Restoring Old Family Photos With AI: Face Accuracy, Colorization, Damage Repair, and Digital Preservation

Restoring Old Family Photos With AI: Face Accuracy, Colorization, Damage Repair, and Digital Preservation

AI photo restoration can remove scratches, fix fading, sharpen blur, and colorize old family photos convincingly. Faces are harder to restore: when detail is missing, tools invent plausible features instead, and the result can look like a stranger wearing your grandmother’s expression. The safe workflow: scan properly, keep the original untouched, do light repairs first, apply heavy AI only where you can verify it against the source, treat colorization as interpretation, and archive both versions.

For family photos, preserving a familiar face matters more than making the image look sharp. I looked at restoration tests, vendor documentation, and archival guidance with that tradeoff in mind.

What AI photo restoration actually does (and why faces are the hard part)

Restoration AI is a prediction engine, not a time machine. When your scan is missing detail, the model fills the gap with the most statistically likely detail. For scratches, dust, and yellowing, that’s exactly what you want. On faces, it’s the one property you don’t want.

Face hallucination is when a restoration model generates facial features that look plausible but were never in the original photo: a jawline the person didn’t have, eyes shaped by the training data instead of by genetics. Adobe’s own product page describes the mechanism plainly: Neural Filters “use algorithms to generate new pixels in your photos.” Fine for a torn background; a different matter for the only photograph of someone.

Builders are surprisingly open about the tradeoff. The README for GFPGAN, Tencent’s widely used open-source face restoration model, lists “slight change in identity” as a known weakness of its most natural-looking version. CodeFormer, a NeurIPS 2022 model, ships with a “fidelity weight” slider from 0 to 1 where, per its own documentation, smaller values produce higher-quality results and larger values produce higher fidelity to the damaged input. When researchers build a control specifically to stop their model drifting from the real person, the drift is real.

A 2025 survey of deep face restoration in ACM Computing Surveys (Wenjie Li and colleagues) calls the task ill-posed, meaning many plausible outputs fit one damaged input, and reviews identity-preservation metrics that exist precisely because models wander.

Which AI photo restoration tools are worth using in 2026?

It depends on the job. Photoshop gives the most manual control, MyHeritage is the easiest for genealogy, Topaz is the strongest local option, Palette.fm does colorization only, and the open-source models cost nothing if you’re comfortable with Python. I based the face-accuracy risk estimates on published tests and vendor documentation.

A few notes the table can’t fit:

• Photoshop’s Photo Restoration filter is still tagged beta even in Photoshop 2026 (version 27.11 shipped in September 2026), and it downloads the model on first use.
• Topaz rebranded Topaz Photo AI to Topaz Photo and moved to subscription-only pricing; Recover Faces 3 arrived in v1.6.0 in May 2026.
• MyHeritage’s photo tools sit inside one of genealogy’s biggest databases; the company added 6.6 billion historical records in 2025 alone, per its Year in Review, and free accounts get only limited uses of the photo features.

What’s the safest way to restore old family photos with AI?

Work from a good scan, keep the master untouched, and never let face AI run at full strength unverified.

1. Scan properly, once. The U.S. National Archives’ technical guidelines specify a recommended pixel array of 4,000 pixels across the long dimension of a photographic print. For a 4x6 print that works out to roughly 600 to 670 ppi, which is why 600 dpi became the standard advice for family prints. The Library of Congress’s personal archiving guidance says 300 to 400 dpi is fine for same-size reprints, but extra resolution gives AI real detail to work with instead of inventing its own. Scan the whole print with a small border, in 24-bit color, saved as TIFF.
2. Lock the master file. Name the scan as your master and never edit it; every restoration output becomes a new file.
3. Do the low-risk repairs first. Dust removal, descratch, and modest sharpening rarely touch identity. Keep the sliders low: CreativePro’s November 2025 walkthrough of Photoshop’s Photo Restoration filter found that pushing the Scratch Reduction slider hard produced mangled faces on a difficult image. Small tears are often better fixed with a manual heal tool than an AI pass.
4. Apply heavy AI to faces only, starting at zero. In Photoshop, set the Enhance Face slider to 0 and raise it gradually. In Topaz Photo, do the same with face recovery strength. In CodeFormer, push the fidelity weight toward 1. Stop at the first result that still looks like the person, not at the prettiest one.
5. Verify against independent evidence. Compare the restored face with any other photos of the same person, and with people who knew them. Check the unglamorous features: ear shape, teeth, eyebrows, hairline, glasses. If only one photo exists, label the AI output as an approximation, because nobody can verify it.
6. Colorize a copy, never the master. Colorization is the most interpretive step in the pipeline, so run it last on a separate file and keep the black-and-white original in the archive.
7. Export for sharing, archive for keeps. Keep the TIFF master, the restored TIFF, and the colorized version; make JPEG copies for sharing. Then apply the 3-2-1 backup rule: three copies, two media types, one offsite. The Library of Congress adds one more habit: organize everything into one collection folder and migrate it to new storage media about every five years.

Does AI photo restoration make up faces?

Yes, sometimes, and the evidence is no longer anecdotal. In March 2026, PetaPixel examined 73 before-and-after examples that ON1 supplied to press for its new Restore AI feature in Photo RAW MAX. One turned a woman’s face into a distorted mess, and PetaPixel counted five outright bad results among 74 images. ON1 asked the outlet not to publish that example and later apologized. What matters more is what ON1’s own VP of product, Dan Harlacher, told PetaPixel: “some faces, especially tiny and or blurry faces have little data to start with so creating a faithful representation is difficult and sometimes it is better to just turn it off.” When a face has too little detail, leaving it alone can preserve more of the original than an AI repair.

PetaPixel’s editor in chief Jaron Schneider followed up days later by running Restore AI over public-domain historical photos. On a group portrait, the restored version changed a man so much he stopped being identifiable. His verdict on another photo doubles as a warning for anyone restoring family pictures: the AI recolored a man’s jacket, and “who is to say his jacket was blue?”

“The ‘restoration’ stole his identity.” — Jaron Schneider, PetaPixel, “‘Restoring’ Old Photos With AI Is a Fundamentally Broken Concept”, March 28, 2026

The pattern keeps repeating across tools. A June 2026 report by Seattle’s KIRO 7 was headlined “AI photo restoration works until it starts inventing what’s not there,” describing tools that stop repairing and start guessing. And genealogists on Google’s own support forums complained through mid-2026 that Gemini’s image model was altering faces even during simple colorization requests.

Three practical rules fall out of the evidence:

• Risk scales with damage. A sharp scan needing dust cleanup is low risk. A tiny, blurry, torn face is the worst case, which is exactly when the tools are marketed hardest.
• Damage repair and face reconstruction are different jobs. Scratches and fading have visual context that constrains the model. Facial detail does not, so that’s where invention happens.
• The absence of information is permanent. No tool, AI or human, can recover pixels that were never captured. As Schneider noted, hand restoration is no different; AI just fills the void with confident-looking invention instead of an honest approximation.

If a restored face matters as a record of the person, treat it like genealogists treat a transcribed record: a derivative source, useful, never the original.

Is AI photo colorization historically accurate?

No. Colorization is a statistical guess dressed up as fact. Scienceline, NYU’s science journalism outlet, put it well in 2021: colorization networks learn associations “in the same way that we might learn that the sky is blue,” but they cannot know what color a specific tie, dress, or wall actually was, because that information isn’t in a black-and-white image. Professional colorizers spend weeks researching uniforms, fabrics, and records for a single image; no consumer tool does that.

Adobe’s own documentation adds a caveat most people miss: the Colorize filter produces “less accurate color predictions” on faded or badly exposed images, which describes a large share of family photos. Critics have also noted that results inherit the demographic biases of training data, so skin tones and period details can drift toward dataset averages rather than the photograph’s reality.

None of this makes colorization useless. A colorized portrait can make a distant relative feel real to younger family members, and that’s legitimate. Treat it as an interpretation layered on the record, not as the record itself: colorize a copy, keep the original black-and-white scan, and label the file as colorized so nobody cites it as evidence in twenty years.

What happens to your family photos when you upload them?

It depends on the service, and the differences are bigger than most guides admit. ON1 states on its Restore AI page that photos “are handled securely and never used to train AI models.” MyHeritage’s terms confirm you retain ownership of photos you upload, while granting the company a license to host and process them as part of the service; its terms page carries an October 14, 2026 update date, so check the current text before uploading anything you care about. Remini is owned by Bending Spoons, an Italian subscription-app giant, which is worth knowing before you hand it a decade of family scans.

The privacy-conscious route is local processing. Photoshop’s Neural Filters mostly run on your machine after the model downloads, Topaz Photo runs locally, and the open-source models run anywhere you like. Local tools also sidestep the consent question that surfaced in 2021 when MyHeritage’s Deep Nostalgia went viral and TechCrunch flagged the ethics of animating faces of people who never agreed to anything, living or dead. MyHeritage cautions against uploading photos of living people without permission, and that applies just as much to restoration uploads.

How do you preserve restored photos for the long term?

Restoration is wasted effort if the files die on a failing drive. The preservation rules are short and boring, which is why they work.

• Use the right formats. The Library of Congress’s Recommended Formats Statement lists TIFF first for digital photographs, ahead of JPEG 2000, PNG, and JPEG, and asks for the highest resolution available, uncompressed or losslessly compressed, 16-bit where possible. NARA’s guidelines are blunter: uncompressed TIFF for production masters, and “do not use JPEG compression within the TIFF format.” JPEG is fine as a sharing copy.
• Follow 3-2-1. Three copies, two different media types, one offsite. A cloud service counts as offsite, but it doesn’t count as your only copy.
• Migrate on a schedule. The Library of Congress suggests moving your collection to fresh storage media about every five years. Drives fail quietly and formats age.
• Attach the context. Put names, dates, and places in filenames and in the files’ embedded metadata, so “IMG_4021_restored_final.jpg” never becomes an orphan.
• Keep the physical originals. Archives treat scans as copies; the print itself remains the artifact.

One last habit: write down which version is which. “Original scan,” “restored,” and “colorized” should be visible in the filename, because a colorized, face-enhanced derivative floating in a family folder with no labels is exactly how invented details become family fact.

Frequently asked questions

Does AI photo restoration make up faces?

Sometimes, yes, especially when the original face is small, blurry, or damaged. ON1’s own product lead told PetaPixel in March 2026 that with tiny or blurry faces it’s sometimes better to turn the feature off, and GFPGAN’s documentation admits a “slight change in identity” as a known weakness. Keep face enhancement at zero until you’ve compared the result against other photos of the person.

What is the best AI photo restoration software in 2026?

No single winner. Photoshop’s Photo Restoration filter offers the most control, MyHeritage bundles restoration, colorization, and animation for genealogy use, Topaz Photo is the strongest offline option at $199/yr, and Palette.fm handles colorization only. At zero budget, GFPGAN and CodeFormer are free but need technical setup.

Can AI fix badly damaged photos?

Partially. Modern tools repair scratches, dust, fading, and small tears well, because surrounding context tells the model what belongs there. Large missing areas, torn-off sections, and badly blurred faces are different: whatever the AI puts there is invention, not recovery. Do the light repairs first and judge any heavy AI pass against the original.

Is colorization historically accurate?

No. AI colorizers produce plausible colors based on statistical patterns in training data, and Adobe’s documentation concedes that faded or badly exposed images yield less accurate color predictions. For family use, that’s usually fine as long as you keep the original black-and-white version and label the colorized copy as an interpretation.

What resolution should I scan old photos at?

About 600 dpi for standard prints is a good default: the National Archives’ digitization guidelines call for around 4,000 pixels across a print’s long dimension, which for a 4x6 print lands near 600 to 670 ppi. Go up to 1,200 dpi for wallet-size photos or planned enlargements, and save the scan as uncompressed TIFF.

Is Remini free?

Partly. The free tier is limited, with watermarks or daily caps, and the full version runs on auto-renewing subscriptions: the US App Store lists Remini Lite at $7.99 per week and Remini Pro at $9.99 per week, with monthly and yearly options. Weekly billing with a free trial that converts automatically is Remini’s most common complaint, so set a reminder to cancel if you’re only restoring one batch.

How do I preserve restored photos long-term?

Keep an uncompressed TIFF master of every version, follow the 3-2-1 backup rule (three copies, two media types, one offsite), and migrate the whole collection to new storage about every five years, per the Library of Congress’s personal archiving advice. Put names and dates in filenames and embedded metadata, and keep the physical prints. The originals outlast every file format so far.