Resources： GAN

Introduction

GAN大领域内的资料整理，会逐渐地更新划分成多个小领域，可能有多个papers包含在不同的领域中。

Unconditional GAN 是最原始的GAN的任务，对目标的图片的分布建模，学习映射 ${z \to x}$ 。

Generative Adversarial Nets. ( GAN )
Deep Generative Image Models using a Laplacian Pyramid of Adversarial Networks ( LPGAN )
Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks. ( DCGAN )
Progressive Growing of GANs for Improved Quality, Stability and Variation. ( PG-GAN )
A Style-Based Generator Architecture for Generative Adversarial Networks. ( Style-GAN )
Analyzing and Improving the Image Quality of StyleGAN. ( Style-GAN2 )

Conditonal GANs 是加入可控隐变量（监督信息，很多情况下是分类标签）的GAN的结构，Conditional GANs 的应用比 Unconditional GANs 的应用范围更广，而且 Conditional GANs 可以用于高质量的多类图像分布建模。

Conditional Generative Adversarial Nets. (cGAN)
InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets. (InfoGAN) [paper]
Conditional Image Synthesis with Auxiliary Classifier GANs. (ACGAN)
Stacked Generative Adversarial Networks. ( StackedGAN )
Spectral Normalization for Generative Adversarial Networks. ( SN-GAN )
cGAN with Projection Discriminator.
Self-Attention Generative Adversarial Networks. ( SA-GAN )
Large Scale GAN Traning for high fidelity natural image synthesis. ( BigGAN )
Large Scale Adversarial Representation Learning. ( BigBiGAN )
LOGAN: Latent Optimisatoin for Generative Adversarial Networks.

改进GAN的技术，考虑GAN最主要的两个问题：mode collapse 和 unstable training。可以考虑的解决方法有：改进loss，改进模型结构等。

Wasserstain Generative Adversarial Networks. ( WGAN )
Improved Traning of Wasserstain GANs ( WGAN-GP )
Least Squares Generative Adversarial Networks ( LSGAN )
f-GAN: Training Generative Neural Samplers using Variational Divergence Minimization.

Pacgan: The power of two samples in generative adversarial networks. 2017.
Minibatch Discriminator: Improved techniques for training GAN. 2016.
Minibatch Stddev Layer: Progressive Growing of GANs for Improved Quality, Stability and Variation. 2017.

Disentangled Representation Learning 和 VAE，Conditional GAN 的联系十分密切。针对的是生成模型的生成效果的语义可控性。

InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets.
Adversarial Feature Learning ( BiGAN )
On the “steerability” of generative adversarial networks. [paper] [codes]
Interpreting the Latent Space of GANs for Semantic Face Editing. [paper] [codes]

GAN可以用于Semi-supervised Learning，对仅含有部分、或者少量标签的数据集。

Image2Image 是GAN中非常重要的一类任务，而且它包含的子任务十分广泛，而且Text2Image、Video2Video等其实都可以归为这一类，它们对GAN的应用具有非常相似的地方。

Image-to-Image Translation with Conditional Generative Adversarial Networks. ( pix2pix )
Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks. ( CycleGAN )
StarGAN: Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation. ( StarGAN )
U-GAT-IT: Unsupervised Generative Attentional Networks with Adaptive Layer Instance Normalization for Image-to-Image Translation.

着重对GAN的某一任务(应用)的研究，这一部分和最后一部分有重合。

Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network. ( SRGAN for super-resolution )
GANimation: Anatomically-aware Facial Animation from a Single Image.
Perceptual Losses for Real-Time Style Transfer and Super-Resolution.
Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization.
Precomputed Real-Time Texture Synthesis with Markovian Generative Adversarial Networks. ( Markovian GAN)
SketchyGAN: Towards Diverse and Realistic Sketch to Image Synthesis.

GAN 和 few-shot learning 的结合，主要关注 few-shot generation。

InGAN: Capturing and Retargeting the “DNA” of a Natural Image.
SinGAN: Learning a Generative Model from a Single Natural Image.
Few-Shot Adversarial Learning of Realistic Neural Talking Head Models.
MetaGAN: An Adversarial Approach to Few-Shot Learning. (实际上这篇主要不是讲GAN，而是将GAN用于few-shot learning的classification)

GAN生成图片的质量的衡量是一个still open problem，但是仅仅针对 GAN Evalutaion 的papers在我看过的papers中不多，大多都是提一嘴的程度。

具体的方法如下：

IS: Improved techniques for training GAN.
Slice-Wasserstain-Distance: Progressive Growing of GANs for Improved Quality, Stability and Variation.
Wasserstain-Distance: Wasserstain Generative Adversarial Networks.
FID: GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium.
MS-SSIM: Conditional Image Synthesis with Auxiliary Classifier GANs. (ACGAN)
PPL: StyleGAN
LPIPS: The unreasonable effectiveness of deep features as a perceptual metric. CVPR2018.

将他们可以分类为：

在GAN中会用到的特殊结构、结构。

Conditional Batch Norm: Modulating early visual processing by language. NIPS2017. CGAN用于注入噪声的方法。

StackGAN: Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks. ( StackGAN )
Boundary Equilibrium GANs. ( BEGAN )
Energy-based Generative Adversarial Networks. ( EBGAN )
Real or not real, that is the question. ( RealnessGAN )
A Review on Generative Adversarial Networks: Algorithms, Theory, and Applications. ( GAN Review in 2020 )
Invertible Conditional GANs for image editing. (IcGAN) [paper] [codes]
加入Encoder结构编码图像的属性，这些属性又以cGAN的方式注入到GAN中生成，可以完成属性修改操作。
Learning to Discover Cross-Domain Relations with Generative Adversarial Networks. (DiscoGAN) [codes] [paper] CycleGAN 同期工作，域到域的Image2Image任务。
DualGAN.
Adversarial Latent Autoencoders (ALAE) [codes] [paper]

pass

特殊的应用，或者是一类的应用任务。

Image2Image：可以用于各种图像处理任务。
画风的图像风格转换GAN. [project] [paper]
style2paints: 线稿上色. [project] Two-stage Sketch Colorization. [paper]
clourise:黑白照片转彩色照片 [homepage]
Virtual Try-On: 虚拟试衣。
- 一篇知乎上的Review. [page]
- Towards Multi-Pose Guided Virtual Try-on Network. ICCV2019. [paper]
- VITON: An Image-Based Virtual Try-On Network. CVPR2018. [paper] 但是好像作者用的是Encoder-Decoder结构哈哈。
- VTNFP: An Image-Based Virtual Try-On Network With Body and Clothing Feature Preservation. [paper]
aiportraits: [homepage] 人脸转绘画。
AnimeGAN: [codes] 轻量级的图像转绘画风格的 GAN。
3D photo Inpaiting: [post] [paper]