Dogway's Filters Packs

[Dogway]

I updated a few scripts today. Wanted to call out Boulder as I updated TransformsPack so now 'SSIM2' works over there or deep_resize().
deep_resize() also supports now ratio resize, so deep_resize(2) will mean to double the size, likewise deep_resize(0.5) is halving the size.

As I said I finally added accumulated scene stats to SoftLimiter, making the limiting work on a per-scene basis.
My plan is to include per-scene filtering in more functions so if you have a request for one of my filters let me know.

Now all the scripts are up-to-date with internal (oops, nope, ExTools needs a minor update). From now on I will add global stats parser to SceneStats, work on Zopti optimizations and finally finish TransformsPack final.

[Shinkiro]

Originally Posted by kedautinh12

Can you share script you were change to expr??

setmemorymax(6000)
ffms2("video.mkv",cache=true)

ContinuityFixer(left=4, top=0, right=5, bottom=0, radius=0)
FillBorders(left=3,top=0, right=4, bottom=0, mode=0, y=2, u=3, v=3)
EdgeCleaner(strength=10, rep=false, rmode=17, smode=0, hot=false)

W=1920
H=1080
nnedi3_rpow2(rfactor=2, nsize=0, nns=4, qual=2, etype=0, pscrn=4, cshift="spline36resize", threads=8)
Blackmanresize(W, H, taps=4,0,0,-0,0)
ConvertTo16bit()

mthr = 16
bi = BitsPerComponent(last)
mthrHBD = ex_bs(mthr, 8, bi, true)
mlight1=last.flatmask(2, scale=5.0, lo=4, MSR=70, invert=false)
mdark1=last.flatmask(3, scale=7.0, lo=4, MSR=50, invert=false)
mask=last.ConditionalFilter(mlight1, mdark1, "AverageLuma()",">","60").ex_lut(Format("x {mthrHBD} <= x 0.5 * x 2 * ?"), UV=1).RemoveGrain((W>960) ? 20 : 11, -1)

deg1 = last.SMDegrain(tr=3,thSAD=340, thSADC=150, thSCD1=250,thSCD2=96, contrasharp=false, refinemotion=true, chroma=true, plane=4)
deg2 = last.SMDegrain(tr=2,thSAD=121, thSADC=50, thSCD1=156,thSCD2=96, contrasharp=false, refinemotion=true, chroma=true, plane=4)
deg=ConditionalFilter(deg1, deg2, "AverageLuma()",">","60")
ex_merge(deg ,last ,mask, luma=true, Y=3, UV=3)

LSFmod(ss_x=1.0,ss_y=1.0,strength=4,Smode=5)
FineDehalo(rx=2.8, ry=2.8, thmi=80, thma=128, thlimi=50, thlima=100, darkstr=0.0, brightstr=1.0, showmask=0, contra=0.0, excl=true)#.aWarpSharp2(depth=2)

db=last.neo_f3kdb(sample_mode=2, Y=64, Cb=64, Cr=64, grainy=44, grainC=0, range=15, dynamic_grain=false)
ex_merge(db, last, mask, luma=true, Y=3, UV=3) # пкрвый клип  (дебандинг) второй клип(чистые линии) 

z_ConvertFormat(colorspace_op="470bg:601:470bg:f=>709:709:709:f",dither_type="none")
ConvertBits(bits=10)
Prefetch(8)

[Boulder]

Originally Posted by Dogway

I updated a few scripts today. Wanted to call out Boulder as I updated TransformsPack so now 'SSIM2' works over there or deep_resize().
deep_resize() also supports now ratio resize, so deep_resize(2) will mean to double the size, likewise deep_resize(0.5) is halving the size.

Thanks, did some eyeballing tests. It's definitely much sharper than a simple Spline36. 'SSIM' looks like it could be a good replacement for that instead. Are there any caveats I should be aware of?

[Dogway]

Originally Posted by Boulder

Thanks, did some eyeballing tests. It's definitely much sharper than a simple Spline36. 'SSIM' looks like it could be a good replacement for that instead. Are there any caveats I should be aware of?

Prerequisites only as it needs SimilarityMetrics, fmtconv, vsTCanny... Speed wise I guess slower but haven't benchmarked. Hopefully I didn't break something/logic.

[Boulder]

Something I noticed - 'Didée' as a string doesn't work correctly as it requires UTF8 support. VDub2 won't put the 'é' there in the script editor.

EDIT: and show=true won't help to see what is really being used since it shows whatever string is given in the function call.

[Dogway]

Originally Posted by Boulder

Something I noticed - 'Didée' as a string doesn't work correctly as it requires UTF8 support. VDub2 won't put the 'é' there in the script editor.

EDIT: and show=true won't help to see what is really being used since it shows whatever string is given in the function call.

Thanks, "Didee" is also supported as alias so I changed default for that.

Do you mean to show also 'b' and 'c' coefficients? I know them by memory mostly, expected people to also familiarize but I can add them no problem.

[Boulder]

Originally Posted by Dogway

Do you mean to show also 'b' and 'c' coefficients? I know them by memory mostly, expected people to also familiarize but I can add them no problem.

Yes, I think it would be good to see what the actual resizer parameters are for each case. There's quite a few of them there

[tormento]

@DogWay

I recently bought the Aniplex version of Full Metal Alchemist Brotherhood (an anime) and (I could easily guess) it's an upsampled version of the broadcast show.

There are many aliased black line arts (with some halo too) and what was some grain (not excessive though) now became pixel blocks, mostly in dark areas.

Is there anything I could do with your tools?

Some examples:

[Julek]

Originally Posted by tormento

@DogWay

I recently bought the Aniplex version of Full Metal Alchemist Brotherhood (an anime) and (I could easily guess) it's an upsampled version of the broadcast show.

There are many aliased black line arts (with some halo too) and what was some grain (not excessive though) now became pixel blocks, mostly in dark areas.

Is there anything I could do with your tools?

FMAB is a 960x540 animation upscaled to 1920x1080 with bicubic 0/1, the correction process (using VS) looks like this:

from vsutil import depth, split, join
import vapoursynth as vs
core = vs.core

src = core.dgdecodenv.DGSource(r"D:\Encodes\FMAB\00000.dgi")
src32 = depth(src, 32)
y, u, v = split(src32)
dsc = core.descale.Debicubic(y, width=960, height=540, b=0, c=1)
w2x = core.caffe.Waifu2x(dsc, noise=-1, scale=2, model=6, cudnn=True, tta=False)
w2x = join([w2x, u, v])
depth(w2x, 10).set_output()

Result: https://slow.pics/c/zSsrC5HW

[ChaosKing]

Lineart looks great now!

[tormento]

Originally Posted by Julek

FMAB is a 960x540 animation upscaled to 1920x1080 with bicubic 0/1, the correction process (using VS) looks like this

Is possible to have a AVS+ script?

I am really a noob with VS. Anyway, is it the CUDA filter version?

P.S: Results are awesome.

[Dogway]

Yes, I was going to recommend de-scalers, probably the aliasing is due to a bad upscaling, so it should be fixed by returning it to the original size and optionally upscaling with a better (AI) upscaler.

Julek example is wonders, not even it cleaned the aliasing but also the ringing.

A port attempt:

src   = DGSource("D:\Encodes\FMAB\00000.dgi")
src32 = ConvertBits(src, 32)
dsc   = DeBicubicResizeMT(ExtractY(src32), 960, 540, b=0, c=1)
# w2x   = Waifu2x(dsc, nr=0, scale=2, models="C:\opt\waifu2xAvisynth-models\models", jobs=4)
w2x   = nnedi3resize(dsc, 1920,1080)
w2x   = CombinePlanes(w2x, src32, planes="YUV",Pixel_Type="YUV420PS")
Adaptive_sharpen(w2x,0.5)
ConvertBits(10, dither=1)

[Julek]

Waifu2x for AVS has the cunet model? It is the only model with good results, the other models damage the image a bit.
The second option is to use just nnedi3_rpow2.

[Dogway]

I don't know, it's this build but only for AVS 2.6 x86.

There are other AI upscalers in AVS+ like FCBI, SuperResXBR or FSRCC. But I don't think any is modeled for anime.

EDIT (alternative with less ringing):

src   = DGSource("D:\Encodes\FMAB\00000.dgi")
src32 = ConvertBits(last, 32)

dsc   = deep_resize(ExtractY(src32), 960, 540)
w2x   = deep_resize(dsc, 1920,1080,grain=0,deconv=0)
w2x   = CombinePlanes(w2x, src32, planes="YUV",Pixel_Type="YUV420PS")
w2x
ConvertBits(10, dither=1)
aWarpSharp2(depth=4,type=1,blur=3,thresh=128,chroma=2)
FastLineDarkenMOD(prot=0, thinning=10, strength=30)
# some denoise
sec  = ex_bs(0.51, 8, 10, fulls=true, flt=true)
ex_lutxy(ex_median("median5"), Format("x y == x dup y x - sgn {sec} * + ?"), UV=3)

[tormento]

Originally Posted by Dogway

but only for AVS 2.6 x86.

We need something that works with GPU of we can get old waiting

[tormento]

Originally Posted by Julek

Waifu2x for AVS has the cunet model? It is the only model with good results, the other models damage the image a bit.
The second option is to use just nnedi3_rpow2.

Can you "take my hand" and lead me to files needed, i.e. urls etc?

I can "stupidly" apply your script if you give me some coordinates.

[Julek]

Originally Posted by tormento

Can you "take my hand" and lead me to files needed, i.e. urls etc?

I can "stupidly" apply your script if you give me some coordinates.

Talk about a VS installation? I am a little busy today, but I can give a very detailed explanation tomorrow.
In advance I can tell you that we have 3 cunet options:
caffe (CUDA ver) which has a brightness bug
w2xnvk (Vulkan ver), which is slower than caffe, but has a bug-free release.
mlrt (CUDA ver) which is the fastest of them, but has the same bug as caffe

It's possible to mitigate the effects of the bug with Expr, or just use w2xnvk R3.2

[tormento]

Originally Posted by Julek

Talk about a VS installation? I am a little busy today, but I can give a very detailed explanation tomorrow.

Thank you in advance. Please cover every aspect so that other users can have benefits too.

[Dogway]

Originally Posted by Julek

You might be interested in this:
https://github.com/dnjulek/jvsfunc/b...xpr.py#L10-L55

I did ccd using only Expr, I saw that AVS does not yet have a native plugin and I think this might help.

To help with vs-expr: https://github.com/AkarinVS/vapoursy...pr-expressions

Finally ported. It acts more like a chroma deblocker so added into DeblockPack. I added a few optimizations and features so you might want to give it a look.

[Boulder]

Originally Posted by Julek

FMAB is a 960x540 animation upscaled to 1920x1080 with bicubic 0/1

Is there a clear method for defining the upscale method? I have some film sources which are clearly rather poor upscales and would like to downscale back to the original resolution for the encode.