The negative takes are mostly correct in all the limitations they talk about, but what they miss is how amazing these things are despite these limitations. These things are remarkably simple and limited yet they can generate realistic photos of myself in places I've never been, wearing clothes I've never worn, doing things I've never done, with a quick text sentence. Or nearly pass the bar exam.
On top of that a lot of the limitations have straightforward ways to address, many of which are already in progress. It is going to get really interesting. StableDiffusion knows nothing about the images it is produces, its just repeated denoising with image targets. It doesn't really understand anything about your text either, its just matching up tags. But both of those things can easily change. Put a big language model in front of it to better understand text. Already variants of these image models have depth information. Next up 3d object information, maybe next models of physics so it can understand how things would actually work in the scene, and so on.
going to get wild.