Generalized Resampled Importance Sampling: Foundations of ReSTIR 

at 13:18 there's some really heavy low frequency noise on the walls. i'm not sure if a denoiser will handle this, i don't think it will unless its specifically designed to, but in my experience this indicates that the spatial resampling range is too small for that case. it's probably possible to use more geometric context in selecting pixels to resample from rather than to only assume closer => better, and allow for a greater distance if the local geometry is similar enough. If there were a convenient way to do this, it would automatically deal with cases where low frequency noise shows up (ie large smooth diffuse surfaces), while leaving other cases with more granular geometric detail alone. otherwise great work! surprised to see an algorithm outperform restir gi in such a wide variety of scenes this soon!

I'm having problems implementing ReSTIR GI... very nice results but too much noise compared to other people's results :

Can Neural Radiance Cache be used after the ReSTIR PT pass for a further boost in image quality?