#conditioning-dropout

Question 1: Single choice

Classifier-free guidance needs a conditional noise prediction $\epsilon_\theta(z, c)$ and an unconditional one $\epsilon_\theta(z)$ at every sampling step. How does the original method obtain the unconditional prediction?

The trained conditional model is reused with a randomly chosen class label in place of $c$ 1
A classifier is trained on noisy images, and its gradient is subtracted from $\epsilon_\theta(z, c)$ 2
The conditional predictions for all classes are averaged at sampling time to form $\epsilon_\theta(z)$ 3
One network does both; in training the label is sometimes replaced by a null token4

#conditioning-dropout

Timed test