We introduce a simple (one line of code) modification to the Generative\nAdversarial Network (GAN) training algorithm that materially improves results\nwith no increase in computational cost: When updating the generator parameters,\nwe simply zero out the gradient contributions from the elements of the batch\nthat the critic scores as `least realistic'. Through experiments on many\ndifferent GAN variants, we show that this `top-k update' procedure is a\ngenerally applicable improvement. In order to understand the nature of the\nimprovement, we conduct extensive analysis on a simple mixture-of-Gaussians\ndataset and discover several interesting phenomena. Among these is that, when\ngradient updates are computed using the worst-scoring batch elements, samples\ncan actually be pushed further away from their nearest mode. We also apply our\nmethod to recent GAN variants and improve state-of-the-art FID for conditional\ngeneration from 9.21 to 8.57 on CIFAR-10.\n