3D Gaussian splatting vs NeRF

Both techniques answer the same question — given photos of a scene, render it from viewpoints nobody photographed — and they answer it in opposite ways. NeRF learns a function. Gaussian splatting builds a pile of objects. That single difference explains almost every practical trade-off between them, including which one you will actually be handed when you export from a world model in 2026.

The two representations

NeRF stores the scene inside a small neural network that maps a 3D position and a viewing direction to colour and density (Mildenhall et al., ECCV 2020). To render a pixel you march a ray through the volume, sample the network many times along it, and integrate. The scene is implicit: there is no list of anything, only weights.

3D Gaussian splatting stores the scene as millions of explicit ellipsoids, each with position, scale, rotation, opacity and view-dependent colour (Kerbl et al., SIGGRAPH 2023). To render, you project them to the screen, sort by depth and alpha-blend. The scene is a data structure you can inspect, and rasterisation is something GPUs already do extremely well.

Where each one wins

	3D Gaussian splatting	NeRF
Render speed	Real time, rasterised	Ray marching, slow without acceleration
Training time	Tens of minutes on one GPU (35–51 min in the 3DGS paper)	Hours originally; minutes with Instant-NGP
File size	Large: 248 bytes per splat as .ply	Small: a network, often tens of MB
Editing	Direct — select, delete, transform splats	Indirect — weights encode the whole scene
Hard optics	Struggles with mirrors, refraction, thin glass	Some variants handle these better
Tooling in 2026	Browsers, engines, exporters, converters	Mostly research code

The row that decides most projects is the first. A splat scene runs in a browser tab at interactive frame rates on hardware people already own. That is what moved capture-to-view work onto splats, and it is why the ecosystem — viewers, formats, engine plugins — grew there.

The row that people underestimate is editing. Because splats are explicit, you can drag-select the ceiling and delete it, transform a chair, or prune floaters to shrink a file. With a NeRF, those operations are research problems. Anyone building a production pipeline feels this difference within a week.

Where NeRF still earns its place

The honest version of “NeRF is dead” is “NeRF is no longer the default.” It remains interesting when:

• Storage or bandwidth dominates. A network is dramatically smaller than a splat cloud of the same capture. If you are shipping thousands of scenes, that ratio is a budget line.
• The optics are hard. Reflection, refraction and thin transparent structures were an early weakness of splatting because a Gaussian is a poor model for a mirror. Several NeRF variants handle these more gracefully, and hybrids exist specifically to combine the two.
• You want continuity. A continuous field is naturally smooth. Splats are discrete and can show popping or aliasing under bad sorting; fixes exist, but they are fixes.

Accelerated variants also changed the picture on training. Instant-NGP brought NeRF training down from hours to seconds-to-minutes with a multiresolution hash encoding, which is why “NeRF is slow to train” is now a statement about the 2020 paper, not about the family. Rendering is where splatting keeps a structural advantage, because rasterising ellipsoids currently has a structural advantage over marching rays through a network.

What this means for world models

Generative world models sit downstream of this choice, and the ones that export 3D have picked splats. World Labs exports Marble worlds as Gaussian splats and publishes Spark, an open-source splat renderer for three.js, which tells you which representation their pipeline is built around. Atlas is expected to produce assets from the same family; see what Atlas is for what the model itself does.

For anyone consuming these outputs, the practical consequences are:

• Your files will be .ply or .spz, not network weights. The format guide covers the difference.
• You view them in a browser or an engine, not in a mesh tool. See how to open a splat file.
• You can edit and prune them, which means asset cleanup is a normal production step rather than a re-training run.
• Mirrors and glass in generated worlds will look wrong more often than matte surfaces do, and that is a property of the representation rather than a bug in the model.

Choosing, in one paragraph

If you are capturing real spaces to look at, present or ship into an engine, use Gaussian splatting; the tooling, the speed and the editability are all on that side, and every world model that exports 3D assets today outputs it. If you are doing research on view synthesis, working under hard storage limits, or fighting reflective and transparent materials, NeRF variants and hybrids are still worth reading. For everyone else, the useful question is not which technique is better but which file you were handed — and in 2026 that file is almost always a splat.

Bottom line

Splatting won the default because explicit primitives rasterise fast and edit easily; NeRF’s compactness and its handling of difficult optics keep it alive in the places where those properties matter more than frame rate. Drop whatever you were given into the browser viewer and you will know within a second which of the two you are dealing with.