77 tokens limit ?

[Cariboost]

Hello

[Edit] since posting this message I understood how to search correctly on GitHub. However I find the answer to be "token limit is on model side". However for the same model I use on A1111 with Google collab, I don't have this limit. And from what I read online this limit can be bypassed by the diffuser

I am using fastsdcpu on Android (btw I am using android 12 (on note 10+) even if compatibility says android 13 at least and it is working pretty well, even if quality is quite low but I didn't deep dive yet on why).

I am facing the limit of 77 tokens in the prompt : everything after 77 characters is ignored (terminal explicitly say that).

I checked on Google and it seem to be quite a common issue that most GUI like A1111 managed to solve. I tried to see if something was feasible on my side without coding but I couldn't find. So I was wondering if a way existed with fastsdcpu ? Or is it something wanted to avoid CPU overload

[rupeshs]

yes current default token limitation is 77.

[mykeehu]

Here are some solutions. I'm no programmer, but maybe someone will make a PR after this: https://github.com/huggingface/diffusers/issues/2136

[kotoji]

"compel" could be a good option, as introduced here:
https://huggingface.co/docs/diffusers/using-diffusers/weighted_prompts#stable-diffusion-xl