Generative adversarial networks (GANs) have become popular generative models. However, the problem of generating instability of network training is hard to eliminate. A new method (SICS-GANs) is proposed by integrating the advantages of previous models. In the improved conditional deep convolutional generative adversarial networks model, spectrally normalized global weights are exploited. Using group normalization to process the input of the hidden layer can speed up training and improve the quality of generated samples. The model employs the global average pooling layer to replace the full connection layer to avoid overfitting. In addition, the improved method is applied to image recognition on two datasets, CIFAR-10 and SVHN. Compared with other stable methods, the experiments consequence show that this approach not only generates better images, but also effectively increases the accuracy of image recognition.