Although GAN has been widely applied in various aspects, training a GAN is not an easy task. During the training process, problems such as mode collapse, non-convergence of the loss, and blurriness of generated samples may occur. This chapter introduces the three most common problems in GAN training, namely gradient vanishing, non-convergence of the objective function, and mode collapse, and analyzes the causes of these problems. Regarding the problem of gradient vanishing, the annealing noise method is introduced. In response to the oscillation and instability of the objective function during GAN training, two methods, spectral regularization SNGAN and consistent optimization, are elaborated in detail. In addition, many GAN training techniques, such as feature matching and historical mean, are also explained. For the problem of mode collapse, from the two perspectives of the objective function and the GAN structure, some algorithms that can effectively alleviate mode collapse are introduced, and specific methods such as unrolledGAN, DRAGAN, MADGAN, VVEGAN, and the minibatch discriminator are provided.