Best performance/quality ratio with DivX Pro 5.1.1

[CiNcH]

Hi friends,

I have just read the official DivX 5.1 Guide (which is just great, thanks to DigitAl56K & Co) and stumbled over two suggestions for a multipass encoding I do not really understand.

Keep in mind that a multipass encoding should only refine rate control strategy.


Suggestion 1:
Use Slowest P/Q mode only in final pass (Standard P/Q mode in previous passes).

Personal comment:
Slowest P/Q mode resp. its Rate-Distortion algorithm should have great impact on rate control strategy. Why isn't it advised to use this algorithm for second-last and final pass? Doesn’t RDA more do a refining of the rate control strategy for the next pass? (when using it in final pass only it has the effect of a 1-pass, doesn’t it? So it would require another pass to effectively take advantage of the RDA using the multipass log...)


Suggestion 2:
Use Psychovisual Enhancements only from second-last pass onwards.

Personal comment:
Wouldn’t it be enough to enable the Psychovisual Enhancements in final pass only? As far as I understand PVE it only reallocates bits for each frame individually (by lowering bit spend in textured and giving those saved bits to flat areas of the picture) and doesn’t have the least impact on rate control strategy so it cannot screw it up.


Any opinions and corrections are welcome, thanks.

[DevilsChild]

As for the first item, I think it's because DivX5 is extremely slow in "Slowest" mode. In theory you should get best quality in 6 passes with Slowest mode, but few would want to spend 2-3 days encoding a single movie.

[SeeMoreDigital]

Agreed,

In fact, when generating the second pass I use 'Standard Mode' as well. As I can't visually see any difference in the encode between 'Standard Mode' and 'Slow and Slowest Modes'

I much prefer to do two 'Nth' passes in 'Standard Mode' which for me does generate a visual improvement!

Cheers

[DevilsChild]

Yep, "Standard" mode is extremely fast and in my experience the quality is only marginally worse than "Slowest".

I just encoded Disc1 of "Pearl Harbor" in DivX 5.1.1 with neutral bicubic resize (640x256, 1045 kb/s) and "Standard" quality mode. The Home Theater profile was selected, and an MV file was generated and reused in 3 passes.

Encoding speed on an Athlon 2600+:

1st pass - 40.348 fps
2nd pass - 49.082 fps
3rd pass - 47.675 fps

Total encoding time was around 3.5 hours for a 2 hr clip. Seems to be pretty fast for an allegedly "slow" codec. In fact, I think DivX 5.1.1 is second only to Nero Digital in terms of encoding speed...

[brute]

please try it with 704x288 Lanzcos filter, disable Profile, and GMC and Bidirecctional encoding enabled!

[SeeMoreDigital]

If you want your encodes to match the sources aspect ratio of 2.35:1 then give the following a try: 640x272, 672x288 or 720x304!

Cheers

[DigitAl56K]

Hi CiNcH,

Let me take a pass at answering your two questions:

1 - When we tested doing 1st Standard, nth Slowest versus 1st Slowest, nth Slowest, on average the results tended towards 0 PSNR difference (PSNR is one mathematical evaluation of the video quality - its not perfect but its a good indication). In the first pass motion and texture complexities are analyzed and recorded for each frame in the multipass log file (divx.log). You can open the log file in Notepad to see this if you like. This leads to the answer to your sub-question: "If you change the p/q mode on the last pass doesn't that make it similar to a 1-pass" - no it doesn't, because the rate control still has access to all the motion and texture complexity information it needs from the log file and the approximate bitspend. The RC in DivX 5.1.1 is very flexible and so changing to Slow or Slowest mode on the last pass brings you the benefits of the RD algorithm without the performance hit of doing Slow/Slowest for every single pass.

2 - Psycho-vis affects both the texture complexity and bit spend. You can tune the RC without psycho-vis if you like during 1st pass, but to best incorporate this into the rate control strategy its wise to run at least one pass before your final pass with the PV on.

So combining both these tips for 2 passes you would have PV on for both passes, p/q would be Standard for 1st pass, and Slow or Slowest for 2nd pass.

I hope this helps - if you find your results differ or you can contribute advice of your own we'd love to hear it for the next guide version :)

-Al

[CiNcH]

Hi DigitAl56K!

Thanks for the answer.

Probably you can help me with another one ;) ...

What is this "Psy-Vis Chroma operation" all about (which seems to be beta at the moment)? From what I have seen in the feedback window it reduces bit spend. Does it take advantage of chroma info in motion search?