← All guides

What enhancement actually recovers

Detail is rarely gone — it is buried under compression, sensor noise and resampling. What comes back, and what does not.

Last updated 4 September 2026

Most people meet image enhancement expecting an enlargement and are surprised by what actually changes. Enlarging is the smaller half of the job. The larger half is undoing what the image has already been through.

What happens to a photograph before you get it

Every file you own has been through something.

Compression. A codec keeps what it judges you will notice and discards the rest. It is reasonable about this at generous file sizes and increasingly wrong at small ones, which is why an image that looked fine on a phone screen falls apart when you open it at full size. What it leaves behind is characteristic: blocking in flat areas, a halo along high-contrast edges, and color that has been sampled at half the resolution of the brightness.

Sensor noise. A small sensor in low light produces a grainy signal, and the camera usually smooths it before you ever see the file. That smoothing cannot tell noise from real fine texture, so it takes fabric weave, hair and foliage with it. The photograph arrives already softened, and no amount of sharpening puts back what was averaged away.

Resampling. Every resize is a resampling, and every platform you upload to does one. A picture that has been through three of them is three generations from the original.

The part that matters

In each of those cases the detail was not removed cleanly. It was buried under something — an artifact, a smoothing pass, a resample — and what is buried can often be dug out. That is what these models do: they have seen enough pairs of damaged and undamaged images to recognize what a given kind of damage looks like and what tends to sit beneath it.

That is also why the result is better than sharpening. Sharpening increases local contrast, including the contrast of the artifacts. Recovery works the other way round: identify the damage, remove it, then resolve what it was covering.

How to judge a result

Not by a percentage. The useful question is binary, and it is about your own work:

• Can the print be made at the size the customer asked for?
• Does the listing clear the marketplace’s resolution requirement?
• Can the animal in the frame be identified?
• Is the text legible enough to read, given that it must not be relied on as a record?

Run your own worst file. A demonstration image chosen by us tells you how the model performs on an image chosen by us. Yours is the only test that decides anything, and it costs nothing to run — a workspace opens for your browser with no signup and no card.

Where to go next

If you are weighing whether it will work for a particular job, the honest boundaries are in what it will not do. If you already know it works and want it in your own pipeline, start with your first API call.