Compression/PSNR improvement Matrix?

[tetsuo55]

There is a lot of discussion about this setting adds more compression than that setting and so forth.

What i(and i think a lot of others) would like to see is matrix.

I know some features are related to eachother but what it would basically boil down to is:

Function, value, compression%. PSNR%
Bframes, 1, 10%, N/A
Bframes, 10, 15%, N/A

then the same can be done for b-frames with pyramids and so forth. i don't have the hard numbers, if someone is willing to provide them i will make the matrix.

The obvious functions are(PSY functions can not be included):

Bframes (all possible combinations of b-frame options)
Ref frames (does any other setting effect this?)
Scene cut (does any other setting effect this?)
--me (does any other setting effect this?)
--subme (does any other setting effect this?)
--trellis (does any other setting effect this?)

[Dark Shikari]

But such things are heavily source dependent.

I have seen B-frames help as little as 3-4% and as much as 20-30%.

I have seen 16 refs help as little as 3-5% compared to 1 ref, and as much as 40%.

I've seen 8x8dct help 5-6% or more, and I've seen cases where its almost never used and helps <1%.

[Sharktooth]

i dont think there could be hard numbers.
it depends very much on the source.
dark shikari once given some directives but things may have changed in the meanwhile.

[tetsuo55]

What we could do is:

take 3 sources and calculate the average from them

-Popular anime
-Popular movie
-Extremely difficult source

[Sharktooth]

nope... it varies wildly even from the same source types.
the only thing that can be estimated is compression/speed ratio...

[tetsuo55]

So basically:

any INSANE setting(maxing out that options value) could lead to massive improvements of either compression or PSNR, but on the other hand it could be a total waste of time hardly improving either.

I assume all these functions are based on math and each function is supposed to have "Fixed" improvement value.

Assuming a perfect source.

bframes 16 might have a 40% increase in compressability over bframes 1
So adding one bframe adds 2,5% to the compression

Now lets say a difficult source has trouble with the bframes. i assume it would still scale the same but look like this:

bframes 16 might have a 4% increase in compressability over bframes 1
So adding one bframe adds 0,25% to the compression

[Sharktooth]

regarding insane settings, yes, but in general there is a tendency to improve compression.
regarding b-frames it's more complex and there may not be a linear or predictable scaling.

[Dark Shikari]

bframes 16 might have a 40% increase in compressability over bframes 1
So adding one bframe adds 2,5% to the compressionThat's an unbelievably misleading conclusion, in reality it would likely be the following:

1 more bframe = 20% improvement
1 more = 10% improvement
1 more = 6% improvement
1 more = 3% improvement
all the rest more = 1% improvement

[tetsuo55]

That's an unbelievably misleading conclusion, in reality it would likely be the following:

1 more bframe = 20% improvement
1 more = 10% improvement
1 more = 6% improvement
1 more = 3% improvement
all the rest more = 1% improvement

It was just a random example of what it could look like.

i need the kind of information you just posted to make it accurate, thanks!

[Sharktooth]

those info are not accurate... they're more a trend... it was an example like yours but more realistic.

[CruNcher]

Tough imho B-frames are still a two sided sword they aren't as bad anymore as in ASP but still to much off them can have a big visual impact especialy with very fine source like Film, and all my tests show a strange thing.
Useing 1 B-frame and highering CRF is PSNR wise allways better then useing 3-Bframes and lower CRF reaching same compression in the end. And the amount of Blur over the time B-frames are couseing is perception wise much better with 1-bframe then when you have 3-bframes and the blur period is longer.
The only Drawback the 1-bframe method has it's slower @ Encoding but imho perceptionwise much better on the other side it's much faster than useing --b-rdo with 3-bframes and the same percepted result.

[tetsuo55]

I think i misunderstand what b-frames do

the way i understand it B-frames and ref frames only aid compression and not image quality at the same bitrate

--

Assuming above is true.

Would a fixed number of b-frames(completely disabled smart placement) always result in the same compression ratio in a perfect source?

[ggf31416]

each function is supposed to have "Fixed" improvement value.


The only option in any lossy compressor that have a "fixed improvement value" over compression ratio is reducing bitrate. All other options depend on the source. Also there is no such thing as a perfect source.

[Sharktooth]

tetsuo: b frames help image quality at a fixed bitrate... since B frames take less space than I and P frames, so at the end you will end up with more bits for I and P frames...
however, read what gg31416... that means there's no fixed matrix or whatsoever...