Having established how diffusion models are trained to predict noise within corrupted data, this chapter focuses on the inverse operation: generating novel data samples. We will examine how to start with pure noise, typically sampled from a Gaussian distribution $x_T \sim \mathcal{N}(0, \mathbf{I})$, and iteratively apply the learned reverse diffusion process to produce a clean data point $x_0$.

You will learn the step-by-step procedure for the standard Denoising Diffusion Probabilistic Models (DDPM) sampling algorithm. We will then introduce Denoising Diffusion Implicit Models (DDIM), a related but often faster sampling technique that offers different trade-offs between speed and sample quality. Finally, we will cover the practical implementation details required to build these sampling loops in code.