Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deconstructing Bias: A Multifaceted Framework for Diagnosing Cultural and Compositional Inequities in Text-to-Image Generative Models

Muna Numan Said, Aarib Zaidi, Rabia Usman, Sonia Okon, Praneeth Medepalli, Kevin Zhu, Vasu Sharma, Sean O'Brien

TL;DR

The paper tackles cultural bias in text-to-image generation by introducing the Component Inclusion Score (CIS), a metric that quantifies compositional fragility and contextual misalignment across prompts. Analyzing 2,400 images, the authors show significant Western–non-Western disparities and identify data imbalance in LAION-5B, elevated cross-attention entropy, and embedding superposition as key drivers. They benchmark multiple models (e.g., Stable Diffusion v2.1, Realistic Vision, Dreamlike Photoreal) and demonstrate that CIS captures biases not reflected by traditional metrics like FID or CLIP scores (with $p<0.001$ for some non-Western prompts). The work offers actionable, architecture- and data-centric guidance to diagnose and mitigate biases in T2I systems, advancing toward more equitable generative AI. CIS serves as a practical diagnostic tool for researchers and practitioners aiming to improve cultural fidelity in AI-generated imagery.

Abstract

The transformative potential of text-to-image (T2I) models hinges on their ability to synthesize culturally diverse, photorealistic images from textual prompts. However, these models often perpetuate cultural biases embedded within their training data, leading to systemic misrepresentations. This paper benchmarks the Component Inclusion Score (CIS), a metric designed to evaluate the fidelity of image generation across cultural contexts. Through extensive analysis involving 2,400 images, we quantify biases in terms of compositional fragility and contextual misalignment, revealing significant performance gaps between Western and non-Western cultural prompts. Our findings underscore the impact of data imbalance, attention entropy, and embedding superposition on model fairness. By benchmarking models like Stable Diffusion with CIS, we provide insights into architectural and data-centric interventions for enhancing cultural inclusivity in AI-generated imagery. This work advances the field by offering a comprehensive tool for diagnosing and mitigating biases in T2I generation, advocating for more equitable AI systems.

Deconstructing Bias: A Multifaceted Framework for Diagnosing Cultural and Compositional Inequities in Text-to-Image Generative Models

TL;DR

The paper tackles cultural bias in text-to-image generation by introducing the Component Inclusion Score (CIS), a metric that quantifies compositional fragility and contextual misalignment across prompts. Analyzing 2,400 images, the authors show significant Western–non-Western disparities and identify data imbalance in LAION-5B, elevated cross-attention entropy, and embedding superposition as key drivers. They benchmark multiple models (e.g., Stable Diffusion v2.1, Realistic Vision, Dreamlike Photoreal) and demonstrate that CIS captures biases not reflected by traditional metrics like FID or CLIP scores (with for some non-Western prompts). The work offers actionable, architecture- and data-centric guidance to diagnose and mitigate biases in T2I systems, advancing toward more equitable generative AI. CIS serves as a practical diagnostic tool for researchers and practitioners aiming to improve cultural fidelity in AI-generated imagery.

Abstract

The transformative potential of text-to-image (T2I) models hinges on their ability to synthesize culturally diverse, photorealistic images from textual prompts. However, these models often perpetuate cultural biases embedded within their training data, leading to systemic misrepresentations. This paper benchmarks the Component Inclusion Score (CIS), a metric designed to evaluate the fidelity of image generation across cultural contexts. Through extensive analysis involving 2,400 images, we quantify biases in terms of compositional fragility and contextual misalignment, revealing significant performance gaps between Western and non-Western cultural prompts. Our findings underscore the impact of data imbalance, attention entropy, and embedding superposition on model fairness. By benchmarking models like Stable Diffusion with CIS, we provide insights into architectural and data-centric interventions for enhancing cultural inclusivity in AI-generated imagery. This work advances the field by offering a comprehensive tool for diagnosing and mitigating biases in T2I generation, advocating for more equitable AI systems.
Paper Structure (24 sections, 2 equations, 3 figures, 3 tables)

This paper contains 24 sections, 2 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. The Bias Amplification Challenge
  3. Root Causes of Bias
  4. Limitations of Current Approaches
  5. Metric Gaps: The Phantom of Objectivity
  6. Architectural Blind Spots: Symptomatic Solutions
  7. The Missed Nexus: Data, Architecture, and Culture
  8. Our Contributions
  9. Related Work
  10. Bias in Generative Models
  11. Evaluation Metrics and Diagnostic Tools
  12. Architectural Drivers of Bias
  13. Methodology
  14. Component Inclusion Score (CIS)
  15. Experimental Design
  16. ...and 9 more sections

Figures (3)

  • Figure 1: Figure showing the cross-attention entropy for mainstream and marginalized pairs across transformer attention layers. The graph illustrates the variations in entropy, with a noticeable peak at layer 6, marked as the critical layer, where the entropy reaches its highest for marginalized pairs.
  • Figure 2: The framework of the CIS metric.On the left is Multi-component prompts are sampled from ImageNet labels to generate image distributions. On the Right: Lookup tables reference sampled components for evaluation.
  • Figure 3: Comparison of images generated by different models of a Bangladeshi flag on a fishing boat and Canadian flag with their respective CIS evaluation