Large-scale vision-language models like CLIP excel in cross-modal understanding but face deployment challenges due to high computational costs. While multi-modal knowledge distillation emerges as a promising compression strategy, existing methods suffer from two critical limitations: over-reliance on image-text alignment at the expense of intra-modal image-image relationships, and insufficient handling of low-quality training data. To address these gaps, in this work, we present Image-to-Image-enhanced Contrastive Learning Distillation (I2CLD), a method that integrates cross-modal (image-text) and intra-modal (image-image) alignment within distillation paradigm. Our approach extends distillation objectives to jointly optimize image-text and image-image semantic coherence. This dual-alignment strategy enables the model to capture fine-grained visual-textual relationships while maintaining accuracy. Furthermore, we design a data augmentation pipeline that filters non-aligned pairs and synthesizes high-quality samples, enhancing image-text pairs in terms of alignment. We validate our method on large-scale industrial image-retrieval datasets that encompass diverse real-world scenarios, alleviating the data sparsity issue in long-tail scenarios. Experimental results demonstrate that I2CLD enhances both cross- and intra-modal alignment accuracy compared to distillation method that only considers cross-modal target. By focusing on cross- and intra-modal alignment, data augmentation, and real-world validation, our work presents a feasible method to compress multi-modal models under contrastive learning distillation paradigm.