We describe a new training methodology for generative adversarial networks. The key idea is to grow both the generator and discriminator progressively: starting from a low resolution, we add new layers that model increasingly fine details as training progresses. This both speeds the training up and greatly stabilizes it, allowing us to produce images of unprecedented quality, e.g., CelebA images at 1024^2. We also propose a simple way to increase the variation in generated images, and achieve a record inception score of 8.80 in unsupervised CIFAR10. Additionally, we describe several implementation details that are important for discouraging unhealthy competition between the generator and discriminator. Finally, we suggest a new metric for evaluating GAN results, both in terms of image quality and variation. As an additional contribution, we construct a higher-quality version of the CelebA dataset.
核心贡献 · Key contributions
渐进式增长 GAN，从低分辨率到高分辨率逐步添加层以建模细节。 Progressive growing of GANs from low to high resolution, adding layers for fine details.
小批量标准差层，无需可学习参数即可增加生成图像的多样性。 Minibatch standard deviation layer to increase variation without learnable parameters.
通过运行时权重缩放实现均衡学习率，平衡各层训练速度。 Equalized learning rate via runtime weight scaling for balanced training.
生成器中逐像素特征向量归一化，防止信号幅度失控。 Pixelwise feature vector normalization in generator to prevent signal escalation.
切片 Wasserstein 距离（SWD）指标，用于评估 GAN 的图像质量和多样性。 Sliced Wasserstein distance (SWD) metric for evaluating GAN quality and variation.
高质量 CelebA-HQ 数据集，分辨率为 1024x1024，用于基准测试。 High-quality CelebA-HQ dataset at 1024x1024 resolution for benchmarking.
局限 · Limitations
渐进式增长可能无法泛化到非图像领域或结构化输出。 Progressive growing may not generalize to non-image domains or structured outputs.
逐像素归一化可能过度约束生成器，限制其表达能力。 Pixelwise normalization can overly constrain generator, limiting expressiveness.
SWD 指标依赖随机投影，可能遗漏细微的模式坍塌。 SWD metric relies on random projections and may miss subtle mode collapses.
该方法需要仔细调整淡入调度和层添加时机。 Method requires careful tuning of fading schedules and layer addition timing.
高分辨率训练仍需要大量计算资源。 High-resolution training still demands significant computational resources.
论文章节 · Sections（共 16）
摘要Abstract
1 引言1 Introduction
2 GAN 的渐进式增长2 Progressive growing of GANs
3 利用小批量标准差增加变化3 Increasing variation using minibatch standard deviation
4 生成器和判别器中的归一化4 Normalization in generator and discriminator
4.1 均衡学习率4.1 Equalized learning rate
4.2 生成器中的逐像素特征向量归一化4.2 Pixelwise feature vector normalization in generator
5 用于评估 GAN 结果的多尺度统计相似性5 Multi-scale statistical similarity for assessing GAN results
6 实验6 Experiments
6.1 各贡献在统计相似性方面的重要性6.1 Importance of individual contributions in terms of statistical similarity
6.2 收敛性与训练速度6.2 Convergence and training speed
6.3 使用 CelebA-HQ 数据集生成高分辨率图像6.3 High-resolution image generation using CelebA-HQ dataset