Inside EvoLight: inferring normals from a single frame

The problem with relighting

Traditional relighting requires geometry. You need a normal map — a per-pixel description of which direction each surface faces — and an albedo map, and ideally a roughness map. With those in hand, you can place virtual lights and compute how they would interact with the scene.

The problem is that most plates do not come with geometry. A raw photographed frame has none of that. You have RGB and nothing else.

EvoLight's contribution is simple to state and hard to execute: infer the geometry from the frame itself.

How the inference works

EvoLight uses a transformer-based architecture trained on paired (plate, geometry) datasets accumulated over years of production work. Given a single frame, it produces three outputs:

These three outputs are enough to place real light sources in the scene and have them interact honestly with the inferred geometry.

Why it is not a filter

A filter re-colours pixels. It looks at the image and shifts hue, brightness, or contrast. EvoLight does not touch the pixels directly. It builds a model of the scene — its geometry and materials — and then relights that model from scratch.

The difference matters when you move a light. A filter has no concept of a light; it can only approximate a look. EvoLight lets you place a key light at 45 degrees and watch the shadow fall in the right direction, because the normal map told it which way the surface faces.

The honest relight

Because the inferred geometry is approximate, the relight is approximate too. But it is approximate in the right direction — the shadows fall where physics says they should, the highlights appear on the right surfaces, and the ambient term fills the shadows gently rather than flattening them.

That honesty is what makes it survive a 16K finishing pass. The relight is not a trick; it is a plausible physical event applied to a plausible scene.