DCGAN is initialized with random weights, so a random code plugged in the network would produce a totally random image. Having said that, while you might imagine, the network has numerous parameters that we could tweak, as well as the aim is to locate a location of these parameters that makes samples created from random codes seem like the training