[SOLVED] Derainbow and Restore extremely bad source

[AOA]

Hello, for a few weeks now I've been trying to restore haruhi's bluray.
As some of you might already know, it was in fact authored so badly, that no one should even attempt to improve it, but I was still able to find a few decent encodes around the internet.

The major problem is the heavy rainbow artifacts as you can see (link below).
Looking at the video I can't even imagine what they had done to make it look so bad. Scenes that are extremely grainy (on purpose) on the DVD, look like someone drew circles on the screen when watching the bluray.

Anyway
I couldn't find any derainbow filter that managed to improve the colors even a little.
To make a dent, you'd have to turn the threshold down so much, that other colors would start fading into each other.

I know, its impossible to completely restore the colors and lineart for that matter, but what do you guys recommend?
Is there a specific plugin I missed, that specialices on these kinds of sources?

https://mega.nz/file/YMF1nJjS#Qjgma8nmeF2Cw4YnPESze0HPFTY5cpC9ckvTeeDEowo

Sorry if this posts seems like a duplicate, but I'm thinking these rainbows have to be fixed a little differently.

[Selur]

I don't see any rainbow artifacts,....
What do you mean if you refer to rainbow artifacts?
No really knowing what the issue is, here a few examples what different filters do to your source:

VSDPIR denoise + LineDarkening: https://imgsli.com/MTM0NzU5
VSDPIR deblock + LineDarkening: https://imgsli.com/MTM0NzYy
CCD(10) + LineDarkening: https://imgsli.com/MTM0NzYz
BasicVSR++ + LineDarkening: https://imgsli.com/MTM0NzY0
RealCUGAN + LineDarkening: https://imgsli.com/MTM0NzY1
VSGAN(AnimeUndeint_Compact) + LineDarkening: https://imgsli.com/MTM0NzY2
=> you might want to clarify a bit clearer what artifacts you want to get rid of.

Cu Selur

[AOA]

I don't see any rainbow artifacts,....
What do you mean if you refer to rainbow artifacts?
No really knowing what the issue is, here a few examples what different filters do to your source:

VSDPIR denoise + LineDarkening: https://imgsli.com/MTM0NzU5
VSDPIR deblock + LineDarkening: https://imgsli.com/MTM0NzYy
CCD(10) + LineDarkening: https://imgsli.com/MTM0NzYz
BasicVSR++ + LineDarkening: https://imgsli.com/MTM0NzY0
RealCUGAN + LineDarkening: https://imgsli.com/MTM0NzY1
VSGAN(AnimeUndeint_Compact) + LineDarkening: https://imgsli.com/MTM0NzY2
=> you might want to clarify a bit clearer what artifacts you want to get rid of.

Cu Selur

oh sorry.
I am still pretty new to filtering so I might have gotten the name wrong.
The artifacts Im talking about is pretty obvious to see on the ear of the character to the left, or the ribbon of the character in the middle.
Its colors look shifted and sometimes (ribbon) completely destroyed.

Looking at the amazing comparison you provided, BasicVSR++ did it already improve it by a long shot.
Much cleaner than what I could come up with, while testing.
Thanks!! :)

[Selur]

That is probably the side effect of some not so good upscaling.(https://silentaperture.gitlab.io/mdbook-guide/filtering/descaling.html and https://guideencodemoe-mkdocs.readthedocs.io/encoding/descaling/ might be interesting.)
Here's what's happening when I downscale to 960x540 and then upscale using w2xncnnvk.Waifu2x(clip=clip, scale=2, model=0, fp32=False): https://imgsli.com/MTM0Nzg1
960x540 is probably still too high of a resolution, but this should show the idea.

[AOA]

That is probably the side effect of some not so good upscaling.(https://silentaperture.gitlab.io/mdbook-guide/filtering/descaling.html and https://guideencodemoe-mkdocs.readthedocs.io/encoding/descaling/ might be interesting.)
Here's what's happening when I downscale to 960x540 and then upscale using w2xncnnvk.Waifu2x(clip=clip, scale=2, model=0, fp32=False): https://imgsli.com/MTM0Nzg1
960x540 is probably still too high of a resolution, but this should show the idea.

So I've been playing around with numerious of these upscalers like BasicVSR++ and Waifu2x. (Are these AI based upscalers?)

I don't feel comfortable reupscaling anime using machine learning, unless it isn't.

Considering your suggestion, I assume there is not really any other (less AI) kind of solution for this?

What do you think of this encode I found online?
https://imgsli.com/MTM0OTA1
Is this possibly waifu2x|... as well?

[Selur]

Could be Waifu or vsgan based models, or just a bunch of filters with proper descaling beforehand. :)
This was just to give an idea of what happens. :)

Also note that you can do some masked filtering, basically any filtering script I have seen from anime encoders has tons of masking.
With anime, it's always unclear what are fine details and what is noise.
=> You have to decide what you like more.

[AOA]

Could be Waifu or vsgan based models, or just a bunch of filters with proper descaling beforehand. :)
This was just to give an idea of what happens. :)

Also note that you can do some masked filtering, basically any filtering script I have seen from anime encoders has tons of masking.
With anime, it's always unclear what are fine details and what is noise.
=> You have to decide what you like more.

Hm, makes sense. Do you know any resources that specialise on these restorations? I've read many guides and pdfs about anime filtering but they all barely touch masks and only explain mainstream stuff like debanding and degrain.

Thank you for helping me out!

[mastrboy]

Hm, makes sense. Do you know any resources that specialise on these restorations? I've read many guides and pdfs about anime filtering but they all barely touch masks and only explain mainstream stuff like debanding and degrain.

Thank you for helping me out!

https://guide.encode.moe/encoding/masking-limiting-etc.html and https://silentaperture.gitlab.io/mdbook-guide/filtering/masking.html should get you in the right direction.

[Selur]

Personally, I would probably go for: DeHalo_Alpha + DPIRDeblock + adaptive sharpening + LineDarkening
Script: https://pastebin.com/3rDakaKC
Image: https://imgsli.com/MTM1MDkx
But with animes, it's always the thing, what one would call noise is details for others. ;)

Cu Selur

[AOA]

https://guide.encode.moe/encoding/masking-limiting-etc.html and https://silentaperture.gitlab.io/mdbook-guide/filtering/masking.html should get you in the right direction.

Oh these look great! Thanks :)

Personally, I would probably go for: DeHalo_Alpha + DPIRDeblock + adaptive sharpening + LineDarkening
Script: https://pastebin.com/3rDakaKC
Image: https://imgsli.com/MTM1MDkx
But with animes, it's always the thing, what one would call noise is details for others. ;)

Cu Selur

This is great. Thank you so much for the example.
It's pretty hard to distinguish whether a filter is an improvement or just a new feature. I think filters should help the picture become what the artists tried to achieve, not what I personally think looks good.
Pretty unfortunate how some members of the encoding community desperately want to distribute content that looks 'special'.
thanks again!

[Selur]

I think filters should help the picture become what the artists tried to achieve, not what I personally think looks good.
Sure, the problem is that you don't know unless you are the artist or the artist documented it. ;)

[AOA]

Sure, the problem is that you don't know unless you are the artist or the artist documented it. ;)

Especially when the source is so bad, you can't even make out the lines properly haha :D