What is a Gaussian Splat? A plain-English guide to 3D Gaussian Splatting
Gaussian Splatting turns ordinary photos or video into a photoreal 3D scene you can walk around in a browser. Here is how it works, and where it beats meshes and point clouds.
· 3 min read · SpatialSense team
A Gaussian Splat is a 3D scene made of millions of tiny, soft, coloured blobs. Each blob is a 3D Gaussian: a fuzzy ellipsoid with a position, a size and orientation, a colour and an opacity. Viewed together from any angle, they blend into an image that looks like a photograph, except you can move the camera wherever you like.
The technique behind it, 3D Gaussian Splatting, was introduced by researchers at Inria in 2023 ("3D Gaussian Splatting for Real-Time Radiance Field Rendering", SIGGRAPH 2023). It quickly became the most practical way to turn real places into explorable 3D.
How a splat is made
- Capture. Take overlapping photos or a video of a subject or site, walking around it. A phone works. So do a drone or a 360° camera.
- Camera poses. Software works out where each photo was taken from (structure-from-motion, for example COLMAP), producing a sparse point cloud.
- Training. Starting from that sparse cloud, an optimiser adds, moves, stretches and recolours Gaussians until rendering them reproduces the original photos as closely as possible.
- Export. The result is saved as a file of Gaussians, most often
.ply, or a compressed format like.spzor.splat.
Training a typical scene takes minutes to an hour on a modern GPU. Viewing it afterwards is light enough for a browser.
Why it looks so real
Each Gaussian can store view-dependent colour (spherical harmonics), so a surface can look different from different angles, just as real materials do. That is how splats keep reflections, sheen, thin structures like foliage and railings, and soft lighting. Photogrammetry meshes usually bake these away or turn them into noise.
Gaussian Splats vs meshes vs point clouds
| Gaussian Splat | Textured mesh | Point cloud | |
|---|---|---|---|
| Looks like | A photo, from any angle | A 3D model with textures | Dots in space |
| Reflections and fine detail | Kept | Often lost or warped | Not represented |
| Made from | Photos or video | Photos or video | Laser scans or photos |
| Good for | Seeing a place as it was | CAD, printing, simulation | Survey-grade measurement |
| Edit as geometry | Limited | Yes | Yes |
They are complementary. Survey teams still rely on laser-scan point clouds for precise measurement. Gaussian Splats are the best way to see and share a real place, and they are increasingly used alongside point clouds and BIM.
Where splats are used
- Construction and property: visual records of a site on each visit, progress updates for clients, and remote inspections.
- Heritage and museums: sculptures, interiors and buildings that people can explore online.
- Products and e-commerce: objects like cars, furniture and instruments, viewable from every side.
- Film, games and VR: photoreal environments captured instead of modelled.
How to view a Gaussian Splat
You don't need to install anything. Open the SpatialSense Splat Viewer, then drop your file onto the page. It opens .ply, .splat, .spz, .ksplat, .sog and .rad files right in your browser, and your file never leaves your device. Or explore the gallery to try some first.
Not sure which file you have? Read our guide to Gaussian Splat file formats.
From one scene to a shared record
A single splat is a great way to look at a place. On a real project, teams need more: sharing captures with clients and subcontractors, comparing visits over time, checking work against the design, and measuring. That is what SpatialSense is built for. See how construction teams use Gaussian Splats.