Currently the configuration only allows users to affect the size of the style_dim, latent_dim and the size of the input. In some cases, users might need to increase the generator capacity (for example by increasing the number of bottleneck steps, or by decreasing the number of down-sampling steps).
Useful in cases where the discriminator wins?