waifu2x texture upscaler

Hi all
First of all thanks for making the best damn doom enhancement around!

I have been looking at the waifu2x texture upscaler. It looks pretty good.
Is it possible to implement this in Doomsday as an alternative to smart2x?
https://github.com/nagadomi/waifu2x
Description.
"Image Super-Resolution for Anime-Style Art using Deep Convolutional Neural Networks. And it supports photo"

I have tried out some tests images of doom textures. Left is original resized 400% and right is waifu2x, then resize 200% with a bit of noise and sharpen to bring out the details (did this as a test so don't be too concerned with the values)

I think it looks like a good enhancement to the original textures. The transparency isn't recognised in some of them so the edges are harsh between the art and the cyan color, not sure if this is easy to fix or not.
So is this possible for the future?

aa45t4e7dssq.png
i6fynsnxiczq.png
dlna0hxsfptz.png
55s0uhi018pg.png
ngw6b0s21f2c.png
phv9xvwquyhy.png
pcrzw2hn0h3s.png
yn22jtm51i9r.png
u9vufms5jmg3.png
rvv6ggjz5joj.png
0lxe8eze25ml.png
wu6q78w7xgac.png
2gzlw1bogfyq.png
dd9lab5ae194.png
8ccrls6691z5.png
0bn0xjy21z1l.png
99hmj28207qb.png

Comments

  • edited 2018 Feb 25
    I took a quick look at the code, but it doesn't appear to be possible to integrate this into Doomsday in any feasible way. Also, the failure to smooth out alpha edges is an issue.

    IMO this looks like something one could use to manually prepare a high-resolution texture pack.
  • Thanks for taking a look skyjake.
    Thats a shame. The alphas probably need somekind of preprocessing first somehow to smooth the edges as you say. I'm not sure about where copywrite stands on making derivative images like this. But otherwise a batch operation that extracts, processes and upscales the textures and packs them into a pack, could work.
  • The guys at doomworld are all over this. Looks interesting.
    https://www.doomworld.com/forum/topic/99021-doom2-neural-upscale-2x/
Sign In or Register to comment.