GAN Photo Editing
Beilei Zhang
Projection
Vanilla Gan with different perceptual loss lambda
The output of the Vanilla Gan is not ideal, It's hard to tell which result is better. The model is fragile and not always converge after 500 iteration.
Style Gan | W space | different perceptual loss lambda
StyleGan has a much better performance than Vanilla GAN. It seems perceptual loss help capture the pose of the target cat. 50% perceptual loss and 50% pixel loss is ideal for the final result
Style Gan | W+ space | different perceptual loss lambda
With mapping z to w+ space, even a small lambda value give decent result. It seems 20% perceptual loss + 80% pixel loss give the best result
The best and fastest result
The styleGAN shows better result especially with w+ embedded space. However, the vanilla GAN is faster to trained because of its structure simplicity
Interpolation
Interpolating between two cats
Interpolating between 0.png, 1.png
Interpolating between two cats
Interpolating between 2.png, 3.png
Scribble to Image
Generate with color constrains
If there are too much color in the user input, the cat tend to be blur and pale