Introduction
On paper, the GAN game ends in a perfect equilibrium. In practice, GAN training is famous for being fragile. Runs can collapse into producing the same image over and over, oscillate without ever settling, or blow up after looking fine for hours. A lot of GAN research has been about making training stable.
Think of two people learning to play tennis by playing only against each other. If one becomes much better too fast, the other never wins a point and can't learn. If they keep exploiting each other's latest weakness, they may circle around forever without either improving overall. GANs face exactly these problems.