Brush

Overview
Brush is an experimental 3D reconstruction engine leveraging the Radiance Field method, Gaussian Splatting. Developed by Arthur Brussee, Brush aims to democratize access to high-quality 3D reconstruction by enabling Radiance Field training on nearly any device, including browsers. This makes it one of the most accessible and portable tools for 3D reconstruction currently available.

What Makes Brush Unique?
Traditional 3D reconstruction often requires powerful hardware and complex setups, limiting accessibility. Brush changes that by running natively on macOS, Windows, Linux, Android, or directly in a browser with WebGPU compatibility. Users can train Radiance Fields without the need for specialized equipment, making 3D modeling more inclusive and versatile.

Key Features of Brush

• Cross-Platform Compatibility: Supports macOS, Windows, Linux, AMD, NVIDIA GPUs, and Android, demonstrating remarkable flexibility.

• High quality: Brush aims for the highest quality training achievable without complicated tweaks.

• Simple Setup: Load pre-assembled Radiance Field scenes or custom datasets with minimal configuration. Compatible with COLMAP transforms and other common formats for easy integration.

• Open Source: Licensed under Apache 2.0, Brush is freely available for any use and customization.

• Browser-Based Training: Train Radiance Fields directly in Chrome, with upcoming support for Safari and Firefox. No installation is needed beyond the browser itself.

Who Can Benefit from Brush?
Brush is an excellent tool for developers, researchers, and hobbyists exploring 3D reconstruction technologies. Its browser-based workflow is ideal for educators introducing students to Radiance Fields without requiring advanced hardware. Additionally, digital artists and VR creators can leverage its portability to prototype and visualize scenes quickly. Brush also plays back dynamic splats as PLY sequences, and NVIDIA Research uses it to power the interactive dynamic 3D Gaussian previews on the project page for World from Motion, its method for generative dynamic Gaussian reconstruction from monocular video.

Getting Started with Brush
Trying Brush is simple. Start by downloading a pre-assembled Radiance Field scene or use your own dataset. Upload your training images and COLMAP transform file directly into the browser to begin training—no complex setup required. For offline workflows, each release ships precompiled desktop builds for Apple Silicon macOS, x64 Windows and x64 Linux, and the latest is v0.3.0. Brush can also be built from source with Rust.

Future Potential
Brush has moved past its proof of concept stage, with v0.3.0 adding MCMC style training that still grows splats automatically, training on datasets larger than RAM and a redesigned interface, and its flexibility opens the door to future expansions, such as integrating with advanced tools like George Kopanas’s Slang.D based splatting. Though Arthur Brussee works at Google, Brush is not an official Google product.

Learn More
To explore Brush further, visit the official GitHub page and try the in-browser demo. Stay tuned for updates as Brush evolves into a more robust platform for 3D reconstruction.