Path planning is a critical component for enabling autonomous navigation in mobile robots. Sampling-based planners are widely adopted due to their strong generality, yet they rely heavily on uniform sampling, which often leads to unstable performance and high computational cost in complex environments. To address this issue, recent studies feed free-space point clouds into neural networks to infer a set of guidance states near the optimal path, thereby enabling non-uniform sampling; however, the accuracy of the guidance set becomes a key bottleneck for further improvement. In this paper, we propose an improved point-cloud neural RRT* framework, termed IPN-RRT*, which achieves fast near-optimal planning via a high-precision guidance state set. Specifically, we develop an Improved PointNeXt-based neural sampling network (IPN) that enhances the geometric representation of free-space point clouds using high-dimensional sinusoidal positional encoding ( $HP{E}_{SIN}$), and further improves cross-scene feature discriminability and robustness through a gated covariance-enhanced channel attention module (GCECA). These designs substantially improve the quality of the guidance state set and accelerate convergence toward better solutions during planning. Extensive experiments on large-scale datasets built from complex random maps demonstrate that IPN significantly outperforms existing point-cloud prediction models in region prediction accuracy. Moreover, across diverse challenging environments, IPN-RRT* finds near-optimal paths with fewer nodes and shorter runtime, while preserving probabilistic completeness and asymptotic optimality.