Yourtale
Start a free session →

How to restore old damaged family photos with AI

What AI photo restoration can and cannot recover, a step-by-step workflow, honest limits on faces and color, real costs, and how to get a restored photo print-ready for a memoir.

By The Yourtale team · Published 14 September 2026 · 16 min read

AI photo restoration removes scratches, tears, stains, and fading, sharpens soft detail, and adds plausible color to a black and white print. It cannot recover information the photo no longer holds. Where damage covers a face, the software predicts what was probably there, and a prediction about a face is a guess wearing the clothes of a memory.

That single distinction decides how you should use these tools. On a torn corner, a water stain, or a sky full of dust specks, AI restoration is close to magic and costs nothing. On the one surviving photo of a grandmother's face, it is the tool most likely to hand you back a stranger who looks a bit like her. This guide covers the workflow, the settings, the money, and the places where the honest answer is to stop.

We do not sell photo restoration as a service. We make hardcover memoirs drafted from recorded interviews, and family photos go into those books, so we spend a lot of time looking at scans of damaged prints and deciding which ones can carry a full page. Most of what follows comes from that.

Key takeaways

  • Restoration is prediction, not recovery. AI fills missing areas with what its training data says probably belongs there. On fabric, walls, skies, and scratches that is almost always right. On faces it is a guess, and the more damaged the face, the bigger the guess.
  • Scan before you restore, and scan properly. The Library of Congress recommends 400 to 600 PPI for paper prints and around 3,000 PPI for 35mm slides and negatives (Library of Congress preservation guidelines). No restoration tool recovers detail a low-resolution scan threw away.
  • Free covers most of it. Google Photos gave Magic Eraser, Photo Unblur, and Portrait Light to every user at no charge in May 2024 (Google), and MyHeritage repairs several photos free before asking for a subscription (MyHeritage).
  • Colorization is the least trustworthy output in the box. The model infers color from context, so a dress it renders blue may have been red, and nobody in the photo is around to correct it.
  • Keep the unretouched scan forever. The National Archives is blunt about this: "Do not throw away your original film and prints after you digitize them. Digitized images are not considered a replacement for originals" (National Archives).
  • A restored photo is a conversation starter before it is a picture. The person who was in the room can tell you what color the dress actually was, who is cropped out of the left edge, and why everyone is standing so far apart. Ask while you still can.

What AI photo restoration actually does to a photo

It is worth knowing the mechanism, because the mechanism explains every limit further down this page.

A restoration model has been trained on millions of pairs: a clean photograph and a deliberately degraded copy of it. From those pairs it learns what degradation looks like and what usually sits underneath it. When you hand it your grandfather's cracked portrait, it is not reading the crack and reconstructing what the emulsion held before the crack. It is looking at everything around the crack and generating the most statistically likely content to put there.

For most of a photograph this works beautifully, because most of a photograph is predictable. A crack across a brick wall has an obvious answer. A scratch through a shirt sleeve has an obvious answer. Fading is a global shift the model can reverse with real confidence. Grain and scanner noise are exactly what these models were built to remove.

Faces are the exception, and they are the exception for a structural reason. A face carries the identity of a specific person, and the model has no access to that person. It has access to faces in general. So it produces a face in general: symmetrical, smooth, well lit, with plausible hair and plausible eyes.

PetaPixel put this in front of readers in March 2026 when ON1 previewed its Restore AI module. Testing it on old family photographs, the site reported that "the AI appears to be trained on what it thinks an attractive human face looks like, and so it will add hair, apply makeup, and enhance facial features to achieve its final result" (PetaPixel, 25 March 2026). ON1 acknowledged the results were early, and noted that where a face carries limited data, "sometimes it is better to just turn it off."

That last sentence is the most useful piece of advice in the entire category, and it comes from a company selling the feature.

Can AI really fix a torn photo?

Yes, and this is the case where it performs best. A tear is a region of missing information surrounded by intact information, which is precisely the problem inpainting models were designed to solve. Tears through background, clothing, furniture, and landscape are usually repaired so cleanly that you cannot find the seam at full zoom.

MyHeritage's Photo Repair advertises exactly this list: scratches, tears, holes, stains, and other physical damage. It ships two models, Gentle Repair as the default and Extensive Repair for heavier damage, and it saves the repair as a separate copy rather than overwriting what you uploaded (MyHeritage).

The results split by what the tear crosses:

There is a simple test for whether an automatic repair is safe: if the damaged area contains nothing that identifies a specific person, let the software handle it.

Scan it properly first, because restoration cannot fix a bad scan

The most common reason a restoration looks soft and plasticky is that the input was a phone snapshot of a print lying on a table.

Every restoration model works from the pixels you give it. Scan at low resolution and you have thrown away the fine detail before the model ever sees the photo, and what comes back is the model's invention rather than your grandmother's face. Scan at high resolution and the model has real evidence to work from, so it invents less.

The numbers, from the Library of Congress preservation guidelines:

If you are digitizing a whole box rather than a handful of photos, the economics of mail-in services against a home scanner are covered in our guide to digitizing old family photos and videos.

One more number matters if the photo is going into a printed book. Lulu, the print-on-demand company most family memoirs run through, requires 300 PPI minimum for interior images and recommends staying under 600 PPI (Lulu image formatting). A restored file that looks sharp on a laptop can still print as mush. Check the pixel dimensions against the size it will occupy on the page before you commit to it.

How to restore an old family photo, step by step

This is the sequence we use. The order matters: every step after the first one gets easier if the earlier ones were done properly.

  1. Photograph or scan the original at full resolution. Flatbed scanner at 600 PPI for a print. If the photo is too fragile or too large for the scanner, lay it flat in even daylight, away from direct sun, and shoot it square-on with a phone in the highest-resolution mode. Turn the flash off.
  2. Save that scan untouched and back it up. Name it something like 1961-08_grandma-lake-cabin_MASTER.tif. This file never gets edited again. Everything else is a copy.
  3. Fix geometry and crop before anything else. Straighten the photo and crop away the scanner bed. Do not crop into the image yet, because the borders sometimes carry a studio stamp or a handwritten date.
  4. Correct exposure and contrast manually. Old prints fade toward flat grey. Pulling the black point and white point back to the ends of the histogram recovers more apparent detail than most AI enhancement does, and it does not invent anything.
  5. Run the automatic damage repair. Adobe's Photo Restoration neural filter sits under Filter, Neural Filters, Photo Restoration, and splits the work across three sliders: photo enhancement, facial enhancement, and scratch reduction (Adobe). MyHeritage, Google Photos, and most consumer tools expose a simpler version of the same thing. Start at the lowest setting that fixes the damage.
  6. Turn facial enhancement off, then decide. Run the restoration once with faces excluded and once with faces included, and compare them at 100%. If the enhanced version changed the shape of an eye, the hairline, or the expression, use the version without it. This is the step almost nobody does, and it is the step that decides whether the photo is still of your relative.
  7. Clean up what is left by hand. A spot healing brush or clone stamp on the three or four remaining marks takes two minutes and introduces nothing invented. Free tools like GIMP do this as well as paid ones.
  8. Export two files. A full-resolution TIFF for the archive, and a JPEG sized for wherever it is going. If the destination is a printed page, size it so the file still lands above 300 PPI at its printed dimensions.

Steps 2 and 6 are the ones that separate a restoration you can trust from a nice-looking image with an unknown relationship to the past.

What AI cannot recover, and the mistakes it makes instead

None of these are bugs. They follow directly from the mechanism described at the top.

Faces in small or heavily damaged areas. In a group photo where each face is forty pixels wide, there is almost no facial information in the file. The model has to invent nearly all of it, and it will. Wedding groups, school classes, and regiment photos are the worst case. Enlarge and inspect every face before you accept a group restoration.

Colorization. A colorizer infers color from context: skin, sky, grass, and wood are usually right, because they vary within a narrow range. Clothing, paint, cars, and furniture are guesses. A colorized photograph is a plausible illustration of a scene, not evidence of what the scene looked like. Label colorized files as colorized, so that a grandchild in forty years does not mistake the guess for a record.

Text and numbers. Shop signs, house numbers, newspaper headlines, and the handwriting on the back of a print come back confidently wrong. Generative models reconstruct the shape of text rather than the content. If the text matters, transcribe it from the original before restoring anything.

People who are not in the frame. Some tools offer to extend a photo beyond its borders. Whatever appears there was never in your family's house. Do not put an outpainted photograph in a family book.

Jewellery, uniforms, insignia, and medals. Small high-detail objects carrying real meaning are exactly the class of thing models smooth into generic versions. A regiment badge becomes a badge-shaped object. For a veteran's photographs this matters, and it is worth checking against the record; our guide to preserving a veteran's war stories and military service covers where those records live.

Motion blur and out-of-focus originals. Unblurring tools produce a sharp image. They do not produce the image that would have existed had the camera been still. The sharp result is a new picture of a similar scene.

The usable rule: AI restoration is trustworthy on materials and untrustworthy on identities.

How much does photo restoration cost?

Free, for most of the work.

Free tier. Google gave Magic Eraser, Photo Unblur, and Portrait Light to every Google Photos user on Android and iOS from 15 May 2024, with no subscription. Magic Editor, the heavier generative tool, is capped at ten saves per month on non-Pixel devices unless you hold a Google One plan of 2 TB or above (Google). For a handful of photos this costs nothing.

Family-history tools. MyHeritage repairs several photos free, and asks for a Complete subscription beyond that (MyHeritage). If you are already paying for a genealogy subscription, the restoration is included in something you own.

Photoshop. The Photo Restoration neural filter ships in the shipping version of Photoshop, so the cost is the Creative Cloud photography subscription rather than a per-photo charge. Worth it if you are restoring dozens of photos and want the manual tools alongside the automatic pass.

A human restorer. Published studio prices give the real range. Damaged Photo Restoration lists $30 per picture for standard work, $45 for cracks and medium scratches, and $60 for heavily damaged prints (Damaged Photo Restoration prices). Scott Hancock Photography bills at $65 per hour, which works out at roughly $45 for a half-hour job and $130 or more for a two-hour rebuild (published pricing). Severely damaged one-of-a-kind photographs run higher at specialist studios.

The sensible allocation for a family archive is to run everything through a free tool, then spend real money on the two or three photographs where the face is damaged and there is no second copy in existence. A human restorer working from reference photographs of the same person can do something no current model can: rebuild a specific face rather than a generic one.

We include AI restoration for up to 100 photos in our own $199 memoir, because a book of family photographs is where most people eventually want these files to end up, and paying separately per photo for that volume would cost more than the book.

Keep the original. The restored file is a copy, not a replacement

This is the part that professional archivists care about most and consumer tools mention least.

The National Archives defines a preservation master as a file created at maximum capture quality that "typically does not undergo significant processing or editing" (National Archives). The point of a master is that it can stand in for the original record. A file that has been through a generative model cannot do that, because parts of it were generated.

So keep three things, not one:

The same storage discipline applies to all three of the digital files: two copies on different media in different locations, migrated to current storage every few years. Our seven tips to digitize and preserve family memories covers the full rule.

Note that MyHeritage already does the right thing here by default, keeping your upload and writing the repair to a separate copy. Tools that silently overwrite the file you gave them are the ones to avoid.

Which photos are actually worth restoring

A box of 2,000 family photographs does not contain 2,000 photographs worth restoring. It contains perhaps sixty.

A printed memoir of 200 pages holds roughly 40 to 80 photographs before the pictures start crowding out the text. That constraint is useful, because it forces the only question that matters: which photographs carry a story somebody can still tell?

Our triage, in order:

  1. Photographs with a living witness. Anyone still alive who was in the room, or who knew the people in the frame. These are the highest-value images in the box and the only ones with a deadline attached.
  2. The only surviving photograph of a person. Restore it, and consider paying a human for it if the face is damaged.
  3. Photographs of places that no longer exist. The farm, the first flat, the shop, the ship. These tend to unlock long stretches of memory.
  4. Photographs where something visibly happened. A wedding, an arrival, a departure, a uniform, a hospital doorway.
  5. Everything else. Leave it digitized and unrestored. Restoration is cheap but attention is not, and a book with sixty considered photographs beats a hard drive with two thousand polished ones.

Restore the shortlist. Leave the rest as scans.

The part everyone skips: show the restored photo to the person who was there

This is where restoration stops being a technical exercise.

A restored photograph is the single most effective memory prompt we have seen. Faded prints are hard for older eyes to read, and a photograph someone cannot see clearly does not produce a story. Bring back the contrast and the detail, print it at a decent size, put it on the table, and the story comes out of them without a question being asked.

It also fixes what the software got wrong. The model guessed the dress was blue; she will tell you it was green and that she made it herself. The model smoothed out a scar; he will tell you where it came from. Only a witness can separate the evidence from the invention, and that opportunity has an expiry date.

Questions that work in front of a restored photo:

Record the answers rather than writing them down afterwards. The exact words and the pauses are the part that cannot be reconstructed later, by any model. Our guide to interviewing your grandparents before it is too late covers the protocol, and questions to ask your grandparents has a longer list to work from.

That recording is the thing AI genuinely cannot fake. A restored photograph is an informed guess about how a moment looked. A recording of the person describing it is the moment itself.

Frequently asked questions

Can AI really fix a torn photo?

Yes, when the tear runs through background, clothing, or scenery, where inpainting models are highly reliable and usually leave no visible seam. Tears that cross a face are a different problem, because the model rebuilds the whole face to make the patch blend and has no information about that specific person. For a tear through a face, either accept the damage or pay a human restorer working from other photographs of the same person.

How much does photo restoration cost?

Most of it is free. Google Photos gave Magic Eraser and Photo Unblur to all users at no cost in May 2024, and MyHeritage repairs several photos free before requiring a subscription. Human restorers publish prices of roughly $30 to $60 per photograph for standard through heavy damage, and hourly studios charge around $65 per hour, which lands at $45 to $130 or more depending on the rebuild.

Does AI photo restoration change what people actually looked like?

It can, and this is the main risk. Restoration models predict missing detail from training data rather than recovering it from the print, so a damaged face is rebuilt as a generic plausible face. PetaPixel reported in March 2026 that one restoration tool added hair, applied makeup, and enhanced features to produce what it treated as an attractive result. Run the restoration with facial enhancement turned off, compare the two versions at full zoom, and keep the one that still looks like your relative.

Is AI colorization of old photos accurate?

Only approximately. Colorizers infer color from context, so skin, sky, grass, and wood are usually close, while clothing, paint, vehicles, and furniture are guesses drawn from what similar scenes usually look like. Treat a colorized photograph as an illustration rather than a record, label the file as colorized, and ask anyone still living what the colors actually were.

What resolution should I scan a damaged photo at before restoring it?

The Library of Congress recommends 400 to 600 PPI for paper prints and around 3,000 PPI for 35mm slides and negatives. Scan at the higher end, in 16-bit color depth if the scanner offers it, and save the master as TIFF. Restoration software cannot recover detail that a low-resolution scan discarded, so a poor scan forces the model to invent more.

Should I keep the original photo after restoring it?

Yes, and the unretouched scan as well. The National Archives states that digitized images are not a replacement for originals and advises against discarding film and prints after digitization. A file that has passed through a generative model cannot serve as a preservation master, because parts of it were generated rather than captured. Keep the physical print, the untouched scan, and the restored copy as three separate things with names that distinguish them.

Will a restored photo print well in a book?

Only if the pixel dimensions hold up at the printed size. Lulu requires interior images at 300 PPI minimum and recommends staying under 600 PPI. A restored file that looks sharp on a screen can still print soft, because screens display around 100 PPI. Calculate the pixel width against the printed width before you place it, and rescan at higher resolution if it falls short.

Where to go next

If the box is still a box, start with how to digitize old family photos and videos, which covers scanners, mail-in services, and the resolution decisions this page depends on. For the wider archive, including video, audio, and paper, seven tips to digitize and preserve family memories covers the storage rules.

If someone in the photographs is still alive, that is the piece of work with a deadline. Start with how to interview your grandparents before it is too late.

Sources cited above