Diffusion

there are many ways to understand diffusion. classically, diffusion is the process by which particles in a closed system move from areas of high concentration/density to lower ones.

diffusion in #compsci is a Generative Modeling process; diffusion modeling is a two step process of adding and removing successive "doses" of gaussian noise to training data. The forward process adds the noise while the reverse process learns how to reconstruct original image; here, the reverse process is maximizing the likelihood of the training data.

diffusion models and Score-matching Generative Models have a lot of similarities and may be two expressions of the same phenomena.

diffusion models benefit heavily from exploration and sampling methods of a variety of flavors due to its Expressiveness. two example are the classes of Guided Sampling Methods and Fast Sampling Methods