This work presents a novel strategy to measure bias in text-to-image models. Using paired prompts that specify a gender and vaguely reference an object (e.g. “a man/woman holding an item”) we can examine whether certain objects are asso...