Xvid min and max quantizers

[hello_hello]

If you're running a single pass encode using a target quantizer, do the minimum and maximum quantizer values have any effect?

Using MeGUI I just ran some encodes of the same video, using a target quantizer of 2.5. For each I ran an encode setting Xvid's minimum and maximum quantizers values to 2 and 2 respectively, then 2 and 3 respectively, then finally 2 and 31. Each encode came out to exactly the same file size and looks exactly the same to me, so I assume the minimum and maximum settings only apply to 2 pass encoding?

Thanks.

[jethro]

Min and max quantizer matter when you set target bitrate not target quantizer. Min/Max is more useful in 2pass mode as it prevents waste of bits (too low quantizer) or bad quality (too high quantizer) when encoder wants to cut bitrate too much. Often custom matrices were designed to have min/max limits.
Also, 2.5 quantizer is 2 and 3 alternating.

[hello_hello]

Cheers.

Has it always been that way?
I ask because I was looking at the XviD settings used by AutoGK and for a single pass encode it sets the min and max quantizers to 2 and 4 respectively, but for a 2 pass encode it sets them to 2 and 3 (They actually change according to the quality, but those are the min and max quantizers it uses for a single pass encode at a 75% quality setting and when the reported quality after a compression test for a 2 pass encode is also 75%). I just assumed it must change the quantizer values when running a single pass encode because they have an effect, but apparently not....

[jethro]

For single pass only, min/max do matter if you set fixed bitrate. But single pass fixed bitrate with limited quantizers does not really make much sense because you are uncertain of encode size. If you want certain quality better use fixed quantizer, h.263 @Q3, for example.

[hello_hello]

Thanks for the info.
I guess after having used AutoGK for so many years and never really thought about Xvid itself too much, I'm just trying to understand how things work and why AutoGK restricts the minimum and maximum quantizers. I made quite few little test encodes using single pass encoding and I think I understand things a little better now, but please correct me if I'm wrong.

When running single pass encodes Xvid seems to use semi fixed quantizers. The target quantizer is always the minimum quantizer, and the maximum quantizer is about twice the target quantizer. So for a Q4 single pass encode all I-frames will be Q4, all B frames Q7 and P frames seem to vary (I think between 3 and 4).
A 2 pass encode using the same file size as the single pass encode will distribute the quality a little differently. The minimum quantizer for I-frames will sometimes be less than four and for B Frames it can be more than 7 (If I'm remembering correctly).

From what I can tell when AutoGK sets the min and max quantizers to 2 and 3 when running a 2 pass encode (at around a 75% reported quality after a compression test) it's reducing the quality variations between I, P and B frames. Looking at some of my AutoGK encodes the minimum I-frame quantizer can still be as low as 2 while the maximum B frame quantizer seems to be 4 (even though the max is set to 3). I guess in theory this would increase the quality, and I guess also increase the file size for a given quality, but whether smaller variations between min and max quantizers improve the perceived quality... well I'm not sure on that one yet.

AutoGK also sets "I-frames tuning" differently to the default values. Instead of the boost and reduction values being 10% and 20% respectively it seems to set them both to 0%. I'm still unclear as to the effect on quality it'd have though.
I'm only looking at what AutoGK does because I always to to achieve around a 75% reported first pass quality when encoding, and I've always been happy with the results. If anything, and I'll have to do some proper testing to make sure it's true, I'd say if anything AutoGK encodes at a given quality do tend to look a little better.... all else being equal.

Thoughts anyone?

[jethro]

You're a bit discovering the wheel again.
In xvid vfw you can emulate AutoGK settings or play with the quantizers, b-frame ratios offsets and matrices yourself to achieve great results.
The most advanced tool for compressing with xvid is located here http://forum.doom9.org/showthread.php?t=119500
See the included Xvid_presets.html

[hello_hello]

You're a bit discovering the wheel again.
In xvid vfw you can emulate AutoGK settings or play with the quantizers, b-frame ratios offsets and matrices yourself to achieve great results.

That's pretty much what I spent several hours doing yesterday, running little encodes with VDubMod, changing Xvid settings, looking at the Xvid stats and comparing the results.
I'm a lot closer to understanding what the various Xvid options do, but I was interested in the opinion of experienced Xvid users at to what they think of the Xvid setting AutoGK uses and the effect on quality.
One thing I don't yet understand is how to use a compression test or first pass to manually calculate the expected quality for a given file size. I'm still reliant on AutoGK for that.

The most advanced tool for compressing with xvid is located here http://forum.doom9.org/showthread.php?t=119500
See the included Xvid_presets.html

Cheers. I'll check it out.

[SilaSurfer]

Hello hello_hello :p

I would strongly suggest that you look into Xvid presets which were created by Teegedeck (Sticky). It helped me to understand Xvid functionality a lot better a couple of years ago.

~SilaSurfer~

[hello_hello]

Hello hello_hello :p

I would strongly suggest that you look into Xvid presets which were created by Teegedeck (Sticky). It helped me to understand Xvid functionality a lot better a couple of years ago.

Thanks, I shall.
And you never know, if they do help me to understand XviD functionality a lot better, then maybe down the track if someone else's asks about the Xvid settings AutoGK uses and for an opinion on them, I might even be able to answer their question. ;)

[SilaSurfer]

I never use AutoGK or Single pass encoding with Xvid. My opinion is that Twopass encoding is superior to Single pass not only because of size prediction and yes you can achieve large filesizes and fantastic quality. You talk about compressibility check that AutoGK performs on your source and how you can simulate that. I think that 75% of compressibility reached is actually where twopass encoding is unnecessary with specific matrix if any is used in AutoGK. What matrices AutoGK uses beside H263 or MPEG quantization in Xvid? You can check that with Mpeg4Modifier on videos that you encoded.

[hello_hello]

I never use AutoGK or Single pass encoding with Xvid. My opinion is that Twopass encoding is superior to Single pass not only because of size prediction and yes you can achieve large filesizes and fantastic quality.

File prediction aside, from what I'm seeing looking at the Xvid stats after an encode, single pass Q2 encoding is as good as it gets. You can't actually improve on a Q2 single pass encode even by running two passes and in fact the only way to achieve the same result when running 2 passes seems to be to pick a larger file size than the one which resulted from the single pass and let Xvid fail trying to achieve it. Then the output will be identical to the Q2 single pass encode.
Of course all that changes as soon as you aim for something less than maximum quality because then a 2 pass encode will distribute the quality differently (higher quantizers for some frames, lower for others etc).

You talk about compressibility check that AutoGK performs on your source and how you can simulate that. I think that 75% of compressibility reached is actually where twopass encoding is unnecessary with specific matrix if any is used in AutoGK. What matrices AutoGK uses beside H263 or MPEG quantization in Xvid? You can check that with Mpeg4Modifier on videos that you encoded.

The way I understand it a 75% quality 2 pass encode is literally a 2 pass encode using a file size which is 75% of what you'd get from a Q2 single pass encode. In other words if a Q2 single pass encode gave you a file size of 1000MB, then a 2 pass encode with a target file size of 750MB would achieve 75% quality. Whether that's completely accurate I'm not sure, but assuming it is I'd also assume it'd be when running a 2 pass encode which uses the same Xvid settings a single pass encode would.
Not that I understand how to obtain the target file size from a compressibility test yet, but I'd assume once you have the information it's just some simple math. For instance if the compressibility test is run over 10% of the video then you'd take the resulting file size, multiply it by ten, and you'd have roughly the file size for a 100% quality encode.

What I don't yet understand is what effect limiting the maximum quantizers (as AutoGK does) would have on those calculations, or whether doing so would be likely to improve the perceived quality. I guess AutoGK must do it for a reason though.

At some stage soon I'll run a Q2 single pass encode, use 75% of the resulting file size for a 2 pass encode and see what quality AutoGK reports after the first pass. I guess if it doesn't report a 75% quality I can assume the Xvid settings it uses increases or decreases the quality compared to the defaults, for a given file size.

AutoGK can use custom matrices, although I don't know anything about them. I've got the standalone compatibility setting enabled which forces it to use the h263 matrix, or if the compression test result is too much over 75% it'll switch to the mpeg matrix to bring it back closer to 75% again.

So when you're running 2 pass encodes, how to you pick the file size you're going to use and determine what the quality will be? Do you run some sort of compression test first or do you just take a guess and hope for the best?

[SilaSurfer]

File prediction aside, from what I'm seeing looking at the Xvid stats after an encode, single pass Q2 encoding is as good as it gets. You can't actually improve on a Q2 single pass encode even by running two passes and in fact the only way to achieve the same result when running 2 passes seems to be to pick a larger file size than the one which resulted from the single pass and let Xvid fail trying to achieve it. Then the output will be identical to the Q2 single pass encode.
Of course all that changes as soon as you aim for something less than maximum quality because then a 2 pass encode will distribute the quality differently (higher quantizers for some frames, lower for others etc).

Single pass encode FQ2 is used for maximum quality. When I sad you can achieve fantastic quality with twopass encoding at specific filesize. How does 6OF9 matrix twopass encode of a particular movie with a lot of high motion scenes reaching 4GB in size sound to you? And doing it at fixed quantizer even bigger but not by much, because above mentioned Twopass encode would reach 90% of quality of FQ3 but still reaching a desired filesize (one would know the filesize of the final output)

some tests:

FixedQ3, Bframe set: 2,1.62,0, VHQ4, 1 minute video 720x304 resolution. Avarage quntizer ~ 3.60

1.Mpeg Standard 7.21 MB

2.6OF9 11.6 MB

3.H263 7.41 MB (a bit strange compared to Mpeg, H263 being actually bigger but that can happen)

You see how CQM Matrix can play a big role in achieveing better quality compared to defaults. I see matrices as sort of additional Xvid tweaking where one would use specific matrix for specific cases. Want more compression - smaller filesize use a low bitrate matrix which is optimized to a achieve maximum quality that you can get out at that high level of compressibility. Want higher quality less compression use high bitrate matrices like 6OF9, EQM V3HR, Heini MR matrix etc.

The way I understand it a 75% quality 2 pass encode is literally a 2 pass encode using a file size which is 75% of what you'd get from a Q2 single pass encode. In other words if a Q2 single pass encode gave you a file size of 1000MB, then a 2 pass encode with a target file size of 750MB would achieve 75% quality. Whether that's completely accurate I'm not sure, but assuming it is I'd also assume it'd be when running a 2 pass encode which uses the same Xvid settings a single pass encode would.
Not that I understand how to obtain the target file size from a compressibility test yet, but I'd assume once you have the information it's just some simple math. For instance if the compressibility test is run over 10% of the video then you'd take the resulting file size, multiply it by ten, and you'd have roughly the file size for a 100% quality encode.

Yes that would be correct but I think that doing a two pass encode where you would limit the quantizers to 2,3 is not needed. In this case it would be better do do a Single pass encode. But know that Mpeg standard Q2 is 6OF9 Q3. Now I would not do a twopass encode where I would use Q2 setting in First pass and then based on the bframe offset limit all quantizers settings to 2,3. It's a waste of time doing a two pass encode.

What I don't yet understand is what effect limiting the maximum quantizers (as AutoGK does) would have on those calculations, or whether doing so would be likely to improve the perceived quality. I guess AutoGK must do it for a reason though.

Xvid defaults should be 2-31. Now limit that to 2-3 what do you get. You force Xvid to use only quants 2,3 and not a range from 2-31. Those are limited to achieve a specific compression. The more quantizers are restricted, less compression is applied. Now as for a quality, picture a frame encoded at Q3.5, next frame encoded at much higher quant like 8.6 this is what you could get at 2-31, you would see the diffrence in quality compared to frame encoded at 3.5 and lets say 4.6, in case where you limited the quants to 3-5. This is why I told you to read the sticky Xvid preset thread created by Teegedeck.



AutoGK can use custom matrices, although I don't know anything about them. I've got the standalone compatibility setting enabled which forces it to use the h263 matrix, or if the compression test result is too much over 75% it'll switch to the mpeg matrix to bring it back closer to 75% again.

So when you're running 2 pass encodes, how to you pick the file size you're going to use and determine what the quality will be? Do you run some sort of compression test first or do you just take a guess and hope for the best?

Are you watching your encodes on standalone, if so you should check which chipset is built in your player.

As far as comp check goes

Read this

http://forum.doom9.org/showpost.php?p=916870&postcount=1

[hello_hello]

Single pass encode FQ2 is used for maximum quality. When I sad you can achieve fantastic quality with twopass encoding at specific filesize. How does 6OF9 matrix twopass encode of a particular movie with a lot of high motion scenes reaching 4GB in size sound to you?

Pointless? :)
I'll have to try it though, just out of curiosity.

You see how CQM Matrix can play a big role in achieveing better quality compared to defaults. I see matrices as sort of additional Xvid tweaking where one would use specific matrix for specific cases. Want more compression - smaller filesize use a low bitrate matrix which is optimized to a achieve maximum quality that you can get out at that high level of compressibility. Want higher quality less compression use high bitrate matrices like 6OF9, EQM V3HR, Heini MR matrix etc.

I understand the principle behind using different matrices, even though I've not played around with them, but for the moment I'm just trying to understand the effects of various Xvid settings for a given matrix. H263 to start with, but I assume the same principles would apply to all of them.

Yes that would be correct but I think that doing a two pass encode where you would limit the quantizers to 2,3 is not needed. In this case it would be better do do a Single pass encode. But know that Mpeg standard Q2 is 6OF9 Q3. Now I would not do a twopass encode where I would use Q2 setting in First pass and then based on the bframe offset limit all quantizers settings to 2,3. It's a waste of time doing a two pass encode.

It wouldn't be a waste of time if the result is an increase in perceived quality for a given file size. I'm just trying to understand if that's the case.

Xvid defaults should be 2-31. Now limit that to 2-3 what do you get. You force Xvid to use only quants 2,3 and not a range from 2-31.

The trouble is, that's not what seems to happen.
For instance when I run a single pass Q2 encode, B frames always use Q4. The same thing happens when running a 2 pass encode while setting the maximum quantizer to 3 (assuming the target file size is the same as the 2 pass encode), B-frames still use Q4. So I assume there's a relationship between I-frame/P-frame and B frame quantizers which can't be reduced by setting a minimum quantizer. I'll have to play around with some more 2 pass encodes using different file sizes but it doesn't seem to be as simple as Xvid obeying the quantizer restrictions unconditionally.
If I remember correctly the same thing happened when limiting the quantizers to values such as 3 and 4. When running a 2 pass encode I'm sure I was sometimes still seeing frames with quantizers less than 3.

Those are limited to achieve a specific compression.

I'll have to try it again myself, but what if I ran a Q2 encode, divided the output file size by 50% and used it as a target file size for a 2 pass encode while setting the quantizer min and max to 2 and 3? I'm pretty sure when I did Xvid seemed to pretty much ignore the min and max values.

The more quantizers are restricted, less compression is applied. Now as for a quality, picture a frame encoded at Q3.5, next frame encoded at much higher quant like 8.6 this is what you could get at 2-31, you would see the diffrence in quality compared to frame encoded at 3.5 and lets say 4.6, in case where you limited the quants to 3-5.

Now that I get.... I did run a couple of encodes with and without restricted quantizers and compared the two encodes frame by frame. There's no doubt when the quantizers were unrestricted there were individual frames which appeared lower in quality (they used higher quantizers) but at the same time the "unrestricted" encode had frames with lower quantizers too.

So for a given file size it's not as simple as "restricting the maximum quantizer reduces the compression and increases the quality", because if the file size is the same then the over-all bitrate is the same which means the difference is in how the quality is distributed throughout the video.
I guess that's what I'm trying to understand.... would AutoGK's 75% quality produce the same file size as a 75% two pass encode using unrestricted quantizers (so the quality is just distributed differently) or would AutoGK's 75% quality result in a larger file size, and would that mean it's 75% quality is actually better? If it's the latter, then I wonder how you'd go about calculating it?

This is why I told you to read the sticky Xvid preset thread created by Teegedeck.

I did.

Are you watching your encodes on standalone, if so you should check which chipset is built in your player.

I never watch my AVI encodes. I encode all my stuff using x264. At the same time though, I also encode to AVI for others in the house to use with their players so I aim for "play on anything" compatibility.
I'm happy to continue using AutoGK as I've always done, but I just started looking at the Xvid settings it uses and wondered "why".

As far as comp check goes
http://forum.doom9.org/showpost.php?p=916870&postcount=1

I couldn't get enc to work. It kept either crashing or giving me errors so I gave up.
I did however try using the compression test in HDConvertToX. I ran it on a file I'd recently encoded using AutoGK and the resulting file size AutoGK produced was 285MB with a reported 75% quality (or close to). According to the HDConvertToX compression test, 100% quality was going to result in a 420MB file which means a 75% quality encode would be 360MB, so at that point I gave up again.

[SilaSurfer]

Maximum quality Fixed Q3 Single pass

Perform a comp check at your desired filesize, % returned

First pass done with FQ3.



* XviD quality range 58%-89% of the FQ3 or comp range 42%-11% of the FQ3
(HQ): MSP=6, VHQ=4, VHQ for b-frames, b-frames, 2/1.62/0, Qpel, SixOfNine, curve-compr. H10, L3, quantizer-restrictions min. 3, max. 5 (4 for I-frames), Overflow Control Strength=0

Maximum quants for P-frames

avg P-Quant 3x1.62+0 = 4.98 rounded up to 5

As for B-frames being higher than I an P frames thats how Xvid operates.

Example.

PBP - you have one B frame which neighbours two P frames. Avg quant of those two P frames is 3.6.

3.6 x 1.62 + 0 = 5.8 rounded up to six so the avarage quant for b-frames in above setting of quantizers would be higher then that of Pframes even if you restrict them to 3-5 you would have b-frames at 3-6, their avarage quant would be 4.2~4,3 something like that and because of this the avarage quant of all frames is not 3 like specified in the first pass but rather something like 3.6

1.62 - B frame quantizer ratio
0 - Bframe offset setting

* XviD quality range 90%-99 quality of FQ3 comp range 10% - 1% of FQ3

(HQ): MSP=4, VHQ=3, VHQ for b-frames, b-frames 1/1.00/1.00, QPel, b-frame sensitivity=-3, SixOfNine, no curve-compression, quantizer-restrictions min. 2, max. 4 (3 for I-frames), Overflow Control Strength=0

In this case to achieve that filesize and quality you have to allow min quant of 2 and restrict max 3 for I frames 4 for P,B frames lower max than in the previous example but also allowing to have min quants 2 to achieve that filesize.

So you see how quants restrictions do play a role in achieving a a specific compressibility level and quality based on that. More tigther min max quants settings less compressibility. Maximum quality is FQ3

Quoting HeadBangeR77 (I couldn't add anything more)

Quantizer distribution gives you imo a better overall impression, because the quality is equal throughout the film/clip. There are no jumps in visual quality, and I have seen many in my life (some scenes look really well, and then suddenly appear blocks by fast motion) - most probably because of not limiting the quantizers.

~SilaSurfer~

[hello_hello]

SilaSurfer,
Thanks for sticking with me on this.

I'm going to run some more test encodes and mess around etc, while reading your last post as many more times as I need to in order to understand it all. In the mean time though....

Quantizer distribution gives you imo a better overall impression, because the quality is equal throughout the film/clip. There are no jumps in visual quality, and I have seen many in my life (some scenes look really well, and then suddenly appear blocks by fast motion) - most probably because of not limiting the quantizers.

To see if I understand this correctly.....
You're saying restricting the quantizers is a good thing because Xvid may otherwise (when encoding to a particular file size) use higher quantizers in places where the drop in quality is noticeable, such as fast motion scenes, and the more you restrict the quantizers, the less compression used and the larger the file size?
Would it also be safe to assume AutoGK's I-frames tuning (0% boost and 0% reduction) is to help achieve the overall "less compression" goal?

Here's where I'm still a little lost.....
Assuming the above is correct, I know a compression test is run using Q2 (same as single pass encoding) and I think I know how the resulting file size/quality is calculated from there when using the default min and max quantizers, but I'm still not sure how it's calculated when restricting the quantizers. Maybe the formula is buried in your last post and I'm just not seeing it yet, but AutoGK changes the min and max quantizer values according to the results of the compression test.... for instance a 75% compression test will have it using 2 and 3, a 65% compression test will have it using 2 and 4 etc. Okay, it's fairly obvious why it does that, it's just the calculations I don't yet understand.

In the case of AutoGK the compression test seems to be always run using a Q2 single pass encode (over 10% of the video) where there doesn't seem to be any control over the min and max quantizers, AutoGK reports the compression test result, runs the first pass using it's restricted quantizers, then reports the final quality after the first pass. I guess what I'm really wanting to know is how to perform the calculations AutoGK does, and how it factors the restricted quantizer values into those calculations. If AutoGK can do it, then there must be a way to do it manually.

Unfortunately I couldn't get Enc to work (I kept getting a "couldn't delete temp\\enc.avi" error message) and I'm still no closer to understanding how the final/actual quality is calculated from the result of the first pass.

[hello_hello]

How does 6OF9 matrix twopass encode of a particular movie with a lot of high motion scenes reaching 4GB in size sound to you? And doing it at fixed quantizer even bigger but not by much, because above mentioned Twopass encode would reach 90% of quality of FQ3 but still reaching a desired filesize (one would know the filesize of the final output)

I just finished playing with the 6OF9 matrix and running some little test encodes. I'm not sure I really like it. I also ran a few other encodes to compare it to, using the h263 matrix and also an anamorphic x264 encode. My little test encodes didn't have a lot of action scenes.

In order of file size:
anamorphic x264 CRF19 (largest), q3 6OF9, q2 h263, 75% AutoGK 2 pass (smallest)
Not a huge difference in file size between q3 6OF9 and q2 h263. The AutoGK 2 pass encodes were much smaller and the anamorphic x264 encodes were much bigger, but to be fair being anamorphic they also had a higher resolution.

In terms of quality, and I'm judging it by which encodes looked the closest to the original vob files:
anamorphic x264 CRF19, q2 h263, q3 6OF9, 75% AutoGK 2 pass.
The anamorphic encodes were easily the closest to the original vob files, q2 h263 and q3 6OF9 looked very close while AutoGK 75% 2 pass probably retained the least detail. Not to say it looked bad at all, I'm referring to a very minor reduction in detail for a 25% reduction in file size.
I really only preferred the q2 h263 encodes over the q3 6OF9 encodes because the latter seemed to have the potential to retain the same amount of detail while sometimes retaining/introducing a little more "noise", and because I didn't restrict the quantizers for any of the encodes (except the AutoGK ones) q3 6OF9 sometimes seemed to have a lower quality than q2 h263 in frames with a lot of movement.
No doubt in that respect, CRF19 x264 leaves Xvid for dead.

Of course all the above was from stopping each encode on identical frames while sitting in front of the monitor and comparing them full screen, then moving on to different frames and comparing them again.
When actually running all the encodes and switching between them as they played, I really couldn't tell which one was which most of the time.

[henryho_hk]

Try my batch script (http://forum.doom9.org/showthread.php?t=119500) for Teegedeck's compressibility test. It defaults at a 3% CT and you can tweak it by changing the first parameter of the "SelectRangeEvery()" call. For example, changing it from 466 to 140 you will have a 10% CT. (I am not sure if AutoGK is using the same CT settings.)

At a compressibility of 58% or above, XviD has comparable or better quality than x264 at a faster speed using the ">58% comp. check (fast)" and ">90% comp. check (fast)" presets which use the 6-Of-9 matrix. For 45~58%, XviD is still not too bad with H.263 @ VHQ=4, Quant=2-31, and better with PSNR-HVS-M if you have XviD 1.3.x. (">45%" presets seem more likely to produce "moving textures"). Under 45%, it's x264's world (even at the same size).

[hello_hello]

henryho_hk,
Trying your batch script is actually on my to do list, in fact I downloaded it a few days ago. Just got to wait till I've got a bit of free time and my brain's alert. The two don't always go together.
Hopefully pretty soon though....

Thanks.

[SilaSurfer]

if you have XviD 1.3.x. (">45%" presets seem more likely to produce "moving textures"). Under 45%, it's x264's world (even at the same size).


@henryho_hk

try using owerflow settings 0/8/4, I've done some testing with low motion - almost static scenes - short encodes and if you are reffering moving textures as floating walls - flat areas then the above settings somewhat reduced that effect and encodes stayed in desired filesizes. Avarage Q in second pass was 3.3 with Teegedeck 90% with the above Owerflow settings and Teegedeck 90% default was 4. But this was tested only on short clips and not whole movies yet.

[henryho_hk]

Another possible reason is that I do not use QPEL at all because it crashes rather frequently.

[SilaSurfer]

You mean Qpel is broken in 1.3.x versions? I use it for both passes but using VirtualDub as encoding tool. On short samples I never had a crash. I'm going to do some regular full movie backups soon, thats why I'm asking.

[henryho_hk]

http://forum.doom9.org/showthread.php?p=1491592#post1491592