3D-scan any room from a phone video with Gaussian splatting
Film your room for two minutes on your phone, and a few steps later you can fly a camera through it in a web browser, photoreal, reflections and all. This is Gaussian splatting, and the whole pipeline runs on free, open-source tools: COLMAP, Brush or nerfstudio to train it, and SuperSplat to clean it up and publish the link.
- A cleaned, compressed splat in SuperSplat, exported as a single web page you can host anywhere
- A kit: capture checklist, dataset scripts for Windows and Mac/Linux, and a point-cloud preview tool
- A phone that records 4K or 1080p video
- A computer with a GPU for local training (NVIDIA, AMD, Intel Arc or Apple silicon), or a few pounds to rent one: see Rent a cloud GPU
- ffmpeg installed: FFmpeg for creators
- Comfortable in a terminal. New to it? Terminal survival kit
What a Gaussian splat is, and the pipeline
Know the four stages, where each runs, and why capture decides the result.
A splat isn't a meshA normal 3D scan is triangles with a texture painted on. A Gaussian splat is millions of tiny, soft, coloured blobs (3D Gaussians), each with a position, size, rotation, opacity and a colour that changes with viewing angle. Rendered together they look like a photo from any viewpoint, including shiny surfaces, glass and fine detail like plants and hair that meshes turn to mush.
The pipeline has four stages:
- Capture: a slow, overlapping video on your phone (step 2).
- Camera poses: COLMAP finds matching features across frames and works out exactly where the phone was for each one, plus a sparse cloud of 3D points (step 3). This runs on any computer's CPU.
- Training: starting from those points, the trainer adjusts millions of Gaussians until renders from each camera match your frames (steps 4 and 5). This needs a GPU.
- Clean and publish: SuperSplat in your browser removes floaters, crops, compresses and exports a web viewer (step 6).
Two ways to trainBrush is a single download that trains on almost any GPU (NVIDIA, AMD, Intel, Apple) through WebGPU-style tech, with a live preview. It's the easy local route on a gaming PC or an Apple silicon Mac. nerfstudio's splatfacto is the research-standard trainer on NVIDIA CUDA; on a rented GPU it's the fast, reliable route for any laptop.
The kitDownload the kit: capture checklist, dataset scripts for Windows and Mac/Linux, and a preview tool.