Indeed, including pure unidirectional sampling via MIS removes the geometric singularities (though I haven't seen this proven formally anywhere yet). Nevertheless, when the light source is small, the variance can still be arbitrarily high. For producing the results for the paper, I haven't used this technique in any method. It would improve all images with area lights (with all methods), but only slightly and mostly in the areas very close to the light sources.Dietger wrote:Tiny question; when you make a comparison with 'traditional path tracing', did you include the contribution of random walks hitting the light sources directly in there with MIS or did you rely completely on next event estimation? Maybe you mentioned it somewhere but I missed it. Including direct light source hits should fix the infinite variance problem as well right?
Right, that obviously only improves things if the probability of actually hitting the light source is significant which is the case close to the lights. But especially in low density volumes it will also get rid of some arbitrarily bright fireflies from paths through any pixel that get too close to a light source after multiple scatter events. Anyhow, with or without unidirectional sampling traditional path tracing still sucks at volumes, so your joint IS is a pretty cool improvement!ingenious wrote:It would improve all images with area lights (with all methods), but only slightly and mostly in the areas very close to the light sources.
Exactly, unidirectional sampling works well for paths whose second last vertex is close to the light source, where there's a reasonably high chance of hitting the light by shooting in a random direction.Dietger wrote:Right, that obviously only improves things if the probability of actually hitting the light source is significant which is the case close to the lights. But especially in low density volumes it will also get rid of some arbitrarily bright fireflies from paths through any pixel that get too close to a light source after multiple scatter events. Anyhow, with or without unidirectional sampling traditional path tracing still sucks at volumes, so your joint IS is a pretty cool improvement!
Thanks for your reply, ingenious. So the more aggressive warp can eliminate variations of geometric term, leaving only the phase function stuff for tabulation. Great idea.ingenious wrote: Assuming I've understood your question the direction transformation correctly, my answer it that yes, this transformation is in fact needed. The standard (\theta, \phi) transformation you mention yields a uniform direction distribution over the sphere. You need to more aggressively warp the \theta angle in order to cancel the singularity around the pole.
