A guess about what colour the dress was

What the model is predicting

A colourisation model takes the luminance channel (the greyscale photograph) and predicts the two chrominance channels that would go with it. Working in a space like Lab, where lightness is separate from colour, means the original tones are preserved exactly and only colour is added.

It was trained by taking millions of colour photographs, discarding their colour, and learning to predict it back. So it has learned strong statistical associations: sky is blue, grass is green, skin is in a particular range, wood is brown, water reflects sky.

Those associations are exactly as reliable as the association is in the world.

What it gets right and wrong, predictably

Reliable: sky, foliage, grass, water, skin in the range the training data covered, wood, brick, sand, stone. These have narrow colour distributions in reality, so the prediction is usually close.

Unreliable: clothing, painted objects, vehicles, signage, flowers, furniture, and anything manufactured. A dress in a 1940s photograph could be any colour. The model produces a plausible one: often a desaturated blue, brown or muted red, because those are the average of the possibilities, and averages are what a model trained with a squared-error objective produces.

That last point explains the characteristic look. Colourised photographs are frequently undersaturated and slightly brown, because when the model is uncertain it hedges towards the mean of the plausible colours, and the mean of all possible dress colours is a muddy neutral. Models trained with adversarial objectives commit more strongly and are more often confidently wrong.

Systematically wrong in one direction you should know about: skin tone prediction inherits the composition of the training data, which as the skin tones lesson described has historically under-represented darker skin. Colourisation of photographs of people whose skin sits outside the well-represented range is measurably less reliable.

Free tools

DeOldify is the best-known open-source colouriser, available as a notebook you can run in Google Colab for free, or locally. It has two models: an “artistic” one that commits to more vivid colour and makes more mistakes, and a “stable” one that is more conservative. It also does video.

There are newer open models, and several image editors and phone apps have a colourise button that wraps one of them.

GIMP has a G’MIC colourise filter, and separately supports the hand method below.

Hand colourising, which is often better

The technique predates the models by a century and it produces a result you can defend.

1. Keep the greyscale photograph as the base layer.
2. Add a new layer set to Colour blend mode above it.
3. Paint flat colour onto that layer with a soft brush. Because the blend mode takes hue and saturation from your paint and luminosity from the photograph, all the tone and texture of the original shows through.
4. Use a separate layer per region (skin, clothing, background, hair) so each can be adjusted independently.
5. Reduce each layer’s opacity to 40 to 70 per cent. Fully saturated hand colouring looks like a postcard.

The great advantage: you decide, and you can be told. A family can say the coat was navy, the car was green, the house was whitewashed. That information exists in people’s memory and in written records, and a model has no access to it. A hand-coloured photograph informed by the family is a reconstruction based on evidence; a model-coloured one is a reconstruction based on statistics.

A sensible hybrid: run the model to get a base, then correct the regions where somebody actually knows, using Colour-mode layers on top.

The honesty question, which is unusually sharp here

Colourisation is popular, moving and widely shared, and it is also the operation most likely to be mistaken for a recovery. People see a colourised 1912 photograph and experience it as seeing the past in colour. What they are seeing is a greyscale photograph with a model’s guess painted over it.

This matters most for historical and documentary material. Historians have raised consistent objections: colourisation removes the viewer’s awareness that they are looking at a mediated document, it asserts specific facts (uniform colours, flag colours, skin tones) that may be wrong, and the confidence of the result discourages checking.

The positions people take:

• For family photographs, where the audience knows what it is and the emotional value is the point, most people consider it entirely fine.
• For museum, archive and educational use, the widespread convention is to present the original alongside, label the colourisation clearly, and document what the colours were based on.
• For anything asserting historical fact (a uniform’s colour, a person’s appearance, a flag) a guess should not be presented as a finding.

The practical habit, as elsewhere in this course: deliver both files, label the colourised one, and write one line about where the colours came from.

The limit

The photograph contains no colour information. None. It was never recorded, and there is no signal to recover: this is not like a faded colour print where a weak dye layer survives.

Everything the colouriser produces comes from outside the photograph, either from a model’s statistics or from your knowledge. That is the entire content of the operation, and it is why the honest description of the output is a colour illustration derived from a photograph rather than a colour photograph.