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. |
22nd November 2016, 10:51 | #40561 | Link | |
Registered User
Join Date: Jan 2016
Posts: 52
|
Quote:
I'm not sure about weeding them out, but maybe condensing it in one category. Or to simply grey out options that do not work/create too many artifacts with NGU(or any other algorithm) but not remove them for other algos. |
|
22nd November 2016, 10:55 | #40562 | Link |
Kid for Today
Join Date: Aug 2004
Posts: 3,477
|
Indeed, I'm also unsure about the downscaler's AR and I think it harms PQ when SR2 LL's AR is engaged, need to run moar tests I guess.
There's also the story of whether feeding C-R with a sharper picture or Bicubic 100+ / SSIM 1D with a softer one, more tests to be conducted again. Last edited by leeperry; 22nd November 2016 at 15:26. |
22nd November 2016, 10:58 | #40563 | Link |
Registered User
Join Date: Jun 2016
Posts: 39
|
As a non-expert user it took me quite awhile to finally be able to use the new algorithm properly and I want to apologize for asking the same question that so many people have asked before, should've read before commenting.
My experience so far is that the new algorithm is freaking awesome. I tried to push my GPU to the very limit and oh boy is it sharp. I use NGU-med for luma doubling to upscale a 720p anime to 1080p. When the animation moves too fast the sharpness is not as obvious but I think that is due to the artifacts and details lost during the downscaling process, but when the animation is still it is so sharp as if it is a 1080p video in the first place. I also tried to use NGU-med for chroma upscaling and I think it looks quite a lot sharper than before when I used NNEDI32+SR1. I even turned off adaptive sharpen and I really don't see the image gets any less sharp than having it on. Thank you mashi for this incredible new algorithm! |
22nd November 2016, 11:48 | #40565 | Link |
Registered User
Join Date: Dec 2011
Posts: 1,812
|
Why would it be too sharp? I think if we did a ground truth comparison for chroma, it should come closer to it than any other non-reconstruction based scaler.
But I agree that there may be other crucial factors in reality with chroma, aka distracting ringing in sources. But the same goes for luma, I personally won't blur everything just to mask the artifacts. |
22nd November 2016, 12:42 | #40568 | Link |
Registered User
Join Date: Sep 2014
Posts: 280
|
Some more testing:
I exchanged my RX480 with my GTX1060 from my gaming pc. With same settings and NGU high I can run 1080p to UHD with 34ms (GTX). With RX480 i got 48ms. So I will stay with the GTX1060 until NGU got a better performance on my RX480.
__________________
Intel i5 6600, 16 GB DDR4, AMD Vega RX56 8 GB, Windows 10 x64, Kodi DS Player 17.6, MadVR (x64), LAV Filters (x64), XySubfilter .746 (x64) LG 4K OLED (65C8D), Denon X-4200 AVR, Dali Zensor 5.1 Set |
22nd November 2016, 12:53 | #40569 | Link |
Registered User
Join Date: May 2012
Posts: 447
|
By the way madshi, do you think you could add a way to toggle rather than just activate a specific profile, or alternatively add a hotkey to switch back to automatic detection? Maybe just turn 'activate' into 'toggle' so it switches the profile off if it's already been manually activated. I've set up a bunch of profiles for my laptop that I activate when I have to unplug it (so watching video doesn't drain power like crazy), but right now the only way to switch back is to reopen the video. Alternatively if you could add detection for when the device has AC power, that would be great too
__________________
Test patterns: Grayscale yuv444p16le perceptually spaced gradient v2.1 (8-bit version), Multicolor yuv444p16le perceptually spaced gradient v2.1 (8-bit version) Last edited by Ver Greeneyes; 22nd November 2016 at 13:37. |
22nd November 2016, 14:25 | #40571 | Link |
Registered User
Join Date: Oct 2015
Posts: 99
|
Ah, I missed the earlier post(s) concerning AR confusion before posting my own.
Well, I'm sort of relieved I'm not the only one who's a bit confused (again). I'd take the "NGU for everything" route just to keep things simple (and due to serious lack of testing time) but different options may of course interact in so many ways. |
22nd November 2016, 15:40 | #40573 | Link | |
Registered User
Join Date: Sep 2016
Posts: 176
|
Quote:
Yesterday I discovered something that is weird. I played a movie that showed 24 fps (source) and after changing some settings in madvr it changed to 59 fps. Why is this? Last edited by HillieSan; 22nd November 2016 at 15:43. |
|
22nd November 2016, 15:49 | #40574 | Link | |
Registered User
Join Date: Jul 2013
Posts: 27
|
Quote:
Increasing queue size will not help in your case, it might help if you rarely get a few dropped frames. |
|
22nd November 2016, 16:02 | #40575 | Link | |
Registered User
Join Date: Sep 2016
Posts: 176
|
Quote:
I use 4:2:0 10 bits (or 12 bits) Last edited by HillieSan; 22nd November 2016 at 16:43. |
|
22nd November 2016, 16:39 | #40576 | Link |
Kid for Today
Join Date: Aug 2004
Posts: 3,477
|
Turns out SR does add bloat(and SR's AB was removed for some reason I didn't understand IIRC) but SSIM 2D 100% LL with 50% AB tackles the issue very nicely, all roads lead to Rome yada yada ^^
I rest my case that SR LL's AR doesn't require downscaler's AR as the latter only worsens PQ, I ran all those tests on lowres SD(who knew 360x288@1080p could look so good!). I think 720p30 looks indeed better with vanilla NGU w/o SR, anyway my HD7850 doesn't quite have the horsepower to run it all together with SSIM on 720p so case closed. I'm reluctant to upgrade coz 280/290 boards provides poor efficiency, suck up a lot of power and 480's need a 64bit OS that I don't have and seem to provide poor NGU performance, anyway PQ is fantastic so the upgraditis monster can go back to where it came from Last edited by leeperry; 22nd November 2016 at 16:52. |
22nd November 2016, 16:41 | #40577 | Link | |
Registered User
Join Date: Sep 2016
Posts: 176
|
Quote:
UPDATE: Of course, this keeps the display to 1080p and no image doubling is performed. My mistake. UPDATE: Changing the frequency of the display from 60p to 24p doesn't help. Last edited by HillieSan; 22nd November 2016 at 17:04. |
|
22nd November 2016, 16:42 | #40578 | Link |
Registered User
Join Date: May 2012
Posts: 447
|
Smooth motion adds a lot of overhead for UltraHD output, thankfully you can detect it using a (smoothMotion) profile rule (it doesn't work without the parentheses) so I've set up a separate profile for it that doesn't use expensive scalers.
__________________
Test patterns: Grayscale yuv444p16le perceptually spaced gradient v2.1 (8-bit version), Multicolor yuv444p16le perceptually spaced gradient v2.1 (8-bit version) |
22nd November 2016, 16:50 | #40579 | Link |
Registered User
Join Date: Sep 2012
Posts: 77
|
RX 480 on Win 7 x64 with 16.10.3, 720x480 movie doubling (luma only), SXBR100 for chroma, No de-ringing and no SR.
NNEDI3-64: ~6.48ms ~8.1% GPU load NGU High: ~8.45ms ~24.1% GPU load Same GPU\Memory clocks in both cases. So, yeah, NGU is quite hungrier than NNEDI3. madshi, I'm interested so see how it performs on D3D11 shaders. |
22nd November 2016, 17:41 | #40580 | Link | |
Registered User
Join Date: Jul 2013
Posts: 27
|
Quote:
|
|
Tags |
direct compute, dithering, error diffusion, madvr, ngu, nnedi3, quality, renderer, scaling, uhd upscaling, upsampling |
Thread Tools | Search this Thread |
Display Modes | |
|
|