Restoring Old Family Photos with Image-to-Image AI: What Works and What Doesn't

Jeson Wu — Thu, 12 Nov 2026 00:51:32 GMT

Last month I scanned a shoebox of prints from my parents' place: faded color shots from the 80s, a few black-and-white portraits with cracks across the faces, and some that were just badly out of focus. I wanted to fix them without spending a weekend in a photo editor.

Image-to-image AI turned out to be the right tool for most of them, with a few caveats worth knowing before you try it.

Why image-to-image instead of text-to-image

Text-to-image models invent a picture from a prompt. That's useless for restoration: you don't want a grandmother, you want your grandmother. Image-to-image models start from the photo you upload and keep the face, pose and composition, changing only what you ask for — removing scratches, bringing back color, sharpening detail.

A simple workflow

I used Image2Img, which I built for exactly this kind of job, but the steps are similar on most tools:

Scan at a decent resolution. 600 dpi for small prints is plenty. A phone photo of a print works too if you shoot straight on with even light and no glare.
Crop out the border and the scanner bed. The model does better when the whole frame is the photo.
Upload and choose a restore style. On image2img.net this is the "old photo restore" preset; you can also describe the fix in plain words, e.g. "remove scratches and dust, keep the original faces".
Compare against the original. Zoom into eyes, teeth and hands — that's where any model drifts first.
Iterate on the result, not the original, if one area still needs work. Small, specific requests ("fix the crease across the left cheek") work better than one big prompt.

Image2Img gives 5 free generations a day with no sign-up and no watermark, which was enough to work through a batch over a few evenings. Uploads accept JPG, PNG, WebP and HEIC up to 20 MB, and results are deleted after 2 days unless you save them.

What worked well

Dust, scratches and small cracks — near-perfect in most cases.
Faded color — the model brought back believable skin tones and sky without making everything oversaturated.
Mild blur — faces got noticeably sharper while still looking like the same person.

Where to be careful

Heavily damaged faces. If half a face is missing, the model has to guess, and the guess may not look like the person. I kept the original next to the restored version for these.
Text and signatures on the back or in the frame can come out garbled. Crop them out first.
Group photos with tiny faces. Crop each person, restore separately, and keep the full shot as-is.

Keep the originals

Whatever tool you use, store the untouched scans. A restoration is an interpretation; the scan is the record. I name files 1986-beach-original.jpg and 1986-beach-restored.jpg so nobody mixes them up later.

If you have a box of prints sitting somewhere, an evening with a scanner and an image-to-image tool gets you surprisingly far.