Restore Old Family Photos with AI: 6 Careful Prompts and an Honest Guide

Digitise the print first: phone scanning that works

Restoration only works with what the scan captured. The print should lie flat under even light, with no reflections and the phone parallel to it. Glare on a glossy print is lost detail that no editor can truly bring back.

A flatbed scanner gives flat, glare-free light by design; 600 dpi is a sensible setting for small prints. On a phone, Google's free PhotoScan app combines several captures to reduce glare, or use your normal camera with these steps:

• •Take the print out of plastic sleeves when safe. If it is stuck to glass, do not force it; photograph it through the glass.
• •Remove dust with a soft, dry brush. Never wipe a print with a wet cloth.
• •Lay it flat on a dark, matte surface in soft window daylight, out of direct sun, with overhead lights off.
• •Hold the phone directly above, fill the frame without digital zoom, tap to focus, and check the capture at full zoom.

What restoration can and cannot honestly recover

AI is good at fixing damage that sits on top of the picture and bad at recovering detail that was never captured. A scratch hides a little, and the area around it shows what probably belongs there. A face that was out of focus in 1965 has no sharp detail to find.

• •Usually fixable: dust, fine scratches, small stains, a crease line, uniform fading, an orange or magenta cast, weak contrast.
• •Fixable with care: a missing corner that showed only plain background, mild softness.
• •Not honestly recoverable: a very blurred or tiny face, an eye lost to a tear, faded handwriting, and the real colours of a black-and-white photo.

Why AI invents faces, and how to limit it

Generative editors redraw the image rather than patching pixels. Where information is missing, they fill the gap with what is typical in their training images, which for old portraits often means a smoother, younger, more symmetrical, modern-looking face.

You can reduce this, though not remove it. Ask for one repair per request, say plainly that faces must not be redrawn, tell the tool to leave unclear areas soft rather than invent detail, and ask it to keep film grain. Then compare original and result at the same zoom, checking eyes, mouth, ears and hands.

If a face changes, use a non-generative tool for that spot: the Healing tool in Snapseed, a free photo editor from Google, or the Heal tool in GIMP, a free, open-source desktop editor. These copy nearby texture instead of drawing new features.

Colourisation is an interpretation, not historical truth

A black-and-white photo records brightness, not colour. A red saree and a green saree of the same lightness become nearly the same grey, so the AI picks colours common in similar images. The result can look natural and still be wrong about every garment.

Skin can drift lighter than the grey tones suggest, so the colourise prompt ties skin to the original tones. If relatives remember real colours, add them. Keep the black-and-white scan as the record and label the coloured version.

Keep the original, and label what you changed

The scan of the print is the real record; every AI version is a redrawn copy, and may be smaller than your scan. Gemini's help page, for example, says images download at 1K resolution without a Google AI plan and 2K with one. Keep the master scan in two places and never overwrite it.

Label restored and colourised copies, for example 'Restored with AI; colours are guessed'. Google says Gemini images carry an invisible SynthID watermark, and OpenAI says ChatGPT images include C2PA metadata, but metadata can be stripped when files are shared, so a written label is still the most reliable.

• •Note names, place and year in a text file beside the scan.
• •Never use an AI-restored photo for official documents or identity purposes.
• •Store the original flat, away from damp walls, direct sun and sticky tape.