This paper presents improvements to the StyleGAN architecture to enhance image quality and address specific artifacts. The authors analyze and modify the generator's normalization, progressive growing, and training methods. They redesign the generator normalization to remove blob-like artifacts, revisit progressive growing to improve resolution utilization, and introduce a path length regularizer that enhances generator invertibility and image quality. The path length regularizer also improves the consistency and stability of generated images. The authors also find that the generator can be more effectively inverted with the new architecture, making it easier to attribute generated images to their source. Additionally, they identify a capacity problem in the model and propose training larger models for further quality improvements. The improved model significantly enhances the state of the art in unconditional image modeling, both in terms of distribution quality metrics and perceived image quality. The paper also discusses the benefits of using a path length regularizer, which helps in achieving smoother generator mappings and better image quality. The authors also explore alternative network architectures and find that skip connections and residual networks improve performance. The paper concludes with a discussion on the practical implications of the improvements, including faster training times and better image quality. The implementation and trained models are available at https://github.com/NVlabs/stylegan2.This paper presents improvements to the StyleGAN architecture to enhance image quality and address specific artifacts. The authors analyze and modify the generator's normalization, progressive growing, and training methods. They redesign the generator normalization to remove blob-like artifacts, revisit progressive growing to improve resolution utilization, and introduce a path length regularizer that enhances generator invertibility and image quality. The path length regularizer also improves the consistency and stability of generated images. The authors also find that the generator can be more effectively inverted with the new architecture, making it easier to attribute generated images to their source. Additionally, they identify a capacity problem in the model and propose training larger models for further quality improvements. The improved model significantly enhances the state of the art in unconditional image modeling, both in terms of distribution quality metrics and perceived image quality. The paper also discusses the benefits of using a path length regularizer, which helps in achieving smoother generator mappings and better image quality. The authors also explore alternative network architectures and find that skip connections and residual networks improve performance. The paper concludes with a discussion on the practical implications of the improvements, including faster training times and better image quality. The implementation and trained models are available at https://github.com/NVlabs/stylegan2.