Object re-identification (ReID) aims to identify instances with the same identity as a given query object from a large image repository. One of the main challenges in ReID is the substantial variability in viewpoints of the same object. To mitigate the detrimental effects of this viewpoint inconsistency, we propose a novel strategy called tight-loose pairs mining. This method is designed to systematically explore and exploit the inner-class discrepancies among samples. Specifically, we introduce a pairwise loss called Tight-Loose loss (TL loss) to fortify the resemblance between distinct subsets of samples belonging to the same identity class. Precisely, we segment the feature space by grouping samples sharing the same class into ‘tight’ and ‘loose’ clusters, depending on their relative proximity in the feature space. The TL loss maximizes the similarity between tight group samples and loose group samples of the same class while minimizing the similarity between group features of different classes. Extensive experiments have been conducted to evaluate the effectiveness of TL loss against prevailing pairwise loss functions. The results demonstrate that TL loss effectively reduces intra-class variations more than other existing methods. The proposed approach has been validated on various object ReID benchmarks including CUHK03, Market-1501, DukeMTMC-reID, MSMT17, VeRi-776, and VehicleID, demonstrating consistent improvements across these datasets. The code is available at https://github.com/CharvinMei/TL-RelD.