A comparison of AVIsynth denoise filters

[hello_hello]

This isn't anywhere near being a definitive comparison and it was only done use the default settings for each denoiser. I have no doubt some of the denoisers could do a better job if I fiddled with their settings but I'm far from being an expert there so I just used the defaults.

The main purpose for doing the test encodes was to compare other denoise filters to the filter I've been using for denoising despite the fact it's not officially a denoiser (QTGMC in progressive mode). I was going to post screenshots of each test encode but they really don't tell the full story so I uploaded small samples instead.

In my opinion, QTGMC does a better job (or at least more pleasing to my eye) than the denoisers even though it's not officially a noise removal filter as such. It didn't necessarily remove the most noise, but it does "stabilise" what it doesn't remove so it no longer seems as "noisy" and therefore it doesn't blur the picture as some noise filters do. In fact QTGMC can have a slight sharpening effect instead. The two script based denoisers I tested did a pretty good job removing a large amount of noise without too much blurring, but for me, QTGMC's output is still more pleasant to look at.

I've included the original video with the samples. If anyone would care to try denoising it with clever filter parameters or a different filter please feel free. Being fairly lazy myself though, I'd prefer to use a "one size fits all" denoiser which I know will do a pretty good job without my needing to spend time making adjustments every time I use it.

Here's my take (alphabetical order). Feel free to tell me if you disagree.

degrainmedian: Seems like a decent compromise between noise removal and blurring, as long as you're prepared to overlook the first frame which it appears not to denoise.

dfttest: Removed a fair percentage of noise while at the same time blurring too much for my taste.

FastDegrain script: Subjectively, I liked the result. It removed a great deal of noise and while it did blur a little, for some reason I didn't find it too objectionable.

fft3dfilter: Caused as much blurring as a couple of other filters without removing as much noise.

FluxSmooth: Seems pretty similar to degrainmedian. Why don't they denoise the first frame?

RemoveGrain: Doesn't seem to do a great deal of noise removal using default settings.

TemporalDegrain script: Removes a large amount of grain/noise using default settings. Considering how effective it is the amount of blurring is acceptable, although it probably blurs a little too much for my taste.

TTempSmooth: Default settings don't seem to do a great deal of noise removal.

VagueDenoiser: Probably belongs in the "lightweight denoiser" category using it's default settings.

VerticalCleaner: Probably belongs in the "lightweight denoiser" category using it's default settings.

Denoisers which don't denoise the first frame would be an issue for me. Now and then I use a denoiser in combination with Trim() to denoise just the sections which need it rather than denoise the entire video.
So far the denoisers I prefer are the two script based ones. FastDegrain and TemporalDegrain. Even though FastDegrain doesn't remove quite as much noise as TemporalDegrain I prefer it's denoising because it doesn't blur quite as much.

QTGMC doesn't actually remove as much noise as the previous scripts, however to me, "noise stabilisation" is generally a more pleasing result than "noise removal" because it blurs the video less. It mightn't be to everyone's taste, but I prefer it so far. QTGMC encodes included in the attached zip file:

QTGMC (default settings): Very slow in single threaded mode.
QTGMC (medium speed preset): Acceptable speed in single threaded mode at 720p.
QTGMC (fast speed preset): On a par with many of the other filters speed-wise.

The faster the QTGMC preset, the less noise removal/stabilisation it does, but I kind of like that, especially when the source video only contains "light noise'. I've deliberately used a faster preset on occasion because I've thought it produced a more "natural" looking result (to me). I guess it's similar to the traditional noise removal/blurring compromise.

Last of all QTGMC with it's "internal" noise filtering enabled (I used it's EZDenoise function). By default, QTGMC's noise removal is disabled.

QTGMC(InputType=1, Preset="fast", EZDenoise=2.0)

Anyway, if anyone would care to look at the samples and offer an alternative opinion I'd be interested. Likewise if anyone would care to re-encode the original sample using their favourite denoise filter/settings and think it does a much better job I'd be keen to have a look.
Keep in mind all of the above is simply my opinion and as a general rule I'm fairly confident I have no idea what I'm talking about.

Edit 02 June 2014:
Well I came back to this thread by accident nearly a year after creating it, and discovered my samples are still being downloaded occasionally. Not too many other posters offering an option though.....
A year later and I'm still using QTGMC for noise removal regularly. Initially I preferred to stabilise the noise rather then remove it but my taste has changed a bit and now I prefer to remove as much noise as possible without noticeable blurring (or very little blurring) and then stabilise any noise remaining. As a result I find myself using QTGMC in a progressive mode like this quite a bit:

QTGMC(InputType=1, Preset="slow", EZDenoise=2.0)

I tend to adjust the speed preset according to the resolution (a faster preset for HD so encoding doesn't take forever). For very light noise I won't add the EzDenoise option. For more noise I add it and increase it as required.

Anyone else tried QTGMC for noise removal?

[sanlyn]

I was in a rush at the time I saw this and didn't thank you for working up this post and the attachment. Most enlightening. Keeping the original in the zip package was something others forget to do with comparisons like this. Thanks again.

[hello_hello]

Thanks for the thanks. I thought this thread had disappeared. So what was your opinion of the results? I guess it's all very subjective. Sometimes I look at yesterday's "perfect filtering and encoding job" and wonder what the hell I was thinking......

I've run some more comparison encoding using various sources since the original post and for stuff I care about I keep going back to QTGMC. For most noise removal I generally go with QTGMC's medium preset and EzDenoise=2.0 as it seems to be a good compromise between removing noise and blurring.
Once you start cranking up the EzDenoise value the noise removal results can be fairly similar to the better denoisers (using their default settings, at least), such as the TemporalDegrain script or dfttest, but for low quality video (ie some old Xvid/AVIs I recently re-encoded), QTGMC seems to do a better job when it comes to also stabilising the picture.... fixing de-interlacing artefacts and "compression wobble" etc.

I've only just started playing around with the denoising presets (they're independent of QTGMC's own speed presets). The slower ones use dfttest for denoising instead of fft3dfilter. So far I'm not convinced the difference is enough to justify the slowdown in encoding speed (at least using my old PC) but I'll need to play around a bit more.

[sanlyn]

Thought it was helpful in comparing the type and amount of the noise in the original with various results. Most of the filters attacked the grain to varying degrees, but I noted some vague gray horizontal bars in the background that seemed to disappear entirely with TemporalDeGrain but were still visible with some other filters. Some of them address chroma noise with the salt-and-pepper grain, some don't. It's been a while since I watched and downloaded the samples. I'll be looking at them again, of course, comparing results and then looking into the parameters of some of those plugins to see what they're doing. For example, everything QTGMC and some of the others do is somewhere in the avsi scripts.

I don't think it's a matter of which is the "best", worst, etc., but rather what kind of noise/grain does each handle well, to what extent, with what possible side effects, and which compromise is easier to live with. I thought the original clip offered enough "disturbances" to serve as a guide to which denoiser(s) to use.

[hello_hello]

You're no doubt correct about there being no one size fits all noise filter. I guess my logic behind using QTGMC in progressive mode, is without noise filtering it stabilises the picture (which is a good starting point), and for light noise that's often enough. Then for noisier sources it's just a matter of enabling QTGMC's denoising and increasing it as much as required.

I've attached some samples of a video I was playing around with today. Sample 2 is QTGMC in progressive mode with no noise filtering. I'm not saying it looks particularly pretty (the original video is a fair example of horrible), but you can see the picture is more stable and an improvement on no filtering at all. I prefer the result to the two samples which follow (dfttest and fft3dfilter on their own).

Next step was to enable QTGMC's noise filtering (I'm still just using the default noise filter speed preset) and slowly cranking it up while comparing it to the other noise filters. I didn't bother with many samples of "standalone" filters. The results weren't great.

So then it came down to comparing QTGMC with the FastDegrain and TemporalDegrain scripts. To me, QTGMC with EzDenoise=4.0 produced a similar noise removal result to FastDegrain while stabalizing the picture more, and QTGMC with EzDenoise=6.0 was very similar to TemporalDegrain. I had trouble picking the winner, although I tend to favour QTGMC with EzDenoise=4.0 (more noise removal tends to start blurring too much for me). TemporalDegrain seemed to do the best job of noise removal at the cost of a little extra blurring, while QTGMC fixed much more of the "bobbing", which I assume was caused by the original de-interlacing (hopefully that's the correct terminology). I guess you can't have everything......

Anyway, this is no doubt a more extreme example, but as a result it might illustrate my motivation for using QTGMC a little better. I'd be interested to hear what you think.

[AEN007]

8Dezember2013
There is Scintilla's Guide to AVISynth Postprocessing Filters ...
RemoveGrain has (for me ... for whatever that is worth ...) REALLY "cleaned up"
some grainy/blocky YouTube downloads ... The output avi(s) have been "better" than the input mp4/flv files ...

There is also a very powerful SmoothD2 filter ...

[hello_hello]

I'd be interested for you to try RemoveGrain on the original video in the comparison encodes I attached to post #5.

There's a few denoisers in the guide you linked to which I haven't tried yet, so I'll give them a spin.

Cheers.

[tyee]

hello_hello
I just tried qtgmc and I agree with you, it does a great job. I just tried a few of my others and I also like "mdegrainN" here. It's very close to qtgmc but qtgmc looks just a little bit better but mdegrainN gets faster than real time (28fps) on my i7 4.3GHz while qtgmc gets only 18fps so there is a tradeoff.

[hello_hello]

Yeah QTGMC is slow, especially if you're like me and don't use an MT version of AVIsynth.
I havn't tried mdegrainN myself but I will shortly, given I have the Dither plugin installed (I assume it's part of that package given your link).

Any tips on usage?

[tyee]

This is what I'm using here. Seems to work good generally with lots of sources.

PS, I updated that mdegrain link in the above post to the most recent version. It's also linked in the link in this post too.

Hand in hand with denoising, I use sharpening too. In the last few weeks I have been going thru many sharpeners also and have come to a conclusion. It seems that depending on when a movie was made will dictate what sharpener to use. I've been into some Clint Eastwood spaghetti westerns and have found that for movies made in the 1960's the best sharpener to my eyes is asharp. For movies made in the 80's, I like unfilter, for current movies I like finesharpen which really brings out fine detail. Must be due to the film type used over the years and the grain structure I guess.

What's your favorite sharpener?

[hello_hello]

I've used asharp on the odd occasion but I've never been keen on sharpening so I've rarely used any. I don't think I've ever tried unfilter or finesharpen. Maybe I should to see if I'm missing out on anything.

I've experimented with SeeSaw a bit and as far as sharpeners go it seems quite good. Probably because it sharpens in combination with denoising and you can use whichever denoiser you prefer. If memory serves me correctly (it's been a while since I've played around with it), the denoiser used has a fair effect on the way it sharpens so the right combination can work quite well.

Now you've got me thinking about it, I might try SeeSaw again myself at some stage.

I'll give mdegrainN a spin using your settings later on (might have to be tomorrow).

[tyee]

Yes, seesaw is good a lot of the time. Actually I tried it last night again using fluxsmooth as denoiser and it was very good but I still like QTGMC better. A year or so ago I tried seesaw with a quite noisy source and it didn't work out very well. I guess my denoiser wasn't good enough. Used degrainmedian back then.

I re-encoded "The Wild Bunch" last night with QTGMC and I really like it. Used unfilter for sharpener and it looks even better than with MdegrainN. Slightly that is but noticeable. I also added spresso to the end of the script for a test segment and it really dropped the bitrate. Size was 100MB without spresso and 70MB with. Big difference, and that's just for a 5 minute segment. There was just a noticeable loss of sharpness, hardly any but I could see it. Personally I like sharpness up to a point which I think is more than a lot of people but hey it's easy to adjust for any taste.

[hello_hello]

I was re-encoding some anamorphic 720p video (960x720) today to convert it to square pixels, so I thought I'd have a play with some more filtering. The videos were fairly clean with just some very light noise, so I didn't bother with QTGMC.

mdegrainN definitely does a nice job with light noise. I'll be trying it again soon when QTGMC is way more than required.

As we'd been discussing sharpening filters I thought I'd give SeeSaw another spin and quite liked the results. For me, the video has to be very clean before I'll sharpen it as I hate sharpened noise, so I tried SeeSaw in combination with fft3dfilter doing the denoising. I'll admit I was quite happy with the result in this case. It mightn't be my thing every time, but for the video I was encoding today the sharpening looked quite natural and I could reduce the resolution to 960x540, yet the encoded version still looked like it had more detail than the original. I'll have to spend some more time experimenting with SeeSaw too.

[tyee]

Did you see the script Didee posted that sharpens only the video portion of a source without touching the noise? I'll try to find the post.

Here it is -> Request - Sharpening Filter.


It's pretty cool.

[hello_hello]

Cheers, I'll give it a spin.

[StrobeLightEpileptic]

Just a little "noise" from me ;o) The problem with the video samples you post is that the video is pretty clean to begin with. I watched all the samples and could find no difference in any of them.

I wish you would post some really problematic video and show your results from that.

I'm about to embark on a project that is going to use some denoising. I did this once before using VirtualDub and NeatVideo. The result was fantastic! But I was watching on a cathode ray tube television. When I watched my output on a flat screen the result looked horribly blurry.

I'm a little more sophisticated now with avisynth, so I'm hoping to find something that will clean the video as well as NeatVideo did without all the blur. This time it has to pass the flat screen test.

[hello_hello]

Just a little "noise" from me ;o) The problem with the video samples you post is that the video is pretty clean to begin with. I watched all the samples and could find no difference in any of them.

Something there seems not quite right to me. The "original" sample is fairly noisy. I can see differences between them even when viewing them on my CRT PC monitor. Using my TV, there's a fair difference. How are you viewing the samples? There's nothing in your playback chain which could be removing noise? My TV has a reasonably effective noise filter and deblocker of it's own (both of which I disable as they also blur a bit).

These are screenshots from the samples in the first post. Can you see the difference between them when they're pics rather than video? "Original" vs "QTGMC, Preset fast, EzDenoise 2".




Plus the file size of the samples (same CRF value and encoder settings each time) range from 2MB to 7.2MB, which is a fair indication they don't have the same amount of noise.

I wish you would post some really problematic video and show your results from that.

Did you look at the torture test samples attached to post #5?

I'm about to embark on a project that is going to use some denoising. I did this once before using VirtualDub and NeatVideo. The result was fantastic! But I was watching on a cathode ray tube television. When I watched my output on a flat screen the result looked horribly blurry.

It's amazing what a CRT can hide.

I'm a little more sophisticated now with avisynth, so I'm hoping to find something that will clean the video as well as NeatVideo did without all the blur. This time it has to pass the flat screen test.

I find QTGMC in progressive mode (default settings) will sharpen a little without a lot of noise removal (EzDenoise less than 2). QTGMC also has noise removal speed presets although I've not played with them much myself.
Unfortunately I'm yet to find the free lunch..... noise removal = blurring..... and it's always a matter of finding the best ratio between the two. QTGMC's EzDenoise with values of 2 or more will definitely start to blurr a little, so I tend not to increase it much more than that unless a video is very noisy.
The problem with noise too, is it can look like picture detail, so removing it can give the impression the noise removal has blurred more than it has..... sometimes.

Anyway...... currently my "go to" denoising when the picture also needs stabilising is:

QTGMC(InputType=1, EZDenoise=2.0)

When the picture is already very stable (no compression wobble, blocking, bad de-interlacing etc) but there's film grain which needs removing, I've been using the script tyee linked to earlier in the thread. It's also quite fast (at least compared to QTGMC). It requires the dither package.

tr = 1 # Temporal radius
mt = true # Internal multithreading
lsb = false # 16-bit
thSAD = 200 # denoising strength
blksize = 16 # block size
overlap = 4 # block overlap
super = MSuper (mt=true)
multi_vec = MAnalyse (super, mt=true, multi=true, blksize=blksize, overlap=overlap, delta=tr)
MDegrainN (super, multi_vec, tr, mt=true, lsb=lsb, thSAD=thSAD, thSAD2=150)

And speaking of the dither package..... if I'm using any denoising I've pretty much learned my lesson when it comes to colour banding and I always add dithering to the end of each script:

gradfun3()

Someone introduced me to the MCTemporalDenoise script a little while back. You might want to take it for a test drive. Lots of experienced Avisynth users seem to swear by it. I'm yet to play around with it much myself but the couple of test encodes I ran indicate it's default settings tend to sharpen..... if anything, a bit too much for my taste (I'm not a fan of sharpening) but it has lots of settings which can be adjusted. I should play around with it more myself at some stage.......
Other than that the TemporalDegrain script removes a lot of noise (although it's default settings do blur a bit).

If you discover any exciting denoise options, please post back.

[killerteengohan]

For heavier noise I use

Deen("c3d", 0, 3, 2) up to Deen("c3d", 0, 7, 2)


For lighter noise removal I use

Convolution3D(0, 3, 4, 3, 2, 2.8, 0)

[Selur]

as a side note: in general combining a 'light' denoiser and a grain handler like removegrain(2) or x264s own -nr normally is a good way for denoising without too much detail loss.
To add some additional clips, I used

x264 --preset slow --crf 18

and created files for:

• PeachSmooth
• Deathray
• Deen
• fft3dgpu

all using their default settings

[hello_hello]

Wow..... this thread's back.

I opened the first video in the folder on my PC containing the original test encodes from the first post, ran MPC-HC full screen on my TV and stepped through them one at a time. Aside from them being in alphabetical order I didn't know which one was playing.
I paused on FastDegrain, which surprised me. After all this time I still liked the result, yet I don't think I've used the script for denoising since. I stopped when I got to "QTGMC defaults" because it stood out to me as different. The noise was suddenly a lot more stable. Maybe it's all just personal taste. I think "QTGMC defaults and EzDenoise=2" would have done the job nicely. I'll have to try it later.

Of your encodes Selur, I think I prefer "Deen". I'll have to experiment with it a little myself. I'm not sure I've ever used it.

I'm not sure I've ever tried Convolution3D either killerteengohan. Tomorrow when I've got some time I'll have a play.

I might even dig out a different video to play with and upload some different samples. The TemporalDegrain script..... I can't quite get my head around that one. A couple of times I've used it in preference to QTGMC for denoising as it's removed a similar amount of noise while blurring less, but as was the case with the sample from the original post, sometimes it blurs too much. Maybe I haven't found the magic settings yet......

I see nobody has attempted noise removal on the sample from post #5 yet. Or at least nobody has attempted it and posted the result yet......

[Selur]

I see nobody has attempted noise removal on the sample from post #5 yet. Or at least nobody has attempted it and posted the result yet......

There is no 'original' inside the .zip from post #5 and re-encoding the '1 re-encode, no filtering.mkv' seems unfair, since the whole reencoding isn't lossless.

Cu Selur

Ps.: As a side note: the encodes before weren't meant to be a solution, they are meant to show how some other denoisers handle the source using their default settings.

[hello_hello]

I see nobody has attempted noise removal on the sample from post #5 yet. Or at least nobody has attempted it and posted the result yet......

There is no 'original' inside the .zip from post #5 and re-encoding the '1 re-encode, no filtering.mkv' seems unfair, since the whole reencoding isn't lossless.

You probably have a point. I can't remember why I did that (included the "no filtering" sample as opposed to just keeping the original). It might have been to reduce the file size, or to fix the aspect ratio. I'm not sure.

[Selur]

-> if you post another sample I can create Deen&Co files for you.

[brusno]

I think that the parameters in filters are important. It can mean the difference of shooting with a tank or a sniper.

I've been using:
MPEGSource(" ", CPU=2)
DeBlock(25)
TTempSmooth()
Undot()

and a sharpener to compensate for blurring:
MSharpen(threshold=3, strength=30)
Sometimes FastLineDarken() too, but I can't remember the settings for this right now.


The dilemma has been to remove noise without removing too much details (compromise). The above filters and settings are pretty mild in noiseremoving. Some film demand stronger denoising. In animation, one can usually denoise harder without loosing critical details. Except for example "Bambi".

The parameters for FastLineDarken that I use is: FastLineDarken(strength=24,thinning=0,luma_cap=100 )

[brusno]

A comparison of original and the result of my filter configuration:


[Attachment 28924 - Click to enlarge]


[Attachment 28925 - Click to enlarge]

[brusno]

The original:
original.mkv

The filter-test:
test.mp4

I think the filters are dealing with the noise adequatlely

[rainmaverick@mail.ru]

I think that parameters for denoising and sharpening must be calculated for every scene of the film. Usually we have the different grain in different scenes (if it shooted on 35mm camera), the dark scenes especially have a more grain/noise
But we have some questions to answer:
1. How to analyze the scene cut from the film for choosing the right preset? (And save analyzing log for next usage)
2. How to create the set of presets for QTGMC based on scenes types?

Any ideas?

[hello_hello]

This thread is back! I seem to have missed the last few posts, although they were a year ago now.

rainmaverick@mail.ru,
I don't know if there's a clever way to analyse a scene from a film and choose the right preset automatically. If I'm really fussy, I'll sometimes used QTGMC followed by LSFMod(strength=50), or strength=75, although QTGMC and LSFMod can make for a pretty slow encode, and I'd only use LSFMod if all the noise is removed as I don't like sharpened noise.

Mostly, I try to pick a "it'll do for the whole video" denoising method and just go with that. There's been a few times where my denoising enthusiasm has been greater, so I've re-encoded a video twice with weaker and stronger noise filtering, then cut and appended the output files for a final version (be sure to use -stitchable in the x264 command line). You can of course create a script that denoises different scenes by different amounts but that's sometimes even more work.
More often than not I encode everything with a single noise filter setting and if when I watch the encode I come across a "could have used more noise removal" scene I'll encode just that scene again.

Back in June 2014 I edited my original post but since then my taste has probably changed a bit again. I tend to use a faster QTGMC speed preset with an appropriate level of EzDenoise, the logic being any noise left behind isn't stabilised completely so it still looks more like natural noise. If you stabilise it too much you can risk a "looking through a flyscreen" effect if too much "static" noise remains. I tend not to use higher than EzDenoise=2 unless I really have to. If I was denoising the sample from the OP today I'd probably go with:

QTGMC(InputType=1, Preset="Medium", EzDenoise=2.5)
gradfun3()

Any denoising is a line between noise removal and blurring, and I guess it depends where the line is for each person.
Following denoising with gradfun3 (from the dither package) tends to be a good idea to fight colour banding, and QTGMC with EzDenoise=1 and gradfun3() seem to do a better job of deblocking than any de blocking filter I've tried.

Looking back at the original denoised samples from post #1, I still think the FastDegrain() script did a pretty good job in this particular case, although I don't use it much. I'll have to try it again on some different sources.

"original.mkv" attached to post #26

[Srivas]

Great thread indeed. I was looking for denoisers comparison test for avisynth and this thread has lot's of valuable information.

Here is my two cents. I didn't test the above original clip but my experience with Qtgmc is that when EzDenoise is enabled, then higher values start to blur too much for my taste. So to avoid this and at the same time remove lots of noise I like to combine it with deen. That gives overall better noise removal than simply Qtgmc and good sharpening at the same time. So that is Qtgmc fast preset with EzDenoise on 2.0 and plus deen (with whatever settings you like best). I haven't really tested lots of filters but I like the results of this combination.

[hello_hello]

Srivas,
I've never been able to run two different noise filters effectively, but I'll try anything. In what order do you run deen and QTGMC and do yo use the default deen settings? There's a sample of deen denoising the original clip attached to post #19. From memory it did a reasonable job (I kept the sample but it's saved to a different PC).

At one stage I tried getting QTGMC to use dfttest for the denoising (it uses FFT3DFilter by default) but from memory it was a fair bit slower and for some reason I preferred the result with FFT3DFilter doing the denoising. I probably should revisit that at some stage.

I'm with you on the blurring thing..... I tend to try not to use anything higher than EzDenoise=2 if I can help it, and I mostly try to stick to EzDenoise=1 or EzDenoise=1.5 if possible. I'd rather leave a bit of noise behind than blur too much, although I keep telling myself for most other noise removal methods the amount of blurring would generally be a fair bit greater for an equivalent amount of noise removal. And noise can do a very good job of impersonating picture detail as far as the human brain is concerned. I've compared individual frames countless times where the noise removal has appeared to only remove noise, but when the video is playing normally the noisy version can still appear to have a little more fine detail.