In plain words
An old black and white photo keeps how light or dark each spot was, but throws away the colour. A red sari and a green one of the same brightness come out as the same grey.
A colourisation model has studied a very large number of colour photos. It has learned that skies are usually blue, leaves green, skin within certain tones, and a road grey. Shown your grandparents' wedding photo, it paints in the most likely colours for every part of it.
The result can look striking. But remember what it is: a plausible guess, not a record. The model has no way to know the sari was really maroon or the car was really cream. Family memory, or other sources, can be the better guide for those details.
Everyday example
Colouring in a newspaper photo of a cricket match
Give a child crayons and a black and white newspaper photo of a cricket match. They will make the grass green, the sky blue and the ball red, because that is what they have always seen. For the crowd's shirts they will just pick something. AI colourisation does the same: confident on common things, guessing on the rest.
Where the example breaks: A child knows they are guessing. The AI does not say which colours are sure and which are guesses: every patch looks equally confident, even when it is wrong.
How it actually works
Colourisation is a classic image-to-image task. Because one grey image fits many colour versions, there is no single right answer to learn.
Split brightness from colour
Many models use a colour system called Lab: one channel for lightness (L) and two for colour (a and b). The grey photo already gives L, so the model only has to predict the two colour channels for every pixel.
Train on colour photos made grey
Training is easy to set up: take a million colour photos, turn them grey, and teach a convolutional network to bring the colour back. The original colours are the answer key, so nobody has to label anything.
Avoid the dull brown average
If a shirt could be red, blue or green, a model trained to minimise average error picks the middle: a greyish, washed-out colour. A 2016 method fixed this by treating each pixel's colour as a choice among 313 colour bins and favouring vivid, rarer ones.
Make results look real
Later tools added a GAN-style critic that judges whether the result looks like a real colour photo. The open-source DeOldify project (2018) used this to colour and restore old photos and film.
- Key paper
- "Colorful Image Colorization", Richard Zhang, Phillip Isola and Alexei Efros, 2016, from UC Berkeley
- Fooling people
- In that paper's test, people mistook its colourised photo for the real one in 32% of trials
- Trained on
- About a million colour photos from the ImageNet collection
- Where you meet it
- Photo restoration apps and tools that colour old photos and film
DeOldify's creator noted his model painted the Golden Gate Bridge towers in colours they likely never had and wrote that "historical accuracy remains a huge challenge".
Questions people ask
Are AI colourised photos historically accurate?
Not reliably. The AI picks colours that are common in modern photos of similar things, not the colours that were actually there. Skies and plants are often close. Clothes, cars, buildings and paint are guesses and can be wrong.
How do I colourise an old family photo with AI?
Scan it at good resolution, then use a photo app or website with a colourise feature. Check the result against what relatives remember about clothes and places, and keep the original black and white file.
Why do AI colourised photos look brownish or faded?
When the model is unsure, averaging many possible colours gives a muddy, desaturated result. Better models choose one vivid option instead of averaging, but uncertain areas can still look washed out.
Is colourisation the same as restoring a photo?
No, though apps often bundle them. Restoration repairs damage such as scratches, tears and blur, and upscaling adds sharpness. Colourisation only adds colour. Both involve the AI guessing what was there.
Was this clear?
1 is confusing, 5 is crystal clear. One tap, no account needed.
Spotted something wrong, or know a better example? Suggest a change