Active manipulation of droplets is crucial in droplet microfluidics. However, droplet polydispersity decreases the accuracy of active manipulation. We develop a microfluidic "droplet filter" that accurately separates droplets by size. The droplet filter has a sharp size cutoff and is capable of distinguishing droplets differing in volume by 20%. A simple model explains the behavior of the droplets as they pass through the filter. We show application of the filter in improving dielectric sorting efficiency. Published by AIP Publishing. [http://dx