Sharpen an Image with AI

How it works

The image is handed to Real-ESRGAN, a reconstruction model trained on pairs of degraded and clean photographs. It enlarges the picture four times, predicting what the missing detail between your pixels would have looked like. The result is then scaled back down to the size you started with, and given a light sharpening pass.

The round trip is the point. You get the detail the model reconstructed, at the dimensions you already had, rather than a larger file you did not ask for.

Scaling back down is also what keeps it usable. Averaging four predicted pixels into one cancels the predictions that disagree with each other and keeps the ones the model was confident about, so the output is noticeably cleaner than the raw four times enlargement it came from.

This does add resolution — which a sharpening slider cannot

An ordinary sharpen increases acutance: how abruptly tone changes across an edge that is already recorded. It cannot separate two fine lines that the camera recorded as one gray smudge, because the information that they were ever separate is not in the file.

A reconstruction model can, and does. It has seen enough photographs to have an opinion about what that smudge most likely was, and it writes that opinion into the output as new pixels.

That is a genuine gain and a genuine risk in the same operation. The detail is real in the sense that it is plausible and usually right. It is not real in the sense that your camera recorded it.

What the model invents

Run this on a portrait and look at the areas of fine repeating texture first. Foliage is the clearest example: a soft dark mass of leaves comes back as crisp individual leaves with bright edges, and those leaves were decided by the model, not photographed. The same happens with fabric weave, brickwork, gravel and distant hair.

Text is the one to check hardest. The model has seen a great deal of text and will produce confident letterforms that are not the characters in your source. Small type on packaging, a sign in the background, a name badge — compare every one against the original before you rely on it.

Faces respond well when they are large in the frame and poorly when they are small. A face occupying a few dozen pixels does not constrain the result, so what comes back is a plausible face rather than that person's face.

Reading the output

Compare at the same displayed size. Putting an enlarged source next to its reconstruction at one hundred percent compares two different things and always flatters the reconstruction.

The effect here is strong by design, because a soft source is the case this exists for. On an image that was already sharp it will look overdone, and the tell is texture that reads as etched rather than photographed — skin that has acquired more pores than the light could have shown, or leaves that are sharper than the branch holding them.

When to use something else

If you only need different dimensions and the source already holds the detail, use the ordinary resizer. Every pixel in its output comes from a pixel you supplied, where this one invents detail that was not in the file.

If the photograph is genuinely out of focus or blurred by movement, that is a different repair with a different tool. And if the file is a heavily compressed download, fixing the compression damage first will get you further than reconstructing on top of it.