← All guides

What it will not do

Detail that was never captured does not come back, and output is not evidence. The cases where the answer is no.

Last updated 7 September 2026

Worth knowing before you spend anything. These are the cases where the answer is no, and saying so is cheaper for both of us than a refund.

Detail that was never captured does not come back

A face forty pixels across, a license plate at the far end of a car park, a sign photographed from too far away. What is recovered is what the file still holds, and below a certain point the file holds nothing about that subject at all.

The distinction that matters is between buried and absent. A compressed photograph of a face has a face in it, under the artifacts. A forty-pixel face does not have the features in it in any form. A model handed the second will still produce something — plausible, sharp, and not the person who was standing there.

Output is not evidence

This follows directly from the above and deserves its own heading, because it is the failure with real consequences.

Text and numbers that have degraded past legibility can be reconstructed into something that reads cleanly and is wrong. A serial number, a date on a document, a plate, a face in a security frame: if the original could not establish the fact, the enhanced version cannot either, no matter how convincing it looks. Do not use a result to support a claim that the source image could not already support.

Keep the original alongside anything you publish or file. If a result is going anywhere that matters, the original is what settles a question about it.

It recovers and repairs; it does not compose

Nothing here removes an object, replaces a background, relights a scene or changes a composition. One file goes in and the same picture comes out. If you need an edit rather than a recovery, this is the wrong tool and will politely do nothing useful.

Restore is the exception worth knowing about, and it is narrower than it sounds: it repaints damage, and asked to, it will bring color to a photograph that never had any. It is still working from your picture rather than making a new one.

There is a ceiling on enlargement

Every model publishes the factor it works to and the largest input it accepts, and those figures are in the API reference rather than in this page for a reason: they are the authoritative ones and they can change. Past that ceiling, feeding a result back in for a second pass compounds whatever the first pass got wrong rather than doubling the gain. One pass at the right setting beats two at the wrong one.

Enhance sharpens; Restore repairs

Physical damage to a print — a tear, a crease, a stain, heavy fading — is a different problem from compression and noise, and the enhancement models are not trained to reconstruct across it. A scan of a torn photograph, enhanced, comes back sharper and still torn. Send that file to Restore instead, which is built for exactly this and repaints the damage away.

What Restore asks in return is worth knowing before you send a portrait. It repairs by repainting the frame rather than by lifting the original pixels, so a face comes back as a close likeness rather than the one that was there. For a photograph of a person that matters, keep the scan.

What to do with a no

If your file falls into one of these, the fastest path is still to run it: a workspace opens with no signup and no card, and you will know in under a minute rather than after an argument with a support queue.