madVR - high quality video renderer (GPU assisted)

[Anima123]

Originally Posted by Shiandow

The aliasing you're seeing seems to be because they've upscaled the images using point sampling. Presumably to make it easier to see differences (otherwise you'd have to look at a really small image). The other algorithms are simply too blurry to make the aliasing noticeable.

Shiandow, it make me curious if it helps with the overall quality once a good downscaler was adopted into SuperRes.

[madshi]

Originally Posted by aufkrawall

Thanks for the performance optimizations.
With GTX 980, now Adaptive Sharpen and sharpen edges roughly seem to be equally fast. Adaptive Sharpen is still slightly faster, I get a little lower GPU usage and rendering time with it. So there seems to be a difference to Kepler.

Yep, but roughly equally fast is still different to Intel & AMD, where sharpen edges is still 50% slower than AdaptiveSharpen.

Originally Posted by aufkrawall

Regarding SuperRes again:
I read here that SR would produce an overly sharp picture to some when doing 1080p -> WQHD. However, I can hardly believe this. I made a sharpness comparison with a game screenshot rendered in WQHD, and despite of SR with strength of 3 on top of Jinc AR scaling, it's still noticeably softer than the original (and much closer to ground truth than just Jinc AR without SR)

Yes, I agree. The only thing that sometimes bothers me is that SuperRes makes the edges sharper but leaves textures untouched. That can make the image look comic like sometimes. But I hope adding a bit of "enhance detail" in the latest madVR build might help there. (Should probably only be used on good sources, though.)

Originally Posted by leeperry

Well, 2X SR improves edges clarity so how about another extra pass on top of it for instance? Most users here don't care about increased GPU load if PQ improves.

You can't run SuperRes without upscaling first. So this would have to be done by using SuperSampling. I'm not sure if doing that would be worth the extra performance hit. I'd rather try the 4 new sharpening algos instead.

Originally Posted by leeperry

Thanks a lot for .17 but apparently 2X is still disabled for SR only? Based on my screenshots comparison, you seemed to agree that 2X SR improves PQ so why not? I got no use for 1X only SR

Do you have the power to make my days have 30 hours instead of 24? If not, then please try to have a bit of patience.

Originally Posted by KoD

The CPU is a i7 4790K (HD4600 iGPU), and the discrete GPU is a GTX 980. OS is Win 8.1 64 bit.

What I looked for was to have QS doing the decoding and deinterlacing, and having the GTX 980 doing the resizing. But decoding + deinterlacing on QS was not working properly when using the dGPU for output (as soon as I skipped in the file, playback started grinding), so I tried to debug what was going on, and connected the TV directly to the motherboard HDMI port. That's how I encountered all of the above.

Ok. The combination of Intel GPU with newer OSs and DX11 is known to sometimes cause this frame jumping in madVR. Unfortunately I can't reproduce it on my development PC with my HD4000. I hope this will change once I upgrade to win10, but I don't know for sure yet.

I'll try to reproduce the QS specific problems later, when I find some time.

Originally Posted by KoD

is it certain that the bat file really resets all the settings? And that it does not have to be run as administrator?

It works for me, but there's no guarantee. You can double check yourself: The bat file should delete the "settings.bin" file in the madVR folder (if there is one), and it should delete the settings in "HKCU\Software\madshi\madVR". So if you want to be certain, double check that the bat file does all that correctly.

Originally Posted by KoD

The frame jumping back and forth is still there though, as are the issues when using the GTX 980 for output and QS for decoding and deinterlacing.

A quick guess on my part: I once had a user reporting problems and after some digging we found out that uploading to the GPU was extremely slow for him, when uploading to RGB textures, but was much faster when uploading to NV12 surfaces. In order to test this, please activate DXVA scaling and zoom into the image a bit, so that DXVA scaling actually gets active. This forces madVR to upload the frames as NV12. When doing that, do the QS problems go away? The subtitle positions are probably a completely different topic.

Originally Posted by ace960

I uploaded zipped log and the freeze reports to http://ulozto.net/xPtByEpB/madvr-zip.
note: I generated the files with mpc-hc 1.7.10.5 and madvr 0.89.17.

Thanks, will have a look at this later...

Originally Posted by Sp00kyFox

can it be that the nnedi3 implementation in madvr doesn't correct the center shift that the algorithm introduces? if I pause a frame and activate luma doubling with nnedi3 (neuron number doesn't matter) I can cleary see that the image content slightly moves.

Fixing the center shift costs performance and quality, and you would never notice the shift when you don't compare to a different algorithm, so that's why madVR does *not* fix the center shift. The situation changes when you use SuperRes because SuperRes requires the center to be unshifted. So in that case madVR does correct the center shift. Also, when using quadrupling, madVR uses NNEDI3 in such a way that quadrupling shifts the center into the opposite direction, so the overall center shift after quadrupling is still there, but smaller than when just using doubling.

Originally Posted by Sp00kyFox

played around with the new version and some settings. I don't really get the negative feedback for the biliteral chroma upscaling. I tried it with some real world movies and TV shows and it does a great job. using the luma information as a guidance seems to me is the best way to do it, so yeah.. it's a good chroma scaler in my eyes.

Originally Posted by har3inger

Bilateral scaling can be really really good, but also runs into some exceptions that make it look like trash. Any time you have solid red shapes on solid black, Bilateral has zero luma information to draw on and gets ugly at all the edges (happens on logo/studio sequences before films, rare scenes from certain movies, anime) Furthermore, it sometimes "fills to edges" inappropriately, giving a dramatically different image (chroma bleed) from what ground truth is.

It's just not consistent enough to be left on for all video content. 80-90% of the time it's great, the other 10% it looks so bad I think something's corrupted or broken.

If it were just the problems with a flat luma channel, I might invest the work to detect that situation and switch to a different algo in that case. But the inappropriate "edge filling" which sometimes occurs is a problem that I don't know how to solve. Which is why I haven't tried to improve Bilateral chroma upscaling yet. Maybe Shiandow has some ideas there? I'm not sure if anything can be done there, though. Maybe it's not possible to solve the inappropriate "edge filling" without removing all the benefits Bilateral has in other situations.

Originally Posted by Shiandow

That looks interesting, bit worried about the 'patent pending' though.

Yep, noticed that, too. Was planning to wait until they have code available (website says "soon") and then to email them to ask about the license/patent situation.

Originally Posted by Shiandow

The aliasing you're seeing seems to be because they've upscaled the images using point sampling. Presumably to make it easier to see differences (otherwise you'd have to look at a really small image). The other algorithms are simply too blurry to make the aliasing noticeable.

Oh, that could be an explanation. In any case, it's worth a try, if the license situation allows it.

Originally Posted by ryrynz

Looks a lot more accurate to the source now, I'm seeing less haloing and overall better picture quality.

Glad to hear that!

So how does it compare to warpsharp now?

Originally Posted by ryrynz

FWIW, applying thinning after upscaling at a setting of 2.0 seems to look basically identical to doing it @ 1.0 every 2x upscale, you save some rendering time doing is this way also.

Yes, I just wish the other algos had the same properties, but they don't. Which makes it hard for me to streamline all the options into something less complex.

Originally Posted by ryrynz

Madshi, do you think crispen edges goes a little too far? > 2.0 starts to look quite bad. > 3.0 makes me wonder why it's even there.

I haven't really done full comparisons myself yet. crispen edges has a somewhat different look to it which I generally like, but with high settings it does start to fall apart. Maybe I can improve it in the future, I don't know...

[ryrynz]

Originally Posted by madshi

Glad to hear that!
So how does it compare to warpsharp now?

It's looking like a pretty good alternative, there's some line warping properties from the last version that look closer to the original source so I think it could still use a little tweaking but this was a step in the right direction. I'll send you some screenshots etc.

[hubblec4]

Hi madshi

I have posted in the LAV thread an issue.

With the new madVR0.89.17 play the DVD Aliens(test dvd without the big-vobs) with 4:3 format (menu and video), but it should be 16:9. With EVR it plays fine with 16:9 format.

[leeperry]

Originally Posted by madshi

The only thing that sometimes bothers me is that SuperRes makes the edges sharper but leaves textures untouched. That can make the image look comic like sometimes. But I hope adding a bit of "enhance detail" in the latest madVR build might help there.

That's especially what I like about it I think as edges are sharper but not the picture itself, that does wonderful things to fast motion but yes you're right as it gives a "comic-like" look indeed but it's sooo smooth and it's not like NNEDI3 didn't look artificial on its own so at some point you have to find an acceptable compromise between sharpness and naturalness I think and as much as the thin look of the latter kinda bothers me SR appears right up my alley

Originally Posted by madshi

You can't run SuperRes without upscaling first. So this would have to be done by using SuperSampling. I'm not sure if doing that would be worth the extra performance hit.

Ouh, wet dreaming of SS SR now and I need an excuse to buy a new GPU anyway ^^

Originally Posted by madshi

Do you have the power to make my days have 30 hours instead of 24? If not, then please try to have a bit of patience.

Yeah so sorry I figured afterwards that a checkbox on the end-user side could mean a lot of changes in the back with those four new options and all

Originally Posted by madshi

Yep, noticed that, too. Was planning to wait until they have code available (website says "soon") and then to email them to ask about the license/patent situation.

Ouh, this would be most promising in mVR if any possible.

[Thunderbolt8]

the perceptual downscaling image comparisons look interesting!

[Shiandow]

Originally Posted by Anima123

Shiandow, it make me curious if it helps with the overall quality once a good downscaler was adopted into SuperRes.

It's kind of difficult to predict what effect a different downscaler will have, especially a non-linear one. It might also be the wrong way to go about it, I think there's a better way to use their ideas in SuperRes.

Basically, the basis of their algorithm is similar to SuperRes; they generates a resized image which is in some way 'similar' to the input. In their case they generate a smaller image, instead of a bigger image (which means that there's a unique solution). But they also use a different way to measure similarity. It would make more sense to see if SuperRes could use their method to measure similarity, which would actually remove the need to use a downscaling algorithm at all.

Coincidentally, if you use SuperRes's method for measuring similarity to downscale an image you'd get the 'box' downscaling algorithm, which isn't that bad, but Mitchell-Nertavalli is easily better.

[XMonarchY]

Will SuperRes for Chroma Upcaling be updated any time soon? Last I heard SuperRes was updated only for Upscaling Refinement, not Chroma Upscaling.

[avinab]

@madashi in the latest build nnedi3 is not working.even if nnedi3 is selected in chroma upscaling or image doubling,it is showing JINC3.
But super-xbr is working.please look into it.

[huhn]

what is your OS, GPU and driver version?

[KoD]

Originally Posted by madshi

a quick guess on my part: I once had a user reporting problems and after some digging we found out that uploading to the gpu was extremely slow for him, when uploading to rgb textures, but was much faster when uploading to nv12 surfaces. In order to test this, please activate dxva scaling and zoom into the image a bit, so that dxva scaling actually gets active. This forces madvr to upload the frames as nv12. When doing that, do the qs problems go away? The subtitle positions are probably a completely different topic.

Hi madshi, you were right, the settings are reset, both the file and the registry entries get deleted, so that's not the reason why the subtitle positioning got fixed. In fact, after doing what you asked me, I can now say with certainty that the subtitle positioning issue happens only when DXVA is used for the scaling of the luma and of the chroma. If DXVA is not used for scaling, then the subtitles are at their proper location in the image. It's as if DXVA is used for scaling, then the subtitles are kept at their original position in the frame in absolute values, which is now somewhere in the middle-left part of the image. And yes, using software decoding + DXVA scaling also shows this issue; it's not related to QuickSync.

When "Enable Adaptive HW deinterlacing = disabled" for QuickSync, then the frames were sent interlaced to the renderer and since "use DXVA chroma upscaling when doing DXVA deinterlacing = enabled" after a madVR reset settings, madVR was using DXVA to do the scaling and that's why the subtitle positioning issue was always happening in this configuration.

These are my results when using QuickSync for decoding, but the resizing and video output are performed by the GTX 980:

1. With QuickSync "Enable Adaptive HW deinterlacing = disabled", DXVA is always used for deinterlacing, and these are the playback results:
- DX9 + "use DXVA chroma upscaling when doing DXVA deinterlacing = disabled" + luma Bicubic + chroma Bicubic -> luma Bicubic + chroma Bicubic are used, and there are no issues with switching fullscreen and back, nor with skipping in the file, nor with subtitle positioning
- DX9 + "use DXVA chroma upscaling when doing DXVA deinterlacing = enabled" + luma Bicubic + chroma Bicubic -> luma DXVA + chroma DXVA are used, and there are issues only with subtitle positioning; I find it strange that luma DXVA is used in this configuration - is this normal?

- DX11 + "use DXVA chroma upscaling when doing DXVA deinterlacing = disabled" + luma Bicubic + chroma Bicubic -> luma Bicubic + chroma Bicubic are used, and just like for DX9 there are no issues whatsoever
- DX11 + "use DXVA chroma upscaling when doing DXVA deinterlacing = enabled" + luma Bicubic + chroma Bicubic -> just like for DX9, luma DXVA + chroma DXVA are used, and there are issues only with subtitle positioning

2. With QuickSync "Enable Adaptive HW deinterlacing = enabled", QuickSync is doing the deinterlacing, and these are the playback results:
- DX9 + "use DXVA chroma upscaling when doing DXVA deinterlacing = disabled" + luma Bicubic + chroma Bicubic -> luma Bicubic + chroma Bicubic are used, and if you just wait after opening the file, the sound plays but no image, the madVR buffers stay at 1; if you get impatient and start skipping in the file, past the scrolling credits part towards the talking heads segment, you might get lucky and get unstuck and playback is smooth as if there's nothing wrong; skipping again might prove lucky, or you might get stuck again on some frame freeze, and you might even get in a state where the sound gets moving but the image is like refreshing one frame every few seconds or so, or not recovering at all, but you can close the playback in the player without issues (so the player does not freeze); no issues with subtitle positioning
- DX9 + "use DXVA chroma upscaling when doing DXVA deinterlacing = enabled" + luma Bicubic + chroma Bicubic -> same as above

- DX11 + "use DXVA chroma upscaling when doing DXVA deinterlacing = disabled" + luma Bicubic + chroma Bicubic -> same as above
- DX11 + "use DXVA chroma upscaling when doing DXVA deinterlacing = enabled" + luma Bicubic + chroma Bicubic -> same as above

- DX9 + "use DXVA chroma upscaling when doing DXVA deinterlacing = disabled" + luma DXVA + chroma Bicubic -> luma DXVA + chroma DXVA are used (again, is this normal?), and there are all the issues above + the issue with subtitle positioning
- DX11 + "use DXVA chroma upscaling when doing DXVA deinterlacing = disabled" + luma DXVA + chroma Bicubic -> same as for DX9


I think it's fair to also say what the results are with EVR CP:
- with QuickSync "Enable Adaptive HW deinterlacing = disabled" -> no issues whatsoever
- with QuickSync "Enable Adaptive HW deinterlacing = enabled" -> I get the same issues as with madVR (in this configuration with the GTX 980 for output; there were no issues when the iGPU was used for output!)

Originally Posted by hubblec4

Hi madshi

I have posted in the LAV thread an issue.

With the new madVR0.89.17 play the DVD Aliens(test dvd without the big-vobs) with 4:3 format (menu and video), but it should be 16:9. With EVR it plays fine with 16:9 format.

I confirm there is an AR issue for DVD playback myself, with a DVD of my own. But this is not new, it's here since at least 0.89.15 (did not test further back).

The copyright notice is 4:3 and is displayed properly, the studio intro is also 4:3 and no issues, and then the menu appears, which is a static image, and should be displayed as 16:9, but madVR displays it as 4:3. In EVR it's 16:9. No sample from me, but I think there needs to be the two videos included as well to reproduce this, and I don't know how to cut the vobs with the intro videos in.

[Knight77]

Warner306 in relation to the new scalers which are the settings you suggest now for 720p to 1080p and 1080p to 1080p?

[Warner306]

Originally Posted by Knight77

Warner306 in relation to the new scalers which are the settings you suggest now for 720p to 1080p and 1080p to 1080p?

I gave a response to this in the Kodi forums. These sharpeners are new to me, so I would ask others instead of relying on settings posted months ago. I will continue to change my preference into the future and find it is harder to recommend specific settings with these shaders.

So I don't know what is best.

I used to use SuperRes alone, but might try adding enhance detail in the future.

[Knight77]

Originally Posted by Warner306

I gave a response to this in the Kodi forums. These sharpeners are new to me, so I would ask others instead of relying on settings posted months ago. I will continue to change my preference into the future and find it is harder to recommend specific settings with these shaders.

So I don't know what is best.

I used to use SuperRes alone, but might try adding enhance detail in the future.

Thanks Warner306 and please don't get my question wrong or feel bothered about it: thanks you in particular I tried and get focused on some very good settings so to me you are like a trustable source. Obviously I'll check other options too, but by now looks like sharpen edges (0.2) and enhance detail (0.5) are a good starting point.

[digitech]

@Madshi can you pls give us a proper MadVR profile for 1.- 1080p, 2.- 720p and 3.- below 720p (SD) profile for all kinds of video files.., i have been trying lots of combinations but always thing get mixed up, sd files playing in hd and viceversa

[JarrettH]

I just updated to the latest from 88.21

Zoom control...a little help?...

Do I need to use this if I'm just playing video files on a computer and watching via computer monitor? I want to maintain the aspect throughout the movie.

Using touch window from inside in MPC

[Ver Greeneyes]

If that new SSIM-based downscaling algorithm can be used, would be interesting to see how it affects SuperRes (which relies on downscaling, right?).

[Warner306]

Originally Posted by Knight77

Thanks Warner306 and please don't get my question wrong or feel bothered about it: thanks you in particular I tried and get focused on some very good settings so to me you are like a trustable source. Obviously I'll check other options too, but by now looks like sharpen edges (0.2) and enhance detail (0.5) are a good starting point.

Low values with image enhancements are important. Higher values should only be used when upscaling by using sharpening applied from upscaling refinement.

[Warner306]

Originally Posted by JarrettH

I just updated to the latest from 88.21

Zoom control...a little help?...

Do I need to use this if I'm just playing video files on a computer and watching via computer monitor? I want to maintain the aspect throughout the movie.

Using touch window from inside in MPC

These settings are primarily aimed at users with Constant Image Height (CIH) projection set-ups, so it would be unlikely you would have to change anything if you didn't use these settings before.

[huhn]

these settings are very useful for 4/3 sources with 16/9 content on a 16/9 screen. just as an example.

screen config is for CIH projectors