Running MiniMax-H3-Image on Apple Silicon Macs (h3metal)
Hi, and thanks for releasing the weights.
h3metal (Apache-2.0) runs MiniMax-H3-Image on Macs with M5/M6-class GPUs, text-to-image and editing with references, from the official checkpoint converted locally. On a Mac mini (M6, 24 GB) a 2048 x 2048 image takes 41 s and a 4096 x 4096 image 244 s.
The transformer runs in int8 (with an int4 MLP on 24 GB Macs) on the GPU's matrix units, and each part is tested against your diffusers and transformers modules; the figures and the method are in the README. The README also states your license's territory and acceptable-use terms, and the PNGs h3metal writes say that they are AI-generated.
It is a community project, not affiliated with MiniMax. I have opened a PR that adds it to the Deployment list of the GitHub README: https://github.com/MiniMax-AI/MiniMax-H3-Image/pull/6
Why do you massacre yourself with Apple problem? Get PC, Windows 11, forget about problems.
Great, I really need it. Is there a workflow that can be used for testing？
@simgon Thanks! There is no ComfyUI workflow (h3metal has no ComfyUI node at the moment); it is a command-line tool with a small Python API, so a first test is four commands:
pip install "h3metal @ git+https://github.com/audenvenn/h3metal@v0.2.0"
h3metal download # the official checkpoint, ~125 GB
h3metal convert ~/.h3metal/checkpoint # once; afterwards the checkpoint can be deleted
h3metal generate -p "A red fox in fresh snow at sunrise" -o fox.png
An edit takes reference images with -i and names them <Picture 1>, <Picture 2>, ... in the prompt:
h3metal generate -p "Put <Picture 1> on a beach at dusk" -i photo.jpg -o edit.png
It needs a Mac with an M5- or M6-class GPU, macOS 26, Python 3.10+ and the Xcode command line tools. 24 GB of memory is enough, and the converted model takes about 44 GB on disk. Please read the model's license first (summary in the README): it limits where and how the model may be used. The README has the other options and the Python API. If something does not work, or a ComfyUI node would help you, please open an issue on GitHub.
An update for anyone following this thread: h3metal 0.6.0 also runs MiniMax-H3-Image in 2, 3 or 4 steps with pDMD's 2-step LoRA (Apache-2.0), merged into the weights at conversion.
On a Mac mini (M6, 24 GB):
| 6 steps | 2 steps | |
|---|---|---|
| 2048 x 2048 | 42 s | 16 s |
| 4096 x 4096 | 240 s | 85 s |
It now installs from a wheel and has a small local web page (h3metal serve):
pip install https://github.com/audenvenn/h3metal/releases/download/v0.6.0/h3metal-0.6.0-py3-none-macosx_26_0_arm64.whl
h3metal download # the official checkpoint
h3metal convert ~/.h3metal/checkpoint # the six-step model
h3metal convert ~/.h3metal/checkpoint --pdmd # optional: 2, 3 and 4 steps
h3metal generate --steps 2 -p "..."
- What 2 steps give: composition and lettering come out right. Fine texture can show a faint crosshatch at full size, and the six-step model stays the default. There is a comparison in docs/few-step.md.
- Hardware: it needs an M5/M6-class Mac.
- License: the images are AI-generated, and this model's license terms apply as before.