"a picture of a girl wearing glasses" / "a picture of a girl with red hair"
It might make sense to optimize those separately an then pick from the joint distribution
https://colab.research.google.com/drive/1lUfuFPDuK1qQInTKj4tGF5gmd7G9pLCA?usp=sharing notebook, based on nagolincs initial CLIP one
Works decently when going for one attribute that is reasonably common, not so well for more complex prompts or rare attributes
Lot of avenues for improvement, still #nc
this run worked nicely, huh
so it can work, if you hit the right spots (or maybe I was hallucinating yday?)
@halcy last row, third from the left
Though it may be Shiro Emiya instead, either way the data might be tainted
@elfi w... who?