Effectiveness of intervention at the prompt-revision stage

Determine whether intervention at the prompt-revision stage is feasible, sufficient on its own, or best combined with image-level debiasing to mitigate cultural stereotyping in text-to-image systems.

Background

The paper identifies the prompt-revision layer as a causal source of cultural stereotyping in commercial text-to-image systems: it can introduce stereotypical content before the image-generation model receives the prompt. This finding implies that debiasing efforts directed solely at the image model may not remove stereotypes already encoded during prompt revision.

The authors therefore identify an unresolved mitigation question concerning the practical viability of intervening at the revision stage, whether such intervention could independently address the observed bias, and whether it should instead be combined with image-level debiasing. The paper’s evaluation methodology is presented as a means for assessing these alternatives in future work.