Restore and colorize old photos with AI, and what it invents

Somewhere in a drawer there is a shoebox: a wedding from the fifties, a school class with one face circled in pen, a beach in faded orange. Scanned, they look worse than they do in the hand, because the scanner faithfully records every scratch, fleck of dust and soft edge. The single-job image tools in the catalog can take a lot of that away. They can also, quietly, put things in that were never there. This post is the order we would work in, and where to stop.

Before any tool: scan well and keep the original

Every tool below works from the pixels you give it, so the scan is the ceiling. Scan the print, not a phone photo of it, if you can; wipe the glass; scan larger than you think you need. Then put the untouched scan somewhere safe and never overwrite it. Each step that follows makes a new file from the last one, and the original is the only version that has not been interpreted by anything.

Step 1: scratches and dust

Search the models page for "restore". Two tools come up that are built for exactly this. One takes the picture and nothing else, and its job is described as removing noise, scratches and blur from old photos. The other has a single restore mode aimed at dust and scratches, with an option to save the result as PNG instead of JPEG. Neither takes a prompt, which is the point: there is nothing to word wrongly, and they are told to repair, not to reimagine.

Run one, then lay the result over the original and flick between them. Dust specks and thin cracks should be gone. If a crease ran through an eye or a mouth, look there first; that is where a repair turns into a guess.

Step 2: grain and softness, gently

If the photo is still grainy or soft, there are separate tools for each. A denoise tool offers Normal, Strong and Extreme settings; start at Normal, because the stronger settings smooth skin and fabric into wax along with the grain, and old film grain is part of what makes a photo look its age. A sharpen tool offers modes rather than a strength: Portrait for faces, Motion Blur for a subject that moved, Refocus for a lens that missed, and a few more. Pick the one that names your problem instead of the strongest one.

Step 3: colour, if you want it

The colourise tool takes the photo and one choice of style: contemporary colour, vivid colour, black and white, or sepia vintage. The last two run the other way, from colour to monochrome, which is handy when a set mixes black and white prints with faded colour ones and you want them to sit together in an album.

Contemporary colour is the calmer of the two colour styles and usually the better start for portraits. Treat whatever comes back as a suggestion. A black and white print holds no colour at all; the tool paints what is typical, so grass comes out green and the sky blue, and a dress that was teal may come out navy. If someone in the family remembers the real colours, a general edit model can fix one thing at a time. Word it as an instruction with a keep clause, as in the guide to edit prompts:

Step 4: size, last

Enlarge only once the picture is clean, or the upscaler will enlarge the scratches too. The upscalers differ in how much they are allowed to add. One describes itself as preserving the original content without regeneration and offers a 2x or 4x setting; that is the safe pick for a face that has to stay that person. Another has a creativity slider where negative values stay closer to the source and positive ones add generated detail. For a family photo, keep it at zero or below: new detail on a stranger's portrait is decoration, but on your grandfather it is a different man.

What it costs

Each tool shows its price on the Generate button in the studio before you press it. The restore and colourise tools have no prompt to tune, so a retry usually means trying the other tool or another setting rather than rewording. A whole bench order still means four or five paid runs per photo; if you are doing a box of fifty, try the order on three photos first and see which steps actually earned their place. The post on the price per usable picture has the arithmetic.

What these tools cannot do

They cannot recover what is not in the scan. A face that is a few pixels wide, half a figure torn away, an eye lost under a crease: whatever fills that space is invented, and it can look perfectly convincing. Colour is always a guess, skin tones and fabric dyes most of all. Handwriting, dates and shop signs in the picture can be smoothed or altered, so check them against the original. None of the tools knows who anyone is or when the photo was taken. And the result is a new picture made from an old one, not a recovered negative: label coloured or heavily repaired copies as such before they go into a family archive, and keep the original scan next to them.

For cut-outs of a single person from an old group shot, see background removal or an edit prompt. And if you want a restored portrait to move, turning a photo into a video covers the next step, with its own limits.