Hi All,
Here are a couple of renders I put together using Cycles. I was trying to see how clean Cycles can get. But as you can see, even after thousands of samples there is still considerable noise.
Part of the problem I see is there is no falloff for light or reflection. So you can not attenuate light with the current nodes as they are. The caustics and reflection seem to generate the most noise.
I appended your noise reduction node group into my main scene. I’m not sure if I hooked it up correctly, however. If you would like to play with the scene, here it is.
26_cycles_puzzle_pieces_NR.blend (846 KB)
@Atom
I doubt you can easy decrease noise, many light sources + at least 1 translucent object boundary = need 3 bounces to reach light, too low probability.
Falloff work as physically based light wave, it fall from light distance based on 1/(dist*dist) formula. Maybe try to render one light at a time and combine result later in composer, and very important you need not only set other light color/strength to 0 but completely delete all light sources except one.
There you go!
Firstly I plugged your translucent shader (which is diffuse transmission as opposed to diffuse reflection) into the color bleed input, and raised up to 1 the filter factor for all nodegroups instances. Then i actually revised the math nodes, so that they now also work with refractive shaders. I tried it at low final samples (300), still noticing noise reduction. Now i’m really curious to see how it behaves on longer rendertimes!
(first picture the trick is off)
26_cycles_puzzle_pieces_NR2.blend (790 KB)
Here is the re-render of the scene with your suggested node fixes.
I put them on top of each other in Photoshop so I could A-B them.
There is not really any noise reductions, that seems to stay the same. What I do see happening is that I lose some of the scattering, generated by the checker pattern, that is apparent in the first image. Your noise reduction might be thought of more as a light smoother/averager. The actual pixels generated, however still seems to be just as noisy as ever.
Notice how the window pattern from the checker texture is more lost in this image.
i find it less noisy and with no fireflies, but of course the big issue here is that checker pattern shadow that’s missing. I’ll take a look, hope it can be tackled. Still this node can turn out good for reflective materials.
It would also be useful to know how soon the convergence goes for each render. I mean, after 9k passes maybe a lot of renders look the same, but what after 500 or 1000, 2000…
So, here is attempt #2!
In order to deal with that checkers shadow, I had to add a mask , which of course uses some inputs already in the setup (the checker pattern and the RGB node)
Attached you’ll find the blend file and 2 samples, the first is filter off, the second is filter on. Notice in the pictures how the modded version cleans faster but with less samples, in the same amount of rendertime. Maybe you should lower the 9600 samples target or, even better, let it run for the same amount of time it took to render your first version, so that you can have a proper Time vs Noise comparison.
Unfortunately it would take ages to me to render this at fullHD, so as usual i’ll tell you… can’t wait to see if this turns out good/acceptable
