<p>Text-to-image person re-identification (TIReID) aims to identify and locate pedestrian images based on given textual description queries. The main challenge of the task is bridging the significant gap between text and image modalities. Previous works primarily utilize cross-modality matching constraints to align the global or local features between samples. However, these methods overlook the relationship inconsistency problem caused by different text descriptions and generate local information redundancy in the local feature extraction process. In this paper, we propose the Granularity-Associated Invariance Features (GAIF) learning strategy to explore potential cross-modality invariant information. Firstly, we propose Global Matching Relationship Improvement (GMRI) with dynamic constraint factors to regulate the matching relationships between different samples. Secondly, we construct the Local Joint Learning Strategy (LJLS) to iteratively optimize fine-grained information from representation learning or metric learning views. Furthermore, we integrate GMRI and LJLS into a unified framework and utilize various constraints to comprehensively optimize global and local associated invariant features. We conduct extensive experiments to assess the proposed GAIF on three TIReID benchmark databases. The experimental results demonstrate that the proposed GAIF outperforms most of the advanced methods in key criteria.</p>