[PYTHON] Line art coloring web service Paints Chainer has been released

http://paintschainer.preferred.tech Here is a demo of line art coloring that I wrote about last month! !! You can try it for everyone who had a great response! !! (I hope the expected value hasn't risen too much, but w)

http://qiita.com/taizan/items/cf77fd37ec3a0bef5d9d Previous article "Beginners tried coloring line art with chainer. I was able to do it. ] Is here. ⇓ And here is an example of using Mojakki-san.

Due to the use of GPU, if access is concentrated, it may slow down or the server may go down, but I hope you will keep an eye on it warmly. ← It has become very heavy because it has spread to many people. Gunununu

The screen looks like this. If you select a line art file, it will be painted automatically for the time being.

• However, gif and transparent png are not currently supported. Please use jpg or ordinary png

After selecting the file for a while, the line art and coloring result 1st row and coloring result 2nd row will be displayed like ↓. The first stage is about 128px and the second stage is about 512px.

Next, select the color you want to color from the color palette and add hints. In many cases, it is better to add a light color than a highly saturated color.

If you give a hint, press the colorize button and it will be colored with reference to the hint. In some cases, a rough hint will work, but in other cases, it will not feel good without giving detailed hints. Once you get the hang of it, you will be able to understand the feelings of neural networks.

In some cases, manually adjusting various parameters for each image will work even better, but this time I will omit it.

Last time it was only a model, but I uploaded the source code including training and UI to github. https://github.com/taizan/PaintsChainer You can also download the trained model data from here. http://paintschainer.preferred.tech/downloads/

So, if you have a GPU, you can easily create a clone environment. There are still some areas where training is insufficient, so please fine-tune it to make it feel better. I think I'll give a memorial service to the models that I didn't use.

Also, since I used a lot of man-hours during work, please make a star or pull request to chainer to become a contributor in order to make an excuse for PR of chainer. Thank you for chainer.

Thanks to mattya and rezoolab for their cooperation in creating this system! !!