Jawor's 1CD Quantization Matrix

[Jawor]

If you'd like to test it, just follow the link in my signature :D

Feedback will be appreciated :cool:

[Blue_MiSfit]

i will give it a try tonight. I'm feeling old-school so I think I'm going to encode breakfast at tiffany's :)

will get back to you on the results

[Jawor]

Thanks for your CPU time. :thanks:

So far I've tested this matrix on The Ninith Gate and The Matrix Revolutions (both R2 Poland). The 8% compressibility gain compared to HVS - Good Picture was observed when encoding with constant quantizers of 2, 3, 4, 5 and 6. The settings were:

Qpel, GMC, AQ, Trellis ON
Packed Bitstream OFF
Max I-frame interval 250
VHQ=4, MSP=6, ChromaME
B-frames 3/1.50/1.00
B-VOP sensitivity 3
Profile AS@l5
and the rest set to defaults

The Ninth Gate doesn't look brilliant even at quant=2 with both HVS - Good Picture and Jawor's 1CD (it seems to me that they overdid the edge enhancement while processing). I did not see any noticeable difference in quality.

The Matrix Revolutions encodes also look very similar to each other.

I had to use
SelectRangeEvery(500,25)
with both movies, so this was kinda compressibility test intended to check if further tests of this matrix have any sense. My computer is quite slow (it's a Celeron 1200 MHz), so further testing will take me a lot of time :( .

Now I'm doing a two-pass encode of The X-Files #8ABX03: Redrum, an analogue TV capture from TVP2 (one of the channels of the so-called public Polish TV, which is not public at all - they ask us to pay for it :devil: ) at desired video size of 180 MB (ca. 43 minutes, 512x384). I did an encode like this with the HVS - Good Picture matrix and this will be my first comparison at the same desired size, so that we could see how good is this matrix when working under difficult conditions. Unfortunately I won't be able to put here some screens or samples, but you can count on my report.

[Teegedeck]

How about comparing it to HVS-best as this one seems the 1-CD matrix of choice at least for bond, JasonFly and me? :)

[Jawor]

I intend to do it, but it will take some time. Of course you can do it yourself :D

[Blue_MiSfit]

Well... I did a test encode, but I am going to hold off passing judgment until I do a couple more encodes. My first test was pretty unsuccessful, with a lot of macroblocking and haloing. Here are the settings (which I think were a little flawed...):

AVS:
LoadPlugin("C:\PROGRA~1\GORDIA~1\mpeg2dec3.dll")
LoadPlugin("C:\PROGRA~1\GORDIA~1\undot.dll")
LoadPlugin("C:\PROGRA~1\GORDIA~1\peachsmoother.dll")

mpeg2source("D:\Tiffanys\Breakfast at Tiffany's.d2v")

crop(6,0,710,476)
LanczosResize(640,352)
Undot()
convertToYUY2()
PeachSmoother(NoiseReduction=45,Stability=15,Spatial=70,NoiseLevel=4.8,Baseline=3.2)
convertToYV12()

XviD RC4:
2 pass - target filesize: 609,612 kb
Adaptive Quant, Qpel
Bvops @ 5/1.0/1.5, Packed
MSP @ 6
VHQ @ 4
Trellis
All other settings at defaults

I got the following for the status window:
Average Quantizer: 5.04
B frame to P Frame ratio: 2.403
Only 1051 I Frames out of 164469 total!!

Whole thing encoded at about 25fps on the second pass with my system.

I think for this extremely grainy source ( Breakfast at Tiffany's ), PeachSmoother did a nice job of temporal smoothing, but even so this is a very difficult movie to encode. It looked nice in the I will do another run of this with some different settings. Exactly what those settings will be I am not totally sure, we'll see. Suggestions are welcome. I will get to the encoding in a few days since I don't want to stress my already highly tempermental power supply until my new 450 watt Antec arrives. I'm running like 400 watts of load on a 300 watt unit... It doesnt spin up the hard drives half of the time on boot now. scarry.

Sorry for the lengthy post, but I think this matrix has promise.

[Jawor]

Thanks for the results.

I would suggest using GMC. I know it's slowing things down and the compressibility gain is not big (or sometimes none), but for 1CD encodes GMC is sometimes helpful, so I think it's important to test this matrix with it.

Tomorrow I'll write something about this X-Files episode.

[Jawor]

As I promised, here are the results of tests performed on The X-Files #8ABX03: "Redrum" (25fps, progressive, 4:3)

XviD settings:
Profile: AS@L5
MSP=6, ChromaME, VHQ=1
Trellis, GMC, Qpel, AQ ON
Packed Bitstream OFF
B-frames 3/1.50/1.00
B-VOP Sensitivity 3 (movie), 10 (credits)
weight=0.4 for the end credits
Quantizers 2-31
maximum I-frame interval 250
"Fast" first pass
desired size 180 MB
I left the rest at their defaults

AviSynth script
AVISource(...)
ConvertToYV12()
Crop(4,4,-4,-2)
Deen("c3d",1,16,20)
UnDot()
LanczosResize(512,384)
UnDot()

I captured this episode using a Bt878A-based card at 768x576. The source was quite noisy, hence the usage of Deen() and 2*UnDot(). The clip was more compressible, than I thought, so artifacts were not so obvious. Jawor's 1CD Matrix achieved an average quantizer of 4.197, HVS - Good Picture gave avg. quant = 4.335.

Quantizers and frame types for Jawor's 1CD Matrix:
I-frames - 1.27 %
P-frames - 22.29 %
B-frames - 61.93 %
S-frames - 14.51 %
2 - 4.9 %
3 - 24.6 %
4 - 22.6 %
5 - 42.0 %
6 - 4.6 %
7 - 0.6 %
8 - 0.4 %
9 - 0.3 %
>9 - 0.1 %

Quantizers and frame types for HVS - Good Picture:
I-frames - 1.28 %
P-frames - 22.58 %
B-frames - 61.88 %
S-frames - 14.27 %
2 - 3.3 %
3 - 24.8 %
4 - 20.1 %
5 - 41.6 %
6 - 7.9 %
7 - 1.3 %
8 - 0.2 %
9 - 0.4 %
>9 - 0.3 %

Take a look at these two screenshots - the first one is Jawor's 1CD Matrix, the second one is from HVS - Good Picture encode:
http://img8.imageshack.us/img8/9069/jawors.png
http://img8.imageshack.us/img8/2426/hvsgood.png
Compare the mosquito noise on the edges - it's slightly more visible with HVS - Good Picture. Besides the mosquito noise, I haven't notice any diffrences between the two clips.

Currently I'm testing the same clip with the HVS - Best Picture matrix.

[SoonUDie]

If you don't mind a little competition, would you mind testing this matrix? I'm sure you'll be interested once you see it ;)
The Evil Matrix - v0.1 (http://students.oxy.edu/tcoldwell/forum_uploads/evil_v0.1.qmatrix)

[Jawor]

I don't mind a little competition :D

I'll test your matrix right after HVS - Best Picture with the same settings using the same X-Files episode.

[SoonUDie]

Thanks! I don't have time to run many tests, but on the material I've done so far (Men in Black II) it outperforms everything else at 1-CD bitrates. I got the values from a program designed to tune DCT matrices based on human perception, and adapted them for use in XviD (inter-frame matrix uses a higher compression tuning than intra-frame). I didn't exhuastively tune or test it though (came up with it a few hours ago, actually), which is why I'm asking for help. Thanks again for testing, I appreciate it. I'll make sure to try your matrix on my material too.

[kilg0r3]

Ther seem to be a few typos in your matrix since the vlues are not allways increasing in the downward and rightward direction, or, is this intended?

http://www.netcologne.de/~nc-allgeife8/1.jpg

And, please remember, the average quant does not tell you anything about image quality.

BTW, what is currently the best _free_ 'objective' method for measuring image quality.

[SoonUDie]

Yes, that's inteded for now, because it's the way the program calculates the values (it's designed to optimize JPEG DTC matrices). I may try to transpose the values, but I never found out exactly what XviD's scanning method was. I think it's right, down left, down, up right, etc. Also, this was optimized for a specific scene, so it's possible the values were affected by that. It's also possible that the numbers might reflect certain human perceptual traits (sensitivity to certain "zones" within the DCT table) since the program generated it according to it's perceptual model.

In any case, typo or not, I was able to run some short tests - and yet again, this matrix came up with more detail and less artifacts that any other matrix I've tested so far (h.263, Jawor's, HVS Best, mpeg, and a few more of my own). I eventually want to test every matrix that was included with LigH's quant editor. I also hope other people can test it, so that it can be exposed to a large number and type of test cases. If people report good results, I may try to come up with a better one based off of the same principles. At the moment, the scene that it's doing really well on is high-motion.

[SoonUDie]

Originally posted by kilg0r3
BTW, what is currently the best _free_ 'objective' method for measuring image quality. I've heard SSIM is pretty good, and at least better than PSNR. There are a couple others available, many of which have been "advertised" in these forums. :search:

The program I used to create these matrices also has a "quality metric" mode. It's called DCTune, and you can get it here: http://vision.arc.nasa.gov/dctune/

[dimzon]

2Jawor
Nice matrix! I've made 1CD DVRRip "Gigli" (2hr3min 640*352) at 670kbps - it's fine and clear
thanx

[Jawor]

@dimzon: I'm glad I could help :D

[Jawor]

Results of tests performed on The X-Files #8ABX03: "Redrum" with HVS - Best Picture:

Average quantizer = 4.194

Quantizers and frame types:
I-frames - 1.28 %
P-frames - 22.81 %
B-frames - 61.81 %
S-frames - 14.10 %
2 - 2.4 %
3 - 24.6 %
4 - 36.5 %
5 - 28.2 %
6 - 6.6 %
7 - 1.2 %
8 - 0.2 %
9 - 0.1 %
>9 - 0.1 %

And here's the screenshot:
http://img8.imageshack.us/img8/5372/hvsbest.png

It seems to me, that these both HVS matrices produced more perceptible mosquito noise and a softer image than my matrix. Is it only a matter of my imperfect sight or do you see it too?

I'm currently testing it with Evil 0.1 matrix by SoonUDie.

[ObiKenobi]

I'll be doing some tests as well on Lupin III - Plot of the Fuma Clan which is a 70 minute ova. Should have some pics posted later this afternoon comparing Jawor's 1CD, Evil .1 and HVS-Best

[SoonUDie]

Jawor: in the shots you've posted so far, what I see is:
1) Your matrix is "rougher", and encodes more noise/"detail"...
2) ...at the price of more blocks, and a
3) loss of definition in some areas (see the red blotches)

Do you think you could post some other screen shots, showing areas with high deatil, and areas with motion, etc.?

At the moment, I'm stuck - in some cases, my matrix looks better than yours, and in others it doesn't. Take a look at the following pairs of shots:

Jawor's
http://students.oxy.edu/tcoldwell/forum_uploads/Doom9/Evil/jawor1.jpg

Evil 0.1
http://students.oxy.edu/tcoldwell/forum_uploads/Doom9/Evil/evil1.jpg

Check out Gandalf's teeth! Evil does a much better job. Also, you can see that Evil has a slightly lower, but more consistent quality across the sleeve. HVS Best also gets it right (though I haven't included the image).

But in the next pair, the tables are turned:

Jawor's
http://students.oxy.edu/tcoldwell/forum_uploads/Doom9/Evil/jawor2.jpg

Evil 0.1
http://students.oxy.edu/tcoldwell/forum_uploads/Doom9/Evil/evil2.jpg

Evil seems to suffer from high quantization here. Non-deblocked, Jawor's has much more blocking than Evil in the red smoke, but deblocking seems to take care of that. HVS Best seems to be on par with Jawor's.

So it looks like I still have some work to do... HVS best may still be the champ for now.

[Jawor]

Do you think you could post some other screen shots, showing areas with high deatil, and areas with motion, etc.?
Yeah, sure... But it can take some days, because I'd have to get more webspace.

[SoonUDie]

Originally posted by Jawor
Yeah, sure... But it can take some days, because I'd have to get more webspace. I can host them. I'll PM you my email.

[Soulhunter]

Originally posted by Jawor
Yeah, sure... But it can take some days, because I'd have to get more webspace. Get some free pic hosting here... (http://www.imageshack.us/) ;)


Bye

[LigH]

Cute - 2 new matrices for my collection... :D

[SoonUDie]

Originally posted by LigH
Cute - 2 new matrices for my collection... :D oi... :rolleyes:

Actually, your collection has been very helpful; I'm learing a lot from playing around. I've been doing some testing, and it seems that most matrices scale *much* worse than I thought they would, so that 1CD matrices suck at HDTV, and high-detail matrices suck at 1CD. As a teaser, here's how things are so far (compared at 1024x576 from HD source, 24fps, 3800kb/s 2-pass, 300 frame sample):

ultimate > bulletproof heavy compression > mb1DV = Maxlt = CCE VLB = andreas VCD ~=KVCD320 ~> h.263 > 4 rooms > evil0.1 > professional ~> mpeg > selur HDR > selur > andreas 99 > 6of9 > 28 days > Semi-insane

I was surprised that Ultimate has come out on top so far, as I've never heard of it before, but there is no denying how amazingly good it is - and how amazingly bad 6of9 and Semi-insane are - in this test.

[LigH]

Indeed, several matrices were made for specific reasons, and may fail in different circumstances. You shall always think of one fact: The lower the factors in a matrix are, the higher must be the per-frame / per-macroblock quantization factor to achieve a similar bitrate. Matrix factors quantize single frequency components, but the MBQF is multiplied to all the matrix factors at once.

[Didée]

SoonUDie:

Dunno for sure what it is, but something must be flawed in your little test. Or you have a rather strange and special source.

You're encoding to approx. 0.27 bits/pixel*frame, and that's not low bitrate, and not really medium anymore.

And you report that all the lumberjacker matrices deliver a better result than all of the high-bitrate matrices - that's confusing.

Quick shot: your sample of 300 frames is rather short, there might have been issues with the rate-control in the 2nd-pass.
Did you force quant=3 in the 1st-pass for the high-bitrate matrices? (I suspect you didn't.)

But perhaps you are right - and I should switch back do DivX 4.01 for encoding - perhaps it looks amazing, in the meantime ;)

- Didée

[Jawor]

Results of tests performed on The X-Files #8ABX03: "Redrum" with the Evil v0.1 matrix by SoonUDie and the MPEG matrix:

Average quantizer for Evil v0.1 = 4.556

Quantizers and frame types for Evil v0.1:
I-frames - 1.28 %
P-frames - 22.98 %
B-frames - 61.88 %
S-frames - 13.87 %
2 - 0.3 %
3 - 23.0 %
4 - 23.7 %
5 - 34.6 %
6 - 14.7 %
7 - 3.2 %
8 - 0.2 %
9 - 0.2 %
>9 - 0.1 %

Average quantizer for MPEG = 4.247

Quantizers and frame types for MPEG:
I-frames - 1.28 %
P-frames - 22.67 %
B-frames - 61.90 %
S-frames - 14.16 %
2 - 1.6 %
3 - 25.6 %
4 - 30.1 %
5 - 34.9 %
6 - 6.4 %
7 - 0.9 %
8 - 0.2 %
9 - 0.2 %
>9 - 0.1 %

And here are the knife screenshots:
http://img8.imageshack.us/img8/6500/evil.png
http://img8.imageshack.us/img8/6057/mpeg.png

I took some more screenshots from this episode. The first series shows the face of an actor named Joe Morton (HVS - Good Picture, HVS - Best Picture, MPEG, Evil v0.1, Jawor's 1CD):
http://img8.imageshack.us/img8/7023/hvsgood1.png
http://img8.imageshack.us/img8/1302/hvsbest1.png
http://img8.imageshack.us/img8/1418/mpeg1.png
http://img8.imageshack.us/img8/266/evil1.png
http://img8.imageshack.us/img8/3239/jawors1.png

The second series comes from a high motion scene:
http://img8.imageshack.us/img8/4616/hvsgood2.png
http://img8.imageshack.us/img8/3483/hvsbest2.png
http://img8.imageshack.us/img8/8527/mpeg2.png
http://img8.imageshack.us/img8/9973/evil2.png
http://img8.imageshack.us/img8/7177/jawors2.png

Again, it seems to me, that my matrix preserves the highest amount of detail in both comparisons above. Tell me, am I right or not?

Currently I'm doing a backup encode of K-PAX (anamorphic, PAL, 2.35:1) with Jawor's 1CD Matrix. The second pass will finish probably tomorrow evening (my system is quite slow), so you'll have to wait until next week to read my report.

@SoonUDie: Thanks for the link.

[SoonUDie]

Jawor,

After comparing your captures in Paint Shop Pro, it looks like your matrix and HVS Best are tied for first place. I would have to see more evidence to determine which one is really better - yours tends to keep more "grain", which is better in some situations, but it also fails in some places where HVS Best doesn't - like on Gandalf's teeth :D

Also, your screen of the man's face with HVS Good is different from the rest in that series. You may want to re-upload the correct frame. :)

[Jawor]

Originally posted by SoonUDie
Also, your screen of the man's face with HVS Good is different from the rest in that series. You may want to re-upload the correct frame. :)
You're right. Shit, I can't do it today. The right screenshot should be uploaded tomorrow ;)

[SoonUDie]

@Didée

My evaluation of different matrices for HD encoding is definately an incomplete, ongoing process, and is prone to error. Please take my evaluations with a grain of salt. File sizes varied by up to 7% (I consider 1% acceptable for "full testing"), so there's a large chance that the matrices with larger file sizes ended up looking better. If we go down to 1% with my previous test, I can confirm relationships between smaller groups of codecs, but not all of them. I hope to fix this in the future, when I have time (mostly by using longer samples). Is there any way to batch the encoding process?

[TheUnforgiven]

Is there any way to batch the encoding process?

in virtualdub[mod] when u save the file tick the chickbox at the bottom this will delay the encoding. repeat that with all encodes and when u want to run the batch job hit f4

its in doom9 guides i think

[ObiKenobi]

Originally posted by SoonUDie
Is there any way to batch the encoding process? [/B]

Yeah virtualdub has had batch encoding as far as I remember, and thats over 5 years. Just click don't run this job now when you save, and then after you save all the passes go to job control and start.

[ObiKenobi]

Sorry in advance for those on dialup for the horrendous load times this will cause.

Okay settings are as follows

QPEL, GMC, MSP=6, VHQ=4 Bframes:2/1.5/1,Chroma Motion,Chroma optimizer.

AVS:

ConvertToYV12()
Trim(515,105060)
Crop(4,0,-4,0)
UnDot()
LanczosResize(640,480)

Used Evil .1, Jawor's 1CD, and HVS-Best, and this is the order the of the pics.

http://img8.imageshack.us/img8/4619/EvilScene01.png
http://img8.imageshack.us/img8/9742/JaworScene01.png
http://img8.imageshack.us/img8/5932/HVSBestScene01.png

http://img8.imageshack.us/img8/6552/EvilScene02.png
http://img8.imageshack.us/img8/3706/JaworScene02.png
http://img8.imageshack.us/img8/6526/HVSBestScene02.png

http://img8.imageshack.us/img8/738/EvilScene03.png
http://img8.imageshack.us/img8/602/JaworScene03.png
http://img8.imageshack.us/img8/3281/HVSBestScene03.png

[ObiKenobi]

http://img10.imageshack.us/img10/1633/EvilScene04.png
http://img10.imageshack.us/img10/7947/JaworScene04.png
http://img10.imageshack.us/img10/1318/HVSBestScene04.png

http://img10.imageshack.us/img10/2386/EvilScene05.png
http://img10.imageshack.us/img10/7969/JaworScene05.png
http://img10.imageshack.us/img10/9176/HVSBestScene05.png

http://img10.imageshack.us/img10/6662/EvilScene06.png
http://img10.imageshack.us/img10/9392/JaworScene06.png
http://img10.imageshack.us/img10/4002/HVSBestScene06.png

http://img10.imageshack.us/img10/6212/EvilScene8.png
http://img10.imageshack.us/img10/7872/JaworScene8.png
http://img10.imageshack.us/img10/6518/HVSBestScene07.png

Overall, my preference in order is HVS-Best, Jawor, and then Evil, but they all perform pretty good overall.

[Soulhunter]

Originally posted by Jawor
@SoonUDie: Thanks for the link.
*cough* Erm... *cough*

[SoonUDie]

@ObiKenobi

Comparing each shot, I get the following rankings:

scene 1
Jawor ~= HVS > evil

scene 2
Jawor ~= HVS > evil

scene 3
HVS > Jawor > evil

scene 4
Jawor > HVS > evil (evil has least blocking, but least detail)

scene 5
evil > HVS > Jawor

scene 6
Jawor ~= HVS > evil

scene 7
Jawor (least blocks) > HVS (most detail, most blocks) > evil


name 1st 2nd 3rd
----------------------------
Jawor 5 1 1
HVS 4 3 0
Evil 1 0 6


I think Jawor's is the winner in this case, but HVS is really close. In most scenes, it tended to have slightly less fine details, slighty more "sharpenss", and slightly less blocks. Evil just sucks (at anime?) :o.

I think I'll do a "re-do" of the evil matrix using a different methodology. Version 0.1 was optimized only for a specific scene (and it kicked ass there), and I was hoping it would do better in general. This time, I'll try averaging the results of the optimization over many different scenes / movies. I would have before, but in general I'm kind of lazy.

*cough* Erm... *cough*I'm sure that was just a typo :devil:

[ObiKenobi]

Well if you do I'll always be happy to run some tests on it, I got about 180 dvds to use :D

[Jawor]

Yeah, that was a typo. Thanks again, Soulhunter.

@ObiKenobi: Actually I never thought of using this matrix for anime, since I designed it for "natural" movies. Thanks for this anime comparison.

The screenshot of Joe Morton's face from HVS - Good Picture encode is OK now.

Those who wait for my report about the K-PAX encode wil have to wait a few days longer, since I had to change the B-VOP settings (it's a normal backup encode, so I don't want to get an undersized file) and now I have to do both passes again. Stats Reader says that I'd get 571 MB after encoding with B-VOPs set to 3/1.50/1.00, so now I set them to 3/1.00/1.00.

Currently I'm encoding my TV capture of The X-Files : "Via Negativa" at constant quantizer of 2 with settings similar to those I chose for "Redrum" (except for VHQ, now I use 4, cause this time it's a backup encode) and the resulting video file will certainly be smaller than 300 MB.

[Selur]

btw. for those who encode animes:
you might also want to check out what happens if you boost b-frame sensivity, set b-frame offset to 0 and ratio to 1.0,.. (oh and set max b-frames to max i-frame distance)

@Jawor: nice matrix

Cu Selur

[atreya2011]

btw. for those who encode animes:
you might also want to check out what happens if you boost b-frame sensivity, set b-frame offset to 0 and ratio to 1.0,.. (oh and set max b-frames to max i-frame distance)


Q1. What value of B Frame Sensitivity ? 5?
Q2. Max B frames = Max I frame distance? Pardon my stupidity but I dont understand that part. I dont remember a Max I frame Distance Option in XviD

Q3. Forgive me if this Off topic but has anyone tried out Soulhunter's Anime Custom Matrix?

Here's the link
Soulhunter's Anime Matrix (http://forum.doom9.org/showthread.php?threadid=72564&perpage=20&highlight=how%20to%20make%20a%20good&pagenumber=4)

[ObiKenobi]

Not yet but I'll set up some passes tonight and test it out. I'll try to post some screenshots either tomorrow or Monday.

[Selur]

@atreya2011:

to Q1: more like 100 ;)
to Q2: ment maximum Iframe interval (Advanced=>Motion=>Other) so, a value like 240 :)

[Jawor]

After backing up two of my recent TV captures (X-Files episodes "Via Negativa" and "El Mundo Gira"), I've decided to use my matrix with all analog captures from now, since I'm really pleased with the results. I won't post any screenshots, since I didn't encode these episodes using other matrices. I've encoded both episodes with a constant quantizer of 2 and VHQ=4, all the other settings were the same as the ones I've used with "Redrum". Resulting filesizes are 195 M ("Via Negativa") and 380 MB ("El Mundo Gira") - no wonder, since the "El Mundo Gira" capture was more noisy than the one of "Via Negativa".

This matrix still needs to be tested with DVD source. The K-PAX encode should be finished until Friday :D .

[Aktan]

Originally posted by Selur
btw. for those who encode animes:
you might also want to check out what happens if you boost b-frame sensivity, set b-frame offset to 0 and ratio to 1.0,.. (oh and set max b-frames to max i-frame distance)



You would sometimes get problems like this: http://forum.doom9.org/showthread.php?s=&threadid=71651

[Jawor]

Guys, can you remember me? I'm the one who started this thread :D

Just kidding. I don't mind this DCTune discussion at all - it's very interesting.

Here's another series of screenshots from "Redrum" (HVS - Good Picture, HVS - Best Picture, MPEG, Evil v0.1, Jawor's 1CD):

http://img10.imageshack.us/img10/2426/hvsgood.png
http://img10.imageshack.us/img10/5372/hvsbest.png
http://img10.imageshack.us/img10/6057/mpeg.png
http://img10.imageshack.us/img10/6500/evil.png
http://img10.imageshack.us/img10/9069/jawors.png

[dimzon]

Originally posted by Jawor

Here's another series of screenshots from "Redrum" (HVS - Good Picture, HVS - Best Picture, MPEG, Evil v0.1, Jawor's 1CD):

Since You have source video can You encode them using DCTune based matrix and post screnshots to compare?

[Jawor]

Yea, sorry about that :D. I've somewhat hijacked the thread in developing my own method for creating matrices.
Oh, I really don't mind.

Also, you can stop using the Evil Matrix now - it's served its purpose as a test, and I think my current method far outstripes it in creating a quality encode.
I don't use it anymore. I've just used it in the "Redrum" comparison.
I would appreciate it if you try my new method :)
I'll let you know if I'd try.

I know that XviD supports (or at least, used to support) changing quantizations like this, but I think it's not mpeg4 compliant.
You're right, it's not :p

[Jawor]

Here is the long-awaited :D report concerning the K-Pax encode.

AviSynth script:
Clip=MPEG2Source("D:\K_PAX\VIDEO_TS\K-Pax.d2v",cpu2="ooooxx")
Clip=Crop(Clip,8,70,-6,-72,align=true)
Clip=Deen(Clip,"c3d",1,6,8)
Clip=UnDot(Clip)
Clip1=Trim(Clip,0,165784)
Clip1=LanczosResize(Clip1,608,256)
Clip2=Trim(Clip,165785,171555)
Clip2=BilinearResize(Clip2,608,256)
Clip3=Trim(Clip,171556,172920)
Clip3=LanczosResize(Clip3,608,256)
Clip=Clip1++Clip2++Clip3
Clip=UnDot(Clip)
Return Clip

I had to use the built-in deringer from mpeg2dec3.dll, because the movie had a lot of visible ringing artifacts (the source DVD is a low-cost edition bundled with a Polish magazine).

XviD Settings:
MSP=6, ChromaME, VHQ=4
Trellis ON, quantizers 2-31
AQ, Qpel, GMC ON
B-frames 3/1.00/1.00
Packed bitstream OFF
"fast" first pass
B-VOP sensitivity 2 (movie), 30 (credits)
Greyscale zone and weight=0.30 for the credits
I left rest of the settings at their default values

Results:
Video filesize 602 MB
Average quantizer = 2.714

quantizers:
2 - 28.1 %
3 - 71.8 %
>8 - 0.1 %

frame types:
I-frames - 0.98 %
P-frames - 17.11 %
B-frames - 63.39 %
S-frames - 18.51 %

And some screenshots (probably they won't tell you much, since tere's nothing to compare :D ):

http://img7.imageshack.us/img7/520/kpax1.png
http://img7.imageshack.us/img7/1073/kpax2.png
http://img7.imageshack.us/img7/980/kpax3.png
http://img7.imageshack.us/img7/6114/kpax4.png
http://img7.imageshack.us/img7/8164/kpax5.png
http://img7.imageshack.us/img7/9400/kpax6.png
http://img7.imageshack.us/img7/3106/kpax7.png

[SoonUDie]

Originally posted by Jawor
Here is the long-awaited :D report concerning the K-Pax encode.Very nice looking! I wish there were comparisons, but with an entire movie I can understand why there aren't :p

[Jawor]

Originally posted by SoonUDie
Very nice looking! I wish there were comparisons, but with an entire movie I can understand why there aren't :p
Well, I'm quite happy with this encode :)

@dimzon: Thanks for testing my matrix.