Debiasing Text-to-Image Evaluation via Implicit Cultural Alignment Reward Modeling

arXiv:2607.15740v3 Announce Type: replace-cross Abstract: As Text-to-Image (T2I) systems rapidly advance, evaluating the cultural authenticity of synthesized content has become increasingly important for fair and trustworthy generative AI. Existing T2I evaluat