TRELLIS-Silicon

Run TRELLIS.2 image-to-3D generation natively on Apple Silicon — no NVIDIA GPU required.

TRELLIS.2 is Microsoft Research’s state-of-the-art image-to-3D model. It ships CUDA-only. This project runs it on Apple Silicon through PyTorch’s MPS backend: a single image in, a textured GLB out, entirely on your Mac’s GPU.

TL;DR

• One command: trellis-silicon photo.png → GLB with baked PBR textures. Web UI: trellis-silicon-web.
• Fast: warm runs take ~20s load + ~1.6-2.0 min generation + ~22s GLB export (default 512 pipeline, ~2.4 min wall, median of four runs on an M5/32GB). The performance pass over the base port cut load 5× (~103s → ~20s) and total warm time roughly in half (~300s → ~145s; the base-port figure was measured on an earlier sample image); --steps 8 fast mode trims generation by another ~20%.
• Where the speed came from (measured, not guessed): skipping a weight init that gets overwritten anyway, loading only the checkpoints the pipeline uses, batching the two CFG forwards, swapping MPS SDPA for a naive attention that is 3.3-4.1× faster at these shapes, running sparse convolution on…