"A Machine Learning Approach for Filtering Monte Carlo Noise" compatible with cylces?

I watched some videos from siggraph 2015 and then there was this http://cvc.ucsb.edu/graphics/Papers/SIGGRAPH2015_LBF/

I don’t know if this video comes from the same group but the method seems to be the same.

Next to some wobbly big noise going on in animations, I can see (in the second video) that the glas shader clearly gives a very blurry results. Ofcourse this comes from the very low sampled noise (8-16 samples) that they are filtering.

Could this kind of methods be implemented in cycles?
Would be very nice for previews (or maybe even high sample renderings?)
I’d like to see some comparisons between different samplecounts for one scene.

I for one would love some sort of interpolating noise filter built into Cycles. So long as it was fast and did not damage the accuracy much.

If you read the paper, they compare against RPF (favorably, especially in terms of speed).

Could this kind of methods be implemented in cycles?

It’s a post process, all of the required “primary features” should be obtainable through render passes in current Cycles.

I’d like to see some comparisons between different samplecounts for one scene.

They have a few comparisons for 64spp images, but not the kind you’d like. (no high-frequency detail)

In general, these methods tend to all have drawbacks that don’t vanish from increasing samples. They all tend to annihilate high-frequency detail, especially when it comes from bumpmaps or high-frequency geometry. Some materials require a lot of samples to really have their appearance “come through”. Methods like this cannot capture that.

Either way, it’s laudable that they produce source code so their results can be verified.

There are two devs currently experimenting with this for Blender/Cycles. I won’t spoil test images or names since they could probably want to do it themselves though.

Still, i’m not really sold about such reconstruction filters, they are really nice “just” for clay renders at current state, IMHO. Details disappearing and getting blurry is something i would not use at all for clients work.

Speaking of artifacts, I prefer a thousand times blotches from radiance caching and similar.