DCGAN is initialized with random weights, so a random code plugged in the network would make a completely random impression. Even so, as you might imagine, the network has countless parameters that we could tweak, as well as the intention is to find a location of these parameters which makes samples generated from random codes appear like the train