My First Spectral rendering

[jbikker]

Been playing with spectral rendering:

spectraldragon.jpg (156.18 KiB) Viewed 4001 times

Completely new world for me so be gentle.
This is with tracing four wavelengths at once; three of them are killed when a wavelength dependent pure specular dielectric is encountered.

The tracer accumulates CIE XYZ colors and converts them to RGB using the approach proposed by Brian Smits (http://www.cs.utah.edu/~bes/papers/color/paper.pdf). I chose not to use the more accurate method of Meng et al., (https://jo.dreggn.org/home/2015_spectrum.pdf) as it requires substantially more involved calculations.

EDIT: another one:

dragon_spectral.jpg (105.32 KiB) Viewed 3980 times

This is unidirectional path tracing so the caustics took a while to converge, and they are still noisy...

[shocker_0x15]

Cool

What is the material?
Do you implement Hero wavelength spectral sampling?

[jbikker]

Quartz. I'm tracing four wavelengths as proposed by the Hero paper, but I have not implemented their MIS scheme yet. In this case it's useless by the way; the quartz will reduce weights for all wavelengths but the hero to zero anyway. For the floor and walls the four wavelengths are very effective, obviously.

[ziu]

Since we're talking secondaries here, did you guys notice this article?
"Building an Orthonormal Basis, Revisited"; Tom Duff, James Burgess, Per Christensen, Christophe Hery, Andrew Kensler,
Max Liani, and Ryusuke Villemin:
http://jcgt.org/published/0006/01/01/paper.pdf

It describes a fast and accurate way to generate an orthonormal basis for things like random sampling. I got the link from a Twitter post by Carmack some time ago but I don't remember seeing it linked here yet.

I watched the Hero presentation at EGSR a couple years back. The results looked really excellent.

[jbikker]

I hadn't seen it, but it looks good.
On a related note: https://www.shadertoy.com/view/lssfD7 , Lambert reflection without tangents.

[friedlinguini]

Whoah, weird. I had thought you could do something very similar, but by adding a random point on the *interior* of a sphere to the normal, rather than on the surface: http://psgraphics.blogspot.com/2014/09/ ... -rays.html. Obviously it would be better to use exactly two random parameters than an unbounded number of them, but I have a hard time wrapping my brain around the possibility that both are correct.