Welcome to Doom9's Forum, THE in-place to be for everyone interested in DVD conversion. Before you start posting please read the forum rules. By posting to this forum you agree to abide by the rules. |
20th November 2016, 04:09 | #40341 | Link | |
Registered User
Join Date: Jan 2012
Posts: 106
|
Quote:
Thanks for the idea, anyhow. Last edited by strumf666; 20th November 2016 at 04:12. |
|
20th November 2016, 07:53 | #40342 | Link | |
Registered User
Join Date: Nov 2011
Posts: 187
|
Quote:
Regarding NGU: NGU Image Doubling looks very impressive. It’s commendable that you have worked on it entirely. As others pointed out, there is a dirty look with the “add grain” option. I believe the grains should be dynamic (changing with every frame). I haven’t had any performance issues or keyboard shortcuts not working problem with the 0.91.0 build on my system with AMD HD8970M card. However, I do notice the loud coil whine with NGU. Regarding the “feedback”, I think NNEDI3 should stay. I don’t care about SuperRes doubling and NEDI when NGU and NNEDI3 are available. Last edited by omarank; 20th November 2016 at 08:07. |
|
20th November 2016, 08:42 | #40343 | Link | |
I'm Siri
Join Date: Oct 2012
Location: void
Posts: 2,633
|
Quote:
|
|
20th November 2016, 09:00 | #40344 | Link |
Kid for Today
Join Date: Aug 2004
Posts: 3,477
|
And it's not my imagination as I double-checked, skipping with NGU in PotP is just as fast as with Jinc....NNEDI3 & sxbr are a lot more sluggish, A/B'ing them against NGU is also laughable, please don't ditch any SR coz I really like them both in combination with NGU and "soften edges"@1. I don't see much point in the new "add grain" static noise option and the older "enhance detail" anymore though.
And so far I still prefer RS soft for chroma with SR@2. Last edited by leeperry; 20th November 2016 at 09:12. |
20th November 2016, 09:01 | #40345 | Link | ||
Registered User
Join Date: Feb 2002
Location: San Jose, California
Posts: 4,406
|
Quote:
Quote:
Maybe a very advanced version could look OK, but unassisted and in real time on video playback is probably a ways off.
__________________
madVR options explained |
||
20th November 2016, 09:20 | #40346 | Link |
Visual Novel Dev.
Join Date: Nov 2015
Location: Bucharest
Posts: 200
|
Tested this a little more in depth, and it's true.
It seems that NGU does produce considerably more coil whine. Most other madVR settings produce no coil whine at all, so it makes me curious why NGU does this..
__________________
Visual Novels! https://www.facebook.com/seventh.heart.studios/ https://audiophile-heaven.blogspot.com/ https://www.youtube.com/watch?v=27VWU3ydRY8 |
20th November 2016, 10:13 | #40348 | Link | ||||||||
Registered Developer
Join Date: Sep 2006
Posts: 9,140
|
Are you sure that NGU is activated, then? If the OSD doesn't work, probably NGU isn't activated, either (!). You're probably getting super-xbr doubling. It's a known problem, and I'm working on it.
Quote:
Quote:
Quote:
Quote:
Quote:
Quote:
Quote:
Quote:
Anyway. Google's RAISR was mentioned in my thread, so I posted my subjective impression. FYI, Shiandow posted a similar impression in the MPDN thread. It's hard to proof anything because as far as I can see, there's no way for me to actually test RAISR. All I can look at is the PDF. They haven't even made the images they used in the PDF available as BMP/PNG files, so we also can't run them through NGU to compare (except those that are well known like the butterfly, see below). However, I think I have some "weak proof" that I can post: If you look at page 21 of the Google RAISR PDF, you can clearly see that RAISR has noticeably worse PSNR & SSIM results than SRCNN. And IMHO anything that can't beat SRCNN is kind of underwhelming for a new algorithm. SRCNN was state of the art when it was released but that was a couple of years ago. Have a look at these two images, with the butterfly also used in the RAISR PDF: butterfly SRCNN (PSNR: 32.181) butterfly NGU (PSNR: 33.689) Please zoom into these images with your favorite image viewer to see how much better NGU is than SRCNN. And Google RAISR is worse than SRCNN, according to their own PDF!! Of course it's possible that RAISR could be dramatically faster than NGU, I can't say for sure. But in terms of quality, NGU should be noticeably better than RAISR. (I should add that I used a custom NGU version which was optimized for best PSNR/SSIM results. I believe all the scientific papers do that. The NGU version used in madVR is tweaked for being as artifact free as possible, which lowers PSNR/SSIM scores but produces much cleaner results. But the NGU version in madVR still beats SRCNN PSNR/SSIM numbers by a noticeable margin.) Just to be safe: Have you checked the madVR OSD to see if NGU is even being activated in your case? It's a known problem with the current madVR version that on some PCs NGU doesn't activate. In that case super-xbr is used instead. You can verify that by activating NGU for image doubling (not for chroma upscaling!!) and then checking the OSD if it's listed there. Since you don't seem to like what you've seen so far, I can only guess that you're one of those for whom NGU doesn't even activate atm? It's not possible atm, I think. I'm not even sure which playback component would be the best to offer such a setting (e.g. subtitle renderer, or subtitle splitter, or madVR, or media player, or movie file splitter etc). |
||||||||
20th November 2016, 10:17 | #40349 | Link | |
Kid for Today
Join Date: Aug 2004
Posts: 3,477
|
Quote:
Another proof is that "edges soften" and "add noise" also work, PQ is quite amazing on 720p@1080p in combination with RS chroma |
|
20th November 2016, 10:20 | #40350 | Link | ||
Registered Developer
Join Date: Sep 2006
Posts: 9,140
|
Quote:
Quote:
Just my 2 cents, of course. |
||
20th November 2016, 10:21 | #40351 | Link | |
Registered Developer
Join Date: Sep 2006
Posts: 9,140
|
Quote:
So what's your opinion on which algos need to stay and which can go (NNEDI3, SuperRes, super-xbr, NEDI)? |
|
20th November 2016, 10:21 | #40352 | Link | |
Registered User
Join Date: Apr 2002
Location: Germany
Posts: 4,926
|
Quote:
The other methods wont produce that low Render Frametimes at all continuously and not stress the Hardware as much even if they utilize it fully. You will hear more GPU ALU Stress these days very fast because everything is Physically on the Edge and under additional Economic Value constraints. Sometimes you can use that even better to measure your code efficiency just listen to what you System does can be more efficient then any Software Performance analysis Because the reaction time through the RTOS is super fast these days that does the Power Management and Frequency Switching bellow the OS.
__________________
all my compares are riddles so please try to decipher them yourselves :) It is about Time Join the Revolution NOW before it is to Late ! http://forum.doom9.org/showthread.php?t=168004 Last edited by CruNcher; 20th November 2016 at 10:31. |
|
20th November 2016, 10:33 | #40353 | Link | |
Kid for Today
Join Date: Aug 2004
Posts: 3,477
|
Quote:
For GPU-hogging 1440x1080p60 footage I will prolly need to fall back to sxbr, whatever you decide to ditch please don't remove any SR coz I really like what they both do on motion-blur Maybe a "soften edges" sub-option would be good for chroma NGU? RS chroma is tough competition you know. |
|
20th November 2016, 10:38 | #40354 | Link |
Registered Developer
Join Date: Sep 2006
Posts: 9,140
|
Are you using NGU + SR? Please try NGU without SR, that actually looks better IMHO. SR kind of reintroduces a little bit of bloating into the image, IMHO.
With RS you mean "recon-soft"? Yeah, I like that one, too. It's my custom tweaked version of feisty2's original VapourSynth chroma reconstruction script. I do hope to replace it with a luma guided NGU version at some point in the future, though. |
20th November 2016, 10:47 | #40355 | Link |
Registered User
Join Date: Jun 2015
Posts: 22
|
I can confirm pretty nasty coil whine with NGU on a Gigabyte G1 1070 + HX750i PSU
Maybe it has something to do with the memory controller being used so much more? In my case, for 720 to 1440 doubling it has 40% load with NGU med chroma and high luma compared to 10% to low teens for NNEDI3 64 luma 16 chroma
__________________
Windows 10 // i7-8700k + GTX 1080 SLI // MPC-BE x64 + LAV + XySubFilter Last edited by zvans18; 20th November 2016 at 10:49. |
20th November 2016, 10:59 | #40357 | Link |
Registered User
Join Date: Jun 2015
Posts: 22
|
Is there any way to lower the memory impact, at least enough for experimentation? I agree it shouldn't be a factor, but despite how good NGU looks, I can't justify using it daily or even extensively testing for that matter because the whine is the loudest part of my computer and it's incredibly bothersome. And this 10 series and even the last 9 series are more susceptible than previous series, and with these newer ones being some of the only ones that can run it comfortably, we're kind of in a catch 22.
__________________
Windows 10 // i7-8700k + GTX 1080 SLI // MPC-BE x64 + LAV + XySubFilter Last edited by zvans18; 20th November 2016 at 13:07. |
20th November 2016, 11:02 | #40358 | Link |
Registered User
Join Date: Apr 2002
Location: Germany
Posts: 4,926
|
Welcome to the rise of the Consoles
First of all you could try to reduce the Pre OC stress from your card Gigabyte should have a application you can also define these modes easily with AIBs call these modes mostly Silent,Game and OC Switching to Silent Mode should lower the stress @madshi that's not really your job @all also because the characteristics are so different you couldn't ensure it anyways to be perfect for everyone with a simple option to chose from
__________________
all my compares are riddles so please try to decipher them yourselves :) It is about Time Join the Revolution NOW before it is to Late ! http://forum.doom9.org/showthread.php?t=168004 Last edited by CruNcher; 20th November 2016 at 11:19. |
20th November 2016, 11:18 | #40359 | Link |
Registered Developer
Join Date: Sep 2006
Posts: 9,140
|
The only way to lower the memory impact would be to use DirectCompute or OpenCL, and NGU doesn't really need that because NGU is mostly arithmetic limited and not memory limited, I think. Going DirectCompute or OpenCL would make everything much more complicated, once more, and open the door to all kinds of driver issues etc. And since NGU doesn't seem to be memory limited, it probably wouldn't actually improve performance in any noticeable way, either.
There are custom GPUs from different manufacturers. Aren't there any without coil whine? Try lowering the GPU clocks or voltages by using some tweak tool, maybe that helps? |
Tags |
direct compute, dithering, error diffusion, madvr, ngu, nnedi3, quality, renderer, scaling, uhd upscaling, upsampling |
Thread Tools | Search this Thread |
Display Modes | |
|
|