Deep generative models are powerful tools for the exploration of chemical space, enabling the on-demand gener- ation of molecules with desired physical, chemical, or biological properties. However, these models are typically thought to require training datasets comprising hundreds of thousands, or even millions, of molecules. This per- ception limits the application of deep generative models in regions of chemical space populated by only a small number of examples. Here, we systematically evaluate and optimize generative models of molecules for low-data settings. We carry out a series of systematic benchmarks, training more than 5,000 deep generative models and evaluating over 2.6 billion generated molecules. We find that robust models can be learned from far fewer examples than has been widely assumed. We further identify strategies that dramatically reduce the number of molecules required to learn a model of equivalent quality, and demonstrate the application of these principles by learning models of chemical structures found in bacterial, plant, and fungal metabolomes. The structure of our experiments also allows us to benchmark the metrics used to evaluate generative models themselves. We find that many of the most widely used metrics in the field fail to capture model quality, but identify a subset of well-behaved metrics that provide a sound basis for model development. Collectively, our work provides a foundation for directly learning generative models in sparsely populated regions of chemical space.