madVR - high quality video renderer (GPU assisted)

[dimitrik]

Originally Posted by onyx76

I do 10 bit output madVR + 12 bit output Nvidia CP -> Sony 10 bit capable TV. Everything works with 0 problems!

My OCed GTX 970 can do Nnedi3 64 neurons doubling & quadrupling + ED2 + sharpen edges 0.2 + enhance detail 0.5 = SD -> 1080P = 18 ms render time.
= 720P -> 1080P = 21 ms render time.

Good choice on the GPU (GTX 970 is almost silent + low power usage), hope the upgrade goes well!

That's great thank you very much!

[Warner306]

Originally Posted by ryrynz

Everything has it's own niche and will work together fine. In some cases it may be better to scale one up rather than have another enabled.. but then someone enabling all of these likely has compared the image quality with things on and off anyway. If you have the performance available might as well use it, I can't imagine someone would just tick things and set values haphazardly.

I can accept that. But over-sharpening is possible, and this is known to reduce not improve picture quality. The purpose of madVR is to improve picture quality. It can be a misnomer to turn everything on because it is purported to improve the quality of the image.

[madshi]

Originally Posted by Warner306

I think I have settled on the following settings for 1080p -> 1080p content using image enhancements:

Sharpen Edges (0.2)
Enhance Detail (0.5)

These settings result in small but noticeable enhancement. Sharpen Edges leads to less noticeable edge enhancement than Crispen Edges due to its thicker lines, which better blend into the background. However, Sharpen Edges is far too expensive for my system, so I will wait for an updated version.

Originally Posted by truexfan81

same here, looks great but Sharpen Edges is too expensive right now

When ignoring performance, do you like "sharpen edges" better than AdaptiveSharpen?

Originally Posted by AngelGraves13

I'm not a fan of artificial sharpness. I even set the sharpness on my TV to 0.

The sharpness control in TVs often has very low quality so it's highly recommend setting it to "neutral" (usually but not always 0). The quality of the madVR sharpening algos is much higher than what your average TV does.

Of course it's all a matter of preference. If you prefer not turning madVR's sharpening algos on, that's perfectly fine.

Originally Posted by MistahBonzai

FYI AMD users. I recall at least one report of queues failing to fill in 'fullscreen exclusive mode' while working fine in windowed mode. While chasing such an issue I made the following discovery the hard way.

If you have set the CCC->Performance-Frame Rate to your display refresh rate (or there abouts) in an effort to reduce power consumption be aware that you have capped the frame rate. Therefore 'fullscreen exclusive mode' can't catch-up when the 'render queue' empties however it continues to work just fine in fullscreen window mode.

Good find!!

Originally Posted by dansrfe

Is there a whitepaper on how the NNEDI3 algorithm roughly works? I'm not sure how to quantify the difference in quality between NNEDI3 32 and NNEDI3 16

With some sources you might not see a difference at all between 16 and 32 neurons. Power consumption is nearly twice as high with 32 neurons. Quality can be higher with 32 neurons, depending on which angles the lines in the source have. 32 neurons produces a bit less artifacts than 16 neurons, and reproduces some line angles better than 16 neurons.

Originally Posted by FreeFall

Sorry about the confusion, I'm using the latest nightly LAV build 0.66.0-33. I have re-tested using madVR 0.89.16 and MPC-HC nightly build 1.7.9.202. and everything is working fine using this setup.

Zoom Player 11.1 beta2 Max behaves strangely for some reason with the new test DVD, set to use the default AR, derived (recommended).
https://www.dropbox.com/s/6ifrmre5xl...tdisc.zip?dl=0

Zoom Control Settings: (these are the only options I have enabled)

move subtitles, to bottom of the screen / window.
automatically detect hard coded black bars.
notify media player about cropped black bars, no more than once every 2 sec.
keep black bars visible if they contain subtitles, forever.

Here is the result.
https://www.dropbox.com/s/ij9asv9tqa...zoom1.jpg?dl=0

If I disable the option notify media player about cropped black bars it works as expected, showing the subtitles rendered into the black bars.
https://www.dropbox.com/s/84h1atoxsd...zoom2.jpg?dl=0

Is Zoom Player causing the problem?

Not sure, will have a look at that next weekend.

Originally Posted by mcn

While I was preparing your sample file I noticed something interesting.

Apparently the issue is related to tags like <i>...</i>.
This happens also with <b> and <u>.
The tags have the desired effect, the text is visually modified as it should.

However when encountering the first line with a tag the subtitles are not shown anymore in the bottom black bar but on the active video area.

Hence if the .srt starts with a line that has tags all the subs are shown in the active video area.

If the line with tags is in the middle then the subtitles are displayed in the bottom black bar until said line is encountered.
Then the subtitles begin to be shown in the active video area.
After this happens if I go back with the video the subtitles that were previously correctly shown are now displayed in the active video area too.

Good catch! Will have a look at that. Probably I'll have to create another XySubFilter patch to fix this...

Originally Posted by Thunderbolt8

I still have stuttering in windowed mode with my mobile nvidia card. no such problems on my AMD laptop though. does anyone know the reason for that?

What does the OSD (Ctrl+J) say? Any queues getting (near) empty? Any frame drops/glitches listed? Have you tried resetting madVR to default settings, as a test to see if that fixes the problem?

Originally Posted by leeperry

I'd swear I prefered having it applied after every 2X step before

Will run more tests but with sxbr25/50 I find the picture less sharp/appealing now, much like having it applied only once at the end in older builds and the new options don't make up for the PQ decrease as far as I can see =(

I guess you expect us to crank up sharpness in sxbr or the new features when I seemingly prefered having SR applied several times in a row, any chance for a dirty hack to have it applied after every 2X step just like before please? something like a "2X step SuperRes" directory in mVR's folder would work if any possible please.

Have you activated any other sharpening algo in addition SuperRes? If yes, running everything at every 2x step might produce different results. But if you're only using SuperRes without any additional sharpening, then (the last time I checked) running SuperRes only once produces identical or slightly better results than running it on every 2x step. If you have a different opinion (when just using SuperRes, no other sharpeners), please show me comparison screenshots that proof your point, thanks.

Originally Posted by iSeries

Quick question regarding MadVR's gamma processing. My TV has 3 gamma modes, 1.8, 2.2 & 2.4. I've used 2.2 and find shadows too bright, however 2.4 is somewhat odd because it comes out of black just as quickly as 2.2 but is much darker overall. I've found a pleasant compromise with setting the TV to the default 2.2, and setting MadVR to apply a gamma change to 2.35. Nothing is clipped and it comes out of black slowly as it should. Is there any drawback to doing this, quality-wise?

If you don't have a meter for proper calibration, I guess all you can do it trust your eyes. From madVR's point of view there's no downside to modfying gamma curves, as long as you don't disable dithering.

Originally Posted by QBhd

the 4 new refinements are working wonderfully here. There is little to no impact on performance with my R9 270X. All content I watch pushes the GPU to it's limits (except 1080p24) and enabling all 4 refinements still fit in without any sacrifices. So far I have only left the 4 at the default setting of 1 and I am very pleased with the results. Everything looks fantastic, crisp (but not artificially so) and clean.

Here is my settings if anyone is curious:

Target Window = 1024x768 (rectangular pixels) @ 72Hz
Content 720p24

Good to hear you like it! Hmmmm... 720p at 1024x768 is half upscaling half downscaling, right? To be honest, I don't know right now if SuperRes and upscaling refinement becomes active in that situation? If you see a difference with these algos turned on, obviously they must be active, of course.

Originally Posted by Warner306

I don't think all of the sharpeners are designed to work together at the same time. You would save a lot of performance by turning some of them off.

Originally Posted by ryrynz

Everything has it's own niche and will work together fine. In some cases it may be better to scale one up rather than have another enabled.. but then someone enabling all of these likely has compared the image quality with things on and off anyway. If you have the performance available might as well use it, I can't imagine someone would just tick things and set values haphazardly.

Originally Posted by Warner306

I can accept that. But over-sharpening is possible, and this is known to reduce not improve picture quality. The purpose of madVR is to improve picture quality. It can be a misnomer to turn everything on because it is purported to improve the quality of the image.

These algos do work together nicely, because they all do slightly different things. In my tests I'm getting the best quality when activating all 4 at the same time - but of course at reasonable strengths! You can oversharpen when using just one algo, if you use too high strength. At the same time you can achieve no oversharpening when using all 4. It all depends on the chosen strengths. Add to that that after upscaling the algos are less effective, so higher strengths after upscaling makes sense compared to the strengths you should use for image enhancement.

I'm not sure yet whether I would activate all 4 together with SuperRes, though. Will need to do some more tests to clarify that question. "thin edges" and "enhance detail" should definitely match well together with SuperRes. Not sure about the other two. From what I've seen SuperRes might already do something similar to a combination of "sharpen + crispen edges". But then, if the source is already rather soft, SuperRes alone might not suffice to produce a reasonably sharp final image because SuperRes' aim is not to produce a sharp image - but to produce an image faithful to the original unscaled source image. If the source image is soft, SuperRes' output will be soft, too. In that case using all 4 algos might make sense once again.

[QBhd]

Originally Posted by madshi

Good to hear you like it! Hmmmm... 720p at 1024x768 is half upscaling half downscaling, right? To be honest, I don't know right now if SuperRes and upscaling refinement becomes active in that situation? If you see a difference with these algos turned on, obviously they must be active, of course.

Unless this is a case of Placebo effect, I would say yes, they are active in my odd-ball situation. I am sure that with the downscaling nature of my settings (ultimately that is the final step) SuperRes and the 4 new refinements do have a noticeable impact since downscaling does tend to make things a little softer (unless I am mistaken in that assumption).

It does seem like overkill but there is a very noticeable improvement when they y is doubled and then downscaled compared to just using Jinc to upscale it directly.

I too am still just testing these settings out. I will see what all 4 look like without SuperRes... heck I will even go way back and look at Jinc again as a direct upscale for the y (it has been a long time since NNEDI3 was introduced)

QB

[iSeries]

Originally Posted by madshi

If you don't have a meter for proper calibration, I guess all you can do it trust your eyes. From madVR's point of view there's no downside to modfying gamma curves, as long as you don't disable dithering.

Thanks Madshi, I checked with the AVSHD patterns, bar 17 barely flashes and it ramps up nicely from there. With MadVR outputting the video 'as is' and setting the TV to 2.4 bar 17 flashed very brightly, in fact 17-20 all flashed as brightly as each other. I was just wondering if MadVR applying a gamma curve, and then the TV applying it's own gamma curve would have a detrimental effect on quality (dithering is enabled in madVR).

[chros]

Originally Posted by iSeries

With MadVR outputting the video 'as is' and setting the TV to 2.4 bar 17 flashed very brightly, in fact 17-20 all flashed as brightly as each other.

Interesting. With my setup, I have to raise the power curve gamma level in MadVR (at least to 2.05) to be able to see bar 17 on the TV.
But this functionality raised new possibilities (I have just discovered not a long time ago): both my TV and my laptop display crush black level.
On the laptop display I have to raise madvr gamma level (power curve) to 1.80 to be able to see the 18 bar
My choice is for the TV: I set gamma on TV to 2.4 and set power curve gamma to 2.15 in madvr. It still clips bar 17 but the rest look good, and since I mostly watch TV during evenings/nights I vote for the darker picture with more depth.
I have tried out on the TV: if I set gamma to 2.4 on TV and 2.00 (power curve) in madVR then the result will be (almost) the same as I had chosen 2.2 on the TV and left gamma on default in madvr.

[huhn]

don't change the gamma to fix crushed blacks. use the custom levels for it. way more effective and it "doesn't" change the gamma of the picture

[Thunderbolt8]

Originally Posted by madshi

What does the OSD (Ctrl+J) say? Any queues getting (near) empty? Any frame drops/glitches listed? Have you tried resetting madVR to default settings, as a test to see if that fixes the problem?

present queue was empty, render queue almost empty and I had lots of dropped/delay frames or glitches. but resetting to default settings fixed it

Originally Posted by huhn

don't change the gamma to fix crushed blacks. use the custom levels for it. way more effective and it "doesn't" change the gamma of the picture

where can I find the custom levels setting?

[iSeries]

Originally Posted by huhn

don't change the gamma to fix crushed blacks. use the custom levels for it. way more effective and it "doesn't" change the gamma of the picture

In my case, on my TV I find 2.2 too bright overall, and 2.4 (even though it comes out of black very quickly, shadow areas are far too bright with both 2.2 and 2.4) a bit too dark overall and a bit too contrasty. Setting the TV to 2.2 and having MadVR apply a 2.35 gamma curve seems just about perfect.

[huhn]

Originally Posted by Thunderbolt8

where can I find the custom levels setting?

devices -> "device name" properties -> the display expects the following RGB output levels.

[har3inger]

Originally Posted by Warner306

I can accept that. But over-sharpening is possible, and this is known to reduce not improve picture quality. The purpose of madVR is to improve picture quality. It can be a misnomer to turn everything on because it is purported to improve the quality of the image.

How much sharpening is good depends also on screen size, view distance, and observer preferences. There is no one-size fits all.


@madshi: The new sharpen algorithm is better than the old adaptive sharpen, because it feels cleaner and more consistent. It's also better than crispen edges because it doesn't alias very much at all. (However, it's wtf expensive right now, hopefully that will change by a lot. It's not worth the performance hit at all, since it costs about the same as switching from superxbr to nnedi3)

[leeperry]

Originally Posted by madshi

Have you activated any other sharpening algo in addition SuperRes? If yes, running everything at every 2x step might produce different results. But if you're only using SuperRes without any additional sharpening, then (the last time I checked) running SuperRes only once produces identical or slightly better results than running it on every 2x step. If you have a different opinion (when just using SuperRes, no other sharpeners), please show me comparison screenshots

Fair enough, well I'm still experimenting with the four new features and I would thoroughly enjoy the ability to try them in combo with 2X SR in order to decide where that gets me so I would still enjoy a way to force it if any possible please

It might just be that 2X SR alone actually oversharpened the picture, which would also explain why sxbr25 was all it took sometimes but there's really only one way to find out how that would react in combo with the new options and I might have already reached videophool territories at this point coz PQ kills anyway heh.

[Warner306]

Sharpen Edges is preferable to AdaptiveSharpen because it is less artificial in appearance.

[Warner306]

So all four sharpeners can be used at the same time? I am already using such low values. Obviously, my threshold for artificial sharpening is very low.

[madshi]

Originally Posted by QBhd

Unless this is a case of Placebo effect, I would say yes, they are active in my odd-ball situation. I am sure that with the downscaling nature of my settings (ultimately that is the final step) SuperRes and the 4 new refinements do have a noticeable impact since downscaling does tend to make things a little softer (unless I am mistaken in that assumption).

It does seem like overkill but there is a very noticeable improvement when they y is doubled and then downscaled compared to just using Jinc to upscale it directly.

I too am still just testing these settings out. I will see what all 4 look like without SuperRes... heck I will even go way back and look at Jinc again as a direct upscale for the y (it has been a long time since NNEDI3 was introduced)

Great, let us know what you find!

Originally Posted by har3inger

The new sharpen algorithm is better than the old adaptive sharpen, because it feels cleaner and more consistent. It's also better than crispen edges because it doesn't alias very much at all. (However, it's wtf expensive right now, hopefully that will change by a lot. It's not worth the performance hit at all, since it costs about the same as switching from superxbr to nnedi3)

Ok, thanks. Yes, I know performance sucks atm, but there's significant room for improvement there. Consider the current build a "proof of concept", not an optimized implementation.

Originally Posted by leeperry

Fair enough, well I'm still experimenting with the four new features and I would thoroughly enjoy the ability to try them in combo with 2X SR in order to decide where that gets me so I would still enjoy a way to force it if any possible please

It might just be that 2X SR alone actually oversharpened the picture, which would also explain why sxbr25 was all it took sometimes but there's really only one way to find out how that would react in combo with the new options and I might have already reached videophool territories at this point coz PQ kills anyway heh.

I think you didn't understand me: Running SuperRes every 2x upscaling step simply doesn't do anything. It looks pretty much the same as running it only once after upscaling is complete. So it's a total waste of performance to run it after every 2x upscaling step. As I said, if you think I'm wrong then please provide screenshot proof, using an older madVR build (please keep AdaptiveSharpen/FineSharp/LumaSharpen disabled for those screenshots).

I'm not willing to spend any time on adding hacks for something that I'm sure is useless. So you'll have to provide some evidence to make me change my mind. Providing screenshot proof should cost you less time than it would cost me to add a hack for you. So if you're not willing to provide proof, then I'm not willing to add a hack. Sounds fair?

Originally Posted by Warner306

Sharpen Edges is preferable to AdaptiveSharpen because it is less artificial in appearance.

Thanks, I like to hear that.

Originally Posted by Warner306

So all four sharpeners can be used at the same time? I am already using such low values. Obviously, my threshold for artificial sharpening is very low.

Yes, you can use them all at the same time. They change different attributes of the image, so they don't compete against each other very much (maybe "crispen edges" competes a bit with "sharpen edges"). It's very important to differ between image enhancement and upscaling refinement, though. Upscaling makes the image rather soft (even when using NNEDI3), so after upscaling you can apply a lot more sharpening than you can before upscaling. I would be careful which algos to enable by default before upscaling (= image enhancements) at which strengths. There's much more room for sharpening in upscaling refinement.

[Thunderbolt8]

so when using both image enhancement and upscaling refinement, are both sharpeners then used twice? one time as general image enhancement and a 2nd time after upscaling? or how would this work?
and if you ticked one sharpener for image enhancement and a different one for upscaling refinement, would both then used? or is one of them disgarded? (e.g. the one with image enhancement in case image enhancement would get disabled if upscaling refinement is used as well at the same time)

[madshi]

Image enhancement is applied before scaling, upscaling refinement is applied after upscaling. So yes, if you activate both, and then upscale, both will be applied separately ("twice"). At least that's the case atm. I have some ideas how to maybe streamline the whole thing, but the algos will need improvements before I can do that.

[huhn]

they "stack". i just totally destroyed a image using 4.0 thin edges 4.0 on both.

for better quality you should stick to upscaling refinement only.

[har3inger]

I think that we should probably not use image enhancements if there's any upscaling (and especially superres) involved, but that it can look pretty good at low strength for native pixel material. I'm using different profiles (srcWidth=1920 or not) to manage which set of settings gets applied.

[DragonQ]

I have a pretty general question about scaling which I think people in this thread are likely well positioned to answer, but apologies if this goes too far off topic. Basically, I'm wondering if you have an HD source which itself has been upscaled from SD but using a poor algorithm (e.g. bilinear), is there likely to be any benefit in downscaling it back to SD and re-upscaling it using a better algorithm? Or would it always look worse?