GAN training about faceswap-gan

[kikoqiu]

warppedA -> encoder -> decoderA <=>inputA
warppedB -> encoder -> decoderB <=>inputB
They share the same encoder. The encoder will transform the input face into a general description of a face, and the decoderA will translate the description back into face A.
When test time, original face B => general description of the face => decoderA => face A

[shaoanlu]

Model A knows little about person B. I say "knows little" because model A and B actually share the same encoder, so at least model A knows the encoding of faces of B, and vise versa. The trick is, in test time, we feed normal (non-warped) faces of person B into model A. Model A treats these faces as if they are warped faces of A, and thus transform face B into a fake face A.