LCC said:
For each destination picture, does the trainer only use source pictures with a similar face angle?
Or is a free-for-all, with any image being chosen?
As per what I have observed if there are a face similar in angle in src and dst then it is pseudo independent of that. Basically it chooses pictures of both faces and learns to draw it simultaneously . later it just switches to creating a src face where it should create dst .
Kind of you learn two languages (src and dst facesets) and all alphabets of them ( here faces) and later just translate what written in second language sentence (dst faces) into your learned first language sentence making use of its alphabets which closely resembles to the meaning to what written in second .
So yes its pseudo independent as if there are no faces of similar angle in src like in dst then It will simply not know how to translate that dst face specific angle into prediction. Like there is no alphabet equivalent of that second language (dst) alphabet in first