DCGAN is initialized with random weights, so a random code plugged in the network would produce a very random image. On the other hand, as you may think, the network has countless parameters that we could tweak, and also the target is to locate a environment of those parameters which makes samples generated from random codes seem like the teaching