WAIFU2X VULKAN MULTITHREAD

>github.com/nihui/waifu2x-ncnn-vulkan
Waifu2x has latest vulkan update with multi-thread support and its looking sexy.

Now its roughly 2x faster than "waifu2x-caffe-cui" in single threaded and 4-5x faster with default multithread and scales well with GPU too.

Features batch folder convert for anime conversion aswell.

Attached: waifu2x.png (1058x2382, 1.19M)

Other urls found in this thread:

geeks3d.com/dlz/#amd-graphics-drivers
twitter.com/AnonBabble

my workplace forced me to be a macfag
now what

Nicesu

Don't use your office computer for this shit.

does it work with amd gpus now

>VULKAN
Any GPU that supports vulkan works. Including AMD and nVidia but not limited to.

>miku, now more of a slut

Still not as fast as cuda I reckon

What does this do again?

2x faster than CUDA.

AFAIK, it uses bit more GPU memory than CUDA implementation but the speed up is worth it.

Can we get some actual numbers on these claims?

Its on the github page.

Looks good. Any idea why it runs the wall at 4k->8k?

So this is the power of weaponized autism.

>wall @ 4K->8K
No idea tbqh. But I don't really need THAT sort of upscaling on my anime. Just 480p->1080p is more than enough.

how to video player

The the jpg output needs to be added. PNG compression is fine for picture perfect fidelity, but it costs a ton of space.

If I needed it, I'd have looked into issues already. More general interest why it happens rather harsh at this stage.

storage is cheap

Only if you brainlet. Time large storage cost time. Time is money.

>windows only
bud why

Works better

Until someone makes a way for me to use it directly on an MKV, I wont bother.

I just can't be bothered to go through the effort of putting my MKV into ffmpeg and extracting every frame, THEN throwing that through waifu2x, and THEN recreating an MKV from the resulting frames.

test

Attached: waifu2x.png.png (2116x4764, 3.87M)

>weebs driving video and image scaling innovation
Who would of thunk it?

Attached: 1554953564372.png (502x360, 204K)

Trannies are annoying, but they're pretty good at actually getting shit done.

Is there a gui for it?

do you actually know any trannies?

>needing a gui

Only seen those on the internet.
Whenever a """female""" actually manages to do something even mildly impressive that can't be trivially attributed to "her" mostly-male team, that's a tranny.

>weeabo == tranny
go fuck yourself.

Non-tranny weebs are trash like pedos instead and pedos are low IQ and can't write software.

Anyone get this to compile on linux?

file names are annoying to call up

What about mpv, dxvk, waifu2x? Go fuck yourself.

Trannies

the programmer behind this looks a bit strange

Attached: 146487878968.gif (167x204, 492K)

Found the nigger

Dumb question, but are there any numbers comparing it with the original waifu2x? Or can it be assumed that waifu2x-caffe is just as fast as waifu2x?

There's video2x. Just be sure to deactivate the strong denoising (especially if it takes more than one upscaling step) or you'll get horribly washed out results.

waifu2x-caffe is refined version with more support/gui/etc for windows.

Not trannies, fucking retard.

Prove it.

Cool, but vulkan support for illustration2vec when?

why is Miku so perfect...?

Attached: 1563168942777.jpg (600x828, 93K)

>slutmiku
>not trash

There's cafe but it doesn't support this vulcan thing.

I know that. I just don't know how Caffe compares to Torch7 in terms of speed.

>ncnn

Nice try, China. I only run American botnet software.

>t. seasonal waifubait enthusiast

>liking the slutty version of a waifu
>not being a total cuck

>not waifu2x-ing that image
It's a low-res noisy image. Would've been a perfect chance to give Miku the respect she deserves.

testing

Attached: 1563150366322-edit.jpg (6348x4680, 1.56M)

more like oilpainting2x

In my experience even -noise-level 3 won't get rid of such severe compression artifacts. Also no point in waifu2x-ing a shitty JPEG, when a high quality (but still low resolution) version can be easily found.
If you really want a higher resolution then run waifu2x on that one.

Attached: 9862158_p0.png (600x828, 399K)

Woops already ran on jpeg.

Attached: distor-chan.jpg (3000x4140, 2.3M)

here is the other image

Attached: somehow even worse.jpg (3000x4140, 2.02M)

>distor-chan.jpg
Autism, but also cute

>somehow even worse.jpg
What are you talking about?
Also what settings did you use for each image?

It's pretty based, but I hate how it adds .png on top of each file instead of directly replacing it in directory mode. Also no progress bar or anything. No option to leave the same file type (which caffe also lacks btw). All in all I love it, at last waifu can finally be free of nvidia chains and work on all gpus. Once aforementioned issues are fixed I'm ditching caffe for good.

Not 100% HQ but it work for what it does.
Used github sample upscale 2x with waifu2x noise=2 scale=2

waifu2x-ncnn-vulkan.exe -i origin.jpg -o output.png -n 2 -s 2

Attached: 75152109_p0.jpg (900x1270, 148K)

Attached: output.png (1800x2540, 3.18M)

>presses Tab
nothing personel

Attached: jaypegbecauseoutputtoolarge.jpg (1800x2540, 642K)

What if I need to improve some 3D photos? Any good alternative?

Shouldn't cunet handle both of those styles well?

tried this but no output file? what going on

You know, if you want to know what's going on, you can try reading the error message on the command prompt. If you don't know what they are, you can always post your command/results.

Random "whats going on?" isn't going to help anyone, least of all, yourself.

>increase image resolution
>save it with lossy compression

there is no error message

Attached: no error message.png (862x737, 42K)

You already had a Waifu2x fork that works with OpenCL for AMD chips

gpu?

And?

what is the point of a picture with more than 2000px on the vertical side if you're never gonna print it as A3

it's not like you fap zoomed in

Attached: adamsface.jpg (395x499, 39K)

>it's not like you fap zoomed in
You don't?

Attached: 1317355572758.png (400x399, 214K)

I need to make eye contact with the animu girl

Its all in the mind. You gotta believe.

So i can only upsample a 128x128 image to 256x256?
what's the point

390X
It was idling the entire time

Does waifu work with actual photos, are the results any good?

try adding "-gpu 0" or "-gpu 1"

Maybe its conflicting if you have an iGPU/not being detected/etc. Or vulkan driver on 390X isn't complete(maybe driver update?). Or the code has some issues with older AMD.

CPU has no IGP though

Also, doesn't work

You need to install vulcan driver? Or the latest drivers are vulcan itself? The latest driver package contains the 2017 drivers

Yes

Attached: Arch.jpg (1080x1032, 298K)

>You need to install vulcan driver? Or the latest drivers are vulcan itself? The latest driver package contains the 2017 drivers
Not sure, but nVidia updates vulkan drivers and comes with latest GPU drivers AFAIK. So maybe its there with latest AMD drivers.

The latest driver package from AMD installs the 2017 drivers though for 390x

I'm sorry I can't take you seriously with a name like that

Wait, do you need the radeon software to use vulcan? Currently running the latest bare driver because the driver software is bloat and cringe

IDK, I dont have Radeon lol

If you can't use it, then the only option left is the opencl version of waifu2x.

>geeks3d.com/dlz/#amd-graphics-drivers

According to AMD the latest vulkan driver is v2.0 and API version 1.1.

So whatever the case, your GPU has outdated Vulkan drivers.

Yes, there are extra photo models. The results aren't as good as for drawn content, but still better than conventional upscaling algorithms.

Can it upscale in real time now?

nice

waifu2x-ncnn-vulkan 4000x4000 8000x8000 400/200/100 22.78/23.78/27.61 2448/654/213
waifu2x-caffe-cui 4000x4000 8000x8000 400/200/100 18.45/21.85/31.82 3325/1652/1236

slower than the caffe build on some cases.

No, but the speed up is happening. Few more optimization layers and it could potentially do real-time. Dandere2x may have some potential to do another speed up trick, the dev there needs bit of work on getting it setup first. The idea is nice in principle.

Attached: fro2.png (800x798, 587K)

Attached: noise2.png (800x798, 625K)

Attached: noise1.png (800x798, 611K)

Just pipe it to something that saves in jpeg? Don't see why the scaler needs jpeg output.

Its an extra computational step. In a program where all the compute time is critical, an option to save as jpg is very useful.

Raw output to a generic jpeg compressor should take virtually no extra time for something like saving an image.

unless you're dealing with hundreds of thousands of images, I just can't see it being a big enough impact to matter.

"Virtually no extra time" is irrelevant because we're talking in ms/seconds in a program where every little ms counts.

A single ~20-30 min anime episode is ~40K worth of images. Batch of 12 episode is 480K. 24 is 960K. I'm making my own GUI for converting anime in batch. Reason I wanted native JPG output is just for file saving measures. Its not too big of a hassle for 1 episode, its still quite a large file size requirement. 480p->960p upscale of 40K files requires ~40GB for PNG. Scaling that up to 12 episodes would give ~ 500 GB. Mind you, the original is only ~150-200MB 480p files. Unless I'm seeking some "videophile" level placebo effects, that sort of file size is just not worth it on a SSD and certainly not worth the wear/tear when a very comparable jpg quality can be had for 1/10th the drive usage.