No one's going to need to generate training data, because image generators will simply setup and run their own artist pages using LLMs to interact with their fans, and then weight their outputs by likes/clicks/web requests.

The idea that complete original "art" is needed to run image generators is a very temporary state of affairs that bootstraps the proof of concept, but is in no way necessary to the process. The entire process can be adversarially trained solely against human aesthetic values - we know this, because image generators are already a derivative technology of building image classifiers - i.e. they functionally are adversarial against a given textual description in a classifier.

The real limitation right now is speed: once we can diffuse at > 1 frame-per-second, data collection can be run solely from the consumer side.