Decomposed evaluations of geographic disparities in text-to-image models

Decomposed Evaluations of Geographic Disparities in Text-to-Image Models

The evaluated paper introduces Decomposed Indicators of Disparities in Image Generation (Decomposed-DIG) as a novel set of metrics that scrutinize geographic disparities in text-to-image generation systems. This work builds on existing metrics by disentangling geographic disparities into object and background representations within generated images.

Introduction and Motivation

Recent advancements in text-to-image generative models, particularly latent diffusion models (LDMs), have revolutionized visual content creation. However, increasing scrutiny has exposed significant geographic biases in these models, such as over-representation of stereotypical imagery and underrepresentation of certain geographic regions. Traditional evaluation metrics fail to attribute these disparities to specific components of generated images, prompting the need for more granular analysis.

Methodology: Decomposed-DIG

The paper proposes Decomposed-DIG, which extends the precision and coverage metrics from earlier works [4] to separately measure disparities in objects and backgrounds. The procedure involves:

1. Object and Background Segmentation: Utilizes the Segment Anything Model (SAM) in conjunction with GroundingDINO's object detection to achieve precise segmentation.
2. Feature Extraction: Employs a Vision Transformer (ViT) for feature extraction, isolating object and background features based on segmented regions.
3. Benchmarking: Computes precision (as a proxy for realism) and coverage (as a proxy for diversity) for object-specific and background-specific segments.

Key Findings

Using Decomposed-DIG, the study evaluates a widely-used LDM and uncovers nuanced patterns of geographic disparities:

1. Realism and Diversity: Generated objects exhibit higher realism compared to backgrounds. Object representations are more consistent across regions than backgrounds, which show approximately 1.7 times larger disparities between geographic regions.
2. Specific Disparities: The paper provides detailed, qualitative insights. For example, backgrounds in African regions often lack modern infrastructure like paved streets, and generated images frequently place objects in unrealistic settings (e.g., cooking pots outdoors in Europe).

Mitigations via Prompt Engineering

Building on these findings, the authors explore an early mitigation strategy through prompt engineering. By modifying the prompt structure from "{object} in {region}" to "{regional adjective} {object}", the study achieves notable improvements:

• A 52% enhancement in background diversity for the worst-performing region (Africa).
• A 20% average improvement in background diversity without significantly compromising realism or object diversity.

Implications and Future Work

The introduction of Decomposed-DIG offers a more granular and interpretable framework for evaluating geographic disparities in text-to-image models. This work underscores the importance of distinguishing between object and background disparities, thus making subsequent bias identification and mitigation strategies more effective.

Future developments could focus on extending this methodology to other aspects of image generation, such as color biases or temporal disparities. Moreover, incorporating Decomposed-DIG into the training process could directly influence model development, promoting fairness and inclusivity in generative AI systems.

By isolating object and background biases, the paper facilitates a deeper understanding of where and why disparities arise, guiding the development of more geographically equitable text-to-image models.

Conclusion

The paper presents a significant step forward in the nuanced evaluation of geographic disparities in text-to-image models. Decomposed-DIG not only clarifies the origins of these disparities but also provides a solid foundation for developing targeted mitigation strategies. This methodology invites further research into decomposed evaluations, potentially extending beyond geographic disparities to other socio-demographic factors, thus broadening the scope of equitable AI.