madVR - high quality video renderer (GPU assisted)

[ryrynz]

Not sure what you're referring to specifically, but likely it's how it is in the source, you can take a look with another scaler. I think you're just seeing NGU doing a good job connecting the dots.
There's a lot of areas where there's drop outs of strength and aliasing along the lines, which makes it a good test for these neural network type smart scalers.

[madshi]

Originally Posted by cork_OS

My bad, I didn't checked that. I have 1680x1050 display and just used "touch window from outside" in player. Additionally, I didnt'n unchecked "add grain".
I redid the comparison with zoom factor around 4x and found that NGU pix output is much closer to NNEDI3 with some presharpen (like LumaSharpen 0.3) added to NGU.
NNEDI3 64+32 vs. NGU pix med+med
NNEDI3 64+32 vs. NGU pix med+med (LumaSharpen 0.3)
So I carefully assume people may prefer NNEDI3 due to its sharpness, not line thinning differences.

Hmmm... I'm not sure why NNEDI3 is somewhat sharper looking, to be honest. But I'm not sure that's too important. The key purpose of both NNEDI3 and NGU is not to be sharp, but to provide aliasing-free images.

Originally Posted by amayra

if i change madvr to high bitdepth output and my display does not support it
there any harm or bad effects for this ?
source : MPDN user + MPDN Options

What does MPDN have to do with this? When using madVR, setting madVR to high bitdepth output, when your display doesn't support it, usually results in the GPU driver converting the bitdepth down behind madVR's back. This is usually not good for image quality.

Originally Posted by omarank

1) NGU Pixart looks very nice and better than NNEDI3 overall, but I find NNEDI3 more natural looking. NGU Pixart gives a flat look to the images as I see it.

Screenshots, please? (with source image)

Originally Posted by Georgel

1 . NGU Pixart looks better overall.

3 . If you really want to - you could.

4. I think yes at this moment.

Thx.

Originally Posted by atcmvr

First off, in the devices tab, for hdr, is it possible/recommended to use the 3dlut used for calibration? Or are they 2 completely separate things, that each need to be generated on their own?

If you activate a 3dlut in the HDR section, that 3dlut is expected to perform both HDR -> SDR conversion and calibrate your display at the same time. The 3dlut in the calibration tab is not used in that case, when playing HDR content. If you tell madVR to convert HDR -> SDR by using pixel shader math, the calibration 3dlut in the calibration tab is used instead, after HDR -> SDR conversion. Both solutions should end up with nearly the same result. Obviously it's easier to just create one 3dlut for calibration and let madVR do the HDR -> SDR conversion. But the GPU consumes a bit more power this way.

Originally Posted by atcmvr

In the scaling algorithms section, for image upscaling, if you choose ngu, what is the difference between the chroma here, and the actual chroma upscaling tab? Does one take precedence, or are they doing entirely different things? And is it sufficient to leave the chroma option in image upscaling, along with the algorithms to use after doubling to "let madvr decide"?

The extra chroma upscaling tab is for converting 4:2:0 YCbCr content to RGB. E.g. a Blu-Ray has an Y channel with the resolution 1920x1080, but the CbCr channels are only 960x540. So the CbCr channels first need to be upscaled to 1920x1080. That's what the extra "chroma upscaling" tab is about. Then, when all 3 channels are at 1920x1080, and if you have a 4K display, the image upscaling needs to convert the image up from 1920x1080 to 3840x2160. When using NNEDI3 or NGU, this is done separately with the Y and CbCr channels again. So you have once more the option how to handle the CbCr channels. The luma channel is more important, which is why it often makes sense to choose different settings. If you don't know what is best, I'd recommend using "let madVR decide".

Originally Posted by atcmvr

Also, I think I read that you don't really have to bother with any of the upscaling refinements if using ngu, since many of them are already applied?

None of them is applied by default when using NGU, but NGU is very sharp on its own, so often no upscaling refinement is needed on top of that. But it's your decision.

Originally Posted by atcmvr

Also, is ngu high/very high "better" than using nnedi with a high neuron count?

It depends on the source. With a high quality source (the resolution doesn't matter, but the quality of each pixel does), NGU is the best algorithm to use. But if the source is bad (e.g. visual aliasing or compression artifacts), the softer NNEDI3 can be better.

Originally Posted by atcmvr

Another question I had was, does bumping the gpu queue size up past 8 do anything? I ask, because on a different machine + prior version of madvr, maybe 2 years ago, increasing that option to 16 from 8 helped me avoid frame drops for the settings I had then.

Using large queues requires a lot of extra RAM on your GPU. If you have enough RAM, it should at least not harm, might even be useful.

Originally Posted by atcmvr

Finally, what combination of settings should I use, in terms of v-sync, g-sync, and smooth motion? The monitor I have is a g-sync monitor, can go up 165hz. Normally leave it at that, with g-sync enabled, and v-sync enabled through the nvidia console as well. But for viewing content that I guess is 23.976fps, at least going by what madvr says, should I turn of g-sync? Turn off v-sync? Drop the refresh rate to 144hz? Does doing any of that even matter/change anything And which setting under smooth motion would I choose, in combination with the above?

That's a complicated topic. Generally, G-Sync is useless for madVR, so it's probably better to disable it for video playback. Don't disable v-sync. Ideally, set your display to a refresh rate which matches the movie framerate. If you can do that, smooth motion FRC isn't needed.

Originally Posted by ryrynz

Depends on what you're viewing. The lines that NGU pixart creates are excellent but the overall look is bland, it's also sometimes not smart enough and removes detail which I've stated as it being a bit "too aggressive"

At this point if that were to happen there would definitely be some unhappy people, me being one of them.

Here's NNEDI3 64 neurons vs NGU very high on a source I use a lot when evaluating upscalers and my own Anime Avisynth profies for low res content.

http://screenshotcomparison.com/comparison/199986

There is reasonably strong ringing and aliasing and quite soft lines, NNEDI3 enhances more of the image creating lines stronger and true to the source. Here the eyes are sharper and the image is more defined, although the lines are not as well enhanced as NGU pix very high. This is NGU's best quality and it bests even NNEDI3 256 neurons at creating these beautiful non aliased lines, notice the top shirt lines just below the neck for a good example of this. Unfortunately like I've said it also removes some things it doesn't consider to be lines, the wedge of hair just above the left eyebrow loses it's blurry left line for example.

Looking at the above you can see NGU pix's somewhat dull, processed appearance, for NGU pix to replace NNEDI3 it needs to do a better job in all areas and it's simply not doing this right now.

Where's the source image (not the video, but the exact frame you used for the comparison)? FWIW, NGU is currently optimized to do one 2x upscale. Although of course you can cascade it to do 2x 2x, I think that if I manage to do direct 4x, results might be better. What happens if you compare NNEDI3 vs NGU at just 2x upscaling?

-------

After some reviewing, I think the pictures I'm currently working with to create NGU algorithms are too consistently high quality with too sharp and thin lines. Basically I'm mostly using very high quality photos and a few sharp anime screenshots.

Maybe I can improve the algorithms further to also handle soft and fat lines if I add some more such images to the mix. So please help me collecting a wider variety of images. What I need are screenshots of sources (mostly anime, but also real world video/photo) which are high-res but have rather fat and soft lines in it. I need images with lots and lots of high contrast (but fat and soft) lines in them. An image with a large blurred area and just a couple strong lines doesn't help. The whole frame should have a lot of lines in them to be helpful. The images need to be aliasing free and have no (visible) compression artifacts. So they need to be high quality, *but* have soft and fat lines. Can anybody find such images for me? The exact resolution doesn't matter.

If any Anime viewer could create maybe a dozen such images for me, that might be a great help improving NGU pixart (maybe also the regular NGU) further - thanks!

(Just to avoid confusion: I need these images *unscaled* in their original resolution.)

P.S: Please zip the images and upload them somewhere for me to download. Please don't attach them to this thread, because attachment approval can take a long time in this forum.

[citrixscu]

Could be somewhat OT here, but I noticed after upgrading to the latest Windows 10 Fast Ring Build (15031), D3D11 presentation mode no longer crashes MPC-BE x64. Running x64 versions of everything and CUVID hardware decoding. Nvidia 378.57 on GTX980.

[omarank]

Originally Posted by madshi

Screenshots, please? (with source image)

I have created screenshot comparisons with images having different perspectives. I find NNEDI3 doubling has more natural depth than NGU Pixart. (Scaling settings used are also mentioned in a text file in the zipped folder)

[kitame]

Originally Posted by ryrynz

Not sure what you're referring to specifically, but likely it's how it is in the source, you can take a look with another scaler. I think you're just seeing NGU doing a good job connecting the dots.
There's a lot of areas where there's drop outs of strength and aliasing along the lines, which makes it a good test for these neural network type smart scalers.

i'm taking about these errors on NNEDI3.
the distortion on the lines as if it was "stretched".

http://imgbox.com/igGCmjpw

[Damien147]

Is there a reason why GPU load isn't linear?I thought it could be a cpu bottleneck but look at HD cpu load,still spikes happen.

4K

HD

If I remember well my old HD4850 had linear load.

[huhn]

the GPU usage in GPU-Z doesn't work properly with polaris GPU.
it is jumping between 0-100 % as it pleases just ignore it.

[Damien147]

HWmonitor does the same thing.I will ignore I guess.

[ryrynz]

Originally Posted by kitame

i'm taking about these errors on NNEDI3.
the distortion on the lines as if it was "stretched".

http://imgbox.com/igGCmjpw

No error, the source is blurry. NNEDI just doesn't really touch it much.

[madshi]

Originally Posted by omarank

I have created screenshot comparisons with images having different perspectives. I find NNEDI3 doubling has more natural depth than NGU Pixart. (Scaling settings used are also mentioned in a text file in the zipped folder)

Thanks, some interesting images in there. I can extract the following conclusions from these images:

1) Both NGU pixart and NNEDI3 sometimes make mistakes in interpreting the images. Maybe it happens slightly more often with NGU pixart, but I'm not completely sure.

2) NGU pixart has a significantly higher ability to smooth out jagged lines in certain angles. E.g. in your image 3, NGU pixart wipes the floor with NNEDI3.

3) There's a certain difference in the "look" NGU pixart and NNEDI3 produce. NGU pixart tends to create images that look finer structured, but maybe ever so slightly more like an oil painting, sometimes. NNEDI3 tends to create images which look a bit sharper but also slightly more bloated. I'm not sure which look is "better". It might be a matter of taste.

Maybe if enough users send me good images I can improve the algo a bit more, I'm not sure.

Originally Posted by kitame

i'm taking about these errors on NNEDI3.
the distortion on the lines as if it was "stretched".

http://imgbox.com/igGCmjpw

I think you have the images reversed. The one you named NNEDI3 is probably actually NGU pixart and vice versa.

[Georgel]

Originally Posted by ryrynz

Depends on what you're viewing. The lines that NGU pixart creates are excellent but the overall look is bland, it's also sometimes not smart enough and removes detail which I've stated as it being a bit "too aggressive"



At this point if that were to happen there would definitely be some unhappy people, me being one of them.

Here's NNEDI3 64 neurons vs NGU very high on a source I use a lot when evaluating upscalers and my own Anime Avisynth profies for low res content.

http://screenshotcomparison.com/comparison/199986

There is reasonably strong ringing and aliasing and quite soft lines, NNEDI3 enhances more of the image creating lines stronger and true to the source. Here the eyes are sharper and the image is more defined, although the lines are not as well enhanced as NGU pix very high. This is NGU's best quality and it bests even NNEDI3 256 neurons at creating these beautiful non aliased lines, notice the top shirt lines just below the neck for a good example of this. Unfortunately like I've said it also removes some things it doesn't consider to be lines, the wedge of hair just above the left eyebrow loses it's blurry left line for example.

Looking at the above you can see NGU pix's somewhat dull, processed appearance, for NGU pix to replace NNEDI3 it needs to do a better job in all areas and it's simply not doing this right now.

I see now!

Very interesting example!

I will try to get a few examples of mine!

@mdashi - what is the best method to get an "original image" example to have what to compare to?

[madshi]

Probably the media player's screenshot functionality. It's important that the original image is unscaled, and not even AR corrected. So e.g. an NTSC DVD screenshot should be 640x480.

[Asmodian]

NTSC DVD should be 720x480 if unscaled and no AR correction or cropping.

[madshi]

You're right, of course.

[huhn]

image like this "looking upscaled" with fat lines: https://abload.de/img/testktuua.png
or images like this DVD fat lines but not artifact free: https://abload.de/img/test2lru2k.png

[madshi]

They could both be helpful, I'll have to try. Maybe you can sort them into two groups?

[huhn]

here some screens:

http://filehorst.de/d/bpvwgDoI
the BD screens are from "low" bit rate (>20 mbit) VC-1 BD there is sometimes some colored noise around the lines

[kitame]

Originally Posted by ryrynz

No error, the source is blurry. NNEDI just doesn't really touch it much.

no, i checked the source, it doesn't get steched like that at those spots.

Originally Posted by madshi

I think you have the images reversed. The one you named NNEDI3 is probably actually NGU pixart and vice versa.

then that makes NNEDI3 poor in aliasing.

you can only see these things when you zoom in real close, pixel errors are hard to notice.

[atcmvr]

Thanks for the reply madshi. Definitely helped clear some things up for me. I still was a bit unclear about refresh rate and v-sync though. You say generally g-sync is useless for madvr, so it's better to disable it. Better to disable because it can potentially mess with things, or just better to turn off? No problem with turning it off, but obviously I would like to avoid doing so every time I watch something if it wasn't really necessary. Also, as an aside, render times ought to be low 30's, correct? Been trying to keep it max 32.

As for setting the refresh rate, the lowest I can set is 24fps. Other than my mouse lagging across my screen, video seems more or less fine. I do notice that smooth motion turns off, whether set to the motion judder option or the refresh rate option, but I also saw that it drops a frame about every 25 seconds. Bumping up to 144hz, a frame drop every 1.5-2 minutes or so, smooth motion off for the judder option, but on for the multiple option(makes sense, 24*6 = 144). When I turn it all the way up to 165hz though, I noticed a couple things. One, it didn't say I would have dropped frames every x anymore. And it turned on smooth motion for both the judder setting, and the multiple setting. Had to look up what judder was, can't say I've really ever noticed it myself, perhaps I never really saw a bad example in regular viewing. But I guess it's happening at 165hz, and smooth motion is on to fix it. Not sure why the multiple setting also works though. Am I correct in thinking that the unequal division of 165/24 causes some sort of judder? Was also on for 60 and 100hz.

I'm unclear as to why frames are being dropped at some levels and not others. At 100hz, 165hz no frames dropped, but as stated above, 144hz does have frame drops, in addition to 24 and 60hz. Also unsure why you would turn on smooth motion if there was no judder, ie. the refresh rate being a multiple.

Basically unsure as to what setting is best. 24hz, frame drops, no judder(at least madvr says there shouldn't be any), 144 hz, less frequent frame drops, no judder, 165hz, no frame drops, but judder(though again, not sure I can notice it), etc. The ideal I assume is neither happening.

[ryrynz]

Originally Posted by kitame

then that makes NNEDI3 poor in aliasing.

You should really look at scaling with Lanczos or whatever to see how bad the aliasing is when anything that's not a neural network scaler isn't working on it..

Yes NNEDI3 loses to NGU on line smoothness as I said, in this area NGU is probably unparalleled, but NNEDI3 is far from poor.. It was nothing short of incredible when released.