AI vs Traditional Photo Restoration: What Changes Once You See Both Outputs

What a retoucher actually does

Comparing ai vs traditional photo restoration only makes sense once you know what the manual route involves. A professional working in Photoshop does not wave a filter over a scan. They spend most of the job on three things, and none of them are automatic.

• Inpainting by hand. A missing corner from a torn print gets rebuilt with the clone stamp and healing brush, borrowing texture from elsewhere in the image and then blending it. The restorer decides what goes in that hole, and can explain the choice afterwards. See water damaged prints for the kind of damage that lands here first.
• Tonal and channel correction. Silver mirroring and magenta casts are corrected per channel before anything is sharpened. Doing it in this order is why colour correction is listed as an early step rather than a finishing touch.
• Printing intent. Working to a known output, whether that is an A4 print or a screen, changes how far sharpening goes. Sharpening for screen and then printing is how you get halos.

That is why manual work is quoted in hours rather than per megabyte. A clean-up pass on a slightly dusty scan might take twenty minutes. Rebuilding a face where the emulsion has lifted can take several hours of the same careful borrowing, and it is priced accordingly. The cost side is laid out in free versus paid.

What the model does instead

A restoration model has seen millions of damaged prints paired with undamaged ones. It infers a plausible clean image. It is genuinely excellent at the common failures: dust, light scratches, mild fade, low contrast, slight softness. Those are statistical problems, and statistics solve them well.

It is weaker at anything that has one right answer. A missing half of a face is filled with a statistically average face, which can look nothing like your grandmother. A uniform gets a plausible colour that may be the wrong regiment. This is the same failure mode written up for colour in how accurate colourisation is: plausible is not the same as correct, and when the picture is someone's mother, plausible is not good enough on its own.

The second practical limit is repeat passes. Running enhancement twice rarely improves anything; skin goes waxy and fine detail turns to plastic. One clean pass from a good scan beats three stacked attempts, which is why a single upscale is usually the last step rather than a first one.

Where each one breaks

• Surface scratches and dust: AI wins outright. A spot healing session can take an hour; the model does it in seconds and usually better.
• Fading and flat contrast: AI wins. Curve correction is mechanical, and a retoucher doing it by hand is spending billable time on something an automated pass handles.
• Water damage and emulsion loss: Manual wins. Filling a void is a judgement call about what should be there, which is exactly what inference gets wrong.
• A badly damaged face: Manual wins, and it is not close. This is the case that sends people to a professional in the first place, and it is where the price is easiest to justify.
• Two hundred prints from one album: AI wins. Nobody hand restores a shoebox, and nobody should pay for it. Triage instead: automated passes for most, real work for the twenty that carry the history. Scanning guidance is in album digitisation.
• Archival or client material: Whichever route keeps the original intact and lets you show your work later.

Time and money, per print

Numbers here vary a lot by region and by whoever answers the phone, so treat the ranges as shapes rather than quotes. Light automated work costs cents in credits and takes seconds. Manual retouching is quoted per hour at professional rates, so a simple dust clean-up lands low while a reconstruction climbs quickly, and turnaround runs from a day to several weeks depending on the queue in front of you. What people are actually willing to pay is dealt with in whether professional work is worth it.

• Free tier or credits: the right answer for most boxes of prints, especially when most of them need the same light pass.
• Hourly manual work: the right answer for one print where the face matters and nowhere else.
• Mixed: not either-or in practice. Most people end up doing the batch automatically and paying a person for the one that argued back.

The order that gets the best out of both

1. Scan first, and scan properly. Nothing downstream recovers detail the capture never had. Dust off the glass, and scan at higher resolution than you think you need.
2. Run the automated pass on the copy. Never the original file. This clears dust, flatness and light scratches in under a minute.
3. Look at it at 100 percent before doing anything else. Eyes, mouth edges and hairline are where invented detail shows up.
4. Send that file to a person if it still needs work. They are now starting from something cleaned rather than something filthy, which is cheaper for everyone involved.
5. Save versions. Original scan, automated pass, hand finished file. Three files, clearly named, so nobody has to guess which one is the record.

Keeping those versions separate also settles the awkward question later about what has been altered. The original scan is the evidence; everything after it is a treatment of that evidence. See archive organisation for a naming scheme that survives contact with a real family.

How to decide about one specific print

Ask one question: does this image need to be defensible, or does it need to look good? A photograph going into a family history along with dates and names is a record, and invented detail in it is a problem someone inherits later. A photograph being reprinted for a birthday is a keepsake, and plausible is plenty. The distinction is unpacked further in app versus desktop workflows, which covers the same split from the tooling side.

FAQ

• Is AI good enough to replace a retoucher? For dust, fade and light scratches, yes, and faster than hiring anyone. For reconstruction where parts of the image are gone, no. Those jobs still go to a person.
• Will the AI invent a face? It will complete one. Missing features get filled with something statistically average, which may read as a different person entirely. Check eyes and mouth at full magnification before you believe the result.
• Why does my second pass look worse? Repeated enhancement smooths skin into something closer to plastic each time it runs. One pass from a clean scan beats several stacked ones.
• Should I colourise before or after repair? After. Colourising damage means colourising the cracks and stains too. Sequence and failure modes are in colour accuracy, and this year's tools are compared in 2026 colour tools.
• Is there a middle option? Yes, and it is the common one: run the cheap pass on everything, then pay for hand work on the two or three prints you intend to frame.

← All articles