Abstract: With the explosive growth in mobile communication demands, large-scale complex scenarios, such as smart campuses with highly heterogeneous user spatial distributions, pose severe challenges to the capacity and coverage performance of wireless network planning. In such environments, user distribution is highly concentrated, spatially differentiated, and embedded in complex and diverse building environments. Consequently, base station (BS) site selection and antenna parameter configuration exhibit strong coupling and multiscale characteristics in signal coverage and interference management, which necessitate coordinated optimization from a global perspective. However, conventional planning methods, heuristic algorithms, and deep learning methods often suffer from inefficient exploration and unstable convergence in large-scale high-dimensional decision spaces, making it difficult to effectively address the joint optimization of BS locations and engineering parameters. To address these challenges, this study proposes a conditional diffusion model-enabled hierarchical reinforcement learning (HRL-DiffNet) framework. In particular, a conditional diffusion model learns high-quality deployment distributions to generate candidates that serve as structured priors to shrink the reinforcement learning (RL) exploration space. Unlike conventional RL methods that rely on random exploration, the diffusion model utilizes historical optimization experience to capture the implicit distribution characteristics of effective BS deployment solutions. By considering environmental states as conditional inputs, the model generates deployments with higher optimization potential, providing effective initialization guidance for the policy, and reducing unnecessary exploration in the large discrete decision space. A hierarchical reinforcement learning architecture is developed, where a high-level policy is responsible for BS site selection, whereas a low-level policy performs fine-grained optimization of antenna parameters, such as downtilt angles. Through this hierarchical RL architecture, the strong coupling between deployment planning and parameter configuration is effectively alleviated, allowing policies to achieve coordinated optimization with improved convergence efficiency. In addition, a hybrid neural architecture that integrates convolutional neural networks (CNNs) and graph convolutional networks (GCNs) is designed to model complex topological dependencies and environmental features, providing conditional constraints for the diffusion model and serving as a unified state representation for reinforcement learning. A CNN is utilized to extract spatial characteristics from environmental information, whereas a GCN captures the relationships among BSs, users, and the surrounding network topology. The combination of these two representations enables the proposed framework to perceive both spatial distribution patterns and topological correlations comprehensively in complex wireless environments. This establishes an end-to-end coupled framework that jointly integrates topological representation, conditional diffusion-based generation, and optimization-driven decision-making. Extensive simulations are conducted in large-scale smart campus network planning scenarios to evaluate the effectiveness of the proposed HRL-DiffNet framework. The proposed method is compared with representative optimization approaches, including conventional RL algorithms and existing intelligent optimization methods. The experimental results demonstrate that HRL-DiffNet achieves superior performance in terms of network coverage, throughput improvement, and optimization stability. Specifically, the proposed approach achieves a network coverage rate of 99.1% and an average throughput of 1.766 Mbps, demonstrating its effectiveness for large-scale deployment optimization. Furthermore, the results validate the effectiveness of the hierarchical decision mechanism in handling the coupled optimization of BS deployment and antenna configuration in large-scale heterogeneous scenarios.