← All articles

Why shiny car paint is hard for Gaussian splatting (and how to get better scans)

Anyone who has made a few Gaussian splats of cars notices the same thing: a matte grey hatchback on a cloudy day looks great, while a black metallic saloon in the afternoon sun can come out cloudy, noisy or full of strange floating blobs. This is not bad luck. Reflective paint is genuinely one of the hardest things for Gaussian splatting (3DGS). In this article we explain why, what you will see in your results, and what you can do to get better car splats.

How 3DGS "sees" a car

Gaussian splatting learns a scene from images. During training it compares rendered views with the real frames from your video and adjusts millions of small Gaussians until they match. Two assumptions help this process:

1. The same point on the car looks roughly the same from nearby angles. 2. The camera positions are correctly reconstructed from matching details in the frames.

Shiny surfaces break both assumptions.

Problem 1: Reflections move when you move

A glossy car is partly a mirror. What you see in the paint is not just the paint, but the sky, the buildings, the trees and you, walking around with your phone. As you move, those reflections slide across the bodywork.

For the training, this is confusing. A bright stripe on the door appears in one frame, moves in the next and disappears in the third. Gaussian splatting can model some view-dependent colour, which is why splats handle gloss far better than photogrammetry (see 3DGS vs. photogrammetry for cars). But strong, sharp reflections are too much. The optimiser often "solves" them by placing Gaussians *behind* or *in front of* the surface, which shows up as depth artefacts and fog.

Problem 2: Dark and chrome surfaces have little texture

Camera reconstruction depends on recognisable details: an edge, a badge, a speck of dirt. A clean black bonnet or a chrome bumper offers almost nothing to match except reflections, which move. Fewer reliable matches mean less accurate camera poses, and less accurate poses mean a blurrier splat everywhere, not just on the shiny parts.

Problem 3: Glass is transparent and reflective

Windows let you see the interior and reflect the environment at the same time. 3DGS can represent this partially, but windows are often where you see the most noise, especially on tinted glass.

Typical artefacts on shiny cars

  • Cloudy or noisy paint where the surface should be smooth
  • Floaters: small blobs hanging in the air around the car
  • Reflections baked in at a fixed position, so the sky appears "painted" on the door
  • Soft or wavy panels instead of crisp body lines
  • Noisy windows and dark interior areas

None of this means the scan is useless. A splat is a visual copy, and for many purposes slight artefacts are acceptable. But you can reduce them a lot.

Tips for scanning glossy cars

Choose soft light

The single biggest improvement: film under an overcast sky or in even shade. Soft, diffuse light creates soft, broad reflections that change slowly, which the training handles much better than hard sun.

Watch your surroundings

The car reflects whatever is around it. A plain wall or open space gives calm reflections; a busy street full of moving cars and people gives chaos. Avoid bright lamps or windows directly facing the car.

Keep your distance constant

Walking a steady circle at the same distance keeps reflections changing gradually. Sudden moves closer and further make them jump.

Film slowly and sharply

Motion blur and reflections together are the worst combination. Walk slowly, especially in the detail round.

Two calm rounds

Use the CarTwin approach: one overview round with the whole car in view, and one detail round where you slowly move the camera up and down. More viewpoints help the training separate the real surface from the moving reflection. Our full guide is how to film a car for 3DGS.

Consider a little dirt

This sounds odd, but a showroom-polished black car is harder than a car with a light layer of road dust, because dust gives the camera reconstruction something to hold on to. We are not saying you should make your car dirty, just that perfectly clean dark paint is the hardest case.

What a car-specific pipeline can do

General 3DGS tools treat a car like any other scene. CarTwin uses a pipeline built specifically for cars. It selects the sharpest frames from your video, reconstructs the camera path around the vehicle, recognises the ground and straightens the car, and trims away background that does not belong to the model. That does not make reflections disappear, but it gives the training the best possible starting point.

It is important to be honest: CarTwin is still fully in development and in an early pilot phase. On very glossy or dark cars, the result will not always be a perfect one-to-one copy. If a scan fails because of an error in our software, you get the credit back. A disappointing result caused by difficult conditions, such as harsh sun on black paint, is not refunded. That is exactly why these tips matter.

Can you clean up artefacts afterwards?

Yes, to a degree. You can download your model as a .ply and remove floaters or stray Gaussians in an editor such as SuperSplat. Read opening your .ply in SuperSplat for a step-by-step workflow.

Summary

  • Reflections move, so they confuse 3DGS training
  • Dark and chrome surfaces offer little texture for camera reconstruction
  • Soft light, calm surroundings and slow, steady movement help the most
  • Small artefacts can be removed afterwards in SuperSplat

Try it with your own car

The fastest way to learn is by doing. Pick a grey day, follow the recording guide, and start a scan. Bundles start at one credit, see pricing, and you can compare your result with our examples.