MD-EGAN: Evolutionary GAN with dynamic latent sampling and relative adaptive discriminator for improved performance

Training instability, mode collapse, vanishing gradients, and high computational cost are significant challenges in generative adversarial networks (GANs), particularly in evolutionary-based GANs. To address these issues, we propose the multi-distribution evolutionary GAN (MD-EGAN), a method aimed at improving training stability, enhancing sample diversity, and accelerating convergence. MD-EGAN leverages multiple latent space priors—including Gaussian, Uniform, Poisson, and Truncated Gaussian distributions paired with a relative adaptive discriminator (RAD) that provides dynamic and comparative feedback. By exploring diverse latent distributions and RAD feedback, MD-EGAN enables more robust population-based generator evolution. Experimental evaluations on the CIFAR-10 and STL-10 demonstrate that MD-EGAN outperforms several baseline GANs models in both image quality and diversity. Specifically, MD-EGAN achieves inception score (IS) = 8.92, and Fréchet inception distance (FID) = 10.08 on CIFAR-10 and IS = 10.31, and FID = 21.93 on STL-10. Meanwhile, MD-EGAN reduces convergence time by up to 43.16 % when compared with cooperative dual evolutionary GAN, demonstrating significant improvement in computational efficiency. These results validate the effectiveness of multi-distribution latent modeling and relative feedback in addressing key limitations of GAN training, leading to improved generative performance.