You are right! There is an L1 loss but only for super-resolution. I am speaking of GANs that don’t do super-resolution but turn noise into realistic samples. I should have linked another paper that has an example of those but I couldn’t find one with a graph of generator and discriminator losses.
Thank you for pointing out the L1 loss I will edit the original question.
However, the adversarial loss is the same as in normal GANs. The difference is that L1 loss is being added to it (weighted sum actually). So I still think that having a good discriminator loss would mean bad generator loss even if we add the L1 loss. I still think there is a trade-off in the adversarial losses!
I did not include the graph of the L1 loss because I believe “Generator loss” is not the weighted sum of the L1 and adversarial loss in Figure 2. Because the L1 loss is also plotted in the original Figure 2.
The GAN I am training also displays this relationship between the generator and discriminator loss:
# disc_loss is WGAN-CT loss
# gen_loss output of discriminator when a fake sample is input
# disc_r output of discriminator when a real sample is input
# disc_f output of discriminator when a fake sample is input
epoch disc_loss gen_loss disc_r disc_f
0 -3.379324 0.885375 -3.702462 -0.813928
1 -3.07398 0.731019 -3.96317 -0.781391
2 -3.121166 0.364389 -4.529734 -0.299867
3 -2.89777 0.721051 -4.004071 -0.712604
As the generator gets better at fooling the discriminator (gen_loss decreases) the discriminator becomes worse at telling them apart (disc_f increases). I do not understand why the output of the discriminator is not put in a sigmoid or tanh either.