Repolode
Submit
zxcHolmes

zxcHolmes/minimax-h3-studio

A self-hosted web studio that drives ComfyUI as a Python library to generate audio-video, pictures, 3D models, music and speech on a single NVIDIA GPU machine.

View on GitHub
Stars
187
Forks
47
Open issues
0
Last commit
2 days ago

What is it?

MiniMax-H3 Studio is a self-hosted web app for running MiniMax-H3 audio-video generation on one GPU machine. It uses ComfyUI as a Python library rather than its node UI or workflow JSON, so it can load quantized ComfyUI checkpoints and ComfyUI's dynamic VRAM offloading, and it places a resident job server and a browser UI in front of them.

The studio covers text-to-video, image-to-video (first frame or first and last frame), and reference-to-video with up to 6 images and 3 videos. It also handles picture generation and editing with MOSS-Image, FLUX.2 Klein 4B and Qwen Image 2.1 Turbo, image-to-3D with Pixal3D that exports GLB, music generation with MiniMax Music 3, video upscaling with FlashVSR or SeedVR2, and speech through a native audio.cpp CUDA server (TTS, voice cloning and Qwen3-ASR transcription). Jobs run one at a time from a persistent queue, and model weights stay loaded between jobs.

Deployment is handled by a script that pushes the server and UI to a GPU machine over SSH and sets up a systemd service, and a separate script downloads the required checkpoints. The user must install ComfyUI and its Python environment on the GPU host.

Key features

  • 01Text, image and reference-driven video generation with optional guide audio for lip sync and music videos.
  • 02Picture generation and editing with three models, with outputs sendable to editing, video or 3D.
  • 03Image-to-3D producing textured GLB files with an in-browser viewer.
  • 04Speech synthesis, voice cloning and transcription with subtitle export via audio.cpp.
  • 05Persistent job queue with weights kept resident between jobs.
  • 06SSH deploy script and a model download script for setup on a GPU machine.

Use cases

  • Generating short AI video clips with sound on a local RTX-class GPU.
  • Making music videos by anchoring a song as the soundtrack for each shot.
  • Creating textured 3D models from single product or object photos.
  • Transcribing and subtitling recorded audio or video on your own hardware.

Quick start

  1. 1

    Clone the repository

    git clone https://github.com/<you>/minimax-h3-studio && cd minimax-h3-studio
  2. 2

    Create your local configuration file and edit it

    cp .env.example .env
  3. 3

    Download the configured model checkpoints to the GPU machine

    ./deploy.sh models
  4. 4

    Build the UI, install the server and start it

    ./deploy.sh

Things to know

  • Requires a Linux machine with an NVIDIA GPU; it was developed on an RTX 4090 with 24 GB VRAM.
  • You must install ComfyUI yourself, since this project only imports it.
  • The model checkpoints total tens of gigabytes, so downloads and disk space are a significant cost.
  • The fastest CUDA backend on PyTorch older than cu130 needs a source build of comfy_kitchen and a patch to ComfyUI.

In collections

Self-hosted appsAI video & audio toolsAI image generators & Stable Diffusion tools

Topics

#ai#video-generation#comfyui#image-generation#text-to-speech#speech-recognition#3d-generation#self-hosted

FAQ

What is minimax-h3-studio?
A self-hosted web studio that drives ComfyUI as a Python library to generate audio-video, pictures, 3D models, music and speech on a single NVIDIA GPU machine. MiniMax-H3 Studio is a self-hosted web app for running MiniMax-H3 audio-video generation on one GPU machine. It uses ComfyUI as a Python library rather than its node UI or workflow JSON, so it can load quantized ComfyUI checkpoints and ComfyUI's dynamic VRAM offloading, and it places a resident job server and a browser UI in front of them.
Is minimax-h3-studio free to use? What license does it have?
minimax-h3-studio is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is minimax-h3-studio still maintained?
Actively maintained. The last commit to minimax-h3-studio was on Nov 6, 2026.
What can I use minimax-h3-studio for?
Generating short AI video clips with sound on a local RTX-class GPU. Making music videos by anchoring a song as the soundtrack for each shot. Creating textured 3D models from single product or object photos. Transcribing and subtitling recorded audio or video on your own hardware.
How do I install minimax-h3-studio?
Clone the repository: git clone https://github.com/<you>/minimax-h3-studio && cd minimax-h3-studio Create your local configuration file and edit it: cp .env.example .env Download the configured model checkpoints to the GPU machine: ./deploy.sh models Build the UI, install the server and start it: ./deploy.sh
Where can I get minimax-h3-studio?
The source code and installation instructions are on GitHub: https://github.com/zxcHolmes/minimax-h3-studio

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects