waifu2x

https://github.com/nagadomi/waifu2x

Image Super-Resolution for anime/fan-art using Deep Convolutional Neural Networks.
Demo-Application can be found at http://waifu2x.udp.jp/ .

Why the fuck wasn't this posted here, you fucking redditor weebs?

It's a magical 2D art scaler/denoiser. Someone should give it one of those 320x420 dakimakura samples and post the results.

for anime

No, thank you.

I don't go on reddit!

>>2
e/g/in post /g/ro

Tried it with pixelart - results are order of magnitude better than with vector-tracing algorithms:

http://tinypic.com/r/1zqscp1/8

A scaler for DOSBox would be appreciated.

give it one of those 320x420 dakimakura samples and post the results

Original:
http://s22.postimg.org/nl2f8otvz/kyou_dakimakurasmalluncensored1.jpg
http://s22.postimg.org/f1j1axljj/kyou_dakimakurasmalluncensored.jpg

Upscale:
http://s12.postimg.org/xjpruxpfv/kyou_dakimakura1.png
http://s12.postimg.org/94hjtvqiz/kyou_dakimakura2.png

Pretty good.

>>6
You're taste in anime girls is terrible!

>>7
Shut the fuck up, you dumb nigger.

More comparisons I made: http://imgur.com/a/2AAYs
It does well over all. It was negligible with lewd Konata over bicubic, but over the photo of fig Konata it was by far the best. The weird one with Miu and Chiyo is also better than bicubic by a rther large margin in my opinion. It also is best on the photo of the gay nigger sucking on the candy cane. It is also the best on the doujin screencap, but the dots are now to big (still better than blurry dots though).

It does well.

>>9
The waifu nigger is less blurry, specifically eyes are sharper. Same for pixel art and many textures: waifu blurs less than xbrz.

Interesting, can you use this algorithm for lossy/loseless compression based on pixel prediction? Say you subtract original image pixels from its downsampled version, which, when image predict original, reduces channel bandwidth.

>>10
Yeah, I was thinking the same thing. This would be a game changer for distributing content to the plebs on mobile phone distribution channels like Apple Store, Android Store, etc. At the moment, Apple has a 100MB limit for cellular distribution, and Google has it set at 50MB for Android. If you're over the limit, plebs will receive a notification that they need to connect to WIFI in order to download, so you end up losing customers, because normie plebs easily forget that they wanted to download your game 10 minutes earlier.

However, from my tests so far, it's pretty computationally expensive, requires a powerful GPU. I think it would be possible to do it to a limited degree on more recent mobile hardware, but you'd need to keep a waifu2x'ized texture cache on disk. Only possible on cutting edge phones with OpenGL ES 3.1 compute shaders or latest Apple devices with Metal. Definitely some challenges to overcome.

>>10
Yeah, I was thinking the same thing. This would be a game changer for distributing content to the plebs on mobile phone distribution channels like Apple Store, Android Store, etc. At the moment, Apple has a 100MB limit for cellular distribution, and Google has it set at 50MB for Android. If you're over the limit, plebs will receive a notification that they need to connect to WIFI in order to download, so you end up losing customers, because normie plebs easily forget that they wanted to download your game 10 minutes earlier.

However, from my tests so far, it's pretty computationally expensive, requires a powerful GPU. I think it would be possible to do it to a limited degree on more recent mobile hardware, but you'd need to keep a waifu2x'ized texture cache on disk. Only possible on cutting edge phones with OpenGL ES 3.1 compute shaders or latest Apple devices with Metal. Definitely some challenges to overcome.

>>9
That nigger's left eye looks messed up. His lower eyelid has melted into the eyeball.

Seeing the word pleb has explained so much of what's going on here as of late.

http://i.imgur.com/ms4lXff.png
http://i.imgur.com/sF5wkKj.png
It added JPG artifacts when there weren't any in the original. WTF?

>>15
Cudder, just shut the fuck up and die already.

>>14
plebs gonna pleb with pleb pleb pleb

>>16
Pixelation isn't artifacts. Besides that, it is a neural net trained to upscale anime porn; the fact that it does so well with text and photos is still impressive.

I'm just glad that someone is doing something useful with NN instead of just holding press conferences saying ``Doh hoh, we got this neural net we paid four hundred million dollars for to identify a cat in a video one time'' like Google does.

>>15
I'm surprised it has preserved pixeled font as pixeled. I.e. no blur. That is the only upscaling algorithm discerning between antialiased and non-antialiased. Also love how it have amplified deficiencies in Linux's font hinting, look at the `x` in "Mentifex"

>>18,19
The ultrasharp scaling of the text in the Konata image provoked me to try that. It's not perfect though - look at the noise it adds around the actual upscaled pixels (simple pixel-quadruple scaling of a piece of the 2nd image in >>15):

http://i.imgur.com/ZRa6iOX.png

That's not Linux font rendering either.

Here's another one, a ТИТ-0249бис test card:

http://i.imgur.com/f58oEyF.png
http://i.imgur.com/kEt6RQL.png

Very impressive.

I want to see how it does with photos after it's been trained with photos instead of anime girls.

I but I definitely don't want to spend time trying to grok all this CUDA nonsense. Can I run it (even slowly) on a $100 graphics card or not?

Don't get me wrong, GPU stuff is neat and all that and I'm sure it lends to this process in some way but hardware companies are awful and all these proprietary APIs are cancer.

>>21
Then run your apps without all that pesky hardware.

>>22
I said they were awful, not unnecessary, you fucking faggot.

>>23
Wrong, I am a virgin faggot.

>>21

Can I run it (even slowly) on a $100 graphics card or not?

In theory, sure. It may take some coding, but this is /prog/, so if if you are afraid of that, maybe you shouldn't be here. A shitty graphics card probably can't be used for GPGPU purposes, but you may be able to hack something together with shaders or something, who knows.

Of course, it probably uses the GPU for a reason. Neural networks are embarrassingly parallel algorithms, so a plain CPU will probably never match it. I'm pretty sure Amazion rents GPU compute clusters, so maybe you can steal your mom's credit card if you want to experiment with it without buying a new computer.

In theory, sure.

You could have just said no.

if you're afraid of that, maybe you sh

"Afraid" and "having enough of a life that digging into an arcane proprietary GPU API that will almost certainly be completely useless in several years is a waste of time" are two different things.

Amazon rents GPU

So I need to rent time on a mainframe just to run software? Terrible! The only reason I asked is that I bought some shitty Nvidia card to play some stupid video games. Nice to know it is already obsolete.

>>26
Back to /g/, please.

>>27
/g/ is going to be even more 3d card manufacturer dependent algorithm apologist than you are being right now. Maybe you need to go back there, faggot.

>>28
You're the one that thinks everything should be handed to you on a silver platter so you don't have to spend any time thinking or making an effort, apper.

>>29
I realize that wasted efforts are the shiniest badges an autist like you can wear, but outside of the basement no one would consider satisfying extremely casual curiosity in this fashion. Just a pro tip, fellow memer.