What actually happens when AI restores a photograph

Ask three people what "AI photo restoration" means and you will get three answers. One is thinking of scratch removal. One is thinking of a black-and-white portrait turned to colour. One is thinking of the upscaler on their phone that makes a blurry face suddenly sharp. Those are three different problems, they fail in three different ways, and the difference explains why some restored photographs come back looking like a stranger.

Damage repair is the easy part

A torn corner, a crack across an emulsion, a coffee ring, the white speckle of dust: these are missing or corrupted pixels surrounded by intact ones. The technical name is inpainting, and it is the most mature part of the stack. A model looks at the neighbourhood around a hole and predicts what belongs there. Because the surrounding context is real, the prediction is tightly constrained. Brick continues as brick, sky continues as sky.

This is why heavy physical damage is often less frightening than it looks. A photograph with a tear straight through the middle can come back cleanly, provided the tear does not run through a face.

Fading and colour are a correction, not an invention

Dye layers in a colour print fade at different rates. Cyan usually goes first, which is why 1970s prints drift orange and 1980s prints drift red. That is a predictable, physical process, and reversing it is closer to a calculation than to a guess. The information is still in the print, just unbalanced.

Colourising a black-and-white photograph is a genuinely different act. There was never any colour data. A model trained on millions of images knows that grass is usually green, that a 1940s military uniform is usually a particular olive, that skin falls in a certain range. It applies the most probable colour. The shapes are real. The exact shade of a dress from 1938 is an educated guess, and anyone who tells you otherwise is selling something.

That does not make colourisation dishonest. It makes it an interpretation, and it should be labelled as one.

Faces are where restoration goes wrong

Here is the part that matters most and gets discussed least. Most general-purpose image models include a face enhancement step. It exists because "sharpen the blurry face" is what most consumers want from a photo app. The model has learned what faces look like in general, and when detail is missing it fills in from that general knowledge.

On a modern snapshot that is harmless. On a fifty-year-old photograph of someone who died in 1994, it is a catastrophe. The model does not know your grandmother. It knows faces. Given a soft, grainy, low-contrast face, it will produce a plausible face — symmetrical, smooth, subtly younger, subtly prettier. It will not be her.

This is the single most common complaint about consumer restoration apps, and it is not a bug in the usual sense. The tool is doing exactly what it was built to do. It was simply built for a different purpose than the one people bring to it.

The constraint worth asking about

The useful question to ask any restoration tool is not "how good is your model" but "what is it not allowed to change".

A restoration should be permitted to repair cracks, tears, stains, missing corners and fading, and to return colour. It should not be permitted to alter the face, the apparent age, the expression, the glasses, or who is in the picture. Those are not aesthetic preferences. They are the reason the photograph exists.

When a result comes back, the check takes ten seconds. Open the original and the restored version side by side and look at one thing: the eyes, and the distance between them. Then the mouth corners. Then anything the person was wearing on their face. If those have moved, you are not looking at a restoration.

What this means in practice

If the photograph is damaged but the face is intact, almost any competent tool will do a reasonable job.

If the face itself is degraded — soft, small in frame, badly exposed — that is where tools diverge sharply, and where you should be most suspicious of a spectacular result. A dramatic improvement in a face that had very little information left in it is a warning sign, not a triumph.

And whatever you use, work on a copy. Photograph or scan the print, keep the original file untouched, and let the software have the duplicate. The print itself should never be the thing you experiment on.

NNC Restore is built on the second constraint above: it repairs damage and returns colour, and it is not permitted to change the person. If a face comes back wrong, the restoration is re-run free and refunded if it still fails. That guarantee is not marketing generosity. It is the only honest response to the one failure mode that actually matters.