We propose an alternative generator architecture for generative adversarial networks, borrowing from style transfer literature. The new architecture leads to an automatically learned, unsupervised separation of high-level attributes (e.g., pose and identity when trained on human faces) and stochastic variation in the generated images (e.g., freckles, hair), and it enables intuitive, scale-specific control of the synthesis. The new generator improves the state-of-the-art in terms of traditional distribution quality metrics, leads to demonstrably better interpolation properties, and also better disentangles the latent factors of variation. To quantify interpolation quality and disentanglement, we propose two new, automated methods that are applicable to any generator architecture. Finally, we introduce a new, highly varied and high-quality dataset of human faces.
核心贡献 · Key contributions
提出一种基于风格的生成器架构，实现高层属性与随机变化的无监督分离。 Proposes a style-based generator architecture for GANs, enabling unsupervised separation of high-level attributes and stochastic variation.
引入中间潜在空间，提升变化因素的解耦性和线性度。 Introduces an intermediate latent space that improves disentanglement and linearity of factors of variation.
在 CelebA-HQ 和 FFHQ 数据集上达到最优 FID 分数，提升近 20%。 Achieves state-of-the-art FID scores on CelebA-HQ and FFHQ datasets, improving by nearly 20%.
开发两种新自动指标——感知路径长度和线性可分性——量化插值质量和解耦性。 Develops two new automated metrics—perceptual path length and linear separability—for quantifying interpolation quality and disentanglement.
引入 FFHQ 数据集，包含 70,000 张 1024x1024 分辨率的高质量人脸图像。 Introduces the FFHQ dataset, a high-quality human face dataset with 70,000 images at 1024x1024 resolution.
通过风格混合和噪声注入展示图像合成的尺度特定控制。 Demonstrates scale-specific control of image synthesis through style mixing and noise injection.
局限 · Limitations
该方法未修改判别器或损失函数，限制了这些组件的潜在改进。 The method does not modify the discriminator or loss function, limiting potential improvements from those components.
所提指标依赖预训练分类器和感知距离，可能引入偏差。 The proposed metrics rely on pre-trained classifiers and perceptual distances, which may introduce bias.
截断技巧减少生成图像的多样性，限制了变化范围。 The truncation trick reduces variation in generated images, limiting diversity.
该方法仅在人脸和场景数据集上评估，未验证对其他领域的泛化性。 The approach is evaluated only on face and scene datasets; generalization to other domains is not verified.
映射网络需要仔细调整学习率以避免训练不稳定。 The mapping network requires careful tuning of learning rates to avoid training instability.
论文章节 · Sections（共 19）
摘要Abstract
1 引言1 Introduction
2 基于样式的生成器2 Style-based generator
2.1 生成图像的质量2.1 Quality of generated images
2.2 现有技术2.2 Prior art
3 基于样式的生成器的特性3 Properties of the style-based generator
3.1 样式混合3.1 Style mixing
3.2 随机变化3.2 Stochastic variation
3.3 全局效果与随机性的分离3.3 Separation of global effects from stochasticity
4 解耦研究4 Disentanglement studies
4.1 感知路径长度4.1 Perceptual path length
4.2 线性可分性4.2 Linear separability
5 结论5 Conclusion
6 致谢6 Acknowledgements
附录 A FFHQ 数据集Appendix A The FFHQ dataset
附录 B 𝒲中的截断技巧Appendix B Truncation trick in 𝒲𝒲\mathcal{W}
附录 C 超参数与训练细节Appendix C Hyperparameters and training details