Generative adversarial networks (GANs) are quickly becoming a ubiquitous\napproach to procedurally generating video game levels. While GAN generated\nlevels are stylistically similar to human-authored examples, human designers\noften want to explore the generative design space of GANs to extract\ninteresting levels. However, human designers find latent vectors opaque and\nwould rather explore along dimensions the designer specifies, such as number of\nenemies or obstacles. We propose using state-of-the-art quality diversity\nalgorithms designed to optimize continuous spaces, i.e. MAP-Elites with a\ndirectional variation operator and Covariance Matrix Adaptation MAP-Elites, to\nefficiently explore the latent space of a GAN to extract levels that vary\nacross a set of specified gameplay measures. In the benchmark domain of Super\nMario Bros, we demonstrate how designers may specify gameplay measures to our\nsystem and extract high-quality (playable) levels with a diverse range of level\nmechanics, while still maintaining stylistic similarity to human authored\nexamples. An online user study shows how the different mechanics of the\nautomatically generated levels affect subjective ratings of their perceived\ndifficulty and appearance.\n