ZimengXiong ports Hunyuan3D to MLX for Apple Silicon and iPhone

Developer ZimengXiong has completed a Swift-MLX and Python MLX port of the Hunyuan3D-Paint and Hunyuan3D-Shape models, enabling local image-to-3D generation on Apple Silicon devices. The work is distributed as the open-source Modelr desktop app for macOS and iOS, along with source code for integrating the technology into Swift applications.

• Benchmarks on an M4 Max show hy3d shape (small) completes in 20.9 seconds using ~5.6 GB of memory, while hy3d shape (large) takes 22.3 seconds using ~7.3 GB.
• The hy3d paint model requires significantly more resources, taking 231 seconds for RGB texturing (~38 GB peak memory) and 344 seconds for PBR texturing (~39 GB peak memory).
• The MLX implementation allows the models to run on recent Macs and iPhones in Q4 or Q8 quantization, avoiding the overhead of PyTorch or CPU execution.

This port makes it possible to run image-to-3D generation locally on consumer Apple hardware with low memory requirements, providing a standalone app for macOS/iOS and libraries for developers.