Faithful, Enriched, and Precise: Benchmarking Natural-Science Illustration Generation by T2I models
Researchers have introduced FEPBench, a new benchmark designed to evaluate Text-to-Image (T2I) models specifically for generating scientific illustrations. The benchmark assesses models on instruction faithfulness, reasoning enrichment, and semantic precision, going beyond holistic evaluations to analyze fine-grained elements. Current state-of-the-art models, including GPT Image 2 and Nano Banana Pro, still face challenges with text rendering, reasoning capabilities, and balancing generation richness with precision. AI
IMPACT Identifies key limitations in current T2I models for scientific illustration, guiding future development for more accurate and contextually rich visual communication.