Generate a short AI video with audio from a text prompt on a Mac
Animate a scene between a supplied first frame and last frame image
Guide a video generation using one or more reference images
Tune speed versus quality with denoising passes and layer settings for a Mac's memory
| antirez/h3.c | pi-hole/ftl | peng-zhihui/peak | |
|---|---|---|---|
| Stars | 1,734 | 1,680 | 1,829 |
| Language | C | C | C |
| Last pushed | — | 2026-07-25 | 2022-12-12 |
| Maintenance | — | Active | Dormant |
| Setup difficulty | hard | moderate | hard |
| Complexity | 4/5 | 3/5 | 4/5 |
| Audience | developer | ops devops | developer |
Figures from each repo's GitHub metadata at analysis time.
Requires downloading the MiniMax-H3 model separately, FFmpeg, and an Apple Silicon Mac.
h3-metal, also called h3.c, is a native video and audio generation engine for Apple computers with M-series chips. It runs the MiniMax-H3 AI model locally on your Mac using Apple's Metal graphics technology, letting you turn a written text prompt into a short generated video, complete with sound, without sending anything to a cloud service. The tool works from the command line. After downloading the MiniMax-H3 model files and building the program, you type a prompt describing a scene, such as a fox walking through snow, and it produces a numbered video file. You can also run it as an interactive session, which keeps the model loaded in memory so that generating several videos in a row is faster than starting fresh each time. Inside a session there are commands for setting a random seed, adjusting the video length, saving output, and checking status. Beyond plain text-to-video, the tool supports several more advanced ways of guiding a generation. You can supply a first frame and a last frame image so the video is built to move between them, or supply one or more reference images that the model treats as visual anchors and can be told to animate, for example making a person in a photo wave at the camera. Generating video this way is computationally demanding, so the project gives you several dials to trade speed against quality: how many denoising passes to run, how many of the model's internal transformer layers to use, and whether to reuse some calculations instead of redoing them every time. A very fast setting can produce a rough result in a few seconds, while a slow, full-quality setting that matches a reference output can take much longer. There is also a memory-saving mode that streams model data from disk instead of keeping the entire model loaded in RAM, which trades some speed for a much smaller memory footprint. This is an actively developing project built specifically for Apple Silicon Macs, and using it requires downloading the underlying MiniMax-H3 model separately along with FFmpeg for video handling. The full README is longer than what was shown.
h3-metal runs the MiniMax-H3 AI model natively on Apple Silicon Macs to turn text prompts into short generated videos with audio.
Mainly C. The stack also includes C, Metal, MiniMax-H3.
Setup difficulty is rated hard, with roughly 1h+ to a first successful run.
Mainly developer.
This repo across BitVibe Labs
Don't trust strangers blindly. Verify against the repo.