Restoring damaged old photos with AI is now one of the most accessible forms of digital preservation available to anyone with a smartphone or a browser. As of August 2026, AI restoration tools can repair scratches and tears, remove dust and mold spots, sharpen blurred faces, upscale low-resolution scans, and add plausible color to black-and-white photographs — often in under a minute per image. This guide covers what these tools actually do, how well they work, which options exist at different price points, the mistakes that ruin otherwise good restorations, and when it makes sense to act rather than wait.
What AI Photo Restoration Actually Does
Also worth reading: How can I restore a photo of my dad that is damaged or faded? · How do you mitigate algorithmic bias when using AI to colorize and restore historical photos? · How to restore vintage black and white photos using AI tools like colorizethis.io?
Modern restoration models are trained on millions of paired images: degraded versions of photographs alongside their clean originals. By learning the statistical patterns of scratches, creases, fading, and blur, the model can predict what a damaged region most likely looked like. NVIDIA demonstrated this approach as early as July 2018, showing an AI system that could fix bad photos by learning from other bad photos, and the technique has matured considerably since then. Today's tools typically bundle several operations into a single pipeline: denoising, scratch removal, face reconstruction, super-resolution upscaling (often 2x to 4x), and colorization.
It is worth understanding what is genuinely happening versus what marketing implies. When a tool 'reconstructs' a face torn down the middle, it is not recovering lost information — no algorithm can do that. It is generating a statistically plausible guess based on thousands of similar faces it has seen during training. The same applies to colorization: the sky will be blue because skies are usually blue, not because the AI knows your grandmother's dress was navy. This distinction matters for anyone restoring historically significant material, where invented detail can quietly replace authentic record.
The results vary enormously by damage type. Surface-level problems — dust, mild scratches, yellowing, low contrast — are handled reliably, often with 90%+ visual fidelity. Structural damage like large tears, missing corners, water damage, or heavily degraded emulsion forces the model to invent much more content, and quality drops sharply. A 2026 Fstoppers review of ON1 Photo RAW 2026.4's Restore AI noted it can fix old photos but has real limits; PetaPixel went further, showing cases where Restore AI turned family portraits into what they called grotesque nightmare fuel. Both observations are accurate: the technology is impressive within its envelope and unsettling outside it.
Why Old Photos Deteriorate and Why Timing Matters
Physical photographs degrade through several mechanisms, and understanding them helps you prioritize. Color prints from the 1960s–1980s are often the worst offenders: the chromogenic dyes used in that era fade rapidly, with magenta typically failing first, which is why so many prints from that period have an orange-cyan cast. Black-and-white silver gelatin prints are far more stable and can last centuries under good conditions, but they still suffer mechanical damage — creases, tears, silvering, and surface abrasion. Humidity drives mold growth that eats into the emulsion permanently, while light exposure bleaches dyes irreversibly.
The practical implication is simple: every year you wait, marginal photos get worse, and some cross the threshold where even AI cannot produce a convincing result. Digitization itself is time-sensitive. A photo scanned today preserves whatever information remains; a photo scanned five years from now may have lost detail to continued fading or mold spread. Restoration can always be re-run later as algorithms improve — the scan is the irreplaceable asset. A reasonable rule of thumb: if a photo shows visible fading, foxing (brown spotting), or any active mold, digitize it within weeks, not years.
There is also a generational argument for acting now. The people who can identify who is in a photo, correct the AI's color guesses ('her eyes were green, not brown'), and confirm dates are aging. Capturing that context alongside the restored image turns a file into a documented family record. Several news stories in 2025–2026 illustrated this well, including Shreveport brothers who used AI to restore historic Mardi Gras photographs while the community members who remembered those events were still alive to verify them.
How to Restore Damaged Photos with AI: Step by Step
The process is straightforward enough that most people complete their first restoration in 15–30 minutes. Start with the best possible source. Scan physical prints at 600 DPI minimum — 1200 DPI if the print is small (wallet-size) or heavily detailed — using a flatbed scanner rather than photographing the print with a phone, which introduces perspective distortion and uneven lighting. If scanning isn't possible, photograph the print outdoors in diffuse daylight on a flat surface, filling the frame completely. Save files as TIFF or maximum-quality PNG/JPEG; never restore from a screenshot or a compressed copy of a copy.
Next, choose a tool matched to your damage type. Upload the image and run scratch/damage removal first, before upscaling or colorization — order matters, because repairing artifacts after enlargement produces cleaner results than enlarging first and trying to fix enlarged defects. Review the output at 100% zoom, paying particular attention to faces, hands, text, jewelry, and architectural details, which are where generative models most often hallucinate. Most tools let you compare before/after side by side; use it ruthlessly.
If the result distorts a face or invents odd features, dial back the restoration strength if the tool offers that control, or mask the problem area and leave it partially unrestored — a slightly imperfect original detail beats a confidently wrong invention. Apply colorization last, since color decisions should sit on top of a structurally sound image. Finally, archive both versions: the raw scan and the restored file, ideally in two locations (a cloud service plus an external drive). Label everything with names, dates, and locations while you still have relatives who can supply them.
Comparing Your Options: Tools, Approaches, and Costs
The 2026 market splits into four broad categories, each with distinct trade-offs. Consumer web apps and mobile apps dominate on convenience; desktop software offers more control; open-source pipelines offer maximum control at the cost of technical skill; and professional services handle the hardest cases. PerfectCorp's testing of 12 AI restoration apps found meaningful quality gaps between leaders and laggards, particularly in face handling, while iLounge's 2025 roundup confirmed that mobile-only tools remain weaker on severe structural damage. MyHeritage added dedicated Photo Repair for scratched or damaged photos alongside its long-running colorization feature, making it a common choice for genealogists.
| Feature | Web/Mobile Apps | Desktop Software | Open-Source Pipeline | Professional Service |
|---|---|---|---|---|
| Typical cost | Free–$10/mo | $80–$130 one-time (e.g., ON1 Photo RAW 2026.4) | Free (GPU recommended) | $5–$50 per photo |
| Skill required | None | Low–moderate | High (Python, command line) | None |
| Severe tear/loss repair | Inconsistent | Moderate | Good with tuning | Best available |
| Face accuracy | Variable; check closely | Better controls | Tunable | Human-reviewed |
| Batch processing | Limited | Yes | Yes | Yes (quoted) |
| Privacy of family photos | Uploaded to servers | Local processing | Fully local | Depends on provider |
| Time per photo | Under 1 minute | 2–5 minutes | 1–10 minutes | Days |
Where AI Restoration Falls Short: Honest Limitations
No honest guide skips the failure modes. The biggest is facial hallucination: when a face occupies few pixels or is obscured by damage, the model fills in features from its training distribution, producing a person who looks plausible but may not resemble the actual subject. For family history this is more dangerous than an obvious glitch, because a subtly wrong face gets accepted as truth and propagates through the family tree. Always compare restored faces against other photos of the same person when any exist.
Text, patterns, and period-specific objects are second-tier problems. AI frequently mangles handwritten inscriptions, signage, newspaper clippings, and uniform insignia, smoothing them into illegible approximations. Clothing styles, military medals, and vehicle badges may be rendered generically rather than accurately. Historians and archivists rightly insist that AI-restored images be labeled as such and kept separate from originals — major institutions treat restoration output as interpretive artwork, not documentary evidence.
Colorization carries its own caveats. Skin tones across different ethnicities are handled unevenly by some models, and era-specific colors (military uniforms, vintage paint schemes, traditional garments) are frequent guesses. The 109-year-old video that was colorized and upscaled with AI drew admiration for its vitality and criticism for its invented palette in equal measure — a fair summary of the whole field. Treat colorized output as one interpretation, and keep the grayscale master alongside it.
Common Mistakes That Ruin Restorations
The most damaging mistake is working from a bad copy. Restoring a low-resolution JPEG that was already screenshotted, compressed, and cropped wastes the entire effort; garbage in, garbage out remains true regardless of how good the model is. Scan properly once, at high resolution, and every future improvement benefits. The second mistake is over-processing: cranking sharpening and enhancement sliders to maximum produces waxy skin, plastic textures, and the uncanny-valley look that PetaPixel mocked. Moderate settings applied to a good scan beat aggressive settings applied to a poor one.
Third, people often delete or overwrite the original scan after restoring. Never do this. Models improve yearly, and a restoration you dislike in 2026 may be redoable beautifully in 2028 from the same archive file. Fourth, skipping verification: family members are your ground truth for identity, colors, and dates, and consulting them costs nothing. Fifth, ignoring privacy — uploading sensitive family photos to unknown web services means trusting their retention policies; read them, or choose a local-processing option for intimate images. Finally, don't expect one pass to fix everything; running damage repair, then targeted touch-ups, then colorization in stages consistently outperforms a single 'enhance everything' button.
Cost Breakdown and What You Actually Need to Spend
Budget-conscious users can accomplish a great deal for zero dollars. Free tiers of several web apps allow a handful of restorations per day, sufficient for a small album. The open-source route costs nothing beyond electricity and optionally a used GPU. Mid-range spenders should consider a one-time desktop purchase around $99–$129 (ON1 Photo RAW sits in this band) if they have hundreds of photos and want batch processing plus local privacy. Subscription services at $5–$12/month make sense for a focused project — say, restoring 200 photos over two months — then cancel.
Professional restoration services charge roughly $5–$20 per photo for standard work and $25–$75+ for severely damaged pieces requiring manual retouching combined with AI assistance. Reserve this tier for genuinely important images: a wedding portrait, a deceased relative's only surviving photo, or anything destined for print display. For bulk albums of snapshots, professional pricing multiplies quickly and rarely justifies itself against competent DIY results. Whatever route you take, budget for storage too — high-resolution scans and restorations accumulate fast, and a 2TB external drive (~$60) covers most family archives for years.
When to Act and How to Build a Lasting Archive
Act now on three categories: photos showing active deterioration (fading color prints, mold, flaking emulsion), photos whose subjects or witnesses are elderly, and photos stored in poor conditions (attics, basements, garages). Everything else can wait months without harm. Set aside a weekend, borrow or buy an inexpensive flatbed scanner ($100–$200 new, less used), and work through collections in priority order rather than trying to digitize everything at once.
Build the archive deliberately: consistent file naming (year_event_person), a spreadsheet or genealogy app linking images to names and stories, and the 3-2-1 backup rule — three copies, two media types, one offsite. Revisit restorations periodically; as models improve through 2026 and beyond, re-running your archived scans through newer tools costs minutes and yields visibly better faces and colors. The combination of a preserved scan, a thoughtful AI restoration, and recorded human context is, realistically, the best possible outcome for a damaged photograph — better than the original will ever be again, and honest about what was recovered versus reconstructed.