382,31 @@ class ThresholdMask: return (mask,) +# 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(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 + + @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) + + NODE_CLASS_MAPPINGS = { "LatentCompositeMasked": LatentCompositeMasked, "ImageCompositeMasked": ImageCompositeMasked, @@ -392,6 +420,7 @@ NODE_CLASS_MAPPINGS = { "FeatherMask": FeatherMask, "GrowMask": GrowMask, "ThresholdMask": ThresholdMask, + "MaskPreview": MaskPreview } NODE_DISPLAY_NAME_MAPPINGS = { diff --git a/comfy_extras/nodes/nodes_model_merging.py b/comfy_extras/nodes/nodes_model_merging.py index 8fa08466f..9e891e8ec 100644 --- a/comfy_extras/nodes/nodes_model_merging.py +++ b/comfy_extras/nodes/nodes_model_merging.py @@ -208,6 +208,9 @@ def save_checkpoint(model, clip=None, vae=None, clip_vision=None, filename_prefi metadata["modelspec.predict_key"] = "epsilon" elif model.model.model_type == model_base.ModelType.V_PREDICTION: metadata["modelspec.predict_key"] = "v" + extra_keys["v_pred"] = torch.tensor([]) + if getattr(model_sampling, "zsnr", False): + extra_keys["ztsnr"] = torch.tensor([]) if not args.disable_metadata: metadata["prompt"] = prompt_info @@ -272,7 +275,7 @@ class CLIPSave: model_management.load_models_gpu([clip.load_model()], force_patch_weights=True) clip_sd = clip.get_sd() - for prefix in ["clip_l.", "clip_g.", ""]: + for prefix in ["clip_l.", "clip_g.", "clip_h.", "t5xxl.", "pile_t5xl.", "mt5xl.", "umt5xxl.", "t5base.", "gemma2_2b.", "llama.", "hydit_clip.", ""]: k = list(filter(lambda a: a.startswith(prefix), clip_sd.keys())) current_clip_sd = {} for x in k: diff --git a/comfy_extras/nodes/nodes_optimalsteps.py b/comfy_extras/nodes/nodes_optimalsteps.py index f6928199b..e7c851ca2 100644 --- a/comfy_extras/nodes/nodes_optimalsteps.py +++ b/comfy_extras/nodes/nodes_optimalsteps.py @@ -20,13 +20,14 @@ def loglinear_interp(t_steps, num_steps): NOISE_LEVELS = {"FLUX": [0.9968, 0.9886, 0.9819, 0.975, 0.966, 0.9471, 0.9158, 0.8287, 0.5512, 0.2808, 0.001], "Wan":[1.0, 0.997, 0.995, 0.993, 0.991, 0.989, 0.987, 0.985, 0.98, 0.975, 0.973, 0.968, 0.96, 0.946, 0.927, 0.902, 0.864, 0.776, 0.539, 0.208, 0.001], +"Chroma": [0.992, 0.99, 0.988, 0.985, 0.982, 0.978, 0.973, 0.968, 0.961, 0.953, 0.943, 0.931, 0.917, 0.9, 0.881, 0.858, 0.832, 0.802, 0.769, 0.731, 0.69, 0.646, 0.599, 0.55, 0.501, 0.451, 0.402, 0.355, 0.311, 0.27, 0.232, 0.199, 0.169, 0.143, 0.12, 0.101, 0.084, 0.07, 0.058, 0.048, 0.001], } class OptimalStepsScheduler: @classmethod def INPUT_TYPES(s): return {"required": - {"model_type": (["FLUX", "Wan"], ), + {"model_type": (["FLUX", "Wan", "Chroma"], ), "steps": ("INT", {"default": 20, "min": 3, "max": 1000}), "denoise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}), } diff --git a/comfy_extras/nodes/nodes_post_processing.py b/comfy_extras/nodes/nodes_post_processing.py index 455923c2b..23d794427 100644 --- a/comfy_extras/nodes/nodes_post_processing.py +++ b/comfy_extras/nodes/nodes_post_processing.py @@ -141,6 +141,7 @@ class Quantize: CATEGORY = "image/postprocessing" + @staticmethod def bayer(im, pal_im, order): def normalized_bayer_matrix(n): if n == 0: diff --git a/comfy_extras/nodes/nodes_primitive.py b/comfy_extras/nodes/nodes_primitive.py index b770104fb..1f93f87a7 100644 --- a/comfy_extras/nodes/nodes_primitive.py +++ b/comfy_extras/nodes/nodes_primitive.py @@ -1,6 +1,8 @@ # Primitive nodes that are evaluated at backend. from __future__ import annotations +import sys + from comfy.comfy_types.node_typing import ComfyNodeABC, InputTypeDict, IO @@ -19,11 +21,26 @@ class String(ComfyNodeABC): return (value,) +class StringMultiline(ComfyNodeABC): + @classmethod + def INPUT_TYPES(cls) -> InputTypeDict: + return { + "required": {"value": (IO.STRING, {"multiline": True,},)}, + } + + RETURN_TYPES = (IO.STRING,) + FUNCTION = "execute" + CATEGORY = "utils/primitive" + + def execute(self, value: str) -> tuple[str]: + return (value,) + + class Int(ComfyNodeABC): @classmethod def INPUT_TYPES(cls) -> InputTypeDict: return { - "required": {"value": (IO.INT, {"control_after_generate": True})}, + "required": {"value": (IO.INT, {"min": -sys.maxsize, "max": sys.maxsize, "control_after_generate": True})}, } RETURN_TYPES = (IO.INT,) @@ -38,7 +55,7 @@ class Float(ComfyNodeABC): @classmethod def INPUT_TYPES(cls) -> InputTypeDict: return { - "required": {"value": (IO.FLOAT, {})}, + "required": {"value": (IO.FLOAT, {"min": -sys.maxsize, "max": sys.maxsize})}, } RETURN_TYPES = (IO.FLOAT,) @@ -66,6 +83,7 @@ class Boolean(ComfyNodeABC): NODE_CLASS_MAPPINGS = { "PrimitiveString": String, + "PrimitiveStringMultiline": StringMultiline, "PrimitiveInt": Int, "PrimitiveFloat": Float, "PrimitiveBoolean": Boolean, @@ -73,6 +91,7 @@ NODE_CLASS_MAPPINGS = { NODE_DISPLAY_NAME_MAPPINGS = { "PrimitiveString": "String", + "PrimitiveStringMultiline": "String (Multiline)", "PrimitiveInt": "Int", "PrimitiveFloat": "Float", "PrimitiveBoolean": "Boolean", diff --git a/comfy_extras/nodes/nodes_torch_compile.py b/comfy_extras/nodes/nodes_torch_compile.py index 0aa6b480e..508515782 100644 --- a/comfy_extras/nodes/nodes_torch_compile.py +++ b/comfy_extras/nodes/nodes_torch_compile.py @@ -11,6 +11,7 @@ from comfy import model_management from comfy.language.transformers_model_management import TransformersManagedModel from comfy.model_patcher import ModelPatcher from comfy.nodes.package_typing import CustomNode, InputTypes +from comfy_api.torch_helpers import set_torch_compile_wrapper logger = logging.getLogger(__name__) @@ -103,6 +104,7 @@ class TorchCompileModel(CustomNode): if move_to_gpu: model_management.unload_all_models() model_management.load_models_gpu([m]) + set_torch_compile_wrapper(m, object_patch=object_patch, **compile_kwargs) m.add_object_patch(object_patch, torch.compile(model=m.get_model_object(object_patch), **compile_kwargs)) # todo: do we want to move something back off the GPU? # if move_to_gpu: diff --git a/comfy_extras/nodes/nodes_video.py b/comfy_extras/nodes/nodes_video.py index 1be2366b6..2e198d0ee 100644 --- a/comfy_extras/nodes/nodes_video.py +++ b/comfy_extras/nodes/nodes_video.py @@ -1,14 +1,18 @@ from __future__ import annotations +import av import json import os -from fractions import Fraction - -import av import torch +from fractions import Fraction +from typing import Optional, Literal +from comfy.cli_args import args from comfy.cmd import folder_paths -from comfy.comfy_types import FileLocator +from comfy.comfy_types import IO, FileLocator, ComfyNodeABC +from comfy_api.input import ImageInput, AudioInput, VideoInput +from comfy_api.input_impl import VideoFromFile, VideoFromComponents +from comfy_api.util import VideoContainer, VideoCodec, VideoComponents class SaveWEBM: @@ -51,13 +55,15 @@ class SaveWEBM: for x in extra_pnginfo: container.metadata[x] = json.dumps(extra_pnginfo[x]) - codec_map = {"vp9": "libvpx-vp9", "av1": "libaom-av1"} + codec_map = {"vp9": "libvpx-vp9", "av1": "libsvtav1"} stream = container.add_stream(codec_map[codec], rate=Fraction(round(fps * 1000), 1000)) stream.width = images.shape[-2] stream.height = images.shape[-3] - stream.pix_fmt = "yuv420p" + stream.pix_fmt = "yuv420p10le" if codec == "av1" else "yuv420p" stream.bit_rate = 0 stream.options = {'crf': str(crf)} + if codec == "av1": + stream.options["preset"] = "6" for frame in images: frame = av.VideoFrame.from_ndarray(torch.clamp(frame[..., :3] * 255, min=0, max=255).to(device=torch.device("cpu"), dtype=torch.uint8).numpy(), format="rgb24") @@ -75,6 +81,169 @@ class SaveWEBM: return {"ui": {"images": results, "animated": (True,)}} # TODO: frontend side +class SaveVideo(ComfyNodeABC): + def __init__(self): + self.output_dir = folder_paths.get_output_directory() + self.type: Literal["output"] = "output" + self.prefix_append = "" + + @classmethod + def INPUT_TYPES(cls): + return { + "required": { + "video": (IO.VIDEO, {"tooltip": "The video to save."}), + "filename_prefix": ("STRING", {"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."}), + "format": (VideoContainer.as_input(), {"default": "auto", "tooltip": "The format to save the video as."}), + "codec": (VideoCodec.as_input(), {"default": "auto", "tooltip": "The codec to use for the video."}), + }, + "hidden": { + "prompt": "PROMPT", + "extra_pnginfo": "EXTRA_PNGINFO" + }, + } + + RETURN_TYPES = () + FUNCTION = "save_video" + + OUTPUT_NODE = True + + CATEGORY = "image/video" + DESCRIPTION = "Saves the input images to your ComfyUI output directory." + + def save_video(self, video: VideoInput, filename_prefix, format, codec, prompt=None, extra_pnginfo=None): + filename_prefix += self.prefix_append + width, height = video.get_dimensions() + full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path( + filename_prefix, + self.output_dir, + width, + height + ) + results: list[FileLocator] = list() + saved_metadata = None + if not args.disable_metadata: + metadata = {} + if extra_pnginfo is not None: + metadata.update(extra_pnginfo) + if prompt is not None: + metadata["prompt"] = prompt + if len(metadata) > 0: + saved_metadata = metadata + file = f"{filename}_{counter:05}_.{VideoContainer.get_extension(format)}" + video.save_to( + os.path.join(full_output_folder, file), + format=format, + codec=codec, + metadata=saved_metadata + ) + + results.append({ + "filename": file, + "subfolder": subfolder, + "type": self.type + }) + counter += 1 + + return {"ui": {"images": results, "animated": (True,)}} + + +class CreateVideo(ComfyNodeABC): + @classmethod + def INPUT_TYPES(cls): + return { + "required": { + "images": (IO.IMAGE, {"tooltip": "The images to create a video from."}), + "fps": ("FLOAT", {"default": 30.0, "min": 1.0, "max": 120.0, "step": 1.0}), + }, + "optional": { + "audio": (IO.AUDIO, {"tooltip": "The audio to add to the video."}), + } + } + + RETURN_TYPES = (IO.VIDEO,) + FUNCTION = "create_video" + + CATEGORY = "image/video" + DESCRIPTION = "Create a video from images." + + def create_video(self, images: ImageInput, fps: float, audio: Optional[AudioInput] = None): + return (VideoFromComponents( + VideoComponents( + images=images, + audio=audio, + frame_rate=Fraction(fps), + ) + ),) + + +class GetVideoComponents(ComfyNodeABC): + @classmethod + def INPUT_TYPES(cls): + return { + "required": { + "video": (IO.VIDEO, {"tooltip": "The video to extract components from."}), + } + } + + RETURN_TYPES = (IO.IMAGE, IO.AUDIO, IO.FLOAT) + RETURN_NAMES = ("images", "audio", "fps") + FUNCTION = "get_components" + + CATEGORY = "image/video" + DESCRIPTION = "Extracts all components from a video: frames, audio, and framerate." + + def get_components(self, video: VideoInput): + components = video.get_components() + + return (components.images, components.audio, float(components.frame_rate)) + + +class LoadVideo(ComfyNodeABC): + @classmethod + def INPUT_TYPES(cls): + input_dir = folder_paths.get_input_directory() + files = [f for f in os.listdir(input_dir) if os.path.isfile(os.path.join(input_dir, f))] + files = folder_paths.filter_files_content_types(files, ["video"]) + return {"required": + {"file": (sorted(files), {"video_upload": True})}, + } + + CATEGORY = "image/video" + + RETURN_TYPES = (IO.VIDEO,) + FUNCTION = "load_video" + + def load_video(self, file): + video_path = folder_paths.get_annotated_filepath(file) + return (VideoFromFile(video_path),) + + @classmethod + def IS_CHANGED(cls, file): + video_path = folder_paths.get_annotated_filepath(file) + mod_time = os.path.getmtime(video_path) + # Instead of hashing the file, we can just use the modification time to avoid + # rehashing large files. + return mod_time + + @classmethod + def VALIDATE_INPUTS(cls, file): + if not folder_paths.exists_annotated_filepath(file): + return "Invalid video file: {}".format(file) + + return True + + NODE_CLASS_MAPPINGS = { "SaveWEBM": SaveWEBM, + "SaveVideo": SaveVideo, + "CreateVideo": CreateVideo, + "GetVideoComponents": GetVideoComponents, + "LoadVideo": LoadVideo, +} + +NODE_DISPLAY_NAME_MAPPINGS = { + "SaveVideo": "Save Video", + "CreateVideo": "Create Video", + "GetVideoComponents": "Get Video Components", + "LoadVideo": "Load Video", } diff --git a/comfy_extras/nodes/nodes_wan.py b/comfy_extras/nodes/nodes_wan.py index a3ece8816..b98f06d88 100644 --- a/comfy_extras/nodes/nodes_wan.py +++ b/comfy_extras/nodes/nodes_wan.py @@ -203,6 +203,7 @@ class WanVaceToVideo: "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}), "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}), "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}), + "strength": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1000.0, "step": 0.01}), }, "optional": {"control_video": ("IMAGE", ), "control_masks": ("MASK", ), @@ -217,7 +218,7 @@ class WanVaceToVideo: EXPERIMENTAL = True - def encode(self, positive, negative, vae, width, height, length, batch_size, control_video=None, control_masks=None, reference_image=None): + def encode(self, positive, negative, vae, width, height, length, batch_size, strength, control_video=None, control_masks=None, reference_image=None): latent_length = ((length - 1) // 4) + 1 if control_video is not None: control_video = comfy.utils.common_upscale(control_video[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1) @@ -267,8 +268,9 @@ class WanVaceToVideo: trim_latent = reference_image.shape[2] mask = mask.unsqueeze(0) - positive = node_helpers.conditioning_set_values(positive, {"vace_frames": control_video_latent, "vace_mask": mask}) - negative = node_helpers.conditioning_set_values(negative, {"vace_frames": control_video_latent, "vace_mask": mask}) + + positive = node_helpers.conditioning_set_values(positive, {"vace_frames": [control_video_latent], "vace_mask": [mask], "vace_strength": [strength]}, append=True) + negative = node_helpers.conditioning_set_values(negative, {"vace_frames": [control_video_latent], "vace_mask": [mask], "vace_strength": [strength]}, append=True) latent = torch.zeros([batch_size, 16, latent_length, height // 8, width // 8], device=comfy.model_management.intermediate_device()) out_latent = {} @@ -296,6 +298,90 @@ class TrimVideoLatent: samples_out["samples"] = s1[:, :, trim_amount:] return (samples_out,) +class WanCameraImageToVideo: + @classmethod + def INPUT_TYPES(s): + return {"required": {"positive": ("CONDITIONING", ), + "negative": ("CONDITIONING", ), + "vae": ("VAE", ), + "width": ("INT", {"default": 832, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}), + "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}), + "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}), + "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}), + }, + "optional": {"clip_vision_output": ("CLIP_VISION_OUTPUT", ), + "start_image": ("IMAGE", ), + "camera_conditions": ("WAN_CAMERA_EMBEDDING", ), + }} + + RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT") + RETURN_NAMES = ("positive", "negative", "latent") + FUNCTION = "encode" + + CATEGORY = "conditioning/video_models" + + def encode(self, positive, negative, vae, width, height, length, batch_size, start_image=None, clip_vision_output=None, camera_conditions=None): + latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device()) + concat_latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device()) + concat_latent = comfy.latent_formats.Wan21().process_out(concat_latent) + + 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) + concat_latent_image = vae.encode(start_image[:, :, :, :3]) + concat_latent[:,:,:concat_latent_image.shape[2]] = concat_latent_image[:,:,:concat_latent.shape[2]] + + positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": concat_latent}) + negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": concat_latent}) + + if camera_conditions is not None: + positive = node_helpers.conditioning_set_values(positive, {'camera_conditions': camera_conditions}) + negative = node_helpers.conditioning_set_values(negative, {'camera_conditions': camera_conditions}) + + 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 (positive, negative, out_latent) + +class WanPhantomSubjectToVideo: + @classmethod + def INPUT_TYPES(s): + return {"required": {"positive": ("CONDITIONING", ), + "negative": ("CONDITIONING", ), + "vae": ("VAE", ), + "width": ("INT", {"default": 832, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}), + "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}), + "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}), + "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}), + }, + "optional": {"images": ("IMAGE", ), + }} + + RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "CONDITIONING", "LATENT") + RETURN_NAMES = ("positive", "negative_text", "negative_img_text", "latent") + FUNCTION = "encode" + + CATEGORY = "conditioning/video_models" + + def encode(self, positive, negative, vae, width, height, length, batch_size, images): + latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device()) + cond2 = negative + if images is not None: + images = comfy.utils.common_upscale(images[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1) + latent_images = [] + for i in images: + latent_images += [vae.encode(i.unsqueeze(0)[:, :, :, :3])] + concat_latent_image = torch.cat(latent_images, dim=2) + + positive = node_helpers.conditioning_set_values(positive, {"time_dim_concat": concat_latent_image}) + cond2 = node_helpers.conditioning_set_values(negative, {"time_dim_concat": concat_latent_image}) + negative = node_helpers.conditioning_set_values(negative, {"time_dim_concat": comfy.latent_formats.Wan21().process_out(torch.zeros_like(concat_latent_image))}) + + out_latent = {} + out_latent["samples"] = latent + return (positive, cond2, negative, out_latent) NODE_CLASS_MAPPINGS = { "WanImageToVideo": WanImageToVideo, @@ -304,4 +390,6 @@ NODE_CLASS_MAPPINGS = { "WanFirstLastFrameToVideo": WanFirstLastFrameToVideo, "WanVaceToVideo": WanVaceToVideo, "TrimVideoLatent": TrimVideoLatent, + "WanCameraImageToVideo": WanCameraImageToVideo, + "WanPhantomSubjectToVideo": WanPhantomSubjectToVideo, } diff --git a/comfy_extras/nodes/nodes_webcam.py b/comfy_extras/nodes/nodes_webcam.py index ab87ded19..40635950a 100644 --- a/comfy_extras/nodes/nodes_webcam.py +++ b/comfy_extras/nodes/nodes_webcam.py @@ -20,9 +20,13 @@ class WebcamCapture(LoadImage): CATEGORY = "image" - def load_capture(s, image, **kwargs): + def load_capture(self, image, **kwargs): return super().load_image(folder_paths.get_annotated_filepath(image)) + @classmethod + def IS_CHANGED(cls, image, width, height, capture_on_queue): + return super().IS_CHANGED(image) + NODE_CLASS_MAPPINGS = { "WebcamCapture": WebcamCapture, diff --git a/comfy_extras/nodes_ace.py b/comfy_extras/nodes_ace.py new file mode 100644 index 000000000..cbfec15a2 --- /dev/null +++ b/comfy_extras/nodes_ace.py @@ -0,0 +1,49 @@ +import torch +import comfy.model_management +import node_helpers + +class TextEncodeAceStepAudio: + @classmethod + def INPUT_TYPES(s): + return {"required": { + "clip": ("CLIP", ), + "tags": ("STRING", {"multiline": True, "dynamicPrompts": True}), + "lyrics": ("STRING", {"multiline": True, "dynamicPrompts": True}), + "lyrics_strength": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}), + }} + RETURN_TYPES = ("CONDITIONING",) + FUNCTION = "encode" + + CATEGORY = "conditioning" + + def encode(self, clip, tags, lyrics, lyrics_strength): + tokens = clip.tokenize(tags, lyrics=lyrics) + conditioning = clip.encode_from_tokens_scheduled(tokens) + conditioning = node_helpers.conditioning_set_values(conditioning, {"lyrics_strength": lyrics_strength}) + return (conditioning, ) + + +class EmptyAceStepLatentAudio: + def __init__(self): + self.device = comfy.model_management.intermediate_device() + + @classmethod + def INPUT_TYPES(s): + return {"required": {"seconds": ("FLOAT", {"default": 120.0, "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): + length = int(seconds * 44100 / 512 / 8) + latent = torch.zeros([batch_size, 8, 16, length], device=self.device) + return ({"samples": latent, "type": "audio"}, ) + + +NODE_CLASS_MAPPINGS = { + "TextEncodeAceStepAudio": TextEncodeAceStepAudio, + "EmptyAceStepLatentAudio": EmptyAceStepLatentAudio, +} diff --git a/comfy_extras/nodes_apg.py b/comfy_extras/nodes_apg.py new file mode 100644 index 000000000..25b21b1b8 --- /dev/null +++ b/comfy_extras/nodes_apg.py @@ -0,0 +1,76 @@ +import torch + +def project(v0, v1): + v1 = torch.nn.functional.normalize(v1, dim=[-1, -2, -3]) + v0_parallel = (v0 * v1).sum(dim=[-1, -2, -3], keepdim=True) * v1 + v0_orthogonal = v0 - v0_parallel + return v0_parallel, v0_orthogonal + +class APG: + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "model": ("MODEL",), + "eta": ("FLOAT", {"default": 1.0, "min": -10.0, "max": 10.0, "step": 0.01, "tooltip": "Controls the scale of the parallel guidance vector. Default CFG behavior at a setting of 1."}), + "norm_threshold": ("FLOAT", {"default": 5.0, "min": 0.0, "max": 50.0, "step": 0.1, "tooltip": "Normalize guidance vector to this value, normalization disable at a setting of 0."}), + "momentum": ("FLOAT", {"default": 0.0, "min": -5.0, "max": 1.0, "step": 0.01, "tooltip":"Controls a running average of guidance during diffusion, disabled at a setting of 0."}), + } + } + RETURN_TYPES = ("MODEL",) + FUNCTION = "patch" + CATEGORY = "sampling/custom_sampling" + + def patch(self, model, eta, norm_threshold, momentum): + running_avg = 0 + prev_sigma = None + + def pre_cfg_function(args): + nonlocal running_avg, prev_sigma + + if len(args["conds_out"]) == 1: return args["conds_out"] + + cond = args["conds_out"][0] + uncond = args["conds_out"][1] + sigma = args["sigma"][0] + cond_scale = args["cond_scale"] + + if prev_sigma is not None and sigma > prev_sigma: + running_avg = 0 + prev_sigma = sigma + + guidance = cond - uncond + + if momentum != 0: + if not torch.is_tensor(running_avg): + running_avg = guidance + else: + running_avg = momentum * running_avg + guidance + guidance = running_avg + + if norm_threshold > 0: + guidance_norm = guidance.norm(p=2, dim=[-1, -2, -3], keepdim=True) + scale = torch.minimum( + torch.ones_like(guidance_norm), + norm_threshold / guidance_norm + ) + guidance = guidance * scale + + guidance_parallel, guidance_orthogonal = project(guidance, cond) + modified_guidance = guidance_orthogonal + eta * guidance_parallel + + modified_cond = (uncond + modified_guidance) + (cond - uncond) / cond_scale + + return [modified_cond, uncond] + args["conds_out"][2:] + + m = model.clone() + m.set_model_sampler_pre_cfg_function(pre_cfg_function) + return (m,) + +NODE_CLASS_MAPPINGS = { + "APG": APG, +} + +NODE_DISPLAY_NAME_MAPPINGS = { + "APG": "Adaptive Projected Guidance", +} diff --git a/comfy_extras/nodes_camera_trajectory.py b/comfy_extras/nodes_camera_trajectory.py new file mode 100644 index 000000000..5e0e39f91 --- /dev/null +++ b/comfy_extras/nodes_camera_trajectory.py @@ -0,0 +1,218 @@ +import nodes +import torch +import numpy as np +from einops import rearrange +import comfy.model_management + + + +MAX_RESOLUTION = nodes.MAX_RESOLUTION + +CAMERA_DICT = { + "base_T_norm": 1.5, + "base_angle": np.pi/3, + "Static": { "angle":[0., 0., 0.], "T":[0., 0., 0.]}, + "Pan Up": { "angle":[0., 0., 0.], "T":[0., -1., 0.]}, + "Pan Down": { "angle":[0., 0., 0.], "T":[0.,1.,0.]}, + "Pan Left": { "angle":[0., 0., 0.], "T":[-1.,0.,0.]}, + "Pan Right": { "angle":[0., 0., 0.], "T": [1.,0.,0.]}, + "Zoom In": { "angle":[0., 0., 0.], "T": [0.,0.,2.]}, + "Zoom Out": { "angle":[0., 0., 0.], "T": [0.,0.,-2.]}, + "Anti Clockwise (ACW)": { "angle": [0., 0., -1.], "T":[0., 0., 0.]}, + "ClockWise (CW)": { "angle": [0., 0., 1.], "T":[0., 0., 0.]}, +} + + +def process_pose_params(cam_params, width=672, height=384, original_pose_width=1280, original_pose_height=720, device='cpu'): + + def get_relative_pose(cam_params): + """Copied from https://github.com/hehao13/CameraCtrl/blob/main/inference.py + """ + abs_w2cs = [cam_param.w2c_mat for cam_param in cam_params] + abs_c2ws = [cam_param.c2w_mat for cam_param in cam_params] + cam_to_origin = 0 + target_cam_c2w = np.array([ + [1, 0, 0, 0], + [0, 1, 0, -cam_to_origin], + [0, 0, 1, 0], + [0, 0, 0, 1] + ]) + abs2rel = target_cam_c2w @ abs_w2cs[0] + ret_poses = [target_cam_c2w, ] + [abs2rel @ abs_c2w for abs_c2w in abs_c2ws[1:]] + ret_poses = np.array(ret_poses, dtype=np.float32) + return ret_poses + + """Modified from https://github.com/hehao13/CameraCtrl/blob/main/inference.py + """ + cam_params = [Camera(cam_param) for cam_param in cam_params] + + sample_wh_ratio = width / height + pose_wh_ratio = original_pose_width / original_pose_height # Assuming placeholder ratios, change as needed + + if pose_wh_ratio > sample_wh_ratio: + resized_ori_w = height * pose_wh_ratio + for cam_param in cam_params: + cam_param.fx = resized_ori_w * cam_param.fx / width + else: + resized_ori_h = width / pose_wh_ratio + for cam_param in cam_params: + cam_param.fy = resized_ori_h * cam_param.fy / height + + intrinsic = np.asarray([[cam_param.fx * width, + cam_param.fy * height, + cam_param.cx * width, + cam_param.cy * height] + for cam_param in cam_params], dtype=np.float32) + + K = torch.as_tensor(intrinsic)[None] # [1, 1, 4] + c2ws = get_relative_pose(cam_params) # Assuming this function is defined elsewhere + c2ws = torch.as_tensor(c2ws)[None] # [1, n_frame, 4, 4] + plucker_embedding = ray_condition(K, c2ws, height, width, device=device)[0].permute(0, 3, 1, 2).contiguous() # V, 6, H, W + plucker_embedding = plucker_embedding[None] + plucker_embedding = rearrange(plucker_embedding, "b f c h w -> b f h w c")[0] + return plucker_embedding + +class Camera(object): + """Copied from https://github.com/hehao13/CameraCtrl/blob/main/inference.py + """ + def __init__(self, entry): + fx, fy, cx, cy = entry[1:5] + self.fx = fx + self.fy = fy + self.cx = cx + self.cy = cy + c2w_mat = np.array(entry[7:]).reshape(4, 4) + self.c2w_mat = c2w_mat + self.w2c_mat = np.linalg.inv(c2w_mat) + +def ray_condition(K, c2w, H, W, device): + """Copied from https://github.com/hehao13/CameraCtrl/blob/main/inference.py + """ + # c2w: B, V, 4, 4 + # K: B, V, 4 + + B = K.shape[0] + + j, i = torch.meshgrid( + torch.linspace(0, H - 1, H, device=device, dtype=c2w.dtype), + torch.linspace(0, W - 1, W, device=device, dtype=c2w.dtype), + indexing='ij' + ) + i = i.reshape([1, 1, H * W]).expand([B, 1, H * W]) + 0.5 # [B, HxW] + j = j.reshape([1, 1, H * W]).expand([B, 1, H * W]) + 0.5 # [B, HxW] + + fx, fy, cx, cy = K.chunk(4, dim=-1) # B,V, 1 + + zs = torch.ones_like(i) # [B, HxW] + xs = (i - cx) / fx * zs + ys = (j - cy) / fy * zs + zs = zs.expand_as(ys) + + directions = torch.stack((xs, ys, zs), dim=-1) # B, V, HW, 3 + directions = directions / directions.norm(dim=-1, keepdim=True) # B, V, HW, 3 + + rays_d = directions @ c2w[..., :3, :3].transpose(-1, -2) # B, V, 3, HW + rays_o = c2w[..., :3, 3] # B, V, 3 + rays_o = rays_o[:, :, None].expand_as(rays_d) # B, V, 3, HW + # c2w @ dirctions + rays_dxo = torch.cross(rays_o, rays_d) + plucker = torch.cat([rays_dxo, rays_d], dim=-1) + plucker = plucker.reshape(B, c2w.shape[1], H, W, 6) # B, V, H, W, 6 + # plucker = plucker.permute(0, 1, 4, 2, 3) + return plucker + +def get_camera_motion(angle, T, speed, n=81): + def compute_R_form_rad_angle(angles): + theta_x, theta_y, theta_z = angles + Rx = np.array([[1, 0, 0], + [0, np.cos(theta_x), -np.sin(theta_x)], + [0, np.sin(theta_x), np.cos(theta_x)]]) + + Ry = np.array([[np.cos(theta_y), 0, np.sin(theta_y)], + [0, 1, 0], + [-np.sin(theta_y), 0, np.cos(theta_y)]]) + + Rz = np.array([[np.cos(theta_z), -np.sin(theta_z), 0], + [np.sin(theta_z), np.cos(theta_z), 0], + [0, 0, 1]]) + + R = np.dot(Rz, np.dot(Ry, Rx)) + return R + RT = [] + for i in range(n): + _angle = (i/n)*speed*(CAMERA_DICT["base_angle"])*angle + R = compute_R_form_rad_angle(_angle) + _T=(i/n)*speed*(CAMERA_DICT["base_T_norm"])*(T.reshape(3,1)) + _RT = np.concatenate([R,_T], axis=1) + RT.append(_RT) + RT = np.stack(RT) + return RT + +class WanCameraEmbedding: + @classmethod + def INPUT_TYPES(cls): + return { + "required": { + "camera_pose":(["Static","Pan Up","Pan Down","Pan Left","Pan Right","Zoom In","Zoom Out","Anti Clockwise (ACW)", "ClockWise (CW)"],{"default":"Static"}), + "width": ("INT", {"default": 832, "min": 16, "max": MAX_RESOLUTION, "step": 16}), + "height": ("INT", {"default": 480, "min": 16, "max": MAX_RESOLUTION, "step": 16}), + "length": ("INT", {"default": 81, "min": 1, "max": MAX_RESOLUTION, "step": 4}), + }, + "optional":{ + "speed":("FLOAT",{"default":1.0, "min": 0, "max": 10.0, "step": 0.1}), + "fx":("FLOAT",{"default":0.5, "min": 0, "max": 1, "step": 0.000000001}), + "fy":("FLOAT",{"default":0.5, "min": 0, "max": 1, "step": 0.000000001}), + "cx":("FLOAT",{"default":0.5, "min": 0, "max": 1, "step": 0.01}), + "cy":("FLOAT",{"default":0.5, "min": 0, "max": 1, "step": 0.01}), + } + + } + + RETURN_TYPES = ("WAN_CAMERA_EMBEDDING","INT","INT","INT") + RETURN_NAMES = ("camera_embedding","width","height","length") + FUNCTION = "run" + CATEGORY = "camera" + + def run(self, camera_pose, width, height, length, speed=1.0, fx=0.5, fy=0.5, cx=0.5, cy=0.5): + """ + Use Camera trajectory as extrinsic parameters to calculate Plücker embeddings (Sitzmannet al., 2021) + Adapted from https://github.com/aigc-apps/VideoX-Fun/blob/main/comfyui/comfyui_nodes.py + """ + motion_list = [camera_pose] + speed = speed + angle = np.array(CAMERA_DICT[motion_list[0]]["angle"]) + T = np.array(CAMERA_DICT[motion_list[0]]["T"]) + RT = get_camera_motion(angle, T, speed, length) + + trajs=[] + for cp in RT.tolist(): + traj=[fx,fy,cx,cy,0,0] + traj.extend(cp[0]) + traj.extend(cp[1]) + traj.extend(cp[2]) + traj.extend([0,0,0,1]) + trajs.append(traj) + + cam_params = np.array([[float(x) for x in pose] for pose in trajs]) + cam_params = np.concatenate([np.zeros_like(cam_params[:, :1]), cam_params], 1) + control_camera_video = process_pose_params(cam_params, width=width, height=height) + control_camera_video = control_camera_video.permute([3, 0, 1, 2]).unsqueeze(0).to(device=comfy.model_management.intermediate_device()) + + control_camera_video = torch.concat( + [ + torch.repeat_interleave(control_camera_video[:, :, 0:1], repeats=4, dim=2), + control_camera_video[:, :, 1:] + ], dim=2 + ).transpose(1, 2) + + # Reshape, transpose, and view into desired shape + b, f, c, h, w = control_camera_video.shape + control_camera_video = control_camera_video.contiguous().view(b, f // 4, 4, c, h, w).transpose(2, 3) + control_camera_video = control_camera_video.contiguous().view(b, f // 4, c * 4, h, w).transpose(1, 2) + + return (control_camera_video, width, height, length) + + +NODE_CLASS_MAPPINGS = { + "WanCameraEmbedding": WanCameraEmbedding, +} diff --git a/comfy_extras/nodes_preview_any.py b/comfy_extras/nodes_preview_any.py new file mode 100644 index 000000000..e6805696f --- /dev/null +++ b/comfy_extras/nodes_preview_any.py @@ -0,0 +1,43 @@ +import json +from comfy.comfy_types.node_typing import IO + +# Preview Any - original implement from +# https://github.com/rgthree/rgthree-comfy/blob/main/py/display_any.py +# upstream requested in https://github.com/Kosinkadink/rfcs/blob/main/rfcs/0000-corenodes.md#preview-nodes +class PreviewAny(): + @classmethod + def INPUT_TYPES(cls): + return { + "required": {"source": (IO.ANY, {})}, + } + + RETURN_TYPES = () + FUNCTION = "main" + OUTPUT_NODE = True + + CATEGORY = "utils" + + def main(self, source=None): + value = 'None' + if isinstance(source, str): + value = source + elif isinstance(source, (int, float, bool)): + value = str(source) + elif source is not None: + try: + value = json.dumps(source) + except Exception: + try: + value = str(source) + except Exception: + value = 'source exists, but could not be serialized.' + + return {"ui": {"text": (value,)}} + +NODE_CLASS_MAPPINGS = { + "PreviewAny": PreviewAny, +} + +NODE_DISPLAY_NAME_MAPPINGS = { + "PreviewAny": "Preview Any", +} diff --git a/comfy_extras/nodes_string.py b/comfy_extras/nodes_string.py new file mode 100644 index 000000000..b1a8ceef0 --- /dev/null +++ b/comfy_extras/nodes_string.py @@ -0,0 +1,360 @@ +import re + +from comfy.comfy_types.node_typing import IO + +class StringConcatenate(): + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "string_a": (IO.STRING, {"multiline": True}), + "string_b": (IO.STRING, {"multiline": True}), + "delimiter": (IO.STRING, {"multiline": False, "default": ""}) + } + } + + RETURN_TYPES = (IO.STRING,) + FUNCTION = "execute" + CATEGORY = "utils/string" + + def execute(self, string_a, string_b, delimiter, **kwargs): + return delimiter.join((string_a, string_b)), + +class StringSubstring(): + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "string": (IO.STRING, {"multiline": True}), + "start": (IO.INT, {}), + "end": (IO.INT, {}), + } + } + + RETURN_TYPES = (IO.STRING,) + FUNCTION = "execute" + CATEGORY = "utils/string" + + def execute(self, string, start, end, **kwargs): + return string[start:end], + +class StringLength(): + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "string": (IO.STRING, {"multiline": True}) + } + } + + RETURN_TYPES = (IO.INT,) + RETURN_NAMES = ("length",) + FUNCTION = "execute" + CATEGORY = "utils/string" + + def execute(self, string, **kwargs): + length = len(string) + + return length, + +class CaseConverter(): + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "string": (IO.STRING, {"multiline": True}), + "mode": (IO.COMBO, {"options": ["UPPERCASE", "lowercase", "Capitalize", "Title Case"]}) + } + } + + RETURN_TYPES = (IO.STRING,) + FUNCTION = "execute" + CATEGORY = "utils/string" + + def execute(self, string, mode, **kwargs): + if mode == "UPPERCASE": + result = string.upper() + elif mode == "lowercase": + result = string.lower() + elif mode == "Capitalize": + result = string.capitalize() + elif mode == "Title Case": + result = string.title() + else: + result = string + + return result, + + +class StringTrim(): + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "string": (IO.STRING, {"multiline": True}), + "mode": (IO.COMBO, {"options": ["Both", "Left", "Right"]}) + } + } + + RETURN_TYPES = (IO.STRING,) + FUNCTION = "execute" + CATEGORY = "utils/string" + + def execute(self, string, mode, **kwargs): + if mode == "Both": + result = string.strip() + elif mode == "Left": + result = string.lstrip() + elif mode == "Right": + result = string.rstrip() + else: + result = string + + return result, + +class StringReplace(): + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "string": (IO.STRING, {"multiline": True}), + "find": (IO.STRING, {"multiline": True}), + "replace": (IO.STRING, {"multiline": True}) + } + } + + RETURN_TYPES = (IO.STRING,) + FUNCTION = "execute" + CATEGORY = "utils/string" + + def execute(self, string, find, replace, **kwargs): + result = string.replace(find, replace) + return result, + + +class StringContains(): + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "string": (IO.STRING, {"multiline": True}), + "substring": (IO.STRING, {"multiline": True}), + "case_sensitive": (IO.BOOLEAN, {"default": True}) + } + } + + RETURN_TYPES = (IO.BOOLEAN,) + RETURN_NAMES = ("contains",) + FUNCTION = "execute" + CATEGORY = "utils/string" + + def execute(self, string, substring, case_sensitive, **kwargs): + if case_sensitive: + contains = substring in string + else: + contains = substring.lower() in string.lower() + + return contains, + + +class StringCompare(): + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "string_a": (IO.STRING, {"multiline": True}), + "string_b": (IO.STRING, {"multiline": True}), + "mode": (IO.COMBO, {"options": ["Starts With", "Ends With", "Equal"]}), + "case_sensitive": (IO.BOOLEAN, {"default": True}) + } + } + + RETURN_TYPES = (IO.BOOLEAN,) + FUNCTION = "execute" + CATEGORY = "utils/string" + + def execute(self, string_a, string_b, mode, case_sensitive, **kwargs): + if case_sensitive: + a = string_a + b = string_b + else: + a = string_a.lower() + b = string_b.lower() + + if mode == "Equal": + return a == b, + elif mode == "Starts With": + return a.startswith(b), + elif mode == "Ends With": + return a.endswith(b), + +class RegexMatch(): + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "string": (IO.STRING, {"multiline": True}), + "regex_pattern": (IO.STRING, {"multiline": True}), + "case_insensitive": (IO.BOOLEAN, {"default": True}), + "multiline": (IO.BOOLEAN, {"default": False}), + "dotall": (IO.BOOLEAN, {"default": False}) + } + } + + RETURN_TYPES = (IO.BOOLEAN,) + RETURN_NAMES = ("matches",) + FUNCTION = "execute" + CATEGORY = "utils/string" + + def execute(self, string, regex_pattern, case_insensitive, multiline, dotall, **kwargs): + flags = 0 + + if case_insensitive: + flags |= re.IGNORECASE + if multiline: + flags |= re.MULTILINE + if dotall: + flags |= re.DOTALL + + try: + match = re.search(regex_pattern, string, flags) + result = match is not None + + except re.error: + result = False + + return result, + + +class RegexExtract(): + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "string": (IO.STRING, {"multiline": True}), + "regex_pattern": (IO.STRING, {"multiline": True}), + "mode": (IO.COMBO, {"options": ["First Match", "All Matches", "First Group", "All Groups"]}), + "case_insensitive": (IO.BOOLEAN, {"default": True}), + "multiline": (IO.BOOLEAN, {"default": False}), + "dotall": (IO.BOOLEAN, {"default": False}), + "group_index": (IO.INT, {"default": 1, "min": 0, "max": 100}) + } + } + + RETURN_TYPES = (IO.STRING,) + FUNCTION = "execute" + CATEGORY = "utils/string" + + def execute(self, string, regex_pattern, mode, case_insensitive, multiline, dotall, group_index, **kwargs): + join_delimiter = "\n" + + flags = 0 + if case_insensitive: + flags |= re.IGNORECASE + if multiline: + flags |= re.MULTILINE + if dotall: + flags |= re.DOTALL + + try: + if mode == "First Match": + match = re.search(regex_pattern, string, flags) + if match: + result = match.group(0) + else: + result = "" + + elif mode == "All Matches": + matches = re.findall(regex_pattern, string, flags) + if matches: + if isinstance(matches[0], tuple): + result = join_delimiter.join([m[0] for m in matches]) + else: + result = join_delimiter.join(matches) + else: + result = "" + + elif mode == "First Group": + match = re.search(regex_pattern, string, flags) + if match and len(match.groups()) >= group_index: + result = match.group(group_index) + else: + result = "" + + elif mode == "All Groups": + matches = re.finditer(regex_pattern, string, flags) + results = [] + for match in matches: + if match.groups() and len(match.groups()) >= group_index: + results.append(match.group(group_index)) + result = join_delimiter.join(results) + else: + result = "" + + except re.error: + result = "" + + return result, + + +class RegexReplace(): + DESCRIPTION = "Find and replace text using regex patterns." + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "string": (IO.STRING, {"multiline": True}), + "regex_pattern": (IO.STRING, {"multiline": True}), + "replace": (IO.STRING, {"multiline": True}), + }, + "optional": { + "case_insensitive": (IO.BOOLEAN, {"default": True}), + "multiline": (IO.BOOLEAN, {"default": False}), + "dotall": (IO.BOOLEAN, {"default": False, "tooltip": "When enabled, the dot (.) character will match any character including newline characters. When disabled, dots won't match newlines."}), + "count": (IO.INT, {"default": 0, "min": 0, "max": 100, "tooltip": "Maximum number of replacements to make. Set to 0 to replace all occurrences (default). Set to 1 to replace only the first match, 2 for the first two matches, etc."}), + } + } + + RETURN_TYPES = (IO.STRING,) + FUNCTION = "execute" + CATEGORY = "utils/string" + + def execute(self, string, regex_pattern, replace, case_insensitive=True, multiline=False, dotall=False, count=0, **kwargs): + flags = 0 + + if case_insensitive: + flags |= re.IGNORECASE + if multiline: + flags |= re.MULTILINE + if dotall: + flags |= re.DOTALL + result = re.sub(regex_pattern, replace, string, count=count, flags=flags) + return result, + +NODE_CLASS_MAPPINGS = { + "StringConcatenate": StringConcatenate, + "StringSubstring": StringSubstring, + "StringLength": StringLength, + "CaseConverter": CaseConverter, + "StringTrim": StringTrim, + "StringReplace": StringReplace, + "StringContains": StringContains, + "StringCompare": StringCompare, + "RegexMatch": RegexMatch, + "RegexExtract": RegexExtract, + "RegexReplace": RegexReplace, +} + +NODE_DISPLAY_NAME_MAPPINGS = { + "StringConcatenate": "Concatenate", + "StringSubstring": "Substring", + "StringLength": "Length", + "CaseConverter": "Case Converter", + "StringTrim": "Trim", + "StringReplace": "Replace", + "StringContains": "Contains", + "StringCompare": "Compare", + "RegexMatch": "Regex Match", + "RegexExtract": "Regex Extract", + "RegexReplace": "Regex Replace", +} diff --git a/comfyui_version.py b/comfyui_version.py new file mode 100644 index 000000000..6962c3661 --- /dev/null +++ b/comfyui_version.py @@ -0,0 +1,3 @@ +# This file is automatically generated by the build process when version is +# updated in pyproject.toml. +__version__ = "0.3.40" diff --git a/hook_breaker_ac10a0.py b/hook_breaker_ac10a0.py new file mode 100644 index 000000000..c3e1c0633 --- /dev/null +++ b/hook_breaker_ac10a0.py @@ -0,0 +1,17 @@ +# Prevent custom nodes from hooking anything important +import comfy.model_management + +HOOK_BREAK = [(comfy.model_management, "cast_to")] + + +SAVED_FUNCTIONS = [] + + +def save_functions(): + for f in HOOK_BREAK: + SAVED_FUNCTIONS.append((f[0], f[1], getattr(f[0], f[1]))) + + +def restore_functions(): + for f in SAVED_FUNCTIONS: + setattr(f[0], f[1], f[2]) diff --git a/pyproject.toml b/pyproject.toml index 76e17cead..b9f56f090 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "comfyui" -version = "0.3.29" +version = "0.3.40" description = "An installable version of ComfyUI" readme = "README.md" authors = [ @@ -23,6 +23,7 @@ dependencies = [ "torchaudio", "comfyui-frontend-package", "comfyui-workflow-templates", + "comfyui-embedded-docs", "torchdiffeq>=0.2.3", "torchsde>=0.2.6", "einops>=0.6.0", @@ -96,7 +97,8 @@ dependencies = [ "flax", "jax", "colour", - "av", + "av>=14.2.0", + "pydantic~=2.0", "typer", "ijson", "scikit-learn>=1.4.1", @@ -234,6 +236,7 @@ comfyui-frontend-package = [ [tool.ruff] lint.select = [ + "N805", # invalid-first-argument-name-for-method "S307", # suspicious-eval-usage "S102", # exec "T", # print-usage diff --git a/tests-unit/comfy_api_nodes_test/mapper_utils_test.py b/tests-unit/comfy_api_nodes_test/mapper_utils_test.py new file mode 100644 index 000000000..69488f691 --- /dev/null +++ b/tests-unit/comfy_api_nodes_test/mapper_utils_test.py @@ -0,0 +1,297 @@ +from typing import Optional +from enum import Enum + +from pydantic import BaseModel, Field + +from comfy.comfy_types.node_typing import IO +from comfy_api_nodes.mapper_utils import model_field_to_node_input + + +def test_model_field_to_float_input(): + """Tests mapping a float field with constraints.""" + + class ModelWithFloatField(BaseModel): + cfg_scale: Optional[float] = Field( + default=0.5, + description="Flexibility in video generation", + ge=0.0, + le=1.0, + multiple_of=0.001, + ) + + expected_output = ( + IO.FLOAT, + { + "default": 0.5, + "tooltip": "Flexibility in video generation", + "min": 0.0, + "max": 1.0, + "step": 0.001, + }, + ) + + actual_output = model_field_to_node_input( + IO.FLOAT, ModelWithFloatField, "cfg_scale" + ) + + assert actual_output[0] == expected_output[0] + assert actual_output[1] == expected_output[1] + + +def test_model_field_to_float_input_no_constraints(): + """Tests mapping a float field with no constraints.""" + + class ModelWithFloatField(BaseModel): + cfg_scale: Optional[float] = Field(default=0.5) + + expected_output = ( + IO.FLOAT, + { + "default": 0.5, + }, + ) + + actual_output = model_field_to_node_input( + IO.FLOAT, ModelWithFloatField, "cfg_scale" + ) + + assert actual_output[0] == expected_output[0] + assert actual_output[1] == expected_output[1] + + +def test_model_field_to_int_input(): + """Tests mapping an int field with constraints.""" + + class ModelWithIntField(BaseModel): + num_frames: Optional[int] = Field( + default=10, + description="Number of frames to generate", + ge=1, + le=100, + multiple_of=1, + ) + + expected_output = ( + IO.INT, + { + "default": 10, + "tooltip": "Number of frames to generate", + "min": 1, + "max": 100, + "step": 1, + }, + ) + + actual_output = model_field_to_node_input(IO.INT, ModelWithIntField, "num_frames") + + assert actual_output[0] == expected_output[0] + assert actual_output[1] == expected_output[1] + + +def test_model_field_to_string_input(): + """Tests mapping a string field.""" + + class ModelWithStringField(BaseModel): + prompt: Optional[str] = Field( + default="A beautiful sunset over a calm ocean", + description="A prompt for the video generation", + ) + + expected_output = ( + IO.STRING, + { + "default": "A beautiful sunset over a calm ocean", + "tooltip": "A prompt for the video generation", + }, + ) + + actual_output = model_field_to_node_input(IO.STRING, ModelWithStringField, "prompt") + + assert actual_output[0] == expected_output[0] + assert actual_output[1] == expected_output[1] + + +def test_model_field_to_string_input_multiline(): + """Tests mapping a string field.""" + + class ModelWithStringField(BaseModel): + prompt: Optional[str] = Field( + default="A beautiful sunset over a calm ocean", + description="A prompt for the video generation", + ) + + expected_output = ( + IO.STRING, + { + "default": "A beautiful sunset over a calm ocean", + "tooltip": "A prompt for the video generation", + "multiline": True, + }, + ) + + actual_output = model_field_to_node_input( + IO.STRING, ModelWithStringField, "prompt", multiline=True + ) + + assert actual_output[0] == expected_output[0] + assert actual_output[1] == expected_output[1] + + +def test_model_field_to_combo_input(): + """Tests mapping a combo field.""" + + class MockEnum(str, Enum): + option_1 = "option 1" + option_2 = "option 2" + option_3 = "option 3" + + class ModelWithComboField(BaseModel): + model_name: Optional[MockEnum] = Field("option 1", description="Model Name") + + expected_output = ( + IO.COMBO, + { + "options": ["option 1", "option 2", "option 3"], + "default": "option 1", + "tooltip": "Model Name", + }, + ) + + actual_output = model_field_to_node_input( + IO.COMBO, ModelWithComboField, "model_name", enum_type=MockEnum + ) + + assert actual_output[0] == expected_output[0] + assert actual_output[1] == expected_output[1] + + +def test_model_field_to_combo_input_no_options(): + """Tests mapping a combo field with no options.""" + + class ModelWithComboField(BaseModel): + model_name: Optional[str] = Field(description="Model Name") + + expected_output = ( + IO.COMBO, + { + "tooltip": "Model Name", + }, + ) + + actual_output = model_field_to_node_input( + IO.COMBO, ModelWithComboField, "model_name" + ) + + assert actual_output[0] == expected_output[0] + assert actual_output[1] == expected_output[1] + + +def test_model_field_to_image_input(): + """Tests mapping an image field.""" + + class ModelWithImageField(BaseModel): + image: Optional[str] = Field( + default=None, + description="An image for the video generation", + ) + + expected_output = ( + IO.IMAGE, + { + "default": None, + "tooltip": "An image for the video generation", + }, + ) + + actual_output = model_field_to_node_input(IO.IMAGE, ModelWithImageField, "image") + + assert actual_output[0] == expected_output[0] + assert actual_output[1] == expected_output[1] + + +def test_model_field_to_node_input_no_description(): + """Tests mapping a field with no description.""" + + class ModelWithNoDescriptionField(BaseModel): + field: Optional[str] = Field(default="default value") + + expected_output = ( + IO.STRING, + { + "default": "default value", + }, + ) + + actual_output = model_field_to_node_input( + IO.STRING, ModelWithNoDescriptionField, "field" + ) + + assert actual_output[0] == expected_output[0] + assert actual_output[1] == expected_output[1] + + +def test_model_field_to_node_input_no_default(): + """Tests mapping a field with no default.""" + + class ModelWithNoDefaultField(BaseModel): + field: Optional[str] = Field(description="A field with no default") + + expected_output = ( + IO.STRING, + { + "tooltip": "A field with no default", + }, + ) + + actual_output = model_field_to_node_input( + IO.STRING, ModelWithNoDefaultField, "field" + ) + + assert actual_output[0] == expected_output[0] + assert actual_output[1] == expected_output[1] + + +def test_model_field_to_node_input_no_metadata(): + """Tests mapping a field with no metadata or properties defined on the schema.""" + + class ModelWithNoMetadataField(BaseModel): + field: Optional[str] = Field() + + expected_output = ( + IO.STRING, + {}, + ) + + actual_output = model_field_to_node_input( + IO.STRING, ModelWithNoMetadataField, "field" + ) + + assert actual_output[0] == expected_output[0] + assert actual_output[1] == expected_output[1] + + +def test_model_field_to_node_input_default_is_none(): + """ + Tests mapping a field with a default of `None`. + I.e., the default field should be included as the schema explicitly sets it to `None`. + """ + + class ModelWithNoneDefaultField(BaseModel): + field: Optional[str] = Field( + default=None, description="A field with a default of None" + ) + + expected_output = ( + IO.STRING, + { + "default": None, + "tooltip": "A field with a default of None", + }, + ) + + actual_output = model_field_to_node_input( + IO.STRING, ModelWithNoneDefaultField, "field" + ) + + assert actual_output[0] == expected_output[0] + assert actual_output[1] == expected_output[1] diff --git a/tests-unit/comfy_api_test/input_impl_test.py b/tests-unit/comfy_api_test/input_impl_test.py new file mode 100644 index 000000000..5fc21a9a7 --- /dev/null +++ b/tests-unit/comfy_api_test/input_impl_test.py @@ -0,0 +1,91 @@ +import io +from comfy_api.input_impl.video_types import ( + container_to_output_format, + get_open_write_kwargs, +) +from comfy_api.util import VideoContainer + + +def test_container_to_output_format_empty_string(): + """Test that an empty string input returns None. `None` arg allows default auto-detection.""" + assert container_to_output_format("") is None + + +def test_container_to_output_format_none(): + """Test that None input returns None.""" + assert container_to_output_format(None) is None + + +def test_container_to_output_format_comma_separated(): + """Test that a comma-separated list returns a valid singular format from the list.""" + comma_separated_format = "mp4,mov,m4a" + output_format = container_to_output_format(comma_separated_format) + assert output_format in comma_separated_format + + +def test_container_to_output_format_single(): + """Test that a single format string (not comma-separated list) is returned as is.""" + assert container_to_output_format("mp4") == "mp4" + + +def test_get_open_write_kwargs_filepath_no_format(): + """Test that 'format' kwarg is NOT set when dest is a file path.""" + kwargs_auto = get_open_write_kwargs("output.mp4", "mp4", VideoContainer.AUTO) + assert "format" not in kwargs_auto, "Format should not be set for file paths (AUTO)" + + kwargs_specific = get_open_write_kwargs("output.avi", "mp4", "avi") + fail_msg = "Format should not be set for file paths (Specific)" + assert "format" not in kwargs_specific, fail_msg + + +def test_get_open_write_kwargs_base_options_mode(): + """Test basic kwargs for file path: mode and movflags.""" + kwargs = get_open_write_kwargs("output.mp4", "mp4", VideoContainer.AUTO) + assert kwargs["mode"] == "w", "mode should be set to write" + + fail_msg = "movflags should be set to preserve custom metadata tags" + assert "movflags" in kwargs["options"], fail_msg + assert kwargs["options"]["movflags"] == "use_metadata_tags", fail_msg + + +def test_get_open_write_kwargs_bytesio_auto_format(): + """Test kwargs for BytesIO dest with AUTO format.""" + dest = io.BytesIO() + container_fmt = "mov,mp4,m4a" + kwargs = get_open_write_kwargs(dest, container_fmt, VideoContainer.AUTO) + + assert kwargs["mode"] == "w" + assert kwargs["options"]["movflags"] == "use_metadata_tags" + + fail_msg = ( + "Format should be a valid format from the container's format list when AUTO" + ) + assert kwargs["format"] in container_fmt, fail_msg + + +def test_get_open_write_kwargs_bytesio_specific_format(): + """Test kwargs for BytesIO dest with a specific single format.""" + dest = io.BytesIO() + container_fmt = "avi" + to_fmt = VideoContainer.MP4 + kwargs = get_open_write_kwargs(dest, container_fmt, to_fmt) + + assert kwargs["mode"] == "w" + assert kwargs["options"]["movflags"] == "use_metadata_tags" + + fail_msg = "Format should be the specified format (lowercased) when output format is not AUTO" + assert kwargs["format"] == "mp4", fail_msg + + +def test_get_open_write_kwargs_bytesio_specific_format_list(): + """Test kwargs for BytesIO dest with a specific comma-separated format.""" + dest = io.BytesIO() + container_fmt = "avi" + to_fmt = "mov,mp4,m4a" # A format string that is a list + kwargs = get_open_write_kwargs(dest, container_fmt, to_fmt) + + assert kwargs["mode"] == "w" + assert kwargs["options"]["movflags"] == "use_metadata_tags" + + fail_msg = "Format should be a valid format from the specified format list when output format is not AUTO" + assert kwargs["format"] in to_fmt, fail_msg diff --git a/tests-unit/comfy_api_test/video_types_test.py b/tests-unit/comfy_api_test/video_types_test.py new file mode 100644 index 000000000..b25fcb1ca --- /dev/null +++ b/tests-unit/comfy_api_test/video_types_test.py @@ -0,0 +1,239 @@ +import pytest +import torch +import tempfile +import os +import av +import io +from fractions import Fraction +from comfy_api.input_impl.video_types import VideoFromFile, VideoFromComponents +from comfy_api.util.video_types import VideoComponents +from comfy_api.input.basic_types import AudioInput +from av.error import InvalidDataError + +EPSILON = 0.0001 + + +@pytest.fixture +def sample_images(): + """3-frame 2x2 RGB video tensor""" + return torch.rand(3, 2, 2, 3) + + +@pytest.fixture +def sample_audio(): + """Stereo audio with 44.1kHz sample rate""" + return AudioInput( + { + "waveform": torch.rand(1, 2, 1000), + "sample_rate": 44100, + } + ) + + +@pytest.fixture +def video_components(sample_images, sample_audio): + """VideoComponents with images, audio, and metadata""" + return VideoComponents( + images=sample_images, + audio=sample_audio, + frame_rate=Fraction(30), + metadata={"test": "metadata"}, + ) + + +def create_test_video(width=4, height=4, frames=3, fps=30): + """Helper to create a temporary video file""" + tmp = tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) + with av.open(tmp.name, mode="w") as container: + stream = container.add_stream("h264", rate=fps) + stream.width = width + stream.height = height + stream.pix_fmt = "yuv420p" + + for i in range(frames): + frame = av.VideoFrame.from_ndarray( + torch.ones(height, width, 3, dtype=torch.uint8).numpy() * (i * 85), + format="rgb24", + ) + frame = frame.reformat(format="yuv420p") + packet = stream.encode(frame) + container.mux(packet) + + # Flush + packet = stream.encode(None) + container.mux(packet) + + return tmp.name + + +@pytest.fixture +def simple_video_file(): + """4x4 video with 3 frames at 30fps""" + file_path = create_test_video() + yield file_path + os.unlink(file_path) + + +def test_video_from_components_get_duration(video_components): + """Duration calculated correctly from frame count and frame rate""" + video = VideoFromComponents(video_components) + duration = video.get_duration() + + expected_duration = 3.0 / 30.0 + assert duration == pytest.approx(expected_duration) + + +def test_video_from_components_get_duration_different_frame_rates(sample_images): + """Duration correct for different frame rates including fractional""" + # Test with 60 fps + components_60fps = VideoComponents(images=sample_images, frame_rate=Fraction(60)) + video_60fps = VideoFromComponents(components_60fps) + assert video_60fps.get_duration() == pytest.approx(3.0 / 60.0) + + # Test with fractional frame rate (23.976fps) + components_frac = VideoComponents( + images=sample_images, frame_rate=Fraction(24000, 1001) + ) + video_frac = VideoFromComponents(components_frac) + expected_frac = 3.0 / (24000.0 / 1001.0) + assert video_frac.get_duration() == pytest.approx(expected_frac) + + +def test_video_from_components_get_duration_empty_video(): + """Duration is zero for empty video""" + empty_components = VideoComponents( + images=torch.zeros(0, 2, 2, 3), frame_rate=Fraction(30) + ) + video = VideoFromComponents(empty_components) + assert video.get_duration() == 0.0 + + +def test_video_from_components_get_dimensions(video_components): + """Dimensions returned correctly from image tensor shape""" + video = VideoFromComponents(video_components) + width, height = video.get_dimensions() + assert width == 2 + assert height == 2 + + +def test_video_from_file_get_duration(simple_video_file): + """Duration extracted from file metadata""" + video = VideoFromFile(simple_video_file) + duration = video.get_duration() + assert duration == pytest.approx(0.1, abs=0.01) + + +def test_video_from_file_get_dimensions(simple_video_file): + """Dimensions read from stream without decoding frames""" + video = VideoFromFile(simple_video_file) + width, height = video.get_dimensions() + assert width == 4 + assert height == 4 + + +def test_video_from_file_bytesio_input(): + """VideoFromFile works with BytesIO input""" + buffer = io.BytesIO() + with av.open(buffer, mode="w", format="mp4") as container: + stream = container.add_stream("h264", rate=30) + stream.width = 2 + stream.height = 2 + stream.pix_fmt = "yuv420p" + + frame = av.VideoFrame.from_ndarray( + torch.zeros(2, 2, 3, dtype=torch.uint8).numpy(), format="rgb24" + ) + frame = frame.reformat(format="yuv420p") + packet = stream.encode(frame) + container.mux(packet) + packet = stream.encode(None) + container.mux(packet) + + buffer.seek(0) + video = VideoFromFile(buffer) + + assert video.get_dimensions() == (2, 2) + assert video.get_duration() == pytest.approx(1 / 30, abs=0.01) + + +def test_video_from_file_invalid_file_error(): + """InvalidDataError raised for non-video files""" + with tempfile.NamedTemporaryFile(suffix=".txt", delete=False) as tmp: + tmp.write(b"not a video file") + tmp.flush() + tmp_name = tmp.name + + try: + with pytest.raises(InvalidDataError): + video = VideoFromFile(tmp_name) + video.get_dimensions() + finally: + os.unlink(tmp_name) + + +def test_video_from_file_audio_only_error(): + """ValueError raised for audio-only files""" + with tempfile.NamedTemporaryFile(suffix=".m4a", delete=False) as tmp: + tmp_name = tmp.name + + try: + with av.open(tmp_name, mode="w") as container: + stream = container.add_stream("aac", rate=44100) + stream.sample_rate = 44100 + stream.format = "fltp" + + audio_data = torch.zeros(1, 1024).numpy() + audio_frame = av.AudioFrame.from_ndarray( + audio_data, format="fltp", layout="mono" + ) + audio_frame.sample_rate = 44100 + audio_frame.pts = 0 + packet = stream.encode(audio_frame) + container.mux(packet) + + for packet in stream.encode(None): + container.mux(packet) + + with pytest.raises(ValueError, match="No video stream found"): + video = VideoFromFile(tmp_name) + video.get_dimensions() + finally: + os.unlink(tmp_name) + + +def test_single_frame_video(): + """Single frame video has correct duration""" + components = VideoComponents( + images=torch.rand(1, 10, 10, 3), frame_rate=Fraction(1) + ) + video = VideoFromComponents(components) + assert video.get_duration() == 1.0 + + +@pytest.mark.parametrize( + "frame_rate,expected_fps", + [ + (Fraction(24000, 1001), 24000 / 1001), + (Fraction(30000, 1001), 30000 / 1001), + (Fraction(25, 1), 25.0), + (Fraction(50, 2), 25.0), + ], +) +def test_fractional_frame_rates(frame_rate, expected_fps): + """Duration calculated correctly for various fractional frame rates""" + components = VideoComponents(images=torch.rand(100, 4, 4, 3), frame_rate=frame_rate) + video = VideoFromComponents(components) + duration = video.get_duration() + expected_duration = 100.0 / expected_fps + assert duration == pytest.approx(expected_duration) + + +def test_duration_consistency(video_components): + """get_duration() consistent with manual calculation from components""" + video = VideoFromComponents(video_components) + + duration = video.get_duration() + components = video.get_components() + manual_duration = float(components.images.shape[0] / components.frame_rate) + + assert duration == pytest.approx(manual_duration) diff --git a/tests-unit/comfy_extras_test/__init__.py b/tests-unit/comfy_extras_test/__init__.py new file mode 100644 index 000000000..e69de29bb diff --git a/tests-unit/comfy_extras_test/image_stitch_test.py b/tests-unit/comfy_extras_test/image_stitch_test.py new file mode 100644 index 000000000..b5a0f022c --- /dev/null +++ b/tests-unit/comfy_extras_test/image_stitch_test.py @@ -0,0 +1,243 @@ +import torch +from unittest.mock import patch, MagicMock + +# Mock nodes module to prevent CUDA initialization during import +mock_nodes = MagicMock() +mock_nodes.MAX_RESOLUTION = 16384 + +# Mock server module for PromptServer +mock_server = MagicMock() + +with patch.dict('sys.modules', {'nodes': mock_nodes, 'server': mock_server}): + from comfy_extras.nodes_images import ImageStitch + + +class TestImageStitch: + + def create_test_image(self, batch_size=1, height=64, width=64, channels=3): + """Helper to create test images with specific dimensions""" + return torch.rand(batch_size, height, width, channels) + + def test_no_image2_passthrough(self): + """Test that when image2 is None, image1 is returned unchanged""" + node = ImageStitch() + image1 = self.create_test_image() + + result = node.stitch(image1, "right", True, 0, "white", image2=None) + + assert len(result) == 1 + assert torch.equal(result[0], image1) + + def test_basic_horizontal_stitch_right(self): + """Test basic horizontal stitching to the right""" + node = ImageStitch() + image1 = self.create_test_image(height=32, width=32) + image2 = self.create_test_image(height=32, width=24) + + result = node.stitch(image1, "right", False, 0, "white", image2) + + assert result[0].shape == (1, 32, 56, 3) # 32 + 24 width + + def test_basic_horizontal_stitch_left(self): + """Test basic horizontal stitching to the left""" + node = ImageStitch() + image1 = self.create_test_image(height=32, width=32) + image2 = self.create_test_image(height=32, width=24) + + result = node.stitch(image1, "left", False, 0, "white", image2) + + assert result[0].shape == (1, 32, 56, 3) # 24 + 32 width + + def test_basic_vertical_stitch_down(self): + """Test basic vertical stitching downward""" + node = ImageStitch() + image1 = self.create_test_image(height=32, width=32) + image2 = self.create_test_image(height=24, width=32) + + result = node.stitch(image1, "down", False, 0, "white", image2) + + assert result[0].shape == (1, 56, 32, 3) # 32 + 24 height + + def test_basic_vertical_stitch_up(self): + """Test basic vertical stitching upward""" + node = ImageStitch() + image1 = self.create_test_image(height=32, width=32) + image2 = self.create_test_image(height=24, width=32) + + result = node.stitch(image1, "up", False, 0, "white", image2) + + assert result[0].shape == (1, 56, 32, 3) # 24 + 32 height + + def test_size_matching_horizontal(self): + """Test size matching for horizontal concatenation""" + node = ImageStitch() + image1 = self.create_test_image(height=64, width=64) + image2 = self.create_test_image(height=32, width=32) # Different aspect ratio + + result = node.stitch(image1, "right", True, 0, "white", image2) + + # image2 should be resized to match image1's height (64) with preserved aspect ratio + expected_width = 64 + 64 # original + resized (32*64/32 = 64) + assert result[0].shape == (1, 64, expected_width, 3) + + def test_size_matching_vertical(self): + """Test size matching for vertical concatenation""" + node = ImageStitch() + image1 = self.create_test_image(height=64, width=64) + image2 = self.create_test_image(height=32, width=32) + + result = node.stitch(image1, "down", True, 0, "white", image2) + + # image2 should be resized to match image1's width (64) with preserved aspect ratio + expected_height = 64 + 64 # original + resized (32*64/32 = 64) + assert result[0].shape == (1, expected_height, 64, 3) + + def test_padding_for_mismatched_heights_horizontal(self): + """Test padding when heights don't match in horizontal concatenation""" + node = ImageStitch() + image1 = self.create_test_image(height=64, width=32) + image2 = self.create_test_image(height=48, width=24) # Shorter height + + result = node.stitch(image1, "right", False, 0, "white", image2) + + # Both images should be padded to height 64 + assert result[0].shape == (1, 64, 56, 3) # 32 + 24 width, max(64,48) height + + def test_padding_for_mismatched_widths_vertical(self): + """Test padding when widths don't match in vertical concatenation""" + node = ImageStitch() + image1 = self.create_test_image(height=32, width=64) + image2 = self.create_test_image(height=24, width=48) # Narrower width + + result = node.stitch(image1, "down", False, 0, "white", image2) + + # Both images should be padded to width 64 + assert result[0].shape == (1, 56, 64, 3) # 32 + 24 height, max(64,48) width + + def test_spacing_horizontal(self): + """Test spacing addition in horizontal concatenation""" + node = ImageStitch() + image1 = self.create_test_image(height=32, width=32) + image2 = self.create_test_image(height=32, width=24) + spacing_width = 16 + + result = node.stitch(image1, "right", False, spacing_width, "white", image2) + + # Expected width: 32 + 16 (spacing) + 24 = 72 + assert result[0].shape == (1, 32, 72, 3) + + def test_spacing_vertical(self): + """Test spacing addition in vertical concatenation""" + node = ImageStitch() + image1 = self.create_test_image(height=32, width=32) + image2 = self.create_test_image(height=24, width=32) + spacing_width = 16 + + result = node.stitch(image1, "down", False, spacing_width, "white", image2) + + # Expected height: 32 + 16 (spacing) + 24 = 72 + assert result[0].shape == (1, 72, 32, 3) + + def test_spacing_color_values(self): + """Test that spacing colors are applied correctly""" + node = ImageStitch() + image1 = self.create_test_image(height=32, width=32) + image2 = self.create_test_image(height=32, width=32) + + # Test white spacing + result_white = node.stitch(image1, "right", False, 16, "white", image2) + # Check that spacing region contains white values (close to 1.0) + spacing_region = result_white[0][:, :, 32:48, :] # Middle 16 pixels + assert torch.all(spacing_region >= 0.9) # Should be close to white + + # Test black spacing + result_black = node.stitch(image1, "right", False, 16, "black", image2) + spacing_region = result_black[0][:, :, 32:48, :] + assert torch.all(spacing_region <= 0.1) # Should be close to black + + def test_odd_spacing_width_made_even(self): + """Test that odd spacing widths are made even""" + node = ImageStitch() + image1 = self.create_test_image(height=32, width=32) + image2 = self.create_test_image(height=32, width=32) + + # Use odd spacing width + result = node.stitch(image1, "right", False, 15, "white", image2) + + # Should be made even (16), so total width = 32 + 16 + 32 = 80 + assert result[0].shape == (1, 32, 80, 3) + + def test_batch_size_matching(self): + """Test that different batch sizes are handled correctly""" + node = ImageStitch() + image1 = self.create_test_image(batch_size=2, height=32, width=32) + image2 = self.create_test_image(batch_size=1, height=32, width=32) + + result = node.stitch(image1, "right", False, 0, "white", image2) + + # Should match larger batch size + assert result[0].shape == (2, 32, 64, 3) + + def test_channel_matching_rgb_to_rgba(self): + """Test that channel differences are handled (RGB + alpha)""" + node = ImageStitch() + image1 = self.create_test_image(channels=3) # RGB + image2 = self.create_test_image(channels=4) # RGBA + + result = node.stitch(image1, "right", False, 0, "white", image2) + + # Should have 4 channels (RGBA) + assert result[0].shape[-1] == 4 + + def test_channel_matching_rgba_to_rgb(self): + """Test that channel differences are handled (RGBA + RGB)""" + node = ImageStitch() + image1 = self.create_test_image(channels=4) # RGBA + image2 = self.create_test_image(channels=3) # RGB + + result = node.stitch(image1, "right", False, 0, "white", image2) + + # Should have 4 channels (RGBA) + assert result[0].shape[-1] == 4 + + def test_all_color_options(self): + """Test all available color options""" + node = ImageStitch() + image1 = self.create_test_image(height=32, width=32) + image2 = self.create_test_image(height=32, width=32) + + colors = ["white", "black", "red", "green", "blue"] + + for color in colors: + result = node.stitch(image1, "right", False, 16, color, image2) + assert result[0].shape == (1, 32, 80, 3) # Basic shape check + + def test_all_directions(self): + """Test all direction options""" + node = ImageStitch() + image1 = self.create_test_image(height=32, width=32) + image2 = self.create_test_image(height=32, width=32) + + directions = ["right", "left", "up", "down"] + + for direction in directions: + result = node.stitch(image1, direction, False, 0, "white", image2) + assert result[0].shape == (1, 32, 64, 3) if direction in ["right", "left"] else (1, 64, 32, 3) + + def test_batch_size_channel_spacing_integration(self): + """Test integration of batch matching, channel matching, size matching, and spacings""" + node = ImageStitch() + image1 = self.create_test_image(batch_size=2, height=64, width=48, channels=3) + image2 = self.create_test_image(batch_size=1, height=32, width=32, channels=4) + + result = node.stitch(image1, "right", True, 8, "red", image2) + + # Should handle: batch matching, size matching, channel matching, spacing + assert result[0].shape[0] == 2 # Batch size matched + assert result[0].shape[-1] == 4 # Channels matched to max + assert result[0].shape[1] == 64 # Height from image1 (size matching) + # Width should be: 48 + 8 (spacing) + resized_image2_width + expected_image2_width = int(64 * (32/32)) # Resized to height 64 + expected_total_width = 48 + 8 + expected_image2_width + assert result[0].shape[2] == expected_total_width + diff --git a/tests/unit/test_user_manager.py b/tests/unit/test_user_manager.py index 11b0f8ef4..cb0206d4d 100644 --- a/tests/unit/test_user_manager.py +++ b/tests/unit/test_user_manager.py @@ -231,3 +231,61 @@ async def test_move_userdata_full_info(aiohttp_client, app, tmp_path): assert not os.path.exists(tmp_path / "source.txt") with open(tmp_path / "dest.txt", "r") as f: assert f.read() == "test content" + + +async def test_listuserdata_v2_empty_root(aiohttp_client, app): + client = await aiohttp_client(app) + resp = await client.get("/v2/userdata") + assert resp.status == 200 + assert await resp.json() == [] + + +async def test_listuserdata_v2_nonexistent_subdirectory(aiohttp_client, app): + client = await aiohttp_client(app) + resp = await client.get("/v2/userdata?path=does_not_exist") + assert resp.status == 404 + + +async def test_listuserdata_v2_default(aiohttp_client, app, tmp_path): + os.makedirs(tmp_path / "test_dir" / "subdir") + (tmp_path / "test_dir" / "file1.txt").write_text("content") + (tmp_path / "test_dir" / "subdir" / "file2.txt").write_text("content") + + client = await aiohttp_client(app) + resp = await client.get("/v2/userdata?path=test_dir") + assert resp.status == 200 + data = await resp.json() + file_paths = {item["path"] for item in data if item["type"] == "file"} + assert file_paths == {"test_dir/file1.txt", "test_dir/subdir/file2.txt"} + + +async def test_listuserdata_v2_normalized_separators(aiohttp_client, app, tmp_path, monkeypatch): + # Force backslash as os separator + monkeypatch.setattr(os, 'sep', '\\') + monkeypatch.setattr(os.path, 'sep', '\\') + os.makedirs(tmp_path / "test_dir" / "subdir") + (tmp_path / "test_dir" / "subdir" / "file1.txt").write_text("x") + + client = await aiohttp_client(app) + resp = await client.get("/v2/userdata?path=test_dir") + assert resp.status == 200 + data = await resp.json() + for item in data: + assert "/" in item["path"] + assert "\\" not in item["path"]\ + +async def test_listuserdata_v2_url_encoded_path(aiohttp_client, app, tmp_path): + # Create a directory with a space in its name and a file inside + os.makedirs(tmp_path / "my dir") + (tmp_path / "my dir" / "file.txt").write_text("content") + + client = await aiohttp_client(app) + # Use URL-encoded space in path parameter + resp = await client.get("/v2/userdata?path=my%20dir&recurse=false") + assert resp.status == 200 + data = await resp.json() + assert len(data) == 1 + entry = data[0] + assert entry["name"] == "file.txt" + # Ensure the path is correctly decoded and uses forward slash + assert entry["path"] == "my dir/file.txt"