Efficient subsampling of realistic images from GANs conditional on a class or a continuous variable

To improve image quality, subsampling or refining images generated from (unconditional) generative adversarial networks (GANs) has received much attention recently. These methods are less effective or inefficient, however, for conditional GANs (cGANs) that either condition on a class (i.e., class-conditional GANs) or a continuous variable (i.e., continuous cGANs or CcGANs). Thus, we introduce a novel subsampling scheme for cGANs: conditional density ratio rejection sampling (cDR-RS). Specifically, cDR-RS comprises an improved feature extraction mechanism and a conditional Softplus loss (cSP). We also derive an error bound for the density ratio model trained with the cSP loss. Fake images are accepted or rejected based on the improved estimated conditional density ratio. An additional filtering scheme further increases fake images’ label consistency without losing diversity when sampling from CcGANs. We extensively test the effectiveness and efficiency of cDR-RS in sampling from both class-conditional GANs and CcGANs on six benchmark datasets. Experimental results demonstrate that the proposed method can achieve state-of-the-art performances in subsampling both types of conditional GANs.