[论文解读] Auditing Gender Presentation Differences in Text-to-Image Models

Yanzhe Zhang, Lu Jiang+2 more|arXiv (Cornell University)|Feb 7, 2023

Human Motion and Animation被引用 8

一句话总结

The paper introduces the GEP metric to quantify attribute-level gender presentation differences in text-to-image models, and proposes automatic GEP estimation via cross-modal classifiers built on CLIP space, showing higher correlation with human annotations than CLIP alone across three models.

ABSTRACT

Text-to-image models, which can generate high-quality images based on textual input, have recently enabled various content-creation tools. Despite significantly affecting a wide range of downstream applications, the distributions of these generated images are still not fully understood, especially when it comes to the potential stereotypical attributes of different genders. In this work, we propose a paradigm (Gender Presentation Differences) that utilizes fine-grained self-presentation attributes to study how gender is presented differently in text-to-image models. By probing gender indicators in the input text (e.g., "a woman" or "a man"), we quantify the frequency differences of presentation-centric attributes (e.g., "a shirt" and "a dress") through human annotation and introduce a novel metric: GEP. Furthermore, we propose an automatic method to estimate such differences. The automatic GEP metric based on our approach yields a higher correlation with human annotations than that based on existing CLIP scores, consistently across three state-of-the-art text-to-image models. Finally, we demonstrate the generalization ability of our metrics in the context of gender stereotypes related to occupations.

研究动机与目标

• Define and quantify gender presentation differences (GEP) at the attribute level in text-to-image generation.
• Introduce a neutral and explicit prompting framework to measure how gendered prompts affect attribute presence in generated images.
• Develop automatic GEP estimation methods that correlate with human annotations and improve over CLIP-based approaches.
• Evaluate GEP across multiple state-of-the-art models to understand model-specific presentation differences and generalize to occupation stereotypes.

提出的方法

• Define GEP vectors where each dimension measures the difference in attribute frequency between images generated from 'a woman' vs 'a man' prompts under neutral and explicit settings.
• Construct attribute set A and context set C from ConceptNet and COCO-derived contexts to analyze presentation attributes.
• Compute GEP as the normalized L1 norm of the GEP vector, enabling cross-model comparisons of overall gender presentation differences.
• Propose automatic estimation of attribute frequencies using CLIP-based similarities with calibration and cross-modal attribute classifiers trained in the CLIP embedding space.
• Train multiple (ensembled) attribute classifiers on text embeddings to predict attribute presence in images, enabling CLS-f_a measurements that correlate with human judgments.
• Evaluate cross-model correlations between auto GEP and human GEP using Kendall’s tau and MCC; compare auto methods (C-f_a, CC-f_a, CLS-f_a) against human annotations.

实验结果

研究问题

• RQ1 How do text-to-image models differ in presenting gendered attributes in generated images when prompted with gender-specific cues?
• RQ2 Can an automatic GEP estimation framework closely approximate human-annotated GEP and serve as a scalable evaluation tool?
• RQ3 Do GEP-based analyses reveal model-specific tendencies and occupation-related stereotypes in generated images?

主要发现

• GEP captures attribute-level gender presentation differences that align with common stereotypes such as dresses for women and suits for men in several models.
• Explicit prompts amplify presentation differences more than neutral prompts across models, with varying magnitudes by model.
• DALL-E 2 tends to show stronger gender-attribute coupling and higher image quality (as indicated by CLIP score) but exhibits different attribute correlations than other models.
• Automatic GEP estimators based on CLIP-derived features (C-f_a, CC-f_a, CLS-f_a) show stronger correlation with human GEP than raw CLIP similarity alone, indicating improved alignment with human judgments.
• GEP-based analysis generalizes to occupation-related stereotypes, demonstrating utility beyond broad gender categorization.