Developer ports Microsoft's TRELLIS.2 image-to-3D model to Apple Silicon

A developer has successfully ported Microsoft's TRELLIS.2 image-to-3D generation model to run natively on Apple Silicon hardware. The original model requires CUDA with specialized libraries that don't work on Mac systems. The port replaces CUDA-specific operations with pure-PyTorch alternatives, including a gather-scatter sparse 3D convolution implementation and Python-based mesh extraction. These changes allow the 4-billion-parameter model to generate textured 3D meshes from single images without NVIDIA GPUs. On an M4 Pro with 24GB of memory, the port generates meshes with approximately 400,000 vertices in about 3.5 minutes. The output includes OBJ and GLB files with physically based rendering materials ready for use in 3D applications. While slower than the CUDA version, which processes images in seconds on H100 hardware, the Apple Silicon port provides an offline alternative with no cloud dependency. The implementation is available on GitHub under the MIT license, though some upstream model weights have separate licensing terms.