How do I restore scratched old photos with AI in 2026?

Restoring scratched, creased, or otherwise damaged old photographs used to mean hours of painstaking manual retouching in Photoshop, and even skilled retouchers charged $30–$100 per image for serious damage repair. In 2026, AI photo restoration tools can handle most of that work automatically — detecting scratches, filling in missing areas, removing dust and stains, sharpening blurred faces, and even colorizing black-and-white originals — often in under a minute per photo. This guide covers exactly how the technology works, which tools are worth your time, where AI still fails badly, and how to get results that look like a restored photograph rather than a plastic, over-smoothed imitation.

What AI Photo Restoration Actually Does

Also worth reading: What's the best way to restore old family photos with AI in 2026 without ruining them? · 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?

Modern restoration tools are built on generative models trained on millions of paired images: damaged photos and their clean counterparts, or high-resolution scans downgraded artificially so the model learns to reverse the degradation. When you upload a scratched photo, the system runs several stages in sequence. First, a defect-detection network identifies scratches, creases, tears, dust spots, and mold patterns — MyHeritage's Photo Repair feature, launched after their earlier colorization success, works this way, flagging damage regions before repair begins. Second, an inpainting model fills those flagged regions by synthesizing plausible content based on surrounding pixels; NVIDIA demonstrated this approach as early as July 2018, when Engadget reported on a system that could "fix bad photos by looking at other bad photos," learning degradation patterns rather than requiring pristine training data.

Third, super-resolution upscaling increases pixel dimensions while hallucinating plausible fine detail — this is why a 600×900 scan from 1953 can come out looking like a 2400×3600 file. Fourth, face enhancement applies specialized models trained specifically on facial features, since faces are what viewers scrutinize most and generic upscalers tend to mangle eyes and teeth. Finally, if you want it, a colorization model assigns plausible colors to black-and-white originals. Each stage is optional and each introduces its own risk of error, which is why understanding the pipeline matters: if your photo only has scratches, running face enhancement anyway can alter your grandmother's actual facial features into something she never looked like.

The Honest Limitations: Where AI Gets It Wrong

No honest guide to restoring scratched old photos with AI should skip the failure modes, because they are real and well documented. PetaPixel's review of ON1 Restore AI described outputs turning family photos into what the reviewer called "grotesque nightmare fuel" — melted facial features, invented jewelry, teeth that look like dentures from a horror film. The core problem is that generative inpainting does not recover information; it invents information that looks statistically plausible. If a scratch crosses someone's eye, the AI will draw a new eye, and that new eye may belong to no one who ever existed.

Three specific risks deserve attention. First, identity drift: heavily damaged faces get "repaired" toward the model's average of human faces, so siblings in one photo can end up looking more alike than they did in reality. Second, texture smoothing: aggressive denoising removes film grain and photographic texture along with noise, producing the waxy, airbrushed look that experienced restorers immediately recognize as AI output. Third, historical inaccuracy in colorization: colorization models guess skin tones, uniform colors, and fabric dyes from context, and they guess wrong regularly — a navy uniform may come out blue-gray, a 1920s dress may receive a color that didn't exist in that decade's dye palette. Treat every AI output as a draft, not a finished restoration, especially for archival or genealogical purposes where accuracy matters more than prettiness.

Step-by-Step: Restoring a Scratched Photo with AI

Start with the best scan you can produce, because garbage input caps everything downstream. Scan physical prints at 600 DPI minimum (1200 DPI for small or severely damaged prints), save as TIFF or PNG rather than JPEG, and avoid applying any scanner-level sharpening or auto-correction. If you only have a phone, photograph the print flat under diffuse daylight with no flash, then crop tightly.

Next, run scratch and damage removal first, before any enhancement. Tools like PicWish — which launched an upgraded old-photo restoration model with Ultra HD support — and MyHeritage Photo Repair detect and fill scratches automatically; check the result at 100% zoom and manually mask any region where the AI filled something incorrectly. Then apply denoising and sharpening conservatively; research published in Electronic Imaging (2022) noted that aggressive AI denoisers significantly enhance detail but also fabricate it, so prefer settings labeled "mild" or adjust strength below maximum. Run face enhancement only if faces are present and blurry, and compare before/after side by side to confirm the person still looks like themselves. Upscale last, targeting 2× to 4× depending on intended use — print needs roughly 300 DPI at final size, so an 8×10 print requires about 2400×3000 pixels. Colorize at the very end if desired, since colorization amplifies any remaining artifacts. Total hands-on time for a single photo is typically 5–15 minutes including quality checks, versus hours for manual retouching.

Comparing the Main Restoration Options in 2026

The market splits into three tiers: consumer web apps, desktop suites, and open-source pipelines. PerfectCorp's testing of twelve AI restoration apps found wide variance in face fidelity and artifact handling, and independent reviews from Fstoppers (testing VideoProc Converter AI as an all-in-one toolkit) reached similar conclusions — convenience varies less than output quality. Here is how the main categories compare:

Remini, developed originally by BigWinePot Inc. and acquired in June 2021, remains one of the most downloaded mobile options and excels at face enhancement, though reviewers consistently note it produces a distinctive smoothed style that some users love and others find artificial. HackerNoon documented a hobbyist building a full open-source pipeline restoring black-and-white photos to 4K color using free models, proving the DIY route is viable if you have patience and ideally a GPU with 8GB+ VRAM. Blockchain Council has also covered prompt-based approaches using general-purpose image generation models, which work but give you far less control over preserving original identity.

Common Mistakes That Ruin Restorations

The most frequent mistake is stacking every enhancement at maximum strength. A photo with minor surface scratches does not need 4× upscaling plus full face reconstruction plus aggressive denoising; each pass compounds artifacts, and the cumulative effect is the uncanny, doll-like result that gave tools like ON1 Restore AI their bad reputation. Apply only the corrections the specific damage requires.

Second, people restore from low-quality copies instead of finding the original. A screenshot of a scanned photocopy of a print contains almost no recoverable detail; the AI will invent nearly all of it, and the result is fiction. Always hunt for the best-generation source — the original negative, the earliest print, the highest-DPI scan — before touching any tool. Third, skipping the before/after comparison at full zoom. Artifacts that vanish at thumbnail size become obvious at 100%, particularly around eyes, hairlines, and text (AI frequently garbles handwriting and printed dates on studio stamps). Fourth, overwriting originals: always keep the untouched scan archived separately, because future tools will be better than today's, and a destroyed original cannot be reprocessed. Fifth, trusting colorization blindly for anything historically sensitive — verify clothing colors against period references if accuracy matters.

What It Costs and Whether Paid Tiers Are Worth It

Costs in 2026 range from genuinely free to moderately expensive. Most web apps offer a free tier with watermarks, resolution caps, or daily limits — typically 2–5 free restorations per day at reduced output resolution. Subscriptions run roughly $5–$15 monthly (Remini's mobile subscription sits near the lower end; MyHeritage bundles Photo Repair and colorization into broader genealogy plans around $129–$189 annually, which makes sense only if you also want the family-tree features). Desktop one-time purchases like ON1 products cost $50–$100 and pay off if you're processing hundreds of photos, since there's no per-image quota. The open-source route costs nothing but your time and electricity; expect a weekend of setup if you're comfortable with Python.

For most families digitizing an album of 50–200 photos, a one-month subscription to whichever web app tested best on your particular photos is the rational move: test the same damaged photo across two or three services' free tiers first, pick the winner, subscribe for one month, batch-process everything, then cancel. Paying annually rarely makes sense unless you restore photos continuously.

When to Use AI Versus When to Call a Professional

AI restoration is the right choice when damage is moderate (surface scratches, fading, mild creases), when volume is high, when budget is limited, and when a slightly idealized result is acceptable. It handles the classic shoebox-album scenario — dozens of 1970s–1990s prints with dust, fading, and light scratching — remarkably well, and colorization of black-and-white family photos for sentimental sharing is now effectively solved at the consumer level.

Call a professional human restorer when the photo is severely torn or missing large sections, when it has significant monetary, historical, or legal value, when accurate identity preservation is non-negotiable (court evidence, archival donation, published biography), or when the AI output repeatedly fails on a particular face. Professional manual restoration still costs $30–$150 per image depending on damage severity, but a skilled retoucher reconstructs from reference knowledge — other photos of the same person, period-accurate details — rather than statistical guessing. A sensible hybrid workflow exists too: use AI for the bulk cleanup, then hand the few hardest images to a professional, or use AI output as a starting layer for manual refinement. And whatever path you choose, act sooner rather than later on physical prints: mold, adhesive degradation, and chemical deterioration of chromogenic color prints (which begin noticeably fading within decades even in dark storage) mean the source material itself degrades every year you wait.

Practical Recommendations for Different Situations

If you have one precious damaged photo, scan it at high DPI, try three different tools' free tiers on the identical file, and pick based on face fidelity rather than overall sharpness — sharpness is easy to add, wrong facial features are not. If you have a large family archive, standardize your scanning first (consistent 600 DPI TIFFs, logical filenames), then batch-process in a desktop tool or a single month of a web subscription, keeping a folder structure of originals, AI drafts, and finals. If you're technically inclined and privacy-conscious, the open-source route with CodeFormer-class face restoration models plus a separate inpainting pass gives you weight controls that let you dial back the "beautification" that consumer apps force on you. And if your goal is colorization specifically — turning black-and-white ancestors into color for a family celebration, say — dedicated colorization tools generally outperform all-in-one restorers on color plausibility, so consider a two-tool workflow: repair first, then colorize with a specialized service. Whichever route you take, the single most valuable habit is skepticism: zoom in, compare against the original, and ask whether the person in the restored photo still looks like the person in the damaged one.