Real-time, cross-platform 3D reconstruction is shifting from research-only code to tools you can run on a phone or in a browser. Brush demonstrates that Gaussian-splatting-based pipelines can be trained interactively and rendered live across desktop, mobile and WASM targets by combining WebGPU-compatible rendering with the Burn machine-learning framework.
What Sets It Apart
- Cross-platform first: targets macOS/Windows/Linux (native), Android, and Web (WASM via WebGPU). This reduces the friction of shipping demos and running experiments on non-CUDA devices. So what: you can prototype and inspect results on the same hardware your users will run.
- Live training visualization: training can run natively or in-browser while you interact with the scene and compare renders to input views. So what: it speeds iteration and debugging of reconstruction pipelines by making training dynamics visible.
- Lightweight deployment: produces dependency-free binaries and a browser demo, avoiding heavy CUDA-only stacks. So what: easier distribution to devices without complex GPU driver or CUDA requirements.
- Practical viewer features: supports streaming .ply/.compressed.ply, .zip animations, masks/transparency, and rerun integration for richer visual diagnostics. Benchmarks in the repo report rendering and training are generally faster than gsplat in common kernels, making it suitable when performance matters.
Who It's For & Trade-offs
Great fit if you want an open-source, experiment-friendly 3D reconstruction engine that you can run and visualize across devices (desktop, Android, and Chrome/Edge browsers) without heavy CUDA dependencies. It’s also useful for researchers and demo builders who need live training feedback and .ply streaming.
Look elsewhere if you need a production-grade, fully managed pipeline with wide browser support and polished cross-browser WebGPU fallbacks—browser usage currently requires recent Chrome/Edge builds (README notes Chrome 134+ on some platforms), and WebGPU/device support can vary across vendors. Also, on-device training still depends on available compute—large datasets or high-resolution scenes may be slow on mobile. Note: this repo is a community fork of a Google Research project and is not an official Google product.
