[Bf-cycles] Feature: Per-Object or Per-Material Samples?

Zauber Paracelsus zauberexonar at gmail.com
Sat Nov 22 00:30:04 CET 2014

Okay, so I've gotten a native 64bit linux build of blender compiled, but 
sadly without GPU acceleration.

I started off with rendering a scene that I knew was fairly intensive, 
so that I could give AS a good stress test.  I used this scene I 
rendered last night ( http://fav.me/d874cm8 ), which is fairly intensive 
due to the atmospheric layers.  Both moons have their atmospheres as 
additional meshes, with the moon in the back having one atmosphere layer 
and the moon in the front having two.

For this, I used the following settings:
Size: 256x256
Samples: 256 Branched PT
Light Paths: Transparency 8,8; Bounces 1,8; Diffuse 2; Glossy 4; 
Transmission 8; No Caustics; Filter Glossy 10
Adaptive Sampling: 90% confidence, map update rate 16x
Tile Size: 16x16

Without Adaptive Sampling:   2:10 minutes
With Adaptive Sampling:   1:25 minutes

So, that's about a 35% speed increase with AS enabled.  Adaptive 
Distribution, of course, isn't useful with CPU rendering, where 
efficiency favors smaller tiles.  Though, that did give me a thought for 
running another test.  In the test above, I used a tile size of 16x16, 
but Blender can go lower than that.  The minimum tile size is 8x8, which 
normally has no worthwhile speed advantage over 16x16.  So, I ran the 
above test again, using 8x8 tiles.

Without Adaptive Sampling:   2:10 minutes
With Adaptive Sampling:   1:18 minutes

So, with 8x8 tiles, there's no real difference without adaptive 
sampling, but with it, there is a small gain.  For comparison sake, I 
fired up a downloaded copy of blender, and rendered the same scene in 
GPU mode with a 256x256 tile.

GPU Rendering, no AS:   1:06 minutes

Normally, my GPU renders twice as fast as the CPU.  But with Adaptive 
Sampling, CPU render times get brought within striking distance of GPU 
rendering.  So, I'm very impressed with this.

More information about the Bf-cycles mailing list