Voyager has released Vauss, a source available gaussian splatting trainer and reconstruction CLI, as a research preview.
Vauss builds on Brush, Arthur Brussee's portable GPU splatting engine, which reached v0.3 in September 2025 with a hybrid MCMC approach. Vauss is native Rust built with Burn, CubeCL and wgpu, and its trainer runs through Vulkan, Metal or DX12 without requiring CUDA.
The vauss run command takes a video through a capture check, frame selection, camera solving, training, held out evaluation and export in one pass, with --fast for a preview and --max for a larger budget. Fisheye and action cam footage can be trained on the fisheye lens itself with --camera-model opencv-fisheye, without an undistorted crop, and existing Nerfstudio transforms.json or COLMAP folders are accepted.
One run writes a standalone browser viewer that opens from disk, a Blender pack with the full splat PLY, surface voxels and a textured mesh where one is available, and a scene.spz with scene.json for the web or for import through File, Import, SPZ in Blender 5.3, which now treats gaussian splats as native objects. A collision mesh, convex pieces and a floor estimate export with a Unity import folder. The same capture also becomes a playable first person voxel game with destructible voxels, where V toggles between splats and voxels. Separate vauss shot and vauss stabilize commands render new orbits and smooth shaky footage.
An installed, logged in Codex or Claude Code agent can drive the pipeline with --driver codex or --driver claude, and vauss serve launches a stdio MCP server for other clients. Vauss also comes preinstalled alongside Blender 5.3 on Voyager Cloud GPU machines that use the creative image.
Voyager reports mean training times of 126.86 seconds for Vauss, 194.96 seconds for VkSplat and 679.37 seconds for gsplat-MCMC 1.5.3 across seven Mip-NeRF 360 scenes, paired on the same L40S machine at a one million splat cap and 30,000 steps. The 5.4x headline is the reciprocal of the mean paired time ratio against gsplat-MCMC, and in that comparison Vauss used 0.43 to 0.56 times its GPU memory, scored lower LPIPS and higher SSIM on all seven scenes, and higher PSNR on six. Voyager states the result is not a universal ranking or an equal quality speedup at arbitrary budgets. Performance has been measured on NVIDIA hardware only, with Apple M4 limited to a smoke test.
Vauss ships under the Vauss Preview License, which is source available but not open source. Use is free during the preview, including commercial work and private modification, while distributing a modified version or offering Vauss as a hosted service requires written permission, and later versions may come under paid terms. The Brush derived training crates remain Apache 2.0. Every Vauss build so far is tagged as a release candidate, the latest being vauss-v0.1.25-rc.
Vauss is available now as a research preview.




