Vauss turns one video into a splat, a Blender pack and a voxel game

Voyager has released Vauss, a Gaussian splat trainer and reconstruction CLI, as a research preview. It builds on Arthur Brussee’s Brush engine and is written in native Rust with Burn, CubeCL and wgpu, so training runs on Vulkan, Metal or DX12 without CUDA. The latest build so far is the release candidate vauss-v0.1.25-rc, tagged on 14 November 2026.
A single vauss run command checks the capture, selects frames, solves the cameras, trains the splat, evaluates held-out views and writes the exports. --fast gives a quick preview and --max a larger training budget. Stitched 360 video is detected automatically, and fisheye or action-cam footage can be trained on the fisheye lens itself with --camera-model opencv-fisheye. Existing Nerfstudio transforms.json and COLMAP folders are accepted too.
One run produces a browser viewer that opens straight from disk, a Blender pack with the full splat PLY, surface voxels and a textured mesh where available, and a scene.spz for the web or for Blender 5.3. A collision mesh, convex pieces and a floor estimate export with a Unity import folder, and the same capture becomes a first-person voxel game with destructible voxels, where V switches between splats and voxels. vauss shot renders new orbits and push-ins, and vauss stabilize re-renders shaky footage along a smooth path.
On seven Mip-NeRF 360 scenes on the same L40S at a one million splat cap and 30,000 steps, Voyager reports mean training times of about 127 seconds for Vauss against about 679 seconds for gsplat-MCMC 1.5.3, with 0.43 to 0.56 times the GPU memory and better LPIPS and SSIM on all seven scenes. Voyager notes this is not a universal ranking, and performance has only been measured on NVIDIA hardware so far. Vauss ships under the source-available Vauss Preview License: use is free during the preview, including commercial work, while redistributing modified versions or offering it as a hosted service needs written permission. The Brush-derived training crates remain Apache 2.0.