AI photo restoration for old, damaged pictures

It is repairing a damaged photograph with a model instead of by hand. You upload a scan, the model removes the physical damage — creases, tears, stains, dust — and corrects fading, then returns a clean version. What used to be a slow job with a healing brush becomes one upload.

You get free credits on signup with no card required, and a restoration costs a few credits, so the free balance covers several photos. After that it is pay-as-you-go; one-time credit packs do not expire.

About fifty seconds on our own demo run, which is the example shown above. Larger uploads take longer. It is fast enough to iterate — most people run a photo twice, once as-is and once cropped tighter, and keep the better one.

That is the failure we prompt hardest against, and the prompt is published above so you can check. Identity drift happens when the source is too damaged in that area for anything to be recovered — the model then has nothing to work from and fills in. Crop closer to the face and re-run if you see it.

It can add colour, but understand what that is: an interpretation, not a recovery. The information was never in the negative. For family history work, many people keep the restored monochrome version as the record and treat a colourised one as a separate creative output.

Only softness, not true blur. If the photo was out of focus or the subject moved when the shutter opened, the detail does not exist anywhere in the file — sharpening it produces invented detail that can look like a different person. Scan resolution problems are fixable; camera-shake is not.

A flatbed scan at 300–600 dpi beats a phone snap of a print every time, because a phone adds glare, keystone distortion and its own noise on top of the damage you are trying to remove. If a phone is all you have, shoot flat-on in even indirect light with no flash.

Generations are tied to your account so you can find them again in your history, and paid plans keep generation private. If you are working with sensitive family material, download what you need and delete the generation when you are done.

Nano Banana 2 by default, the same image model the rest of the site runs on, driven by the published prompt. GPT Image 2 is also available if you want a second opinion on a difficult photo — they fail in different ways, and on badly damaged sources it is worth trying both.

A human retoucher still wins on the hardest cases: a photo torn in half, a face largely missing, or an heirloom where every detail must be defensible. They cost per photo and take days. AI photo restoration costs a few credits and takes under a minute, and on the common case — creases, stains, fading, dust — the gap has closed enough that most people never need the retoucher. The sane order is to run it here first, and only pay a human for the handful it cannot save.

One photo per run here, because each one needs its own look at the damage. If you are working through a box of family photos, the practical rhythm is to scan a batch first, then run them one at a time and keep the tab open — a restoration finishes in under a minute, so a stack of twenty is an evening, not a project.

Often better than people expect, because the damage on a tintype or an albumen print — foxing, silvering, heavy contrast loss — is exactly the kind of surface damage AI photo restoration handles well. The limit is the same as anywhere else: if a face has faded past the point of being visible, nothing can recover it, and what comes back is an invention. Judge the result against any other surviving photo of the same person before you treat it as the record.