Motivation
- unconditional GANs 在生成图片方面限制于数据集
Contribution
- an unconditional generative model that can be learned from a single natural image
- capture the internal distribution of patches within the image
- This allows generating new samples of arbitrary size and aspect ratio, that have significant variability, yet maintain both the global structure and the fine textures of the training image
Method
金字塔结构
底层为常见GAN模型的纯生成过程，随后顺序通过生成图片上采样与噪音结合生成下一尺度图片。如下所示：
可看出噪音通过卷积层生成图像高频细节信息（残差）
损失函数
由底向上训练GAN，底层训练好后权重固定训练上层GAN，各层损失函数如下：
Adversarial loss
实验中是用WGAN-GP作为损对抗失函数
Reconstruction loss
个人不是很理解为何除了初始噪音，后续噪音均为0
Experiment
Generation from different scales
Number of scales
Application
应该是通过改变输入（原图下采样）来进行pix2pix图像生成，整体效果确实很惊艳。
SR
效果确实很强，比ZSSR效果还好…
Paint-to-Image
Harmonization
Editing
