I think the clear analogy is copyright in photographs. There is no a priori reason to believe that copyright should apply to either photographs or model outputs. In the US the copyright act was extended by congress to allow for copyright in photographs. When this was challenged in Burrow-Giles Lithographic Co. v. Sarony, the argument was made that a photograph only mechanistically captures the world and doesn't involve authorship and therefore doesn't fall in the scope of copyright outlined in the constitution. The court disagreed and found there is sufficient authorship in photography through the choice of lighting, framing, camera settings, posing of subjects, etc.

Similarly, one could argue that there is authorship in the choice of topic, prompting, framing, etc when using a generative model. Look at some of the very complex prompts on Midjourney's example page. I've spent a fair amount of time trying to make images which are just as striking and failing. Just as in photography, I think an extension of copyright in this case would probably need to involve an act of congress rather than a development in case law (though I'm outside the US and not that familiar with the act and relevant case law).

If you haven't read the Zarya of the Dawn letter from the US copyright office, I think it's a very cogent explanation of where things currently stand. An artist generated images for a comic book. They get no copyright in the images because they didn't sufficiently have authorship in them. They do get copyright in the entire comic book which involves a particular selection and arrangement of images. https://www.copyright.gov/docs/zarya-of-the-dawn.pdf