Super-Resolution Generative Adversarial Networks

Loss Function

For Gen loss

[optional] VGG19 BottleNeck feature loss (content loss)
MSE loss (content loss)
Adversarial GAN loss with sigmoid

For Disc loss

Adversarial GAN loss with sigmoid

Architecture Networks

Same as SRGAN paper

DIFFS	SRGAN Paper	ME
Weight initializer	`normal dist`	`HE initializer`
image scaling	`LR[0,1] HR[-1,1]`	`L/HR[-1,1]`
G loss	`content loss with vgg19`	`just MSE`
global steps	`2e5`	`1e5`

Tensorboard

Elapsed time : 1d 9h 30m 52s with GTX Titan X 12GB x 1 (maxwell)

Result

VALID HR image	VALID LR image

Global Step 10k	Global Step 25k	Global Step 55k

To-Do

Not good performance...
on-editing...