How to check an AI restored photo before you keep it
A six point check at full zoom, eyes, teeth, hairline, hands, fabric and text, plus the side by side comparison that catches an invented face before it becomes the only version you have left.

A restored photograph can look finished and still be wrong. Not blurry wrong or oddly coloured wrong, the kind of wrong you would catch in a second. The kind that matters here is quieter: a face that is close but not quite right, a hand with the wrong number of fingers folded into shadow, a pattern on a dress that has resolved into something that looks like detail but was never there. These results pass a glance. They do not pass a look.
This happens because a restoration model is doing two different jobs at once, and nothing on the page tells you which job it just did. Where the picture survives but is degraded, the model is separating the image from the damage, and what comes back genuinely was there. Where the picture is missing, the model is generating a plausible fill based on what it has learned from other photographs, not from yours. Both operations produce the same clean, confident output. Our longer look at free AI restoration covers that distinction in full. This guide is about what to do with a result once you already have it in hand: six places to check, in order, and the comparison habit that makes the check reliable.
Why this matters more than it looks like it should
A restoration you do not check becomes the record. Print it, share it with family, file it as the photograph, and the quiet errors travel with it from then on. Nobody opens an old family photo at a hundred percent to re-examine it years later. The version you keep is the version that survives, mistakes included.
The stakes are highest on portraits of people, because a face is the one part of a photograph everyone can judge without being told what to look for. Get a fence post or a background building subtly wrong and almost nobody notices. Get a grandmother’s mouth subtly wrong and someone who knew her will feel it before they can explain why.
The six point check
Do this at full size, not the thumbnail the tool shows you on its results page. Zoom to one hundred percent, or as close to it as the screen allows, and work through each area in order.
Eyes. Compare the shape, the spacing and the direction of the gaze against the original. Eyes are small, high in information, and one of the first things a viewer’s own eye goes to, which makes them both easy to get subtly wrong and easy to catch.
Teeth. A soft or degraded original rarely shows much tooth detail at all. A restoration that hands back a full, even, bright smile where the source barely showed parted lips has invented a smile, not recovered one.
Hairline. Hair is dense, irregular and hard for a model to reconstruct convincingly, so the hairline often shows the seam between recovered and generated detail more clearly than the face itself does. Look for texture that is too uniform, or a line that sits in a different place than the original.
Hands. The single most reliable tell across almost every image generation model. Count fingers. Check that a hand resting on a shoulder or holding an object still makes physical sense. Hands in deep shadow or partly out of frame in the original are exactly where a model is most likely to guess.
Patterned fabric. A lace collar, a tartan, a woven basket. Models tend to resolve irregular, hand-made pattern into something that looks orderly and repeats too cleanly. Real fabric has small irregularities that a generated version usually smooths away.
Text and signage. Lettering on a shopfront, a badge, a street sign, a newspaper someone is holding. If the photograph matters for what it records rather than only for who is in it, this is the detail most likely to be quietly wrong, because the model is drawing letterforms it finds plausible rather than reading what was actually printed there.
If a photograph has no faces, hands or fabric in frame, the check still holds. Apply it to whatever carries the most information: a building’s brickwork, a vehicle’s badge, the fine structure of a tree line. The principle is the same everywhere. Anything dense and specific is a place a model can quietly get it wrong.
The invented detail test
Checking a restoration on its own tells you whether it looks convincing. Checking it against the original tells you whether it is right, and those are different questions. A face can look completely natural and still be the wrong face, because a model trained on millions of other faces is perfectly capable of producing a natural one that simply is not this person’s.
Open both files side by side, at the same zoom level, and work through the six points above on each image in turn rather than judging the restoration in isolation. Do not trust your memory of what the original looked like. Memory is generous, it fills gaps the same way a model does, and it will wave through exactly the kind of change you are trying to catch.
Three signs that a detail has been invented rather than recovered, once you have both images in front of you:
It is too clean for the photograph’s age. A century-old print that suddenly shows crisp individual eyelashes or sharp fabric weave has been given detail the original resolution could never have held in the first place.
It repeats. Generated texture, hair strands, lace, brickwork, often shows a pattern that is just a little too regular, the same motif recurring in a way real, worn material rarely does.
It is plausible but unfamiliar. This is the hardest one to describe and the one that matters most on a face. The restored version looks like a real person, just not quite the person in the original. If something feels slightly off and you cannot name why, trust that feeling enough to look closer rather than dismissing it.
If you can show the comparison to someone who actually knew the person, that remains the best likeness test available, and it costs nothing beyond asking.
Keep the original, always
This is the one habit that makes every mistake above recoverable. Never let a restored file replace the scan it came from. Save the restoration as a clearly named copy, keep the untouched original in its own place, and treat the restoration as a version, not as the record.
This protects you two ways. If a result turns out to have invented a face or a detail you only notice later, you still have the source to work from or to send to a different tool or a human retoucher. And restoration models keep improving. What a model produces today will look different from what one produces in two or three years, and you will want to try again from the original scan, not from a copy another model already guessed at.
If you have not scanned the photograph properly yet, start there. A restoration is only as good as what it is given to work from, and our scan resolution calculator will tell you the right setting before anything touches a restoration tool at all. Once a file exists, backing up scanned family photos covers where the original and the restored copy should each live so neither one is ever the only surviving version.
When the check fails
If a restoration does not hold up under this check, that is useful information, not a dead end. On structural damage specifically, torn paper, chunks missing from a corner, heavy scratching across a face, a one-click model is working with very little to go on, and inventing conservatively is often the best it can do. Our page on scratches on old photographs shows the difference between a narrow scratch a model can reconstruct with real confidence and heavier damage where it is mostly composing.
When a face specifically has drifted, the honest options are to accept the restoration as an interpretation rather than a record and say so when you share it, try a different tool and compare results against each other as well as against the original, or hand the photograph to a human retoucher who can make considered judgement calls about what a missing area should plausibly contain, informed by the rest of the photograph rather than by a general idea of faces. Our guide on what photo restoration actually costs sets out what that tier looks like and when it earns its price.
A five minute habit worth keeping
Checking a restoration properly takes five minutes: open both files, zoom to full size, work through eyes, teeth, hairline, hands, fabric and text, and compare rather than recall. That is a small amount of time against a photograph you cannot get a second original of.
The tools keep getting better at this, which sounds like reason to check less. It is the opposite. A beautiful, convincing restoration is harder to doubt than an obviously wrong one, and that is exactly why the five minutes matters more each year the models improve, not less.
Common questions
- How do I know if AI restoration changed my photo's face?
- Open the original and the restored version side by side at full size and compare the eyes, the shape of the mouth, the hairline and the set of the jaw one at a time. A changed face usually still looks like a plausible person, just not quite the right one, so judge it against the original rather than against your memory.
- What should I check first on a restored photo?
- The face, before anything else. It is the area a model is most likely to reconstruct rather than recover, and it is the area where a wrong guess matters most. Only after the face holds up is it worth checking hands, fabric, text and background detail.
- Can AI restoration invent detail that was not in the original photo?
- Yes, and it does this constantly in places where the original picture is missing or badly degraded. The model fills gaps with a plausible guess based on what it has learned from other photographs, not from the one in front of it. The result looks confident and finished either way, so the invention is not labelled.
- Should I keep the original scan after restoring a photo?
- Always, and never overwrite it. A restoration is one interpretation of the photograph made by one model at one point in time. Keeping the untouched scan means you can try again later, check a result against the source at any point, and recover if a restoration turns out to be wrong.


