Oh holy shit: deepsense.ai
Nvidia basically took this: github.com
And implemented as API/driver logic via its Tensor Cores with Convoluted Networks.
Required Reading: deepsense.ai
Oh holy shit:
Matthew Morgan
Luis Williams
Secondary reading: arxiv.org
Kevin Jackson
Can I use this to upscale Naruto to 4k?
Christopher Hall
Technically, if SVP added functionality to take advantage of DLSS while engaging in frame interpolation. You could theoretically waifu2x Naruto to 4K60 (anime) with very low noise and artifacts.
So yes, needs some work, but yes.
Andrew Roberts
Too bad we'll never see it outside of some expensive proprietary software for professionals
Noah White
No, there's a good chance we'll see it consumer market. Waifu2x was released in late 2015/early 2016. We're looking at GTX 980 generation. If mid-range and high end consumer has access to fixed function Tensors; there's a much greater incentive to push this kind of stuff out. Also, SVP is more Nvidia specialized than AMD specialized.
High chance we'll see DLSS with video interpolation to higher framerate within the next 3-5 years. Nvidia will never go back from this point on to a purely shader uArch model. Its now going to be Shader + RT + Tensor + any other future fixed function they can leverage to do more things all at once.
Blake Morales
I can't believe there are weebs that are actually good for something.
Camden Williams
x264 was lead by a weeb that eventually became a trap
Anthony Thompson
waifu2x is shit
Carson Turner
mentally ill people make some good programmers.
Justin King
Except, that's exactly what DLSS is doing but a bit further. Its take a low resolution image and then using Convoluted Networks (DNN), scaling it up by analazing the color boundaries and increasing them exponentially.
Basically: image is fed into the DNN algorithm, it uses tensor cores to do an elder god tier levels of fast analysis of all color boundries and stores it into memory. Then with shader operations in parallel, draws a super sampled version of the data while mainting the color boundary and addressing any noise & artifacts as they crop up.
The end result is an image that at resolution & proximity to the player's focus (screen & distance between eyes and ability to discern pixels), can result in a final image that is of a significantly higher quality. Pic related is primary example, I'll upload a secondary & its comparison shortly after.
Elijah Foster
Source image.
Ryder Rogers
Comparison image.
This 2nd picture went through the following process:
>Waifu2x @ 1.5x magnify & denoise
>2-D illust (Y Model)
>png
>output depth 8 bits
>split size: 512
Each spat out image was ran through HoneyView3's sharpening/blurring filters and saves, followed by feeding it back into Waifu2x for another one to three passes of 1.25 to 1.5x increases with each subsequent passes going through more sharpening/blurring of boundaries to assit the DNN algorithm running in CUDA.
Final image was then fed BACK into Waifu2x and down-sampled to >4MB size.
Everything I did above, is essentially what DLSS is doing at the driver level. Its a whole lot of upscaling, reshading, blurring & sharpening with multiple passes to get a final image that at distance looks superior to the original.
Also, when you view this image at 100% and compare it to the original, you can see a decent loss in quality. But when you view it as a thumbnail (which inherently is larger than the original picture resolution), you get a "final" image of SIGNIFICANTLY higher quality.
Luke Clark
Where and when can I use it?
Tyler Kelly
>1+ years old articles
>nothing about Tensor Cores in them
The fuck is this bullshit thread?
Carson Roberts
-D illust (Y Model)
Can you explain what Y Model is? I just slap UpRGB on everything.
Another good effect that I noticed from using waifu2x is debanding. How does that work?
Noah Turner
You can already use these meme networks with madvr or mpv.
Grayson Lee
>smudge2x
It's alright but pretty overrated
Samuel Harris
Wait a minute wait a minute
hold on
Does this mean the old and widely mocked "enhance the image" trope will eventually work in real life?
Nicholas Clark
yes, use mpv
github.com
Brayden Clark
No, not really. It works best with drawings relying on clearly defined line art (chink cartoons).
Colton Perry
Only if you have supercomputer as one frame takes several seconds or even minutes to render.
Nolan Peterson
You drank the novideo kool aid hard dude. you can train smaller networks that work just fine for real time use RIGHT NOW.
Dominic Robinson
I have no idea what you just said.
Hunter James
That's on CPU, on GPU it takes several milliseconds.
Aiden White
There are many upscaling convolutional neural network models, some of which can already run in real time on cheap consumer graphics processing units. Despite the lies Papa Jensen fed you, you do not need a supercomputer for this shit.
Christopher Gonzalez
Read the damn article you linked. FSRCNN (which is a fast, optimized implementation of SRCNN that the article discussed) is already mature and is used in NGU / FSRCNNX in madVR/mpv for real-time frame upscaling in video. Whatever novideo is cooking up might eventually be "better" but it's definitely too slow for real-time.
Gabriel Morales
Nah, with Tensor Cores it should be fast, but it will be nvidia only.
Grayson Cruz
These are actually plenty fast on current GPUs without tensor cores.
I really think the tensor cores shit is a mistake.
Noah Foster
>Nah, with Tensor Cores it should be fast, but it will be nvidia only.
The article literally said the problem with SRRN's is that its slow as shit, but whatever you say man.
Mason Gomez
Care to give a link to some actually working ones for madvr? I tried a literal implementation of waifu2x in avisynth once and it was unusable. Also I don't know why the fuck are you acting like I'm op.
Jeremiah Rodriguez
The article is 1 year old and not about nvidia or tensor cores, OP is a fag.
Jayden Jenkins
install mpv
Brody Thompson
madVR uses one under the hood. All variants of NGU are basically FSRCNN.
Lincoln Campbell
Oh well, it's not even close to waifu2x quality. I don't know what this thread is about anymore.
Jack Rivera
You tend to notice these things less when you're not pixel peeping between static images and actually watching something in motion. If it's going to be used in gaymes it has to be fast. And even then the ones around are already quite good compared to most linear scalers.
Jacob Evans
>not even close to waifu2x quality
It's the same shit only smaller, waifu2x is a huge network.
Anthony Gonzalez
user pls. Tensor cores specialize exclusively in Convoluted Networks. That's their entire purpose. Nvidia's been experimenting with Tensors in their Quadros for years now. But they were too expensive to scale into the consumer market due to needing several years of research in designing and writing software that could leverage them at 16ms or less requirements.
Tensor = DNN. Prior to this HARDWARE, all DNN was done IN CUDA using traditional shaders. A shader doing DNN and a tensor doing DNN is night and fucking day levels of performance. The former is trash and the latter is practically magic. Just because the article doesn't mention Tensors in no way means its bullshit. Way to fail to see the forest for the trees.
>Explain Y Model
I honestly don't know the details here, but my best guess is that he various types of rasters and vectors and fed them into his DNN model, and optimized that model until he got the best possible outcome he could get at the time, with the hardware he had; and the end result is the models we see. That said, I believe Y model refers to a vector/raster combination model and I've found it to generate the best possible results amongst all available options. Take this explanation with a grain of salt, googling the topic further might help or just parse through the source code, hope there's comments and try to understand what he wrote.
Ayden Bailey
The whole point of the article was to give an explanation on what DLSS is you massive faggot, not what its running on or supposed to run on. Jesus christ.
Blake Gutierrez
>random bullshit again
I'm telling you that hhese articles have nothing to do with nvidia developing a super resolution. Where did you get that info?
Bentley Reed
>night and day
>fucking trash
Nvidia's own marketing charts gives your tensor cores a 3x (at best) gain from the previous generation. I don't think you know how slow any of these DNN shits are - even with a 3x speedup they're nowhere close to doing anything fast. They trained SRRN for 7 days on multiple Titan Xs for that miniscule gain over a network that's already usable in real time (SRCNN). Cut that into 2.5 days and that's still shit.
John Ortiz
Dude do you ever watch anime? It's static pictures 80% time. Waifu2x provides nearly perfect upscale from 720p while ngu is blurry. it might be better than other upscale algorithms, but so is waifu2x compared to this.
Here is ngu.
Jack Powell
And here is waifu2x. The difference is like between youtube video and bd rip.
Lucas Cooper
source pic?
Luke Peterson
you absolute moron
Asher Rogers
it's only slightly sharper. You need a 10 times bigger network for that? Just turn on a sharpener in madvr. Pathetic.
Robert Taylor
where is the source pic, retard
Ayden Sanders
And so is ngu compared to more lightweight algorithms.
Jeremiah Carter
no, it also removes artifacts like compression, ringing, aliasing.
Justin Walker
You really picked a wrong subject to shill your nvidia garbage
Eli Gonzalez
>SVP
SVP uses motion vectors. A better solution is to use a convolutional neural net, eg.
web.cecs.pdx.edu