Must read and other references.
4 posts • Page 1 of 1
I was initially excited by the title of the paper but I soon found a deal-breaker. With primary-/path-space path guiding, the dimensionality of the ideal cache is infinite. But the radiance field in the scene in world space is only a 4D or max 5D function (surface/volume + direction). Prior path guiding methods utilize a world-space 4D/5D cache and work really well. With path-space guiding, the curse of dimensionality kills everything.
It seemed to have a lot of the same ideas as multidimensional adaptive sampling (https://www.ci.i.u-tokyo.ac.jp/~hachisuka/mdas.pdf), and also the same weaknesses.