madVR - high quality video renderer (GPU assisted)

[zoyd]

I've managed to reproduce something similar to the Calman result by keeping saturation the same and removing all excess luminance from each pixel. Which is pretty much the opposite of what madVR 0.99.22 does. The next madVR build will compromise by walking middle ground: I'm going to move such out-of-gamut pixels in-gamut by reducing both (perceived) luminance and saturation in a similar amount. It's not directly a dE formula I'm using, but it should be a decent approximation.


That should improve things, will test when ready.

I forgot to ask, what does madVR do with BT.2020 input when you specify

1. "this display is already calibrated" for BT.709 primaries
vs.
2. "this display is already calibrated" for BT.2020 primaries

and is it the same whether HDR processing is on or off?

[leeperry]

I've downloaded both your mega files, and looked at all your screenshot comparisons. To be honest, I'm not fully sure which image was made with which settings. The files are not named with the full settings in the file name. E.g. they should have been named "SR1 no AB" or "SR2 AB75". But instead they have names such as "1.png", "2.png", "3 orignal 4X.png" etc. How am I supposed to understand which SR level and which AB percent those were made with?

Because all the names are not clear, it's hard to draw final conclusions. But e.g. if I compare your "3 SR1 no AB 4X.png" image with "3 AB75 4X.png", then I think the first one (without AB) looks better with less noise around the edges, especially at the bottom left part of the image.

1 is the full size 1080p picture, 2 the 300x300 isolated bit and 3 the 4X magnified version of the 300x300 bit. I initially used SR2 before you asked for SR1. Here are the fully renamed versions of the 1080p files on mega (https://mega.nz/#!W1xART6C!xhdH28qEyDhph3DuWIXvs_4sAtgMVHWor2K4yS3HHdg).

I don't really understand how you could disagree that the top left corner on this picture with SR@1 doesn't look edgier without AB25: http://screenshotcomparison.com/comparison/178293

Ultimately, I guess the previous posts with the guitar make a better point than my pictures as SR without AB in mVR looks way too sharp and edgy to me and I might not be the only one who thinks so for that matter. (http://forum.doom9.org/showpost.php?p=1773757&postcount=38665)

I use SSIM 2D 100% for downscale and I was refining SR after every 2X upscale. I fully realize that SR is still a work in progress but I have never found any use to its AR subsetting when OTOH the AB one is vital to me.

But I must admit that Shiandow's result with his own version of SR looks very impressive as in not overly sharp and very natural looking, I'd love to give it a shot if any possible please :o

[Iznogûd]

Hi, Madshi, my apologies for the delay in posting these:

madVR v09.90.20 (https://www.dropbox.com/s/gcd1kxi6s2lwyag/madVR%20v0.90.20.PNG?dl=0)
madVR v09.90.21 (https://www.dropbox.com/s/x5lkpapzaw3isej/madVR%20v0.90.21.PNG?dl=0)

Now, the settings are the same in both cases:

Chroma upscaling: super-xbr 150, AR on, SR 3.
Image downscaling: SSIM 1D 50%, AR on, linear light on, anti-bloating on 100%.
Image doubling: always if needed, super-xbr anti-bloating 100
Image upscaling: Jinc, AR on
Upscaling refinement: sharpen edges 1, thin edges 1, enhance detail 1, anti-bloating 100%, AR on, SR 1 linear light AR on.


I hope this will be useful to you. Please let me know if you need more info from me.

[madshi]

I forgot to ask, what does madVR do with BT.2020 input when you specify

1. "this display is already calibrated" for BT.709 primaries
vs.
2. "this display is already calibrated" for BT.2020 primaries

and is it the same whether HDR processing is on or off?
With SDR BT.2020 content and "this display is already calibrated to BT.709", madVR converts the gamut from BT.2020 to BT.709, but colors are simply clipped in the most simple way to BT.709, which is not optimal. With SDR content and "this display is already calibrated to BT.2020", no processing is necessary/done.

With HDR BT.2020 content, it's almost the same. However, the BT.2020 -> BT.709 gamut mapping is performed in much higher quality.

It's on my to do list to perform the higher quality processing also for SDR gamut down-conversions.

I don't really understand how you could disagree that the top left corner on this picture with SR@1 doesn't look edgier without AB25: http://screenshotcomparison.com/comparison/178293
Comparing SR1 with or without AB was not the intention of this whole operation. I've asked for SR1 screenshots because my suspicion is that AB for SR does nothing but simply decrease SR strength. If my suspicion is right than the images to compare would be 1) SR2 with AB and 2) SR1 without AB (or with less AB). If both images are nearly the same then that would indicate that AB's only benefit is in reducing SR strength.

So here we go:

SR2 AB75 (http://madshi.net/madVR/leeperry_SR2AB75.png)
SR1 no AB (http://madshi.net/madVR/leeperry_SR1AB0.png)

Although the images don't have exactly the same look, they appear reasonably close, with the SR1 image having less noisy edges. Based on this one test image, I'd prefer SR1 over SR2 AB75. Wouldn't you?

Ultimately, I guess the previous posts with the guitar make a better point than my pictures as SR without AB in mVR looks way too sharp and edgy to me
Who's to judge which image is better there? We don't have a groundtruth for that one.

But I must admit that Shiandow's result with his own version of SR looks very impressive as in not overly sharp and very natural looking, I'd love to give it a shot if any possible please :o
No. Every time I compare it, it looks worse to my eyes than what madVR produces. But you're welcome to try yourself. If you do, please use an image where you have a ground truth, and be careful that you pick similar algorithms everywhere, so that you really only test the different SuperRes algos and not differences in other algos.

Hi, Madshi, my apologies for the delay in posting these:

madVR v09.90.20 (https://www.dropbox.com/s/gcd1kxi6s2lwyag/madVR%20v0.90.20.PNG?dl=0)
madVR v09.90.21 (https://www.dropbox.com/s/x5lkpapzaw3isej/madVR%20v0.90.21.PNG?dl=0)

Now, the settings are the same in both cases:

Chroma upscaling: super-xbr 150, AR on, SR 3.
Image downscaling: SSIM 1D 50%, AR on, linear light on, anti-bloating on 100%.
Image doubling: always if needed, super-xbr anti-bloating 100
Image upscaling: Jinc, AR on
Upscaling refinement: sharpen edges 1, thin edges 1, enhance detail 1, anti-bloating 100%, AR on, SR 1 linear light AR on.

I hope this will be useful to you. Please let me know if you need more info from me.
Thanks. That's looks like quite a strong difference. Is it possible that with 0.90.20 the upscaling refinement was not applied? Please try unchecking those options and see if it makes a difference ot the render times.

There was a user who complained about upscaling refinement settings not becoming active when upscaling by a relatively small scaling factor. I've changed that, which will of course slow things down if upscaling refinement suddenly becomes active when it didn't before.

[zoyd]

With SDR BT.2020 content and "this display is already calibrated to BT.709", madVR converts the gamut from BT.2020 to BT.709, but colors are simply clipped in the most simple way to BT.709, which is not optimal. With SDR content and "this display is already calibrated to BT.2020", no processing is necessary/done.

With HDR BT.2020 content, it's almost the same. However, the BT.2020 -> BT.709 gamut mapping is performed in much higher quality.


Thanks! So if I use instead "calibrate with 3dlut" and load a BT.2020->BT.709 LUT in the BT.2020 slot and enable HDR processing, madVR will use the LUT for gamut mapping and HDR algos for tone mapping?

[leeperry]

Comparing SR1 with or without AB was not the intention of this whole operation. I've asked for SR1 screenshots because my suspicion is that AB for SR does nothing but simply decrease SR strength. If my suspicion is right than the images to compare would be 1) SR2 with AB and 2) SR1 without AB (or with less AB). If both images are nearly the same then that would indicate that AB's only benefit is in reducing SR strength.

So here we go:

SR2 AB75 (http://madshi.net/madVR/leeperry_SR2AB75.png)
SR1 no AB (http://madshi.net/madVR/leeperry_SR1AB0.png)

Although the images don't have exactly the same look, they appear reasonably close, with the SR1 image having less noisy edges. Based on this one test image, I'd prefer SR1 over SR2 AB75. Wouldn't you?

I understood your point that AB might just decrease SR's strength but my point was to compare SR1 with/without AB and AB does make the top left edge smoother and less noisy.

Well, screenshots are one thing, how AB fares in combination with FRC in 60Hz motion another, basically SR1 doesn't look sharp enough and SR2 too sharp. AB gives SR an analog look and does exactly what you initially said: "The new anti-bloating tries to concentrate sharpening on higher frequencies and to remove lower frequencies. I like the look that produces, but your mileage may vary".

I -and others- like the look that it produces, we like how it sharpens higher frequencies and removes lower ones :)

BUT if you really wanna compare SR1 and SR2 AB75, only 4X zooms make sense to me as I run a 32" TV from a 1 meter distance: http://screenshotcomparison.com/comparison/178799

Apart from the top left corner I far prefer the SR2 AB75, it even makes the shadow in the bottom left corner lighter & closer to the ground truth, very obviously less bloated™:

http://screenshotcomparison.com/comparison/178801

http://screenshotcomparison.com/comparison/178802

Who's to judge which image is better there? We don't have a groundtruth for that one.

Oh so now a ground truth has become mandatory? So we went from "I like the look that produces, but your mileage may vary" to "got ground truth?" :p

Maybe the OP would be able to provide it, but it seems rather hard to deny that mVR's SR looks way oversharpened in his example IMHO. I'm certainly not interested in such a sharp image myself, and neither would be others as it would appear.

No. Every time I compare it, it looks worse to my eyes than what madVR produces. But you're welcome to try yourself.
Apparently the only way to try it would be that MPDN thingie but I've never used and I have no interest in learning as it seems tied to MPC.

[pose]

Any RX 480 users here? Id like to compare my rendering times on 7870 to the new card! Mine could do 1280x720 24FPS video to 1080p with luma doubling 64 NNEDI3, smooth motion, error diffusion dithering and low/high debanding with around 35ms rendering time. Trade quality for performance - all unchecked. Win 7 Catalyst 13.12 driver.

[ShiftyFella]

Just tried to reproduce this, but it seems to work just fine on my PC. Can you please make a screenshot of your Ctrl+J OSD, when the problem occurs?
Ok, here it is:

SuperRes turned on:
http://i.imgur.com/3khJ82K.png
http://i.imgur.com/p6JGGyZ.png
SuperRes turned off:
http://i.imgur.com/57jtmWP.png
http://i.imgur.com/IwiMxG6.png

heck, even did a gif switching sr back and forth to show changes:
https://gfycat.com/FloweryIdolizedCaterpillar

In HDR processing everything turned on, preserve hue is on high and peak luminance is at 400 nits. Having different chroma upscaling, image downscaling, doubling, image upscaling algos has no effect on image color shift including turning on or off superres in upscaling refinement, only turning off superres for chroma gets rid of color shit.

[QBhd]

Sorry for the no screen shots madshi... but as Iznogûd posted, the screen shots wouldn't have made much of a difference since they are the same except for the performance hit... I plan on doing some testing of the upscaling refinements myself but work has not let me... and I am about to leave soon for my first Saturday night out in what seems like years! LOL... Hopefully tomorrow, I will have plenty of time to test this performance hit. I'm just glad I'm not the only one who sees it. Although I think I'm the first to report it since I always push the GPU to the max on my 720p content (which is what I mostly watch) and the slightest performance hit will always push my GPU over the edge.

QB

EDIT:

p.s. Is there anyway to create a debug OSD that will show every setting being used? Even if it covers the entire screen... something we could use to narrow things down for you, and provide more detailed SS's. It could also be useful for hardcore users to tweak their own settings

[har3inger]

No screenshots for the OSD for now, but literally everything in it is the same, except render time (35ish -> 45ish) , and the render queue emptying slowly over time.

I have no upscale refinement algo's checked except SR (which is definitely on in both cases), so it shouldn't be that, unless one became forced on regardless of the UI setting.

[ryrynz]

p.s. Is there anyway to create a debug OSD that will show every setting being used? Even if it covers the entire screen... something we could use to narrow things down for you, and provide more detailed SS's. It could also be useful for hardcore users to tweak their own settings

There is not. The best way to narrow things down for him is to simply reset to defaults and test each option in your profile on both versions until you find what is causing it. You ideally want the coder to spend as little time hunting things down as possible so more time can be spent on adding new features or optimization.

[x7007]

The render queue is the first queue in the list which is nearly empty, so that's the bottleneck. It seems that your GPU can't keep up. Might have to do with the high refresh rate (100Hz). You may have to lower your scaling settings. E.g. use 16 neurons instead of 32 for chroma upscaling.


Hi,

I will test the 100Hz, but I tried everything on lowest, not NNEDI3, just everything was "worst quality" let's call it, and still it the rendered didn't go up at all. again, it worked well with FSE but not windowed mode.

[nevcairiel]

BUT if you really wanna compare SR1 and SR2 AB75, only 4X zooms make sense to me as I run a 32" TV from a 1 meter distance: http://screenshotcomparison.com/comparison/178799

SR2 AB75 looks terrible, so many artifacts with higher contrast then the SR1 image.

[K2mil]

Does anyone have that bug when double click into full screen mode is not instant there is some sort of video distortion / corruption for half a sec before full screen plays.

I'm on i5 2500K @ 4.5 and 1070 GTX with latest drivers Had it also on my old 680 GTX


Question two any recommended settings for 1070 GTX ?? Thanks

[ryrynz]

Does anyone have that bug when double click into full screen mode is not instant there is some sort of video distortion / corruption for half a sec before full screen plays.

I'm on i5 2500K @ 4.5 and 1070 GTX with latest drivers Had it also on my old 680 GTX


Don't think anyone else does. Is it FSE mode? Are you using D3D11? Windowed overlay? Are you changing refresh rates? You could always just stay in windowed mode. Did you do a fresh install of your drivers of did you just install the 1070 straight in after taking out another Nvidia card?

[Iznogûd]

Is it possible that with 0.90.20 the upscaling refinement was not applied? Please try unchecking those options and see if it makes a difference ot the render times.


Well, that seems to be it. For testing purposes, I've cranked the upscaling refinements all the way up.

With 0.90.20, the render times are virtually the same, comfortably sitting around 30ms, whether they are engaged or not. Also, there is no appreciable visual difference at all.

With 0.90.21, the performance drop is huge, and the visual difference is quite evident, so it would seem that doubling and refining simultaneously is something my GPU (GTX 750TI) can't cope with.

Thanks for looking into it.

[madshi]

Well, screenshots are one thing, how AB fares in combination with FRC in 60Hz motion another, basically SR1 doesn't look sharp enough and SR2 too sharp.
Which would again only suggest that you use AB to fine tune SR strength, nothing else.

AB gives SR an analog look and does exactly what you initially said: "The new anti-bloating tries to concentrate sharpening on higher frequencies and to remove lower frequencies. I like the look that produces, but your mileage may vary".
I can't see much of this in the screenshots. The madVR SR implementation by design doesn't bloat that much, if at all.

BUT if you really wanna compare SR1 and SR2 AB75, only 4X zooms make sense to me as I run a 32" TV from a 1 meter distance: http://screenshotcomparison.com/comparison/178799

Apart from the top left corner I far prefer the SR2 AB75, it even makes the shadow in the bottom left corner lighter & closer to the ground truth, very obviously less bloated™
Don't you see all those extra artifacts introduced by SR2 AB75? I consider them far more serious than any ever so slight improvement in the "look" of the image.

I'm sorry to say, but based on this screenshot comparison I'd say SR1 is clearly better than SR2AB75. Maybe you have two or three other images where SR2AB75 looks better than SR1?

Oh so now a ground truth has become mandatory? So we went from "I like the look that produces, but your mileage may vary" to "got ground truth?" :p

Maybe the OP would be able to provide it, but it seems rather hard to deny that mVR's SR looks way oversharpened in his example IMHO. I'm certainly not interested in such a sharp image myself, and neither would be others as it would appear.
Having an objective ground truth is very helpful, but not always a requirement. Have you actually carefully compared the guitar image in madVR yourself, toggling between different SR settings? Or have you just quickly looked at two screenshots, and noticed that one is way softer than the other (which should be no surprise if one is SR3 with no AB and the other one AB100, which removes almost all of what SR did)?

FWIW, I had downloaded the original guitar image and had done several tests myself before I wrote my previous comment. I had compared the guitar image in madVR with different settings, and had also compared it to Shiandow's new algorithm, before I came to any conclusions. And the conclusion I ended up with was that with this specific image I still prefer my SuperRes over Shiandow's, and I prefer sharper over softer, but it's hard to do a definite judgement without having a ground truth.

Apparently the only way to try it would be that MPDN thingie but I've never used and I have no interest in learning as it seems tied to MPC.
So you've spent 5 seconds looking at 2 images, which don't even have the same scaling factor, without having a ground truth to compare to, and you're not willing to do some tests yourself to at least get comparison images with the same scaling factor? But you still think you are in a position to judge which algo looks better, and to request that I spend several hours of my time to implement it?

Ok, here it is:

SuperRes turned on:
http://i.imgur.com/3khJ82K.png
http://i.imgur.com/p6JGGyZ.png
SuperRes turned off:
http://i.imgur.com/57jtmWP.png
http://i.imgur.com/IwiMxG6.png

heck, even did a gif switching sr back and forth to show changes:
https://gfycat.com/FloweryIdolizedCaterpillar

In HDR processing everything turned on, preserve hue is on high and peak luminance is at 400 nits. Having different chroma upscaling, image downscaling, doubling, image upscaling algos has no effect on image color shift including turning on or off superres in upscaling refinement, only turning off superres for chroma gets rid of color shit.
Ah, thanks, I've found out that the problem only occurs if the luminance measurement feature is activated at the same time as SuperRes for chroma upscaling!

No screenshots for the OSD for now, but literally everything in it is the same, except render time (35ish -> 45ish) , and the render queue emptying slowly over time.

I have no upscale refinement algo's checked except SR (which is definitely on in both cases), so it shouldn't be that, unless one became forced on regardless of the UI setting.
p.s. Is there anyway to create a debug OSD that will show every setting being used?
Not exactly, but something that approaches that: Create an empty file named "ShowRenderSteps" in the madVR folder. Then make a screenshot of the OSD with the 2 different madVR versions. That should allow us to see if there's any difference in the configuration of the render steps (shader passes).

I will test the 100Hz, but I tried everything on lowest, not NNEDI3, just everything was "worst quality" let's call it, and still it the rendered didn't go up at all. again, it worked well with FSE but not windowed mode.
Can you please make a screenshot of the Ctrl+J OSD with these "worst quality" settings where the render queue doesn't fill in windowed mode?

Does anyone have that bug when double click into full screen mode is not instant there is some sort of video distortion / corruption for half a sec before full screen plays.
Switching to fullscreen exclusive mode does take some time, there's not much I can do about that. You can try if disabling fullscreen exclusive mode helps, but it's the most reliable playback mode, so I'm not sure I would recommend sticking with that configuration, even if it helps get rid of the distortion.

[madshi]

Well, that seems to be it. For testing purposes, I've cranked the upscaling refinements all the way up.

With 0.90.20, the render times are virtually the same, comfortably sitting around 30ms, whether they are engaged or not. Also, there is no appreciable visual difference at all.

With 0.90.21, the performance drop is huge, and the visual difference is quite evident, so it would seem that doubling and refining simultaneously is something my GPU (GTX 750TI) can't cope with.
Good to hear we've solved the mystery, in your case at least.

Don't think anyone else does.
I think he just means the short moment when the video can appear slightly stretched on screen when going into FSE mode.

[jerryleungwh]

Now I am having another problem that when I am doing audio bitstreaming with the latest lav audio filter(nightly), and enters D3D11 FSE mode, there is no audio and the video freezes, even though the seek bar is still moving

[madshi]

Now I am having another problem that when I am doing audio bitstreaming with the latest lav audio filter(nightly), and enters D3D11 FSE mode, there is no audio and the video freezes, even though the seek bar is still moving
This problem only occurs with audio bitstreaming but not with PCM? And it only occurs in FSE mode? Sounds strange...

-------

Got a reply from Graeme when asking about whether I should reduce more luminance or more saturation:

it's very dependent on the situation. If it's out of
gamut due to there being an "exposure adjustment" problem,
then you would reduce luminance more than saturation.

If its luminance is about right, but it is simply more
vivid than the destination gamut can accommodate, then
reducing both luminance and saturation equal perceptual
amounts might be a good first approximation. The exactly
ratio depends on the role the color plays in the overall image.
So the next build is going to have a new option where you can choose how much luminance vs saturation reduction you want to have. Not sure if the option will stay long term, but it should allow us to do some experiments to maybe find a good setting for all content.

[jerryleungwh]

This problem only occurs with audio bitstreaming but not with PCM? And it only occurs in FSE mode? Sounds strange...

-------

Got a reply from Graeme when asking about whether I should reduce more luminance or more saturation:


So the next build is going to have a new option where you can choose how much luminance vs saturation reduction you want to have. Not sure if the option will stay long term, but it should allow us to do some experiments to maybe find a good setting for all content.
I'll need to run more tests to say for sure as I'm switching between 2 TVs and I'm not sure if that matters. When I uncheck the audio formats on LAV audio filter for bitstreaming and use the TV headphone socket as output, it works fine in D3D11 FSE mode, but when I switch to another setup which I can use an AV receiver for HD audio bitstream(computer->AV receiver->TV), I check the audio formats for bitstreaming in LAV audio filter and theres is no audio as soon as I switch to FSE mode. The audio doesn't return even after I switch back to windowed mode without quitting the player.

[ShiftyFella]

Ah, thanks, I've found out that the problem only occurs if the luminance measurement feature is activated at the same time as SuperRes for chroma upscaling!
I see, that's the only thing I didn't try turning off in hdr. I tried turning off compress highlights but that made no difference to the image and because gui greys out other options, I just assumed they were turned off as well, so never gave second thought to play with them them.


Any RX 480 users here? Id like to compare my rendering times on 7870 to the new card! Mine could do 1280x720 24FPS video to 1080p with luma doubling 64 NNEDI3, smooth motion, error diffusion dithering and low/high debanding with around 35ms rendering time. Trade quality for performance - all unchecked. Win 7 Catalyst 13.12 driver.
According to benches, 480 should be roughly equal to my 290x, so this might be good reference point for you, as I can do your settings at around 16ms but you can crank it up to Chroma ReconSoft@SR1, Doubling NNEDI3 128 luma and 32 for chroma, CE 0.6 AR in upscaling refinement and end up at around 32ms with smooth motion, error diffusion 2 http://i.imgur.com/SJO2IF8.png
If I go for super-xbr for doubling instead of nnedi3 render time drops to 18ms with almost non existent load on the gpu, which is pretty nice if you ask me

[pose]

ShiftyFella, performance of your 290x is amazing! I really hope that RX 480 will perform as well but i doubt it after seeing Sunset1982 test. He compared RX 480 to GTX 970 and they are pretty close in terms of performance. If i remember correctly, back in the days NNEDI3 worked faster on AMD cards and my 7870 were faster than 970. Maybe things changed since then but this is really concerns me.
Excel File:
MadVRComparison.xlsx (http://www.file-upload.net/download-11746774/MadVRComparison.xlsx.html)

[madshi]

I'll need to run more tests to say for sure as I'm switching between 2 TVs and I'm not sure if that matters. When I uncheck the audio formats on LAV audio filter for bitstreaming and use the TV headphone socket as output, it works fine in D3D11 FSE mode, but when I switch to another setup which I can use an AV receiver for HD audio bitstream(computer->AV receiver->TV), I check the audio formats for bitstreaming in LAV audio filter and theres is no audio as soon as I switch to FSE mode. The audio doesn't return even after I switch back to windowed mode without quitting the player.
Don't know, might be that in your Optimus system one GPU doesn't pass the bitstream properly to the other GPU or something like that. I don't think there's anything I can do about it.

[madshi]

madVR v0.90.23 released

http://madshi.net/madVR.zip

* updated AdaptiveSharpen to version 2016-07-10
* custom output levels can now be adjusted in 0.25 steps
* HDR: added option to define how to handle too bright & saturated pixels
* fixed: HDR luminance measurements + chroma SuperRes produced green screen
* fixed: #420: chroma was doubled despite being below scaling factor threshold

[leeperry]

I'm sorry to say, but based on this screenshot comparison I'd say SR1 is clearly better than SR2AB75. Maybe you have two or three other images where SR2AB75 looks better than SR1?
I'm afraid that whatever screenshots I would provide they would never be to your taste as you just said yourself that you "prefer sharper over softer" when I'm looking for the best sharp/natural compromise and again I use SSIM 2D 100% downscaling which is very sharp. SR's AB looks stunning to me in 60Hz FRC on low-res low-bitrate footage.

Have you actually carefully compared the guitar image in madVR yourself, toggling between different SR settings?
I just like the look of Shiandow's, it looks very natural as in neither too sharp or too soft. My goal is to have the feeling of looking through a window and his screenshot seems more prone to fool my brain.

I had downloaded the original guitar image and had done several tests myself before I wrote my previous comment. I had compared the guitar image in madVR with different settings, and had also compared it to Shiandow's new algorithm, before I came to any conclusions. And the conclusion I ended up with was that with this specific image I still prefer my SuperRes over Shiandow's, and I prefer sharper over softer, but it's hard to do a definite judgement without having a ground truth.
Right, so even though several people in this thread made clear that they thoroughly enjoyed SR's AB you're still not keen on bringing it back and again you just said that you "prefer sharper over softer", I see little point in me continuing to waste your time about it. I rest my case that I especially like SR's AB in combination with 60Hz FRC and 2X SR refinement. I'll just stick to 0.90.21 for a (long) while then because as others in this thread we like the look that it produces, but our mileage may very well vary indeed ;)

So you've spent 5 seconds looking at 2 images, which don't even have the same scaling factor, without having a ground truth to compare to, and you're not willing to do some tests yourself to at least get comparison images with the same scaling factor? But you still think you are in a position to judge which algo looks better, and to request that I spend several hours of my time to implement it?
I figured that at this point it would only be a matter of rolling a script in mVR as the implementation would appear to already be there but tbh I must admit that I'm kinda getting tired of constantly whining in this thread for my favorite settings to stay and you must feel even more irritated reading it. Please allow me to apologize for this situation as I sincerely do not mean to abuse your patience and I certainly do not take anything for granted, we're already extremely lucky to have you sharing your most tremendous work with us so let's call it a day then http://forum-images.hardware.fr/icones/smilies/jap.gif

Too bad stupid MPC can't do seamless playback and provide a D3D GUI, the W95 WMP GUI really nails it huh....it would feel like a Formula 1 engine in a Trabant or something, oh well c'est la vie otherwise I would happily give MPDN a shot as scripts would appear to be user-customizable for a change.

[James Freeman]

madVR v0.90.23 released


* HDR: added option to define how to handle too bright & saturated pixels


Looks beautiful, Thank You!
Samsung and Panasonic eating their hats with jealousy. ;)

EDIT:
Bug: Luminance/Saturation % takes effect even if "preserve hue in" is unchecked.

[madshi]

I'm afraid that whatever screenshots I would provide they would never be to your taste
Taste got nothing to do with it. I've proven in the past that if users provide objective screenshot proof, I'm very willing to change my mind. Unfortunately your screenshot comparison didn't convince me. Which again has nothing to do with taste, but everything to do with that your preferred settings produced very visible artifacts.

I just like the look of Shiandow's, it looks very natural as in neither too sharp or too soft. My goal is to have the feeling of looking through a window and his screenshot seems more prone to fool my brain.
Shiandow's image might have a soft vs sharp balance that you like. But it is clearly more "bloated" looking than anything madVR's SuperRes algo produces. Which you would have seen if you'd have done a proper comparison.

Right, so even though several people in this thread made clear that they thoroughly enjoyed SR's AB
Huh? Please give me a list of those several people saying they "thoroughly enjoyed SR's AB". And no, ryrynz writing a short comment that he prefers AB in one specific image doesn't satisfy your claim.

Too bad stupid MPC can't do seamless playback and provide a D3D GUI, the W95 WMP GUI really nails it huh....it would feel like a Formula 1 engine in a Trabant or something, oh well c'est la vie otherwise I would happily give MPDN a shot as scripts would appear to be user-customizable for a change.
MPDN got nothing to do with MPC.

[madshi]

Looks beautiful, Thank You!
Samsung and Panasonic eating their hats with jealousy. ;)
Hehe - thanks! So which luminance vs saturation setting do you prefer (see new option in the HDR settings page)? Would be great if you could check with multiple different demos, though, because what looks good in one scene/video might not look so well in another scene/video.

[Uoppi]

Thanks. That's looks like quite a strong difference. Is it possible that with 0.90.20 the upscaling refinement was not applied? Please try unchecking those options and see if it makes a difference ot the render times.

There was a user who complained about upscaling refinement settings not becoming active when upscaling by a relatively small scaling factor. I've changed that, which will of course slow things down if upscaling refinement suddenly becomes active when it didn't before.

EDIT: Oops, I missed the post above where this was just explained. So SR was not activated previously on v.20, most likely.

(Hmm, me personally, I use SR1 or SR2 for every scaling profile and no other refinements. So SR might not have kicked in previously?

All I know is that I've needed to revert from SSIM 2D to SSIM1D in most of my profiles after v.20, so yeah, the difference must be several ms at least. Not a huge deal to me but was just intrigued if anyone else was having the issue. So hence no real testing or screenshots, sorry for that.)

[James Freeman]

This scene from the Chess video looks absolutely wrong with anything but 0% luminance reduction.
Reminds me of the Black/Blue Gold/White dress controversy image that was on the internet lately.

http://www.mediafire.com/convkey/e089/q8yyokc2ng8xdsgzg.jpg

I'm still experimenting.

[leeperry]

Taste got nothing to do with it. I've proven in the past that if users provide objective screenshot proof, I'm very willing to change my mind. Unfortunately your screenshot comparison didn't convince me. Which again has nothing to do with taste, but everything to do with that your preferred settings produced very visible artifacts.
Yes, you were kind enough to bring back SR's 2x refinement after I managed to show you screenshots that changed your mind and you're right maybe my good ole lighthouse is not up to snuff this time.


Huh? Please give me a list of those several people saying they "thoroughly enjoyed SR's AB".
Well I got one here:
SR AB was nice especially for upscaled MPEG-2 DVDs. It really smoothed out the rough edges in the transfers and made SR look smoother.
^ What the man said, it smoothes out noisy edges and makes SR look way smoother. Couldn't have said it better, I love it in 60Hz FRC :)

Just how pathetic does it get when it takes forum goers to backup your point lol

MPDN got nothing to do with MPC.
Screenshots in its own dedicated thread make it look like MPC and I haven't seen an option in PotP to run MPDN as a VR. Anyway MPDN is explained to be a "modern extensible media player" when I'm seeking a standalone VR to run in PotP because the latter provides seamless playback and a D3D GUI. If it's not based on MPC whatsoever it might have deemed worthy to use a more modern GUI.

[Ver Greeneyes]

This scene from the Chess video looks absolutely wrong with anything but 0% luminance reduction.
I haven't tested any other scenes, but on this one I lean toward 25% luminance reduction and 75% saturation. 0% luminance reduction looks too exaggerated to me (on the bits where there's a visible difference), but anything above 25% looks too reduced (like the light is taking on the color of the fabric rather than being white).

Which other demos are good to test with by the way? I know some of them have bad/missing metadata and such.

[har3inger]

Not exactly, but something that approaches that: Create an empty file named "ShowRenderSteps" in the madVR folder. Then make a screenshot of the OSD with the 2 different madVR versions. That should allow us to see if there's any difference in the configuration of the render steps (shader passes).


Can you please make a screenshot of the Ctrl+J OSD with these "worst quality" settings where the render queue doesn't fill in windowed mode?



Very cool. I didn't know this was a feature.

I've tried to do this, but the OSD text size is too large to see whats going on beyond 11 steps that involve chroma upscaling and 2 passes of chroma SR, and 1 deband angle step. The rest of it is off the screen. Increasing window height makes the font larger, ever since the OSD font update.

Render times for either .20 or .23 go up to ~55 and ~58 ms respectively (~35 and ~41 ms w/o) when ShowRenderSteps is on.

Edit: screenshots of what little there is (nothing), taken at exactly the same frame in same video.

http://imgur.com/a/qrT7O

[madshi]

Hmm, me personally, I use SR1 or SR2 for every scaling profile and no other refinements. So SR might not have kicked in previously?
Might have. You can test disabling it manually and see if doing that affects render times in either 90.20 or 90.23.

There's one more thing I might have changed: In older versions when there was image doubling -> downscaling, IIRC I ran upscaling refinement on the downscaled final res. I might have changed that to run on the doubled image before downscaling. Which of course would also increase render times, but should also improve image quality.

This scene from the Chess video looks absolutely wrong with anything but 0% luminance reduction.

[...]

I'm still experimenting.
I haven't tested any other scenes, but on this one I lean toward 25% luminance reduction and 75% saturation. 0% luminance reduction looks too exaggerated to me (on the bits where there's a visible difference), but anything above 25% looks too reduced (like the light is taking on the color of the fabric rather than being white).

Which other demos are good to test with by the way? I know some of them have bad/missing metadata and such.
Keep the feedback coming! Would be great if we could find one setting that works as a decent compromise for most/all demos.

@Ver Greeneyes, the Sony HDR demo makes for a good test, too. I don't know if this new setting affects all HDR demos or just some. Haven't done too much testing myself yet.

Well I got one here: [...]
Ok, so it's AngelGraves13 and you. But there were more users voicing their opinion against SR AB, myself included.

Anyway, if anyone wants SR AB returned, I'm willing to look at further screenshot comparisons.

Very cool. I didn't know this was a feature.

I've tried to do this, but the OSD text size is too large to see whats going on beyond 11 steps that involve chroma upscaling and 2 passes of chroma SR, and 1 deband angle step. The rest of it is off the screen. Increasing window height makes the font larger, ever since the OSD font update.
Oh yes, I often have the same problem. I guess I should make the font size smaller if this special feature is enabled.

Maybe you can try setting chroma upscaling to Bilinear without any extra stuff? I'm quite sure the rendering time slowdown got nothing to do with chroma upscaling. That way we'd see more useful information in the screenshots.

Alternatively you could also create a debug log with 0.90.20 and 0.90.23. That would also help.

Edit: Too much cut off. Could you create a debug log for me with both versions, please?

[zoyd]

Hehe - thanks! So which luminance vs saturation setting do you prefer (see new option in the HDR settings page)? Would be great if you could check with multiple different demos, though, because what looks good in one scene/video might not look so well in another scene/video.

Yes, a definite improvement but the optimal setting is color dependent. For very saturated colors 75%-90% luminance reduction looks good but for more unsaturated colors, especially to retain gradients in whites above 100 nits, 0% luminance reduction is needed, and you probably want to maintain current saturation in these cases. Perhaps there is a simple way to weight the adjustment like a 1/r^2 type distance from fully saturated?

I also noticed that the compression level has a significant effect on the resulting appearance so it's important to set that to your true peak white value. In my case the current options 265 or 400 nits straddle my display peak white of 350 nits.

[CarlosCaco]

There was a user who complained about upscaling refinement settings not becoming active when upscaling by a relatively small scaling factor. I've changed that, which will of course slow things down if upscaling refinement suddenly becomes active when it didn't before.

Hey madshi that was me! Thanks a lot! Did you changed this on .23? Because in .22 upscallinf refinements for that small scalinh factor only activated with supersampling , Gone test tonight!!!

[madshi]

Yes, a definite improvement but the optimal setting is color dependent. For very saturated colors 75%-90% luminance reduction looks good but for more unsaturated colors, especially to retain gradients in whites above 100 nits, 0% luminance reduction is needed, and you probably want to maintain current saturation in these cases. Perhaps there is a simple way to weight the adjustment like a 1/r^2 type distance from fully saturated?
I'm not sure I fully understand. You say "for more unsaturated colors [...] 0% luminance reduction is needed" and "you probably want to maintain current saturation in these cases". That sounds like I should neither touch luminance nor saturation for "more unsaturated colors". But if I don't touch either of those, the pixel will stay out of gamut! I think I'm misunderstanding something.

Hey madshi that was me! Thanks a lot! Did you changed this on .23? Because in .22 upscallinf refinements for that small scalinh factor only activated with supersampling , Gone test tonight!!!
No, I didn't change anything in that area from .22 to .23.

[zoyd]

I'm not sure I fully understand. You say "for more unsaturated colors [...] 0% luminance reduction is needed" and "you probably want to maintain current saturation in these cases". That sounds like I should neither touch luminance nor saturation for "more unsaturated colors". But if I don't touch either of those, the pixel will stay out of gamut! I think I'm misunderstanding something.

How does the luminance/saturation adjustment interact with "compress highlights"?

What I see when there is an out-of-gamut white (with some gradients e.g. between 100-4000 nits) is that compress highlights and 0% luminance reduction maintains the gradients in the highlights while 100% luminance reduction does not.

[madshi]

How does the luminance/saturation adjustment interact with "compress highlights"?

What I see when there is an out-of-gamut white (with some gradients e.g. between 100-4000 nits) is that compress highlights and 0% luminance reduction maintains the gradients in the highlights while 100% luminance reduction does not.
Oh well. It's all relatively complex. I can't do everything in perfect quality in real time in a pixel shader, so I'm compromising accuracy a little bit (but hopefully not much).

The calculation of the amount of luminance reduction is based on a gamut mapping "estimate" performed in xyY. This estimate will probably be slightly off, but it shouldn't be off much. But consequently also the luminance reduction amount might be slightly off. Maybe this results in the lost gradients at 100% luminance reduction, but I'm not sure.

Do the gradients also get lost with 75% luminance reduction? Which HDR demo file (and runtime in the file) are you testing this with?

[leeperry]

Anyway, if anyone wants SR AB returned, I'm willing to look at further screenshot comparisons.
OK fair enough, what would you expect to see from screenshots then? Obviously AB will make the picture softer so I'm not sure what would change your mind, I'm certainly not saying that SR AB is perfect but upscaling is all about compromises and I like the smoothness SR AB provides in motion. Maybe SR strengths going from 1 to 10 instead of 1 to 4 would hit the spot, especially when 4 is unusable.

[Ver Greeneyes]

Tried the camp demo (http://demo-uhd3d.com/fiche.php?cat=uhd&id=144), still think 25% luminance reduction and 75% saturation looks best. On the tents in the dark it's hard to choose between 25/75 and 50/50 though. I felt like 25/75 left the red tent a bit too bright, and 50/50 darkened it a bit too much. 33.3/66.7 might be a good compromise for that particular scene. 100/0 is right out by the way, causes awful banding on the sun at the beginning of the scene.

Edit: Oh, I should mention I tested this using 120 nits for the display luminance to match my calibration, but 120 nits might still be overbrightening things. 180 nits looks less washed out, though it's hard to say what the 'right' brightness should be.

[zoyd]

Do the gradients also get lost with 75% luminance reduction? Which HDR demo file (and runtime in the file) are you testing this with?

0% was best for this case and not much difference with 25% but started losing detail by 50%. pm'd about clip.

@Ver Greeneyes, I don't get any banding at any of the settings. That tent scene amber headlight renders best at 100%/0% while the white headlight renders best at 0%/100% on my display. Those renderings also best match the full CalMAN LUT mapping.

[ShiftyFella]

* HDR: added option to define how to handle too bright & saturated pixels
did quick test

sony camp:
100% luminance 0% saturation: http://i.imgur.com/xTousOt.jpg
50%: http://i.imgur.com/55ztgDJ.jpg
0% luminance 100% saturation: http://i.imgur.com/YWe8Txv.jpg

chess demo:

100% Luminance 0% saturation: http://i.imgur.com/VwRWAI9.jpg
50%: http://i.imgur.com/v0S7SSw.jpg
0% luminance 100% saturation: http://i.imgur.com/KfpXHHJ.jpg

100% Luminance 0% saturation: http://i.imgur.com/z0mJ5hj.jpg
50%: http://i.imgur.com/OA13aqp.jpg
0% luminance 100% saturation: http://i.imgur.com/59CMxPp.jpg

To my eyes camp looked better with 100% luminance but for chess demo it was exact opposite and only 100% saturation looked like it should. I think even with 'looked best' option it's still not quite ideal and i can see in some places of the image where other options work better but rest of the picture looks wrong with them. I'm happy that we have options for now

[zoyd]

The chess demo is encoded incorrectly- I wouldn't use it for eval.

[Warner306]

madVR v0.90.23 released

http://madshi.net/madVR.zip


* custom output levels can now be adjusted in 0.25

What does this option do?

[Warner306]

madVR v0.90.23 released

http://madshi.net/madVR.zip


* HDR: added option to define how to handle too bright & saturated pixels

Can you also define this?

[Ver Greeneyes]

@Ver Greeneyes, I don't get any banding at any of the settings. That tent scene amber headlight renders best at 100%/0% while the white headlight renders best at 0%/100% on my display. Those renderings also best match the full CalMAN LUT mapping.
Maybe banding was the wrong word, but the sun looks odd to me with 100% luminance reduction. Here's a comparison:

100% luminance reduction and 0% saturation (http://i.imgur.com/bybPZH3.jpg)
75% luminance reduction and 25% saturation (http://i.imgur.com/FsEiyrF.jpg)
50% luminance reduction and 50% saturation (http://i.imgur.com/seKZrg6.jpg)
25% luminance reduction and 75% saturation (http://i.imgur.com/Zr9lm8A.jpg)
0% luminance reduction and 100% saturation (http://i.imgur.com/XlZSu0X.jpg)

(I normally use a 3DLUT but I turned it off for this)

[ShiftyFella]

The chess demo is encoded incorrectly- I wouldn't use it for eval.
Are you referring to values being higher then defined peak luminance in metadata? Shouldn't madvr compression account for this or are they being clipped. I don't know how this is being handled by madvr, as for lg clip, I think metadata is wrong rather then encode but I could be wrong.

[zoyd]

Are you referring to values being higher then defined peak luminance in metadata? Shouldn't madvr compression account for this or are they being clipped. I don't know how this is being handled by madvr, as for lg clip, I think metadata is wrong rather then encode but I could be wrong.

The encode has values up to 10,000 nits which is clearly wrong. madVR has no idea how to handle this situation since it has no idea what the values should be. It makes no sense to tune an algorithm to faulty input data.