madVR - high quality video renderer (GPU assisted)

[pankov]

Hmmmmm... Can you try whether pressing F2 works? You know, F2 is available if you toggle some of the settings. F2 is handled identically in v0.84.0 vs. v0.83.7, so I'm wondering if that still works. Also, can you try changing the settings to use some very simple keyboard shortcuts? E.g. just "Ctrl+Something"? Does that also not work?
Yes, I've just tried it and F2 does work.
When I initially reported the problem I was using exactly such "simple" shortcuts - Ctrl+J and Ctrl+R to show and reset the OSD stats.
I was just about planning to add a new one (toggle the different deinterlacing states) when I discovered that none of the shortcuts worked with Girder.
btw
can you tell me more about how you hook the keyboard events?
which process is actually doing it? madHcCtrl.exe or the madVR.ax itself and which "window" inside of them? I'm asking this because Girder has the ability to "target" a specific window (found by window Name and/or ClassName and/or executable name) not only the foreground one and it's possible that there might be a solution that doesn't require you to change anything.

[madshi]

Update: GPU usage is anywhere between 20%-60% on a 720p video with Jinc3 AR. I guess it depends on the video's complexity.
GPU usage should not change much with different video files or scenes. It might change if it has to fill up the queue. So I guess those 60% is when the queue is empty.

Yes, I've just tried it and F2 does work.
When I initially reported the problem I was using exactly such "simple" shortcuts - Ctrl+J and Ctrl+R to show and reset the OSD stats.
Can you please create a log, then press Ctrl+J, Ctrl+R with Girder?

can you tell me more about how you hook the keyboard events?
which process is actually doing it? madHcCtrl.exe or the madVR.ax itself and which "window" inside of them? I'm asking this because Girder has the ability to "target" a specific window (found by window Name and/or ClassName and/or executable name) not only the foreground one and it's possible that there might be a solution that doesn't require you to change anything.
madVR.ax inside of the media player process is doing the hooking, by using SetWindowsHookEx(WH_KEYBOARD_LL).

[6233638]

When the framerate of the video get detected, reclock reinitialize or something similar and the audio cut for a fraction of a second.

I deactivated the portion of reclock that change the fps since it bothered me too much and only kept it for the WASAPI audio.

As far as I know, it's not possible to avoid this little interruption but it would be great if there was a way.I can't say that I have experienced that, but ReClock is detecting the framerate on videos almost instantaneously for me. By the time I can open up the dialogue box, the framerate has been detected.

I have madVR set up with both the "delay playback start" options enabled though. I'm not sure if that will affect it, but I suggest trying that.

ReClock is set up like this for me:
http://www.abload.de/thumb/reclocktpjxc.png (http://www.abload.de/img/reclocktpjxc.png)

I'm going straight from my PC to the display using HDMI though, and am not trying to bitstream. (bitstreaming is detrimental to audio playback on a PC)

It seems like this discussion belongs in the ReClock forums (http://forum.slysoft.com/forumdisplay.php?&f=85) rather than madVR though.

[pankov]

Can you please create a log, then press Ctrl+J, Ctrl+R with Girder?

madVR.ax inside of the media player process is doing the hooking, by using SetWindowsHookEx(WH_KEYBOARD_LL).
Here is the log
http://www.mediafire.com/file/2j7403jel8zdxst/madVR_-_log_(Girder_-_keyboard).rar
I've opened it myself and found the problem
madVR gets Ctrl+Shift+J instead of Ctrl+J. This reminded me that I've configured a capital J in Girder. When I changed it to the small letter it started working. I guess this is not something you can fix on your end but you never stop to surprise me so I wait for your conclusion.

[SamKook]

I can't say that I have experienced that, but ReClock is detecting the framerate on videos almost instantaneously for me. By the time I can open up the dialogue box, the framerate has been detected.

I have madVR set up with both the "delay playback start" options enabled though. I'm not sure if that will affect it, but I suggest trying that.

ReClock is set up like this for me:
http://www.abload.de/thumb/reclocktpjxc.png (http://www.abload.de/img/reclocktpjxc.png)

I'm going straight from my PC to the display using HDMI though, and am not trying to bitstream. (bitstreaming is detrimental to audio playback on a PC)

It seems like this discussion belongs in the ReClock forums (http://forum.slysoft.com/forumdisplay.php?&f=85) rather than madVR though.

He did say it was off-topic(and this will be my last reply on the subject here). I tried your settings and I still get a delay of about 4 seconds before detection(unless it detects previously set framerate even though I cleared the DB, then it's instantaneous). If you display the icon in the tray, you can see when it gets detected since it will change color.

I also have both delay options enabled in madvr(it doesn't change a thing), my pc hooked to my tv via hdmi(dvi to hdmi to be exact) and not trying to bitstream the audio.

Since my TV only support 60Hz and most of the stuff I watch is 23.976, I never bothered asking in the reclock forum since it wouldn't change much for me anyway.

[leeperry]

framerate multiplication [..] will it be possible with custom shaders?
I believe Seb.26(an HLSL PS script coder from the HCFR forum) made very clear that frame interpolation wasn't possible in MPC because it wouldn't allow access to N-1 frames (http://translate.google.com/translate?hl=fr&sl=fr&tl=en&u=http%3A%2F%2Fwww.homecinema-fr.com%2Fforum%2Fviewtopic.php%3Fp%3D171661553%23p171661553), and Casimir666 never found the time to implement it either.

Hopefully, madshi will polish the PS script capabilities with the usual excellence he has now used us to, allowing before/after scaling processing, automatic rules based on resolution/frame rate and N-1 frames access(which would allow all kinds of temporal manipulations) :cool:

BTW, chambolle made a script that cuts the picture in 6 zones so you can finetune their respective convergence (http://translate.google.com/translate?hl=fr&sl=fr&tl=en&u=http%3A%2F%2Fwww.homecinema-fr.com%2Fforum%2Fviewtopic.php%3Fp%3D171658787%23p171658787) in order to "fix" a misconverged projector(anything with 3 discrete panels is misconverged to some extent) for instance(HR also provides this feature). Really cool stuff when you have a semi-busted display! Seb.26 improved it by adding even more zones but I can't find it atm.

Needless to say that I'm eagerly looking forward PS scripts support in mVR(any rough ETA please? http://www.forum-audiophile.fr/images/smilies/44.gif ), so I can do levels conversion/mirroring/hardcore deinterlacing/super resolution upscale and what-not in full blown 32fp \o/

yesgrey3 also repeatedly told me that post-processing should preferably be applied on a color corrected picture and not prior to gamut mapping, so PS scripts support would pretty much kill all birds with one stone :)

Pixel shaders are only a small part of madVR, unfortunately. I wish I could write pixel shaders all day
PS scripts coders definitely seem to really enjoy writing them: they get all the good sides of coding without being nagged by a million of other annoying things that essentially end up killing the fun :mad:

There's no real benefit over LAV at the moment.
OK, thank you for the detailed explanation. I thought it might have had to do with jitter and so, but CUVID LAV + mVR truly make up for a stunning duo! And processing jitter doesn't seem to be as much of an issue with video as it can be with audio.

PS: bah, now I crave a new GPU so I can use jinc3 AR for 720p to 1080p luma upscale...my 96SP 8800GS only allows jinc3 AR for chroma and lanczos/spline AR for luma. This said, lanczos3 AR looks quite good and maybe you'll manage to optimize jinc3 AR for luma at some point :p

[jmone]

If you compare the data, the 660 is clearly on another level:

550Ti: 28.8 GT/s, 98.5 GB/s, 691 GFLOPS
660: 78.5 GT/s, 144.2 GB/s, 1881 GFLOPS

Great - got a Asus ENGTX660-DC2-2GD5 GF GTX660 DirectCU II PCI-E 3.0, 2GB on order (hope it gets here for the weekend) and will repost the 60i results.

[cyberbeing]

@SamKook

Reclock's DirectShow Frame-rate detection is extremely fragile (it uses a special source filter to automatically build a temporary Directshow graph at the beginning of playback). Certain combinations of splitters, decoders, and transform filters will cause it to stop working. When Reclock's DirectShow Frame-rate detection breaks, it falls back to Built-in Estimator which takes a few seconds of playback to determine the frame-rate. It would be nice if James @ Slysoft would fix this, but it seems he's finished doing improvements to Reclock.

If you set a default frame-rate like '23.976' instead of 'unknown' in Reclock, you should be able to prevent an initialization pop/gap when the estimator and your default frame-rate match. Depending on your audio hardware, if you set Reclock's 'Sound pre-buffer' low enough, you may also be able to reduce or eliminate the initialization pop/gap effect.

[MasterMeNL]

{snip}Post about ReClock{/snip}
{snip}Post about ReClock{/snip}
I took over your settings except for the buffer and a few other options and assuming a 23.976 frame rate did it.
It assumes 23.976 and after a few seconds it has measured the frame rate which is 99% of the time 23.976 fps so nothing changes. Thanks a lot for your info.

I'm really surprised by this forum, it's very active and a lot of helpful people in here. I've been following this for a long time but I only started posting yesterday. Great stuff!

[madshi]

madVR gets Ctrl+Shift+J instead of Ctrl+J. This reminded me that I've configured a capital J in Girder. When I changed it to the small letter it started working. I guess this is not something you can fix on your end but you never stop to surprise me so I wait for your conclusion.
I could "fix" it, but I don't think I should. After all there's a difference between Ctrl+J and Ctrl+Shift+J, and we want to have the option to use these two key combinations for different things, don't we?

Hopefully, madshi will polish the PS script capabilities with the usual excellence he has now used us to, allowing before/after scaling processing, automatic rules based on resolution/frame rate and N-1 frames access(which would allow all kinds of temporal manipulations) :cool:
In the first step I'll probably just duplicate what MPC-HC offers. A more complete solution might come at a later date. Similar to how madVR's screenshot functionality is currently very simplistic and will get more features sometime in the future.

yesgrey3 also repeatedly told me that post-processing should preferably be applied on a color corrected picture and not prior to gamut mapping
I don't think I agree with that.

This said, lanczos3 AR looks quite good and maybe you'll manage to optimize jinc3 AR for luma at some point :p
I wouldn't expect speed improvements for Jinc3 AR. It might even get a little slower, after some more tweaks to the AR algorithm.

[sunnah]

PS: bah, now I crave a new GPU so I can use jinc3 AR for 720p to 1080p luma upscale...my 96SP 8800GS only allows jinc3 AR for chroma and lanczos/spline AR for luma. This said, lanczos3 AR looks quite good and maybe you'll manage to optimize jinc3 AR for luma at some point :p

Try to set in chroma softcubic 80 and for luma upscaling jinc3 or jinc4. The supposed difference what the human eye could detect from chroma upsampling is the color if you zoom-in on the image on a red color and some other colors, you will see it becomes a little brighter with jinc. But I think, this difference is not too important if you can't use jinc for scaling, because of that.

So try that mate :)

Cheers.!

[SamKook]

If you set a default frame-rate like '23.976' instead of 'unknown' in Reclock, you should be able to prevent an initialization pop/gap when the estimator and your default frame-rate match.

Looks like I got bad result because of some manual framerate that were set a while back when I did my test, because that did get rid of the problem.

But now I remember why I set it to unknown in the first place instead of 23.976. It's because it can't detect properly the frame rate of some of my own videos with 59.94 fps so I wanted to avoid it using 23.976 in those cases.

Thanks for the information.

[leeperry]

In the first step I'll probably just duplicate what MPC-HC offers. A more complete solution might come at a later date.
Alright, sounds good! I see that Seb.26 released condition scripts that allow some commands to only be processed depending on the movie dimensions, so if we can process the PS scripts prior to scaling that should work fine and we could process different commands for SD and HD :)

I'm not sure how we could roll scripts, though? through hotkeys again I presume.

I don't think I agree with that.
Oh, well it made sense to me to, for instance, sharpen onto the right colors instead of rolling colors afterwards.

I wouldn't expect speed improvements for Jinc3 AR. It might even get a little slower, after some more tweaks to the AR algorithm.
OK, I need to read up more on Jinc coz I dunno whether it was placebo but it seemed to be a far superior algorithm than the others from my limited testing. Uber sharpness and zero ringing :eek: Hopefully I could get away with a $100-150 GPU.

Try to set in chroma softcubic 80 and for luma upscaling jinc3 or jinc4. The supposed difference what the human eye could detect from chroma upsampling is the color if you zoom-in on the image on a red color and some other colors, you will see it becomes a little brighter with jinc. But I think, this difference is not too important if you can't use jinc for scaling, because of that.
Actually I ran a bunch of tests yesterday and IIRC I can use jinc3 for both luma and chroma when upscaling 720p to 1080p, I can even enable AR for chroma but AR for luma is a big no-no....98% GPU load on my 8800GS and massive frames drop. Jinc3 w/o AR is an halo feast so that's a no-go.

This said, I just asked for a refund on that 46" CCFL TV coz the last ten vertical percents on the right end had some slight colorimetry drift that was almost impossible to see in 50fps digital TV but utterly annoying after a while in slow 24p pans...dunno if it came from a backlight homogeneity issue, some problem with the polarizing filters or whatever else. Strangely, everything looked fine in Nokia Monitor Test, the problem was pretty much only visible in 24p pans :o

And as 6233638 said, CCFL's make nasty bright green RBE's that are absolutely unbearable to me on anime, duh again :scared:

So I'm back on the market for a new TV....and the only company that doesn't con its customers with the panel lottery is Sharp so I guess a LC-40LE630E with the killer UV²A panel & 24Hz support is in good order :devil:

[madshi]

OK, I need to read up more on Jinc coz I dunno whether it was placebo but it seemed to be a far superior algorithm than the others from my limited testing. Uber sharpness and zero ringing :eek: Hopefully I could get away with a $100-150 GPU.
The main positive thing about Jinc is that it has a very pleasent almost analog like look with no aliasing, and a decent sharpness (though a bit less sharp than Lanczos). In comparison all other resamplers look more digital and have more aliasing. With AR turned on, there should be no difference in ringing. If you're a sharpness junky, you may want to stay with Lanczos, though, or wait until you can post-sharpen the Jinc output. A great GPU would be the 660, if you can swing it.

So I'm back on the market for a new TV....and the only company that doesn't con its customers with the panel lottery is Sharp so I guess a LC-40LE630E with the killer UV²A panel is in good order :devil:
If you plan to use it for movie playback, only, and not for any computer stuff, then I believe plasma (e.g. try Panasonic) gives the best image quality of the currently available TVs, at least if you can half way control ambient light. LCD is better when there's lots of ambient light. But we're getting a bit OT here.

[leeperry]

The main positive thing about Jinc is that it has a very pleasent almost analog like look with no aliasing, and a decent sharpness (though a bit less sharp than Lanczos). In comparison all other resamplers look more digital and have more aliasing. With AR turned on, there should be no difference in ringing. If you're a sharpness junky, you may want to stay with Lanczos, though, or wait until you can post-sharpen the Jinc output. A great GPU would be the 660, if you can swing it.
Analog, exactly! It doesn't have the digital look of Lanczos :)

Post-sharpening the Jinc output? I've grown tired of sharpening tbh.

If you plan to use it for movie playback, only, and not for any computer stuff, then I believe plasma (e.g. try Panasonic) gives the best image quality of the currently available TVs, at least if you can half way control ambient light. LCD is better when there's lots of ambient light. But we're getting a bit OT here.
Yeah, well a friend of mine who doesn't see any RBE's on his 32" CCFL Sony LCD(5K:1 CR in HCFR) tells me that he sees them very clearly on Plasma's.....he currently owns the same 4X DLP projector I used to have so I trust his eyes.

Well, if the Sharp fails as well I will try my luck with a Sammy in a brick & mortar shop that also provides a cooling-off period. Argyll's CR measurement will make the call :D

I'm terribly sorry for hijacking the thread, but you'll have to admit that it's still more or less mVR related because we all want the best bang for bucks gear to enjoy mVR....and it's quite a jungle out there if you want to buy a proper flat screen. But I think I've got my mind made up now, and it might help other mVR users as well :)

[madshi]

Post-sharpening the Jinc output? I've grown tired of sharpening tbh.
I don't like conventional sharpening, either. But have a look at FineSharp, that looks quite promising to me.

Yeah, well a friend of mine who doesn't see any RBE's on his 32" CCFL Sony LCD tells me that he sees them very clearly on Plasma's.....he currently owns the same 4X DLP projector I used to have so I trust his eyes.
Did he have a plasma at his home? Plasmas at shops are usually put into torch mode, so they can half way compete with the super bright LCDs. Torch mode doesn't really do any favors to plasmas. You know, stores are usually brightly lit (much brighter than your living room is), and there plasmas don't look so good compared to LCDs. It's the same as with audio volume: Louder sounds "better". In the same way brighter looks "better". But in your living room, with controlled lighting, everything is very different. I've had a plasma myself and I couldn't see any hint of RBE. Plasmas do not draw images sequentially like DLPs do. 1-chip DLPs draw color by color. Plasmas have red/green/blue subpixels and all 3 are active at the same time. So there's no reason plasmas should have any RBE. What they do have is dithering noise, because plasma pixels can only be on or off, but they can't be half on. So they simulate half on by pulsating very quickly. Why don't you go to a shop and check yourself whether you can see any hint of RBE with plasmas? They really do have the best image quality, IMHO, if you can control ambient light.

[NicolasRobidoux]

... I wouldn't expect speed improvements for Jinc3 AR. It might even get a little slower, after some more tweaks to the AR algorithm.Dr. Minglun Gong of Memorial University will confirm that I can work miracles refactoring HLSL code (which I assume is what you're using) when paired with an expert.

[madshi]

Dr. Minglun Gong of Memorial University will confirm that I can work miracles refactoring HLSL code (which I assume is what you're using) when paired with an expert.
Maybe I can hire you when I earned my first million with madVR. That might be hard to achieve with freeware, though... :D I fear the main performance cost is reading all the source pixels, and you can't really optimize that away with HLSL code. There might be a higher chance of optimization by using CUDA/OpenCL. For now my main priority is on adding all the missing features, though, not on squeezing out 10% more performance.

[NicolasRobidoux]

Put a flattr button on your download site. This may get you to your first thousand faster than you think.
(And of course it's features first.)
P.S.... or somewhere obvious. (I just realized the download site is this thread, basically.)
P.S.... and it makes complete sense that moving things through memory is the bottleneck, not operations. But the fact that Jinc is slower than, say, Lanczos4, suggests that there is headroom.

[NicolasRobidoux]

Not totally sure this is useful, but here are some news, courtesy of the GIMP dev. mailing list:
Vladimir Yuzhikov has release his SmartDeblur program under the GPL at
https://github.com/Y-Vladimir/SmartDeblur

Yuzhikov worked on two common cases: blurring due to incorrect focus and blurring due to movement. "Each of you knows very well," he points out, they "are very difficult to repair." In fact, the examples used as samples by Yuzhikov are impossible to fix using current commercial technology. You can see these amazing results here:
http://gizmodo.com/5953601/incredible-csi-enhance-technology-fixes-unfocused-photos

[madshi]

Put a flattr button on your download site. This may get you to your first thousand faster than you think.
(And of course it's features first.)
I'm planning to do that (or maybe releasing a "pro" version with additional features) when madVR is feature complete, but not before.

P.S.... and it makes complete sense that moving things through memory is the bottleneck, not operations. But the fact that Jinc is slower than, say, Lanczos4, suggests that there is headroom.
Roughly calculated, when upscaling, Lanczos4 reads 16 source pixels for every destination pixel, while Jinc3 reads 36 source pixels to do the same. So it's pretty clear that Jinc3 must be slower than Lanczos4.

Ok, reading a square of 36 pixels is not optimal for Jinc3. It should be possible to shave off some of the corner pixels because they'll definitely lie outside of the Jinc circle. But anyway, we'll never get down to those 16 source pixels Lanczos4 is happy with. The simple reason why Jinc3 is slower than Lanczos4 is that we can't 2-pass Jinc, as we can do with Lanczos.

One problem with HLSL is that a true branching "if" instruction is usually slower than doing extra work. That can make optimizations quite difficult.

Not totally sure this is useful, but here are some news, courtesy of the GIMP dev. mailing list:
Vladimir Yuzhikov has release his SmartDeblur program under the GPL at
https://github.com/Y-Vladimir/SmartDeblur

Yuzhikov worked on two common cases: blurring due to incorrect focus and blurring due to movement. "Each of you knows very well," he points out, they "are very difficult to repair." In fact, the examples used as samples by Yuzhikov are impossible to fix using current commercial technology. You can see these amazing results here:
http://gizmodo.com/5953601/incredible-csi-enhance-technology-fixes-unfocused-photos
This looks very cool, but he definitely needs an AR filter... :D And it's probably too slow for realtime video playback. Another problem is that you probably need to play with the sliders to get the right parameters for every image, while with video you just want to playback, you don't want to play with sliders all the time. Anyway, looks quite interesting, and maybe in the long run it can be used to create better sharpening algorithms.

[leeperry]

Did he have a plasma at his home? Plasmas at shops are usually put into torch mode
I think he did actually, and there are the usual burn-in issue, highly glossy panel and brightness loss IIRC?

Also, some plasma's output ridiculous CR, such as this 40" LG with 800:1 CR (http://translate.google.fr/translate?sl=fr&tl=en&js=n&prev=_t&hl=fr&ie=UTF-8&layout=2&eotf=1&u=http%3A%2F%2Fwww.lesnumeriques.com%2Ftv-televiseur%2Flg-42lm660s-p13285%2Ftest.html) yay, even worse than a TN.

But anyway, I forgot to try mirroring the picture this morning(and it was the last day I could ask for a refund), but apparently many 24p movies have the last 10 vertical percents on the right hand marked as hell, such as this sample: test.mkv (http://www.mediafire.com/file/k8ki1y6ryejnpf4/test.mkv)

I can mirror it, roll VR's or decoders, move the window in the middle of my CRT and it's still terribly marked, duh! And many/most movies, whatever from the US or EU seem to have this problem IME.

I never noticed how noisy some telecine's were on my previous DLP pj, I guess LCD is indeed merciless with defects :devil:

I still don't regret asking for a refund coz the RBE's and backlight buzzing noise were darn annoying anyway. This said, apparently edge LED LCD's also suffer from buzzing noise when their backlight is dimmed but at least their colorimetry should be more stable :)

[madshi]

Every current Panasonic plasma has a *much* better native contrast than any LCD. Of course LCD uses some tricks to level the playing field like e.g. local dimming, but these tricks can introduce artifacts on their own. LCD backlights lose light, just as much as plasmas do. Current Panasonic plasmas are rated to reach half brightness after 100.000 hours of use. If you leave on your plasma 8 hours every day, it will reach half brightness after 34 years of use. That's much better than CRTs used to be. And LCD backlights won't hold any longer than this, either. There are valid and good reasons for choosing LCD over plasma, but lifetime is not one of them. Just reading reviews in the latest "audiovision" print magazine right now:

LG LCD "65 LM 620 S":
- on/off contrast: 1522:1
- brightness at 10-60° viewing angle: 50%
- contrast at 10-60° viewing angle: 36% (= 548:1 !!)
- 24p quality: 2/3
- motion sharpness: 2/4
- cost: 2800 EUR

Panasonic Plasma TX-P 65 VT 50E:
- on/off contrast: 10963:1
- brightness at 10-60° viewing angle: 80%
- contrast at 10-60° viewing angle: 95% (= 10415:1)
- 24p quality: 3/3
- motion sharpness: 4/4
- cost: 4000 EUR

Ok, price wise this is not a fair comparison, but these are the most comparable reviews in the current issue of "audiovision". You should get an LCD instead of a plasma if (and only if, IMHO):

- you want to use the display as a computer monitor, too (word, programming, browsing)
- you have lots of ambient light in your room and no way (or will) to control it
- power consumption is extremely important to you

If none of these apply, and if you truly value image quality, then you should seriously consider giving Plasma another chance. 'nough said. This is my last comment on LCD vs Plasma in this thread.

[SamuriHL]

I've already decided on a Panasonic plasma for my b-day present in February. :) Absolutely wonderful consumer products. I can't wait to replace my dying SXRD.

[NicolasRobidoux]

...
One problem with HLSL is that a true branching "if" instruction is usually slower than doing extra work. That can make optimizations quite difficult...I know all about this, and it makes optimization more interesting :)

How about this: When I run out of wealthy consulting clients, I'll let you know that I'm ready to take improbably repaid IOUs based on an internet handshake.

Note: With Jinc deblurred with the value that I most recommend with AR (0.981...), you need a 7x7 square to make sure to hit all the locations with nonzero weights. Within this 7x7 square, at most 38 pixels actually have nonzero weights, and it is possible to only use those by exploiting symmetry and arithmetic branching. If you deblur just a little more, so that the disc has radius sqrt(10), which is not a bad thing given that Jinc3 is a little soft anyway, at most 36 pixels actually have nonzero weight.

But of course, the ones that stick out of the 6x6 have minuscule weights, so there is no harm in getting rid of them, esp. since you are actually already approximating via a LUT. If you actually get rid of all those which are at a distance of 3 or more, and use symmetry and arithmetic branching, you are down to 33.

OK, this is not a big reduction.

However, what you can do, is write efficient code that increases the density by 4 in both directions, and either finish off by the built in bilinear interpolation, or with cheap Lagrange-type interpolation. Because increasing the density by 4 in both directions involves only three types of relative positions w.r.t. the original ones when you exploit symmetry, you can program this cheaply (you know the coefficients ahead of time, as well as the relevant pixels, for each type), and then you can use arithmetic branching to "select" the values you need to assemble.

Just sayin'...

[leeperry]

I'm planning to do that
Good, coz it's about time we show you our gratitude :)

OK, got it! I will check plasma's but I'm not shelling out 4 grand :D

Anyway, I'm still shocked to see that my problem came from marked telecine's.......darn it, even the new Spiderman movie had this very same problem in the exact same location and several french movies too....the bigger your LCD screen, the more merciless it gets, I can now understand why many ppl hate 24p played in 60Hz. This marked telecine issue is already extremely distracting in a 24Hz multiple..

And for some reason, mVR doesn't want to work in FSW anymore on my CRT....will try to find out why :o

[madshi]

I've already decided on a Panasonic plasma for my b-day present in February. :) Absolutely wonderful consumer products. I can't wait to replace my dying SXRD.
Yeah, if I weren't a projector guy, I'd get one of those 65" Panasonic plasmas, myself, for serious movie watching. However, maybe for the living room, for sports and news watching, an LCD might be better suited (due to ambient light). Argh, I wanted to *stop* talking about plasma vs. lcd. <sigh>

How about this: When I run out of wealthy consulting clients, I'll let you know that I'm ready to take improbably repaid IOUs based on an internet handshake.
Haha, thanks, that sounds good to me... :D I have some funny ideas about some other video processing algorithms, too, where I lack the necessary math skills...

Note: With Jinc deblurred with the value that I most recommend with AR (0.981...), you need a 7x7 square to make sure to hit all the locations with nonzero weights. Within this 7x7 square, at most 38 pixels actually have nonzero weights, and it is possible to only use those by exploiting symmetry and arithmetic branching. If you deblur just a little more, so that the disc has radius sqrt(10), which is not a bad thing given that Jinc3 is a little soft anyway, at most 36 pixels actually have nonzero weight.

But of course, the ones that stick out of the 6x6 have minuscule weights, so there is no harm in getting rid of them, esp. since you are actually already approximating via a LUT. If you actually get rid of all those which are at a distance of 3 or more, and use symmetry and arithmetic branching, you are down to 33.
I think I'd prefer to keep using 0.981... to keep anti-aliasing power at max. I'm already limiting myself to 36x36 right now. It seems that ignoring those border pixels doesn't harm image quality much.

However, what you can do, is write efficient code that increases the density by 4 in both directions, and either finish off by the built in bilinear interpolation, or with cheap Lagrange-type interpolation. Because increasing the density by 4 in both directions involves only three types of relative positions w.r.t. the original ones when you exploit symmetry, you can program this cheaply (you know the coefficients ahead of time, as well as the relevant pixels, for each type), and then you can use arithmetic branching to "select" the values you need to assemble.
My math skills are not good enough for this... :o

[NicolasRobidoux]

...
My math skills are not good enough for this... :oMine are :D

[madshi]

Mine are :D
Ok, but can you also speak Elvish? http://s4.postimage.org/30093uxno/smiley_tongue.gif

[NicolasRobidoux]

...
I think I'd prefer to keep using 0.981... to keep anti-aliasing power at max. I'm already limiting myself to 36x36 right now. It seems that ignoring those border pixels doesn't harm image quality much.Yes, but because you restrict the filter to a 6x6, this value may have drifted a little.
The "radius sqrt(10)" deblur is not that different anyway: 0.97651932791303247285

[NicolasRobidoux]

Non, mais je parle le joual. http://en.wikipedia.org/wiki/Joual

[NicolasRobidoux]

When I have a bit of time, I'll recompute the "optimal" deblur taking into account that you restrict yourself to a 6x6. My guess is that it will go down just a little.

[NicolasRobidoux]

Ah! With the optimization criterion that gives 0.981..., it does not make a difference if you restrict yourself to a 6x6. (It makes a difference with the other criterion.)
So: Just keep 0.981..., the deblur I would guess works best for Jinc3 with AR.

[madshi]

Ah! With the optimization criterion that gives 0.981..., it does not make a difference if you restrict yourself to a 6x6. (It makes a difference with the other criterion.)
So: Just keep 0.981..., the deblur I would guess works best for Jinc3 with AR.
Ok, so all is good - thanks!

[DragonQ]

madshi, can you figure out why this file won't deinterlace with any field order setting in MadVR? I have a feeling MKV Merge is to blame but I'm not sure. EVR struggles with it too - stepping frame-by-frame shows it deinterlaces properly but plays at 46-48 fps instead of 50 and has a bit of a dodgy graph, resulting in jerkiness.

Sample Clip (http://www.mediafire.com/?ho71875bvmy8m5d)

The original TS file plays fine using both EVR and MadVR.

[madshi]

I'm sorry, but if the original file plays fine then it can't be madVR's fault, can it? I really don't have time to look into how mkvmerge might have screwed things up. It's already more than enough work to only work on my own bugs...

[Xaurus]

Every current Panasonic plasma has a *much* better native contrast than any LCD. Of course LCD uses some tricks to level the playing field like e.g. local dimming, but these tricks can introduce artifacts on their own. LCD backlights lose light, just as much as plasmas do. Current Panasonic plasmas are rated to reach half brightness after 100.000 hours of use. If you leave on your plasma 8 hours every day, it will reach half brightness after 34 years of use. That's much better than CRTs used to be. And LCD backlights won't hold any longer than this, either. There are valid and good reasons for choosing LCD over plasma, but lifetime is not one of them.
As a previous owner of the following plasma brands:

LG
Pioneer (kuro)
Panasonic

I bought the 60" LG plasma in the fall of 2010. By the spring of this year (1,5 years) I was unable to get it to display the blacks that it initially had during the first period. It became so bad (black was literally grey) that I simply decided to sell it and got Sony's top-model LCD in 55", and now after half a year I still think it's the best TV I ever owned. Ok, 5" down in screen size but the 65" model was too expensive, he he.

Bottom line: For plasmas, take the manufacturer claimed brightness reduction time with a grain of salt. I know the previous gen (2011) Panasonics also had this problem.

[DragonQ]

Alright, I'll ask Mosu.

[6233638]

I've had a plasma myself and I couldn't see any hint of RBE. Plasmas do not draw images sequentially like DLPs do. 1-chip DLPs draw color by color. Plasmas have red/green/blue subpixels and all 3 are active at the same time. So there's no reason plasmas should have any RBE. What they do have is dithering noise, because plasma pixels can only be on or off, but they can't be half on. So they simulate half on by pulsating very quickly. Why don't you go to a shop and check yourself whether you can see any hint of RBE with plasmas? They really do have the best image quality, IMHO, if you can control ambient light.Actually, Plasmas do draw their image in quite a similar way to DLP. The pixels can only be turned on or off (similar to DLP) and gradation is built up over several subframes by pulsing the pixels on/off for varying lengths of time to control brightness. Because each phosphor has different response/decay characteristics, this does essentially draw the image in sequential colour, even though all the pixels are technically being addressed at the same time.

This video illustrates how Plasma draws its image: http://www.youtube.com/watch?v=tGpdah32n3c

Every current Panasonic plasma has a *much* better native contrast than any LCD. Of course LCD uses some tricks to level the playing field like e.g. local dimming, but these tricks can introduce artifacts on their own. LCD backlights lose light, just as much as plasmas do. Current Panasonic plasmas are rated to reach half brightness after 100.000 hours of use. If you leave on your plasma 8 hours every day, it will reach half brightness after 34 years of use. That's much better than CRTs used to be. And LCD backlights won't hold any longer than this, either. There are valid and good reasons for choosing LCD over plasma, but lifetime is not one of them.

...

You should get an LCD instead of a plasma if (and only if, IMHO):

- you want to use the display as a computer monitor, too (word, programming, browsing)
- you have lots of ambient light in your room and no way (or will) to control it
- power consumption is extremely important to you

If none of these apply, and if you truly value image quality, then you should seriously consider giving Plasma another chance. 'nough said. This is my last comment on LCD vs Plasma in this thread.I would argue that if you value image quality, you would give Plasmas a wide berth. They measure well on native contrast, viewing angle and color accuracy (nowadays) but the way they draw their image due to the heavy use of dithering, motion artefacts, the ABL circuit, flicker, crosstalk, poor gradation capabilities (particularly above 60Hz) all result in decidedly poor image quality in my experience.


For what it's worth, some of the best LCD panels out there are approaching 5,000:1 native contrast, and it is only this year that Panasonic has managed to exceed that.

Edge LED and "local dimming" (ha!) Edge LED sets really aren't any good in my opinion. However, the best local-dimming sets that use a full array backlight with hundreds of zones combined with a high contrast panel (for example LG try it with 1,000:1 IPS panels and it looks terrible) will exceed Plasma contrast easily with very little haloing. Examples would be the Sharp Elite and Sony's UV2A HX900. (I hear the newer Sonys with Samsung panels fare worse)

Even the Kuros with their 30,000:1 CR (when calibrated) weren't nearly enough for me, because they still never look black in a dark room. Local dimming achieves that with ease.

Image retention is still a very real issue with Plasma despite what people claim online. (just try only watching 2.37:1 films for a while, or getting hooked on a game for a week without watching anything else and see how "resistant" they really are...)

Plasma still doesn't handle ambient light very well, though they are getting better with it these days.

The thing that kills Plasma for me though, is the "ABL circuit" which starts to dim the screen when the image gets over a certain brightness. With most plasmas you lose over half their brightness when you go from a small white window pattern and a full white screen. It makes bright daylight scenes in films look dull/washed-out, and is a considerable problem if you play games or want to use the screen with a PC. (not that I recommend either due to image retention) It's not an issue of the image not being able to go bright enough - I only watch at reference levels (100cd/m²) but that to achieve that brightness with "brighter" images, you need to calibrate the panel to 200cd/m² or more, which is far too bright for a dark room, or try and compromise somewhere in-between where the image is too bright in some scenes but not bright enough in others etc. LCD does not have this problem at all, and while it affected CRTs, most were well within 10% of their calibrated brightness, not 50% or more.

Motion handling is moot between the two technologies these days. Now that scanning backlights have been introduced, LCDs can match or exceed Plasma motion handling due to the reduced persistence on the retina. (but it may introduce some degree of flicker if you are sensitive to that)


On the plus side, Plasmas actually have a good image when viewed off to the side, if that is a concern for you (I never sit to the side of my TV) because even the best IPS panels still look terrible when you aren't sitting directly in front of them.

They also update the panel progressively, so fast motion won't skew like it does on LCDs.

I agree with your sentiment on the life of these displays - it isn't something you should be concerned about with either LED sets or Plasma, though Plasmas did have a tendency to die early due to power supply failures. Modern Plasma is considerably more efficient now though, so that should be less of a concern.


Another thing is that Plasmas are significantly cheaper to get a well calibrated, high contrast image than LCDs. Panasonic's Plasmas are very cheap compared to LED backlit LCDs. (not edge LED)

If Plasmas image quality problems or other potential issues are not a concern for you, Panasonic's Plasmas have by far the best price/performance ratio out there today.


Yeah, if I weren't a projector guy, I'd get one of those 65" Panasonic plasmas, myself, for serious movie watching.I miss having a good projection setup. :o

I keep seeing the price of a good DLP projector drop and thinking I should buy one again, but I don't really have the space for one now. :( (there's no way I would buy another SXRD - they're far too soft, and motion handling sucks, JVC's D-ILA is far too expensive, and has similar problems)

[Qaq]

Every current Panasonic plasma has a *much* better native contrast than any LCD.
Yep, but they don't support RGB 4:4:4 yet. So say goodbye to the "high quality chroma upsampling".

[madshi]

Actually, Plasmas do draw their image in quite a similar way to DLP. The pixels can only be turned on or off (similar to DLP) and gradation is built up over several subframes by pulsing the pixels on/off for varying lengths of time to control brightness.
Red/green/blue are driven independently and simultanously which is a key difference to how single-chip DLP works.

Because each phosphor has different response/decay characteristics, this does essentially draw the image in sequential colour, even though all the pixels are technically being addressed at the same time.

This video illustrates how Plasma draws its image: http://www.youtube.com/watch?v=tGpdah32n3c
This video illustrates how one specific LG plasma model draws its image. If you buy a plasma from LG then that's your own fault. Different plasma manufacturers are using *totally* different panels, different phosphors and different driving schemes. It's quite possible to optimize phosphors to get nearer in response/decay characteristics. And it's also quite possible to modify the driving scheme to reduce artifacts. Now I haven't seen Pioneer Kuro and Panasonic 1200Hz captures, so I can't say for sure whether they have color separation problems or not, but I'd like to first see that before I believe it. I know for a fact that Pioneer was using a very different driving scheme than Panasonic a few years back.

Strange enough, people here are clever enough to know that one bad LCD model doesn't mean all LCD models are bad, and that there are different LCD panels with very different properties and problems. But then, 3 different people now mentioned bad LG plasmas, without really acknowledging that maybe not every plasma is the same.

I would argue that if you value image quality, you would give Plasmas a wide berth. They measure well on native contrast, viewing angle and color accuracy (nowadays) but the way they draw their image due to the heavy use of dithering, motion artefacts, the ABL circuit, flicker, crosstalk, poor gradation capabilities (particularly above 60Hz) all result in decidedly poor image quality in my experience.
Well, I guess we will have to agree to disagree here. Whenever I go into a shop and compare (Panasonic) Plasmas and LCDs, I wonder how anybody could buy an LCD for movie playback. They just don't cut it in comparison, from my point of view.

Anyway, we've now heard enough pro and con views about LCD vs. plasma, I think. So maybe we should really give it a rest now?

[Pat357]

I've got a lot of 25fps SD files, some are DVD encodes, some aren't
Those ones that aren't DVD encodes, what are these ?
Exactly those might end up wrong coulored in MadVR when Madshi would implement EBU for *all* 25fps SD content, even with additional check in WxH.

because encoding upscaled pixels is a terrible idea to begin with.
Uh ? Why is that ?
Does this mean we should abandon all (up)scaling from our avisynth scripts if we use these to encode !!? :eek:
Strong statements you make here, I must say... :rolleyes:

All I know is that mastering houses usually have their CRT/CCFL LCD screens.

Mastering houses don't use CCFL backlights anymore since a long time. Almost all switched to LED, because off the much better performance & stability over time.
CCFL has a higher black-point and a much less stable curve at low Y.
From the physical POV, controlling a CCFL at low output is much harder than walking over a cord !
The only reason that CCFL is still in the market is price €/m2 compared to LED and other.

[6233638]

This video illustrates how one specific LG plasma model draws its image. If you buy a plasma from LG then that's your own fault. Different plasma manufacturers are using *totally* different panels, different phosphors and different driving schemes. It's quite possible to optimize phosphors to get nearer in response/decay characteristics. And it's also quite possible to modify the driving scheme to reduce artifacts. Now I haven't seen Pioneer Kuro and Panasonic 1200Hz captures, so I can't say for sure whether they have color separation problems or not, but I'd like to first see that before I believe it. I know for a fact that Pioneer was using a very different driving scheme than Panasonic a few years back.It's true that they use different driving schemes (for example when Panasonic introduced their 3D panels, they switched to displaying the brightest subframe first, rather than the dimmest) but they fundamentally need to build up the image over several subframes, and all suffer from these kind of artefacts. Panasonic have done some interesting things with their "2500Hz" driving scheme (which refers to displaying each frame within 0.4ms, not refreshing at 2500Hz) but they still suffer from it: http://www.youtube.com/watch?v=a-03ENXkcs8

For what it's worth, Pioneer's driving scheme meant that they had very limited gradation, and had to use a lot of dithering. At anything other than 60Hz, gradation was extremely poor.

Blue phosphors are still considerably quicker, and red/green still have mis-matched responses. Here is a chart from Panasonic a couple of years ago when they switched to newly developed phospors to better meet the requirements of 3D:

http://i.imgur.com/LxoiB.jpg

Strange enough, people here are clever enough to know that one bad LCD model doesn't mean all LCD models are bad, and that there are different LCD panels with very different properties and problems. But then, 3 different people now mentioned bad LG plasmas, without really acknowledging that maybe not every plasma is the same.If you can find another slow motion video of a Plasma, it will show similar results, with the image being built up over several subframes, and the mis-matched phosphor response resulting in the image essentially being drawn in sequential colour. It just happened to be an LG model being used in that video on YouTube. (there was another video I found, but it doesn't mention what TV was used)

As someone that has owned, tested and reviewed Plasmas from a number of manufacturers, I can tell you that it affects them all.

Well, I guess we will have to agree to disagree here. Whenever I go into a shop and compare (Panasonic) Plasmas and LCDs, I wonder how anybody could buy an LCD for movie playback. They just don't cut it in comparison, from my point of view.I don't blame you, LCDs look terrible at their default store settings. (just like Plasmas look awfully dim in a store in comparison)

With a well-calibrated full array LED set, I think your opinion would change. I don't like edge LED sets at all though, the uniformity is too bad, and they have far too few zones for dimming to be effective. (16-32 at best I think?)

[egur]

Not totally sure this is useful, but here are some news, courtesy of the GIMP dev. mailing list:
Vladimir Yuzhikov has release his SmartDeblur program under the GPL at
https://github.com/Y-Vladimir/SmartDeblur

Yuzhikov worked on two common cases: blurring due to incorrect focus and blurring due to movement. "Each of you knows very well," he points out, they "are very difficult to repair." In fact, the examples used as samples by Yuzhikov are impossible to fix using current commercial technology. You can see these amazing results here:
http://gizmodo.com/5953601/incredible-csi-enhance-technology-fixes-unfocused-photos

Actually not a new idea. Usually referred to as inverse filtering. I've seen nice results 14 years ago in an image processing course.
BTW, all the samples he showed belong to the "flat world" class - focus distance is identical to all pixels. Since (out of) focus blurring can be measured and approximated, finding the inverse filter is doable. It gets really hard when there're multiple areas in the image with different distances.

The CSI style operation is usually done by applying weak deblurring kernels again and again to smooth parts of the image. The trick is when to stop. It never looks really good in end but some details can surface. So the CSI use case is valid but fixing your home photos will not work right.

I've read what you wrote on the Jinc function in your website. Basically an ellipse (2D) shaped Sinc like function (right?). But I didn't see any head to head compares against Lanczos.

Is the Jinc AR algorithm proprietary?

[leeperry]

those might end up wrongly colored in MadVR when Madshi would implement EBU for *all* 25fps SD content, even with additional check in WxH.
I don't see any better technical choice than to use EBU for 25fps SD files. There's a far higher probability that a 25fps SD file is PAL than anything else IMO..be it a WEB-DL, a DVD encode or home made stuff.

Uh ? Why is that ?
Does this mean we should abandon all (up)scaling from our avisynth scripts if we use these to encode !!?
Well, I was mostly talking about 1024x576 PAL DVD rips. The best way is to remux the untouched 720x576 stream and set a DAR of 16/9 or 4:3.

Avisynth? processed in 8bit? Yes, this will temper with the PQ I'm afraid and encoding upscaled pixels is a waste of bits. We can easily upscale at playback by setting the right DAR.

[NicolasRobidoux]

...
I've read what you wrote on the Jinc function in your website. Basically an ellipse (2D) shaped Sinc like function (right?). But I didn't see any head to head compares against Lanczos.

Is the Jinc AR algorithm proprietary?I have a website? You mean the parent site of http://www.imagemagick.org/Usage/filter/nicolas/?
-----
Basically:
Take the inverse Fourier transform of the characteristic function on a centered interval (a.k.a. the "ideal low pass filter") in 1D -> Sinc
Take the inverse Fourier transform of the characteristic function on a centered disc in 2D -> Jinc.
-----
All my stuff is academic research with no patents or what not. (I am an ex-prof.) This includes the million variants of EWA LanczosSharp = Elliptical Weighted Averaging Jinc-windowed Jinc, the basis of madVR's Jinc method. It builds on ideas of, among other people, Paul Heckbert, Andreas Gustafson (and, I would guess, his thesis advisor Heikki Saikkonen), Craig deForest and Anthony Thyssen.
AR is Mathias Rauen's baby, and I'm not in a position to comment on legalese.

[NicolasRobidoux]

@Eric Gur: I don't have a ready made head to head comparison of plain vanilla tensor Lanczos and EWA LanczosSharp (=madVR's Jinc3 AR minus the AntiRipple, which again, is Mathias' private baby).
But you can easily do one yourself. (Warning: I'm a *n*x snob.)
Install HDRI ImageMagick http://www.imagemagick.org/script/high-dynamic-range.php, for otherwise you get a version of tensor Lanczos that has clipping artifacts.
Once it's done:
convert input.png -filter Lanczos -resize 800% output.png
gives you plain tensor Lanczos 3-lobe, while
convert input.png -colorspace RGB -filter Lanczos -resize 800% -colorspace sRGB output.png
gives you gamma corrected results.
To get EWA LanczosSharp:
convert input.png -filter LanczosSharp -distort Resize 800% output.png
convert input.png -colorspace RGB -filter LanczosSharp -distort Resize 800% -colorspace sRGB output.png
If you have a reasonably standard 16-bit (not HDRI) version of ImageMagick, the above code will give you correct EWA LanczosSharp results, so if you generate the Lanczos results with some other piece of software, you can compare.
----
If you want to discuss the EWA Jinc part of this (but not AR, which again is linked to madVR), I suggest you take it to the ImageMagick Forums http://www.imagemagick.org/discourse-server/index.php.
P.S. You may want to use a non-bleeding edge version of ImageMagick: A png conversion bug appeared sometime in the last few releases (2 weeks ago? IM updates about 6 times/month). It affects, among other things, 3-channel greyscale. :(

[NicolasRobidoux]

The disc becomes an ellipse when you warp. For aspect ratio preserving resize, it's always a perfectly round disc.
Elliptical Weighted Averaging was invented to deal with perspective transformations, and is commonly used for splatting. Using it to resize is a recent development.
To make it competitive required "a few tweaks".

[madshi]

Blue phosphors are still considerably quicker, and red/green still have mis-matched responses. Here is a chart from Panasonic a couple of years ago when they switched to newly developed phospors to better meet the requirements of 3D
They seem to at least have brought red and green nearer to each other. So color separation is reduced. And that was a couple of years ago. They might have improved it further in the meanwhile. What I'm saying is that with 1-chip-DLP, as far as I remember, each color is drawn totally separately, in equal time distances, with black in between. Plasma at least *tries* to draw at the same time. Maybe it doesn't work perfectly, due to different phosphor behaviour, but I don't think you can claim that color separation is anywhere near as bad as with 1-chip-DLP. Even with that LG video, although the colors appear to be shown somewhat sequentually, they are relatively near to each other in time, with longer black phases in between.

As someone that has owned, tested and reviewed Plasmas from a number of manufacturers
That doesn't make you the one and only judge, though. Different reviewers still have different opinions. From those comparison reviews I've read from print magazines I actually trust, usually (Panasonic) Plasma is crowned winner, as long as ambient light is low. Does that mean they are 100% right and nobody in his right mind could possibly prefer LCD? Definitely not. Different people are perceptable to different kind of artifacts, and maybe even "looks".

I've read what you wrote on the Jinc function in your website. Basically an ellipse (2D) shaped Sinc like function (right?). But I didn't see any head to head compares against Lanczos.
If you have a couple of test images you want to see scaled with Lanczos vs. Jinc, I can do that and upload the results. Or you could also do that yourself by using the latest madVR version. The key difference for me is that Jinc has virtually no aliasing, no matter which angle the edges are. However, Lanczos is a bit sharper.

Is the Jinc AR algorithm proprietary?
Jinc itself is available in open source software ImageMagic (they call it "EWA LanczosSharp", IIRC). The madVR AR filter is an add-on and is my private invention. madVR is closed source and I've not fully explained the AR algorithm anywhere in all detail. I'm still constantly tweaking the algorithm, anyway, so it's not final yet. I might be willing to explain how it works in concept, but I don't plan to publish the source/shader code. I don't want to make it too easy for my renderer competitors to simply copy my work with zero effort. FWIW, I've already explained some of the basics in this thread:

http://forum.doom9.org/showthread.php?t=145358

[egur]

Any more details?
Do you scale using 1 pass in 2D? 2 passes in 1D? 2 passes in 2D?
Does it work well in YCbCr?
Is it any better or worse in linear light?