Rendering with adaptive refinement sampling

[aldozang]

Hi guys,
I come to bring a new discussion. I hope we get developers interested in working on this idea, which is widely used in other renders. As always we are talking about performance, and many of us like to render large images where generally only a few small regions (about 10%) have a noise difficult to resolve with normal sampling. Let me tell you if we make a system adaptive refinement of sampling for the Luxrender. I think it's possible without too much effort to adapt Luxrender to make a calculation of the variance in each stage of sampling and so to reach a variance set by the user.
Thus, the pixels that already have a low variance will not be sampled again, and the pixels that have not achieved the desired level of variance will continue to be sampled.

I hope you understand something of the idea. Excuse me for my terrible english, as you can see I'm not so good.

Thanks!

[jeanphi]

Hi,

Good idea! We are looking into such algorithms for a long time, but we didn't had time to implement one of them until now.
I had already tested a pixel variance estimate which gave a pretty good information about the noise level while being really easy to compute.

Jeanphi

[disgustingangel]

Hi guys,
I think it's possible without too much effort to adapt Luxrender to make a calculation of the variance in each stage of sampling and so to reach a variance set by the user.
Thus, the pixels that already have a low variance will not be sampled again, and the pixels that have not achieved the desired level of variance will continue to be sampled.

Perhaps this could be useful even without the automatic variance detection... If I'm not wrong I've seen, in some versions of luxrender, a button in the top right corner of the GUI which should have centered the sampling on the selected portion of the image...
I think this could be very useful, cause sometimes there is a noticeable, even obvious, "variance in the variance" ... maybe there could be some kind of a "sampling importance map", where we could paint the areas which are more problematic...
I think the automagic way is sampler work, or nor?

[aldozang]

that's good! I have seen that the Luxrender code has a unused variable "variance" in the ContributionBuffer class.

[aldozang]

.. If I'm not wrong I've seen, in some versions of luxrender, a button in the top right corner of the GUI which should have centered the sampling on the selected portion of the image...

I do not remember seeing this button.

Well, I think we can do multiple implementations, using different metric for measuring the variance, and other controls of noise. In particular, these ideas would be good in scenes where the noise is concentrated in small areas (very common in several scenes, see the gallery), since the renderer would concentrate its efforts on sampling over these parts.

[jeanphi]

that's good! I have seen that the Luxrender code has a unused variable "variance" in the ContributionBuffer class.

It is used by the metropolis sampler to bend sampling, but it has never been really worked on.
For the zone rendering, there was a button, but it did nothing and was removed to avoid confusion. Zone rendering is extremely difficult to implement with some samplers because they rely on whole picture normalization and zone rendering would destroy the render outside the zone.

Jeanphi

[patro]

Hi,

any news on this?