I don't see the big deal with either DALL-E or Stable Diffusion. It's compositional value is nil. At best it's text-to-img mapping may help some with getting to

I don't see the big deal with either DALL-E or Stable Diffusion. It's compositional value is nil. At best it's text-to-img mapping may help some with getting to learn what different artistic styles look like, but I see it as the same dynamic as GitHub CoPilot. Garbage In, Garbage Out. The more it's used, the more it's weights will be skewed toward SD'ness, and to be frank, I'm not entirely convinced that there won't be rapidly discovered "verbatim" reproductions of things as outputs that'll turn into legal challenges a la license enforcement. Further, the user experience is extremely disappointing. One is quickly left feeling as if one's time may be better spent articulating one's vision enough such that a human artist can make it happen. It's much like what I call "the hair stylist" problem. The hard part around communicating is figuring out the hairstylist vocabulary. DALL-E may make learning that easier. Then again, my track record with "meh, never happens" has been abysmal lately.