x264 CQVBR still pumping?

[NoX1911]

I use x264 (518) CLI with following command line:

x264.exe --crf 30 --keyint 500 --ref 5 --mixed-refs --bframes 3 --b-pyramid --b-rdo --bime --weightb --direct auto --subme 6 --trellis 2 --analyse all --8x8dct --me umh --merange 32 --progress --no-psnr --output "" ""

Source material is RGB lossless 25fps (->ConvertToYV12), animated (amiga game, frame capturing). The final video shows "pumping" like used to be seen on CBRs. After a complete scene change (for example a white flash covering the whole screen) the picture needs around a second to get sharp. It smoothly fades from low to high quality. Shouldn't this be prevented by 'Constant Quality' mode?

Example video:
x264 pumping.zip (http://tinyurl.com/rcgzh) 259kb

[akupenguin]

In short: yes, the rate of change of qp is limited by --qpstep.

Also complicating your video: x264 sees the speckles all over the objects warping in, and says "this frame is high complexity, give it a high qp". Only once everything is motionless does it decrease the qp.

Yes, a per-macroblock qcomp would be better (try the --tcplx patch, combined with increased --qcomp). But as far as per-frame ratecontrol goes, it's working as indended: The alternative (at the same bitrate) is to give all frames the same low quality, and never turn sharp.

[Oline 61]

Using --qp would eliminate the pumping by giving each frame the same quantizer, wouldn't it?

[akupenguin]

That's what I said. But do you think a constant low quality looks better than some lq frames and some hq stills?

[GodofaGap]

Perhaps lowering --qpmax could help a bit? I don't know how big the fluctuation is.

[NoX1911]

I played around with --aq-tcplx, --qpstep and --qcomp but achieved best results with --qpmax 30 (at --crf 30) with just a slightly gain in filesize.

Kinda better now.

:thanks: