DCGAN is initialized with random weights, so a random code plugged in the network would crank out a very random impression. Even so, as you may think, the network has a lot of parameters that we are able to tweak, along with the aim is to find a setting of such parameters which makes samples generated from random codes look like the training detail