CGE-GAN: Contrastive-guided evolutionary generative adversarial networks with dynamic adaptive weight sharing

Generative adversarial networks (GANs) have achieved remarkable success in image synthesis but faces major challenges, including mode collapse, training instability, and inefficient architecture search. Existing evolutionary GANs partially address these issues but lack semantic alignment with real data, effective weight reuse, and knowledge transfer between model generations. To overcome these limitations, we propose contrastive-guided evolutionary GANs (CGE-GAN)-a unified method introduces a novel hybrid Wasserstein-Contrastive loss function that drives generators to align semantically with real data while maintaining adversarial competitiveness. Besides, we incorporated dynamic adaptive weight sharing (DAWS) for efficient training and knowledge distillation-based crossover to preserve useful features across generations. The CGE-GAN is evaluated on CIFAR-10 and STL-10, and it achieves an Inception Score (IS) of 8.99 and 10.46, and fréchet inception distance (FID) of 9.74 and 21.86, respectively. Compared to strong baselines, CGE-GAN reduces FID by up to 1.74 points while maintaining high semantic diversity and convergence efficiency with only 0.36 GPU days. These results highlight the effectiveness of contrastive-driven evolution for generating stable and high-fidelity outputs.