Colourising black-and-white photos with AI
A colourisation model does not know what colour your grandmother's dress was: it decides from millions of similar photos. This guide explains which colours are reliable, which are not, and how to use intensity so the result looks like a photograph rather than a painted postcard.
How the AI decides colours
DDColor, the model Airpic uses, learns the correlation between luminance, texture and context: skin, sky, vegetation, wood, asphalt or sea have statistically stable colours and come out right. Clothing, cars and painted walls are ambiguous: the model picks a plausible colour consistent with the scene, which may not be the real one.
Natural or vivid
Natural produces muted tones typical of period film: usually the most believable for portraits and streets. Vivid saturates more and works well for landscapes, markets and bright scenes. The intensity control blends the generated colour with the original photo: at 60–70 % the result keeps the look of an old photograph.
Before colourising: clean and sharpen
Colour relies on edges and textures. If the photo has grain or compression, run Denoise first or Restore (which already offers colourising as an option). A 300–600 dpi scan as PNG gives better boundaries between objects than a phone capture.
Step by step on Airpic
- Upload the black-and-white photo (sepia works too).
- Choose natural or vivid and set the intensity.
- Compare with the divider; if an object comes out the wrong colour, process again with less intensity or fix it afterwards in an editor.
- Download as PNG or JPEG.
FAQ
Does the AI know the real colour of clothes in an old photo?
No. It gets skin, skies and vegetation right; for clothing, cars and walls it picks a plausible colour. If you know the real colour, fix it afterwards in an editor with a colour layer.
Can a sepia photo be colourised?
Yes. The model works on luminance, so the sepia tint does not affect it; the result is a colour photo.
How long does it take and how much does it cost?
10 to 20 seconds when warm (the first model load can take a minute). Previewing is free; downloading uses the free download of the month or one credit from €0.099.