Geometry-Guided Latent Diffusion Model for Static Point Cloud Color Attribute Denoising

Point cloud denoising is a challenging task due to the high-dimensional nature of data. To remove the compression artifact in the point cloud, this letter presents a static point cloud color attribute denoising method based on the geometry-guided Latent Diffusion Model (LDM). In specific, to better mimic realistic noise in LDM, we propose a color noise simulation strategy that interpolates distortion levels from Geometry-Based Point Cloud Compression (G-PCC) into the continuous noise distribution. Additionally, a dynamic sampling step control mechanism adaptively allocates diffusion steps based on local texture complexity, ensuring detailed denoising in complex regions while efficiently processing smoother areas. In the reverse denoising process, the geometry information is enabled as condition to guide color attribute recovery. Experimental results indicate that our method outperforms the state-of-the-art work, effectively balancing denoising quality and computational efficiency.