DCGAN is initialized with random weights, so a random code plugged to the network would produce a completely random graphic. On the other hand, as you may think, the network has many parameters that we could tweak, along with the aim is to find a location of those parameters that makes samples created from random codes appear like the teaching fact