Where the detail in a restored face comes from
Search for this and you get two kinds of page. One is a research summary written for the people building these models, and the other is a list of tools with a before and after strip attached. Neither answers the question somebody holding a photograph actually has, which is whether their own photograph has enough face left in it to be worth restoring at all.
That question can be settled by looking, and it matters more than which tool you pick, because every tool in this category behaves the same way at the point where it fails. The failure is not that the output comes back soft. It is that the output comes back looking like somebody else, and looking confident about it.
A restoration model is not a sharpener
Sharpening works on detail that is present but low in contrast. It finds an edge that survived and makes it more definite. Everything in the result was already in the file, which is why sharpening can disappoint but cannot really lie to you.
Face restoration does something different in kind. It reads whatever survived of the face, works out what a face consistent with that would look like, and draws that. The eyelashes in the output were not uncovered. They were supplied, because a face without eyelashes is not a plausible face, and plausibility is the thing being optimised.
That is why the category produces results so much better than sharpening on a damaged face, and it is also why it has a failure mode sharpening cannot have. A sharpener given almost nothing returns a crisper version of almost nothing. A restoration model given almost nothing returns a clean, well lit, entirely invented face.
Identity drift, the failure worth naming
What goes wrong has a direction to it. As the amount of surviving face shrinks, the output moves away from the person and towards an average. Features that were never recorded get filled in with the most ordinary version of themselves: the nose straightens, the eyes even up, and the small asymmetries that made the face recognisable quietly leave.
On a clear photograph this is invisible, because there was enough information to pin every feature down and the model had almost no room to invent. On a face that is a dozen pixels across there is nothing to pin, so nearly all of the output is invention, and it arrives looking sharper and more certain than the original rather than looking like a guess.
This is the thing worth understanding before restoring a photograph of somebody you knew. A soft photograph of your grandmother is still a photograph of your grandmother. A sharp photograph of a plausible woman is not, and it is very hard to unsee once you have it on screen.
The test: how much face is actually left
This takes under a minute and needs no tool. Open the photograph, zoom to full size on the face rather than fitting the whole frame on screen, and look for four specific things.
- An edge to the iris. Not a dark patch where the eye is, but a real boundary between the iris and the white around it.
- A line where the lip meets the skin, rather than one colour fading into the other with no border.
- Separation at the lashes or the brow, so you can see that the hairs are many things rather than one solid shape.
- Any texture at all in the skin. A pore, a crease, a freckle, a shadow with a definite shape to it.
Reading the result of that test
Count how many of the four survived. Three or four means the face is well recorded and a restoration has real material to work from, so what comes back will be an improvement on the original rather than a substitute for it.
Two puts you in the middle, where the result is usually worth keeping and always worth comparing against the original before anything gets deleted.
One or none means the face is below the line. A tool will still return something, that something will look better than what you fed it, and it will not be dependable as a likeness. If the point of the photograph is to remember what somebody looked like, this is the case to leave alone.
What you are counting matters more than the pixel dimensions of the file. A face can be large in a big photograph and still hold none of the four, because the photograph was out of focus or because it was compressed hard somewhere along the way. The test reads the face, not the file size.
Group photographs are the hard case
Nearly every photograph that fails this test fails it for the same reason, and the reason is not damage. It is that the face is small because the photograph was never about that face. A wedding group, a school row, a street scene with a relative somewhere in it.
Those faces occupy very few pixels even in a careful scan, because the scanner resolves the whole frame and the face is a small part of the frame. Scanning the print again at a higher setting genuinely helps here, since the detail is still in the physical print and it was the scan that threw it away. Enlarging the scan you already have does not help, because by then the detail has already gone.
So the order is worth getting right. Rescan first, restore second, if the print still exists. The scanning guide on this site covers the settings worth using, and rescanning is the one step in this whole subject that adds real information rather than plausible information.
How to judge a restored face honestly
Compare it against another photograph of the same person rather than against the input. The input is soft, so anything sharp beats it and your eye will say yes to a stranger. A second photograph of that face, from any year of their life, settles the question in seconds.
Keep the original file as well, and not merely as a backup habit. The original is the evidence and the restored version is an interpretation of it. They are two different objects and an archive wants both, particularly if the photograph is ever going to be shared with family who will take it as a record.
Look at the eyes first, then the corners of the mouth, then the nose. Those carry most of what makes a face identifiable, and they are also where invention shows up first, in that order.
Where Clara sits in this
Clara enlarges, sharpens and restores detail, faces included, and the honest description of what that means is everything above. It is the right instrument for a face that still holds most of those four markers and the wrong instrument for a face that holds none of them, and it cannot tell you which one you have. You can, in under a minute, with the test above.
The repair side of this subject has its own pages here rather than being restated: what an upscaler can and cannot add, and where to start on a print that is damaged as well as soft. This page is only about the face, because the face is the part where being wrong matters.
More guides
Questions
What does AI face restoration actually do?
It reads what survived of a face and draws a face consistent with that, rather than finding detail that was already there. The output is a plausible reconstruction, which is why it looks so much better than sharpening on a damaged photograph and why it cannot be trusted as a likeness once there is very little face left to read.
Why does a restored face look like a different person?
Because the features that were not recorded had to be filled in, and what fills them is the most ordinary version of each feature. The nose straightens, the eyes even up, the asymmetries that made the face recognisable go. The less face there was to begin with, the more of the output is invention, and invention averages.
How do I know if my photo has enough face left to restore?
Zoom to full size on the face and look for four things: an edge to the iris, a border where the lip meets the skin, separation at the lashes or brow, and any real texture in the skin. Three or four of those means restoration has something to work with. One or none means it will be inventing the face.
Does scanning the photo again at a higher setting help?
Yes, if you still have the print, and it is the only step here that adds genuine detail. The detail is in the print, and a low scan setting is what discarded it. Enlarging a scan you already have cannot bring it back, which is why rescanning comes before restoring rather than after.
Is face restoration the same as sharpening?
No, and the difference decides what can go wrong. Sharpening only ever works with detail already in the file, so it can underwhelm but it cannot replace a feature. Restoration generates detail that is consistent with what survived, so it can transform a damaged face and it can also quietly change whose face it is.
Should I keep the original after restoring a photo?
Always. The original is the record and the restored copy is a reading of it, and only one of the two can be checked later. Keep both side by side, especially if the photograph is going to be passed around a family, where a sharp version tends to become the version everybody remembers.