this is one of my experiments for colorizing graphical images - it uses a concept I call "color graphs", or color palettes in the form of undirected, weighted graphs. The main motivation for this is to abstract away color design from the form or structure of the image.

some more details on how it works server-side:

- the input color graph is matched with an existing design from a dataset

- nodes on the graph that misalign with the input are knocked out and re-generated image-completion style with a GAN

I think it works pretty well, but like a lot of my experiences with ML projects it's not 100% accurate.