A community port of Microsoft TRELLIS.2 now runs natively on Apple Silicon, enabling image-to-3D generation on Mac hardware without an Nvidia GPU. Developer Shivam Kumar published the port on April 19, 2026, replacing five CUDA-dependent components with Metal and PyTorch equivalents that run on M-series chips.

What Happened

The original TRELLIS.2 from Microsoft required CUDA, limiting 3D generation to Nvidia hardware. Kumar ported it to Apple Silicon by replacing sparse convolution with PyTorch gather-scatter operations, swapping CUDA attention for PyTorch SDPA, and using Metal-accelerated texture baking. The result generates approximately 400,000-vertex 3D meshes from a single input image, with physically-based rendering textures output in GLB format ready for import into Blender, Unity, or any PBR-compatible renderer.

Why It Matters

TRELLIS.2 is one of the strongest open image-to-3D models available, capable of generating detailed, textured meshes from a photograph. Until now, Mac users had no practical way to run it locally. This port brings that capability to anyone with an M1 or later Mac, without cloud inference costs or Nvidia hardware requirements.

This fits a broader pattern in open-source 3D AI. Earlier this month, Tencent open-sourced HY-World 2.0 for 3D world generation, and Alibaba released Happy Oyster for interactive 3D video environments. The tools for 3D creation are becoming more accessible and less GPU-dependent with each week.

Key Details

  • Output quality: approximately 400,000-vertex meshes with PBR textures
  • Output format: GLB, compatible with Blender, Unity, Unreal, and most 3D tools
  • Hardware: Apple Silicon M-series (any Mac with M1 or later)
  • Performance: approximately 10x slower than the original CUDA version due to pure-PyTorch sparse convolution
  • Known limitation: hole filling is currently disabled in the port
  • Models used: TRELLIS.2 for geometry and texture, DINOv3 for image features, RMBG-2.0 for background removal
  • Repository: github.com/shivampkumar/trellis-mac
  • Created: April 19, 2026

What to Do Next

Clone the repository from the repository and follow the setup instructions. You will need Python and the documented dependencies. The first run will download the model weights, which are several gigabytes. Start with a clean, well-lit photo against a simple background for the best initial mesh quality.

The 10x speed difference versus CUDA is real, so expect generation times in the range of minutes rather than seconds. For production pipelines where speed matters, pair this with cloud-based TRELLIS runs. For local experimentation, concept validation, and one-off 3D asset creation from reference photos, this port is now a practical option for every Mac studio setup.