[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / r / s / t / u / v / vg / vr / w / wg] [i / ic] [r9k / s4s / vip / qa] [cm / hm / lgbt / y] [3 / aco / adv / an / asp / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / qst / sci / soc / sp / tg / toy / trv / tv / vp / wsg / wsr / x] [Settings] [Search] [Home]
Board
Settings Home
/mlp/ - Pony


Thread archived.
You cannot reply anymore.



https://www.youtube.com/watch?v=XTgRlYLbc44
ZX did a 4K upscale test with the iTunes rip and waifu2x.

>This video is a test demonstration of what My Little Pony: Friendship is Magic could look like if it were properly upscaled into 3840x2160 (4K) resolution.
>To begin, I exported an image sequence from Adobe After Effects. This sequence contained 844 PNG files, totaling 1.16GB in size.
>Once all 844 image files were exported to a directory, that directory was then imported into an image upscaling software called waifu2x.
>Each image was then magnified and de-noised with the above settings, outputting a clean, sharp 3840x2160 image.
>Once all 844 images were successfully upscaled through waifu2x, I then imported the image sequence back into Adobe After Effects. The composition was set to 3840x2160, with a framerate of 23.976.
>Overall, upscaling this 35 second clip took me around two hours of total time investment.
>>
File: WANT.png (165 KB, 880x800)
165 KB
165 KB PNG
>>33848333
Whole series when?! Hasbro will probably nuke us for such kind of piracy.
Fuck you, though. I've just bought FullHD projector for watching pone and can't afford 4K shit.
>>
>>33848399
Apparently one season would take 72 days with his hardware.
>>
>>33848444
We can wait.
>>
It's garbage worse than original
good job
>>
>>33848444
What hardware does he have?

I’m pretty sure you could rent an EC2 instance with nuclear levels of power and do the whole series in a matter of hours
>>
>>33848333
>all that hassle when he could have just simply used the waifu2x plugin for Avisynth or Vapoursynth
in fact, I think I'm gonna test it right now
>>
>>33848444
I believe that is why you made this thread? So >we can do this together?
Also, stop playing with numbers, you scaring me.
>>
>>33848333
When was this uploaded? I'm phone posting like a filthy nigger right now and can't see.

I feel like this was inspired by something I did not too long ago. Though maybe I'm inflating my ego a bit here.
>>
>>33848577
today
>>
>>33848333
>upscale
Into the trash it goes
>>
>>33848614
Interesting. Was there any prior knowledge or indication of such a thing happening? By him or anyone else?
>>
>>33848465
GTX 1080 Ti
>>33848626
fuck if I know, I just saw it in my sub feed
>>
>>33848631
I just bring it up because I literally did this exact process with some porn gifs and webm's a couple weeks ago (with some differences, like I used ffmpeg instead), and I can't help but think he saw those posts and got inspired to do similarly.

Not that I would mind if he did, but I'm dying to know now if he did.
>>
File: mpv-shot0001.jpg (808 KB, 3840x2160)
808 KB
808 KB JPG
>>33848618
>>upscale
What's wrong with it?
>>
>>33848642
What hardware did you use?
I don't know too much about video rendering but I wonder if something like a 2950X would be better suited.
>>
>>33848657
I also used a GTX 1080Ti. W2X is better utilized on the GPU than the CPU since it takes advantage of CUDA cores in the GPU. It still takes a long time, though, especially if you're going over 4k.
>>
>>33848681
Is that using the NVENC hardware encoder or just the GPU cores to accelerate encoding?
>>
>>33848643
He shouldn't have uploaded the video to Youtube since YT uses it's own compression algorithms which cause artifacting to the source, nor should you have taken that screenshot in jpg which also adds another layer of compression and artifacting. You can see how bad it looks up close.

>>33848701
I believe it's just the CUDA cores for the *acceleration* portion. I think you can use the GPU still on its own without them, but it might take a little while longer. I don't know all the technicaly specifics. Honestly, my knowledge is about as much as yours, maybe only marginally more.
>>
>>33848642
>some porn gifs and webm's
pony ones?
>>
>>33848701
Not him. It's not much about video rendering but about neuro-math crap that upscaler uses so you graphic card have to support math acceleration (CUDA). Otherwise it takes long time to upscale frames.
>>
>>33848719
Eeeeuhhh, *a* pony one. The rest of them EqG and Humanized

Here is the pony one.
https://derpibooru.org/2021904

I designed them so they could be posted on 4chan size-wise, which means I was restricted to using libvpx for the webm encoding process, but then it turned out it was too big anyways just due to the length of the source. It came out pretty damn good anyways, so I said aight.
>>
>>33848333
>Upscaled to 4k
>Only option is 720p
I hope it's not just my phone here.
>>33848643
>JPG
>Blur and jagged edges even when looking up close.
Next time use PNG. Not trying to be a jerk but JPG and JPEG stretch and compress when saved and even sharing it makes it worse as it degrades over the span of time. There is a reason why images and certain memes of 00's look really bad or even worse than it used to be. PNG has no compression which results in a bigger file size and can stand up much longer for archival and viewing purposes.
>>
>>33848708
>nor should you have taken that screenshot in jpg which also adds another layer of compression and artifacting
>muh jpeg boogeyman
I'll have you know that jpeg != automatically bad quality
>>
File: 1375724636599.png (39 KB, 500x476)
39 KB
39 KB PNG
>>33848708
>>33848723
Okay I misread, I thought he was talking about using CUDA cores for the actual rendering of the video file - whereas W2X uses CUDA to run the upscaler.

I wonder if they could use the tensor cores on the RTX GPUs to accelerate it?

t. machine learning brainlet
>>
File: 220.jpg (36 KB, 400x506)
36 KB
36 KB JPG
>>33848740
>PNG has no compression
>>
>>33848740
>Next time use PNG. Not trying to be a jerk but JPG and JPEG stretch and compress when saved and even sharing it makes it worse as it degrades over the span of time. There is a reason why images and certain memes of 00's look really bad or even worse than it used to be. PNG has no compression which results in a bigger file size and can stand up much longer for archival and viewing purposes.
Literally all of this is wrong, bruh.

>>33848744
It doesn't if you export it as such from the raw source, but default saving methods use a compression level that do cause noticeable degradation of the source.
>>
>>33848708
Correct YouTube does use slight compression algorithms that are not the best and Google is trying to use Webm video standards like AV1 and HEVC that would have greater compression without loss of quality.
>>
>>33848749
It has barely any compression compared to JPEG and JPG. I dare you to prove me wrong.
>>33848752
If I am wrong then prove it. PNG is the better standard of quality.
>>
>>33848734
derpibooru.org/2021904
neat
>>
>>33848754
Yeah, but the thing is there's still a lot of people who use hardware that doesn't support HEVC hardware acceleration (like me lole) which means it's going to lag and break and basically not work right.

It's like fiber. One would think every household would have it by now, but amazingly, DSL and even fucking dial-up is still in common usage.

>>33848757
I want you to do me a favor. Go onto /g/ and say the exact same thing you said here. Report back after a while with your findings.
>>
File: always Png.png.jpg (86 KB, 600x896)
86 KB
86 KB JPG
>>
>>33848749
>>33848752
https://youtu.be/ww12lImOJ38
False. PNG is lossless.
>>
>>33848765
>Says to use PNG when saving as a shitty JPG
LMAO this never gets old.
>>
>>33848769
Just because it's a lossless format doesn't mean that it doesn't use compression. It absolutely does, but it's LOSSLESS compression. Still compression.

Dude just stop.
>>
>>33848772
>the_joke.png
>>
File: Untitled.png (38 KB, 585x864)
38 KB
38 KB PNG
>>33848765
>always Png.png.jpg
kek

Yes, you can see artifacts on jpeg sceenshot, but I believe that jpeg is enough to see difference between 4K and 1080. Picrelated is same screenshot in jpeg and png.
Also, png exceeds image size limit.
>>
>>33848760
>Go to /g/
How about you go on there and tell them that you use JPG and that there is no issues with JPEG at all compared to PNG.
>>33848776
Yes but it's much better than JPG which was the entire point of my argument. It's a better compression and it does not try to stretch the image and blur everything.
>>33848777
Good at least you used PNG.
>>
File: Untitled.png (12 KB, 10000x10000)
12 KB
12 KB PNG
>>33848757
286MB as PPM
>~24414:1 compression ratio
>barely any compression
Based retard
>>
Don't let the retards itt discover that JPEG standard supports lossless compression or they're going to have a meltdown
>>
>>33848783
>but I believe that jpeg is enough to see difference between 4K and 1080. Picrelated is same screenshot in jpeg and png.
This is true. It's only when you start to zoom in you begin to see the issues or those with extremely good eyes to see the difference.
>>
File: intro.png (1.05 MB, 3840x2160)
1.05 MB
1.05 MB PNG
>>
File: intro.jpg (503 KB, 3840x2160)
503 KB
503 KB JPG
ABjpeglets btfo'd
>>
>>33848796
now if the standard also supported transparency it would actually be useable
>>
>>33848788
What you intended was not what you said. What you said was demonstrably wrong, even if you meant something else.

>How about you go on there and tell them that you use JPG and that there is no issues with JPEG at all compared to PNG.
You dare to put words in my mouth? Why is a graceful acknowledgement of wrongness so impossible with you zoomers?
>>
File: 1994118.png (447 KB, 791x707)
447 KB
447 KB PNG
>>33848788
>it does not try to stretch the image
What did he mean by this?
>>
File: 1498266007114.png (1.31 MB, 882x767)
1.31 MB
1.31 MB PNG
>>33848783
>Also, png exceeds image size limit.
>he doesn't optimize his .png's before uploading
>>
I really pity that 4chan removed the security flaw allowing one to upload 800MB+ jpegs to the site
>>
>>33848444
I explained in the text to speech in the VA's voice thread how to get tons of computing power for free.
But you fags dont listen.
So im going to write a TLDR version:
1: register a new M$ account, apply for free azure credits
2: allocate a VM with 4 nvidia V100 GPU (faster than a 2080ti and has 3 times as much ram(32GB/gpu, so 128GB in total)), the machine also has 24 CPU cores and 448Gb of RAM.
3: ????
4: profit

You can run this machine for 3 hours for free. I bet that you can easily upscale a single episode in 3h with that machine. If not, then convert waifu2x to fp16 which will result in 2-8x speedup while barely reducing the quality (will not be noticable by human eyes).
>>
File: AJ Doubt.jpg (91 KB, 600x1025)
91 KB
91 KB JPG
>>33848891
>will not be noticable by human eyes
>>
>>33848891
But what if I have augmented eyes? Your argument goes out the window right there.
>>
>>33848819
>>33848824
What's your point?
>>
File: 1498279666666.jpg (83 KB, 723x724)
83 KB
83 KB JPG
>>33848929
Stop it.
>>
>>33848333
You know there are leaked flash assets to S8 theme song. You can even render 16k if you want without need to upscale.
Plus I have access to some flash files you won't find. Including episode clips and more. I won't share tho. I have my reasons.
>>
>>33848950
Big Hackintosh? Is that you?
>>
>>33848955
No.
>>
>>33848950
>You can even render
Well why don't you do this so we can compare results?
>>
>>33848974
my PC is crap
>>
>>33848994
Too bad. I probably have a decent PC but I'm Adobe Brainlet and don't know if it even run on my OS (Linux).
Hope some other anon would do this. I'm really curious to see differences between rendered from source and upscaled versions.
>>
>>33848333
May I suggest using this instead?
https://github.com/k4yt3x/video2x
The tool is free, and you don't need to use After Effects. Hell, split the job up among several users, and we can convert all of MLP to 4K in two weeks, tops.
>>
>>33848950
>Plus I have access to some flash files you won't find
Yeah, sure bro, and I have access to Hasbro's FTP server, and already downloaded Season 9 finale on my hard drive
>>
File: snap.jpg (70 KB, 1073x499)
70 KB
70 KB JPG
>>33849111
what about that?
>>
>>33849121
>ayyqg trash
I'm actually glad you're refusing to share this.
>>
File: nani the fuck.png (1.35 MB, 2290x2348)
1.35 MB
1.35 MB PNG
>>33849121
Bruh, dude, share that shit right the fuck now.
>>
>>33849128
not only EG but that was first file I decided to open.
>>
>>33849137
[Citation needed]
>>
>>33848915
>>33848919
I bet you faggots dont even know what fp16 is.
>>
>>33849032
This looks interesting.
Unfortunately I think both my R9 280X and E3-1231 v3 are a bit outdated.
>>
>>33848783
>>33848929
How are these comparisons different?
>>
>>33849241
On the first you can see more "dirt" on borders of solid colors (see right white stain). On the second you can see overall quality decrease because it was saved in unsuitable for comparison format so please disregard it.
>>
>>33848838
>What you intended was not what you said. What you said was demonstrably wrong, even if you meant something else.
well I was wrong then and I will just ignore your attempt to label me a Zoomer.
>>33848847
If you watched the video I linked it would make sense to you on how image data works. It's too complex for you to understand.
>>
>>33849032
Looks interesting. But what’s the best source material? iTunes 1080P?
>>
>>33849572
Go for Amazon/Netflix rips. iTunes has shitty H.264 encoding. In any case, 1080p is a must.
>>
>>33849609
Netflix has better audio quality but their video encoding is noticeably worse than the iTunes.
>>
>>33849363
>If you watched the video
I did. Nowhere does it mention stretching. If you meant ringing artifacts (of course zoomer like you wouldn't know how they're called), then hooo boy, are you retarded.
>>
File: tempest.waifu2x.webm (2.88 MB, 2048x1616)
2.88 MB
2.88 MB WEBM
>>33848467
This took forever. I used the German Blu-ray because it looked better and had to crop it because of resolution limits. Vapoursynth script:

import vapoursynth as vs
core = vs.get_core()
clip = core.ffms2.Source('mlp.de.mkv')
clip = clip.std.Trim(93406,93406).std.Crop(0, 0, 136, 136).fmtc.bitdepth(bits=32).w2xc.Waifu2x(0,2,512,False,0,-1,False,True)#.fmtc.bitdepth(bits=8)
clip.set_output()

Anyway, any vector-based animation should benefit from a 4K upscale if the HDR filters were properly applied, but that requires special tone mapping and calibration for the BT.2020 colorspace which I couldn't do myself because I only have regular SDR screens. On the other hand, celluloid animation doesn't come up looking as good unless you clean up the shit out of it like those Disney Blu-rays, which ends up messing up a lot of the picture.
>>
>>33848333
holy buck nice trips but also holy buck please keep doing this
>>
>>33849912
>I did. Nowhere does it mention stretching.
You can not get into reading comprehension LMAO!!!
>If you meant ringing artifacts (of course zoomer like you wouldn't know how they're called), then hooo boy, are you retarded.
LMAO a horse can't understand artifacts.
>>
File: rd-happy.png (160 KB, 900x888)
160 KB
160 KB PNG
>>33849320
Thanks anon
>>
>>33850292
Since you brought up the movie, the high bitrate leak should help tremendously, shouldn't it?
>>
File: waifu2x leak.png (1.39 MB, 1920x1080)
1.39 MB
1.39 MB PNG
>>33851425
The improvement is slight but I think it's there. This is from the actual 128GB file, I don't have enough space to uncompress the whole thing so I'm just working with the intro. The big issue with the leak is that is missing post-processing and some VFX.
>>
File: waifu2x bluray.png (1.89 MB, 1920x1080)
1.89 MB
1.89 MB PNG
>>33852351
And this is from the German Blu-ray. For some reason, the picture is ending up brighter. I thought at first it was the colorspace conversion but it has to do with waifu2x.
>>
>>33852351
>>33852407
So the colours are not from missing post then, interesting.
Did anyone check if show colours change too?
>>
Is there any way to retrain network specifically on MLP? Would it benefit result somehow?
What about vectorization upscale algorithms? Will they work better or worse?
>>
>>33854341
>retrain network specifically
Neural networks? That would probably take even longer.
>>
>>33848643
https://www.youtube.com/watch?v=QEzhxP-pdos
>>
File: Spoiler Image (556 KB, 1320x1026)
556 KB
556 KB PNG
>>33848333
This is some good shit. Thanks for doing it.
>>33848470
You want to see some scary numbers?
>>
bump
>>
For those with the hardware, and the time to set it up, you can get a pretty good live 4k playback from madVR + LAVFilters.
https://www.videohelp.com/software/LAV-Filters-Megamix

I mean, I assume at least. I'm getting about 2 frames per sec on the high-end settings, but they're damn good lookin frames



Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.