DCGAN is initialized with random weights, so a random code plugged into your network would crank out a totally random picture. Nonetheless, as you may think, the network has many parameters that we could tweak, along with the aim is to find a location of those parameters that makes samples created from random codes appear like the instruction data.