Temporal blurring of quantizer

[check]

Hi all, I was wondering if someone could take the time to explain why x264 temporally blurs the quantizer curve, why blurring is applied twice, and what units (or scale) the two settings are in.

On a somewhat related note, as I understand it, the default --qcomp setting moves the bitrate distribution from what the encoder feels is optimal to a more constant distribution. How does it do this, and why is this not set to 1.?

[akupenguin]

Hi all, I was wondering if someone could take the time to explain why x264 temporally blurs the quantizer curve[QUOTE]
It makes quality more stable. Especially in anime, where the framerate of the content is often less than the nominal framerate of the video, so in two consecutive frames one has motion and the next doesn't. Without any temporal blur they'd have very different qp, but the optimal encode is to make them the same.

[QUOTE]why blurring is applied twice
The ratecontrol I stole from libavcodec already did one, and I added another algorithm I liked better.

and what units (or scale) the two settings are in.
frames.

On a somewhat related note, as I understand it, the default --qcomp setting moves the bitrate distribution from what the encoder feels is optimal to a more constant distribution. How does it do this, and why is this not set to 1.?
It's hard to define optimal. But no, the default qcomp is my attempt at optimal, not a compromise between optimal and something else.
qcomp=1 would be identical to cqp.

[check]

Thanks for the info. A few more questions from your answers:
o if the normalised t.blur setting is 1, will that average all frames, or just everything within IDR range or similar.
o do either of the blurs take into account (presumeably by ignoring) scene changes?

o with --qcomp, was 'optimal' something that you could measure, or is it by a large part perceptual?
Thanks again.

[akupenguin]

if the normalised t.blur setting is 1, will that average all frames, or just everything within IDR range or similar.
No, it's not normalized. The radius is in frames. qblur=1 is a gaussian blur of radius 1.

do either of the blurs take into account (presumeably by ignoring) scene changes?
Yes, in addition to limitting the area of effect based on distance, both of them stop at intra frames. (and that's any frame containing mostly intra blocks, even if scenecut detection was disabled.)

with --qcomp, was 'optimal' something that you could measure, or is it by a large part perceptual?
Thanks again.
perceptual