On the Cultural Gap in Text-to-Image Generation