The paper proposes two ways of improving flow-based models of images such as Glow: 1. introducing invertible (locally) autoregressive layers implemented using masked convolutions and 2. making the multi-scale architecture more fine-grained by "factoring" out the variables in stages rather than in one go. While the reviewers found both of these contributions quite incremental, the experimental section is quite strong, with informative ablation studies and state-of-the-art results. The paper really needs substantial reworking however, as the primary contributions are not described with sufficient clarity and detail, and too much space is used for explaining data dequantization, which is not novel and could be covered much more concisely.