How do I restore old family photos with AI without ruining them?

Restoring old family photos with AI is now one of the most practical applications of consumer image technology, but it is also one of the easiest to get wrong. The short answer: scan your originals at high resolution, run them through an AI restoration or colorization tool that preserves facial structure, review the output critically against the original, and stop before the tool starts inventing details that were never there. Tools like Photoshop's Neural Filters and Restore features, ON1 Photo RAW's Restore AI, Photomyne, MyHeritage, and dedicated colorization services such as colorizethis.io can all do the job, but each has distinct failure modes. The difference between a restored photo that looks like your grandmother and a plastic-faced stranger is almost always in how you use the tool, not which tool you pick.

Why Old Photos Degrade and What AI Can Actually Fix

Also worth reading: How can I restore faded old photos for free in 2026? · Should you restore old photos before colorizing them? Here's what actually works in 2026? · How do you mitigate algorithmic bias when using AI to colorize and restore historical photos?

Physical photographs degrade through predictable mechanisms: silver-based black-and-white prints fade and develop yellowing from sulfur compounds in the air; color prints from the 1960s through 1980s suffer dye instability, with cyan dyes fading fastest and leaving orange-tinted images; and physical damage — creases, tears, water stains, mold spots, and emulsion cracking — accumulates from handling and poor storage. Scans of newspaper clippings and low-quality reprints add halftone patterns and blur on top of all that.

AI restoration tools address these problems in three broad categories. First, denoising and deblurring models remove grain, scanner artifacts, and softness, often trained on paired datasets of degraded and clean images — a technique NVIDIA demonstrated as early as July 2018 when it showed an AI that could fix bad photos by learning from other bad photos. Second, inpainting models fill in damaged regions: a torn corner, a scratch across a face, a missing chunk of background. Third, colorization models assign plausible colors to grayscale images by inferring what skies, skin, foliage, and clothing most likely looked like based on millions of training examples.

The critical thing to understand is that categories two and three involve guesswork. Inpainting does not recover lost information; it synthesizes new pixels that look statistically plausible. Colorization does not know whether your grandfather's suit was navy or charcoal; it picks the most probable answer. This is why professional archivists treat AI output as an interpretation, not a restoration in the strict conservation sense. For family use, interpretation is usually fine — but you should know which parts of the result are recovered data and which are educated invention.

How AI Restoration Actually Works Under the Hood

Most modern photo restoration tools are built on diffusion models or GAN-style architectures trained on large image datasets. When you upload a faded portrait, the model performs several operations in sequence. It detects faces using landmark detection, segments the image into regions (skin, hair, eyes, clothing, background), estimates and reverses degradation like noise and blur, then optionally regenerates fine detail — eyelashes, fabric texture, individual hairs — that was never captured in the original scan.

That last step is where quality diverges sharply between tools. A conservative model enhances contrast and removes scratches while keeping the underlying pixel structure intact. An aggressive model hallucinates detail: it will give everyone symmetrical teeth, smooth out period-accurate wrinkles, and occasionally change eye shape or jawline enough that family members say "that doesn't look like her." PetaPixel documented exactly this problem in its coverage of ON1's Restore AI, showing how aggressive settings turned ordinary ancestors into what reviewers called nightmare fuel — waxy, over-smoothed faces with uncanny features. Fstoppers' follow-up testing of ON1 Photo RAW 2026.4 reached a similar conclusion: the tool fixes old photos, but it has real limits, particularly on heavily damaged originals.

Colorization works differently. Models like those used in DeOldify and commercial services predict per-pixel color distributions conditioned on luminance and context. Skin tones are usually reliable because training data contains millions of portraits; uniforms, flags, and brand-specific items are where errors concentrate, because the model defaults to common colors rather than historically correct ones. A World War I uniform might come out olive-drab when it was actually a different national shade. Treat colorized output as a plausible visualization, and if historical accuracy matters, verify against written descriptions or surviving garments.

Practical Steps: From Shoebox to Restored Digital Archive

Start with the best possible source. Scan prints at 600 DPI minimum (1200 DPI for anything smaller than 3x5 inches) using a flatbed scanner rather than a phone camera when possible. Phone scanning apps like Photomyne — which raised $2.6 million in May 2016 specifically to bring old photos online and now offers built-in AI colorization — are convenient for bulk digitizing shoeboxes, and their quality is genuinely good for sharing purposes, but flatbed scans preserve more tonal information for restoration algorithms to work with. Handle prints by the edges, wear cotton gloves for fragile originals, and never attempt to uncurl or flatten brittle photos yourself; that is conservator work.

Second, organize before you restore. Sort into three piles: lightly faded but intact, physically damaged, and severely deteriorated. Lightly faded photos need only color correction and maybe mild colorization. Damaged photos benefit most from AI inpainting. Severely deteriorated ones may be beyond meaningful recovery, and attempting full regeneration on them produces the worst hallucination artifacts.

Third, work non-destructively. Always keep the original scan untouched in a separate folder. Run restoration on copies, save multiple versions (denoised only, denoised plus scratch removal, fully colorized), and compare them side by side. Every serious tool lets you adjust strength sliders — start low. A 30–40% face-restoration strength usually looks natural; pushing past 70% is where the wax-figure effect begins.

Fourth, review with fresh eyes. Zoom to 100% and check eyes, teeth, hands, and jewelry — the areas where generative models make mistakes. Then zoom out and ask whether the person still looks like themselves. Show results to older relatives who remember the people; they catch identity drift faster than any algorithmic metric.

Fifth, archive properly. Save masters as TIFF or high-quality PNG, create JPEG derivatives for sharing, back up to at least two locations, and embed names, dates, and places in filenames or metadata while someone alive still knows them. Restoration is worthless if the next generation cannot identify who is in the picture.

Comparing Your Options: Tools, Strengths, and Failure Modes

The market splits into desktop software, mobile apps, web services, and manual editing. Here is how the main approaches compare:

Desktop tools give you control but demand judgment: WSJ's testing of AI photo fixing concluded that knowing when to stop matters more than the software choice, because every tool will happily over-process if you let it. Web services trade control for convenience — upload, wait seconds, download — and purpose-built colorization sites generally produce more natural skin tones than general-purpose editors running default settings. Mobile apps excel at volume: Photomyne's workflow of scanning, auto-enhancing, and colorizing hundreds of prints in an afternoon is unmatched, though fine detail suffers versus a flatbed. Manual retouching remains the gold standard for irreplaceable heirlooms, at a cost of roughly $50 to $150 per photograph from a skilled retoucher, with complex reconstruction projects running higher.

A reasonable strategy for most families: use a web service or app for the bulk of the collection, reserve manual restoration or careful desktop work for the ten or twenty most important images, and keep every original scan regardless.

Common Mistakes That Ruin Otherwise Good Restorations

The single most common error is over-processing faces. Aggressive face restoration smooths skin, regularizes features, and erases the specific asymmetries that make a face recognizable. If a restored portrait could pass for a different person, it has failed even if it looks pretty. Reduce the enhancement slider until imperfections return slightly — a little authentic grain beats a lot of synthetic perfection.

The second mistake is restoring from bad scans. Running AI on a 72 DPI screenshot or a photo of a photo amplifies existing problems, because the model interprets compression artifacts and moiré patterns as image content and may hallucinate texture to match. Garbage in, plausible-looking garbage out. Rescan properly first; it takes minutes and changes everything.

Third, people destroy originals in the name of improvement: cropping off borders that contain studio stamps and dates, cutting photos out of albums with adhesive backing, taping tears with pressure-sensitive tape that yellows and lifts emulsion, or writing on prints with ballpoint pen. Do no physical harm. All corrections should happen digitally.

Fourth, trusting colorization blindly. AI assigns statistically likely colors, not verified ones. Eyes come out blue by default in some models; military uniforms default to generic khaki; a red dress might become green. If you know the true colors from family accounts, note them and manually correct, or choose tools that accept color hints.

Fifth, discarding the black-and-white version. Some family members prefer the original aesthetic, and future colorization technology will improve. Keep both versions permanently.

When to Act: Timing, Preservation Urgency, and Realistic Expectations

Act sooner than feels necessary. Color prints from the 1970s and 1980s are actively fading right now — dye degradation is chemical and continuous, and typical chromogenic prints show visible loss within 20 to 50 years depending on storage. Magnetic photo albums with sticky pages, popular from the 1960s onward, accelerate decay through acid migration and make removal risky. Every year of delay means less recoverable information, and AI cannot restore detail that was never scanned.

Set realistic expectations by damage type. Fading, yellowing, mild scratches, dust, and moderate blur respond excellently to current AI — expect near-magical results on 80% of typical family collections. Heavy creasing across faces, large missing sections, severe mold damage, and extreme blur produce mixed results: the model fills gaps plausibly but invents specifics. Completely destroyed areas require either acceptance of imperfection or explicit artistic reconstruction, which should be labeled as such. As Fstoppers and PetaPixel both found in their 2024–2026 testing cycles, the technology plateaued at "very good with limits" rather than reaching flawless reconstruction, and marketing claims of effortless perfection oversell what heavily damaged originals can yield.

Budget time realistically: digitizing a 500-print collection takes a weekend or two with a sheet-fed scanner or scanning app, restoration via web services adds seconds per image, but reviewing and correcting outputs carefully takes five to fifteen minutes per important photo. Plan accordingly rather than rushing everything in one sitting.

Cost Breakdown and What You Actually Need to Spend

You can do this for free or spend hundreds, and the middle path suits most people. Free options include trial tiers of web colorization services, GIMP with open-source plugins, DeOldify via free notebooks, and basic scanning apps. These handle light restoration and simple colorization well.

Subscription tools run $10 to $25 monthly for Adobe Photography plans including Photoshop's Neural Filters, or around $100 for a perpetual ON1 Photo RAW license with Restore AI included. If you have more than about fifty photos to process, a one-month subscription to whichever tool fits your workflow costs less than having five photos professionally restored.

Professional services charge roughly $20 to $60 per photo for standard restoration and $75 to $200+ for heavy reconstruction, with specialist colorization priced per image similarly. For a handful of irreplaceable portraits — a wedding photo, the only surviving image of a great-grandparent — this is money well spent, since human retouchers make judgment calls AI cannot. For hundreds of casual snapshots, per-photo pricing makes no sense; batch tools win decisively.

The hidden cost worth avoiding is redo: skipping proper scanning or keeping no originals forces you to repeat everything when standards improve, and they will improve. Storage costs pennies; a 2 TB drive holds millions of scans.

Preserving Results for the Next Generation

Restoration only matters if the files survive. Follow the 3-2-1 rule: three copies, two different media types, one off-site. Store masters as uncompressed TIFF or PNG, name files descriptively ("1948-06-chicago-wedding-grandma-rose.jpg"), and record who, when, and where in embedded metadata or a companion document. Print the most important restorations on archival paper — inkjet prints rated for 100+ years cost little more than drugstore prints — because digital formats obsolesce while paper endures. Share copies with siblings and cousins so the archive exists in multiple households. And revisit the collection every few years: as AI restoration improves, rerunning your untouched original scans through newer tools will yield better results than trying to enhance already-processed files. The original scan is your master asset; protect it above all else.