You Can Now Run State-of-the-Art Image-to-3D on Your Mac — No GPU, No Cloud, No Subscription

Show HN is, at its best, a reliable signal for "something in your workflow just got materially cheaper." Today's example is trellis-mac, a port of Microsoft's TRELLIS.2 image-to-3D model from CUDA-only to Apple Silicon via PyTorch MPS. It hit the Hacker News front page overnight. If you ship anything 3D-adjacent as a solo operator — indie games, product viz, AR try-ons, generative asset pipelines — your unit economics just shifted.

What TRELLIS.2 Actually Does

TRELLIS.2 is a 4-billion-parameter model from Microsoft Research that takes a single 2D image and produces a textured 3D asset. Not a point cloud, not a rough mesh — a proper PBR-textured model at up to 1536³ resolution. It's built on native 3D VAEs with 16× spatial compression, and the output is the kind of thing you can drop into Blender, Unity, or Three.js with minimal cleanup.

State of the art, in the "you can pay fal.ai or Tripo to run this in the cloud" sense, has been in this neighborhood for about a year. The new thing isn't the model. The new thing is where it can run.

The Port: Grunt Work, Not Magic

The interesting part of trellis-mac is how unglamorous it is. TRELLIS.2 depends on several CUDA-only libraries — the kinds of things that work beautifully on an H100 and segfault on anything else. The port replaces them with pure-PyTorch or pure-Python alternatives, patches every hardcoded .cuda() call to use the active device, and makes sure everything still lines up numerically.

No model surgery. No clever distillation. Just a careful read through the dependency tree, a handful of replacement implementations, and a bunch of testing. This is the part of the indie AI ecosystem that doesn't get written up enough. The interesting work is often "make the expensive thing run on consumer hardware" rather than "train a new thing."

If you've ever wondered what the gap between "research code" and "usable code" looks like in practice, go read the patch. It's the whole playbook: pin a working commit, identify the CUDA-specific call sites, replace them with MPS-compatible equivalents, handle the places where MPS is subtly different from CUDA (there are more than you'd expect), and write a test that actually renders a mesh.

The Numbers That Matter

On an M4 Pro with 24 GB of unified memory, TRELLIS.2 generates roughly 400,000-vertex meshes from a single photo in about 3.5 minutes. On an H100 the same task takes seconds. That gap sounds huge and it is huge, but the unit economics point in an unexpected direction.

An H100 rents for roughly $2–$4 per hour depending on provider. At three and a half minutes per asset, you're paying somewhere around $0.12 to $0.23 per asset on cloud hardware, plus bandwidth if you're shipping images up and meshes down. An M4 Pro you already own is zero marginal cost per asset. The only real cost is the wall-clock time, which matters for interactive workloads and doesn't matter at all for batch generation done overnight.

Put differently: cloud 3D generation is better if you need a single asset now. Local is better if you need a hundred assets by tomorrow, and much better if privacy or offline capability is a product requirement.

What This Unlocks for Solo Projects

A few concrete scenarios where the math flips:

Indie game asset pipelines. A small team or solo developer shipping a game with a few hundred unique props has been stuck between hand-modeling everything (slow) and paying a cloud provider per asset (scales linearly with the game's content volume). Local generation overnight on a Mac you already own turns the asset pipeline into a fixed cost.

E-commerce product-to-3D. If your SaaS lets a merchant upload a product photo and generate a 3D view for their store, your per-user cost used to be a real number. Now you have a choice: keep cloud generation for speed-sensitive flows, and offer a "batch process overnight, cheaper" tier that runs on your Mac Studio or a merchant's own hardware.

AR try-on and preview tools. These are basically dead without fast 3D generation. They're still hard live, but they're now plausible as "upload a photo, get an AR view in a few minutes." Previously the cloud bill made them untenable for any solo product without a premium price point.

Private use cases. Anyone working with proprietary product photography, architectural assets, or IP-sensitive imagery that can't legally leave the company's infrastructure — you're covered now, on consumer hardware. That's genuinely new.

The Broader Pattern

"Native on Mac" is quietly becoming the default for hobbyist and solo AI workflows. The trajectory is obvious if you list what now runs locally with reasonable performance:

llama.cpp turned multi-billion-parameter language models into something that runs on a laptop. MLX did the same, with better Mac-native performance, for anyone doing serious research on Apple hardware. whisper.cpp makes transcription trivially local. Stable Diffusion has been local-first for years. Now TRELLIS.2 for 3D generation.

The pattern is the same every time: a CUDA-first research model gets ported, the port is imperfect but usable, and a class of products that was previously only economical on cloud GPU becomes economical on hardware you already own. The solo operator who bets on this trend saves a meaningful amount of money. The solo operator who ignores it is paying a recurring cloud tax they don't need to pay.

None of this makes cloud obsolete. H100-class speed is real, and for interactive or time-sensitive workloads the cloud wins. But for solo builders who are bandwidth-constrained on money rather than time — which is most of us — the local-first path is quietly winning.

How to Try It

Clone shivampkumar/trellis-mac from GitHub. You'll need macOS with Apple Silicon (M1 or later; M4 Pro or M4 Max is the sweet spot), a recent Python, and enough patience to let the first dependency install finish. Pull the model weights from the original Microsoft TRELLIS.2 release. Feed it a single image, wait three and a half minutes, open the output .glb in Blender.

You'll generate your first asset locally for free. That's the whole demo. The interesting question is the one that comes next: what product can you ship now that was previously priced out?

For a solo operator, that's the real question. The model got cheaper. The product opportunity just showed up. What you do with it is the part that still matters.