Improved Precision and Recall Metric for Assessing Generative Models

We present an evaluation metric that can separately and reliably measure both of these aspects in image generation tasks by forming explicit, non-parametric representations of the manifolds of real and generated data.