Looking for a fast denoiser with good compressibililty.

[vigi_lante]

Speed and compressibililty is more important for me than anything else.

Since I will use it with true-interlaced sources, it's very important that the filter also support interlaced.

Thanks.

[Didée]

FluxSmoothT()

[vigi_lante]

Looks like FluxSmoothT dont have interlaced support...

So, using this code would give me the same results... ?

function ApplyInterlacedFilter(clip v1, string filter) {

v2 = separatefields(v1)
selecteven(v2)
even = Eval(filter)
selectodd(v2)
odd = Eval(filter)
interleave(even,odd)
return weave()

}

ApplyInterlacedFilter("FluxSmoothT()")


Maybe it would be usefull to mention what kind of source I'm working...

It's recorded teleclasses (dvb digital captured - low bitrate: about 2300 kbps ) So, most of the time it's basically a teacher, in front of a whiteboard.

I'm also testing degrainmedian. It supports interlaced video, not as fast as FluxSmoothT(), but gave me better compression.

[Didée]

FluxSmoothT is a purely temporal filter, hence doesn't need to explicitely "support" interlaced footage. Result with or without the wrapper you posted would be exactly the same. (But of course it's faster without the wrapper).

FluxSmoothST is spatio-temporal, and would require such a "un-weave interlaced fields" wrapper.

If it's a clip with so little motion, compressability probably could be improved way more than DeGrainMedian alone can do. But then things will get slower, too ... more filters, more processing time.
Where your personal sweetspot for (speed)/(compressibility) ratio actually is located, we can't tell.

[frodeste]

How does FluxSmoothT compare with undot(), the standard denoiser used by MeGUI/Avisynth script creator?

[Blue_MiSfit]

I'm almost positive you mean UnDot(), which is really really fast. It's modern SSE enhanced version RemoveGrain(mode=1) is even faster.

How fast do you need geez... :)

~MiSfit

[vigi_lante]

Using degrainmedian() + FluxSmoothT() is actually faster than using degrainmedian alone. I dont know if this is a good filter combination, but I guess it would be fine to combine two different filters as long as both are faster.

[Chainmax]

Didée, how would you say FluxSmoothST() compares to DeGrainMedian() and FFT3DFilter(sigma=1,bw=32,bh=32,bt=3,ow=16,oh=16) in terms of denoising strength and speed?

[Pookie]

Tested 300 Frames @720x400 compressing to Lagarith codec

FluxSmoothT() - 42 seconds - 40,008K
http://fileserver1.jpghosting.com/images/tn_fluxsmootht_1c7f916d20a834a52699c829d1eb6350.png (http://fileserver1.jpghosting.com/images/fluxsmootht_1c7f916d20a834a52699c829d1eb6350.png)
FFT3dfilter(sigma=1.5,degrid=1) - 61 seconds - 34,322K
http://fileserver1.jpghosting.com/images/tn_fft_ee2707b8fe0bbe4d696cf837a1982ac2.png (http://fileserver1.jpghosting.com/images/fft_ee2707b8fe0bbe4d696cf837a1982ac2.png)
DegrainMedian(mode=1) - 39 seconds - 40,710K
http://fileserver1.jpghosting.com/images/tn_degrainmedian_e283d9ff35e740597173644660b4936b.png (http://fileserver1.jpghosting.com/images/degrainmedian_e283d9ff35e740597173644660b4936b.png)

All looked good - I liked the results of fft3d the most (kinder to the vertical blinds in the back), but there's the performance penalty.

[Prettz]

FluxSmoothST() is one of the most versatile denoisers out there in my experience, and certainly one that is useful on the largest number of different kinds of sources, which is remarkable because of how simple it is. It almost always tends to give a result with much less tendency to produce ugly artifacts after encoding than other general denoisers (like say Convolution3D, which always tends to turn noise into swimming blobs when I use it). And it's really, really fast, too.