AI Photo Restoration

100% Free

Upload the old photo

…or press Ctrl+V to paste a scan

Drag & drop a scanned or faded photograph

JPG, PNG, WebP — up to 30MB

Faces are rebuilt on your own device — the photo never leaves it.

What to restore

Rebuild faces GFPGAN reconstructs each detected face at 512px. This is the slow part — about 20 seconds per face on your device, plus a one-time 330MB model download. Correct fading and colour cast Measures exposure, contrast, cast and noise, then corrects exactly what it measured. Instant, and it tells you what it changed.

— downloaded once, then cached by your browser.

What was done

Before / after

100% Free

All tools are completely free to use.

Privacy First

Your files are never uploaded or stored on our servers.

Super Fast

Compress and process files in seconds.

High Quality

Best results with minimal quality loss.

What is AI Photo Restoration?

AI Photo Restoration does two things to an old photograph: it rebuilds the faces, and it corrects what age did to the whole image. Both run on your own device. A scan that came out soft, grainy or washed out goes in; a photograph with recognisable, detailed faces and sane contrast comes out.

Faces are found, aligned, rebuilt and blended back

Face rebuilding is done by GFPGAN, a network trained on millions of portraits, and the quality depends almost entirely on a step most people never hear about. The model only works on faces positioned exactly the way it was trained to expect them, so every face in your photograph is first located along with five landmarks — eyes, nose and mouth corners — and mathematically rotated and scaled onto that standard position. It is rebuilt there, then warped back into your photograph through a soft-edged mask so no rectangle appears around anyone's head. Up to six faces are handled per run, at roughly twenty seconds each.

Fading is measured, not guessed

The second half is ordinary, explainable image processing driven by real measurements of your scan: exposure from the luminance histogram, contrast from the percentile spread, colour cast from the channel means, noise from a standard estimator. Whatever it measures as wrong is corrected in proportion, and the page lists every step it took with the numbers behind it. A photograph that measures fine is left alone.

What it will not do, stated plainly

It does not repair scratches, tears, stains or missing corners — that is physical damage in parts of the image the face model never looks at, and the AI Object Remover erases those properly. It does not add colour; that is AI Colorize Photo. And a rebuilt face is a plausible reconstruction rather than recovered truth: on a badly degraded photograph the model fills in detail that was never in the scan, and the result can drift from the person you remember. Compare before and after carefully on photographs that matter.

Nothing is uploaded. Tools on this page that use a neural network download the model to your browser once, cache it, and run it on your own machine from then on. That is why there is no account, no queue, no watermark and no daily limit — there is no server doing the work, so there is nothing for us to meter and no copy of your photo anywhere but your device.

How to Use AI Photo Restoration

1

Upload the old photo

Drag in a scan, paste with Ctrl+V or choose a file. It never leaves your device.
2

Choose what to restore

Rebuild faces, correct fading and colour cast, or both. The measured problems with your scan are listed above the buttons.
3

Press Restore

The first run downloads about 330MB of models with a real progress bar, then each face takes roughly twenty seconds of on-device compute.
4

Read what was done

The page lists every correction with its numbers, and says how many faces it rebuilt — or that it found none.
5

Compare, then continue

Check the before/after slider and download, or hand the photo on to the colouriser, the object remover or the upscaler.

Frequently Asked Questions

On a soft, grainy or low-resolution scan the difference is dramatic — eyes, lips and hair come back with real detail because the model rebuilds them from what it learned across millions of portraits. On a photo that is already sharp there is little to gain. The honest caveat is that this is reconstruction, not recovery: the detail is plausible rather than the detail your camera captured.

No, and no restoration model that only looks at faces can. Scratches, tears and stains are damage across the whole image; the AI Object Remover erases them properly by regenerating the background behind them. One click hands the photo between the two tools with no upload.

That is the size of the face detector plus GFPGAN, and downloading them is what keeps your family photographs on your own machine. It happens once; your browser caches both models for every future visit.

Yes — every face is detected, aligned and rebuilt individually, up to six per run, at roughly twenty seconds each. If more are found the page tells you how many it skipped rather than silently ignoring them.

GFPGAN was trained only on faces at a fixed position and scale, so feeding it a raw crop produces a smeared, uncanny result. Each face here is rotated and scaled onto that standard position using five detected landmarks, rebuilt, then warped back. That step is most of the quality difference between this and a naive implementation.

The whole-photo corrections still run and the page says plainly that no face was found. Very small, heavily obscured or strongly profile faces are the usual causes; cropping closer to the person often helps.

No. Both models are downloaded to your browser and every step runs on your own processor. For inherited family photographs, that is the difference between using a tool and handing your archive to a company.

Remove scratches first, then restore faces, then colourise. Each tool hands the working photo to the next in one click, so the order is easy to follow and nothing is re-uploaded at any point.

Output format

Quality

What this restores — and what it does not

✓ Faces: blur, grain, JPEG damage and low scan resolution are genuinely rebuilt by a network trained on millions of portraits.
✓ Fading, colour casts and flat contrast across the whole photo, from real measurements.
i Not scratches, tears or stains — those are physical damage in other parts of the image. The object remover erases them properly.
i A rebuilt face is a plausible reconstruction, not recovered truth. On a very degraded photo it may drift from the person you remember.