Abstract: Point cloud upsampling plays a critical role in enhancing the resolution and quality of 3D data captured by sensors. However, existing methods often struggle to simultaneously preserve global topological structures and recover local geometric details, frequently leading to non-uniform point distributions and outlier artifacts. To overcome these limitations, this paper proposes a novel upsampling network, termed PointMamba-UPS (Effective Point Cloud Upsampling Network based on Mamba and Self-attention), which integrates the Mamba (Linear-Time Sequence Modeling with Selective State Spaces)architecture with self-attention mechanisms. The method employs a dual-branch feature extraction framework. Specifically, a global feature extractor built upon Mamba blocks is designed to capture the overall shape and semantic information of objects. By utilizing a feature reordering strategy to accommodate the disorder inherent in point clouds, this branch establishes long-range dependencies while maintaining linear computational complexity. In parallel, a local feature extractor based on multi-head self-attention mechanism is constructed to adaptively aggregate geometrically similar features within neighborhoods, thereby precisely characterizing edge and curvature variations. Furthermore, a symmetric dense connection fusion structure is adopted to achieve deep integration of global semantic priors and local geometric details. Experimental results demonstrate that, compared with PU-Transformer (Point Cloud Upsampling Transformer), the proposed method reduces the Chamfer Distance, Hausdorff Distance, and Point-to-Surface Distance by 5.1%, 3.3%, and 6.9%, respectively, on the PU1K dataset, while decreasing the inference time by 16.1%. The proposed method outperforms existing mainstream methods across multiple evaluation metrics, fully validating the effectiveness and practicality of the dual-branch collaborative architecture for point cloud upsampling.