DCGAN

DCGAN (Radford, Metz, Chintala) showed that carefully constrained convolutional designs make GANs both stable and expressive. The generator maps a latent vector to an image using fractionally strided convolutions, ReLU inside and tanh at the output, no dense layers. The discriminator mirrors this with strided convolutions, LeakyReLU activations, and no pooling, while batch normalization is applied across most layers to control covariate shift. With these choices and Adam optimization, the model learns hierarchical features, produces sharper textures than earlier GANs, and yields internal representations that transfer well to tasks like classification or object discovery. DCGAN set a practical recipe for image synthesis, inpainting, and unsupervised representation learning, and it influenced many later models that pushed GAN fidelity and stability even further.