Render Samples Tweet

[maxwood]

Hi all,

I've been doing quite a few test renders with direct light and path tracing and I was just wondering if someone could help me clear up the questions about the max samples.

Is there a better way of approximating how many samples you'll need to eliminate visible noise in a render? I noticed that 16000 is the default but that seems awfully high, from what I can see on this project almost all of the noise is gone by 1000 samples. I know that this will of course differ between project.

More importantly, is there a way of setting a noise threshold so that the render will finish when the threshold has been met rather than oversampling the image and wasting time?

Cheers,

Max

[kavorka]

seems to be the more bounced light and indirect illumination there is, the longer it takes. So an interior scene will take longer than a product shot

[bepeg4d]

hi,
the noise is not only project dependent, but also camera dependent. In the same scene, you can have exterior cameras that render quickly or cameras with dof, interior cameras with lots of specular materials and so on that render very slowly. There are so many aspects that bring noise that is impossible to predefine what is the right sampling. But you can render at lower resolution to have an idea of what sampling level you need to reach and then, set it for the final render
A technique for speeding up the noise cleaning is to render at double size but with a 1/4 of sampling and then downsize to the correct dimensions. In this way, any bright alone single pixel will be lost drastically reducing the sampling needed. Unfortunately, is not the best technique to use with cube-vr because of the huge frame buffer request, unless you have some titan X under the hood
ciao beppe

[maxwood]

Thanks guys. Unfortunately no Titan X under the hood haha

Does a noise threshold exist? Is that even possible with GPU rendering? What I mean is if it's possible and I set a noise threshold value that each pixel is checked to see if it needs more samples, that way it saves time by not resampling noise free areas and frees up more power to sample the more complex areas.

[aoktar]

Please watch for some possible tricks:
https://vimeo.com/groups/209276/videos/124066262