Help with artifact clean up

[killerteengohan]

I am using a filter called debilinear because it is not only helping to fix line aliasing, but it's improving sharpness, and detail by a lot.

source = last
Debilinear(1280,720)
nnedi3_rpow2(2, cshift="LanczosResize", fwidth=1920, fheight=1080, nns=4)
MergeChroma(last, source)

Unfortunately this is coming out rather filthy looking.

Look at the woman on the left. She ends up covered in artifacts.
https://slow.pics/c/SNcYGS3w

Usually in this case I just add MergeLuma(last, source, weight=0.5) after resizing back up on sources like this to help lessen the sharpening the filter does, and it solves my problem by kind of mixing the before and after half and half. I can't do that with this source though because doing so brings the messed up double lines, and aliasing partially back. It's not really working for this source anyways for whatever reason. The artifacts are still appearing.

The typical denoise filter I use which is mctemporaldenoise didn't really help much at all. I had to set it to such overkill to see it catch the spots all over the grays, and I was not willing to sacrifice that much detail.

I tried cleaning it with a bit of dithering and TNLMeans on a lighter setting.

LoadPlugin("C:\Program Files (x86)\AviSynth 2.5\plugins\Dither\Dither 1.28.0\win32\dither.dll")
f3kdb(range=3, grainy=32, grainc=32, y=56, cb=56, cr=56)
source5 = last
emask = mt_edge(mode="cartoon", thY1=2, thY2=2, chroma="-128")
TNLMeans(ax=1, ay=1, az=0, sx=20, sy=20, h=1.0, a=0.8)
Overlay(last, source5, mask=emask)
MergeLuma(last, source5, weight=0.4)

It does help in that above example quite a bit and it looks much better. However it looks like the very thing it cleans in that first example, makes it appear or look worse in others. It makes her hair and forehead in this example looks like its gained spots all over it, yet they are not there when I remove the TNLMeans or use mctemporaldenoise.
https://slow.pics/c/IUnVrYzs


Anyone have any cleaning suggestions I can try that will not smudge a bunch of detail out?

[killerteengohan]

I also tried a really strong dehalo filter as well. It didn't seem to help much either.

[Selur]

DPIR Deblock + Adaptive Sharpening Anime does this https://slow.pics/c/BvOwOp2i when using your unknown_Source as basis.
Vaguedenoiser with some masking might help when using Avisynth.

[killerteengohan]

DPIR Deblock + Adaptive Sharpening Anime does this https://slow.pics/c/BvOwOp2i when using your unknown_Source as basis.
Vaguedenoiser with some masking might help when using Avisynth.

That smeared the crap out of it. Thats overkill for me. It certainly did clean it up though. Just did it in a way that looks awful to me is all.

[Selur]

Yeah, Vaguesdenoise should be combined with some adjusted masking.
In case there is a filter in Avisynth which allows using GAN models, there are quite a few for animes&co over at https://upscale.wiki/wiki/Model_Database .
Maybe some temporal filtering will also work, but that at least I can't say from a still image.

[jagabo]

Maybe something simple like BilinearResize(1920,576).Spline36Resize(1920,1080)

ImageSource("gTQV49gQ.png") # 1920x1080 source image
BilinearResize(1920,576).Spline36Resize(1920,1080)

That cleans up the aliasing without increasing grain by much. Try different up/downscale methods and sizes until you get what you like. Maybe add a mild vertical sharpen while downscaled.