Restoring Kodomo no Omocha

[RGMOfficial]

Also, coming back to this thread to talk about ANOTHER issue that some episodes have:

Chromatic Aberration

Yes, that's right.

Just to know, what appropriate filters and settings i can use to fix this?

[Selur]

for the fun of it:
tackling the Chromatic Aberration using VSGAN(BroadcastToStudioLite) :

(used: QTGMC + Crop + DeHalo_alpha + VSGAN(BroadcastToStudioLite) )
-> still there but a lot weaker

Cu Selur

Ps.: Sorry, overlooked that you are using AviSynth, so it's probably not helpful, but I thought I post it anyways

PPs.: try "FFT3DFilter(sigma=40.00,plane=2)", smoothing the UV plane helps at least with the chromatic aberration on the right, see: https://imgsli.com/MTEzMDE4

[jagabo]

Those are rainbow artifacts. Caused by high frequency luma interfering with the chroma carrier of a composite signal. Use a DeRainbow filter.

http://avisynth.nl/index.php/Category:Rainbow_%26_Dot_Crawl_Removal

Maybe add some temporal noise reduction to the chroma after it.

[RGMOfficial]

for the fun of it:
tackling the Chromatic Aberration using VSGAN(BroadcastToStudioLite) :

(used: QTGMC + Crop + DeHalo_alpha + VSGAN(BroadcastToStudioLite) )
-> still there but a lot weaker

Cu Selur

Ps.: Sorry, overlooked that you are using AviSynth, so it's probably not helpful, but I thought I post it anyways

PPs.: try "FFT3DFilter(sigma=40.00,plane=2)", smoothing the UV plane helps at least with the chromatic aberration on the right, see: https://imgsli.com/MTEzMDE4

Very interesting...

I did try FFT3DFilter now with some other frames...

FFVideoSource("EPISODE_FILE.mkv")

AnimeIVTC(mode=2,bbob=4,degrain=3,normconv=False)

og = last

FFT3DFilter(sigma=40.00,plane=2)
FineDehalo(rx=2.0, ry=2.0, darkstr=0.8, brightstr=1.0, showmask=0)

Levels(14,1,245,0,255)

Crop(6, 0, -10, -0)
Spline64Resize(640,480)
Subtitle("Restored")

og = og.Crop(6, 0, -10, -0)
og = og.Spline64Resize(640,480)
og = og.Subtitle("Original")

StackHorizontal(og,last)

[Selur]

I would have used plane=3 to filter uv not just v

[RGMOfficial]

Also, yeah, i also noticed you use VapourSynth.

This is because, i think, AnimeIVTC isn't ported yet to VS.

[Selur]

Yup seems like nobody had the time and motivation to port that massive script yet.

[RGMOfficial]

Yup seems like nobody had the time and motivation to port that massive script yet.

So, since you use VSGAN, there's some models i have found so far.

2X_DigitalFilmV5_Lite.pth and 2x_AnimeClassics_UltraLite_510K

You can try these, since you are in the models's Discord.

[Selur]

Sure, here you go:
2X_DigitalFilmV5_Lite.pth: https://imgsli.com/MTEzMTUw doesn't touch the chromatic aberration
DeHalo_Alpha + 2X_DigitalFilmV5_Lite.pth: https://imgsli.com/MTEzMTUz
2x_AnimeClassics_UltraLite_510K: https://imgsli.com/MTEzMTUx does a fine job.
DeHalo_Alpha + 2x_AnimeClassics_UltraLite_510K: https://imgsli.com/MTEzMTUy

You can try these, since you are in the models's Discord.

the models are all available over at https://upscale.wiki/wiki/Model_Database no need to be in the discord channel.

Cu Selur

[RGMOfficial]

Great!
Not exactly like the original cels we have on the net, but very interessing and great, regarding we are using an early version of an AI model recently-released.

the models are all available over at https://upscale.wiki/wiki/Model_Database no need to be in the discord channel.

Really. Oh, well, that's some time since i last looked the Model Database page.

[jagabo]

I did try FFT3DFilter now with some other frames...

FFVideoSource("EPISODE_FILE.mkv")

AnimeIVTC(mode=2,bbob=4,degrain=3,normconv=False)

og = last

FFT3DFilter(sigma=40.00,plane=2)
FineDehalo(rx=2.0, ry=2.0, darkstr=0.8, brightstr=1.0, showmask=0)

Levels(14,1,245,0,255)

Crop(6, 0, -10, -0)
Spline64Resize(640,480)
Subtitle("Restored")

og = og.Crop(6, 0, -10, -0)
og = og.Spline64Resize(640,480)
og = og.Subtitle("Original")

StackHorizontal(og,last)

That's totally destroying the chroma. Clip1.mkv, U (top) and V (bottom) planes, original on left, after FFT3DFilter(sigma=40.00,plane=3) on the right:


[Attachment 65474 - Click to enlarge]

[Selur]

yup

[jagabo]

Here's what I came up with:

LWLibavVideoSource("clip1.mkv") 
AnimeIVTC(mode=2,bbob=4,degrain=3,normconv=False)
BilinearResize(480, height)
DeRainBow(threshold=20, bfactor=1)
MergeChroma(aWarpSharp2(depth=5), aWarpSharp2(depth=15))
SMDegrain (tr=2, thSAD=200, refinemotion=true, contrasharp=false, PreFilter=4, mode=0, truemotion=true, plane=4, chroma=false)
FineDehalo(rx=2.5, ry=2.0, BrightStr=1.5, DarkStr=1.5)
nnedi3_rpow2(2, cshift="Spline36Resize", fwidth=720, fheight=480)
aWarpSharp(depth=5)
Sharpen(0.5, 0.3)

Here's the old version of DeRainbow.avs I used:

##############################################################################
#
#	DeRainbow(clip c, int "threshold", int "bfactor")
#  
#	Rainbows are caused by high frequency luma interfering with the chroma
#	carrier.  So luma edges are where you will get rainbows (high frequency
#	noise in the luma channels).  They are removed here by blurring the chroma.
#	Only areas near edges are blurred.
#
#	threshold default 40
#	Edge detection threshold. This basically controls what parts of the picture
# 	will be processed.  Lower values will result in removal of rainbows in
#	more and wider areas but potentially blurring of the chroma in those areas.
#	Valid values are 0 to 255.  Typical values are probably 20 to 60.  Use the
#	highest value you can and still get removal of rainbows in the areas you 
#	want.
#
#	bfactor, default 4
#	Blur factor.  Higher values more effectivly remove rainbows but may
#	also result in noticably blurred colors and loss of saturation of small
#	patches and edges.  Valid values are 1 and above. Typical values 
#	are 4 to 8.  Use the lowest values that gives effective rainbow removal.
#
##############################################################################


function DeRainbow(clip c, int "threshold", int "bfactor")
{
	threshold = default(threshold, 40)
	bfactor = default(bfactor, 4)

	bwidth = (c.width/bfactor/8)*8
	bwidth = max(bwidth, 16)

	bheight = (c.height/bfactor/8)*8
	bheight = max(bheight, 16)

	umask = c.UtoY().mt_edge(mode="hprewitt", thy1=threshold, thy2=threshold)
	vmask = c.VtoY().mt_edge(mode="hprewitt", thy1=threshold, thy2=threshold)
	rmask = Overlay(umask, vmask, mode="add").mt_expand().Blur(1.4).Blur(1.4).Spline36Resize(c.width,c.height)
	#return(rmask) # if you want to see the edge mask

	cblur = c.BilinearResize(bwidth, bheight).KNLMeansCL(d=1, a=2, h=3, channels= "UV").Spline36Resize(c.width,c.height)
	# return(rblur) # if you want to see the blurred video
	cblur = Overlay(c, cblur, mask=rmask)
	MergeChroma(c, cblur)
}

##############################################################################

There's much less damage to the chroma. Clip1.mkv, U (top) and V (bottom) planes, original on left, after FFT3DFilter(sigma=40.00,plane=3) on the right:


[Attachment 65480 - Click to enlarge]

[Selur]

Ah, you got quite different version of DeRainbow than the one from http://avisynth.nl/index.php/DeRainbow no wonder you get different results than I do with DeRainbow. (but I think we are aiming to remove differenc chroma artifacts )

[jagabo]

I thought he was talking about this (8x point upscale):


[Attachment 65482 - Click to enlarge]

That's back and forth between two frames from the boy's white shirt with black stripes. The green/magenta discolorations are from high frequency luma interfering with the chroma carrier of a composite signal.

I'm not absolutely sure, but I think I wrote the derainbow filter I used above.

[Selur]

Yup and I was looking at these:




Cu Selur

[jagabo]

Yes, very different phenomenon. I guess we'll have to wait for the OP to clarify.

[Selur]

Yup.

[Selur]

btw. for the still existing ghosting ExBlend (https://forum.doom9.org/showthread.php?t=175948) might be worth a try.
Attache the source and the output I got from Exblend + mclean. (still not perfect, but a huge step I think)

Cu Selur