Gender Bias Evaluation in Text-to-image Generation: A Survey

Abstract: The rapid development of text-to-image generation has brought rising ethical considerations, especially regarding gender bias. Given a text prompt as input, text-to-image models generate images according to the prompt. Pioneering models such as Stable Diffusion and DALL-E 2 have demonstrated remarkable capabilities in producing high-fidelity images from natural language prompts. However, these models often exhibit gender bias, as studied by the tendency of generating man from prompts such as "a photo of a software developer". Given the widespread application and increasing accessibility of these models, bias evaluation is crucial for regulating the development of text-to-image generation. Unlike well-established metrics for evaluating image quality or fidelity, the evaluation of bias presents challenges and lacks standard approaches. Although biases related to other factors, such as skin tone, have been explored, gender bias remains the most extensively studied. In this paper, we review recent work on gender bias evaluation in text-to-image generation, involving bias evaluation setup, bias evaluation metrics, and findings and trends. We primarily focus on the evaluation of recent popular models such as Stable Diffusion, a diffusion model operating in the latent space and using CLIP text embedding, and DALL-E 2, a diffusion model leveraging Seq2Seq architectures like BART. By analyzing recent work and discussing trends, we aim to provide insights for future work.

• The paper provides a comprehensive survey of evaluation methodologies that assess gender bias in text-to-image models.
• It details the use of prompt design, attribute classification, and various bias metrics to analyze both context-to-gender and gender-to-context biases.
• Findings reveal a male-dominant trend in professional depictions while underscoring the need for robust bias mitigation strategies.

Gender Bias Evaluation in Text-to-Image Generation: A Survey

Introduction

The paper "Gender Bias Evaluation in Text-to-image Generation: A Survey" investigates the ethical considerations surrounding text-to-image generation models, particularly with respect to gender bias. As prominent models like Stable Diffusion and DALL-E 2 continue to advance, they face significant scrutiny for perpetuating biases, notably gender biases, evident through tendencies like recurring associations of particular genders with professions. This survey critiques the existing literature on gender bias evaluation, focusing on the setup of these studies, the metrics used, and prevailing findings, aiming to illuminate paths for future endeavors.

Bias Evaluation Setup

The evaluation of gender bias within text-to-image models involves key methodological considerations: the definitions of gender and bias, prompt design, and attribute classification.

Gender Definition: Most research dichotomizes gender into binary categories—female/woman and male/man. Nonetheless, some investigations expand this to include non-binary and neutral genders, addressing an otherwise overlooked demographic.

Bias Definition: Two primary types of gender bias are identified—context-to-gender bias and gender-to-context bias. Context-to-gender bias surfaces when gender-neutral prompts disproportionately yield images of certain genders. Gender-to-context bias emerges when gendered words influence contextual elements like backgrounds or objects.

Prompt Design: Template-based prompts, such as "a photo of [DESCRIPTION]", dominate the evaluation methods. Prompts may encapsulate professions, adjectives, or activities, enabling comprehensive bias investigations. Additionally, LLMs are becoming instrumental in generating diverse prompts.

Attribute Classification: Assigning gender to generated images often involves gender classifiers focused on facial features, or embeddings evaluated against text sentences like "a photo of a woman/man". Human annotations play a supplementary role, especially when automated methods fall short.

Bias Evaluation Metrics

Metrics employed to evaluate gender bias are categorized into distribution metrics, bias tendency metrics, and quality metrics.

Distribution Metrics: Measures like the Mean Absolute Deviation and chi-square tests assess disparities between detected and idealized attribute distributions. These tools are pivotal for quantifying context-to-gender bias.

Bias Tendency Metrics: These metrics ascertain whether attributes disproportionately favor a gender. Proportion calculations vis-à-vis real-world data reveal amplification or mitigation of societal biases. Novel approaches like Stereotype and Neutrality Scores expand traditional binary assessments.

Quality Metrics: While bias metrics are critical, image generation quality metrics like CLIPScore and FID ensure that the generated images maintain semantic coherence and visual fidelity, independent of bias discussions.

Findings and Trends

Repeated evaluations show text-to-image models frequently produce images skewed towards male representation in professional settings. Bias extends to attire, suggesting deeper rooted societal stereotypes. Furthermore, emerging research highlights the proliferation of such biases beyond human depiction, influencing contextual elements in generated images.

A notable trend is the expanding scope of evaluations, incorporating varied models and more nuanced axes of bias assessment. This advancement aims to provide comprehensive insights, fostering development of bias mitigation strategies.

Conclusion

The paper successfully surveys current methodologies and findings on gender bias in text-to-image generation models. It accentuates the necessity of robust evaluation frameworks, precise metrics, and continuous examination of prevailing biases, to inform future research and the ethical deployment of such models. Continued efforts in standardized definitions and cross-disciplinary collaborations could greatly elevate the fairness and inclusivity of these generative technologies.