tHis tour is about a paper that started a new era of generative deep learning in computer vision and many other fields afterward: the era of diffusion models. Its titled “Probabilistic Denoising Diffusion Models” and introduces a new framework known as DDPM, short for the article's title.
While the general idea of diffusion models may seem intuitive, the mathematics behind it is not, and you may find it difficult to understand articles on that topic. At least I did. At the same time, many of the current generative models such as DALL-E3, Image, SORAand Stable diffusion 3 They are based on diffusion models. Therefore, it is important to understand the basics.
So buckle up because today is the day we will build a solid intuition on the basics of diffusion models. We'll put the DDPM into a broader context and remove the equations, tables, and illustrations from the article, add some additional annotations, and find out what they're really about.
