Toksvig AA

[devshgraphicsprogramming]

Mipmaps of Roughness and Bumpmap derived maps (Dot3 Normal Map or Derivative Map) need to be calculated together.

I do not know yet, if Reflectance and Metallic Maps need to be taken into account.

This needs to be done properly for each BRDF (GGX, Orren Nayar).

As well as have separate encodings for Dot3 Normal Map and Derivative Maps respectively.

Note: This is completely not necessary for BRDF Explorer.

[devshgraphicsprogramming]

It seems that the filter (box, gauss, lanczos, etc) is orthogonal to the technique (have to double check)

http://www.selfshadow.com/sandbox/toksvig.html

[devshgraphicsprogramming]

For reference, although this stuff only works for Blinn-Phong, will have to be mathematically transformed for Cook-Torrance and GGX.
https://blog.selfshadow.com/2011/07/22/specular-showdown/

Old comment from a long while ago:

A sane approach to derivative map filtering should also be taken, maybe something akin to normal maps
http://developer.download.nvidia.com/whitepapers/2006/Mipmapping_Normal_Maps.pdf
https://kosmonautblog.wordpress.com/2018/09/17/roughness-mip-maps-based-on-normal-maps/
The ideal approach would be to mip-map the derivative map just like a normal texture (this part doesn't have to change), but also keep track of the variance introduced (just like in the Nvidia paper) by averaging multiple normals (implicit, can work out from derivatives) and what the gaussian distribution of normals would be. That gaussian distribution of normals can be used to increase the roughness in roughness (specular) map's mipmaps. Basically while mip-mapping the derivative map, we should mip-map the roughness/specular at the same time.

[devshgraphicsprogramming]

Toksvig (but was derived for non GGX NDF, so needs different factors to work):
https://developer.download.nvidia.com/whitepapers/2006/Mipmapping_Normal_Maps.pdf

Also CoD WWII improves on this by storing an explicit LUT of Roughness to Normal Lenght Mappings (that one can invert or binary search).

Olano (derived for Beckmann, will need work to adapt for GGX):
https://pdfs.semanticscholar.org/8329/b0aec576d19ef1d338524e8cd3628912ea4b.pdf
https://www.csee.umbc.edu/~olano/papers/lean/

LEADR:
https://hal.inria.fr/hal-00858220v1/document

Latest (any work by Anton Kaplanyan makes me very happy):
https://research.nvidia.com/publication/filtering-distributions-normals-shading-antialiasing

Old + interesting stuff:
http://www.graphics.stanford.edu/courses/cs448-05-winter/Schilling-1997-TechRep.pdf
https://www.cs.sfu.ca/~pingtan/Papers/egsr05.pdf
https://blog.selfshadow.com/publications/s2013-shading-course/rad/s2013_pbs_rad_slides.pdf
https://shuangz.com/projects/nmfilter-egsr17/nmfilter-main-ld.pdf
http://www.unilim.fr/pages_perso/xavier.chermain/assets/pdf/revised_version.pdf

[devshgraphicsprogramming]

@Przemog1 we might do this for the derivative mapping stuff

[devshgraphicsprogramming]

folded into Devsh-Graphics-Programming/Nabla#57