It's interesting to see that Latent Diffusion + LAION 400M does much better with text, it's actually a challenge sometimes to get it to not write the parts of the prompt meant to guide the style on the resulting image.
Try it yourself here: https://colab.research.google.com/github/multimodalart/laten...
EDIT: Here's 36 uncurated renders for "Hacker News - Dang's Dance. R&B Album Artwork" https://johan-nordberg.com/tmp/hn-laion400m.png