MaSC: A Masked Similarity Metric for Evaluating Concept-Driven Generation
Researchers have developed MaSC, a new metric for evaluating concept-driven image generation, which improves upon existing methods by spatially decomposing image analysis. Unlike previous metrics that use global embeddings, MaSC utilizes foreground masks to separately assess concept preservation and prompt following. This approach demonstrates superior performance on benchmarks like DreamBench++ and ORIDa, outperforming models such as GPT-4V and approaching GPT-4o in human-rated evaluations. AI
IMPACT Provides a more accurate evaluation framework for text-to-image models, potentially guiding future development and benchmarking.