MulCLIP: A Multi-level Alignment Framework for Enhancing Fine-grained Long-context CLIP

Vision-language models such as CLIP excel at image-text alignment but struggle with long, detailed descriptions due to training on short captions. Recent methods address this limitation using region proposals to align visual regions with sentences, but at substantial deployment cost. We present MulCLIP, an end-to-end multi-level alignment framework that directly exploits natural long-text structure without region proposals. Instead of simply stacking objectives, MulCLIP aligns different textual granularities through compatible training signals: global contrastive alignment for long captions and short summaries, Word–Patch Reconstruction over locally calibrated features for within-sample word–patch semantics, and Subcaption–Aggregated Patch alignment for context-rich subcaption grounding. Experiments on benchmarks spanning varying caption lengths show consistent gains, and ablations confirm that the proposed objectives are complementary and jointly improve the overall framework.