bring-shrubbery · 2026-05-03 · notable
ml-sharp-web — Apple's SHARP 3D Gaussian Splat Model Runs Entirely In Your Browser
Show HN demo exports Apple's SHARP photo-to-3D model to ONNX and runs it client-side via WebAssembly. Drop in an image, get a downloadable .ply Gaussian splat, no server.
Apple's one-shot photo-to-3D Gaussian splat model, ported to ONNX and shipped as a 100% in-browser web app.
Key specs
| Parameters | 2.4 GB ONNX |
|---|---|
| GitHub stars | 201 |
| Hn points | 171 |
| Hn comments | 43 |
What is it?
ml-sharp-web is a TypeScript web app that runs Apple's SHARP — a 3D Gaussian splatting model that turns a single photo into a navigable 3D scene — entirely in the browser via ONNX Runtime Web. The author (bring-shrubbery) exported SHARP from Apple's official ml-sharp repo to ONNX and wrapped it in a Vite/React UI with a live splat viewer. A Show HN posted on May 3 hit 171 points and 43 comments.
How does it work?
SHARP itself is a feedforward neural network that regresses a 3D Gaussian scene representation from a single image in one pass. The author exports the PyTorch model to ONNX (sharp_web_predictor.onnx, plus a ~2.4 GB weights sidecar), then loads it with onnxruntime-web compiled to WebAssembly. Inference runs locally on the user's CPU/GPU; the resulting splats are rendered with the GaussianSplats3D library and exportable as a standard .ply file.
Why does it matter?
Browser inference for a 2.4 GB scene-synthesis model removes the cloud-GPU dependency that has gated most splatting demos. Anyone with Chrome and enough RAM can drop a photo in and walk around a 3D reconstruction. It is also a reference implementation for shipping large research models as static web apps via ONNX Runtime Web — relevant beyond view synthesis.
Who is it for?
graphics researchers, web developers, and anyone exploring on-device 3D reconstruction
Try it
git clone https://github.com/bring-shrubbery/ml-sharp-web && bun install && bun dev