Despite recent progress in generative image modeling, successfully generating high-resolution, diverse samples from complex datasets such as ImageNet remains an elusive goal. To this end, we train Generative Adversarial Networks at the largest scale yet attempted, and study the instabilities specific to such scale. We find that applying orthogonal regularization to the generator renders it amenable to a simple "truncation trick," allowing fine control over the trade-off between sample fidelity and variety by reducing the variance of the Generator's input. Our modifications lead to models which set the new state of the art in class-conditional image synthesis. When trained on ImageNet at 128x128 resolution, our models (BigGANs) achieve an Inception Score (IS) of 166.5 and Frechet Inception Distance (FID) of 7.4, improving over the previous best IS of 52.52 and FID of 18.6.
核心贡献 · Key contributions
证明 GAN 从规模扩张中显著受益，训练了参数 2-4 倍、批量大小 8 倍的模型。 Demonstrated that GANs benefit dramatically from scaling, training models with 2-4x parameters and 8x batch size.
引入两种简单架构变化（共享嵌入、跳跃连接 z），提升可扩展性和性能。 Introduced two simple architectural changes (shared embedding, skip-z) that improve scalability and performance.
对生成器应用正交正则化，实现截断技巧以精细控制保真度与多样性权衡。 Applied orthogonal regularization to the generator, enabling the truncation trick for fine control over fidelity-variety trade-off.
发现并刻画了大规模 GAN 特有的不稳定性，将其与生成器层的谱爆炸关联。 Discovered and characterized instabilities specific to large-scale GANs, linking them to spectral explosion in generator layers.
在 ImageNet 128x128 上实现 166.5 的 Inception Score 和 7.4 的 FID，大幅超越此前最优。 Achieved state-of-the-art Inception Score of 166.5 and FID of 7.4 on ImageNet 128x128, improving over previous best by large margins.
将发现扩展到 JFT-300M 数据集，表明设计选择可迁移，更大数据集可能缓解稳定性问题。 Extended findings to JFT-300M dataset, showing design choices transfer and larger datasets may alleviate stability issues.
局限 · Limitations
大规模训练仍会出现崩溃，尽管有正则化仍需早停。 Training collapse still occurs at large scale, requiring early stopping despite regularization.
通过对判别器施加强约束可确保稳定，但性能代价巨大（如 IS 降低 45%）。 Stability can be enforced by strong constraints on discriminator, but at a dramatic cost to performance (e.g., 45% IS reduction).
截断技巧降低样本多样性；最优阈值因数据集而异（如 JFT-300M vs ImageNet）。 Truncation trick reduces sample variety; optimal threshold varies across datasets (e.g., JFT-300M vs ImageNet).
判别器记忆训练集（训练准确率 98% vs 验证集 50-55%），限制泛化。 Discriminator memorizes training set (98% training accuracy vs 50-55% validation), limiting generalization.
模型在数据少或结构复杂的类别上表现不佳（如 ImageNet 上人群 vs 狗）。 Model struggles with classes having less data or more structure (e.g., crowds vs dogs on ImageNet).
论文章节 · Sections（共 14）
摘要Abstract
1 引言1 Introduction
2 背景2 Background
3 扩展 GANs3 Scaling Up GANs
3.1 用截断技巧权衡多样性与保真度3.1 Trading off variety and fidelity with the Truncation Trick
3.2 小结3.2 Summary
4 分析4 Analysis
4.1 表征不稳定性：生成器4.1 Characterizing Instability: The Generator
4.2 表征不稳定性：判别器4.2 Characterizing Instability: The Discriminator
4.3 小结4.3 Summary
5 实验5 Experiments
5.1 在 ImageNet 上的评估5.1 Evaluation on ImageNet
5.2 在 JFT-300M 上的额外评估5.2 Additional evaluation on JFT-300M