The Principles of Diffusion Models - New Course Launch

Video Description

Two years back, we suddenly saw that the quality of images that were generated using AI has drastically increased. We could have almost realistic-looking images generated by AI tools within a very short amount of time. All of us take this for granted, but very few of us really understand how this sudden improvement in the quality of images was possible. I remember the first tool I started using to create images was Midjourney, and it used to create beautiful images. That's when I truly realized the power and scope of generative AI. There was another tool, however, which was becoming popular. It was called Stable Diffusion. It was a premier product of Stability AI. The best part about Stable Diffusion was that it was open source, which made high-quality text-to-image generation accessible to everyone, sparking the boom in AI art. I always wondered, while using these tools, where did the word “Diffusion” in “Stable Diffusion” come from? I had only learnt about diffusion in the subject of physics, where it was taught to me that particles diffuse when there is a concentration gradient. No way this was related to an AI technique. When I started studying about where exactly did the word diffusion and stable diffusion come from, I stumbled upon a beautiful research arc. It turned out that my experience using Midjourney was so awesome because of 10 years of dedicated research where every work built upon the previous work. I also saw that this was a field where mathematics, physics, and AI all intersected together very beautifully, and the diffusion in the sense of “physics” which I had learnt in college, was actually the same technique which was used to create images. This really caught my attention, and I wanted to completely understand the fundamentals of diffusion models from scratch. The rich intersection of diverse domains is the first part which got my attention. The second part is the applications of diffusion models. There are numerous applications of diffusion models starting from image generation, protein structure synthesis to video diffusion models, robotics policies and diffusion models to generate music and speech. Diffusion Large Language Models are also emerging as a serious alternative to the autoregressive LLMs (most of the modern LLMs) which generate tokens one after another. These models are ultrafast, e.g: Mercury LLMs. I am very excited to start a new lecture series called “The Principles of Diffusion Models”. The main objective of this series will be to understand the fundamental concepts behind diffusion models from scratch while also giving equal focus to the practical applications. All the lectures will be uploaded on YouTube for free.