I spent some time using PaintsChainer, a ML tool to autofill colors based on a few starting choices and the results were... rough.
The problem with ML art, is that it lacks the ability to polish. 90% of the impact of art is in the the last 10% of work, where the artist meticulously refines the piece to turn it from a loosely colored sketch, into a cohesive and complete picture.
Many of the tricks and tools to create this polish are heuristics that are not quantifiable or teachable via image sets. They come from an understanding of the "Gestalt" of a picture, or what the "gestalt" should be and then doing the necessary steps to get it there.
The other problem is you can't teach an ML algorithm about the hidden volumes in art. Much of drawing/painting is about tricking the eye into perceiving volume when none exists. An ML algorithm can perceive volume and identity after color/light have been applied, because those categories carry data. However, an ML algorithm can't infer what color/light a circle should have to give it the correct volume/perspective. A circle can be a doorknob, a pie, a ball, an eye. Each needing different data applied to it, which the ML algorithm doesn't (and won't ever) have.
Personally, I'd welcome a tool to make painting easier. It'd be amazing. However, I don't think we are all that close to a machine creating polished artwork.