Text-to-Image Person Re-Identification (TIReID) addresses the challenging task of retrieving target pedestrian images from textual descriptions, requiring precise fine-grained visual-language alignment while overcoming interference from hard negative samples. To tackle these challenges, we propose a novel framework named Hard Negative-aware and Frequency-enhanced Asymmetric Learning (HF-AL). First, we introduce a Hard Negative-aware Contrastive Loss (HNCL) that focuses on non-matching pairs with high similarity, improving feature discrimination near decision boundaries. Second, we propose Frequency-enhanced Teacher–Student Asymmetric Learning (FTAL), which integrates Teacher–Student Asymmetric Learning (TSAL) and Frequency-domain Augmentation (FDA) to enhance cross-modal feature robustness. Extensive experiments on CUHK-PEDES, ICFG-PEDES, and RSTPReid demonstrate that HF-AL achieves remarkable and consistent performance, validating the effectiveness of our approach.