This isn’t completely accurate. The generator sees the training data in the same way supervised learning might, because the discriminator sees the data, and the

This isn’t completely accurate. The generator sees the training data in the same way supervised learning might, because the discriminator sees the data, and the generator shares gradients with the discriminator. Your point stands though, that it’s obviously not overfitting, and no scientist would publish a result that was just overfitting face generation.