MLX Backend Port for Apple Silicon

[CodeWithBehnam]

Hi @yakhyo,

I've created a fork of UniFace with native MLX backend support for Apple Silicon Macs (M1/M2/M3/M4):

🔗 Repository: https://github.com/CodeWithBehnam/mlx-uniface
📦 PyPI: https://pypi.org/project/mlx-uniface/

What's New

• Full MLX implementation of all models (RetinaFace, SCRFD, YOLOv5Face, ArcFace, etc.)
• Automatic backend selection (MLX on Apple Silicon, ONNX fallback elsewhere)
• Numerical parity with ONNX outputs (correlation > 0.999)
• Significant performance gains on Apple Silicon due to unified memory

Credits

All original work is credited to you in the README and pyproject.toml. This fork builds on your excellent library architecture.

Would you be interested in:

1. Merging MLX support into the main repo?
2. Linking to this fork from your README?

Happy to discuss either approach. Thanks for creating such a well-structured library!

Best,
Behnam

[yakhyo]

Hi Behnam @CodeWithBehnam,

Thanks for adding mlx support, I really appreciate your work.
At this moment, I am thinking of keeping the repo minimal and support only ONNX Runtime (except emotion detection).
I will acknowledge your work in the documentation soon.

thank you,
Yakhyo

Edit:

I have carefully checked your updates yet. I will check and if I change my mind, I may request for PR in future if you do not mind.