ml-sharp-web: A browser-based Gaussian splat generator using Apple's SHARP model.

ml-sharp-web is a highly practical, browser-native implementation of the SHARP (Single-image High-resolution Appearance Reconstruction Program) principle. At its core, it solves the challenging task of inferring a 3D scene representation (specifically, Gaussian splats) from a single 2D photograph. Its strength lies in its complete client-side execution model, reducing reliance on backend API calls and making it immediately accessible to any user with a modern browser. From a technical standpoint, the stack is robust and modern. It combines React/TypeScript for the UI, Vite for bundling, and critically, ONNX Runtime Web for the computationally intensive AI inference. The design correctly handles the complexities of model deployment, noting that the SHARP export requires serving both the `.onnx` graph and the necessary `.onnx.data` weights sidecar, which is a common pitfall in web ML deployment. The inclusion of GaussianSplats3D suggests the necessary visualization plumbing is already integrated. The development process outlined is detailed, spanning from dependency installation (using Bun) to the necessary upstream export steps using Python and specific exporters. This level of detail, including explicit warnings about memory usage and required browser capabilities (WebGPU/WASM), speaks to the author's deep understanding of the operational constraints of running large models in a constrained environment. However, this very complexity presents a limitation: deployment and optimal performance are highly sensitive to browser hardware and current WASM/WebGPU support, meaning the user experience could be inconsistent. In summary, this is not merely a demo; it's a working, well-documented proof-of-concept for deploying advanced academic computer vision models (like SHARP) into a readily consumable, self-contained web application. It serves as an excellent reference architecture for developers looking to push the boundaries of client-side AI, provided they are willing to contend with the performance overheads of large ONNX models in the browser environment.