The paper introduces a novel approach to deep unsupervised learning using non-equilibrium thermodynamics. The method involves two main steps: first, a forward diffusion process is used to gradually transform a complex data distribution into a simple, tractable distribution. Second, a reverse diffusion process is learned to restore the structure in the data, resulting in a flexible and tractable generative model. This approach allows for efficient sampling, evaluation of probabilities, and manipulation of conditional and posterior distributions. The authors demonstrate the effectiveness of their method on various datasets, including toy problems and natural images, showing superior performance compared to existing techniques. The paper also includes an open-source implementation of the algorithm.The paper introduces a novel approach to deep unsupervised learning using non-equilibrium thermodynamics. The method involves two main steps: first, a forward diffusion process is used to gradually transform a complex data distribution into a simple, tractable distribution. Second, a reverse diffusion process is learned to restore the structure in the data, resulting in a flexible and tractable generative model. This approach allows for efficient sampling, evaluation of probabilities, and manipulation of conditional and posterior distributions. The authors demonstrate the effectiveness of their method on various datasets, including toy problems and natural images, showing superior performance compared to existing techniques. The paper also includes an open-source implementation of the algorithm.