Conditioning Dropout for Classifier Free Guidance
#1
by
pyone
- opened
How about adding conditioning dropout (let's say 10%) in training just like original stable diffusion?
Thank you Justin and lambdalabs for sharing the model. It's really awesome.
Would it be better for conditioning dropout? Or there is no significant difference for fine-tuning.
I would appreciate if you could share your take on it.
I found the dropout in FrozenCLIPImageEmbedder class. Now I know image conditioning also follows text conditioning dropout. Thank you anyway.
pyone
changed discussion status to
closed