Abstract: This paper proposes a new framework, called "Dual-Cycle GAN(DUC-GAN)" to enhance the stability of training in generative adversarial networks (GAN). The framework addresses the issue of training instability in GAN by introducing an additional cycle between the generator and discriminator. This new cycle is composed of a frozen original discriminator and a new discriminator, both of which are trained together with the generator and switched based on the generator's performance. Testing on multiple datasets has shown that the proposed framework significantly improves the performance and training stability of GAN compared to existing methods, achieving faster convergence and better generation quality. The framework is also applicable to other GAN variants and is expected to become an important tool for future GAN research.

Key words: generative adversarial networks, dual-cycle structure, training stability, mode collapse