Does AI upscaling add real detail?

What the model is actually doing

Ordinary enlargement averages neighbouring pixels. It cannot invent anything, so the result is a larger, softer version of what you had — never sharper.

An upscaling model does something different. It was trained on millions of pairs of images, each a large one and its shrunken twin, until it learned what a shrunken eyelash, brick wall or fabric weave tends to look like. Given a small image, it writes back a plausible large one.

The word doing the work there is plausible. The detail it adds is not recovered — the information was thrown away when the image was made small, and it is gone. What comes back is a confident guess that resembles the sort of thing that was probably there.

We measured it losing to plain resizing

We took a public-domain photograph from 1896, shrank it to 240 pixels, then enlarged it four times both with the model and with ordinary high-quality resampling.

The model lost. A checked fabric in the picture had its weave erased completely — the model read that fine, regular texture as noise and smoothed it away — and the face came out looking like wax. The plain enlargement was blurrier and more faithful.

And then measured it winning clearly

We saved the same photograph as a quality-35 JPEG first, which is roughly the state of a great many images that have been passed around the internet, and ran the comparison again. This time the model won by a wide margin: the blocky compression artefacts vanished and the edges came back.

The reason is that this class of model is built to repair damage, not to magnify. It removes what it reads as noise. Compression blocking is noise, so it goes and the picture improves. Film grain and fine fabric are also read as noise, so they go too and the picture gets worse.

That gives you a usable rule. If the problem with your image is that it has been compressed, screenshotted, or re-saved to death, the model will help. If the problem is only that it is small but otherwise clean, plain enlargement may be the more honest result.

The size wall is not arbitrary

Four times the width is sixteen times the pixels. A one-megapixel input becomes a sixteen-megapixel output, and every one of those pixels is produced by a neural network rather than copied.

That is why there is a limit on input size, and why we disable the button and explain it rather than letting you try. Freezing the browser tab for several minutes and then producing nothing is the worst possible outcome, and it is what happens if the limit is not enforced.

The 2× option is made by producing the 4× result and halving it, rather than by asking for 2× directly. Invented detail tidies up when it is shrunk, so the result is better. It takes the same time, because the expensive part happened either way.

The sharpest model is not the right model

We compared two candidates with acceptable licences. The transformer-based one is visibly sharper, and it took 9.7 seconds on a 128×128 image. The compact convolutional one took 16.5 seconds on a 512×512 image — sixteen times the pixels. Per pixel, that is roughly a sixtyfold difference.

A graphics card does not rescue the slow one. We measured WebGPU at about 3.4× faster than CPU here, not the twenty or fifty times people expect, because shader compilation and moving data to the card cost real time. Three times faster is worth having, but a model that is unusable on a CPU is generally still unusable on a GPU.

So the sharper model is the one nobody would wait for, and the one that ships is the one that finishes. This is the same judgement made for every model on this site.