This paper introduces Glow, a new type of generative flow model that uses invertible 1×1 convolutions. Glow improves upon existing flow-based generative models by achieving significant improvements in log-likelihood on standard benchmarks. The model is capable of generating realistic-looking images and performing efficient manipulation of large images. Glow is based on the NICE and RealNVP flows, but introduces new elements such as actnorm, invertible 1×1 convolutions, and affine coupling layers. The model is trained with a multi-scale architecture and is efficient in both inference and synthesis. The model is evaluated on various datasets, including CIFAR-10, ImageNet, and LSUN, and shows significant improvements in log-likelihood and image quality. The model is also able to perform semantic manipulation of images by modifying attributes such as facial features. Glow is the first likelihood-based model that can efficiently synthesize high-resolution natural images.This paper introduces Glow, a new type of generative flow model that uses invertible 1×1 convolutions. Glow improves upon existing flow-based generative models by achieving significant improvements in log-likelihood on standard benchmarks. The model is capable of generating realistic-looking images and performing efficient manipulation of large images. Glow is based on the NICE and RealNVP flows, but introduces new elements such as actnorm, invertible 1×1 convolutions, and affine coupling layers. The model is trained with a multi-scale architecture and is efficient in both inference and synthesis. The model is evaluated on various datasets, including CIFAR-10, ImageNet, and LSUN, and shows significant improvements in log-likelihood and image quality. The model is also able to perform semantic manipulation of images by modifying attributes such as facial features. Glow is the first likelihood-based model that can efficiently synthesize high-resolution natural images.