Free tool

Photo restorer, for old and blurry faces.

Restore an old family photo, a soft scan or a small blurry portrait. The page finds every face, rebuilds it with an open AI model, sharpens the rest of the picture, and can add colour to a black and white photo or paint out scratches you point at. It runs in this tab, so your photo is never uploaded. No sign-up, no watermark.

This runs in your browser. Your photos are never uploaded.

How it works

Four parts, each one optional.

01 / Find

Faces are found

Each face gets a number on the photo. Click a number to leave that face out.

02 / Rebuild

An AI model rebuilds them

Every face is turned upright, rebuilt, and blended back into the photo. Drag the line to compare.

03 / Finish

Sharpen, colour, save

Clean the background, add colour to a black and white photo, then download the full size picture.

Guide

How to restore an old photo, and what the AI can and cannot fix.

To restore an old photo online, you need something that repairs three different problems: soft or blurry faces, a dull or noisy picture, and physical damage such as scratches and tears. This photo restorer handles the first two on its own and the third with a brush you control. It runs on your computer inside the browser tab, so a family photo never goes to a server.

How to use the photo restorer

  1. Drop your photo on the box, paste it, or click to choose. A scan or a phone picture of a print both work. Crop away the white border first if there is one.
  2. Wait a moment while the page looks for faces. The first visit downloads the AI engine, about 27 MB, and your browser keeps it. Each face found gets a number on the photo. Click a number to leave that face out, for example a spot that is not really a face.
  3. Check the settings on the right. Face strength, the clean-up of the rest, the size and colour are all explained below.
  4. If the photo has scratches, press Paint over scratches, set the Brush size and paint over them. This is optional.
  5. Press Restore the photo. The first time, the face model downloads, 74 MB, and is kept in your browser. The page then rebuilds each face one at a time and shows how long it took.
  6. Drag the line across the picture to compare Before and After. Below it, each face is shown before and rebuilt, side by side, so you can look closely.
  7. Move Face strength to taste, then press Download the photo. The file name ends in -restored.

What each step does

The face step works on each face separately. The page finds the eyes, nose and mouth corners, turns and scales the face to a fixed upright size, and gives it to RestoreFormer++, an open face restoration model. It returns a clean face, which the page matches to the colour cast of your photo and blends back through a soft oval, so there is no hard edge. Because the model rebuilds the face from what it has learned about faces, it can add detail that was never in the photo.

The clean-up step uses Real-ESRGAN, a small open model that removes noise and sharpens the whole picture. With Size set to 2x or 4x it also enlarges the picture, and the rebuilt faces are placed on top at that larger size so they stay crisp. Large photos are limited to about 16.7 million pixels in total.

The colour step uses DDColor, the same model as the colorize photo page. It guesses natural colours for a black and white photo. The guess is not a record of the real colours.

Face strength

SettingWhat you getGood for
100%The rebuilt face as the model made itVery soft or tiny faces, where there is little to keep
60% to 85%The rebuilt face mixed with the originalFaces that look too smooth or not quite like the person
0%Your original facesChecking what the model changed

Honest limits

  • Faces are rebuilt, not recovered. A blurry face holds less information than the model puts back. The result is a likely face. It can look smoothed, a little too young or clean, or not quite like the person. Compare with the original and lower Face strength if it does.
  • Scratches, tears and stains are not found automatically. The face model also copies marks that lie across a face. Paint over them first, using the scratch brush, so they are filled in before the face is rebuilt.
  • Small faces can be missed. Faces smaller than about 20 pixels across can be missed, and even when one is found, a face that small gains little. Crop closer or use a larger scan.
  • Only faces are rebuilt. Hands, writing, clothes and background are sharpened and cleaned, not rebuilt. Tiny text can come out as made-up shapes.
  • Turned or covered faces, such as a profile, sunglasses or a hat pulled low, give weaker results or are not found.

Tips for a better restoration

  • Scan prints at 600 dpi or more, or photograph them flat in even light with no glare. A better starting file always gives a better result.
  • Fix a very dark or washed out picture before restoring it, for example with the AI photo enhancer.
  • For a heavy blur with no faces, use unblur image. To enlarge only, use the AI image upscaler.
  • To lift a person out of the background afterwards, try remove background. To take a stain or a stranger out, use remove objects from a photo, which uses the same scratch model.
  • Keep the original scan. Save the restored photo as a new file and note that it was restored, so family history is not confused by a face the AI helped to build.

What the first visit downloads

PartSizeWhen
AI engine (ONNX Runtime)about 27 MBAlways, first visit
Face finder0.2 MBAlways
Face model (RestoreFormer++)74 MBWhen you press Restore the photo with a face found
Sharpening model (Real-ESRGAN)5 MBWhen Sharpen and clean the rest is on
Colour model (DDColor)115 MBOnly if you turn colour on
Scratch model (LaMa)208 MBOnly if you paint over scratches

Each file is checked against its published checksum before it is used, then kept in your browser so it downloads once. Nothing about your photo is sent anywhere.

Questions

Good to know.

Is my photo uploaded?

No. The AI models are downloaded to your browser and run on your computer. The photo never leaves it. Only the model files come from the internet.

Which AI models does it use?

YuNet finds the faces (MIT licence). RestoreFormer++ rebuilds them (Apache 2.0). Real-ESRGAN sharpens the rest (BSD 3-Clause). DDColor adds colour and LaMa fills painted marks, both Apache 2.0. The licence notes sit with the files on this site.

Why does a restored face sometimes not look like the person?

The model does not know the person. It rebuilds a face that fits the eyes, nose and mouth it can still see, using what it learned from many faces. With a very blurry source, several faces fit, and it picks a likely one. Lower Face strength to bring back some of the original, or leave that face out.

Can it remove scratches and tears?

Only the ones you paint over. Press Paint over scratches, cover the marks and press Restore the photo. A model fills the painted area from what is around it. Thin scratches work best. A large tear across a face is hard, and the fill can look wrong.

Can it colorize black and white photos?

Yes. Turn on Add colour to a black and white photo. The colours are a guess, so a dress or a car may not match the real one. Colour strength lets you tone it down.

Why was a face not found?

Very small faces, faces turned away, and faces covered by sunglasses, a hat or a hand can be missed. Crop closer to the face or enlarge the photo first, then try again.

Why is it slow on my computer?

The page uses your graphics card through WebGPU when the browser supports it. When it does not, the models run on the processor, which is much slower, especially for faces and for the sharpening step. A recent Chrome or Edge on a computer with a graphics card gives the faster path. You can press Stop at any time.

Does it change the size of my photo?

Not unless you choose 2x or 4x. Photos above about 16.7 million pixels, or 4096 pixels on the long side, are scaled down before they are worked on, and the page says so.

Is it free?

Yes. There is no sign-up, no watermark and no limit on photos. The cost is the one-time model download.

Is it a good way to restore a photo of a relative?

It can make a faded or blurry portrait much easier to look at. Treat the result as a careful reconstruction, not as proof of what the person looked like. Keep the original scan next to it.

Listed on

Listed on JustMacAppsMotionShine on Mac App IndexMotionShine on ScrollLaunchMotionShine featured on StartupBaseFeatured on FazierFeatured on Twelve ToolsMotionShine - Featured on Smol Launch