Enlarging a picture invents the detail it appears to recover

Added detail is generated, not retrieved

Enlarge a small image and something has to fill the new pixels. Older methods spread the existing ones and produced a soft result that everyone correctly understood as blurry. Current methods produce sharp edges, plausible skin texture, believable fabric weave and crisp foliage, none of which was present in the original file.

That is the whole thing to understand, and everything practical follows from it. The output is a picture of what an image like this usually looks like at that size, conditioned on the pixels you supplied. It is a reconstruction with the confidence of a photograph.

Most of the time this is fine and genuinely useful. A low-resolution logo, an old photograph for a family album, a thumbnail that needs to fill a slide — in each case the detail is decorative and nobody is going to interrogate it.

The failure is that it looks better than it is

A blurry photograph carries its own warning. An upscaled one doesn’t, and the sharpness reads as recovered information to anyone who didn’t enlarge it themselves. That inversion is the reason this technique deserves care rather than a note in a list of tips.

The invented content clusters where the tool has strong expectations. Faces gain features that plausibly belong to a person and do not necessarily belong to this person. Text becomes crisp letters that may not be the original letters. Fine repeated patterns come back regularised in a way the real object wasn’t.

So the rule for anything where the detail carries meaning is simple: do not enlarge it, and if it has been enlarged, say so. Registration plates, serial numbers, documents, faces in any context where identification matters, medical or technical imagery. None of these should be read from a reconstruction, and treating a sharpened output as evidence is a serious mistake rather than a technical shortcut.

This is one of the clearest cases in the whole subject where the tool is the wrong instrument. If you need to know what a small region actually contains, the answer lies in a better source file, not in a better enlargement.

Restoration is the same trade in a friendlier setting

Repairing damaged photographs sits in the same territory and is mostly benign, because the goal is a picture that looks right rather than a record of what was there. Scratches removed, tears filled, faded colour corrected — the result is a better artefact than the damaged original.

It becomes less benign when the damage covered something meaningful. A face partly obscured, a date on a document, a background that has been reconstructed rather than restored. The output will not show any uncertainty about which parts it invented, so the person doing the work has to remember.

A practical habit: keep the original file untouched alongside the restored one, and mark the restored version as restored. This costs nothing and it is the difference between a family photograph that has been repaired and a family photograph nobody can verify.

Get the resolution at the start instead

Most enlargement is compensation for a decision made earlier. If a picture is going to appear large, generate or shoot it large, since resolution decided at the point of creation is real and resolution added afterwards is not.

When a set of images is being made for mixed uses, produce at the largest size any of them will need and reduce for the smaller placements. Reduction loses nothing that matters and it is the direction in which the arithmetic works honestly.

Moderate enlargement of an already reasonable file is the safe case and generally invisible. Large multiples applied to a small file are where the output stops being a version of the original and becomes a new image that resembles it.

Judge it at the size it will be used

The common review error is to inspect the enlarged image at full magnification, where every invented texture is obvious and it looks worse than it will ever appear in use. The opposite error is to check it only as a thumbnail, where nothing is visible at all.

Look at it at the size and distance of the actual placement. A poster is viewed from two metres. A slide is seen for thirty seconds. A print in the hand is inspected closely, and that is the case where invented texture around eyes and hair tends to give itself away.

And if the image will be examined by someone with a reason to look carefully, the safe assumption is that they will notice. That is a good reason to reserve enlargement for material where nothing depends on what the new pixels claim to show.