Guide · Gaussian splatting

What is gaussian splatting?

Gaussian splatting reconstructs a scene from photos, video or renders and returns it as a cloud of coloured points. It looks close to a photograph, opens in a browser and on a phone, and it is not an editable mesh: you cannot measure it or swap materials.

Mori Maison team · Reading 8 min · Updated: November 2026

What it is and what you see on screen

It is a visual representation with volume, not a polygon model. Every point stores position, size, colour and opacity; the viewer blends them on each frame and gets real light and texture without loading a mesh.

The name comes from the method: the scene is described by thousands or millions of gaussians. Each one carries a position, a scale, an orientation, a colour and a degree of transparency. Projected on screen they blend into the appearance of the object. There is no texture stretched over triangles, only points that together form the image.

Visitors notice it in the light. A room with soft shadows, fabric and plants looks like the photograph, without the plastic feel a mesh sometimes has when materials are approximated. In exchange, you cannot select a part, change a fabric or pull a dimension out of the model.

How it is captured and rebuilt

You shoot photos or video around the object or space, with enough overlap between shots. A process estimates where each camera was and rebuilds the scene as a splat; then it is cropped, cleaned and optimised so it opens fast.

Capture matters more than software. You need coverage from several angles, stable light and no area that never appears in a shot. A small object is handled with a phone and a turntable; an industrial hall needs a walkthrough, a tripod and patience.

The reconstruction returns the look of the scene, with noise on edges and on surfaces that shine. The next job is cleaning: crop, fill holes, set the starting view and reduce the point count until the file has a sensible weight for the web.

Two ways in: a real capture or CG renders

You can start from something that exists or from a 3D model you already own. In the second case the model is rendered from many viewpoints and rebuilt as a splat. It is fidelity compression: the look of the render travels light to the web.

A real capture makes sense when the product is manufactured and the authentic texture is part of the argument: a finish that differs slightly between units, a patina, a fabric with body. Also when the space exists and you want to show how it turned out, not how you imagine it.

CG renders cover the opposite case. There is a well-lit 3D model, but it is too heavy to open on a phone or to move smoothly. Rendering it from dozens of cameras and rebuilding the result keeps the render light for a fraction of the viewing cost.

Where it breaks: the limits worth knowing

It breaks where reconstruction cannot invent: reflections and glass come out with artefacts, thin objects lose definition and scale is not reliable. It cannot change materials or be measured. These are limits of the technique, not of the project.

A mirror or a glass table is reproduced as seen from the camera, not as a material that reacts to light. Move the viewpoint and you get smears and odd edges.

Thin objects are the other problem. Chair legs, lamp rods, cables or blind slats lose thickness and look translucent or incomplete. If the product is defined by those lines, a splat is the wrong format and a mesh model is the better answer.

Then there is weight. An unoptimised capture can run to hundreds of megabytes and take seconds to open. Splatting is not light by nature: it becomes light when someone spends time cutting points and compressing without destroying the look.

When it pays off and when it does not

It pays off when the scene already exists and what you need to show is how it looks: a finished product, a furnished room, a prototype, a plant or an environment for testing perception. It does not pay off when the buyer needs to configure, measure or compare finishes.

The opposite case is common too. A manufacturer showing a collection with twelve finishes does not need to capture each sample: they need a 3D model with swappable materials. And if a distributor will take dimensions for a project, a splat is no help because the scale is not reliable.

  • The piece exists and the real texture sells better than a clean model.
  • You want to show a space as it turned out, without recreating it.
  • The CG render is too heavy and the buyer only needs to look and turn.
  • You are preparing a prototype or a machine for internal training.
  • You need to show a plant, a warehouse or an installation as it is.
  • You want one link that opens on any phone during a trade fair.
  • No variant configuration, measurement or geometry editing is needed.

How it behaves on the web and on a phone

It runs in the browser with WebGL, mobile included, but weight depends on the scene and has to be worked on. Our target is 60 fps on a mid-range phone, and it is not always reached without cutting detail.

Nothing has to be installed: the scene loads inside the page. That simplifies access and moves the problem to the file. The first visit downloads the whole scene, so initial weight decides whether someone sees it or closes the tab.

Optimisation happens in layers: remove points where detail is invisible, compress attributes, split the scene into zones loaded around the camera and cap resolution on small screens. It is fine-tuning work and should be budgeted, because it does not come included with the capture.

Splatting, a 3D model or a 360 video

All three show a product in the browser and they do not compete on the same ground. A splat wins on visual fidelity, a 3D model wins on control and a 360 video wins on cost when you only need a panorama. What the visitor must do decides it.

It can also live alongside the other two: a space in splat and products in mesh for finish changes. The important thing is deciding which part needs interaction and which part only needs to look right.

Three ways to show a scene on the web
CriterionGaussian splatting3D model (glTF)360 video
What it isReconstruction with photographic lookEditable geometry with materialsSpherical recording
Light fidelityVery high: it is captured lightDepends on render and engineHigh, but flat
Swap finishesNoYes, per materialNo
Reliable dimensionsNoYes, if built to scaleNo
WeightNeeds optimisationTarget under 5 MB per productHigh per video
When it fitsShowing how something real looksConfiguring, measuring, comparingA quick panorama

What a project looks like, in four steps

The work runs in four stages: capture or render, reconstruction, optimisation and publishing. The schedule is fixed after seeing the source material, because the quality of the shots decides half the result.

  • Capture: photos or video of the object, or rendering the 3D model from the required viewpoints.
  • Reconstruction: aligning shots, generating the splat and reviewing edges, reflections and holes.
  • Optimisation: reducing points, compressing and testing on a real phone, not just a laptop.
  • Publishing: a viewer on your site, in your colours, with the scene files handed over.

What to ask before you commission one

Ask for a test with your own object and look at it on a mid-range phone, not on the supplier’s laptop. Ask how much it weighs, what is out of scope, who owns the files and what happens when you want to change the scene later.

  • A demo with your product or space, not a showroom piece.
  • The real file weight and load time on 4G.
  • The accepted limits in writing: reflections, glass, thin objects.
  • What optimisation includes and what is charged separately.
  • Who owns the viewer code and the scene files.
  • How the scene is updated when the product or space changes.

Frequently asked questions

Is gaussian splatting the same as photogrammetry?

No. Photogrammetry returns a textured mesh you can measure and edit. Splatting returns a photographic-looking representation with no mesh: it handles light and materials better, but cannot be used for dimensions or finish changes.

Can it be used to sell product?

Yes, when the product exists and its real appearance is part of the argument: finishes, fabrics, wear. It cannot swap variants or materials. That needs a 3D configurator, and the two often live on the same site.

Do I need a special scanner?

No. Photos or video from a camera or phone are enough, as long as coverage and light are good. What changes the result is not the device but the number of shots, the overlap and steady lighting.

Does it work on mobile?

Yes, with WebGL. The target is 60 fps on a mid-range phone, but it depends on scene weight. If that is not reached, detail is reduced or render resolution is capped on small screens.

Can I measure a splatting scene?

Not reliably. The reconstruction guarantees neither scale nor precision, and a plan taken from it can be wrong. If you need dimensions, the starting point has to be a model built to scale.

Can it show a product that does not exist yet?

Yes, if there is already a 3D model. It is rendered from many viewpoints and rebuilt as a splat. What you cannot do is capture a prototype that has not been made.

What does it cost and how long does it take?

It depends on the source material, the size of the scene and the detail you ask for. There is no single rate or standard schedule: scope and timing are fixed in each proposal, after reviewing the shots.

Sources and references

Related service

Photoreal scenes that open in any browser

We prepare the capture or the renders, rebuild the scene, optimise the weight and publish it on your site with our own viewer. No monthly fee, and the code and files stay with you.