3D point cloud technology have grown significantly recently, and its compression technology enables real-time transmission in bandwidth-constrained environments. However, compression artifacts affect signal fidelity and visual quality, further limiting the performances of downstream point cloud tasks. To tackle this challenge, deep learning-based artifact removal techniques have shown remarkable promise, particularly in enhancing geometry accuracy and attribute quality. This chapter delves into deep learning-based methods for geometry and attribute artifacts removal, analyzing their advantages in point cloud data organization, feature extraction paradigms, inference forms, and loss functions, while showcasing their adaptability to varying sparsity levels of point clouds. Specifically, it first categorizes and compares point cloud geometry and attributes artifact removal techniques, highlighting the performance and applicability of methods such as sparse convolution, 3D convolution, and graph convolution in feature extraction. Second, it summarizes and compares representative methods, including U-Net, Grnet, MS-GAT, and GPCC++, outlining their strengths, limitations, and applicable scenarios. Afterward, it examines the uniqueness and challenges of joint geometry and attribute artifact removal tasks, introducing innovative frameworks like GPCC++ for handling hybrid distortions.