Colourising old photographs: what works and what does not

There is a photograph of your grandparents on their wedding day. Black and white, a little faded, one corner torn. You have looked at it a hundred times. And yet, when you see it in colour, something happens. The dress takes on a shade, the sky becomes sky, the faces come closer. It is not a miracle. It is an estimate. That difference is what this article is about.

What colourising actually does

A black and white photograph contains no colour. It contains light and dark. A program that adds colour looks at the shapes in the picture and compares them with millions of other pictures it has seen. Grass is usually green. A sky is usually blue. A face usually has a skin tone within a certain range. From that, the program guesses, patch by patch, which colour most probably belonged there.

So it is an estimate, not a memory. The program does not know that your grandmother's dress was pale blue rather than white, because blue was cheaper to dye at the time. It does not know that your father's bicycle was red. It only knows what is usually the case. If you keep that clearly in mind, you can enjoy a colourised photograph without being misled by it.

Start with a good scan

The result is never better than the scan you begin with. A print photographed with a phone, under a lamp, with the shadow of your hand across it, gives you a colourised picture with that shadow in it. So take your time over the scan.

• Resolution: 600 dpi is a good measure for an ordinary print. Small pictures, such as passport photographs or contact prints, can go higher. At 300 dpi you lose detail in faces that you will miss later.
• File format: save the scan as TIFF or PNG. Those formats throw nothing away. JPEG is convenient for sharing, but every time you edit a JPEG and save it again, a little disappears.
• The master: make one scan first that you never touch. That is your master. Everything you do afterwards, repairing, colouring, cropping, you do on a copy. That way you can always go back.
• Clean glass, no correction: wipe the scanner glass and switch off automatic enhancements. You want the photograph as it is, not as the scanner thinks it ought to be.

If you have no scanner, a phone will do, with limits. Lay the print flat near a window, out of direct sunlight, hold the phone straight above it and turn the flash off. It is not the same as a 600 dpi scan, but it is better than nothing, and sometimes nothing is the alternative.

Where it goes wrong

There are three places where colourising tends to go astray, and it helps to know them so that you recognise them.

Skin tones. A program picks an average skin. With a light-skinned family that often works reasonably, but the cheeks come out too pink, or everyone gets the same tone, while your uncle always worked outdoors and your aunt never did. With darker skin it goes wrong more often: too grey, too orange, or uneven across the face.

Uniforms and work clothes. Here the estimate is most dangerous, because here a colour means something. A program does not know which army, which period, which rank. It often turns a uniform into a brown or green that never existed. The same goes for club colours, a school uniform or a nurse's apron.

Fabrics and patterns. A checked dress, floral wallpaper, a knitted jumper: the program sees light and dark areas and fills them with what is usual. Red becomes brown, yellow becomes white, and a pattern that really had two colours ends up with one. Old interior photographs suffer from this too, because a room in 1955 had different colours from a room today.

And there is one more thing: a colourised photograph looks convincing. That is the difficulty. A wrong colour does not look like a mistake, it looks like a fact. Thirty years from now a grandchild may believe the dress really was blue.

Pass on what your family knows

The good news is that you have something the program does not: you know these people. Or you know someone who knew them. That knowledge is the difference between a guess and a grounded reconstruction.

Before you start, write down what you know. What colour was the front door? What colour was the car? Did grandfather always wear blue overalls? Was the wedding dress ivory, or actually pale pink, as your mother always said? Ask the people who still remember, and ask them now. This is the kind of knowledge that disappears with a generation.

Many tools let you point out colours: you click on the dress and say what colour it was. Do that where you are certain, and let the program estimate the rest. Where you do not know, a neutral, restrained colour is more honest than a bold choice. And keep the black and white scan next to the coloured version. The scan is the source. The colour is a note beside it.

Repair first, then colour

Colourising is usually not the first step. A torn corner, stains, scratches and fading all disturb the estimate. A scratch across a face is read by the program as a line that belongs there, and given a colour. So repair first, and colour afterwards.

If you want to try a single photograph, without installing anything or creating an account, you can do that through the restore door at thememorypalace.ai/restore. You give one photograph, you see the result, and you decide for yourself what to do with it. It is a quiet way to find out what is still in an old print.

What you get out of it

A colourised photograph is not a replacement for the original. It is an invitation. Children who walk past a black and white picture stop at one with colour. They ask who that is. And then you can tell them: that is your great-grandmother, on her wedding day, and that dress was pale blue, because she always said so herself.

That is how it works best: the scan as the source, the colour as an estimate, and your story as the thing that makes it true.