MiniMax H3 – Minimal I2V + Audio Workflow for ComfyUI

A compact MiniMax H3 image-to-video workflow for ComfyUI with synchronized audio generation.

This package contains a tested workflow JSON, documentation, and direct model download references.
Model weights are not included. Please download them from the original Comfy-Org repository.

Features

• Image-to-video from a single starting image

• Joint video + audio generation

• 124 frames at 24 fps (about 5.17 seconds)

• res_multistep sampler

• 20 steps

• MiniMax H3 sigma shift:

  • Video: 12

  • Audio: 3

• Standard ComfyUI VAE loaders

• Uses the official Comfy-Org Audio VAE with latent statistics included

• VideoHelperSuite output to H.264 MP4 with audio

Required model files

1. Diffusion model

minimax_h3_fl2va_pruned_int8_convrot.safetensors

Direct download:
Download diffusion model

Place in:
ComfyUI/models/diffusion_models/
or your configured diffusion_models folder.

2. Text encoder

qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors

Direct download:
Download text encoder

Place in:
ComfyUI/models/text_encoders/
or your configured text_encoders folder.

3. Video VAE

minimax_h3_video_vae_fp16.safetensors

Direct download:
Download video VAE

Place in:
ComfyUI/models/vae/
or your configured vae folder.

4. Audio VAE

minimax_h3_audio_vae_fp32.safetensors

Direct download:
Download audio VAE

Place in:
ComfyUI/models/vae/
or your configured vae folder.

Required custom node

• ComfyUI-VideoHelperSuite

MiniMax H3 model / conditioning nodes used by this workflow are ComfyUI core nodes in a ComfyUI build with MiniMax H3 support.

Sample prompt – English speech

A woman looks directly at the camera and speaks naturally in English: "Hello! I'm testing the audio generation. Can you hear my voice clearly?" Her voice is clear and easy to understand. Natural synchronized English speech, accurate lip movements, and subtle room ambience.

Workflow settings

• Resolution in included tested sample: 512 × 512

• Length: 124 frames

• Frame rate: 24 fps

• Sampler: res_multistep

• Scheduler: normal

• Steps: 20

• Denoise: 1.0

• Video sigma shift: 12

• Audio sigma shift: 3

You can change the prompt, source image, and resolution as needed.

Included files

• MiniMax_H3_I2V_test.json

  • Tested 512×512 English speech example

  • Source image connected to first_frame

• README.md

• CIVITAI_DESCRIPTION.md

• MODEL_DOWNLOADS.md

• MANIFEST.txt

Important note

Use the official Comfy-Org Audio VAE listed above.

A differently packaged Audio VAE that lacks latents_mean / latents_std, or uses unfused weight_g / weight_v decoder weights, may decode to:

• silence

• noise

• a constant tone

This package does not redistribute model weights.
Please follow the licenses and terms of the original model repositories.