NMM-HD

From what you described, I guess that's something about nnedi3, not about this script.
nnedi3 is an edge-directed interpolator, which means it interpolates image guided by edges(line, fine structure, texture, etc.). While in general edi can produce much better results than simple isotropic interpolator such as B-Spline/Lanczos, it will also produce artifacts in some regions such as the ones with complex textures.
As for nsize, nns and pscrn, they are the parameters of nnedi3, so you can look into the readme of it. For example, nsize=0/4 is suitable for upscaling, nsize3/6 is suitable for AA since it connects more lines(as well as produces more artifacts).

mawen1250 写了：From what you described, I guess that's something about nnedi3, not about this script.
nnedi3 is an edge-directed interpolator, which means it interpolates image guided by edges(line, fine structure, texture, etc.). While in general edi can produce much better results than simple isotropic interpolator such as B-Spline/Lanczos, it will also produce artifacts in some regions such as the ones with complex textures.
As for nsize, nns and pscrn, they are the parameters of nnedi3, so you can look into the readme of it. For example, nsize=0/4 is suitable for upscaling, nsize3/6 is suitable for AA since it connects more lines(as well as produces more artifacts).

Nah, man. I know that much about nnedi.
But I guess it's my bad - I didn't mention that using just nnedi3_rpow2 (without your wrapper) with the same parameters (nns=4, nsize=3, pscrn=0) don't cause the mentioned artifacts.
Just in case, I keep my plugins updated and use avisynth+.
Anyway, I guess If you can't reproduce this bug the problem is on my end... Still, maybe you take a quick look at least.

Hoodlum 写了：

mawen1250 写了：From what you described, I guess that's something about nnedi3, not about this script.
nnedi3 is an edge-directed interpolator, which means it interpolates image guided by edges(line, fine structure, texture, etc.). While in general edi can produce much better results than simple isotropic interpolator such as B-Spline/Lanczos, it will also produce artifacts in some regions such as the ones with complex textures.
As for nsize, nns and pscrn, they are the parameters of nnedi3, so you can look into the readme of it. For example, nsize=0/4 is suitable for upscaling, nsize3/6 is suitable for AA since it connects more lines(as well as produces more artifacts).

Nah, man. I know that much about nnedi.
But I guess it's my bad - I didn't mention that using just nnedi3_rpow2 (without your wrapper) with the same parameters (nns=4, nsize=3, pscrn=0) don't cause the mentioned artifacts.
Just in case, I keep my plugins updated and use avisynth+.
Anyway, I guess If you can't reproduce this bug the problem is on my end... Still, maybe you take a quick look at least.

I asked the same thing before, it's just simple nnedi3 artifacts, well, I'd say...it's like kinda related to interpolate direction order, to reproduce the exact issue with nnedi3_rpow2 (), turnleft () your clip first and turn it right back after nnedi3_rpow2 (), the same skeleton in the closet will be revealed

feisty2 写了： I asked the same thing before, it's just simple nnedi3 artifacts, well, I'd say...it's like kinda related to interpolate direction order, to reproduce the exact issue with nnedi3_rpow2 (), turnleft () your clip first and turn it right back after nnedi3_rpow2 (), the same skeleton in the closet will be revealed

This sounded like it could be the case so I went ahead and did a comparison.



But apparently it's not...

feisty2 写了：http://forum.doom9.org/showthread.php?t=170968

Oh, I see now. It didn't occur to me that it was using simple nnedi3() as well.
Thanks for your help!

请问如果使用eedi3_resize16做SD->HD的转换，color matrix这点应该怎么处理？601->709？是否需要另加ColorMatrix()？

可以在x26x的参数里手动指定--colormatrix --transfer --colorprim
至于转换的话：
1. 对于420的源，需要先转到444才能做matrix转换，带来了额外的麻烦（而且如果最后转回420就多了两次resample的损失）。
2. 光是转换matrix是不符合定义要求的，因为BT.601、BT.709、BT.2020定义了一整套颜色标准，而不单单只是matrix，所以要转换的话就应该同时转换matrix、transfer和primary。而AVS里似乎并没有方便的函数可以做到（就我所知只有3D-LUT之类的可以实现，然而使用麻烦而且只支持8bit）；VS倒是有zimg的z.ColorSpace可以一步完成所有的转换，fmtconv未来也会加入primary转换的支持。

For the nnedi_resize16 script, is it ok to upgrade nnedi3 to jpsdr's 0.9.4.17 or should I be using the original 0.9.4?

Izuchi 写了：For the nnedi_resize16 script, is it ok to upgrade nnedi3 to jpsdr's 0.9.4.17 or should I be using the original 0.9.4?

Both are OK (as long as jpsdr's version doesn't introduce new bugs), personally I use the original one.