Short answer: modern AI colourisation is very good at faces, sky, foliage and natural light. It is unreliable for clothing colours, period dyes, and military regalia. Use the colourised version as a starting point and fact-check anything historically important — uniforms, flags, religious garments, brand colours of cars and signage. For 95% of family photos, the default result is good enough to share immediately.
What AI colourisation actually does
Colourisation models are trained on millions of pairs of B&W and colour photos. The model learns conditional probabilities — "if I see a face with this skin texture and this shadow pattern, it's probably this colour"; "if I see sky with this cloud structure, it's probably this blue." The output is a plausible guess, not a recovered ground-truth — the colour information genuinely isn't in the B&W original.
This is the most important thing to internalise: colourisation is interpretation, not restoration. Two passes of the same model on the same photo can produce slightly different colours, and both can be "right" in the sense of plausible.
Where AI colourisation is reliable
Faces and skin tones
This is what the models are best at. Skin tone, lip colour, hair colour, eye colour (when the eyes are sharp enough) — accurate in 90%+ of cases. The model picks up on melanin gradients, vein shadows, and lighting cues that humans can't consciously articulate. Trust this.
Sky, water, foliage
Equally reliable. Sky is sky, summer leaves are summer leaves, autumn leaves are autumn leaves. AI gets seasonal context from the rest of the scene.
Natural light direction and warmth
Morning light, evening light, indoor incandescent vs daylight — the AI reads these from the shadows and applies a consistent global warmth. Often better than amateur manual colourisation, where people forget to colour the highlights and shadows differently.
Where AI colourisation is unreliable
Uniform colours
A British army battledress is khaki. The AI might guess olive, sand, or muddy grey-green. Same problem for police uniforms, school uniforms, religious habits, sports kit. If a uniform is in the photo, plan to override the colour manually.
Period dyes
Victorian and Edwardian fabrics used dyes (mauveine, alizarin, indigo) that the AI's training data underrepresents. A photo from 1895 will often come back in 1950s colour-palette because that's what the training set was richest in. Same for 1920s Art Deco patterns and 1970s avocado-and-mustard interiors.
Brand and signage colours
Coca-Cola red, Royal Mail red, John Deere green, NYPD blue — all "factual" colours the AI doesn't know without context. If the photo includes a recognisable brand or vehicle, fact-check it.
Religious garments
Catholic vestments, Sikh dastār, Jewish tallit, Buddhist robes — colour is theologically meaningful and varies by region, denomination and liturgical season. AI defaults will often be wrong. Worth overriding.
Step-by-step: getting a good colourised result
- Start with a good B&W scan. Scan at 600 DPI minimum. Restore damage first using our restoration guide, then colourise — never the other way round.
- Run the default colourise mode. No prompt, no overrides. See what the AI returns.
- Audit faces first. If the faces look right, you've cleared the highest bar. If a face has gone yellow or green, the photo had unusual lighting; try once more.
- Audit clothing and uniforms. Use Creative Modifications mode with a prompt like "navy blue wool suit, dark red tie" if you know what the actual colours were.
- Sanity-check period and place. Was that wallpaper colour in fashion in 1932? Did your great-grandfather's regiment wear that shade? A 5-minute Google search saves a permanent inaccuracy.
- Keep both files. The B&W original is the historical record. The colourised version is the share-on-social version.
When you should not colourise
Two categories of photo we suggest leaving in B&W:
- Photos taken specifically as B&W art — Ansel Adams-style landscapes, mid-century portrait studio work, fashion photography intended in mono. Colourisation here destroys the original creative intent.
- News and documentary photographs where the B&W medium is the historical record — war photography, civil-rights-era documentation, mid-century photojournalism. Colourisation isn't wrong here; it's just a different image.
Where to start
Pick the most colourful B&W photo in your archive — a wedding, a holiday snap, a child in a bright outfit. That's the easiest result for the AI and the most rewarding for you. Run it through Photo Repair Lab's colourise tool with the 20 free credits we give every new account. Once you've calibrated your eye on an easy one, work your way toward the photos with uniforms, regalia and period-specific dyes.