Interpoint Inception Distance: Gaussian-Free Evaluation of Deep Generative Models

This paper introduces the Interpoint Inception Distance (IID) as a new approach for evaluating deep generative models. It is based on reducing the measurement of discrepancy between multidimensional feature distributions to one-dimensional interpoint comparisons. Our method provides a general tool for deriving a wide range of evaluation measures. The Cramér Interpoint Inception Distance (CIID) is notable for its theoretical properties, including a Gaussian-free structure of feature distribution and a strongly consistent estimator with unbiased gradients. Our experiments, conducted on both synthetic and large-scale real or generated data, suggest that CIID is a promising competitor to the Fréchet Inception Distance (FID), which is currently the primary metric for evaluating deep generative models.