Skip to content
GPU-assistedUpload required

Move the Light on a Portrait That Is Already Taken

Relighting is the hardest thing on this list of portrait pages, and the reason is that light is not local. Move the key and every shadow moves, every highlight moves, the color temperature across the whole frame changes, and the falloff onto the background changes with it. Get any one of those out of step with the others and a viewer knows something is wrong immediately, even in a photograph they have never seen before.

Use this without the search next time. Prathom Workbench puts Prathom's tools in your toolbar.

Add to Chrome — free

Drop an image here, or click to browse

Up to 25 MB. Uploaded temporarily, deleted after processing.

GPU-assisted — results vary slightly between runs.

What it does

  • Whole-frame shadow direction moved together
  • Specular highlights matched to the new key
  • Background falloff kept consistent
  • Highest denoise of the portrait pages

How to use AI Face Relight

  1. 1

    Check how much of the face was dark

    Whatever sat in shadow was not recorded. The larger that region, the more of the relit face is generated rather than moved.

  2. 2

    Run the relight

    A soft key from the front and slightly above is applied, with shadows on the face, neck, and collar moved to agree with it.

  3. 3

    Audit every shadow in the frame

    Nose, chin, collar, glasses, hair, and anything behind the subject must all point the same way. One survivor from the old lighting ruins it.

How it works

Your portrait is scaled to about one and a fifth megapixels and repainted at the highest denoise of the portrait pages, with a fresh seed each run. The instruction specifies a soft key from the front and slightly above, and then asks for every shadow on the face, the neck, and the collar to agree with that direction, and for the specular highlights and the color of the light to match across the whole frame.

The negative prompt is dominated by consistency failures — contradictory shadows, two light directions — because those are what break a relit image, far more than any local artifact.

Light is a global constraint

Most edits are local. A spot, a stray hair, a logo on a shirt: change it and nothing else in the image needs to know.

Lighting is the opposite. A single key light determines the shadow under the nose, the shadow under the chin, the shape of the shadow on the neck, the bright edge on the cheekbone, the position of the reflection in each eye, the shadow cast by the frame of a pair of glasses, the falloff across the background, and the color cast on everything. All of those are consequences of one fact, and a viewer's visual system knows they should agree even when they have never consciously thought about it.

Which means a relight has to change all of them together and correctly. Any single survivor from the original lighting reads as an error, and it is generally noticed as a feeling that the picture is composited rather than as a specific observation about a shadow.

That is why the denoise here is high. There is no version of this operation that touches a small part of the frame.

What was dark was never recorded

The second difficulty is more fundamental and cannot be engineered around.

A face lit hard from one side has a lit half and a dark half. In the dark half the sensor collected very little, and what it did collect is mostly noise. There is no detail hiding in there waiting to be brought out — the information was not captured.

So relighting to a frontal key means the previously dark cheek, jaw, and eye socket have to be produced from scratch, informed by the lit half and by general knowledge of faces. For a symmetrical face that is a reasonable bet. For a face with a scar, a mole, a distinctive asymmetry, or an unusual jaw line on the dark side, it is a bet the model will lose without any indication that it did.

The practical guide: the more of the face was dark, the more of the result is invention, and it scales smoothly rather than switching over at a threshold.

The audit that takes thirty seconds

Relit portraits fail in a way that is easy to check once you have a list.

Look at the shadow under the nose and confirm it points where the new light implies. Then the chin, then the shadow on the neck, then the shadow cast by any glasses, then the hair against the forehead, then anything in the background — the wall behind the shoulder, a chair, a door frame. Finally look at the reflection in each eye and check it matches the new key rather than the old one.

Six or seven items, and any disagreement between them is the whole failure. If they all agree, the image will hold up even where parts of it were generated.

Publication gate

This page ships once the workflow has been run against a hard side lit portrait, a flat on-camera flash frame, a subject in mixed warm and cool light, and a portrait wearing glasses, with every shadow in the frame audited for direction in each output.

Examples

Harsh side lit portrait

window.jpg - 1600x2000, one side lit, one side dark
prathom-ai-face-relight.png - 980x1225

The interesting case. Opening the dark side means generating a cheek and jaw that the camera recorded almost nothing about.

Flat on-camera flash

flash.jpg - 1800x2400, even, shadowless, harsh
flash-relit.png - 954x1272

Easier. Everything is recorded, so adding gentle modeling is a redistribution rather than a reconstruction of missing material.

Frequently asked questions

Why is relighting harder than retouching?

Because a retouch is local and lighting is not. Removing a spot affects a few dozen pixels and nothing else in the frame needs to know. Changing the key light means every shadow, every highlight, the color of the light, and the way it falls off behind the subject all have to change consistently, and consistency across a whole frame is a much larger constraint.

Can it recover the dark side of a face?

It can produce one. Where a region was in deep shadow, the sensor recorded almost nothing about the skin, so the cheek and jaw that appear in the relit version are generated from the visible half and from what faces look like in general. That is a reconstruction, and asymmetric faces suffer most.

What is the most common failure?

A shadow that did not move. The nose and chin usually update convincingly, and then a shadow under a collar, behind a pair of glasses, or on the wall behind the subject stays pointing the old way. Viewers cannot always name it, but the portrait stops looking like a photograph.

Will it fix a photo taken under mixed lighting?

Partially, and this is one of its better uses. Warm lamp on one side and cool daylight on the other produces skin that is two different colors, which is very hard to correct by ordinary means. Relighting to a single soft key gives the model license to unify it, though check the background did not go strange.

Is the AI Face Relight tool free, and do I need an account?

No sign-up. The ai face relight workflow takes an image up to 25 MB, runs it on the GPU host and hands back a file you can download directly. Nothing is charged and nothing is stamped onto the output. Because no account exists, nothing keeps your history either, so save the result within the thirty minutes the job is retained.