Two different things under one name
One of them is raising the resolution: turning few pixels into more pixels while keeping the result sharp. The other is restoration: turning a scratched, faded, blurred or damaged picture into a usable one. The same technology solves both jobs today, and the same button starts them, but their risk profiles are very different.
The old methods simply smeared the existing pixels around — the picture got bigger and blurrier. Today's tools are generative models instead: out of a blurry patch they manufacture plausible detail. That is where the crisp eyelash, the readable sign and the fabric weave come from — things that are physically not in the original file.
The machine does not enlarge, it invents
That sentence is not pedantry, it is the heart of the subject. The model works by the same mechanism as the image generators: it builds a complete picture out of a noisy, incomplete input, and puts in whatever detail its training data says is most likely. Exactly how that happens is covered by how a diffusion model works, which is worth reading first.
The resulting error has a name of its own: hallucination. Not blur, not noise, but confidently drawn-in detail that never existed. An eight becomes a six, wrinkles appear on a face, a distant sign resolves into a meaningful word — just not the one that was actually there.
How much does it invent? There are numbers
This gets measured, and the measurements are sobering.
- In an evaluation built specifically for this, the restoration models tested scored between 2.5 and 3 for hallucination on a five-point scale — meaning invented detail is not the exception but the norm.
- With text content it is measurable: at four times enlargement the legibility score for lettering ranged between 39 and 60 percent across the models tested.
- Which is to say that on an enlarged number plate, account number or house number every second character may be doubtful.
- And the worst part: the wrong character looks back at you just as sharply and just as confidently as the right one.
Faces are the most dangerous part
Restoring faces is a research field in its own right, because this is where the human eye is most sensitive — and because here the error is not aesthetic but a matter of identity. At a professional challenge in 2026 the best entries scored between 0.68 and 0.85 on a face identity metric running from zero to one; one entry simply failed the identity test on 82 images. The system produced a handsome face, in other words — just not that person's.
The emblematic case in the field is a 2020 research system that made high-resolution pictures out of 16×16 pixel inputs. From a pixelated photo of a well-known politician it generated a white man, and analysis of the system found that 72.6 percent of its faces came out white regardless of the input. That was not deliberate bias but the imprint of the training data — and it shows precisely where the "detail" the machine adds comes from.
Where this stops being an aesthetic question
If the image is evidence, a document or the basis of an identification, enlarging it is not improvement but a risk of forgery. There are three instructive cases.
A photocopier fault from 2013 went unnoticed for years: while compressing, the machine swapped similarly shaped digits for one another on construction blueprints, so the copy looked sharp and flawless — it simply had different numbers on it. The other case is recent and unambiguous: in 2024 a US court excluded from evidence a video that had been AI-upscaled, on the grounds that the processing had added sixteen times as many pixels to the footage as the original contained. The 2023 examination of a phone maker's moon photographs showed the same boundary from another angle: the machine put crisp craters onto a deliberately blurred, tiny image of the moon, because it knew what the moon looks like.
Colourisation is its own topic
Colourising black-and-white photographs belongs here technically, but it deserves separate thought. The machine does not know what colour the dress was; it knows what colour a dress like that usually is in the pictures it has seen. The guess is often convincing, but it is always a guess.
The field itself does not treat this as neutral: as one art-historical analysis puts it, a colourised archive photograph is no longer a witness to the past but a reinterpretation of it. With a family photo that is entirely fine — there the experience is the point. In an archive, an exhibition or a publication, though, it ought to be labelled.
What to do in practice
Four rules that head off most of the trouble.
- Keep the original in a separate file, and name the enlarged version recognisably. That is your only way back.
- Look at it at 100 percent zoom wherever there is a number, a caption or a face. That is where the damage happens.
- Do not expect more than two to four times. One professional tool deliberately allows only double the side length, and only once — a good rule of thumb elsewhere too.
- Do not use it on documents, evidence or identification. There a blurry true image is worth more than a sharp invented one.
Common misconceptions
- "The detail was in there, it just had to be brought out." It was not in there. Whatever is missing from the file, the model adds from what it has learned.
- "If the result is plausible, it is fine." Plausibility is the problem here, not the solution: invented detail is dangerous precisely because it does not look invented.
- "Run it twice and it gets even better." Running it again compounds the invented detail, and the second pass amplifies the first pass's inventions.
- "Restoration and editing are the same thing." They are relatives, but the goals differ. Targeted correction is covered in editing an existing image, where the risks look rather different too.
What to take away
Enlargement and restoration are excellent tools — for family photographs, old prints, low-resolution archive images. One thing has to stay in mind: added detail is not a discovery, it is a guess. Wherever that carries stakes — a number, a face, a document, a piece of evidence — the blurry original is the more truthful one.