Large Scale GAN Training for High Fidelity Natural Image Synthesis
Andrew Brock Heriot-Watt University、Jeff Donahue DeepMind、Karen Simonyan DeepMind Work done at DeepMindEqual contribution
Abstract
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 128 $\times$ 128 resolution, our models (BigGANs) achieve an Inception Score (IS) of 166.5 and Fréchet Inception Distance (FID) of 7.4, improving over the previous best IS of 52.52 and FID of 18.65.
中文速览
生成对抗网络（GAN）在ImageNet这类大规模复杂数据集上生成高分辨率、多样化图像一直是个难题。研究者通过大幅扩大模型参数量（增至原来2–4倍）、增大批量大小（增至8倍），并引入共享类别嵌入、噪声跳跃连接等架构改进，同时加入正交正则化（Orthogonal Regularization）来改善生成器的平滑性，训练出了名为BigGAN的大规模条件图像生成模型。正交正则化的一个关键副产品是让模型天然支持"截断技巧"（truncation trick）——在采样时对噪声向量进行截断，从而在图像质量与多样性之间实现灵活可控的权衡。最终，BigGAN在128×128分辨率的ImageNet上将Inception Score从52.5大幅提升至166.5，Fréchet Inception Distance从18.65降至7.4，在256×256和512×512分辨率上同样刷新了记录，证明了合理的规模化策略能够显著缩小生成图像与真实图像之间的差距，为后续大规模图像生成研究提供了重要基准和实践经验。
原文 arXiv:1809.11096;中英对照 + 大白话阅读 https://aha.fim.ai/paper/1809.11096v2