Improving Composed Image Retrieval via Contrastive Learning with Scaling Positives and Negatives