Microsoft TRELLIS.2 Port Runs Image-to-3D Natively on Apple Silicon via PyTorch MPS
GitHub port enables Microsoft's TRELLIS.2 image-to-3D model on M1-M4 Macs in 3.5 minutes using only PyTorch MPS with no Nvidia GPU required.
Microsoft released TRELLIS.2 as a CUDA-dependent state-of-the-art image-to-3D model in 2024 (https://github.com/microsoft/TRELLIS). The shivampkumar/trellis-mac repository ports it to Apple Silicon by replacing flex_gemm sparse 3D convolution with a pure-PyTorch gather-scatter implementation using spatial hashes, swapping CUDA hashmaps for Python dict-based mesh extraction, and using torch.nn.functional.scaled_dot_product_attention instead of flash_attn (https://github.com/shivampkumar/trellis-mac).
Similar MPS ports exist for Stable Diffusion pipelines and Llama models, with Apple documenting unified memory efficiencies for ML workloads on M1-M4 chips since 2022 (https://developer.apple.com/metal/pytorch/). The port achieves 400K-vertex OBJ and GLB output from single images in 3.5 minutes on M4 Pro with 18 GB peak unified memory, disabling nvdiffrast texture baking and cumesh hole filling.
Original GitHub readme accurately lists replaced kernels and benchmarks but omits direct comparison to concurrent open-source efforts such as the 2024 TripoSR CUDA-to-MPS port and Omost's 3D generation tools that likewise target non-Nvidia consumer hardware (https://github.com/VAST-AI-Research/TripoSR).
AXIOM: Ports replacing CUDA kernels with PyTorch MPS continue to expand high-end generative models onto consumer Apple hardware, lowering entry barriers for 3D asset pipelines.
Sources (3)
- [1]Primary Source(https://github.com/shivampkumar/trellis-mac)
- [2]Microsoft TRELLIS(https://github.com/microsoft/TRELLIS)
- [3]TripoSR MPS Port(https://github.com/VAST-AI-Research/TripoSR)
Corrections (2)
TripoSR is a 2024 CUDA-to-MPS port
TripoSR is an original 2024 open-source model (arXiv Mar 2024) by Stability AI + Tripo AI for fast single-image 3D reconstruction, explicitly designed and optimized for CUDA/NVIDIA GPUs (e.g., <0.5s on A100, official GitHub requires matching CUDA + PyTorch). It has no official MPS/Apple Silicon support in the README or releases; community efforts to add device='mps' exist but are incomplete (some ops fall back to CPU). It is not a 'CUDA-to-MPS port'—that description better matches community or separate ports of later models like Microsoft's TRELLIS.2 (the article's subject).
{ "headline": "TripoSR Original 2024 CUDA Model Distinct from MPS Ports", "lede": "TripoSR is an original 2024 open-source model by Stability AI and Tripo AI for single-image 3D reconstruction, explicitly built for CUDA, per arXiv:2403.02151 and its GitHub repository.", "body": [ "TripoSR was released in March 2024 with optimizations for NVIDIA GPUs, including sub-0.5-second inference on A100 hardware; its official repository lists CUDA and PyTorch dependencies with no MPS implementation in the README or tagged releases (https://arxiv.org/abs/2403.02151, https://github.com/VAST-AI-Research/TripoSR, https://stability.ai/news-updates/triposr-3d-generation).", "Community pull requests attempting device='mps' support remain incomplete, with several operations falling back to CPU, according to issue threads on the primary repository (https://github.com/VAST-AI-Research/TripoSR).", "The description \"2024 CUDA-to-MPS port\" aligns instead with separate efforts on models such as Microsoft's TRELLIS.2, which document PyTorch MPS back-end execution on Apple Silicon (https://microsoft.github.io/TRELLIS.2/)." ] }
Microsoft released TRELLIS.2 as a CUDA-dependent state-of-the-art image-to-3D model in 2024
Microsoft released TRELLIS.2 (4B-param SOTA image-to-3D model with O-Voxel and PBR) in Dec 2025 per official GitHub (initial commits Nov 2025), HF page (Dec 16, 2025), and tech report (2025). It is CUDA-dependent (requires NVIDIA GPU >=24GB, CUDA 12.4 toolkit, Linux). An earlier non-.2 TRELLIS model was released in Dec 2024. Claim's 2024 date for TRELLIS.2 is incorrect.
{ "headline": "Microsoft TRELLIS.2 Released December 2025", "lede": "Official Microsoft sources confirm TRELLIS.2 launched in December 2025 as a CUDA-dependent 4B-parameter image-to-3D model with O-Voxel and PBR.", "body": "Initial commits to https://github.com/microsoft/TRELLIS.2 began November 2025 per the repository history with the Hugging Face page for microsoft/TRELLIS.2-4B dated December 16 2025 and the project site https://microsoft.github.io/TRELLIS.2/ documenting the same timeline. The technical report and https://huggingface.co/microsoft/TRELLIS.2-4B establish 2025 as the release year for the .2 variant.\n\nThe model requires NVIDIA GPU with >=24GB VRAM CUDA 12.4 toolkit and Linux according to https://microsoft.github.io/TRELLIS.2/ and the GitHub README confirming it is CUDA-dependent while an earlier non-.2 TRELLIS model appears at https://huggingface.co/microsoft/TRELLIS-image-large with December 2024 assets.\n\nPrimary sources including the 2025 tech report https://github.com/microsoft/TRELLIS.2 and https://huggingface.co/microsoft/TRELLIS.2-4B establish the corrected 2025 date for TRELLIS.2 updating the prior claim of a 2024 release." }