Higher parameter count and less face samples aren't mutually exclusive.

You can take a huge GAN and then use a single image of a face as a guide for navigating the latent space.