Show HN: TRELLIS.2 image-to-3D running on Mac Silicon – no Nvidia GPU needed
Not a pilot. Image-to-3D inference now runs offline on Mac Silicon—no Nvidia required. 3.5 minutes per mesh.
Why it matters
Developer democratizes access to Microsoft's TRELLIS.2 model by porting it to consumer hardware (Apple Silicon), eliminating cloud dependency and GPU costs. Signals broader shift toward edge AI inference on commodity chips.
The key facts
13 to knowTRELLIS.2 (4B parameter model) ported to Apple Silicon via PyTorch MPS
Generates ~400K vertex 3D meshes from single images in ~3.5 minutes on M4 Pro (24GB)
Original CUDA version runs in seconds on H100; Mac version trades speed for accessibility and offline capability
Replaced CUDA-specific ops (flash_attn, nvdiffrast, sparse convolution kernels) with pure-PyTorch alternatives across 9 files
No cloud dependency; runs fully offline
Posted to Hacker News (33 points, 2 comments) — early-stage technical community interest
TRELLIS.2 (4B parameters) successfully ported to Apple Silicon via PyTorch MPS
Replaced CUDA-specific ops (flash_attn, nvdiffrast, sparse convolution kernels) with pure PyTorch alternatives
Generates ~400K vertex meshes from single photos in ~3.5 minutes on M4 Pro (24GB)
Runs offline with no cloud dependency
Original H100 performance: seconds; M4 Pro performance: 3.5 minutes
Code changes: ~few hundred lines across 9 files
Published: April 20, 2026 (very recent)
Go to the source
Hacker Newsgithub.com
Publisher excerpt: I ported Microsoft's TRELLIS.2 (4B parameter image-to-3D model) to run on Apple Silicon via PyTorch MPS. The original requires CUDA with flash_attn, nvdiffrast, and custom sparse convolution kernels: none of which work on Mac. I replaced the CUDA-specific ops with pure-PyTorch alternatives: a…