So I am currently attempting to implement the Beckmann Microfacet BSDF model with the new cvex_bsdf functionality. I am almost there but am having some issues with bias between bsdf and light sampling.

I have setup a Veach test using the default ‘blinn’ pbr specular model and pbr diffuse to see how the pathtracer mis bias is working..

Attached are the results. The question i have is about the variance and noise in the ‘bsdf only’ -1 bias image….
The cvex_bsdf docs state…. “If the rendered results are the same (apart from noise) with values of -1, 0, and 1, your shader is bias-free.”

Is the variance from that image compared to the the bias 0 image within the expected threshold? I seems pretty good apart from the high roughness and small light panel in the bottom left.

I am having the opposite affect….

My surface bsdf sampling is fine but light sampling is very noisy…..

Setting the ‘pdf’ export in the bsdf_sample shader to any random value does not affect the rendered result… After looking in the path_tracer code i see that ‘hitpdf’ is only used in the shadowing function and does not seem to affect the MIS bias…
Wondering where that value is used?