Pull down to go back
TRELLIS.2 image-to-3D now runs on Mac (Apple Silicon) - no NVIDIA GPU needed
opensourcereddit

TRELLIS.2 image-to-3D now runs on Mac (Apple Silicon) - no NVIDIA GPU needed

TRELLIS.2 圖像轉 3D 現在可在 Mac（Apple Silicon）上運行 - 無需 NVIDIA GPU

I ported Microsoft's TRELLIS.2 to run on Apple Silicon via PyTorch MPS. The original depends on five CUDA-only compiled extensions (flex_gemm, flash_attn, o_voxel, cumesh, nvdiffrast) that have no Mac equivalent. Wrote replacement backends from scratch: Pure-PyTorch sparse 3D convolution (replacing flex_gemm), Python mesh extraction using spatial hashing (replacing CUDA hashmap ops in o_voxel), SDPA attention for sparse transformers (replacing flash_attn), and GPU-accelerated trilinear interpolation (replacing cumesh and nvdiffrast).

Tech Blogger Take

Someone just freed Microsoft's best 3D AI from NVIDIA jail — and your MacBook is about to become a 3D modeling powerhouse

Microsoft's TRELLIS.2 can turn any image into a 3D model, but it was locked behind five CUDA-only libraries that made it Windows-with-NVIDIA-only. Some absolute legend just spent weeks rewriting every single one from scratch to work on Apple Silicon. We're talking pure-PyTorch sparse 3D convolution, custom mesh extraction using spatial hashing, and GPU-accelerated attention that doesn't need CUDA. This isn't just a port — it's a complete reimagining of how 3D AI should work. The fact that one person could reverse-engineer and replace Microsoft's entire CUDA stack shows how artificial these hardware barriers really are. Your M1 MacBook Pro is about to do things that required a $5000 workstation last month.

VerdictDownload this immediately and watch your laptop turn photos into 3D models — the democratization of 3D AI just hit warp speed.
9/10

Action

馬上試用
https://github.com/microsoft/TRELLIS
Open SourceMacCLI
1Clone the Mac-compatible fork from the GitHub repository
2Install PyTorch with MPS support and required dependencies
3Run the demo script with any image to generate your first 3D model
Before

Needing a $3000+ NVIDIA workstation to run cutting-edge image-to-3D AI models

After

Generating 3D models from photos directly on your MacBook Pro using Apple Silicon

AI Analysis

3D Content Creation

high
Action Required

Test TRELLIS.2 on your Mac immediately — this could replace your entire 3D modeling pipeline for concept work

Key Insight

Someone just made Microsoft's cutting-edge image-to-3D tech work on laptops by rewriting five entire CUDA libraries from scratch

Why It Matters

You no longer need a $3000 NVIDIA workstation to turn sketches into 3D models — your MacBook Pro can do it

Game Development

high
Action Required

Download this port and start prototyping 3D assets from concept art — could revolutionize your asset pipeline

Key Insight

The technical barriers between Mac developers and AI-powered 3D generation just disappeared overnight

Why It Matters

Your indie game studio can now generate 3D models without buying expensive hardware or switching to Windows

Job Impact Analysis

3D Artist

Role Shift
Why It Impacts

Image-to-3D generation is now accessible on standard Mac hardware, democratizing advanced 3D creation tools

How to Adapt

Learn this workflow now — clients will expect faster turnaround times when they discover you can generate base models from references

Indie Game Developer

Opportunity
Why It Impacts

Mac developers can now access Microsoft's state-of-the-art 3D generation without needing NVIDIA GPUs

How to Adapt

Integrate this into your asset pipeline immediately — prototype faster and spend budget on polish instead of modeling

AI Researcher

Opportunity
Why It Impacts

Someone proved you can replace complex CUDA operations with pure PyTorch, making research more accessible

How to Adapt

Study this port's architecture — it's a masterclass in making GPU-locked research universally accessible

Read original →

Keywords

image-to-3DMac compatibilityCUDA replacementPyTorch MPSsparse convolutionmesh extractionflash attention alternative

Glossary

TRELLIS.2（TRELLIS.2）
Microsoft's state-of-the-art AI model that converts 2D images into 3D models, originally requiring expensive NVIDIA hardware until this Mac port broke it free.
PyTorch MPS（PyTorch MPS）
Apple's Metal Performance Shaders backend for PyTorch, letting AI models run on Mac GPUs instead of requiring NVIDIA CUDA.
Sparse 3D Convolution（稀疏3D卷積）
A way to process 3D data efficiently by only computing on non-empty voxels, which this port reimplemented without CUDA dependencies.
Flash Attention（Flash注意力機制）
Memory-efficient attention mechanism that was CUDA-only until this port replaced it with PyTorch's SDPA for Mac compatibility.
Mesh Extraction（網格提取）
Converting 3D voxel data into triangle meshes, which this port achieved using spatial hashing instead of CUDA hashmaps.

Related Articles

Show HN: Run TRELLIS.2 Image-to-3D generation natively on Apple Silicon
Open SourceRead

Show HN: Run TRELLIS.2 Image-to-3D generation natively on Apple Silicon

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 gather-scatter sparse 3D convolution, SDPA attention for sparse transformers, and a Python-based mesh extraction replacing CUDA hashmap operations. Total changes are a few hundred lines across 9 files.

Hacker News
157
Which Qwen 3.6 Quantization Works Best for M5 Pro with 24GB RAM?
Open SourceRead

Which Qwen 3.6 Quantization Works Best for M5 Pro with 24GB RAM?

A user with an M5 Pro machine featuring 24GB of RAM is trying to figure out the best quantization level for running Qwen 3.6 via Ollama. They're unsure whether to go with Q4 (4-bit quantization) but can't find a solid Q3 (3-bit) alternative. They're looking for recommendations on which version would run smoothly on their setup without compromising too much on quality.

Reddit