When a restored face is not quite them
Old photo faces sometimes come back unrecognizable after AI restoration. What identity drift is, why bad damage causes it, and how to check a result.
A search for old photo faces unrecognizable restore usually starts the same way: a scan came back from an AI tool, the scratches and stains are gone, and the person in the photo is someone else. Not a small change — a different face wearing the right clothes. This happens often enough to deserve a name — identity drift — and it happens for a specific, explainable reason rather than because the software is simply bad.
Why the model starts guessing
Restoring a face works from evidence sitting in the pixels: the exact spacing between the eyes, the particular fold of an eyelid, the angle of a jaw, a faint asymmetry that makes this face this face rather than a face in general. Where that evidence survives in the scan, a careful restoration is anchored to it — repairing a scratch across a cheek does not touch the eyes six centimeters away, because nothing about that scratch tells the model anything about the eyes.
The trouble starts where the evidence itself is gone. A crease running through both eyes, a chunk of print missing over the mouth, a face so faded that skin and background sit at nearly the same tone, or a print so small and soft that a face occupies a few dozen pixels — all of these remove the specific information a restoration needs and leave a face-shaped gap that still has to be filled with something. The model fills it with the most statistically plausible face for that pose, lighting, and apparent age, drawn from patterns learned across a great many photographs, not from anything unique to the person who was actually there. A little bit of that guessing is invisible — it is the same soft smoothing restoration does automatically to fine texture. A lot of it produces a complete, coherent, entirely convincing stranger.
Drift is not only about torn or missing faces
It is tempting to assume identity drift only happens where a piece of the photo is literally gone — a tear, a hole, a chemical stain that ate through the emulsion. That is one cause, and it is covered in detail, including how to judge how much of a face has to survive for a reconstruction to be trustworthy, in the guide to restoring a photo with missing pieces.
But a face does not need a physical hole in it to run short on evidence. A photo copied through window glass at an angle, a snapshot taken from across a room and cropped in tight afterward, a heavily faded print where the whole face has gone flat and grey, or a scan so low-resolution that a face is a handful of soft pixels — none of these are missing anything in the sense of a tear, and all of them can leave the model with too thin a signal to trust. The mechanism is the same either way: less real evidence in, more invention required to fill the frame, higher risk that the invented part looks like somebody else.
What an honest restoration does differently
The difference between a tool that quietly guesses and one that is honest about guessing is not in how good the guess looks — a confident wrong face and a confident right one can look equally polished at first glance. The difference is in what happens before and after the guess. Before generating anything, an honest process looks at how much of the face actually survived and says so plainly when too little did, rather than letting the first surprise arrive after payment. After a result comes back, you compare the restored face against whatever of the original remains, checking the things that make a face specific rather than generic. If the comparison shows drift, the honest options are to try again anchored more tightly to the surviving evidence, or to say directly that this particular face is a reconstruction rather than a record.
That last option matters more than it sounds. Sometimes the correct, if unwelcome, answer really is that a face cannot be restored without inventing it — both eyes gone, or damage covering essentially the whole face — and the more useful thing a restoration can do at that point is say so rather than hand back a stranger with confidence. Epoha works this way on purpose: faces are told in detail, before repair, what may not change, and every result is something you look at next to the original before paying for it, specifically so a drifted face is your decision to reject, not a surprise you discover afterward.
None of this argues for AI over a human retoucher — that broader choice is its own question, answered in the comparison of AI and manual photo restoration. This page is narrower: one specific failure inside AI restoration, and how to catch it before trusting the result.
How to check a restored face against the original
- Distance between the eyes — hold the two photos side by side and compare eye spacing relative to the width of the face; a shift here is one of the fastest tells that a face has drifted
- Jawline — the same width and the same angle where it meets the ear, not a narrower or squarer chin than the original had
- Ears and hairline — ear shape rarely gets attention in a quick check, but it is highly individual, and so is exactly where hair starts at the forehead and temples
- Apparent age — restoration should not make someone visibly younger or older than the original photo's own skin, light, and posture show
- Expression — a closed mouth should stay closed, a slight smile should stay slight, not sharpen into a different one
- Asymmetry — a mole, an uneven brow, a faint tilt to the head: these small imperfections are personal signatures, and a version that has quietly smoothed them into perfect symmetry is worth a second look
Faces changing during restoration — questions
Why did AI restoration make my grandmother look like a stranger?
Most likely because too little real evidence survived in the damaged area of her face for the model to reconstruct it from — a crease, heavy fading, or very low resolution left it filling the gap with a statistically plausible face rather than her actual features. The more of a face is damaged or unclear, the higher that risk gets.
What is identity drift in AI photo restoration?
It is the point where a restored face stops matching the person in the original and becomes a different, if convincing, face. It happens because the model has to fill damaged or unclear areas with an estimate, and a large enough estimate can drift away from the specific person entirely rather than staying anchored to their real features.
How can I tell if a restored face is accurate before I trust it?
Compare it against the original point by point rather than glancing at the overall impression: eye spacing, jawline, ear shape and hairline, apparent age, expression, and any small asymmetry the person actually had. A result that gets the overall vibe right but changes several of those specific details has likely drifted.
Can a badly damaged face always be restored accurately?
No, and an honest answer says so. Where both eyes are gone or damage covers essentially the whole face, there is no real evidence left to anchor a reconstruction to, and any result is a plausible face for that pose and era rather than the actual person. In that case, restoring everything else in the photo well and being upfront about that one area is more honest than a confident guess.
See the face before you decide
Upload the scan and look at the restored face next to the original yourself. In early access the restoration is free, and you decide whether the likeness holds before you rely on it.
Epoha is in early access: restoration is free and there is nothing to pay yet.