QC-T2I-Bench: Scalable Text-to-Image Evaluation and Routing.

Evaluating modern text-to-image (T2I) models is challenging because models with similar overall scores often have distinct strengths and weaknesses, making practical selection difficult. Existing fine-grained benchmarks often revert to prompt-level scores or fixed categories, obscuring specific failure points and ignoring prompt complexity. This research presents "QC-T2I-Bench," a question-centric framework designed for scalable T2I evaluation. It transforms open prompts into attributed atomic questions and maps their dependencies using Davidsonian Scene Graphs (DSGs). The framework uses hierarchy-constrained question aggregation to prevent simple and complex prompts from being weighted equally and to exclude downstream questions if prerequisites fail. The DSG structure also allows for measuring joint success within prompts and comparing repeated entities, distinguishing basic realization failures from those under additional requirements. Experiments on open-source T2I models with English and Chinese prompts demonstrated that joint completion rates significantly decrease as prompt complexity increases. Furthermore, the framework's records can be reused for training-free, cost-aware routing, achieving comparable performance to a high-performing model with substantial GPU-s/MP savings.