Guidance scale more than 1 gives non sensical results?
Thanks for the hard work, this is a really cool model. However, whenever I try and set the guidance to about 8 (which is standard), I get really bad results like this:
Guidance = 8:
Guidance = 2:
Is something broken with guidance scale with LCM?
Please check out the blog post for recommendations around guidance_scale:
https://huggingface.co/blog/lcm_lora
it says in the info to stay between 1 and 2 doesn't it?
@dipstik Yeah but is there a good explanation for why this is broken? For normal LCM guidance = 8 doesn't give broken results.
@asdaweqw12 Their report says that the model is already "trained" with CFG of w=7.5, so the model already emphasizes your text without two feed-forwards as done in previous CFG-based inferences.
haas anyone made it work with a1111?