x264: Film Grain Optimization

[Dark Shikari]

example:
with, without

fprofiled win32 build
patch (updated to work on latest revision and on 64-bit)
explanation (how it works, etc)

usage:

--fgo, see --help for more info. Requires RDO, i.e. --subme 6 or higher and --b-rdo. Obviously doesn't work well on non-grainy sources.

[Sagittaire]

You use nssd?
Come from libavcodec?

[Dark Shikari]

Originally Posted by Sagittaire

You use nssd?
Come from libavcodec?

Yes, I ported NSSD from libavcodec and drastically simplified the assembly. Other changes include some modifications to RD thresholds.

[IgorC]

Is it good idea if I try to encode with fgo enabled and trellis 2 + deblock -1:-1?

[Dark Shikari]

Originally Posted by IgorC

Is it good idea if I try to encode with fgo enabled and trellis 2 + deblock -1:-1?

It should work fine with trellis 2. I haven't experimented much, so feel free to do whatever you want and see how well it works.

[Adub]

So, would this be used instead of your Grain Optimizer? Or in conjunction with?

[Dark Shikari]

Originally Posted by Merlin7777

So, would this be used instead of your Grain Optimizer? Or in conjunction with?

Unrelated, though I really don't think grainoptimizer is that effective nor that reliable.

[ToS_Maverick]

now that you created another mysteriously patch, would you please explain how it is supposed to work? i'm very interested, and can only trust a patch if i know what it does

[Blue_MiSfit]

<3!!!!

I've been waiting for this one for a long time!!

I was under the impression that FGM had to be supported decoder side as well?

~misfit

[Dark Shikari]

Originally Posted by ToS_Maverick

now that you created another mysteriously patch, would you please explain how it is supposed to work? i'm very interested, and can only trust a patch if i know what it does

RDO normally uses sum of squared differences as its metric. That is, it rewards modes which approximate the source the most accurately according to that metric, and which don't cost too many bits.

FGO uses NSSD instead. NSSD consists of two components:

1) SSD as before.

2) abs( Noise(Reference Block) - Noise(Coded Block) ) * Weight, where "weight" is the --fgo parameter (this parameter also controls a few other things in FGO).

The resulting score is SSD + ND (Noise Difference).

Noise is measured as follows:

If X(i,j) is a pixel in a block, the noise score for that pixel is abs(X(i,j) - X(i+1,j) - X(i,j+1) + X(i+1,j+1)). This is summed up over all pixels in the block.

In other words, NSSD rewards modes which have a similar amount of noise and approximate the source reasonably well. This noise need not be exactly like the source's noise; it merely needs to be good enough.

Originally Posted by Blue_MiSfit

I was under the impression that FGM had to be supported decoder side as well?

This isn't FGM, its FGO.

[Blue_MiSfit]

ooooo...

right

I guess I mis-read it.

I like the results, at any rate!

[mahsah]

The pictures and Binary are down... anyone got a mirror?

Also should this be used with AQ or not...?

and for that matter, how exactly does it work (in layman's terms perhaps? )

[Dark Shikari]

Originally Posted by mahsah

The pictures and Binary are down... anyone got a mirror?

Also should this be used with AQ or not...?

and for that matter, how exactly does it work (in layman's terms perhaps? )

It should be used with AQ, the server isn't down (works fine here), and a somewhat simple explanation was posted above.

In incredibly simple terms: it measures only that the amount of noise in a candidate block is the same as it should be, not that its the same noise as in the source.

[K0zi]

What CQM's would you recommend to use with FGO?
How about prestige?

[Dark Shikari]

Originally Posted by K0zi

What CQM's would you recommend to use with FGO?
How about prestige?

I have no idea; I haven't tried any.

I've never really had good experiences with Prestige.

[ToS_Maverick]

OMG Dark Shikari, you just wrote the XviD-KILLER!!!

everything that was better with xvid is GONE! with FGO i simply got the best encode of the BlackPearl-Sample. looks like 60 mb, at only 32 mb!!!

and forget about matrices, who needs matrices now?

if you don't believe it, here are some screens
Left AQ (29,4 MB), Right AQ + FGO (32,1 MB)





and the most amazing image-pair:


commandline:

 --crf 20.0 --level 4.1 --keyint 100 --min-keyint 1 --ref 3 --mixed-refs --no-fast-pskip --bframes 2 --b-pyramid --b-rdo --bime --weightb --filter -2,-2 --subme 6 --trellis 1 --analyse p8x8,b8x8,i4x4,i8x8 --8x8dct --vbv-bufsize 9781 --vbv-maxrate 29400 --threads auto --thread-input --progress --no-dct-decimate --output "output" "input" --fgo 10

[Sagittaire]

Originally Posted by Dark Shikari

--fgo, see --help for more info. Requires RDO, i.e. --subme 6 or higher and --b-rdo. Obviously doesn't work well on non-grainy sources.

The resulting score is SSD + ND (Noise Difference).

Perhaps SSD problem here (with non-grainy source certainely that ND ~ 0). Why not choose other metric like SAD/SATD if ND ~ 0 or if ND < Threshold. Why not SAD + ND or SATD + ND too.

[Dark Shikari]

Originally Posted by Sagittaire

Perhaps SSD problem here (with non-grainy source certainely that ND ~ 0). Why not choose other metric like SAD/SATD if ND ~ 0 or if ND < Threshold. Why not SAD + ND or SATD + ND too.

SAD/SATD are really measures of bit cost, not distortion.

[Inventive Software]

Wow, Dark Shikari with another wonder-patch that'll surely make it's way to SVN within a matter of weeks!

How optimised is it? And do I now have to re-encode most of my movies to fit?

[Dark Shikari]

Originally Posted by Inventive Software

Wow, Dark Shikari with another wonder-patch that'll surely make it's way to SVN within a matter of weeks!

How optimised is it? And do I now have to re-encode most of my movies to fit?

How optimized?

Its probably erring on the side of caution in the disabling of certain RD thresholds; I'm guessing I might want to reinstate them, just at a higher level than normal.

The "noise" operation is SIMD-optimized but its not nearly as optimized as it could be; it can definitely get faster.