MSU Subjective Comparison of Modern Video Codecs [Archive]

DmitriyV2

6th February 2006, 14:21

We publish results of

MSU Subjective Comparison of Modern Video Codecs

Goals of our assessment are subjective (visual) comparison of popular videocodecs, comparison of results with objective metrics and subjective assessment technology testing.

2 bitrates (690 kbps and 1024 kbps)
4 tested codecs:
* DivX 6.0
* Xvid 1.1.0
* x264
* WMV 9.0
4 test sequences
50 experts
Testing method SAMVIQ (see description in ITU-T BT.500)

Main comparison parts:

1. Subjective comparison of videocodecs.
2. Comparison of results with results of objective metrics PSNR, VQM and SSIM.

SELECTED MAIN RESULTS:

Following graph illustrates average MOS for each codec and bitrate and its' 95% confidence intervals (MOS is Mean Opinion Score, the higher the better).
http://research.graphicon.ru/images/stories/video/subjective_testing/image001.gif_

On a graph below you can see average MOS for all codecs:
http://www-win.compression.graphicon.ru/video/codec_comparison/images/mos06_mos_eng.png

Main idea:
* Looks like XviD extrely needed for better filtering.
* x264 has the same results on 690kbit as WMV 9.0 on 1024 kbit - really good job!

SETECTED METRICS COMPARISON CHARTS:

We measured a number of objective metrics (PSNR, VQM and SSIM). To evaluate their precision, their values were mapped on the scale of MOS (0-9) using a fitting function.

http://research.graphicon.ru/images/stories/video/subjective_testing/image002.gif_
http://research.graphicon.ru/images/stories/video/subjective_testing/image003.gif_
http://research.graphicon.ru/images/stories/video/subjective_testing/image004.gif_

So it's possible to calculate error in % in measurement of PSNR, VQM and SSIM in comparison with visual results (SSIM has better results, VQM - worst).

See detailes in report (32 pages):
* On report page http://www.compression.ru/video/codec_comparison/subjective_codecs_comparison_en.html
* Or in report PDF (852 kB, 32 pages with all details)
http://www.compression.ru/video/codec_comparison/pdf/msu_subjective_codecs_comparison_en.pdf

Enjoy! :)

arch_angel16

6th February 2006, 15:50

I wonder about the PSNR results. In both bitrates, x264 evaluates higher than its PSNR value, while Xvid evaluates significantly lower than its PSNR value. Also, I don't understand why, for the compressions, the measured PSNR values differed between repetitions.

Finally, though I understand your reasoning in not altering default encoder settings, the xvid beta is at a disadvantage due to it having deblocking disabled by default. I believe this is the factor that caused the seemingly abnormally low MoS score in xvid.

*.mp4 guy

6th February 2006, 16:08

Sagittaire

6th February 2006, 16:12

Well unfortunaly this codec test is totaly useless because:
- you don't test quality codec here but simply Rate Control quality
- Nobody use these codec for make ABR encoding : it's always multipass encoding or 1 pass quality mode
- ABR 1 pass encoding is a bad RC mode : bit repartition is never good and will be never good with this particular RC mode¨
- Post Process PP4 is enable for DivX and WMV9 but not for XviD

Conclusion : your test say simply that "WMV9 PP4 and DivX PP4 ABR mode are really better than XviD PP0 ABR mode". If I make test with VLC player (use internal deconding) then PP for all MPEG4 ASP codec (XviD and DivX) will be identical and conclusion will not the same for XviD and DivX.

If you want make real quality test:
- Use always best possible RC mode with best possible quality setting
- Use always best decoding procedure

Anyway second part "comparison of results with objective metrics and subjective assessment technology testing" is very interessing.

akupenguin

6th February 2006, 18:08

DivX 690 kbps, 1024 kbps
XviD 690 kbps, 1024 kbps
x264 690 kbps, 1024 kbps
WMV 700000 bps, 1048576 bps
In all of those codecs, k=1000. So WMV should get 690000 bps, 1024000 bps.

Sagittaire

6th February 2006, 18:49

In all of those codecs, k=1000. So WMV should get 690000 bps, 1024000 bps.

in fact wmv9 codec don't use the same unit ... 1024 for wmv9 and 1000 for xvid, divx, x264 done the same target size.

bkman

6th February 2006, 19:39

This gets me thinking: Why don't we (doom 9 community) have a "blind" test of various codecs at optimal (2-pass) settings? I mean a blind test just as audio codecs are tested over at hydrogenaudio, Ie. Compress a source, then decompress (with deblocking?) and encode losslessly (ffhuff?), name it something not obvious so that no-one knows precisely which codec was used, and present the samples for download and relative scoring by readers. Then tally the results.

What do you guys think?

bond

6th February 2006, 20:40

interesting but also not really surprising results, as i never had a doubt x264 would beat wmv9 clearly. in fact i am more surprised that xvid did not good against wmv9, cause in my tests wmv9 was not able to beat xvid

i assume this might be caused by two things:
1) the already mentioned useage of 1 pass abr, which xvid surely isnt tuned for (everyone uses it with 2pass)
2) the small sample number nearly only focusing on high motion (3 of 4). its known xvid tries to remove details from high motion scenes as the eye doesnt recognize details there anyways, which surely hurt it in this test setup

therefore my propsal for future tests:
- make a public discussion about the test setup prior to the test
- use 2pass for optimally testing a reallife situation
- use clearly more samples with only a few samples focusing on high motion, as the eye doesnt recognize details in high motion anyways

DmitriyV2

6th February 2006, 21:05

Finally, though I understand your reasoning in not altering default encoder settings, the xvid beta is at a disadvantage due to it having deblocking disabled by default. I believe this is the factor that caused the seemingly abnormally low MoS score in xvid.
Not only:
* We several times write that it's necessary to turn on deblocking in XviD. Common users does not so deep interested in setting and as a result prefer DivX. :) And they are right - it's results are really better on default settings.
* We make some invertigations and situation not only in deblocking (RC, prefiltering and etc).

I hope XviD developers will work in this direction.

bond

6th February 2006, 21:09

DmitriyV2

6th February 2006, 21:20

The reason Xvid Scored so horribly is because its deblocking filter is disabled by default. The people at msu used the default settings for everything in the comparison, so they didn't enable xvids deblocking filter.
* We did not know reasons to use clearly worst setting as defaults in codecs.
* We began to use non-default settings now. In last annual H.264 comparison (http://compression.ru/video/codec_comparison/mpeg-4_avc_h264_2005_en.html) we use settings, received from developers. Some next big our comparisons will have such rules (see FAQ (http://compression.ru/video/codec_comparison/h264_2005_comparison_faq_en.html)).

Revgen

6th February 2006, 21:26

@DmitriyV2

First of all, I understand how hard it is to test codecs, and I appreciate the comparison.

The decoder is just as important as the encoder when it comes to codecs. Anybody who knows any thing about xvid knows that postprocessing is very important. Granted that it's not on by default, it would at least be good to mention the advantages of postprocessing in Xvid so that non-enthusiasts can be informed.

I assume that you're comparison was meant more for businesses and companies and not so much for enthusiasts. If so then at least mentioning the postprocessing advantages for XVID will help these people make a better informed decision.

DmitriyV2

6th February 2006, 21:34

Conclusion : your test say simply that "WMV9 PP4 and DivX PP4 ABR mode are really better than XviD PP0 ABR mode". If I make test with VLC player (use internal deconding) then PP for all MPEG4 ASP codec (XviD and DivX) will be identical and conclusion will not the same for XviD and DivX.

1) One pass comparison with MS WM9 was interesting for us. In next comparisons we will measure also another situations.
2) Looks like you did not catch sight of x264. :)

Anyway second part "comparison of results with objective metrics and subjective assessment technology testing" is very interessing.
Thank you!

DmitriyV2

6th February 2006, 22:11

interesting but also not really surprising results, as i never had a doubt x264 would beat wmv9 clearly. in fact i am more surprised that xvid did not good against wmv9, cause in my tests wmv9 was not able to beat xvid

We show this comparison to some persons during last month reviewing (not only to codecs experts) - and situation with WMV9 and x264 was the most unexpected news for them (most persons believe that WMV9 is better, than DivX and news that x264 on 690 is better than WMV9 on 1024 - also was impressive sometimes).

i assume this might be caused by two things:
1) the already mentioned useage of 1 pass abr, which xvid surely isnt tuned for (everyone uses it with 2pass)
2) the small sample number nearly only focusing on high motion (3 of 4). its known xvid tries to remove details from high motion scenes as the eye doesnt recognize details there anyways, which surely hurt it in this test setup

Secret information - we have research project of one pass RC (based on one pass Xvid optimization), so current situation with codecs was interested for us. We planning to publish results with sources. This is our non-commercial and not so fast fan project, but hope it will help.

therefore my propsal for future tests:
- make a public discussion about the test setup prior to the test
- use 2pass for optimally testing a reallife situation
- use clearly more samples with only a few samples focusing on high motion, as the eye doesnt recognize details in high motion anyways

Thanks for good ideas!

DmitriyV2

6th February 2006, 22:21

this decoder issue is actually the same as i tried to point it out in your h.264 comparison:

1) why dont you use the same decoder for all streams that can be decoded by the same decoder? eg an independant one would be libavcodec

2) why do you enable postprocessing anyways? are you testing the encoders here or the performance of the decoders (postprocessing done by the decoder is definitely not an encoder feature)?
when your test aims at testing the encoders you should use the same decoder for all streams where possible and disable postprocessing for all streams
Our test aims was in testing of popular CoDecs - COders and DECoders. It was impossible to use the same decoder for any pair from x264, WMV9 and DivX. And XviD was used to understand situation with is without special efforts of professional user.

Sagittaire

6th February 2006, 23:04

Secret information - we have research project of one pass RC (based on one pass Xvid optimization), so current situation with codecs was interested for us. We planning to publish results with sources. This is our non-commercial and not so fast fan project, but hope it will help.

why ... ???

1 pass ABR RC isn't and will not never better than multipass or quality mode with same bitrate simply because 1 pass mode can't predict complexity for next frame. At this time all codec can use turbo first pass. Nero (ateme) use ultra turbo mode and time for first pass is less to 10% of total time encoding for example ...

DmitriyV2

6th February 2006, 23:52

why ... ???
1 pass ABR RC isn't and will not never better than multipass or quality mode with same bitrate simply because 1 pass mode can't predict complexity for next frame. At this time all codec can use turbo first pass. Nero (ateme) use ultra turbo mode and time for first pass is less to 10% of total time encoding for example ...
Maybe this sounds strange, but codecs also used not only for ripping and not only on PC. :)

bond

7th February 2006, 01:05

We show this comparison to some persons during last month reviewing (not only to codecs experts) - and situation with WMV9 and x264 was the most unexpected news for them (most persons believe that WMV9 is better, than DivX and news that x264 on 690 is better than WMV9 on 1024 - also was impressive sometimes).well thats obviously the typical believe following the marketing being done by microsoft pushing their format
i have always stated that wmv9 is not able to beat xvid, and therefore isnt better than the best mpeg-4 asp codecs. your results backed this up again via divx (which seems to have been improved a lot recently).
following the typical marketing blabla wmv9 should stumb mpeg-4 asp to the ground but obviously it doesnt

we will see if vc-1 as a format will ever be able to keep up with avc, especially as current implementations dont seem to be able to even clearly beat asp, but thats a different story

Secret information - we have research project of one pass RC (based on one pass Xvid optimization), so current situation with codecs was interested for us. We planning to publish results with sources. This is our non-commercial and not so fast fan project, but hope it will help.interesting, tough i am not sure whether the xvid 1pass was ever tuned in any way, cause basically xvid is mainly used for dvd backups and that means 2pass all the way

Our test aims was in testing of popular CoDecs - COders and DECoders. It was impossible to use the same decoder for any pair from x264, WMV9 and DivX. And XviD was used to understand situation with is without special efforts of professional user.but this also means that you are mixing up decoder and encoder which are two totally different things

i still believe that you should use the same decoder for both xvid and divx (eg ffdshow/libavcodec), but thats also a matter of taste

DmitriyV2

7th February 2006, 01:09

@DmitriyV2
First of all, I understand how hard it is to test codecs, and I appreciate the comparison.
Thank you! :)

The decoder is just as important as the encoder when it comes to codecs. Anybody who knows any thing about xvid knows that postprocessing is very important. Granted that it's not on by default, it would at least be good to mention the advantages of postprocessing in Xvid so that non-enthusiasts can be informed.

I assume that you're comparison was meant more for businesses and companies and not so much for enthusiasts. If so then at least mentioning the postprocessing advantages for XVID will help these people make a better informed decision.
Ok, we can note this on site. BTW - last time this topic was discussed shortly after our MSU MPEG-4 SP/ASP Codecs Comparison (http://www.compression.ru/video/codec_comparison/mpeg-4_en.html) (March 2005). Hope that before our next comparison everything will be ok with XviD default settings.

akupenguin

7th February 2006, 02:05

I'm curious how you managed to get x264 to produce bitrates so far from the target on Matrix sc1. (603 vs 690)
Using the matrix clip you sent after the H.264 comparison (720x416, 239 frames), I ran
cat matrix.yuv matrix.yuv matrix.yuv matrix.yuv matrix.yuv | x264.352 - 720x416 --bitrate 690 -o test.264
and it resulted in 695 kbps for the last copy.

CruNcher

7th February 2006, 02:27

Secret information - we have research project of one pass RC (based on one pass Xvid optimization), so current situation with codecs was interested for us. We planning to publish results with sources. This is our non-commercial and not so fast fan project, but hope it will help.

will be very apreciated for streaming thx :) DmitriyV2 i would be very interested about a comparsion of your and Samsungs Deinterlaceing vs some new aproaches with Tdeint and Nedi2 done here :)

dragongodz

7th February 2006, 04:37

- Nobody use these codec for make ABR encoding : it's always multipass encoding or 1 pass quality mode
false. some people do sometimes use 1 pass for a quick encode where they are not after optimal quality. please dont assume things based on what you do. also its quant mode not quality.

- ABR 1 pass encoding is a bad RC mode : bit repartition is never good and will be never good with this particular RC mode
of course 1 pass ABR wont be as good as 2 pass VBR. that doesnt make it BAD but sub-optimal(meaning not the best).

- Post Process PP4 is enable for DivX and WMV9 but not for XviD
agree but on the same hand couldnt the Xvid guys set atleast partial PP, such as PP2, as default aswell ?

Well unfortunaly this codec test is totaly useless
no other way to say this but ... rubbish. this test is to test 1 pass mode only true but since the codecs give the option of this then why shouldnt they be tested ? thats like saying Xvid has 4 VHQ settings but we will only ever test with VHQ4 since its the best and ignore the others exist.
of course this doesnt reflect on what 2 pass results would give and should not be taken as such. however having tests of as many different settings, including modes, is IMHO a good thing. it gives a more complete image of how and where each codec performs.

Mug Funky

7th February 2006, 05:31

1 pass ABR RC isn't and will not never better than multipass or quality mode

though it's true it wont be better than 2-pass (naturally), it can often be better than constant quant or constant quality modes in some cases.

an example - you have a static scene. encode the I-frame at q1, and the rest at a much higher quant (say linear quant 8). you'll get a much better result than the same GOP with all frames at quant 6, and a similar (if not lower) bitrate. [edit] a practical example of this is image galleries or storyboards - 1-pass VBR with avg bitrate of 1000 kbps is very watchable, where CQ 6 is not so.

1-pass VBR also respects max bitrates where CQ doesn't always do this (though there's a lot of encoders that do). this means 1-pass approaches the quality of 2-pass as the average approaches the max allowed rate (and they both approach CBR). in a limited bandwidth situation (rather than limited filesize), there's no reason not to use 1-pass VBR. it has a place :)

of course, when fitting to filesize and where avg rate is significantly less than max rate (or there is no max rate), then 2-pass is always better.

@ DmitriyV2:

thanks for the results. i'm ambivalent about the postprocessing settings (the argument of testing CoDec's as a whole doesn't really cut it as a large percentage of divx/xvid files actually get played back on standalones), but the objective-vs-subjective results are fascinating. also, using 1-pass, though not "optimal", has value in that it shows where xvid can be improved (come on, guys - no codec is perfect...).

Sagittaire

7th February 2006, 09:27

false. some people do sometimes use 1 pass for a quick encode where they are not after optimal quality. please dont assume things based on what you do. also its quant mode not quality.

- ABR 1 pass it's only for direct streaming ... some people here use XviD for direct streaming ... ???

- If you want make real streaming test mode you must use the same buffer constrain for all codec : I don't think that buffer are XviD, DivX or WMV9 are the same in this test. In fact I don't think that DivX ABR 1 pass is real streaming mode. Default buffer setting for DivX is vbv "home theater profil" ... !!? For streaming 2 pass CBR (with constant buffer) or better 2 pass VBR (with variable buffer) is very better solution and XviD or DivX can't make that. In fact IMO in this test only WM9 (WMV9 codec specification + *wmv container specification) can make streaming with really good control.

- "constant quant mode" is "quality mode" for DivX6. Quality mode (crf) for x264 is not constant quantizer mode.

however having tests of as many different settings, including modes, is IMHO a good thing. it gives a more complete image of how and where each codec performs.

certainely but you must make test with comparable setting:

- 1 pass ABR mode for XviD, DivX or WMV9 are totaly different.
- you cant' compare XviD PP0 and DivX PP4 : compare CCE with 15" cathodic screen and TMPGEnc with TFT 32" screen is not good test too ... ;-)

dragongodz

7th February 2006, 11:19

ABR 1 pass it's only for direct streaming
you have to be kidding. its quite amazing that i can do a 1 pass ABR encode then and produce a file that i can play quite fine not for streaming.
so tell me where exactly you get the idea that 1 pass ABR is for streaming ONLY.

If you want make real streaming test
its NOT a streaming test. its a test of the 1 pass modes.

"constant quant mode" is "quality mode" for DivX6
constant quant does not equel constant quality. different footage at the same quant will vary in quality. i will have to put divx 6 back on and see exactly what it calls it but if it is a constant quant then it most definatly is not a constant quality mode no matter what they label it.
as for x264, well i just did a single pass quantizer test. with ffdshow set to show mean frame quants I and P frames where always what i had set and B frames were always 2 higher. so while not all frame types are the same quant its pretty clear it is a constant quant type and nothing to do with a quality measure.

certainely but you must make test with comparable setting:
of course thats reasonable but on the same hand why shouldnt someone test whats the defaults aswell ? surelly these should be a reasonable tradeoff for quality versus speed. meaning not the absolute best possible but also not the slowest etc. either. it may even help show where the default settings could or should be changed.
as i already said this shouldnt be taken as meaning anything more than the limited scope of itself and more tests with other settings and modes etc would give a fuller picture of what codecs are really capable of.

1 pass ABR mode for XviD, DivX or WMV9 are totaly different.
they are all 1 pass target bitrate so how are the different ? you mean because they use different RC's ? well i can show you mpeg2 encoders that all use different RC's for 1 pass and 2 pass etc. doesnt mean they can not be compared as 1 pass or 2 pass modes.

you cant' compare XviD PP0 and DivX PP4
didnt i already agree with that ? the only extra thing i said was that if people were going to use or test with default settings, applies especially to newbies trying these codecs for the first time, then surely xvid should have atleast PP2 set as default(as in hit defaults button or first time install etc) rather than PP0.

akupenguin

7th February 2006, 17:42

an example - you have a static scene. encode the I-frame at q1, and the rest at a much higher quant (say linear quant 8). you'll get a much better result than the same GOP with all frames at quant 6, and a similar (if not lower) bitrate. [edit] a practical example of this is image galleries or storyboards - 1-pass VBR with avg bitrate of 1000 kbps is very watchable, where CQ 6 is not so.
In a static scene, I-quant=1 + P-quant=8 is the same bitrate as CQ1. If nothing changes in the P-frames, they'll take no bits regardless of quantizer. And the reason CQ6 is unwatchable is because it's much lower bitrate.

Sagittaire

7th February 2006, 17:49

as for x264, well i just did a single pass quantizer test. with ffdshow set to show mean frame quants I and P frames where always what i had set and B frames were always 2 higher. so while not all frame types are the same quant its pretty clear it is a constant quant type and nothing to do with a quality measure.

I make the same error (thx Manao ...) but quality mode for x264 (crf mode) is not constant quant mode. Reference for x264's RC is Pframe (Bframe and Iframe are ratio/offset) and quant for PFrame are not constant in x264 crf quality mode.

they are all 1 pass target bitrate so how are the different ? you mean because they use different RC's ?

Default buffer are simply not the same. with exactly the same codec if I make encoding with long buffer (60000 ms for example) and short buffer (1000 ms for exemple) the result will not the same. You can't compare these ABR 1 pass mode simply baecause objective are not the same ...

well i can show you mpeg2 encoders that all use different RC's for 1 pass and 2 pass etc. doesnt mean they can not be compared as 1 pass or 2 pass modes.

Really ... make comparison with 2 pass VBR unconstrained and 1 pass ABR mode with little buffer (I have perfect sample for that ... lol). Buffer specification are very important for quality ...

MfA

7th February 2006, 17:56

Maybe this sounds strange, but codecs also used not only for ripping and not only on PC. :)So what? Most of these codecs are not suited to online coding (no bounds on compression time). With offline coding you can always do better with multiple passes. Absolute bandwith limitations are no easier to guarantuee in ABR either.

I don't see how this arguement holds up.

stax76

7th February 2006, 21:09

the already mentioned useage of 1 pass abr, which xvid surely isnt tuned for (everyone uses it with 2pass)

interesting, tough i am not sure whether the xvid 1pass was ever tuned in any way, cause basically xvid is mainly used for dvd backups and that means 2pass all the way

IMHO multipass makes hardly sense for DVD backups. There are many that think the same as me and much likely there are much more that are little to nothing educated on this complex topic and only use multipass because they read it in tutorials that were written five years ago were mostly single CD rips were made.

mg262

7th February 2006, 22:09

stax,
IMHO multipass makes hardly sense for DVD backups. There are many that think the same as me and much likely there are much more that are little to nothing educated on this complex topic...
I am really interested in this... when you have a moment, could you elaborate?

stax76

7th February 2006, 22:13

I am really interested in this... when you have a moment, could you elaborate?

http://forum.doom9.org/showthread.php?t=104548

http://forum.doom9.org/showthread.php?t=106949

dragongodz

8th February 2006, 01:38

I make the same error (thx Manao ...) but quality mode for x264 (crf mode) is not constant quant mode.
ok i tried crf mode and indeed the quants vary more around. however they do not vary by great deals. lets just say this is quality based for now anyway.
however with the VFW version, which is the only way all of these can be used the same, it is "single pass quantizer" and just has different set quants for the different types of frames. so basically still a frame type dependant constant quant.
i also installed Divx 6.1 and yes under "quality mode" it has "target quant" which is again just staggered constant quant by frame type. a constant quality mode would actually vary the quants greatly to try and keep frames at a constant quality since different types of footage would require different quants.
so ye call it "quality" all you like but the staggered and constant quants modes are not really quality modes at all since the quality will vary greatly based on the footage/scene.
Reference for x264's RC is Pframe (Bframe and Iframe are ratio/offset)
please dont try and tell me about the RC, i DO know where it comes from and what its based on already.

however this is getting off topic, which is the comparison of 1 pass ABR/VBR modes.

Default buffer are simply not the same. with exactly the same codec if I make encoding with long buffer (60000 ms for example) and short buffer (1000 ms for exemple) the result will not the same. You can't compare these ABR 1 pass mode simply baecause objective are not the same ...
and i can name other things that are not the same either. yes it can be argued that changing all things to be equell would give far different results but the fact is these are not the same technologies. you are basically saying we have an apple and an orange and you want to make the orange as much like the apple as possible so its fairer to compare them. there are more things than just the buffer size that can not be simply or always altered(yes xvid and divx have profiles and wmv can have buffer set in CBR mode and x264 can with CLI i know) so they should never be compared because of that ? hmm you better tell Doom9 his codec comparisons are pointless aswell then.

so really its you that is missing the objective. that is what kind of quality does each codec give for their 1 pass bitrate modes at these bitrates at their default settings. you dont have to agree with how useful that is to you or certainly say where you think it could have been done better but please do not simply go calling a test totally useless because you dont use that mode.

foxyshadis

8th February 2006, 04:51

More info on crf (http://forum.doom9.org/showthread.php?t=101551), particularly omion's great explanation near the end. It's definitely different from constant quant, although it acts as a sort of average quant rate (ending up with nearly the same average quant as cq). It most definitely takes scene complexity into account.

sysKin

8th February 2006, 08:14

Anyone against me activating XviD PP by default?

stax76

8th February 2006, 08:18

A constant quality mode would actually vary the quants greatly to try and keep frames at a constant quality since different types of footage would require different quants.

In depth understanding of codec technology is far beyond my interest which leads to a problem in understanding more complicated topics. I came to the conclusion XviD at quant 2.5 produces constant quality regardless of the source meaning no matter what type of footage it is I got equal quality which is exactly the simplicity I always strive for. Similar like a comp. check would help me to get equal quality, after all q2 is used to determine the compressibility.

Manao

8th February 2006, 11:15

Actually, quantizer 2.5 saturates the perceptual quality, that's why you don't see it vary. But, using Q 2.5, you're doing an overkill : you could have got the same quality on some parts of the movie with a higher quantizer.

Also, if you used Q 4 or more, you would see that quality does vary with a constant quantizer.

dragongodz

8th February 2006, 11:23

Anyone against me activating XviD PP by default?
maybe start a poll in the Xvid section ? i am definatly for PP2 being on as default.

More info on crf, particularly omion's great explanation near the end. It's definitely different from constant quant
i already said it was different. i also said its only available from CLI and the other codecs do not have this mode. thanks for the link though as it clearly says it isnt really constant quality but more a target average quant with quant variance towards visual quality improvement. or to quote from that link
the result is more constant visual quality than in constant QP mode.

I came to the conclusion XviD at quant 2.5 produces constant quality regardless of the source meaning no matter what type of footage it is I got equal quality
at a low quant you would probably 99% of the time be unable to really see any difference true. however that doesnt mean it isnt there. try running a mixxed footage type clip through some sort of comparison(SSIM or such) and notice how the values change for the frames and scenes. its not constant at all though the varience should be smaller with low quants than if you used higher quants. however try it at higher quants and you should start to see the differences for yourself.
since its harder to see at low quants thats probably why its generally called a "quality mode". since most people wont see a great deal of difference then its seems some people take it as being the same thing.

and finally all of that is waaaaay of topic. this thread is about the 1 pass target bitrate tests done. if people really want to keep discussing these things wouldnt it be better to start other threads ?