Description
Want to run MiniMax-H3 audio-video generation on your own GPU without building ComfyUI node graphs? MiniMax-H3 Studio is a self-hosted web UI that puts text, image and reference-to-video, plus picture generation and editing, speech, transcription, music, image-to-3D and upscaling on one job queue.
It drives ComfyUI as a Python library, with no ComfyUI interface. It needs Linux and an NVIDIA GPU (developed on an RTX 4090 24 GB, 64 GB+ RAM recommended) and you install ComfyUI yourself. On a 4090 a 5-second clip with sound takes roughly 30–45 s at 544p. Apache-2.0.
Video modes: Text to video, image to video (first or first and last frame), and up to 6 images and 3 videos as references.
Pictures and 3D: Image generation and editing, and a single image to a textured GLB you can rotate, zoom and download.
Speech and music: Speech generation and cloning, audio and video transcription with subtitle export, and songs up to 5 minutes from lyrics.
Upscaling and guide audio: Upscale finished videos to 1080p or 1440p keeping sound, or anchor a song or recording as a shot's soundtrack.
Job queue: Weights stay loaded between jobs, and jobs run one at a time from a persistent queue.
It drives ComfyUI as a Python library, with no ComfyUI interface. It needs Linux and an NVIDIA GPU (developed on an RTX 4090 24 GB, 64 GB+ RAM recommended) and you install ComfyUI yourself. On a 4090 a 5-second clip with sound takes roughly 30–45 s at 544p. Apache-2.0.
Features
Video modes: Text to video, image to video (first or first and last frame), and up to 6 images and 3 videos as references.
Pictures and 3D: Image generation and editing, and a single image to a textured GLB you can rotate, zoom and download.
Speech and music: Speech generation and cloning, audio and video transcription with subtitle export, and songs up to 5 minutes from lyrics.
Upscaling and guide audio: Upscale finished videos to 1080p or 1440p keeping sound, or anchor a song or recording as a shot's soundtrack.
Job queue: Weights stay loaded between jobs, and jobs run one at a time from a persistent queue.
