A new open‑source project called ml-sharp-web lets developers upload an image and generate Gaussian splats entirely in the browser. Built on top of Apple's SHARP model, the tool renders a preview and offers a downloadable .ply file. The interface runs on React, TypeScript and Vite, using ONNX Runtime Web for on‑device inference.

To function, the app requires the SHARP model exported to ONNX, which consists of a ~2.4 GB .onnx.data sidecar file alongside the graph definition. Users must clone Apple’s upstream repository, run a Python export script, and host both files together. The setup script also copies WebAssembly assets into the public folder, enabling fast inference on Chrome or Edge.

Performance hinges on browser WebGPU and available RAM; initial model loading can be slow, and large gauss‑splat counts may cause crashes. Despite these constraints, the prototype demonstrates that high‑fidelity point‑cloud generation can be achieved without server resources, opening possibilities for client‑side visual effects and rapid prototyping.