Question about Pose Encoder and the Add operation.

[Zheng-Chong]

Lightweight pose encoder is quite impressive !
I'd like to know if the denoising UNet is trainable to adapt to a pose-conditioned noise?
If not, is the pose features strong enough to guide the portrait generation ? Thanks