What colourisation can infer, and what it cannot know
A black and white photograph does not hold colour waiting to be brought back out. It records brightness. Two objects that were entirely different colours but equally bright landed on the film as the same grey, and nothing in the print separates them afterwards.
So colourisation is not recovery, it is inference. A plausible colour is proposed for each region based on what things of that shape and context usually are. Often that is right, and the result is genuinely moving. The difficulty is that it looks exactly as confident when it is wrong.
Where inference is reliable
A guess is only as good as the range it is guessing inside. Skin, sky, grass, timber and brick all sit in narrow ranges, so a plausible answer within them is usually close to the true one.
Those regions also happen to occupy most of the area of most photographs, which is why a colourised result reads as convincing at a glance. The eye checks the faces and the sky first, agrees with both, and extends the benefit of the doubt to everything else in the frame.
Where it is simply inventing
The moment a region had a choice behind it rather than a nature, brightness stops constraining the answer and the result becomes fiction that happens to look like a photograph.
- Clothing. A dress has no natural colour and its grey tone narrows nothing down. This is invention every time.
- Paint. Doors, cars, walls and signage were decisions somebody made, and decisions are not recoverable from brightness.
- Anything that carried meaning. Uniforms, flags, ribbons, team colours. These are the details people care most about and the ones a model has least ability to know.
- Objects whose whole significance was that they were an unusual colour, which is precisely the case a plausible guess handles worst.
Say which parts are inferred
The answer to this is not better software, it is labelling. A colourised print shown next to the original, with the invented regions named, is honest in a way the colourised print on its own is not.
That matters most for family photographs, because they become the record. A grandchild who only ever sees the colourised version will remember a blue coat that nobody ever established was blue, and the grey original is no longer beside it to say otherwise.
So keep the untouched scan. Everything after it is a version, and versions should stay reversible.
When colourising is the right call
For a picture meant to be looked at rather than cited, it is often the better version to hang on a wall. Monochrome puts distance between a viewer and a face, and wanting that distance gone is a perfectly good reason.
For a photograph doing documentary work, where somebody may later ask what colour something actually was, it is the wrong version to circulate. Those two uses only conflict if one file replaces the other, which is an argument for keeping both and labelling which is which.
More guides
Questions
Can colourisation recover the original colours of a photo?
No. A black and white negative records brightness only, so the original colours are not present in the image to recover. Colourisation proposes colours consistent with that brightness, which is a different thing, and it is right far more often for skin and sky than for clothing or paint.
Why do colourised photos sometimes look obviously wrong?
Usually because a region had no constraint on it. Skin and sky have narrow plausible ranges. A painted door or a printed dress has none, so the answer is a guess, and a guess sitting inside an otherwise convincing image is exactly what the eye catches.
Should I colourise a family photograph?
Do it if you want a version to look at, and keep the untouched scan as the record. The risk is not the colour, it is the colourised copy quietly becoming the only version anyone still has.