madVR - high quality video renderer (GPU assisted)

[madshi]

Originally Posted by Anime Viewer

Given both the performance difference and the small difference I see between the two pictures Anima posted I vote for the higher strength with less passes. In my testing I see very little ( on my sources) in the low, medium, high, and ultra settings you noted before. I could see myself using the low or medium settings, but don't think I'd every use the high or ultra settings given how much of a performance hit they are with not enough significant image improvement. (Pretty much the same reason why I choose to use Super-xBR instead of NNEDI3).

Edit: I like the way it combines with image enhancements (not the upscaling refinement version) of Adaptive Sharpening (not sure on the strength...0.3 or 0.5 maybe), but I'll hold back on further commenting on that until you want to talk about combining effects.

Ok, thanks.

Originally Posted by Eyldebrandt

Well, SuperRes is, again, quite better than the previous build.
But it creates a huge amount of aliasing.

Here is a sample.
that's the Gone Girl Blu-ray (which is what we can call a very high quality content) upscaled to 3440x1440 résolution in 2.39 ratio.
Chroma is NNEDI3 64, image Doubling is NNEDI3 64, downscaling is CR AR LL.

Look at the police car, the framing of the rear door.

Could I have a small sample, please? Maybe a PNG of the original Blu-Ray frame might already be good enough, if enlarging the PNG produces the same aliasing artifacts (I think it will).

Originally Posted by Ver Greeneyes

Like this? (top is processed, bottom is original) Of course there are some problem areas, and there are way more examples where this neural network generates extremely weird things, but I thought its performance on this particular image was pretty incredible.

Euwh, weird!!

Originally Posted by Eyldebrandt

The rear window of the car.
The sample with AS @ 1.5 is ugly as hell. Wax on trees, loss details on fines structures, luma pixels oversharpened.

Which is the opposite of what SuperRes does. Which is why I like SuperRes. Although it currently still has its problems. But maybe we can solve them.

Originally Posted by Eyldebrandt

To my opinion, i have more to gives to Mathias with this kind of sample, rarer than 480p to 1080p content. I can post hundred of samples

If you can find more samples where you see problems with SuperRes, by all means, hit me! The more samples we have, the better we can work on tuning to SuperRes to get rid of the artifacts.

Originally Posted by tFWo

Rings of Saturn are gone. Rings of Uranus remain. (madshi don't hate me, mods please don't ban me )

AR filter helps a lot. Alternative color space (acs) is just bad. I prefer algo 1.

But tonight I found out that s-xbr100+ and SuperRes don't like each other. Something is seriously wrong when you use s-xbr with sharpness 100 or higher and SR.

[...] (This is what I called dirt in the last post. SR actually blurs/darkens the numbers on the plate.)

SuperRes tries to make sure that the upscaled image is a "correct" interpretation of the sources. Obviously in this case SuperRes believes that the numbers on the plate are too bright. And if you compare the super-xbr numbers to the NNEDI3 numbers prior to SuperRes, they *are* brighter in the super-xbr image. That's why SuperRes tries to tone them down again. Maybe it doesn't do a very good job at that, but generally that's the concept of SuperRes.

Originally Posted by aufkrawall

With Hawaii GPU, image doubling with 64 neurons can be done for 1080p30 via DirectCompute

And probably with OpenCL, too. Unless you're stuck with PCIe 1.x or something. Btw, it's not DirectCompute, but PS5.0, I think.

Originally Posted by MS-DOS

Much less ringing and probably other artifacts. The difference is bigger when more passes are used.

Cool, thanks.

Originally Posted by leeperry

I've had time to read the few pages I missed while I was away and it would appear that the "majority of users" you would be mentioning is actually two guys in this thread, maybe three.

I think it was four, plus my own strong preference, against only one clear and reliable vote (yours) for HQ off. Yeah, not a lot of votes, but I have to work with what I'm given. Of course I could just tweak the algos myself, based on my own preferences. But just to be safe, I'm asking for feedback. And if the majority vote of the feedback happens to agree with my own eyes, then I think things are pretty clear.

Originally Posted by leeperry

I rest my case that I'm quite sure nobody uses the "don't use linear light for dithering" debug option anymore and I would happily trade it for "don't use HQ downscaling for SuperRes" if any possible please.

The "trade quality for performance" options are there to trade quality for performance. What you're asking for has nothing to do with performance. So the option you want would not belong into the "trade quality for performance" section.

Originally Posted by Schwartz

True. I've been digging into the subject a bit since yesterday and benchmarked some scalers. I ended up just running Mitchell. I do a lot of picture downsizing with Mitchell and it turns out that it's pretty damn good for video upsizing as well. Probably nothing for the 'sharper' crowd but I find it's a nice compromise.

I've never really seen NNEDI3 ring, but then again I don't see sxbr 50-75 ring all that much either. But all of my image doubling happens to SD content where it can 'do some work', a bit of smoothing is usually desired. Pixel-perfect encodes I leave alone.

I second the user who said he'd enjoy more sxbr options. 60-65 might be the sweet spot for me. Maybe an input field?

An input field doesn't fit into the current settings dialog, at least not without a redesign. Maybe I'll add finer options between 50 and 100, but I'm really tired of changing the settings dialog all over again all the time, so I'll probably not do that soon.

Originally Posted by SithUK

I've given up on using a 3D LUT for now. I've tried turning the LAV filter decoding to None (instead of DXVA native), and using the Bilinear chrome upscaler. I tried ticking all boxes in the "trade quality for performance" section. Unfortunately my average render speed sometimes exceeds 50ms on busy action scenes, with 40ms being the minimum for a 25fps film.

It looks like a 5 year old laptop (i5 2.4Ghz/4Gb/HD 4500) is too old tech to manage a 3D LUT.

That's too bad.

[aufkrawall]

Originally Posted by madshi

And probably with OpenCL, too. Unless you're stuck with PCIe 1.x or something. Btw, it's not DirectCompute, but PS5.0, I think.

Thanks for that clarification. Can results of DirectCompute even be shared with D3D9?
Problem is the extreme divergent results of copyback performance, some users have far less than half performance of PCIe 3.0 with 2.0, what shouldn't be.
I couldn't use NNEDI3 at all when I had a R9 290X due to this. Since I really like a lot NNEDI3, I'm somewhat tied to Nvidia when I want to use it in madVR.

[Anime Viewer]

Originally Posted by leeperry

I didn't realize it had any effect on performance, I think you would be surprised how much banding this option adds to any gray ramp.......to the point that enabling debanding in mVR and checking this option would be utterly counterproductive IMHO. It might very well also affect your judgment on other post-processing settings.

I'm not sure why, but the rendering hit I used to encounter with it unchecked is no longer occurring (perhaps a bug was fixed, or I had a different setting enabled at the time). I'm not noticing any improvement in banding with it unchecked, but maybe the source files I'm testing never had much banding to begin with...

Originally Posted by QBhd

Not sure what the rest of you are looking at (including the one who posted this) but the last image is stunning. I have been reading a lot of the posts lately, and it seems everyone wants a soft image. Soft=blurry.... I would love a crisp and clear image (with no artifacts, ringing or aliasing of course) And I must say this last image is just that.
QB

Those are my impressions as well. The last image that was clearer and crisper than the others looked better to me too. Some of the people may be having issue with the strength setting that was apparently used with the last image, and they may view it as overkill since its being labeled at strength 1.5. If the same image was posted with a strength listing of .5 or 1 it might not get as much complaint. Perhaps the same image with a setting of .5, 1, or something in-between would give similar clarity. Regardless I see Adaptive Sharpening as beneficial.

[Ver Greeneyes]

Originally Posted by madshi

Euwh, weird!!

Haha, I know. It's kind of fascinating what specialized neural networks can do when given free reign though. I'm sure we'll get to human levels of recognition one day (and it'll stop adding weird eyes all over the place). For a bit more context, there was a google research blog post about their neutral networks here (and this follow-up for the code.

[madshi]

Originally Posted by aufkrawall

Thanks for that clarification. Can results of DirectCompute even be shared with D3D9?

Yes.

Originally Posted by Anime Viewer

Those are my impressions as well. The last image that was clearer and crisper than the others looked better to me too. Some of the people may be having issue with the strength setting that was apparently used with the last image, and they may view it as overkill since its being labeled at strength 1.5. If the same image was posted with a strength listing of .5 or 1 it might not get as much complaint. Perhaps the same image with a setting of .5, 1, or something in-between would give similar clarity. Regardless I see Adaptive Sharpening as beneficial.

The house, roof and many edges look better in that image, but everything nature (grass, bushes) in front of the house looks absolutely terrible. Looks like a water color / oil painting instead of a photo.

Originally Posted by Ver Greeneyes

Haha, I know. It's kind of fascinating what specialized neural networks can do when given free reign though. I'm sure we'll get to human levels of recognition one day (and it'll stop adding weird eyes all over the place).

Maybe, but NNEDI3 will be ultra fast compared to such kind of algorithms.

[leeperry]

Originally Posted by madshi

The "trade quality for performance" options are there to trade quality for performance. What you're asking for has nothing to do with performance. So the option you want would not belong into the "trade quality for performance" section.

I would also be cool with creating a folder in mVR's folder such as "SuperRes HQ OFF" or something if that's not too much trouble please. You allowed this kind of kludge in the past to force PC levels and a few other things, not the end of the world apparently as it's not adding clutter to the GUI and won't raise questions from newbies either?

Originally Posted by madshi

I think it was four, plus my own strong preference, against only one clear and reliable vote (yours) for HQ off. Yeah, not a lot of votes, but I have to work with what I'm given. Of course I could just tweak the algos myself, based on my own preferences. But just to be safe, I'm asking for feedback. And if the majority vote of the feedback happens to agree with my own eyes, then I think things are pretty clear.

Duly noted, still the middle of summer in the US and EU is more than likely not the best time to take decisions of such paramount importance(to me at least ) when it comes to an option that adds such a nasty veil to the picture IMHO. Talk about one step forward, two steps backwards

That's usually how they do it in my country when they wanna pass laws nobody would agree on, they pass them at night during summer, case closed

Either way, HQ is a complete showstopper for me so if .15 is seriously the last version providing a switch for it then I'm more than likely through with updates till mVR supports resolution-changing PS scripts so I could go nuts and mess with ALL the SR/sxbr settings......any very rough ETA for this please? Apart from "not soon"?

Or maybe I'll get bored of SR at some point as I do realize that it's extremely sharp and I've never been a fan of sharpening to begin with, but its effect on 24p motion blur with HQ off is pure magic to me in combination with monostatic ED2@8bit

Originally Posted by Anime Viewer

I'm not sure why, but the rendering hit I used to encounter with it unchecked is no longer occurring

If you were using it in combination with any kind of ED, run a gray ramp and you'll both be be horrified and happy to have it unchecked now

[iSunrise]

Originally Posted by madshi

Thanks. Maybe I can add an AR filter to FineSharp in the future.

Extremely interesting idea. Would love to try it out.

But now to my follow-up regarding Finesharp vs. AdaptiveSharpen, which I now had a bit more time to dig deeper, after I did some more tests with different samples.

On a more general note, from the samples I looked at, Finesharp is a lot more uniform with it's sharpening, which especially benefits real-world content, since it doesn't selectively sharpen like AdaptiveSharpen. Which means that, if you want to sharpen the whole image for basically the same value, Finesharp is definitely the way to go. It also looks way more natural due to this behaviour.

AdaptiveSharpen is definitely not a replacement for Finesharp, but more like an additional sharpening effect, which can be beneficial on some content, but (see below) you have to be very careful with higher values.

Also, there are huge (and I mean HUGE) differences when you compare both algorithms using image enhancements and upsaling refinement, more below.

Here's an example, why I would completely stay away from AdaptiveSharpen on the image enhancements tab, no matter what values you set.

Image enhancements - Finesharp 1.0



Image enhancements - AdaptiveSharpen 0.3



I think I don't need to explain the results, they speak for themselves. Finesharp can be used with very high values (I still would recommend not to go much higher than 1.0), while AdaptiveSharpen even beginning from values like 0.1 already rings and 0.3 already rings like crazy and it also adds some very strange look to the image. On game recordings, anime and real-world content, this should be avoided at all cost. For image enhancements, Finesharp is the clear winner on all the content I have looked at and tested. With an additional AR algorithm applied, Finesharp would be very close to perfect.

The situation with the higher ringing that AdaptiveSharpen shows when used in the image enhancement tab, completely reverses when you use AdaptiveSharpen on the upscaling refinement tab. Using the same values from above, AdaptiveSharpen rings a lot less than FineSharp in this tab, however, it still retains that oil painting look that worsens the higher you go with the values.

So, for me, personally, I would stay away from AdaptiveSharpen in the image enhancements tab altogether and only use it in the upscaling refinements tab and also preferably with VERY low values (like 0.1 or maybe even lower).

[aufkrawall]

Mostly agreed.
I really like the effect of AS as an UR with low values (e.g. 0.2) for NNEDI3. Everything's still well reconstructed, but that softness disappears for a huge amount. Especially if sharpened after each 2x upscaling step. Thanks to its adaptive nature, it doesn't oversharp already sharpened areas and there is very little aliasing introduced.
Maybe this works a bit better with naturalistic content than with artificial one due to that thickening of black lines. But without sharpening, those lines are very soft with NNEDI3 though.

I still get a ton of presentation glitches on my 120Hz monitor, even if I set it to 60Hz. The same exact settings work perfectly fine on my TV @ 23Hz. I don't get it. Average rendering times are identical on both TV and monitor - about 27-28ms (same as max stats). I can reduce presentation glitches on 120Hz monitor by enabling "present a frame for every VSync", but then still occur (1 every 10 seconds or so). However, enabling this option during playback on my TV @ 23Hz creates dropped frames and presentation glitches. Setting "present several frames in advance" to 1 also helped to reduce presentation glitches (1 every 15-30 seconds). Using less GPU-taxing settings made no difference.

The second issue I get on my TV is that at times when I pause playback and resume it, I get severe playback stuttering, which I can fix by exiting fullscreen exclusive mode to window mode and then going back to fullscreen exclusive mode. That does not happen on my 120Hz monitor though...

I reset madVR settings to default, uninstall it as admin, restarted, install madVR as admin, reset settings back to default, and applied settings I use. That did not help.

EDIT: Could these issues because I set "Maximum Pre-Rendered Frames" to 1 in NVidia Control Panel? I set it so because it reduces stuttering associated with using Borderless Window mode in PC games.

[nevcairiel]

High Refresh Rate screens can be a bit fiddly. You may need to increase the queue sizes quite a bit to accomodate for the higher refresh rate, that tends to help for me to reduce the glitches, and "present a frame every vsync" is pretty much mandatory.
Personally, its not like I see the glitches, it just piles up on the counter.

PS:
There will never be settings that work 100% perfect on every system, so having to change a setting here or there depending if you're playing 23.976 on a 120Hz screen, or having it play 1:1 on a 23p TV is not uncommon.

[kasper93]

@madshi: Something weird is going on. If I make MPC-HC window size equal to screen size madVR will go into exclusive mode and position window in center even though it weren't before. If I open context menu to disable exclusive mode it window go back to its original place. Also seekbar doesn't work in this "exclusive mode" probably it is not reported correctly.

To reproduce open any file in MPC-HC. And stretch the window to display size (not work area size). Basically this means to Make MPC-HC max size. To do that you need manual grab window borders and stretch it.

I will look into details tommorow. But it looks like madVR have "if (windowSize == displaySize) GoToRetardedExclusiveMode();" logic

Originally Posted by nevcairiel

High Refresh Rate screens can be a bit fiddly. You may need to increase the queue sizes quite a bit to accomodate for the higher refresh rate, that tends to help for me to reduce the glitches, and "present a frame every vsync" is pretty much mandatory.
Personally, its not like I see the glitches, it just piles up on the counter.

PS:
There will never be settings that work 100% perfect on every system, so having to change a setting here or there depending if you're playing 23.976 on a 120Hz screen, or having it play 1:1 on a 23p TV is not uncommon.

Erm, but enabling "present several frames in advance" and setting "how many video frames shall be presented in advance" to 1 helped a lot. Anything higher than 1 and I get more frequent presentation glitches.

[nevcairiel]

Originally Posted by XMonarchY

Erm, but enabling "present several frames in advance" and setting "how many video frames shall be presented in advance" to 1 helped a lot. Anything higher than 1 and I get more frequent presentation glitches.

Don't set pre-presented frames in the NVIDIA control panel then, it messes madVR up quite considerably.
Luckily the NVIDIA panel has per-app settings!

[Magik Mark]

Originally Posted by nevcairiel

Don't set pre-presented frames in the NVIDIA control panel then, it messes madVR up quite considerably.
Luckily the NVIDIA panel has per-app settings!

Thanks for this tip. Are there any other settings that would degrade the performance of madvr or any other renderers in the Nvidia control panel? Is it better to just tick "Let 3d application decide" option?

[huhn]

Originally Posted by Magik Mark

Thanks for this tip. Are there any other settings that would degrade the performance of madvr or any other renderers in the Nvidia control panel? Is it better to just tick "Let 3d application decide" option?

just leave them at default and you are fine.

[aufkrawall]

Vsync, prerenderlimit, power management mode and Optimus settings are the only driver-settings that should affect madVR. Everything should be kept at driver default (app-controlled) for the profile of your media player.
If you change settings globally, you have to set them back for each profile.

[Anime Viewer]

Originally Posted by leeperry

If you were using it in combination with any kind of ED, run a gray ramp and you'll both be be horrified and happy to have it unchecked now

I realized why I had it checked, and it wasn't because of render rate stats, but present stats. With "don't use linear light for dithering" checked and the video full screen my present rates are in the 3ms area, but with it unchecked and the video full screen they are in the 11ms area.

Edit: Windowed Full Screen (D3D11) runs with the 11ms presents, but if I run in D3D11 Exclusive (10 bit) its back down to 3ms again. Edit #2: Unchecking "present a frame for every VSync" also seems to be a way to bring it back down to 3ms while in (D3D11) Windowed Full Screen mode.

[6ari8]

Originally Posted by Anime Viewer

I realized why I had it checked, and it wasn't because of render rate stats, but present stats. With "don't use linear light for dithering" checked and the video full screen my present rates are in the 3ms area, but with it unchecked and the video full screen they are in the 11ms area.

Edit: Windowed Full Screen (D3D11) runs with the 11ms presents, but if I run in D3D11 Exclusive (10 bit) its back down to 3ms again. Edit #2: Unchecking "present a frame for every VSync" also seems to be a way to bring it back down to 3ms while in (D3D11) Windowed Full Screen mode.

Are you using an HDMI port? I think you can probably bring down the present stats to below 1ms by using the DisplayPort (though that depends on the laptop). You can see which port you have that's connected to the discreet GPU from Nvidia's control panel. Mine has the mDP only but there are different cases for laptops like these:

[Warner306]

Today, I attempted to embrace image sharpening as tool to improve an image rather than enhance its artifacts. I found the following settings beneficial in improving the quality of high-definition video:

1080p -> 1080p
Image Enhancements - FineSharp (strength 0.6)

720p -> 1080p
Upscaling Refinement - SuperRes (medium: strength=1.0; passes=1)

SuperRes is my favourite of the two. Its ringing is not obvious or bothersome compared to other sharpeners. I'm curious if the luma and chroma are doubled together or just the luma like the other shaders?

An AR filter for FineSharp would be great.

[Anime Viewer]

Originally Posted by 6ari8

Are you using an HDMI port? I think you can probably bring down the present stats to below 1ms by using the DisplayPort (though that depends on the laptop). You can see which port you have that's connected to the discreet GPU from Nvidia's control panel. Mine has the mDP only but there are different cases for laptops like these:

Yes, I have an HDMI port that I connect my TV to, but both the TV and the laptop screen are shown to go through the Intel GPU like the second image you posted. Here is the PhysX display diagram on my system: