DCGAN is initialized with random weights, so a random code plugged into the network would deliver a totally random image. However, while you might imagine, the network has millions of parameters that we could tweak, as well as the goal is to locate a setting of such parameters that makes samples created from random codes appear like the instruction