QTGMC Deinterlacing Script (v3.32)

[Stormborec]

Btw. what type of sharpening is used? It will be something with removegrain, but - is it nonlinear? Because the text shouldn't be much sharpened, rather details in flat areas ...

[Dogway]

I see this is a very static grain free source, so that changes it all. Maybe you would prefer to set lower grain restore values or remove it completely.
Use higher temporal radius, maybe tr2=1 could be a nice idea to remove some mosquito noise.
Also the internal used sharpening method is temporal limited, so you will see more differences on motion (much blurrier in the lossless method), although you can already see some detail gain on the hair, dress pattern and monitors on the desk. You can try to set the limiting to spatial SLMode=3.

If you want upload a chunk so I can play and test.

[Stormborec]

Let's play ...
http://uloz.to/x5F8hLQT/zpravy-3-10-2013-mpg
http://uloz.to/xfiXkzNX/zpravy-7-11-2013-mpg

I use dgmpgdec and then mpeg2source("1.d2v") ...

My idea is: no need of change of QTGMC parameters according the scene change.

Input vs. output:
No change in static areas, no sharpening, no grain adding nor no noise removing ... But perhaps, if the sharpening would work very well, then it is perhaps better, than adding e.g. seesaw after deinterlacing.

Also I have no idea, how much is "rounding steps" in the QTGMC proccess - if it does something like: nnedi => rounding to intiger => sharpen => rounding to intiger => TR2 => rounding to intiger ... then could be better to do as a few steps as possible

[Dogway]

QTGMC doesn't add grain, it only recovers, but sometimes you underestimate the amount of grain/noise in source and would prefer some slight safe denoising. For a theoretical approach you would recover all grain/noise and turn off sharpening, but my opinion is there are better options. So it comes down to a balance between grain/noise restore and denoising strength, for either lossless or standard deinterlacing.

Here are comparisons using similar settings for lossless and standard:

In the left the best lossless I could get, and standard QTGMC on the right


Lossless is good for tiny detail patterns like the teeth and collar in the image, things like grass, etc. For everything else I find standard a bit better, it retains textures, check neck, forehead, and dark section with monitors. So the lossless mode has a bit more definition on very small details, but muddy on everything else (normally what a dumb sharpener does), while standard has a better balance of overall sharpening, it gives that unprocessed "source feeling" in my opinion. It's true that lossless retains better text definition, but this is a very specific situation, if all your source was like this; on-screen text, detailed patterns, no grain or lack of texture maybe you could go with lossless, but for a general approach I think standard is better.

Here is another example with motion scenes using same settings. Standard brings more temporal and spatial cohesion so kinda acting like motion blur while lossless tries to retain original blocking/artifacts that are no relevant to visual information.

[Stormborec]

Uf, it's tricky ...

mpeg2source("1.d2v")
crop(0,0,-464,-448)
x=last.trim(115,115).sharpen(1)
a=QTGMC("very slow",tr2=0).trim(230,230).sharpen(1)
b=QTGMC("very slow",sourcematch=2,Lossless=2,smode=0,tr2=0).trim(230,230).sharpen(1)


interleave(x,a,b,x)

pointresize(width*4,height*4)

[Stormborec]

I was reading the forum, and I found this two posts from -Vit-, concerning my question:

http://forum.doom9.org/showpost.php?...postcount=1280

http://forum.doom9.org/showpost.php?...postcount=1264

Maybe it's the solution ...

[Dogway]

yeah... what I said, for theoretical source reproduction you restore all grain and do no sharpness, while on lossless mode. But the point is... what for? If you want a master you already got the master, make a backup of it, it's already compressed. There is little point on having a deinterlaced lossless copy of your master since it takes loads of disk space.

If you want to do a remaster, you are supposed to enhance visuals in the safest form possible, like not being too artsy or subject to interpretation. I think standard mode with grain restore is a good candidate for remaster purposes in most cases... but that's me. I like to hear opinions.

[Stormborec]

I found still one possibility here:

tdeint(mode=1,edeint=nnedi3(field=-2))

but it's a bit exterme ...

If "lossless" - then would be hard to find something better than:

qtgmc(sourcematch=2,lossless=2,smode=0) ...

[yup]

Hi all!
I try QTGMC with YV16 source, script simple

AVISource("samplevhs.avi")# YUV 422 source
AssumeTFF()
ConvertToYV16(interlaced=true)
QTGMC()

At bottom half I see green or pink background or pink and green dig.
Not all frame, not regular, but frequently.
I use last Set avisynth build and plugin from
http://forum.doom9.org/showthread.ph...50#post1699950
Please advice.
yup.

[Reel.Deel]

QTGMC and MVTools still uses the old and ugly planar YUY2 hack from the SSE*Tools plugin. So if you want to process YUY2 you need to use the older plugins like MaskTools2 for AviSynth 2.5, RemoveGrain, Repair, VerticalCleaner (if needed) and you will also need the SSE*Tools plugin. I really wish cretindesalpes would add native support for YV16 in MVTools. According to jackoneill the code already supports YV16 it just needs to be used. It would make processing YUY2/YV16 way easier and less convoluted.

[TheSkiller]

While comparing between YUY2 and normal YV12 processing I found that the luma plane is in fact not identical but very slightly different between the two. Is this normal? I don't see why the luma processing would differ if I'm running QTGMC in YUY2 mode.

I also found that in YUY2 mode there sometimes seem to be problems with residual chroma combs if the video dimensions are not mod16 (no, it's not a cropping mistake).


I think we really need proper YV16 support for QTGMC.

[feisty2]

Originally Posted by TheSkiller

I found that the luma plane is in fact not identical but very slightly different between the two. Is this normal?

yup, normal, cuz motion estimation refers to all planes by default (chroma=true), set chroma=false in msuper and manalyze, the results will be identical.

[LouieChuckyMerry]

Hello. As I was reading this thread from the beginning, in preparation to attack my early-seasons NTSC Simpsons discs (thinking I could deinterlace and denoise in a single step, given the limited quality of the sources), I came upon several comments to the effect that QTGMC wasn't necessarily the best option for animated sources (these comments are quite early in the thread). Please, is this still the case? I ask because the results of QTGMC with default settings on a couple of test clips was vastly superior to any of MeGUI's 3 default options (after analyzing the clips), TIVTC, TIVTC + TDeint(EDI), and Decomb IVTC (plus I'd really appreciate not having to read another seventy-some pages of this thread if it's not necessary ;-). Thanks in advance for any help.

[LigH]

Because the reason won't change, the result won't change either: QTGMC is a deinterlacer which relies on the assumption that there are fields recorded in equal temporal distances, so from field to field, the time when it was shot progresses regularly, more or less linearly.

Real cartoons don't have the same progress for every frame, because their low frame rate (probably about 8 fps) causes repeated content over several frames, then suddenly the content changes from one field to another. This is no "regular motion", therefore a deinterlacer would calculate wrong interpolations, assuming that the previous and following fields continue with about the same amount of motion.

Similar case for Telecine: There is no regular progress through the fields either, instead, there has been a progressive content before, which can be restored almost losslessly; deinterlacing with interpolation would already be worse due to the interpolation, and on top even more worse because there is no (more or less) "linear" progress in motion following the field order.

But cartoons are certainly the harder case, because their progress pattern is harder to detect due to the many duplicates, and when there is a "hybrid animation" (with computer generated or at least computer assisted cartoons), the optimal technique to recover progressive frames would even change scene-wise... Deinterlacers are simply not meant to handle that. Neither of them. Not even QTGMC.

[creaothceann]

To summarize:

• source: interlaced 30/1.001 fps VIDEO (e.g. TV recordings, most DVD content)
• deinterlacer (e.g. QTGMC) → progressive 30/1.001 fps VIDEO
• IVTC (inverse telecine/pulldown) → progressive 24/1.001 fps FILM (movies, anime), tricky in case of irregular patterns or interlaced content (e.g. credits)

[TheSkiller]

Originally Posted by creaothceann

• deinterlacer (e.g. QTGMC) → progressive 30/1.001 fps VIDEO

No, you get progressive 60000/1001 (59.94) fps video after QTGMC, unless you throw away half of the video (every other frame) using FPSDivisor=2.

[LigH]

But IVTC can only revert Telecine if your material was processed with 3:2 pulldown at all. You will always have to test if either IVTC or Deinterlace would work. Never assume, always test, by watching the content field-wise and slowly.

[Boulder]

And the next round to make things even more complicated: material that looks like it's interlaced but it's actually fieldblended

[LouieChuckyMerry]

Thank you all for your informative replies .

Originally Posted by creaothceann

To summarize:

• source: interlaced 30/1.001 fps VIDEO (e.g. TV recordings, most DVD content)
• deinterlacer (e.g. QTGMC) → progressive 30/1.001 fps VIDEO
• IVTC (inverse telecine/pulldown) → progressive 24/1.001 fps FILM (movies, anime), tricky in case of irregular patterns or interlaced content (e.g. credits)

Would this be considered the best method of attack for material such as mine? For example, is there a better deinterlacer than QTGMC for SD animation? Would I output this stage to lossless or is it stage one of a two-stage script? I apologize for my ignorance; I'm rather new to this realm, but capable of learning and have a functional MT version of AviSynth, and would very much appreciate suggestions. Thanks again for your time.

[LigH]

Cartoons are often victims of the strangest kinds of norm conversions where at least one person in the video studio had no clue how to handle the material with care. Therefore it will be hard to recommend any recovery method without a sample scene. But quite certainly, a real deinterlacer (like QTGMC and many others) is not the recommended choice here. Because a cartoon has no linear motion, any filter relying on linear motion will fail.