In this article
The idea, simply put
Think of a pointillist painting: thousands of dots of color that mean nothing up close and form a picture from a distance. Gaussian splatting does something similar in three dimensions: it represents a room with millions of blobs of color suspended in space.
Each blob is a Gaussian: a small cloud, soft at the edges, with a position, a size, a shape (round or stretched), an orientation, a color and a degree of transparency. A single one looks like nothing. Millions together reproduce the wall, the floor, the light coming through the window and the sheen on the countertop, and you can look at them from any side.
The name comes from a computer graphics approach where each element is projected onto the screen as a blob, or splat.
Where it came from
The technique was introduced in 2023 in 3D Gaussian Splatting for Real-Time Radiance Field Rendering, a paper by researchers at Inria, Université Côte d'Azur and the Max Planck Institute for Informatics, published at SIGGRAPH (the project page). There were already methods for reconstructing a scene from photos, but displaying the result in high quality and smoothly was very expensive. This work made it possible to show them in real time, in high quality.
From a video to a tour
1. Locating the camera in every frame
Frames are pulled from the video, and software looks for recognizable points in them: a corner, an outlet, the grain of the wood. By following those points from frame to frame, it works out where the phone was and where it was pointing at every moment. Along the way it gets a first point cloud of the room.
That is why recording slowly in a still room matters so much: if the phone moves fast or something moves in the scene, locating the camera gets harder.
2. Training the Gaussians
With the cameras located, the software scatters Gaussians through the room and starts adjusting them. It draws the scene from the viewpoint of each frame, compares it with the real frame, and corrects position, shape, color and transparency so they match better. It repeats this thousands of times, adding Gaussians where detail is missing and removing the ones it doesn't need. This step runs on a powerful graphics card and is the one that takes longest.
3. Showing it in the browser
The result is millions of Gaussians. They are compressed into a file, and a viewer in the browser draws them in real time as the visitor moves, on a phone or a computer, with nothing to install.
Why it looks so real
Because it isn't a hand-built model: everything comes from what the camera saw. Textures, shadows, the way light falls and even some of the reflections are stored in the Gaussians. That is why a tour made this way looks more like a photo than an architectural render.
Where it still struggles
- What the camera never saw. There is no way to make it up: it stays empty or blurry. That is why the recording guide asks for two loops.
- Mirrors and glass. A mirror shows a "room" that doesn't exist, and the technique rebuilds it as best it can. Film them at an angle.
- Smooth, featureless surfaces. A bare white wall offers few recognizable points for locating the camera.
- Anything that moves. Someone walking past leaves a ghost.
- Floaters. Sometimes loose Gaussians are left hanging in the air. You delete them afterwards with an editor; in Hausplat, in the studio.
Why it matters for showing properties
Until recently, a 3D walkthrough of a property meant dedicated gear or a lot of modeling work. With Gaussian splatting, a well-recorded video is enough, and the result can be explored freely instead of jumping between fixed points. The full comparison with panoramic photo tours is in 3D tour vs. 360 tour.
The industry is already using it. In July 2025, Zillow, the US real estate marketplace, introduced SkyTour: a 3D view of a home's exterior built with Gaussian splatting from drone footage (Zillow's announcement).
Hausplat applies the same technique to interiors, from a phone video: you film the room, upload it and get a link to the tour. To try it, start with how to make a virtual tour with your phone.
Quick glossary
- Gaussian: each of the blobs of color, with a shape and a transparency, that make up the scene.
- Camera locating: working out where each frame was filmed from, often called structure from motion.
- Training: the process of adjusting the Gaussians until the scene matches the frames of the video.
- Floater: loose Gaussians left in the air that don't belong to any surface.