Gaussian Splatting in Games: 4 Smart Questions Every Artist Should Ask

Gaussian splatting has spent two years as the thing that looks astonishing in a demo video and then quietly disappears. It is now creeping into real pipelines, and it is worth understanding before someone hands you a scan and asks you to build a level out of it.

Two small releases this month tell the story better than any keynote. Jawset shipped a year-one update to Postshot, its standalone scanning app, adding SPZ 4 support, GPS metadata and point cloud import. Chaos added splat relighting to V-Ray. Neither is exciting. Both are plumbing, and plumbing is how a technique becomes a tool.

1. What is Gaussian splatting actually good at?

Photoreal reconstruction from very little effort. A single video walkthrough or a set of stills becomes a scene you can fly a camera through, with lighting and material response that would take an environment artist days to hand-author.

It is strongest on the things that are miserable to model: dense vegetation, weathered surfaces, cluttered real-world interiors, anything organic and irregular. If your project needs a believable forest floor and you have a camera, that maths is suddenly very different.

2. Why does Gaussian splatting fight your engine?

Because engines are built on triangles and it is not. Unreal, Unity and Godot assume vertices, faces, UVs and normals. A splat is a cloud of oriented blobs with no topology underneath it at all.

That mismatch has consequences you feel immediately. No bone skinning, so nothing you capture can be animated conventionally. No collision by default. No normal maps or UV-driven material system, because there are no UVs. You get a beautiful thing your engine does not know how to treat as geometry.

3. Can you light it and edit it yet?

This is where the last few months matter. Captured splats have historically arrived with the original lighting baked in, which makes them useless in any scene with a day-night cycle or a dynamic key light.

V-Ray shipping splat relighting is a real signal that this is being solved rather than worked around. Tooling for editing is following the same curve — slowly, and mostly outside the engines rather than inside them.

4. What does Gaussian splatting cost you?

GPU time and disk. Splats render in real time but they are hungry, and raw captures are large. Cloud processing can compress a model by roughly 50 to 60 percent, which helps, but shipping one inside a build still needs deliberate optimisation rather than hope.

There is also a hidden cost worth naming: nobody on your team knows how to fix a bad splat. A bad mesh has thirty years of accumulated craft behind repairing it. A bad capture usually means going back and shooting again.

How to try Gaussian splatting this week

You do not need a scanning rig. Walk a phone slowly around a subject for ninety seconds, keeping it evenly lit and avoiding reflective or transparent surfaces, then run the footage through one of the consumer capture apps. The first result will be flawed in instructive ways.

Pick something with real depth and clutter rather than a clean product shot. The whole argument for Gaussian splatting rests on it handling mess that would take days to model, so a tidy subject tells you nothing useful about whether it belongs in your pipeline.

Then try to get it into your engine of choice. That second step is where most people discover the difference between an impressive capture and a usable asset, and it is a far cheaper lesson to learn on a weekend test than on a milestone.

Where Gaussian splatting fits right now

Previz, set dressing reference, distant background geometry, and scanned hero props you intend to remodel properly. In other words: everywhere the audience will not interact with it, and everywhere an artist still gets the final pass.

That is a smaller claim than the demos make, and a much larger one than the technique had eighteen months ago. It is on the same trajectory as every other capture technology before it, including photogrammetry, which took roughly a decade to go from novelty to default.

If you are weighing which of these shifts deserve your attention, which AI tools are actually worth learning covers the same question from the software side, and there is a related look at where AI 3D model generators currently stand. Broader industry context sits in why this is a good time to be a game artist.

Reference points if you want to dig further: CG Channel tracks the release cadence, and KIRI Engine has a solid breakdown of the engine integration problem.

If breakdowns like this are useful, subscribe and the next one comes to you.