Can AI remove all types of scratches from old photos?

AI can remove most surface scratches and light physical damage from photo prints. The model identifies scratches as linear anomalies that do not match the surrounding image content and fills them based on adjacent pixels. Deep scratches that have physically removed the emulsion layer may leave faint traces after AI restoration, as there is no image data to reconstruct. The AI can approximate what should be there but cannot recreate lost detail perfectly.

Is it better to remove scratches before or after upscaling a photo?

Remove scratches first, then upscale. If you upscale first, the scratches become larger and the AI has more scratch area to fill, increasing the chance of visible artifacts. Removing scratches at the original resolution produces cleaner results. After scratch removal, a 2x or 4x upscale sharpens the restored image without amplifying damage.

What is the difference between scratches on prints vs. scratches on scanned negatives?

Scratches on prints appear as white or light lines where the physical surface was damaged. Scratches on scanned negatives appear as dark lines. AI restoration handles both, but scratches on prints are generally easier to fix because the damage is on the surface layer. Scratches on negatives remove image data entirely and the AI must reconstruct what was lost.

How to Remove Scratches from Old Photos with AI

AI restoration removes surface scratches from old photos almost perfectly. The model identifies fine white scratch lines as damage and fills them seamlessly from surrounding pixels. Deep scratches that have physically removed the photo emulsion leave faint traces even after AI repair. For best results, remove scratches first, then upscale the clean image.

My grandmother kept the family photos in a shoebox. No albums. No sleeves. Just a cardboard shoebox where photos rubbed against each other every time someone opened it. After about sixty years of this treatment, every photo had scratches. Some had just a few light surface marks. Some looked like they had been dragged across a sidewalk.

Scratches are the most common form of damage on old photo prints, and fortunately they are the easiest for AI to fix. But the approach depends on the type of scratch, and getting this wrong wastes time. Here is what I learned from fixing scratches on about 200 photos from that shoebox.

The three types of scratches and how to handle each

Surface scratches are the most common. These are fine white lines where the top layer of the print was lightly abraded. They do not penetrate the image layer. They just scatter light and appear white. AI restoration removes these almost perfectly. The model identifies the fine white line as damage, looks at the pixels on either side, and fills the scratch with a seamless continuation of the surrounding image. I processed about 150 photos with surface scratches and the AI got every single one clean.

Deep scratches are less common but more problematic. These go through the emulsion into the paper below. The image layer is physically gone along the scratch line. A deep scratch appears as a white line with a darker edge on each side. The AI can still handle these, but the results are less perfect. It fills the scratch based on the surrounding image, but without the original image data, the fill is an approximation. On complex areas like faces or patterned clothing, a deep scratch may leave a faint visible line after restoration. On uniform areas like sky or plain backgrounds, the fill is invisible.

Pattern scratches are the rarest and most annoying. These happen when a photo was dragged across a textured surface, leaving parallel lines rather than a single scratch. Think of the marks a vinyl record makes on a needle. The AI tries to fill each line individually but sometimes the pattern confuses it. The model sees multiple parallel anomalies and does not know whether they are damage or part of the image. For these, I had to run the AI restoration twice. The first pass removes the most obvious lines. The second pass cleans up the fainter ones that remain.

The right order of operations

I made the mistake of upscaling scratched photos first. Do not do this. When you upscale a scratched photo from 600 pixels to 2400 pixels, the scratches get bigger too. A one pixel scratch becomes a four pixel scratch. The AI now has to fill four times as much area, and the fill quality drops noticeably.

The correct order is scratch removal first, then upscaling. Remove the scratches at the original resolution. The AI fills one or two pixels of damage with high accuracy. Then upscale the clean image. The upscaler works with clean data and produces a sharp, scratch free result.

When to use manual cleanup instead of AI

For one or two prominent scratches on an otherwise clean photo, manual cleanup in an image editor might be faster than running AI processing. The healing brush in Photoshop or GIMP handles single scratches perfectly. Click, drag, done. Five seconds per scratch.

AI scratch removal becomes worth it when you have many photos with many scratches, or when the scratches are fine and numerous. The shoebox photos averaged about twenty to thirty fine scratches each. Manual cleanup would have taken maybe five minutes per photo. AI processing took about fifteen seconds per photo. For 200 photos, that is the difference between sixteen hours of manual work and about fifty minutes of letting the AI run.

That shoebox of scratched photos is now a clean digital archive. Every photo in the family has a copy. The scratches are gone. All it cost was time and a few credits per photo. If you have a shoebox of your own, start with the scratched ones. They are the easiest to fix and the results are the most dramatic.