Question about Encoder in cLR-GAN

[z-linlinlin]

As introduced in the paper, we want the Encoder to recover the latent vector from the output image in the cLR-GAN (Conditional Latent Regressor GAN). But we only update parameters in Generator. Why did you design the network like this ?

[junyanz]

We don't want the generator and encoder to cooperate and cheat. It is likely that the generator can produce a hidden signal and encoder can decode it. Therefore, we only update one network in the implementation. Encoder is still trained to reconstruct the photo in the cVAE-GAN branch.