Diffusion Posterior Sampling

Diffusion posterior sampling leverages pre-trained diffusion models as powerful priors to efficiently solve inverse problems, aiming to generate high-quality samples from a posterior distribution given noisy or incomplete data. Current research focuses on improving the accuracy and efficiency of sampling algorithms, including modifications to Langevin dynamics, Markov Chain Monte Carlo methods, and the development of novel likelihood approximations and optimized ODE solvers. This approach holds significant promise for various applications, such as image restoration, compressed sensing, and material decomposition, by offering a flexible and computationally efficient framework for Bayesian inference in high-dimensional spaces.