madVR - high quality video renderer (GPU assisted)

[ryrynz]

Originally Posted by omarank

1. I prefer “detailed” over “clean”. The former may have more aliasing but from a distance it looks more pleasing than the latter. The difference is subtle though.

Yup, my thoughts too.

[mogli]

Since v0.90.5 my former problems with D3D11 *without* presenting a frame for every VSync are gone!
(Generally playback was fine before, but in rare cases out of nowhere a massive amount of stuttering and frame drops started. Restarting the movie fixed it, until it happened randomly the next time.)

However can it be that D3D11 *without* presenting a frame for every VSync is a tiny bit less smooth? That's my impression with slow pans and stuff. It's really minor though, but constant and regular nano stuttering.

[ryrynz]

Originally Posted by mogli

However can it be that D3D11 *without* presenting a frame for every VSync is a tiny bit less smooth? That's my impression with slow pans and stuff. It's really minor though, but constant and regular nano stuttering.

Watch some scrolling text (end credits) also if you're playing through a TV enable frame interpolation if you have it (Smooth Motion, Intelligent Frame creation etc) that can be unforgiving to presentation issues and might help you determine if one is better than the other for your setup.

[madshi]

Originally Posted by leeperry

I far prefer SSIM LL AR in .4 than in .6 as either of the two new options in the latter are too sharp and aliased for my other settings(both SR/RS/NNEDI3/FRC/etc). I understand that you want to remove as many options as possible but they are all very much dependent on each other, I mean I personally already find strength 1 of both SR's too sharp so in the end a very sharp picture through an über-sharp downscaler ends up aliased and oversharpened

Well, as I said, SSIM in .6+ does what it's supposed to do, while in .4 it was half broken. I'm not going to put a half broken algorithm back in. There's no sharpness knob for SSIM because it uses math to calculate the exact sharpness level needed to correctly reproduce the original image in lower resolution.

If you find the image too sharp, you could try using a softer upscaling algorithm, or to dial down the upscaling refinement options.

It would be possible to reduce the "effectiveness" of SSIM, but so far you're the only user who asks for that, and I'm not going to add option just for one user.

In any case, we're still working on the algorithm itself. The algorithm might be improved/changed a little bit, depending on user feedback. Even if we were to add some sort of sharpness control, it would be too early now, because it's not clear yet whether the algo is 100% done or not.

Originally Posted by Warner306

Could someone tell me the advantages of turning off 3D in the operating system during playback of 2D or 3D videos?

In 3D mode some GPUs stay in frame packed output mode. Which means the display stays in 3D mode all the time. 3D mode may have a different calibration (e.g. a different gamma curve to make up for the light loss caused by using active shutter). Also it may flicker a bit when using active shutter.

Another reason is that NVidia GPUs like to crash when staying in 3D mode, and then playing back 2D content through D3D11 with madVR.

Originally Posted by Codex0nz

I've been looking around for a while now but can't find a good answer in regards to madvr, I get quite bad jitter at times with playback even with smoothing enabled, tried reclock and all sorts.

How do you define "bad jitter"? Can you explain how it looks (or sounds) to you? Are you talking about audio or video? Does the madVR OSD (Ctrl+J) show any dropped, delayed, repeated frames or presentation glitches while that jitter appears? Does your refresh rate match your video frame rate?

Originally Posted by ryrynz

This occurs when I have SSIM downscale detailed only with quadrupling and when there's either nothing ticked in upscaling refinement or when a refinement is ticked but only applied when image scaling is complete, no issues when every 2x upscaling step is ticked (applies to any enhancement/combination)

Originally Posted by Anima123

Madshi, with 0.90.7, if double a 720p video with NNEDI3 32 then using SSIM for downscaler, the video will be 'greenized' if I didn't use 'double chroma resolution'.

Finally was able to reproduce this issue. Should be fixed in this test build:

http://madshi.net/madVR907b.rar

Originally Posted by ryrynz

I prefer the detail level.. but yes edges do seem a touch over aliased although clean isn't *that* much better.
madshi is it possible to slightly lower the sharpness in some way of SSIM so that detailed could be preferred?
I like the extra detail throughout the picture but I'm not a huge fan of the extra edge sharpness that comes with it.

Originally Posted by omarank

1. I prefer “detailed” over “clean”. The former may have more aliasing but from a distance it looks more pleasing than the latter. The difference is subtle though.

@omarank, can you comment on whether your preference mostly applies to edges or detailed image areas (if there is a difference)?

One thought Shiandow and I were having is that maybe it would make sense to use the "clean" method for edges and the "detailed" method for non-edge areas? That would make things more complicated and quite a bit slower, though.

Originally Posted by FreeFall

I'm having a similar problem.

Windows7 x64 Pro (GeForce GTX 980Ti, 356.04), Zoom Player Max 11.1, latest LAV nightly build and madVR 0.90.7.

Using the default d3d9 FSE mode, to reproduce play a DVD and then stop playback (s key), the last frame stays on screen until you exit the player (alt + x). If I enable the option use a seperate device for presentation (Vista and newer) playback works without any problems.

If I understand this correctly, the "only" problem is that in stopped state an old frame stays on screen? No other problem? While not being "nice" it doesn't sound like a very big problem? Just a cosmetical glitch?

Originally Posted by MokrySedeS

I have two very similar issues.
1. When I drag mpc-hc window from one display to another the playback pauses and I have to press Play/Pause button twice to resume. It was introduced in the latest .7 version. It happens every time in both D3D9 and D3D11.
2. When the player picks up another file from the playlist and madVR switches to different refresh rate - the playback pauses. What's interesting is that in D3D9 I only need to press Play/Pause once to resume, twice in D3D11. Also, it happens only from time to time, more often in D3D11 than D3D9. This was introduced earlier but I'm not sure in which version exactly.

I'll have to check if I can reproduce those issues here.

Originally Posted by omarank

2. 1.6 linear looks more like Gamma light to me. I prefer the look of 2.2 linear. For the past few years, Catmull Rom AR LL (with 2.2 gamma) has been considered as the best downscaling setting. I would suggest to not replace the 2.2 gamma with 1.6 gamma LL, so that one can always go back to CR AR 2.2LL.

Ok.

Originally Posted by omarank

I think you were using 32 bit float texture format specifically for the downscaling steps when Linear Light downscaling was used. Have you changed that to even less than 16 bit texture format?

No. The banding was caused by a simple logic bug.

Originally Posted by omarank

As you explained it earlier, Linear Light downscaling is done on RGB image so chroma upscaling must be done first to get the RGB image. Now in the latest build you are scaling luma and chroma separately (perhaps for large downscaling factors?). How is it done now?

After some tests I found that I could not see any difference in image quality if I convert Y' to Y directly and scale that, instead of going the long way through R'G'B' -> RGB -> YCbCr.

Originally Posted by omarank

Well, I seem to like the previous downscaling method especially for 4k to FHD when the chroma was upscaled to 4k and then image was downscaled to FHD. I think the supersampling effect on chroma affects the perception of depth in the images. I am not sure if this can be shown through screenshots. Anyway, I would prefer if you could add a trade quality option to restore the previous downscaling method.

I guess I can add such an option. Although when you don't see something like this in a screenshot (even with carefully selected test images/scenes) it can be a hint that it might be some sort of placebo effect. Anyway, adding a trade quality option shouldn't hurt.

Originally Posted by mogli

Since v0.90.5 my former problems with D3D11 *without* presenting a frame for every VSync are gone!
(Generally playback was fine before, but in rare cases out of nowhere a massive amount of stuttering and frame drops started. Restarting the movie fixed it, until it happened randomly the next time.)

However can it be that D3D11 *without* presenting a frame for every VSync is a tiny bit less smooth? That's my impression with slow pans and stuff. It's really minor though, but constant and regular nano stuttering.

If all works correctly, it should be just as smooth either way. If it isn't, then simply keep that option checked.

[madshi]

Here's a small comparison of SSIM "detailed" vs "clean" (using Linear Light 2.2 downscaling):

1) lighthouse:



2) asian letters:



I think the first image shows that "detailed" can become rather nasty in some situations. It not only produces more aliasing but can sometimes also create rather wild artifacts. However, in the 2nd image "detailed" looks clearly (less blurry) better than "clean".

So which method to use? What do you guys think?

[Manni]

Originally Posted by madshi

So which method to use? What do you guys think?

The artifacts in detailed are not really acceptable, so for me I'd vote for clear. I'd take slight blurriness over obvious occasional artifacts anytime.

[Thunderbolt8]

Originally Posted by ryrynz

I prefer the detail level.. but yes edges do seem a touch over aliased although clean isn't *that* much better.
madshi is it possible to slightly lower the sharpness in some way of SSIM so that detailed could be preferred?
I like the extra detail throughout the picture but I'm not a huge fan of the extra edge sharpness that comes with it.

perhaps it would be helpful to add the 0.90.4 way of SSIM leeperry talked about again so that we can compare that one to the two current ones.

edit: nvm, read that it was broken then. but as other said, perhaps ist possible to get rid of more aliasing.

[Thunderbolt8]

Originally Posted by madshi

So which method to use? What do you guys think?

Originally Posted by madshi

One thought Shiandow and I were having is that maybe it would make sense to use the "clean" method for edges and the "detailed" method for non-edge areas? That would make things more complicated and quite a bit slower, though.

Id like to see this tested.

aside from that: is it possible to estimate which of the two should do better with regular UHD BD Content scaled down to 1080p or regular Blu-ray 1080p scaled down to 720p? I guess that should be the most common and useful formats the algorithm will be/should be used for. so perhaps simply take that one which looks better with these sources.

[Shiandow]

Originally Posted by madshi

After some tests I found that I could not see any difference in image quality if I convert Y' to Y directly and scale that, instead of going the long way through R'G'B' -> RGB -> YCbCr.

I wonder if you're not going to run into problems when downscaling colours. Scaling colours in gamma light tends to cause some nasty problems if you're not careful. That said, I don't think this will become a problem until the final size is significantly smaller than the chroma size. And of course you don't really know the colours to begin with...

Sometimes I really wonder why people even bother with chroma sub-sampling.

[FreeFall]

If I understand this correctly, the "only" problem is that in stopped state an old frame stays on screen? No other problem? While not being "nice" it doesn't sound like a very big problem? Just a cosmetical glitch?

The problem is that the last frame stays on the screen and you can't use the player until you exit and reload it. The frame is stuck there on screen when switching to windowed mode or trying to interact with the player, using the keyboard commands play a chime sound through the speakers.

I tried re-setting madVR back to defaults and stopping playback worked (no stuck frame) but when I disabled all of the trade quality for performance options the problem returned again (strange). The problem only seems to happen with DVD playback using FSE mode and Zoom Player, windowed mode works without any problems.

[leeperry]

Originally Posted by madshi

Well, as I said, SSIM in .6+ does what it's supposed to do, while in .4 it was half broken. I'm not going to put a half broken algorithm back in. There's no sharpness knob for SSIM because it uses math to calculate the exact sharpness level needed to correctly reproduce the original image in lower resolution.

If you find the image too sharp, you could try using a softer upscaling algorithm, or to dial down the upscaling refinement options.

It would be possible to reduce the "effectiveness" of SSIM, but so far you're the only user who asks for that, and I'm not going to add option just for one user.

In any case, we're still working on the algorithm itself. The algorithm might be improved/changed a little bit, depending on user feedback. Even if we were to add some sort of sharpness control, it would be too early now, because it's not clear yet whether the algo is 100% done or not.

All I'm using is both SR's with a strength of 1 and sxbr25/NNEDI3, I got no other knob to mess around with and it's all VERY sharp to begin with ending up way oversharpened through the new SSIM

I see ppl using sxbr150 in here, possibly with SSIM on top oh wow that's what I call sharpness lol..are they wearing prescription glasses made of polycarbonate(that comes with an astonishingly low constringence that makes everything blurry), TN panels with a coarse and grainy antiglare coating or possibly massively misconverged videoprojectors? I wonder.

Originally Posted by madshi

"detailed" can become rather nasty in some situations. It not only produces more aliasing but can sometimes also create rather wild artifacts.

Exactly, completely unusable IME and "clean" isn't all that better due to its extreme sharpness

[omarank]

Originally Posted by madshi

@omarank, can you comment on whether your preference mostly applies to edges or detailed image areas (if there is a difference)?

One thought Shiandow and I were having is that maybe it would make sense to use the "clean" method for edges and the "detailed" method for non-edge areas? That would make things more complicated and quite a bit slower, though.

My preference applies to detailed image areas like skin texture, texture of clothing etc.

The idea of using the "clean" method for edges and the "detailed" method for non-edge areas sounds promising to me.

Originally Posted by madshi

I guess I can add such an option. Although when you don't see something like this in a screenshot (even with carefully selected test images/scenes) it can be a hint that it might be some sort of placebo effect. Anyway, adding a trade quality option shouldn't hurt.

Ok, I have got the screenshots. Please download from the below link a zipped folder with three files: the Original 4k image and its downscaled images, one with 0.90.7 build and the other with 0.90.4 build. Please see the edges of the melons.

http://www.mediafire.com/download/bs...ownscaling.rar

Now I believe you will be convinced to add a trade quality option to restore the previous Linear Light downscaling method.

Originally Posted by madshi

So which method to use? What do you guys think?

The artifacts with the “detailed” algo in the lighthouse image are sort of unacceptable. So I think more tweaking is required for SSIM. Perhaps using a combination of “clean” and “detailed”, as you thought, may lead in the right direction.

Edit 1: You may also want to check the downscaling results of the Lighthouse image using the previous Linear Light downscaling method where chroma was upscaled first.
Edit 2: Oh the image is in RGB itself, so please ignore Edit 1.

[ryrynz]

Originally Posted by madshi

Finally was able to reproduce this issue. Should be fixed in this test build: http://madshi.net/madVR907b.rar

Yup, fixed.

Originally Posted by madshi

I think the first image shows that "detailed" can become rather nasty in some situations. It not only produces more aliasing but can sometimes also create rather wild artifacts. However, in the 2nd image "detailed" looks clearly (less blurry) better than "clean".

So which method to use? What do you guys think?

I think we need to find a middle ground.

[iSunrise]

Originally Posted by madshi

So which method to use? What do you guys think?

I would always go for the clean one, since the lighthouse basically shows that detailed looks so nasty that it's completely unacceptable. It's way overtuned, not anything that anyone could want if you're really about accurate downscaling. Detailed could probably work fine for very rare cases where your complete image is very unsharp to begin with or your downscaling factor is not too high, and the lighthouse picture basically is already pretty sharp anyway.

Just look at the lighthouse's top railing, where the clean picture looks just right, while detailed almost does not look like a rail anymore, but some edged mess that the detailed algorithm created out of just straight lines. It applies way too much aliasing. The same can be seen when you look at the house (garage) on the left of the image, where the fine structures create a moving aliased mess, again. The detailed algorithm also goes against anything that looks natural to me, because when you think about it, an object that is farther away is always going to look a bit more blurred, but that has nothing at all to do with a blurry picture in the first place, it's just that fine detail on more distant objects will naturally melt together. If you apply detailed to the image, it will over-pronounce the detail that is in front the most, therefore clean is a lot more natural, because it will work nice with the detailed areas in the front, but will also apply that to more distant objects to a natural degree.

These two examples in the lighthouse picture are already enough to say that detailed can never be used without actually knowing, which content you feed madVR and you would also have to analyse all the content you feed it, on whether or not detailed would destroy the image (and quite possibly also create some really nasty shimmering scenes in a movie), which is just not the way you would want madVR to work.

Sharpening is one of those post-processing "effects" that need to be applied only so much that you can perceive a slight sharpness increase, but you never dial it up so much that you can "see a completely new picture", erasing some of the old detail altogether and introducing shimmering or other nasty artifacts when we're not only talking about a still image. We want to "downscale accurately" based on the content we have and not downscale and create a new movie along with it.

Maybe that can be optimized and we would need some reference pictures with already good fine detail, so we can figure out when exactly we are crossing the barrier. I would however vote against different SSIM parameters, we should choose and optimize one method and not concentrate on more complicated things.

Clean looks pretty good to me as is, already, because you can see that even clean already has a slight edgy look to it, but that is just because we have a lower resolution and the algorithm tries to fit as much in as possible.

PS: Your lighthouse example is rather low in resolution anyway, since both lighthouses are just about a bare ~128x128 pixels, which is WAY over-pronouncing the algorithm anyway, since the original has 1024x1536 pixels which is a really huge downscaling factor. Detailed probably looks better when your downscaling factor is not that stark to begin with. But even that won't change the fact that it just looks like detailed is way overtuned.

[aufkrawall]

Originally Posted by leeperry

Exactly, completely unusable IME and "clean" isn't all that better due to its extreme sharpness

"Extreme sharpness" is not per se afflicted by artifacts, what's the problem? I really like the sharpness and there's not haloing and not much aliasing (at least with the clean SSIM configuration).

I btw. found out that bicubic 150 can be considered as sharper in general than SSIM for 4k -> WQHD, but the difference is that SSIM doesn't have an anti aliasing effect for aliasing of the source. This mostly occurs with game screenshots, but I think it's totally fine by definition.

I think SSIM clean is still sharp enough. Detailed hardly has any advantages for typical downscaling factors, as our examples show, but its artifacts can be obtrusive.

[6233638]

Originally Posted by madshi

I suppose you're using linear light downscaling? Earlier madVR builds first upscaled chroma in that situation and then converted the image to RGB and then downscaled. The latest build may opt to scale luma and chroma separately for performance reasons. In that situation madVR only uses NNEDI3 for chroma upscaling if the upscaling factor needed for chroma is at least 1.5x. In your case it's 768/540 = 1.42, so madVR decided not to use NNEDI3 because it probably wouldn't bring any visual benefit.

I remember you made a similar change a while ago, but if I recall correctly, you kept it so that linear light downscaling would still upscale chroma to match the luma resolution first, instead of downscaling chroma directly.
Has that decision now been reversed in 0.90.x?

In which case, I'm going to have to try and find some time to start testing chroma scaling again, because I don't believe that this is going to produce the best quality results, and should be a "trade quality for performance" option instead. (perhaps even one enabled by default)

I realize that I still need to get back to you on some of the issues that I reported with a 3D output, but unfortunately I just don't have the time for that at the moment, enabling 3D causes a few issues in my setup so I've had to remove it from the TV for now, and frankly my interest in 3D video is practically nil - at least while I still own a TV that uses active 3D instead of a 4K passive 3DTV.

Originally Posted by madshi

Here's a small comparison of SSIM "detailed" vs "clean" (using Linear Light 2.2 downscaling)
[...]
So which method to use? What do you guys think?

Clean gets my vote. That aliasing is terrible.
Aren't there a ton of sharpening options in madVR now for people that want a sharper image anyway?

Originally Posted by huhn

film is not really mastered for gamma 2.4 or at gamma 2.2.
the trick is to assume the gamma is 2.2 and to change it to 2.4 for a bat cave to get the "correct" gamma of 2.4.

Where are you getting this information?

I don't really want to take things too off-topic here, but I think you are confusing "end-to-end" or "system gamma" where content is not intended to be viewed at the inverse of the BT.709 transfer function, with how mastering is performed.
And if you are, then you are also mistaken in thinking that the best-fit curve for the BT.709 transfer function is 2.2 gamma - it's 1.96
Check page 21 in this EBU Tech paper for some detail on that.

CRTs had an approximate 2.4 gamma response (2.35-2.50 depending on the source) and the new specification for master monitors (BT.1886) is to use 2.4 gamma when the display's contrast ratio exceeds 10,000:1 - which any good mastering monitor will.
So I don't know where people are getting 2.2 from, other than the fact that PC monitors use 2.2 gamma. (to my knowledge, the sRGB transfer function was never actually in use, and only appeared as an option in calibration packages in the past five years)

[madshi]

Ok, thanks for the SSIM feedback, everyone.

Originally Posted by omarank

Ok, I have got the screenshots. Please download from the below link a zipped folder with three files: the Original 4k image and its downscaled images, one with 0.90.7 build and the other with 0.90.4 build. Please see the edges of the melons.

http://www.mediafire.com/download/bs...ownscaling.rar

Now I believe you will be convinced to add a trade quality option to restore the previous Linear Light downscaling method.

I'm confused. I can't reproduce anything near to your 0.90.7 downscaling result on my PC. It looks very aliased in your screenshot, but it looks totally fine on my PC. Can you double check? Maybe the problem only shows with the video source and not with the jpg source image or something? I did use AviSynth to try to get the jpg to 4:2:0 so I could get near to what you might have been testing with, but I still can't reproduce the problem.

Maybe the problem only shows in certain situations, e.g. with certain trade quality options set? Or with image enhancements activated? Or something else like that?

Originally Posted by iSunrise

PS: Your lighthouse example is rather low in resolution anyway, since both lighthouses are just about a bare ~128x128 pixels, which is WAY over-pronouncing the algorithm anyway, since the original has 1024x1536 pixels which is a really huge downscaling factor.

That's true. FWIW, the problem already shows with a smaller downscaling factor, but it gets worse the bigger the downscaling factor is. It should also be noted that the problem doesn't show in a lot of other test images. In most cases SSIM "detailed" looks just fine. But there are some corner cases like this.

Originally Posted by 6233638

I remember you made a similar change a while ago, but if I recall correctly, you kept it so that linear light downscaling would still upscale chroma to match the luma resolution first, instead of downscaling chroma directly.
Has that decision now been reversed in 0.90.x?

If you read through my last couple of posts, I've already explained it in more detail.

Originally Posted by 6233638

In which case, I'm going to have to try and find some time to start testing chroma scaling again, because I don't believe that this is going to produce the best quality results, and should be a "trade quality for performance" option instead. (perhaps even one enabled by default)

Originally Posted by Shiandow

I wonder if you're not going to run into problems when downscaling colours. Scaling colours in gamma light tends to cause some nasty problems if you're not careful. That said, I don't think this will become a problem until the final size is significantly smaller than the chroma size. And of course you don't really know the colours to begin with...

In my tests I didn't find any problems. But my tests were rather limited so I might have missed something. In any case, I'll probably add a trade quality for performance option.

Originally Posted by FreeFall

The problem is that the last frame stays on the screen and you can't use the player until you exit and reload it. The frame is stuck there on screen when switching to windowed mode or trying to interact with the player, using the keyboard commands play a chime sound through the speakers.

I tried re-setting madVR back to defaults and stopping playback worked (no stuck frame) but when I disabled all of the trade quality for performance options the problem returned again (strange). The problem only seems to happen with DVD playback using FSE mode and Zoom Player, windowed mode works without any problems.

Ok. In that case can you please check which exact madVR build introduced the problem for you? Also: Are you using a multi-monitor setup? Finally: In the problem situation: Does the media player still react to anything (e.g. mouse clicks) at all? Or is it totally frozen? In that case please create a freeze report by pressing Ctrl+Alt+Shift+Pause/Break, in case that still works.

[Ver Greeneyes]

The artifacts for 'sharp' on the fence of the lighthouse images look pretty bad, but ignoring the bottom half of the image for a moment, I do think 'sharp' makes the lighthouse itself look better (though it's difficult to say whether it's more faithful to the original sharpness). The potential drawbacks of 'sharp' are too big right now, though (in addition to the problems with the fence, the roofs of the houses also look pretty jagged).

[omarank]

Originally Posted by madshi

I'm confused. I can't reproduce anything near to your 0.90.7 downscaling result on my PC. It looks very aliased in your screenshot, but it looks totally fine on my PC. Can you double check? Maybe the problem only shows with the video source and not with the jpg source image or something? I did use AviSynth to try to get the jpg to 4:2:0 so I could get near to what you might have been testing with, but I still can't reproduce the problem.

Maybe the problem only shows in certain situations, e.g. with certain trade quality options set? Or with image enhancements activated? Or something else like that?

The jpg is already 4:2:0. Please make sure that LAV video decoder is used to decode the jpg as it outputs NV12, 8bit, 4:2:0 for this file. If your system is forcing Still image decoder, try changing the extension of the jpg file to png then LAV video decoder should automatically be used in MPC-HC.

By the way, I am not using any trade quality option or image enhancements.

[madshi]

Originally Posted by omarank

The jpg is already 4:2:0. Please make sure that LAV video decoder is used to decode the jpg as it outputs NV12, 8bit, 4:2:0 for this file. If your system is forcing Still image decoder, try changing the extension of the jpg file to png then LAV video decoder should automatically be used in MPC-HC.

By the way, I am not using any trade quality option or image enhancements.

Ok, I've reset madVR to default settings, and then changed downscaling to SSIM "clean", with AR and LL turned on. Also I've made sure LAV is used as the "decoder" and it sends NV12. And here's what I'm getting:

http://madshi.net/melon907.png

Totally different to your result. Please also try to reset to default settings. Are your screenshots created with the jpg as the source? Or did you use the video file?