Start from a native template
Update ComfyUI, restart it, then open the template library and search for MiniMax H3. Choose T2V for a text-only test, I2V when one or two keyframes control the clip, or R2V when references control identity, style, motion, camera, or voice. The templates are a safer first run than assembling the graph from memory because they connect H3-specific video and audio components correctly.
- T2V and I2V share the FL2VA diffusion weights.
- R2V uses the separate Ref2VA diffusion weights.
- All three workflows require the text encoder, visual VAE, and audio VAE.
Check the model folders
The current ComfyUI templates use pruned INT8 H3 diffusion weights and a quantized Qwen3-VL encoder distributed through the Comfy-Org model repository. These filenames differ from the original BF16 checkpoint layout on the MiniMax repository.
ComfyUI/
└── models/
├── diffusion_models/
│ └── minimax_h3_fl2va_pruned_int8_convrot.safetensors
├── text_encoders/
│ └── qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
└── vae/
├── minimax_h3_video_vae_fp16.safetensors
└── minimax_h3_audio_vae_fp32.safetensorsUse a conservative first run
Keep the default fast preview resolution for the first request. Use a short text-only prompt, one output, and the template step count. Confirm that ComfyUI produces a playable MP4 with synchronized audio before increasing dimensions or duration.
- Keep width and height on the 32-pixel grid used by H3.
- After the preview succeeds, target about 1.0 megapixel at 16:9, roughly 1344 x 768.
- H3 duration follows its frame-block grid at 24 fps, so the rendered length may not equal a hand-entered decimal exactly.
- Watch the console during model swaps; low-VRAM systems may spend more time moving weights than denoising.
Add image or reference conditioning
For I2V, connect a first frame, a last frame, or both to the MiniMax H3 image-to-video node. For R2V, connect materials in a stable order and name them by the same one-based order inside the prompt. Give each item one explicit job.
| Input | Prompt role example | Common failure |
|---|---|---|
| Picture 1 | Owns character identity | Competes with another face reference |
| Video 1 | Owns movement and camera rhythm | Accidentally becomes the visual style |
| Audio 1 | Owns the speaking voice | Dialogue is too long for the shot |
Speed tuning comes after a baseline
ComfyUI documents an optional SageAttention path through either a KJNodes patch node or the global --use-sage-attention flag. Install only a wheel that matches the active Python, PyTorch, CUDA, and GPU architecture. Save a baseline clip before enabling it, then compare generation time, faces, fine motion, text, and stereo audio with the same seed.