default noise filter for GK

[len0x]

At the momemt we have TemporalSoftener (I've changed that from TemporalSmoother in order to be YV12 compatible).

Which YV12 filter should we use by default?

I saw lots of ppl using Convolution3d - what would be the filter options for three choices on saveAVS window then?

[vinks]

since there is now an option for lanczos resize, perhaps adding undot and unfilter would be a usefull addition as a noise filter(s).

[OvERaCiD23]

I like FluxSmooth about the same as C3D. FluxSmooth is a bit faster and yields slightly better compression ratios.

[Colyptic]

I like deen, it works alot like Convolution3D with at about the same speed. It handles smoke and dust very well, which is a known issue with Convolution3D and it has about 10-15% more compression than Convolution3D. Just a thought, I don't hear it mentioned by very many people but it was the winner (as far as speed/compression and personal visual quality) in the few tests I did, which included Convolution3D, FluxSmooth, STMedianFilter, SmoothHiQ (all YV12) and even added Dust (RGB24) in cause I heard seen alot of good things about it on this forum.
Tests included a few approx. 1 min clips from different movies, from a few smoke/water scenes in Swamp Thing to the moving wall in The Matrix. Deen did the best average job (in my opinion) in all the tests, if I remember right Convolution3D and, think it was FluxSmooth, maybe STMedian looked fugly on the smoke/water scenes.

There will always be different opinions, so adding a couple good filters might be the best idea. Even though they can easily be manually entered in the script, if you are getting into the filters, you should know how to add them.

I'll stop rambling now...

[bkam]

What about leaving a listbox or some radio buttons for customizable filters, with defaults as c3d, flux, deen, etc, but totally customizable. Let them rename it, and set what it adds, and have them also be governed by the check that decides whether the filter is applied before or after the resizing. I.e., I click customize, it asks for a name, I say Flux, asks for the line to put in the script, I say FluxSmooth(8,4) or some such, then from then on I can use it in preview, etc, without having to set it up again. (Have these settings, as well as the before/after resize check, be sticky). Also, it would be nice if you made deinterlacing settings stick between times you open save&encode window the way the vobsub is sticky. (I mean the program remembers your settings between each time you open the window--I'm not sure if sticky is the right terminology).

EDIT: You could even give them a textbox with multi-lines to do a sort of "profile" for scripts, in case they want to use multiple filters. Hey, maybe have a %resize% in there or something so you can show where filters go with regard to the resizing. Now that would be awesome! Or make the filter profiles checkboxes instead of radio so they could use multiple filters if they felt like it.

Or you could just take common filters and build them into the dialogue, e.g., sliders for the thresholds of flux or c3d.

[len0x]

I think we all understand that customization is good, but what I'm more interested in for 0.28 is just default filter (i.e. not even showing it's name for user)
When you don't really know what's the difference between different filters, but just want to have three options for improving compressability: little, medium, heavy noise.

Any ideas for universal solution for average user ? (YV12 of course...)

[Colyptic]

some suggestions if you are going with Convolution3D, the internal presets work pretty good.

For light noise, preset=movieHQ

Medium noise, preset=movieLQ (you'll also want to add noise with blockbuster which runs very slow, but works great) a function like

Blockbuster(method="sharpen",block_size=3,luma_offset=-2,luma_threshold=30,detail_min=1, detail_max=30)

works pretty good.

Heavy noise, you could have preset=vhsBQ and then add noise

I haven't tested much with Convolution3D besides the preset's but they seem to work pretty well when mixed with blockbuster, but its very slow, 4-5fps on XP 1800+. Someone else might have a better suggestion, just something to start with.

[bkam]

These are settings I use:

(Very) Light Noise: Convolution3d(preset="movieHQ")
Medium Noise: FluxSmooth(10,5)
Heavy Noise: Not sure, I might use Convolution3d(preset="movieLQ") and FluxSmooth(12,8) or something along those lines

Even with the last of these I think I get around 22fps on a 2200+ a little overclocked. Not too sure I mostly run them overnight and don't watch (plus stuff like deinterlacing will take a big hit out of it too).

Hope that is more helpful.

[cyberyeye]

Fluxsmooth should be a good choice, beacause there's less parameters to configure and so people could easily change them by themselves "if they would edit the avs script into notepad", especially for beginners/medium users fluxsmooth is easier to "understand".

param i'm using for dvdrip with fluxsmooth:

light: FluxSmooth(5,7)
medium: FluxSmooth(7,7)
strong: FluxSmooth(10,15)

and for C3D:

light: Convolution3d(0,2,3,2,2,2.8,0)
medium: Convolution3d(0,3,4,3,4,2.8,0)
strong: Convolution3D (1,6,8,6,8,2.8,0)

> I think "lighter" settings should be good by default:

Convolution3d(0,2,3,2,2,2.8,0) or FluxSmooth(5,5)

Of course everybody would says this is not optimum for compressibilty and that's right ! But this "light" settings of c3d or fluxsmooth used in combination with pre-processing fonction from mpeg2dec3.dll (cpu=2) should be very good for most of dvdrip, i mean when using d2v file of course !

In case of avi file opened in Gk instead of d2v file, mpeg2dec3.dll cannot be used, cause, mpeg2source is not avisource, in this case, probably, a medium settings from fluxsmooth in the best solution imho for re-editing avifile.

ps: sorry for my bad english but i'm french and from the old europe lol

[len0x]

So far:
- Convolution3D
or
- FluxSmooth

I've read lots of good words about Dust, but it's not YV12 compatible yet...

[irasnyd]

I'm gonna vote for deen. I just did a bunch of tests today with 0.28alpha9 and Divx 5.04b3.

Using deen() looked the best for me when cleaning up some really bad stuff. In very extreme noise, edeen() worked even better.

[N_F]

Originally posted by len0x
I've read lots of good words about Dust, but it's not YV12 compatible yet...

I don't know if it's really a good idea, but you could combine it with warpsharp's loadpluginex (http://forum.doom9.org/showthread.php?s=&threadid=48160). Dust can work wonders on noisy sources where Convolution is a bit to delicate IMO.

Light: Convolution3d("moviehq")
Medium: Convolution3d("movielq")
Heavy: Pixiedust() (or perhaps even stronger)

[manono]

Hi-

In the original GKnot, the light noise filter was fast, the medium was a bit slower, and the heavy filter was very slow. Why saddle yourself with a slow noise filter such as C3D when all you want is some light noise filtering? I vote for TemporalCleaner or TemporalSoften for light filtering, FluxSmooth for medium and C3D for heavy. And Dust is so slow that no matter how good, I'd never use it. And I haven't used Deen, so I can't comment on speed or quality, but I've read good things about it.

We're trying to make this easy for beginners, aren't we? The rest of us can switch around with no trouble anyway. So I agree with len0x:

I think we all understand that customization is good, but what I'm more interested in for 0.28 is just default filter (i.e. not even showing it's name for user)

[N_F]

I may remember completely wrong, but isn't temporalsoften very slow? I remember doing some tests some weeks ago and I think one of the conclusions was that C3D was faster than temporalsoften, but again, I may way out there, perhaps I'm confusing it with something else.

[Wilbert]

I agree that Dust is way to slow (I use it only for short analog caps). I also agree the conv3d(HQ) is to slow as a light version. I propose the following:

Light:
mergechroma(blur(1.3))

Median:
Convolution3d("moviehq")

Heavy:
Convolution3d("movielq")

[N_F]

I did a quick test last night:

temporalsoften(2,5,5,10,2) - 48s
fluxsmooth() - 53s
convolution3d("movielq") - 54s
temporalsoften(3,5,5,10,2) - 59s
pixiedust() - 156s


Dust is indeed a bit slow...

[Wilbert]

Could you include

Light:
mergechroma(blur(1.3))

in your test? Or is it to light?

[N_F]

I don't have the files on the harddrive anymore, so no, not for this comparision. I'll throw togethere a more thorough (sp?) test with some more options.

Never actually tried mergechroma(blur(1.3)), to be honest I think convolution3d("moviehq") is too light for anything but quite clean movies with some very small noise. I guess mergechroma(blur(1.3)) is even lighter? Anyway, I'll include it.

[len0x]

next version will have almost all settings already there (commented)
For defaults ones: mergechroma,fluxsmooth,convolution3d

[cyberyeye]

Just one word.....

Great :-)

[N_F]

Originally posted by N_F
I'll throw togethere a more thorough (sp?) test with some more options.
http://forum.doom9.org/showthread.php?s=&threadid=51181

[N_F]

Originally posted by len0x
next version will have almost all settings already there (commented)

Great! :)

For defaults ones: mergechroma,fluxsmooth,convolution3d I still say C3D is way to delicate to handle heavy noise (at least the presets "moviehq" and "movielq") and I think my test shows it (see above link). On the other hand, maybe we just see light/medium/heavy differently. For me Clerks, Seven Samurai and Pi are movies with heavy noise; something that should very rarely be required, say 1 in 200 DVDs.