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. |
31st May 2009, 11:43 | #1262 | Link | |
Registered User
Join Date: Jun 2006
Posts: 133
|
Quote:
Anyway, you are right there is no point outputing YUY2 with Dscaler when the input is YV12 material since we might lose CPU time and quality for nothing. |
|
31st May 2009, 12:40 | #1263 | Link |
Registered User
Join Date: Jul 2005
Posts: 103
|
I saw the comparison screens in the first post and they looked quite impressive, but I certainly couldn't reproduce those. See this screenshot:
The first is Nvidia VMR9, second Haali, third MadVR. I wasn't able to spot a difference even when compating single pixels and definitely not when watching a video. So what's the benefit of different renderers? Or do I get something wrong? Shots were taken on a slightly upscaled video, VMR9 resizing set to Bicubic A=-1. Last edited by bur; 31st May 2009 at 12:43. |
31st May 2009, 14:36 | #1264 | Link | |
3 eyed CRT supporter
Join Date: Jan 2008
Location: Or-strayl-ya
Posts: 563
|
Quote:
Sorry, my mistake, I just checked how I have it setup. I do have Dscaler outputting YV12. I just have ffdshow after it for stuff like deinterlacing when needed. |
|
31st May 2009, 16:59 | #1265 | Link | |
Registered User
Join Date: Jun 2005
Posts: 630
|
Quote:
You can have lots of blur both in Haali and mVR, depending on the settings. A=-1 iirc is the sharpest bicubic setting in VMR9. |
|
31st May 2009, 18:09 | #1266 | Link |
Registered User
Join Date: Jun 2005
Posts: 630
|
Let's make then a bit more methodologically sound comparison, shall we?
Same source, same frame, all shots done via print screen. DVD untouched source, famous touhou Presented in original resolution + upscale to apply correct A/R (i.e. around 853 width instead of 720, plus normal chroma upscale). VMR9 3D with A=-1: Haali with -0.8: mVR with SoftCubic50 on both luma and chroma: mVR with Lanc4 for luma and CMR for chroma: Of course, since I used ffdshow for all of the snapshots, in all cases but with mVR, initial (or whole) chroma upscale is performed by the decoder. Softcubic's version is the softest one. In terms of filesize in PNG, VMR9 << Haali << Softcubic50 << Lanc4+CMR Numbers don't lie )) Last edited by Egh; 31st May 2009 at 18:12. |
31st May 2009, 21:26 | #1267 | Link |
Registered User
Join Date: Jul 2005
Posts: 103
|
I used the default settings for Haali and MadVR (whatever those are, I just enabled them in MPC and there were no settings available - maybe that's what I did wrong?).
Other than that, the frames are still nearly 100% equal. I don't know what Egh means by "more methodologically sound", but of course I used the same frames for all three shots. Pictures were saved as BMP through MPC and later converted to PNG with Paint.Net. I have to admit though there are some differences to be seen in Egh's shots. But still nowhere as drastic as in the thread's first post where it looks like MadVR was far superior to any other renderer. And even though there are some differences if you compare the pictures posted by Egh on a per pixel basis, I'm not sure if you would notice them when watching the video. Especially if it's not drawn, but "real" life. I doubt you'd really notice anything much between those four settings/renderers, VMR9 and MadVR with lanczos4 look very much the same. It would make an interesting experiment, having someone else set up the video with different renderers and then watch them to compare quality. PS: Where can I change the MadVR settings? Last edited by bur; 31st May 2009 at 21:32. |
1st June 2009, 00:48 | #1268 | Link | |
Registered User
Join Date: Feb 2006
Posts: 293
|
Quote:
You can change madVR settings by go to Play->Filters->madVR (in MPC-HC)
__________________
Spec: Intel Core i5-3570K, 8g ram, Intel HD4000, Samsung U28D590 4k monitor+1080p Projector, Windows 10. Last edited by Hypernova; 1st June 2009 at 00:50. |
|
1st June 2009, 11:42 | #1269 | Link |
quack quack
Join Date: Apr 2009
Posts: 259
|
You can also change settings with a DirectShow filter manager.
|
1st June 2009, 16:42 | #1270 | Link | ||||||||||||||||
Registered Developer
Join Date: Sep 2006
Posts: 9,140
|
Quote:
Quote:
Quote:
Quote:
Quote:
Quote:
Quote:
Quote:
Quote:
Higher output resolutions are more demanding than lower res. Downscaling is more demanding than upscaling. Why don't you guys search this thread for "x64"? Quote:
Quote:
However, I'd not be sad at all if the .icm color profiles didn't work because the plan is to use cr3dlut for complete display calibration. Quote:
Careful. Are you talking about input or output bitdepth? cr3dlut uses 8bit input and 16bit output! I doubt CLUT can compete with that. Maybe eventually CLUT is 10bit input 10bit output. But I rather think it's probably 8bit input 10bit output. So worse in every way compared to cr3dlut. Also the next question would be: Does the CLUT round higher bitdepth input data down to the CLUT bitdepth? Or does it interpolate? madVR does trilinear interpolation between the 3D 8bit 3dlut. Interpolating an 8bit lut is probably better than rounding to a 10bit lut. Quote:
Quote:
madVR 0.10 is optimized to achieve smooth motion on capable graphics cards. It seems that the rendering approach used by madVR 0.10 does not play nice with slower/older graphics cards, unfortunately. I don't think perfect smooth playback will ever be possible on such slower/older GPUs *in windowed mode*. I expect, however, to achieve quite good (maybe perfect) results with fullscreen exclusive mode even on some older cards. Quote:
Quote:
(1) as mathematically accurate rendering as possible (2) being independent of stupid driver behavior and driver bugs (3) included display calibration Some other renderers may come close to madVR image quality in specific scenes, and then fail in other scenes. In every day life scenes the difference between madVR and other renderers can be very small. But there are specific scenes where other renderers sometimes stumble. The comparison screens on the first page show such specific scenes. If you want to reproduce such big differences on your PC, try to find scenes that are similar. E.g. look for scenes with lots of red on black background ("chroma upsampling"). Or look for scenes with smooth color gradients ("dithering"). Or look for Hypernova's posts in this thread. He's posted some nice real life comparison screenshots where madVR produces visibly better results. Some of these image quality differences are harder to see in motion while other differences are actually *easier* to see in motion. E.g. banding artifacts produced by not using proper dithering can be extremely annoying in motion. One very big annoyance factor I find with other renderers is that you never know exactly what you will get. Depending on OS, graphics card model, renderer, driver revision, connection type and even display you can get different results. E.g. I've found that sometimes simply asking PowerStrip to change refresh rate can result in the GPU switching between video <-> PC levels. Or some people seem to get good chroma upsampling quality from ATI cards, while other people don't and nobody knows why exactly. With madVR you don't have any such problems. You get 100% the same image quality on every graphics card and every OS because madVR simply doesn't leave the GPU any room for interpretation... |
||||||||||||||||
1st June 2009, 18:04 | #1273 | Link | |
Kid for Today
Join Date: Aug 2004
Posts: 3,477
|
Quote:
the only way to get a LUT calibration that works all the time is w/ a more complicated ICC 4.0 file using the <LUT> tags, or w/ ARGYLLCMS. ARGYLLCMS will play zillions of test patterns automatically(for 25 minutes in HQ mode), and in the end will make a 256 levels 1D LUT for R/G/B that will automatically make the graphic card match perfect calibration(D65/2.2 mostly, and you can actually check the calibration through Color.HCFR afterwards...it's stellar! ΔE is usually 1 on the whole IRE scale ) if someday cr3dlut enables to import the .cal files from ARGYLLCMS, this might indeed become an all-in-one solution 2)well yeah, true...I know the CLUT works in 10bit on CRT, so mostly it's 8bit input > 10bit CLUT > 10bit output over VGA(ARGYLLCMS can measure the CLUT accuracy). but DVI is 8bit anyway, so how do the 10bit get encoded to 8bit TMDS? no idea, but it's prolly ugly anyway, I only use the graphic card's CLUT over VGA on my CRT, on DVI I leave the CLUT untouched. Last edited by leeperry; 1st June 2009 at 18:13. |
|
1st June 2009, 18:24 | #1274 | Link | ||
Broadband Junkie
Join Date: Oct 2005
Posts: 1,859
|
Quote:
If you have another idea of why, when the decoder and/or render queues drop to only 15/16 and/or 7/8 respectively, I always get dropped/delayed frames, I'm open for suggestions about how it can be fixed. Edit: Also is there any chance for you to quickly make a stop-gap release (something like 0.10b) that only fixes the remembering of luma scaling settings? It's been two weeks now, and having to manually set my scaling settings for every single video I open, is getting very annoying, very fast. I hope we don't have to wait another month or something just to see this minor thing fixed. Last edited by cyberbeing; 1st June 2009 at 18:37. |
||
1st June 2009, 19:36 | #1276 | Link | |
Broadband Junkie
Join Date: Oct 2005
Posts: 1,859
|
Quote:
My preferred: Upscaling: Spline36 Downscaling: Spline64 Chroma: Mitchell-Netravali Last edited by cyberbeing; 1st June 2009 at 19:50. |
|
1st June 2009, 22:46 | #1277 | Link | ||||
Registered Developer
Join Date: Sep 2006
Posts: 9,140
|
Quote:
But is 3x 1D LUT good enough? I think R/G/B influence each other. Why else would cr3dlut use one big 3D LUT instead of 3x 1D LUT? Quote:
Quote:
If I decreased the queue to only 4 entries, you'd have your queue oscillating between 3/4 and 4/4. If I increased the queue to 16 entries, you'd have your queue oscillating between 15/16 and 16/16. Quote:
Do you have a real-world sample where the loss in chroma detail is visible? Having such a sample here on my PC might help madVR development! |
||||
1st June 2009, 23:18 | #1278 | Link | |
Kid for Today
Join Date: Aug 2004
Posts: 3,477
|
Quote:
2)3x1D LUT is perfectly fine for calibration...we just want the curves to meet properly to reach perfect D65/2.2 levels...as I understand it a 3D LUT is only necessary for gamut transfer functions(which you cannot do w/ 1D LUT's) : http://www.color.org/ICC_Chiba_07-06..._DMP_Float.pdf this PDF is simple, yet very instructive the graphic card's CLUT does not support 3D LUT's, only 3x1D LUT's...they always make sure to do it as sloppy as possible as you know a .cal file looks like this : Code:
KEYWORD "RGB_I" NUMBER_OF_FIELDS 4 BEGIN_DATA_FORMAT RGB_I RGB_R RGB_G RGB_B END_DATA_FORMAT NUMBER_OF_SETS 256 BEGIN_DATA 0.0000 0.083189 0.021229 0.023934 3.9216e-003 0.087305 0.047484 0.028051 7.8431e-003 0.091516 0.061404 0.032320 0.011765 0.095825 0.071292 0.036751 0.015686 0.10023 0.079941 0.041354 0.019608 0.10475 0.088256 0.046137 0.023529 0.10937 0.096318 0.051114 0.027451 0.11409 0.10400 0.056294 0.031373 0.11882 0.11121 0.061668 0.035294 0.12350 0.11782 0.067159 0.039216 0.12807 0.12387 0.072689 0.043137 0.13247 0.12948 0.078181 0.047059 0.13671 0.13473 0.083563 0.050980 0.14080 0.13966 0.088789 0.054902 0.14475 0.14437 0.093852 0.058824 0.14860 0.14889 0.098760 ~ I'm sure he could even output 3dlut files if we asked nicely that's my stock CRT : that's after ARGYLLCMS in HQ mode for 25 mins(in a pitch black room w/ the i1d2 stuck to the CRT) : Last edited by leeperry; 1st June 2009 at 23:58. |
|
2nd June 2009, 00:02 | #1279 | Link | |
Registered Developer
Join Date: Sep 2006
Posts: 9,140
|
Quote:
Sounds like a good idea to me, but that's really up for yesgrey3 to decide. |
|
2nd June 2009, 00:46 | #1280 | Link | ||
Broadband Junkie
Join Date: Oct 2005
Posts: 1,859
|
Quote:
I was just assuming it was something like Haali Renderer's queue, where if you set it too high and you're decoding difficult material, the jitter will go through the roof and stay there. Reducing the Haali Renderer frame queue to something more reasonable will fix these situations when they happen. Since you wrote madVR, and you're saying it's not the case, I'll believe you. I just hope you are able to fix whatever you believe really is causing delayed/dropped with only small drops in the queues, since you seem to be implying in previous posts that small queue drops like these should not be a major issue, but they obviously are on my PC. Quote:
I believe it was from some anime DVD source, but it was awhile ago, so I can't remember exactly which I tested it on. Some previously noisy chroma became nearly flat from the blurring when upscaled with chroma set to SoftCubic100. Mitchell-Netravali was the lowest common denominator I could find which didn't seem to overly blur the chroma, and isn't overly sharp either. I'll see if I can track down another source where it is extremely noticeable. Last edited by cyberbeing; 2nd June 2009 at 00:50. |
||
Tags |
direct compute, dithering, error diffusion, madvr, ngu, nnedi3, quality, renderer, scaling, uhd upscaling, upsampling |
Thread Tools | Search this Thread |
Display Modes | |
|
|