Open FlashlightET opened 1 year ago
Example:
input: nacho cheese
output: cheese trail, potato wedges, orangette, fast food, oreo, ikura \(food\), food on breasts, food on penis, kiwi slice, cat burger, chocolate cornet, loaf of bread, mint chocolate chip, drink, cream puff, ujikintoki, grilled corn, swiss cheese, cookie, food wrapper
I can add explicit as a negative prompt but sometimes that just causes it to no longer have a clue what its trying to generate
Full model seems to be more specific and less horny
9 times out of 10 it seems to generate porn prompts.
I think it is because much of the vocabularies in danbooru are such words.
The following histogram is the distribution of cosine similarities between tag descriptions and the sentence "sexual or genital".
And the Negative needs more "explicit" words because it is tokenized by "general" text tokenizer; it is interpreted as general English text which is not including the meanings only in danbooru.
Full model seems to be more specific and less horny
This is histogram for "strict" model (previous one is for full model).
Full model is including the tags related to works and characters, and I think they dilutes explicit words.
(frequency between 0.0 and 0.15 seems to decrease in strict model, so the general words may have these similarities.)
(frequency between 0.0 and 0.15 seems to decrease in strict model, so the general words may have these similarities.)
9 times out of 10 it seems to generate porn prompts.