HAvsFunc

[tormento]

Originally Posted by Sharc

Difficult to advise when we don't know your source.....

Mostly old reels with lot of grain, the type of movies that drive you crazy to denoise.

[Sharc]

Originally Posted by tormento

Mostly old reels with lot of grain, the type of movies that drive you crazy to denoise.

KNLMeansCL? dfttest? MDegrain? TemporalDegrain2(degrainTR=8)? Any of these possibly followed by a pure temporal denoiser ... Have you tried all these?
Much depends on the particular noise/grain of the source. There is no one fits all filter.

[Boulder]

KNLMeansCL could be one to test. Maybe add some motion compensation with MVTools.

[tormento]

Originally Posted by Boulder

KNLMeansCL could be one to test. Maybe add some motion compensation with MVTools.

I.e. SMDegrain that I correctly use.

On some reels I have to use tr>3 and it’s where VS stops being of any use.

[tormento]

Originally Posted by Boulder

I'd say that any details will be long gone with radius 3 at strong settings anyway so you could do with something faster.

The higher tr is, the better the results, mostly on fast speed scenes. Strong settings is all another matter. On really old BW movies I have to use tr=9 with lot less detail loss than tr=3.

[Sharc]

Originally Posted by tormento

I.e. SMDegrain that I correctly use.

On some reels I have to use tr>3 and it’s where VS stops being of any use.

Try Avisynth TemporalDegrain2(degrainTR=xx). xx=16 for example. It's slow though.

[Boulder]

Originally Posted by tormento

I.e. SMDegrain that I correctly use.

On some reels I have to use tr>3 and it’s where VS stops being of any use.

SMDegrain uses MDegrain to denoise as far as I know, a very different method compared to KNLMeans.

It would be interesting to have sample clips of before and after filtering of one severe case where a high temporal radius is needed. I personally never go above 1 when denoising, but then again, I prefer keeping things as close to the original as it's possible.

[tormento]

Originally Posted by Boulder

SMDegrain uses MDegrain to denoise as far as I know, a very different method compared to KNLMeans.

KNLMeans is applied as prefilter.

Again: higher tr means more precise results as motion vectors.

High thSAD is where details start be lost.

[Sharc]

Maybe tormento's original has very low details?
For VHS tapes a higher TR can be beneficial.

[tormento]

Originally Posted by Sharc

Maybe tormento's original has very low details?
For VHS tapes a higher TR can be beneficial.

Old BW movies have way lot more details than anything produced now.

On very high quality scans, i.e. 4k or more, you can easily see the single silver particle grain.

Anyway we are going OT. All i need is a VS MVTools with support for tr>3

[Boulder]

Originally Posted by tormento

KNLMeans is applied as prefilter.

Only affects the detection of motion, indirectly to the results.

Again: higher tr means more precise results as motion vectors.

Now there is a dilemma we all often have to deal with. How do you know the motion vector predicted to for example 10 frames in the future is any good? Often the detection is difficult already for 2-3 frames from the current frame.

[tormento]

Originally Posted by Boulder

Now there is a dilemma we all often have to deal with. How do you know the motion vector predicted to for example 10 frames in the future is any good? Often the detection is difficult already for 2-3 frames from the current frame.

Unless scene change (and it is the main limit of MVTools), object mostly tend to move in a way that the more frames you sample, the best interpolation you can get. I don't know exactly the MV algorithm but at least on math, to have good interpolation and extrapolation, you need data, the more, the better.

[ChaosKing]

Also every mvtools based deboiser can't really denoise single frames with sudden changes like flashes, explosions etc (typically found in action animes) becouse it does not have any reference frames to work with.

[tormento]

Originally Posted by ChaosKing

Also every mvtools based demoiser can't really denoise single frames with sudden changes like flashes, explosions etc (typically found in action animes) becouse it does not have any reference frames to work with.

Nothing last one frame only, unless it is flash from a photo. The more frames you can sample, the better results. With very grainy noise, it’s absolutely impossible to denoise a fast moving object with tr=3. You could have to rise until 9 on some material. I did many many times. I have a collection of almost 3 thousand DVD and BD, plus digital transfers of reels. I am not rich, I simply collect things people want to throw. I know what I am talking about.

[Boulder]

By the way, changing SMDegrain to support tr 1-6 should not be very hard. If you take a look at MCTemporalDenoise inside havsfunc, you can see how it is done there. I didn't check the function thoroughly, but there are a few things that need to be done, basically copying the functionality which is already done for deltas 1-3 and then do the degraining with weighted merging.

[feisty2]

Originally Posted by Boulder

By the way, changing SMDegrain to support tr 1-6 should not be very hard. If you take a look at MCTemporalDenoise inside havsfunc, you can see how it is done there. I didn't check the function thoroughly, but there are a few things that need to be done, basically copying the functionality which is already done for deltas 1-3 and then do the degraining with weighted merging.

the result is different from native MDegrain(tr=6), the mdegrain algorithm is here, do the math.

[Boulder]

Originally Posted by feisty2

the result is different from native MDegrain(tr=6), the mdegrain algorithm is here, do the math.

Too bad I suck at maths. What is the actual real-world difference? Just a different output but impossible to tell which one looks better or something that can be pointed out as a real flaw in the workaround?

[feisty2]

Originally Posted by Boulder

Too bad I suck at maths. What is the actual real-world difference? Just a different output but impossible to tell which one looks better or something that can be pointed out as a real flaw in the workaround?

the difference is pretty notable, the home-assembled version in MCTD is usually notably blurrier. there's no better or worse, just two completely different algorithms

[Boulder]

I don't think tormento would mind blurring And the limit parameter could be used to control it quite well as the default seems to be 255..

[feisty2]

this is insane, just grab dubhater's code and do some copy-paste and you got a legit MDegrain6 in less than 5 minutes, why bother with all that homegrown mess.