MSG-GAN: Multi-Scale Gradients for Generative Adversarial Networks

MSG-GAN: 生成对抗网络的多尺度梯度

code：GitHub - akanimax/BMSG-GAN: [MSG-GAN] Any body can GAN! Highly stable and robust architecture. Requires little to no hyperparameter tuning. Pytorch Implementation

paper：MSG-GAN: Multi-Scale Gradients for Generative Adversarial Networks | Papers With Code 

资源下载

---

解决了什么问题？

虽然生成对抗网络在图像合成任务中取得了巨大的成功，但众所周知，它们很难适应不同的数据集，部分原因是训练过程中的不稳定性和对超参数的敏感性。本文提出了多尺度梯度生成对抗网络(MSG-GAN)，

整体架构

提出的方法基于PGGAN，但是不是像PGGAN似的渐进式增长，而是通过连接G和D的每一个中间层，输入到鉴别器D的多尺度图像被转换成空间体（spatial volumes），该空间体与从卷积层的主路径获得的相应激活体连接。

 Figure 2: Architecture of MSG-GAN, shown here on the base model proposed in ProGANs [13].

实验

MSG-GAN比PGGAN的稳定性强

 Figure 6: Image stability during training. These plots show the MSE between images generated from the same latent code at the beginning of sequential epochs (averaged over 36 latent samples) on the CelebA-HQ dataset. MSG-ProGAN converges stably over time while ProGANs [13] continues to vary significantly across epochs