DCGAN is initialized with random weights, so a random code plugged into the network would produce a completely random image. Even so, as you may think, the network has millions of parameters that we can easily tweak, plus the purpose is to find a location of these parameters that makes samples generated from random codes appear to be the education