Looking for subtle purely-temporal filter

[Xodarap]

I'm looking for a filter that is purely a temporal denoiser. I don't need any (spatial) blurring. I'm using high-quality (DVD) sources, and want to minimize the "pixel scattering" effect common to DVDs prior to encoding.

TemporalSmoother blurs the image too much at (2), and, though I haven't really tried it, I'm worried that TemporalSoften(2,3,3,6,2) or especially Convolution3d will do the same.

If any of those are actually a good bet, please let me know. If it matters at all, I'm just putting together an avs profile for MeGUI's one-click thingermajigger. It'll only be for DVD VOBs.

[Morte66]

Try TTempsmooth from here: http://bengal.missouri.edu/~kes25c/ Note that it needs an explicit LoadPlugin, it doesn't autoload.

Then try a motion compensated temporal denoiser like mvdegrain2

[Dark Shikari]

My GrainOptimizer filter is 100% temporal, since its built to stabilize noise over a short period of time (a few frames) rather than remove it.

[Xodarap]

Try TTempsmooth from here: http://bengal.missouri.edu/~kes25c/ Note that it needs an explicit LoadPlugin, it doesn't autoload.

Then try a motion compensated temporal denoiser like mvdegrain2

Are you suggesting that I run one and then the other in the same script? In the past, I've had bad results with running two temporal denoisers in the same script, although these two look like they work off of different principles.

TTempSmooth says it's "motion adaptive" -- I take it that's as opposed to "motion compensated"?

I will be making two catch-all avs scripts: one for anime and one for live-action. But all input will be high-quality dvd vob sources, and all encoding will be at 1200kbps ABR, 2-pass x264, "insane quality." All I care about is subjective visual quality. I don't care about speed (within reason) or filesize. I want my live-action to be without blocks (I've come to hate the color red), and my anime to be visually stationary and smooth.

I'm going purely temporal (as opposed to 3d) because I want to preserve as much detail as possible. But the "dust" or "grain scatter" in DVD video is (1) subjectively annoying, and (2) an impediment to efficient encoding...

[Unearthly]

I have had very good results with TTempSmooth, it should be able to do what you want. About the adaptive/compensated, the readme says:

TTempSmooth is a motion adaptive (it only works on stationary parts of the picture), temporal smoothing filter.

I've used it on anime and found it removes temporal noise well without harming details much if at all.

[Xodarap]

My GrainOptimizer filter is 100% temporal, since its built to stabilize noise over a short period of time (a few frames) rather than remove it.

It currently requires running two clips and comparing, correct? You said to use a "heavy" denoiser for the denoised clip -- should I go 3d or use a heavy temporal denoiser (like TempSmooth(1,8,8,6,15)?)?

[Dark Shikari]

It currently requires running two clips and comparing, correct? You said to use a "heavy" denoiser for the denoised clip -- should I go 3d or use a heavy temporal denoiser (like TempSmooth(1,8,8,6,15)?)?Yes. The heavy denoiser's purpose is to eliminate all variation that isn't actually motion, to tell GrainOptimizer what is moving and what isn't. It also serves to tell GrainOptimizer how much grain there is--it compares the two to get a feel for roughly how much noise there is in the clip.

FFT3dfilter is fine, as is dfttest, ttempsmooth, etc.

[Xodarap]

I have had very good results with TTempSmooth, it should be able to do what you want. About the adaptive/compensated, the readme says:



I've used it on anime and found it removes temporal noise well without harming details much if at all.

That's exactly what I was quoting. But above, someone said to use TTempSmooth "and then use a motion-compensated filter..." Since I knew TTempSmooth is "motion adaptive," I assume those were opposed properties, and maybe complementary...

[Xodarap]

Yes. The heavy denoiser's purpose is to eliminate all variation that isn't actually motion, to tell GrainOptimizer what is moving and what isn't. It also serves to tell GrainOptimizer how much grain there is--it compares the two to get a feel for roughly how much noise there is in the clip.

FFT3dfilter is fine, as is dfttest, ttempsmooth, etc.

I read all about grainoptimizer and it sounds perfect for what I want! While I don't care about preserving grain as an end, I don't mind preserving it, and in theory, a filter designed to preserve grain should preserve more detail. :) I'm gonna run a bunch of tests with it over the next week. :thanks:

[Sagekilla]

Yes. The heavy denoiser's purpose is to eliminate all variation that isn't actually motion, to tell GrainOptimizer what is moving and what isn't. It also serves to tell GrainOptimizer how much grain there is--it compares the two to get a feel for roughly how much noise there is in the clip.

FFT3dfilter is fine, as is dfttest, ttempsmooth, etc.

Sounds a lot like the degraining filter (http://forum.doom9.org/showthread.php?p=1076276#post1076276) turned function (http://forum.doom9.org/showpost.php?p=1078534&postcount=116) that Didee created a while back, except that this actually removes the grain rather than stabilizing it.

Same concept: Heavily denoise a clip (Provide your own or have a built in fft3d call do it, and also an additional hqdn3d call for a bit more) to look for motion and then, instead of just stabilizing, do a temporal degrain thru MVDegrain2.