But Flux.2 seems to read only the first paragraph

#6
by Tehom - opened

After a little experimentation it seems like Flux.2 only reads about a paragraph (used to be 77 tokens for Flux.1) but the prompt expander writes much more. Thus prompt upsampling generates these splendid elaborations that get lost in the ether.

That's the way it goes. Would be great if they'd tune the VLM more for this demo..

edit: I was never able to get it to work with the 'upsampling' with all of the two times I tried anyway.

Sign up or log in to comment