search

nagadomi / waifu2x

Image Super-Resolution for Anime-Style Art
http://waifu2x.udp.jp/
MIT License
27.44k stars 2.71k forks source link

[Question] How to use on windows? #162

Open BennyAlex opened 7 years ago

BennyAlex commented 7 years ago

Hello, I know there is a version for windows (https://github.com/lltcggie/waifu2x-caffe) but this version dont use the the original waifu2x code, or? It is just a rework?

The problem with waifu2x-caffe is that it dosnt get updated often and it is mostly in Japanese and I have errors running it in english...

So is there a way to use your (the original waifu2x) code and run it on windows?

nagadomi commented 7 years ago

Probably this code does not work on windows. I haven't tried that. The latest waifu2x-caffe is compatible with my pretrained models and supports multilingual UI.

BennyAlex commented 7 years ago

@nagadomi theres a english UI, but when I select some pictures, the path contains japanes letters and the application crashes when I want to start upscaling...

I am wondering about the waifu2x.lua code. As far I understand, waifu2x-caffe uses an own code? So there a probably worse results then with yours? Sorry I dont understand how the hole machine-learning upscaling work so please forgive me to ask this questions.

nagadomi commented 7 years ago

that error maybe related to the locale settings. but I am not familiar with windows env.

As far I understand, waifu2x-caffe uses an own code? So there a probably worse results then with yours?

The big difference between waifu2x-caffe and my implementation is that it uses a different neural networks framework. waifu2x-caffe uses Caffe, my implementation uses Torch. But the principle part of waifu2x is a set of coefficient and formulas to execute it (we called it pretrained model. that files can be found at ./models/*/*.json ). waifu2x-caffe uses same pretrained model as my implementation so the upscaled result by waifu2x-caffe is same as my implementation of that. The different parts are image loading, saving, UI, and small numerical error.