Suggest me a denoiser that...

[FlimsyFeet]

1. Works in YUY2 colourspace.
2. Works in the temporal domain but does a light denoising only, prseserving some of the existing noise, and more importantly most of the existing detail.
3. Minimal motion blurring/ghosting artefacts.
4. No too slow!

Thinking TTempSmooth or FFT3DFilter?

[Nikos]

FluxSmoothT(2).

[fibbingbear]

I don't think FFT3DFilter is going to satisfy requirement #4 :-p

[Adub]

Fine then. FFT3DGPU.

[nurbs]

What would be good sigma values for FFT3DGPU if I only want light denoising? I am planning on experimenting a bit and some starting points would help. Is it true that the the higher sigmas should be lower (sigma1 > ... > sigma4) because they tend to destroy more detail?

[Sagekilla]

sigma= high frequency detail
sigma2= medium frequency detail
sigma3= low frequency detail
sigma4= very low frequency detail


high frequency detail is stuff that's very fine detail, like individual hairs. Low frequency is stuff like a blob of a gradient. You'll probably want to use something like sigma=1 for general "light" denoising. You can tweak sigma2,3,4 if necessary but you probably won't have to.

Edit: the low sigmas correspond to what we consider "detail," the very fine stuff like hairs. So yes, you can set this lower to prevent excessive detail killing.

[halsboss]

Convolution3d ?

[g-force]

source = last
sigma = 2 #play around with this value
filt = source.fft3dfilter(sigma=sigma,sigma2=sigma*.75,sigma3=sigma*.5,sigma4=sigma*.25,sharpen=.3)
filt.Repair(source,9)

-G

[Sagekilla]

If you have a fast GPU, you can try FFT3DGPU()

[nurbs]

Thanks for the sugesstions Sagekilla and g-force.