DivX Plasma Codec: Alpha release

[LordRPI]

This is the moment that a lot of us have been waiting for at DivX. In the past few weeks, we've been doing encodes with an alpha version of a new DivX codec - and we thought that you all should have the chance too.

DivX Plasma (tritium) offers improved compression over DivX 5.2.1 by 8-34% depending on the source and bitrate. It also boasts a few more under the hood improvements while still being fully compliant with current DivX Certified devices. Read more and get your own copy! (http://labs.divx.com/archives/000035.html)

[bond]

cool stuff seeing a new divx5 testing version :)

trying to point out the plain changes, i think from the technical point of view this means that divx plasma (the upcoming 5.3?) adds to divx 5.2.1:
- 4MV
- RDO for h.263 quants (but not for mpeg quants i assume)
- motion estimation speed ups on the encoder side
- new deblocking on the decoder side
- the rest are gui changes

did i oversee something?
edit: maybe/hopefully an enhanced gmc or qpel (eg working with slower modes?) :)

[DigitAl56K]

We're tracking known-issues here:
http://forums.divx.com/viewtopic.php?topic=62463&forum=23

Please remember that even if you post here on Doom9 it will help us enormously if you can post bug reports on the DivX Beta Release Feedback forum:
http://forums.divx.com/viewforum.php?forum=23

[DigitAl56K]

bond: There are a few GUI bugs right now. Sometimes options which should be enabled get disabled, sometimes options which should be disabled are shown as enabled.

We're working on all of these but we wanted to give you something to play with in the meantime :)

[akupenguin]

Rate-distortion optimized H.263 quantization Does that differ from trellis (as implemented in lavc or XviD)? Or is it a completey different feature that just sounds the same?

[Sharktooth]

Bugs apart i have only 1 word: WOW!

[obieobieobie]

Very cool. Gonna try it. Insane mode sounds cool. ;)

[LordRPI]

Originally posted by Sharktooth
Bugs apart i have only 1 word: WOW!

Hehe, thanks :D

[SeeMoreDigital]

Thanks guys,

This GUI is much, much better: -

http://img110.exs.cx/img110/1018/DivXPlasma_GUI.gif

Simple, easy to use, and to the point... what a relief :)

But why no proper "anamorphic" encoding yet? Is it on DXN's agenda at all?


Cheers

[nicco]

first impression on the fly...

- the new GUI is very cool (some bugs of course)
- I can not see such a high improvements in quality/filesize even in insane mode, but I did only a little test last night (so I'm not 100% sure).
- Picture semms to be a little smoother than 5.2.1
- I can not acess new plasma postprocessing

And now..let's go testing!!!

[WorBry]

Just tried DivX Plasma out with some PAL DV test clips, using the following path:

AVS Script:
avisource("E:\test file.avi")
ConvertToYUY2(interlaced=true)
KernelDeint(order=0,sharp=true)
LanczosResize(720,528)
Undot()

VirtualDub:
Fast Recompress
DivX 5.2.1Pro settings - Home Theatre Profile, Standard, Two-pass, 2000kbps, key frame 250, single B-frames, no GMC or QPel

Tried the same settings with DivX Plasma. Regardless of the quality level (Balanced through to Insane) it just kept defaulting back to 780kbps and the encoded DivX file was blurred and mosaic, and the actual bitrate (VDub File Info) was 528kbps. Repeated using the same settings in the "unconstrained" profile and tried uninstalling DivX Plasma, resetting DivX5.2.1 to default and then reinstalling, but all with the same result.

Obviously, something is amiss. To my relief however, DivX 5.2.1 functionality was completely restored after uninstalling Plasma.

[stephanV]

turn off the feedback window under the advanced button (4th screenshot of SeeMoreDigital) and the bitrate modes work again...

(thx to Sagittaire :) )

[IgorC]

There´s a bug. VirtualDub 1.5.10 ( of GK) closes when i try to open first window of codec.

Maybe i have a litle older version Divx 5.2.1 (build b1328 WafleDay Sep 3 2004)

Now i dowloading a last Divx 5.2.1 frow official site

P.S. I cheked from divx feedback site. The problem is in Mode insane + Qp.

[nicco]

Test:
DV-Avi PAL
AVS:

tomsmocomp(0,3,1)
bicubicresize(640,464)
undot()
spatialsoften(2,2,4)
converttoyv12()

decoder:ffdshow postprocessing enabled

5.2.1
1st pass:
standard, 1200kbps, multiple Bframes, GMC, h263
2nd:
standard, 1200kbps, multiple Bframes, GMC, h263
3rd:
slow, 1200kbps, multiple Bframes, GMC, h263

NO psv
final filesize 4,913 KB

http://img31.exs.cx/img31/1941/15521.jpg


plasma
1st pass:
better, 1200kbps, multiple Bframes, GMC, h263-optimized
2nd pass:
better, 1200kbps, multiple Bframes, GMC, h263-optimized
3rd pass:
insane, 1200kbps, multiple Bframes, GMC, h263-optimized

NO psv
final filesize 4,961 KB

http://img9.exs.cx/img9/4146/plasma.jpg



impressions
Plasma seems to give a higher quantization than 5.2.1, with no apparent quality improvements. I don't know if something in my plasma settings is wrong.

question
why *.log file has changed to map.txt and I can control it only in advanced setting?

[WorBry]

Hi StephanV. I disabled the Feedback Window as you suggested (it was already disabled in 5.2.1) but now I'm consistently getting 1879kbps (VirtualDub File Info) from a 2-pass setting at 2000kbps (same other settings as my first post above. Checked previous encodes at this bitrate and they are accurate. Reinstalled 5.2.1, leaving the settings at default, and then Plasma, but still got only 1879kbps. Uninstalled Plasma again and encoded with 5.2.1 and the bitrate was back to 2000kbps (actually 2027kbps 0.24% overhead). Weird.

[stephanV]

what happens if you choose another bitrate? like 4000?

[WorBry]

Bitrates:

Setting/ Encoded

2000/ 1879 (0.25%)
2500/ 2428 (0.20%)
3000/ 2946 (0.16%)
4000/ 4042 (0.12%)

These were using the "Balanced" quality settings. Just done one at "Extreme" setting for good measure:

2000/ 1882 (0.25%)

I think I'll go back to 5.2.1Pro for now !

[DigitAl56K]

Originally posted by WorBry
I think I'll go back to 5.2.1Pro for now !

Please remember this is an alpha, not a release candidate. It is likely to have many defects and you should not even be thinking about using it as a production product at this stage.

However, the more time you spend playing with the alpha, and the more defects you report, the better our next alpha build is going to be! Try not to tell us "this doesn't work, I'm going back to 5.2.1". Do some digging around and tell us in as much detail as you can where it doesn't work and the cause of the problem (if you can find it). The information we request on our DivX Labs Plasma post is a good place to start.

This is your chance to help make the next DivX release really shine. After the 5.2 release many people requested a beta program so that they could help improve the codec. There is always concern when releasing a beta that some people will try it, find a bug (which is only to be expected), and give up on it altogether. That concern is even greater when releasing alpha versions, but the codec team agreed that the community would love to experiment with bleeding-edge DivX technology and contribute defect reports back to us. In this light, please don't give up on testing. If you're having problems with bitrate-based encoding then report it and try quality-based encoding out instead. Experiment with the new feedback window. Tell us about problems with the GUI. The more defects you find the more we can fix :)

Thanks to everyone who has contributed defect reports so far - we've already started to work on many of them.

[Sagittaire]

For better target bitrate and better quality make N pass:

- First and N-1 pass with "better quality"
- Last Pass with "insame quality"

For me 3 pass is very good ...

[CruNcher]

2605 frames 25 fps Pal 720x576

DivX Plasma Multiple B-frames Better Quality H263 optimized 900 kbps 2pass
1pass = 6:50 min
2pass = 4:48 min
both = 11:38 min

Result = 11,1 MB (11.687.042 bytes) = 895 kbps

XviD 1.1 2-bframes VHQ=2 ME=6 H263 trellis 900 kbps 2pass
1pass = 1:53 min
2pass = 3:20 min
both = 5:13 min

Result = 11,1 MB (11.705.650 bytes) = 896 kbps

both were Visualy Checked (decoded)with Libavcodec and the DivX Plasma Picture Stability is awefull shitty i mean it looks like as the Macroblocks would party :P And this wasn't even compared vs my Tweaked Extra Detail XviD Build :D

[SeeMoreDigital]

Originally posted by CruNcher
both were Visualy Checked (decoded)with Libavcodec and the DivX Plasma Picture Stability is awefull shitty i mean it looks like as the Macroblocks would party :P You've beaten me to it!

I've just finished a similar test (without B-VOP). And was not all that impressed either!

I find it most alarming that there's such a difference between the two Mpeg4 codecs :eek:


Cheers

[WorBry]

Digital56K.

Sorry, I did'nt intend to offend. My comment about going back to DivX 5.2.1 was simply because I had reached the depth of my knowledge of the subject. I was not even sure if the problems were a genuine bug or whether there was something obvious I was missing in the set ups - wouldnt be the first time. I will leave the testing to those more qualified than I and look forward to being able to use the advanced features in due course.

Thanks.

[nicco]

And was not all that impressed either!

do you agree with me?
I still can't see improvements

one bug:
-GMC option automatically unceck when moving from 1st to nth pass

suggestion
"Plasma contains a new Insane Quality mode that enables all applicable quality enhancements"
-why don't you let the GUI shows what are the enhancements activated/deactivated in the various quality settings?

[LordRPI]

Originally posted by CruNcher

both were Visualy Checked (decoded)with Libavcodec and the DivX Plasma Picture Stability is awefull shitty i mean it looks like as the Macroblocks would party :P And this wasn't even compared vs my Tweaked Extra Detail XviD Build :D [/B]

CruNcher,

Does the libavcodec give different visual results than the DivX(tm) decoder?

[BBugsBunny]

I think one big problem is that people use PSNR far too much to measure perceived quality.
A german computer journal that I like very much had a "codec shootout" back in 2003 (10) where they used the Sarnoffs JNDmetrix to measure the perceived quality (but it's not free).
The JND metrix is a lot better to measure if a codec has got a good perceived quality or a bad one.
JND mextix: http://www.sarnoff.com/products_services/video_vision/jndmetrix/index.asp
C't (in german): http://www.heise.de/ct/03/10/146/default.shtml
There was a whitepaper about the jndmetrix (JND_whitepaper_0106.pdf) that is no longer online.
"The JND metrix takes account of coding artefacts, that mathematical analyse (PSNR) attributes little relevance, but the observer perceiving them with inconvenience."
I think the "macroblocks would party" is exactly what that means: you got a good PSNR but the perceived quality is bad.
Another quality measurement tool is here:
http://www.its.bldrdoc.gov/n3/video/

[IgorC]

It´s beter to use Divx decoder instead of ffdshow to play Divx´s. Divx´s decoder gives more details and not so smooth like ffdshow.
For me quality of Plasma DIVX and of XVID (1.1 Beta) is (less or more) iqual.

[CruNcher]

@Igorc ffdshow only smooths if PP is used
and yes for the not trained Eye those samples would look the same
but the DivX picture is hell unstable (bugy)

DivX Plasma Tritium
PSNR= Y:44.6601 U:47.3138 V:48.0535

XviD 1.1
PSNR= Y:44.9149 U:47.4825 V:48.6036

both rendered through the native decoders :)

NO PP was used in both

[LordRPI]

Originally posted by CruNcher

DivX Plasma Tritium
PSNR= Y:44.6601 U:47.3138 V:48.0535

XviD 1.1
PSNR= Y:44.9149 U:47.4825 V:48.6036

both rendered through the native decoders :)

NO PP was used in both [/B]

Thanks CruNcher.

I've actually gone home and tested a few clips of both Plasma and XviD using FFusion (libavcodec) and VLC on my G5 at home. Macs having a default target gamma of 1.8 (compared to windows target gamma of 2.2) show a little bit more detail when it comes to blocking. Playing them side by side simultaneously, I will say that Plasma has some more "well defined" blocks, but it also seems that XviD tends to smudge around those blocks when there's motion quite a bit more than Plasma does. I'm using an "Animated" clip for this.

I'm stressing both codecs with my test clips. I'll try to get in some higher bitrate clips.

Any chance somebody can do the above mentioned test and/or SSIM?

[nicco]

--DOES GMC WORK?

when I play plasma movie I don't see "frame type:GMC" (ffdshow osd)
Becouse in the gui the GMC option is buggy (it uncecks each time I acess the config dialog when -nth pass is enabled, and I have to ceck it each time) my doubt is that it would not work!

*in 5.2.1 I could see "frame type:GMC"*:
http://img26.exs.cx/img26/4950/521_2.jpg

*in plasma "frame type:GMC" does not appear at all!*:
http://img28.exs.cx/img28/687/plasma2.jpg


--How can I change te place/name of the new "map.txt" file?

[nicco]

Plasma Block-Party

http://img27.exs.cx/img27/1248/plasma_blockparty.jpg

http://img123.exs.cx/img123/5049/plasma_blockpart2y.jpg

[SeeMoreDigital]

Hi nicco,

That's nasty alright :eek:

On a different note... Where was this video shot? as it Italy or somewhere in northern Spain (beginning with G)?


Cheers

[CruNcher]

XviD 1.1 2-bframes VHQB=1 VHQ=2 ME=6 H263 trellis 900 kbps 2pass
1pass = 2:00 min
2pass = 4:00 min
both = 6:00 min

Result = 11,1 MB (11.736.890 bytes) = 899 kbps
PSNR = Y:44.9782 U:47.4984 V:48.5886


DivX Plasma Multiple B-frames Insane Quality H263 optimized 900 kbps 2pass
1pass = 20:10 min
2pass = 10:30 min
both = 30:40 min

Result = 11,1 MB (11.679.912 bytes) = 894 kbps
PSNR = Y:44.7439 U:47.3569 V:48.0983

still the macroblock party problem (that will be some nasty bugsearching nights for DARC ;))

[Sharktooth]

Originally posted by SeeMoreDigital
On a different note... Where was this video shot? as it Italy or somewhere in northern Spain (beginning with G)?
Uhm... i think thats Italy...

However, i cant manage to get GMC working thru the GUI too... btw the codec looks promising, but still slow compared to xvid.

[nicco]

btw the codec looks promising

the first step is a working GMC,
the second is no "block-party",

then it will look more promising for me when I'll see a really consisent quality improvement compared to 5.2.1

I hope our tests will help you divx team!

the shots are from....I'll tell you later!!!:D

[LordRPI]

Originally posted by CruNcher


still the macroblock party problem (that will be some nasty bugsearching nights for DARC ;))


@ nicco & CruNcher

Could you re-run those with Single B-frames? I've run through a couple of encoded clips and can't reproduce that and the only thing I see different is that you're using multiple bframes - I'm away from a windows machine for the weekend...

Edit - Thanks for the screenshot, nicco. I misunderstood a bit what you guys meant by "macroblock party."

[CiNcH]

What about the Rate Distortion Algorithm now? Is it QPel-aware? Are all modes that perform Motion Search (from Balanced onwards) QPel-aware?

[nicco]

Could you re-run those with Single B-frames
YES!
my tests was only with multiple b frames.
In fact they seem to appear only in B-frames.

the place is in...SPAIN!
where?....

[SeeMoreDigital]

Originally posted by nicco
the place is in...SPAIN!
where?.... Arhhhh... just put me out of my misery!

Is it somewhere in the north east? I'm sure I've been there!


Cheers

[nicco]

New test:
1st pass:
better, 1200kbps, single Bframes, GMC, h263-optimized
2nd pass:
better, 1200kbps, single Bframes, GMC, h263-optimized
3rd pass:
insane, 1200kbps, single Bframes, GMC, h263-optimized

NO psv

http://img120.exs.cx/img120/5532/plasma_blockparty_singleb.jpg

still block party!!!
If you look these bloks in detail you can notice that they seem vertical flipped...

the man who gave me this movie told me this place is Santiago de Compostela, Spain ;)

[SeeMoreDigital]

Originally posted by nicco
the man who gave me this movie told me this place is Santiago de Compostela, Spain ;) Jeez...

This is miles away from where I thought it was (got the right country though)... Can't say I've been to this region.... Unless it was in another life!


Cheers

[crnagora]

Well,

is it just me or is the n-th pass function broken ?

Because I've done a 6-Pass Encode of a short clip and after viewing the map.txt it showes only: "Passes 1".

http://img42.exs.cx/img42/8238/maptxt.jpg

Here is my CLI:

-bvnn 500000 -vbv 4854000,3145728,1376256 -dir c:\ -pq 8610 -nf -profile=3 -quantization=2

Settings:
500k Insane Quality
B-Frames (Single)
No GMC no QPEL
H.263 opt.

Hm...maybe something went wrong because I choosed the root directory instead of c:\DivX...

Oh...and FrameLevelControl.txt also doesn't work.

[Sharktooth]

Nicco, are you sure you have ALL ffdshow postprocessing disabled?
Cause "Accurate deblocking" in the OSD means...

[Sagittaire]

Non it's a real bug with bframe: I have the same problem with all MPEG4 decoder or player ...

[nicco]

Nicco, are you sure you have ALL ffdshow postprocessing disabled?
Cause "Accurate deblocking" in the OSD means...

I've never said postprocessing was disabled:p
In fact it's enabled, but this problem don't appear in Xvid and Divx5.2.1 movies which are ffdshow-decoded too, and it appears only in B-frames!


Non it's a real bug with bframe: I have the same problem with all MPEG4 decoder or player

I'm not alone!:D

[Doom9]

umm.. am I the only one who only gets crashes from this release?

I've been encoding this same movie using 4 other codecs, all of them pre-release (one being xvid and the rest codecs I can't mention right now), all using the same avisynth script and vdub 1.6.1, but as soon as I start encoding in plasma, vdub closes itself without warning and any error or debug message. I figured it could be vdub so I went back to v1.5.10 but that didn't solve anything.

settings are pretty standard: unconstrained with multiple b-frames, qpel, h.263 and the rest of the defaults. The script also doesn't contain any specialties.. just the lanczos resize filter and that's it.

[SeeMoreDigital]

Jeez...

Personally I don't use VirtualDuB or VirtualDubMod. But I seems to me, just about everytime DivX releases a new codec, loads of vdub users start complaining of crashes!


Cheers

[thrivx my appendix]

Doom9, I have the same problem. There is a post about this error in their feedback forum. No updates yet :(

[DigitAl56K]

Mr Doom9 sir: Chances are it's a defect in Plasma :) What you have in your hands is a very early alpha. We'll be throwing up new builds on DivX Labs as we put them together. Keep reporting the bugs, we'll keep fixing them.

[Sagittaire]

@ Doom9

Disable feedback windows, QPel, GMC and psy mode ...

[Doom9]

disabling QPel now gives me a C++ runtime error from vdub once encoding started. The feedback window seems to have no effect.

[LigH]

Originally posted by WorBry
Regardless of the quality level (Balanced through to Insane) it just kept defaulting back to 780kbps
Originally posted by Doom9
umm.. am I the only one who only gets crashes from this release?

Agree: It either encodes with 780 kbps, or VirtualDub(Mod) crashes (either without further notice, or with "Abnormal program termination").

AMD Duron 800 (MMX, ISSE, 3Dnow!), 256 MB RAM; DivX Build 1394.

[Gannjunior]

Hi,
i've done a very fast test on a short clip to see if MPEG matrix bug has been solved. But i've had the usually negative result i found in 5.20 and 5.21: MPEG matrix has broken and it works in opposite way it should do.


1200 frames encoded at Q=2 (saturation), balanced mode.

- h263 -> 16.920 Kb
- h263 optimized -> 17.060 Kb
- MPEG -> 9.368 Kb

....So i don't understand why divx networks doesn't solve this bug. I, and and other ( Lenox, Sharktooth...) had shown the problem some months ago but without obtain any answer...
i pefectly know Plasma is just an alpha release, but the mentioned problem exists since 5.2x : i'd like only to know if the idea to solve the problem has been taken in account.

ciao!

[DigitAl56K]

There is some recognized oddity there.

[thrivx my appendix]

Off topic: Digital, do you have Nordic roots?

[calinb]

Al's roots are red...and that's about all he's got with his (short) hairstyle. :D

http://cv.divx.com/2004/10/divx-hq-revealed/

[DigitAl56K]

Uncalled for!!! *pouts*

;)

[SeeMoreDigital]

Al... Why arn't you wearing "national dress" at work?


Cheers

[LordRPI]

He does, he just wore pants that day to hide the chain from his desk.