Neural nets can churn out color sets if you give them a loss that prefers high entropy but still respects human legibility. The real challenge is designing that loss so the output isn’t just random noise. If you want a style, encode constraints or a palette prior. It’s a neat experiment, but you’ll need to sift the chaos for something useful.

Sure, give it a regularization term that penalises extreme saturation jumps and a smoothness loss on the hue wheel. That’ll keep the palette from spiralling into a migraine. I’ll keep an eye on the loss curves and let you know if it starts diverging. Good luck with the doodles.