Sharpen a Soft Photo

How it works

The image goes through three stages, none of which involve guessing what a blurred shape used to be.

First a super-resolution model, RealESRGAN, redraws the image at four times its width and height. Models like this are trained on pairs of images — a high-resolution original and a degraded copy — and learn to invert that specific degradation. They are good at reconstructing texture and edges that are implied by what survived.

Second, an unsharp pass raises contrast in a small radius around edges. This is the step that makes an image read as sharp. It adds no information; it makes the information already there more visible.

Third, the result is scaled back to twice the original size rather than all the way down, because averaging pixels during a downscale is itself a mild sharpener and stopping halfway keeps some of the model's added detail.

The distinction this page keeps making

There are two very different situations that both get called blur.

In the first, the detail is present but low in contrast — a slightly missed focus, a soft lens, an image that has been resized down and back up. The information survived. Raising local contrast makes it visible again, and the result is a genuine improvement.

In the second, the detail was destroyed at capture. A subject that moved during the exposure, or a focal plane that landed somewhere else entirely, does not have soft detail; it has no detail. Every pixel is an average of things that were never separated. There is nothing for a contrast pass to bring back, because contrast operates on differences and the differences are gone.

How to tell which you have

Open the original at one hundred per cent and look for the smallest thing you can identify. Eyelashes, the weave in fabric, text on a label, individual leaves.

If you can just about make it out, you have the first case and this pass will help. If it has become a smooth gradient with no internal structure, you have the second, and no amount of sharpening produces detail — it produces a harder edge around the blur, which is the characteristic over-sharpened look.

Reading the result

Judge at full size, never zoomed out. A downscaled preview hides both the improvement and the damage.

Watch three places. Skin, where over-sharpening shows as texture that was not there. Text, where it shows as bright fringes on either side of each stroke. And any large flat area — sky, a painted wall — where amplified sensor noise appears as grain. If those three look right, the pass has done what it can.

Publication gate

This page ships once the workflow has been run against a mildly soft image, a genuinely out-of-focus image, and a motion-blurred image, with all three outputs compared to their sources at full pixel size.