x264: qp2bitrate, bit2qp ?

ncahammer · 22nd October 2008, 13:38

If I encode a two pass encode, using a given bitrate A and the x264 log reports an achieved mean QP B, is some kind of function that will give me the estimated QP B' if I reencode (using the same video and parameters of course) using a bitrate A'

is the opposite also predictable ?
I mean if I want to achieve a QP B' what is the bitrate A' that I should give to x264 ?

I searched the x264 source code and I did find in the ratecontrol.c some functions (qscale2bits,qp2qscale) but I am not sure that I am in the correct path, I suspect that I need to read and reevaluate each frame from the .stats file

Sagekilla · 22nd October 2008, 15:37

The simple fact is you can't determine how many bits will be given for a certain QP and the number of bits to achieve a certain QP.

What happens when you have a completely flat frame, i.e. one color? It could have a QP of 0 and take up almost no bits. On the other hand, you can have a scene from parkrun that has a QP of 20 but takes up tons of bits.

akupenguin · 22nd October 2008, 15:59

Bitrate is approximately proportional to 2^(-qp/6)
Yes you'll get a slightly more accurate estimate if you run qscale2bits on each frame, which scales mvs and residuals by different amounts.
But if you want a target quality, use CRF and not 2pass.

ncahammer · 22nd October 2008, 16:40

Quote:

Originally Posted by Sagekilla

What happens when you have a completely flat frame, i.e. one color? It could have a QP of 0 and take up almost no bits. On the other hand, you can have a scene from parkrun that has a QP of 20 but takes up tons of bits.

The x264 has given somehow the 0 and 24 QP based only on the .stats and the target bitrate I specified in the command line. How are those computations will be altered if I specify a diffent bitrate and run again just the second pass using the same .stat ? Obviously the x264 will give different QP to those frames and my question is: can I predict it ?

Quote:

Originally Posted by akupenguin

Bitrate is approximately proportional to 2^(-qp/6)
Yes you'll get a slightly more accurate estimate if you run qscale2bits on each frame, which scales mvs and residuals by different amounts.
But if you want a target quality, use CRF and not 2pass.

Thanks I will try both !
about CRF specially if you use AQ or psy is not very accurate. The final QP varies several units on different videos I encoded with the same CRF value.

akupenguin · 22nd October 2008, 16:47

So you don't want constant quality, but really constant QP (which is just some number used internally by the encoder and not very related to quality). Why?!

ncahammer · 22nd October 2008, 17:21

I do want constant quality but also I do not trust so much the CRF mode with AQ and psy.
Also QP is the only number that x264 reports (in the log), so I was wondering if I could predict it.
If it had an average RF value (quality factor) I could use that, but it is not available
Unless you implying I should use SSIM instead of QP, but again SSIM also varies on different videos encoded with the same CRF.

Dark Shikari · 22nd October 2008, 17:36

Quote:

Originally Posted by ncahammer

I do want constant quality but also I do not trust so much the CRF mode with AQ and psy.

Then why do you trust the quantizer, which is even more arbitrary with regard to quality?

ncahammer · 22nd October 2008, 18:03

Quote:

Originally Posted by Dark Shikari

Then why do you trust the quantizer, which is even more arbitrary with regard to quality?

I do not trust it more. I want to predict it.
if the result of an encode doesn't match on what I observe as transparent, I want a factor than I can increase or decrease and Bitrate and QP are the only numbers I have. I was wondering if there was a formula where altering one I could predict the other

Dark Shikari · 22nd October 2008, 18:06

Quote:

Originally Posted by ncahammer

I do not trust it more. I want to predict it.
if the result of an encode doesn't match on what I observe as transparent, I want a factor than I can increase or decrease and Bitrate and QP are the only numbers I have. I was wondering if there was a formula where altering one I could predict the other

Decreasing QP by 6 doubles bitrate. Approximately. This is not always guaranteed to be true.

Sagekilla · 22nd October 2008, 21:17

ncahammer, if you do several encodes to find the highest CRF that'll produce "transparent" quality for you, you can rest assured you can reuse that CRF over many encodes and the quality will be very similar.

Generally speaking, CRF 18 should produce transparency in almost every condition (I've yet to encounter a situation where it hasn't..)

22nd October 2008, 13:38	#1 \| Link
ncahammer Registered User Join Date: Feb 2007 Location: Athens, Greece Posts: 43	x264: qp2bitrate, bit2qp ? If I encode a two pass encode, using a given bitrate A and the x264 log reports an achieved mean QP B, is some kind of function that will give me the estimated QP B' if I reencode (using the same video and parameters of course) using a bitrate A' is the opposite also predictable ? I mean if I want to achieve a QP B' what is the bitrate A' that I should give to x264 ? I searched the x264 source code and I did find in the ratecontrol.c some functions (qscale2bits,qp2qscale) but I am not sure that I am in the correct path, I suspect that I need to read and reevaluate each frame from the .stats file

22nd October 2008, 15:37	#2 \| Link
Sagekilla x264aholic Join Date: Jul 2007 Location: New York Posts: 1,752	The simple fact is you can't determine how many bits will be given for a certain QP and the number of bits to achieve a certain QP. What happens when you have a completely flat frame, i.e. one color? It could have a QP of 0 and take up almost no bits. On the other hand, you can have a scene from parkrun that has a QP of 20 but takes up tons of bits. __________________ You can't call your encoding speed slow until you start measuring in seconds per frame.

22nd October 2008, 21:17	#10 \| Link
Sagekilla x264aholic Join Date: Jul 2007 Location: New York Posts: 1,752	ncahammer, if you do several encodes to find the highest CRF that'll produce "transparent" quality for you, you can rest assured you can reuse that CRF over many encodes and the quality will be very similar. Generally speaking, CRF 18 should produce transparency in almost every condition (I've yet to encounter a situation where it hasn't..) __________________ You can't call your encoding speed slow until you start measuring in seconds per frame.

22nd October 2008, 15:59	#3 \| Link
akupenguin x264 developer Join Date: Sep 2004 Posts: 2,392	Bitrate is approximately proportional to 2^(-qp/6) Yes you'll get a slightly more accurate estimate if you run qscale2bits on each frame, which scales mvs and residuals by different amounts. But if you want a target quality, use CRF and not 2pass.

22nd October 2008, 16:47	#5 \| Link
akupenguin x264 developer Join Date: Sep 2004 Posts: 2,392	So you don't want constant quality, but really constant QP (which is just some number used internally by the encoder and not very related to quality). Why?!

22nd October 2008, 17:21	#6 \| Link
ncahammer Registered User Join Date: Feb 2007 Location: Athens, Greece Posts: 43	I do want constant quality but also I do not trust so much the CRF mode with AQ and psy. Also QP is the only number that x264 reports (in the log), so I was wondering if I could predict it. If it had an average RF value (quality factor) I could use that, but it is not available Unless you implying I should use SSIM instead of QP, but again SSIM also varies on different videos encoded with the same CRF.