Overview
WebGPU 3D Gaussian Splatting Viewer — A real-time renderer implementing the state-of-the-art 3D Gaussian Splatting technique using WebGPU. This project enables interactive visualization of neural radiance field scenes directly in the browser with GPU-accelerated performance.
Description
Implemented a real-time 3D Gaussian Splatting renderer using WebGPU, enabling high-fidelity neural radiance field visualization directly in the browser. The project transforms millions of 3D Gaussians into smooth, photorealistic images through GPU-accelerated compute and render pipelines.
The rendering pipeline consists of a preprocessing compute shader that performs view frustum culling, computes 3D-to-2D covariance projection using EWA splatting, and evaluates Spherical Harmonics (up to degree 3) for view-dependent color. A parallel radix sort orders Gaussians by depth for correct alpha blending, while indirect draw calls dynamically render only visible splats.
- Implemented EWA splatting algorithm to project 3D Gaussian covariances into 2D screen-space ellipses using Jacobian-based perspective projection
- Computed view-dependent colors using Spherical Harmonics (SH) coefficients up to degree 3 for realistic lighting
- Developed GPU-based radix sort for depth ordering to enable correct back-to-front alpha compositing
- Optimized rendering with view frustum culling, reducing GPU workload by discarding off-screen Gaussians
- Utilized WebGPU compute shaders for parallel preprocessing of millions of Gaussian primitives
- Implemented premultiplied alpha blending with indirect draw calls for efficient splat rasterization