DCGAN is initialized with random weights, so a random code plugged in to the network would deliver a totally random image. On the other hand, as you may think, the network has many parameters that we will tweak, and the purpose is to locate a location of these parameters which makes samples created from random codes seem like the training knowledge