Beyond the Surface: A Global-Scale Analysis of Visual Stereotypes in Text-to-Image Generation