CRF ratio calculation

[jeffy]

Is there a way to calculate a good approximation of the filesize of a file encoded at eg. CRF 18 when I know it has a certain size at CRF 20? In other words, is there some equation for a simple ratio calculation, eg. the filesize at CRF 18 equals 1.15 × the filesize at CRF 20? All settings remain the same.

I'm sorry if this was answered before, if yes, just give me the link, please. Thank you.

[stax76]

I would guess difference between 18 and 20 will be 20%-30% percent.

[Dark Shikari]

Dropping CRF by 6 will double filesize. Dropping CRF by 1 will raise filesize by 12.5%.

Roughly. Its quite possible to have a case where this guideline is totally wrong.

[ajp_anton]

If dropping by 6 means 2x increase, shouldn't dropping by 1 increase filesize by [2^(1/6)]x ~ 12.2% ?
Yeah I know, who cares... =)

[Dark Shikari]

If dropping by 6 means 2x increase, shouldn't dropping by 1 increase filesize by [2^(1/6)]x ~ 12.2% ?
Yeah I know, who cares... =)The 2x increase is an approximation, the 12.5% is an exact definition from the quantizer definition in the H.264 standard.

Of course, neither would guarantee exactly that change in bit cost.

[jeffy]

Thank you to all, and especially to Dark Shikari for the exact definition.

[Sergey A. Sablin]

12.5% actually came from Ian Richardson (not from standard) - http://www.vcodex.com/files/h264_transform.pdf
"Note that Qstep doubles in size for every increment of 6 in QP; Qstep increases by 12.5% for each increment of 1 in QP."

But if you'll look at the Table 2-1 Quantization step sizes in H.264 CODEC

QP 0 1 2 3 4 5 6 ...
QStep 0.625 0.6875 0.8125 0.875 1 1.125 1.25 ...


it's clear that quant step exactly doubles at each increment QP by 6 and increment QP by 1 changes quant step by 7.6658% .. 18.1818% depending on actual QP value.

While in general increase of quant step by X% should drop file size by ~X% - it's only valid for medium and high bitrates, cause at low bitrates this model behaves non-linearly because of many macroblocks without coefficients.

[jeffy]

@Sergey A. Sablin: Thank you for the link.