madVR - high quality video renderer (GPU assisted)

[thighhighs]

Originally Posted by Gopa

Always open to suggestions for alternative settings.

What you think about NGU Sharp without any extra sharpness? IMHO:
NGU Sharp look very sharp by default, edges looks thin and sharp in focus. Another situation when you use NGU AA for image upscaling. NGU AA + post sharp can be better than NGU Sharp for low quality sources, because NGU AA can fix some edges, where NGU Sharp\Standard just sharp it.

Try something like this:

Artifact removal - Debanding: Low\Low or med\med
Artifact removal - Deringing: Off
Artifact removal - RCA: What you like
Image enhancements: Off
Chroma: NGU AA High
Image doubling: NGU Sharp High (looks like you prefer sharp image)
Chroma doubling: Off\ let madvr decide
Upscaling refinement: Off
Dithering: Orderer

edit:

Originally Posted by ABDO

well, no Surprise here, NGU Sharp look, best, crisp and closer to
Blu-Ray

+

[huhn]

Originally Posted by madshi

In case you guys wonder how NGU Sharp compares to mpv's latest FSRCNN(X), here's a little comparison:

Blu-Ray screenshot
downscaled (PNG) | (JPG, 100 quality)
latest FSRCNN32
latest FSRCNNX32
NGU Sharp - High
NGU Sharp - Very High

To make things as fair as possible I've downscaled the image with Bicubic/Catrom, which is exactly what FSRCNN and FSRCNNX were trained for.

are you sure everything is correct i mean FSRCNN has chroma bleeding...

but in term of anti alasing (ignoring chroma) i give FSRCNN more points but i would nearly for sure use NGU AA on this source.
i'm not a fan of NGU sharp on anime in general lines can look really strange.

[madshi]

Originally Posted by huhn

are you sure everything is correct i mean FSRCNN has chroma bleeding...

Well, I don't know how to activate FSRCNN(X) for chroma doubling. I don't know if it's possible. If anyone knows, let me know and I'll redo the FSRCNN(X) screenshots.

Originally Posted by huhn

but in term of anti alasing (ignoring chroma) i give FSRCNN more points but i would nearly for sure use NGU AA on this source.
i'm not a fan of NGU sharp on anime in general lines can look really strange.

Can you point me to where in the image exactly you see FSRCNN having lower aliasing than NGU Sharp?

[huhn]

Originally Posted by madshi

FWIW, the key problem with Optimus is that there are really 2 totally separate GPUs, each of which has their own HDMI/DVI driver inside. And what is worse: The actual driving of the display (and HDMI output) is done by the Intel GPU, while the rendering is done by the Nvidia GPU, so the 2 GPUs have to work together. Basically Nvidia has to transport the rendered frames to the Intel GPU, to be sent to the display. Practically this means, both drivers have to work together. That's like driver bugs ^ 2!

If the new Intel CPUs with integrated AMD GPUs only have one GPU instead of two, most (or even all) of the Optimus problems should not occur.

using the intel iGPU for desktop and co makes sadly a lot of sense. the hole system is clearly made for battery time.
i mean the price of HBM is about 3x GDDR5 and currently the major benfit of HBM is powerconsumption.
AMD evne has a zero power GHPU function and not using this woudl be quite a waste.
i'm just guessing here so whatever will take a some time until we see these devices anyway.

These are very weird benchmark results, to say the least. We know that NGU doesn't run as well as it should on AMD Polaris GPUs. But FSRCNN(X) running (ever so slighty) faster on my AMD 560 than on my Nvidia 1070 is just plain weird.

do they use openCL.
i eman even with the handbrake copyback issues on AMD side these cards where usually still a lot faster than nvidia cards back in the nnedi 3 days.

[huhn]

Originally Posted by madshi

Can you point me to where in the image exactly you see FSRCNN having lower aliasing than NGU Sharp?

the bigs babys head for example.

the red hat top center (where FSRCNN has the huge chroma issue)

BTW. the chroma issue is visible on the JPG screen.

[Neo-XP]

Originally Posted by madshi

Well, LL on/off sharpen different kinds of edges with different strengths, so in some parts of the image LL off will sharpen more than LL on, and vice versa. I'm surprised you got halos, though. Do you happen to have a couple good test images where I can see why you prefer LL off? That might help with development.

A quick comparison with The Neon Demon ( Bluray / UHD Bluray ):



Let's try to upscale the Bluray ( NGU very high / NGU very high + AS 0.5 / NGU very high + AS 0.5 LL ):



Zoomed (AS with LL off and on) : http://screenshotcomparison.com/comparison/124629

I prefer with LL off, because it is sharper for the same value. Higher values for LL only accentuate artifacts around the edges so it is not fair to compare with different values.

The old NGU is also closer to the UHD Bluray :



Zoomed : http://screenshotcomparison.com/comparison/124639

Old NGU looks a lot better also for me for this movie, but I do not know if it was downscaled with LL or not.

[madshi]

Originally Posted by huhn

are you sure everything is correct i mean FSRCNN has chroma bleeding...

Updated mpv screenshots, they have less chroma bleeding now.

Originally Posted by huhn

do they use openCL.

No, OpenGL/GLSL.

[madshi]

Originally Posted by huhn

the bigs babys head for example.

the red hat top center (where FSRCNN has the huge chroma issue)

I'm sorry, I'm having trouble seeing it. Where do you see aliasing there? Do you mean the top line of the big baby's head? If so, there's some minor aliasing there, but it's also in the Hi-Res image! Shouldn't you give NGU bonus points for reproducing the Hi-Res image more faithfully, warts and all?

Talking about the big baby's head: How about the more obvious differences, like the big baby's mouth, the eye brows, the nose?

Originally Posted by Neo-XP

A quick comparison with The Neon Demon ( Bluray / UHD Bluray )

Old NGU looks a lot better also for me for this movie, but I do not know if it was downscaled with LL or not.

Thanks!

[leeperry]

Originally Posted by madshi

Is that 4x NGU Sharp in one step, followed by SR? Or is it 2X NGU Sharp + SR + 2x NGU Sharp + SR?

Direct 4X NGU sharp > SR@2 LL for both luma & chroma > SSIM 2D 100% AB25%, hard to believe how good VCD looks now with all the bells and whistles enabled huh!

I still wish you would properly align chroma for MPEG-1 especially as nevcariel said that he was willing to cooperate but it still looks great when outputting properly aligned RGB32 off ffdshow so no biggy I guess

Originally Posted by madshi

I wonder if I should maybe remove the "quality" option and auto pick, based on the strength?

Please don't, I can afford "high" on some videos but not on others and I have some blocky ones that actually appear to look better in "medium". I see "high" as the icing on the cake if I have GPU cycles to spare and PQ improves.

For the record I only use RCA@1, anything higher is too strong to my taste.

And somewhat unrelated to mVR but as much as we care about shielding of A/V cables, it seems that power cables are vastly overlooked and a triple-shielded garden-hose one has tremendously improved sharpness on my Sammy TV. Identical cable actually did the same on my G3-550 powered PC but then again I pickup +40 wifi networks here so not having them interfering with my A/V gear does appear to do the magic

[huhn]

Originally Posted by madshi

I'm sorry, I'm having trouble seeing it. Where do you see aliasing there? Do you mean the top line of the big baby's head?

that line

If so, there's some minor aliasing there, but it's also in the Hi-Res image! Shouldn't you give NGU bonus points for reproducing the Hi-Res image more faithfully, warts and all?

well that topic again.
well i guess we want the image close to the original but one bonus of upscaling was anti alaising. NGU is just so sharp that it is hard to totally avoid alaising.

Talking about the big baby's head: How about the more obvious differences, like the big baby's mouth, the eye brows, the nose?

NGU is generally sharper and the line thickness is closer to the original.
but i just gave FSRCNN(and i through waifux2 is a bad name...) more points for anti aliasing and nothing else and the new image isn't that much better now the chroma bleeding was hiding the aliasing quite a bit. but the thin lines from NGU have a bad part to at the bottom line of the red hat top center. the line is nearly completely gone now while FSRCNN makes it relative big now but the source has this line.

is there a reason you have to use a JPG with mpv don't tell me they don't support PNG? but why else should you add a jpg and a PNG as source...

[Soulnight]

Originally Posted by Polopretress

I agree. That would be great to make the option "command line to execute when this profile is activated/deactivated "works

Madshi: any idea when you could make this function work?

http://bugs.madshi.net/view.php?id=210

[Fabulist]

Originally Posted by madshi

Sharpen Edges internally does some supersampling which reduces bloating. It's all somewhat subjective, please trust your eyes and pick values that look good to you. This is all not really scientific. If you want scientific, you'd have to use deconvolution instead of sharpening, but even then, which deconvolution kernel would you be using. Gaussian or other? Linear light or gamma light?

I understand, thank you!

[madshi]

Originally Posted by huhn

well i guess we want the image close to the original but one bonus of upscaling was anti alaising.

Yes. The problem is that any change you do to an algorithm usually improves something, but makes something else worse. So if I designed NGU to intentionally modify the image to reduce aliasing, it would also result in a softer look and loss of detail at the same time. The whole design approach of NGU Sharp is that it tries to get as close to the original Hi-Res image as possible. And if the original Hi-Res image has aliasing, then NGU Sharp unfortunately sometimes also reproduces that aliasing faithfully.

That said, most aliasing that was introduced through downscaling is removed by NGU Sharp very effectively. Which all the conventional algos like Bicubic, Lanczos or even Jinc don't do.

Originally Posted by huhn

NGU is generally sharper and the line thickness is closer to the original.

And NGU is virtually ringing free, while FSRCNN(X) have visible ringing.

Originally Posted by huhn

but the thin lines from NGU have a bad part to at the bottom line of the red hat top center. the line is nearly completely gone now while FSRCNN makes it relative big now but the source has this line.

Yes. Unless one algo totally sucks, when comparing 2 different algos you'll almost always find something which algo 1 does better than algo 2 and vice versa. The important thing is which algo does better overall, of course.

Originally Posted by huhn

is there a reason you have to use a JPG with mpv don't tell me they don't support PNG? but why else should you add a jpg and a PNG as source...

mpv does support PNG, but it doesn't activate FSRCNN(X) when I try to "play" a PNG or BMP file. I'm not completely sure why, maybe the source has to be YCbCr instead of RGB, I don't know. It's weird, but that's the way it is. I've used the highest possible JPG quality level, though, which is without any visible compression artifacts, even when zooming in. So it should be ok. I would have compared with PNG, if I could, of course.

[huhn]

Originally Posted by madshi

Yes. Unless one algo totally sucks, when comparing 2 different algos you'll almost always find something which algo 1 does better than algo 2 and vice versa. The important thing is which algo does better overall, of course.

i noticed in general that NGU sharp has a tendency to thin really thin lines a little bit to much.

and it looks like NGU sharp has problems with aliased sources too not only "bad" sources.

And NGU is virtually ringing free, while FSRCNN(X) have visible ringing.

if you mean the holoing around lines and discolouring of small image parts than yes i can see this ringing too else i don't really notice it.

another issue i notice with FSRCNN is the "noise" reproduction.
this old source has some colored noise in it and this noise is sometimes reproduced in a checkboard pattern which is really ugly.

[Patrik G]

Originally Posted by madshi

madVR does what you ask it to do. If you use "convert HDR content to SDR by using pixel shader math", and if you also set the "calibration" tab to "this display is already calibrated", then madVR will convert the HDR content to the gamut and transfer function you selected in the calibration tab. That can be BT.709 or DCI-P3 or BT.2020, whatever you choose.

thanks!
it must be a pain in the ass to answer all questions i guess

so that was easy fix
since this Kuro have a native gamut close to DCI P3 i just select this display is already calibrated and choose DCI P3 then.
thats the closest.

but even then
downconverting uhd hdr content to srgb colors with a display that has a wider gamut doent look bad.
colors gets slight more oversaturated but not in a bad way as you would have with blu ray content

[madshi]

Originally Posted by Soulnight

Madshi: any idea when you could make this function work?

I'll move it up in the priority. It's really annoying that the functionality *seems* to be there, but simply doesn't work.

Originally Posted by huhn

if you mean the holoing around lines and discolouring of small image parts than yes i can see this ringing too else i don't really notice it.

another issue i notice with FSRCNN is the "noise" reproduction.
this old source has some colored noise in it and this noise is sometimes reproduced in a checkboard pattern which is really ugly.

It's interesting. Ringing and bloated lines stick out like a sore thumb to my eyes. I can spot that from a mile away. As does banding (at least in motion). I don't like aliasing, either, but it's not as dramatic for me as it seems to be for you. And I didn't even notice the checkboard pattern. So it really seems every person has different sensivities to different types of artifacts.

Originally Posted by Patrik G

so that was easy fix
since this Kuro have a native gamut close to DCI P3 i just select this display is already calibrated and choose DCI P3 then.
thats the closest.

but even then
downconverting uhd hdr content to srgb colors with a display that has a wider gamut doent look bad.
colors gets slight more oversaturated but not in a bad way as you would have with blu ray content

So everything's fine now?

[Gopa]

Originally Posted by thighhighs

What you think about NGU Sharp without any extra sharpness? IMHO:
NGU Sharp look very sharp by default, edges looks thin and sharp in focus. Another situation when you use NGU AA for image upscaling. NGU AA + post sharp can be better than NGU Sharp for low quality sources, because NGU AA can fix some edges, where NGU Sharp\Standard just sharp it.

Try something like this:

Artifact removal - Debanding: Low\Low or med\med
Artifact removal - Deringing: Off
Artifact removal - RCA: What you like
Image enhancements: Off
Chroma: NGU AA High
Image doubling: NGU Sharp High (looks like you prefer sharp image)
Chroma doubling: Off\ let madvr decide
Upscaling refinement: Off
Dithering: Orderer

edit: +

2160p: Tried all of above + thin edges (refinements x2), because this refinement is always my go-to preference for anime. Not sure yet if debanding helped. Will look more closely. Chroma NGU AA High DID help! Reduce random noise (RRN?), instead of RCA, provided same results, with lower rendertime (mostly, I think, because of chroma NGU AA high). Still not quite sharp enough, though & getting very close to rendertime limit. Experimenting, with ffdshow sharpeners. I think, msharpen helps quite a bit. Lots of other possibilities in ffdshow. Not very user-friendly, though. Have yet to figure-out how to add shader pre-sets, in mpc-hc, so that is on my to-do list. Will try lower chroma and/or luma settings & see if I can use another madVR sharpener, instead of ffdshow, because I can easily see realtime results & also, because I can use anti-ringing and anti-bloating, as needed. And yes, I know, less is more, when it comes to sharpeners. I can use NGU Sharp Direct x4, so medium, might be okay, instead of high. Also, experimenting, with 1440p, instead of 2160p. I can use very high settings (anything, everything), at 1440p & images, look a little better, to me, than 2160p (significantly lower madVR settings). Thank you for your input. Almost there!

[madshi]

So guys, sorry to say, but no new madVR build this week. I tried, but didn't manage to complete a new build...

To wet your appetite, here's what I'm working on: I'm trying to fuse "NGU Sharp" + "Reduce Compression Artifacts" into one. Well, not really, but partially. Uhm. Ok. So if you don't use NGU Sharp, everything will stay as it is now. But if you activate both NGU Sharp + Reduce Compression Artifacts at the same time, basically RCA will come for free. So you get a nice speed reward for combining NGU Sharp + RCA. However, you can still use both separately, of course.

First results seem promising. I'll need your tests to check if there's any quality loss when using the fused algorithm instead of doing both separately. I hope there'll be no quality loss, but I'll let you test it, once it's ready to go. Maybe ready next weekend, maybe the weekend after that, but probably no later then 14 days from now.

For a start, it will only be available for NGU Sharp, maybe I'll also offer it for NGU Standard/Soft/AA, don't know yet. It's all a lot of work, so...

---------------------

In the meanwhile, I'd like your FEEDBACK on the following 3 questions:

1) Do we need "double again/twice" for NGU Sharp/Standard/Soft? Or is everybody happy with "direct quadruple"? I'd really like to remove the "double again" option, because it's dramatically bad for performance and IMHO also looks worse. But here's your chance to object, before I remove it.

2) Anybody still using NNEDI3 for chroma upscaling? If so, is there a good reason why you prefer it over NGU AA? I'd really like to remove NNEDI3 chroma upscaling, because it's simply worse than NGU AA, IMHO, and slower, so removing the option should be all good, and make the settings dialog simpler.

3) Anybody still using NNEDI3 for image upscaling? If so, is there a good reason why you prefer it over NGU AA? I've recently seen an image comparison where NGU AA Medium was significantly better than NNEDI3-128 (sharper, more detailed *and* more natural), so I really wonder if there's any need for NNEDI3, anymore?

If I could remove NNEDI3 completely, that would be a relieve, because I could completely get rid of OpenCL, with all the driver issues etc.

[ashlar42]

My opinion is that NNEDI3 can go the way of the dodo.

[3x3cUt0r]

I stopped using NNEDI3, so i vote yes, go ahead and remove it.