DCGAN is initialized with random weights, so a random code plugged in the network would generate a totally random impression. However, while you may think, the network has millions of parameters that we can tweak, and the intention is to locate a environment of these parameters that makes samples created from random codes look like the training