MiniMax H3 ComfyUI Guide

MiniMax H3 ComfyUI at a Glance

Which MiniMax H3 ComfyUI Workflow Should You Use?

Choose the workflow from the type of input you already have, not from the visual style you want to create.

If you only create text-to-video or image-to-video generations, you do not need to download the Ref2VA diffusion model. If you only use reference-to-video, you need Ref2VA; FL2VA is only necessary if you also want T2V or I2V.

How to Install MiniMax H3 in ComfyUI

MiniMax H3 is supported through native ComfyUI workflows, so a basic H3 setup does not require a third-party H3 custom node.

1. Update ComfyUI to version 0.30.0 or later.
2. Open ComfyUI and go to Template Library.
3. Choose Video.
4. Search for MiniMax H3.
5. Choose MiniMax H3 T2V, MiniMax H3 I2V, or MiniMax H3 R2V.
6. Download the model files required by that workflow.
7. Place each model in the folder shown below.
8. Restart ComfyUI or refresh the available model list if the new files are not detected.
9. Load the workflow and confirm that every loader points to the correct H3 file.
10. Set the aspect ratio, megapixel target, duration, prompt, and references.
11. Queue the workflow.

MiniMax H3 T2V and I2V Model Files

Text-to-video and image-to-video use the same FL2VA diffusion model. The difference between T2V and I2V is the conditioning input, not a different set of H3 diffusion weights.

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.safetensors

If T2V or I2V cannot find the H3 diffusion model, first confirm that the FL2VA safetensors file is inside models/diffusion_models rather than models/checkpoints.

Want every H3 download in one place? See the Download MiniMax H3 guide.

MiniMax H3 Reference-to-Video Model Files

Reference-to-video uses a different diffusion model from T2V and I2V. Replace the FL2VA diffusion weights with Ref2VA when the workflow needs reference images, reference videos, reference audio, character identity, motion, camera movement, or voice conditioning.

ComfyUI/
└── models/
    ├── diffusion_models/
    │   └── minimax_h3_ref2va_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.safetensors

FL2VA vs Ref2VA

FL2VA answers "how should the video begin or end?" Ref2VA answers "what identity, style, motion, camera behavior, or voice should the new video borrow from reference material?"

Best Resolution Settings for MiniMax H3 in ComfyUI

ComfyUI's native H3 workflows use a Resolution Selector with Aspect Ratio, Megapixels, and Multiple controls.

MiniMax H3 Duration and Frame Rate

MiniMax H3 produces video at 24 FPS and supports output durations from 4 to 15 seconds. In the native ComfyUI workflow, the duration control maps the requested duration onto H3's internal 17-frame-per-block grid.

For normal users, choose the intended duration in the workflow UI and let the native H3 node handle the frame-grid conversion. Do not manually calculate frame counts unless you are modifying the workflow at node level.

Can MiniMax H3 Run 2K Completely Locally?

Not with the complete official MiniMax H3 2K pipeline as currently released. The full H3 system has three stages: H3-Context-IR, H3-Base, and H3-Regenerate-2K. H3-Base is the open-weight generation model and produces the 768p base result. H3-Context-IR is a hosted preprocessing and orchestration system. H3-Regenerate-2K feeds the base result and original context back into H3 to regenerate a 2K result, and MiniMax states that this regeneration module is not yet open-sourced.

This means that running H3-Base locally in ComfyUI is not the same as reproducing MiniMax's complete official 2K pipeline offline. The current open-weight ComfyUI path should be treated as the local H3-Base workflow, while the complete official 2K validation workflow still uses MiniMax-hosted components or APIs.

Complete MiniMax H3 pipeline

Multimodal input
        ↓
H3-Context-IR
        ↓
H3-Base
        ↓
768p audiovisual result
        ↓
H3-Regenerate-2K
        ↓
2K audiovisual result

This distinction matters when comparing a local H3 result with a 2K result produced through MiniMax's hosted H3 workflow.

MiniMax H3 Reference Limits

Reference audio cannot be the only input to Ref2VA. MiniMax requires audio references to be accompanied by image or video input.

These are H3 model limits. A third-party web interface may enforce lower upload limits.

How to Assign References in H3 R2V

H3 reference inputs are identified in connection order with labels such as <Picture 1>, <Video 1>, and <Audio 1>. Do not simply attach references and expect H3 to infer the role of every file. State what each reference should control.

<Picture 1> → character identity
<Picture 2> → clothing and visual style
<Video 1>   → body motion and camera movement
<Audio 1>   → voice timbre

Use <Picture 1> for the main character's facial identity and hairstyle. Use <Video 1> only for the walking motion and handheld camera movement. Use <Audio 1> as the voice-timbre reference. Keep the new scene set inside a rainy neon-lit train station.

ref_image_size: match vs max

The native R2V workflow exposes a ref_image_size choice. Use match when you want reference images scaled toward the generation resolution for faster processing. Use max when identity detail matters more: ComfyUI's H3 documentation states that max can preserve reference images up to a 2048-pixel short edge, at the cost of additional processing.

Optional: Speed Up H3 with Sage Attention

Sage Attention is optional. The standard MiniMax H3 ComfyUI templates work without it.

ComfyUI's official H3 documentation states that Sage Attention can roughly double generation speed with minimal quality loss in its workflow setup. The improvement is not guaranteed on every machine, and H3 layers using unsupported dtypes can automatically fall back to standard PyTorch attention.

Do not install Sage Attention until the standard H3 workflow works correctly. First confirm that the model files, folders, and native workflow are functioning; then optimize performance.

Common MiniMax H3 ComfyUI Problems

MiniMax H3 ComfyUI vs Online Generator

Use ComfyUI when you want direct control over the open-weight H3-Base workflow and node graph. Use an online generator when you want to generate without managing model files, Python environments, CUDA dependencies, or local model storage.

MiniMax H3 ComfyUI FAQ

Related MiniMax H3 Guides

Official Sources