Ornstein-Uhlenbeck Noise Driven Diffusion Model

This paper proposes a new method called OU-Diff to improve current diffusion models. Diffusion models typically generate images by gradually adding and removing noise, but existing approaches have some issues—either the initial noise injection is insufficient, or the computational efficiency is not optimal. Our solution introduces a dynamic noise mechanism based on the Ornstein-Uhlenbeck (OU) process, allowing the model to more flexibly control noise addition and removal. Theoretical analysis shows that this method effectively adjusts noise levels, leading to better generation results. Experiments demonstrate that, without modifying the model architecture, OU-Diff performs better in image generation tasks.