Abstract: Traditional and deep learning denoising techniques exhibit shortcomings in handling specific types of noise and data generalization validation in Electrocardiogram (ECG) signals. This paper proposes a generative ECG denoising model based on diffusion models, which leverages simulated data to learn the score function of clean ECG distribution and generates and separates ECG and noise based on the Euler method for solving Ordinary Differential Equations (ODE). The model is trained on simulated data and validated on an independent real dataset. Compared to other relevant methods, the obtained results demonstrate that this model has significant advantages in removing diverse noise and maintaining consistency in ECG waveforms of different amplitude features.