How AI colourisation of old photos works

How can an AI colourise a black and white photo?

It predicts plausible colours from what it recognises: a face, a sky, grass, brickwork, fabric. It learned those associations by looking at a very large number of colour photos, not by reading some hidden colour data in your print.

When an object could be any colour, such as a dress, a car or a front door, the model picks the most likely option based on what it has seen. That is why skin and foliage often look right while a jumper may come out in a different shade from the real one. We look at this in more detail in the real colours of old photos.

Which model does Souvenotte use?

Souvenotte uses DDColor, an automatic colourisation model that came out of academic research. It runs at the very end of the restoration, once the photo has been cleaned and the faces rebuilt.

The order matters: colourising a grainy or blurred image makes colours bleed. Coming last, the colour lands on a clean picture. The steps, in order:

• gentle contrast recovery, on brightness only;
• 2× enlargement with Real-ESRGAN, which also cleans noise, grain, JPEG blocks and light blur;
• every detected face rebuilt with GFPGAN, then blended back with a soft edge;
• colourisation with DDColor, if the photo has no colour or you asked for it.

Why is the colour laid on the brightness of the photo?

To keep every detail of the restoration: the model only supplies the colour, while light, shade and line come from the restored photo. Rebuilt faces and recovered sharpness are left untouched.

Any image can be split into luminance (light and dark, which carries all the detail) and chrominance (the hue). A black and white photo is luminance only; colourisation adds the hue without redrawing anything else. If you are worried about what AI does to people's features, read does AI change faces?.

Why do the colours look a little soft?

Because Souvenotte deliberately lowers saturation to 85 % after colourising. Colourisation models can produce over-bright tones, skin especially, and slightly restrained colour tends to look more natural on an old photo.

If you want more punch for a print, the Studio offers five looks: natural, vivid, soft, sepia and black and white. Vivid lifts the colours without changing the restoration itself.

When does Souvenotte colourise automatically?

On the automatic colour setting, a photo is colourised only if it is black and white, sepia, or so faded that it is down to almost a single tint. A photo that still has its colours, like many prints from the 1970s, is restored without being recoloured.

When you upload, you pick one of three options:

What if you do not like the colours?

Redo the photo with the other option from its page: it is free, and a photo already unlocked in HD stays unlocked. You can keep one colour version and one black and white version.

There is no way to pick a colour by hand, say the exact shade of a regimental uniform or a bridesmaid's dress. If one particular colour matters, a retoucher can paint it area by area: see AI or a photo retoucher.

Which photos colourise best?

Sharp, well-exposed portraits and outdoor scenes with sky and greenery usually give the most natural results. Very dark photos, tiny faces in a crowd and unusual objects are harder.

Your starting file makes a big difference: a clean, large enough scan gives the model more to work with. Check what resolution to scan at before you begin.