"quantizer value jitter"

[OUTPinged_]

ugly topic, eh? :-)


what i want to ask, is a (drf value sequences follow) jumping quantizer value stream look preduce better quality output than the constant quanztier(closest to average) value stream.

i mean, will a sequence 3-3-3-3-3-KF-5-5-5-5-5 look better than 2-3-4-2-4-KF-4-5-6-4-6?

interesting to know if the outcome will change if it will be:

1.static scene
2.hi-mo scene
3. panning scene.

some moments: jitter will occur only with nearest values. small motion begins(ends?) on peak (lower drf). high changes are eaten by KF.

why this question has risen.

the new CC algo has that amount of "drf jitter" controlled less or more by curve aggressivnes.

if that thing is usefull, CC could be adapted for it. it wouldnt reduce the filesize but can produce better looking scenes :-)

Foxer, it is you who is supposed to know that, Teegedeck said that himself :P

[Teegedeck]

Hm, on second thoughts (the two of us have talked about this before ;)), I don't think a sequence like 2-3-4 would have a benefit over 3-3-3 because it would take an I-frame after the '4' in order to recover from the degraded quality of a P-frame encoded at quant 4. So I think a VBR encode at almost constant quantizer (only the highest bitrates cropped with a higher quant and the smallest frames compressed less) should be preferable.

hoyoyo

[EDIT:] Err - I didn't mean to say Foxer should know the answer, I only meant he's the one who can implement stuff in curve algos.

[Koepi]

*sigh*

it's known since the beginning of the nandub days that the _visual achieved_ quality during watching the _sequence_ (not single pictures, it must be the running movie) is better when having a slightly fluctuating quantizer.
It's something that some other folks around here (forgot his name, sorry) try to achieve with adding noise to the output picture.

You could explain it like with ISDN. If noone speaks on such a phone, there is no noise - that's why people thought the line was dead, so now there gets artificial noise added.
When you watch some scenes that are perfectly noiseless, you will very likely see that the picture looks very artificial and strange.

This fluctuating quantizer helps here a little as the level of quantization noise varies and gives a "more vivid" picture.

I hope this explains the effect a little.

Regards,
Koepi

[Teegedeck]

Fine, then I'm reliefed that I always get a slight fluctuation... But I don't think it would do good to introduce extremely 'jumping quantizers' or would it?

One could still use GKnot to reach a more 'vivid' picture but I've come to dislike that a little. Perhaps I'll do some comparisons on Easter weekend, IF I have time. A matter of taste, anyway. Perhaps we're just not used to clean sources...

[OUTPinged_]

no one was talking about too much jitter there.

maybe just adapt the CC algo for constantly using jitter?

[Teegedeck]

What kind of jitter do you have in mind? Is the constant change between two adjacent quantizers that we usually get now not enough (3-4-3-4-3)?

BTW, B-frames would be better suited for an intentional 'jitter-effect' because a sudden raise of quantizer in a B-frame shouldn't hurt the following frame. So, perhaps there's no need to change anything in curve-modulation at all.

[OUTPinged_]

in that case, yes.