tiya3dd

[Marc FD]

i packed beta one.
i rewrote ugly code and added a emms at the end ^^

i'd like to hear your impressions.

and of course, enjoy.

(attached "tiya3dd beta 1.zip")

[NoLogo]

It's for YU12, right ?
I'll try this as soon the attachment will be ok.
Thx for the good work :).

[CruNcher]

Great looks like its faster and compresses even more then c3d i compresed an 48 MB VOB 6.5 Mbits with tiya3dd and cpu 6 down to 12 MB with XviD H263 Quantizer 2 and it looks great :)

[Marc FD]

> looks like its faster

LOL. really, test it seriously ^^.

just a question, do you think VDub's 2Dcleaner is a fast filter ??

[wotef]

i'd rate as relatively fast for a vdub spatial filter, but to be an effective denoiser, it tends to be too brutal in its blurring

two questions, marc fd, are you going to develop s3dc anymore? and, will there be a yuy2 version of tiya3d? ^__^

[Marc FD]

>i'd rate as relatively fast for a vdub spatial filter,

do you think it's faster than a (any) 3d convolution ? (with the minimal radius : 3x3)

i dunno why, for me tiya3dd is as fast as VDub's 2dcleaner (K7 opt. radius 1 -> 3x3).

>and, will there be a yuy2 version of tiya3d? ^__^

i don't think. ConvertToYV12().tiya3d().ConvertToYUY2() will be faster, because it's exactly what i would do internally if i implement YUY2 in this filter. that's why yuy2 will never be implemented. (BTW, when you do filtering, what's bad to switch to YV12 ? it'll make encoding/playback faster anyway)

>are you going to develop s3dc anymore?

i dunno at all.

in a hand, i like the s3dc concept, because it's much more aggressive, (so better for anime) and i could make it hell fast. (i didn't tried, but that should be something... fast enough to make direct filtering faster)

on the other hand, it seems nobody will use it.

when i saw i could make a full 3d convolution 4x faster than the existing filter (i mean, tiya3dd().tiya3dd().tiya3dd().tiya3dd() is as fast as a single c3d call on my computer) i thought it may be interressing.

but it seems my computer is strange, or the concolution3d i downloaded is much slower than all others. i dunno at all.

when i think of it, i had a fun idea, i'm almost sure i could do 125-point (5x5x5) 3d convolution with a ridiculous speed penalty.
no.... more funny : maybe even faster than tiya3dd, depending on cpu/ram balance. :rolleyes: .
and when i say ridiculous speed penality, i mean that with this indirect approach, it _may_ (it's just possible) achieve, in streaming, a 729 point (9x9x9) 3d convolution with the same kind of speed than c3d. (it's 27x more checkpoints, but who cares...)

but that's only quick estimates, and since nobody seems to like speed... even when quality is the same...

i sometimes wonder why i try to code fast stuff.....
it was fun anyway.

slowly,
MarcFD

[Guest]

Originally posted by Marc FD
ConvertToYV12().tiya3d().ConvertToYUY2() will be faster, because it's exactly what i would do internally if i implement YUY2 in this filter. that's why yuy2 will never be implemented. (BTW, when you do filtering, what's bad to switch to YV12 ConvertToYV12() currently assumes progressive pictures. If you convert a combed picture, chroma will bleed between fields. That is BAD. :rolleyes:

[ajp]

Marc,

don't despair I'm we all want to see new fast and good filters. I know I certainly do!

Personally I thought 'sc3d' showed real promise. I'm afraid I can't really test YV12 filters myself yet because I'm not ready to switch to 2.5. Still too many issues, especially for me doing DV to mpeg2 conversions. Unfortunately all new filters seem to be YV12... :(

Keep up the good work, :)

@jp

[Marc FD]

>ConvertToYV12() currently assumes progressive pictures. If you convert a
>combed picture, chroma will bleed between fields. That is BAD.

try YV12toYUY2(interlaced=true). it's in MPEG2Dec3 filter pack.

>Personally I thought 'sc3d' showed real promise...

speedwise ? maybe. s3dc was about as fast as c3d in C.

with low-lvl opt., tiya3dd is even faster, but not "simple", because it does the full work directly.

[ajp]

Marc,

I meant the sc3d results showed promise. Pretty good at denoising. I didn't pay much attention to speed when testing it since it was an early release...

Oddly enough tiya3d didn't denoise very well at all for me, didn't even remove all the flickering etc. Maybe I did something wrong? C3D has much more effect...?

[Marc FD]

>Oddly enough tiya3d didn't denoise very well at all for me, didn't even
>remove all the flickering etc. Maybe I did something wrong? C3D has much
> more effect...?

i'll look at it.
i'm sure of one thing : tiya3dd(1,30,30,30) is doing denoising. i dunno if it's as much as c3d does. i don't see how they could be different.

PS : are you sure you didn't used 3 as temporal influence ? it's about 10x bigger than c3d one, so recheck it, because you may end with simple 2d convolution all the time. (not noticably faster, and far less effective)

[ajp]

Marc,

I will try it with those parameters and see if it's any better. If I want to move to 2.5/YV12 I will have to have a denoiser as good as Peachsmoother to convince me... ;)

I used the standard (0,10,10,30) parameters but the source was pretty noisy, analogue tv cap...

I only wish I had the time and knowledge to make filters myself, sounds kind of fun.... :D

[Guest]

Originally posted by Marc FD
>ConvertToYV12() currently assumes progressive pictures. If you convert a
>combed picture, chroma will bleed between fields. That is BAD.

try YV12toYUY2(interlaced=true). it's in MPEG2Dec3 filter pack.
Umm, you are saying YV12toYUY2() is an equivalent of ConvertToYV12()? One makes YUY2 and one makes YV12. Hello?

Chroma downsampling is the problem I am referring to.

[ajp]

Nope...(1,30,30,30) didn't denoise very well either on a noisy tv cap... Don't know what I'm doing wrong...?

[Marc FD]

>Umm, you are saying YV12toYUY2() is an equivalent of ConvertToYV12()?
>One makes YUY2 and one makes YV12. Hello?
>Chroma downsampling is the problem I am referring to.

oups i totally missed that.
i thought Sh0dan already implemented it....
but i just checked and he missed it.

i'll add that tomorow afternoon.

@Sh0dan :
it's the interlaced YV12->YUY2 and YUY2->YV12 functions in XviD colorspaces convertions. they finish with _i_<cpu> or something like that.

[Marc FD]

>Nope...(1,30,30,30) didn't denoise very well either on a noisy tv cap... >Don't know what I'm doing wrong...?

you tested c3d too ? because they are like twins when i test. tweak a bit and you can achieve the same effect.

try tiya3dd(1,250,250,8000) too. you'll see the max denoising possible. it wouldn't denoise more than that (spatially).

PS: BTW, i may create my little monster tomorow afternoon. i like too much the idea, so i'll code it, even if it's only for me. you'll get a copy, simply pm me your e-mail. (with 125-points and more matrixes, trust me, you'll get denoising ^^)

[Guest]

It's not enough to have a ConvertToYV12() that supports interlaced source. The problem is with hybrid material, which you must know is quite common. If you set progressive, you ruin the interlaced parts and vice versa. By staying in YUY2 you avoid the problem entirely. It is just not realistic to think of YV12 as the panacea for everything.

I capture YUY2 MJPEG off air. Why should I downsample the chroma just to filter it when I don't have to?

[ajp]

Marc,

to tell you the truth C3D doesn't do a perfect job denoising this particular clip either but still better. I will try tweaking tiya3dd a little more and also try the max values for fun... :)

I'll pm you with my email, I'd love to try the monster! :) Right now I use Peachsmoother for those really noisy caps, best I've found so far. This whole YUY2->YV12 issue is also making me hesitate to convert to AVS 2.5, might be more trouble than it's worth. Unfortunately everybody seem to be concentrating on YV12 now and forgetting about us YUY2 users. Seems to me YV12 is only good for those ripping DVD's to Xvid...? :(

Cheers,

@jp

[Guest]

Originally posted by ajp
This whole YUY2->YV12 issue is also making me hesitate to convert to AVS 2.5, might be more trouble than it's worth. Unfortunately everybody seem to be concentrating on YV12 now and forgetting about us YUY2 users. Seems to me YV12 is only good for those ripping DVD's to Xvid...? :(
It isn't the YUY2->YV12 issue that should make you hesitate, because Avisynth 2.5 still supports YUY2, you just need the filters recompiled. What might make you hesitate is that Avisynth 2.5 is still alpha and many filters have not yet been recompiled and made available.

[cjv]

@ajp:
You can just grab the source (if offered) for the filters in question and recompile for 2.5. Takes about 10 seconds for Decomb and MSharpen :)

cjv

[ajp]

neuron2,

thanks I realize that but the point of converting to AVS2.5 was taking advantage of the new filters being written for it. I guess YUY2 support will be more common when 2.5 enters beta though and hopefully the issues will be worked out. I'm just to damn impatient....must..resist..trying..new...stuff...arrrgh... :D

cjv,

thanks for the suggestion and while I'm sure it's easy if you've done if before I feel it's more trouble than it's worth for me. I'll wait for the authors to do it themselves... :)


I'll stick with 2.0x for 'production' use for now and experiment with 2.5a when feeling brave. I set up a 'dual system' last night, copying a few files will switch between the two versions...

[Wilbert]

This whole YUY2->YV12 issue is also making me hesitate to convert to AVS 2.5, might be more trouble than it's worth. Unfortunately everybody seem to be concentrating on YV12 now and forgetting about us YUY2 users. Seems to me YV12 is only good for those ripping DVD's to Xvid...?
Soon there will be a HuffYV12 and then you can capture in YV12 (I hope)!

[ajp]

Thanks but my ADVC-100 captures and compresses to DV in hardware so HuffYV12 isn't an option. The capture program recieves a DV stream compressed and ready... :)

[ErMaC]

Actually I have setup a "clean environment" for AVIsynth 2.5 - I have an AVISynth_2.5_DEVEL directory where I put the alpha DLL, VirtualDubMOD, and the various filters all in it and when I want to use the alpha I just run the VDubMod in that directory and it uses 2.5's version. Meanwhile the rest of my programs all still use AVISynth 2.07, with plugin autoloading even, and I haven't encountered any problems at all.
Works great for testing purposes, I highly recommend it for those who want to try things out and not possibly mess up their system (although I seriously doubt just installing 2.5 will mess up anything).

As for this particular filter, I'll give it a look next time I encode a DVD but currently I'm waiting on a disc to arrive before I do that.

Also I really really want a YV12 version of VobSub :( Cause I've got an import anime disc I wanna encode with Decomb+Vobsub in YV12 as a test.

[vlad59]

Originally posted by Marc FD
when i saw i could make a full 3d convolution 4x faster than the existing filter (i mean, tiya3dd().tiya3dd().tiya3dd().tiya3dd() is as fast as a single c3d call on my computer) i thought it may be interressing.

but it seems my computer is strange, or the concolution3d i downloaded is much slower than all others. i dunno at all.



That's very strange. On my computer (PIII750) there is only is only 1fps between tiya3dd and C3D. I will try on my Duron 1.2 as soon as I get back home (this weekend).

After making some test, this speed difference should be due to the fact that I was working with word (and not bytes) for threshold check and temporal influence. I made a little hack to test it and it seems to be it.

But my biggest question was why is the compressibility better with your filter ..... I don't have the code to look at but I imagine that you're doing the threshold check like me ....

I can find some explanation for the speed bonus but for the compressibility gain :confused: . I gonna check quality tonight to find why.

Anyway good work as always Marc.

[Marc FD]

>It's not enough to have a ConvertToYV12() that supports interlaced
>source. The problem is with hybrid material, which you must know is
>quite common.

yes, true.

>If you set progressive, you ruin the interlaced parts

on chroma, true. luma untouched.

>and vice versa.

lol. i don't think so.
let me explain because i think i know what i'm speking about.
As you may know, in earlier versions of MPEG2Dec3 i wondered if i could get the interlaced flag from the MPEG-2 stream to make interlaced convertion/postprocessing automatic. and i saw one think : never rely on MPEG-2 interlaced flags : i've even got a stream with progressive frames (24 fps FILM !!) who was tagged as interlaced...

do you really think you can see the differences between interlaced downsampling and progressive downsampling on a progressive source ??

mine couldn't. so i'll first add interlaced support in avisynth, then i mail you the dll, and if you don't like it, i'll implement hybrid mode, who will use a combing detector to choose the method of up/down sampling in realtime. okay ? ^__^

>By staying in YUY2 you avoid the problem entirely. It is
>just not realistic to think of YV12 as the panacea for everything.

YUY2 is usefull on interlaced material, but planar is better.
BTW, 4:2:2 planar uses the same quality/bandwith than YUY2.
we use 4:2:0 planar because it's the far most used colorspace in video compression.

> I capture YUY2 MJPEG off air. Why should I downsample the chroma just > to filter it when I don't have to?

i don't understand. avisynth is a frameserver. you output for encoding or to playback, no ?

[kilg0r3]

@Marc

Which settings do you use for hq dvd sources?

[Marc FD]

>That's very strange. On my computer (PIII750) there is only is only 1fps
>between tiya3dd and C3D. I will try on my Duron 1.2 as soon as I get
>back home (this weekend).

very strange indeed. i'll load last c3d version. maybe the one i downloaded was a buggy one.

>After making some test, this speed difference should be due to the fact
>that I was working with word (and not bytes) for threshold check and
>temporal influence. I made a little hack to test it and it seems to be
>it.

yes i use byte vectors, in fact my code is ultra-simple. it's why the bin is so small.

>But my biggest question was why is the compressibility better with
>your filter ..... I don't have the code to look at but I imagine that
>you're doing the threshold check like me ....

i think so : if (abs<thr) {blend pixel with} else {keep current value} that's how fixed convolution works. dynamic convolution is much more aggressive, but harder to code in SIMD.

i do it 27x by pixel or 9x if 2d mode is triggered (no 18 points mode. i was too lazy to code it)

>I can find some explanation for the speed bonus but for the
>compressibility gain . I gonna check quality tonight to find why.

i just noticed my rounding was better in 111 mode.

>Anyway good work as always Marc.

thx :)
i'm gonna code the serious one now.

french powa ^_^

[Marc FD]

>Which settings do you use for hq dvd sources?

simply try what fits to you the best. tiya3dd(0,4,4,20) may be a good start point. not aggressive at all.

say me your favourite values with c3d.

[kilg0r3]

i usually use the movie hq (?) preset

[Marc FD]

from c3d's readme :
>Convolution3d (preset="movieHQ") // Movie Hi Quality (good DVD source)
> is an alias for Convolution3D (0, 3, 4, 3, 4, 2.8, 0)

i think tiya3dd(0,3,4,28) will be a good start point to tweak, no ??

as you see it don't use temporal/spatial thresolds because i only work in 3d/2d, i don't do spatial/temporal stuff. my "temporal influence" code is very aggressive, and does the job very well as far as i can see. if you use huge thresolds, simply lower it around 20 and less to kill all possible ghosting effects.

[kilg0r3]

thx

[Marc FD]

> thx

no pb, i'd like to here the result (quality&speed)

@all

regarding speed, here are precise results :

base script :

MPEG2source("xxx.d2v")
ac3source("xxx.ac3")
TomsMoComp(1,5,1)

speed : 41-42 fps

with tiya3dd(0,30,30,30) (beta 1)
speed : 28-29 fps

with Convolution3d(0,30,30,30,30,3) (c3dYV12, last version)
speed : 14-15 fps

so you could think tiya3dd is 2x faster ?
not so fast. the last result speak for himself :

with tiya3dd(0,30,30,30).tiya3dd(0,30,30,30).tiya3dd(0,30,30,30).tiya3dd(0,30,30,30) (beta 1)
speed : 13-14 fps

[jang0]

One general questions as i'm a newbie:

How do you measure the FPS values?
Do you output the video with HuffYUV and look at the Stats window in vdub or is there any method that is more exact? (as writing to disks reduces speed and affects the frames per second value)

[vlad59]

@Marc

Your results are very strange :

No filter -> 10,23 fps

Using my beta3 of C3DYV12 -> 7,21 fps
Convolution3D (0, 8, 8, 8, 8, Whatever, 0)

With tiya -> 8,46
Tiya3dd (0, 8, 8)

With five tiya -> 5,22

Really strange !!!!!!

I encode with a PIII750. I gonna recheck my SMP code to be sure it isn't used in your case.

[MaTTeR]

I ran a few short speed tests and I only seen a 3-4 FPS difference between C3D and tiya. I didn't check visuals at all but remember I'm also on an SMP box.

mpeg2source("D:\DIvX RIPs\Diablique\test\diab.d2v",idct=2)
Telecide()
Decimate(cycle=5)
crop(4,2,712,476)
#convolution3d(preset="moviehq")
tiya3dd(0,3,4,28)
BicubicResize(640,360,0,0.5)

I compared the two settings above and got this-

C3D/HQ= 20FPS and filesize was 40,540K

Tiya= 24FPS and filesize was 43,240K

As you can see C3D/HQ obviously compressed alot better with a minor speed penalty. I'll check the clips visually on TV later and post some feedback. Let me know if I'm not comparing apples to apples here:D

[Wilbert]

@Marc FD,
Could you provide a short description (how is it different from Conv3d, etc.) for the faq?

[Marc FD]

>Really strange !!!!!!

yes, as you said ^^

>I encode with a PIII750. I gonna recheck my SMP code to be sure it isn't
>used in your case.

i don't think so. c3d was always slow, even version without SMP.

>As you can see C3D/HQ obviously compressed alot better with a minor >speed penalty. I'll check the clips visually on TV later and post some
>feedback. Let me know if I'm not comparing apples to apples here

1) you should try to tweak the settings to get a very close result from the both filters. BTW, c3d-YV12 produce much more ghosting (than c3d-YUY2 or tiya3dd) maybe it helps compressibility more.
2) if you want to test speed, chain several tiya3dds. because your cpus are doing other stuff (decoding, IVTC, resize, encoding), so encoding speed is not proportionnal to the filter speed.

>Could you provide a short description

ok

>how is it different from Conv3d

i'd like to know ^^

seriously :

"tiya3dd is a 27-points 3d convolution denoising filter, exactly like c3d. it's faster, depending on you box (lol ^^). i think the temporal influence code is very different too."

BTW, they gived me my site back. the problem was, it was written in english, and my host said it was a french community, so i need to write it in french.
lol, doom9 don't want me to speak french and my ISP don't want me to speak english....

[frank]

@ marc FD
A fine test sequence is the begin of Jurassic Park. There are noise, dark szenes, some action, blocks, and much nebulous areas.
The Avisynth script results I compare with VirtualDub on my sharp tube monitor resizing to 800x600 - not on tv.

I have the feeling that your code is much less agressive than C3D.
tiya3dd(0,30,40,28) == Convolution3d(0,3,4,3,4,2.8)
The parameters must be set 10x higher to C3D. In this case you'll get equal smoothing. That are my visual test results.
Can that be right? Can't look at the sources.

Speed:
The same like Vlad59.
tiya3dd is about 20 % faster.

My equipment: P4, 2 GHz, 256MB RAM

[Marc FD]

>A fine test sequence is the begin of Jurassic Park. There are noise,
>dark szenes, some action, blocks, and much nebulous areas.

thx. real tests, that's why i wanted. i get my smile back, see : :)

>My tests are for SVCD creation and about 1800-2000 kbit/s.
>The results I compare with VirtualDub on my sharp tube monitor - not on
>tv.

>I have the feeling that your code is much less agressive than C3D.
>tiya3dd(0,30,40,28) == Covolution3d(0,3,4,3,4,2.8)

WOW !! 0,30,40,28 !! that's huge !!

>The parameters must be set 10x higher to C3D. In this case you'll get >equal smoothing.

are you sure ?!?

>That are my visual test results.
>Can that be right? Can't look at the sources.

well, i dunno how vlad does the stuff, but i do it the most classic/simple way.

BTW, when you write "c3d(0,3,4,3,4,2.8)", forget the 2.8 parameter, if i remember, it's disabled in c3d-YV12 (vlad, can you confirm that?).

frank, can you please run a quick test of compressibility with both settings ? take DivX/XviD, encode 2000-3000 frames at quant 2 and report the results. (info about speed would be cool too).

lol, vlad, it's weirder and weirder !!
MaTTer, maybe it explains you result, you may tweak a lot yourself too, thx ^_^

BTW, i've finished the 2d stuff of the big brother. but in C only :-/
so speed means nothing (10 fps in 3x3)

[frank]

quick guy :)
I will make some tests if I am back at home.

[MaTTeR]

Originally posted by Marc FD
MaTTer, maybe it explains you result, you may tweak a lot yourself too, thx ^_^ I think your right Marc. In terms of compressibility, I was obviously using much too conservative variables with tiya and that explains why I seen slightly more details in the clip compared to C3D.

I'm going to run a few Quant 2 tests on my torture clip of Pitch Black and I'll post some results a little later. Thx again Marc!

[Marc FD]

i released beta 2 on my site.

now, i think everybody will find what he wants :
it's able to filter very light (at least it seems)
and i've added a new mode (2) where it's a bit boosted on the aggresive side.

enjoy ^_^

[jang0]

Some speed test:

1) Plain MPEG2 Decoding with MPEG2Dec3 v0.94 (32bit MMX SSE IDCT, no pp):

24-26 FPS

2) same settings as in 1) + tiya3dd(0,6,12,28) (this should correspond to the animeHQ setting of conv3d):

14-15 FPS

3) same settings as in 1) + convolution3d(preset="animeHQ"):

8-9 FPS

4) same settings as in 1) + 4 calls in a row of tiya3dd(0,6,12,28):

6-7 FPS

5) same settings as in 1) + 4 calls of convolution3d(preset="animeHQ"):

3-4 FPS


Used Versions: VdubMod 1.4.12, Avisynth Alpha 2.5, MPEG2Dec3 v0.94, tiya3dd beta 2, convolution3dYV12 beta 3

CPU: Duron 800, OS: win2k, Graph: TNT2, RAM: 256 MB PC133

Another thing I noticed:

After uninstalling divx 5.02, in the file information menu of vdubmod, it says that xvid mpeg4 codec is now used for decompression (before: DivX 5.02). It seems that Xvid is approx. 4-5 FPS faster than divx :devil: so i really recommend using xvid for this (the tests above were done with xvid of course :rolleyes: )

jang0

[jang0]

I did some calculations to see how big difference really is by evaluating how long 1 frame needs to be drawn

Used Filter FPS -> 1 frame takes -> filtering takes -> 1 filter call takes

none 25 0.040 s - -
tiya3dd 14.5 0.069 s 0.029 s 0.029 s
conv3d 8.5 0.118 s 0.078 s 0.078 s
4* tiya3dd 6.5 0.154 s 0.114 s 0.029 s
4* conv3d 3.5 0.286 s 0.246 s 0.062 s

The results of 4 * conv3d are very inaccurate though, perhaps the deviation is caused by the very simple calculation method.

According to the results, 1 call of conv3d takes 2 - 2.5 times longer than 1 call of tiya3dd.
That doesn't mean, of course that rendering 1 frame lasts twice as long because the decoding time of 0.04 seconds remains constant.
Nevertheless you can certainly see a speed difference, at least on my machine. perhaps it's different when you're running it on an intel machine.


jang0

[Marc FD]

thx for testing jang0

that's very close to the results i had.
i'm running on a XP 1600+ / 256 Mo DDR, and the assembly in tiya3dd is really optimised (global fast coded stuff, no fine tunning ^^), so it may depends on differents cpu and ram systems.

your approach : time/filter call is the good one.
the interrest to use several times the same filter is to expand the cpu time it's demanding. (makes calculation more accurate)
that's how i tested my MMX blur, with

Blur2(1.5,1.5).Blur2(1.5,1.5).Blur2(1.5,1.5).Blur2(1.5,1.5)
Blur2(1.5,1.5).Blur2(1.5,1.5).Blur2(1.5,1.5).Blur2(1.5,1.5)
Blur2(1.5,1.5).Blur2(1.5,1.5).Blur2(1.5,1.5).Blur2(1.5,1.5)
Blur2(1.5,1.5).Blur2(1.5,1.5).Blur2(1.5,1.5).Blur2(1.5,1.5)

scripts ^____^

(it's faster than Blur(1.5,1.5).Blur(1.5,1.5).Blur(1.5,1.5).Blur(1.5,1.5))

[jang0]

Originally posted by Marc FD
the interrest to use several times the same filter is to expand the cpu time it's demanding. (makes calculation more accurate)
that's how i tested my MMX blur, with

Blur2(1.5,1.5).Blur2(1.5,1.5).Blur2(1.5,1.5).Blur2(1.5,1.5)
Blur2(1.5,1.5).Blur2(1.5,1.5).Blur2(1.5,1.5).Blur2(1.5,1.5)
Blur2(1.5,1.5).Blur2(1.5,1.5).Blur2(1.5,1.5).Blur2(1.5,1.5)
Blur2(1.5,1.5).Blur2(1.5,1.5).Blur2(1.5,1.5).Blur2(1.5,1.5)

scripts ^____^

yeah, thats right, when adding more filter calls, you can increase the influence of the filters themselves on the test results while decreasing the influence of the constant stuff (decoding etc.), perhaps i'll do some more testing later on. but you can definately say that tiya3dd runs faster than conv3d (of course on MY system with the versions posted above, i'm too careful to take general conclusions, especially concerning intel/amd differences, different chipsets, os's etc.)

(it's faster than Blur(1.5,1.5).Blur(1.5,1.5).Blur(1.5,1.5).Blur(1.5,1.5))

wow, nice

[bilu]

@Marc

Will you release this filter for Avisynth 2.0x ?
:rolleyes:

[Marc FD]

>Will you release this filter for Avisynth 2.0x ?

as soon as there is YV12 in, yes :D

lol, never mind, answer is : no.

avisynth 2.5 will go to beta sooner and later, and i use avisynth 2.5 without any problems for several weeks now. i'm not going to switch anymore ^^

[ErMaC]

Isn't the only time Divx/xvid is used for the decompression is when you're actually viewing it in the VDub preview window? I think that when you actually do any sort of compression with fast recompress neither codec touches the video stream. Or am I missing something obvious here?