What is Gaussian splatting?

A plain-language guide, from photos to a walkable 3D scene · last updated 15 August 2026

Gaussian splatting is a way to turn ordinary photos or video of a place into a photoreal 3D scene you can move through freely. Instead of building the scene from triangles like a video game, it builds it from millions of tiny, soft, colored blobs. Each blob is see-through at the edges and solid in the middle, like a small puff of colored mist. Stack enough of them in the right places, with the right colors and sizes, and together they reproduce a room, a statue or a bicycle so faithfully that screenshots are hard to tell from photos.

How a splat is built

Every Gaussian splat starts as a set of photos of the same subject from many angles. Four things happen on the way to a finished scene:

1. Camera poses. The computer works out where each photo was taken from and in which direction the camera pointed. Classic desktop pipelines spend a long time on this step. On a phone, motion sensors and visual tracking provide a strong starting point that is then refined mathematically, a process called bundle adjustment.

2. A starting guess. Matching points that appear in several photos are triangulated into a sparse 3D point cloud, a rough skeleton of the scene. Each point becomes a first blob.

3. Optimization. This is the heart of the method, and it is honest, brute-force learning. The computer renders its current cloud of blobs from one photo's viewpoint, compares the rendering with the real photo, and nudges every blob to shrink the difference: position, size, shape, orientation, color, transparency. Where detail is missing, blobs split and multiply. Where blobs contribute nothing, they are removed. Repeat thousands of times across all photos, and the blur of blobs converges into a sharp scene.

4. Viewing. The finished splat is just a long list of blob parameters. Drawing it means sorting the blobs back to front and blending them onto the screen, which modern GPUs do fast enough for real-time viewing on a phone or in a browser.

Why splats look better than meshes for real places

Photogrammetry, the older approach, produces a mesh: a shell of triangles wrapped in photo textures. Meshes are great for solid, matte objects, and they remain the right choice when you need geometry for 3D printing or physics. But real places are full of things meshes handle badly: plant leaves, railings, thin cables, glass, mirrors, polished stone, hair. A mesh must decide where the surface is, and for a window or a bush there is no single good answer, so it tears holes or melts detail into blobs of texture.

Splats never have to decide where the surface is. A semi-transparent blob can sit in the volume of a bush or hover as a highlight on a glossy floor. Blobs can also change their color slightly depending on the viewing direction, which is how a splat reproduces reflections and shine that move with you. That is why walking through a good splat feels like being there in a way textured meshes rarely achieve.

A short history

The problem of rebuilding a scene from photos is old, but two breakthroughs made it photoreal. In 2020, NeRFs (neural radiance fields) showed that a neural network could learn a scene so well that it rendered convincing new viewpoints. NeRFs were stunning but slow, both to train and to display. In 2023, researchers at Inria in France published 3D Gaussian splatting, which replaced the neural network with the explicit cloud of blobs described above. Quality stayed comparable, but rendering became hundreds of times faster, fast enough for real time. That single property moved the field from research demos to products, viewers and phone apps within two years.

Gaussian splatting on an iPhone

Until recently, making a splat meant a desktop computer with a gaming GPU: hours of camera-pose solving, then an hour or more of optimization. Phones change the recipe in two ways. First, they know a lot about their own camera. Motion tracking gives good initial poses while you record, so the expensive desktop pose step shrinks to a quick refinement. Second, Apple Silicon chips share one pool of memory between CPU and GPU, which is exactly what the optimization step needs to handle a room-sized scene on a device that fits in a pocket.

Scantic builds on both: it captures video and motion data while you walk around your subject, then runs the full optimization directly on the iPhone, in a couple of minutes, offline, with nothing uploaded. The result is a real Gaussian splat that you can step inside on the phone, share as an end-to-end encrypted link, or export as a standard splat file for desktop tools.

How to capture a good splat

The optimization can only reproduce what your photos show, so capture is where quality is decided:

Move in a slow circle. Walk around your subject and try to see every side. Surfaces that no photo saw cannot be reconstructed and will show as holes or fog.

Change your height. A second pass at a higher or lower angle fills in tops and undersides.

Light is quality. More light means sharper photos and cleaner detail. Dim rooms produce blurrier, noisier splats.

Slow beats smooth. Fast panning blurs frames. Moving your feet instead of swinging the phone gives the method the parallax it needs.

What you can do with a finished splat

View it from anywhere, including positions you never stood in. Share it: Scantic creates links that open in any modern browser, end-to-end encrypted if you choose. Export it: the PLY and SPZ files that Scantic writes open in the standard splat ecosystem, for example the SuperSplat editor for cleanup and publishing, or Blender with a splat add-on for compositing. Game engines are adding native splat support as well.

Frequently asked questions

Do I need LiDAR? No. Splatting works from regular photos. Camera positions come from motion tracking plus mathematical refinement. Scantic runs on iPhones with and without LiDAR.

How long does it take on a phone? With Scantic, a couple of minutes from the end of capture to a finished scene. Objects are faster than rooms, and newer iPhones are faster than older ones.

Splatting vs. photogrammetry? Meshes for solid matte objects and when you need printable geometry; splats for faithful, walkable reproductions of real places with difficult materials.

Splatting vs. NeRF? Same goal, different representation. NeRFs compute every pixel through a neural network and are slow to display. Splats are explicit geometry that GPUs draw in real time.

Which file formats? PLY (the original research format) and SPZ (compressed, much smaller at similar quality). Scantic exports both.

What does it cost? Scantic is free on the App Store for iPhones running iOS 17 or later. An optional Premium tier adds full-quality export and extra share links.