Image glossary

Super-resolution

Also known as AI upscaling, neural upscaling

Super-resolution is the use of a trained model to increase an image's resolution by generating plausible detail consistent with its training data, rather than by interpolating between existing pixels.

What it does differently

A super-resolution model is trained on millions of pairs: a high-resolution image and a degraded version of it. Over that corpus it learns the statistical relationship between what a downscaled eyelash looks like and what the eyelash looked like.

At inference it applies that learned mapping. Given forty pixels that look like a downscaled eye, it generates the several hundred pixels that a real eye would most plausibly have produced.

The key word is plausibly. The model is not recovering your detail. It is generating detail consistent with its training data, conditioned on the low-resolution evidence in your file.

Where that is fine and where it is not

On texture it is excellent. Skin, hair, fabric, foliage, stone, water — the eye judges these on whether the texture is plausible, not on whether individual elements are correct, and nobody can tell that leaf 4,182 is in the wrong place. Portraits are the strongest case because faces are the most heavily represented category in every training set.

On information it is a liability. Asked to enlarge illegible text, a model produces sharp, confident, fabricated letters — which is far worse than remaining blurry, because a crisp result invites you to read it. The same applies to serial numbers, licence plates, logos with specific letterforms, and any face that has to be identified rather than merely admired.

The question to ask first

Before running an upscale: would a plausible-but-different detail here cause a problem? A jumper's knit pattern invented differently is fine. A serial number invented differently is not. Leaves on a tree, fine. A face in a legal exhibit, not.

That single question resolves nearly every real case, and it reduces to a rule: texture is safe, information is not.

Practically: start from the best source you have rather than a compressed copy, prefer 2x over 4x where you can because more evidence per generated pixel means fewer inventions, do not run output back through the model, and check the result at 100% on any text, faces, and fine repeated patterns.

Tools that work with this

Back to the full glossary.