madVR - high quality video renderer (GPU assisted)

[James Freeman]

Still, we must not forget that all this is in 4-bit, In 8-bit the discern-ability will be several times harder.
Its much easier to discern step 10 from 11 in 4-bit (0-16), than step 175 from 176 in 8-bit (0-255).

Here is a simple example:
http://www.mediafire.com/convkey/f26f/h3a63gtxpgm04agfg.jpg

Top is 100% bottom 400%.
The left square is 160 & 176 (one step in 4-bit).
The right square is 175 & 176 (one step in 8-bit).

This is a very organized (perfectly straight lines) static example, far from real content (in fact, its the exact opposite).
Even now I don't see the difference in shades in the 100% 8-bit square.

Another one with chess pattern:
http://www.mediafire.com/convkey/9bad/2vuj2c96tvnpw1vfg.jpg

Now the Left 4-bit 100% is harder to discern.


Madshi's dithering is endlessly better in blending these shades.
So take that into consideration when selecting the best looking build in 4-bit, because in 8-bit it'll be several times better.

[iSunrise]

...Basically what madVR's random dithering does is identical to "multiColor dynamic", just by using random dithering instead of error diffusion.
Unfortunately, 4bit mode doesn´t work when RD is active (I already tried) so we´re not able to do comparisons ourselves.

mono_static vs. multi_dynamic vs. opposite_dynamic

http://abload.de/thumb/bwramp_mono_static_gig6kbu.png (http://abload.de/image.php?img=bwramp_mono_static_gig6kbu.png)http://abload.de/thumb/bwramp_multi_dynamic_mpk29.png (http://abload.de/image.php?img=bwramp_multi_dynamic_mpk29.png)http://abload.de/thumb/bwramp_opposite_dynamevk8v.png (http://abload.de/image.php?img=bwramp_opposite_dynamevk8v.png)

Why exactly is it beneficial for us, when we insert even more noise (which is colored, too) into black areas, where there should only be gray dots. I just don´t get it. I find it extremely distracting, when I see pink, green and blue dots (which there are a lot of) in my shadow details, which was exactly the reason why we wanted another algorithm, that is more accurate than RD.

[JarrettH]

The development in the last 5 pages sounds interesting, except I don't understand it :D

[James Freeman]

Why exactly is it beneficial for us, when we insert even more noise (which is colored, too) into black areas, where there should only be gray dots. I just don´t get it.

It looks better to the human eye.
Isn't that the whole point of dithering?

[Aikibana]

I think from a hardware point of view AMD probably has the better compute hardware atm, so I don't think this is something NVidia can easily fix with a new driver build. AMD OpenCL interop is costly. This is something that *could* be fixed with a new driver. But I would not count on that ever happening. FWIW, probably I'll use DirectCompute for most new compute features and DirectCompute has no interop cost for AMD hardware. However, it's possible that NNEDI3 may stay with OpenCL. In that case I don't really know whether a high-end NVidia GPU would be better or worse than an AMD GPU.

Thanks for the info!

By the way: why use D3D9?
I've read AMD's interop proces with D3D10 and D3D11 is way better optimised. I guess you must have a very good reason...

[iSunrise]

It looks better to the human eye.
Isn't that the whole point of dithering.
No, because you´re not able to judge what is more accurate with your own eyes. Not in 8bit mode. And if you look at my shots, it doesn´t look better, it actually looks worse.

[nevcairiel]

No, because you´re not able to judge that with your own eyes. Not in 8bit mode. And if you look at my shots, it doesn´t look better, it actually looks worse.

This is obviously an enhanced image, as usually you would only see difference of +1, which isn't a "pink dot", but a slightly more red grey dot.

[lhan]

Does GPU-Z indicate that OpenCL is working correctly? (checkbox flagged)

On paper there's no obvious reason why this system config shouldn't work, right?
Yes OpenCL is checked in GPU-Z. It just runs slow and lags.
But I think I have found the problem.

When viewing interlaced source including DVDs I usually use de-interlacing with bob2x or similiar method to double the framerate, so that frames have to process is 60.(am I right?) My gpu seems just can't handle that much. Total output frames are only about 15fps.

After I turn de-interlace off or use non double-frame method, it can output 30fps, the gpu usage is still nearly full though. I think it works normally now(?

Edit:Never mind, it slows again after rebooting computer, dont know why.

[iSunrise]

This is obviously an enhanced image, as usually you would only see difference of +1, which isn't a "pink dot", but a slightly more red grey dot.
They were done in 4bit mode and they are completely untouched. I just exchanged the builds, nothing more, nothing less.

[nevcairiel]

4-bit is a virtual testing mode in itself, which amplifies such artifacts greatly. Look at 8-bit if you want to judge real world.

[leeperry]

Makes no sense to me. You've been watching content with madVR's random dithering for years now, and madVR's random dithering has a much higher chroma noise level than any of the multiColor/oppositeColor builds. Basically what madVR's random dithering does is identical to "multiColor dynamic", just by using random dithering instead of error diffusion.
I was using SmoothL for quite some time but maybe it's a bad case of rolling VR builds at 1AM duh...or maybe it's because the noise used to be totally random and identical on every channel when now it's doing that "wave cancellation" thingie. Fair enough, I shall try again.

You still haven't told us your opinion about the oppositeColor builds, btw. To my eyes it looks clearly lower noise than monoColor. See here:

| -- monoColor (http://madshi.net/monoColor.png) -- | -- multiColor (http://madshi.net/multiColor.png) -- | -- oppositeColor (http://madshi.net/oppositeColor.png) -- |
I find mono clearer looking, noise in the last one is lower but it seems to be hiding the original grain of the source? Seriously that big round stuff behind him is sharper with mono and even blurrier with opposite, isn't it :confused:

I don't like that "soft" look, it looks blurry in motion but alright, I'll compare them again on 12 grams.

PS: I'm using a 4:2:2 TV with BFI enabled, its "sharpness" setting set to neutral and low debanding forced in mVR BTW.

[iSunrise]

4-bit is a virtual testing mode in itself, which amplifies such artifacts greatly. Look at 8-bit if you want to judge real world.
Of course they are amplified (also because of the bit-difference), but that´s the whole point of it, isn´t it? If the dithering-algorithm itself is not changed between 4bit and 8bit though, the 4bit mode results should also translate when in 8bit and higher. They are just way less obvious.

[James Freeman]

4-bit is a virtual testing mode in itself, which amplifies such artifacts greatly. Look at 8-bit if you want to judge real world.

Different people look for different things.

I think we should look for "Smoothness" & "Noise" in 4-bit, NOT "Extra Colors" or "Color Artifacts".
Extra colors or color artifacts will be invisible in 8-bit (look at this post (http://forum.doom9.org/showpost.php?p=1669757&postcount=23608)).

I think the key is to judge what you think would look better in 8-bit.

[iSunrise]

Different people look for different things.

I think we should look for "smoothness" & "Noise" in 4-bit and not "extra colors".
Color difference will be invisible in 8-bit (look at my square examples).
If you add extra colors into the mix, you are adding more noise. And there´s a threshold somewhere, where adding more noise is not what we want. And I believe we just passed that.

A linear light build with the new changes would be interesting to see and compare against.

[James Freeman]

If you add extra colors into the mix, you are adding more noise. And there´s a threshold somewhere, where adding more noise is not what we want. And I believe we just passed that.

Fair argument.

I agree that for 8-bit we could have stopped on NL6 or even earlier... but NO... we decided to dig deeper to achieve perfection.
The lower we go in bit depth the better we can judge the Noise & Smoothness aspects, and the worse we can judge the Color aspect.
Maybe adding more colors is adding more noise theoretically, but it appears the opposite, even in 4-bit where color distance is exaggerated.
The color builds just look smoother vs the mono builds.

Based on that, the OppositeColor Static build will also do perfectly fine in 8-bit, it will NOT look like dirt on the screen.
It also has less apparent noise.

[flashmozzg]

I liked mono static. Multi dynamic ads "random dif" like noise, so there is no point in it. With normal videos dithering patterns are rarely even noticeable, especially in motion.

[iSunrise]

Maybe adding more colors is adding more noise theoretically, but it appears the opposite, even in 4-bit where color distance is exaggerated.
The problem for us is that we were given the 4bit mode as an opportunity to more easily see differences instead of having to assume them or make image adjustments ourselves (at a point where we already have a very hard time of seeing improvements anymore), and if the algorithm doesn´t change and the 4bit mode works correctly (by reducing the input bitdepth) then it should translate to all other bitdepths as well, the only difference being that it´s just WAY less obvious to us.

So if I don´t find it to be more accurate in 4bit mode, why should I suddenly like it in 8bit more?

[*Touche*]

Why exactly is it beneficial for us, when we insert even more noise (which is colored, too) into black areas, where there should only be gray dots. I just don´t get it. I find it extremely distracting, when I see pink, green and blue dots (which there are a lot of) in my shadow details, which was exactly the reason why we wanted another algorithm, that is more accurate than RD.

Because we don't watch enhanced gray gradients, but actual content.

[iSunrise]

Because we don't watch enhanced gray gradients, but actual content.
You can see the colored dots when you make screenshots in movies, too. The gradients are just there to illustrate it better.
And again, the gradients are not enhanced, they are direct madVR windowed captures from the 4bit mode with the different builds.

[Shiandow]

I'm not sure I understand what you mean. Are you suggesting that I calculate the green random value from the red and blue random values? But doing "g = 1.39782 (0.072 b + 0.2126 r)" would have a much lower peak noise level for green than for red and blue. That means dithering for the green channel would not be very good. The lower noise level would reintroduce worm artifacts for the green ramp, maybe even some banding. I believe every channel needs exactly the full peak noise to make gradients reliably smooth and pattern-free.

I suppose it would indeed be a problem if the green channel didn't get enough noise, conversely if you raise the noise such that the green channel does have enough noise then this will probably make the red and green channels to noisy. Anyway, I'm beginning to suspect that the method you used in the "oppositeColor" algorithm actually is optimal, in the sense that of all the 'reasonable' ways to create noise it is the one that causes the least amount of Y noise. Although I'm not exactly sure how you'd need to define reasonable to actually prove that.

All that aside, oppositeColor looks great so it may not be necessary to look too hard for slight improvements.

[James Freeman]

The problem for us is that we were given the 4bit mode as an opportunity to more easily see differences instead of having to assume them or make image adjustments ourselves, and if the algorithm doesn´t change and the 4bit mode works correctly, then it should translate to all other bitdepths as well, the only difference being that it´s just WAY less obvious to us.

So if I don´t find it to be more accurate in 4bit mode, why should I suddenly like it in 8bit more?

You are right, every choice is valid.

What I'm trying to convey here is the human eye is less sensitive to color and more sensitive to light changes.
That why I think our judgement should be based on Noise & Smoothness, and less on exaggerated Colors.

Find an HD movie with clear skin tones where the character does not move much, and compare the builds.
The test patterns will not reflect real content.

I like the OppositeColor Dynamic build but,
I also like the OppositeColor Static build, for the reason the the "dirt on screen" effect will completely disappear in 8-bit,
It has significantly less noise compared to the dynamic build, and it has the benefits of less apparent Luma noise compared to the MonoColored builds.

[*Touche*]

You can see the colored dots when you make screenshots in movies, too. The gradients are just there to illustrate it better.
And again, the gradients are not enhanced, they are direct madVR windowed captures from the 4bit mode with the different builds.

I don't see them while watching, though. Luma noise on colored images makes more of a jumping color noise effect to me.

[iSunrise]

I don't see them while watching, though. Luma noise on colored images makes more of a jumping color noise effect to me.
What builds are you comparing? If you can show me a sample of that, I would be thankful.

[nevcairiel]

You can see the colored dots when you make screenshots in movies, too. The gradients are just there to illustrate it better.
And again, the gradients are not enhanced, they are direct madVR windowed captures from the 4bit mode with the different builds.

I also don't want to see singular bright grey dots on black instead of a gradient, which your 4bit screenshots show *very* clearly.
I think you need to take a step back and realize the extreme exaggeration the 4bit mode adds. In actual movie contetnt 4bit doesn't look too bad, but especially with greyramps or color ramps, it will just look much worse then 8-bit, and nothing can make that better. It does not translate 1:1 to 8-bit the way you think it does.

[iSunrise]

It does not translate 1:1 to 8-bit the way you think it does.
That´s the whole problem I´m having. Assumptions are not my thing.

Test patterns, blind tests and subjective tests in a healthy mix are. But suddenly, everyone is posting from their subjective POV.

[SamKook]

Well you can use madVR on SLI systems, it is just much slower than if you disable SLI. I updated my drivers again so I do not use NNEDI3 (I usually watch blurays and 1080p->1440p with NNEDI3 isn't much better) so I do not notice or care that rendering is slower. EDIT: I will probably switch back soon though, NNEDI3 chroma upsampling is nice all the time. :o

It is probably only an issue when using NNEDI3 or if you have SLI'ed two low end cards.

I would love to hear reports from other SLI and Crossfire users as to rendering times with and without SLI/Crossfire. Use NNEDI3 to put a real load on the GPU, otherwise the cards sit in a lower power state and ramp up their clock speeds more in SLI instead of taking longer to render.

You mean madvr works with GPU in SLI? I haven't tested it a lot with my 2 GTX760 with updated drivers, but I never could get a picture to display(with the new stuff disabled obviously), it only display what's under the player and keep that image imprinted and everything freeze if I leave it on too long.
I did most of my testing with Nvidia surround activated though(which I think was mentionned at some point wasn't working, but I haven't done my research yet. It's something I wanted to confirm and add a bug report of it) and only tried it disabled with the keyboard shortcut.
I'll have to do a bit more testing.

[Anime Viewer]

Madshi, do these builds use OpenCL or DirectCompute, because the performance is about the same as v0.87.4 for me (A10-5800k)?

To be more specific its using DirectCompute for Error Diffusion, but OpenCL for NNEDI3. (Correct madshi?)
Meaning error diffusion should work across the board for everyone while NNEDI3 will only work for ATI users which includes you with your A10. (Nvidia hasn't put out a fix / corrected drivers, and its so slow on Intel GPU that its unusable). :(

[leeperry]

Alright, I got fresh eyes and feel great, thought it would be time for a shoot-out again :)

madVRadaptive4 : glorious PQ, everything looks natural, the impression of depth is unbelievable, I don't feel like I'm watching a movie at all. This looks extremely real.

madVR_monoColor_static_temporal : like a nasty constant rehashing of the dither grain, its cadence is annoying...it messes with the movie original grain, the sense of depth of the source is partly lost and the picture looks fuzzier with no added value. This is way too noisy and too predictable.

madVR_monoColor_dynamic : feels like a poorly executed x264 encode of madVRadaptive4, clarity is gone and the dithering grain disobeys to what the original movie was conveying.

madVR_monoColor_dynamic_temporal : feels like the dithering grain has learned new dance steps, the darn thing is doing 3 steps tango on top of my movies now.

madVR_multiColor_static: I still find the picture uneasy to look at but alright, maybe it's nocebo somehow...noise is lower but movements look blurry just like those three screenshots of yours look to me. The "natural" and "glorious" PQ of Oblivion is now replaced by softness, way not sharp enough to my taste.

madVR_multiColor_dynamic : The infamous dither grain dance on top of the motion bluriness, yay.

madVR_oppositeColor_static : Impression of depth is very poor, not impressed at all anymore. Still very(more?) blurry movements, still uneasy to look at(even more?).

madVR_oppositeColor_dynamic : Sharpness & sense of depth are much improved over the latter, but still not natural looking or impressive at all for that matter. There's stll that artificial grain that keeps dancing around on top of the movie grain, that's annoying and kills the the reel grain to my eyes. It's the least worse of the color builds to my eyes.

And I'm finally back to madVRadaptive4, I feel like I'm watching through to the TV instead of at the TV itself, my eyes instantly feel at home and can focus on that good ole monochrome movie grain, all those color builds feel like a bad dream and I'm stunned all over again :)

[iSunrise]

madVRadaptive4 : glorious PQ, everything looks natural, the impression of depth is unbelievable, I don't feel like I'm watching a movie at all. This looks extremely real.
Clarity is subjectively the best with the Adaptive 4 build for me, too. The other builds have the same effect on me that I described some pages ago, which is that added grain layer, which makes especially highly compressed movies look smoother. But that smoother appearance comes with the cost of sharpness and lost shadow details and I guess that´s what you´re describing as depth of the image. And that´s also what our eyes can see in the real world.

If madshi would add a grain effect that would work in the same way as multicolor_opposite works now, it would basically be the same result, but that´s not what I look for in a dithering algorithm, I want transparency above anything else and not added pixelated clouds.

I´m also not sure if that´s my panel specifically, but especially in fast camera pans, when using the multicolor_dynamic builds, it almost seems like there´s so much pixel change going on that I can barely focus the image at all.

[bacondither]

Seeing motion blur due to dither... that's something new. :D

[leeperry]

Well, that big round stuff behind the guy is sharper with mono and even blurrier with opposite, isn't it :confused:
| -- monoColor (http://madshi.net/monoColor.png) -- | -- multiColor (http://madshi.net/multiColor.png) -- | -- oppositeColor (http://madshi.net/oppositeColor.png) -- |

Oh well, maybe my 4:2:2 TV, its BFI and/or my lack of red cones all play their part in that problem...not sure.

[iSunrise]

Seeing motion blur due to dither... that's something new. :D
If you have a sharp image and move it, your eyes can focus way easier on the contents. I don´t think that this is new at all. That´s also the reason some people prefer sharper algorithms like Lanczos over SoftCubic. And since the noise is static in ED4, your eyes can easily concentrate purely on the content, rather than the noise, which changes in every frame in every channel, which will distract the focus point when there´s panning.

Well, that big round stuff behind the guy is sharper with mono and even blurrier with opposite, isn't it :confused:
What I don´t like about madshi´s example is that it´s almost completely out of focus. Very hard to see differences. It´s probably a lot better in motion, though.

[bacondither]

OppositeColor dynamic looks to have a tiny bit lower subjective noise level then multiColor dynamic. 4-bits are almost enough to look good!. My preference lean towards oppositeColor dynamic now.

Letting different color channels cancel out some noise and some temporal smoothing(dynamic thresholding noise) to lower the subjective noise further is looking like a good idea.

If you pause the playback when using a dynamic build, the subjective noise level rises to the same as the static builds. Which is only logical, because the "dynamic" part has become static when paused.

[cyberbeing]

Tried that, but it doesn't look better than dymamic to me. Here are 2 test builds for you with this solution:

http://madshi.net/madVRtemporal.rar

I think I agree. madVR_monoColor_dynamic_temporal I found hard to determine if it was a noticeable improvement over madVR_monoColor_dynamic. madVR_monoColor_static_temporal was different, as it strobes the dither between those two states as expected and could possibly be helpful, but only if viewed from an ample distance. I expect the apperance of either of these temporal builds would probably largely depend on the display, so probably a bad idea to have such a thing without a knob to toggle it. If no one else likes it, I'd say just scrap it.

Unfortunately, I can't say I prefer either of these over a normal static dither.

You still haven't told us your opinion about the oppositeColor builds, btw. To my eyes it looks clearly lower noise than monoColor.

I took some screenshots of my own, and I now see this luma noise difference you were talking about. That said...

oppositeColor was horrible to my eyes, it produces way too much strangely colored chroma noise. Not worth the trade-off at all, and seems to harm video quality greatly.

multiColor was more balanced between luma and chroma noise, but the chroma noise wasn't strange like oppositeColor. Could probably live with this if necessary.

monoColor while it did have the most luma noise, I see this as actually beneficial. The dither indeed has a more neutral impact on colors, while the luma noise has a side effect of improving sharpness and definition of the image slightly, compared to multiColor which has a softer appearance.

On a scale of 1 (worst) to 10 (best), I'd rate them as follows for the static builds:

monoColor = 8
multiColor = 6
oppositeColor = 1

[wolfman2791]

Here's a completely non-technical opinion: For me, the perfect picture is a blend of sharpness and vividness of color. When the image gets too sharp, the frames move robotically and you lose the "natural feel" of the video. Similarly, if the colors are too vivid, everything looks like anime.

For me, the "opposite color-dynamic" version seems to capture the balance the best out of the last test series. But i did prefer the multicolor-dynamic over opposite color... this is just an opinion on the opposite color batch.

I would suggest releasing 3 builds btw:
- One that focuses on the vividness of the colors, which would be great for anime.
- One that focuses on the sharpness of the image, which would be great for documentaries, sports, etc.
- One that focuses on balance, which is the best for viewing moves and other videos - basically the one that's going to be the most widely used.

It would be great if the 3 could be incoporated in one build and have a toggle mechanism... but i know that's asking too much :lol:

Just my 2 cents.

[GREG1292]

And I'm finally back to madVRadaptive4, I feel like I'm watching through to the TV instead of at the TV itself, my eyes instantly feel at home and can focus on that good ole monochrome movie grain, all those color builds feel like a bad dream and I'm stunned all over again

I can't seem to go back to A4 as to me it looks like something is missing? I'm so far into the picture with ODC build
I don't need to concentrate on the grain but the movie it's self. A4 i'm alway's trying to make the picture look better.
Very distracting indeed.

[Asmodian]

Oh well, maybe my 4:2:2 TV, its BFI and/or my lack of red cones all play their part in that problem...not sure.

My father is very colorblind (both red/green and blue/yellow). The most amazing difference I experienced growing up with him was the "which is more colorful" test. There are two images one made up of reds and greens and the same image but in browns and tans. To me the red and green is obviously much more colorful while he thought the browns and tans were. He also doesn't notice red flowers on a green bush. I bet the colored noise looks very different to you than to us. :(

I could also see how 4:2:2 mode doesn't play nice with colored noise. I wonder if you could try some tests in a 4:4:4 mode so your opinion is less tied to how your TV processes RGB -> YUV -> RGB? What if you get a better display that can do everything you want while keeping RGB.

If we will also have a low noise dither option maybe lowering noise beyond the level in Adaptive4 isn't necessary for the medium noise algorithm? I must say the difference in luma noise level is all I can notice between the different algorithms (1000:1 display, glossy 1440 IPS - none of those matte anti-glare screens). :p

[*Touche*]

What builds are you comparing? If you can show me a sample of that, I would be thankful.

monoColor_dynamic vs multi/oppositeColor_dynamic, 8bit mode

I've worded it wrong. It is not so much colored noise, as it is noise on areas of single/similar color.

I don't have small samples but this is what I use:

There is a scene in Thor 2 with a closeup of a book @0:32:20. It has areas with lots of details, areas with larger surfaces of single/similar color, rotating monocolored strands...
Another test sample that I use is True Detective s0e3 around 8th minute with large static background sufaces, human skin, hair, shirts...

I've just tried Adaptive4 and find it also inferior to oppositeColor_dynamic.

[leeperry]

Clarity is subjectively the best with the Adaptive 4 build for me, too. The other builds have the same effect on me that I described some pages ago, which is that added grain layer, which makes especially highly compressed movies look smoother. But that smoother appearance comes with the cost of sharpness and lost shadow details and I guess that´s what you´re describing as depth of the image. [..]
I´m also not sure if that´s my panel specifically, but especially in fast camera pans, when using the multicolor_dynamic builds, it almost seems like there´s so much pixel change going on that I can barely focus the image at all.
As I said since the beginning, I'm here to watch the movie and I expect the dithering to look as "transparent" as possible meaning that I don't want it to mess with clarity, motion, movie grain, sharpness, colorimetry and try to become the center of attraction.

Oblivion has gone through tremendous video production(every BD review I've found agrees) and the only grain I wanna see is the stunning work of the engineers who worked on it.

If you have a sharp image and move it, your eyes can focus way easier on the contents. [..] since the noise is static in ED4, your eyes can easily concentrate purely on the content, rather than the noise, which changes in every frame in every channel, which will distract the focus point when there´s panning.
Exactly, a wild analogy would be that ED4 is like J3AR...naturally sharp and very clear looking, http://forum-images.hardware.fr/images/perso/loomy.gif it.

way too much strangely colored chroma noise.
Good to know that it's not my imagination, I see very faint color fringing and my brain is wondering whether I'm drunk, high or if my eyes or visual cortex cannot be trusted anymore...it creates quite a bit of a kernel panic, now to understand why nobody else is seeing it :o

A4 i'm alway's trying to make the picture look better. Very distracting
Weird, what's the native CR of your display? Is it perfectly REC709/D65 calibrated? Is its "sharpness" setting set to neutral?

When the image gets too sharp, the frames move robotically and you lose the "natural feel" of the video.
Are you watching 24/25p in 60Hz? I use Reclock in a multiple refresh rate and everything's butter smooth here. I hate artificial sharpening and mVR has never done any of that IME.

One that focuses on the vividness of the colors
The only proper way to watch videos is D65 calibration within the right gamut(mapped if necessary), if you want to play around with color saturation I'm sure JanWillem has a few PS scripts for you (http://forum.doom9.org/showthread.php?t=157634).

I bet the colored noise looks very different to you than to us.
I only lack some red cones, it's not so bad as some colorblinds are wayyy worse.

I could also see how 4:2:2 mode doesn't play nice with colored noise. I wonder if you could try some tests in a 4:4:4 mode so your opinion is less tied to how your TV processes RGB -> YUV -> RGB? What if you get a better display that can do everything you want while keeping RGB.
I cannot stand LCD without BFI and there are only <=27" computer monitors that allow 4:4:4 BFI in my price range(<$1K roughly) AFAIK, I really do need a 32 incher and at least 3K:1 CR. Without BFI, my 32" 3500:1 A-MVA panel is motion-blurry and sluggish as clearly explained here (http://www.tftcentral.co.uk/articles/content/benq_fp241w_3.htm).

YV12 is 4:2:0, 4:4:4 kinda seems overkill IMHO...and I already find 32x NNEDI chroma artificial looking in 4:2:2, OTOH J3AR looks fantastic(and the TV engine runs in 10/12bit) :)

[fairchild]

I've just gone back to vanilla random dithering until this new dithering pattern is hashed out. I honestly have gotten dizzy from all the dithering talk and different methods being used and people flip-flopping on their favorite version. ;)

I can't even see a difference between random dithering and no dithering on good content. (own 166 blu-ray discs currently) This is on a calibrated 55" VT60 plasma using a quality meter to rec709/d65. (colormunki photo spectrometer) I'm sure I'll be happy with whatever the consensus is and what madshi decides ultimately.

[6233638]

Tried that, but it doesn't look better than dymamic to me. Here are 2 test builds for you with this solution:

http://madshi.net/madVRtemporal.rarThis seems to be worse than the "opposite" builds. The noise pattern is a lot more obvious, and it doesn't seem to do as good a job blending tones.

Why exactly is it beneficial for us, when we insert even more noise (which is colored, too) into black areas, where there should only be gray dots. I just don´t get it. I find it extremely distracting, when I see pink, green and blue dots (which there are a lot of) in my shadow details, which was exactly the reason why we wanted another algorithm, that is more accurate than RD.If you look at your images, you will see that a lot more of the image is "filled in" with the colored builds. With monochrome dither there are larger gaps in the dither pattern which add texture/noise.

When you are looking at an 8-bit image (or even a 4-bit image) rather than a contrast-boosted one, it should blend together to create smoother tones.

I find mono clearer looking, noise in the last one is lower but it seems to be hiding the original grain of the source? Seriously that big round stuff behind him is sharper with mono and even blurrier with opposite, isn't it :confused:"sharpness" differences like that are due to how our perception works. Higher noise is often perceived as being a sharper image due to the texture it adds.

PS: I'm using a 4:2:2 TV with BFI enabled, its "sharpness" setting set to neutral and low debanding forced in mVR BTW.There's a very good chance that viewing in 4:2:2 will be affecting the results. I do all my testing in 4:4:4.

I wonder if some people are testing with displays which are less than 8-bit native too.

And I'm finally back to madVRadaptive4, I feel like I'm watching through to the TV instead of at the TV itself, my eyes instantly feel at home and can focus on that good ole monochrome movie grain, all those color builds feel like a bad dream and I'm stunned all over again :)How do you feel "adaptive4" compares to "mono static" ? (not mono static temporal)



EDIT: I have just done some more testing with the "static" builds, and I know what it is that I don't like about them.

When the camera moves, e.g. in a hand-held shot where there is a subtle swaying motion, because the dither pattern is static it looks like there is something on the lens, as it moves with the camera.

With the dynamic builds, it's randomized so this does not happen.
Yes the noise level is somewhat higher, but I just don't see it being an issue when viewing at 8-bit.

[Asmodian]

YV12 is 4:2:0, 4:4:4 kinda seems overkill IMHO...and I already find 32x NNEDI chroma artificial looking in 4:2:2, OTOH J3AR looks fantastic(and the TV engine runs in 10/12bit) :)

But you are judging the results of dithering which is absolutely in native 4:4:4 (RGB even). What does the conversion to 4:2:2 do to the error that has been spread out by the dithering? I am sure no BFI looks worse but I would hate to end up with a dither algorithm that is optimized for a 4:2:2 screen. 4:2:2 is a complete deal breaker for me. :(

I wish I still had a bigger screen so I could do my own testing. :o

[madshi]

Why exactly is it beneficial for us, when we insert even more noise (which is colored, too) into black areas, where there should only be gray dots. I just don´t get it. I find it extremely distracting, when I see pink, green and blue dots (which there are a lot of) in my shadow details, which was exactly the reason why we wanted another algorithm, that is more accurate than RD.
No, the reason why we wanted another algorithm got nothing to do with colored versus monochromatic noise. I could modify the random dithering to be mono colored, too. And static, if I wanted to. The key difference between error diffusion and random dithering is that error diffusion can achieve the same (or better) smoothness with a noticeably lower noise floor. Whether we use mono colored noise or multi colored noise, or whether we use static or dynamic noise patterns, is independent of whether we use random dithering or error diffusion.

I explicitly asked you guys to concentrate more on real video content instead of gray ramps with the latest test builds, because, and I *hope* we all agree on this, real movie content is what counts in the end, not gray ramps.

And it's incorrect that you say that we add more noise. That's simply not true. If you look at real video content, the oppositeColor build has the lowest subjective noise level. Look here:

| -- monoColor (http://madshi.net/monoColor.png) -- | -- multiColor (http://madshi.net/multiColor.png) -- | -- oppositeColor (http://madshi.net/oppositeColor.png) -- | -- original (http://madshi.net/edTestOrg.png) -- |

Now which of the three 4bit images looks closest to the 8bit original (I think there's no big difference there)? Which of the 4bit images has the lowest subjective noise level (definitely oppositeColor, IMHO)? Tell me, do you want to judge image quality by staring at gray ramps? Or by testing with actual movie content? And then, if you do want to test with a gray ramp, use one which moves (gets darker/brighter) and then compare in motion. I think you'll not see colored pixels. (Well, maybe leeperry will, what do I know).

If you can find real world material (e.g. a black & white movie) where you can see colored dither dots while watching the movie in motion then *that* is something I would consider a good argument for at least offering an option to use monochromatic noise instead of colored noise.

By the way: why use D3D9?
I've read AMD's interop proces with D3D10 and D3D11 is way better optimised. I guess you must have a very good reason...
I use D3D9 because I want to support XP, too, and because all current DXVA decoders are still based on D3D9 (AFAIK). Besides, the way the current madVR windowed mode works, wouldn't be possible with D3D11. So I believe if I tried to use D3D11 for windowed mode, windowed mode playback would get worse. I doubt that D3D11 <-> OpenCL interop works faster than D3D9 <-> OpenCL, but I haven't tested it.

I suppose it would indeed be a problem if the green channel didn't get enough noise, conversely if you raise the noise such that the green channel does have enough noise then this will probably make the red and green channels to noisy. Anyway, I'm beginning to suspect that the method you used in the "oppositeColor" algorithm actually is optimal, in the sense that of all the 'reasonable' ways to create noise it is the one that causes the least amount of Y noise.
I think so, too. At least I see no better solution right now.

I find mono clearer looking, noise in the last one is lower but it seems to be hiding the original grain of the source? Seriously that big round stuff behind him is sharper with mono and even blurrier with opposite, isn't it :confused:
Let me add the original image to make the comparison more helpful:

| -- monoColor (http://madshi.net/monoColor.png) -- | -- multiColor (http://madshi.net/multiColor.png) -- | -- oppositeColor (http://madshi.net/oppositeColor.png) -- | -- original (http://madshi.net/edTestOrg.png) -- |

As you can see, there's no grain in the original image, and the background is supposed to be blurred. Ok, so this test image does not help us decide which of the error diffusion algorithms reproduces film grain in the best possible way. We need a different video/image for that. But this image does show us which error diffusion algorithm produces the lowest subjective luma noise levels. Agreed?

To be more specific its using DirectCompute for Error Diffusion, but OpenCL for NNEDI3. (Correct madshi?)
Meaning error diffusion should work across the board for everyone while NNEDI3 will only work for ATI users which includes you with your A10. (Nvidia hasn't put out a fix / corrected drivers, and its so slow on Intel GPU that its unusable). :(
Correct.

madVRadaptive4 : glorious PQ, everything looks natural, the impression of depth is unbelievable, I don't feel like I'm watching a movie at all. This looks extremely real.

madVR_monoColor_static_temporal : like a nasty constant rehashing of the dither grain, its cadence is annoying...it messes with the movie original grain, the sense of depth of the source is partly lost and the picture looks fuzzier with no added value. This is way too noisy and too predictable.

madVR_monoColor_dynamic : feels like a poorly executed x264 encode of madVRadaptive4, clarity is gone and the dithering grain disobeys to what the original movie was conveying.

madVR_monoColor_dynamic_temporal : feels like the dithering grain has learned new dance steps, the darn thing is doing 3 steps tango on top of my movies now.

madVR_multiColor_static: I still find the picture uneasy to look at but alright, maybe it's nocebo somehow...noise is lower but movements look blurry just like those three screenshots of yours look to me. The "natural" and "glorious" PQ of Oblivion is now replaced by softness, way not sharp enough to my taste.

madVR_multiColor_dynamic : The infamous dither grain dance on top of the motion bluriness, yay.

madVR_oppositeColor_static : Impression of depth is very poor, not impressed at all anymore. Still very(more?) blurry movements, still uneasy to look at(even more?).

madVR_oppositeColor_dynamic : Sharpness & sense of depth are much improved over the latter, but still not natural looking or impressive at all for that matter. There's stll that artificial grain that keeps dancing around on top of the movie grain, that's annoying and kills the the reel grain to my eyes. It's the least worse of the color builds to my eyes.

And I'm finally back to madVRadaptive4, I feel like I'm watching through to the TV instead of at the TV itself, my eyes instantly feel at home and can focus on that good ole monochrome movie grain, all those color builds feel like a bad dream and I'm stunned all over again :)
Ok, thanks!

Well, that big round stuff behind the guy is sharper with mono and even blurrier with opposite, isn't it :confused:
It's noisier, not sharper. See the original image where that big round stuff is very blurry.

OppositeColor dynamic looks to have a tiny bit lower subjective noise level then multiColor dynamic. 4-bits are almost enough to look good!. My preference lean towards oppositeColor dynamic now.

Letting different color channels cancel out some noise and some temporal smoothing(dynamic thresholding noise) to lower the subjective noise further is looking like a good idea.
The temporal tries didn't work out so well for now, unfortunately.

oppositeColor was horrible to my eyes, it produces way too much strangely colored chroma noise. Not worth the trade-off at all, and seems to harm video quality greatly.
Can you upload a small sample where this is visible in motion?

monoColor while it did have the most luma noise, I see this as actually beneficial. The dither indeed has a more neutral impact on colors, while the luma noise has a side effect of improving sharpness and definition of the image slightly, compared to multiColor which has a softer appearance.
Yesterday I watched Americal Idol in 4bit and I found the image to be nicely "crisp". Then I switched to 8bit, and it was subjectively so blurry! In the same way watching the movies in 10bit will look (a tiny bit) softer than watching them in 8bit. Having a higher luma noise floor can make the image appear more detailed. But that's most definitely *not* the purpose of error diffusion. The main purpose of error diffusion is to simulate a higher bitdepth as much as possible. So the better the error diffusion works, the less crisp the image might actually appear, compared to a higher-noise alternative. So what we should be aiming for is not crispness, but low noise smoothness. If you want that small amount of added crispness, you can add artificial grain, or sharpen the image, or tell madVR that your display can only do 7bit or 6bit.

On a scale of 1 (worst) to 10 (best), I'd rate them as follows for the static builds:

monoColor = 8
multiColor = 6
oppositeColor = 1
:( It seems some users hate oppositeColor, while the majority prefers it over the other solutions. That is bad...

I would suggest releasing 3 builds btw:
- One that focuses on the vividness of the colors, which would be great for anime.
- One that focuses on the sharpness of the image, which would be great for documentaries, sports, etc.
- One that focuses on balance, which is the best for viewing moves and other videos - basically the one that's going to be the most widely used.
This would only be possible if it were objectively possible to figure out which build does which best. And we're not all in agreement about this. So I don't think an option like this is going to happen.

And I'm finally back to madVRadaptive4, I feel like I'm watching through to the TV instead of at the TV itself, my eyes instantly feel at home and can focus on that good ole monochrome movie grain, all those color builds feel like a bad dream and I'm stunned all over again

I can't seem to go back to A4 as to me it looks like something is missing? I'm so far into the picture with ODC build
I don't need to concentrate on the grain but the movie it's self. A4 i'm alway's trying to make the picture look better.
Very distracting indeed.
I'm confused. Didn't you say you prefered oppositeColor_dynamic? So you changed your mind again and now you prefer adaptive4 again? At least your first paragraph reads like that. But then in your 2nd paragraph you're saying adaptive4 looks like there's something missing, which is a negative judgement? So I'm not sure right now whether you like adaptive4 or not.

"sharpness" differences like that are due to how our perception works. Higher noise is often perceived as being a sharper image due to the texture it adds.
Yep.

There's a very good chance that viewing in 4:2:2 will be affecting the results. I do all my testing in 4:4:4.
I'm not sure about this. I think having the TV downscale the chroma from 4:4:4 to 4:2:2 should actually smooth some of the chroma noise away. Ok, maybe with 4bit dithering the chroma downscaling could produce somewhat wrong results because the colors are so far away (the downscaling should be done in linear light, but isn't). But this shouldn't be a problem with higher bitdepths, where the colors are rather near to each other.

EDIT: I have just done some more testing with the "static" builds, and I know what it is that I don't like about them.

When the camera moves, e.g. in a hand-held shot where there is a subtle swaying motion, because the dither pattern is static it looks like there is something on the lens, as it moves with the camera.

With the dynamic builds, it's randomized so this does not happen.
Yes the noise level is somewhat higher, but I just don't see it being an issue when viewing at 8-bit.
Yes, this effect is what I earlier called the "dirty screen effect". You're describing it as a "dirty lens effect", which is basically the same thing. I do wonder, though, whether this effect is visible/distracting at all at 8bit. If it's not, maybe it'd be better to stick to the static builds to keep the subjective noise as low as possible? I'm on the fence on this one. It's definitely a trade between lower subjective noise floor (pro static) and dirty screen/lens effect (contra static).

[GREG1292]

Leeperry not a projector calibration issue a noise issue and lack of depth and shadow detail. I to am only concerned about the picture but 92" from
10 feet it is very easy to see the differences in all these test builds. You have great eyes and I have learned a lot from your expertise about what to
look for in an image.


Sent from my iPhone using Tapatalk

[madshi]

http://madshi.net/edClock.png

To my eyes this looks something like 4bit, 4.5bit, 5bit and 8bit. Please note how the images with error diffusion look "crisper" than the original. This is why I'm saying that when judging the error diffusion algorithms we shouldn't aim for crispness, but for the smoothest result with the lowest noise level. That should actually bring out the original source detail the best, even though it doesn't look as crisp.

Anybody willing to do a similar test with a movie which has lots of grain and detail?

@leeperry, do you like ClearType? I think that's somewhat similar to Multi/OppositeColor. I've read that some people are bothered by ClearType and prefer it turned off, while the majority of people prefer it turned on.

[6233638]

If you can find real world material (e.g. a black & white movie) where you can see colored dither dots while watching the movie in motion then *that* is something I would consider a good argument for at least offering an option to use monochromatic noise instead of colored noise.For what it's worth, some of my testing has included black and white content, and some films which appear to be black and white, but actually have a color tint to them, and I still found the colored builds to have the lowest subjective noise level, compared to the monochromatic ones.

I use D3D9 because I want to support XPI understand that you might still want to continue supporting it, but even Microsoft is dropping XP support in 7 weeks time.
Considering the increasing hardware requirements for madVR, I have to wonder who is using modern hardware and still running XP.

Yesterday I watched Americal Idol in 4bit and I found the image to be nicely "crisp". Then I switched to 8bit, and it was subjectively so blurry! In the same way watching the movies in 10bit will look (a tiny bit) softer than watching them in 8bit. Having a higher luma noise floor can make the image appear more detailed. But that's most definitely *not* the purpose of error diffusion. The main purpose of error diffusion is to simulate a higher bitdepth as much as possible. So the better the error diffusion works, the less crisp the image might actually appear, compared to a higher-noise alternative. So what we should be aiming for is not crispness, but low noise smoothness. If you want that small amount of added crispness, you can add artificial grain, or sharpen the image, or tell madVR that your display can only do 7bit or 6bit.It's funny, as I noticed the same thing. It makes me think that it might be worthwhile for madVR to have some sort of way to add grain to the image at some point. (but obviously not through the dithering process)

I'm not sure about this. I think having the TV downscale the chroma from 4:4:4 to 4:2:2 should actually smooth some of the chroma noise away. Ok, maybe with 4bit dithering the chroma downscaling could produce somewhat wrong results because the colors are so far away (the downscaling should be done in linear light, but isn't). But this shouldn't be a problem with higher bitdepths, where the colors are rather near to each other.When the RGB > 4:2:2 conversion is out of our control, it's difficult to predict what the results will be. I agree that the blurring could possibly be beneficial, but we can't say that it will be for certain, on all displays.
This is why I make sure that my display is operating in a 4:4:4 mode when testing.

Yes, this effect is what I earlier called the "dirty screen effect". You're describing it as a "dirty lens effect", which is basically the same thing. I do wonder, though, whether this effect is visible/distracting at all at 8bit. If it's not, maybe it'd be better to stick to the static builds to keep the subjective noise as low as possible? I'm on the fence on this one. It's definitely a trade between lower subjective noise floor (pro static) and dirty screen/lens effect (contra static).Not that I'm saying you should bring it back, but I wonder if it's the same issue where dithered areas next to undithered areas may be more noticeable than if the whole image were dithered. (i.e. limited vs unlimited builds) Obviously dithering everything would add noise though.

I still haven't been able to find anything where the dynamic builds add noticeable noise or "energy" to the image when viewing in 8-bit, compared to the static builds.
And while it does add noise or "energy" to the image in 4-bit, I think it is very much preferable to the "dirty screen effect".

@leeperry, do you like ClearType? I think that's somewhat similar to Multi/OppositeColor. I've read that some people are bothered by ClearType and prefer it turned off, while the majority of people prefer it turned on.I actually really dislike the default ClearType settings. I find that the subpixel rendering is far too saturated, and color fringing is very noticeable to me.
But I prefer using the less saturated version of ClearType (available via the ClearType tuner) to using monochrome rendering, as this still gives you most of the resolution benefits without obvious color fringing.

[madshi]

For what it's worth, some of my testing has included black and white content, and some films which appear to be black and white, but actually have a color tint to them, and I still found the colored builds to have the lowest subjective noise level, compared to the monochromatic ones.
Ok, good to know.

I understand that you might still want to continue supporting it, but even Microsoft is dropping XP support in 7 weeks time.
Considering the increasing hardware requirements for madVR, I have to wonder who is using modern hardware and still running XP.
True. But XP support is only one out of several reasons why I'm using D3D9.

Not that I'm saying you should bring it back, but I wonder if it's the same issue where dithered areas next to undithered areas may be more noticeable than if the whole image were dithered. (i.e. limited vs unlimited builds) Obviously dithering everything would add noise though.
The limiter doesn't disable dithering, it just limits the noise to a reasonable range. E.g. for a 10.2 pixel value the limiter only allows error diffusion to use pixel values of 10 and 11, but not 9 or 12.

I still haven't been able to find anything where the dynamic builds add noticeable noise or "energy" to the image when viewing in 8-bit, compared to the static builds.
And while it does add noise or "energy" to the image in 4-bit, I think it is very much preferable to the "dirty screen effect".
I think I do see a slightly higher noise floor in 8bit with the dynamic builds (at least when using monoColor), and I'm not sure I see the dirty screen effect at 8bit, so I'm not sure whether the static builds wouldn't be better in 8bit. But in 4bit I definitely prefer the dynamic builds. So I'm still on the fence about whether to use static or dynamic at 8bit.

[6233638]

I think I do see a slightly higher noise floor in 8bit with the dynamic builds (at least when using monoColor), and I'm not sure I see the dirty screen effect at 8bit, so I'm not sure whether the static builds wouldn't be better in 8bit. But in 4bit I definitely prefer the dynamic builds. So I'm still on the fence about whether to use static or dynamic at 8bit.On images where the dynamic noise is visible (though I'm still not really seeing it in 8-bit images - I wonder if it's because people are testing on displays which are less than 8-bit?) it should be safe to assume that the static noise pattern would be equally visible.

I would much rather see the dynamic noise than the "dirty screen" effect.

In my testing using the 4-bit mode though, I find that there are images where additional noise from the dynamic builds is very difficult to see, but the "dirty screen effect" is always obvious.

To my eyes this looks something like 4bit, 4.5bit, 5bit and 8bit.I agree. Here's another example taken from a random patch of sky:

http://abload.de/img/ditheringffu1i.gif

The colored and opposite builds show much less noise than the mono one.

[James Freeman]

I now see this luma noise difference you were talking about. That said...

oppositeColor was horrible to my eyes, it produces way too much strangely colored chroma noise.

monoColor while it did have the most luma noise, I see this as actually beneficial.



So what we should be aiming for is not crispness, but low noise smoothness.

Well there's you problem.

cyberbeing is taking chroma noise heavily into account, which is pushed to the extreme in 4-bit, and does not translate 1:1 to 8-bit.
Its been said by madshi & other people, you should concentrate on subjective/apparent Smoothness & Noise level, and less on Color...

I would much rather see the dynamic noise than the "dirty screen" effect.
I think I do see a slightly higher noise floor in 8bit with the dynamic builds (at least when using monoColor), and I'm not sure I see the dirty screen effect at 8bit,
so I'm not sure whether the static builds wouldn't be better in 8bit. But in 4bit I definitely prefer the dynamic builds. So I'm still on the fence about whether to use static or dynamic at 8bit.

I have the same thoughts,

Should we fear the Static builds "dirt on screen" effect that much, or will it be insignificant in 8-bit?
The random noise of the dynamic builds helps to cover the 4-bit big step distance to our human eyes thus looking smoother, but will it be that beneficial in 8-bit?

Is the dynamic build worth the extra noise in 8-bit?... That's the question.
In 4-bit, Yes.
In 8-bit, Hmm.

[6233638]

Is the dynamic build worth the extra noise in 8-bit?... That's the question.If you can see increased noise in the dynamic builds with 8-bit images (I've still to find any samples where that is the case, rather than noise encoded into the source) then you will be able to see the effects of static dither patterns too.

In the 4-bit builds where the effects of either are clear, dynamic is less harmful to the image than static.
Static dither patterns are very obvious, dynamic adds noise, but blends in with the image more.

I can't think of any reason why that would change as you increase the bit-depth.