I think I figured out the GPU problem.
I also tested a couple more models of the new type. But I still couldn't achieve the stylization as in my other images. This model has more capabilities, it understands more tags and characters, the dataset is much larger. But everything rests on the visual part (too bad.