DCGAN

Radford et al. stabilized Generative Adversarial Networks by replacing pooling layers with strided convolutions and implementing batch normalization.

Deep Convolutional Generative Adversarial Networks (DCGANs) bridge the gap between CNNs and unsupervised learning. By enforcing specific architectural constraints (removing fully connected hidden layers and using ReLU activation in the generator), DCGANs solve the instability issues inherent in Goodfellow’s original 2014 GAN framework. This model excels at high-resolution image synthesis: it can learn a hierarchy of features from object parts to entire scenes. In the landmark 2015 paper, researchers demonstrated vector arithmetic on face samples (e.g., 'Smiling Woman' minus 'Neutral Woman' plus 'Neutral Man' equals 'Smiling Man'), proving the latent space captures meaningful semantic structures.

https://arxiv.org/abs/1511.06434