I think the diffusion versus llm distinction is nonsense. I commented on the Cully Hill Boys post yesterday (and got downvoted today) because I'm genuinely impressed with the tech, and the fidelity and coherence that is currently achievable. Uncanny valley issues from previous model generations seem mostly gone, and the results can be "engaging." Looks like there's an actual process behind this kind of generative output, and the quality of the final product will by the way still depend on an "artistic" idea and vision, and technical execution, i.e. how much one invests in preparation (asset generation, prompting) and post-processing (editing, kill your darlings). Of course, there's also utter crap to be found on that higgs website, loads of it. There will always be low-effort slop, especially when the tech gets democratized like this, but in general I think the current sota is a big deal. It is unfortunate that it will probably cause the death of a large segment of a creative industry, if it hasn't already. We'll see how it plays out, but if my name was Boris, I'd probably say that "entertainment is largely solved."
Once again, “AI is bad at things I’m very good at, but good at things I am not very good at (or don’t care about).”