An indie developer has successfully ported Apple's SHARP model for single-image 3D reconstruction to run entirely in web browsers using ONNX Runtime Web. The project, posted to Hacker News on May 3, 2026 by user bring-shrubbery, allows users to upload images and receive downloadable 3D models through client-side processing, though it has encountered significant technical limitations.
Technical Implementation Enables Client-Side 3D Reconstruction
The developer exported Apple's SHARP model (single-image 3D Gaussian splatting) to ONNX format and deployed it using ONNX Runtime Web with WebGPU support. The implementation processes images entirely within the browser, maintaining user privacy by keeping photos local rather than uploading them to servers. Inference typically takes a few seconds on recent Mac hardware.
Key technical achievements include:
- Successfully running a 2.4 GB model in browser environments
- WebGPU acceleration for improved performance
- No server-side processing required
- Direct 3D model downloads from browser-processed results
Browser Compatibility and Memory Issues Limit Accessibility
Despite the technical accomplishment, users reported significant practical limitations. Multiple commenters experienced browser tab crashes due to the model's high memory requirements. The 2.4 GB model size pushes the boundaries of what current browsers can handle reliably.
Browser compatibility proved inconsistent. The project failed to run in Firefox on Linux, though it functioned in Chrome. This reflects the cutting-edge nature of WebGPU, which only recently became available across browsers and remains incompletely supported.
Model Accuracy Questioned for Real-World Photos
While the technical port succeeded, users noted that results from personal photos were noticeably worse than the demonstration images. This sparked broader discussion about fundamental limitations of monocular depth estimation.
Several commenters debated whether single-image 3D reconstruction inherently "makes up" depth information that isn't physically present in 2D images. Multi-view approaches that triangulate actual geometry from multiple photos were cited as producing more accurate results, though they require additional input data.
Privacy Benefits Offset by Technical Constraints
The community appreciated the privacy advantages of client-side processing, with several expressing interest in similar browser-based AI imagery applications. By keeping all processing local, users avoid uploading potentially sensitive photos to external servers.
However, the practical limitations around memory consumption, browser compatibility, and model accuracy tempered enthusiasm. The project demonstrates both the expanding capabilities of in-browser AI and the real-world constraints that still limit widespread adoption.
Key Takeaways
- Developer successfully ported Apple's 2.4 GB SHARP 3D reconstruction model to run in browsers using ONNX Runtime Web and WebGPU
- Client-side processing maintains user privacy but causes tab crashes due to high memory requirements for many users
- Browser compatibility remains limited, with the project failing in Firefox on Linux while working in Chrome
- Single-image 3D reconstruction inherently involves estimating depth information not present in 2D photos, leading to accuracy questions for real-world images
- The project highlights both the potential and current limitations of in-browser AI, with cutting-edge capabilities offset by practical constraints