Guide · 3D capture

Gaussian splatting or photogrammetry?

Photogrammetry rebuilds a textured mesh: it can be measured, edited and relit, within error. Gaussian splatting rebuilds how the scene looks without a mesh: it is more faithful in light and materials, but cannot be measured or re-materialled. What you need to do with the model decides.

Mori Maison team · Reading 7 min · Updated: November 2026

The difference in one sentence

Photogrammetry returns geometry; splatting returns appearance. A mesh can be measured, cut, animated and dressed with new materials. A cloud of gaussians reproduces how the scene looks from the camera and does not support those operations.

Both start from photos and both end up in a browser, and that is where the similarity stops. A table rebuilt by photogrammetry is a surface with a texture on top: move it to another engine and the light changes. The same table as a splat is a set of points that already carry captured colour and shine: it looks right, but you cannot ask it for the dimensions of the top.

That difference decides the project before any comparison of quality. If the destination is a configurator, a BIM sheet, a drawing or an animation, you need a mesh. If the destination is showing how a product or a space looks, a splat usually wins on realism for the same effort.

How each one is captured

Both need overlapping photos around the object, but photogrammetry demands more control: matte surfaces, diffuse light and full coverage. Splatting handles reflections no better in principle, yet it also rewards steady light and plenty of shots.

In photogrammetry the software finds common points across images and solves camera positions. It then builds a mesh and projects the photos onto it. If an area is never sharp or the material is transparent, the algorithm finds no matches and the hole stays.

In splatting the cameras are also estimated, but instead of closing a surface it fits a set of gaussians until the scene, seen from the same viewpoints, matches the photos. It does not have to decide where the surface exactly is, which lets it reproduce fabric, leaves and soft shadows that a mesh struggles with.

A point-by-point comparison

Splatting wins on light realism; the mesh wins on control and measurement. Weight depends on the scene and on optimisation work, not on the technique. Both can run on mobile, but a splat needs more care with point count.

The last row matters more than it seems. A glTF mesh opens in any engine, store or viewer. A splat needs a viewer that understands the format; it is common on the web today, but not universal. If the model has to travel across tools, the mesh is the shared language.

Photogrammetry and gaussian splatting on the same object
CriterionPhotogrammetryGaussian splatting
What you getTextured meshCloud of gaussians
MeasureApproximate scale, with errorNot reliable
Edit geometryYes: cut, move, simplifyNo: it is rebuilt
Change materialsYes, with new lightingNo: colour is captured
Reflections and glassShine breaks upVisible, with artefacts
Thin objectsDeform or disappearLose definition
Light and shadowsDepend on the engine that opens itAre the captured ones
Weight on the webHigh in large scenesHigh without optimisation
Engine requiredAnything that reads glTFA splat-compatible viewer

Realism: where the difference shows

Splatting reproduces the lighting as captured, so a room with mixed light or a fabric with body looks convincing without building a render setup. Photogrammetry needs new materials and light, and some of the likeness is lost there.

In an interior the difference is clear. A living room with window light, soft shadows and plants has a complexity that a textured mesh tends to simplify. A splat keeps those shadows because it does not calculate them: it copies them.

Product behaves in a similar way with fabric and micro-texture. Bouclé or irregular wood grain look better in a splat than in a mesh with a repeated texture. On the other hand, a hard-edged object can look cleaner as a well-built 3D model.

Geometry, scale and measurement

A photogrammetric mesh can be measured with an error that depends on capture and calibration; it is good for approximate dimensions, not for manufacturing. A splat has no surface to measure: any dimension taken from it is an interpretation.

Neither technique replaces CAD when the number has manufacturing consequences. Photogrammetry can give an idea of volume and check whether a piece fits a gap, but two centimetres of error in a fitted cabinet is a problem.

Splatting is even less suited to measuring because its representation defines densities, not edges. Two viewpoints can give different readings of the same edge. Scaling the scene against a known reference makes the viewer behave, but does not turn it into a drawing.

Weight, performance and publishing

Both end up as a file the visitor downloads, and in both weight is the enemy. A mesh is optimised by cutting polygons and compressing textures; a splat by cutting points. Our target is the same: 60 fps on a mid-range phone.

A badly optimised mesh can carry twenty million triangles and 8K textures. An unworked splat can run to hundreds of megabytes. In both cases nobody waits eight seconds for a product page to open.

To publish, a mesh is served as glTF and opens in any WebGL viewer. A splat needs its own viewer and a specific format. Either way, a fallback image for the first load and a clear message when the device cannot move it are worth having.

When to choose each one

Choose photogrammetry if you need to measure, edit, integrate the model into other tools or change materials. Choose splatting if what matters is how the scene looks and the visitor only has to watch, turn and walk through. If you need both, they can be combined.

  • Photogrammetry: BIM sheet, approximate drawing, AR with occlusion or later editing.
  • Photogrammetry: the model has to open in different engines and tools.
  • Splatting: a finished product with real texture and no variants to configure.
  • Splatting: a furnished interior, stand, plant or prototype shown as it is.
  • Splatting: a CG render that is too heavy and must be lightened without losing its look.
  • Splatting: a realistic environment for testing perception or for training.
  • Both: the space in splat and the product in mesh inside the same scene.

They can be used together

Yes, and it is a common mix: the space as a splat, because light and atmosphere rule, and the products as meshes, because finishes have to change and spec sheets have to be shown. Each format does its job and the seam disappears if the framing is cared for.

The fine work is in the integration: matching scale, lighting the mesh to resemble the splat and keeping the camera from giving the switch away. Performance matters too, because two kinds of asset are now downloaded. Loading the scene first and the products as the visitor approaches solves it.

The alternative is to choose one and accept its limits: a single-collection showroom fits entirely in mesh, and a museum piece fits entirely in splat. The mix makes sense when the project asks for both.

How to decide in one meeting

Ask three questions: what the visitor has to do, what has to come out of the model and how much it can weigh. If the answer includes measuring, editing or swapping materials, you need a mesh. If it only includes looking, a splat gives more realism for the same effort.

  • Does the visitor only look, or also configure and compare?
  • Will anyone take dimensions or drawings from the model?
  • Does the model have to enter another tool or engine?
  • Are there thin objects, glass or mirrors that may fail?
  • What weight can the page carry and on which device will it be tested?

Frequently asked questions

Which is more realistic, gaussian splatting or photogrammetry?

Splatting is usually more faithful in light and materials because it copies the captured lighting. Photogrammetry depends on the materials and light applied afterwards. On matte, well-shot surfaces the gap narrows.

Which weighs less on the web?

It depends on the scene and on optimisation, not on the technique. A mesh with many polygons and large textures is heavy; so is a splat that has not had its points reduced. Either way, weight is worked on before publishing.

Can I measure with photogrammetry?

You can get approximate dimensions if the capture is well calibrated and there are scale references. It is not a manufacturing method and does not replace a drawing. For exact dimensions you start from CAD or model to scale.

And with splatting?

No. The scene has no defined surface to measure, so dimensions are interpretations. It can be scaled so the viewer behaves, but that does not make it a metric model.

Which works better on mobile?

Both can work. A mesh depends on polygon count and texture size; a splat on point count and compression. Test on a mid-range phone and offer a fallback image.

Can I change the colour of a captured product?

Not in splatting: the colour is captured and there is no material to edit. In photogrammetry you can, if the mesh and UVs allow it, though the result may need retouching. If the goal is configuring finishes, the route is a 3D model built for that.

Which costs more?

There is no single answer. Photogrammetry adds retopology and material work; splatting adds reconstruction and optimisation. Cost depends mostly on capture coverage and the detail you ask for, not on the technique.

Sources and references

Related service

We tell you which technique your case can take

We review the photos or the model you have and propose splatting, mesh or a mix of both. Not the fashionable technique: the one that solves what the visitor has to do.