DCGAN is initialized with random weights, so a random code plugged to the network would create a completely random image. Having said that, when you may think, the network has countless parameters that we can tweak, and the objective is to locate a placing of these parameters that makes samples generated from random codes look like the schooling in