Modern text‑to‑image (T2I) models often achieve similar overall scores while excelling in different capabilities, making practical selection challenging. Fine‑grained benchmarks usually decompose prompts into questions, but then map them back to prompt scores or fixed categories, weakening attribution and overlooking question complexity. Related requirements are frequently scored separately or merged into a single total, obscuring basic realization errors from compositional ones.
We introduce QC‑T2I‑Bench, a question‑centric framework that converts open prompts into attributed atomic questions and organizes their dependencies with Davidsonian Scene Graphs (DSGs). Hierarchy‑constrained aggregation excludes downstream questions when a prerequisite fails and prevents simple and complex prompts from receiving identical total weight.
Leveraging the DSG structure, we measure joint success within a prompt and compare repeated entities across prompts, separating basic realization failures from those under additional requirements. We evaluate several open‑source T2I models on English and Chinese prompts. Results show joint completion drops from 80.7% for components with two capabilities to 37.2% for those with seven or more.
Finally, we reuse the same records for a training‑free router. The cost‑aware router matches ERNIE’s 89.51‑point estimate while using 21.3% less GPU‑s/MP.
Blogger's Review: QC‑T2I‑Bench’s question‑level granularity and dependency graph provide clearer insight for T2I model selection and diagnosis, and its cost‑aware routing balances performance with efficiency, making it a valuable contribution to both research and practice.