How to Install MiniMax H3 in ComfyUI on Windows

Install MiniMax H3 in ComfyUI on Windows to generate short videos with stereo audio from text, images or reference media. This guide covers the native templates, model folders, a first-generation check and the fixes to try when a model is missing or a workflow will not load.

Start with the native MiniMax H3 T2V template, download the files it requests and queue one short clip with turbo mode off. Once that works, use the I2V template for image inputs or the separate Ref2VA checkpoint for reference media.

Tested during editorial review: Updated 10 August 2026 against first-party documentation covering filenames, folders, modes, limits and versions. MiniMax and ComfyUI do not publish a universal consumer-GPU minimum; community low-VRAM results are labelled separately. Use the included generation check to record speed and VRAM use on your GPU.

The 30-second version

• Use the current stable ComfyUI release and a fresh native H3 template. H3 first appeared in v0.30.0; v0.31.0 added important fixes, but current templates include newer nodes and options.
• Choose the native MiniMax H3 T2V or I2V template for your first run.
• T2V and I2V use the fl2va diffusion checkpoint. R2V uses a different ref2va checkpoint.
• The listed T2V/I2V core files and 8-step LoRA total about 44.43 GB. Adding the Ref2VA checkpoint and its 4-step LoRA brings them to about 67.35 GB. Allow extra space for ComfyUI, temporary downloads, optional embeddings and outputs.
• Local H3-Base works on a 768-pixel-short-edge canvas. MiniMax’s complete 2K workflow also uses hosted modules that are not part of the open release.
• Read the community licence before downloading. It has meaningful territorial and use restrictions.

What MiniMax H3 can do locally

MiniMax describes H3 as an omni-modal generation system. It accepts text and, depending on the checkpoint, image, video and audio references. It produces 24fps video with 32kHz stereo audio. The published duration range is 4 to 15 seconds.

The open H3-Base checkpoint generates on a native canvas with a 768-pixel short edge, capped at 768 × 1344 and rounded to a multiple of 32. MiniMax’s headline “up to 2K” result comes from H3-Regenerate-2K; its full workflow also uses the hosted H3-Context-IR service. Neither module is included in the initial open-weight release. In plain English: local 768p is real, but fully local native 2K is not what this guide is promising.

Before you download anything

Licence warning: H3 uses the MiniMax H3 Community License Agreement, not Apache 2.0 or an OSI-approved open-source licence. As published on 2 August 2026, its applicable territory excludes the United States, European Union, United Kingdom and Republic of Korea unless separate permission is obtained. It also contains commercial, distribution, disclosure and acceptable-use conditions. Check the current agreement for your location and use case; this paragraph is not legal advice.

• A 64-bit Windows 10 or Windows 11 computer.
• Comfy Desktop or ComfyUI Portable. The portable build is the easiest folder layout to follow in this guide.
• A supported dedicated GPU. NVIDIA is the most established Windows route, although current ComfyUI Portable packages also exist for AMD and Intel hardware.
• A fast SSD with space for the selected model files plus ComfyUI, temporary downloads and generated video. The model totals below are not the complete disk requirement.
• A current graphics driver and enough system memory to support offloading.

Do not buy a graphics card because one post says “H3 needs exactly X GB.” Duration, resolution, checkpoint format, offloading, system RAM and optional custom nodes can all change the result. Start with the official pruned and quantised components below, then measure your own machine.

Step 1: Install or update ComfyUI

If you would rather let Claude Code or Cursor validate and run a known workflow, the Comfy MCP setup guide connects an AI client to this local ComfyUI installation without exposing it publicly.

Update ComfyUI before opening a fresh H3 template. Version 0.31.0 fixed several early H3 problems, but it is not a compatibility guarantee for today’s templates. The Comfy-Org model card recommends int8_convrot when PyTorch cu130 is available; it lists fp8_scaled diffusion checkpoints as the fallback when that path cannot be used. The NVFP4 text encoder does not require a Blackwell GPU. Check ComfyUI’s own PyTorch/CUDA environment before choosing a different checkpoint.

Comfy Desktop

Install the official Windows app from the Comfy Desktop guide. If it is already installed, use Desktop Update Ready or open Desktop Settings → Updates → Check for updates, then restart.

ComfyUI Portable

Close ComfyUI, open the portable installation’s update folder and run update_comfyui_stable.bat. The official documentation reserves update_comfyui_and_python_dependencies.bat for runtime problems; it is not the first button to press because dependency changes can break custom nodes.

Step 2: Choose the right H3 workflow

Open ComfyUI, go to Template Library → Video, and search for MiniMax H3. Current ComfyUI builds ship three native templates:

• MiniMax H3 T2V: the simplest first test and the best place to prove the installation works.
• MiniMax H3 I2V: accepts an optional first frame, last frame, or both.
• MiniMax H3 R2V: accepts reference images, videos and audio, but needs the separate Ref2VA diffusion checkpoint.

Choose T2V for now. When the template opens, ComfyUI may offer to fetch missing models. You can use that prompt or download the files manually from the official Comfy-Org/MiniMax-H3 repository.

Step 3: Download the exact model files

The four core T2V/I2V files below total about 42.47 GB. The current template’s 8-step LoRA adds about 1.96 GB, giving 44.43 GB in decimal units. R2V shares the text encoder and both VAEs, but needs the separate Ref2VA checkpoint and its own LoRA: about 22.93 GB extra if you keep the T2V files too. These are model-file totals; software, optional embeddings, caches and outputs need additional space.

For T2V, download the four core files above, skipping Ref2VA, and the additional files requested by the opened template. The current T2V/I2V templates list minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors in ComfyUI/models/loras/. Use the template’s download link or the official H3 download list. This file can be required by the model scan while turbo mode remains off. Restart ComfyUI after moving files, then check the missing-model prompt and loader selections again. If a download looks wrong, compare its full filename and size with the repository entry.

Step 4: Generate and check your first clip

1. Open the native MiniMax H3 T2V template.
2. Confirm that the FL2VA checkpoint, Qwen text encoder, video VAE and audio VAE appear in their loader nodes.
3. Use the template’s fast preview size for the first queue. For the native 16:9 canvas later, set 0.98 megapixels with a multiple of 32, or set 1344 × 768 directly. Avoid 1.0 megapixel: it rounds to 1376 × 768 and exceeds H3’s pixel-area limit.
4. Keep the resolution multiple at 32.
5. Choose a short duration. H3 supports 4–15 seconds at 24fps; ComfyUI snaps duration to the model’s valid 17-frame-block grid.
6. Leave the supplied sampler and model settings alone until one output succeeds.
7. Use a fixed seed while troubleshooting.

Queue one short T2V clip with turbo_mode off and the supplied sampling settings unchanged. Keep the template’s requested files installed. Leave optional attention patches and custom-node changes until this baseline completes, so there are fewer variables to check if it fails.

Step 5: Prompt the picture and the sound

H3 generates the audio and video together, so a silent visual prompt leaves part of the model under-directed. Describe the scene, subject, movement, camera, lighting and sound in one block. For multi-shot clips, add rough timestamps.

A close shot of a small red robot repairing a radio on a rain-soaked workbench at night. The camera slowly pushes forward while the robot turns a brass dial. Warm workshop light, realistic reflections and shallow depth of field. Stereo audio: rain tapping on the metal roof, a soft electrical hum, tiny tool clicks from the centre, and distant thunder moving from left to right. No speech, captions or logos.

Queue the prompt and keep the ComfyUI console visible. After completion, open the saved video and check that it plays, contains audible sound and uses the requested dimensions. Record the ComfyUI version, checkpoint, resolution, duration, seed and observed result before changing settings. If generation fails, save the first relevant console error rather than repeatedly changing several options at once.

Step 6: Move to I2V or R2V

Image-to-video and first/last frames

Open the I2V template and connect a first frame, last frame, or both to the MiniMaxH3ImageToVideo node. The FL2VA checkpoint remains correct. Describe the movement between the frames rather than repeating a static description of the image.

Reference-to-video

Open the R2V template and select minimax_h3_ref2va_pruned_int8_convrot.safetensors. Its current model list also includes minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors in ComfyUI/models/loras/; keep the Lightning LoRA switch off for the first check. The limits are nine images, three videos and three audio clips, with 12 files maximum overall. Each video or audio clip must be 2–15 seconds, and the combined duration must not exceed 15 seconds for video or 15 seconds for audio. Audio cannot be the only reference type.

Refer to inputs in connection order (such as <Picture 1>, <Video 1> and <Audio 1>) and give each one a job: identity, art style, motion, camera or voice. “Use Picture 1 for the character and Video 1 for the camera movement” is much clearer than dumping references into the graph and hoping the model reads your mind.

Optional: Sage Attention after the native workflow works

ComfyUI’s official H3 page says Sage Attention can roughly double generation speed with minimal quality loss. This is an upstream estimate; measure the effect on your own workflow. It is optional and adds a Python wheel plus KJNodes, so it belongs after the stable baseline, not before it.

If you try it, use the SageAttention wheel matching your installed PyTorch and CUDA versions. The official route adds KJNodes’ Patch Sage Attention KJ between UNETLoader and BasicGuider, or launches ComfyUI with --use-sage-attention. Audit third-party nodes before installing them. If the workflow fails, remove the patch and reproduce the problem with the untouched native template first.

How to confirm H3 is really local

After the model downloads, H3-Base generation runs through your local ComfyUI instance. You can see the local process using GPU, system memory and disk in Task Manager, and the result appears in the local output folder. That does not make MiniMax’s entire production pipeline local: H3-Context-IR and H3-Regenerate-2K are hosted services in the current release.

MiniMax H3 troubleshooting

The first load looks frozen

Watch the console rather than the browser spinner. If disk, CPU or GPU activity continues, the model may still be loading. If the console stops on an error, copy the first meaningful error message (not merely the final stack-trace line) before searching for a fix.

ComfyUI broke after an update

Test the native template in a separate clean portable ComfyUI copy, without optional custom nodes. If it works there, compare the original installation’s version, dependencies, settings and custom nodes. That result narrows the investigation; it does not identify the cause by itself. Add optional components back one at a time.

Frequently asked questions

Is MiniMax H3 open source?

It is open weight, but it is distributed under the custom MiniMax H3 Community License Agreement rather than a standard OSI-approved licence. The current agreement has territorial, commercial, distribution and use restrictions, so “downloadable” should not be confused with “unrestricted.”

Can MiniMax H3 run on 12GB or 6GB VRAM?

Community workflows report runs on 12GB and even 6GB cards using GGUF, aggressive offloading, reduced resolution and extra nodes. Those are useful experiments, not official minimums and not the native setup documented here. Expect compromises in speed, resolution, complexity or all three.

Can H3 generate audio locally?

Yes. H3-Base jointly predicts video and 32kHz stereo audio, and the native ComfyUI templates use a separate H3 audio VAE to decode it.

Is MiniMax H3 2K fully local?

No, not in the initial release. The open H3-Base path produces 768p-class output. The official 2K validation workflow also calls MiniMax’s hosted Context-IR and Regenerate-2K services.

Should I disable Dynamic VRAM?

Not as a first fix. Dynamic VRAM lets ComfyUI offload components when memory is tight. Disabling it can replace a slow reload with an out-of-memory failure. Update, reduce the job and reproduce the issue with the clean native workflow before changing memory flags.

Related local-AI guides

For a smaller still-image workflow focused on anime illustration, see the Anima-2.9B ComfyUI setup. If you need a first frame before animating it, try the Leonardo AI image-generation guide. For a smaller audio-only local project, see the MusicGen guide or the Tortoise TTS setup guide. You can also browse more practical AI tutorials.

Official sources

• MiniMax H3 model repository and system documentation
• MiniMax H3 model card and weights
• MiniMax H3 Community License Agreement
• Official MiniMax H3 ComfyUI workflows
• ComfyUI v0.31.0 release notes and H3 fixes
• ComfyUI RAM-aware model-cache change
• Official Comfy Desktop Windows installation guide
• Official ComfyUI Portable Windows guide

Final thoughts

Once a short T2V clip completes, move to I2V or R2V with one change at a time. Keep the working workflow and its settings so you can return to them when troubleshooting. Source review, 6 September 2026: checked the official templates, model card and documentation for file requirements, compatibility and generation limits.