Skip to content

COMFYUI_SETUP_AND_INFERENCE.md

Latest commit

 

History

History
82 lines (59 loc) · 2.9 KB

COMFYUI_SETUP_AND_INFERENCE.md

File metadata and controls

82 lines (59 loc) · 2.9 KB
Raw

ComfyUI Setup and Inference

Environment

ComfyUI 0.31.0 or newer is required. The example graphs use ComfyUI core nodes only, including the built-in MiniMax-H3 subgraph (MiniMaxH3ImageToVideo).

Required models

Download the base MiniMax-H3 components from Comfy-Org/MiniMax-H3 and place them in the corresponding ComfyUI model directories.

Download the ComfyUI LoRAs from lightx2v/Minimax-h3-Turbo into ComfyUI/models/loras/:

hf download lightx2v/Minimax-h3-Turbo \
  minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors \
  minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors \
  minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors \
  --local-dir ComfyUI/models/loras
ComfyUI/
└── models/
    └── loras/
        ├── minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors
        ├── minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors
        └── minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors

Example workflows

Ready-to-import graphs are listed in the main README. The T2VA and I2VA graphs default to FL2VA Turbo 8-step v1.0; the Ref2VA graph uses Ref2VA Turbo 4-step v0.1.

Inputs

On the subgraph, set lora_name, steps, shift_video, and shift_audio together to match the model table. Sampler defaults to euler.

Resolution. Width and height must be multiples of 32. Change the aspect ratio or megapixel value on Resolution Selector when another size is needed. The I2VA graph also includes an unused Use Image Size group (ImageScaleToTotalPixels → GetImageSize); wire those outputs to the subgraph width / height if you want the video canvas to follow the input image instead of the selector.

Duration. Defaults to 5 seconds. MiniMax-H3 runs at 24 FPS and the frame count must be 17 * n + 5, so a 5-second request is snapped to 124 frames (about 5.17 seconds).

Prompts. Use three sections:

integrated_multimodal_description: [Shot 1] Describe the visual style, subject, action, camera, lighting, and dialogue.

overall_soundscape: Describe dialogue, ambient sound, and synchronized sound effects.

non_diegetic_music: Describe the background score, or use N/A for no score.

Dialogue can be written as <d>[English] Dialogue text.</d>.

For I2VA / FL2VA, keep first-frame identity with <Picture 1> (and <Picture 2> if last_frame is connected):

integrated_multimodal_description: For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.

[Shot 1] Preserve the subject, clothing, and scene from <Picture 1>, then describe the motion.

Run

Drag a workflow JSON onto the ComfyUI canvas, set the inputs above, and queue it. Output is written under video/MiniMax_H3.