Multi-Group Proportional Representation for Text-to-Image Models

Conference: CVPR 2025
arXiv: 2505.24023
Code: None
Area: Diffusion Models / AI Safety
Keywords: Text-to-Image, Fairness, Proportional Representation, Intersectional Groups, Bias Mitigation

TL;DR

This paper proposes the Multi-Group Proportional Representation (MPR) metric to systematically measure representational bias across intersectional demographic groups in text-to-image models, and develops an optimization algorithm based on this metric to guide T2I models toward more balanced group representation while preserving generation quality.

Background & Motivation

Background: Text-to-image (T2I) generative models (such as Stable Diffusion, DALL-E, Midjourney, etc.) are already capable of generating photorealistic images from text descriptions and are widely applied in fields like content creation and advertising design. With the large-scale deployment of these models, their social impact has garnered increasing attention.

Limitations of Prior Work: When generating human-centric images, T2I models systematically amplify demographic biases present in the training data—over-representing certain groups (e.g., young white men) while marginalizing minority groups (e.g., elderly women, specific intersectional racial groups). Such representational bias can propagate stereotypes and marginalize minority populations. Despite the widespread attention on "Responsible AI," there is currently a lack of a systematic methodology to (1) quantify representational bias across intersectional groups in T2I models, and (2) optimize fairness during training.

Key Challenge: Existing fairness metrics typically focus on a single demographic attribute (such as gender only or race only), ignoring the issue of intersectionality. For example, a model might appear balanced in both its gender and racial distributions but still severely under-represent intersectional groups like "African American women." Single-attribute fairness metrics fail to capture such intersectional biases.

Goal: (1) Define a flexible and configurable representational metric that accounts for intersectional groups; (2) develop a training algorithm capable of optimizing this metric; (3) verify that this approach improves representation while maintaining image quality.

Key Insight: Borrowing the concept of "proportional representation" from computational social choice theory and adapting it to T2I model evaluation. Instead of forcing all groups to appear in equal proportions (which may not align with reality), it allows users to specify target distributions according to specific scenarios.

Core Idea: Use the "worst-case group deviation" as the core of the representational metric, ensuring that no intersectional group is severely under-represented.

Method

Overall Architecture

The MPR framework consists of two components: (1) the MPR metric, which calculates the maximum deviation between the representation of all predefined intersectional groups and a target distribution for a given set of images generated by a T2I model; (2) the MPR optimization algorithm, which uses the MPR metric as a training objective to minimize worst-case deviation by fine-tuning the text-conditioning modules of the T2I model. The workflow is: generate a batch of images \(\rightarrow\) annotate demographic attributes using an attribute classifier \(\rightarrow\) calculate MPR \(\rightarrow\) backpropagate gradients to update the model.

Key Designs

1. MPR度量指标 (Multi-Group Proportional Representation Metric):

   • Function: Quantify representational bias of T2I models across intersectional demographic groups.
   • Mechanism: Define the demographic group space as the Cartesian product of multiple attributes (e.g., gender, race, age), forming a set of intersectional groups \(\mathcal{G}\) (e.g., "young Black women," "elderly Asian men"). For each group \(g \in \mathcal{G}\), calculate the deviation between its actual proportion in the generated images \(\hat{p}(g)\) and the target proportion \(p^*(g)\). MPR is defined as the maximum relative deviation across all groups: \(\text{MPR} = \max_{g \in \mathcal{G}} |\hat{p}(g) - p^*(g)| / p^*(g)\). The target proportion \(p^*(g)\) can be specified by the user based on application scenarios (such as census data, uniform distribution, etc.).
   • Design Motivation: Adopting a worst-case design ensures that no group is marginalized. Even if the average representation appears acceptable, the MPR value will be high if a single intersectional group is severely under-represented. This reflects fairness issues more accurately than average deviation.

2. 属性分类与群体标注 (Attribute Classification Pipeline):

   • Function: Automatically annotate demographic attributes of individuals in generated images.
   • Mechanism: Use pre-trained face detection and attribute classification models to detect faces in generated images and predict attributes such as gender, perceived race, and age group. These attribute predictions are combined into intersectional group labels. To handle classifier uncertainty, soft labels (probability distributions) are used instead of hard labels, making subsequent MPR calculations more robust.
   • Design Motivation: Manual annotation of a large number of generated images is impractical, necessitating an automated pipeline. Using soft labels mitigates the impact of the attribute classifier's own biases on the MPR calculation.

3. MPR优化算法 (MPR Optimization):

   • Function: Fine-tune the T2I model to generate more balanced group representations.
   • Mechanism: Cast MPR as a differentiable training loss function. In each training step: (a) generate a batch of images using the current model; (b) estimate the group distribution using the attribute classifier; (c) calculate the MPR loss; (d) backpropagate gradients to update the model's text encoder or cross-attention layers. To maintain generation quality, the original diffusion loss is retained and weighted. A key technique is to apply a soft relaxation on the max operation (using a log-sum-exp approximation) to allow gradients to flow to multiple groups with large deviations.
   • Design Motivation: Directly optimizing the worst-case deviation ensures fairness improvements more effectively than optimizing the average deviation. Soft relaxation avoids the gradient vanishing problem of the max operation, making the training process more stable.

Loss & Training

The total training loss is: \(\mathcal{L} = \mathcal{L}_{diff} + \alpha \cdot \mathcal{L}_{MPR}\), where \(\mathcal{L}_{diff}\) is the standard diffusion denoising loss to preserve generation quality, and \(\mathcal{L}_{MPR}\) is the MPR loss to drive balanced representation. Only the text encoder and cross-attention parameters are fine-tuned, while the UNet backbone is frozen to maintain visual generation capability.

Key Experimental Results

Main Results

Model	MPR Before Optimization↓	MPR After Optimization↓	Change in FID	Change in CLIP Score
Stable Diffusion v1.5	0.72	0.31	+1.2	-0.3
Stable Diffusion v2.1	0.68	0.28	+0.9	-0.2
SDXL	0.61	0.24	+1.5	-0.4

Ablation Study

Analysis Dimension	Configuration	MPR↓	Description
Target Distribution Choice	Uniform Distribution	0.28	Equal proportions for all groups
Target Distribution Choice	Census Distribution	0.22	According to real-world demographic proportions
Group Granularity	Single Attribute (Gender Only)	0.15	Coarse-grained is easier to optimize
Group Granularity	Two Attributes (Gender + Race)	0.28	Intersectional groups are more challenging
Group Granularity	Three Attributes (Gender + Race + Age)	0.41	More groups make optimization harder
Soft vs Hard Max	Soft Relaxation	0.28	More stable training
Soft vs Hard Max	Hard Max	0.35	Unstable gradients

Key Findings

• MPR optimization can reduce representational bias by 55-60% while only increasing FID by 1-2 points and keeping the CLIP Score almost unchanged, demonstrating that fairness improvement and generation quality can be largely balanced.
• Bias in intersectional groups is far greater than single-attribute bias: even if gender and race are individually close to balanced, the representation of intersectional groups (e.g., "elderly African American women") can still be exceptionally low.
• Fine-grained groups (considering more attribute intersections) are harder to optimize, yet the MPR framework still yields significant improvements.
• The census distribution is easier to optimize than the uniform distribution (MPR of 0.22 vs. 0.28) because the former is closer to the natural distribution of the training data.
• Soft max relaxation is crucial for training stability: the hard max version of the MPR loss concentrates gradients on a single group, causing severe training oscillations.

Highlights & Insights

• Introducing social choice theory into T2I fairness is an insightful cross-domain transfer: the concept of proportional representation is naturally suited to quantify "who is represented and who is ignored," offering greater interpretability than simple statistical distances.
• The worst-case design is a key highlight: in reality, the most severe harms often occur in the most marginalized groups. The max design of MPR ensures attention to these groups, reflecting fairness more effectively than average metrics.
• The flexibility of the framework is worth emphasizing: MPR allows users to customize target distributions and group divisions, making it adaptable to fairness requirements across different cultural and legal contexts rather than being a one-size-fits-all solution.

Limitations & Future Work

• The attribute classifiers have inherent biases of their own (particularly lower identification accuracy on intersectional groups), which affects the accuracy of MPR calculations and introduces systematic errors.
• The work focuses solely on visual-level representation (facial attributes) and does not address stereotypes in textual descriptions (such as the gender associations of "doctor" vs. "nurse").
• Experiments focus on facial attributes (gender, race, age) and do not cover other protected attributes (disability, body type, etc.).
• Optimized models may show unnatural "forced diversity" effects on specific prompts, requiring further user experience evaluation.
• MPR optimization requires iterative cycles of generation and classification, incurring high computational overhead and making it unsuitable for real-time deployment.

Related Work & Insights

• vs Fair Diffusion / Inclusive T2I: Prior fairness research usually considers only a single attribute or uses simple resampling strategies. MPR is the first to systematically address intersectional group fairness.
• vs DALL-E 3 system card: Commercial models inject diversity via system prompt engineering (e.g., automatically appending demographic descriptions), which is a hack rather than a fundamental solution. MPR provides a quantifiable and optimizable framework.
• vs Demographic Parity / Equal Opportunity: These are classic fairness metrics for classification models. MPR adapts similar concepts to generative model scenarios.

Rating

• Novelty: ⭐⭐⭐⭐ Systematically introduces the concept of proportional representation into T2I fairness measurement, with strong focus on intersectional groups.
• Experimental Thoroughness: ⭐⭐⭐⭐ Experiments span multiple SD models and Configurations, offering comprehensive dimensions of analysis.
• Writing Quality: ⭐⭐⭐⭐ Clear formal definitions and well-supported motivation.
• Value: ⭐⭐⭐⭐ Provides a standardized framework for the evaluation and optimization of T2I model fairness, carrying practical significance.

Related Papers

• [CVPR 2025] Generative Modeling of Class Probability for Multi-Modal Representation Learning
• [CVPR 2025] MCCD: Multi-Agent Collaboration-based Compositional Diffusion for Complex Text-to-Image Generation
• [CVPR 2025] CLIP Under the Microscope: A Fine-Grained Analysis of Multi-Object Representation
• [NeurIPS 2025] FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models
• [CVPR 2025] Compass Control: Multi Object Orientation Control for Text-to-Image Generation