If we are going to go technical I would argue that artists generate a highly derivative work, i.e. they input the "training data" via their eyes, experience of the world, ruminating (is that the word?) about art... and then output motions of their hands, words, and so on which create new artistic work.

Diffusion models tune a series of numbers via a complicated but essentially straightforward and "pipeline-like" process from the training data, so their work is significantly less derivative, even if visually it looks different. I guess we all agree that say, a encrypted zip file of a song is not a derivative, it's the song after some data processing.

Maybe we should improve the definition of derivative work to express how many layers of indirection lie between training material and output? If I trace over a painting I guess most would agree I've barely done any art, and I think that the process is more or less "pipeline-like" (even if complicated): input data is immediately converted to "art" by my brain following the lines. An artist that studies Goya for years and slowly improves his Goya-like portraits is another kind of level of indirection between training material and output!