Pedestrian Open-Attribute Recognition (POAR) involves identifying appearance attributes within an open-world setting. In this work, we focus on improving the performance of the model in identifying both seen and unseen attributes. To achieve this, we propose a novel POAR framework based on Dynamic Semantic Masking (DSM). The proposed framework first employs an Adaptive Coupled Localization (ACL) module, synergizing OpenPose and SAM outputs for precise localization of fine-grained attribute parts. Subsequently, a Keypoint-driven Dynamic Attention Masking (KDAM) mechanism produces attribute-specific image masks. The model is then optimized via a vision-language multimodal contrastive loss. Finally, to enhance generalization towards unseen attributes, the knowledge distillation method is employed to transfer the generalization capabilities of CLIP to our model. The experimental results demonstrate the effectiveness of the proposed method on three public PAR datasets.