DivX 5.05 - Multipass or Two Pass?

[BluDChyLD]

Hi,
I've done a few tests using nth pass encoding with 3 passes. This is alot more time consuming and compared 2 nth pass (first, nth) there is no noticeable differance....

Has anyone test the quality differance between the "original" 2 pass and an nth pass at 2 passes? Which provides the best quality? I would test myself but I haven't got much time at the mo...
Cheers

[bkam]

Will run tests tomorrow on extremely uncompressible TV caps. I don't know if it's my card or my cables or my provider but for some reason BilinearResize(480,yyy).msmooth.cnr2.fluxsmooth(12,12).c3d(vhsBQ).dup (no joke) only lands me about 55% compressibility for 25 minutes of cartoon at 1800 bitrate (350mb including 2ch alt-preset standard mp3). Surprisingly the details aren't too obliterated and I get like 10 fps, on this laptop I would get like -50! Anyway, sort of OT, but I will try out the multipass with less filtering on this source. I figure if the encoding is at 50fps with no filtering I can do five passes instead of 2 passes with that much filtering.

[Acaila]

I haven't really tested them side by side, but I doubt there's even supposed to be a difference between original 2-pass and multipass at 2 passes.

[crusty]

I think they introduced multipass primarily too make use of the EKG-tool. I mean, the difference in quality beyond 4 passes is negligable.

But with the EKG-tool you can adapt your settings to visual quality between every pass. This was offcourse not possible with the old 2-pass.

[binky the stunt cat]

Personaly i'm using 3 passes. It does seem to give a better result, but that could also be due to the fact that Psychovisual enhancements have been turned on recently in my method, and have been fixed from the last time i tried them.

[Babeinternational]

I think the number of passes needed for a file, as well as its compressibility, depends on the file itself (complexity, motion...)
So I usually do more passes on a file with a high compressibility test number.
As for when in the sequence to use EKG, I think it's good to do more than 2 passes before to be sure the artefacts left couldn't be suppressed by more passes. So you avoid messing with quantizers for nothing.

[crusty]

So you avoid messing with quantizers for nothing.
it's not dangerous to mess with your quantizers....it's a good feature actually.

[Babeinternational]

I didn't say it wasn't a good feature to be able to change quantizers (this is why EKG was invented), but it is indeed dangerous (less in EKG than in the encoder settings I admit, but still, those bits you allow somewhere are taken elsewhere). That's why EKG only allows relatively small (compared to min-max quantizers) quantizer changes.
So if you can prevent using EKG by doing one more pass, then if time is not an issue I would recommend it.

[bond]

i encoded a small test clip (matrix - same as in doom9s test) in original 2-pass and in multipass (3th - 5th pass)

i went for a 1cd bitrate (650 kb/s)
i compared the clips with avscompare (great tool!!!)

settings:
divx505: psy normal, b-frames, gmc, qpel; bitrate modulation: 0
avs: neutralbicubic (no additional filtering)

filesizes:
original_2-pass: 23.484.416 Bytes
n-pass_2-pass: 23.474.176 Bytes
n-pass_3-pass: 23.476.224 Bytes
n-pass_4-pass: 23.474.176 Bytes
n-pass_5-pass: 23.474.176 Bytes (2-, 4- and 5-pass are not the same files!)

1) original_2-pass doesnt create the same file as n-pass_2-pass!

2) n-pass/original_2-pass vs. n-pass_3-pass:
i observed that the colors in n-pass_3-pass or more looked much more colorful and nicer whereas in both 2-passes they were a little bit pale
3) nth-pass_3-pass vs. nth-pass_5-pass:
in my opinion there are as good as no improvements with more than 3 passes! some scenes doesnt even change a little bit any more in more than 2 passes (the more passes you use, the less things are going to be changed between the passes)...

4) Comparison:
i compared original_2-pass, n-pass_2-pass and n-pass_3-pass in 5 different situations similar to the ones used in doom9s codec comparison:

[i]a) Morpheus sitting in front of agent smith
original_2-pass: shows the least blocks in the walls
n-pass_2-pass: shows a little bit more details in morpheus, nothing what you can see if you watch the clip, ugly walls
n-pass_3-pass: exactly the same as n-pass_2-pass

b) Neo using his mobilephone
original_2-pass: wow, real sharp image as known from xvid, especially in areas like his chin, eyes and nose
n-pass_2-pass: much less details than compared to original_2-pass! the right cheek also shows much more blocks
n-pass_3-pass: helps a little bit with the sharpness and blocks but not comparable to original_2-pass

c) Neo standing in front of the weapon shelf talking to trinity
original_2-pass: least sharp and many blocks especially on the rightern part of his face
n-pass_2-pass: middle
n-pass_3-pass: most details especially in neo's face and his left shoulder. less blocks

d) agent smith waiting for the other agents to leave the room
original_2-pass: more details and less blocks in the walls than the other clips but altogehter there were not many differences
n-pass_2-pass: worst
n-pass_3-pass: improves the 2-pass a little bit but still more blocks and less details than original_2-pass

e) lobby scene - many spalls flying around
original_2-pass: clearly worst, many blocks (if you compare the still pictures, not many differences can be seen if you watch the clip)
n-pass_2-pass: best
n-pass_3-pass: surprisingly shows less details than n-pass_2-pass

5) Conclusion:
imho the thing you have to decide is if you want to use original_2-pass or n-pass_3-pass!
if there wasnt c) i would be sure to use original_2-pass...
so decide for yourself :D

although i think that people with few time and/or weak hardware should use original_2-pass

and i would advise divxnetworks to improve their n-pass_2-pass to the level of their original_2-pass! weird that they arent the same already...

[temporance]

How did you set bitrate modulation when doing nth pass? If you set it slightly to the right, the encoder will behave more like original two pass and give priority to low-motion scenes. I'm sure you would be able to find a setting for the modulation slider in which nth pass beats original 2nd pass. Try 0.15.

[bond]

interesting, but why did divxnetworks change the default settings between original 2-pass and nth pass?

[CaptainCarrot]

I did some logfile-comparisons between 1st, 2nd, 3rd, 4th, and 5th-pass logfiles and already posted the results in another post (http://forum.doom9.org/showthread.php?s=&threadid=52160).
The results were that beyond 3rd pass there won't be many changes.
Btw. It's quite normal that some scenes look worse after another pass, since the codec tries to get the best average quality, so if it thinks that a scene looks below average, it puts some more bits into it, but has to take them from somewhere else. This can result in quite funny behavior, i.e. when I did the test, there were a couple of scene pairs where the quality was changing back and forth from pass to pass (like 2nd pass: 1st scene good quality, 2nd scene bad. 3rd pass: 1st scene bad quality, 2nd scene good. 4th pass: like 2nd. 5th pass: like third ...).

[EagleDM]

I found out EXACTLY the same thing !!!!

pretty irritating if you ask me...

Trying to encode matrix in 2 cd rip with a 78% compresibility and some scenes like (Morpheous being captured and the fire-alarms went on and rain all over the place) and this particular scene looks TERRIBLE WRONG ! not even mention the walls, that got a blocky effect that IS NOT THERE with the old 2-pass method..

Well, I don't know, after doing a Multipass 2-pass, 3-pass and even 5-pass !! I have to say that I STILL keep with the old 2-pass method.


Too bad, If I'm doing something wrong, someone please tell me, I'm using GordianKnot 0.28.2 and I still got better quality from the old 2-pass method.

[temporance]

@EagleDM,

What settings did you use?
Resolution, bitrate, max bitrate and modulation?

In Home Theater profile, DivX 5.0.5 enforces a maximum bitrate of 4000kbps - even this might not be enough for the fast scenes in The Matrix.

[OvERaCiD23]

Originally posted by EagleDM
Well, I don't know, after doing a Multipass 2-pass, 3-pass and even 5-pass !! I have to say that I STILL keep with the old 2-pass method.

Too bad, If I'm doing something wrong, someone please tell me, I'm using GordianKnot 0.28.2 and I still got better quality from the old 2-pass method.

I have the exact same results with movies I've done lately. Original 2-pass produces noticeably better results than any amount of passes using multi-pass. I upgraded to 5.0.5 to try multi-pass again (after being discourged by 5.0.3), and it still sucks. I'm still using 5.0.5 for original 2-pass because I see no change between original 2-pass from 5.0.2. My encodes were on a variety of movies, from 'Goodfellas' to 'Sleepers' to 'Catch Me If You Can'. Multi-pass sucked in all 3, and I did original 2-passes and was happy again (2 CD rips with 5.1 AC3).

[EagleDM]

704x288 with 1218 kbps (notice that it is high enough for a compresibility rating of 89% !!! )

the bitrate modulation is at the center and the profile were de-activated (no enforced bitrate max) using B-frames and Psy enh set on Ligh (this settings are OPTIMAL for the old 2-pass method and gives me a LOT better image quality on the old-2pass, but are a NO GO for this multipass method).

[temporance]

Originally posted by EagleDM
the bitrate modulation is at the center and the profile were de-activated (no enforced bitrate max)
So the max bitrate was 12180 kbps, right (10 * 1218kbps)? Sometimes the GUI sets max bitrate same as target bitrate - this makes very bad quality.

[EagleDM]

yes it was 12180 kbps which is "i think" sufficient.

not, I double-checked, the max bitrate was at 12180, also, take note that the compresibility test was done with 15% of the movie ! and gave me 89%, there is no excuse for the codec to behave like this.. with this kind of bitrate and compresibility the image should be top-notch, and is not the case.
Anyway, the old 2-pass method gave me DVD-Quality image at these settings, it is soo much different that I am really considering the whole multipass method a complete waste of time. but, again, this is my experience, all people should try and find they own findings...

[jeremymacmull]

hey there,

Ive encoded The Bodyguard with Gordian Knot .28.2 4 times now

all 1185 k and 5.1 ac3

here are my findings these are my opinion so dont take it the wrong way

ALL methods had same settings these were

640X352
neutral bicubic , no phy , no b frames , no gmc , no qpel credits done at 20

bit rate was 1185 which gave me a compressibility check of 78% for all

1) Multipass: 3 passes noticed blocks appearing in many scenes not very sharp would not be happy with this encode.

2) Original 2 pass method noticibaly less blocks much higher overall quality. Not very sharp tho in some scenes Would be happier with this encode

3) Multipass 6 passes (i did it to the limit of gknot to test) overall good quality few blocks appearing quite sharp would be happy with this encode : HOWEVER it took damn 16 hours to encode (p3 850)

4) Multipass 5 passes not much diffrerent to the 6 passes method could not really notice much difference but possibly 6 passes method picture was sharper.

SINCE this ive done it again with the same settings except i used phy settings on normal

5) Original Two Pass encode very good quality would be happiest with this encode the phy settings make quite a difference to the overall quality of the film it seems. This is the final version i kept

So in conclusion i found (and this might only hold true to this movie and this movie alone) if you are going to do 3-4 passes dont bother better off and cosiderably faster doing original 2 pass method. however if you have the time and inclination to do 5-6 passes probably better overall quality.

Me personally for such a small difference in quality and such an increase in the time needed to get it (2passes vs 6!!) i think for now ill stick with original 2 pass method for nearly all my encodes

The exception being those with low compressibility checks and 1 CD encodes. where ill go for 6 passes.

another point to note the phy options work so far very very well with 5.05 considerably better than 5.02 and should be used on light for most encodes possibly normal for 1 CD encodes they seem to make a big diff.

B frames only use if u got 40-60% compressibility i find

i never use GMC or Qpel i find they introduce artifacts.( the exception being i use gmc and b frames for the credits and turn phy to strong for credits)(seems to lower size a bit and get rid of high quant errors for credits)(as i like to make the max possible bit rate available for the film while still having reasonable credits)

hope this helps in the discussion over 2 pass or multipass and how many etc

JEREMY

[BluDChyLD]

Thats wierd, I would of thought extra passes would definetley increase quality... I did a few brief comparisons and the 3nth pass option was a little better. I think it distributes more bitrate to t
high motion scenes, hence the blocks in still and slow motion clips... Ah well, I'm not bothered as I watch my rips on tv out :) I'll probably got back to the old 2 pass method - rips are taking 8 hours instead of 5, it's getting a little tedious...

[CdinT]

I found one certain gain from the use of 3-pass.
The quality seems to be the same as in 2-pass BUT with 3-pass I can always reach the exact desired movie size. Even with a comp. test at 95%!!! That was not possible with 2-pass.

[dTb]

It's interesting that a lot of you have found 2-pass to be better. I don't really have the time for comparisons but the quality I get with 5.0.5 multipass is definately better than what I was getting with 2-pass in 5.0.2. There's a possibility some of the improvements could be due to refinements in my methods but I think the codec has played a part.
I read a post by Gej stating that constant bitrate modulation actually favours low motion slightly and you might find your results better overall with a small bias towards high motion (-0.01 to -0.02). I can't remember now however how this compares to 5.0.2.
I never use post-processing on playback of quality vids and what I've noticed with extra passes is a reduction in macroblocks, for me this is enough to warrant a few extra passes (lucky I have a fast system I guess).
I've also noticed that quant distribution still changes fairly significantly even after 4 or 5 passes.

[Matthaeus]

Hi everyone!

Here's a trick how to use Nandub-like high quality two-pass encode with DivX5.
1. Disable profiles to get old features (in Divx 5.0.5)
2. Select Quality based mode with Quantizer 2 (97%)
3. Check 'Write log file' and do a (first) pass
4. Select Multipass n-th pass and Set Bitrate modulation to 0.20
5. Do a (second) pass and be happy :)

[cjaar]

@Matthaeus

what about the file size, do u hit the right file size????

cjaar

[Matthaeus]

I have hit exactly the right file size :)
But why don't you give it a try?

[dTb]

I've never heard of this method before Mattheus, I might give it a try. The one thing I'm not sure of though is the bitrate modulation at 0.20, this is pretty heavily biased towards low motion, is this just what you prefer or does it somehow act differently when using this method? Generally I use -0.02 when doing conventional multipass.

[Matthaeus]

No!!!!!!!!! Trust me! Use positive values! Since the 1st pass is quality based (not "filesize"-based, cbr) you get bigger frames at high motion scenes and smaller frames at low motion scenes. You must regroup the bits to the low-motion frames. With a standard 1st pass (CBR) you can get undersized high motion frames but with a quality based 1st pass you can't! (every frame gets the same quantizer of 2)
If you are using negative values with this method you can get quant < 2 on high-motion scenes. It's a waste of bits!
The other reason is to use higher values that the bitrate is modulated only once (since you have two passes), and there can be high differences between low and high motion frames(because of the constant quant 2 first pass).

Use at least + 0.10 bitrate modulation with this method !

[temporance]

Thanks for posting that Matthaeus - you are right :)

I tried a couple of trailers with positive modulation settings, three passes. They look better for the bitrate than any rips I've done recently.

You should post your results over on forums.divx.com too.

On whole movies though I find that divx505 spends a lot of bits on the credits, especially with positive modulation. It's doing the right thing, making all parts of the movie look good, but making credits look good takes about 15% of the bits for the whole movie!!

One way to fix this is to encode credits first at q=20, subtract the size of this encoding from the size for the whole movie and then use nth pass (2 or 3 passes) for the rest of the movie. Then join movie to credits.

Of course with EKG we should be able to take bits away from the credits without cutting and joining the movie but I haven't worked out how to do this yet. Anyone know how this can be done?

[crusty]

@Temporance, try the EKG howto at www.divx.com/support/divx/guide.php...it tells you exactly what to do.
Read it two times through before even touching EKG and then read it a third time while doing the EKG. You'll get it.
@Matthaeus, I'll give it a try soon.

[windsor]

Anyone else who tried Matthaeus' method enjoy the quality more than original 2-pass? (I personally haven't tried it yet.)

[windsor]

After doing comparison tests, I have to conclude that original 2-pass still offers the best quality output.

[magomago]

hrmm...I do six pass but all i encode are animes and I have noticed better quality compared to original 2 pass....especially in darker scenes

[SeeMoreDigital]

Eeee gods, not this old chestnut again!

In my experience, there really is no benifit to go beyond 3 passes. Especially if the source is a 2.35:1 movie (there are around 311'040 active pixels for a PAL source and 264'960 active pixels for a NTSC source).

Animes may produce different results I really have not tested any.

Matthaeus looks to have cracked it though!

[StoneRoses]

I think Matthaeus' method is very similar to RoBa's method for CCE encoding (MPEG-2)

http://www.doom9.org/mpg/cce-roba.htm

[temporance]

Continuing discussion from this (http://forum.doom9.org/showthread.php?s=&postid=347389#post347389) thread....

Originally posted by bond
yes, because it shows more details in still scenes and i dont have the time to make three passes (and i think that 2 passes should be enough for a good codec to encode a movie normally)

but why do you think n-th pass is better?
I usually only use two passes.

With a good positive modulation, I don't find any problems with the nth pass still scenes. Kauehi was a big improvement, to my eyes. Ok, it's hard to say that it's actually better than original 2-pass, but it's certainly no worse, IMHO.

[bond]

hm, i really have to try bm 0.25 now (never thought about using the max. value :eek: )

i like to have a sharp picture (that's why i use qpel too) in low motion scenes (i tend to smooth fast motion with the qmf filter as i dont see the details there anyway) and original 2-pass is a good way to reach that goal imho (until now ;) )
so if i can reach the same results with n-pass bm 0.25 as with original 2-pass then i have nothing to say against n-pass anymore :D

[aaronp]

Matthaeus (or anyone else who's in the know), would you still recommend using "Quality Based - 1st pass" when doing more than 2 passes?

[bond]

aaronp
plz read more about the codec basics!

quality-based is 1-pass, not 1st-pass!
thats a big difference, quality based is for encoding a file in only one pass not two or more...

[aaronp]

bond, did you see the Matthaeus' method above? Or are you saying that his method wouldn't work at all? He recommended using Quality Based mode for the first pass, to prevent getting undersized high motion frames. If you write a log file, I don't see why you couldn't use quality based for more than one pass.

[bond]

no, i didnt read it, sorry my fault

[Twilight_Fandom]

hmm..here it looks like there is a way to go the "old" way of 2 pass if im right.
Are they different,old 2-pass and new multipass-2pass?
I thought it was the same.

How can u use the old way with divx 5.0.5 and gknot 0.28.5?
thanx in advance

[BluDChyLD]

Originally posted by Twilight_Fandom
hmm..here it looks like there is a way to go the "old" way of 2 pass if im right.
Are they different,old 2-pass and new multipass-2pass?
I thought it was the same.

How can u use the old way with divx 5.0.5 and gknot 0.28.5?
thanx in advance

Hey,
Adjust the DivX 5 codec paramters by clicking the first pass and nth pass buttons in the add job dialogue - then change "mulipass 1st pass" to "original 2 pass 1st pass" and "nth pass" to "original 2 pass 2nd pass". I usually encode in bulk and the nth pass doesn't seem to warrant any quality increase over the old 2 pass mode but it takes a lot longer!