SalArt-VQA: Diagnosing Whether VLMs Understand Salient Artifacts in Generated Images
SALART-VQA is a diagnostic benchmark created to test how well vision-language models understand subtle artifacts in AI-generated images. It moves beyond simple detection by asking chained questions about presence, location, and description. The results show that while models can often detect an artifact, they frequently fail at the more complex tasks of correctly localizing and describing it.