Noise filter to replace FFT3DGPU

[gelo333]

Please tell me the replacement of FFT3DGPU. I need an easy fast noise filter for Avisynth. Without perfect processing, just as universal for all files.

FFT3DGPU started to work with bugs, greatly increasing the encoding time, and the video card driver (AMD) gives an error that causes the system to freeze. I think the reason is that FFT3DGPU has not been updated for a long time, and the latest version of the video card driver is not compatible.

[pandy]

If you need something GPU accelerated then probably ffmpeg can be alternative https://ffmpeg.org/ffmpeg-filters.html#OpenCL-Video-Filters .
Filters made many years ago are mostly tailored for NVidia GPU.

[gelo333]

Not a lot of advice. Okay, I'll ask differently.
Is the KNLMeansCL filter similar to the FFT3DGPU filter? I heard that in addition to the KNLMeansCL filter, another noise filter needs to be applied.

[lollo]

Is the KNLMeansCL filter similar to the FFT3DGPU filter?

Why don't you try by yourself and post here the results of a comparison between FFT3DGPU and KNLMeansCL on your source?

[gelo333]

Is the KNLMeansCL filter similar to the FFT3DGPU filter?

Why don't you try by yourself and post here the results of a comparison between FFT3DGPU and KNLMeansCL on your source?

I tried. And I still don't see the difference. Maybe because I don't care about perfect results. But I know little about the KNLMeansCL to draw conclusions, and to take it for permanent use.

[lollo]

In term of spatial denoise is excellent and a good alternative to FFT3DGPU, but it depends on source. Here a comparison with a temporal filter like TemporalDegrain2.

The second is better, but I suspect not adequate to your "fast" requirement.

https://imgsli.com/MTc5MDk5

https://imgsli.com/MTc5MTAw

High frequency removal based filters suffer in general of detail loss.

I would switch from FFT3DGPU to KNLMeansCL if its speed is ok for your requirements. YMMV according yo your source.

[jagabo]

Is the KNLMeansCL filter similar to the FFT3DGPU filter?

Why don't you try by yourself and post here the results of a comparison between FFT3DGPU and KNLMeansCL on your source?

I tried. And I still don't see the difference.

The difference between what? The source and the filtered videos? The FFT3DGPU and KNLMeansCL videos? What settings are you using? Both of those filters can be set to do almost nothing to completely obliterating a video. What kind of noise are you looking at removing?

[gelo333]

Here a comparison with a temporal filter like TemporalDegrain2.

The second is better, but I suspect not adequate to your "fast" requirement.

Using TemporalDegrain2 on my computer takes 10 times longer than using FFT3DGPU. It was 15 seconds, now it's 2.5 minutes. This is with settings:
TemporalDegrain2(degrainTR=0, postFFT=1, postSigma=1, grainLevel=0, fftThreads=2)

Although I read that you can connect the KNLMeansCL filter to it (instead of FFT3D?), And even the FFT3DGPU itself. Maybe this will speed up the video processing time, but I need to know how to connect them.

[gelo333]

https://imgsli.com/MTc5MDk5

on your first example I saw that KNLMeansCL breaks the image. Or is this the source image?

[gelo333]

Is the KNLMeansCL filter similar to the FFT3DGPU filter?

Why don't you try by yourself and post here the results of a comparison between FFT3DGPU and KNLMeansCL on your source?

I tried. And I still don't see the difference.

The difference between what? The source and the filtered videos? The FFT3DGPU and KNLMeansCL videos? What settings are you using? Both of those filters can be set to do almost nothing to completely obliterating a video. What kind of noise are you looking at removing?

difference The FFT3DGPU and KNLMeansCL videos. My settings:
FFT3DGPU(sigma=1.5, bt=4, bw=32, bh=32, ow=16, oh=16, sharpen=0, NVPerf=False, precision=2)
On different sources, mostly blocks and aliasing. And some stains and dirt. Little grain. I can't say for sure. I don't need professional quality processing.

[lollo]

TemporalDegrain2 is slow, and not adequate for your purposes. KNLMeansCL can be added inside TD2 as further denoiser at the end, but then it will be slower!

I used TD2 (one of the best denoiser) just to show that KNLMeansCL is lower quality but not too bad compared to it.

My understanding (without looking to your source) is that you can replace FFT3DGPU with KNLMeansCL stand-alone. If you have noise to be removed across frames (temporal radius) in an accurate way both are less effective, but you have no choice because speed is a requirement and in any case you are not sensitive to small differences.

[poisondeathray]

FFT3DGPU started to work with bugs, greatly increasing the encoding time, and the video card driver (AMD) gives an error that causes the system to freeze. I think the reason is that FFT3DGPU has not been updated for a long time, and the latest version of the video card driver is not compatible.

Which version were you using ?

https://github.com/pinterf/FFT3dGPU
pinterf fork 20210104 , so just over 2 years actually more recent update than KNLMeansCL

https://github.com/pinterf/KNLMeansCL/releases
20201112

[gelo333]

FFT3DGPU started to work with bugs, greatly increasing the encoding time, and the video card driver (AMD) gives an error that causes the system to freeze. I think the reason is that FFT3DGPU has not been updated for a long time, and the latest version of the video card driver is not compatible.

Which version were you using ?

https://github.com/pinterf/FFT3dGPU
pinterf fork 20210104

i am using this latest version

[gelo333]

thanks for answers.

I used TD2 (one of the best denoiser) just to show that KNLMeansCL is lower quality but not too bad compared to it.

Why did defects appear on the lips, teeth, eyes in example-1? Is it from the source, or do they appear after KNLMeansCL?

[lollo]

Here the comparison with the source (the full flow also include some color correction, but should be not relevant);

https://imgsli.com/MTc5MTU5

judge yourself (imo the defects are already in the source, better cleaned by TD2)

[gelo333]

I understood you. There are defects in the source, but KNLMeansCL strengthens them. Maybe fix changing KNLMeansCL settings.

Last question. I saw that other noise filters are used in tandem with KNLMeansCL, probably also fast ones.

[lollo]

Last question. I saw that other noise filters are used in tandem with KNLMeansCL, probably also fast ones.

After a temporal denoise, KNLMeancCL may help to reduce residual noise intra-frame.

You can try a simple MDegrain with low temporal radius (2), and a KNL after, and check if speed stays reasonable.

[pandy]

You can give a chance to HQ3dn for fast temporal filter (but probably it is not implemented by GPU, at least i'm not aware of this... strangely there is not many filters even today implemented for modern GPU - even if there are fast GPU FFT implementations).
Seem there is at least few frameworks based around Python for fast signal processing (OpenCL) but not used in Avisynth/Vapoursynth world either.

[jagabo]

HQDN3D() (CPU only) runs about 10 times faster than FFT3DGPU (GPU) on my computer, a i9 9900K. And KNLMeansCL() is much slower than FFT3DGPU.

[rgr]

I also used HQ3DN -- praise for speed.

[pandy]

HQDN3D() (CPU only) runs about 10 times faster than FFT3DGPU (GPU) on my computer, a i9 9900K. And KNLMeansCL() is much slower than FFT3DGPU.

Yes, HQ3DN is fast on CPU but if you could implement whole signal chain on GPU avoiding cycle costly transfers then it would be way faster. but that's why i've proposed HQ3DN as temporal noise removal - it is fast without GPU.

[gelo333]

My HQ3DN breaks the image. Clearly visible on medium-dark details:
https://imgsli.com/MTc5NTYw
FFT3DGPU does not give strong interference. HQ3DN version 1.1.1, settings:
HQDN3D(ls=4.0, cs=3.0, lt=6.0, ct=4.5, restart=7)

[pandy]

My HQ3DN breaks the image. Clearly visible on medium-dark details:
https://imgsli.com/MTc5NTYw
FFT3DGPU does not give strong interference. HQ3DN version 1.1.1, settings:
HQDN3D(ls=4.0, cs=3.0, lt=6.0, ct=4.5, restart=7)

You can reduce strength, you can reduce strength and stack another (or more) iteration of hq3dn in series - it allow for finer control of filtering.
Also you can try to remove gamma correction (but probably you need to operate in 16 bit depth) - as crude workaround for gamma problem you can try to invert levels (negate) filter and re-apply negate operand - this may help with gamma (non linear) videos.

[gelo333]

Why does HQDN3D break the image? And the distortion is on default settings!

By the way, I found another high-speed filter. This is MiniDean.

[pandy]

Why does HQDN3D break the image? And the distortion is on default settings!

By the way, I found another high-speed filter. This is MiniDean.

hq3dn was created when mostly VCR captures was dominating - defaults can be minor for VCR capture but for digital capture they may be to high, side to this hq3dn is very fast so there is always trade off between speed of algorithm and it's complexity (capability to deal with difficult problems).

There are also wavelet based filters - for example http://avisynth.nl/index.php/VagueDenoiser .

This thread can be interesting for you https://forum.videohelp.com/threads/358043-A-comparison-of-AVIsynth-denoise-filters , probably there are also other interesting information's on doom9 forum.

[jagabo]

Why does HQDN3D break the image?

It's not broken, it's blurred. All noise reduction filters blur the image in some way -- it's the fundamental nature of the process. The stronger the noise reduction the more they blur. If you find the values you're using blur too much use lower values.

[gelo333]

Why does HQDN3D break the image?

It's not broken, it's blurred.

I have tried several denoisers. All of them are close to the original. Only HQDN3D changes the image a lot.

[davexnet]

LS and CS too high. Try max of 2 and 2.5 respectively

[rgr]

There are also wavelet based filters - for example http://avisynth.nl/index.php/VagueDenoiser

VagueDenoiser is single-thread so very slow.

[jagabo]

LS and CS too high. Try max of 2 and 2.5 respectively

ls=1.0 and cs=1.0 are a lot closer to the FFT3DGPU settings he gave in post #10.