↓

How AI Copies Lighting From Your Photo

Oct 29, 2026

The light in your photo is the light in your video. Not a copy of it, not an interpretation — the model reads the direction and quality of light off the source image and carries it forward, because that's the cheapest way to keep the frame plausible.

Which means the most important lighting decision you make happens before you upload anything.

What "copying the lighting" actually means

The model doesn't have a lighting rig. It has a portrait where the left side of a face is brighter than the right, and it infers: key light from the left, roughly 45 degrees, moderately soft. Then it generates frames that stay consistent with that inference.

It's inference, not transfer. Nothing is measured in lux. The model is pattern-matching against what lit portraits look like.

That's why the result is usually faithful on direction and vague on quality. A soft window light and a softbox both read as "soft, from the left" and the model treats them as the same thing.

The three properties it reads

Direction. Which side of the face is lit, and how high the light sits. This is the strongest signal and the one it holds most reliably.

Hardness. Are the shadow edges crisp or gradual? Hard light creates a strong, graphic look; soft light is forgiving. The model reproduces the feel of this well.

Colour. Warm tungsten, cool daylight, greenish fluorescent. Colour casts get carried and amplified — a photo shot under a warm lamp comes back warmer than you remember it.

Everything else — falloff, bounce, specular highlights on skin — is approximated, sometimes badly.

Why mixed light breaks first

Two light sources of different colour temperature are the single most common cause of a clip that looks wrong without an obvious reason.

A subject lit by a window on one side and a ceiling bulb on the other gives the model contradictory evidence. It picks one, and then the shadow on the losing side flickers between generations.

Phone flash is the other reliable failure. On-camera flash flattens the face and kills the shadow that tells the model where light comes from. With no directional evidence, it guesses, and it guesses differently in different clips.

Practical rule: one dominant light source, one direction, no flash.

What makes a photo easy to relight

Look for these four things in the source:

  1. A visible shadow on one side of the face. That shadow is the information the model needs.
  2. Even exposure. A blown-out forehead has no texture to animate.
  3. A neutral background. A strongly coloured wall bounces colour onto skin and the model carries it.
  4. Both faces lit the same way. If one person is by a window and the other isn't, the clip has to reconcile two different worlds.

The last one is specific to couple videos and it's where most sources fail. Two portraits taken in different rooms at different times of day are two different lighting setups. The model can merge them, but the result tends to look like a composite — because it is one.

The thing you can't fix afterwards

You can't relight a generated clip. There's no RAW, no shadow detail to lift. If the source had flat, sourceless light, the clip inherits flat, sourceless light, and no amount of grading adds a key light that was never there.

This is the argument for spending your effort on the photograph rather than on the prompt. "Dramatic side lighting" in a prompt is a weak request. A photo already lit from one side is a command.

Getting it right in the source

If you're shooting specifically for this, the setup is unglamorous and takes two minutes:

  • One window, or one lamp, off to one side
  • Face turned slightly toward it, not square on
  • No overhead light, no flash
  • Same spot, same time, for both people

That's it. The clip won't look cinematic because of it, but it also won't fight you, and a stable plain clip beats an unstable dramatic one every time.

Try it at the generator; price at pricing.

Frequently asked questions

Does the AI add lighting that isn't in the photo?

It extends what's there — a room gets plausible shadow, a background gets depth. It doesn't invent a new key light direction.

Can I ask for different lighting in the prompt?

You can, and it will nudge the result, but a text request is much weaker than the visual evidence already in the photo.

Why does one face look brighter than the other?

Because it was. If the two portraits were shot under different light, the model carries both setups and reconciles them imperfectly.

Is phone flash really that bad?

Yes, for this specific purpose. It removes the shadow that encodes light direction, leaving the model to guess.

Does colour grading fix a mismatch?

It can pull two clips closer, but it can't create direction that was never recorded.

Zombie Couple Video