What a Gaussian splat actually is
3D Gaussian splatting is a way of reconstructing a real place from ordinary photographs. Rather than building a mesh and painting textures onto it, the scene is stored as a very large number of soft, coloured, oriented blobs — each with a position, a size, a rotation, an opacity and a colour that shifts depending on where you look from.
That last property is why splats feel so different to older 3D scans. Gloss on a floor, light through leaves and the sheen on a car bonnet all change as you move, because the representation stores view-dependent colour rather than one flat texture. And because the blobs are drawn by rasterising, the whole scene runs smoothly in a browser on ordinary hardware.
The pipeline is: take photographs or video, solve where each camera was, then optimise the cloud of blobs until it reproduces those photographs. Everything that goes wrong later can be traced back to one of those three steps.
Decide what the scene is for before you press record
Capture quality is a means, not the goal. A splat that looks beautiful and answers no question is an expensive screensaver. Before filming, decide what a visitor should be able to do: understand a layout, judge the atmosphere of a venue, see how a site changed since last month, inspect a piece of plant, or walk a route they cannot physically get to.
That decision changes the capture. A marketing walkthrough wants flattering light and a clear route. A construction progress record wants consistent angles you can repeat next month. An inspection scene wants dense coverage of one asset and does not care about the car park.
The five things that ruin a capture
Almost every disappointing scan fails for one of these reasons, and all five are decided on site rather than in software.
- Changing light. Sun going in and out mid-capture is the most common killer. The reconstruction cannot reconcile the same surface being two brightnesses, and you get haze and ghosting. Overcast is your friend.
- Automatic exposure. If the camera re-exposes as you turn, every frame disagrees about what colour the wall is. Lock exposure, white balance and focus before you start.
- Moving too fast. Motion blur destroys the detail the solver needs, and phones with rolling shutter smear straight edges. Walk slower than feels necessary.
- Things that move. People walking through, traffic, trees in wind, and water all appear in some frames and not others. They come out as floating fragments.
- Mirrors and glass. A reflection is not a surface, and the reconstruction will try to place geometry behind the mirror. Expect artefacts and plan the route so mirrors are not the subject.
How to move through a space
The reconstruction needs to see every surface from several directions. That means orbiting subjects rather than pointing at them, and covering the same area at more than one height — roughly chest height for the natural view, plus a lower and a higher pass to give the solver parallax.
Keep generous overlap between consecutive frames, keep the subject in shot as you move around it rather than panning from a fixed spot, and close the loop by returning to where you started. Panning from one position gives the solver rotation with no parallax, which is the single most common reason a capture fails to reconstruct at all.
Corners, undersides and anything you only saw once will be the weak points. If a surface matters, walk round it twice.
Splats, meshes and choosing between them
The honest comparison is about what happens after capture. Photogrammetry gives you geometry: something you can measure, cut sections through, take into CAD, collide against in a game engine, or hand to a surveyor. Gaussian splats generally look better, capture difficult materials that defeat photogrammetry, and process faster — but the output is not a mesh, so measurement and physics are not free.
In practice a lot of projects want both: a splat for people to look at and understand, and a mesh or point cloud underneath for the measurements. Deciding that up front is much cheaper than discovering it after the site visit.
Getting it in front of people is half the job
A trained scene straight out of processing is not a deliverable. It is typically far too large to send to a phone, it contains floating debris around the edges of the capture, and it has no navigation, labels or reason to explore.
The work that turns it into a product is cleaning out the floaters, cropping to what matters, compressing for the web, and wrapping it in a viewer with a sensible starting camera, a route people can follow, and hotspots that explain what they are looking at. That final stage is what separates a scan somebody watched once from a tool a sales team opens every week.
You can try both halves free with our own tools: the Gaussian splat editor for cleaning and framing a scene, and the studio viewer for publishing it. There are finished examples in the Gaussian worlds gallery.
Common questions
What is a Gaussian splat?
A way of representing a real place captured from photographs. Instead of building a mesh with textures, the scene is stored as a large cloud of soft, coloured, oriented blobs, each with its own size, angle, transparency and view-dependent colour. Because they are drawn by rasterising rather than ray tracing, a scene renders smoothly in an ordinary browser.
Do I need a special camera?
No. A recent phone shot carefully will beat an expensive camera swung around quickly. What matters is steady movement, locked exposure, even lighting and full coverage — not sensor size.
How is this different from photogrammetry?
Photogrammetry produces a mesh you can measure, edit and drop into CAD or a game engine. Gaussian splatting produces something that usually looks considerably more convincing, especially for foliage, fine detail and reflective surfaces, but it is not a mesh. Choose by what you need to do with the result, not by which looks better in a video.
How long does a capture take?
A single room is a few minutes of filming. A house or a small site is under an hour on location. Processing is the longer part, and preparing the result for the web — cleaning, compressing, building the viewer — usually takes longer than either.
Can visitors view it on a phone?
Yes, provided the scene has been optimised for delivery. An unprocessed training output can be hundreds of megabytes, which no phone visitor will wait for. Compression and level-of-detail work are what make the difference between a demo and a product.
Related guides
- Gaussian splat capture kit and workflow — gear, settings and the repeat-capture pipeline
- Gaussian worlds gallery — finished scenes you can walk through
- Best Gaussian splat virtual tour workflows — turning captures into tours that sell
- Gaussian splat applications by industry — where this is being used commercially