Enhance a Photo with AI

How it works

Three steps on a GPU. The photograph goes through Real-ESRGAN, which reconstructs detail while enlarging four times; it is scaled back to its original dimensions; then contrast is lifted slightly and a very light sharpen finishes it.

The contrast step is what separates this from a pure detail pass. A tired photograph is usually two problems at once — it is soft, and it is flat — and recovering the texture without addressing the flatness leaves it looking sharper but no better.

The lift is deliberately small. A larger one produces the punchy look that photographs acquire when someone has enhanced them, which is a different thing from a photograph that looks like it was taken on a better day.

What "enhanced" honestly covers

Detail and tonal separation. That is the whole list.

It does not correct exposure, so a dark photograph comes back dark with more contrast, which can make it read darker still. It does not fix a color cast, so an indoor shot under tungsten light comes back orange. It does not recover clipped highlights, straighten anything, or change the crop.

Worth saying plainly, because "enhance" invites people to expect a general repair and then to conclude the tool is broken when their underexposed photo comes back underexposed with deeper shadows.

The recovered detail is a prediction

The model is not revealing detail hidden in your file. It is predicting what detail probably belonged there, based on the photographs it was trained on, and writing that prediction into the output.

Most of the time the prediction is good and the result is a better photograph. The places it goes wrong are consistent enough to check: repeating fine texture gets tidier and crisper than reality — leaves, weave, gravel, distant hair — and small text gains confident letterforms that may not be the right characters.

On a portrait, watch the shadow under the eyes. The contrast lift deepens it, and combined with recovered skin texture that can age a face slightly. If that happens, the source was flatter than it looked and this is not the right treatment for it.

The photographs this suits

Scans, which are uniformly and mildly soft in exactly the way the model handles best, and usually flat as well.

Photographs taken through glass, in haze, or in humid conditions — a thin veil over an otherwise well recorded image.

Older digital photos from small sensors, where there is real detail present but not much of it, and the contrast has never been touched.

When it will not help

An out of focus subject, camera shake, a very dark frame, or a subject occupying a small part of the picture. In those cases the model has too little to work from and the output is mostly invention.

A photograph that is technically fine and simply uninteresting is also outside this. No amount of recovered texture fixes composition, and running it anyway tends to produce something that looks processed rather than better.