In StarGAN, the D model has 2 different output, judging the image from dataset or G and predicting the domain vector. My question is that since we have the predicted domain vector to direct the G, why we need the D to judging Ture or False? Or in other words, how, in StarGAN, T/F judgment can help the whole model capture more feature information than the n dimension domain vector?
I guess the T/F judgment is the essence of GAN model which build a bridge between D and G and keep the dynamic game. And discrete domain vector will make D concentrate on the specific image features while the T/F judgment score gives D the maximum degree of ‘freedom’ to capture the feature information.