nnedi2/nnedi3/eedi3 - Page 8

Undead Sega · 19th February 2010, 16:30

well my script so far is:

Code:

LoadPlugin("C:\Downloads\VirtualDub-1.8.6\Requirements\nnedi2\nnedi2.dll")
Import("C:\Downloads\VirtualDub-1.8.6\TempGaussMC_beta1mod.avsi")
AviSource("D:\***.avi")
ConvertToYV12()
nnedi2_rpow2(rfactor=4)
Spline36resize(720,576)

and i think i have the latest Avisynth installed, is there anything special i need to install or add perhaps?

scharfis_brain · 19th February 2010, 16:35

nnedi2 already is MT. what's the point in trying to MT it further?

Terka · 24th February 2010, 11:16

Tritical,
"I'm actually close to an nnedi3"...
Sorry for my bad english. Does this mean you will continue working on nnedi3?
Will be the final speed +- like nnedi2?
Will you work on nnedi3 to improve the results?
Thank you, Terka

kool · 24th February 2010, 18:27

I apology for my very noobi questions.
is the nnedi resizer meant for upsizing & downsizing?
is it possible to use nnedi2_rpow2 resize for 720p & 1080p?
what would be the rfactor for 1080p (1920x816) out of (1920x1080 blu-ray)

julius666 · 25th February 2010, 08:22

Quote:

Originally Posted by kool

I apology for my very noobi questions.
is the nnedi resizer meant for upsizing & downsizing?
is it possible to use nnedi2_rpow2 resize for 720p & 1080p?
what would be the rfactor for 1080p (1920x816) out of (1920x1080 blu-ray)

It's only for upsizing. And it's most useful when the resolution that you are aiming for is near to 2× of the original resolution.
In your case I would simply use spline36resize.
Oh, and

Inspector.Gadget · 25th February 2010, 14:22

kool, if you're keeping the hoizontal resolution the same, that suggests you want to crop, rather than resize.

kool · 25th February 2010, 15:40

IG, do you mean that I don't resize at all only crop the 1920x1080 > 1920x816 and encode it?

Inspector.Gadget · 25th February 2010, 15:56

Yup. 1920x1080 BDs are 1:1 PAR, no need to resize. Don't crop before deinterlacing/IVTC, though, if your source is 1080i true interlaced or hard telecine. To avoid taking this thread further off topic from nnedi2, make a new thread about anything you can't find an answer for and we'll help you there.

kool · 25th February 2010, 16:03

OK, thank you very much.

Dogway · 26th February 2010, 10:22

Quote:

Originally Posted by tritical

See if this: nnedi2.dll works on that processor.

Thanks for that, if it wasnt for your input I would still be "guessing" whats wrong with my script... and my CPU is an intel core 2 duo (T9400)

By the way, do you know if theres a problem on TempGaussMC when using qual=above 1? It makes crash my encoding on 2nd pass...

Stephen R. Savage · 28th February 2010, 02:16

@tritical: Since JoshyD seems to be serious about developing AVS64, perhaps you could make a 64-bit build of nnedi2? Since it's closed-source, I can't ask him to undertake the task! Then again, it seems that the use of threading is incompatible with JoshyD's AVS64.

Terka · 29th March 2010, 10:13

Tritical,
are you still working on nnedi3, or is it a dead project?

tritical · 1st April 2010, 07:09

I am still working on it. Training a 150000 parameter model on 50 million 288 dimensional vectors takes a little while even using stochastic gradient descent. I will post some pictures and a test dll sometime soon.

EDIT: Actually, I'm kind of exaggerating. The training only takes 4-5 days, and that is the largest model. The smaller models (smaller local neighborhood around center pixel) only take 1/6-1/2 that time. Most of the time since February (up to about 1 week ago) was been spent tweaking the model structure and making the training algorithm faster and more memory efficient in order to use more training data. 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.

Terka · 1st April 2010, 07:12

Thank you, T.

3D!

wonkey_monkey · 1st April 2010, 17:53

Quote:

Extension to 3D is only a trivial change at this point.

Oh, hey, what? Have I not been paying attention closely enough? Is a motion interpolating NNEDI in the pipeline?

David

tritical · 1st April 2010, 22:31

It isn't that advanced. It's sort of like training a model to do motion-adaptive deinterlacing... giving it the 3D local neighborhood of a pixel instead of just the 2D neighborhood. Motion compensation could be performed separately beforehand to create a truly (large scale motion) motion compensated deinterlacer. I said trivial, but it will actually take a while to create new training/testing data sets, run experiments, etc... But I am fairly convinced that nnedi3 exploits the spatial domain for all it's worth (without introducing higher level processing like object recognition/scene understanding)

tritical · 2nd April 2010, 08:22

Here is a beta version: nnedi3.dll. I hacked it together quickly from nnedi2 so please report any problems.

Parameters are the same as nnedi2/nnedi2_rpow2, but named nnedi3/nnedi3_rpow2. There are two differences:

1.) 'nsize' now controls the size of the 2D neighborhood. Possible settings for now (default is 2):

0: 8x6
1: 16x6
2: 32x6
3: 48x6

First value is the horizontal diameter. Second is the vertical diameter. Larger horizontal diameter allows connecting lines of smaller slope, but also has higher probability of artifacts. I would recommend always using 0 for image enlarging.

2.) 'qual' can only be set to 1 or 2. It works the same as in nnedi2 (default is 1).

In the future there will be more neighborhood size options, and a new parameter 'nns' to select model complexity (the number of neurons) for a fixed 'nsize' value. All of the models in the beta dll use nns=256, which is the largest I will probably go. Smaller values in the future will allow trading quality for speed.

Further notes. There is only one nsize=3 model in the beta dll so qual=2 w/ nsize=3 will give the same result as qual=1. That nsize=3 model is also not the final version as it was only trained on a small data set for test purposes. It will be replaced in the future (which should give better results).

aegisofrime · 2nd April 2010, 09:13

tritical, I don't know if it's just me but I'm getting a 0 byte file when I download nnedi3.dll.

AVIL · 2nd April 2010, 10:54

I've downloaded a 0 byte dll too.

D3C0D3R · 2nd April 2010, 12:08

Huge RESPECT to Tritical. I've use NNEDI (EEDI) 1,2 + TFM + TDecimate about 5 years.
Now when i regged on forum, I would like to thank for your great work.

as link for nnedi3 still not works, I wanna ask How has the speed changed compare to nnedi2?

19th February 2010, 16:30	#141 \| Link
Undead Sega Registered User Join Date: Oct 2007 Posts: 713	well my script so far is: Code: LoadPlugin("C:\Downloads\VirtualDub-1.8.6\Requirements\nnedi2\nnedi2.dll") Import("C:\Downloads\VirtualDub-1.8.6\TempGaussMC_beta1mod.avsi") AviSource("D:\***.avi") ConvertToYV12() nnedi2_rpow2(rfactor=4) Spline36resize(720,576) and i think i have the latest Avisynth installed, is there anything special i need to install or add perhaps?

19th February 2010, 16:35	#142 \| Link
scharfis_brain brainless Join Date: Mar 2003 Location: Germany Posts: 3,650	nnedi2 already is MT. what's the point in trying to MT it further? __________________ Don't forget the 'c'! Don't PM me for technical support, please.

24th February 2010, 18:27	#144 \| Link
kool Registered User Join Date: Nov 2009 Posts: 79	I apology for my very noobi questions. is the nnedi resizer meant for upsizing & downsizing? is it possible to use nnedi2_rpow2 resize for 720p & 1080p? what would be the rfactor for 1080p (1920x816) out of (1920x1080 blu-ray) Last edited by kool; 25th February 2010 at 00:04.

1st April 2010, 07:09	#153 \| Link
tritical Registered User Join Date: Dec 2003 Location: MO, US Posts: 999	I am still working on it. Training a 150000 parameter model on 50 million 288 dimensional vectors takes a little while even using stochastic gradient descent. I will post some pictures and a test dll sometime soon. EDIT: Actually, I'm kind of exaggerating. The training only takes 4-5 days, and that is the largest model. The smaller models (smaller local neighborhood around center pixel) only take 1/6-1/2 that time. Most of the time since February (up to about 1 week ago) was been spent tweaking the model structure and making the training algorithm faster and more memory efficient in order to use more training data. 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. Last edited by tritical; 1st April 2010 at 07:36.

1st April 2010, 07:12	#154 \| Link
Terka Registered User Join Date: Jan 2005 Location: cz Posts: 703	Thank you, T. 3D! Last edited by Terka; 1st April 2010 at 18:29.

24th February 2010, 11:16	#143 \| Link
Terka Registered User Join Date: Jan 2005 Location: cz Posts: 703	Tritical, "I'm actually close to an nnedi3"... Sorry for my bad english. Does this mean you will continue working on nnedi3? Will be the final speed +- like nnedi2? Will you work on nnedi3 to improve the results? Thank you, Terka

25th February 2010, 14:22	#146 \| Link
Inspector.Gadget Registered User Join Date: May 2008 Posts: 1,618	kool, if you're keeping the hoizontal resolution the same, that suggests you want to crop, rather than resize.

25th February 2010, 15:40	#147 \| Link
kool Registered User Join Date: Nov 2009 Posts: 79	IG, do you mean that I don't resize at all only crop the 1920x1080 > 1920x816 and encode it?

25th February 2010, 15:56	#148 \| Link
Inspector.Gadget Registered User Join Date: May 2008 Posts: 1,618	Yup. 1920x1080 BDs are 1:1 PAR, no need to resize. Don't crop before deinterlacing/IVTC, though, if your source is 1080i true interlaced or hard telecine. To avoid taking this thread further off topic from nnedi2, make a new thread about anything you can't find an answer for and we'll help you there.

25th February 2010, 16:03	#149 \| Link
kool Registered User Join Date: Nov 2009 Posts: 79	OK, thank you very much.

28th February 2010, 02:16	#151 \| Link
Stephen R. Savage Registered User Join Date: Nov 2009 Posts: 327	@tritical: Since JoshyD seems to be serious about developing AVS64, perhaps you could make a 64-bit build of nnedi2? Since it's closed-source, I can't ask him to undertake the task! Then again, it seems that the use of threading is incompatible with JoshyD's AVS64.

29th March 2010, 10:13	#152 \| Link
Terka Registered User Join Date: Jan 2005 Location: cz Posts: 703	Tritical, are you still working on nnedi3, or is it a dead project?

1st April 2010, 22:31	#156 \| Link
tritical Registered User Join Date: Dec 2003 Location: MO, US Posts: 999	It isn't that advanced. It's sort of like training a model to do motion-adaptive deinterlacing... giving it the 3D local neighborhood of a pixel instead of just the 2D neighborhood. Motion compensation could be performed separately beforehand to create a truly (large scale motion) motion compensated deinterlacer. I said trivial, but it will actually take a while to create new training/testing data sets, run experiments, etc... But I am fairly convinced that nnedi3 exploits the spatial domain for all it's worth (without introducing higher level processing like object recognition/scene understanding) Last edited by tritical; 1st April 2010 at 22:37.

2nd April 2010, 08:22	#157 \| Link
tritical Registered User Join Date: Dec 2003 Location: MO, US Posts: 999	Here is a beta version: nnedi3.dll. I hacked it together quickly from nnedi2 so please report any problems. Parameters are the same as nnedi2/nnedi2_rpow2, but named nnedi3/nnedi3_rpow2. There are two differences: 1.) 'nsize' now controls the size of the 2D neighborhood. Possible settings for now (default is 2): 0: 8x6 1: 16x6 2: 32x6 3: 48x6 First value is the horizontal diameter. Second is the vertical diameter. Larger horizontal diameter allows connecting lines of smaller slope, but also has higher probability of artifacts. I would recommend always using 0 for image enlarging. 2.) 'qual' can only be set to 1 or 2. It works the same as in nnedi2 (default is 1). In the future there will be more neighborhood size options, and a new parameter 'nns' to select model complexity (the number of neurons) for a fixed 'nsize' value. All of the models in the beta dll use nns=256, which is the largest I will probably go. Smaller values in the future will allow trading quality for speed. Further notes. There is only one nsize=3 model in the beta dll so qual=2 w/ nsize=3 will give the same result as qual=1. That nsize=3 model is also not the final version as it was only trained on a small data set for test purposes. It will be replaced in the future (which should give better results). Last edited by tritical; 2nd April 2010 at 08:29.

2nd April 2010, 09:13	#158 \| Link
aegisofrime Registered User Join Date: Apr 2009 Posts: 478	tritical, I don't know if it's just me but I'm getting a 0 byte file when I download nnedi3.dll.

2nd April 2010, 10:54	#159 \| Link
AVIL Registered User Join Date: Nov 2004 Location: Spain Posts: 404	I've downloaded a 0 byte dll too.

2nd April 2010, 12:08	#160 \| Link
D3C0D3R Registered User Join Date: Mar 2010 Location: Ukraine Posts: 50	Huge RESPECT to Tritical. I've use NNEDI (EEDI) 1,2 + TFM + TDecimate about 5 years. Now when i regged on forum, I would like to thank for your great work. as link for nnedi3 still not works, I wanna ask How has the speed changed compare to nnedi2?

Thread Tools	Search this Thread
Show Printable Version Email this Page	Search this Thread: Advanced Search
Display Modes
Linear Mode Switch to Hybrid Mode Switch to Threaded Mode