Existing decentralized learning of Generative Adversarial Network (GAN) suffers from a slower convergence rate and training instability due to the changes in the gradient-sharing approach among workers. It requires more iterations to achieve convergence and deteriorates the GAN training, consequently leading to slow down training speed and accuracy degradation. We propose JediGAN, the novel distributed GAN system that achieves optimal benefits by balancing accuracy and communication overheads through adaptive scheduling strategies. JediGAN reduces overall consensus variance by incorporating adaptive global updates for the discriminators. Additionally, JediGAN effectively leverages the internal characteristics of the GAN to decouple the training processes of the discriminators and generators during the search. It employs the concept of evolutionary selection to update the generators, thereby improving training stability. Through extensive experiments, we show that JediGAN outperforms state-of-the-art decentralized GAN systems, with 13.1% accuracy and 36.6% training speed improvement.