Comment by Scene_Cast2

I've noticed that image models are particularly bad at modifying popular concepts in novel ways (way worse "generalization" than what I observe in language models).