緩解文字生成圖像系統中的刻板偏見
Mitigating stereotypical biases in text to image generative systems
Runway 的研究團隊提出多樣性微調(DFT)方法,透過涵蓋不同膚色、性別、職業及年齡組別的合成數據,緩解文字生成圖像模型中的社會偏見。DFT 將感知膚色的羣體公平性指標提升 150%,將感知性別的指標提升 97.7%;相較基準模型,生成結果包含更多深膚色人士及女性。團隊表示將公開所有提示詞及生成訓練圖像的程式碼。
Authors
Piero Esposito, Parmida Atighehchian, Anastasis Germanidis, Deepti Ghadiyaram 1
State-of-the-art generative text-to-image models are known to exhibit social biases and over-represent certain groups like people of perceived lighter skin tones and men in their outcomes. In this work, we propose a method to mitigate such biases and ensure that the outcomes are fair across different groups of people. We do this by finetuning text-to-image models on synthetic data that varies in perceived skin tones and genders constructed from diverse text prompts. These text prompts are constructed from multiplicative combinations of ethnicities, genders, professions, age groups, and so on, resulting in diverse synthetic data. Our diversity finetuned (DFT) model improves the group fairness metric by 150% for perceived skin tone and 97.7% for perceived gender. Compared to baselines, DFT models generate more people with perceived darker skin tone and more women. To foster open research, we will release all text prompts and code to generate training images.
Video: Learn more about bias in AI and our efforts with diversity finetuning.

Figure 1: We mitigate stereotypical biases by finetuning Stable Diffusion-1.5 (26) and Stable Diffusion-XL (22) on synthetic data that varies across perceived skin tones, genders, professions, and age groups. For the same prompt and seed, notice that our diversity finetuned (DFT) models generate more inclusive results..

Table 1: Effect of prompt qualifiers on group fairness. Given lighter skin tone and perceived male gender are the sub-groups TTI models default to, we measure disparate impact (Eqn. 1) relative to these categories (i.e., X2 = light, male). L/M/D refers to distribution of the predicted skin tones into light, medium and dark categories and F/M refers to distribution into predicted female and male categories.

Table 2: Effect of data composition. Note that a balanced distribution of perceived skin tones yields better performance. L/M/D refers to distribution of the predicted skin tones into light, medium and dark categories.
來源:Runway:News(網頁) · runway.com