What is the difference between a generative adversarial network (GAN) and a diffusion model?

The fundamental difference between a Generative Adversarial Network (GAN) and a Diffusion Model lies in their approach to generating images. A GAN consists of two neural networks, a generator and a discriminator, which are trained in an adversarial manner. The generator tries to create realistic images from random noise, while the discriminator tries to distinguish between real images and those generated by the generator. This adversarial process forces the generator to produce increasingly realistic images. The generator directly maps random noise to an image. Diffusion models, on the other hand, operate through a process of progressively adding noise to an image until it becomes pure noise, and then learning to reverse this process to generate images from noise. This involves two phases: a forward diffusion process where noise is gradually added, and a reverse diffusion process where the model learns to remove the noise step by step, ultimately creating an image. Unlike GANs that directly generate images, diffusion models learn to model the probability distribution of images by learning how to denoise. GANs are known for their speed but can suffer from training instability and mode collapse (generating a limited variety of images). Diffusion models are generally slower but produce higher quality, more diverse images and are more stable to train. In essence, GANs involve a competitive game between two networks, while diffusion models involve a process of learning to reverse a gradual corruption of data.