DCGAN is initialized with random weights, so a random code plugged into the network would create a completely random image. However, when you may think, the network has many parameters that we are able to tweak, as well as the aim is to find a location of such parameters which makes samples generated from random codes look like the schooling facts.