AI Briefing3D generation on a Mac
Key point
TRELLIS.2 was ported to Apple Silicon's MPS, enabling 3D conversion without an NVIDIA GPU.
Details
Microsoft's TRELLIS.2 image-to-3D was ported to run on Apple Silicon / PyTorch MPS.
The original architecture, which was tied to CUDA-only extensions, was replaced with the following alternative implementations.
- Pure-PyTorch sparse 3D convolution instead of flex_gemm
- Python-based spatial hashing mesh extraction instead of o_voxel's CUDA hashmap operations
- SDPA for the sparse transformer instead of flash_attn
- MPS GPU-based trilinear voxel sampling using torch.grid_sample
In terms of performance, on an M4 Pro 24GB, a mesh of about 400K vertices (vertex 400K) was generated from a single photo in 3.5 minutes.
Texture baking took about 18 seconds, and while it's not H100-level speed, the key point is that it enables local offline execution without an NVIDIA GPU.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
OpenArch Released: A Collection of PyTorch Implementations for Modern LLM Architectures
ZeroModels Released, Compatible with JAX, PyTorch, and TensorFlowGoogle Releases Gemini-Based AI Agent Tools to Accelerate TPU Development