DCGAN is initialized with random weights, so a random code plugged into your network would create a totally random picture. Even so, as you might imagine, the network has millions of parameters that we can tweak, as well as aim is to find a setting of these parameters that makes samples produced from random codes appear like the training details.