Ryan Murdock/advadnoun here: Glad to see some interest in the project!
I've written a few follow-ups to this, with some public notebooks as well that produce qualitatively different results. If anyone is interested, it should be pretty easy to find the BigSleep method, which steers BigGAN in a very similar way to this, as well as the Aleph notebooks, which use the DALL-E decoder or Taming Transformers VQGAN to generate images with CLIP, depending on the version of the notebook.