super-resolution

Super-resolution takes a small or blurry image and produces a larger, sharper version with more pixels. A naive enlargement just stretches each pixel into a fuzzy block; super-resolution instead invents convincing fine detail — crisper edges, plausible texture — that was not present in the original. Modern methods are trained by taking sharp photos, shrinking them, and learning to undo the shrinking, so the model absorbs what real detail tends to look like and paints it back in.

This is genuinely useful. It rescues old low-resolution photos and videos, lets a camera capture more apparent detail than its sensor truly resolves, sharpens medical and satellite imagery, and lets streaming services send a small file that is upscaled on your device. When the missing detail is statistically predictable — skin pores, brick textures, foliage — the results can look remarkably better than the input, and for everyday photos that is often exactly what people want.

Here is the caveat that cannot be repeated often enough: the added detail is plausible, not recovered. Information that was never captured cannot be brought back; the model is making an educated guess about what was probably there. Usually a fine guess — but it can hallucinate. The infamous case is faces: an upscaler can turn a blurry mugshot into a sharp, confident face that is simply not the real person. This is why super-resolution must never be trusted for forensic identification, license-plate reading from CCTV, or any setting where the invented pixels could be mistaken for evidence. The "enhance!" trope from television is, in the strict sense, impossible.