Text-based Person Search (TPS) grapples with formidable obstacles in seasonal bias generalization due to marked disparities in apparel characteristics across seasonal domains. This paper proposes a lightweight framework, TPS-SG, to address these issues. Traditional Mix-based methods suffer from poor generalization across seasons because they ignore fine-grained alignment of clothing attributes. Meanwhile, existing diffusion models such as Text2Human and Difffashion have problems with low computational efficiency and identity distortion. To try to solve these problems, we propose TPS-SG, a lightweight framework composed of the following components: Semantic-Aware Region Alignment (SARA): Textual attributes are extracted using a pretrained LLM(GPT) and mapped to corresponding body regions to guide visual-text alignment. Region-Specific Replacement Module (RSRM): Clothing regions are selectively replaced based on cosine similarity thresholds to enable controllable cross-season attribute substitution. Multi-Granularity Identity Preservation (MGIP): Identity-attribute disentanglement loss is applied to suppress irrelevant variations and maintain identity consistency. We experimentally demonstrate the effectiveness of TPS-SG on the CUHK-PEDES and ICFG-PEDES datasets. It achieves a Rank-1 accuracy of 73.5% (±0.3%), with mAP and mINP reaching 68.3% and 54.1%, respectively.