In the same way a normal fractal is a recursive application of some drawing function, this is a recursive application of different generation or "recognition -> generation" drawing functions built on top of the CNN.
So I believe that, given a random noise image, these networks don't generate the crazy trippy fractal patterns directly. Instead, that happens by feeding the generated image back to the network over and over again (with e.g. zooming in between).
Think of it a bit like a Rorschach test. But instead of ink blots, we'd use random noise and an artificial neural network. And instead of switching to the next Rorschach card after someone thinks they see a pattern, you continuously move the ink blot around until it looks more and more like the image the person thinks they see.
But because we're dealing with ink, and we're just randomly scattering it around, you'd start to see more and more of your original guess, or other recognized patterns, throughout the different parts of the scattered ink. Repeat this over and over again and you have these amazing fractals!