DCGAN is initialized with random weights, so a random code plugged to the network would deliver a completely random graphic. However, while you may think, the network has a lot of parameters that we are able to tweak, plus the target is to find a setting of those parameters which makes samples generated from random codes seem like the teaching info