Is there any improvement in accuracy if the activation function of styleGAN is changed? Or will it be less accurate?
It depends what you mean when you say accuracy. If you mean the accuracy of the discriminator then perhaps? You actually have to test the activation in order to find out. ReLU/Leaky ReLU and in some scenarios PReLU have been shown to work very well due to their properties.
However, it has been shown over and over again that The loss/accuracy of discriminator has nothing to do with image quality, so keep that in mind.
Do you mind further elaborating?