Fantasy novel covers are exceedingly easy to do because of the mountain of examples in the training data

It's also because they are basically nonsense (fantasy) so the results in the style look more plausible.

I tried using a few AI generators to get some basic placeholder images for products and it's utterly shit, it's clear that networks dont understand what they are generating. Like I tried to generate bycicles and I would get components sticking to ground, floating components, stupid proportions, visual artifacts.