Point cloud data is integral to various application scenarios due to its provision of extensive semantic information. Although density stands as a vital attribute of point clouds, many prevailing loss functions concentrate solely on inter-point distances, disregarding the distribution and density structure of the point cloud during the evaluation of similarity between two sets of points. To address this, we propose a density loss based on voxelization (VD). This density loss is obtained by summing the squared differences in the number of points within corresponding voxel grids derived from the original data and the model’s reconstructed results, utilizing the same voxelization process as the original data. Our proposed loss enhances the model’s sensitivity to point cloud density while quantifying the similarity between point clouds, thereby leading to improved performance in model optimization. Empirical experiments demonstrate that incorporating the VD loss into the Point-M2AE and I2P-MAE point cloud processing models effectively enhances feature representation and significantly improves the performance of the point cloud MAE models.