Avisynth: Smootherhiq vs. Convolution3D

Jellygoose · #1 03-27-2003, 01:19 PM

Hi all...

I made some tests last night, comparing Convolution3d(preset MovieHQ) and the SmootherHiQ Plugin by Klaus Post.
I like to use FaeryDust as my Temporal Cleaner, because for me it still works best. But using FaeryDust and Convolution3d makes the picture just too blurry for my taste. Now I've come to a good solution for this. SmootherHiQ.

The filesize for my sample with SmootherHiQ was 200kb lower than the one made with C3d... and in my eyes the picture was a lot less blurry. here's my script and the settings I used for SmoothHiQ...

Code:

LoadPlugin("E:\MPEG-Tools\FitCD\MPEG2DEC.dll")
LoadPlugin("E:\MPEG-Tools\FitCD\dustv5.dll")
LoadPlugin("E:\MPEG-Tools\FitCD\legalclip.dll")
LoadPlugin("E:\MPEG-Tools\FitCD\sampler.dll")
LoadPlugin("E:\MPEG-Tools\FitCD\convolution3d.dll")
LoadPlugin("E:\MPEG-Tools\FitCD\gripfit_preview.dll")
LoadPlugin("E:\MPEG-Tools\FitCD\fluxsmooth.dll")
LoadPlugin("E:\MPEG-Tools\FitCD\dctfilter_YUY2.dll")
LoadPlugin("E:\MPEG-Tools\FitCD\smoothhiq.dll")


Mpeg2Source("F:\Genug\genug.d2v")

LegalClip()

GripCrop (544, 576, Overscan=0)

Gripsize()

FaeryDust()
smoothHIQ(3,15,25,192,10)

mergechroma(blur(1.58))
mergeluma(blur(0.05))

GripBorders()
Letterbox(0,0,20,20)

DctFilter(1,1,1,1,1,.5,.5,0) 

LegalClip()

Sampler(length=24)

ConvertToRGB24

try it out and tell me what you think... of course this is only if you don't mind the slowness of FaeryDust.

btw: SmoothHiQ is not noticable faster than C3d...

You can download the filter here:

http://cultact-server.novi.dk/kpo/av...th_hiq_as.html

PyRoMaNiA · #2 03-27-2003, 03:45 PM

Trying it now! I'll let you know my results!

Jellygoose · #3 03-29-2003, 05:26 AM

did you try it pyro?

PyRoMaNiA · #4 03-29-2003, 05:40 AM

Yes I did.

Sorry I'm a bit late though, I got distracted as ToK kept screwing up. I havent actually done a straight comparison yet, but on my results so far, I have to say I agree with you.

I think there might be a bit of an issue when using it with TemporalCleaner() as the output looks much more noisy than it should have been, and I removed TemporalCleaner() and the filesize dropped by about a megabyte in size.

After removing that the output looked very clean and sharp.

I modified your settings a bit though, and used

Code:

SmoothHiQ(5,15,25,200,10)

because I normally use LQ instead of HQ with Convolution3D, so I thought I'd be able to get away with a bit of extra smoothing with SmoothHiQ.

I'm off to encode the same clip with Convolution3D now, I'll be back in about an hour (I hope

) to let you know my results.

Jellygoose · #5 03-29-2003, 07:51 AM

Well using FaeryDust already gives you plenty of temporal cleaning I think, so there's probably no need for TemporalCleaner() in your script.
Since SmoothHiQ is a Spatial Smoother only, I combined it with FaeryDust, to have one good Temporal smoothing filter and one good spatial. Convolution3d is a spatio-temporal cleaner, as I recall.

Kane · #6 03-29-2003, 08:46 AM

i tested your settings too, jellygoose.
the video indeed is a bit less blurry, not much, but i could see it.
i now will test pyro´s suggestion, too

PyRoMaNiA · #7 03-29-2003, 10:31 AM

It seems to me that SmoothHiQ, although it is indeed a sharper picture, causes quite a lot more mosquito noise than Convolution3D, enough to make me favour Convolution3D. But I havent tried it with DCTFilter yet, so I'll do that now.

Jellygoose, any particular reason for having mergechroma and mergeluma after FaeryDust and SmoothHiQ? I normally put them before...will this make a difference?

Jellygoose · #8 03-29-2003, 10:44 AM

Quote:

Originally Posted by PyRoMaNiA

Jellygoose, any particular reason for having mergechroma and mergeluma after FaeryDust and SmoothHiQ? I normally put them before...will this make a difference?

Well my reason for doing this was that I wanted to smooth the original picture before blurring it. the difference would be, that the picture would be blurred, and after that the blurred picture would be smoothed. try it out, and let me know what you think looks better... i'm encoding right now, so I can't test it...
As for the mosquito noise: I cannot affirm that, but you know it's all a question of taste. however I looked at the 2 sample clips in BitrateViewer, and the Q-Level of the SmoothHiQ Clip was .01 points below the one made with C3D... that's not a lot, but still is another point for SmoothHiQ in my eyes!

PyRoMaNiA · #9 03-29-2003, 05:10 PM

Turns out most of the noise was in the source. Man, it seems like noisy DVDs are getting more and more common... Well I also did some very close up noise comparisons with AvsCompare and, to me at least, there is really NO visible difference between Convolution3D(preset="movieLQ") and SmoothHiQ(5,15,25,200,10) BUT the sample with SmoothHiQ is smaller!! So with slightly moderated SmoothHiQ settings (for example the settings you used, 3,15,25,192,10) we can get a sharper picture for a much smaller filesize!

I wonder what kwag has to say on this... (Hint, hint!

)

EDiT: I put the mergechroma and mergeluma back at the beginning of the script, as I thought it looked very slightly better...I didn't really see much difference though.

kwag · #10 03-29-2003, 05:18 PM

Quote:

Originally Posted by PyRoMaNiA

I wonder what kwag has to say on this... (Hint, hint!

)

Haven't had a chance to compare yet! I'll see if I can try it later today

-kwag

kwag · #11 03-29-2003, 05:51 PM

Well, here's my comparison:

As you can see, there's barely any visual difference between the two filters.
My 45 second sample file size with C3D was 7,465KB, and the one with SmoothHQ was 7,351KB. A difference of 115KB, which is VERY significant for the time of the sample. But the big difference was the time to encode. For the SHQ, it was 6:36, and for the C3D, it was 5:13. That's over one minute

So if you want to cram the most quality/space, and you don't mind waiting an extra ~20% encoding time, go for SHQ

-kwag

PyRoMaNiA · #12 03-30-2003, 03:31 AM

I think I'll go for SmoothHiQ then.

I also found you can increase the values to (5,25,35,200,10) with no noticable difference!

jorel · #13 03-30-2003, 09:09 AM

Quote:

Originally Posted by PyRoMaNiA

I think I'll go for SmoothHiQ then.

I also found you can increase the values to (5,25,35,200,10) with no noticable difference!

"I think I'll go for SmoothHiQ then"
i'm going too pyro!

Kwag,

in all my tests for months,i got
more compression(less size) take more time and
less compression(more size) take less time...ever!

Jelly, thanks for this news!

kwag · #14 03-30-2003, 09:51 AM

@PyRoMaNiA,
You might as well just go with something this:

Code:

LegalClip()
mergechroma(blur(1.58))
mergeluma(blur(0.05))
GripCrop(528, 480, overscan=2, source_anamorphic=false)
GripSize(resizer="BicubicResize")
SpaceDust()
TemporalSmoother(radius=2, strength=3)
TemporalCleaner(ythresh=5, cthresh=10)
Convolution3D(preset="movieHQ")
GripBorders()
LegalClip()

The file size is slightly (~80KB) larger than with SHQ, but the quality is far better. I noticed that SHQ looses A LOT of detail, viewed on a HDTV. Here's what I mean. Look carefully at the grill of the TV:

Look at the details lost with SHQ ( also lost with C3D with movieLQ preset ). SHQ kills too much details

Edit: Forgot to mention that encoding speed for the sample with the script above was 5:00

. Far better that C3D(LQ) or SHQ

-kwag

jorel · #15 03-30-2003, 10:07 AM

hey Kwag,
really fantastic diference!

thanks for the script too!

kwag · #16 03-30-2003, 10:14 AM

Thanks jorel,

Also, correction to the above, you may use Convolution3D(preset="movieLQ") and the time difference (for me) was only 2 seconds longer that using HQ. The file size also dropped, and the difference to SHQ is now only 23KB larger that SHQ

-kwag

jorel · #17 03-30-2003, 10:30 AM

Quote:

Originally Posted by kwag

Thanks jorel,

Also, correction to the above, you may use Convolution3D(preset="movieLQ") and the time difference (for me) was only 2 seconds longer that using HQ. The file size also dropped, and the difference to SHQ is now only 23KB larger that SHQ

-kwag

thank you again friend!!!!!

ps:
would you create a thread for "thanks"?
i will be the leader!!!

Jellygoose · #18 03-30-2003, 11:27 AM

@kwag:

that is indeed a big difference! I haven't noticed that yet... however don't you think you should put the temporal cleaners and smoothers before the spatial ones? test again, and see if the filesize or quality changes...

Yeah right kwag! you washed that grill out with MS-Paint on the first picture, just because you wanna show how great YOUR scripts are!!

just kidding buddy, 6 beers is just too much at this heat

muaddib · #19 03-30-2003, 11:48 AM

Try breaking those images and examining them in PicSwith...

You will see a huge difference.

Jellygoose · #20 03-30-2003, 12:05 PM

I believe you. I'm sure there's more room to improve. there are many great filters out there!