I tried to get SD-XL to generate an image of a frog with its eyes closed. It refused. I even cranked up the attention on closed to an absurd level, and it seemed to get sassy with me.
I tried to get SD-XL to generate an image of a frog with its eyes closed. It refused. I even cranked up the attention on closed to an absurd level, and it seemed to get sassy with me.
The prompt, by the way, was
frog with (eyes closed:3).Did you try putting (eyes open) in the negative prompt instead? I find that when it doesn’t have a strong understanding of a compound phrase, it sometimes focuses more on the individual words. So, “eyes closed” may have been impeded by a stronger influence from “eyes”.
The problem here is that you have the token “eyes” with very heavy weighting, and it’s showing you eyes. Another way of thinking about it is…
What do you see when somebody closes their eyes? Eyelids