Closed kovalexal closed 8 months ago
thanks a lot
Interesting and great job discovering this. Any idea what effect this change has on the image results? I'm guessing it meant that it always used the default value for guidance scale. But I am not sure what the guidance scale actually does. :)
@Arcitec the default should work, but you can also try other values
Thanks. Yeah I guess it's this Stable Diffusion parameter:
Default should indeed be fine and we won't see any major improvements with this fix. Good discovery though!
Edit: Looks like guidance_scale
might be their term for the famous "CFG" (classifier-free guidance) prompt adherence parameter. If so, I get it now. :P
Thanks. Yeah I guess it's this Stable Diffusion parameter:
Default should indeed be fine and we won't see any major improvements with this fix. Good discovery though!
Edit: Looks like
guidance_scale
might be their term for the famous "CFG" (classifier-free guidance) prompt adherence parameter. If so, I get it now. :P
yes, it is the original parameter of sd unet not ip adapter
@Arcitec, this parameter influences the whole generation output. It was proposed in Classifier-Free Diffusion Guidance as an approach to increase the quality of output generation.
Simply speaking, this parameters influences on how far your generation would be from the provided negative text prompt.
Hi @xiaohu2015!
I noticed that there was a missing
guidance_scale
parameter passing toStableDiffusionXLPipeline
fromIPAdapterFaceIDPlusXL
, so I added it to these pipelines.