Hey people, got some screens for the ones wanting a AHR vs VXGI comparaison
Well, took a test at the new VXGI build, loaded the sci fi hallway scene, and while it looked quite good, was surprised by the perfomance, or the lack of it really 
Here are some screens of that scene in VXGI and AHR:
AHR
http://i.imgur.com/nH9N358.png
VXGI
http://i.imgur.com/weMfwJ3.png
I do recognize that VXGI looks better here, but keep in mind that is quite a stressful scene for AHR quality, as the normal maps are mostly flat, and I count on them to break the noise. I’m thinking of doing something as micro-variations of the normal based on roughness. We are using a microfacets model after all…
Really need to work on improving the bilateral blur, looks quite bad, too much bleeding. Not sure why I can’t get it to work properly, that’s quite a standard thing…
Anyway, got some thinks on the pipeline to improve quality and performance. One of them is something I have tried before, and worked quite good, but had some problems I’m yet to solve, and that is progressive refinement. In short, changing the ray kernel every frame, and if the pixel haven’t moved, use the prev results. That has the direct effect of increasing the ray count. With a few frames, you can increase it quite a bit and still have a dynamic scene, but there are some subtleties that I need to address.
The other thing is something I like to call “Weighed Bicubic Push-Pull”. It’s the good ol’ push-pull upsampling, but modified so that it uses bicubic filtering on the push phase and takes into account depth discontinuities. The idea with that is to not compute some random pixels on the screen, and then use that filter to fill the empty spaces. idea it’s something of a long shot compared to the others, but we’ll see.