Federated Learning (FL) enables collaborative model training across decentralized clients while preserving data privacy. However, traditional FL methods often struggle to adapt to heterogeneous, non-IID data distributions across clients. In this paper, we propose PFedDyFilter, a novel personalized federated learning framework that leverages Decoupled Dynamic Filters (DDF) to improve model adaptability under non-IID conditions. Specifically, each client generates a low-dimensional client descriptor via a local embedding network, which is transmitted to a server-side hypernetwork to produce personalized model parameters. These descriptors and parameters are then jointly used to modulate the DDF module, dynamically generating convolutional kernels along both spatial and channel dimensions. This design enables lightweight and fine-grained personalization for each client. Extensive experiments on CIFAR10, CIFAR100, and FEMNIST demonstrate that PFedDyFilter consistently outperforms state-of-the-art baselines, maintaining strong generalization even on unseen clients. Additional evaluations on more expressive architectures (e.g., ResNet) further verify the robustness and generalization ability of our approach across diverse client distributions.