Welcome to Doom9's Forum, THE in-place to be for everyone interested in DVD conversion.

Before you start posting please read the forum rules. By posting to this forum you agree to abide by the rules.

 

Go Back   Doom9's Forum > Capturing and Editing Video > Avisynth Usage

Reply
 
Thread Tools Search this Thread Display Modes
Old 3rd April 2010, 02:13   #161  |  Link
tritical
Registered User
 
Join Date: Dec 2003
Location: MO, US
Posts: 999
Sorry, should work now. I put up some image enlargements here http://bengal.missouri.edu/~kes25c/nnedi2_images (using nsize=0/qual=2). Right now speed is slower than nnedi2 except for maybe nsize=0. For reference nnedi2 uses a 12x4 local neighborhood. The speed is basically linearly dependent with local neighborhood size and with nns. It should get better when more nsize options are available and smaller nns values (64 and 128) are available. Going to nns=128 is twice as fast as 256, and shouldn't result in much if any quality loss based on tests so far.

Last edited by tritical; 3rd April 2010 at 02:16.
tritical is offline   Reply With Quote
Old 3rd April 2010, 03:03   #162  |  Link
osgZach
Registered User
 
Join Date: Feb 2009
Location: USA
Posts: 676
Those pics look pretty nice, wow..

I'm beginning to wonder if this might be useful for upsizing some Final Fantasy VII (PC) assets to try and correct some of the idiocy that happened when they ported it.. (low res movies.. not holding out much hope for those, but the low res rendered backdrops could maybe be helped)
osgZach is offline   Reply With Quote
Old 4th April 2010, 22:28   #163  |  Link
markanini
Registered User
 
Join Date: Apr 2006
Posts: 299
Fantastic! Can't wait to use nnedi3 for image upsizing and deinterlacing.
markanini is offline   Reply With Quote
Old 5th April 2010, 08:07   #164  |  Link
jpsdr
Registered User
 
Join Date: Oct 2002
Location: France
Posts: 2,308
I've made a test with nnedi3 (for upsizing), with nsize=0/qual=2, speed is around 3 time slower than nnedi2 with nsize=2 (default)/qual=3...
jpsdr is offline   Reply With Quote
Old 7th April 2010, 07:45   #165  |  Link
D3C0D3R
Registered User
 
Join Date: Mar 2010
Location: Ukraine
Posts: 50
i've made quick tests too.
on lines looks like eedi2, but better (no artifacts, smoother in good meaning)
little bit sharper than nnedi2, faster than incredibly slow eedi3, but still - slow.

source - interlaced. i've deint it & upscale

nnedi2(field=-1,nsize=2,qual=1)
nnedi2_rpow2(rfactor=2, cshift="spline64resize",nsize=2,qual=1)
http://i43.tinypic.com/25a7ygw.jpg

nnedi3(field=-1,nsize=2,qual=1)
nnedi3_rpow2(rfactor=2, cshift="spline64resize",nsize=2,qual=1)
http://i41.tinypic.com/258wjyd.jpg

nnedi2
http://i40.tinypic.com/352ivd5.jpg

nnedi3
http://i40.tinypic.com/99gtc1.jpg

i've got 7 fps by nnedi2 script & about 1 fps by nnedi3.
on my C2Duo @ 4Ghz. but looks very good & finally on nnedi we have excelent lines.

Last edited by D3C0D3R; 9th April 2010 at 11:54.
D3C0D3R is offline   Reply With Quote
Old 7th April 2010, 07:57   #166  |  Link
Frank K Abbott
Registered User
 
Join Date: Mar 2010
Posts: 69
May i ask why you want to upscale an SD source by so much?
Frank K Abbott is offline   Reply With Quote
Old 7th April 2010, 08:25   #167  |  Link
D3C0D3R
Registered User
 
Join Date: Mar 2010
Location: Ukraine
Posts: 50
not so much - just only in 2 times.
read this thread. it upscales only powers of 2.
and in real script after nnedi I've always use spline64resize to downscale imgae to resonable resolution.

Last edited by D3C0D3R; 7th April 2010 at 08:28.
D3C0D3R is offline   Reply With Quote
Old 7th April 2010, 08:54   #168  |  Link
julius666
Registered User
 
julius666's Avatar
 
Join Date: May 2009
Location: Hungary
Posts: 79
Quote:
Originally Posted by D3C0D3R View Post
nnedi3(field=-1,nsize=2,qual=1)
nnedi3_rpow2(rfactor=2, cshift="spline64resize",nsize=2,qual=1)
http://i43.tinypic.com/25a7ygw.jpg

nnedi2(field=-1,nsize=2,qual=1)
nnedi2_rpow2(rfactor=2, cshift="spline64resize",nsize=2,qual=1)
http://i41.tinypic.com/258wjyd.jpg
I'm pretty sure that the pictures you posted are in wrong order.
julius666 is offline   Reply With Quote
Old 7th April 2010, 08:59   #169  |  Link
D3C0D3R
Registered User
 
Join Date: Mar 2010
Location: Ukraine
Posts: 50
Quote:
Originally Posted by julius666 View Post
I'm pretty sure that the pictures you posted are in wrong order.
yeah, you are right

Last edited by D3C0D3R; 8th April 2010 at 14:46.
D3C0D3R is offline   Reply With Quote
Old 24th April 2010, 14:54   #170  |  Link
Skauneboy
Registered User
 
Join Date: Mar 2010
Location: Sweden
Posts: 13
How is nnedi3 coming along tritical? I'm looking forward to the addition of nns options.
Skauneboy is offline   Reply With Quote
Old 3rd May 2010, 02:04   #171  |  Link
tritical
Registered User
 
Join Date: Dec 2003
Location: MO, US
Posts: 999
I put up a new version at the same link. It adds an 'nns' parameter, which can be set to 0 (64), 1 (128), or 2 (256, only available setting in first version). Default is 1. Based on my observations the difference between 1 and 2 is usually small, and 1 is roughly twice as fast. 0 starts to noticeably sacrifice quality. In terms of MSE on my test data, 2 (at all nsize values) is better than 1, and 1 is better than 0, but my perceived visual quality doesn't always follow those numbers. I've started training new nsize options: 8x4, 16x4, 32x4, 48x4, which should be ~33% faster than the x6 versions. Not expecting much difference MSE wise or visually there.
tritical is offline   Reply With Quote
Old 3rd May 2010, 09:28   #172  |  Link
julius666
Registered User
 
julius666's Avatar
 
Join Date: May 2009
Location: Hungary
Posts: 79
Thanks a lot tritical, it's really faster with nns=1, and I can't see any difference between nns=2 and 1 (although I can't see between 0 and 1 either - I'm using nnedi3 for resizing).

Quote:
Originally Posted by tritical View Post
Extension to 3D is only a trivial change at this point. I'm planning to go that route once I have all of the 2D models trained.
Will you do it? And what effects would it have? "Just" less aliasing artifacts, or would it be some kind of superresolution?
julius666 is offline   Reply With Quote
Old 3rd May 2010, 11:47   #173  |  Link
thanhtu5013
Registered User
 
Join Date: Jan 2009
Location: Lyon. France
Posts: 40
Nice job Tritical.
This one works a lot faster than your previous beta release.

Not fast as Nnedi2, obviously.
thanhtu5013 is offline   Reply With Quote
Old 3rd May 2010, 19:45   #174  |  Link
Skauneboy
Registered User
 
Join Date: Mar 2010
Location: Sweden
Posts: 13
Thanks for your awesome work tritical. Your interpolators have become quite essential for many of us Avisynthers.
Skauneboy is offline   Reply With Quote
Old 12th May 2010, 03:44   #175  |  Link
Undead Sega
Registered User
 
Join Date: Oct 2007
Posts: 713
Quote:
Originally Posted by tritical View Post
Motion compensation could be performed separately beforehand to create a truly (large scale motion) motion compensated deinterlacer.
Does this exist in the Avisynth world?

TGMC, MCBob by any chance?
Undead Sega is offline   Reply With Quote
Old 16th May 2010, 18:50   #176  |  Link
tritical
Registered User
 
Join Date: Dec 2003
Location: MO, US
Posts: 999
New version up... with help file and link on my site finally. I also released version 1.5 of nnedi2 which is just 1.4 with the paddq fix for sse.

The new version of nnedi3 adds another nns option: nns=0 (32). nns=0/1/2 from the previous version are now 1/2/3. It also adds nsize=4 (8x4). The defaults for nsize/nns when using nnedi3() are now 2/1 (so should be faster than previous version - 2/2). When using nnedi3_rpow2 the defaults are 0/3.

Quote:
Originally Posted by julius666
Will you do it? And what effects would it have? "Just" less aliasing artifacts, or would it be some kind of superresolution?
I will eventually. The basic idea is the same as that of tdeint or any other motion-adaptive deinterlacer. Namely, weave in areas of no motion and interpolate in areas of motion. The difference is that the weave/interpolate reasoning would be via neural network fed with the local 3D neighborhood instead of human coded rules/logic (like eedi2 vs nnedi3 for interpolation).

Quote:
Originally Posted by Undead Sega
Does this exist in the Avisynth world?

TGMC, MCBob by any chance?
Not that I'm aware. I know that Terka was working on this idea or something similar a few months ago, but I don't know what came of it. The idea being to combine a motion-adaptive deinterlacer with mvtools to make it motion-compensated... the same way you can combine fft3dfilter/dfttest with mvtools to make a motion-compensated denoiser.

Last edited by tritical; 16th May 2010 at 18:59.
tritical is offline   Reply With Quote
Old 16th May 2010, 19:12   #177  |  Link
LoRd_MuldeR
Software Developer
 
LoRd_MuldeR's Avatar
 
Join Date: Jun 2005
Location: Last House on Slunk Street
Posts: 13,248
Quote:
Originally Posted by tritical View Post
New version up... with help file and link on my site finally. I also released version 1.5 of nnedi2 which is just 1.4 with the paddq fix for sse.

The new version of nnedi3 adds another nns option: nns=0 (32). nns=0/1/2 from the previous version are now 1/2/3. It also adds nsize=4 (8x4). The defaults for nsize/nns when using nnedi3() are now 2/1 (so should be faster than previous version - 2/2). When using nnedi3_rpow2 the defaults are 0/3.
__________________
Go to https://standforukraine.com/ to find legitimate Ukrainian Charities 🇺🇦✊
LoRd_MuldeR is offline   Reply With Quote
Old 16th May 2010, 20:56   #178  |  Link
Terka
Registered User
 
Join Date: Jan 2005
Location: cz
Posts: 704
Tritical, dont want to make this thread longer, but nevermind, i must say: thank you! Your filters are one of the top.
Terka is offline   Reply With Quote
Old 16th May 2010, 20:58   #179  |  Link
Terranigma
*Space Reserved*
 
Terranigma's Avatar
 
Join Date: May 2006
Posts: 953
Thanks for the first official release.
I guess there's no need to Update mcbob again at the moment... not just yet, but TempGaussMC otoh.....


btw, wouldn't it be fitting to update this thread's title to something more fitting such as "nnedi3" or "nnedi2/nnedi3" to place emphasis upon the release?

Last edited by Terranigma; 16th May 2010 at 21:02.
Terranigma is offline   Reply With Quote
Old 17th May 2010, 03:38   #180  |  Link
Undead Sega
Registered User
 
Join Date: Oct 2007
Posts: 713
Quote:
Originally Posted by tritical View Post
Not that I'm aware. I know that Terka was working on this idea or something similar a few months ago, but I don't know what came of it. The idea being to combine a motion-adaptive deinterlacer with mvtools to make it motion-compensated... the same way you can combine fft3dfilter/dfttest with mvtools to make a motion-compensated denoiser.
Hmmm, thats very interesting, i wonder what has had happened to that. However, i thought that a full scale motion compensation deinterlacer existed (maybe as a fraction) in the Avisynth world, seems im wrong on that.

But if one was to make one, it could just put an end to all other deinterlacing filter for those who want high quality outputs. At the same time, (even though im not sure how it works) but the algorithm of this motion-compensated deinterlacer can be used for other purposes as well, one example could be that we could finally have a true 'Super-resolution' filter.

As ambitious i may be talking, the one thing i believe wat would let this down is the speed, however if this was made to fully work with 64-bit systems and even CUDA, this would be a small minor for some or many. So if one was to make one, i would suggest that it should be made standalone but based on Avisynth. I hope you get what i mean
Undead Sega is offline   Reply With Quote
Reply

Thread Tools Search this Thread
Search this Thread:

Advanced Search
Display Modes

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off

Forum Jump


All times are GMT +1. The time now is 12:33.


Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2024, vBulletin Solutions Inc.