Publication
Multilingual Text-to-Image Generation Magnifies Gender Stereotypes and Prompt Engineering May Not Help You
Felix Friedrich; Katharina Hämmerl; Patrick Schramowski; Jindrich Libovický; Kristian Kersting; Alexander Fraser
In: Computing Research Repository eprint Journal (CoRR), Vol. abs/2401.16092, Pages 1-24, arXiv, 2024.
Abstract
Text-to-image (T2I) generation models have
achieved great results in image quality, flexibil-
ity, and text alignment, leading to widespread
use. Through improvements in multilingual
abilities, a larger community can access this
technology. Yet, we show that multilingual
models suffer from substantial gender bias. Fur-
thermore, the expectation that results should be
similar across languages does not hold. We
introduce MAGBIG, a controlled benchmark de-
signed to study gender bias in multilingual T2I
models, and use it to assess the impact of multi-
lingualism on gender bias. To this end, we
construct a set of multilingual prompts that
offers a carefully controlled setting account-
ing for the complex grammatical differences
influencing gender across languages. Our re-
sults show strong gender biases and notable
language-specific differences across models.
While we explore prompt engineering strate-
gies to mitigate these biases, we find them
largely ineffective and sometimes even detri-
mental to text-to-image alignment. Our analy-
sis highlights the need for research on diverse
language representations and greater control
over bias in T2I models.