arxivcs.CVcs.AIcs.LGcs.MM2026-07-17
Debiasing Text-to-Image Evaluation via Implicit Cultural Alignment Reward Modeling
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 evaluation metrics and multimodal judges often rely on visual-semantic representation…