TIPS: A Text-Image Pairs Synthesis Framework for Robust Text-based Person Retrieval

ICLR 2026 Conference Submission20172 Authors

Keywords: Text-based Person Retrieval, Text-Image Pairs Synthesis, Diffusion Model, Identity Preservation, Test-Time Augmentation

Abstract: Text-based Person Retrieval (TPR) faces critical challenges in practical applications, including zero-shot adaptation, few-shot adaptation, and robustness issues. To address these challenges, we propose a Text-Image Pairs Synthesis (TIPS) framework, which is capable of generating high-fidelity and diverse pedestrian text-image pairs in various real-world scenarios. Firstly, two efficient diffusion-model fine-tuning strategies are proposed to develop a Seed Person Image Generator (SPG) and an Identity Preservation Generator (IDPG), thus generating person image sets that preserve the same identity. Secondly, a general TIPS approach utilizing LLM-driven text prompt synthesis is constructed to produce person images in conjunction with SPG and IDPG. Meanwhile, a Multi-modal Large Language Model (MLLM) is employed to filter images to ensure data quality and generate diverse captions. Furthermore, a Test-Time Augmentation (TTA) strategy is introduced, which combines textual and visual features via dual-encoder inference to consistently improve performance without architectural modifications. Extensive experiments conducted on TPR datasets demonstrate consistent performance improvements of three representative TPR methods across zero-shot, few-shot, and generalization settings.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 20172