SANEval introduces open-vocabulary compositional benchmarks using LLM-based prompt understanding and open-vocabulary detection to diagnose text-to-image failure modes. Its automated metric correlates more faithfully with human judgments across attribute binding, spatial relations, and numeracy than