Image generators can make a polished picture that quietly uses the wrong reference photo, attaches a detail to the wrong object, or copies something the user never asked for. Everyday quality scores can't catch that.
This gives each task many reference images plus a graph of which entities must appear, how they relate, and scene-level rules. Each node, link, and scene rule becomes a small checklist item, making it possible to see exactly where an image failed rather than just whether it looks good. Expert labels checked the checklist scoring.
The main finding is about workflow: wrapping a fixed image generator in an , a controller that selects evidence and revises requests, improved aggregate results over asking for the image directly. No single workflow won: the best scores for particular subcategories were spread across direct generation, simple harnesses, and a richer ReAct-Plan-Exec-Reflect (RPR) workflow that retrieves evidence, plans, executes, and revises, and RPR helped only when the surrounding harness used its evidence and revision well, otherwise it drifted. The benchmark offers an auditable way to measure dense visual evidence use.