Remove Scratches From a Photo

How it works

The scan is scaled to roughly one megapixel and passed to an image-editing model together with a written instruction: repair fine scratches and surface marks, preserve the people, objects, and composition.

The pass runs at a low edit strength. In a generative model, that setting controls how far the output is allowed to move from the input — a high value produces a reinterpretation, a low one produces a corrected copy. Style tools on this site run high. This one runs low on purpose, because the goal is a photograph that has been mended rather than a new photograph of the same scene.

There is also a negative instruction listing what the result must not contain: changed identity, altered composition, plastic skin, added or missing objects.

Why surface damage is the tractable case

Not all photograph damage is equally repairable, and the distinction is about evidence rather than severity.

A scratch is a narrow line where the emulsion was removed. The picture continues on both sides of it, usually only a few pixels apart. Rebuilding it is a question of continuing a pattern across a small gap, and both the model and a human retoucher do the same thing: read what is on either side and extend it.

Fading, color shift, and low contrast are also tractable, because the information is still present and merely compressed into a narrow range.

What is not tractable is a region that is gone. A torn-off corner, a section bleached to white, a face obscured by mould — there is no surrounding evidence for what was there. A model asked to fill it will fill it, plausibly, and you have a photograph of something that never happened.

The fixed seed

This is the decision most likely to look like a bug, so it is worth being direct about it.

Generative models start from random noise. Change the noise and you get a different image. Most tools on this site take a fresh random value each run, so that a result you dislike can be replaced by pressing the button again.

This one does not. The same photograph always produces the same repair.

The reason is that a repair makes a claim about what was there. If two runs rebuilt the same damaged cheek two different ways, both would be guesses and you would be choosing the more flattering one — which is exactly the failure mode photograph restoration should avoid. One answer per input keeps the tool accountable to the original.

Reading the result

Look at people first, before checking whether the scratches went. Damage crossing a face is where a plausible reconstruction is most likely and least acceptable, so compare the eyes, mouth, and jawline against the scan at full size.

Then look at a flat area — sky, a wall, a plain garment — to see whether grain and paper texture survived, and whether the repair introduced smoothness that was not there before.

Publication gate

This page ships once the workflow has been run against a creased scan, a scratch crossing a face, and a print with heavy grain, with faces and flat areas compared against their sources at full size.