First of all, let me make it clear that I have a very basic knowledge of computers & know nothing about machine learning, neural network etc.
I am trying Faceswap just for fun. The only thing I understood by reading different articles from various websites is that
Encoder - is something that reads a face
Decoder - is something that draws a face based on the information it gets from an encoder
So why is that Faceswap is learning to draw both from A->B & B->A, when in most of the cases (I am sure in more than 95% cases) we only want to do one way swapping. I understand that an application needs to read both the faces but why it has learn to draw both A & B faces.
So is it possible that Faceswap reads both A & B faces but it learns only to draw face B on face A?
If so will it result in faster training & less VRAM usage and is there a chance that Faceswap developers will include this option of one sided training?
Again, I want to end this post by saying that please forgive me if I am asking some stupid question but I think many non technical people like me would be having the same doubt.
Thank you.