dcgan

DCGAN was the model that made GANs actually trainable on images by writing down a recipe of architectural rules. Before it, GANs on images were notoriously unstable; DCGAN (Radford, Metz & Chintala, 2015) found a set of convolutional-network design choices that reliably produced sharp 64×64 images and stable training, and those choices became the default backbone for years of later GANs.

The core rules: replace all pooling with strided convolutions (in the discriminator) and fractionally-strided / transposed convolutions (in the generator) so the network learns its own up- and down-sampling; use batch normalization in both networks, but not on the generator's output layer nor the discriminator's input layer (which destabilizes); remove fully-connected hidden layers for deeper architectures (a fully-convolutional design); use ReLU in the generator for all layers except a tanh output; and use LeakyReLU in the discriminator for all layers. The generator takes a 100-dimensional Gaussian z, projects and reshapes it into a small spatial tensor, then upsamples through transposed convolutions to the final image.

Beyond stable training, DCGAN demonstrated that the learned latent space is structured and semantically meaningful: interpolating between two z vectors smoothly morphs the generated image, and vector arithmetic in z works (the famous "smiling woman − neutral woman + neutral man ≈ smiling man" on faces). It also showed the discriminator's learned features are good for downstream classification, evidencing useful representation learning. DCGAN is now a baseline and teaching model rather than state of the art, but its conventions echo through nearly every later image GAN.