Generative Adversarial Networks (Goodfellow et al. 2014) train a Generator G and Discriminator D in a minimax game: min_G max_D E[log D(x)] + E[log(1 − D(G(z)))]. At the Nash equilibrium, G captures the true distribution and D outputs 1/2 everywhere. In practice training is unstable: if D is too strong, G gets no gradient; if G is too strong, D cannot guide it. Key stabilizations: Wasserstein GAN (WGAN) replaces JS divergence with Earth Mover distance for smoother gradients; spectral normalization constrains D's Lipschitz constant. The visualization shows the loss surface and generator distribution evolving toward the real data manifold.