Metaphor-based Jailbreaking Attacks on Text-to-Image Models