Get ready for an exciting moment.
The future of Radiance Field-related technologies continues to shine with breakthroughs that lower barriers to creating hyper real 3D content. One exciting development came through marked by a single innocuous Twitter post from Arthur Brussee announcing Brush—a 3D reconstruction engine built on the Radiance Field method, Gaussian splatting. With Brush, the dream of high-quality 3D reconstruction accessible on nearly any device moves closer to reality.
Training a Radiance Field has never been more accessible than today and now you’re able to train natively in a browser, on any computer architecture. Whether you’re using an NVIDIA GPU, an AMD card, an Apple device, or an Android phone, Brush functions seamlessly—even in a browser! Current compatibility is limited to Chrome (with Safari and Firefox support on the way), making it one of the most accessible 3D reconstruction tools available.
Arthur will also be our guest on this week’s episode of the View Dependent Podcast on Thursday. If you’d like to learn more about how Brush works, its current features, and its upcoming ones, I highly recommend you subscribe to the channel!
Brush is designed to be highly portable, flexible, and fast, enabling 3D reconstruction without specialized hardware. Supporting a wide range of systems—macOS, Windows, Linux, AMD, Nvidia GPUs, Android, and even browsers—Brush demonstrates the potential of WebGPU-compatible technology and Rust. It leverages the Burn framework with a portable wgpu backend, bringing cross-platform compatibility for users to render and train in nearly any environment. Although still a proof of concept, Brush has already proven impressive, even without advanced optimizations or extensions to Gaussian Splatting.
While Brush is labeled an experimental demo, you can try it now. Start by downloading a pre-assembled Radiance Field scene used for benchmarking. Simply click the load file containing a ZIP of your training images, paired with the necessary transform file from COLMAP), to begin training directly in the browser—no complex setup required. Additionally, with the release of 0.0.1, there are now precompiled binaries for both Mac and Windows!.
Brush does work with custom data sets, but it’s a little easier to replicate your data into the same structure as the reference examples I’ve found.
While it hasn’t been fully explored yet, it could be possible to extend George Kopanas’s Slang.D with Brush. Similar to Kopanas, despite Arthur Brussee’s affiliation with Google, Brush is not an official Google product. It’s an independent project—one that showcases how the barriers to adoption continue to fall.
More information about the Apache 2.0 licensed project can be found on Brush’s GitHub page and the in Browser demo can be found here.
