DCGAN is initialized with random weights, so a random code plugged to the network would deliver a totally random image. On the other hand, as you may think, the network has millions of parameters that we will tweak, and the purpose is to locate a environment of such parameters that makes samples created from random codes seem like the teaching info