[crazygringo]

Very interesting... seems to basically learn that:

Faces -> some variety of flesh-colored from light to dark

Fabric/clothing -> blue

Sky -> blue

Vegetation -> green

Wood -> brown

Blank -> turquoise or tan

Small details -> fascinating variety of colors, but often a brilliant red

Which all seems fairly reasonable. For many things (like wood or skin) it seems accurate.

Obviously things like clothes come in such a variety of colors that there's simply no way at all to predict accurately, zero meaningful signal -- so if it settles on whatever the most common color is, it doesn't surprise me that would be blue.

[citnaj]

Author here. Yeah you're basically right. GANs vastly improve the situation though because being safe with "green for grass, blue for sky, brown as default" doesn't work in the Generative-Adversarial setting. The critic will assign lower scores if the generator keeps outputting brown. Now I'd think the generator would get more creative than going for blue constantly but that might just be a matter or more/better training (...?)

[wbhart]

Like other colorisers I've seen, it does seem to have non-uniform colouring of clothing, with bleed from surrounding areas into clothing being fairly common. That's seems slightly weird, as I am sure few training images would have that.