The style-based GAN architecture (StyleGAN) yields state-of-the-art results in data-driven unconditional generative image modeling. We expose and analyze several of its characteristic artifacts, and propose changes in both model architecture and training methods to address them. In particular, we redesign the generator normalization, revisit progressive growing, and regularize the generator to encourage good conditioning in the mapping from latent codes to images. In addition to improving image quality, this path length regularizer yields the additional benefit that the generator becomes significantly easier to invert. This makes it possible to reliably attribute a generated image to a particular network. We furthermore visualize how well the generator utilizes its output resolution, and identify a capacity problem, motivating us to train larger models for additional quality improvements. Overall, our improved model redefines the state of the art in unconditional image modeling, both in terms of existing distribution quality metrics as well as perceived image quality.
核心贡献 · Key contributions
重新设计生成器归一化，通过权重解调去除水滴伪影。 Redesigns generator normalization to remove droplet artifacts via weight demodulation.
引入路径长度正则化器，促进潜在空间到图像的平滑映射，提升质量和可逆性。 Introduces path length regularizer to encourage smooth latent-to-image mapping, improving quality and invertibility.
用跳跃生成器和残差判别器替代渐进式增长，消除相关伪影。 Replaces progressive growing with skip generator and residual discriminator, eliminating related artifacts.
识别高分辨率层的容量问题，增加特征图以改善细节。 Identifies capacity issue in high-resolution layers and increases feature maps to improve detail.
提出惰性正则化，在不影响效果的前提下降低正则化器的计算成本。 Proposes lazy regularization to reduce computational cost of regularizers without harming effectiveness.
通过 StyleGAN2 的潜在空间投影，展示改进的源归属能力。 Demonstrates improved source attribution via latent space projection with StyleGAN2.
局限 · Limitations
路径长度正则化器在结构较少的数据集上引入 FID 和 PPL 之间的权衡。 Path length regularizer introduces tradeoff between FID and PPL on less structured datasets.
解调依赖于信号的统计假设，而非实际特征图内容。 Demodulation relies on statistical assumptions about signal, not actual feature map contents.
FID 和 P&R 指标偏向纹理而非形状，可能遗漏质量方面。 FID and P&R metrics are biased towards texture over shape, potentially missing quality aspects.
增加容量导致更高的计算成本和训练时间。 Increased capacity leads to higher computational cost and training time.
方法仍需大数据集训练，限制了在数据稀缺场景下的适用性。 Method still requires large datasets for training, limiting applicability in data-scarce scenarios.
论文章节 · Sections（共 14）
摘要Abstract
1 引言1 Introduction
2 去除归一化伪影2 Removing normalization artifacts
2.1 生成器架构再探讨2.1 Generator architecture revisited
2.2 实例归一化再探讨2.2 Instance normalization revisited
3 图像质量与生成器平滑性3 Image quality and generator smoothness
3.1 延迟正则化3.1 Lazy regularization
3.2 路径长度正则化3.2 Path length regularization
4 渐进式增长再探讨4 Progressive growing revisited
4.1 替代网络架构4.1 Alternative network architectures
4.2 分辨率使用4.2 Resolution usage
5 图像到潜在空间的投影5 Projection of images to latent space