A tensorflow implementation of Christian et al's "Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network" paper. ( See : https://arxiv.org/abs/1609.04802 ) This implementation is quite different from original paper. The differences are as followings:

1. MNIST data set is used for convenience. ( It'll be straight-forward applying this scheme to large image data set like Urban 100 )
2. I've completely replace MSE loss with GAN using tuple input for discriminator.( see training source code )
3. I've used ESPCN ( sub-pixel CNN ) instead of deconvolution. ( see : https://www.cv-foundation.org/openaccess/content_cvpr_2016/papers/Shi_Real-Time_Single_Image_CVPR_2016_paper.pdf )

The existing CNN based super-resolution skill mainly use MSE loss and this makes super-resolved images look blurry. If we replace MSE loss with gradients from GAN, we may prevent the blurry artifacts of the super-resolved images and this is the key idea of this paper. I think this idea looks promising and my experiment result using MNIST data set looks good.

1. tensorflow >= rc0.11
2. sugartensor >= 0.0.1.7

Execute

python train.py

to train the network. You can see the result ckpt files and log files in the 'asset/train' directory. Launch tensorboard --logdir asset/train/log to monitor training process.

Execute

python generate.py

to generate sample image. The 'sample.png' file will be generated in the 'asset/train' directory.

This image was generated by SRGAN.

1. Original GAN tensorflow implementation
2. InfoGAN tensorflow implementation
3. Supervised InfoGAN tensorflow implementation
4. EBGAN tensorflow implementation
5. Time-series InfoGAN tensorflow implementation

Namju Kim (buriburisuri@gmail.com) at Jamonglabs Co., Ltd.