Restore Detail in a Face Photo

How it works

Two stages, and the separation between them is what makes this safer than processing the whole photograph.

First a detector locates faces in the image and produces a mask for each one. Nothing outside those regions is touched at any point.

Then each detected face is repainted by RealVis XL, a photographic image model, working only inside its mask. The instruction it is given asks for natural facial detail, faithful restoration, realistic skin texture, and preserved identity and expression. A second instruction rules out the failure modes of this category by name: changed identity, a distorted face, extra eyes, plastic skin, and oversharpening.

The rest of the photograph — background, clothing, hands, everything — is the original file, unmodified.

Identity is the thing at risk

A model repainting a face is making decisions about a specific person's appearance, and the constraint asking it to preserve identity is a request, not a guarantee.

How much risk there is depends almost entirely on how many pixels the face occupies. A large, well lit, moderately soft face contains enough information to constrain the result, and what comes back is recognizably a cleaner version of the same person. A small face — one of a group, someone in the background, a distant subject — does not constrain much, and the output is a plausible face assembled around a few dozen pixels of evidence.

The second case is the dangerous one, precisely because the result looks sharp and definite. Confidence in the output is not evidence about the input.

What to check, in order

Compare the output with the original at the same size, on the face, at full resolution.

Start with the features that carry identity: the shape of the nose, the distance between the eyes, the line of the jaw, the shape of the mouth at rest. These should be the same. If any of them has moved, discard the result — no amount of improvement elsewhere is worth a face that is not the right face.

Then look at asymmetry. Real faces are asymmetric, and a common failure is a face that has been quietly tidied into symmetry, which reads as subtly wrong without being locatable.

Then skin. The texture should still have pores and fine lines. A face that has come back perfectly smooth has been replaced rather than restored, even if it is the same person.

Finally glasses, jewelry, and teeth, which sit at the boundary of what the model handles well and often show the first artifacts.

Where you should not use it

Anything that functions as identification. A passport photograph, a visa application, a driving license, an employee badge, an official form — these are documents where the image is supposed to be a record, and submitting a model-modified face is at best invalid and at worst a misrepresentation.

Anything evidential, for the same reason. A restored face is not a better look at what the camera saw; it is a new image informed by what the camera saw.

Journalism and documentary work, where the standard is what was recorded.

Where it earns its place

A soft family photograph you want to print larger. A group shot where one face came out slightly out of focus. An old scan destined for a frame rather than an archive. A portrait for personal use where the goal is a good picture and everyone involved knows how it was made.

In all of these the original stays untouched and available, which is what makes running the workflow a low cost thing to try.