comfyui

Export the workflow from ComfyUI in API-format JSON, then pass it via workflow_json or workflow_path to comfyui_image or comfyui_video. You must also set output_node to the node ID of the terminal saver node, such as SaveImage or VHS_VideoCombine.

For 8GB-12GB GPUs, prefer lower-footprint options like Wan 2.1 1.3B, LTXV FP8, or Wan 2.2 GGUF/quantized community workflows. The bundled Wan 2.2 14B FP8 video workflows target 16GB-class hardware.

The workflow's loader nodes reference models not present locally. Read the structured missing_models payload for each file's name, role, destination hint, and download URL, then place files under ComfyUI/models/ subfolders like diffusion_models, text_encoders, vae, or loras.

No. The current tools do not inject LoRAs into arbitrary graphs. Use a workflow that already contains a LoraLoader or LoraLoaderModelOnly chain, and record each LoRA name with strength_model and strength_clip values.

Start ComfyUI locally or set COMFYUI_SERVER_URL to point at a running server; the default is http://localhost:8188. If a long render times out, check ComfyUI history before retrying since the server may have completed the prompt.