Comprehensive Survey of Biases in Text-to-Image Generation Models: Definitions, Evaluations, and Mitigation Strategies

Core Concepts

Text-to-image generation models exhibit significant biases in their outputs, leading to harmful stereotyping and under-representation of minority groups. This survey comprehensively reviews prior studies on defining, evaluating, and mitigating biases in these models across three key dimensions: gender, skin tone, and geo-cultural representation.

Abstract

This survey provides an extensive review of prior research on biases in text-to-image (T2I) generation models. The authors categorize bias definitions along three primary dimensions studied in previous works: gender presentation bias, skin tone bias, and geo-cultural bias.
Gender Bias:

Default Generation Bias: Models tend to depict a particular gender when given gender-neutral prompts.
Occupational Association Bias: Models over-represent or under-represent a particular gender for certain occupations.
Characteristics and Interests Bias: Models associate certain characteristics or interests with specific genders.
Stereotypical Objects Bias: Models generate gender-stereotypical objects.
Image Quality Bias: Models generate higher-quality images for a particular gender.
Power Dynamics Bias: Models depict stereotypical genders for prompts indicating power dynamics.
NSFW & Explicit Content Bias: Models disproportionately generate NSFW or explicit content for non-cisgender individuals.
Skin Tone Bias:

Default Generation Bias: Models tend to depict individuals of a certain skin tone when skin tone is not specified.
Occupational Association Bias: Models over-represent or under-represent skin tone groups for certain occupations.
Characteristics and Interests Bias: Models associate certain characteristics with specific skin tones.
Geo-Cultural Bias:

Geo-Cultural Norms Bias: Models over-represent specific cultures in default generations while under-representing others.
Characteristics and Interests Bias: Models depict certain cultures with harmful stereotypes.
The survey also discusses the landscape of evaluation datasets and metrics used in prior works, highlighting the lack of unified frameworks. Additionally, it reviews current mitigation approaches, including model weight refinement and inference-time/data-based methods, and identifies their limitations.
Based on these insights, the authors propose future research directions towards human-centric bias definition, evaluation, and mitigation approaches to build fair and trustworthy T2I technologies.

Stats

"Text-to-Image (T2I) models generate accurate images according to textual prompts. As modern T2I systems such as OpenAI's DALLE-3 [74] quickly advance in generation quality and prompt-image alignment, many applications in real-world scenarios have been made possible."
"Prior works have unveiled severe biases in this depiction [21, 11, 33]. For instance, a version of Stable Diffusion [85] portrayed the world as being run by white masculine CEOs; dark-skinned men are depicted to be committing crimes, while dark-skinned women are delineated to be flipping burgers [72]."
"T2I models' worldviews are extremely biased, failing to represent the world's diversity of genders, racial groups, and cultures [60]."

Quotes

"Such bias might induce false convictions in real world, if the model is applied to help sketch suspected offenders [72, 79]."
"Biases can be propagated by users [68] who are unaware of T2I models' underlying issues and trust their output [31]."
"Noticing the problem with bias in T2I models, researchers have made efforts to identify different aspects of bias-related risks [11, 46], as well as develop methods for evaluating and mitigating such issues."

Key Insights Distilled From

Survey of Bias In Text-to-Image Generation

by Yixin Wan,Ar... at arxiv.org 04-02-2024

https://arxiv.org/pdf/2404.01030.pdf

Deeper Inquiries

How can we ensure that bias mitigation approaches in text-to-image generation models are robust, controllable, and resource-friendly, while also being aligned with diverse human and community preferences?

To ensure that bias mitigation approaches in text-to-image generation models are robust, controllable, and resource-friendly, while also aligning with diverse human and community preferences, several strategies can be implemented:

Robustness:

Implementing diverse mitigation techniques to address a wide range of biases, including gender, skintone, and geo-cultural biases.
Conducting thorough testing and validation to ensure that the mitigation strategies are effective across various scenarios and datasets.

Controllability:

Providing users with options to customize the bias mitigation process based on their preferences and requirements.
Incorporating adjustable parameters in the mitigation algorithms to allow for fine-tuning and control over the mitigation process.

Resource-Friendly:

Developing lightweight algorithms that do not require excessive computational resources for bias mitigation.
Optimizing the mitigation process to be efficient and scalable, especially when dealing with large datasets and complex models.

Alignment with Diverse Preferences:

Engaging with diverse communities and stakeholders to understand their preferences and concerns regarding bias in text-to-image generation.
Incorporating feedback mechanisms to continuously adapt the mitigation strategies based on evolving community needs and values.

How can we address the potential risks and limitations of using automated classification methods (e.g., VQA models) for evaluating biases in text-to-image generation?

Addressing the risks and limitations of using automated classification methods for evaluating biases in text-to-image generation models involves the following steps:

Bias Awareness:

Acknowledging the inherent biases in automated classification models and understanding their potential impact on bias evaluation.

Transparency:

Providing transparency on the limitations and biases of the classification models used for bias evaluation.
Clearly communicating the assumptions and constraints of the automated classification methods to stakeholders.

Validation:

Validating the performance of automated classification methods against diverse and representative datasets to ensure unbiased evaluations.
Conducting regular audits and checks to identify and mitigate any biases introduced by the classification models.

Human Oversight:

Incorporating human oversight and validation in the bias evaluation process to complement the automated classification methods.
Leveraging human annotators to verify and validate the results obtained from automated classification models.

How can we design adaptive and lifelong bias mitigation strategies that evolve based on continuous feedback and the changing landscape of societal norms and values, as well as community needs?

Designing adaptive and lifelong bias mitigation strategies involves the following approaches:

Continuous Monitoring:

Implementing systems to continuously monitor model outputs and detect biases in real-time.
Collecting feedback from users and stakeholders to identify emerging biases and societal norms.

Feedback Loops:

Establishing feedback loops to gather input from diverse communities and incorporate their preferences into the mitigation strategies.
Using feedback mechanisms to adapt and refine bias mitigation techniques based on changing societal values and community needs.

Dynamic Adjustment:

Developing algorithms that can dynamically adjust bias mitigation strategies based on feedback and evolving societal norms.
Incorporating mechanisms for self-correction and adaptation to ensure that the mitigation strategies remain effective over time.

Community Engagement:

Engaging with communities and stakeholders to understand their evolving needs and values related to bias mitigation.
Collaborating with diverse groups to co-create and refine bias mitigation strategies that are aligned with community preferences and values.