Microsoft's TRELLIS.2 image-to-3D model now runs natively on Apple Silicon Macs via PyTorch MPS, eliminating NVIDIA GPU dependency. The port replaces CUDA-specific operations with pure-Python alternatives, including a gather-scatter sparse convolution and SDPA attention, enabling offline 3D mesh generation from single photos.**

Developed by Shivam Kumar, the project converts TRELLIS.2's 4B-parameter architecture to work on M1/M2/M3 chips. Performance on an M4 Pro (24GB) achieves ~3.5 minutes per 400K-vertex mesh, with memory usage peaking at 18GB unified memory. While 10x slower than CUDA implementations, it enables texture export and mesh simplification without cloud dependencies.

Key technical innovations include Python-based mesh extraction replacing CUDA hashmap operations and differentiable rasterization stubs. The port required rewriting nine core files, with sparse 3D convolution implemented through spatial hashing and neighbor map caching. Gated models like DINOv3 and RMBG-2.0 integration adds licensing complexity for commercial use.

Available via [GitHub](https://github.com/shivampkumar/trellis-mac), the tool targets developers needing offline 3D pipeline capabilities. Limitations persist in texture baking (requires nvdiffrast) and hole filling (needs cumesh), but the open-source approach advances accessibility for Mac-based 3D workflows. The project demonstrates Apple Silicon's viability for running large AI models, challenging traditional GPU-centric AI development paradigms.