Faceswap Forum

Sometimes when training, frontal faces train faster than profile faces. In preview profiles are more blurry than frontal ones. I've made new faceset from those hard training profile faces to train only them more to catch up frontal faces. Does this harm model anyway?

I would leave some frontal faces in.

Most datasets will have a bias towards front facing faces, which is why these train faster. Having more difficult angles in your training set will definitely help though.