What AI enhancement invents

An AI enhancer does not find hidden detail in a small photo. It paints in detail that looks like what photos of that kind usually contain. Often that is exactly what you want — and sometimes it is the one thing you must not have.

Recovered or drawn?

A classical enlargement — Lanczos, bicubic — makes each new pixel out of its neighbours. Nothing appears that was not implied by the pixels around it, which is why the result looks soft: the information is simply not there. A super-resolution model is trained on millions of pairs of small and large pictures and learns what large pictures tend to look like. Given your small photo, it produces a large one that is consistent with it and looks like a photograph. The texture in brick, hair, grass and fabric in that result is the model's, not the camera's.

That is not a flaw; it is what the model is for. But it changes what the picture is evidence of. A model-enlarged face is a plausible face, not a better record of the one in the original.

Where models invent most

Typical behaviour of a general super-resolution model
Content	What usually happens	Risk
Repeating texture (brick, grass, fabric)	convincing new texture	low for a picture you will look at; misleading as evidence
Text and numbers	letters sharpened into the nearest familiar shape	high: a 3 can become an 8
Faces	skin smoothed, features firmed up	high: identity can drift
Flat areas (sky, walls)	left mostly alone	low
Noise and JPEG blocks	sometimes cleaned, sometimes turned into texture	medium

How this site shows you

The model stages on this site run only after the classical result is already on screen, and the saved file is named and labelled as containing reconstructed detail. The difference view overlays a heat map of how far the model's brightness differs from the classical result, and prints one sentence: how many pixels moved by more than four levels, and in which part of the frame the largest changes are. Use the classical result is one click away.

When not to use a model

Do not use a model-enlarged picture where the content is the point: a document, a number plate, a receipt, an ID photo for a form, or anything you might need to rely on later. For forms in particular, portals check faces, and a smoothed face is the wrong thing to send. For those, the classical stages — denoise, deblock, deconvolution, Lanczos — change only what they measure and invent nothing.

A model result is labelled reconstructed in the save card and the receipt.
Face restoration is not offered: no model was found that is commercially licensed, small enough to run in a browser and measured to work.

Take it back to the picture

Corrected on your own machine — we never get a copy.