The point is that the source black and white image is not truthful about skin colour. The film locks in a level of lightness but that lightness may be very wrong (depending on the red and blue sensitivity of the film, the colour of the light, the time of day, the print, whether a filter was being sued).

So if you colourise an image of someone who appears to be a light-skinned 1930s African-American with colours that appear to conform to our contemporary understanding of light-skinned Black people of our era, you might be getting it right, of course.

But you might be getting it quite, quite wrong, in a way that matters.

So we're using a color space that has two channels dedicated entirely to color, which is the only thing the model needs to learn.

The model doesn't need to touch the lightness channel at all, only predict the noised added to the color channels at train time.

At inference time, we start with a real lightness channel (b/w image), and initialize the color channels to random noise. The model iteratively denoises the color channels while keeping the lightness channel locked.