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. |
21st November 2016, 08:08 | #40482 | Link |
Registered User
Join Date: May 2012
Posts: 447
|
I think I missed some important posts before, going to try turning off the deringing pre-processing next time and switch to a different option for image upscaling to minimize the overhead.
I wish there was a way to use a different regular upscaler for just the chroma doubling part though - if you have to double and then upscale more (e.g. 720p->1440p->2160p) I don't know if CatmullAR is the best option for the last part. I don't know how you'd work it into the UI though - the list for image doubling options is already pretty long. Maybe you could add at least one 'good' lower quality upscaler to the list of chroma doubling options? By the way, I always feel weird neglecting Chroma like this, but I guess I really can't tell the difference anyway :P Oh, and I hope that question about whether SSIM after doubling makes sense or if bicubic150 is enough at that point gets answered - pretty easy to add a profile that just switches to bicubic150 if (targetWidth > srcWidth) || (targetHeight > srcHeight).
__________________
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; 21st November 2016 at 08:13. |
21st November 2016, 08:32 | #40483 | Link |
Registered User
Join Date: Oct 2016
Posts: 31
|
good morning everyone ... with the new release it all works fine on win 10 madshi thanks ..
boys whom I recommended settings with the new NUG algorithms for upscaling from 1080-> 4K I have a video card 1070 GTX. I want canditare Madshi the oscar award |
21st November 2016, 08:58 | #40484 | Link |
Registered User
Join Date: Jun 2005
Posts: 503
|
For me NGU is not just sharp, it is also clear in the edge of faces in video, even in motion.
SuperRes, IMHO, tend to add some kind of attachments to the edge of faces, which can be really painfully annoying. It's quite hard to sescribe the effect, maybe the algorithm was puzzled distinguishing the face edge and the background. I would've ask madVR to keep SuperRes if it didn't suffer that kind of artefacts. |
21st November 2016, 10:26 | #40485 | Link |
Registered User
Join Date: Jul 2013
Posts: 27
|
I have an EVGA GTX 1070 on my HTPC, outputting to a 4k LG OLED.
I've replaced NNEDI64 (+SR2) for luma doubling of 720p/1080p to 2160p with NGU High and the performance is slightly better. There's a little bit of coil whine but it's inaudible with any audio on. In case of 720p material what do you guys prefer, doubling 720p to 1440p with NGU High and using Jinc AR to scale to 2160p or quadrupling to 2880p with NGU High+Med and downscaling to 2160p using SSIM1D with LL and AR? Currently I'm using sxbr 125 for all chroma upscaling. Does anyone else have another strong preference? Last edited by Gagorian; 21st November 2016 at 10:29. |
21st November 2016, 11:15 | #40487 | Link | |
Registered User
Join Date: Jan 2014
Posts: 216
|
Quote:
I really don't see any difference between low-med-high when switching between them with normal live action video content. High might look ever so slightly sharper than low on things like text, but that could just be placebo effect. So, please tell us madshi so I don't lose my mind trying to spot the differences. |
|
21st November 2016, 11:58 | #40488 | Link |
Registered User
Join Date: Sep 2016
Posts: 176
|
I have done another test with my RX 480.
Image upscaling: Calmull-Rom Chroma doubling: off Luma doubling: NGU low 11 ms NGU med 19 ms NGU high 42 ms NNEDI3 16 12 ms NNEDI3 32 16 ms NNEDI3 64 22 ms NNEDI3 128 34 ms NNEDI3 256 59 ms Image upscaling to DXVA2 disables doubling and performs at 3-8 ms. Enabling “run custom pixel shades in video levels instead of PC levels” saves 1-2 ms. MPC-BE saves 1-2 ms compared to MPC-HC. (latest versions) The strange thing is that I see just minor differences between DXVA2 and NGU low (and other scalers). The smoothness of DXVA2 (8 ms) makes the image better. Perhaps it is the noise in the movie that kills the improvements of the scalers. Or perhaps this is caused by the e-shift of the JVC X5000. Dunno. BTW. The fan goes to high when using NGU low-high. Last edited by HillieSan; 21st November 2016 at 12:33. |
21st November 2016, 13:25 | #40489 | Link | |
Is this for real?
Join Date: Mar 2016
Location: Norway
Posts: 168
|
Quote:
So I will keep using NNEDI3 for now @ madshi
__________________
My HTPC : i9 10900K | nVidia RTX 4070 Super | TV : Samsung 75Q9FN QLED Last edited by Betroz; 21st November 2016 at 13:27. |
|
21st November 2016, 13:43 | #40490 | Link | |
Registered User
Join Date: Dec 2011
Posts: 1,812
|
Quote:
And still no one has provided images why super-xbr 125 would be the king for chroma. Did you try NGU low? |
|
21st November 2016, 14:37 | #40491 | Link | |
Registered User
Join Date: Feb 2014
Posts: 94
|
Quote:
I did upload a comparison showing the differences a few posts ago. This is the link: http://imgur.com/a/JSJBn I compared: SXBR 100 AR, Recon Soft, NNEDI 256 and NGU High. From them all NGU is the worst to my eyes. NNEDI 256 is the best to my eyes but the performance cost is not acceptable for chroma, plus in motion is harder to see differences in chroma. Where to look: http://imgur.com/SwKvVMg Last edited by fedpul; 21st November 2016 at 14:45. Reason: Where to look added! |
|
21st November 2016, 15:25 | #40492 | Link |
Registered User
Join Date: Dec 2011
Posts: 1,812
|
Thank you, I somehow missed that.
I suspect that we see the same problem here we see with really bad sources regarding NGU and luma doubling, the excessive ringing leads to quite unpleasant results. In my American Dad cartoon example with not such excessive ringing, NGU clearly leads to better reconstructed chroma by even less ringing than s-XBR. Can you try also NGU low? It's a bit softer than high. |
21st November 2016, 15:39 | #40495 | Link | |
Registered User
Join Date: Feb 2014
Posts: 94
|
Quote:
|
|
21st November 2016, 15:59 | #40496 | Link |
Registered User
Join Date: Dec 2011
Posts: 1,812
|
Can be a bit tricky to jump to the exact frames in mediaplayer, even when you tell MPC HC/BE to jump exactly to frame x. You may have to seek a bit frame by frame with ctrl + arrow keys.
A more "secure" way would probably be AviSynth with the Trim function. |
21st November 2016, 16:51 | #40497 | Link | |
Registered User
Join Date: Jul 2013
Posts: 27
|
Quote:
|
|
21st November 2016, 16:53 | #40498 | Link | |||
Registered Developer
Join Date: Sep 2006
Posts: 9,140
|
Quote:
Quote:
No. NGU is *significantly* sharper than NNEDI3 64/128. Quote:
That's your choice, of course. But I think your comparison was probably flawed. Did you use SuperRes in your comparison? If so, did you activate it for both NGU and NNEDI3? Do *not* do that. I've said it many times now: NGU does not benefit from SuperRes, results can actually get somewhat worse when switching SuperRes on with NGU. So use NGU alone, without SuperRes. On the other hand, if you want to compare NNEDI3 to NGU, you almost have no other choice than to use combine NNEDI3+SuperRes, because otherwise the difference in sharpness is simply too big to do any useful comparisons. So IMHO the proper comparison would be "NGU without SuperRes" and "NNEDI3 with SuperRes". I don't know your settings, so I can't say for sure. I see nothing wrong with the screenshot, though. |
|||
21st November 2016, 16:55 | #40499 | Link | |
Registered Developer
Join Date: Sep 2006
Posts: 9,140
|
RX480 users
I've received an email from a non-doom9 madVR user. He had big trouble with NGU performance with his AMD GPU, but found a way to fix it. Maybe it helps the RX480 users? Quote:
|
|
21st November 2016, 17:21 | #40500 | Link |
Registered User
Join Date: Jan 2014
Posts: 21
|
Something of my test.
GTX 1060 6GB (driver 373.06) Monitor 1920x1080 60Hz win7 x64 MPC-BE x64 The power state of videocard is troublesome, I have to set mpc-be64.exe to maximum performance on NV control panel, otherwise I would get variable rendering time(it would become very high) and many dropped frames. madVR settings: zoom control off debanding on (low/high) deringing off smooth motion on dithering - ED1 D3D11 FSE mode chroma upscaling - Bilateral soft image downscaling - SSIM 1D+AR+LL image upscaling - Jinc+AR image doubling - check double luma only except super-xbr 720x480 > 1920x1080 super-xbr 100 6ms nnedi3 16 9.5ms nnedi3 32 10ms nnedi3 64 11ms nnedi3 64+SR4 14.5ms nnedi3 128 13ms NGU low 9ms NGU low 6ms(image upscaling - Catmull-Rom+AR) NGU med 9.5ms NGU high 11ms 640x480 > 1920x1080 super-xbr 100 6ms nnedi3 16 8ms nnedi3 32 9ms nnedi3 64 10.5ms nnedi3 64+SR4 12.5ms nnedi3 128 14ms NGU low 7ms NGU low 5.5ms(image upscaling - Catmull-Rom+AR) NGU med 8ms NGU high 11ms 1280x720 > 2560x1440 > 1920x1080 super-xbr 100 8ms nnedi3 16 11.5ms nnedi3 32 13ms nnedi3 64 17ms nnedi3 64+SR4 24ms nnedi3 128 24.5ms NGU low 11ms NGU low 9ms(image upscaling - Catmull-Rom+AR) NGU med 13ms NGU high 20ms 1920x1080 > 3840x2160 > 1920x1080 (image downscaling - bilinear) super-xbr 100 10.5ms nnedi3 16 19ms nnedi3 32 23ms nnedi3 64 31ms nnedi3 64+SR4 52ms nnedi3 128 50ms NGU low 18ms NGU low 12ms(image upscaling - Catmull-Rom+AR) NGU med 22ms NGU high 39ms Acrodding to my test, for performance, NGU-high is between NNEDI3-64 and NNEDI3-128, NGU-med is as much as NNEDI3-32, NGU-low is less than NNEDI3-16 but more than super-xbr 100. In my opinion, NGU is amazing and useful, I think it is good for real-life content that I can raplace NNEDI3 and super-xbr, it brings the sharpness similar to NNEDI3+SR4 and less artifacts, and faster. But for anime content I still prefer NNEDI3(no SR) over NGU especially for low-res and high compression content. For anime, NNEDI3 did a good job of reducing aliasing, NGU is a little too sharp that aliasing and artifacts become more visible, NNEDI3 is more clean even though NGU might closer to original picture.(NGU-high has less artifacts than NGU-low, but still more than NNEDI3) For anime, personally, I like NNEDI3 + thin edge1.0 + adaptive sharpen1.5(AB100%+AR). A set of example: originalhttp://i.imgur.com/D2y3ULU.png NNEDI3 64http://i.imgur.com/obCswUn.jpg NGU highhttp://i.imgur.com/GXZB2M7.jpg NNEDI3 64+TE1.0+AS1.5 http://i.imgur.com/bRXAoKr.jpg Last edited by Neet009; 21st November 2016 at 18:04. |
Tags |
direct compute, dithering, error diffusion, madvr, ngu, nnedi3, quality, renderer, scaling, uhd upscaling, upsampling |
Thread Tools | Search this Thread |
Display Modes | |
|
|