Mixing LSF, SeeSaw and MultiSWAR for upsizing. Which picture do you like the most?

[Chainmax]

I finally got a hold of the Hackers DVD and am setting out to make a 1280x544 version out of it. For that task, I'll use some combination of these three functions and in order to know how to cimbine them I'd like to ask you to take a look at the following pictures and tell me which do you like the most and why:


http://img361.imageshack.us/img361/9190/sourcebilinearug8.png (http://imageshack.us)
http://img338.imageshack.us/img338/1672/seesawmultiswarzd9.png (http://imageshack.us)
http://img338.imageshack.us/img338/231/lsfmultiswarpe7.png (http://imageshack.us)
http://img338.imageshack.us/img338/6645/multiswarseesawza1.png (http://imageshack.us)
http://img411.imageshack.us/img411/592/multiswarlsfhx5.png (http://imageshack.us)
http://img411.imageshack.us/img411/2567/seesawmultiswarlsfad2.png (http://imageshack.us)
http://img402.imageshack.us/img402/8855/lsfmultiswarseesawpa1.png (http://imageshack.us)
http://img209.imageshack.us/img209/8246/seesawmultiswarseesawvl8.png (http://imageshack.us)
http://img341.imageshack.us/img341/192/lsfmultiswarlsfya5.png (http://imageshack.us)


Once you select a picture, you can right-click on it to find out which combination it corresponds to. I think MultiSWAR introduces a lot of noise, so either further parameter adjustment is needed or some denoising has to be done right after it.

[legoman666]

The first pic, the others all look horrid. Why on earth are you upscaling it anyway? Nothing good can come from that.

[Chainmax]

It's just something I want to try once to see how it looks. In any case, I'd appreciate broader feedback. Why are the rest of the pics horrible to you?

[DarkT]

They are hottible because of the noise, or whatever the heck it is ;).

What did the movie ever do to you to deserve such butchery? I love that movie... Awww... ;).

Anyway, maybe apply some other filters besides the sharpening to lower the noising and all that...

http://img411.imageshack.us/img411/592/multiswarlsfhx5.png

that's the one I'd go with... 5th one I think... Anyway...

[foxyshadis]

Second one, definitely, but it's not sharp enough. Use two more LSFs at strength 2000, another xsharpen, and don't forget to change Multiswar's lanczos to use taps=12.

[legoman666]

well, viewing them on an lcd monitor makes every one of them besides the first look noisy. However, Im sitting in my basement viewing them on my 720p hdtv and I'd have to go with either the 2nd or 3rd. They're a lot sharper than the frist, you can really see the difference when you look at the zipper. there is also less banding in the 2nd and 3rd pics than in the first.

Of course, like I said, they all look like crap except the first one when viewed on an LCD, however, I imagine thats not what youre going to do.

[R3Z]

Second one, definitely, but it's not sharp enough. Use two more LSFs at strength 2000, another xsharpen, and don't forget to change Multiswar's lanczos to use taps=12.

:devil:

[Chainmax]

So, for now the preferred ones are Nº2, Nº3 and Nº5.

DarkT, I also love this movie, which is why it was chosen for this little experment :).

Everyone: once you help me decide on a filter order, of course I'll tweak the filterchain so that it doesn't look that bad. I'll also post a couple of pictures with MultiSWAR being replaced by Spline36 and/or EEDI2. If you want, I can post a 10s second sample for you to toy with.

[*.mp4 guy]

I would like to see the sample. I would go with no2, or 3 depending upon how sharp you wan't it, I like 3 better, but its definately not as sharp as 2.

[G_M_C]

I'd say the second one, and in conbtrary to the others above me i'd leave it as it is for now and use a FFT3DGPU with sigma=1 (and maybe sharpen=1.1) to clean it up a little. Think it'll be fine then.

[Terranigma]

I agree with *.mp4 guy. I find no.3 most pleasing.

[anton_foy]

Maybe try HybridFuPP to remove some artifacts. I recognise this type of noise and I think HybridFuPP(MP_Mode=1) will do some good.

[Pookie]

#1 - Everything else is sharper, but not necessarily more pleasing.

[Chainmax]

Thanks for the answers so far, they will be very helpful in deciding which way to go :).

*.mp4 guy: I'll post a 10s sample clip with that scene by tomorrow.

anton_foy: thanks for the suggestion, but I think I'd rather use FFT3DFilter with a sigma of 1 or default DeGrainMedian instead as they are faster and probably as effective.


Everyone: I thought about an extra alternative, but I need you to recommend me a filter that would be about as effective as FFT3DFilter with a sigma of 1. Would MVDeGrain2 be a good choice?

[morsa]

I'd choose Number 2, anyway take notice there are some artifacts over the head of the guy behind the the character on the left.

[*.mp4 guy]

Just to give you an Idea of the best improvement you could hope for.

Standard bicubic upsizing (http://img402.imageshack.us/img402/6244/sharpnks5.png)
0.08fps upsizing (http://img257.imageshack.us/img257/9986/sharpgk7.png)

You will probably want to aim somewhere inbetween. I used bicubic upsizing as the reference, because it is what will be used by the majority of video cards.

[foxyshadis]

I didn't seriously mean #2, thus my extra comments... >.> Of all of them, #1 is least objectionable, but I probably wouldn't use any of them if that was all the source I had.

[*.mp4 guy]

I didn't seriously mean #2, thus my extra comments... >.> Of all of them, #1 is least objectionable, but I probably wouldn't use any of them if that was all the source I had.

I agree, high quality spatial interpolation methods are only worthwhile when you have a lot of information in the source, that would be lost when resized by a standard interpolator to fit your screen.

[Chainmax]

Two more alternatives:

http://img108.imageshack.us/img108/8337/seesawmultiswarn2fft3dana1.png (http://imageshack.us)
http://img82.imageshack.us/img82/214/seesawitersp36fft3dlsfrl1.png (http://imageshack.us)

[Blue_MiSfit]

Second one, definitely, but it's not sharp enough. Use two more LSFs at strength 2000, another xsharpen, and don't forget to change Multiswar's lanczos to use taps=12.

ROFL!

I agree with what others have been saying, but your new samples look promising, especially number 2.

~MiSfit

[*.mp4 guy]

Thats a lot better, a little bit of detail was lost, but it was more then worth it, I like the first one since its sharper, and the second one doesn't have any less artifacts (they are just blurier).

[Chainmax]

I see, I'll try to reduce noise and artifacting on the first one. Here's the sample:

http://www.bestsharing.com/files/IW4wc272217/Sample.demuxed.m2v.html

[Jeremy Duncan]

Just resized using Chainmax's sample.
http://www.imagehosting.com/out.php/t560267_JustResize.png (http://www.imagehosting.com/out.php/i560267_JustResize.png)

Run through the Seesaw code below in realtime with Media Player classic
http://www.imagehosting.com/out.php/t560272_SeesawAvisynth.png (http://www.imagehosting.com/out.php/i560272_SeesawAvisynth.png)

SetMemoryMax(1024)
FluxSmoothT(3)
ColorMatrix(Mode="Rec.601->Rec.709",scaling=2,threads=2)
MT("SeeSaw(NRlimit=0, NRlimit2=5, Sstr=2.40, Szp=1, Slimit=50, Sdamplo=16, Spower=1, sootheT=95, SdampHi=39)",3)
lanczosresize(1280, 544)

[Chainmax]

Here's the filterchain for the first picture:

MPEG2Source("D:\Hackers\Hackers.d2v",info=3)

Colormatrix(hints=true,interlaced=true)

AssumeTFF()
Deinted=TDeint(order=1,field=1,type=1,emask=TMM(order=1,field=1))
TFM(mode=6,order=1,PP=7,slow=2,mChroma=true,Clip2=Deinted)
TDecimate(mode=1)

Crop(0,54,720,372,align=true)

a = last
b=a.degrainmedian(mode=2).VagueDenoiser(threshold=0.8, method=1, nsteps=6, chromaT=0.8)
SeeSaw(a,b, NRlimit=6, NRlimit2=7, Sstr=1.5, Slimit=5, Spower=5, Sdamplo=6, Szp=16)

MultiSWAR(DestX=1280,DestY=544,lanczos=false)

FFT3DFilter(sigma=2,bw=32,bh=32,bt=3,ow=16,oh=16,sharpen=0.3)

AddGrain(3,0,0)


Here's the filterchain for the second picture:

MPEG2Source("D:\Hackers\Hackers.d2v",info=3)

Colormatrix(hints=true,interlaced=true)

AssumeTFF()
Deinted=TDeint(order=1,field=1,type=1,emask=TMM(order=1,field=1))
TFM(mode=6,order=1,PP=7,slow=2,mChroma=true,Clip2=Deinted)
TDecimate(mode=1)

Crop(0,54,720,372,align=true)

a = last
b=a.degrainmedian(mode=2).VagueDenoiser(threshold=0.8, method=1, nsteps=6, chromaT=0.8)
SeeSaw(a,b, NRlimit=6, NRlimit2=7, Sstr=1.5, Slimit=5, Spower=5, Sdamplo=6, Szp=16)

Spline36Resize(896,384)

FFT3DFilter(sigma=1,bw=32,bh=32,bt=3,ow=16,oh=16)

Spline36Resize(1008,432)

dull=last
sharp=dull.LimitedSharpenFaster(SMode=4)
Soothe(sharp,dull,25)

Spline36Resize(1136,480)

FFT3DFilter(sigma=1,bw=32,bh=32,bt=3,ow=16,oh=16)

Spline36Resize(1280,544)

FFT3DFilter(sigma=1,bw=32,bh=32,bt=3,ow=16,oh=16)

dull=last
sharp=dull.LimitedSharpenFaster(SMode=4)
Soothe(sharp,dull,25)

AddGrain(3,0,0)

Any tips on how to further improve the results on either is welcome.

[Jeremy Duncan]

Chainmax,

I fiddled around with some of my configs and got similiar results to the ones on the first page, post one.
The problem was the denoiser.
It was set too high.

To fix this, either lower the denoising strength, or use fluxsmootht(3) as the only denoiser.

[Jeremy Duncan]

Chainmax,

The reason I posted a different picture to the one in the first post is it shows the guy drinking.
It shows his sideburn, in particular.

With the denoiser set too high, the sideburd disappears, or gets washed out.
Just calibrate your denoiser settings by looing at that frame and comparing the sideburn to see if it's ugly.
if it doesn't look right, the denoiser is set too high.

[Chainmax]

Your resizing is wrong, you should have cropped first as the movie has a 2.35:1 AR.

[Jeremy Duncan]

SetMemoryMax(1024)
FluxSmoothT(3)
Crop(0,54,720,372,align=true)
ColorMatrix(Mode="Rec.601->Rec.709",scaling=2,threads=2)
MT("SeeSaw(NRlimit=0, NRlimit2=5, Sstr=2.40, Szp=1, Slimit=50, Sdamplo=16, Spower=1, sootheT=95, SdampHi=39)",3)
lanczosresize(1280, 544)

SetMemoryMax(1024)
Crop(0,54,720,372,align=true)
lanczosresize(1280, 544)

Just Resize
http://www.imagehosting.com/out.php/t566984_JustResizeandCrop.png (http://www.imagehosting.com/out.php/i566984_JustResizeandCrop.png)

Seesaw
http://www.imagehosting.com/out.php/t566986_Seesaw.png (http://www.imagehosting.com/out.php/i566986_Seesaw.png)

Just Resize
http://imgplace.com/directory/dir2133/1178139814_7558.png (http://www.imgplace.com/getcode.php?img=directory/dir2133/1178139814_7558.png)

Seesaw
http://imgplace.com/directory/dir2133/1178139817_4920.png (http://www.imgplace.com/getcode.php?img=directory/dir2133/1178139817_4920.png)

Just resize
http://imgplace.com/directory/dir2132/1178139561_2805.png (http://www.imgplace.com/getcode.php?img=directory/dir2132/1178139561_2805.png)

Seesaw
http://imgplace.com/directory/dir2132/1178139565_9382.png (http://www.imgplace.com/getcode.php?img=directory/dir2132/1178139565_9382.png)

[Chainmax]

Those look pretty good. So could anyone tell me where to download MVDeGrain2 (is that the latest version?) from and which are its dependencies? What other denoisers would you recommend me?

[foxyshadis]

It's part of MVTools (http://avisynth.org.ru/mvtools/mvtools.html), no other external dependencies.

[Chainmax]

I see, thanks for the clarification. In which cases would using the previous and next frames in an exclusive manner as MVDenoise does would yield better results than using them in an inclusive manner as MVDeGrain2 does? Also, at which parameters do you think MVDeGrain2 would perform similarly as FFTDFilter with a sigma of 1, 2 and 3?