Federated learning is a distributed machine learning framework that enables local participants to collaboratively train a global model without sharing their data. However, the presence of heterogeneous data among clients, coupled with the growing demand for user-specific customization, renders a single global model inadequate for such scenarios. This makes personalized federated learning (pFL) a promising research direction. Existing pFL methods often prioritize model personalization at the expense of generalization. To address this issue, we propose a pFL framework based on adaptive bilateral distillation with diffusion models, termed AD \(^{2}\) -pFed, which aims to balance the personalization and generalization capabilities of client models. AD \(^{2}\) -pFed employs adaptive mutual learning and ensemble learning on the client side to facilitate knowledge transfer between private and shared models. On the server side, global pseudo-data, generated from aggregated local generators trained using diffusion models, is used as distillation samples to dynamically fine-tune the initial global model in each iteration, capturing diverse knowledge from clients. Extensive experiments on three benchmark datasets demonstrate that AD \(^{2}\) -pFed consistently outperforms baseline methods under varying data heterogeneity scenarios.