This work introduces Conditional Image Retrieval (CIR) systems: IR methods that can efficiently specialize to specific subsets of images on the fly. These systems broaden the class of queries IR systems support, and eliminate the need for expensive re-fitting to specific subsets of data. Specifically, we adapt tree-based K-Nearest Neighbor (KNN) data-structures to the conditional setting by introducing additional inverted-index data-structures. This speeds conditional queries and does not slow queries without conditioning. We present two new datasets for evaluating the performance of CIR systems and evaluate a variety of design choices. As a motivating application, we present an algorithm that can explore shared semantic content between works of art of vastly different media and cultural origin. Finally, we demonstrate that CIR data-structures can identify Generative Adversarial Network (GAN) ``blind spots'': areas where GANs fail to properly model the true data distribution.