The Rise of Diffusion Models — A new Era of Generative Deep Learning

This walkthrough is about a paper that kicked off a new era of generative deep learning in computer vision and many other fields subsequently: the era of diffusion models. It’s titled “Denoising Diffusion Probabilistic Models” and it introduces a new framework known as DDPM, the abbreviation of the paper’s title.

While the general idea of diffusion models might seem intuitive, the math behind it is not, and you might find yourself having a hard time trying to understand papers on that topic. At least I did. At the same time, many of today’s generative models like DALL-E3, Imagen, SORA, and Stable Diffusion 3 are built upon diffusion models. Hence, it is important to understand the basics.

🚀 So, fasten your seatbelt because today is the day where we’ll build a solid intuition on the basics of diffusion models. We’ll put the DDPM into some broader context and we’ll strip down the equations, tables and illustrations from the paper, add some extra annotations and uncover what they are truely about.