Show HN: Apple's Sharp Running in the Browser via ONNX Runtime Web

A new open-source project, ml-sharp-web, brings Apple's SHARP (Single-image High-quality 3D Asset Representation) model directly into the browser. Built by developer bring-shrubbery, it uses ONNX Runtime Web to run the 2.4GB SHARP model on the client side, enabling Gaussian splat generation from a single image without any cloud backend. Users simply upload an image, click "Generate Splat", preview the result, and download a .ply file.

The project requires Bun for local development and a modern desktop browser (Chrome or Edge recommended). It exports SHARP to ONNX format (two files: .onnx and .onnx.data) and performs all inference and postprocessing—including NDC-to-metric Gaussian conversion and PLY writing—inside the browser via a web worker. While still experimental, ml-sharp-web demonstrates the feasibility of running large 3D generative models on consumer hardware, though performance can be slow and memory-intensive. The code is licensed separately from Apple's SHARP model weights, which carry research-use restrictions.