When an AI upscaler turns a face into someone else
Generative upscalers invent detail that was never in the frame, which is why a face can come back as someone else.
You enlarged a photograph, the result came back impressively sharp, and the person in it is not quite the person you know. This is not a bug in the tool. It is what that class of tool does, and understanding the distinction tells you which one to reach for.
Two different jobs sharing one name
"AI upscaling" covers two approaches that behave nothing alike.
Generative. The model produces plausible detail based on what it learned from training images. Given a blurry cheek, it produces skin texture that would be convincing on some face. The result looks sharp, often dramatically so.
Reconstructive. The model infers what the original pixels most likely were, based on the surrounding structure that survived. Given the same blurry cheek, it recovers what the smudge implies and stops there. The result is more restrained.
On a landscape, the difference rarely matters — invented leaves are still leaves. On a face, it matters completely, because a face is a specific arrangement that a viewer has memorised. Getting the eye spacing a little wrong makes it someone else.
Why faces break the hardest
Human perception is extraordinarily sensitive to faces. Shifts of a pixel or two in eye position, a change in the shape of a lip corner, or slightly different nostril width all register immediately, even when the viewer cannot say what changed.
The information available in a degraded face is also unusually thin. A face occupying 40 pixels in a group photo simply does not contain the identity-carrying detail, so a generative model has nothing to anchor its invention to, and fills in a statistically average face instead. The result is a sharp, confident, plausible stranger.
When invention is unacceptable
- Photographs of real people, particularly anyone who will see the result.
- Family and archival photographs, where the point is the record itself.
- Anything documentary or evidential. A generated detail is a false statement about what was in front of the camera.
- Character artwork that has to stay on-model. Same problem, different domain.
For all of these, a result that stays slightly soft is better than one that is sharp and wrong.
Getting a conservative result
- Open Pixmend and drop the image in. The models here are Real-ESRGAN family models, which reconstruct rather than generate.
- Choose x2 rather than x4. A smaller factor asks for less inference.
- Match the model to the content — the photo model for photographs.
- Compare at 100%, looking specifically at eyes, mouth corners, and the nose.
If the face still does not look right, the useful conclusion is not "try a stronger setting". It is that the detail is not in the file. A different source image is the only real fix.
How to check whether a face changed
Put the before and after side by side at the same display size and look at the relationships rather than the sharpness: distance between the eyes, the width of the mouth relative to the nose, the shape of the jawline. Sharpness draws the eye and disguises identity drift, which is exactly why generated results are convincing at first glance and unsettling on the second.
Anyone who knows the person will notice within a second or two, and will usually struggle to articulate what is wrong. That reaction is the signal.
The trade you are making
A reconstructive model will sometimes leave a face soft when a generative one would have made it sharp. That is the trade, stated plainly: you get a smaller, honest improvement instead of a large, invented one.
Which is correct depends on the job. For a background texture, invention is harmless and often preferable. For a person, the sharp version is a picture of someone who does not exist, and no amount of image quality makes that the better outcome.