Upscale Images

Resampling, not invention

Enlarging an image is a guess about pixels that were never recorded, and there are two honest schools of guessing. Resampling works out each new pixel from the real ones around it using a mathematical filter, and the one used here weighs a wide neighbourhood with a windowed sinc function — which is why it keeps edges and fine lines far better than the simple bilinear enlargement a browser does by default. An unsharp mask tuned for enlargement then restores the local contrast that any resampling softens. Nothing is invented at any stage: every feature in the result corresponds to something in your original.

The other school is generative upscaling, which recognises what a region probably depicts and draws a plausible high-resolution version of it. It produces dramatically sharper results and it is making things up — the classic demonstration being a face that comes back sharp, detailed, and not the same person. For a product photograph, a poster or anything decorative, that is often exactly what you want. For evidence, documents, identification or anything a decision rests on, it is dangerous in a way that resampling is not, and the distinction is worth holding on to.

What upscales well follows directly from having no invention available. Logos, diagrams, screenshots and clean photographs have real structure to interpolate between and come out convincingly larger. A small, heavily compressed JPG does not: its artefacts enlarge along with everything else, so the blocking and ringing become four times the size and considerably more obvious. Starting from the largest original you can find beats any amount of enlargement.

Output is PNG, because writing a lossy file immediately after carefully reconstructing edges would undo part of the work. That makes the result large — a 4× enlargement holds sixteen times the pixels — so compress it afterwards if it is going onto a web page. The ceiling is memory rather than a fixed rule: roughly a 6000×6000 output on a typical laptop, and less on a phone.

When you want something else

For photographs where invented detail is acceptable, the generative upscalers are genuinely in a different class and it would be dishonest to pretend otherwise. Real-ESRGAN and waifu2x run locally, are free, and produce results this cannot approach on faces, fur, foliage and anime artwork. What they produce is a plausible photograph, not your photograph at a higher resolution.

For scanned text, upscaling is usually the wrong move altogether — recognising the words and re-setting them gives you something infinitely sharp rather than a larger blur. And for anything scripted or in bulk, ImageMagick -filter Lanczos -resize 400% and libvips do the same resampling from the command line with no memory ceiling worth worrying about.