madVR - high quality video renderer (GPU assisted)

[CruNcher]

How much more efficient is NEEDI 3 compared to Intels Lanczos 4 Hardware implementation you can access via their Media SDK did somebody compared this ?
Im currently using Bay Trail for HTPC usage and i think it's unbeatable currently @ 4W so changing to ~250W is a huge step so the benefits needs to be huge as well with NEEDI 3 as a DirectCompute solution

[nevcairiel]

I'm afraid that it is not that you have chroma noise instead of luma noise, but rather that you have chroma noise on top of the luma noise

Since we're dealing in RGB here, you can't really separate the two anymore.

[huhn]

How intensive is OpenCL Error Diffusion? My HTPC is decently powerful (Radeon 5850), but when that option is active, my videos play back at what seems like one or two frames per second.

Also, does anyone use a Kuro for 24 Hz playback? I'm trying to figure out why going into full-screen mode in MPC-HC makes all my Blu-ray .mkv rips appear letterboxed when my HTPC is set for 24 Hz output. It doesn't happen when I'm running the panel at 60 Hz or when I'm playing back an .mkv in a window on the desktop.

have you check the overscan settings for all resukotion x refresh rate combos? you have to disable all the overscan for all of them.

and double check 23 hz the real bd refreshrate you are maybe using windows 8?

[madshi]

dc/4/5 my old pc(ati 3650) full black screen dont work madvr...
DirectCompute requires a DirectX 11 compatible graphics card, which means at least 5000 series for ATI/AMD, I believe.
Yes, the DirectCompute (and probably also the OpenCL stuff) will not work with the 3650. However, madVR should detect that and silently disable those features. So a full black screen should not happen. Don't know why that happens. Does v0.87.4 show a correct image?

I don't think that's a good idea, chroma noise is very visible on on blacks. It looks more brown than black in those cases. I think this is because the colors are not perceptually uniform. Varying blue levels doesn't do much, but adding some more red when there shouldn't be any makes it brownish.
I didn't mean to artificially introduce chroma noise. My point was that if it so happens that the current error diffusion algorithm produces chroma noise, I don't think it should be much of a problem.

To me the flat, grey bars still look smoother in build 3. In iSunrise's enhanced screenshots this is especially visible on bar #19. Build 3 produces a fine, uniform lattice which gives it a smooth appearance since it visually blurs together. The new random weights on the other hand create clusters of the same color which are more noticeable.
Well, test build 3 has worm artifacts, test build 5 has the dots more randomly placed. You can't have both. Either one or the other. So you prefer worm artifacts over the randomly placed dots of test build 5?

Much better than Test Build 4 (http://abload.de/img/directcompute-4ocji7.png), but still enough to prevent the display from turning off dimming zones.
Well, I just had an idea: Custom shaders are called before the 3dlut. Custom output levels are applied after the 3dlut. So you could use a custom shader to stretch the video from 16-235 to maybe 16.75-235. And then you could use custom output levels to stretch the content back to 15.75-235 (by using output levels 15-235). I think this should make your dimming zones happy without losing shadow detail.

By the way does "The problem is not even the math instructions, but the added registers the compiler would need to spend on this." mean that you could do the calculations but this would require you to write and read things to memory which takes more time? Because with some minor approximations I think it might actually be possible to do some of the calculations beforehand, in parallel, although if it then takes too much time to read the answers then that obviously doesn't help.
No, writing things to memory would be very much slower than using registers. The problem is that the more code I add to the Error Diffusion algorithm, the more registers the DirectCompute compiler will spend. And if the number of needed registers exceeds a certain value, the GPU may only run at half the speed. The exact number of registers depends on the exact GPU architecture, though, so it's hard for me to test whether I've reached a critical number of registers or not.

Maybe you could allow us to choose between several dithering algorithms? I would more than likely stick with DC3 over 4/5.
Also I think that in cases where the worm patterns themselves aren't visible it is actually beneficial to use those patterns instead of a randomized version since those patterns are closest to the actual colour and any attempt to remove the pattern will make those colours more 'noisy'. So normal error diffusion will probably be better for screens with a high pixel density whereas randomized error diffusion will be better for screens with a lower pixel density. I suspect that the latter case will profit the most from using error diffusion, so randomized error diffusion should probably be the default but it would be nice if you could still switch to normal error diffusion.
I don't really like the idea of adding many new options for dithering. Most users won't even see a difference between random dithering and error diffusion. I already know that there will be dozens of users asking in this thread whether they should use error diffusion or not. If I now add an option to use error diffusion variant A or B, that will only make things worse.

Is there no clear preference for either test build 3 or 5?

How much more efficient is NEEDI 3 compared to Intels Lanczos 4 Hardware implementation you can access via their Media SDK did somebody compared this ?
Im currently using Bay Trail for HTPC usage and i think it's unbeatable currently @ 4W so changing to ~250W is a huge step so the benefits needs to be huge as well with NEEDI 3 as a DirectCompute solution
How do you define "efficient"? Are you talking about performance? Or quality? Or some sort of quality per performance ratio? NNEDI3 is very slow, if that's what you're asking.

Some users did compare Intel's Lanczos 4 hardware implementation to madVR's own Lanczos 4 Anti-Ringing implementation and found madVR's algorithm to have a better image quality (but lower performance). Jinc is IMHO better than Lanczos, and NNEDI3 is better still. But performance drops quite noticeably from Lanczos to Jinc and from Jinc to NNEDI3.

The difference in quality when using NNEDI3 can be very large or it can be non-visible. A lot depends on the source material and the scaling factor. If we're talking about 720->1080p upscaling, the benefit of NNEDI3 will be smaller than when doing 480->1080p. And if your source material is rather soft, NNEDI3 will do nothing for you. NNEDI3 shines with heavily aliased sources (e.g. some SD Animes), and it also shines with really sharp high quality sources when using rather large upscaling factors. Whether the image quality benefit is worth the cost (in dollars and power consumption) is not a question anybody else can answer for you.

Look here for a short discussion about NNEDI3 image quality in comparison to Lanczos and Jinc:

http://forum.doom9.org/showpost.php?p=1663611&postcount=21592

[iSunrise]

Are you sure you've compared the same frame? In theory this is static content, but it's encoded with a lossy video codec, so it's not really static, anymore. I think it's quite probable that the white line is there in some frames and not in others. Not sure about the brightness of bar 17. The error diffusion is now slightly more exact in test build 5 than it was before. But the difference is pretty small so it would surprise me if that would explain the different bar 17 brightness level. Maybe it's also a frame-by-frame difference in the video file?
You´re actually right, for some reason this one was not the same frame, even though I always carefully jumped manually to 00:00:04.389 before I captured. So now I´ve double-check everything, even my madVR settings (disabled smooth motion) and here are the new shots (yes, I did everything again from scratch):

DC3 vs. DC4 vs. DC5 vs. RD

http://abload.de/thumb/directcompute3_enhanc7xkhh.png (http://abload.de/image.php?img=directcompute3_enhanc7xkhh.png)http://abload.de/thumb/directcompute4_enhancckkii.png (http://abload.de/image.php?img=directcompute4_enhancckkii.png)http://abload.de/thumb/directcompute5_enhanc5bjts.png (http://abload.de/image.php?img=directcompute5_enhanc5bjts.png)http://abload.de/thumb/random_dither_enhancemwj3u.png (http://abload.de/image.php?img=random_dither_enhancemwj3u.png)

OK, so the white line is a non-issue now, it is actually where it should be. I still see the slight brightness difference on bar 17 though (it actually also looks a bit thinner) and that is only happening with DC5, all others are fine. And by looking at all example shots, DC3 still looks smoother to me. That´s completely subjective, though, but others seems to agree. I base that on the much nearer abbreviation of the bars greys, while DC5 seems to add too much noise for no benefit at all. Also, 6233638 would not have to manually tinker with his settings because of DC5. But your decision is the final one, of course.

[CruNcher]

Quality/Perf/Power ratio yes i see thx madshi and great work on the optimization :)

is their any Hardware implementation of NNEDI 3 existing yet FPGA or DSP ?

What are the results with Chaotic Movements like Water and Fire ?

[madshi]

So it seems most people prefer the original test build 3 implementation, despite the worm artifacts? Anybody prefering test build 5?

[GREG1292]

I'm sorry, I'm mostly using mVR to watch movies from a 1 or 2 meters distance, I don't primarly use mVR to zoom at 3200% on nasty test patterns and come whine in here for a fix on stuff nobody would ever see from a distance(j/k :p).

The primary thing I've always been looking for in this hobby is the subjective pop effect and I've compared DC3/4/5 several times.....to my brain DC3 provides a much more pronounced depth impression(dead obvious on that sample really). But then again, many things come at play here: how my TV dithers its 10/12bit internal processing to 8bit, the fact that I'm colorblind on red, how its antiglare layer pearly pattern will react, how noisy the encode is, etc etc....

Maybe you could allow us to choose between several dithering algorithms? I would more than likely stick with DC3 over 4/5.

There's no one-size-fits all scaler, there might very well not be no "best" dithering algorithm either, especially if you start adding noise to chroma. DC3 looks stunning, I see the same pop effect as I get from SmoothL FWIW :cool:


I agree using a projector and for movie and gaming only I'm staying
with 3 also. I to notice more the svp 3.15 effect "pop" using gtx770 with
I7 1230v3 and 327 drivers. Sorry can't be more scientific. Just another
happy user of MadVr.




Sent from my iPhone using Tapatalk

[madshi]

is their any Hardware implementation of NNEDI 3 existing yet FPGA or DSP ?
Not that I know of.

[Shiandow]

I don't really like the idea of adding many new options for dithering. Most users won't even see a difference between random dithering and error diffusion. I already know that there will be dozens of users asking in this thread whether they should use error diffusion or not. If I now add an option to use error diffusion variant A or B, that will only make things worse.

Is there no clear preference for either test build 3 or 5?


I personally prefer build 3 since it seems to have the least amount of noise, by far. Which is precisely what error diffusion was meant to do.

[bacondither]

If you want the least amount of noise then disable dithering...

[James Freeman]

Is there no clear preference for either test build 3 or 5?

I personally prefer build 3 since it seems to have the least amount of noise, by far. Which is precisely what error diffusion was meant to do.

Agreed.
DC3 for me too... Silky Smooth, no added Noise, No visible artifacts/patterns in normal viewing conditions (no extreme Photoshop to enhance the invisible), no performance hit, no dithered blacks problems.

Madshi, why not just add a small dropdown box with "Static" or "Random" options when ED selected like you've mentioned (not A,B)?

If you want the least amount of noise then disable dithering...

No.

[madshi]

So far I'm counting 6 votes for test build 3, and none yet for test build 5.

[nevcairiel]

I agree that clean ED is preferable. I don't think you can actually see the pattern on a real world screen.

[Ver Greeneyes]

I kind of like the random look better in the brightened images that were posted, but considering it causes those issues for 6233638 (which I'm worried will also impact my monitor) and does add some level of noise regardless, I don't think it's worth the trouble. So 3 for me as the least dangerous option, but thanks for experimenting with it :)

[The 8472]

If i understand correctly all the test pictures posted so far were from 8bit sources.

When I get home I'll try generating some colored and grey high bit depth shallow gradients and feed them to madvr. Although I have none of the required toolchains set up, so it'll take a while. I'll try avisynth or imagemagick + ffmpeg.
If someone else has everything configured already and could assemble some test-cases faster than me that would be fine too.

[6233638]

Well, I just had an idea: Custom shaders are called before the 3dlut. Custom output levels are applied after the 3dlut. So you could use a custom shader to stretch the video from 16-235 to maybe 16.75-235. And then you could use custom output levels to stretch the content back to 15.75-235 (by using output levels 15-235). I think this should make your dimming zones happy without losing shadow detail.I suppose if it's done with enough precision it's probably not going to harm image quality, but I don't really like the idea of all these different conversion steps.

How would I do something like that, to try it?
madVR is outputting 0-255 to a PC level display.

So it seems most people prefer the original test build 3 implementation, despite the worm artifacts? Anybody prefering test build 5?There's definitely more visible noise once you start adding randomness, though it does do a good job avoiding the "worm artifacts".

Unfortunately I don't have an example I can refer back to, but when I was editing values near black in the 3DLUT, it did produce some visible artifacts on full field patterns - even at a distance.

I haven't noticed anything like that from actual use so far though.

[bacondither]

Plain simple error diffusion has greatly lower noise level then random dithering but it adds ugly patterns which random dithering is free from.

A good compromise then is to add a bit of randomness to the weights of the error diffusion code and eliminating these patterns and at the same time keep the noise much lower then random dithering but with no other artifacts.

It´s for the most times a good compromise.

6233638s problem is some bug with 3dlut implementation and should be solved separately.

[leeperry]

OK, good to read that I'm not all that crazy after all :)

Well, the recipe to pop-effect IME is:
1) as high of a native CR as possible
2) a 2.4'ish gamma curve
3) dead smooth playback with as low of an interframe jitter as possible(FRC with a high refresh rate helps a lot on CRT but LCD is too slow for this trick, even with BFI)
4) as little (random) noise as possible, as this will add a veil of noise that will give a "uniform" look to the picture and decrease dynamics....I just ran a whole bunch of comparisons on that very clean looking danish movie I used for the aforementioned sample again and 3 is king at this game to my eyes, kudos²!

[DarkSpace]

Test build 3 for me also.
If I had to rate, I'd choose 3, then 5, and then 4. I also noticed that I find the "random" noise to somewhat alter the display of colors or brightness, while with test build 3, the colors look more uniform (which, looking at the images posted here so far, seems to be just how it is supposed to be).

[bacondither]

1-bit dithered images.

Floyd Steinberg ramp (http://i.imgur.com/JNYm325.png)

Floyd Steinberg random weights ramp (http://i.imgur.com/4rZ172G.png)

Your specific implementation may vary.

[madshi]

How would I do something like that, to try it?
madVR is outputting 0-255 to a PC level display.
Oh, in that case it'll not be possible. You'd have to use -1-255 output levels, which madVR doesn't support.

1-bit dithered images.

Floyd Steinberg ramp (http://i.imgur.com/JNYm325.png)

Floyd Steinberg random weights ramp (http://i.imgur.com/4rZ172G.png)

Your specific implementation may vary.
FWIW, instead of the Floyd-Steinberg weights 7/16, 3/16, 5/16 and 1/16, test build 3 is using the weights 0.7770, 0.0090, 0.4861, -0.3098. Yes, that's a negative weight there, and the sum of the weights is a bit less than 1.0. Both of this is intentional...

[nevcairiel]

Now post the same test without extreme Photoshop mutilation.

Its 1-bit dithered, you don't need photoshop to enhance it to make it visible. If its a useful comparison is another question however.

You should really stop trolling this thread.

[James Freeman]

I have never trolled on the internet and never will.
It was not my intention even if I may appear like it.

It is important for me that MadVR will have the best picture, and DC3 is in favor by almost everyone including me.
I still don't understand why bacondither pushing so hard to get random ED and posting endless comparison images of non realistic scenarios.
It does not matter, the crowd has already spoken.

EDIT:
There, I even removed the avatar to be a little more low profile...

[leeperry]

FWIW, instead of the Floyd-Steinberg weights 7/16, 3/16, 5/16 and 1/16, test build 3 is using the weights 0.7770, 0.0090, 0.4861, -0.3098. Yes, that's a negative weight there, and the sum of the weights is a bit less than 1.0. Both of this is intentional...
And vanilla FS wouldn't be worth a shot I guess?

[bacondither]

FWIW, instead of the Floyd-Steinberg weights 7/16, 3/16, 5/16 and 1/16, test build 3 is using the weights 0.7770, 0.0090, 0.4861, -0.3098. Yes, that's a negative weight there, and the sum of the weights is a bit less than 1.0. Both of this is intentional...

I still don't think it looks as good as FS or filter line

madshi weights no random corrected (http://i.imgur.com/8dN9qel.png)

madshi weights +random corrected (http://i.imgur.com/3NrManE.png)

Are you using alternating direction for every row?


//Edit

Filter line gives very similiar results to Floyd Steinberg

http://caca.zoy.org/export/4875/web/trunk/www/study/out/fig3-2-7.png

[madshi]

Eeeeh using your weights it looked really crappy...

madshi weights no random (http://i.imgur.com/RJIFDGu.png)

madshi weights +random (http://i.imgur.com/USrJnM1.png)

Are you using alternating direction for every row?
There must be something wrong, I think? Yes, I'm using alternating direction for every row.

I did try the Floyd-Steinberg weights when implementing the OpenCL error diffusion code, but I didn't like the results much.

[bacondither]

My bad, i mixes some weights up... updating my earlier post.
I need tea!


madshi weights no random corrected (http://i.imgur.com/8dN9qel.png)

madshi weights +random corrected (http://i.imgur.com/3NrManE.png)

[6233638]

1-bit dithered images.

Floyd Steinberg ramp (http://i.imgur.com/JNYm325.png)
Floyd Steinberg random weights ramp (http://i.imgur.com/4rZ172G.png)

Your specific implementation may vary.While this does a much better job avoiding obvious patterns, it does seem to introduce a lot of texture/noise compared to regular Floyd Steinberg.

[bacondither]

While this does a much better job avoiding obvious patterns, it does seem to introduce a lot of texture/noise compared to regular Floyd Steinberg.

You can add less noise and still avoid patterns to a degree but it's a compromise.

I added a high amount of noise to show the effect.

[n3w813]

So it seems most people prefer the original test build 3 implementation, despite the worm artifacts? Anybody prefering test build 5?

Another vote for DC3 here also

OK, good to read that I'm not all that crazy after all :)

Well, the recipe to pop-effect IME is:
1) as high of a native CR as possible
2) a 2.4'ish gamma curve
3) dead smooth playback with as low of an interframe jitter as possible(FRC with a high refresh rate helps a lot on CRT but LCD is too slow for this trick, even with BFI)
4) as little (random) noise as possible, as this will add a veil of noise that will give a "uniform" look to the picture and decrease dynamics....I just ran a whole bunch of comparisons on that very clean looking danish movie I used for the aforementioned sample again and 3 is king at this game to my eyes, kudos²!

I know the 'pop-effect' is very subjective but I need to agree with leeperry here about the increased sense of depth with DC3. When I first switched to DC3 from 0.87.4 without changing any options, the depth perception was so pronounced that even my wife asked if we were watching 3D. :D

After some testing with the new options, I had to lower my upscaling algorithms for chroma and luma from Jinc3AR to Lanczos3AR since my GTX 650 could not handle ED, deband, SM (24>60), and 3DLUT. After changing to Lanczos3AR, the depth perception was less pronounced but still better than 0.87.4. Can't explain it...is gamma/contrast somehow affected with DC3?

[fairchild]

If you want the least amount of noise then disable dithering...

I found that in my case in order for me to match the output of my blu-ray player (a Sony BDP-S5100) I had to disable dithering completely to get the same exact black level setting (brightness). Any dithering increases the black levels very slightly at the 17th step and doesn't allow both settings to match. It doesn't actually affect any measurements that I could tell though once the calibration was done.

[bacondither]

I have modified the Filter Lite error diffusion weights to propagate 97% of the error. It's using alternating direction per row and is not using fully randomised error weights.

Modified Filter Lite with noise (http://i.imgur.com/qxTLWsl.png)

out = double(img);
sz = size(out);

for ii=1:2:sz(1)
for jj=1:sz(2)
old = out(ii,jj);
new = 255*(old >= 128+(rand-0.5)*80);
out(ii,jj) = new;
err = new-old;
if jj<sz(2)
% right
out(ii ,jj+1) = out(ii ,jj+1)-err*(0.485);
end
if ii<sz(1)
% down
out(ii+1,jj ) = out(ii+1,jj )-err*(0.2425);
if jj>1
% left-down
out(ii+1,jj-1) = out(ii+1,jj-1)-err*(0.2425);
end
end
end
ii = ii+1; % even rows go right to left
for jj=sz(2):-1:1
old = out(ii,jj);
new = 255*(old >= 128+(rand-0.5)*80);
out(ii,jj) = new;
err = new-old;
if jj>1
% right
out(ii ,jj-1) = out(ii ,jj-1)-err*(0.485);
end
if ii<sz(1)
% down
out(ii+1,jj ) = out(ii+1,jj )-err*(0.2425);
if jj<sz(2)
% left-down
out(ii+1,jj+1) = out(ii+1,jj+1)-err*(0.2425);
end
end
end
end

out = dip_image(out)

[wolfman2791]

Getting a handle on this slowly lol.... how about display calibration? Should i leave it disabled? If not, what do you choose??

[vivan]

I don't think you can actually see the pattern on a real world screen.Well, there was image (unprocessed) in this thread here (http://forum.doom9.org/showthread.php?p=1666286#post1666286) (now it's deleted, wth somebody is using mf for images hosting?). On my display (crappy laptop TN) there was cleraly visible vertical line, under 22 (or 24).
Not that I really care (that laptop can't handle DC stuff anyway), just saying...

[James Freeman]

I use MF for images, its quicker.
I did some cleaning so all the huge images I've posted are gone.

The vertical line exists in 18, 22 & 24 its part of the test pattern.
Nothing to do with dithering I think.

wolfman2791,
If you have Pal & NTSC dvds or watch SD content, choose Already calibrated to rec.709.
Otherwise leave off.

[wolfman2791]

Thanks... i'll leave it off then.

[leeperry]

When I first switched to DC3 from 0.87.4 without changing any options, the depth perception was so pronounced that even my wife asked if we were watching 3D. :D
The good ole "my wife asked me if I changed anything" argument :D

Quite frankly I don't see the point for 3D when you can get a stunning pop effect from a well tuned 2D rig, especially as you would appear to be forced to wear plastic glasses that kill sharpness(organic glass comes very poor constringence specs), that there are so little titles available and that many of them would be poor 2D to 3D conversions. Even display manufacturers appear to take it easy with 3D, mostly because the Avatar hype is over.

After some testing with the new options, I had to lower my upscaling algorithms for chroma and luma from Jinc3AR to Lanczos3AR since my GTX 650 could not handle ED, deband, SM (24>60), and 3DLUT. After changing to Lanczos3AR, the depth perception was less pronounced but still better than 0.87.4. Can't explain it...is gamma/contrast somehow affected with DC3?
Well, if you want some serious pop effect you gotta try 32/64 NNEDI....I would guess that sharpness also tends to improve the pop effect and Lanczos doesn't look natural at all IME, but FWIR FRC needs quite some horsepower to go through NNEDI.

[Anime Viewer]

Bad news: Currently my DirectCompute NNEDI3 implementation is about 10 times slower than OpenCL. I'm not sure yet why. I'm trying to improve performance, but I'm not hopeful right now. It's very well possible (or rather probable at this point) that we'll have to stick with OpenCL for NNEDI3. So we probably do need that NVidia driver fix, after all...

Is it still the case that your DirectCompute NNEDI3 tests show it to be 10 times slower than with OpenCL?

Would you elaborate on what was slower? Are you referring to the rendering ms speed? What was the hardware (Intel, AMD, Nvivida), and Madvr settings (including downscaling choice)?

Perhaps a different combination of settings would work better with the DirectCompute coding compared to OpenCL, or perhaps results would differ when tested with an Nvidia equipped system compared to an AMD.

With the extreme amount of bugginess the Nvidia systems have with the OpenCL coding it may still be better off (even after Nvidia patches the problem) for Nvidia systems to use the DirectCompute version. While it would undoubtedly increase file size would it be plausible to offer an option where checking (or unchecking) a box in a settings area could switch between the two versions? Something like "use OpenCL to process NNEDI3" and/or "use DirectCompute to process NNEDI3" similar to how you currently have the "use OpenCL to process DXVA NV12 surfaces (Intel, AMD)" option in the rendering - general settings area?

[elraffa]

Which of this settings are better?
chroma upscaling NNEDI3 32neurons and in image doubling, double luma channel 1.5x scaling factor 32 neurons...
or
chroma uscaling NNEDI3 16neurons and in image doubling, double luma channel 1.5x scaling factor 64neurons

[Thunderbolt8]

both only have 81-88 gb/s bandwith I guess thats too low for all that stuff like NEDI etc.

[XMonarchY]

madshi,

Could you be so kind to provide some input to my following questions?

1. Is there a way to render JPEG files with madVR to improve/sharpen them as if they were just single film frames?

2. My Samsung LN40B550 supports full 4:4:4 only in 60Hz mode, which is not as smooth (even with Smooth Motion) as the 48Hz 4:2:2 judder-free mode that I prefer. However, the 48Hz mode has 2 versions - Movie and Standard. Movie 4:2:2 mode creates nasty text discoloration and on your 4:2:2/4:4:4 test image I can see 4:2:2 clearly, but I can also see 4:4:4 with similar nasty/ugly discolorated edges around it. Then there is the Standard 4:2:2 mode that I use because unlike Movie 4:2:2 mode, it actually displays text properly (even with sharpness set to 0)! Using this 4:2:2 Standard mode, the 4:2:2/4:4:4 test image shows both 4:2:2 and 4:4:4, but 4:4:4 edges have just a barely noticeable discoloration. Chroma upscaling/doubling, Luma doubling, image upscaling/downscaling are all targeting 4:2:0 - 4:4:4 chroma subsampling conversion. However, I can most definitely see a huge improvement in Standard 4:2:2 mode in video quality and sharpness when using error diffusion, NNEDI3 chroma and luma doubling/upscaling, etc. So, what exactly am I losing?

3. Should I just not bother with some of these settings since they have no effect?

4. Any ideas why Standard mode shows only a barely noticeable text and edge discoloration, while Movie 4:2:2 mode shows a drastically visible one?

5. I know TotalMedia Theatre does not support madVR, but does PowerDVD 13 support it? Can it be used with madVR to render and playback 3D content?

Thank you so much for your hard work - madVR is truly the one and only solution for HTPC high-end video reproduction that improves the image beyond its native-level quality. Nothing else compares.

[Mangix]

Which of this settings are better?
chroma upscaling NNEDI3 32neurons and in image doubling, double luma channel 1.5x scaling factor 32 neurons...
or
chroma uscaling NNEDI3 16neurons and in image doubling, double luma channel 1.5x scaling factor 64neurons

the latter

[Shiandow]

I have modified the filter line error diffusion weights to propagate 97% of the error. It's using alternating direction per row and is not using fully randomised error weights.

Modified Filter line with noise (http://i.imgur.com/qxTLWsl.png)

out = double(img);
...

What do you think of something like this (http://i.imgur.com/PRJsczo.png)? This is a modified version of the Floyd-Steinberg weights combined with Madshi's weights.

[The 8472]

1. Is there a way to render JPEG files with madVR to improve/sharpen them as if they were just single film frames? You can probably wire up avisynth and nnedi to achieve similar results.

[cyberbeing]

In order to update the OSD, I have to re-render the frame. IIRC I do a full redraw, including all OpenCL/DirectCompute algorithms. Obviously, if the OSD is disabled, I don't have to do that.

Okay, thanks. That's what I'd been looking for confirmation on. Does that mean the OSD actually increases GPU load from re-rendering frames during playback, or is this behavior just something you added by-design to get render stats on paused video?

Is there no clear preference for either test build 3 or 5?
So far I'm counting 6 votes for test build 3, and none yet for test build 5.

Personally, I prefer "random dither" over the noiseless patterns and artifacts of "error diffusion" from test build 3. Test build 5 feels like a decent compromise. I'd rate them as:

build 5 > build 4 > random dither > build 3 > ordered dither

FWIW, "error diffusion" build 5 has less noise on black compared to build 3 when using my 3DLUT on the Grayscale Steps video from the AVS HD 709 MP4-2c.

Enhanced screenshots (identical crops, frame 100): build 3 (http://i1.someimage.com/ckQFvf5.png) & build 5 (http://i1.someimage.com/6qkJ6Uc.png)
On the left is 0-0-0 Fullrange BTB in the source video, and 16-16-16 Video Black on the right.

Looking at other samples, it also appears that build 5 has reduced crosstalk compared to build 3, and less arbitrary errors on the edges of madVR's 16x16 dither blocks.

[XMonarchY]

BTW, I pick Version 3 because it has little to no noise and the overall effect is much cleaner (on an LCD at least).

[ryrynz]

16-16-16 Video Black on the right.


How come 16-16-16 is not completely black?

So far I'm counting 6 votes for test build 3, and none yet for test build 5.

I vote build 3 too, just wish there was some minor adjustment to avoid the obvious pattern without changing how well the dithering blends in.

[cyberbeing]

With everyone voting for build 3, it sound like I'm going to need to archive build 5 somewhere safe in case madshi reverts everything. I would propose keeping some kind of reduced artifact w/ random element error diffusion mode as a trade quality for performance option or similar. Overall, I see great value in having an error diffusion mode resistance to patterns and artifacts.

With anime in particular, there is much higher chance you'll notice dithering patterns with all its flat surfaces, gradients, and tendency for banding (i.e. source errors). All my tests on anime so far seem to show build 5 actually has a less noisy image because of the increased tendency to produce flat colors where build 3 produces little checkerboard patterns and has crosstalk in random locations which are much more visible than minor low-level noise.


How come 16-16-16 is not completely black?

It is, except when using an ArgyllCMS generated 3DLUT.

8bit 16-16-16 = 16bit 2056-2056-2056

For example, if the 3DLUT maps 16-16-16 (8bit) -> 2057-2057-2057 (16bit) you'll end up with 1-1-1 (16bit) after TV->PC expansion. When conversion back to 8bit occurs, you end up with a fractional 8bit value of ~0.008 which cannot exist. The only way to simulate an 8bit fractional value is by dithering. In many ways, it would be preferred if madVR did not perform dithering on 16bit values which end up as pure black when shifted to 8bit without dithering.

[turbojet]

Does anyone else with a 24p display get dropped frames when playing this file? http://www.mediafire.com/download/oxasi9427dogajm/24hz_issue.mkv

It works fine at 48 and 72hz though. I wonder if it's a madvr issue or something else.