Tests using Psy. Visual Enhancements in 5.1 = Larger Files!

[BluDChyLD]

Hey guys,
Ain't done much testing yet, but I can confirm that using "slow" or "slowest" peformance setting will dramatically decrease FPS! On my tests the nth pass speed dropped down from 12-13 fps with standard to around 2-5 using slow or slowest.

Anyway, just fiddled with some quality based tests and the new psychovisual modes and here are the results -

(all tests using 20 second clip@quantizer 2)

1. requiem_nopsy.avi - 2.30mb
Total Encoding Time: 31 seconds

2. requiem_fast.avi - 2.93mb
Total Encoding Time: 46 seconds

3. requiem_slow.avi - 4.93mb (!)
Total Encoding Time: 1 minute, 34 seconds

As you can, the higher the psy vis setting the slower the encoding time and the higher the file size! This is very different from earlier divx versions - psy vis settings always lowered the filesize in quality based mode. Anyoneelse want to do anymore tests to confirm?

Hope this helps :)

[temporance]

I haven't seen this (file sizes at CQ are different, but not that different). Was this a noisy clip?

To my eye, these settings do improve quality (using PV during final pass of a multipass encode), especially skin tones, fine details and dark areas.

Having said that, I haven't tried a huge amount yet and improvements seems to be content dependent.

Anyone done any PSNR or HVS test with these settings?

EDIT: try using the Feedback Window to look at effects of PV. Select "Encoded Frame", "PV level" (bottom combo on RHS) and select "Full Macroblock". You can also pause the encoder and try the effects of changing various settings.

[shirka]

I tested at 700 kbits/s psychovisual (slow) + GMC + QPEL + BFrames -> It adds uggly noise, but without psychovisual and gmc, just QPEL + BFrames, -> no noise (but no such quality than Koepi's Xvid 06242003 buid with H.263, Chroma, Qpel, BFrames, VHQ4, TrellisQuant)
I will keep using Xvid for the moment ;)

[temporance]

@shirka: Don't give up so quick! Try multipass, "Slowest", B-frames, no GMC, no Qpel and PV fast (IMHO best settings).

[shirka]

Here is what I used : psycho slow and encoding slowest, 3 passes
Why no QPel, is it still buggy ? As for me, it produces a sharper image, with more details. You think psycho fast produces better quality ?

[temporance]

> You think psycho fast produces better quality ?

Not sure yet, but I think it's a little sharper than "slow" for the same bitrate. Both are better than PV off, IMHO.

[DigitAl56K]

Fast PV is stronger than Slow PV, and according to the guide more advanced because it better places noise when redistributing bits. Slow is aimed primarily at PV with minimal artifacting.

Please don't run PSNR tests with PV enabled - the very function of PV is to change the image which naturally degrades PSNR scores.

[acrespo]

Hi DigitAl56K,

I saw your message in divx forum about peformance/quality settings.
You said:
"Do the first pass on Standard, then move up to Slow or Slowest on the next pass. When you change modes you must disable the MV file."

And about PV settings? Can I disable in 1st pass and enable "fast" in 2nd pass or I need to enable "fast" in both passes?

[LordIntruder]

acrespo:

you can do it, it works fine and it will reduce your encoding time.

[LordIntruder]

I copy-paste divx info about PV:

The "slow" mode analyzes each block and decides whether hiding some noise there can save some bits without degrading appearance. It uses one metric per 8x8 block, so is a pretty straightforward algorithm. It is slower because it is more conservative and is very careful to avoid any blocking or ringing artefacts.

The "fast" mode is, in theory at least, more sophisticated as it is capable of controlling where noise appears within the pixels of an 8x8 block. So, for example, in an anime source, it attempts to "shape" the quantization noise is situated only on the pixels that make up the cell borders.

I thought slow was the best for movies but some people think it may not be the best choice. I'll make some tests and change just the psy option from one test to the next to see what happens.

[DigitAl56K]

You can change settings between passes, but when you do so you must disable the MV file at that point.

Regarding the tests BluDChyLD has conducted here, there is a mirror-thread running over at DivX.com in which eagle (DARC member) has responded:

http://forums.divx.com/viewtopic.php?topic=53458&forum=6

[LordIntruder]

Ah yes but people that experience different filesize between no PV and PV fast or slow, seems to not encode using multipass.

I never experienced such things while my encodes. I use Gknot only to make my own calculations and to do the compression test. If it gives me say 805K to obtain 699Mb, when I open my VDM and put 805K, some hours later I get a 699Mb avi. Of course if you play with filters inside your avisynth script the size may vary but that is another story.

Now those using constant quantizers, etc... have an increased filesize but those that use multipass have no problem.

Didn't got it at first and was making some short tests with and without PV in multipass without notice any change, I was a bit confused. hehe.

[Kheffer]

Originally posted by DigitAl56K
You can change settings between passes, but when you do so you must disable the MV file at that point.


The question is All settings or only "standard", "slow", "slowest"?