How to Run MiniMax H3 Locally in ComfyUI

TL;DR

Run MiniMax H3 locally by updating ComfyUI, opening a non-API MiniMax workflow, and downloading the required diffusion model, Qwen 3 VL 32B text encoder, audio VAE, and video VAE into their matching model folders. Compressed models reduce hardware requirements, while the available workflows support text, images, and references as inputs with built-in audio output.

Transcript

We have a new open-source video generator that you can run locally on your computer. It's called Miniax H3, and this is by far the best one available. The best thing is you can even run this with as low as 5 to 6 GB of VRAM. Plus, it has audio builtin, and it's insanely good at character consistency, world knowledge, instruction following, and even... Read More

Key Insights

• MiniMax H3 is an open-source multimodal generator that can use text, images, video, and audio as input references. Its outputs can include synchronized audio, allowing local workflows to produce both the visual sequence and its accompanying sound.
• MiniMax H3 demonstrates strong instruction following across realistic scenes, cartoons, clay animation, anime, dialogue, unusual actions, and combinations of realistic imagery with hand-drawn elements. It also preserves referenced characters consistently when placing multiple subjects into one generated sequence.
• Reference-to-video is the most powerful workflow presented because it can combine reference characters and backgrounds, transfer actions from an existing video, transform scenes with new visual elements, and add, clone, remove, or replace subjects and objects.
• Local MiniMax H3 workflows are available through ComfyUI templates for text-to-video, image-to-video, and reference-to-video. Templates carrying an API tag use the cloud, so users seeking local generation must choose versions without that label.
• MiniMax H3 installation requires a compatible diffusion model, the Qwen 3 VL 32B text encoder, an audio VAE, and a video VAE. These files belong in ComfyUI's diffusion_models, text_encoders, and VAE folders respectively.
• Compressed diffusion models trade some output quality for lower storage and hardware requirements. The presented options range from a 66 GB full model to 34 GB int8 and 21 GB pruned versions, with some users reporting operation on 12 GB of VRAM.
• Text-encoder memory can be supplied by system RAM instead of fitting entirely within VRAM. Users should select a Qwen 3 VL 32B compression level suited to their device, then refresh ComfyUI and choose the downloaded files in the workflow nodes.
• Generation settings include aspect ratio, duration, resolution, sampler, scheduler, and step count. The tutorial associates 480p output with 0.4 megapixels, recommends keeping the default sampler and scheduler, and initially retains the default setting of 20 steps.
• More videos from this channel:
  • Kim K2 Thinking Open Source AI Review and Demos
  • How to use Gemini 2.0 Flash for image edits
  • How to upsample video 4x and colorize art with AI tools
• More from this channel: AI Search

Questions & Answers

Q: How do you install MiniMax H3 in ComfyUI?

Update ComfyUI first by opening its update folder and running the provided update file. Start ComfyUI, open Templates, search for MiniMax, and select a workflow without an API tag. Download the required diffusion model, Qwen 3 VL 32B text encoder, audio VAE, and video VAE. Place them in their matching model folders, refresh the model list by pressing R, and select each file in the workflow.

Q: Which MiniMax H3 workflow runs locally instead of through the cloud?

Choose one of the MiniMax text-to-video, image-to-video, or reference-to-video templates that does not display an API tag. The API label indicates that generation runs through the cloud rather than locally. If the templates are unavailable inside ComfyUI, the workflow can be downloaded manually from the linked resource page and dragged onto the ComfyUI interface.

Q: Which models are required to run MiniMax H3 locally?

A local workflow requires a MiniMax H3 diffusion model, the Qwen 3 VL 32B text encoder, an audio VAE, and a video VAE. The FL2VA diffusion-model family is used for text-to-video and image-to-video, while reference-to-video uses its designated Ref2VA model. Choose compressed versions that match the available hardware and place every file in its corresponding ComfyUI model directory.

Q: Where should MiniMax H3 model files be placed in ComfyUI?

Place the selected MiniMax diffusion model inside ComfyUI's models/diffusion_models directory. Put the Qwen 3 VL 32B encoder in models/text_encoders. Both the audio VAE and video VAE belong in models/VAE. After all downloads finish, return to the workflow, press R to refresh the available models, and select the newly installed files from the relevant dropdown menus.

Q: How much storage and VRAM do MiniMax H3 models require?

The full diffusion model shown in the tutorial is 66 GB, the int8 version is 34 GB, and the pruned FP8 or pruned int8 versions are 21 GB. Larger files provide the highest quality, while stronger compression introduces some quality loss. Some users reportedly run the 21 GB models with 12 GB of VRAM, and other low-VRAM options are referenced later in the tutorial.

Q: How do you generate text-to-video with MiniMax H3?

Open the local text-to-video MiniMax template, select the installed diffusion model, Qwen 3 VL 32B text encoder, video VAE, and audio VAE, then configure the prompt, aspect ratio, megapixel setting, and duration. The tutorial uses 0.4 megapixels for 480p and leaves the sampler, scheduler, and 20-step setting at their defaults before pressing Run. Completed videos are saved under the output folder's video directory.

Q: How do you animate an image with MiniMax H3?

Open the image-to-video template without an API tag and select the previously downloaded model files. Upload the source image, choose an aspect ratio and resolution, set the desired duration, and enter a motion or scene prompt. The demonstrated workflow uses a 16:9 aspect ratio, 480p output, and the default nearest-exact rescaling method before the generation is started.

Q: What can MiniMax H3 reference-to-video do?

Reference-to-video can combine characters and backgrounds from supplied images, create advertisements from product references, and build short dramatic scenes with dialogue. It can also transform an existing video using new visual elements, copy actions onto different referenced characters, clone a detected character, and replace scenery after a specified event. This workflow requires a designated reference-to-video diffusion model.

Summary & Key Takeaways

• MiniMax H3 is an open-source, multimodal video generator designed to run locally through ComfyUI. It accepts text, image, video, and audio references, generates built-in audio, follows unusual instructions, and maintains characters across scenes. Demonstrated uses include animation, commercials, dialogue, gameplay concepts, scene transformations, action transfer, and object or character editing.

• Installation begins by updating ComfyUI and selecting a MiniMax template without an API tag. The required files include a diffusion model, the Qwen 3 VL 32B text encoder, an audio VAE, and a video VAE. Each file must be placed in the corresponding diffusion_models, text_encoders, or VAE directory before refreshing ComfyUI.

• Text-to-video and image-to-video share the same primary diffusion-model family, while reference-to-video requires its designated model. Compressed versions reduce storage and memory requirements but sacrifice some quality. The demonstrated settings use 480p, 0.4 megapixels, the default sampler and scheduler, 20 generation steps, and approximately one minute of processing per output second.