DCGAN is initialized with random weights, so a random code plugged in the network would crank out a very random impression. On the other hand, when you may think, the network has a lot of parameters that we could tweak, plus the purpose is to locate a placing of those parameters that makes samples generated from random codes look like the training