Can You Use Compressability Check For Xvid in Gnot?

[Dali Lama]

Just wondering...

Is Divx 3 or Divx 4 better as a compressability check?

Thanks,

Dali

[Teegedeck]

You can do a compressibility-check as you used to with DivX3.11.

[rui]

Originally posted by Teegedeck
You can do a compressibility-check as you used to with DivX3.11.

And the conclusions will be the same that doom9 descrives in his Gnot divx3 guide?? (Values between 60% and 100% mean good quality...)
Wouldn't be more logical to use the divx4 comp check, since XviD uses 1k=1000 like divx4?

Thanks

[rui]

Could someone clarify this?

[Doom9]

as with divx4.. the values are actually lower... but afaik nobody tested enough to give you a reliable range of values

[rui]

I am going to try to encode my first XivD movie today (just been reading all you guys posts to try to learn as much as i can before steeping into it), and i will use comp check from divx4. For start, i will use the values Doom9 gave at his divx4 Gnot guide.
Wish me luck :p

[sierrafoxtrot]

@rui i know you've done like a million tests ;) ,what have you found?

thought this thread was dead but hopefully this'll get ppl sharing their experience ... :)

so far i've done about 10 or so very decent rips with Xvid using 2-pass int and found that the compression values that the divx3 comp check gives are kind of accurate, well, more later, but ...

here's what i've found (all values div3 check, no noise filtering, soft or neutral bicubic resize)

memento 60% of 0.290 (640x272)
mexican 50.1% of 0.417 (560x240)
bound 84% of 0.170 (640x352)
saving private ryan 26.3% of 0.812 (592x336)
girl on the bridge 52.3% of 0.564 (560x240)
bullit 69.6% of 0.249 (560x320)
sweet november 57.8% of 0.330 (528x288)
being john malkovich 70.3% of 0.218 (608x336)

the encodes were done using:

2-pass internal
credits at either 10% or quant 20 (depending on build used)
H263
Interframe lock min 2 max6
Intraframe lock min2 max4, quantizer smooth enabled
no luma masking in either pass

and watched using div4 post-processing set at level 3, and the quality is ... not always what i expected.

memento (cinema release) was brilliant, everything was crisp and clear and there were hardly any macroblocks worth mentioning. but it gets interesting when i encoded the special DVD edition where the movie is shown in reverse (to help low watt bulbs like myself). there is a portion of the movie in the beginnning where all the B/W scenes are shown together, and the quality was decent, but some frames were pretty appalling, so much so i had to manually correct the curve in GKnot to get a decent rip.

the mexican was fine, still scenes show some *slight* blocking, but nothing to write home about. bound was brilliant, perfect encode, as was being john malkovich.

SPR was the only case where i disregarded the resizing value given by GKnot comp check, going as low as 26.3%, but seeing that this was what doom9 did for his codec comparisons .. i thought it would be fine. and yeah, the encode was grainy and more than a little blocky during certain scenes, but well within my expectations, considering the difficulty of the movie.

girl on the bridge, french film, B/W, moderate motion. pretty good quality but there were ringing artifacts around the sharp edges due to really high contrast (probably because i used neutral bicubic) but no real complaints about quality.

bullit was pretty decent, but some moderate motion scenes still exhibited blockiness (oddly enough the high motion scenes were pretty crisp.

and sweet november (sappy, but this damn film almost made me weep, guess i wasn't emotionally prepared :angry: ) was once again, fine for the high motion, high contrast scenes, but was really blocky when trying to encode low-contrast dark scenes (despite having disable luma masking). and it is wierd because despite using a size/check % of 57, it seems that most of the movie was pretty good, and then some of it was duff.

i'm not sure what this means, but could it be that the curve modulation is not as effective as Gknot's calculations at maintaining constant quality? (i can't wait for the implementation of foxer's code :p )

so in summary, what i've found is that most movies will look good at % values down to about 50%, i know i know, it's an arbitrary value midway between div3's 60-70% and div4's 40-60%, but it seemed like a good place to start. however, sometimes certain scenes (high conrast, medium motion) will be favoured over others. does this mean larger frames are over-favoured over smaller ones?

i haven't had time to encode the same movie with different resolutions (and hence % comp check values) and that would've been the most scientific method. might do that sometime, but i guess i'm trying to get everyone's input and come to a consensus without too much hard work :p

phew (what a spiel) ... well?

[rui]

Well, I must say that your post must be one of the most complete posts I have ever read :)
I must admit that I still haven’t made any tests using divx3 comp check.
I have always used divx4 comp check, with some positive results.
I started using divx4 because, first, no one would tell me a reason not to, and second, after reading doom9 post in this same thread, where he says that XviD would be like divx4, but with lower values, in spite that no one had made enough tests to know that for sure.

After reading your post, I came to the notion that this is an area that still needs some investigation.
I am about to install divx 3.11 codec, to be able to make some divx3 comp checks.
I will be using the same movie to make a divx3 comp check, and then using divx 4 to make another comp check.
After that I will encode that movie using XviD, and both the resolutions that the either divx3 and divx4 comp checks advise.

But another problem arises: we are assuming that, like you said in your post, for divx4 the values should be in the 40-60 range, and for divx3 they should be in the 60-70 range.
But for XviD this could be false. Even if one decides that divx4 comp check is best, should we still maintain those values (40-60) as the target values?

[sierrafoxtrot]

@rui

yeah, i know what you mean about being unsure about whether to use divx3 or divx4, and with what values. i've found that div3 compcheck values above 50 are okay, with some macroblocks, but all my really good rips were >60%, so it could be something like div3. don't know. might try to run some more tests when i find the time.

i'm trying more for >55% nowadays, hopefully it'll show some improvement ... :)

[rui]

Well, after reading your last post, here what we have:
You have made tests using divx3 comp check, using as target values the ones that doom9 indicates for divx3 (60-70). And you are getting good results.

I have been using divx4 comp check, again using as targets doom9 values for divx4 (40-60), and i am getting aceptable results too.

So, this means that either divx3 and 4 comp checks can be used, if we respect the target values for each one, as determined by doom9?
:confused:

That would mean that my previous doubts about the target values weren't correct ?

[philippas]

Well the best values are not 60-70 % for any of the codecs. If you can reach 100 without compromising a lot the resolution is the best setting to fit a DVD rip to 1-2 cd's.
In the case of Xvid which can use quant = 1 calculating the compressibility test is slightly more difficult and relative what you consider perfect rip (the real perfect rip would be quant = 1-1 which is almost lossless transfer from the DVD-->divx). Given though the space limitations you can't use quant 1-1. Locking the quants with a min=2 will match the assumption of the current compressibility test, which is 100% => all frames encoded with DRF=2 or average DRF = 2 (for Xvid).

If you use Gnot and do a comp test with Xvid the values will be analogous with divx4, divx3. The only difference will be that for the same movie with the same test, divx3.11 will give a lower indication and divx4 will give a higher indication. Xvid will be at the middle.
Now if you reach with Xvid >100% you can either increase the resolution (as with the rest codecs) or lower the min quant from 2 to 1.