Enhanced super resolution generative adverserial networks

how do I use "ESRGAN" on images on my computer? is there an existing (windows) program that lets me do that?

Attached: 1530002190396.gif (593x539, 435K)

Other urls found in this thread:

github.com/xinntao/ESRGAN
youtube.com/watch?v=cIXOH1tJJu4
youtube.com/watch?v=pZXFXtfd-Ak
tumblr.com/tagged/esrgan
youtube.com/watch?v=PupePmY9OA8
twitter.com/SFWRedditVideos

>windows

Attached: 1529056863233.jpg (1280x720, 99K)

install gentoo and emerge from the mud

Imagine using this with PS1 games that have pre-rendered background

Is GNU/Linux a better OS for this use case? If so can you recommend a GNU/Linux program and why it's better than Windows for this?
As you know, being a educated literate computer user, you use the least worst tool for a job.

no thanks.

Attached: 1529748637291.jpg (1920x1080, 255K)

no, because C# has ML.NET. stick to windows, don't fall for the guhnoo loonix meme.

Works fine under LSW, just enable LWS under Windows features and fire up i.e. Ubuntu and do everything the github page tells you to. Can't be any simpler.

>Ubuntu
so did I make a mistake?

Attached: 1518063847656.png (576x231, 15K)

You clearly don't understand what this is.

(You)

Attached: 1516017175241.jpg (2048x1536, 579K)

Oh fuck off libtard.

What is this ESRGAN this and what is its purpose?

github.com/xinntao/ESRGAN

github.com/xinntao/ESRGAN
PyTorch should work on Windows.

Fucking impressive.

Attached: qualitative_cmp_03.jpg (995x793, 291K)

>github.com/xinntao/ESRGAN

the real-life ZOOM IN AND ENHANCE

>As you know, being a educated literate computer user, you use the least worst tool for a job.
then why do you still use the most useless OS?

Zoom and ESRGAN

youtube.com/watch?v=cIXOH1tJJu4

>literally making a computer halucinate for you instead of using your own imagination

>not giving your imagination more substance to work with
Surface level thinker detected

Attached: 1536729017743.jpg (2048x725, 257K)

This coupled with Waifu2x is where the real future is. The ability to use learned models to add in details is immense, but the issue is that you end up with a lot of noise and also an equal amount of smearing. Waifu2x works on a similar principle, but the tool has the capability to denoise/magnify.

If tthe two can be merged into one model, wherein it would train to add in detail, but denoise at each interval group (say 2 rounds, denoise, 2 rounds, denoise, etc), and simultaneously do incremental upscaling at 1.25x rates. You'd be able to come out with some crazy impressive stuff.

But I think it will take roughly another decade to reach that point. The tools and knowledge to do it is all there, but the bottleneck is tensor capability for the consumer market. You need a massive amount of shaders and an equally massive amount of tensors for all the inferencing that needs to take place.

Running a raster through Waifu2x, even on my 1080Ti @ stock clocks, at say a base resolution of 1440p and bumping it up 4K, can take anywhere from 15 to 30s; especially if I'm doing denoising, then magnifying. If in 10 years, we can reach the capability of something like 2-4x1080Tis in a single GPU with roughly 2-3x increase in tensors from the 2080Ti, we'd be able to do this realtime in games, movies, tv shows, anime, and even run old manga scans through and have the network understand what its seeing and then based on that scale up and detail.

So wait, 'enhancing' images is actually doable now?
>tfw all of those detective shows can't be memed anymore

What? You're being sarcastic right? The results are completely inaccurate and unrepresentive of the actual objects. From the picture of the windows it looks like it substituted some images of windows from various third world countries.

As a human I can tell what's going on much better and get greater detail out of the thumbnail than the results of the 'AI'. No idea what's going on here, but maybe that should be the approach taken since it clearly does not preserve macroscopic details with the noise it generates.

The doom and Morrowind mods are fucking incredible. I honestly hope some autists of texture and sprite updates to every old game. I'm so fucking hyped.

Its still basically making shit up, just very convincingly. So this cannot be used for evidence.

Blade Runner is set in Current_Year+1 after all

>coupled with Waifu2x
Waifu2x is fucking garbage and this makes it irrelevant

Christ almighty this place is a joke sometimes

ENHANCE

from blade runner

the future is now

It'll be pretty trippy cuz every frame will be rendered differently.

>As a human I can tell what's going on much better and get greater detail out of the thumbnail than the results of the 'AI
>t. scared meatbag

So stupid. There isn't any more detail to get. They're all working with the same source. Any alteration is corruption and modification of the original. This is just a more convincing fake.

just tried the Doom version of this
It's pretty good, but it's certainly not magic, it basically just looks like a better version of those SNES emulator upscaling filters

Detail is not the right word; information, I can get more information by looking at the thumbnail because I can recognize macroscopic features like the stone window columns and individual strands of grass seed highlighted in the foreground. The algorithm should do edge detection first to preserve obvious macroscopic features that humans would recognize before filling in the picture with noise.

You as a person can do a comparison between the low res picture and the enhanced picture and easily determine that they are not the same while determining that the high res version is the same as the source. If you can do it then the goal should be to get the AI to do the same but in reverse to create a convincing detailed fake.

>This is just a more convincing fake.
Of course. That's the entire point. Your brain does the same thing when it extrapolates what the windows should look like from a low-res image. The difference is that your brain knows more about castles and windows than the neural network. For now.

>For now.

Attached: 1530763180960.jpg (1825x1825, 263K)

>can take anywhere from 15 to 30s
With or without cuDNN?

The best GPU I have is a 750 Ti. Will it be enough to experiment with machine learning or will it be painfully slow?

>ENHANCE
The future is now

How do you even go about implementing something like this on your own machine? have a 1080Ti and Ubuntu and I know basic math and programming. On a scale of 1 to 10 how hard is it implement state of the art papers into actual working code?

you can start by reading the thread first, and then checking the github link that has been posted multiple times already.

Attached: ydG7itRzjdFLkEYA6AUTKX-650-80.png (650x317, 229K)

It's cool and all, and I've seen the DOOM and Morrowind texture packs, but until someone upgrades Daggerfall I won't sufficiently care.

download it from the github link and updrage daggerfall yourself.

The low res version fed to the network is not shown there. What you are seeing is the actual high res version that the nets are supposed to guess.

It's not enhancement, it's just making up stuff that looks good enough

The high res is labelled HR. I'm talking about the thumbnails on the left. What I'm saying is that my brain can get more useful information out of the thumbnail by looking at it from far away than using the NN on it after enlarging because it just generates noise and fills the picture with garbage information.

Pic related; even without the whole picture for context, you can look at the scaled down image and recognize features that the NN just turns into mud in the scaled up image. I get that NN's are a new field of study and capable of doing some amazing things, but in this case an old fashioned hand written program that does edge detection on the scaled down image and uses it for weighting on the noise generated in the scaled up image would produce much better results.

Attached: 1545174549189.jpg (995x793, 294K)

I'm on a NUC. If I had something more powerful than a potato, I wouldn't be playing Daggerfall.

then you will not be able to play it with the upgraded textures anyway.

that thumbnail is not what the neural network sees. the NN sees roughly what you see in bicubic

If true that is amazing and the comparison image was produced by a retard. Why would they not include the original?

mpv devs look this shit

Exactly what I was thinking. Imagine FF7, Parasite Eve 1 and 2 and others...

>LSW

LSW?

TecoGAN
youtube.com/watch?v=pZXFXtfd-Ak

Add another neural network that tries to undo the damage of lossy compression and better remasters of games like grim fandango, where original assets were lost would become a lot more viable.

why every late 90s-early 2000s pre-rendered backround lost their asset??

tumblr.com/tagged/esrgan

OH MY FUCKING GOD

FF7 from tumblr

Attached: tumblr_peyvwpef2U1xvyxl5o2_1280.jpg (1280x1200, 530K)

Tried with Waifu, the results aren't bad.

I upscaled final fantasy FMVs in waifu to 960p with a GTX 760 and each one took like 10-14 hours.

Because Square are bunch of incompetents.

>I upscaled final fantasy FMVs in waifu to 960p
Cool. Can you post a screenshot or a WebM clip? I'm particularly curious how consistent it looks in motion.

that looks pretty bad

Looks a bit funky with the hatching artifacts but still not bad. If you trained the network specifically on 90s CG you would probably get much better results.

Imagine if Squeenix were competent. They'd start developing this right the fuck, make $$$ on enhanced ports and make their fans happy.

Attached: tumblr_penyfqaQpc1xvyxl5o3_1280.jpg (1280x960, 837K)

At the time is was allegedly company policy to delete assets in order to prevent them from falling into the hands of the competition.

because mathematicians and computer scientists understand how to read papers and they know their subject. that's why they write papers instead of shitposting on 4channel pretending you know a thing about anything

Mathematicians definitely shitpost on 4channel.

i know i'm a mathfag
but they don't pretend to know things they don't (generally, at least not good ones)

That blog is full of horrors.

Attached: doom dream.png (600x288, 436K)

this is Dario's house, right?

>The results are completely inaccurate and unrepresentive of the actual objects.

In other words, a LITERALLY MAGIC image processor would be needed to impress you.

garbage in = garbage out. You can't reclaim information that isn't there. You can only insert fake information. Which is fine for the majority of applications. You don't need to arrest someone or do an architectural history report based off these enhanced images.

>As a human I can tell what's going on much better and get greater detail out of the thumbnail than the results of the 'AI'.

No, you see about the same details as the bicubic filter, which looks adequate because it's at a small scale. No one wants "adequate" low resolution content if it means you have to render the image at a small size not to throw up.

>but in this case an old fashioned hand written program that does edge detection on the scaled down image and uses it for weighting on the noise generated in the scaled up image would produce much better results.

So you re-conceived the concept of the sharpen filter, which looks like shit compared to this. GJ

when will this technology be retard proof so i can use it by just double clicking a fucking .exe on muh windows 7?

Attached: 1482199979138.gif (500x375, 2.78M)

>A single exe with a gui you can't do anything at all with, source code is either not available or a visual studio mess
Get good

You can tweak tons of shit this way and actually apply it to something real.

Probably when nvidia releases their version

Also FYI, this isn't worth it unless you have a lot of VRAM, 6gb video cards can't even upscale a 720p image. Just consider it a near future technology for now.

Yup, I knew it looked familiar.

Attached: 1-masa42.jpg (640x480, 52K)

>Get Good
id love to, but i have no time because WE LIVE IN A SOCIETY

oh i see, these networks shits really look promising

>tfw machine learning is going to usher us into an age where we just type our fetishes in and hit enter and our computers are just going to create from nothing the perfect porn for us

Attached: 1544129094901.jpg (731x469, 106K)

Most likely meant "Linux Subsystem of Windows" I usually see people call it WSL though "Windows Subsystem for Linux".
It's an ABI compatibility layer to run Linux binaries like they used to do for Unix.

I've been messing around with topaz AI gigapixel, not quite the same but great for upscaling images

Holy shit shill somewhere else
/3/
/v/
Jow Forums
We get it you're proud of your software I'm so fucking tired of seeing you here. At least you haven't posted the downscaled image as an original next to the original saying that your software made the image perfect like you did elsewhere yet (pic related)

Attached: 1545272695019.png (1067x867, 182K)

kys autistic fag

I fancy these en·shill·adas
at least looks delicious

useless low hanging fruit. just uses existing high resolution art, downsamples, and reproduces original, will just reproduct artifacts of original data set.

creating new high res artwork or 3d boned models from poorly drawn sketches is still an unsolved problem. note that all this super sampling machine learning uses art which by itself already looks pretty good. i await anyone to show me generating high quality art from sketches, that would truly revolutionize the media industry

>Imagine jade cocoon

Attached: 82f02f04a0d114a09649ca8e4f18cff8--the-spider-cocoon.jpg (236x177, 13K)

is that yakkub?

>impressing a mentally ill
lol dude who cares, he's there in the basement for life for a reason, let him rot in the most absolute solitude and virginity until he just kill himself, no one will give a fuck lmao

Attached: 1528625693540.jpg (640x479, 62K)

youtube.com/watch?v=PupePmY9OA8