<p>Random noise is present in seismic data acquisition and affects subsequent data processing and interpretation. Traditionally, it is difficult to completely suppress or separate the random noise using time domain or frequency domain methods. Although some deep learning methods can improve computational efficiency and parameter adaptation, they still fail to completely separate random noise from effective signals. Obtaining sufficient training samples remains a problem. We introduce the Diffusion Model, which adds noise to clean data in a step by step forward diffusion process, and removes noise in the backward direction based on a neural network. The Diffusion Model predicts the noise and reconstructs clean effective signals. Combined with stationary-phase migration, we can obtain training samples and successfully remove the random noise while preserving effective signals. This improves the accuracy of subsequent processing and structure interpretation. The effectiveness of our method has been verified by field imaging profile and synthetic common-shot datasets.</p>