Generative modeling has undergone a major shift in recent years, moving from adversarial training paradigms to probabilistic, noise-driven approaches. For a long time, Generative Adversarial Networks (GANs) dominated the landscape of image synthesis and generative tasks. They produced highly realistic outputs and powered applications ranging from face generation to style transfer. However, despite their success, GANs came with fundamental limitations that made them difficult to scale, unstable to train, and hard to control. This is where diffusion models emerged as a more robust and scalable alternative, gradually replacing GANs in many state-of-the-art systems. At a technical level, GANs operate through a two-player game between a generator and a discriminator. The generator tries to produce realistic data samples, while the discriminator attempts to distinguish between real and generated data. This adversarial setup creates a minimax optimization problem that is notoriously unstable.…