Skip to content
Hacker News front page

Show HN: TRELLIS.2 image-to-3D running on Apple Silicon, no Nvidia GPU needed

Show HN: TRELLIS.2 image-to-3D running on Mac Silicon – no Nvidia GPU needed

Developer shivampkumar ported Microsoft's 4B-parameter TRELLIS.2 to Apple Silicon with PyTorch MPS for single-image 3D generation. He replaced flash_attn, nvdiffrast, and custom sparse conv kernels with pure PyTorch sparse 3D conv, SDPA attention, and Python mesh extraction. On an M4 Pro with 24GB, it generates ~400K-vertex meshes in about 3.5 minutes; slower than H100 seconds, but fully offline.

Why it matters: Strong on all HKR axes: a clear hook, concrete implementation details, and benchmark-like numbers. This is not a Microsoft model launch, but a reproducible local port with real practitioner relevance, so it lands in featured rather than p1.

Read the original ↗Export Markdown