Comment by timnetworks

Generate something you like the style of, then run it back through a vision model and ask it to describe the image without describing the subject, append that to a batch of icon descriptions, easy custom icons. (figure out transparency at this step) Then ask it for an imagemagick script that will convert them into needed sizes and formats.

very manual yes, but gives a lot of control

[edit] the images seem to be made with a transformers model rather than diffusion (e.g. DallE3 vs MidJourney) which is mostly proprietary for now.