Abstract: Leveraging powerful semantic understanding and generation capabilities, Vision-Language Pre-trained (VLP) large models have demonstrated remarkable potential in cross-modal retrieval.