in PyTorch
Diffusion models are a popular new method in the field of generative deep learning. They are used in a lot of SOTA architectures and recently had breakthroughs in text to image generation (DALL-E 2, IMAGEN…). In order to fully understand these models, I thought it might be helpful to build one from scratch. In this video I implement a DDPM (Denosing Diffusion Probabilistic Model) and explain some of the theory on the way. Hope you enjoy it!