Harnessing Diffusion-Generated Synthetic Images for Fair Image Classification

Dataset bias stands as a fundamental data imbalance problem, where uneven group distributions cause image classification models to rely on misleading correlations rather than true target features. When a training dataset disproportionately associates specific attributes with certain demographic groups, classifiers inherit these biases and perform poorly on underrepresented populations. While traditional algorithmic debiasing methods attempt to correct this during model training, their effectiveness degrades significantly as the severity of the data imbalance increases, highlighting the need for a solution that addresses the data distribution itself.

To mitigate this limitation, a team of researchers—Abhipsa Basu and Venkatesh Babu Radhakrishnan (from the Vision & AI Lab in the Department of Computational and Data Sciences, Indian Institute of Science), Aviral Gupta (from Birla Institute of Technology & Science Pilani), and Abhijnya Bhat (from Stanford University)—introduced a generative framework that tackles the root cause of dataset bias by using fine-tuned diffusion models to create group-balanced training datasets. Instead of relying on vanilla stable diffusion, which often fails to capture the precise characteristics of a specific dataset, the methodology leverages personalised text-to-image techniques like LoRA and DreamBooth to generate high-quality synthetic images for minority groups. To handle heavy intra-group variation such as men with both long and short hair, the researchers developed Clustered DreamBooth, a technique that groups similar images within a biased category prior to fine-tuning, allowing the diffusion model to learn distinct sub-distributions more effectively.

The framework operates via a robust two-stage training pipeline. First, an image classifier is pre-trained on the group-balanced synthetic dataset to establish unbiased feature representations across all demographic groups. Second, the model is fine-tuned on the original real-world data to ensure its feature extraction aligns with authentic real-world distributions. This strategy effectively combines the structural fairness of generative data augmentation with the high-fidelity nuances of real samples.

Extensive experimental evaluations demonstrate that this fine-tuned generative approach significantly outperforms standard text-to-image baselines. Furthermore, while the method performs comparably to state-of-the-art algorithmic debiasing techniques under normal dataset imbalances, it vastly outperforms them as the severity of the dataset bias increases, proving highly resilient in extreme, data-starved scenarios.

This research work was presented at the AAAI Conference on Artificial Intelligence (AAAI 2026).

Paper: https://arxiv.org/pdf/2511.08711

MENU