XviD and very very low bitrate

[pogo stick]

Is there custom matrix for a very very low bitrate?
I am trying to encode Simpsons to fit 9-10 episodes on CD at resolution 384x288. With XviD resulting video is always too blocky for me even if cartoon mode is used. I tried ultra low bitrate matrix that came with Koepi's installer, but it doesn't help. And with RV10 video is very soft and almost blockless compared to XviD at the same filesizes. Details are killed, but I would prefer this way than to have many blocks. And Simpsons don't have that much details. So is it possible to have this kind of video in XviD using custom matrix? I much prefer XviD to RV10 because of the encoding speed and using comfort (I am kinda slow and it took a day to read, to understand how RV's jobfiles work and to set everything right). But when it comes to low bitrates I can't get XviD to work right.
And I noticed that in last Doom9's codecs comparison bframes were not used in XviD recommended settings for Futurama. Is it better for animation? If so then why?
Thanks for any help in advance.

[ObiKenobi]

Well to help you you may want to bump up the bframe sensitivity and or doing some prefiltering, otherwise you are just gonna either have to accept blocking or soft video as there is only so much you can compress a video without it suffering problems.

[bond]

definitely people have to use preprocessing at very low bitrates

i am not the cartoon specialist but start with the avisynth filter "unfilter()"
its pretty easy to handle and fast

[SoonUDie]

h.263 is generally pretty good for cartoons. You might also want to try some of the HVS matrices; they seem to work decently at low bitrates.

You may also want to take a look at >this (http://forum.doom9.org/showthread.php?s=&threadid=74463)< thread and try out my matrix and Jawor's matrix.

[Soulhunter]

Or try this (http://forum.doom9.org/showthread.php?s=&postid=468340#post468340) as well... ;)


Bye

[Malevolent]

XviD with:
3 or more B-frames & B-VOP sensitivity 5-10
Adaptive Quantization
Trellis
Recommended Matrix: HSV-good

Avisynth with:

ConvertToYUY2()
PixieDust()
ConvertToYV12)
Undot()
MipSmooth(Preset="AnimeHQ")
Deen("a3d",Mode=1,Thry=8,Thruv=10,TI=3)
HQdering()

It'll still be a pain, but considerably better picture this way.
You can also get a few mb's extra to play with by reducing saturation & contrast:
Tweak(0.0,0.99,0.0,0.99)
This would be (nearly) non-detectable & take out 2-4mb from size ;)

[Chainmax]

Isn't adaptive quantization a bad idea for cartoons since it removes detail in large, uniformly coloured areas. AFAIK, it should only be used on movies.

[Nibor]

Originally posted by Chainmax
... it removes detail in large, uniformly coloured areas. ...
Why should this be bad?
Remove details where you can't see them anyway -> higher compression -> more bits to spend for other more complex areas like faces and the like...
Especially in cartoons, where there are a lot uniformly coloured areas, adaptive quantisation could help.

Or not?

[SoonUDie]

From my personal experience, Adaptive Quant tends to kill details in some areas that are definately visible, and doesn't add noticeably to the quality of the other areas.

[RadicalEd]

I'm pretty sure AQ only works on very light/dark parts of an image, hence the prior title 'lumimasking'.

[Teegedeck]

AQ doesn't look good on animation for all I can tell.

We should keep in mind that the whole concept of adaptive quantization is psychovisual and meant for natural images. It's takes adavantage of human visual perception which normally doesn't value every part of an image evenly.

Animation seems to be perceived differently; we seem to value every part of the picture pretty much the same, more like the way we look at an abstract or cubist painting. This is how I explain it for myself.

@SoonUDie: Of course it does something for those other areas - it lowers the quantizers for them (not in each frame and also this redistribution of bitrate goes for the whole clip, not for individual frames). This can be measured or seen when you use ffdshows 'visualization' feature. The bad thing is that it doesn't look better subjectively in the end IMHO when it redistributes bits within the frame like that.

[SoonUDie]

Originally posted by Teegedeck
The bad thing is that it doesn't look better subjectively in the end IMHO when it redistributes bits within the frame like that. That's what I was saying. I'm glad to see that we concur :D

[pogo stick]

I did preprocessing. Maybe not very strong one.

Well, I tried a few matrixes. As I said, H.263 and Ultra Low Bitrate.
And Jawor 1CD, Evil v0.1, Soulhunters Anime later with the same settings.

AVS was:
SeparateFields().SelectEven()
crop(16,2,-8,-2)
BicubicResize(384,288,0,0.5)
Undot()
FluxSmooth(10,15)

First Simpsons episode (The Simpsons Christmas Special #7G08)
Running time: 22:16
Video size: 69977 KB
Bitrate: 429 KBit/s

All default XviD settings except:
Packed bitstream disabled
Chroma Optimizer enabled
Cartoon mode enabled
Trellis enabled
Min Quantizers at 2

1pass: 15 min
2pass: 20 min

Matrix / Total video size from stat file / 5% Compress. check from Gordian Knot

H.263 / 195 409 KBytes / 36.2
Ultra Low Bitrate / 172 821 KBytes / 39.7
Jawor 1CD / 164 031 KBytes / 41.1
evil v0.1 / 165 622 KBytes / 40.5
Soulhunter Anime / 199 590 KBytes / 36.9

Also I encoded the same video in RealVideo 10 'Elysian' 2 times.
New RC wasn't recommended for low bitrate, but I decide to try.
inloopCutOffQuant at 0 (always use) first time and 31 (disable) second time.
Same target video size: 69977 KB. But they came about 1 MB higher.
Total video size from stat file: 206 566 KBytes both files.
1pass: 19 min
2pass: 44 min

I guess it's not exactly quality test. Bitrates are too low to compare details.
I want to find out settings at which XviD would look like RV10.
But it seems like only strong preprocessing can help.
Every XviD sample looks too blocky and too dirty.
Some of them more, some less.
And info from 'RealVideo 9 Information' thread: "It was not optimized for low bitrates, but to avoid compression artifacts at any bitrate." seems to be true.
Even with inloop disabled it produces less artifacts then XviD.
I made 5 snapshots for each sample. I don't know if you can find it helpful, but you can see it here:

1.rar (http://pogo-stick.narod.ru/1.rar) 1033 KB
2.rar (http://pogo-stick.narod.ru/2.rar) 1891 KB
3.rar (http://pogo-stick.narod.ru/3.rar) 1544 KB
4.rar (http://pogo-stick.narod.ru/4.rar) 1112 KB
5.rar (http://pogo-stick.narod.ru/5.rar) 1544 KB

I should try some heavy denoising later.
And what codec settings should I change?
I will use 3 B-frames and higher B-VOP sensitivity.
What about GMC and QPixels for low bitrats?
Can lowering min Quantizers help?

[ObiKenobi]

Well you have to remember that RV codec does alot of detail smoothing to try to avoid blocking in the foreground which is why it will seem better, which is about the equivalent of doing alot of smoothing then encoding with xvid. And why are you trying to cram so much on a cd anyway? Why not just do like 4-5 eps a disc and give it a bit more bitrate so it won't look bad. And BTW, the link to those rars doesn't work.

[Soulhunter]

Originally posted by ObiKenobi
And BTW, the link to those rars doesn't work. Try -> Save target as... ;)


Bye

[ObiKenobi]

Originally posted by Soulhunter
Try -> Save target as... ;)


Bye

I did, it just tries to save a html file.

Edit: Well now it will try to save the rar, but it just sits at 0k/s.

[Soulhunter]

Originally posted by ObiKenobi
I did, it just tries to save a html file.

Edit: Well now it will try to save the rar, but it just sits at 0k/s. For me it works, but only slow... :(


@pogo stick

Why are the samples in BMP format ???

If you wanna stay lossless use much smaller PNG's instead... ;)


Bye

[ObiKenobi]

Yeah these 22 meg rars are a pain to download when it goes really slow.

[pogo stick]

Originally posted by ObiKenobi
And why are you trying to cram so much on a cd anyway? Why not just do like 4-5 eps a disc and give it a bit more bitrate so it won't look bad.
I have a reason. I think it will not be interesting to read about it. :)
Personally I never encode at bitrates like that. But now I need to.
In the most cases I would agree with you.
Originally posted by ObiKenobi
And BTW, the link to those rars doesn't work.
Or try download managers, like Flashget.
I don't know where else I can put them.:(
Originally posted by Soulhunter
Why are the samples in BMP format ???
If you wanna stay lossless use much smaller PNG's instead... ;)
Didn't know about that.
But I guess RAR compressed BMPs well.

[Jaques]

Why don't you try VP6? It's optimized for low bitrates.

[SoonUDie]

Originally posted by Jaques
Why don't you try VP6? It's optimized for low bitrates. It would be interesting to see, at least. From my expieriences, VP6 tends to be blockier than XviD...

[Soulhunter]

Originally posted by pogo stick
But I guess RAR compressed BMPs well.
PNG could save 25% or much more (content depending)... ;)


Bye

[Didée]

Regarding the PNG <--> RAR'ed BMP issue:

About a year ago, I did some thourough comparison of exactly this : PNG vs. RAR'ed BMPs.

In the end, it turned out equal: sometimes PNG was better, sometimes RAR was better, when using its BMP optimized compression routines. But both were quite similar most of the times, execpt for few-colored things like e.g. WindoZe screenshots: there PNG was clearly better.

- Didée

[Tommy Carrot]

Well, for natural images, lossless jpeg2000 is usually 20-30% smaller than PNG. Of course most people couldn't watch it. :p

[Soulhunter]

Offtopic:


Small test with 8 random pics (mixed content)... ;)


BMP files = 22.275.504 Bytes

PNG files = 09.273.133 Bytes

JP2 files = 07.455.139 Bytes

JPG files = 03.727.141 Bytes


BMP in ZIP = 14.368.646 Bytes

PNG in ZIP = 09.272.321 Bytes

JP2 in ZIP = 07.458.026 Bytes

JPG in ZIP = 03.663.550 Bytes


Note:

JPG = Quality 100% / 4:4:4 (no subsampling)

JP2 = Lossless compression

PNG = Maximum compression

ZIP = Maximum compression



Bye

[pogo stick]

Originally posted by Jaques
Why don't you try VP6? It's optimized for low bitrates.
I just tried latest VP6 (sharpness=0) with the same source. And it seems to make less artifacts then XviD at low bitrate, but more then RV10 with inloop=0.
In case with XviD and more preprocessing, encoding time is increasing proportionally to filtering strength. So at comparable encoding time XviD picture can get more blurry, but not less noisy then RV10 picture.
And, in my opinion, Real Video is the best choice for low bitrate animation video.
Please, don't take it as advertising.

[Sirber]

I use now 500kbps (RV10) for anime content instead of 450kbps. Please don't deacivate inloop at this bitrate :)