Local setup, models, workflows, and troubleshooting

MiniMax H3 ComfyUI Workflow & Setup Guide

Run MiniMax H3 locally with ComfyUI. Choose a T2V, I2V, or R2V workflow. Then install the required models and run your first test.

Download Workflows View Required Models

Comfy-Org T2V, I2V, and R2V workflows

Model files and folders

Local setup commands

Errors and VRAM guidance

On this page

Comfy-Org workflow templates

Download T2V, I2V, and R2V Workflow JSON Files

Comfy-Org provides MiniMax H3 workflow templates for Text-to-Video, Image-to-Video, and Reference-to-Video.

This guide covers the local Comfy-Org workflow. Settings and limits may differ from the online MiniMax H3 generator.

Download Workflows from the Command Line

If you use GitHub CLI, clone the repository and copy all three templates.

gh repo clone Comfy-Org/workflow_templates
cd workflow_templates/templates

cp video_minimax_h3_t2v.json ~/Downloads/
cp video_minimax_h3_i2v.json ~/Downloads/
cp video_minimax_h3_r2v.json ~/Downloads/

On Windows PowerShell:

Copy-Item video_minimax_h3_t2v.json "$HOME\Downloads\"
Copy-Item video_minimax_h3_i2v.json "$HOME\Downloads\"
Copy-Item video_minimax_h3_r2v.json "$HOME\Downloads\"
  1. 01Open ComfyUI.
  2. 02Open Workflows.
  3. 03Choose Open.
  4. 04Select the JSON file.

You can also drag the workflow JSON onto the ComfyUI canvas.

First local run

How to Run MiniMax H3 in ComfyUI

Already have ComfyUI installed? Skip directly to Required Model Files. For a new setup, follow this order.

1

Install or update ComfyUI.

2

Download the required MiniMax H3 model files.

3

Put each file in the correct model folder.

4

Open the T2V, I2V, or R2V workflow.

5

Add your prompt or reference files.

6

Start with a short 5-second test.

7

Queue the workflow.

8

Review the saved video.

The goal of your first run

ComfyUI starts
MiniMax H3 models load
Sampling begins
Video and audio decode
The final video is saved

Installation

Install or Update ComfyUI

Use comfy-cli for the managed setup, or install ComfyUI manually when you want to control the Python environment yourself.

Recommended Method: comfy-cli

The CLI can install and launch ComfyUI and manage the workspace for you.

pip install comfy-cli
comfy install
comfy launch

Launch the most recently installed workspace and check which workspace is active:

comfy --recent launch
comfy which

Manual ComfyUI Installation

Clone ComfyUI, install the PyTorch build for your hardware, install dependencies, and start the app.

gh repo clone Comfy-Org/ComfyUI
cd ComfyUI

# NVIDIA
pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130

# Intel Arc
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/xpu

pip install -r requirements.txt
python main.py

AMD and Apple Silicon require platform-specific PyTorch installation. Follow the current ComfyUI instructions for your hardware.

Update an Existing Installation

cd ComfyUI
git pull
pip install -r requirements.txt
python main.py

If you use comfy-cli, update the target workspace through the CLI instead. Update ComfyUI before troubleshooting missing MiniMax H3 nodes.

Model files

MiniMax H3 ComfyUI Model Download

T2V and I2V share the FL2VA diffusion checkpoint. Reference-to-Video uses a separate Ref2VA checkpoint, while the text encoder and VAE files remain the same.

Required Models for T2V and I2V

minimax_h3_fl2va_pruned_int8_convrot.safetensors

qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors

minimax_h3_video_vae_fp16.safetensors

minimax_h3_audio_vae_fp32.safetensors

Comfy-Org recommends int8_convrot diffusion weights when your PyTorch/CUDA environment supports them.

Folder Structure

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

The Comfy-Org repository uses these model directories.

Download T2V and I2V Models

python -m pip install -U huggingface_hub
cd ComfyUI

hf download Comfy-Org/MiniMax-H3 \
  --include "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" \
  "vae/minimax_h3_audio_vae_fp32.safetensors" \
  --local-dir models

R2V Uses a Different Diffusion Model

Reference-to-Video does not use the FL2VA checkpoint. Use minimax_h3_ref2va_pruned_int8_convrot.safetensors. The Qwen3-VL text encoder and both VAE files remain the same.

hf download Comfy-Org/MiniMax-H3 \
  --include "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" \
  "vae/minimax_h3_audio_vae_fp32.safetensors" \
  --local-dir models

Choose the right graph

Which MiniMax H3 ComfyUI Workflow Should You Use?

Start from the materials you already have and the type of control you need.

Text-to-Video

Choose T2V when you only have a prompt.

Text Prompt

Good for

  • •New scene concepts
  • •Cinematic shots
  • •Product ideas
  • •Game scenes
  • •Short social videos

Image-to-Video

Choose I2V when an image should define the starting visual state.

Image + Prompt

Good for

  • •Character animation
  • •Portrait movement
  • •Product animation
  • •Artwork animation
  • •Starting-frame control

Reference-to-Video

Choose R2V when different reference files should control different parts of the result.

Images + Video + Audio + Prompt

Good for

  • •Character identity
  • •Motion transfer
  • •Camera reference
  • •Visual style
  • •Speech and audio cues
  • •Motion timing
  • •Multi-asset scenes

Workflow tutorials

Run Text-to-Video, Image-to-Video, and Reference-to-Video

Load the matching JSON, confirm the checkpoint, add your inputs, and begin with a short test.

How to Run Text-to-Video

  1. 1. Open: video_minimax_h3_t2v.json
  2. 2. Confirm: FL2VA diffusion weights and the Qwen3-VL text encoder.
  3. 3. Enter a prompt: start simple.
  4. 4. Use a short duration: begin with 5 seconds.
  5. 5. Queue: confirm video and audio decode successfully.
A black running shoe sits on a reflective studio platform.

A narrow light moves slowly across the upper material.

The camera makes a slow push-in.

Keep the shoe centered.

Subtle studio ambience.

How to Run Image-to-Video

  1. 1. Open: video_minimax_h3_i2v.json
  2. 2. Upload: use a clear source image without several unrelated subjects.
  3. 3. Describe: explain the subject and camera movement.
  4. 4. Test: check identity, motion, camera, background stability, and audio.
The woman slowly turns toward the camera.

Her hair moves lightly in the wind.

The camera pushes forward at a steady speed.

Keep the movement natural and restrained.

How to Run Reference-to-Video

  1. 1. Open: video_minimax_h3_r2v.json
  2. 2. Add references: up to 9 images, 3 videos, and 3 separate audio files. Audio cannot be the only reference. Include at least one image or video.
  3. 3. Assign roles: identify what each reference controls.
  4. 4. Use numbered tags: follow the order in which references are connected.
<Picture 1>
<Video 1>
<Audio 1>
<Picture 1> defines the character's face, hair, and clothing.

Use the movement timing and body motion from <Video 1>.

Use the speech and audio cues from <Audio 1>.

The character walks through a neon-lit street while the camera tracks slowly from the side.

Avoid “Use all references.” Give each input one clear role. Ref2VA is sensitive to prompt wording.

Safe starting values

Reference Image Quality and First Test Settings

Confirm the base installation with a simple generation before increasing duration, reference count, or image detail.

Reference Image Quality

ref_image_size = match

Scales images closer to generation resolution. Use it for testing, faster processing, and lower reference detail needs.

ref_image_size = max

Keeps images up to a 2048px short edge. It adds more reference tokens and can increase processing cost.

For your first test, use match.

Recommended First Test

Duration: 5 seconds
Resolution: workflow default
References: none or one image
Camera: static or slow push-in
Main actions: one
Prompt length: short and clear

Do not start with a complicated 15-second reference-heavy workflow. First confirm that models load, sampling begins, and video and audio decode.

Local performance

MiniMax H3 ComfyUI Hardware, VRAM, and Output Limits

Memory use depends on the complete workflow. Begin with conservative settings and measure your own hardware before making performance assumptions.

How Much VRAM Does MiniMax H3 Need?

There is no single official VRAM number for every setup. VRAM use depends on your GPU, model precision, resolution, and video length. Reference count, model offloading, ComfyUI, and PyTorch versions also affect memory use.

If you have limited VRAM

  1. 1.Use the pruned INT8 diffusion model.
  2. 2.Start with a 5-second video.
  3. 3.Keep the workflow's default resolution.
  4. 4.Use fewer references.
  5. 5.Keep ref_image_size on match.
  6. 6.Close other GPU-heavy applications.
  7. 7.Confirm the base workflow before adding LoRAs.

Check System Information

Use comfy-cli to inspect system statistics and record your own test data.

comfy system-stats
GPU:
VRAM:
System RAM:
ComfyUI version:
PyTorch version:
Workflow:
Resolution:
Duration:
Generation time:

Can MiniMax H3 Generate 2K in ComfyUI?

MiniMax H3 supports output up to 2K in some workflows. The current Comfy-Org local workflow does not generate native 2K by default. Its base canvas uses a 768px short edge, with an area limit of about 768 × 1344.

Model support for 2K does not mean every local ComfyUI workflow generates native 2K.

Only describe a local workflow as native 2K after you verify that specific workflow.

Duration and Optional Turbo LoRA

The current Comfy-Org MiniMax H3 node generates video at 24 fps. It converts the selected duration into a valid frame count automatically.

Start with: 5 seconds

Base workflow
↓
Confirm generation works
↓
Test your prompt
↓
Add Turbo LoRA
↓
Compare quality and speed

Store compatible LoRAs under ComfyUI/models/loras/. Turbo LoRAs are optional and are not required for the base workflow.

Fix one layer at a time

Common MiniMax H3 ComfyUI Errors and Fixes

Update first, confirm the correct files and folders, and reproduce the issue with a short base workflow before changing several settings.

MiniMax H3 nodes are missing+

Your ComfyUI installation may be outdated. Update ComfyUI and its requirements before adding third-party replacements.

cd ComfyUI
git pull
pip install -r requirements.txt
python main.py
Model not found+

Check that diffusion models, text encoders, VAEs, and LoRAs are in their matching folders. Restart ComfyUI after moving files.

Diffusion → ComfyUI/models/diffusion_models/
Text encoder → ComfyUI/models/text_encoders/
Video VAE / Audio VAE → ComfyUI/models/vae/
LoRA → ComfyUI/models/loras/
Wrong model in R2V+

Do not use the FL2VA checkpoint for Reference-to-Video. The Comfy-Org R2V template uses separate Ref2VA weights.

Use: minimax_h3_ref2va_pruned_int8_convrot.safetensors
Reference image or motion is weak+

Use the correct numbered tag for each file. Give every reference one clear role, such as appearance, motion, camera, audio, or style.

<Picture 1> defines the character's appearance.
Use the body movement and timing from <Video 1>.
Keep the camera static.
Out of memory+

Reduce one layer of complexity at a time, then retry the workflow.

Shorter duration → Default resolution → Fewer references → ref_image_size = match → INT8 diffusion model
Torch not compiled with CUDA enabled+

Remove the incorrect PyTorch packages, then install the current GPU build recommended by ComfyUI for your hardware.

pip uninstall torch torchvision torchaudio
Custom node conflict+

Test the core installation without custom nodes. If H3 then works, troubleshoot custom packages separately.

python main.py --disable-all-custom-nodes

Recommended Troubleshooting Order

1.Update ComfyUI

2.Confirm the correct workflow

3.Confirm model filenames

4.Confirm model folders

5.Confirm FL2VA vs Ref2VA

6.Test a 5-second generation

7.Remove unnecessary references

8.Disable custom nodes

9.Check logs and system information

Do not change several settings at once.

Choose your setup

Local ComfyUI vs Online MiniMax H3

Choose based on how much workflow control you need and whether you want to manage large model files and GPU setup.

Local ComfyUI

  • •Runs on your own hardware
  • •Requires large model files
  • •Gives node-level workflow control
  • •Supports custom workflow experiments
  • •Performance depends on your hardware
  • •Requires installation and troubleshooting

Online Generator

  • •Runs through a browser
  • •No local model download
  • •Simpler setup
  • •Faster to start
  • •Generation runs on remote infrastructure
  • •No ComfyUI troubleshooting
Run MiniMax H3 Online

Choose ComfyUI for local control, node-level customization, model experiments, and custom pipelines. Choose the online generator when you mainly want to create videos without managing a local model environment.

The online generator uses a separate hosted workflow. Its resolution, pricing, and input options may differ from local ComfyUI.

Frequently asked questions

MiniMax H3 ComfyUI FAQ

Does MiniMax H3 work with ComfyUI?+

Yes. Comfy-Org provides MiniMax H3 T2V, I2V, and R2V workflow templates with ComfyUI-ready model files.

What MiniMax H3 workflows are available in ComfyUI?+

The current core templates include Text-to-Video, Image-to-Video, and Reference-to-Video.

Which model should I use for MiniMax H3 Text-to-Video?+

Use an FL2VA diffusion model with the T2V workflow. The current recommended quantized file is minimax_h3_fl2va_pruned_int8_convrot.safetensors.

Which model should I use for Reference-to-Video?+

Use minimax_h3_ref2va_pruned_int8_convrot.safetensors with the R2V workflow.

Do T2V and R2V use the same checkpoint?+

No. T2V and I2V use FL2VA weights, while R2V uses Ref2VA weights.

Can MiniMax H3 use reference images, videos, and audio in ComfyUI?+

Yes. The R2V workflow accepts images, videos, and separate audio files. Audio cannot be the only reference. Include at least one image or video.

How do I reference files in an R2V prompt?+

Use numbered tags such as <Picture 1>, <Video 1>, and <Audio 1>, then explain what each reference should guide.

How many references can I use?+

The current R2V workflow supports up to 9 images, 3 videos, and 3 audio files. Audio must be paired with at least one image or video.

Does the default local workflow generate native 2K?+

No, not by default. The current Comfy-Org local workflow uses a 768px short-edge canvas with an area limit of about 768 × 1344.

How much VRAM does MiniMax H3 need?+

There is no single official minimum for every setup. Usage depends on model precision, hardware, duration, resolution, reference count, and memory management.

Do I need custom nodes?+

The current Comfy-Org templates use ComfyUI core support. Update ComfyUI first if MiniMax H3 nodes are missing.

Where are generated videos saved?+

The Comfy-Org T2V workflow uses ComfyUI's SaveVideo node and a MiniMax H3 video filename prefix.

Local control or instant access

Run MiniMax H3 Your Way

Use ComfyUI when you want local control and custom workflows. Skip model downloads, GPU setup, and troubleshooting when you want to generate online.

Run MiniMax H3 Online Download Workflows