ml-sharp-web

This project, known as ml-sharp-web, serves as a browser-based Gaussian splat generator utilizing Apple's SHARP model. Designed for ease of use, it empowers users to generate Gaussian splats directly in their web browsers, enabling fast and interactive visualization of results. Here are the key features of the application:

• Upload a single image to begin the processing.
• Generate Gaussian splats in real-time with a simple click.
• Preview the resulting visual output directly in the browser.
• Download the generated output as a .ply file for further usage.

Key Links

• Project repository: ml-sharp-web
• Follow the author on X: @bringshrubberyy
• Explore the upstream SHARP repository by Apple: apple/ml-sharp
• Documentation for SHARP: apple.github.io/ml-sharp
• Reference paper: arXiv:2512.10685

Requirements

To efficiently utilize this project, the following prerequisites are essential:

• Bun: Ensure the Bun environment is installed.
• Browser Compatibility: Use a modern desktop browser with recommended versions like Chrome or Edge.
• System Resources: Adequate disk space and RAM to handle the SHARP model, noting that it might require approximately 2.4 GB due to the large exported ONNX sidecar.

Important Model File Consideration

The SHARP model generates two essential files for proper operation:

• sharp_web_predictor.onnx (the main model)
• sharp_web_predictor.onnx.data (contains model weights)

Both files must be hosted together in the same directory for the application to function correctly, preferably using the default hosted model to avoid errors.

Quick Start

Users can quickly get the application up and running with minimal effort. After installation, the app can be started, allowing image uploads and Gaussian splat generation in a seamless user experience.

Technical Overview

The application's architecture incorporates:

• React + TypeScript for a responsive user interface.
• ONNX Runtime Web Worker for efficient inference execution.
• Browser-side conversions and processing for optimized performance.

Status and Performance

This application is currently classified as a working prototype. While it functions effectively, performance may vary based on the browser's WebGPU/WASM support and the machine's available resources. Users may encounter performance constraints, particularly with larger models or higher generation demands.