Abstract

SynC refines noisy synthetic image-caption datasets through a one-to-many mapping that re-aligns each image with its best-matching captions from the generated pool. This data-centric refinement boosts zero-shot image captioning performance without any additional human annotation.