qcomp quality tests

[-Habanero-]

I did some comparisons of different quantizer curve compression settings with x264 when I recently found out that my use of 0.7 instead of the default might've not been a cost-effective choice.

qcomp SSIM+ standard deviation
0.1 95.79 2.41
0.2 96.09 2.09
0.3 96.40 1.74
0.4 96.63 1.42
0.5 96.72 1.15
0.6 96.72 0.96
0.7 96.54 0.81
0.8 96.24 0.84
0.9 95.66 0.79
1.0 95.19 0.77

So the extremes are the worst, no surprise there. 0.5 and 0.6 are the same average quality but 0.6 has less fluctuation. Again, no surprise, and the screenshots confirm all this. With a capitalist qcomp, the backgrounds were flawless but anything that moved was garbage. And with 1.0 everything truly was constant quality but look at the wooden posts and windows. Smearing, details gone and noise around windows.
So I'd have to agree with the metric that equal sharing of misery does look worse than the other extreme.

Looks like the center leaning a little towards equality wins. The truth is always in the middle.

https://postimg.org/image/jxmoqpl0b/
https://postimg.org/image/m5hxof15j/
https://postimg.org/image/uz8u5io47/
https://postimg.org/image/gerraot5j/

EDIT: Regular SSIM sucked really bad here. There was a linear downward trend from 0.1 to 1.0. The more moving parts were trashed, the better the quality according to it. SSIM+ really seems to have improved upon its predecessor.

[El Heggunte]

But you haven't explored all the possibilities yet.

Namely: zones, qpfile, and (yes!) custom quantization matrices

[-Habanero-]

In conjunction with quantizer curve comp or on their own?