Zhu J Y, Park T, Isola P, et al. Unpaired image-to-image translation using cycle-consistent adversarial networks[C]//Proceedings of the IEEE international conference on computer vision. 2017: 2223-2232.

论文提出Cycle GAN结构，基于unpaired data (Figure 2)，学习domain X到domain Y的映射关系 (Figure 1)。

Cycle GAN能够应用到不同任务上：style transfer, object transfiguration, attribute transfer, and photo enhancement等。
Cycle GAN包含adversarial loss, cycle consistency loss.
对于某些特定任务，Cycle GAN额外包含一个identity loss.

论文假设两个不同domain之间存在underlying relationships，Cycle GAN (Figure 3)从一个image collections X中学习到一些特征，并将这些特征转换到另一个image collections Y上。

仅使用adversarial loss，无法保证生成的图像是有意义的，例如，G可能生成rubbish fool D。此外，生成图像不一定是desired。另一方面，标准的GAN过程可能会导致mode collapse问题：所有输入图像都会被映射到同一个输出图像。因此，模型引入了cycle consistent loss.
对于painting->photo的任务，为了保留输入- painting的颜色 (Figure 9)，模型引入了identiy loss.

Circle GAN可看作是两个auto-encoder：GF和FG。

Image-to-Image Translation (CGAN)
1. I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial nets. In NIPS, 2014
2. P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros. Image-to-image translation with conditional adversarial networks. arXiv preprint arXiv:1611.07004, 2016.
Unpaired Image-to-Image Translation (VAE+GAN)
1. M.Y. Liu, T. Breuel, and J. Kautz. Unsupervised image-to-image translation networks. arXiv preprint arXiv:1703.00848, 2017.
Neural Style Transfer
学习两张特定图片之间的映射，Cycle GAN学习的是两个domain之间的映射。

2. Loss

Adsersarial Loss

Loss

Identity Loss

3. Implementation

D使用70*70 PatchGAN：更少参数，能判别任意大小图像。
将adversarial loss从negative log likelihood改为least square loss.

History Buffer of generated images.
batch size 1 of scratch.

4. Experiments

CoGAN
M.-. Liu and O. Tuzel. Coupled generative adversarial networks. In NIPS, pages 469–477, 2016.
Pixel loss + GAN
SimGAN (self-regularision loss).
Feature loss + GAN
SimGAN (vgg16 feature loss, instead of RGB loss).
BiGAN
V. Dumoulin, I. Belghazi, B. Poole, A. Lamb, M. Arjovsky, O. Mastropietro, and A. Courville. Adversarially learned inference. arXiv preprint arXiv 2016.
Pix2pix
CGAN.
AMT