// the find
ArthurBrussee/brush
3D Reconstruction for all
Brush is a Gaussian splatting training and viewing engine written in Rust, built on the Burn ML framework and WebGPU. It runs natively on Windows/macOS/Linux/Android and in the browser from a single codebase. Aimed at 3D reconstruction practitioners who want something that actually ships cross-platform without a CUDA dependency chain.
WebGPU-first design means one renderer targets everything — desktop GPUs, mobile, and browser — without maintaining separate backends. Using Burn instead of PyTorch cuts the CUDA requirement entirely, producing standalone binaries with no environment setup. The live training viewer with per-view comparison during optimization is genuinely useful for debugging, not just a demo gimmick. Benchmark results claiming faster rendering and training than gsplat, with `cargo bench` to verify, is the right way to make that claim.
WebGPU browser support is still Chrome/Edge-only; Firefox and Safari are listed as 'hopefully soon,' which is not a timeline you can plan around if you need broad web reach. Android builds require manually rebuilding the Rust code before running from Android Studio, a friction point that will catch out anyone used to IDE-managed builds. The Burn ML framework is considerably less mature than PyTorch — you are betting on it not having critical regressions in your training kernels, and the ecosystem for debugging or extending it is thin. COLMAP and Nerfstudio formats are supported but RealityCapture support appears partially present in the tree without much documentation, so format coverage may be incomplete.