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

Zauber Paracelsus zauberexonar at gmail.com
Thu Nov 20 01:42:00 CET 2014

Okay, I just redid my tests.

Here are the settings I used for each:
Image Size:   512x512
Samples:   256 Branched PT (all samples at 1, Sample All Direct, Sample 
All Indirect)
Light Paths:   8,8 transparency; min/max bounces 1,8; Diffuse 2, Glossy 
4, Transmission 8, Caustics enabled, Filter Glossy 10.0
Adaptive Sampling: Map update rate 16, confidence 95%
Compute Device: AMD Phenom II 970 quad-core, 3.5ghz
Render: 4 threads, 16x16 tile size

As a note, I had to start the tests over, because I was using two 
instances of blender to run the tests, one for adaptive sampling, and 
one without.  However, the one without AS was a linux build while the 
one with AS was a windows build running under wine. Because of this and 
unusual results on the second test, I chose to scrap my results and 
start over, doing both tests under the windows build, so as to eliminate 
performance bias from Wine.

The First was a simple scene with just a head shot of a character model, 
rendered to 512x512.
 > Render time without Adaptive Sampling:   3:36.19 minutes
 > Render time with Adaptive Sampling:   54.12 seconds

So, that first one got a substantial speed up, running 3-4 times 
faster.  This next test is a much more complex indoor scene with 
furniture, multiple lights, one character, windows with curtains, and an 
environment map.  This one was rendered to a smaller resolution of 
200x300, and had the bounce count cut in half to save time.
 > Render time without Adaptive Sampling:   7:28.05 minutes
 > Render time with Adaptive Sampling:   5:19.05 minutes

The second test had a less dramatic difference in speed, most likely due 
to the increased complexity of the scene.  Once I can get my hands on a 
64bit Linux build, I'll re-run the tests.  Under Wine, some textures 
wouldn't load and performance was reduced by at least 10-20%.  Once I do 
those tests, I'll redo the tests and post the results.  Also, is this 
working with GPU mode?  Because that's where I do most of my rendering.

Also, where can I find instructions on compiling patched blender builds 
under Linux?

On 11/19/2014 04:51 PM, Colby Klein wrote:
> Lukas: This is doing wonders for my scene. Awesome work! 2:41 @ 
> confidence 75 / rate 30, samples capped at 3000 + adaptive 
> distribution vs 4:05 @ 1000 samples and no adaptive sampling, and the 
> result even looks a little better! I can't wait for this to be in an 
> official release.
> Zauber: Crank up the number of samples in your scene and it'll only 
> use as many as it needs to be confident the tile is noise-free.

More information about the Bf-cycles mailing list