Payback

[NuclearFusi0n]

What's the difference between Payback proportionally and payback with bias? :confused:

[iago]

@NuclearFusi0n

"Payback with bias" symphatizes and favours small frames ;), whereas "Payback proportionally" behaves justly ;) towards all frames and without any prejudice pays back what each frame deserves according to their sizes, thus gives more to big frames and less to small frames! ;)

regards,
iago

[MaTTeR]

iago,

I'm still undecided on which I like to be honest. Perhaps it's my imagination but "bias" seems to cut down on DCT blocking in general. Lately I've been playing with "proportion" on a few low motion flicks with pretty good results though.

Maybe a stupid question considering who I'm talking to but have you done much testing in comparing the two?

[iago]

@MaTTeR

I feel the same as you do actually, since I'm not sure either. I haven't done "exact" and "to-the-point" tests on it, comparing the quantizer distribution and frame sizes, etc. However, I've been using payback proportionally for a long time and especially in high-motion scenes or in highly detailed scenes requiring much bitrate I've had almost no problems for a long time. Or at least I feel like that; and to be honest, just like you, it may be just my imagination too! ;)

regards,
iago

[vinouz]

Well, atm I'm encoding Ronin and I noticed while small frames aren't bad that's mostly big frames that I don't feel big enough. But I didn't help the codec as I did my tests on a part which contains an important proportion of hi motion scenes (car chase in Arles (which btw is a really nice city I live not far from :) )).
So I would prefer linear, and even bias toward big frames. At least for this movie (which comports many car chases, and doesn't have a very good source quality in my opinion [is it due to too much filtering, bad encoding or - I'm rather considering this - bad acquisition])...
Btw I tested Divx5 (gmc+bframes+Qpel, as my Xvid setup [I know gmc and Qpel are not fully implemented but that's the fairer comparison settings]) too, and noticed it gave overall lower quants and a quality distribution I prefer (I won't use better quality, as it isn't proper in my mind) : in hi motion scenes it has lower quants then more bytes thus pactically no artifacts and a smooth image, and in other cases, it's a little blurrier (only a little) and has that boring 'moving walls' artifacts on the road ground.
I have an idea about 'moving walls' artifacts. It's like having the motion vector, but no difference DCT data associated to it, when the difference is almost linear. Only the mv. So I suppose it's done (no diff data) using dquants, and it's an artifact due to lowering too much quants in some regions... Is that right ? And does DivX5.0.2 use dquants ?

Cheers,
Vinouz