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. |
6th October 2012, 13:53 | #14341 | Link | ||
Registered Developer
Join Date: Sep 2006
Posts: 9,140
|
Quote:
Lanczos3 reads 12 source pixels to create 1 output pixel Lanczos4 reads 16 source pixels to create 1 output pixel Lanczos8 reads 32 source pixels to create 1 output pixel Jinc3 reads 36 source pixels to create 1 output pixel Jinc4 reads 64 source pixels to create 1 output pixel Jinc8 reads 256 source pixels to create 1 output pixel Basically doubling taps for Lanczos means double memory bandwidth consumption. Doubling taps for Jinc results in a 4x increase in memory bandwidth consumption. Quote:
|
||
6th October 2012, 14:02 | #14342 | Link | |
Registered User
Join Date: Mar 2009
Posts: 3,646
|
Quote:
This is why I continue to use Mitchell, it's the sharpest resizer that doesn't create ringing halos on animated content. I'll take a close look at how anti-ringing compares with Jinc vs Lanczos at a later since they're using different techniques? I've just been looking at the ringing(haloing) at this point not how it effects the lines when zoomed in, although you did grab that sample from me, did you look any further into it there? Your thoughts on the OSD addition for scaling I suggested? I thought it would be handy to see that although people can always work this out themselves. Last edited by ryrynz; 6th October 2012 at 14:10. |
|
6th October 2012, 14:24 | #14344 | Link | ||
Registered Developer
Join Date: Sep 2006
Posts: 9,140
|
It's the same logic, but due to how Jinc works, madVR can look at 4 source pixels to control anti-ringing, while with all other algorithms, madVR can only look at 2 source pixels. That should make anti-ringing less prone to introducing aliasing/stepping artifacts. Plus, it also allowed me to make the ringing-removal stronger without losing contrast/detail.
Quote:
If it's really the ringing itself which worries you now, Jinc with anti-ringing enabled should really get rid of most ringing. If you find it's not working as well as expected, please send me a sample. Please note, though, that Lanczos and Jinc do tend do make already existing ringing more obvious, which Mitchell might not does as much. So maybe the issue isn't really that Jinc adds ringing by itself but instead that the source already has ringing and Jinc just makes it more obvious compared to Mitchell? You might be able to work around this issue by running a de-ringing avisynth script in ffdshow. Quote:
I've told the compiler to use "." as decimal separator very early during the initialization of madHcCtrl.exe. Somehow when Overlay was used, the decimal separator was switched back to "," for some reason. I don't understand (at all) why and where. I've now simply force the decimal separator back to "." when the settings dialog is shown. |
||
6th October 2012, 15:25 | #14346 | Link | |||
Registered User
Join Date: Mar 2009
Posts: 3,646
|
Quote:
Quote:
Quote:
Source resolution=640x480 Output resolution=1920x1080 Scaling= 3x/2.25x Reason for mentioning it was that I was wondering how much work my resizing work my GPU was doing for it to be running so slow and by displaying the scaling it saves looking at the source and output resolution and dividing it to find out.. for the lazy I guess, thought it would be nice. *shrug* I kept mentioning it every "few months" |
|||
6th October 2012, 16:04 | #14347 | Link |
Registered User
Join Date: Apr 2009
Posts: 1,019
|
Wow, you weren't kidding about Jinc being demanding. It's 9–12ms depending on the video being played and whether or not it needs deinterlacing.
If I play a 720p video at maximum zoom in MPC-HC I've had it hit 140ms! (something I do when evaluating algorithms, not an issue in the real-world) Looking forward to comparing it with everything else though. Sounds like it could be a winner if your system is up to the task. There does seem to be more aliasing than Lanczos at high frequencies though. (Lanczos 4 and Jinc 4 are similar, Lanczos 8 has less) P.S. I love that you now save settings for algorithm sharpness/taps/etc. Would it be possible to save the anti-ringing filter option on a per-filter basis as well? Some filters seem to be better off without it. (SoftCubic for example) Last edited by 6233638; 6th October 2012 at 16:20. |
6th October 2012, 16:25 | #14348 | Link |
Registered User
Join Date: Jun 2011
Posts: 288
|
madshi,
Thanks for the new version with Jinc, which I believe you made just for me since I bought the GTX 660 the other day.
__________________
SETUP: Win 10/MPC-HC/LAV/MadVR HARDWARE: Fractal Design Node 804 | Xeon E3-1260L v5 | Supermicro X11SSZ-TLN4F | Samsung 2x8GB DDR4 ECC | Samsung 850 EVO 1TB | MSI GTX 1650 Super | EVGA G2 750 |
6th October 2012, 17:04 | #14349 | Link | |
Registered User
Join Date: Aug 2008
Posts: 176
|
Quote:
Thanks for .84 Please have a look at those logs: https://dl.dropbox.com/u/43376972/madvr%20-%20logs%2023%20-%2059%20freez.zip |
|
6th October 2012, 17:24 | #14353 | Link |
Registered User
Join Date: Apr 2009
Posts: 1,019
|
Wow, testing just got so much faster. I just had a look at the keyboard shortcuts section, and there are shortcuts to directly switch between scaling algorithms!
Previously I would have to manually take screenshots of an image with each setting, and paste them into Photoshop as layers to compare them. Now I can just hit a key to switch between them and compare inside madVR! So far Jinc seems pretty interesting in my testing. Jinc 3 seems to be sharper than Jinc 4 at high frequencies, but Jinc 4 is sharper at lower frequencies, and has less aliasing. This is probably a good thing, as most high frequency detail is going to be noise unless you are upscaling Blu-rays. Strangely, Jinc shows even less high frequency detail than SoftCubic in some tests—presumably because it's a "circular" filter rather than a "square" one. Jinc 4 with anti-ringing enabled: SoftCubic 80 with anti-ringing disabled: I've now been able to confirm that pretty much universally, anti-ringing adds aliasing to all the Bicubic variants. (Mitchell-Netravali through to SoftCubic) Bicubic 100 also has a serious amount of aliasing with or without it, for some reason. Drastically more than any other Bicubic option. If you're a fan of sharper algorithms, it definitely seems worthwhile to give Jinc a try. I'd probably suggest Jinc 4 with anti-ringing for now, and Jinc 3 if your system can't handle it. (Jinc 3 also seemed OK without anti-ringing, to further reduce the demands) I wouldn't recommend Jinc 4 without anti-ringing. I still like SoftCubic 80 without anti-ringing for highly compressed sources though. But now thanks to the keyboard shortcuts, you can easily assign keys for Jinc, SoftCubic, and to toggle anti-ringing. As before, Bicubic 75 with anti-ringing is still my preference for Chroma right now. (while anti-ringing increases the chance of aliasing, the ringing reduction is more important for Chroma) |
6th October 2012, 17:26 | #14354 | Link |
Registered User
Join Date: Dec 2010
Posts: 62
|
advice on chroma upsampling
Thank you very much for 0.84, madshi! Works great with JRiver 18.
I have one question regarding chroma upsampling, as this is the only relevant scaling for me most of the time (watch 99% 1080p content on a native 1080p projector): With the new "Jinc" resampling and your anti-ringing filter what is your current recommendation for "best quality" chroma upsampling. I did some comparisons with EVR (on a AMD 5870 with latest Catalyst) and to my big surprise EVR had "better quality" chroma upsampling than I was able to achieve with madVR. The sharpness was comparable to madVR with Lanczos 3 but without any obvious ringing. I tested this with the chroma charts on Stacy Spears Benchmark Blu-Ray. I know that with real content the differences in chroma upsampling are rather hard to spot. So EVR on ATI actually had "better quality" in those 1080p chroma test charts than madVR, no matter which option was selected. Now with the anti-ringing filter this might change. Would you advice to use Jinc with anti-ringing for the "best" (sharpest, while keeping artifacts in control) chroma upsampling madVR is able to? THANK YOU! |
6th October 2012, 17:48 | #14358 | Link | |
Registered User
Join Date: Apr 2009
Posts: 1,019
|
Quote:
It loses to others in artificial testing (such as a chroma zone plate, where Lanczos always wins) but tends to win 90+% of the time in everything else I test. |
|
6th October 2012, 18:04 | #14359 | Link | |
Registered User
Join Date: Dec 2010
Posts: 62
|
Quote:
Maybe I should add that my question is for high bitrate Blu-Ray 1080p content only - so no additional luma scaling. I was hoping there is a better choice than Bicubic 75 for that - something that tests as well as EVR AMD (which uses the graphic cards chroma upsampling)? What about Lanczos/Jinc with anti-ringing? btw try a chroma zone plate with lanczos 8 - it is the only algorithm that gets this pattern right. But the artifacts/multiple ringing is much too strong for general use. I will have to test with anti-ringing enabled. |
|
6th October 2012, 18:11 | #14360 | Link | |||||||||||
Registered Developer
Join Date: Sep 2006
Posts: 9,140
|
Yes, for now, unless I get reports that it's not working well, then I can try to tweak the parameters a bit, like I did with the anti-ringing algorithm for the non-Jinc algorithms a couple of weeks ago.
Quote:
Quote:
Quote:
Quote:
Quote:
Quote:
Quote:
What are your impressions with real video content, though? Quote:
Quote:
Quote:
Anyway, are those Stacey Spears chroma charts available for free somewhere? I'd really like to try them out here. Furthermore: Would you mind making a screenshot with EVR compared to madVR? I'd really like to see what you mean with my own eyes. FWIW, on my PC, my Radeon 3850 uses Nearest Neighbor upsampling for chroma, when using EVR, which of course is terrible. Personally, I still like SoftCubic for chroma upsampling, but it admittedly is quite soft. For some test content that's just what the doctor has ordered. For other content a sharper algorithm may look better. If you like sharper chroma upsampling algorithms, I believe the Anti-ringing filter should help a lot. Just give it a try. In the end, let your eyes decide. Quote:
Yes. I'm not even sure if I will implement Jinc downscaling or Jinc chroma upscaling. Maybe. Maybe not. Depends a bit on your feedback on Jinc image upscaling, I guess. |
|||||||||||
Tags |
direct compute, dithering, error diffusion, madvr, ngu, nnedi3, quality, renderer, scaling, uhd upscaling, upsampling |
Thread Tools | Search this Thread |
Display Modes | |
|
|