- Introduced at SIGGRAPH 2023 by Kerbl, Kopanas, Leimkühler, and Drettakis.
- What it is: a 3D scene stored as millions of tiny colored 3D blobs (Gaussians) instead of a polygon mesh.
- Why it matters: trains in minutes-to-hours, renders in real time in the browser, looks photoreal.
- Where you see it: modern walkable 3D tours, AR/VR scenes, film previs, short-video reconstructions.
Table of contents
The plain-English definition
Gaussian splatting is a 3D scene representation that stores the scene as a cloud of millions of small colored 3D ellipsoids — "Gaussians" — instead of as a polygon mesh.
Each Gaussian carries a position, a color, an opacity, and a 3D shape. To render the scene, software projects every Gaussian onto the screen and blends them together. The result is a photoreal image that updates as the viewer moves the camera, all running inside a browser tab.
It is a fundamentally different way of describing a 3D space. Polygon meshes describe surfaces. Splatting describes the visual field directly. That shift is what makes the technique fast and visually accurate at the same time.
From photos to a splat, in three steps
1. Camera poses
The pipeline figures out where the camera was for each input photo. Classical pipelines use COLMAP for structure-from-motion. Newer single-image methods estimate pose directly from the input.
2. Initialize the Gaussian cloud
Start with a rough point cloud — often from COLMAP — and place a Gaussian at each point. Each one gets an initial color, opacity, and scale.
3. Optimize against the inputs
The trainer renders the current Gaussian cloud from each input camera angle, compares it to the real photo, and nudges the Gaussians to reduce the error. It also adds new Gaussians where detail is missing and prunes ones that contribute nothing. After tens of thousands of iterations, the cloud reconstructs the scene from any viewpoint.
Why this changed walkable 3D tours
Before 2023, photoreal 3D scenes on the web meant either heavy polygon meshes (slow to load, visually flat) or NeRF (Neural Radiance Fields — beautiful but compute-heavy and slow to render in a browser). Gaussian splatting was the first technique to deliver photoreal quality and real-time browser rendering at the same time.
For a buyer, that translates to a tour that loads in a few seconds, runs smoothly on a phone, and looks like a photograph rather than a video-game render. For a property — a residential listing, a medspa room, a restaurant interior — that visual fidelity is the difference between a tour someone uses and a tour they bounce from.
"Polygon meshes describe surfaces. Splatting describes the visual field directly."
Gaussian splatting vs the closest neighbor
The closest neighbor is NeRF (Neural Radiance Fields). NeRF stores the scene inside the weights of a neural network and queries that network for every pixel at render time. The output is beautiful but slow — minutes of training per scene used to mean tens of seconds per rendered frame.
Gaussian splatting keeps the scene as explicit, sortable data — the Gaussians themselves — so a GPU can rasterize them in milliseconds. Same visual ceiling, fraction of the runtime cost. That is why almost every commercial photoreal 3D tour shipping today uses splatting.
For the older mesh-based path, see our photogrammetry explainer. For the broader category, start at what is a 3D virtual tour.
How to get a Gaussian splat of your space
Two practical paths in 2026.
Do it yourself. Capture a video of your space, run it through an open-source pipeline (gsplat, Nerfstudio, or the original INRIA code), then host the result with a viewer like SuperSplat. Total cost: a GPU rental on RunPod or Lambda for a few dollars per scene, plus an afternoon of operator time per capture.
Use a hosted short-video pipeline. TourReady's $99 walkable 3D tour uses a hosted Gaussian-splatting pipeline that takes a short video of a room and produces a walkable scene, delivered in about two minutes. Hosting is free for the life of the tour.