Advice on XviD Quant Matrices

[Satlover]

I normally go for a 1.9GB conversion including AC3. After reading the Sticky:"Newbie" Settings I get really good results but I was wondering If it is overkill in size compared to quality if some of the setting are not being used correctly . Should I use a custom XviD Quant Matrice instead of the default one? any recommendations?

Here is what I use
AviSynth 2.53, MPEG2Dec3.dll V1.10,VirtualDubmod 1.5.10.1 (2424), XviD-1.0-RC2-07022004.exe, Video size:same as DVD. VHQ-4, All settings in the profile AS@L5 are switched off.

Anything I should be aware off :rolleyes:

[iago]

Satlover,

First thing you should be aware of is upgrade to XviD-RC3!

Second, you might wanna downgrade to VirtualDubMod 1.5.4.1.

Third, you may upgrade to Avisynth 2.54.

Also, (as a reply to your post in another thread,) latest release of ffdshow is "ffdshow-20040227".

Last, you might try using high bitrate matrices such as "hvs-best-picture" or one of "andreas" matrices that comes packed with XviD.

And, as far as I remember, I had told some of these before in another thread too, hadn't I? :rolleyes:

-iago

[Bogalvator]

Also ;)

You should upgrade from mpeg2dec3 v1.10 to the bugfixed DVD2AVI/MPEG2Dec3 combo - available from www.avisynth.org/warpenterprises/ or http://neuron2.net/mine.html

Also, have your B-frames set at 1/1/0, sensitivity 0, and turn Q-Pel on.

I can second the recomendation for the HVS-Best matrix, gives good results with low quantizers IMO.

[Teegedeck]

As HVS best-picture doesn't even require a very high bitrate, I think Didées 'six of nine' matrix deserves a mention, too.

[Satlover]

@everyone,

Frist thank you all for taking the time,

I am really pissed off with my self:angry: I am using the latest software that iago and Bogalvator mention except for RC3 which I will install today. I was just to damn tired and lazy last night to check my post so I copied a pasted from my last post without seeing the mistake :rolleyes: Really sorry about that. I will look forward to trying the HVS best-picture and 'six of nine' matrix. Will also try the B-frames set at 1/1/0, sensitivity 0, and turn Q-Pel on thank you.

Lesson learnt here.

[Satlover]

Originally posted by Teegedeck
As HVS best-picture doesn't even require a very high bitrate, I think Didées 'six of nine' matrix deserves a mention, too.

I have searched the Forum and Google for Didées 'six of nine' matrix but I cannot seem to find it.

Thanks

[Satlover]

Originally posted by Bogalvator
have your B-frames set at 1/1/0, sensitivity 0, and turn Q-Pel on.


So you mean

Max c. BVOPS 2
Quantizer Ratio 1
Quantizer offset 0 ??

I am using XVID RC3

Thanks

[atreya2011]

Originally posted by Satlover
I have searched the Forum and Google for Didées 'six of nine' matrix but I cannot seem to find it.
Thanks

Here you go, Didée's SixOfNine Matrix, I have'nt tried it yet though.
Is it good to try with bitrates lower than 590kbps?(the source is denoised ofcourse)


The SixOfNine Matrix (http://forum.doom9.org/showthread.php?threadid=54147&perpage=20&highlight=SixOfNine&pagenumber=2)

Just scroll down a bit and you will find it ;)

Originally posted by Satlover
So you mean

Max c. BVOPS 2
Quantizer Ratio 1
Quantizer offset 0 ??

I am using XVID RC3

Thanks

He means
Max BVOPs = 1
Ratio = 1
Offset = 0

[Satlover]

Originally posted by atreya2011
Is it good to try with bitrates lower than 590kbps?(the source is denoised ofcourse)


Thanks allot, allthough I go for High bitrates (total size of 1.9GB)
but I will give it a try. If anyone has it to import I will be glad if it could be mailed to me ?

Thanks for all the help

[Teegedeck]

6 of 9 ist definitely not a low-bitrate matrix.

[Satlover]

Originally posted by Teegedeck
6 of 9 ist definitely not a low-bitrate matrix.

Teegedeck, would you be so kind and mail it to me to import. If not I will just manually type the figures in.
Thanks

[Prettz]

I have a question not related to satlover's, but it is related to quant matrices. If you're using a custom matrix like, say, hvs-better, in place of h.263 for a 1 CD rip, is using trellis quant recommended?

[Teegedeck]

Hopefully no typos in it.

[Didée]

Since some talk is going on about SixOfNine, perhaps I should say something, too ;)

Originally posted by Teegedeck
6 of 9 ist definitely not a low-bitrate matrix.

Definetly not!

There are two things I'd like to mention about the SixOfNine-matrix that's posted in abovementioned thread:

1. The resulting bitrate might be too high on occasion

2. That matrix is not completely safe for playback through ffdshow - not currently, at least. Perhaps libavcodec will get updated somewhen, in that aspect.

These days, personally I use a modified version of SixOfNine, that

- produces a little lower bitrate
- is safe to use along with ffdshow
- has gotten a little of the "characteristics" of the HVS-matrices plugged on top of it (just a little).

Hence, it's now called "SixOfNine-HVS", and looks like this:
Quant-Matrix "SixOfNine - HVS"

08 11 12 12 13 15 16 17 12 11 12 12 13 14 15 16
11 11 12 12 14 15 16 17 11 11 12 12 13 14 14 16
12 12 13 14 15 16 17 18 12 12 12 13 14 15 16 17
12 12 14 16 17 18 19 19 12 12 13 15 16 17 18 18
13 14 15 17 19 20 20 20 13 13 14 16 18 19 19 20
15 15 16 18 20 21 22 22 14 14 15 17 19 20 22 22
16 16 17 19 21 22 23 24 15 14 16 18 19 22 23 24
17 17 18 19 20 22 24 24 16 16 17 18 20 22 24 24

Happy typing ;)

Didée

[Satlover]

Didée

Thanks allot,

ohhhh - any chance to download it? my eyes tend to go a bit funny when I have to type all this in :rolleyes:

[atreya2011]

Is there any other custom matrix I should experiment with for 1CD Rips or low bitrate rips, other than the default H.263???

[Teegedeck]

Thanks, Didée! This will be fun testing. Though I don't give much about ffdshow compatibility; Andreas' 78er doesn't play back cleanly with it, either.

@atreya2011: Depends on your taste; try the HVS matrices. Personally, I like HVS best-picture at low bitrates, too.

@Satlover: You can't be so keen on it, otherwise you'd have typed it by now. :p

[Satlover]

I am looking forward to test it along with all the other adivce I got, but im going to type it all in tommorow. Maybe someone might have uploaded it by then ;)

[Teegedeck]

atreya2011, you PM'ed me that you wanted to post an attachment (don't PM for that) but I don't see anything to validate. Of coures I'd be happy to validate maybe a quantization matrix? ;)

[jang0]

would like to post the 6of9-hvs, but

"You may not post attachments"

[ViCroié]

Hey,

4 all lazy ppl out there i uploaded 6of9-hvs on my serv;)

URL ::: http://home.planet.nl/~donse113/6of9-hvs.txt

Edit:: Corrected typo,
thanks Satlover

(uploaded corrected version)


Have Fun :D

[Satlover]

Thanks ViCroié,

Small typo in Intra Matrix at the bottom 184
thanks again

[Bogalvator]

Originally posted by Prettz
I have a question not related to satlover's, but it is related to quant matrices. If you're using a custom matrix like, say, hvs-better, in place of h.263 for a 1 CD rip, is using trellis quant recommended?

Well, I personally WOULD recommend it, especially for 1 CD rips, but to use the old cliche - test for yourself and see what's best for you yada yada yada ;)

I must confess I've never heard of this 6 of 9 matrix, I shall have to give it a whirl........

[atreya2011]

Originally posted by Teegedeck
atreya2011, you PM'ed me that you wanted to post an attachment (don't PM for that) but I don't see anything to validate. Of coures I'd be happy to validate maybe a quantization matrix? ;)

Yup I wanted to attach SixOfNine-HVS, but someone has already done that. :)

EDIT:
I did a 2 CD rip of X2 with that matrix. Quality is pretty good better that MPEG default matrix. I have compare with HVS-Best and see the results for myself.

[dapipa]

Hi!Just 1 question:what matrix do you recommend for bitrates around/above 2Mbps?I did some serious testing(took a month,or so:))with a few DVD sources(e.g.Solaris/R2,only UnDot filter used and then some Lanczos resizing)and some DVB sources(e.g.Record of Lodoss War/VOX,deinterlaced with DeComb,then UnDot,Convolution3DYV12 and Lanczos resizing were used),I tested H.263,MPEG,HVS-Best and Andreas-78er with various XviD settings(RC2,2-Pass,QPel on,BVOPs1/1/1(or off),Chroma motion on,MSP6,VHQ1(or 4),default Quantizers(or min.Q.set to 2),Trellis on,for animes:Cartoon mode turned on and,of course,bitrate was set to 1920kbps,or length to(1920kbps*length of source in seconds)/8,all encoding was done in VirtualDubMod1.5.4.1),then i compared the results using CompareYV12(or SSIM)and...H.263,BVOPs on and VHQ4 yielded the best PSNR/SSIM results and I was unable to see any visual difference between various encodes(and I'm not visually challenged,for sure:))...So,are those custom matrices any good for high bitrates?P.S.I can post exact results(and maybe some short samples)if anybody wants:)

[Teegedeck]

Well, for one thing I don't see much need to resize at all at really high bitrates. I would not even use undot as it removes a wee bit of detail (noticeable on skin-texture).

And yes, the matrices we're discussing are good for high-bitrate encodes. I can see a difference provided that the source itself has a high quality. It not only depends on your eyes but also on your screen of course. Using a decent TFT is a good idea.

[Satlover]

@Teegedeck

It was interesting what you said about not using undot. I will try without it for high bitrates (1.9GB) Thanks for the tip :D

[dapipa]

Originally posted by Teegedeck
Well,for one thing I don't see much need to resize at all at really high bitrates.

hi!i was resizing just the vertical axis to get correct aspect ratio...

p.s.another(theoretical)question:if you do a fixed q.2 encode using H.263,you'll end up with e.g.125MB file and the same encode using MPEG matrix'll yield to 150MB file,which 1 will look better?i suppose the MPEG one,because it keeps more detail...then,you'll do 2-pass encode of the same source with target file size set to e.g.100MB,which one will look better now?H.263 using quants higher than 2,to hit the target FS,or MPEG,using even higher quants to shrink the 150MB file to 100MB?

p.p.s.file sizes in the example are taken from here (http://forum.doom9.org/showthread.php?s=&threadid=57096).
:)

[dapipa]

no1 willing 2 answer?
:)

[Didée]

We're all sitting here, getting blue faces from holding breath, whilst eagerly awaitening the results from the tests that *you* did on that topic. ;)

This is the old, long-bearded story of the h.263 <-> mpeg debate. So far, the common conclusion is as follows:

The "better" quantization type is exactly that one that looks better to *your* eyes.

Remark: It's not true each-and-everytime that h.263 is delivering smaller filesizes at the same quant. You might stumble over content where it's just the other way round.

- Didée

[frodoontop]

When deciding if you're gonna use a matrix "better" than h263, you can just do a 2pass on a sample of the movie and see in xvid statistics if the average quant gets a lot higher. (Just to be clear: with better I mean that it keeps more information of the source.)

If difference if quite acceptable I tend to use hvs-best matrix (my favourite) since I really like the picture detail I get. And the "better" the matrix, the less visible will the higher quants be.

Use avisynth to set up such a sample, for example add this line to your script to get a 10% sample:

Selectrangeevery(1000,100)

[dapipa]

ook,will post the results tomorrow...

[kilg0r3]

Just in case any body cares, I learned from another thread that streams encoded with the 6f9 matrx are not playable on the elta 8883 standalone players.

[dapipa]

Originally posted by Didée
We're all sitting here, getting blue faces from holding breath, whilst eagerly awaitening the results from the tests that *you* did on that topic. ;)

i wrote i'd post them if anybody wants,but no1 did.
:(

Originally posted by Didée
This is the old, long-bearded story of the h.263 <-> mpeg debate.

i think you didn't understand my question:forget h.263 <-> mpeg,it could have been mpeg <-> hvs_best,or andreas_78 <-> 6of9...the question is,when some matrix X keeps more detail than matrix Y,it(usually)produces larger file(say Z k/M/whatever bytes)in fixed quant encode,right?but if you do 2 pass encode and your desired target file size is(significantly?)lower than Z,codec has to use higher quants in order to reach it,right?so,what will be the impact on overall quality of the encode then?i suppose,when i get undersized file with h.263,it's a good idea to use 'better' matrix to fill desired space(CD/DVD)and thus gain more quality from the source,but how does it work the other way around?

now to my tests:i've chosen to redo them with the new RC3,and here (http://www.fi.muni.cz/~xlukac/solaris.csv) are the first results(solaris/R2)in csv format(columns represent:first encoded frame of the source,last frame,xvid settings,PSNR,SSIM,first pass time,second pass time and overall time)...will do more encodes if it's of any use.
:)

p.s.notice to the encoded ranges i've chosen:first i encoded about a minute long scene(spaceship flying over solaris),which(i think)was quite problematic,then i encoded first 10 minutes of the movie using those settings for each matrix,which yielded the best results in the first step...due to lack of free time i didn't do step 3:full encodes of the movie like the first time(it took me about a month with various sources then),but i can make some selectrangeevery(1000,100) encodes,proposed by frodoontop...

[edit]
redone step 2 with selectrangeevery(1000,100) instead of trim(0,14999) in the script,plus h.263 encoded using fpu/ssemmx IDCT instead of ref,csv is here (http://www.fi.muni.cz/~xlukac/solaris.sre.csv)...

p.s.i'm going abroad tomorrow for about a week,that means:i'll post more results after that,so you can release your breath now and stand up.
:)