Abstract

We introduce autoregressive implicit quantile networks (AIQN), afundamentally different approach to generative modeling than those commonlyused, that implicitly captures the distribution using quantile regression. AIQNis able to achieve superior perceptual quality and improvements in evaluationmetrics, without incurring a loss of sample diversity. The method can beapplied to many existing models and architectures. In this work we extend thePixelCNN model with AIQN and demonstrate results on CIFAR-10 and ImageNet usingInception score, FID, non-cherry-picked samples, and inpainting results. Weconsistently observe that AIQN yields a highly stable algorithm that improvesperceptual quality while maintaining a highly diverse distribution.

 

Quick Read (beta)