Visualizing light paths differentials (actually, ray footprint / spread in this case). These are used in path tracing to compute texture filtering, among other things... and are too easy to botch in corner cases, resulting in bugs and performance issues that are hard to pin back to the cause.... Here I play with specular roughness of two parallel mirrors (the closest of which is made invisible to camera rays) to see how that affects the propagation of the light path footprint. I think it's pretty cool to see.show more

This is so cool Max.

🙏 I am trying to get back in the rhythm.

Great to see you are back on workbench 😁

Quick question about the ray "replay": do you actually store the "hit tree" for all pixels or do you re-compute it when requested ? If you store it, I'd be interested to see how large it is in memory!

I recompute the exact same paths of the selected pixel with the instrumented version of the integrator kernel. Doing it for the whole image would be GBs and likely wouldn’t fit in memory. The way I do it I can store abundant information per path vertex without much to worry.

Ship it.

How do you account for the BRDF and surface roughness when computing spread?

Integrators tend to be implemented to sample solid angles. The solid angle of a cone half angle is: Ω = 2 π(1-cos θ) The probability density of sampling a BRDF is to a factor reciprocal to the sampled solid angle. You probably need to clamp the pdf for values lower than 1. This is an approximation, but a reasonable starting point.

Cool and practical work, the elusive lighting makes it concrete