Can AI photo restoration change what people look like?

Many people who restore an old photo with AI for the first time get a strange feeling: the photo is sharper, but the person seems to have changed. It isn't your imagination. It follows from how generative restoration works.

Why faces change

Faces in old photos are often only a few dozen pixels wide, and details like eyes and lips have long since blurred or disappeared. Generative AI can't "recover" information that's gone. Instead, it infers what those areas most likely looked like, based on the many faces it has seen, and draws them in.

The more it infers, the sharper the result looks, and the more likely it is to drift from the real person. Sharp doesn't mean accurate.

What's most likely to go wrong

• Eyes: eye shape, eyelid folds, direction of gaze
• Mouth and teeth: the curve of a smile, the shape and alignment of teeth
• Glasses: frame shape, lens reflections
• Hairline and hairstyle
• Apparent age: with wrinkles smoothed away, people can look younger
• Text in the photo: signs, dates and handwriting may be "restored" into words that never existed

How to reduce the difference

1. Start with a good original: the sharper the original, the less the AI has to guess. For details, see "How should you scan or photograph old photos for the best restoration?"
2. Restore important photos conservatively: remove scratches and fading only, without enhancing face detail.
3. Check with people who knew them: put the restored version next to the original and ask family members who knew the people in the photo to confirm.
4. Always keep the original: the restoration is a keepsake; the original is the record.

How TimeMend handles it

In TimeMend's restoration studio, "Enhance face detail" is a separate switch. It's off by default and warns that features may change. Results are shown side by side with the original, and the original is never overwritten.