From what I understand though, generators always have an advantage because the generator is allowed to "see" the discriminator's gradients during training. [0]

>The model for the discriminator is usually more complex than the generator (more filters and more layers) and a good discriminator gives quality information. In many GAN applications, we may run into bottlenecks where increasing generator capacity shows no quality improvement. Until we identify the bottlenecks and resolve them, increasing generator capacity does not seem to be a priority for many partitioners. [1]

Put another way: GAN training ends when the discriminator can no longer meaningfully distinguish real from fake. By definition then, the best generator will have no useful discriminator that can distinguish its output from real data. (conversely, if you did have such a discriminator, you could use it to train a better generator)

[0] https://developers.google.com/machine-learning/gan/generator

[1] https://towardsdatascience.com/gan-ways-to-improve-gan-perfo...