Denoising Diffusion Probabilistic Models

The paper presents diffusion probabilistic models, a class of latent variable models inspired by nonequilibrium thermodynamics, for high quality image synthesis. The core method trains these models on a weighted variational bound whose design is guided by a novel connection the authors establish between diffusion probabilistic models and denoising score matching with Langevin dynamics. The resulting models also naturally support a progressive lossy decompression scheme, which the authors interpret as a generalization of autoregressive decoding.

On the unconditional CIFAR10 benchmark the approach achieves an Inception score of 9.46 and a state-of-the-art FID score of 3.17, and on 256x256 LSUN it produces sample quality comparable to ProgressiveGAN. This mattered because it established diffusion models, grounded in a principled connection to score matching, as a competitive and eventually dominant approach for high-fidelity image generation, with an openly available implementation.