Why restored photos
so often look fake.

The uncanny look has specific causes

“Looks AI” is a real perception with identifiable mechanics behind it. There are four common ones, and they are largely independent — a restoration can fall down on any of them on its own.

1. The face was generated, not restored

Most face-restoration models are generative. Handed a soft, damaged face, they do not sharpen it; they produce a new, plausible face conditioned on what they were given. On a clear photograph the result is close enough that nobody notices. On a small, blurred or partly damaged face the model has little to work from and fills the gap with the average of everything it has ever seen. That is the mechanism behind “it does not look like her any more”.

2. Sharpness is rewarded, even when it is invented

The obvious way to measure improvement is to measure detail, and invented detail scores just as well as recovered detail — often better, because it is cleaner. Optimising for a sharpness number therefore actively selects for making things up. A restoration can be measurably sharper and less faithful at the same time.

3. The face and the rest of the photograph get different treatment

Faces are usually processed by one model and the rest of the frame by another. If nothing reconciles them, you get a crisp face pasted onto a soft background, or a face noticeably brighter or more saturated than the photograph it sits in. The eye reads that mismatch instantly, even when it cannot name it.

4. Grain is treated as damage

Film grain is part of what makes a photograph look like a photograph. Aggressive denoising removes it completely, leaving flat areas of tone with hard edges between them. That posterised, painted look is one of the strongest tells, and it is the hardest to undo afterwards.

How to judge a restoration

You do not need any special tools to assess one — only a deliberate look at the right things.

• Compare at the same size. Most before/after presentations show you a small “before” and a large “after”. Put them at identical dimensions before deciding anything.
• Go to the eyes, teeth and jewellery. These are where generative models invent most confidently. Teeth that are individually distinct in a photograph too soft to resolve them are invented teeth.
• Check text and pattern. Lettering on a sign, a book spine, the weave of a fabric. Invented detail tends to be plausible-looking nonsense at close range.
• Look for grain. If every trace has gone, detail has probably gone with it.
• Ask someone who knew them. The only test that matters for a face is whether the person still reads as themselves to somebody who remembers them.

What Revelo does about it

Revelo takes the same generative tools and constrains them, because the constraints are the difficult part rather than the restoration itself.

• Faces are cropped from the original photograph, never from an already-enlarged version, so face pixels pass through a generative model once rather than twice.
• Restored faces are reconciled against the frame they came from — matched in per-channel levels and capped on how much extra detail energy they may bring back, so a face cannot arrive brighter, more saturated or sharper than the photograph around it.
• Drift from the original is measured, against the original photograph rather than an already-processed version of it. A face that has moved too far from where it started is a failure, not a result.
• A face that cannot be restored honestly is declined rather than restored badly.
• Grain is restored deliberately after denoising, because a photograph with no grain at all does not read as a photograph.

BeforeAfter Revelo↔A man in glasses holding a sleeping baby, photographed indoors in the 1970s — drag to reveal more of the restored version