Diffusion models are generative models that produce data through a stepwise process in which noise is first added and then gradually removed.
The training process involves training a model that learns to perform the reverse diffusion process, i.e., removing noise to reconstruct a realistic original from a noisy image.
Typically, a Markov process is used that increases the noise in the data over several steps, and the model learns to reverse this process.
This technique enables the generation of high-quality and diverse data, such as images, audio, or text.
Well-known examples are Denoising Diffusion Probabilistic Models (DDPM) and Score-based Generative Models.