The main purpose of this project is to create anime image only using line drawing.
Tensorflow is utilized for neural network of the work to transplant some environment and devices.
build up ios painting application
deploy the web services
I've been experienced some loss function to get most effective result.
The exprience list is this
only L1 Distance
L1 Distance + DCGAN ( this is same pix2pix project)
Wasserstein gan not added L1 Distance
use Selu , activation function, for DCGAN
The most better result in the above list is lead by L1 Distance + DCGAN. And some experience show the result that the coefficent with DCGAN is abount 1.1.
See below. https://qiita.com/latte_zero/items/2dd7bf55efd5a8d9a215
Some code of this project is on
https://github.com/yenchenlin/pix2pix-tensorflow
The paper, pix2pix
https://arxiv.org/pdf/1611.07004v1.pdf