TRELLIS.2 Brings Image-to-3D to Apple Silicon Without Nvidia

Microsoft's TRELLIS.2 image-to-3D model now runs on Apple Silicon Macs. No NVIDIA GPU required. The port, available on GitHub, uses PyTorch's MPS backend to swap CUDA-only libraries for pure PyTorch alternatives. On an M4 Pro, it generates meshes with 400,000+ vertices from a single image in about 3.5 minutes. Peak memory sits around 18GB.

The trade-offs are real. Texture export is gone. Hole filling is disabled. The sparse convolution replacement runs 10x slower than CUDA. You'll want 24GB of unified memory, minimum.

NVIDIA's CUDA ecosystem has purpose-built tools for 3D AI work. Things like nvdiffrast with optimized kernels for differentiable rendering, the specialized operations that 3D neural networks need constantly. Apple doesn't have an equivalent. Developers can get models running on MPS, but they're stuck with generic PyTorch fallbacks that don't use the hardware properly.

The pipeline uses Meta's DINOv3 for image feature extraction and BRIA AI's RMBG-2.0 for background removal. Watch the licensing: RMBG-2.0 is non-commercial under CC BY-NC 4.0, so you'd need a separate agreement with BRIA for commercial use.