Restoring old black and white photos has changed dramatically over the past decade. What once required hours of manual work in Photoshop — cloning out scratches, rebuilding torn corners, and hand-painting color layer by layer — can now be done in seconds with AI-based tools. As of August 2026, the process typically involves three stages: digitizing the physical print, repairing damage (scratches, tears, fading, blur), and optionally adding realistic color through AI colorization. This guide walks through the entire workflow, compares the main tools available today, and flags the mistakes that ruin otherwise good restorations.
The Short Answer: The Modern Restoration Workflow
Also worth reading: What is the best way to restore damaged old photos with AI 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?
To restore an old black and white photo today, you scan or photograph it at high resolution (at least 600 DPI for small prints), run it through an AI restoration tool that handles scratch removal, denoising, sharpening, and upscaling, then apply AI colorization if you want a color version. Tools like MyHeritage's Photo Repair, Photomyne, PicWish, ON1 Photo RAW 2026.4's Restore AI, Photoshop's neural filters, and browser-based services such as ColorizeThis.io all follow this general pattern. A one-click result takes under a minute; a careful manual-plus-AI hybrid workflow might take 15 to 45 minutes per photo.
The key insight for 2026 is that no single tool does everything well. Repair models are good at removing scratches and noise but often hallucinate faces. Colorization models produce plausible skin tones but frequently get uniforms, flags, and period-specific colors wrong. Upscalers can push a 1930s snapshot to 4K, but they invent detail that was never captured on film. Understanding these limits matters more than picking any particular brand.
Why Old Photos Degrade — And What You're Actually Fixing
Black and white photographs from roughly 1850 to 1980 were made on paper supports coated with silver-based emulsions. Over time, several failure modes appear. Silver particles migrate and cause yellowing or mirroring. Paper acids break down fibers, causing brittleness and foxing (brown spots). Humidity swings cause emulsion cracking. Adhesives in old albums leave staining. Prints stored in attics fade toward low contrast, while negatives suffer base fog and scratches.
Each of these problems maps to a different digital fix. Fading is corrected with contrast curves and levels adjustment. Scratches and dust are removed by inpainting algorithms that fill damaged pixels using surrounding texture. Blur and softness are addressed by deconvolution-style sharpening or generative upscaling. Torn or missing areas require generative fill, which reconstructs plausible content. Knowing which problem you have tells you which tool setting to use — running a heavy generative model on a photo that only needs a contrast boost will degrade it rather than improve it.
There's also a philosophical question worth acknowledging: restoration changes the artifact. A heavily AI-processed image is no longer a faithful copy of the original print. Serious archivists therefore recommend always preserving the untouched scan as a master file and treating every restored version as a derivative. This costs nothing to do and protects you if a future tool does a better job.
Step-by-Step: Digitizing Your Photos Correctly
Restoration quality is capped by your input quality, so digitization deserves real attention. For prints smaller than about 8x10 inches, use a flatbed scanner at 600 DPI minimum; 1200 DPI gives more headroom for cropping and upscaling. Photographing works too — place the print on a flat surface near a window with indirect daylight, shoot straight down, disable flash, and use your phone's highest resolution mode. Apps like Photomyne automate this capture-and-crop process and can batch multiple photos per shot, which matters when you face a shoebox of 500 family prints.
Handle originals carefully. Wear cotton gloves or clean dry hands, support fragile prints on a rigid surface, never bend cracked emulsion, and remove photos from magnetic albums slowly — heat from a hairdryer on low can soften old adhesive without damaging most prints. If a photo is genuinely valuable or historically significant, consider professional scanning services before attempting anything yourself.
Save scans as TIFF or maximum-quality PNG rather than JPEG. JPEG compression artifacts compound when an AI model processes the file afterward, producing smudged textures and ringing around edges. File sizes will be large — a 600 DPI scan of a 4x6 print runs 20–50 MB — but storage is cheap and re-scanning is not always possible.
AI Repair: What the Current Tools Actually Do Well
The repair stage uses machine learning models trained on pairs of degraded and clean images. NVIDIA demonstrated this approach publicly as early as July 2018 with a system that could fix bad photos by learning from other bad photos, and the technique has matured steadily since. In 2025–2026, consumer-facing options include MyHeritage Photo Repair (strong on scratches), Photoshop's Neural Filters and generative fill (most control, steepest learning curve), ON1 Photo RAW 2026.4's Restore AI (integrated into a full raw editor), PicWish (mobile-friendly, added Ultra HD support recently), and Flux Kontext-based pipelines that hobbyists build with open-source tools.
Reviews of these tools show a consistent pattern worth internalizing. They excel at mechanical damage: dust specks, hairline scratches, mild creases, surface stains, and moderate noise. They struggle with large missing regions, severe facial damage, and structural reconstruction. PetaPixel's testing of ON1's Restore AI found it sometimes turned family portraits into what they described as grotesque results — smooth, waxy, uncanny faces — when pushed beyond its limits. Fstoppers' review of the same release concluded it fixes old photos but has real limits. The lesson: run repair at conservative strength first, inspect faces at 100% zoom, and fall back to manual cloning for anything the model mangles.
A practical rule of thumb: if damage covers less than roughly 5–10% of the image area and doesn't cross a face, automated repair will usually look good. If a face itself is scratched or blurred, expect to either accept some artificial smoothing or invest manual effort.
AI Colorization: Adding Plausible Color Without Lying About History
Colorization is where expectations need the most calibration. Modern models predict plausible colors based on learned statistics — grass is green, sky is blue, brick is red — but they cannot know actual historical facts. A soldier's uniform might come out olive when it was navy. A 1920s dress might render purple when it was black. MyHeritage introduced its colorization feature years ago and it remains one of the most popular implementations because it produces natural-looking skin tones quickly. Dedicated services like ColorizeThis.io focus specifically on this step, letting you upload a repaired grayscale image and receive a colorized version without installing software.
Treat colorized output as an artistic interpretation, not documentation. If genealogical accuracy matters — say, identifying a regiment by uniform color — verify against written records rather than trusting the model. Some practitioners keep both versions: a cleaned black and white master plus a colorized interpretation clearly labeled as such. This dual-file approach is increasingly standard among family historians sharing results online.
One useful trick: colorize after repair, not before. Colorization models perform better on clean, high-contrast inputs, and scratches left in place tend to get colored along with everything else, making them more visible, not less.
Comparing Your Options: Tools, Costs, and Trade-offs
Choosing between tools depends on volume, budget, technical comfort, and how much control you want. Here is how the main categories compare:
| Feature | One-click web/mobile apps | Desktop suites (Photoshop, ON1) | Open-source DIY pipeline |
|---|---|---|---|
| Typical cost | Free trials; $5–$15/mo subscriptions | $10–$55/mo or perpetual licenses (~$100) | Free (GPU hardware cost only) |
| Time per photo | Under 1 minute | 5–30 minutes | Setup hours, then minutes |
| Control over output | Minimal | High (layers, masks, manual retouch) | High but technical |
| Best damage types handled | Scratches, fading, mild blur | Everything, including manual fixes | Whatever models you chain together |
| Risk of uncanny faces | Moderate | Low if used conservatively | Moderate-high depending on model |
| Batch processing | Often yes (Photomyne, PicWish) | Yes via actions/scripts | Yes, scriptable |
| Privacy consideration | Uploads to third-party servers | Local processing | Fully local |
Common Mistakes That Ruin Restorations
The most frequent error is over-processing. Running multiple enhancement passes — denoise, then sharpen, then upscale, then colorize, then upscale again — compounds artifacts until skin looks plastic and fabric looks melted. Apply each correction once, at the lowest effective strength. If a result looks obviously artificial at arm's length, dial it back.
Second mistake: destroying the original. Always archive the raw scan untouched. Several reviewers and archivists emphasize this because AI tools occasionally produce worse output than the input, and there is no undo once you've overwritten the source file. Keep a folder structure like /originals, /restored, /colorized so versions stay separated.
Third mistake: trusting AI faces blindly. Generative models reconstruct faces from statistical averages, meaning a badly damaged portrait may come back looking like a slightly different person. Compare restored faces against other known photos of the same individual. If features don't match — eye spacing, nose shape, jawline — the model invented them, and the image should not be presented as an accurate likeness.
Fourth mistake: ignoring resolution math. Upscaling a tiny 2x3 inch print scanned at 300 DPI to poster size multiplies guesswork. A 600 DPI scan of that same print contains four times the pixel data, giving the AI far more real information to work with. Scan properly first; upscale second.
Fifth mistake: colorizing without labeling. Sharing a colorized photo online without noting it's AI-colored misleads relatives who may treat invented colors as historical fact. A simple caption solves this.
When to Act: Timing, Preservation Urgency, and Prioritization
Physical degradation is ongoing, so digitization has genuine urgency. Prints from the 1880s–1950s era, especially those on acidic paper or stored in attics and basements, lose quality every year. Emulsion cracking and silver mirroring are irreversible — no software fully reverses information that has physically vanished from the print. Scanning now preserves whatever remains. If you have inherited boxes of photos, prioritize the oldest and most fragile items first, and photograph them even if you don't have time to restore them yet; a quick phone capture beats waiting for ideal conditions that never arrive.
For the restoration work itself, timing is flexible. Tools improve yearly — comparing 2026 models against those reviewed in 2024 shows measurable gains in face handling and Ultra HD output — so non-urgent projects can wait for better technology. But don't wait indefinitely on the scanning side. A reasonable plan: digitize everything within the next few months, restore the ten most meaningful images immediately, and revisit the rest as tools improve.
Seasonal timing also matters practically. Family gatherings around holidays are the best moments to identify who appears in unlabeled photos and to collect higher-quality originals from relatives. An unidentified face restored beautifully is still an unidentified face.
Cost Breakdown: What Restoration Should Actually Cost You
Budget options dominate this space in 2026. Free tiers exist on most web apps with watermarks or limited credits. Paid plans cluster between $5 and $15 monthly — Photomyne, PicWish, and similar apps sit in this band, typically unlocking unlimited restorations, Ultra HD exports, and ad removal. MyHeritage bundles photo tools into broader genealogy subscriptions. Professional desktop software ranges from subscription pricing around $10–$26 monthly (Adobe) to perpetual licenses near $100 (ON1 Photo RAW). Photoshop's built-in AI features require a Creative Cloud subscription but include generative fill credits.
Professional human restoration services charge $30–$150 per photo for standard work and $200–$500+ for severely damaged or historically significant pieces. Given that AI handles perhaps 80% of typical cases adequately, paying a professional makes sense mainly for: museum-grade archival work, photos destined for publication, severe damage crossing faces, or when legal/estate accuracy matters.
The hidden cost worth weighing is time. Restoring 300 photos manually at 20 minutes each equals 100 hours. Batch-processing the same set through a mobile app might take two evenings. Match the method to the stakes of each individual image rather than applying the most expensive approach uniformly.
Final Recommendations
Start by scanning everything at 600 DPI and archiving untouched masters. Run your bulk collection through a reputable one-click app for quick wins, keeping strengths conservative. Escalate the handful of photos that matter most — wedding portraits, the only surviving image of a grandparent — to a desktop workflow where you control every adjustment, or to a specialist service if damage is severe. Add colorization last, label it honestly, and preserve the black and white version alongside it. Check faces against known references before sharing results. Done this way, AI becomes a powerful assistant rather than a replacement for judgment, and your family's photographic record survives both physically and digitally.