When Generative Augmentation Hurts: A Benchmark Study of GAN and Diffusion Models for Bias Correction in AI Classification Systems

Generative models are widely used to compensate for class imbalance in AI training pipelines, yet their failure modes under low-data conditions are poorly understood. This paper reports a controlled benchmark comparing three augmentation strategies applied to a fine-grained animal classification task: traditional transforms, FastGAN, and Stable Diffusion 1.5 fine-tuned with Low-Rank Adaptation (LoRA). Using the Oxford-IIIT Pet Dataset with eight artificially underrepresented breeds, we find that FastGAN augmentation does not merely underperform at very low training set sizes but actively increases classifier bias, with a statistically significant large effect across three random seeds (bias gap increase: +20.7%, Cohen's d = +5.03, p = 0.013). The effect size here is large enough to give confidence in the direction of the finding despite the small number of seeds. Feature embedding analysis using t-distributed Stochastic Neighbor Embedding reveals that FastGAN images for severe-minority breeds form tight isolated clusters outside the real image distribution, a pattern consistent with mode collapse. Stable Diffusion with Low-Rank Adaptation produced the best results overall, achieving the highest macro F1 (0.9125 plus or minus 0.0047) and a 13.1% reduction in the bias gap relative to the unaugmented baseline. The data suggest a sample-size boundary somewhere between 20 and 50 training images per class below which GAN augmentation becomes harmful in this setting, though further work across additional domains is needed to establish where that boundary sits more precisely. All experiments run on a consumer-grade GPU with 6 to 8 GB of memory, with no cloud compute required.

Adjusting for Bias with Procedural Data

3D softwares are now capable of producing highly realistic images that look nearly indistinguishable from the real images. This raises the question: can real datasets be enhanced with 3D rendered data? We investigate this question. In this paper we demonstrate the use of 3D rendered data, procedural, data for the adjustment of bias in image datasets. We perform error analysis of images of animals which shows that the misclassification of some animal breeds is largely a data issue. We then create procedural images of the poorly classified breeds and that model further trained on procedural data can better classify poorly performing breeds on real data. We believe that this approach can be used for the enhancement of visual data for any underrepresented group, including rare diseases, or any data bias potentially improving the accuracy and fairness of models. We find that the resulting representations rival or even out-perform those learned directly from real data, but that good performance requires care in the 3D rendered procedural data generation. 3D image dataset can be viewed as a compressed and organized copy of a real dataset, and we envision a future where more and more procedural data proliferate while datasets become increasingly unwieldy, missing, or private. This paper suggests several techniques for dealing with visual representation learning in such a future.