Radiance Field-based technologies and platforms continue to accelerate at breakneck speed, and Brush is no exception. Back in our previous coverage, we touched on how Brush, created by Arthur Brussee, harnesses the power of Gaussian Splatting to deliver accessible, cross-platform 3D reconstructions. Specifically, if you’re on Mac, Brush is a great tool to train your captures locally.
Now, with the release of Brush 0.2, the project takes a large step forward, transforming from a experimental demo into a more mature tool capable of tackling real-world data with refined performance, higher quality, and a smooth workflow.
While still implementing what Brussee calls the “basics” of Gaussian Splatting, Brush 0.2 has undergone a significant evolution. According to early tests, training now runs just as fast as competing methods, if not slightly faster, while delivering a bump in visual fidelity. On my M2, it’s pulling about 17 steps a second. With new executables, installing Brush onto Mac was honestly very easy.
Key Features in Brush 0.2
Higher Reconstruction Fidelity
Brush now measures improved image metrics on mipnerf360 scenes, showcasing that it’s no longer merely a playful tech demo. While tuning gsplat could potentially yield close numbers, Brush’s performance is undeniably made substantial progress.Faster Training & Memory Optimization
Significant kernel and memory optimizations provide a more streamlined training process. Users will likely notice quicker results and a smoother overall experience—especially on less powerful hardware.Command Line Interface (CLI)
For those who prefer working outside of a web browser, Brush now supports a comprehensive CLI. Simply invoketo see available commands, which can be combined with
to monitor training interactively.
Enhanced Navigation & Flythrough Controls
A new suite of camera controls—orbital, first-person (FPS), flythrough, and panning—gives users precise command over scene exploration. The “controls” popout in the scene viewer details all these modes for intuitive manipulation.Web-Friendly Improvements
Data Streaming: Load data directly from a URL
‘Zen’ Mode: Add
to the URL for a simplified, fullscreen interface—perfect for demos or immersive viewing sessions.
Dynamic Splats: Import sequences of PLY files or Brush’s custom “PLY with delta frames” format to visualize objects in motion. This functionality has already been used for capturing 4D data like Cat4D and Cap4D scenes.
Support for Transparent Images & Masks
Handling of alpha channels means Brush can respect transparency in source images—ideal for training data that includes backgrounds you want to ignore. Alternatively, users can supply a dedicated folder of mask images to exclude unwanted regions.Flexible COLMAP & Nerfstudio Integration
Brush 0.2 offers broader compatibility with standard dataset formats. If your data includes an initial PLY point cloud, Brush will take advantage of it for improved reconstructions. There’s also a new logic that rotates the up-axis so that “ground is ground,” making final visualizations more intuitive.
For those who like to train 3DGS in the browser, Chrome remains the primary supported platform—though support for other browsers is still on the roadmap. Several demo scenes are available, including the 650MB Garden scene and a small Mushroom dataset.
If you prefer native installs, you can pick up the Brush 0.2 release for macOS, Windows, or Linux. Just choose the file matching your system architecture from the official GitHub releases page. Each platform bundle includes the new CLI and all the latest enhancements.
The next version of Brush is gearing up to introduce more advanced Gaussian Splatting techniques, pushing the platform’s quality, speed, and resilience beyond its already impressive capabilities. These upcoming changes may affect compatibility with other Gaussian Splat viewers, but Arthur Brussee’s retooled web viewer aims to deliver a frictionless experience for creators and enthusiasts alike.
Brush remains Apache 2.0 Licensed and can be accessed from its Github page. If you’re aiming to train Gaussian Splatting on Mac locally, you will be hard-pressed to find a better option out there.
