/ic/ - Artwork/Critique

>>6869541
that's just not true.
see pic related.
i wanted to get this, i used a specific model because i knew it would get me a certain look, same with the keywords. same with how i configured the two cnets i used. you absolutely can follow a vision to some degree.

you're not always throwing shit at a wall because if you experimented a lot with the model, you KNOW how it will respond to certain things.

>and we had this argument countless times already here.
yes and you're still wrong. in fact i've been trying to learn things more in depth recently in order to explain it better, but it's pretty hard to understand, especially the cross-attention stuff where all this actually matters.

but the AI doesn't learn just images, it learns images tied to words. and when you prompt it, you're using those same words. words that are not tied to any image, but to thousands if not millions of images.
in the end, the word you prompt is simply the models representation of all the training data. not any particular image, or even any particular part of an image, but but a far more wholistic understanding of the word and what it represents.
that, on its own is already plenty of proof that the AI is learning.