Straightened or Caricatured: How Text-to-Image Models Portray Queer Gender Identities

We analyze text-to-image generated iamges depicting queer gender identities, and how these images are perceived by members of the community. Through in-depth interviews with community members, we highlight several idiosyncratic behaviors of these models, grounding the harms within a queer theory framework. We further substantiate our findings with quantitative analysis. We show that these models demonstrate a very narrow perspective of queerness: participants notice that these models often portray queer gender identities as white, young, and feminine. The participants also observe that models portray binary transgender identities very differently compared to non-binary identities: `straightening' binary identities by depicting them closer to heteronormative, while picturing non-binary identities as unhappy and oversexualized caricatures. We reflect on implications these have on society, informing future research and development of these models.