← All guides · 2026-10-14 · Updated 2026-11-11 · PhotoUnfade team

How to Fix Scratched Old Photos Without Losing Detail

Pull a stack of family prints out of a shoebox and the most common injury you will see is the scratch: white hairlines from decades of sliding against other photos, long marks from a drawer, a cluster near the edges where fingers gripped. Scratches look simple, but they sit in different layers of the photograph, and which layer they sit in decides whether software can erase them cleanly or barely touch them. Here is how to tell what you have, how to digitize a scratched print without adding new damage, the repair order that actually works, and — measured on our own pipeline, not borrowed from marketing — what AI can and cannot do about scratches.

What kind of scratches does your photo actually have?

Three kinds, and they do not repair equally. Identify yours on the scan before you pick a tool, because the prognosis ranges from "disappears completely" to "needs a human retoucher."

One more "type" is not a scratch at all: scanner dust. A hair or fiber on the scanner glass draws a perfectly straight line down every scan of the session, and owners reasonably conclude the prints are covered in scratches. If the same line appears in the same position on every photo, clean the glass — the prints are innocent.

Why can you not fix scratches on the physical print?

Because a scratch is missing or displaced emulsion — the gelatin layer that carries the image — and no home technique puts emulsion back. Every physical "fix" removes material instead: rubbing, polishing or burnishing smooths the surface by sanding away a microscopic layer of the photograph itself, taking image detail with it and leaving a permanent glossy patch. The scratch becomes slightly less visible and the photo becomes slightly less of a photograph, forever.

The one safe exception is removing what was never part of the photo: loose dust, lint or surface grime. A blower bulb (the kind sold for camera lenses) is safe. Anything more — erasers, cloths, water, tape — risks the emulsion. Gelatin swells in water within seconds, and a damp wipe on an old print can lift the image layer clean off. If the print is genuinely valuable and physically damaged, that is a job for a conservator (the American Institute for Conservation keeps a referral directory); their tools are magnification, stable adhesives and Japanese tissue, not elbow grease. For everyone else, the rule is simple: the physical print gets cleaned gently and preserved properly — our preservation guide covers materials and environment — and the scratch gets repaired on the scan, where nothing irreversible can happen.

How do you scan a scratched photo without making it worse?

Flatbed scanner, clean glass, 600 dpi, TIFF, no pressure on the lid, and the scanner's own automatic repair features turned off. Each of those prevents a specific, common way people add damage while trying to remove it.

1. Clean the glass first. Microfiber cloth, a breath of moisture, then a dry pass. Do this before every session, not once per project — it takes ten seconds and prevents the phantom-scratch problem described above.
2. Never use a document feeder. Sheet-fed scanners pull prints through rollers, and rollers are scratch machines. Brittle, curled or cracked prints can be destroyed outright. Flatbed only — and if the lid would press hard on a cracked print, leave the lid slightly open and cover the scanner with a dark cloth instead.
3. Scan at 600 dpi and save as TIFF. 600 dpi gives repair software enough surrounding pixels to reconstruct each scratch from; TIFF keeps every pixel the scanner captured. The full settings and the reasoning behind them are in our digitizing guide.
4. Turn off the scanner's automatic "dust and scratch removal." On prints these are software blur filters that smear fine detail along with the dust — you lose eyelashes and fabric weave to save a few specks. Infrared-based dust removal (Digital ICE) works only on color film, not on silver-based black-and-white film and not on paper prints at all, so it cannot help here either. Capture the damage faithfully and repair it in software made for the job.

What is the right order to repair scratches digitally?

Damage first, tone second, sharpening last — always. Each step amplifies whatever the previous step left behind, so doing them out of order does not merely fail to help; it makes the scratch harder to remove than it was in the file you started with.

The failure mode is easy to see. Sharpening works by raising contrast along edges, and a scratch is nothing but edges. Sharpen a scratched scan and every hairline gains a bright halo on one side and a dark one on the other — the scratch is now more visible, and the halo pixels have overwritten the image data a repair tool would have used to close it. Correcting tone and color before removing the damage has a milder version of the same problem: it bakes the scratch's brightness error into the new tone curve. This ordering is not folklore; it is how our own pipeline is built. Color correction runs strictly before sharpening, precisely because sharpening first would amplify the edges of any color cast. And the sharpening itself is calibrated, not maxed out: in our published pixel measurements our two modes raise sharpness by +29% (subtle) and +77% (full), while overshoot artifacts — the bright and dark halos — actually fall to 8.6% and 7.4% of edge pixels, from 26.1% in the unprocessed original. Detail goes up; edge artifacts go down. The full method and numbers are in our face accuracy test.

How well does AI actually fix scratches?

Fine scratches: nearly perfectly. Thick, heavy ones: honestly, this is the weak spot — and we can say that with numbers, because we measured it on ourselves.

The good case is physics, not luck. A hairline scratch is a gap one or two pixels wide with the true image on both sides. Inpainting — the technique AI tools use to fill damaged regions — reconstructs what belongs in the gap from its surroundings, and when the gap is narrower than the texture it crosses (a coat, a sky, a wall), the reconstruction is essentially exact. This is the same reason thin tear lines close so well, as our torn photo guide explains. When we calibrated the aging process for our demo photos, we deliberately used five thin, faint scratches — because that is the damage class automated repair genuinely erases.

The honest case deserves equal airtime. In our aging experiments we found a hard boundary: AI restoration is strongest at reversing global damage — overall fogging, fading, yellowing — and weakest at heavy local damage sitting on top of the image. A photo we artificially aged with severe localized staining and heavy marks came through our subtle mode nearly unchanged: a measured pixel-level transformation of 2.8, against a baseline of 20.6 for photos that repair well. The model reads a thick, dark, confident mark as picture content — a tree branch, a shadow, a fold — and politely preserves it. The practical rule: the more a mark looks like it "belongs" in the photo, the worse automated repair does, and the more a human retoucher's clone-and-heal judgment matters. The full three-category measurement — decay, added damage, missing content — is in our measured guide to which photo damage AI can actually fix.

Scratches across faces sit in between. The line itself closes cleanly — but if the scratch destroyed tissue rather than merely marking it, what fills the gap is a statistical guess at a face, not the person. Our standing threshold from the tear guide applies here too: once missing material approaches 30% of a face, you are looking at invention, not restoration.

When should you stop and call a professional?

Four cases: when the emulsion is lifting or flaking (physical conservation first, scanning second); when the scratch network is so dense that every patch would be a guess; when the damaged area covers unique identifying detail — a badge, an inscription, a face — that a guess would falsify; or when the print is the only surviving original and its physical condition keeps worsening. A conservator stabilizes the object; a skilled retoucher then works from the scan with reference photos of the person. Our AI versus professional comparison lays out what each path costs and when the human is worth the money.

The bottom line

Identify the scratch type, scan it right at 600 dpi, and let software work in the correct order: damage, tone, sharpen. Fine scratches — the kind on most shoebox prints — are exactly the class of damage AI erases best, and they cost seconds, not hours. The test is free: PhotoUnfade gives you three restorations, no account, no credit card, full resolution, with the original always beside the result and a slider between them. Take the most scratched face you have, zoom to 100%, and check the scratch lines and the eyes. If a thick mark survives untouched, that is the honest limit described above — a limit, not a malfunction. If the scratches came with yellowing or fading, as they usually do, our yellowing and fading guide covers the tone half of the job, and the step-by-step restoration guide puts the whole sequence in one place. And if the real problem on your print is softness rather than scratches, our blur repair guide covers what sharpening can genuinely recover and where AI starts inventing detail.