Colourise black & white photos automatically

A black & white photo still holds all its light and shade: what's missing is the colour. PicReviver uses a colourisation model (DDColor) trained on millions of colour photographs to predict what colour each part of the scene most likely had, then paints it back over the original tones. Because it has learned how the world usually looks, it gets skies, water, foliage, skin and everyday objects convincingly right, rather than adding flat or random tints.

• Old family portraits and vintage black & white prints
• Historical and archive photographs
• Monochrome scans you'd love to see in colour
• Faded pictures: first restore, then colourise
• Small images: upscale them larger too

Colourising is not restoration. A black-and-white negative never recorded colour, so nothing is being recovered: DDColor is predicting the most likely colours for the things it recognises. That prediction is remarkably good for anything with a natural range and unreliable for anything specific.

• Usually convincing: skin, hair, sky, grass, wood, brick, water, most food. The plausible range is narrow, so a plausible guess is often right.
• Frequently wrong: clothing. A dress that was emerald green and a dress that was burgundy can leave identical grey tones, and the model will pick whichever it has seen more of.
• Basically unknowable: a military uniform's exact shade, a football kit, a flag, a company livery, a car's paint colour, a painted front door. If it matters, that colour needs a human who remembers, not a model.
• Awkward: mixed lighting. A photo lit by both a window and a tungsten lamp can come back with skin tones that shift across one face.

Look at four things at 100% zoom, in this order:

• Edges. Colour bleeding past an outline (hair colour leaking into the background, a jacket colour crossing onto a hand) is the most common artefact and the easiest to spot.
• Skin. Patchiness, or a face where one cheek is warmer than the other, means the model was unsure. Faces are also where a viewer notices wrongness instantly.
• Your control objects. Find something in the frame whose colour you know for certain — a familiar coat, a post box, a brand's packaging, a national flag. If the known colour is wrong, treat every other colour in that photo as decoration.
• Consistency across a set. Colourise several photos from the same day and compare: if the same jumper comes out differently, you're looking at a guess, not a record.

The practical workflow: restore and sharpen first, then colourise, because cleaner edges give the colour model better-defined regions. Then ask someone who was there. A ninety-year-old who remembers that the coat was navy will beat any model, and the two together produce something worth framing.

Above all, keep the black-and-white original as the archive copy. A colourised print is an interpretation of a photograph, in the same way a colourised film is an interpretation of a film. It's lovely to look at and it should never be the only version that survives.