The Generated Adversarial Network (GAN) is commonly used to learn to generate a wide variety of images. The Wasserstein GAN improves the stability of GAN, but there are also deficiencies that do not have controllable conditions. This paper proposes an improved GAN network model, which we call CWGAN. CWGAN achieves the goal of improving the training stability and controllability of GAN by adding condition information to WGAN generators and discriminators. The experiment results show that CWGAN improves the training stability, solves the problem of gradient disappearance, and produces images more clearly, and there is no obvious mode collapse problem.