Classifier free guidance for prior model

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!