Multivariate distribution fitting for GANs via introducing variable correlations

Mode collapse is a major unsolved problem in generative adversarial networks. In this study, we base our proposal on the distribution fitting method and explore methods to suppress mode collapse for multivariate data. We incorporate the covariance constraints that enforce similar linear correlations among the variables. This approach may mitigate the nonuniform sampling issue more effectively for multivariate data, thereby suppressing mode collapse. For images, we also offer a scheme for incorporating covariances by utilizing the difference matrices. The method could handle images better since it considers the distances between pixels and possesses a better tolerance for errors like offsets. The proposed methods inherit the benefits of the distribution fitting method, which circumvents reliance on regularization or network modules, enhancing compatibility and facilitating its practical application. Experiments demonstrate the effectiveness and competitive performance of the proposed method.