QR-GAN: Generative Adversarial Networks meet Quantile Regression

This study proposes a novel Quantile Regression Generative Adversarial Networks (QR-GAN), which uses quantile regression to improve GAN performance by mitigating issues of GAN such as non-convergence, mode collapse, and gradient problems. Instead of directly minimizing the 1-Wasserstein distance between real and generated data distributions, as in WGANs, the proposed method leverages the quantile loss to encourage the generator to produce a diverse range of outputs. Additionally, we conduct an in-depth analysis of the output space of the discriminator and the gradients of fake samples to identify the root causes of mode collapse. Our study reveals that constraining the discriminator to specific values can lead to mode collapse. The proposed QR-GAN method exhibits a high level of robustness against mode collapse and significantly improves the generation performance, as measured by the Frechet Inception Distance (FID) metric, compared to other GAN variants. These findings suggest that the proposed QR-GAN method has the potential to advance the state-of-the-art in GAN research and improve their applicability in a variety of domains.