DCGAN is initialized with random weights, so a random code plugged to the network would generate a totally random graphic. Even so, as you might imagine, the network has millions of parameters that we could tweak, and also the intention is to find a placing of such parameters that makes samples produced from random codes appear to be the coaching f