How to Restore and Colorize Old Family Photos With AI

Restoring an old photograph is the one AI editing task where the stakes feel personal. The photo is usually the only copy, the people in it may be gone, and the result is going to be shown to relatives who remember the original. That is worth doing carefully, and it starts before the AI is involved at all.

What restoration can recover, and what it invents

A restoration model does two different things and it helps to keep them separate in your head.

Repair is removing damage that sits on top of the image: scratches, dust, creases, stains, the yellow cast of a faded print, the softness of a slightly out-of-focus lens. In each case the underlying content is still partly there, and the model recovers it by filling the damaged pixels with what the surrounding pixels imply. This is reliable, and on a face it is often startlingly good.

Reconstruction is filling in content that is genuinely gone: a torn corner, a face lost to water damage, a half of the photo that was folded away for fifty years. Here the model invents. It produces something that fits the surroundings and looks like a photograph, but there is no information to recover, so the result is a plausible fiction. That can be acceptable for a background; it is a different matter for a person.

Colorization is a form of reconstruction. The photo never contained colour; the model adds what colour usually is. Skin, sky, foliage and wood come out plausibly because they are predictable. Clothing, cars, painted walls and eyes are guesses.

The tool in Kitana's Photo Restore offers four presets that map onto this: Full Restore, Fix Scratches, Sharpen Faces and Colorize B&W. The first three are repair; the fourth is reconstruction, and it is worth doing last.

Scanning: where most restorations are won or lost

The model can only work with what the scan contains. Most disappointing results trace back to a phone photo of a print taken under a ceiling light.

1. Use a flatbed scanner if you possibly can. A phone photo adds perspective distortion, uneven light, reflections and the phone's own sharpening. Every one of those is a signal the model has to work around.
2. Scan at 600 dpi. For a standard 10 by 15 print that gives a file large enough for every detail on the paper. Go to 1200 dpi only for very small prints.
3. Turn off every automatic enhancement. Scanner software loves to apply sharpening, dust removal and colour restoration. Each one destroys a little information the AI would have handled better.
4. Scan in colour, even for black and white prints. The tint of a faded print is information about how it faded, and a colour scan preserves it.
5. Clean the glass and the print gently. A soft dry cloth on the glass; a soft brush on the print. Never a wet cloth on an old print.
6. Keep the original scan untouched. Restoration is an interpretation. Save the raw scan separately from anything you generate.

If a phone is your only option, shoot in daylight from a window, never under a lamp, hold the phone parallel to the print, turn off any flash, and take several frames so you can choose the sharpest.

Order of operations

Do repair before reconstruction, and do the least destructive thing first.

• Fix Scratches or Full Restore first. This clears the damage so later steps are working on content, not on tears.
• Sharpen Faces if the faces are still soft after the first pass. It is a targeted step and it is better applied to a clean image.
• Colorize B&W last, and only if you want colour. Colorizing a scratched image asks the model to colour the scratches.

Each pass is a generation and costs one creation from your allowance, so a three-step restoration of one photo is three creations. For most prints, Full Restore in one pass is enough, and you only go step by step when the first result shows a specific problem.

Telling the model what you know

Restoration tools take a prompt, and it is the one place you can supply information the photo does not contain. Use it.

• "The dress is dark green" fixes the most common colorization complaint.
• "This is a 1962 photo taken in Istanbul" nudges period-appropriate colour and tone.
• "Do not change the faces" is worth stating explicitly when the faces are already sharp and you only want the damage removed.
• "Keep the border" if the print has a decorative edge you want preserved.

What you should not do is ask for improvements that are really changes: removing a person, straightening a smile, opening closed eyes. That stops being restoration and becomes a new photograph, and the family member looking at it will notice. The same distinction, between repairing what was there and inventing what was not, is the reason AI object removal fails in the situations it fails in.

Judging the result

Look at three things, in this order.

Faces first. Compare the restored face to the original scan side by side. The eyes, the mouth and any distinguishing mark should be the same. If the model has "improved" a face into a slightly different person, that is a failure, however clean it looks.

Then texture. A good repair keeps the grain and softness of a real print. A bad one produces skin like plastic and a background like a render. If the result looks newer than a modern photo, it has been over-restored.

Colour last. Ask whether it is plausible, not whether it is beautiful. Saturated, evenly warm colour is the model's default and is usually too much for a photo that was originally shot in flat daylight.

If two of the three are wrong, run it again from the untouched scan rather than trying to fix the output. If the faces are wrong, add "keep the faces exactly as they are" and run it again.

What to do with the result

The restored file is usually far larger than the print ever was, and a further pass through the upscaler gives you enough resolution to reprint at a bigger size than the original. How that step works and where it stops helping is in AI photo upscaling: fix blurry photos in seconds.

Two practical notes. First, the photo you upload is used to produce your result and for nothing else; the privacy policy covers what is stored and for how long. Second, an old family photo is exactly the kind of image you should keep a copy of somewhere that is not a phone. Restoration is a good moment to do that.

Photo Restore is one of the fourteen tools in the Kitana apps and uses the same allowance as the rest: two free creations a day, one every four hours.