Abstract: In the field of point cloud registration, the accuracy of the traditional transformer method will decline in the case of low overlap due to the lack of geometric information perception and the loss of local details. In this research, a point cloud registration network based on multi-scale diffusion (MSD) model is proposed, which improves the robustness of the model through cooperative geometric coding and probability optimization strategy. Firstly, the network uses deformable convolution to construct feature pyramid, and combines with dynamic radius search technology to adapt to non-uniform distribution of point cloud data, so as to enhance the modeling ability of local surface continuity. Secondly, the triple geometry embedding mechanism is fused, and the hierarchical moving window attention mode is used to realize the effective coordination of local and global features. Finally, the transformation matrix is iteratively optimized in the double random matrix space, and the sinkhorn constraint is introduced to improve the stability. The experimental results show that the recall rate of MSD point cloud registration network on 3DMatch and 3DLoMatch datasets is significantly higher than other existing methods, which proves the superiority of this method. This study shows that by combining the multi-scale window attention mechanism and diffusion model, the matching accuracy and robustness in the point cloud registration task can be effectively improved.