Awesome
Frameworks
- GAN, DCGAN
- cGAN - Label-conditioning generation
- AcGAN - Supervised generation (auxiliary classifier with labels)
- SGAN - D outputs [CLASS-1, CLASS-2, . . . CLASS-N, FAKE]
- InfoGAN - Unsupervised lantern space disentangling
- ALI, BiGAN - match p(G(z),z) and q(x,E(x))
- GMAN - Multiple discriminator models <!-- TODO -->
- AdaGAN - Multiple generative models <!-- TODO -->
- MGGAN - Multiple generative models <!-- TODO -->
High Resolution
- LAPGAN - Coarse-to-fine generation
- StackGAN - Two-step generation
- PGGAN
Function ('~' means 'match')
g(Z) ~ X
- GAN, DCGAN
g(Z|Y) ~ X|Y, where Y is label or lantern variable
- cGAN
- AcGAN
- InfoGAN
- CFGAN
g(Z),Z ~ X,E(X)
- ALI, BiGAN
Objective Function
f-Divergence
- GAN, DCGAN - JS divergence
- LSGAN (Least Squares GAN) - Pearson χ2 divergence
- f-GAN - Variational divergence minimization
- f-GANs
Integral Probability Metrics (IPM)
- WGAN - Wasserstein distance
- WGAN-GP - Gradient penalty, less capacity compromise
- McGAN - Mean and covariance matching
- Geometric GAN
- Fisher GAN - Chi-squared distance
Maximum Mean Discrepancy (MMD)
- GMMN
- MMD nets
Others
- CatGAN - Entropy of P(Y|X)
- EBGAN - D(x) = ||Dec(Enc(x)) − x||
- LS-GAN (Loss-Sensitive GAN)
- BEGAN - Wasserstein distance between the distribution of real/fake auto-encoder loss
- AGE - $max_emin_g\Delta(e(g(Z))||Y)-\Delta(e(X)||Y)$ and encoder-generator reciprocity (bidirectional mapping)
- Softmax GAN
- Cramér GAN - Cramér Distance
- LDGAN
Regularized GAN
- DRGAN - Vanilla GAN with gradient penalty
- Cramér GAN - Cramér Distance
- Regularized GAN
- CT-GAN
- Varying k-Lipschitz Constraint for Generative Adversarial Networks
Representation Learning
Lantern Space Disentangling
- InfoGAN - Unsupervised lantern space disentangling
- AcGAN - Supervised lantern space disentangling
Specifying Lantern Space Distribution
- AAE
- ALI, BiGAN - Match p(G(z),z) and q(x,E(x)), simultaneously learn an encoder and decoder
- AGE
Semi-supervised Learning
- CatGAN
- SGAN - D outputs [CLASS-1, CLASS-2, . . . CLASS-N, FAKE]
- Improved GAN
- Triple-GAN
- CT-GAN
Evaluations of GANs <!-- TODO -->
- Inception Score - Improved GAN
- FCN Score - pix2pix
- AMT Perceptual Studies - pix2pix
- Semantic Segmentation Metrics - CycleGAN
- FID, Precision, Recall and F1 Score - Are GANs Created Equal? A Large-Scale Study
Applications
Data Augmentation
- Adversarial Generation of Training Examples for Vehicle License Plate Recognition
Domain Adaptation
- Unsupervised Pixel-Level Domain Adaptation with Generative Adversarial Networks
- DTN
- UNIT
- CoGAN
Image Denoising
- Deep Semantic Face Deblurring
Image-to-Image Translation
Unpaired
- CoGAN
- DTN
- CycleGAN, DiscoGAN, DualGAN
- UNIT
- Face Transfer with Generative Adversarial Network
- XGAN - Semantic consistency
Paired
- pix2pix/PatchGAN
- Scribbler
- PAN - Perceptual adversarial loss
- Cross-View Image Synthesis using Conditional GANs
Inpainting
- Context Encoder
- PAN
Super-Resolution
- SRGAN
Text-to-Image Synthesis
- GAN-INT-CLS <!-- TODO -->
- GAWWN <!-- TODO -->
- StackGAN
Face Editing
- VAE/GAN - Visual attribute vectors
- IcGAN
- DIAT
- Learning Residual Images for Face Attribute Manipulation
- DistanceGAN
- CFGAN
- Age-cGAN
- UNIT
- SL-GAN
- IAN
- Neural Face Editing with Intrinsic Image Disentangling
- GeneGAN - Object transfiguration
- Fader Networks
- Unsupervised Visual Attribute Transfer with Reconfigurable Generative Adversarial Networks
- ExprGAN
- StarGAN
- GLCA-GAN
Face Frontalization/Profiling
- DR-GAN
- TP-GAN
Others
- iGAN - Image manipulation
- TVSN - 3D view synthesis <!-- TODO -->
- ID-CGAN - Image de-raining
- Perceptual GAN - Small object detection
Survey
- Generative Adversarial Networks: An Overview
- How Generative Adversarial Nets and its variants Work
- Comparative Study on Generative Adversarial Networks
- An Introduction to Image Synthesis with Generative Adversarial Nets
Unclassified
- DeePSiM
- Unrooled GAN
- SGAN
- AM-GAN
- DeLiGAN - Mixture Gaussian prior distribution
- TTUR
Datasets
- MNIST
- CIFAR10/CIFAR100
- SVHN
- CelebA