Pull down to go back
Show HN: Run TRELLIS.2 Image-to-3D generation natively on Apple Silicon
opensourcehackernews

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

Show HN: 在 Apple Silicon 上原生執行 TRELLIS.2 圖像轉 3D 生成模型

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.

Tech Blogger Take

Someone just cracked the code on running Microsoft's 4B-parameter 3D AI on MacBooks. This changes everything.

Microsoft dropped TRELLIS.2 — a beast of a model that turns any photo into a 3D object — but it was locked to NVIDIA GPUs with all their CUDA-specific magic. Flash attention, custom sparse convolutions, CUDA hashmaps — the works. Then some absolute legend on Show HN said 'hold my coffee' and ported the entire thing to run natively on Apple Silicon. We're talking a few hundred lines of code that replaced CUDA-only operations with pure PyTorch alternatives. The kicker? It actually works. You can now run a 4-billion parameter image-to-3D model on your M3 MacBook Pro. No cloud, no NVIDIA tax, no waiting in line for GPU time. This is the kind of democratization that makes entire industries pivot overnight.

VerdictIf you have an M-series Mac, clone this repo right now and watch your laptop turn photos into 3D models — the future just became portable.
9/10

Action

馬上試用
https://github.com/microsoft/TRELLIS
Open SourceMacPython
1Clone the Apple Silicon port repository from the Show HN post
2Install PyTorch with MPS support on your M-series Mac
3Run the image-to-3D generation script with a test photo
Before

Needing expensive NVIDIA GPUs and cloud services to generate 3D models from photos

After

Running 4B-parameter 3D generation locally on your MacBook Pro in minutes

AI Analysis

3D Content Creation

high
Action Required

Download this port immediately and test it on your M-series Mac — this could replace your entire 3D modeling pipeline for concept work

Key Insight

A 4-billion parameter model that turns photos into 3D objects is now running on laptops, not just $10K GPU rigs

Why It Matters

You can now prototype 3D assets during client meetings instead of waiting days for your render farm

Game Development

high
Action Required

Test this for rapid prototyping — imagine turning concept art into playable 3D assets in minutes

Key Insight

The technical barriers that kept AI 3D generation locked to NVIDIA cards just got demolished with pure PyTorch alternatives

Why It Matters

Your indie game studio can now compete with AAA asset pipelines using nothing but MacBooks

Job Impact Analysis

3D Artist

Role Shift
Why It Impacts

Microsoft's TRELLIS.2 can generate 3D models from photos, and now it runs on the MacBook Pro you already own

How to Adapt

Start experimenting with this for concept work — it won't replace your skills but it'll 10x your iteration speed

iOS Developer

Opportunity
Why It Impacts

Apple Silicon optimization means you can now generate 3D content locally for AR apps without cloud dependencies

How to Adapt

Fork this repo and explore integrating 3D generation directly into your ARKit workflows

Read original →

Keywords

image-to-3Dmodel portingApple SiliconMPSCUDA alternativesparse convolutionmesh extraction

Glossary

TRELLIS.2（TRELLIS.2模型）
Microsoft's 4-billion parameter AI model that generates 3D objects from 2D images, originally requiring expensive NVIDIA GPUs but now ported to run on Apple Silicon.
PyTorch MPS（PyTorch金屬性能著色器）
Apple's Metal Performance Shaders backend for PyTorch that lets AI models run efficiently on Apple Silicon chips, the key technology enabling this CUDA-to-Mac port.
Sparse Convolution（稀疏卷積）
A memory-efficient neural network operation that only processes non-zero values, crucial for 3D generation but originally implemented with CUDA-specific code that this port replaced with pure PyTorch.
Flash Attention（快閃注意力機制）
A memory-efficient attention mechanism typically requiring CUDA, but this port replaced it with PyTorch's SDPA attention to make the 3D generation model work on Macs.

Related Articles

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

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

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).

Reddit
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