DCGAN is initialized with random weights, so a random code plugged into your network would make a totally random image. Nevertheless, as you may think, the network has a lot of parameters that we could tweak, along with the aim is to find a location of those parameters which makes samples generated from random codes appear like the instruction info