Log in

View Full Version : Usefullness of "lower-than-average" quantizers in xvid encodes.


OUTPinged_
17th April 2002, 11:48
Here is the topic: is excessive(5+%) the usage of "much lower than average"(read quant2&1) efficient(noticebly increasing quality) in encodes with "high" (4 and up) average quantizer?


the answer was "yes" for divx311 as it wasnt very good with very small frames on 5+ quant. i did not yet run into that problem in xvid (guess it is fixed :-), and i dont see a difference between a "10" and "90" for low distance value (virtually no/lots of quant2 frames) in a test encode besides a expected quality decrease in small frames due to higher quantizer.

however those bits that are saved when that 20% of frames get their quantizers 1-2 point higher, do affect other frames. i was able to tell a difference on some hi-mo scenes due to their
lower quantizer (some frame's quantizers were lower by 2(!)) and small-frame scenes looked about the same (they all were in quant3-4 range) in motion.


so now to me it looks like an overall better-looking movie will be produced when not using loads of "average_quant-2 or lower" quantizer frames.

please, if somebody has a good reason against not using lots of quant.2 frames in xvid, post here why.


PS> all above is assuming average quantizer is in 4-5 range.

Teegedeck
17th April 2002, 12:10
Nope, I agree with you on this one. :)

Besides, quant=1 really, really is a waste, it just consumes a lot of bits. I'd set min. quants=2.

Quant=3 starting at pretty low bitrates seems OK, too. While it is a bit extreme to leave the low distance at 300% as I did (yeah, I've come to that conclusion at long last), setting it lower than 90% shouldn't improve anything.

OUTPinged_
17th April 2002, 12:53
from my XP a good range of low distance lies within 50-100 range (definitely not higher). 300 is overkill imo.

[edit] that is for high aggression, lower agg levels can operate with more wide ranges pretty good.

btw Teegedeck, where can i find a luma masking code in cvs? no trace of it's algo in codec.c/2pass.c :/

-h
17th April 2002, 12:57
Lumi masking code is in core itself - see encoder.c and adapt_quant.c for specifics.

-h

OUTPinged_
17th April 2002, 12:59
roger that sir.