Does AI change faces when it restores a photo?
It is the most common worry: that Grandad will no longer look like Grandad. Here is what AI actually does to a face, when it is likely to change it, and how to check before you pay.
What does AI do to a damaged face?
It rebuilds missing detail from what is still visible: the outline of the eyes and mouth, skin texture, hair. The model has learnt what a sharp face looks like and recomposes one from the features it can see.
On Souvenotte, each detected face is cropped, rebuilt by a dedicated model, then blended back into the photo with a soft edge. The rest of the image is cleaned and enlarged separately, before this step.
Can the face actually change?
Yes, slightly, mainly when the original is very blurred or very small, because the model then has to invent part of the detail. The eyes, wrinkles, teeth or the shape of a smile can drift a little from the real person.
On a sharp, front-facing face, changes are usually limited to smoother skin and crisper eyes. The table below sums up the usual cases.
| Original face | What usually happens | Risk to the likeness |
|---|---|---|
| Sharp, facing camera, fairly large | detail refined, skin smoothed | low |
| Blurred or grainy | detail partly recreated | medium: check the eyes |
| Very small or badly damaged | a lot of detail invented | higher |
| Too small, in profile or hidden | not detected, only cleaned | low, but less sharp |
Which faces are not rebuilt?
The ones the detector does not find: very small faces (under roughly 24 pixels), faces turned well away, or partly hidden ones. They are only enlarged and cleaned along with the rest of the photo, with no reconstruction.
So they stay more faithful, but less sharp. On a school photo or group shot, that is common for people at the back. Scanning at a higher resolution helps small faces get detected: see what resolution to scan photos at.
How do you check the likeness before paying?
Use the before/after slider on the free preview, looking mostly at the eyes, the mouth and the shape of the face. If the person is no longer recognisable, don't pay: the preview is the full restoration, just watermarked.
Show the preview to someone who knew the person: that is the best test. If a face has changed too much, a professional retoucher can work by hand from the original.
Are the skin and eye colours real?
No: in a black-and-white photo, the AI knows neither eye colour nor hair colour. It suggests believable tints, with saturation deliberately kept moderate so skin looks natural.
For a family portrait, keep the black-and-white version as well; redoing the photo with the other colour choice is free from its page. To go further, read the real colours of old photos.
How can you limit changes to faces?
By giving it the best possible file: a sharp, reasonably large scan, with no phone filter or smoothing. The more real detail the AI has, the less it needs to invent.
Reconstruction applies to every detected face and is not adjustable photo by photo. So always keep your original scan alongside the restored version, especially for family history.
Frequently asked questions
Can I see the result before paying?
Yes: the preview is free. It is the full restoration with a watermark, at about 1,100 pixels, with a before/after slider.
Can AI rebuild a face covered by a stain?
No, a stain or tear across a face is not filled in. A retoucher is needed to rebuild that area.
Are my photos used to train the AI?
No, never. They are deleted automatically 7 days after upload.