The gallery of diminishing returns

Just a test to see how quickly noise is reduced over time in Cycles.
Final result at 1000 samples:

Two details at various sampling levels:

The good news is that there’s really little point in sampling above 300. The bad news is that the noise seems to go away pretty slow as well. No new insights of course, but always nice to have some actual examples.

ps. model made with Kangaroo, equi-sized sphere stacking under gravity constrained within the bigger sphere.


Nice, once I get around to adding view properties panel you’ll have access to more settings that control render quality. For now there aren’t many was to fiddle with these. In theory RhinoCycles_SetRenderOptions, but I haven’t checked it in a long time, so maybe it doesn’t work properly.

Meh, one more.


how did you get those peas in there? it doesnt seem to have an opening :wink:

1 Like

To me it seems like the black uncalculated space (pixles) in the first passes are causing more noice than it should be.
Would it be possible to replace black with gray instead? Or would that mess up the light calculation? (and prevent the calculation of true black areas)

If so then I guess you could work with black as alpha and interpolate the calculated data to cover the not yet calculated areas. That way black calculated pixles would also interpolate with light calculated pixles.

I also think that if you interpolate the black (uncalculated areas) in the first pass and use this as a image background then there would be less contrast in the noice and it would be less visible.

In other words could the calculated pixles voronoi blend ?

I did a quick google search on “voronoi render pass” and came across this whitepaper:

@nathanletwory, your area of expertise I think.

Peas were injected through the fourth dimension.

1 Like

That inject command is really powerful.—Mark

1 Like

Just bumping the max amount of bounces should already give better convergence, for this scene likely the transparency bounces, and diffuse bounces.

@DavidRutten, what quality setting did you start from? If you started with i.e. Draft quality you’ll have pretty low bounce settings. Just bumping sample count won’t give that much improved convergence. Final quality will give you more bounces, you can bump sample count down once switched to that.

Use hdr instead of white background. I mean - white hdr, or any hdr but lower the gamma.
Also - from my experience two different hdr maps can have gigantic difference in the noise amount - so you need to find a good one and this is done only by trial and error.

(i use octane btw)


I used one of the few environment HDR maps that ship with Rhino6. @andy, @BrianJ are we shipping ones with a known low noise characteristic?

The HDR used for SkyLighting is hugely downsampled with lots of blur. They themselves shouldn’t generate much noise anymore. But as said tweaking of the integrator settings is necessary instead.

@Maciej_Suder, it looks like the setup @DavidRutten has is a solid background color and separate HDR for reflection and skylight. This is handled by a world shader that has three background shader nodes that get picked based on light path type (camera for bg, etc.)

Maybe lowering the gamma of hdr - makes it more equalized - so more unified lighting form every angle - and less noise in shorter time.

On the other hand - in outdoor scenes noise shouldn’t be a problem - it should vanish in no time.

Another thing that I use is to crank bounces a little bit - I know it may increase render time - but more light could reach dark places sooner = less noise.

Hdr maps can be different in shapes and sizes - how much noise it would give you can see only by testing it :slight_smile:

All of the above are just my assumptions :slight_smile:


This will be possible when I get the custom view properties panel done.

I have a ton I could add to the library, thanks for the reminder. In my experience, blurring the HDRI reduces GI noise.

1 Like

As I already mentioned for skylight environment texture gets heavily blurred already (and scaled down).