TemporalDegrain2 - so slow?

[rgr]

I have a few-minute-long video recorded with a static camera. Noisy, but I was happy with Neat Video's denoising.
However, I have seen better results on such videos (where 99% of the image is still) using TemporalDegrain2 (I can't find the thread at the moment).
So I used:

ffms2("joined.mp4")
assumefps(30)
convertbits(10)
converttoyuv444(matrix="PC.709")
levels(10,1,900,0,1020,coring=false)
TemporalDegrain2(degrainTR=4,grainLevel=3,postFFT=3,postSigma=1)

And I get a "stunning" 0.01fps for 4k footage -- that's several hours for 200 frames! For comparison, Neat Video = 0.30fps.
Is there any way to speed this up? Am I doing something wrong?

[poisondeathray]

You can try adding prefetch(x) , where x is some number >1 for multithreading. But it will still be slow.

Neat video should be much faster with a decent GPU

Is 99% of the image still in the same section of the frame with no change ? If so, crop to the region of interest and process that with the denoiser, composite with a still frame

[Selur]

a. depending on your setup LWLibavVideoSoure with enabled hardware decoding might speed up the decoding of the source
b. using mulithreading (see: http://avisynth.nl/index.php/SetFilterMTMode; set SetFilterMTMode at the start and Prefetch at the end of the script)

[rgr]

You can try adding prefetch(x) , where x is some number >1 for multithreading. But it will still be slow.

Neat video should be much faster with a decent GPU

Is 99% of the image still in the same section of the frame with no change ? If so, crop to the region of interest and process that with the denoiser, composite with a still frame

Prefetch(4) gave a boost from 0.01fps to 0.05fps Well, still very slow...

[johnmeyer]

Prefetch(4) gave a boost from 0.01fps to 0.05fps Well, still very slow...

Hey, its a 5x improvement!! That's actually pretty darn good.

1. Have you tried it with the default settings?

2. Have you read all the comments in the AVSI file? For instance, it says, "currently just postFFT=0, 1 and 4 are supported for high bit depths." You are using postFFT=3. Not sure what that might do. I only skimmed the AVSI just now, and did verify that the default for DCT in MVTools2 (the one plugin I know very well) is 0. If it gets set to anything other than that, speed can go down to about 10% of what it is at DCT=0.

3. Try different Prefetch numbers. Try setting it to about 1/3 the number of cores you have. I often get negative speed results when Prefetch > 1/2 number of cores.

4. What resolution are you feeding it? Try adding a resize to half resolution in each direction as the first line in the script, and then run the script for a minute. See what sort of speed you get. I'm not suggesting that you permanently want to degrade your video, but am merely trying to help pinpoint what factors might affect speed. I have found MVTools2 to be very unforgiving when you feed it really high resolution video.

What version of MVTool2 are you using? The older the version, the more likely you are to have speed issues.

[edit]I suggest searching the long TemporalDegrain2 thread at Doom9.org. Someone there posted a much faster version called "Temporaldegrain2_fast". Look for that and see if it is faster, but without compromising the quality.

https://forum.doom9.org/showthread.php?p=1982594#post1982594

[johns0]

What computer are you using as in specs?

[rgr]

Ryzen 4650G (6 cores). Video is 4K.

[rgr]

[edit]I suggest searching the long TemporalDegrain2 thread at Doom9.org. Someone there posted a much faster version called "Temporaldegrain2_fast". Look for that and see if it is faster, but without compromising the quality.

https://forum.doom9.org/showthread.php?p=1982594#post1982594

"RadT = Default(RadT, 1) # Temporal Radius of frame analysis. Value 1 or 2"

This script seems to have major limitations where I want it to be very powerful.

[DTL2023]

It is complex script with several different denoisers.

At least vsTTempSmooth got update in performance using AVX2 in latest builds in Asd-g github in 202x years - recommended to use lastest version.

Also MAnalyse using dct > 0 (from fftw library) may run significantly slower in comparison with spatial modes (with dct=0).

If you use Win10 or later and GPU with MPEG-encoder hardware chip with DirectX12 Motion Estimation capable driver you can try mvtools of post-2.7.45 builds (typically blocksize of 8x8 only run stable). https://forum.doom9.org/showthread.php?t=183517 . See examples of hardware accelerated scripts in that thread with optSearchOption=5 or 6. It should support 4K resolution.

And for possibly much better performance with mvtools only denoise much simple scripts may be tested like EZdenoise.

[jagabo]

Prefetch(4) gave a boost from 0.01fps to 0.05fps Well, still very slow...

Why 4? You have a 6 core, 12 thread CPU. Try 6 to 12. Maybe even more.

[davexnet]

I think even a Cray supercomputer would meet its match running Temporaldegrain...

[rgr]

Prefetch(4) gave a boost from 0.01fps to 0.05fps Well, still very slow...

Why 4? You have a 6 core, 12 thread CPU. Try 6 to 12. Maybe even more.

Yes, but then it was consuming all my RAM. In the end, only Prefetch(3) didn't kill the computer. (4k)

[sophisticles]

I have a few-minute-long video recorded with a static camera. Noisy, but I was happy with Neat Video's denoising.
However, I have seen better results on such videos (where 99% of the image is still) using TemporalDegrain2 (I can't find the thread at the moment).
So I used:

ffms2("joined.mp4")
assumefps(30)
convertbits(10)
converttoyuv444(matrix="PC.709")
levels(10,1,900,0,1020,coring=false)
TemporalDegrain2(degrainTR=4,grainLevel=3,postFFT=3,postSigma=1)

And I get a "stunning" 0.01fps for 4k footage -- that's several hours for 200 frames! For comparison, Neat Video = 0.30fps.
Is there any way to speed this up? Am I doing something wrong?

Fair warning, I am going to say a bunch of "stupid" things.

Have you read the guide related to this script:

http://avisynth.nl/index.php/TemporalDegrain2

Under Suggested settings for removing the grain, sorted by increasing source grain it lists the following settings for the most grainy sources:

degrainTR=2,grainLevel=3,postFFT=1,postSigma=value between 0 and 2

Now compare to your settings:

degrainTR=4,grainLevel=3,postFFT=3,postSigma=1

I would also try to use FFT3D directly and eliminate everything else.

http://avisynth.nl/index.php/FFT3DFilter

I have always felt that people filter their stuff way too much. To me filters are like spices for food, a bit of salt, some pepper, maybe oregano or something a pinch of something else, and that's it.

But many people seem to want to add everything in the spice rack and it becomes overpowering.

Keep it as simple as you can.

The other thing is you mentioned that when you change the settings your ram usage goes way up. This is a sign that if you intend to do this type of work you need more ram.

[rgr]

The other thing is you mentioned that when you change the settings your ram usage goes way up. This is a sign that if you intend to do this type of work you need more ram.

After a few attempts with the settings, I stayed with Neat Video. I achieved better denoising using TD2, but at the cost of losing details, especially with moving parts (and I only used degrainTR, the rest at 0).