The paper introduces a novel generator architecture for Generative Adversarial Networks (GANs) inspired by style transfer literature. This architecture separates high-level attributes (e.g., pose and identity) from stochastic variations (e.g., freckles, hair) in the generated images, enabling intuitive and scale-specific control over the synthesis process. The generator uses a learned constant input and adjusts the "style" of the image at each convolution layer based on the latent code, allowing for automatic and unsupervised separation of attributes and variations. The paper also proposes two new automated methods, perceptual path length and linear separability, to quantify the quality of interpolation and the disentanglement of latent factors. Additionally, it introduces a new dataset of human faces, Flickr-Faces-HQ (FFHQ), which offers higher quality and wider variation compared to existing datasets. The generator's architecture and methods improve state-of-the-art performance in traditional distribution quality metrics and disentanglement, while maintaining or improving image quality.The paper introduces a novel generator architecture for Generative Adversarial Networks (GANs) inspired by style transfer literature. This architecture separates high-level attributes (e.g., pose and identity) from stochastic variations (e.g., freckles, hair) in the generated images, enabling intuitive and scale-specific control over the synthesis process. The generator uses a learned constant input and adjusts the "style" of the image at each convolution layer based on the latent code, allowing for automatic and unsupervised separation of attributes and variations. The paper also proposes two new automated methods, perceptual path length and linear separability, to quantify the quality of interpolation and the disentanglement of latent factors. Additionally, it introduces a new dataset of human faces, Flickr-Faces-HQ (FFHQ), which offers higher quality and wider variation compared to existing datasets. The generator's architecture and methods improve state-of-the-art performance in traditional distribution quality metrics and disentanglement, while maintaining or improving image quality.