Gaussian splatting
Also known as: 3D Gaussian splatting, 3DGS, Gaussian splats, splatting
Gaussian splatting is a technique for turning a set of photos or a video of a real place or object into a realistic 3D scene you can move around in.
It represents the scene as millions of tiny, soft, semi-transparent blobs, called Gaussians, each with a position, colour, and shape. While it produces striking results that render in real time, the scenes are hard to edit like normal 3D models, and file sizes can be large.
Gaussian splatting, explained at three levels
At home
Imagine rebuilding a room out of millions of tiny, coloured, see-through puffs of mist, each slightly stretched and tilted. Seen up close, it's a cloud of blobs; from a distance, they blend into a convincing picture. Gaussian splatting does this from ordinary photos or video. The computer works out where each blob goes and what colour it should be, so you can then move around the scene freely.
At work
Gaussian splatting is a 3D capture technique that turns photos or video of a real place into a scene made of millions of soft, coloured blobs called Gaussians. It renders in real time with photographic realism, even on phones and in browsers. Businesses use it for real estate tours, tourism, film set capture, and digital twins, because capture needs only a camera and a few minutes of footage.
At home
Each blob in a splat scene, called a Gaussian, stores a position, a stretched oval shape, a rotation, an opacity, and a colour that can shift with viewing angle. To draw the scene, the computer sorts blobs from front to back and 'splats' each one onto the screen, blending overlapping ones. Because blobs can be tiny and semi-transparent, they handle foliage, hair, and reflections that trip up traditional 3D scans.
At work
A splatting pipeline starts with structure-from-motion software such as COLMAP, which estimates camera positions and a sparse point cloud from the photos. Each point becomes a Gaussian. Training then renders the scene from each camera, compares it with the real photo, and adjusts every Gaussian's position, shape, colour, and opacity, adding blobs where detail is missing and pruning useless ones. Rendering uses GPU sorting and alpha blending, so it runs in real time.
At home
Gaussian splatting is trained to match photos, not to model objects. The blobs that make a table look solid don't describe its surface; they just reproduce its appearance from the angles you captured. Lighting is baked into their colours, and views far from the original camera path can show floating smudges. In my opinion, that's the main limitation: you can't easily measure, relight, or animate a splat like a regular 3D model.
At work
Production use depends on capture discipline: consistent lighting, sharp frames, and full coverage, since gaps become artifacts. Raw scenes can reach hundreds of megabytes, so compression, level of detail, and streaming are active areas. Relighting is limited because lighting is baked in, and physics or collision still require a mesh. The Khronos Group is working on a glTF extension for splats, and support in Unreal, Unity, and web viewers continues to mature.
Real-world examples
At home
Marcus captured a video walking around his grandparents' cottage on Lake Muskoka before it was sold. A splatting app turned it into a 3D scene the family can still explore on their phones.
At work
Nadia runs a real estate media company in Calgary. Her team uses Gaussian splatting to create explorable 3D tours of homes from a few minutes of video, faster than traditional 3D scanning.
How it compares
| feature | Gaussian splatting | Neural radiance field (NeRF) | Photogrammetry |
|---|---|---|---|
| Representation | Millions of soft blobs | Neural network | Textured mesh |
| Rendering speed | Real time | Slower | Real time |
| Realism | Very high | Very high | Good, struggles with shiny or thin objects |
| Editability | Limited | Limited | High |
Splatting combines NeRF-like realism with real-time speed, but meshes remain easier to edit.
For analysts
For 3D capture projects, ask whether you need to view the space or work with it. I'd choose splatting for realistic viewing and photogrammetry or laser scanning when you need measurements or editable models.
History & origin
Researchers have long reconstructed 3D scenes from photos using photogrammetry, which produces meshes. Neural radiance fields (NeRFs), introduced in 2020, created far more realistic views but were slow to render. In 2023, Bernhard Kerbl and colleagues at Inria and the Max Planck Institute presented '3D Gaussian Splatting for Real-Time Radiance Field Rendering' at SIGGRAPH, achieving similar quality at real-time speeds. The approach spread quickly through research and apps, with phone apps letting users capture splats and tools adding support in game engines, web viewers, and VR headsets. In my view, Gaussian splatting is the fastest-moving graphics technique I've seen go from paper to phone app.