Abstract
Generative models are increasingly used to produce synthetic training data, yet they often replicate or amplify biases found in their training sets. We present Quality-Diversity Generative Sampling (QDGS), a model-agnostic framework designed to search a generator's latent space for diverse, intersectionally balanced datasets using language prompts. By framing data generation as a Quality-Diversity optimization problem, QDGS identifies high-quality samples across a multi-dimensional attribute space without requiring model fine-tuning. Our experiments demonstrate that training on QDGS-curated synthetic data significantly enhances downstream classifier fairness and accuracy, particularly for minority groups.
Introduction
Using synthetic data from generative models is an attractive, low-cost way to train downstream machine learning models. However, generative models inherit the biases of their training data (e.g., internet scraping), often over-representing dominant classes and under-representing minority groups. If a classifier is trained on this biased synthetic data, it inevitably inherits these biases, leading to poor performance on minority demographics.
While prior methods exist to debias synthetic data, they typically focus on a single dimension of diversity at a time (e.g., just skin tone or just age). Real-world fairness, however, requires intersectional balance—ensuring adequate representation of complex combinations like older, dark-skinned females. Randomly sampling from a generative model's latent space usually collapses into the dominant modes, missing these intersectional combinations entirely. This paper introduces Quality-Diversity Generative Sampling (QDGS) to address this gap.
Methodology: Quality-Diversity Generative Sampling
The authors frame synthetic data generation as a Quality-Diversity (QD) optimization problem. QDGS explores the latent space of a frozen generative model (such as StyleGAN2) to find a diverse set of high-quality outputs without updating the model's weights.
The framework operates through three primary components:
- Objective Function (Quality): The system uses CLIP to score how well a generated image matches a positive language prompt (e.g., "A detailed photo of an individual...") while penalizing matches to negative prompts (e.g., "An obscure, fake, or discolored photo...").
- Measure Functions (Diversity): Users define multiple dimensions of diversity using CLIP prompts (e.g., skin tone, age, hair length, gender presentation). These prompts serve as coordinates in a multi-dimensional grid.
- Latent Space Illumination: Using an evolutionary algorithm (CMA-MAEGA), QDGS searches the generator's latent space. It attempts to find latent vectors that fill every cell in the multi-dimensional grid (the "archive"), representing every possible combination of the defined diversity measures, while simultaneously maximizing the quality objective.
Experimental Setup
The researchers evaluated QDGS across several domains:
- Proof-of-Concept: A custom color-biased shapes dataset (e.g., heavily skewed toward red triangles and blue squares).
- Face Generation: Using StyleGAN2 trained on FFHQ (Flickr-Faces-HQ).
- Downstream Tasks: Training ResNet101 classifiers for facial recognition using datasets like CASIA-WebFace and BUPT-Balancedface, then evaluating on the Racial Faces in-the-Wild (RFW) benchmark.
Key Insights and Results
Latent Space Mining
Even if a generator is heavily biased, its high-dimensional latent space usually contains underrepresented attributes. They are simply statistically unlikely to be found via random sampling. QDGS actively mines these regions. As shown in the shapes experiment, QDGS provides far superior coverage of the attribute space compared to random sampling.
Performance and Fairness
- Shapes: QDGS successfully generated a balanced dataset from a heavily biased shape generator, improving downstream classifier accuracy on minority shape-color combinations by up to ~27%.
- Face Distribution: QDGS dramatically shifted the distribution of generated faces. Images recognized as having dark skin tones increased from 9.4% to 25.2%, and "old" faces increased from 37.0% to 45.7%.
- Fairness Benchmarks: When pretraining on the imbalanced CASIA dataset, adding QDGS synthetic data improved accuracy on dark-skinned faces in the RFW benchmark from 88.08% to 88.94%, while maintaining high accuracy on light-skinned faces. Models pretrained with QDGS achieved the highest average accuracy across five standard facial recognition benchmarks (LFW, CFPFP, CPLFW, CALFW, and AgeDB).
Discussion and Limitations
QDGS acts as a form of unmeasured data augmentation. By pushing the boundaries of diversity to fill the archive, it often generates images with extreme visual features (e.g., heavy neon lighting), which helps downstream models generalize to unseen domains.
However, the method is limited by the underlying generator; if a demographic is entirely absent from the training data, QDGS cannot recover it. Furthermore, the reliance on CLIP prompts introduces the risk of injecting linguistic biases or reinforcing stereotypes during the diversity definition phase. Future work will explore targeting specific, non-uniform distributions to match real-world population demographics more accurately.
Summary updated