5 Diffusion Models: Reverse Diffusion
This chapter covers
- The mathematics and intuition behind Reverse Diffusion
- The role of the U-Net architecture in noise prediction
- The importance of time step conditioning in the denoising process
- Practical implementation of Denoising Diffusion Probabilistic Models (DDPMs)
In the previous chapter, we explored the Forward Diffusion process, which forms the foundation of Diffusion-based generative models. We learned how these models gradually transform structured data into unstructured noise through a series of small noising steps. Now, we turn our attention to the other half of this powerful framework: the Reverse Diffusion process.
Reverse Diffusion is the heart of generative capabilities in Diffusion models. It is the process that allows us to start with pure noise and progressively refine it into structured, meaningful data. This chapter will explore the mechanisms, mathematics, and practical implementations that make this seemingly magical transformation possible.
The U-Net architecture, originally developed for biomedical image segmentation, has proven to be remarkably effective for the task of noise prediction in Diffusion models. We will explore why this is the case and how U-Nets are adapted for use in Diffusion models.