SciDraw-Bench — Testing AI on Scientific Figure Quality
Text-to-image models now generate mechanism diagrams, lab schematics, and graphical abstracts. Existing benchmarks (GenEval, T2I-CompBench) measure photorealism and object counting, but miss what makes a scientific figure work: legible labels, faithful entity depiction, and disciplinary conventions.
SciDraw-Bench fills that gap with 32 structured tasks across eight figure types and ten disciplines, rated on text clarity, diagrammatic coherence, and scientific accuracy.