This document proposes a method called Diffusion Posterior Sampling (DPS) to perform posterior sampling for inverse problems using diffusion models. DPS approximates the intractable posterior gradient using a denoising score model and Tweedie's formula. Specifically, it replaces the likelihood gradient with the gradient of the denoised data, allowing efficient posterior sampling via Langevin dynamics. The method is demonstrated on various inverse problems like super-resolution, inpainting, deblurring, achieving high-quality reconstructions from noisy measurements.