Meshes, point clouds, NeRFs and Gaussian splats: how to choose a 3D representation

For decades, a 3D model of a real place meant a mesh: a skin of triangles with images pasted on it. That is still the workhorse of games, engineering and mapping. But in the past few years, new ways of representing a captured scene have appeared that can look far more realistic from a handful of photographs, and have changed what is possible for tourism, heritage and virtual tours.

This article puts the four main options side by side, so that you can choose with your eyes open instead of following the latest demonstration.

The big idea: what are we actually storing?

Every 3D representation answers the same question: how do we describe what a space looks like from any viewpoint? The answers differ in what they store, and that decides what each can and cannot do.

• A mesh stores surfaces, as triangles, with materials and textures.
• A point cloud stores positions of measured points, often with colour.
• A neural radiance field (NeRF) stores a small neural network that, given a position and a viewing direction, predicts colour and density.
• A Gaussian splat scene stores millions of tiny, soft, coloured blobs, each with a position, shape, colour and transparency.

Meshes

A mesh is the standard currency of 3D. Every game engine, headset runtime and CAD or GIS tool understands it.

• Strengths: easy to edit, animate and light; supports collisions, physics and measurement; small and fast once optimised; works everywhere.
• Weaknesses: fine, thin or fuzzy things, such as foliage, hair, wires and glass, are hard to capture; shiny and reflective surfaces look flat or wrong; photogrammetry meshes can look waxy and need clean-up.
• Best for: anything interactive, anything that must be measured or simulated, and content that must run on modest devices.

Meshes keep improving. Modern engines can draw enormous numbers of triangles by loading detail only where the viewer needs it, and AI tools help with clean-up and retopology. The mesh is not obsolete. It is the most dependable format we have.

Point clouds

A point cloud is a measurement first and a picture second. Laser scanners and photogrammetry software produce it, and it is the most direct record of what was there.

• Strengths: accurate, simple, ideal for measurement and for archiving; can be huge without losing precision.
• Weaknesses: no surfaces, so it looks like a swarm of dots up close; heavy to display; hard to light or interact with.
• Best for: surveying, change detection, and as the source for meshes and digital twins.

Neural radiance fields (NeRFs)

Introduced in 2020, NeRFs showed that a neural network could learn a scene from photographs and then render new views with striking realism, including reflections and soft effects that meshes struggle with. Later work made training much faster.

• Strengths: very realistic novel views; handles complex lighting, transparency and fine detail; the whole scene is a compact model.
• Weaknesses: slower to render than meshes and splats; the scene is a neural network, so it is hard to edit; no explicit surface for physics or measurement.
• Best for: research and high-quality offline views. For real-time use, splats have largely taken over.

Gaussian splats

Gaussian splatting appeared in 2023 and spread quickly because it combined NeRF-like realism with real-time rendering. A scene is stored as millions of soft blobs, and the graphics chip draws them directly, fast enough for smooth interaction in a browser or headset.

• Strengths: photorealistic, including shine and reflections; captured from ordinary photos or video; trains in minutes to an hour on a good graphics card; renders in real time.
• Weaknesses: no clean surface, so collisions, measurement and editing are difficult; lighting is baked in from the moment of capture; quality drops when you look from angles far from the photos; files can be large; floating specks called floaters need cleaning.
• Best for: virtual tours, heritage and tourism showcases, and visually rich scenes where realism matters more than interaction.

A side-by-side view

• Looks real from photos: splat and NeRF are strongest; mesh is good with effort; point cloud is weakest.
• Edit and relight: mesh is best; the others are limited.
• Collisions and physics: mesh only, in practice.
• Measurement: point cloud and mesh, with control points.
• Real-time on phones and headsets: mesh is easiest; splats work with optimisation; NeRFs are heavy.
• File size: optimised meshes are smallest; splats are medium to large; point clouds can be huge.
• Maturity of tools and standards: mesh is the most mature; splats are developing fast.

Hybrids are normal

The most practical designs combine representations. A museum can show a splat for the look of a hall, with a simple, invisible mesh underneath for walking and collisions. A survey can keep the point cloud for measurement, a mesh for the engine and a splat for a public tour. Researchers are also developing methods that extract surfaces from splats, which narrows the gap between the two.

How to choose: five questions

1. 01Does anyone need to measure or simulate it? If yes, you need a mesh or point cloud, with control points.
2. 02Does it need to be interactive, with objects that move or can be picked up? Use meshes.
3. 03Is the main goal to show a place as realistically as possible? Consider a splat.
4. 04What device will show it? Weak devices favour meshes, or heavily optimised splats.
5. 05How long must it last? Meshes and point clouds in standard formats are safest for the long term. Keep your source photos, since a better method may come along.

Worked examples

• A heritage record of a temple: point cloud and textured mesh as the archive, with a splat for a public virtual visit.
• A hotel showing rooms to overseas guests: splats for realistic tours, with a mesh floor plan for navigation.
• A training simulator for a plant: clean meshes with physics, because trainees must interact.
• A flood model of a town: point cloud and elevation mesh, because it must be measured.

Frequently asked questions

Have Gaussian splats replaced photogrammetry?

No. They complement it. Photogrammetry still produces measurable, editable meshes. Splats excel at realistic viewing. Many teams capture once and produce both.

Can I use a splat in a game engine?

Yes, through plug-ins and viewers, although support is newer and less uniform than for meshes. You will usually still need a simple mesh for physics.

Are NeRFs obsolete?

They remain important in research and for some high-quality results, but splats have become the common choice for real-time viewing.

Which is cheapest to produce?

A mesh from photogrammetry and a splat from the same photos cost about the same to capture. Cleaning and optimising for your audience is where the effort goes.

There is no single best way to store a place. There is only the right representation for the job. Ask what people will do with the model, and the choice usually makes itself.