DCGAN is the convolutional version of the original GAN. The original paper used multilayer perceptrons, and DCGAN replaces them with convolutional networks.
Design points
- it is an all-convolutional architecture, and never a fully connected one
- it uses batch normalization in most layers of both the generator and the discriminator
- pooling layers are replaced by strided convolutions, so the networks learn their own downsampling and upsampling
Everything about the two player game itself stays the same as in Generative Adversarial Networks (GAN).