madVR - high quality video renderer (GPU assisted)

[njfoses]

Originally Posted by njfoses

With all the new versions out in the last few weeks i have not really been keeping up. Im in the same boat as you as i cant stand ringing. I have a dlp and prefer the film look. What would you recommend at this point for chroma, image up and image down on the latest madvr?

Would anybody be willing to share their opinion on the quoted text for somebody that want's as little ringing as possible?

[jokerb47]

Originally Posted by njfoses

Would anybody be willing to share their opinion on the quoted text for somebody that want's as little ringing as possible?

As i get from this thread, if you have enough GPU, you may go with Jinc3 AR for everything. If problems, try first change chroma to SoftCubic 80/100 and test, then luma to Lancos3 AR.

[6233638]

Originally Posted by njfoses

Would anybody be willing to share their opinion on the quoted text for somebody that want's as little ringing as possible?

For the lowest amount of ringing, SoftCubic 80 still seems to be the best. (without the anti-ringing filter - that actually makes SoftCubic worse) SoftCubic has the lowest amount of aliasing, tied with Jinc. (Jinc wins on some tests, SoftCubic wins on others)

For good sources (HD or ringing/artefact-free DVD) that's probably going to be too soft though.

The next option would be Mitchell-Netravali with the anti-ringing filter which has almost no ringing, is quite a bit sharper, but can suffer from aliasing. I would still call that a "soft" filter though - it's just nowhere near as soft as SoftCubic is.

If you want sharper than that, Jinc 3 with the anti-ringing filter seems to be the best option. It does introduce some ringing, but only a small amount, and it has excellent aliasing properties. It's also quite a sharp filter, which is good for HD or good SD sources.

Right now, I find myself switching between SoftCubic 80 and Jinc 3 depending on the quality of the source, which is easy to do now thanks to being able to customise shortcuts.


For Chroma upscaling, I am liking Jinc 3 with the anti-ringing filter enabled for sharper Luma scalers (basically anything above Mitchell-Netravali) and Spline 3 with the anti-ringing filter for the softer luma scalers.


For Luma downscaling, which I should point out is not something I use often, Catmull-Rom with both the anti-ringing filter and linear light options enabled seems to be the best choice in my testing so far. Most of the scaling algorithms look really bad in linear light to be honest, but Catmull-Rom combined with the anti-ringing filter seems to avoid most of those problems.

Originally Posted by jokerb47

As i get from this thread, if you have enough GPU, you may go with Jinc3 AR for everything. If problems, try first change chroma to SoftCubic 80/100 and test, then luma to Lancos3 AR.

You should definitely not use SoftCubic on Chroma - at least that is my opinion. Even SoftCubic 50 is far too soft, and reduces the brightness of, or desaturates Chroma.

[jmone]

Originally Posted by jmone

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.

FYI - 1080/60i material dropped to 52% GPU utalisation on this card with Jin3 AR

After almost an hour the temp is in the low 70c and fan speed is low 60% mark. It is quiet as Nevcairel suggested but I'm not one that is annoyed by some fan noise (and I think most of the noise is the Shuttle Case Fan anyway). It was a squeeze getting this card in a Shuttle SFF case!

[vivan]

Originally Posted by 6233638

LG are using RGBW subpixels, and Samsung appear to be using a vertical subpixel arrangement. Neither of these are suited to use as a monitor, which expects horizontally arranged RGB stripes.

This is true only if you're talking about Windows and ClearType without any modifications.

[nx6]

Originally Posted by vivan

This is true only if you're talking about Windows and ClearType without any modifications.

Keeping in mind this is a MadVR thread, and MadVR is Windows only.

[vivan]

Originally Posted by nx6

Keeping in mind this is a MadVR thread, and MadVR is Windows only.

Keeping in mind MadVR doesn't use ClearType, and thus subpixel arrangment doesn't make any difference when watching video.
So, talking about using something as display not for watching video was offtopic in the first place.

[Keiyakusha]

I believe even windows's ClearType is adjustable to any subpixel arrangement.

[6233638]

Originally Posted by vivan

This is true only if you're talking about Windows and ClearType without any modifications.

It applies to any kind of subpixel rendering, not just Microsoft's ClearType.
It can also create aliasing issues in games and other computer graphics.

Originally Posted by vivan

Keeping in mind MadVR doesn't use ClearType, and thus subpixel arrangment doesn't make any difference when watching video.

Actually, it can affect video quality. Sharp's UV2A panels can show aliasing or other artefacts (such as moiré) due to their non-standard subpixel arrangement/addressing, for example.

Anything other than RGB stripes is detrimental to image quality in general. It also results in a lower fill-factor when only some of the subpixels are lit. (as each subpixel has to be smaller to fit in the same space)

Originally Posted by Keiyakusha

I believe even windows's ClearType is adjustable to any subpixel arrangement.

This does not work as well as one might think. Switching ClearType to BGR results in color fringing on a lot of text on displays which use that subpixel arrangement, for example. I think it only has RGB and BGR options though, not anything else, like vertical subpixel arrangement (which wouldn't really help text rendering much anyway) and there certainly aren't RGBW or RGBY options.

ClearType is not the only subpixel rendering used with computers though. Adobe does its own subpixel font rendering, so does Apple, and subpixel rendering is not only used for text.


This is starting to get wildly off-topic now though, and I apologise for contributing to that.

[njfoses]

Originally Posted by 6233638

For the lowest amount of ringing, SoftCubic 80 still seems to be the best. (without the anti-ringing filter - that actually makes SoftCubic worse) SoftCubic has the lowest amount of aliasing, tied with Jinc. (Jinc wins on some tests, SoftCubic wins on others)

For good sources (HD or ringing/artefact-free DVD) that's probably going to be too soft though.

The next option would be Mitchell-Netravali with the anti-ringing filter which has almost no ringing, is quite a bit sharper, but can suffer from aliasing. I would still call that a "soft" filter though - it's just nowhere near as soft as SoftCubic is.

If you want sharper than that, Jinc 3 with the anti-ringing filter seems to be the best option. It does introduce some ringing, but only a small amount, and it has excellent aliasing properties. It's also quite a sharp filter, which is good for HD or good SD sources.

Right now, I find myself switching between SoftCubic 80 and Jinc 3 depending on the quality of the source, which is easy to do now thanks to being able to customise shortcuts.


For Chroma upscaling, I am liking Jinc 3 with the anti-ringing filter enabled for sharper Luma scalers (basically anything above Mitchell-Netravali) and Spline 3 with the anti-ringing filter for the softer luma scalers.


For Luma downscaling, which I should point out is not something I use often, Catmull-Rom with both the anti-ringing filter and linear light options enabled seems to be the best choice in my testing so far. Most of the scaling algorithms look really bad in linear light to be honest, but Catmull-Rom combined with the anti-ringing filter seems to avoid most of those problems.

You should definitely not use SoftCubic on Chroma - at least that is my opinion. Even SoftCubic 50 is far too soft, and reduces the brightness of, or desaturates Chroma.

Thanks for the input i appreciate it. 99% of my material is hd so i will try jinc 3 with ar for chroma up, jinc 3 with ar for image up, and catmull-rom ar and linear for image down.

[NicolasRobidoux]

The most natural looking downsampler I know, for DSLR image downsampling, is EWA quadratic B-spline-windowed Jinc 3 (with no deblur, although the same deblur as madVR's Jinc 3 does not break anything), through linear light.
The first image I try when comparing downsampling methods is http://upload.wikimedia.org/wikipedi...p_Portrait.jpg which I downsize to 403x600.
Often, this is enough to disqualify a downsampler right there and then.
Also, anything but linear light is deadly.
Of course, making this run reasonably fast when downsampling would take a good amount of "craft".
And I have not checked how well it does with things like text. This being said, this is a scheme with fairly low ringing (without AR).
P.S. Examples of enlargements with this scheme, unfortunately with sigmoidization, which possibly only works well when the source material can "nicely" be converted to linear light without gamut "mismatch", are here: http://www.imagemagick.org/discourse...p?f=22&t=21933.
P.S.2 Downsampling examples are here: http://www.imagemagick.org/discourse...art=225#p90394

[leeperry]

Originally Posted by Xaurus

I would love to test it for you but a few minutes after my testing the SSD with Windows on it in my HTPC died. [..] you can discard my numbers for now.

Ah, bummer! I indeed read many horror stories with SSD's that would completely wipe themselves out of the blue, so I'll be fine with a ramdisk for my A/V apps until these things become more reliable.

Anyway, I'm getting this golden sample 650Ti, it's not that far from the 660 and it's $100 cheaper too. I'll post my tests results but I'm quite sure it'll deliver

[egur]

Originally Posted by NicolasRobidoux

Eric:
Here is an example of EWA LanczosSharp (a.k.a. Jinc 3) used with an alternate method, namely sigmoidization, that reduces halos and that Mathias found does not work well in some circumnstances. (And I'm going to have to double check on this, because it's worrisome. I'm actually not sure at all it would work well with Y'CbCr. And, after Mathias comments, I'm not even sure it always works well with sRGB pngs.)
http://web.cs.laurentian.ca/nrobidou...sSharp.7p5.png

Were can I get the original?
BTW, I want to publish a set of results of my scaler and I want to make sure I use the same images as Madshi\ and yourself.
Can you/Madshi post links and tell me how much to scale?

[NicolasRobidoux]

@Eric:
http://madshi.net/MiniOrg.png

[NicolasRobidoux]

@Eric:
I really like the test images and enlargement ratios discussed here (the end of the thread drifts; stop at the dragon): http://www.imagemagick.org/discourse...p?f=22&t=21804.
But Mathias or somebody else may feel that this is not representative of video content?

[6233638]

Originally Posted by NicolasRobidoux

The most natural looking downsampler I know, for DSLR image downsampling, is EWA quadratic B-spline-windowed Jinc 3 (with no deblur, although the same deblur as madVR's Jinc 3 does not break anything), through linear light.
The first image I try when comparing downsampling methods is http://upload.wikimedia.org/wikipedi...p_Portrait.jpg which I downsize to 403x600.
Often, this is enough to disqualify a downsampler right there and then.
Also, anything but linear light is deadly.
Of course, making this run reasonably fast when downsampling would take a good amount of "craft".
And I have not checked how well it does with things like text. This being said, this is a scheme with fairly low ringing (without AR.

It certainly is awkward to get a specific size when using MPC-HC. The scaling options are limited, so I can only scale to 406x605 or 398x593 using the zoom options, and resizing the window to get a 600px tall image resulted in 402x600.

I had to disable the aspect ratio correction and manually resize the window to get 403x600:

http://www.abload.de/img/4039kopd.png

There does seem to be somewhat more aliasing/moire, but I don't think you are likely to run into that problem when scaling video at all.

With all the other "sharp" scaling algorithms in madVR, there is nasty black ringing around bright objects when using linear light downscaling (even with the anti-ringing filter) at least with the images I tested (didn't try with this) which is a big part in why I settled on Catmull-Rom with the anti-ringing filter enabled.

[NicolasRobidoux]

Originally Posted by 6233638

There does seem to be somewhat more aliasing/moire, but I don't think you are likely to run into that problem when scaling video at all. ... I settled on Catmull-Rom with the anti-ringing filter enabled.

This is a good looking result. Sure, EWA QuadraticJinc 3 has less moire, but again this is a rather extreme test case. Actually, if anything is not quite right with a scheme, this test image and downsampling ratio will make it jump at you. Schemes considered "reasonable" will produce absolutely awful results.
It certainly illustrates how well AR works (Good job Mathias!) because the Catmull-Rom result manages a nudge less halo than EWA QuadraticJinc 3, and it is also a nudge sharper.
I'm impressed.
P.S. I'm VERY impressed. Squeezing this good a result out of Catmull-Rom filtering... This is a better result than with the overwhelming majority of schemes I've tested. And believe me, me and others have looked at a lot of flies.
P.S. This being said, EWA QuadraticJinc clearly wins the moire contest.

[mzso]

Originally Posted by 6233638

The majority of motion blur seen on LCDs is now caused by the human vision system due to LCDs being sample-and-hold type displays rather than being impulse-driven like CRTs were. (and Plasmas somewhat are)

This is why modern LCDs with scanning backlights can have essentially zero motion blur.

OLED is similar to LCD in that it is a sample-and-hold display technology if it's not using black frame insertion.


Interpolation will still be necessary because 24p is not nearly a high enough framerate to perceive fluid motion. (native 24p judders a lot on panning)

Wholeheartedly agree on the last part. Sceptical of the preceding parts. 24p is a blur and choppyness fest no matter what you do with it. Also I have my doubts that BFI does any gooda at 48p+.

Originally Posted by 6233638

There's no reason OLED couldn't be more like LCDs with LED backlight scanning, rather than being more like CRT/Plasma.

I see no reason why it couldn't be both. Meaning black frame insertion as a feature that could be enabled.

Originally Posted by 6233638

LG are using RGBW subpixels, and Samsung appear to be using a vertical subpixel arrangement. Neither of these are suited to use as a monitor, which expects horizontally arranged RGB stripes.

http://roricon.egloos.com/m/3810249

This could only be a problem with sub pixel anti-aliasing. And I think it's trivial to update d2d and whatnot to support new arrangement. Actually I seem to recall faintly that every order/direction of RGB is supported.

[NicolasRobidoux]

Here's the other "downsampler disqualifier": The "Backpack" http://www.mediafire.com/view/?uz8dphpj6q0y6n5#.
I normally downsample it to 1200x801. The result with QuadraticJinc 3 (through linear light, always) is here: http://www.imagemagick.org/discourse...art=225#p90394

[6233638]

Originally Posted by NicolasRobidoux

Here's the other "downsampler disqualifier": The "Backpack" http://www.mediafire.com/view/?uz8dphpj6q0y6n5#.
I normally downsample it to 1200x801. The result with QuadraticJinc 3 (through linear light, always) is here: http://www.imagemagick.org/discourse...art=225#p90394

http://www.abload.de/img/1200k5st5.png

I keep forgetting to disable my calibration settings and dithering when taking these, meaning I have to do it several times.
Again, I don't think you are going to see high-frequency detail like that, even with most Blu-ray discs, so the increased aliasing isn't too much of a concern for video.

Originally Posted by mzso

Wholeheartedly agree on the last part. Sceptical of the preceding parts. 24p is a blur and choppyness fest no matter what you do with it. Also I have my doubts that BFI does any gooda at 48p+.

24p shouldn't have any motion blur other than what is in the source.

I play games at 60p, and there's still a huge difference with backlight scanning on/off with my LCD. (not a response time issue - high speed photographs don't show any motion blur on the panel)
It's an issue of perception, and the same reason why when you use a strobe light, you are able to see things clearly that would otherwise have been a blur - it's the same thing happening here with LCD/OLED. (that doesn't use BFI/Backlight Scanning)

Originally Posted by mzso

This could only be a problem with sub pixel anti-aliasing. And I think it's trivial to update d2d and whatnot to support new arrangement. Actually I seem to recall faintly that every order/direction of RGB is supported.

Pixel structure definitely affects more than just subpixel font rendering. Any kind of computer graphics tend to have problems with non-standard layouts. There are displays out there which modulate pixel brightness by only adjusting half a subpixel (or even less) and they show obvious artefacts when used as a computer display.