Auto Gordian Knot: current version 2.55

[len0x]

Originally posted by Corence
The dll file is there. I'm using an AMD Athlon XP 1700+ CPU.

I have Athlon XP as well and didn't encounter this problem.
OS?

[Corence]

Windows XP Pro

SP2

[Sharktooth]

Originally posted by len0x
btw, what is the difference between V2 and V3HR?
V2 has been tuned starting from my (personal) studies on human eye perceivable frequencies while V3HR has been tuned scaling the ReferenceDivx HVS matrix (taken from IEEE studies on human eye visual system) coefficients.
V2 has no inter matrix coefficients below 16 while V3HR starts with 15 (some standalones doesnt like that - for example ESS chipset based standalones).
V2 is also a little bit "softer" than V3HR that means V3HR preserves more fine details at the cost of compressibility (not so much...).

[Taurus]

--------------------------------------------------------------
Originally posted by Corence
The dll file is there. I'm using an AMD Athlon XP 1700+ CPU.
--------------------------------------------------------------
On my machine everythings works fine. Athlon XP 2400, winXP SP2.
Did you install AutoGk Ver. 1.60 before updating to 1.70?
If not, you get the exact error you described.
Tested it on another machine.

@len0x

I found my time to give lanczos4resize/removegrain/EQM V2 a try =AutoGK beta 1.70.
Using an avs script generated by AutoGk and the *.vcf.
Tried it on three different sources. Artificial Colouring: Nemo
Dark Smoky Scenes: Matrix3, Noisy: Buffy 2&5
I didn't count any numbers this time, just my subjective impressions.
Compared it to the "traditional" sheme: Lanczos/undot/HVS Best.
Hard to say, who's the winner. On some scenes it's easy to see, the new combination is ahead, especially on motion scenes.
Sometimes a little rough around the edges. On more motionless scenes I
prefer the old combinations, they washed a little, give a more static view.
But the overall winner: Tada! New combination.
I know it's hard to decide which matrice for this or that purpose,
but I think it's a step in the right direction.
I remember times when I got upset with the somtimes blurry encodes I made with AutoGK. But this is far, far away now ;)
My eyes are hurting and I wish you good luck and a good hand in your decision which to take.
Everyone else should take into account that this is still beta
and could be changed tomorrow.

Rest my tired eyes

Cheers

Taurus

[evil ryu]

Originally posted by Corence
I'm getting the same error as bourtzovlakas. The VirtualDub Error is:

Avisynth open failure:
LoadPlugin: unable to load "C:\PROGRA~1\AutoGK\filters\RemoveGrain.dll"
(F:\Burn\Comedy\agk_temp\Eddie Izzard - Dress to Kill_comptest.avs, line 3)

I hope this helps. I've never had this error using previous versions.

I have the same error but with "...sse2.dll".

I did a fresh installation of version 1.60 and then I have upgraded to 1.70b .

[therealjoeblow]

Originally posted by len0x

Anyway it all comes from the fact that motivation is these projects tend to go away over time. I think I reached the point when AutoGK is almost everything it is supposed to be originally (this is not saying there is not much to do though) and most ppl are happy with it. So one day might wake up with a question: "Do I really need all this? Don't I have something else to do? It's time to move on!". And financial motivation can be the only one that may postpone this morning... And I already see that moment from time to time :)

I hear you - I got to the same point with the Hidden Options tool, and that was just a *very* simple addon to your bigger picture here. For any freeware developer, the application gets to where you wanted it to, life changes or gets busy on other fronts, and demands on the application naturaly grow beyond what you envisioned. That's the point where you either quit or go commercial. And i'm pretty sure noone was going to pay me for Hidden Options.exe ;-)

[therealjoeblow]

Originally posted by Thomas Davie
I've already found Auto GK useful enough that I would purchase it. I also am lazy enough that I won't bother with a donation (I know, a fault on my part). I'm also unable to use Pay Pal. Do you have a physical address where I could send a money order?

thanks

Tom

I second that - My credit card number was stolen last fall. After going through 4 months of sorting out fraudulent credit card charges, making police reports, statuatory declarations, etc, plus the stress of wondering whether I was eventually going to have to cover thousands of dollars in unauthorized charges, I can tell you, I'll never use my credit card online again. Ever. Period. If you had a snail-mail address, I would, however, send you a postal money order. I did so for Alessio Vito with eXtremeMovieManager in Italy.

[len0x]

Originally posted by evil ryu
I have the same error but with "...sse2.dll".


OK, I read up the thread of RemoveGrain and found out that this error is because you don't have "msvcr70.dll" installed in your system (it should be in windows/system32 dir). I'll try to package it with next version of AutoGK, but in the meantime you can just find this dll over the internet and put in your system directory.

[PlazzTT]

I'm trying to find a definitive list of:

1) What resize methods, and
2) What XviD Quant Matrices

AutoGK uses, and the relevent thresholds it uses for deciding which of the above to use and when.

Does anyone have a general idea?

[Rizlaw]

Originally posted by len0x
OK, I read up the thread of RemoveGrain and found out that this error is because you don't have "msvcr70.dll" installed in your system (it should be in windows/system32 dir). I'll try to package it with next version of AutoGK , but in the meantime you can just find this dll over the internet and put in your system directory.

lenOX,

If you package "msvcr70.dll" isn't it possible that you could cause problems for Windows Operating Systems that do have the correct version of this dll installed by overwriting the user's working installed version with the version you provide? I imagine there are different versions for different OS's. I know XP, and to a lesser degree, W2K, are supposed to stop this problem from occuring (installation of the wrong or older dll files.), but wouldn't it be safer to let the user do this - perhaps with a pop-up window from within AutoGK advising the dll is missing and required for proper operation?

[evil ryu]

Originally posted by len0x
OK, I read up the thread of RemoveGrain and found out that this error is because you don't have "msvcr70.dll" installed in your system (it should be in windows/system32 dir). I'll try to package it with next version of AutoGK, but in the meantime you can just find this dll over the internet and put in your system directory.

I have "msvcr71.dll" in WINDOWS\system32\URTTemp and in WINDOWS\Microsoft.NET\Framework\v1.1.4322

[Taurus]

Originally posted by len0x
OK, I read up the thread of RemoveGrain and found out that this error is because you don't have "msvcr70.dll" installed in your system (it should be in windows/system32 dir). I'll try to package it with next version of AutoGK, but in the meantime you can just find this dll over the internet and put in your system directory.
This DLL is applied by the Microsoft® Visual Studio .NET package.
On my machines even some real old versions from the year 2000 doing allright.
It's also part of the Microsoft C Runtime Library.
Maybe its integrated in the net framework, supplied by MS.
Some people don't install net framework by default during their XP setups.
This might explain, why some people are missing this file.
On all of my PC's (3) they exist, but dunno if it's some kind of voodoo:cool:
It should not harm your system if you add it manually, if it's not there, nothing to overwrite = no trouble.
A little google shows up: one version from 29.11.2000, another one fom
05.01.2002.
Hope this shed some light on the issue.

Cheers

Taurus

[len0x]

Originally posted by Sharktooth

V2 is also a little bit "softer" than V3HR that means V3HR preserves more fine details at the cost of compressibility (not so much...).

In my tests V3 has 5-6% less compressibility which is quite a lot in terms of resolution (at least one mod16 step down). Otherwise I would consider it as well (sometimes i think about switching to 3 matrixes in which case things might change).

[len0x]

Originally posted by PlazzTT

1) What resize methods, and


lancsoz/lanczos4/neutral bicubic

Originally posted by PlazzTT

2) What XviD Quant Matrices


its being changed all the time and history reflects that (now its Jawor's/EQM V2)

Originally posted by PlazzTT

and the relevent thresholds it uses for deciding which of the above to use and when.


that is a "know how" of AutoGK - its just tries to get compressibility around 70%...

[len0x]

Originally posted by Rizlaw
If you package "msvcr70.dll" isn't it possible that you could cause problems for Windows Operating Systems that do have the correct version of this dll installed by overwriting the user's working installed version with the version you provide?

First, as Taurus said its not a system dll but MS Visual C++ runtime library - this doesn't differ from OS to OS, but from one Visual Studio service pack/patch to another and should not have any major differences between them.
Second, you think very little of me if you imagine me overwriting this dll by default - I'm not a rookie :)

Anyway - in the next version both dlls will be compiled staticly (I managed to do that from the source), so that no external libraries are needed.

[Rizlaw]

Originally posted by len0x
Second, you think very little of me if you imagine me overwriting this dll by default - I'm not a rookie :)


lenOx,

Actually, I think highly of your programming abilities. And no, I never considered you to be a rookie, and I don't think my post suggested any such thing. My question was posted before Taurus posted his response concerning the nature of the dll file in question.

Continue with the excellent work.:)

[MediaMage]

just confirming:
i have .net framework and sp2 installed,
-->error in RemoveGrain.dll

downloaded msvcr70.dll
(just googled it and use the first link, dll-files.com)
--> working!

[Yo]

Originally posted by len0x
[B]You probably don't realize but I'm starting to think that the idea of "forcing" ppl to donate is only happening because quite a few of them are ready to do so, but don't because its not needed (its quite suprising but I saw many times ppl saying that they would gladly buy AutoGK if they could - amazing fact it is, especially in the light of not so many donations). Also it might be the case that donations are considered to be not worthy - buying is much better justification of money spending.


I have seen some programs that call themselves "freeware", but then force people to "donate", in order to download it. (Such as the program "Fair Use". That seems rather strange to me.

It's perfectly understandable to charge for your work, everyone needs to make a living. But if the user has to pay for the software, it should not be called "freeware" any more, and the money paid should not be called a "donation".

So, if you wish to start charging for the product, it really shouldn't be called "Freeware" any more, and the money paid should be called a "purchase", not a "donation".

I'm not sure how it works with a product like yours, that bundles freeware products with it, that your program is a front end for. If you charged money for Autogk, do the developers of the bundled software (Vdubmod, xvid, etc., etc.) get part of the money earned?

[Sharktooth]

Originally posted by len0x
In my tests V3 has 5-6% less compressibility which is quite a lot in terms of resolution (at least one mod16 step down). Otherwise I would consider it as well (sometimes i think about switching to 3 matrixes in which case things might change).
I could produce some dedicated custom matrices exclusively for AutoGK...
I just need some data to start with :)

Originally posted by len0x
Anyway - in the next version both dlls will be compiled staticly (I managed to do that from the source), so that no external libraries are needed.
That's not strictly necessary.
Maybe you could re-compile only the removegrain source with MS VC++ Toolkit 2003 so the required DLL will be MSVCR71.DLL that's included in the MS .NET Framework.

[Andee]

For ages now when I point AutoGK to VOB files it says that the audio is "Unknown Audio" where I previously would get Dolby mix and stuff like that.

Does anyone know why this is?

[rantingotaku]

hmm.. the versions just keep getting better and better, and faster too, what more could anyone ask for? *does a happy dance*

I do have one problem though..

I finally got off my butt and started encoding my .hack//sign episodes again (after finally getting my delayed vol.6 :mad: ). anyways, after batch encoding the last 3 episodes, I did a quality check on them, and I noticed the overall image quality is less than my original encodes (around autoGK v.1.43 I think).

The edges of characters faces, clothes etc is very jagged with autoGK 1.7, while the originals (1.43) are almost the same quality as the source (NTSC dvd). I tried it on episode 1 as well so I had something to actually compare to frame by frame (1.43 to 1.7) with the same results. :(

I don't know if it's something with xvid, or the new matrix's/filter combo.. I guess I'm throwing this out there for any suggestions.

encoding settings: xvid, 173mb, one audio/sub stream (128 vbr), auto or fixed width of 720 (some eps looked better with auto, some not).

Summary

older version gave better results for this source
is it possible to keep some of latest version and mix with some of old?
if at all possible, can I install 2 versions without conflict? (some sources work better with new, some don't)


sorry for the long-winded post, and take a break Len0x :D

[NMachiavelli]

Originally posted by len0x
You probably don't realize but I'm starting to think that the idea of "forcing" ppl to donate is only happening because quite a few of them are ready to do so, but don't because its not needed (its quite suprising but I saw many times ppl saying that they would gladly buy AutoGK if they could - amazing fact it is, especially in the light of not so many donations). Also it might be the case that donations are considered to be not worthy - buying is much better justification of money spending.

Have you considered a model with a pay-as-you-develop kind of approach? I ran across this recently, which seemed intriguing to me.

The philosophy:
http://hcsoftware.sourceforge.net/jason-rohrer/freeDistribution.html

An example:
http://mute-net.sourceforge.net/index.shtml

It would also seem to address your lack of motivation or uncertainty as to how much more development you want to do with it.

[len0x]

Originally posted by Sharktooth
I could produce some dedicated custom matrices exclusively for AutoGK...
I just need some data to start with :)


Interesting :) How about having HVS-tuned EQMV2 (or V3HR) that is no more that 2-3% percent less compressible than V2? Ideally we want to preserve as much details as possible (within a comressibility range), but only those details that are visible.

Originally posted by Sharktooth
That's not strictly necessary.
Maybe you could re-compile only the removegrain source with MS VC++ Toolkit 2003 so the required DLL will be MSVCR71.DLL that's included in the MS .NET Framework.

Not everyone has .NET framework anyway (but I'm indeed compiling it with VS.NET), so its simpler to use static linkage.

[len0x]

Originally posted by Andee
For ages now when I point AutoGK to VOB files it says that the audio is "Unknown Audio" where I previously would get Dolby mix and stuff like that.Does anyone know why this is?

FAQ Q 5.9

[len0x]

Originally posted by rantingotaku
Summary

older version gave better results for this source
is it possible to keep some of latest version and mix with some of old?
if at all possible, can I install 2 versions without conflict? (some sources work better with new, some don't)



You can install two versions with no problems (you have to create shortcuts to them manually as there is only one AutoGK group) as long as they are using the same versions of Avisynth, XviD, VobSub (coz those are the programs that are shared.) The rest is pretty much contained within AutoGK forder.

Regarding the difference where 1.70 is producing worse results - can we start investigating by comparing two log files from encodings of 1.60 and 1.70? We might be able to see different compressibility results, so that its still possible to adjust 1.70 to get the same results as in 1.60.

[len0x]

Originally posted by NMachiavelli
Have you considered a model with a pay-as-you-develop kind of approach? I ran across this recently, which seemed intriguing to me.
It would also seem to address your lack of motivation or uncertainty as to how much more development you want to do with it.

Great stuff, I have to say! It does look very appealing, although requires a bit of reorganizing of development process as current bugfixes should go out anyway while waiting/developing the next milestone. I can already see one of the milestones - DV/AVI support :)
The only thing to find out is how quick ppl actually start being involved in this (we have to be ready for the project to be stalled a bit).

[Sharktooth]

Originally posted by len0x
Interesting :) How about having HVS-tuned EQMV2 (or V3HR) that is no more that 2-3% percent less compressible than V2? Ideally we want to preserve as much details as possible (within a comressibility range), but only those details that are visible.
That's not a problem. Actually i can parametrically HVS-optimize any matrix...

[len0x]

Originally posted by Sharktooth
That's not a problem. Actually i can parametrically HVS-optimize any matrix...

Biggest problem is testing of them - we should have quite a few ppl to have a go on them to make sure new matrixes are better than old ones. Its always a problem with having something custom :)

[Sharktooth]

Well i can start optimizing matrices from metrics tests (SSIM and PSNR) results with a given compressibility.
Then, when matrices are ready, a blind test could be a good idea... :)

[Exoskeletor]

Autogk will support in the future the creation of mpg files or only avi?

[therealjoeblow]

I did my copy of http://www.imdb.com/title/tt0120815/ at a resolution of 608x352, 3078Mb target size (full DVD minus 700mb for 1 set of extras, and another 700mb for another extra) c/w AC3(5.1) in both AGK 1.60 and 1.70 (AutoGK selected that resolution in automode).

1.60 reports:
-------------
[10/25/2004 2:35:45 AM] Compressibility percentage is: 44.34
[10/25/2004 2:35:45 AM] Using softer matrix.
[10/25/2004 2:35:45 AM] Chosen resolution is: 608x352 ( AR: 1.73 )
[10/25/2004 2:35:45 AM] Predicted comptest value is: 58.35
.....
[10/25/2004 4:40:38 AM] Expected quality of first pass size: 69.37


1.70 reports:
-------------
[10/23/2004 2:41:29 PM] Compressibility percentage is: 44.28
[10/23/2004 2:41:29 PM] Using softer matrix.
[10/23/2004 2:41:29 PM] Chosen resolution is: 608x352 ( AR: 1.73 )
[10/23/2004 2:41:29 PM] Predicted comptest value is: 58.27
.....
[10/23/2004 4:47:08 PM] Expected quality of first pass size: 74.31



I also did forced it to encode @ 640x368, 3078Mb target size with AGK 1.70 to see if there was any visual difference, which reported:

[10/25/2004 8:26:44 AM] Compressibility percentage is: 44.28
[10/25/2004 8:26:44 AM] Using softer matrix.
[10/25/2004 8:26:44 AM] Chosen resolution is: 640x368 ( AR: 1.74 )
[10/25/2004 8:26:44 AM] Predicted comptest value is: 52.95
.....
[10/25/2004 10:43:04 AM] Expected quality of first pass size: 68.20


To visually test, I then queued them all up in ZoomPlayer with ffdshow (Oct 12, 2004) as the decoder (all post processing and other enhancements OFF), and used ZP's "Next Track" command to cycle back-to-back through exactly the same 3~5 second scenes for each of the 3 encodes at about 10 different points in the movie, and compared relative perceptual quality and overall detail.

Both of the v-1.70's look the same to me, I couldn't tell any difference. However, the v-1.60 marginally wins in terms of overall quality - it has slightly more detail in facial closeups (you can marginally see more wrinkles and hair detail, etc). In action scenes, there's really no precievable difference in any of them.

It's really a close tossup, but visually I prefer v-1.60 on this film, just for the marginal increase in quality in static shots.

And, that said, this test was done watching on the computer monitor. Watching them on the 32" Sony HDTV from the couch, you can't really see any difference in any of them. I'll be trying v-1.60 @ 640x368 later tonight just for a complete comparison.

Also for interest, DRF Analyzer reports as follows (but the numbers don't really mean anything if visually one looks better than the others):

v-1.60 @608x352:
----------------
DRF=1&2: 31186 12.9%
DRF=3: 100880 41.8%
DRF=4: 40910 16.9%
DRF=5: 68486 28.4%
DRF=6: 0 0.0%
DRF=7: 0 0.0%
DRF=8: 0 0.0%
DRF=9: 0 0.0%
DRF>9: 0 0.0%
KeyF/DeltaF: 0.96%
KeyDRF<4: 2308
KeyDRF=4: 0
KeyDRF>4: 0

AverageKeyDRF: 2.54
MAXDRF: 5
AverageDRF: 3.59
Deviation: 1.03


v-1.70 @ 608x352:
-----------------
DRF=1&2: 48297 20.0%
DRF=3: 82827 34.3%
DRF=4: 60261 25.0%
DRF=5: 50082 20.7%
DRF=6: 0 0.0%
DRF=7: 0 0.0%
DRF=8: 0 0.0%
DRF=9: 0 0.0%
DRF>9: 0 0.0%
KeyF/DeltaF: 0.95%
KeyDRF<4: 2303
KeyDRF=4: 0
KeyDRF>4: 0

AverageKeyDRF: 2.63
MAXDRF: 5
AverageDRF: 3.44
Deviation: 1.03

v-1.70 @640x368:
----------------
DRF=1&2: 28203 11.7%
DRF=3: 103702 43.0%
DRF=4: 36670 15.2%
DRF=5: 72851 30.2%
DRF=6: 0 0.0%
DRF=7: 0 0.0%
DRF=8: 0 0.0%
DRF=9: 0 0.0%
DRF>9: 0 0.0%
KeyF/DeltaF: 0.97%
KeyDRF<4: 2344
KeyDRF=4: 0
KeyDRF>4: 0

AverageKeyDRF: 2.83
MAXDRF: 5
AverageDRF: 3.62
Deviation: 1.03



Just as a final point of interest, all encodes switched to the softer matrix, even at 1227kbps. I would have thought the sharp matrix would be used at such a high target file size (4.5 CD's).

[len0x]

Originally posted by Sharktooth
Then, when matrices are ready, a blind test could be a good idea... :)

That's doable if we have not more than 5 of them in total (including existing V2&V3). We can organize it in this forum.

[len0x]

Originally posted by Exoskeletor
Autogk will support in the future the creation of mpg files or only avi?

You mean mpeg2? No, I don't see the purpose of it and there are tools like DVD2SVCD already...

[len0x]

Originally posted by therealjoeblow
I did my copy of http://www.imdb.com/title/tt0120815/

Well, first its very tough movie to compress. Its known to be very noisy and not very representative of what kind of movies are out there.

Originally posted by therealjoeblow
Both of the v-1.70's look the same to me, I couldn't tell any difference. However, the v-1.60 marginally wins in terms of overall quality - it has slightly more detail in facial closeups (you can marginally see more wrinkles and hair detail, etc). In action scenes, there's really no precievable difference in any of them.


Second, its very unstandable - since only soft matrix was used (that is the same in 1.60 and 1.70) then the only factor is RemoveGrain(mode=2) vs Undot (or RemoveGrain(mode=1)). Mode 2 obviously removes more details (especially on very noisy sources) and therefore more compressible. Hence larger resolutions are possible. However this was expected - I would be more interested in comparing current sharper matrixes - coz that is what makes real difference (especially in high motion scenes). Also since RemoveGrain mode 2 is artifact free mode then on cleaner sources the difference should not be really visible.

Originally posted by therealjoeblow
And, that said, this test was done watching on the computer monitor. Watching them on the 32" Sony HDTV from the couch, you can't really see any difference in any of them.


That's another point - if you're apply any kind of noise filter then you will always be able to see the difference in details when examining pixel by pixel. Watching it in real life is another story. The whole point of this exercise is to decide - if loss of details can be justified for trying to increase the resolution or not.

P.S. Also I'm not convinced in usefulness of lanczos4 yet...

[Sharktooth]

@lenox: i'm ready for producing matrices for AutoGK.
If i understood you need about the same compressibility of V2. Right?
... EQM AGK-HR is coming

[len0x]

Originally posted by Sharktooth

If i understood you need about the same compressibility of V2. Right?


yes, that's right.

[Taurus]

Originally posted by len0x
P.S. Also I'm not convinced in usefulness of lanczos4 yet... [/B]

Is it just me, or do others have the same visual aspect on lanczos4resize?
It seems to me, as if the edges and lines get something overdrawn.
Sure this will make up for a sharper image, but sometimes it looks quite unnatural. Tested it by up- and down-scaling.
On a TV screen its okay, cos it is blurring anyway.
But on PC Monitor sometimes it looks a little distorted.
Just my 2 cents.
Maybe this is what therealjoeblow means.

[len0x]

I've changed resizer back to normal Lanczos. Now it should be easy to see the effects of new matrix and noise filter. Feedback is much appreciated.

[Sharktooth]

EQM AGK-HR (first beta):

8 10 10 10 11 11 13 15
10 10 10 10 11 12 14 16
10 10 11 11 13 14 16 18
10 10 11 13 15 17 19 23
11 11 13 15 19 22 26 29
11 12 14 17 22 28 34 41
13 14 16 19 26 34 44 55
15 16 18 23 29 41 55 72

16 16 16 16 17 18 21 24
16 16 16 16 17 19 22 25
16 16 17 18 20 22 25 29
16 16 18 21 24 27 31 36
17 17 20 24 30 35 41 47
18 19 22 27 35 44 54 65
21 22 25 31 41 54 71 88
24 25 29 36 47 65 88 115

Direct download link: http://ebola.gamersrevolt.it/len0x/eqm_AGKHR_b1.xcm
Notes: Compressibility is 1-2% lower than EQM V2. HVS optimized.

[len0x]

Originally posted by Sharktooth
EQM AGK-HR (first beta):


great - it would be useful to have it in an export file form.

[Sharktooth]

Just added the direct download link.

[len0x]

cool, I've queued small sample of AutoGK's 1.72 avs with 4 matrixes: HVSBest, EQM V2, EQM V3HR, and AGKHR - will have a look tonight.

[Sharktooth]

Nice, btw EQM AGK-HR looks better than EQM V2... :)

[len0x]

Originally posted by Sharktooth
Nice, btw EQM AGK-HR looks better than EQM V2... :)

any screenshots to prove that ? :)

[Sharktooth]

No AR correction, no resize-
V2(B-Frame):
http://ebola.gamersrevolt.it/len0x/v2.jpg
AGK-HR (B-Frame):
http://ebola.gamersrevolt.it/len0x/agk.jpg

[len0x]

OK, I have to agree out of 4 matrixes AGK-HR looks he best in terms of shapness and block artifacts! Good job!

P.S. Compressibility is not the same as V2 on my source, but rather the same as V3HR.

[Sharktooth]

Uhm... that's impossible coz AGK-HR has higher inter-frames coefficients than V3HR.

However heres the filesizes of my test encodes:
132.136.960 agk.avi
129.826.816 v2.avi
138.418.176 v3.avi

EDIT: if you need more compressibility there's still some room for tweaking

[len0x]

well, I have:

AGK - 39,100,416
V2 - 36,808,704
V3 - 39,133,184
HVSB - 34,459,648

[Sharktooth]

Did you used 1 or 2 max consecutive b-frames?

[len0x]

1/150/100