Classifier free guidance for prior model

[macrohuang1993]

Hello everyone,
For the "DiffusionPriorNetwork" "forward_with_cond_scale" method:
null_logits = self.forward(*args, text_cond_drop_prob = 1., image_cond_drop_prob = 1, **kwargs)

Could someone explain why we also drop the image_embed via image_cond_drop_prob? I think "image_embed" should be kept even in the unconditional generation since that is the noisy CLIP image embedding, and the classifier free guidance should only drop the text_condition part for calculating null_logits.

Thanks!

[tengteng95]

@macrohuang1993 Hello, have you found the reason? I have the same question. Thanks