The generation of 3D point clouds with controllable part-level details remains a challenging task in the field of computer graphics and vision. Existing methods often produce monolithic point clouds, lacking the flexibility for part editing and transformation necessary for various applications such as object modeling, animation, and scene reconstruction. To address this limitation, we introduce the Controllable Point Cloud Generation Network (CPNet), a novel probabilistic generative model tailored for part-level point cloud generation. Our approach consists of three main components: a Segment Paerts (SP) module that establishes the foundational learning for parts, a Part selector and Transformation sampler that models the distribution of points probabilistically, and a Cross-Attention Diffusion (CAD) module with special attention mechanism designed to denoise and refine the generated point cloud. By decomposing the point cloud at the part-level, CPNet enables the generation of point clouds that can be controlled through part-level editing and replacement. The proposed method achieves state-of-the-art (SOTA) results in terms of point cloud quality and offers unparalleled control over the generation process. We demonstrate the effectiveness of CPNet through extensive experiments and comparisons with existing methods, highlighting its potential for applications in computer graphics, robotics, and other domains requiring high-quality, controllable 3D point cloud data.