In the adversarial scheme, we need to make sure the discriminator can actually tell between batches (or has a structure that is adequate for learning the difference between batches) before we go about trying to fool the discriminator with the autoencoder.