What are Diffusion Models?
Lil'Log
Diffusion models are generative models that learn to reverse a process of gradually adding noise to data, starting from random noise and reconstructing samples through a Markov chain of denoising steps. Key models include denoising diffusion probabilistic models (DDPM) introduced around 2020, which use a fixed training procedure with high-dimensional latent variables matching the original data distribution. The approach enables stable training without the limitations of GANs, VAEs, or flow-based models, with applications expanded to include classifier-free guidance, latent diffusion, and consistency models through 2024.
Why it matters
[Updated on 2021-09-19: Highly recommend this blog post on score-based generative modeling by Yang Song (author of several key papers in the references)]. [Updated on 2022-08-27: Added classifier-free guidance, GLIDE, unCLIP and Imagen. [Updated on 2022-08-31: Added latent diffusion model. [Updated on 2024-04-13: Added progressive distillation, consistency models, and the Model Architecture section.