Step 1 – Update ComfyUI
Use a current ComfyUI build compatible with current H3 support.
Installation and support
MiniMax H3 ComfyUI Setup, VRAM & Troubleshooting Guide
MiniMax H3 can run locally in current ComfyUI workflows for text-to-video, image-to-video and reference-based generation. Local setup requires the correct H3 model files, a compatible ComfyUI build and enough GPU and system memory for the selected workflow.
This guide focuses on installation, model placement, hardware constraints and troubleshooting. For ready-to-import JSON files, use MiniMax H3 workflow JSON downloads.
Last reviewed: September 15, 2026
Can MiniMax H3 Run in ComfyUI?
Yes. MiniMax H3 can run locally through current ComfyUI-compatible workflows for text-to-video, image-to-video and reference-based generation.
The main practical constraint is hardware. MiniMax does not publish one universal VRAM requirement for every H3 setup because memory use depends on model format, workflow, resolution, video length, system RAM and offloading. If you only need JSON files, use the MiniMax H3 workflow JSON downloads; if you need to make a local graph run, continue below.
How Do I Install MiniMax H3 in ComfyUI?
Install a current ComfyUI build, download the required H3 diffusion model, text encoder and video/audio VAE components, place them in the correct ComfyUI model directories, then import a compatible H3 workflow JSON. Hardware requirements depend on workflow, resolution, duration, model precision and offloading.
Where Do MiniMax H3 Models Go in ComfyUI?
ComfyUI/
└── models/
├── diffusion_models/
├── text_encoders/
└── vae/diffusion_models
Place the required MiniMax H3 diffusion model here.
text_encoders
Place the required H3-compatible text encoder here.
vae
Place the required video and audio VAE components here.
Exact filenames may change as official or community packages evolve. Verify the current model repository rather than relying on historical filenames. See the MiniMax H3 model guide for model-family definitions.
How to Install MiniMax H3 in ComfyUI
Step 2 – Choose Your Workflow Family
Choose T2V, I2V, FL2VA or R2V / Ref2VA from the separate workflow library. Do not download every graph by default. Download workflow JSON files from the MiniMax H3 Workflow Library.
Step 3 – Download the Required Models
Obtain the H3 diffusion model, compatible text encoder, video VAE and audio VAE required by the selected workflow.
Step 4 – Place Models in the Correct Folders
Place each component in its matching diffusion_models, text_encoders or vae directory.
Step 5 – Import the Workflow JSON
Open ComfyUI and load the JSON from the MiniMax H3 workflow library.
Step 6 – Resolve Missing Nodes
Use ComfyUI Manager or the appropriate package installation process, restart ComfyUI, and reload the graph.
Step 7 – Start at Conservative Settings
Start with lower resolution and shorter duration before increasing either setting.
Step 8 – Run and Monitor Memory
Observe GPU VRAM, system RAM and the VAE decoding stage during the first run.
Which Workflow Should I Start With?
| Input | Workflow |
|---|---|
| Text | T2V |
| One image | I2V |
| First + last frame | FL2VA |
| Multimodal references | R2V / Ref2VA |
Download all currently verified JSON files in the MiniMax H3 Workflow Library.
MiniMax H3 ComfyUI Troubleshooting
Configuration-dependent
Out of Memory During Sampling
Likely causes include insufficient VRAM, high resolution, long duration or a high-precision model.
First actions
- Lower resolution
- Shorten video duration
- Close other GPU workloads
Then consider
- Use lower precision where supported
- Try a verified quantized build
- Enable CPU offloading
Configuration-dependent
Text Encoder OOM
The text encoder may create memory pressure before generation starts.
First actions
- CPU-offload the encoder where supported
- Use lower precision where supported
- Reduce other GPU memory use
Configuration-dependent
VAE Decode OOM
Sampling can finish successfully and the job can still fail while decoding video.
First actions
- Reduce output resolution
- Shorten duration
- Increase system-RAM or offloading capacity
Setup issue
Missing Nodes
The workflow may require a newer ComfyUI build or node package.
First actions
- Update ComfyUI
- Open ComfyUI Manager
- Install the required node packs
- Restart and reload the workflow
Setup issue
Model Not Found
A model component is absent, placed in the wrong directory or not selected by the loader.
First actions
- Check diffusion_models, text_encoders and vae
- Verify filenames
- Check every workflow loader selection
Community-reported issue
Black or Dark Frames
Some users report dark output in particular workflow, precision, VAE or sampler combinations. This is not established as a universal H3 defect.
First actions
- Reproduce with a standard current graph
- Avoid changing several variables at once
- Verify the matching model family and VAE
Configuration-dependent
Very Slow Generation
Video diffusion is compute-intensive; CPU offloading, VRAM pressure, precision, resolution and duration can all add latency.
First actions
- Reduce resolution
- Shorten duration
- Measure a baseline before testing experimental acceleration
MiniMax H3 VRAM and Hardware Questions
These answers distinguish official requirements from MiniMax3 tests, community measurements and inference. For the primary hardware analysis, use MiniMax H3 local hardware requirements.
Unverified / evidence-scoped
How much VRAM does MiniMax H3 need?
No single official minimum covers every setup. Model format, workflow, resolution, duration, system RAM and offloading determine memory use.
Unverified / evidence-scoped
Can 12GB run H3?
Unverified as a general recommendation. Treat any successful 12GB report as configuration-specific community evidence, not an official requirement.
MiniMax3 tested
Is 16GB enough?
MiniMax3 has recorded a reproducible optimized 16GB configuration, but that result does not prove every 16GB GPU or workflow will run.
Inference
Can RTX 4090 run H3?
A 24GB RTX 4090 is a practical candidate, but this is an inference from capacity and community configurations rather than an official compatibility guarantee.
Inference
Can RTX 5090 run H3?
A 32GB RTX 5090 is a practical candidate, but workflow settings can still cause OOM and no universal performance figure applies.
Unverified / evidence-scoped
Can Apple Silicon run H3?
Do not assume parity with CUDA workflows. Current framework, precision and memory-support details must be checked for the exact Apple Silicon configuration.