Page 13 of 15

Re: Adaptive sampling improvements

Posted: Fri May 17, 2019 8:41 pm
by epilectrolytics
lacilaci wrote: Fri May 17, 2019 10:38 am I have a complex area - that has noise which never goes away
Just wildly guessing here -
wherever there are glossy/coated/metal materials in a scene they throw noise from reflective caustics all over the place and this is only slowly to resolve by path tracing.
Now the converging speed on a certain surface may not be correlated with the noise level - depending on what reflective item creates the caustics and where.
And when there are less noisy areas taking actually longer to converge adaptivity won't help.

I'm waiting for glossy and metal getting included in caustic cache - theoretically that might help solving this.
Blurry reflections work well with big merge radius = small caustic cache not slowing down things too much.
But maybe I'm wrong about this...

Re: Adaptive sampling improvements

Posted: Sat May 18, 2019 5:14 am
by lacilaci
Maybe, maybe the ability to clamp indirect rays only would help too. Clamping helps in my case but the very strong direct light begins to look weird right away.

Re: Adaptive sampling improvements

Posted: Sun Oct 13, 2019 4:02 am
by alpistinho
Hi,

I am working on getting the convergence and noise tests to use arbitrary image pipelines. (https://github.com/Alpistinho/LuxCore/c ... 82c7dc99b2)
I will try then to get OCIO integrated so we can replicate the Blender filmic tonemapping internally and get the adaptive sampling to work as intended inside Blender.

Currently, the adaptive is working directly on the output without gamma correction and tonemapping, causing the overexposed samples to be sampled continuously. Without tonemapping, there is no actual overexposure

Re: Adaptive sampling improvements

Posted: Sun Oct 13, 2019 7:19 am
by lacilaci
alpistinho wrote: Sun Oct 13, 2019 4:02 am Hi,

I am working on getting the convergence and noise tests to use arbitrary image pipelines. (https://github.com/Alpistinho/LuxCore/c ... 82c7dc99b2)
I will try then to get OCIO integrated so we can replicate the Blender filmic tonemapping internally and get the adaptive sampling to work as intended inside Blender.

Currently, the adaptive is working directly on the output without gamma correction and tonemapping, causing the overexposed samples to be sampled continuously. Without tonemapping, there is no actual overexposure

Oh, wow... So finaly a proper adaptive sampling. Can't wait!

Re: Adaptive sampling improvements

Posted: Sun Oct 13, 2019 8:25 am
by Dade
In theory (because run OIDN can be expansive and it is also a totally uncharted territory), you can use an image pipeline with the denoiser and factor denoising in sampling adaptiveness.

More samples will be spent where the Denoiser needs (not always the denoiser needs more samples where there is more noise) ... which is border line science fiction :!:

Re: Adaptive sampling improvements

Posted: Fri Oct 18, 2019 5:24 pm
by alpistinho
This is a bit on hold, since I am without access to my usual environment and getting OCIO to compile on my Windows notebook has been a struggle.

However, I guess I could try implementing the BlendLuxCore part to enable better working adaptive rendering when using some of the current imagepipeline plugins. It would not work with the filmic ones, but I guess that would be a good first step.

Re: Adaptive sampling improvements

Posted: Fri Oct 25, 2019 3:31 am
by alpistinho
I've fixed a bug in the film pipeline selection and implemented a hard coded extra image pipeline on BlendLuxCore (with a big helping hand from B.Y.O.B)

Would you be able to test to see if the adaptiveness is working better now?

Notes:
  • Please disable Blender's color management (Filmic), so what you see is the same Lux is using for calculating the noise
  • You have to use a build generated after my last commit to GitHub

Re: Adaptive sampling improvements

Posted: Fri Oct 25, 2019 6:51 am
by marcatore
Sorry for the dumb question.
With this film pipeline, could we have the same color results between blender output and network rendering output?
At the moment, it seems that Blender applies something that changes colors and the output are different.

Could your develop solve this issue? Or am I completely wrong?

Re: Adaptive sampling improvements

Posted: Fri Oct 25, 2019 7:50 am
by lacilaci
Gotta lot of work...

But I'm doing some quicktests right now... I don't need to enable anything besides adaptivity in your blendluxcore right?

so far it seems that adaptivity is DEFINITELY picking up visible noise(in this case gi/shadow areas) much much better.
Annotation 2019-10-25 095202.jpg
unless there's visible sky... what?
Annotation 2019-10-25 100225.jpg

Re: Adaptive sampling improvements

Posted: Fri Oct 25, 2019 2:33 pm
by alpistinho
Thanks for testing!

I will try to replicate that. Of you have some scene that shows up the issue, feel free to share.