This paper proposes a rather simple but useful neural network module called "Slot Attention", which interfaces with perceptual representations such as the output of a convolutional neural network and produces task-dependent abstract representations (slots). The corresponding encoding method enjoys low computational cost and good empirical results on unsupervised object discovery. The authors' response also nicely addressed the main concerns in the original reviews. All reviewers agree that this is an interesting and important contribution to the field of object-centric representation learning.