DebiasPI: Inference-time Debiasing by Prompt Iteration of a Text-to-Image Generative Model
Sarah Bonna, Yu-Cheng Huang, Ekaterina Novozhilova, Sejin Paik, Zhengyang Shan, Michelle Yilin Feng, Ge Gao, Yonish Tayal, Rushil Kulkarni, Jialin Yu, Nupur Divekar, Deepti Ghadiyaram, Derry Wijaya, Margrit Betke · 2025
Read DebiasPI: Inference-time Debiasing by Prompt Iteration of a Text-to-Image Generative Model by Sarah Bonna, Yu-Cheng Huang, Ekaterina Novozhilova, Sejin Paik, Zhengyang Shan, Michelle Yilin Feng, Ge Gao, Yonish Tayal, Rushil Kulkarni, Jialin Yu, Nupur Divekar, Deepti Ghadiyaram, Derry Wijaya, Margrit Betke (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-18642.txt.
Description
Sarah Bonna and a large team bend text-to-image generation toward fair crowd shares at generation time, with no retraining. Bias by sex and race in generated images is a known flaw, and past fixes often miss the target shares. Their 2025 loop, DebiasPI, tracks what has been generated so far, by probes inside the model or by side classifier tools. It then steers the next prompt toward the thin shares.
Images with even race and sex shares came out for hard news-headline themes. Age, build, job, and skin tone got the same loop. Cross effects show a catch: fix race and sex balance, yet skin spread thins, and the palest tones stay hard to generate. Cut the loop and the images skew young and male.
Catalog record
| CREATOR | Sarah Bonna, Yu-Cheng Huang, Ekaterina Novozhilova, Sejin Paik, Zhengyang Shan, Michelle Yilin Feng, Ge Gao, Yonish Tayal, Rushil Kulkarni, Jialin Yu, Nupur Divekar, Deepti Ghadiyaram, Derry Wijaya, Margrit Betke |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 35 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-18642