Generative Models

• This repository is for implementation of Generative Models using Tensorflow 1.12

• Description of the Papers (Korean)

• The structure of the code is based on the Hwalsuk Lee's Generative Model github repository

Contributors

MMC Lab GAN Study Group members

• Kyushik Min, Gihoon Kim, Hyukju Sohn, Yoonyong Ahn, Seungwon-Choi, Jongyoon Baek

Implemented Paper List (20 Papers)

GAN

1. [GAN] Generative Adversarial Networks
2. [DCGAN] Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks
3. [LSGAN] Least Squares Generative Adversarial Networks
4. [WGAN] Wasserstein GAN
5. [WGAN_GP] Improved Training of Wasserstein GANs
6. [CGAN] Conditional Generative Adversarial Nets
7. [InfoGAN] Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets
8. [HoloGAN] Unsupervised Learning of 3D Representations From Natural Images
9. [SinGAN] Learning a Generative Model from a Single Natural Image
10. [PGGAN] Progressive Growing of GANs for Improved Quality, Stability, and Variation
11. [StyleGAN] A Style-Based Generator Architecture for Generative Adversarial Networks

Image-to-Image Translation

1. [CycleGAN] Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks
2. [AGGAN] Attention-Guided Generative Adversarial Networks for Unsupervised Image-to-Image Translation
3. [StarGAN] Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation
4. [DMIT] Multi-mapping Image-to-Image Translation via Learning Disentanglement

Interpretable GAN Latent

1. Unsupervised Discovery of Interpretable Directions in the GAN Latent Space

VAE

1. Auto-Encoding Variational Bayes (VAE)
2. Beta-VAE: Learning Basic Visual Concepts with a Constrained Variational Framework
3. Neural Discrete Representation Learning(VQ-VAE)

Application

1. Adherent Raindrop Removal with Self-Supervised Attention Maps andSpatio-Temporal Generative Adversarial Networks

Our Results

GAN Results

1. GAN

MNIST

2. DCGAN

MNIST	CelebA

3. LSGAN

MNIST	CelebA

4. WGAN

MNIST	CelebA

5. WGAN-GP

MNIST	CelebA

6. Conditional GAN

MNIST

7. InfoGAN

MNIST

8. HoloGAN

CelebA

9. SinGAN

Balloon

Mountain

Starry Night

10. PGGAN

It took about 2 weeks on TITAN RTX and trained 600k images per stage.

1024x1024 images

Cherry picked images

Latent interpolation

Fixed latent

No cherry picked images

11. StyleGAN

CelebA HQ (512x512 images)

Selected images

Style Mixing with Latent Codes

Random Images

AFHQ (512x512 images)

Selected images

Style Mixing with Latent Codes

Random Images

Image-to-Image Translation Results

1. CycleGAN

Monet to Photo	Photo to Monet

Horse to Zebra	Zebra to Horse

2. AGGAN

Horse to Zebra	Zebra to Horse

3. StarGAN

CelebA

4. DMIT

Summer2Winter

Interpretable GAN Latent

1. Unsupervised Discovery of Interpretable Directions in the GAN Latent Space

1) MNIST

VAE Results

1. VAE

Reconstruction

MNIST	CelebA

Latent Space Interpolation (MNIST)

Latent Space Interpolation (CelebA)

2. Beta-VAE

Latent Space Interpolation: Beta = 10 (CelebA)

Latent Space Interpolation: Beta = 200 (CelebA)

3. VQ-VAE

Reconstruction (MNIST)

Input	Reconstruction

Reconstruction (CelebA)

Input	Reconstruction

PixelCNN Trained Latent Decoding

MNIST	CelebA

Application Results

1. Raindrop Removal