Generators are recognizers. Like how speakers and microphones are the same mechanism, backwards. ImageNet was only a classifier for a couple thousand single-word labels. Dall-E used the same approach to walk an image toward those labels. That’s how you get “avocado chair.”

Yet no one is paying for those training data

They don’t have to, because training is transformative. It’s fair use. If permission is not a factor, why would money be required? Your rights are not a loophole. This is fine for the same reason you’re free to reference, parody, or quote commercial works guarded by flesh-eating lawyers.

Even a vegan model trained on bespoke data could reproduce trademark-infringing characters. If you can describe what Naruto looks like, the model only needs to know what anime means.