<p>Point clouds acquired via LiDAR or depth cameras are often sparse, non-uniform, and incomplete due to limitations in resolution, occlusion, and lighting conditions, severely hindering downstream 3D tasks. While recent deep learning-based upsampling methods partition sparse inputs into patches for separate processing, the inherent inconsistency during patch merging inevitably generates discrete outliers. To address this, we propose PU-iGNN, a novel point cloud upsampling network that eliminates the need for explicit patch merging by leveraging implicit graph neural networks to capture long-range dependencies, thereby suppressing outliers and enhancing distribution uniformity. Our key innovations include: (1) A dual-space feature extraction module integrating implicit and Euclidean spaces for multi-dimensional representation learning; (2) An edge-aware feature expansion structure that propagates features from patch centroids to neighbors, explicitly enforcing local geometric consistency through center-neighbor constraints. Extensive quantitative and qualitative experiments demonstrate that PU-iGNN significantly outperforms state-of-the-art methods in suppressing outliers while achieving superior uniformity and fidelity.</p>