An introduction to diffusion models
In this section, we will explore diffusion models. We will compare them to Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs), which we covered in Chapter 7. This will help you to gain a holistic and comprehensive understanding of generative models. Additionally, it will make comparing and contrasting the architectures, training procedures, and data flow of these methods straightforward. Furthermore, we will also learn how to train a typical diffusion model.
Diffusion Models (DMs) are generative models that were recently proposed as a clever solution to generate images, audio, videos, time series, and texts. DMs are excellent at modeling complex probability distributions, structures, temporal dependencies, and correlations in data. The initial mathematical model behind DMs was first proposed and applied in the field of statistical mechanics to study the random motion of particles in gases and liquids. As we will see later...
