Gaussian splatting explained: photoreal 3D from photos, in real time

If you have seen a recent virtual tour of a building, a statue or a hillside that looks as real as a video, but that you can orbit and walk through at will, you may have been looking at a Gaussian splat. The technique was introduced by a research team in 2023, and within a year it moved from academic papers into everyday tools, viewers in web browsers and product demonstrations.

This article explains how it works, with no mathematics, and what the strengths and limits mean for real projects.

The idea in one picture

Imagine sprinkling millions of tiny, translucent, coloured clouds into a room. Each cloud is a soft blob, stretched in some direction, with a colour that can change slightly depending on where you look from. From a given viewpoint, the blobs in front partly cover those behind, and the sum looks like a photograph. Now imagine you could adjust every blob until the picture it produces matches the real photographs of the room from all the angles they were taken. That is Gaussian splatting.

What is in a splat

• Position: where the centre of the blob is in 3D space.
• Shape: how large it is and which way it is stretched, which lets a thin blob lie along an edge or flat against a wall.
• Colour: stored so that it can vary with the viewing direction, which is how shine and reflections are reproduced.
• Opacity: how see-through it is.

The name comes from the bell-curve, or Gaussian, shape of each blob, which fades smoothly at its edges, and from the way it is drawn by splatting it onto the screen, like a splash of paint.

How a splat is made

1. 01Capture photographs or video of the scene from many viewpoints, with plenty of overlap.
2. 02Recover the camera positions with a structure-from-motion step, the same technique photogrammetry uses. This also gives a sparse cloud of points.
3. 03Initialise one blob at each of those points.
4. 04Optimise. The system renders the scene from each photo’s viewpoint, compares it with the real photo and nudges every blob’s position, shape, colour and opacity to reduce the difference. It repeats this thousands of times.
5. 05Adapt. During training, it splits blobs where detail is needed, clones them where the scene is under-covered, and removes ones that do nothing.
6. 06Export the result as a file that a viewer can open.

On a modern graphics card, training typically takes minutes to an hour for a room-sized scene. That speed, compared with the hours of earlier methods, is a large part of why the technique caught on.

How it is drawn so quickly

Unlike a NeRF, which has to ask a neural network thousands of questions per pixel, a splat scene is drawn by projecting each blob onto the screen, sorting them from near to far and blending them. This is something graphics hardware is built to do fast. As a result, scenes can be viewed in real time, in a web browser, on a laptop or on a headset, though large scenes need optimisation on weak devices.

What splats do especially well

• Realism. Soft effects, fine detail, reflections and view-dependent shine come out naturally.
• Difficult subjects. Foliage, fur, lace, glass edges and cluttered interiors that confuse mesh reconstruction often look good as splats.
• Speed of capture. A walk-around video can be enough.
• Immediacy. The result looks like the place, with no artist needed for texturing and lighting.

Where they struggle

• No surface. There is no solid skin to bump into, measure or cut, so physics and precise measurement need a companion mesh.
• Baked lighting. The light of the day you captured is part of the colours. You cannot easily change the time of day or move a lamp.
• Limited viewpoints. Look from far outside the area covered by your photos and it falls apart into streaks and blobs.
• Floaters. Stray blobs hang in the air where the system misinterpreted something, and need to be cleaned out.
• Size. A high-quality scene can be hundreds of megabytes, so compression and streaming matter.
• Moving things. People, cars and waving trees confuse the process, because it assumes a still scene.

Capture tips that make the biggest difference

• Move slowly and steadily, and keep every frame sharp. Blur cannot be fixed later.
• Cover the scene at several heights and from many angles, with strong overlap between frames.
• Keep exposure and white balance fixed, and capture in even light.
• Avoid people, vehicles and anything that moves, or capture when the place is quiet.
• Add texture where surfaces are bare, or take extra care around plain walls, glass and mirrors, which are hard for any camera-based method.
• Capture more than you think you need, particularly around important details.

Files, formats and tools

Splats are commonly stored as large point-like files, often using the PLY format, with compressed variants such as SPLAT, KSPLAT or SPZ that reduce the size. Viewers and editors exist for browsers and desktops, and plug-ins bring splats into game engines. Standards bodies are working on a common way to carry splats inside glTF, so that they can travel like other 3D assets. This area is still changing, so check the current state of support in the tools that you plan to use.

Splats and the rest of your pipeline

• Use a splat for the beautiful view, and a simple mesh for walking, collisions and measurement.
• Keep the original photos. Methods are improving, and re-processing later may give a better result.
• Clean the splat: crop away the background, remove floaters and set a sensible origin and scale.
• Compress it for the web, and test on target devices, as with any 3D asset.
• Say what it is. A splat is a reconstruction of how a place looked on one day, not a survey.

Frequently asked questions

Is Gaussian splatting the same as photogrammetry?

No. Both start from photos and camera positions, but photogrammetry builds a surface mesh, and splatting fits millions of blobs. They answer different needs.

Do I need an expensive camera?

No. Many good splats come from phone video. A steady hand and good light matter more than the camera.

Can I edit a splat?

You can crop, delete and clean it in editors, and in some tools move or scale parts. Detailed editing, like reshaping a wall, is hard.

Can splats be used for measurement?

Not reliably by themselves. Use a point cloud or mesh with control points when accuracy matters.

Gaussian splatting is the clearest sign of how quickly 3D capture is changing. It does not replace meshes or surveys, but for showing a real place so that people feel they are standing in it, it has become one of the best tools we have.