Abstract: Typhoon path prediction serves as a core technology for early warning of meteorological disasters. Its accuracy improvement holds significant scientific value for disaster prevention and mitigation. However, in the modeling of temporal features of typhoon satellite images, existing deep learning models face challenges such as insufficient dynamic weight allocation and loss of high-frequency details. Hence, this study proposed a typhoon generative diffusion model (TGDiff) based on the denoising diffusion probabilistic model (DDPM): the DDPM-TGDiff model. Specifically, a channel prior convolutional attention (CPCA) mechanism was employed to dynamically decouple channel dependencies and spatiotemporal evolution patterns in typhoon images, enabling precise modeling of typhoon motion characteristics. Combined with the joint optimization in time and frequency domains through a high-frequency perceptual loss (HFPL) function, the DDPM-TGDiff model can more effectively capture the details of eyewall motion trajectories and cloud system evolution. Considering the temporal discontinuities in satellite data, a bidirectional dynamic interpolation module was incorporated to achieve physically consistent reconstruction of missing time-series data within the diffusion framework. Experimental results demonstrate that the DDPM-TGDiff model yielded a mean error of 63.36 km in predicting the 6 h typhoon path on the typhoon dataset of the northwest Pacific. The generated prediction image showed a peak signal-to-noise ratio (PSNR) of 16.21, a structural similarity index measure (SSIM) of 0.254, and a learned perceptual image patch similarity (LPIPS) of 0.149. Compared to existing deep learning models, the DDPM-TGDiff model significantly improves the structural fidelity of generated typhoon features and prediction accuracy.