Turn a Phone Video of a Room Into a 3D Blender Set: Vauss, the Free Gaussian Splat Tool Where Filming Is the Real Skill
The most important command in Vauss does not build anything. It watches your video for a few seconds and tells you whether the room you filmed is worth building at all.
That is an odd thing to lead with for a tool whose headline is speed. Vauss, released on GitHub on November 12, 2026 by a company called Nullframe, takes a one-minute phone video of a room and turns it into a 3D scene you can walk through in a browser, drop into Blender as a set, or play as a little first-person voxel game. The authors say training takes three to six minutes per scene on one graphics card. But read the documentation end to end and a different story comes out. The software part of making a 3D set from footage is close to solved. What decides whether you get a usable set is how you held the phone.
That changes who this is for. If you shoot video, you already have the skill that matters most here.
What you actually get
Vauss builds what is called a Gaussian splat. The README describes it well: "a scene made from millions of tiny coloured 3D brushstrokes." Instead of modeling walls and furniture as hard surfaces, the tool learns where to put millions of small, soft, colored blobs so that, seen from any of your camera positions, they add up to what your phone saw. Splats are good at the things polygon models struggle with: soft light, reflections, clutter, fabric, plants.
One command, vauss run, does the whole job. It checks your video, picks frames, works out where the camera was for each frame, trains the scene, tests it against frames it held back, and writes the exports. A normal run gives you:
- a browser viewer you open by double-clicking a file,
- a compact web file (
scene.spz) for web scenes and Blender, - a Blender pack with the full splat, a textured mesh when one can be made, and colored voxel cubes,
- a first-person game version where you walk with WASD and knock out blocks.
After that, vauss shot renders new camera moves, an orbit or a push-in, from the finished scene. The README's car example rendered and encoded an eight-second orbit in 19 seconds after training. vauss stabilize re-renders shaky footage along a smooth path.
For a filmmaker, that last pair is the part to sit with. You filmed a room once. Now you can find a camera move you never shot.
Why the capture check matters more than the speed
The team publishes a filming guide that reads like a note from a camera assistant. It opens with the physics in one sentence: "A 3D room is built from how near and far things shift against each other as the camera moves." Spin on the spot and nothing shifts, so there is no depth to recover.
The guide's rules are specific:
- Stand one to two meters from a wall and walk a slow sideways arc in front of it, then turn to the next wall.
- Do the loop twice: once at eye height, once at chest height with the phone tilted down about 30 degrees to catch the floor and tabletops.
- Give every wall at least four seconds, including bare ones, and keep a door frame or picture edge in view on plain walls.
- Move about one step a second and turn no faster than ten degrees a second, so a full turn of the room takes 40 seconds or more.
- Use the main 1x lens, not the ultra-wide. Turn on every light, keep direct sun out, and clear out people and pets, because "anything that moves turns into a ghost."
Then comes the part I like most. vauss check room.mp4 --room runs in a few seconds without touching the graphics card, per the README, and grades the clip on length, light, sharpness, movement, detail, speed and coverage, with time codes for the bad stretches. A failing result stops the full run before it wastes your time. The guide separates warnings that its tests linked to worse results from tips that did not measurably hurt. That is honest in a way most tools are not. It tells you which advice it has evidence for.
If you have ever lost an afternoon to a reconstruction that came out as fog, you know why this matters. The expensive mistake in 3D capture happens before the software starts, and nothing downstream can fix it.
The speed claims, in plain terms
The headline says Vauss trains 5.4 times faster than gsplat-MCMC, a widely used open trainer, using about half the graphics memory. Those are the authors' own numbers, measured on October 9, 2026, on seven standard test scenes, three runs each, on NVIDIA L40S server cards with both trainers capped at the same scene size. Their benchmark notes are unusually careful: they say the result holds "within this comparison," not as a universal ranking. The three-to-six-minute training time comes from a different setup on an RTX PRO 6000, another server-class card.
What that means for you: the speed is real enough to change how you work, because a few minutes per attempt means you can reshoot and retry the same afternoon. But no timing is published for a home card. The README suggests a card with 8 GB of memory, 16 GB of system memory and several processor cores as "a good starting point." Performance has been measured on NVIDIA; Apple M4 has only a basic run test, and AMD, Intel and full Windows testing are listed as still open. Expect slower than the headline on your own machine.
The engine runs through Vulkan, Metal or DirectX 12, so the trainer does not require NVIDIA's CUDA. It builds on Brush, Arthur Brussee's open splatting engine, and credits it prominently.
Put this into practice
Here is a first session that should take an evening.
- Install. On Linux or an Apple Silicon Mac, the README gives a one-line install script for the preview build; Windows users download the archive from the releases page. Run
vauss doctorto see what it makes of your graphics card. - Pick a small, well-lit room. A bedroom or office with some furniture beats an empty hallway. Turn every light on.
- Film one take of about a minute, two loops, slow arcs, both hands on the phone, elbows in.
- Run the check:
vauss check room.mp4 --room. Fix anything marked FAIL or WARN and reshoot. It is cheaper than a failed build. - Build a preview:
vauss run room.mp4 --goal "walk-through of my office" --driver fixed --fast --run-dir runs/office. The fixed driver needs no AI account. - Look before you commit. Open
runs/office/exports/viewer/index.htmland fly around. Note where it goes soft or patchy; those spots map to the parts of your video the check flagged or you rushed. - Bring it into Blender. Run
vauss export-web runs/office --out exports/office. The README says Blender 5.3 imports the result natively through File, Import, SPZ; older Blender versions need the KIRI 3DGS Render add-on for the full splat file. Add a camera and compose a shot. For relit scenes, the authors used a custom material, new lights and cleanup of stray blobs, rendered in Cycles. - Find a move you did not shoot. Try
vauss shotfor an orbit or push-in.
If you use Claude Code or Codex, --driver claude or --driver codex lets your agent run the same steps and inspect the results as it goes. It is optional, and it uses your own agent login.
Where it falls short
It is not open source. The code is "source-available" under the Vauss Preview License. Reading the license itself: you may use it for any purpose, including commercial work, modify it for your own use, and share unmodified copies. You may not sell it, distribute a modified version, or run it as a hosted service for other people's jobs without permission. Rendering for your own company and delivering finished media to a client is explicitly allowed. The license says plainly that "later versions may come under different terms, including paid terms," though the version you have keeps its terms. The parts derived from Brush stay Apache 2.0. If you build a business on this, keep a copy of the version you rely on.
You own what you make, with a caveat about what you film. The license says Nullframe claims "no rights in the scenes, code or media you create." It cannot clear rights in what is in front of your lens. A splat of a client's shop or a friend's home is still their space.
Home-card performance is unknown. Every published timing is on server hardware.
It is new. The repository was created two days before this was written and had single-digit stars. Expect rough edges, especially on Windows.
Some features cost money. vauss ad, which plans and renders a five-shot product film, uses paid generation through your own fal.ai account. Voyager Cloud, the company's hosted machines with Vauss and Blender preinstalled, is billed by the second.
A splat is not a model. You get a beautiful scene to look at and film inside, not clean geometry you can easily edit. The mesh export exists, but it needs a successful run, and the outputs guide says naming a mesh in your goal gives cleaner surfaces.
The camera is the hard part again
There is something reassuring in where Vauss puts its effort. The machine learning is fast and mostly invisible. The documentation spends its words on light, pace, lens choice and coverage, the things a good camera operator already thinks about. For once, an AI tool rewards the craft you have instead of asking you to learn a new one.
Film one room this week, slowly, with every light on. Run the check before you run anything else. Whatever it tells you about your footage is the most useful lesson you will get from this tool.
- Sources: Vauss on GitHub · Filming guide · Benchmarks · Outputs guide · Vauss Preview License · Brush