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.

 

Go Back   Doom9's Forum > Hardware & Software > Software players

Reply
 
Thread Tools Search this Thread Display Modes
Old 21st February 2017, 09:57   #42641  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
Quote:
Originally Posted by Werewolfy View Post
Only based on your comparison pictures : I prefer pixart 2, it seems thats this one has better anti-aliasing abilities on the aliasedAnime image. 3 is just a little bit sharper on some pictures but it has more aliasing.

Pixart beats NNEDI3 on most of your pictures, I'll try to test on some Anime your test build to see if that's still the case on my sources.
Thanks. Looking forward to your impressions with your own sources.

Quote:
Originally Posted by burfadel View Post
If pixart2 is between pixart 3 and old, then pixart3 is the winner, however a sharpen pass will help tremendously. I believe the sharpen of irfanview is really basic, like sharpen in avisynth where it sharpens the whole image without using edge detection etc.
Quote:
Originally Posted by burfadel View Post
From what I can find, the sharpen as used in Irfanview applies unsharp sharpening to the image. Unsharp sharpenining is also available from the effects list (ctrl + E). Unsharp sharpening to 1 (weak) produces nice results on all the pixart3 images. Using limitedsharpen or adaptivesharpen won't be as 'good' for this blurring because they operate on different principles.
You seem to think that unsharp masking would be some special high quality sharpening. It's not. It's the most simple sharpening you can think of. Actually LumaSharpen is doing exactly the same, just optimized for highest GPU speed, and with some simple limiting to reduce ringing artifacts.

Quote:
Originally Posted by burfadel View Post
Pixart3 is preferred over the original pixart because it may be fractionally clearer, but like in the case of the outdoor barbecue handle it has bad aliasing. That type of aliasing isn't noticeable in the other pics, but I can think of many scenarios where the aliasing would be noticeable. For example, maybe car racing? The old pixart was far better with antialiasing.

Now once unsharp (weak) is used, the old pixart is the winner because it is sharper and clearer than the others without the aliasing!
The main reason why I tuned the original pixart algo is not the missing sharpness, but real interpretation errors and a less natural look to it. These problems can't be fixed with sharpening.

It's true that the original pixart had a great anti-aliasing capability, but overall I believe pixart 2/3 are clearly better.

Quote:
Originally Posted by wasper View Post
It's kind of a two times upscaling process. Don't know or not if I should avoid this.
Generally it's better to avoid duplicate processing. However, if you upscale from DVD to FullHD resolution with madVR, that's the most important part, so I think it's still better to do that than to send the low-res image to the TV.

Quote:
Originally Posted by wasper View Post
Anyway, my question is about color space. If I'm using film mode, the color space seems to work fine for all the image upsacler available (dxva, jinc, spline, etc...). In video mode, color space is fine only if I'm using DXVA. As soon, as I'm changing to software upscaler, the color space seems to get compressed between 16-235 even though everything is set to 0-255(gpu, madvr, tv). Is this a bug or I'm missing something? I have an old DVD movie in 4:3 native but shooted in 16:9. The black bar in the movie are totally black in film mode but grey in video mode if using dxva deinterlacer without dxva image upscaler. To resume, color space seems wrong if I'm using deinterlacing with CPU image upscaler.
madVR tries to avoid using DXVA processing if at all possible, because often the GPU drivers do fancy things to the video when using DXVA. Avoiding DXVA is possible when using film mode, but not when using DXVA deinterlacing, obviously. This is the cause of those color shifts. As huhn already suggested, check the GPU control panel. There are many options there. Probably one related to "dynamic range" or "content type" or similar might be responsible for this problem.

Quote:
Originally Posted by wasper View Post
Second question is about IVTC. How does a dvd player or a blu-ray will play a film telecined to 29.97? Will it deintrelace like video mode and play it at 29.97p or it will use kind of ITVC to play it back at 24fps (probably not as tv must accept this refresh rate)?

I'm having a DVD that will play ok in video mode but with some jitter on some sequences. It will not jitter on film mode. However, with my hardware blu-ray player, it is also playing fine... Could the jitter in video mode related to some clock deviation that are not present in film mode and totally inexistant on hardware player?
madVR currently only supports film mode or DXVA deinterlacing. DXVA deint is *not* video mode. It's a black box. madVR sends the interlaced frames/fields into DXVA and gets progressive frames back. What DXVA does internally, madVR doesn't know. Most GPU manufacturers try to auto detect if the source is video or film, often even per pixel, and try to apply a proper deinterlacing method for that. Unfortunately due to how the DXVA interface is designed, the output is usually double frame rate, even for film sources, which means you can't play the result at 24Hz, and you often have 3:2 pulldown judder baked into the DXVA output.

Hardware DVD/Blu-Ray players have a processing chip built in which does the deinterlacing. Depending on the chip, deinterlacing quality can vary between good and bad. Better chips are often able to deinterlacing film content correctly, some can even IVTC for 24fps output. How good that works will vary from player to player.

Generally, applying IVTC can be very easy, or it can be very tricky, depending on the content. Some content is really 24fps, but just with flags for 60i output. Such content is extremely easy to output in 24fps. Other content is encoded as 60 different fields, either with or without flags that indicate which fields belong together. Yet other content might constantly switch between different encoding modes. madVR's film mode understands every of these variants and tries to produce perfect 24fps. It usually works pretty well. Hardware DVD/Blu-Ray players usually have no problem with the easier encoding variants, but might stumble with more difficult situations.

DXVA deinterlacing is implemented by the GPU manufacturer. So it will vary, depending on whether your GPU is Intel, AMD or NVidia. I think Intel used to have the worst deinterlacer, but I'm not sure if that's still true. In any case, DXVA can't decimate. So if you know that the source is telecined film, it's strongly recommended to use film mode.

Quote:
Originally Posted by Magik Mark View Post
We welcome innovation on image fidelity. I have been an avid Nnedi3 user. I watch a lot of HD sources. So for, nothing beats Nnedi3 64 chroma upscaling.

Comparing it to NGU, all variants, Nnedi3 still sharper, finer and depthful. Image may appear a little bit smaller because of finer lines.

It is more natural and faithful to the original.

I think what NGU lacks at the moment is depth. Maybe we can achieve this by making the lines finer
Are you talking about regular NGU or about the new NGU pixart 2/3 builds? If you're talking about the new pixart builds, can you show some screenshots which demonstrate what you're saying?

Quote:
Originally Posted by Arm3nian View Post
I like pixart 3.

I also tried pixart 3 with luma sharpen in upscaling refinement, and think it looks pretty good. Gives added sharpness without the nasty artifacts that regular NGU produces with low quality sources, and does better than using luma sharpen with nnedi3. Maybe those who wanted a sharpness slider on ngu pixart can use this as an alternative.
Good to hear!

Quote:
Originally Posted by ryrynz View Post
Should 704x480 upscaled to 1920x1200 when using NGU, show the following? NGU very high > NGU High < SSIM1D100 AR? Just wondering why there's downscaling if its supposed to be doubling.. I've reset madVR to defaults. Is the OSD correct?
Quote:
Originally Posted by ryrynz View Post
Ah so it's all baked into the presets. Anything medium and above is actually quadrupling..

Madshi, when looking at the doubled images of Suzumiya it's near impossible to spot any differences between them and Beyond Compare also backs this up with very little changes so I won't refer to those examples.
What I see on my own screen is quite different, what screen res do you use? Here's another comparison below.

This one is from the anime Full Metal Panic! and shows the levels of sharpness difference between NNEDI3 256 and NGU newpix3, screencapped at 1920x1200
It doesn't really show much else with with regards to handling aliasing or blurry lines etc but you'll see all the lines here are much more in focus, especially the eyes and the radio, knobs etc. with no detriment to the image quality that sharpening afterwards would usually add in the form of haloing.

In this case after using sharpen edges refinement of 1.5 after upscale gives a somewhat similar image, but even after that NNEDI3 is still sharper in other background areas. Even at 2.0 the background lines are still sharper using NNEDI3 256 and unfortunately sharpen edges is now oversharpening some lines now which doesn't make it a worthwhile alternative choice to NNEDI3 for anime IMO.
I think your test is probably flawed because your processing chain when comparing NNEDI3 and NGU pixart probably differs. It's kind of my "fault", though. The problem is that currently madVR starts using NGU pixart quadrupling earlier (at a smaller scaling factor) than with NNEDI3. Furthermore, only the NGU very high preset currently uses the new tuned pixart algo. So if you use quadrupling, you're getting the tuned pixart only for the doubling, but not for the quadrupling step, which will still be done with the old pixart medium setting, only. I think it's very likely that these differences are causing the pretty dramatic difference in sharpness you're seeing. Which is actually a good thing because it means when solving the flawed test conditions, there's a good chance you might change your mind about NGU pixart.

As I said before, I'm going to redesign the settings once more in the next build, which should then hopefully make it easier for you to compare the algos in a fair way. Right now you need to take extra care that you're comparing apples to apples.

Can you please try the following:

1) Either just double the image instead of scaling to 1920x1080.
2) Or alternatively: Double the image, make a screenshot, then load the screenshot into MPC to double another time, or this time to fit the screen.

Only if you carefully double check that the processing chains are identical (except for switching NGU and NNEDI3), you're doing a true apples to apples comparison. Also you have to avoid NGU quadrupling right now because you'll not get the new tuned pixart algo for the 4x step right now. You can double check the processing chain by checking the OSD (Ctrl+J). The up and downscaling algos must be perfectly identical, and for NGU every time "very high" must be used. For NNEDI3 every time 256 taps should be used.

Last edited by madshi; 21st February 2017 at 10:00.
madshi is offline   Reply With Quote
Old 21st February 2017, 10:00   #42642  |  Link
Cinemancave
Registered User
 
Join Date: Dec 2012
Posts: 33
Quote:
Originally Posted by huhn View Post
change the power setting in the nvidia control panel to adaptive not optimal
Thank you sir!
Cinemancave is offline   Reply With Quote
Old 21st February 2017, 10:19   #42643  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
Quote:
Originally Posted by ryrynz View Post
Should 704x480 upscaled to 1920x1200 when using NGU, show the following? NGU very high > NGU High < SSIM1D100 AR? Just wondering why there's downscaling if its supposed to be doubling.. I've reset madVR to defaults. Is the OSD correct?
P.S: I've double checked: With your sample, when upscaling to 1920x1200, the current madVR builds gives you the following processing chains:

NNEDI3-256 > Jinc AR
NGU-veryHigh > NGU-high < SSIM1D100 AR

So basically you're comparing apples to oranges.
madshi is offline   Reply With Quote
Old 21st February 2017, 10:29   #42644  |  Link
ryrynz
Registered User
 
ryrynz's Avatar
 
Join Date: Mar 2009
Posts: 3,646
Quote:
Originally Posted by madshi View Post
So basically you're comparing apples to oranges.
Good to know. How far off is this new build? I might just wait for that.
ryrynz is offline   Reply With Quote
Old 21st February 2017, 10:40   #42645  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
I don't know, could be a week, or 2 or 3 or 4. Can't you switch MPC to "Double Size" for a quick comparison test? That will give you apples to apples.
madshi is offline   Reply With Quote
Old 21st February 2017, 15:13   #42646  |  Link
Werewolfy
Registered User
 
Join Date: Feb 2013
Posts: 137
Quote:
Originally Posted by madshi View Post
Thanks. Looking forward to your impressions with your own sources.
I did some tests and I find NGU pixart way better than NNEDI 3 256 neurons on most videos. It's sharper, more focus and sometimes it seems that NNEDI3 produce some artefacts.

But I find it very tricky to choose between pixart2 and 3... Pixart 2 has better anti-aliasing capabilities and pixart 3 is sharper and a little bit more natural sometimes. On some scenes I prefer 2 and on other scenes I prefer 3... I don't really know which one to choose honestly.
__________________
Windows 8.1 and 10 x64 - Intel Core i5-4670K (4.2 GHz) - 8 GB DDR3 - MSI Geforce GTX 1080 8 GB - Sony KD-55A1 - Denon AVR-X3600H
Werewolfy is offline   Reply With Quote
Old 21st February 2017, 15:36   #42647  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
That's what I like to hear! I've had a hard time myself deciding between pixart 2 and 3, that's why I've asked for feedback. FWIW, I may have found a way to create a hybrid which gets near to the best of both pixart 2 and 3. Not sure yet, though.
madshi is offline   Reply With Quote
Old 21st February 2017, 15:43   #42648  |  Link
xyndv
Registered User
 
Join Date: Jan 2013
Posts: 9
Can the maximum limits in image enhancements/upscaling refinement be raised? I find that anti bloating softens the image quite a bit, and I can't compensate enough with more sharpness.
xyndv is offline   Reply With Quote
Old 22nd February 2017, 02:54   #42649  |  Link
Warner306
Registered User
 
Join Date: Dec 2014
Posts: 1,127
Quote:
Originally Posted by xyndv View Post
Can the maximum limits in image enhancements/upscaling refinement be raised? I find that anti bloating softens the image quite a bit, and I can't compensate enough with more sharpness.
Have you tried using a lower value of anti-bloating (e.g. 25%)?
Warner306 is offline   Reply With Quote
Old 22nd February 2017, 03:23   #42650  |  Link
andybkma
Registered User
 
Join Date: Sep 2006
Posts: 212
madshi, does mVR work more efficient and/or use less gpu load/resources if aero is on or off on a Win7 64 bit system with a low powered laptop gpu (Nvidia 650M)? I have been using mVR for years & years with aero turned off but would like to know if turning aero on would help out mVR. Thanks
andybkma is offline   Reply With Quote
Old 22nd February 2017, 03:57   #42651  |  Link
ryrynz
Registered User
 
ryrynz's Avatar
 
Join Date: Mar 2009
Posts: 3,646
Apples to apples what I've said still holds at least on the Suzuimya image. When checking the FMP image there's almost nothing in it vs NNEDI 64 neurons, a slight sharpness edge is there with NGU newpix3 which was a little surprising, it would still lose out to 256 neurons though. With Suzumiya though, NNEDI3 64 neurons is sharper and by a decent margin too and of course 256 neurons slightly more so. Here I've cut the face away and pixel enlarged by 3x. Newpix's Lines are looking pretty good to me at least.

NGU newpix 3 vs NNEDI3 64 neurons

NGU newpix 3 vs NNEDI3 256 neurons.

On another video I'm looking at (Rurouni Kenshin encode off DVD) NGU newpix 3 is ever so slightly sharper than NNEDI3 64 neurons, the images are very close in appearance to each other as well.
Kinda strange how NGU's sharpness seems to be all over the place in comparison to NNEDI3. Didn't some say it's sharper than NNEDI3 256 in some situations?

Last edited by ryrynz; 22nd February 2017 at 05:01.
ryrynz is offline   Reply With Quote
Old 22nd February 2017, 04:57   #42652  |  Link
70MM
X Cinema Projectionist NZ
 
Join Date: Feb 2006
Location: Auckland NZ
Posts: 310
Its always been said here that NNEDI is better on lesser quality material than NGU, is that because its softer?

So what is better on lesser quality material, NGU or the new pixart?
70MM is offline   Reply With Quote
Old 22nd February 2017, 05:10   #42653  |  Link
ryrynz
Registered User
 
ryrynz's Avatar
 
Join Date: Mar 2009
Posts: 3,646
I'm guessing preferences could vary depending on the material. I just tested 640x480 content and it looks near identical to NNEDI3 64 neurons but with much better performance..
I'd rather evaluate whatever madshi cooks up next with this mix of 2/3 or whatever..ATM I still prefer NNEDI3's over all sharper image.
Having sharper fainter lines is important to me, otherwise you get sharp edges with everything else inside looking rather soft & dull.

I think at this point I may end up switching between NNEDI3 and NGU newpix on the fly and observe the differences.. I'm not sure stills will be enough to make a decision.
Still I'd rather not sink time into something that's going to change soon.

Last edited by ryrynz; 22nd February 2017 at 05:16.
ryrynz is offline   Reply With Quote
Old 22nd February 2017, 05:30   #42654  |  Link
70MM
X Cinema Projectionist NZ
 
Join Date: Feb 2006
Location: Auckland NZ
Posts: 310
I suppose it also depends on the size of the screen one uses....
Everything we show on these pages on a monitor to me can look great, blow them up to my 150" diag scope screen is a very different story.
Here is where I want sharp without showing artefacts and noise.

I love NNEDI but some material can look soft.
Im my experiments with NGU I have to be very careful as its easy to introduce noise to the image if I push sharpness settings...
I have no idea what pixart looks like yet as I havent added it.

But my movies look a 100% better with madvr and they ever did with my Oppos, thanks to madshi!
70MM is offline   Reply With Quote
Old 22nd February 2017, 10:22   #42655  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
Quote:
Originally Posted by xyndv View Post
Can the maximum limits in image enhancements/upscaling refinement be raised? I find that anti bloating softens the image quite a bit, and I can't compensate enough with more sharpness.
Which sharpness algos are we talking about? And which maximum limit would you like to have?

Quote:
Originally Posted by andybkma View Post
madshi, does mVR work more efficient and/or use less gpu load/resources if aero is on or off on a Win7 64 bit system with a low powered laptop gpu (Nvidia 650M)? I have been using mVR for years & years with aero turned off but would like to know if turning aero on would help out mVR. Thanks
In fullscreen exclusive mode it should not make a difference. In windowed mode, having aero on can help because madVR can switch into "present several frames in advance" mode which isn't working with areo turned off. If you want to use areo, I'd advise to use Windows 8.1, though, because aero is *much* improved in 8.1. E.g. aero has a lot of trouble with multi monitor setups in win7.

Quote:
Originally Posted by ryrynz View Post
Apples to apples what I've said still holds at least on the Suzuimya image. When checking the FMP image there's almost nothing in it vs NNEDI 64 neurons, a slight sharpness edge is there with NGU newpix3 which was a little surprising, it would still lose out to 256 neurons though. With Suzumiya though, NNEDI3 64 neurons is sharper and by a decent margin too and of course 256 neurons slightly more so. Here I've cut the face away and pixel enlarged by 3x. Newpix's Lines are looking pretty good to me at least.

NGU newpix 3 vs NNEDI3 64 neurons

NGU newpix 3 vs NNEDI3 256 neurons.

On another video I'm looking at (Rurouni Kenshin encode off DVD) NGU newpix 3 is ever so slightly sharper than NNEDI3 64 neurons, the images are very close in appearance to each other as well.
Kinda strange how NGU's sharpness seems to be all over the place in comparison to NNEDI3. Didn't some say it's sharper than NNEDI3 256 in some situations?
Try the castle image to see an example of where NGU pixart is *noticeably* sharper (or rather more "focused") than NNEDI3, and the difference is much larger in the castle image compared to the Suzumiya image. It's hard to say if it's NGU pixart or NNEDI3 which is all over the place. Depends on which algo you set as "reference", I suppose.

If I look at your screenshot comparison I can agree that NNEDI3 is ever so slightly sharper with this particular image, which of course becomes more noticeable if you stack 2 doublings on top of each other. However, is a tiny difference in sharpness now the one and only thing that is of interest to you? How about good quality lines and natural look?

If you look at the two white dots/circles in the left eye, in the NNEDI3 image the smaller circle seems to be more rectangular shaped compared to the NGU image. This is more obvious when you view the images in 100% instead of zoomed in with nearest neighbor. Or if you look at the left border of the face: There is a faint vertical shadow right next to the left face border in the bottom 3rd of the image. This shadow looks clearly better with NGU than with NNEDI3-256. With NNEDI3-64 this shadow is totally terrible. Or the hair which "meets" the top left corner of the left eye brow is also much better in the NGU image. It seems to me that the NGU pixart result, while overall being ever so slightly softer in this image, overall has the better looking and more natural lines.

Try adding 0.1 AdaptiveSharpen Upscaling Refinement after NGU pixart (after both doubling and quadrupling). The result will look noticeably sharper than NNEDI3-256 while still looking more natural to my eyes. I don't think a tiny sharpness difference should be the one and only factor you're looking at, when judging which algo does better.

What do you think?

Quote:
Originally Posted by 70MM View Post
Its always been said here that NNEDI is better on lesser quality material than NGU, is that because its softer?

So what is better on lesser quality material, NGU or the new pixart?
NGU pixart and NNEDI3 are very very similar to each other. Standard/classic NGU is totally different. Generally, NNEDI3 and NGU pixart are very good at removing aliasing from the source, while being rather soft. Standard/classic NGU is much sharper, but doesn't handle source artifacts well.
madshi is offline   Reply With Quote
Old 22nd February 2017, 10:34   #42656  |  Link
burfadel
Registered User
 
Join Date: Aug 2006
Posts: 2,229
Quote:
Originally Posted by madshi View Post
Which sharpness algos are we talking about? And which maximum limit would you like to have?

Try the castle image to see an example of where NGU pixart is *noticeably* sharper (or rather more "focused") than NNEDI3, and the difference is much larger in the castle image compared to the Suzumiya image. It's hard to say if it's NGU pixart or NNEDI3 which is all over the place. Depends on which algo you set as "reference", I suppose.

If I look at your screenshot comparison I can agree that NNEDI3 is ever so slightly sharper with this particular image, which of course becomes more noticeable if you stack 2 doublings on top of each other. However, is a tiny difference in sharpness now the one and only thing that is of interest to you? How about good quality lines and natural look?

If you look at the two white dots/circles in the left eye, in the NNEDI3 image the smaller circle seems to be more rectangular shaped compared to the NGU image. This is more obvious when you view the images in 100% instead of zoomed in with nearest neighbor. Or if you look at the left border of the face: There is a faint vertical shadow right next to the left face border in the bottom 3rd of the image. This shadow looks clearly better with NGU than with NNEDI3-256. With NNEDI3-64 this shadow is totally terrible. Or the hair which "meets" the top left corner of the left eye brow is also much better in the NGU image. It seems to me that the NGU pixart result, while overall being ever so slightly softer in this image, overall has the better looking and more natural lines.

Try adding 0.1 AdaptiveSharpen Upscaling Refinement after NGU pixart (after both doubling and quadrupling). The result will look noticeably sharper than NNEDI3-256 while still looking more natural to my eyes. I don't think a tiny sharpness difference should be the one and only factor you're looking at, when judging which algo does better.

What do you think?

NGU pixart and NNEDI3 are very very similar to each other. Standard/classic NGU is totally different. Generally, NNEDI3 and NGU pixart are very good at removing aliasing from the source, while being rather soft. Standard/classic NGU is much sharper, but doesn't handle source artifacts well.
It's why I was thinking that adding a sharpening pass to the algorithm would be beneficial, as most people are noticing the softness. This it the reason why I was suggesting a weak unsharp sharpening pass, it might be one of the simplest sharpening filters, but for most of the algorithms it basically just brings it into focus. Supposedly that was the whole point of unsharp to begin with, to reverse the softness caused by digital imagery processing. Since NGU is digital image processing and causes softness...

At least, that's the principle described on Gimp where I believe unsharp originated https://docs.gimp.org/en/plug-in-unsharp-mask.html

Of course the user can still add adaptive sharpen etc after, the whole point is just to overcome the softness more than actively sharpening the image. Of course anyone can add sharpening after, but the additional step is seen as a fallacy of the algorithm, whereas if the simple (but very effective in this case) sharpening was applied within the algorithm itself, perceptively it would be superior to other options, including NNEDI3 256.

I add a comparison of American Dad, Castle, and the Potatoes of PixartNew3 with Unsharp and NNEDI.

American Dad (accidentally called it Potatoes )
http://screenshotcomparison.com/comparison/201374

Castle:
http://screenshotcomparison.com/comparison/201375

Real Potatoes:
http://screenshotcomparison.com/comparison/201377

Last edited by burfadel; 22nd February 2017 at 10:59.
burfadel is offline   Reply With Quote
Old 22nd February 2017, 10:45   #42657  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
You want extra sharpening? Simply activate sharpening in Upscaling Refinement. There's no need to complicate things by adding an extra option/pass to NGU pixart when it's already available right now via Upscaling Refinement.
madshi is offline   Reply With Quote
Old 22nd February 2017, 11:16   #42658  |  Link
ryrynz
Registered User
 
ryrynz's Avatar
 
Join Date: Mar 2009
Posts: 3,646
Quote:
Originally Posted by madshi View Post
If I look at your screenshot comparison I can agree that NNEDI3 is ever so slightly sharper with this particular image, which of course becomes more noticeable if you stack 2 doublings on top of each other. However, is a tiny difference in sharpness now the one and only thing that is of interest to you? How about good quality lines and natural look?
Lines look good, given the level of performance required and the output achieving NNEDI3 256ish quality, I'm pretty happy with it.

Quote:
Originally Posted by madshi View Post
If you look at the two white dots/circles in the left eye..
Or if you look at the left border of the face..
Or the hair which "meets" the top left corner of the left eye brow..
Yeah, I'm aware of all these areas, definite pluses for NGU newpix.


Quote:
Originally Posted by madshi View Post
Try adding 0.1 AdaptiveSharpen Upscaling Refinement after NGU pixart (after both doubling and quadrupling). The result will look noticeably sharper than NNEDI3-256 while still looking more natural to my eyes.
0.2 looks a bit heavy handed, 0.1 is okay (after upscale) still not quite as sharp as NNEDI3 256 in some areas but TBH I don't run anything over NNEDI3 64 neurons with the GTX 960 anyway, so likely this is sharper than that. Almost feel there needs to be another sharpener that can pick up on stuff other than just the more major lines.. but I guess that's what I've got Avisynth for.

Quote:
Originally Posted by madshi View Post
I don't think a tiny sharpness difference should be the one and only factor you're looking at, when judging which algo does better.
Wasn't really judging what was better as such, just my preference and I wanted to focus on what I felt was the most important considering NGU newpix had closed the gap in all the other areas.
ryrynz is offline   Reply With Quote
Old 22nd February 2017, 12:56   #42659  |  Link
Damien147
Registered User
 
Join Date: Mar 2011
Posts: 380
@ Madshi

Can you do something with the black screen and HEVC in dxva(native)?User is gonna install mpc hc,madvr and is gonna be looking what's goin' on with the black screen as it is the default.
I had hardware decoder to none as suggested but I found problem with high bitrate HEVC in 4k.My pc couldn't handle it(cpu load,dropped frames) and I lowered chroma even to Bilinear.
I can run low bitrate HEVC in 4k with NGU med chroma but for high bitrate it's a no go.It seems that dxva(native) could solve that.
At the current state with high bitrate HEVC and hardware decoder to none it's not worth it to use Madvr at 4k.For everything else it's ''godlike'' but I found this small ''hole''.
I know that your priorities are different now but keep it in mind please.

EDIT:


Madvr with bilinear chroma and hardware decoder to none.



EVR(custom presenter) with YV12 chroma upsampling shader and hardware decoder to dxva(native).


You get this ridiculously high cpu load leading to dropped frames with no real life image quality benefits.

Last edited by Damien147; 22nd February 2017 at 13:50.
Damien147 is offline   Reply With Quote
Old 22nd February 2017, 15:29   #42660  |  Link
CruNcher
Registered User
 
CruNcher's Avatar
 
Join Date: Apr 2002
Location: Germany
Posts: 4,926
Even on older Mid Range 4 Core Systems you can reach acceptable 4K decoding results (if you arent afraid of heating your room with decoding already) and benefit from the overall efficiency improvements depending on the overall bitstream complexity 25Mbits is a good overall bitrate but you shouldn't go higher then lanczos3/4 then if you want additional quality through post.
__________________
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; 22nd February 2017 at 15:35.
CruNcher is offline   Reply With Quote
Reply

Tags
direct compute, dithering, error diffusion, madvr, ngu, nnedi3, quality, renderer, scaling, uhd upscaling, upsampling

Thread Tools Search this Thread
Search this Thread:

Advanced Search
Display Modes

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off

Forum Jump


All times are GMT +1. The time now is 13:19.


Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2024, vBulletin Solutions Inc.