h3-studio 🎬

A bilingual (EN/DE) web GUI for antirez/h3.c focused on the music-video workflow: lip-synced video from reference image + audio in a single pass.

Note: There's also PR #45 "H3 Studio" – a native macOS desktop app with hardware-aware presets. This project (h3-studio) is a separate, web-based GUI specializing in audio+video conditioning and multilingual UI.

Unofficial community project. Not affiliated with antirez or MiniMax.

What is this?

h3.c is Salvatore Sanfilippo's (antirez) native C/Metal inference engine for MiniMax H3 – the fastest way to run H3 on a Mac. But it is CLI-only. h3-studio wraps it in a widescreen browser dashboard:

• 📝 Simple mode (free prompt) and Advanced mode (Context-IR fields: Scene / Action / Camera / Look / Audio with smart defaults for empty fields)
• 🖼️ Up to 9 reference images, 🎞️ 3 video refs, 🎵 3 audio refs (Ref2VA)
• 🥇 First/last-frame conditioning (FL2VA) with built-in conflict protection
• 🎚️ Resolution presets, frame/step/layer sliders, seed, reuse, token reduction
• 🌐 DE/EN toggle (remembered per browser)
• 📟 Live terminal log + automatic video preview

Perfect for music videos: reference image + song → lip-synced video with native stereo audio in a single pass. No more "silent render → LatentSync → merge" ping-pong.

Requirements

• Apple Silicon Mac (tested on M4 Max, 128 GB)
• ~210 GB free SSD space
• Compiled h3.c (make -j) with official weights (see below)
• ffmpeg on PATH (brew install ffmpeg)
• Python 3 + Flask

Quick start

1. Build h3.c, then download the weights surgically (~210 GB instead of ~340 GB):

hf download MiniMaxAI/MiniMax-H3 --local-dir ./MiniMax-H3 \
  --include "FL2VA/*" \
  --include "Ref2VA/transformer/*" \
  --include "Ref2VA/tokenizer/*" \
  --include "Ref2VA/model_index.json"
  
  2. Text encoder and both VAEs are byte-identical in both trees – save ~80 GB with symlinks:
  
  cd MiniMax-H3
ln -s ../FL2VA/text_encoder Ref2VA/text_encoder
ln -s ../FL2VA/audio_vae   Ref2VA/audio_vae
ln -s ../FL2VA/video_vae   Ref2VA/video_vae
cd ..

3. Put h3-studio next to the h3.c folder, then:

pip install -r requirements.txt
python3 app.py

4. Open http://127.0.0.1:5000 and direct your first take!

Benchmarks (real-world, end-to-end)
M4 Max, 128 GB (this repo's author)
Resolution      Frames      Steps/Layers    Reference       Time
512x512         243         20/45           1 Image+Audio   32min (incl.cold weight load)
512x512         243         20/45           4 Images+Audio  21min (warm cache) 

Community data (h3.c issues #5, #29, #30)
Machine             Resolution      Duration        Time
M4 Max 64 GB        960x544         10s             35min
M1 Max 64 GB        256x256         15s             6min
MacBook Air M5 32GB 512x512         1s              101sec

Prompt guide (hard-won lessons)

The model follows the described action, not keywords: write "sings the vocals" instead of "lip sync".
Clothing: state it explicitly (wearing a red dress) – or pin it with several consistent reference images.
References carry pose cues! A back-view picture may make the model turn the person around mid-video. Add "stays facing the camera, does not turn around" and "Pictures 2–4 are identity references only".
Audio: 2–15 s, pick a segment with continuous vocals for best sync.
Ref2VA (images/videos/audio) and FL2VA (first/last frame) cannot be mixed.

Roadmap
Chunked 5 s workflow: first-frame chaining + FFmpeg concat button
Progress bar parsed from the live log
Turbo-LoRA fast preset (once h3.c PR #14 lands)
768p quality presets

Credits & License
h3.c © Salvatore Sanfilippo – MIT
MiniMax-H3 © MiniMax – own license; weights are not redistributed here
h3-studio – MIT