DreamSim: Learning New Dimensions of Human Visual Similarity using Synthetic Data NeurIPS 2023 Spotlight * Equal contribution, order decided by random seed Which image, A or B, is most similar to the reference? We generate a new benchmark of synthetic image triplets that span a wide range of mid-level variations, labeled with human similarity judgments. The dots below each image indicate which image is considered most similar to the reference by humans vs several existing metrics and our new met...