An interesting dive into what makes diffusion models work. The summary is that diffusion models are models trained on data with noise to find the original data, at various level of noise. They eventually learn the probability distribution of the images in the space of all possible pixel arrangements. You can then iteratively denoise a pure Gaussian noise picture until you generate a new image: this is like sampling the learned probability distribution.

Some interesting thoughts on diffusion models.

An in-depth description of the connections between diffusion models and Riemannian geometry.