So this was posted to, of all places, the programming subreddit: github.com/nagadomi/waifu2x (or a slowish web version at waifu2x.udp.jp/)

The TL;DR is Japanese computer scientists found a really smart image denoise/upscale algorithm specifically tuned for anime artworks through the use of neural networks. The results in my testing are super impressive compared to simple 2x upscale in a image editor. You can honestly barely tell it's upscaled at all, the lines are as sharp as they were at original size. The denoise is a little less impressive, but I think I tested it with an extreme case, it might work better on simpler scans/images.

You can get a similar effect from opening the image in firefox and zooming in.

No, really, you can't. Not unless your monitor is really bad.

For the interested parties it is worth noting that this Waifu2x stuff apparently has a command-line C++ port for Win32 (called waifu2x-converter) and also a moemoe GUI called Koroshell, using said Win32 port, to make laifu 2x easier. Sorry, had to... So yeah, check it here:

inatsuka.com/extra/koroshell/

Is this the CPU-faking-CUDA one? i tried this and its absurdly slow. A image takes like 5 min to upscale compared to 3-5 seconds on linux.

Well, for me, the standalone Waifu2x Win32 port doesn't work, it just displays an error message as if it was a corrupt or non-standard PE executable. Even after I actually installed (me silly!) the Visual C++ 2013 Distributable. (Not using Linux these years means I'm stuck with Win32 ports of anything useful from the Linux/UNIX ecosystem.)

If using the GUI, Koroshell, it appears to throw no error, but I don't see the Waifu2x port in the process list, only the Koroshell executable's CPU usage jumps up and down periodically. As far as my admittedly amateur programming knowledge go, one process cannot invoke a standalone command-line executable without it appearing in memory as a separate process. It would only work with DLLs. After several minutes I decided to stop testing because I don't know if it is even working or not.

Of course, one can question my sanity when I want to test a neural network upscaler/denoiser originally using CUDA on a mid-low MSI laptop which is more than 3 years old, is powered by an 1.8 GHz Celeron Mobile and still has good ol' 32 bit Windows XP on it, without a single reinstall. But hey, at least my GPU isn't an Intel GMA, but an nVidia GeForce 8200M G which ran Flatout flawlessly when I had it installed. Also, I didn't want to fed Koroshell with highres wallpapers. #firstworldproblems

But yeah, I saw people writing about Waifu2x being painfully slow without the right rig and right implementation, which is a shame. I'm stuck with the web demo version for now.

Its slow because its complex, however the uses for it are for pictures with lower then 1024< width or height. I converted all of the MGQ paradox CG set (3589 images) in 3 hours with a Geforce 680. It was +/-3 sec per image, so its fine for small sized images.

Would having a faster GPU help in conversion speed?

I have a GTX Titan X (with overclock).

That depends on where it gets the speed from.

If it's from more cores, yes. If it's from a higher clock rate with the same number of cores, probably not so much.

hmm, I tried and it didn't use the GPU at all, used CPU instead.

5820K @4.2 / 1.168v