Restore an old family photo when only one copy is left

Repairing a torn photo? First, keep the scan untouched

Most people come to this with one print: a portrait from a drawer, a photo on a gravestone card, the only picture of a great-grandfather that survived a house move. Before any restoration, the most valuable thing you will make is not the restored version. It is a clean scan of the damaged one, saved and never edited again.

The reason is simple. Every restoration is an interpretation. If you overwrite the scan with the first result you like, and later find a better tool, a sharper reference photo or a relative who remembers the colour of the dress, you have nothing left to start from. The print itself will keep fading, and a torn photo tends to lose fragments every time it is handled.

What to do, in order:

1. Scan the print flat on a flatbed scanner at 600 ppi for anything the size of a postcard or larger, 1200 ppi for small prints and passport-sized portraits. If a scanner is not an option, photograph it in daylight from directly above, with the phone parallel to the table to avoid keystone distortion.
2. Save the scan as PNG or TIFF, not as a heavily compressed JPEG. Compression artefacts look like texture to a restoration model, and it will happily sharpen them.
3. Do not crop, straighten or adjust levels on this file. Name it something like grandmother-1952-original.png and keep it in two places.
4. Work only on copies. Every experiment starts from the original, never from a previous result.

If the photo is torn into pieces, place them on the scanner glass together, as close as they fit, and scan them in one pass. Separate scans of each fragment are much harder to line up later. If you also have slides or negatives of the same family, our guide on how to scan slides and negatives at home covers the settings for film.

Fix a damaged photo: what comes back and what is invented

When you fix a damaged photo with AI, the work splits into two very different kinds of result, and it helps to know which is which before you judge the output.

What restoration genuinely recovers is information that is still in the picture, just buried:

• Facial features that are soft but present: the shape of the eyes, the line of the mouth, the nose. A face restoration model is trained on faces and rebuilds detail around those landmarks.
• Sharpness and resolution. The background and clothes are upscaled with a separate model, so a small print can be enlarged without turning to mush.
• Grain and fine noise from the original film and the paper texture.
• Light stains, dust and faint marks, especially small ones on even areas such as sky, walls or skin.
• Faded or yellowed colour, and, if you ask for it, colour on a black-and-white print.

What it cannot recover is information that was never there or has been physically destroyed: a missing corner, a burn, a hole, a crease that has torn through the emulsion, or a face that was out of focus when the shutter fired in 1940. In those areas the model does not restore anything. It fills the gap with something plausible, based on what faces and fabrics usually look like. Sometimes the guess is good. It is still a guess, and it is not your relative.

Two practical consequences follow. First, for a large missing patch in the background, it is often better to use a tool built to remove an object from a photo on that area before restoring, so the fill blends with the surroundings instead of being reinvented at face level. Second, for anything that touches the face (a crease across an eye, a stain on the mouth), keep your expectations modest and check the result closely, as described in the next section.

In our tests on 19 September 2026, an old photo of 1000 × 1289 px was restored in 9 seconds, in three stages: cleaning, faces and colour. Seeing those stages separately is useful: if the cleaned version already looks right and the problem only appears at the face stage, you know exactly which setting to change.

The honest limit when you restore an old portrait

Here is where the usual advice goes wrong. Most guides say to restore an old portrait at full strength for the sharpest result. With a heavily degraded original, full strength is exactly what produces a face that is sharp, clean and wrong.

The reason is how face restoration works. When the input is clear, the model sharpens what it sees. When the input is a blur of a few dozen pixels, a screen capture of a print, or a face covered in scratches, the model has very little to go on, so it leans on what it has learnt about faces in general. The result drifts towards an average face: symmetrical, smooth skin, generic eyes. It looks like a good photo. It does not look like her.

How to spot it:

1. Open the result and view it at 100 %, not fitted to the screen. Reduced views hide the problem.
2. Look at the eyes first. Check the shape of the eyelids, the distance between the eyes and whether the gaze still points the same way as in the original. A changed eye shape is the clearest sign of an invented face.
3. Then look at the mouth: the width of the smile, whether teeth that were visible are still visible, and the corners of the lips.
4. Compare side by side with the original. Ask someone who knew the person. They will notice in a second what you might talk yourself out of.

How to correct it: lower the fidelity slider. Fidelity controls how much of the reconstructed face is blended over the original. At 100 % you get the model's version entirely. Around 40-60 % you keep the original's proportions and expression with clearer detail on top. It will look slightly softer, and it will look much more like the person. If even 40 % drifts, the face in the scan simply does not carry enough information on its own, and the next section is the fix.

One more limit worth stating plainly: running the restoration again on its own output does not increase likeness. Each pass adds more of the model's idea of a face, not more of your relative. If a photo arrived through a messaging app, our article on what can be recovered from a pixelated WhatsApp photo explains why those are the hardest case.

Reference photos: how to keep the likeness

If you have other photos of the same person, the restoration can use them. With one to three reference photos, the face is rebuilt guided by how that person actually looked, instead of by faces in general, and the resemblance improves markedly on badly damaged prints. One good reference is often enough; up to three helps when the original is very poor.

References that work:

• The face clearly visible and reasonably large in the frame, so the features are not just a few pixels.
• Facing the camera, or close to it. A three-quarter view helps less, a profile hardly at all.
• Even light on the face, without hard shadows across the eyes or a hat brim hiding the forehead.
• Sharp: a later, better photo is ideal, even if it was taken years apart, as long as the person is recognisable.
• Close in age to the portrait you are restoring. A wedding photo is a better reference for a portrait at 25 than a photo taken at 80.

References that hurt:

• Group photos where the face is tiny. Crop to the person before uploading.
• Sunglasses, heavy shadow, strong backlight or motion blur.
• Photos with a very different expression if the expression matters to you. Reference photos tend to be serious; if the original is smiling, check that the smile survives in the result.
• Photos of a sibling or parent who looks similar. The model will happily blend two people.

One honest caveat: references guide the face, not the rest of the picture. They do not restore the background, the clothes or the hands, and they do not supply what a torn corner lost. And the result is still a reconstruction. Present it to the family as a restored version, and keep the original scan next to it.

For a broader look at the techniques behind this, see our guide on how to restore old photos with AI.

Step by step: restore an old family photo in Airpic

This is the exact process, with the settings to choose and what you should see at each point. Processing and viewing the result is free, with no account; you only decide whether to download the full file at the end.

1. Prepare a copy of the scan. JPG, PNG (including 16-bit), TIFF, HEIC, WebP and AVIF are accepted, up to 20 megapixels. A 600 ppi scan of a postcard-sized print fits comfortably. EXIF orientation is applied and the colour profile is kept.
2. Open the Old Photo Restorer and drop the file in. You will see a thumbnail with the size in pixels. If the scan is slightly crooked, tick the straightening option: in our tests on 19 September 2026, a scan tilted by 2.5 degrees was straightened automatically, with a correction of 2.48 degrees.
3. Mark any serious damage. If there is a crease or scratch you want repaired, tick the repair option and paint over it in the mask editor. Paint only the damage, not the face around it.
4. Decide on faces and colour. Leave face restoration on for portraits. Turn colour on only if you want a colourised version; if the print is already in colour and simply faded, the cleaning stage handles that. For colour work on its own, the separate tool to colourise a black-and-white photo gives you an intensity slider.
5. Set fidelity. Start at 60 % without references. If you add references, add one to three photos of the same person in the reference block; the slider then becomes fidelity to those references, and 85 % is a sensible starting point.
6. Process and inspect. Use the comparison slider, then view at 100 %. Click through the stage chips (cleaning, faces, colour) to see where any problem appears. Check the eyes and mouth as described above.
7. Adjust once, deliberately. If the face looks generic, lower fidelity by 15-20 points and run again from the same original. If it looks too soft, raise it in smaller steps. Do not stack passes.
8. Download and file it next to the original. Keep both, clearly named. Files are deleted one hour after download, one day at most, and are never used for training; processing runs on GPUs in the European Union.

If the restored portrait is going to be printed large for a frame or a memorial, run the result through the tool to upscale an image for print afterwards rather than restoring at a bigger scale. When you are ready, restore old photos with your own scan and compare it with the original at 100 % before you share it with the family.

FAQ

Keep reading