An adversarial non-volume preserving flow model with Boltzmann priors

Abstract

Flow-based generative models (flow models) are conceptually attractive due to tractability of the exact log-likelihood and the exact latent-variable inference. In order to generate sharper images and extend the Gaussian prior of Flow models to other discrete forms, we propose an adversarial non-volume preserving flow model with Boltzmann priors (ANVP) for modeling complex high-dimensional densities. In order to generate sharper images, an ANVP model introduces an adversarial regularizer into the loss function to penalize the condition that it places a high probability in regions where the training data distribution has a low density. Moreover, we show that the Gaussian prior can be extended to other forms such as the Boltzmann prior in the proposed ANVP model, and we use multi-scale transformations and Boltzmann priors to model the data distribution. The experiments show that proposed model is effective in image generation task.