Since augmentation is what makes #stylegan2ada so much better I keep wondering if adding a pseudo-depth augmentation might be worth a try? (This example from 2018 is not even using a real depth map)
Follow
Redid the old sketch with a neural monocular depth model. As long as one keeps the displacements within certain bounds it might work. Since I did not find any prior mention of using this pseudo depth for data augmentation I put it here under a creative commons licence.
Server run by the main developers of the project 
It is not focused on any particular niche interest - everyone is welcome as long as you follow our code of conduct!