In the classifier-free guidance paper, the formulation is as follow
However, it is implemented in DALL-2 as
null_logits + (logits - null_logits) * cond_scale
According to the formula it should be
logits + (logits - null_logits) * cond_scale
where the logits are the conditional score and null_logits are the unconditional ones.
Can you please clarify the reason for this modification?
Thank you!
In the classifier-free guidance paper, the formulation is as follow
However, it is implemented in DALL-2 as
null_logits + (logits - null_logits) * cond_scale
According to the formula it should belogits + (logits - null_logits) * cond_scale
where the logits are the conditional score and null_logits are the unconditional ones.Can you please clarify the reason for this modification? Thank you!