[Bf-committers] Super-sampling and anti-aliasing
tevhadij at gmail.com
Tue Sep 21 21:45:55 CEST 2010
On Tue, Sep 21, 2010 at 2:02 AM, Damir Prebeg <blend.factory at gmail.com> wrote:
> Use http://pasteall.org/. There you can post your code, blends and images.
This is a very nice tool!
As I said, I would first like to make sure I tuned the parameters the
best way possible in Blender so we can move on and compare with other
supersampling alternatives. Here is a .blend file containing a large
plane covered by a checkerboard texture (created using nodes).
I admit that in my ignorance this was the only way I found to produce
a procedural checkerboard. If there are better procedural
alternatives, please let me know. Texture images are not a valid
alternative because we are analyzing supersampling, and we do not want
texture filtering to interfere.
I then proceeded to render this file on a fresh build of Blender 2.5
from svn (revision 32043).
Even though I selected the "Anti-Aliasing" and "Full Sample"
checkboxes, chose 16 "samples per pixel", and tried every filter
available, renderings look as though the procedural texture is being
sampled only once per pixel. Is this the case? If so, how can I change
this? Here are sample renderings:
Catmull-Rom (Blender) http://www.pasteall.org/pic/5910
Gaussian (Blender) http://www.pasteall.org/pic/5911
I was expecting something similar to this:
Catmul-Rom (16spp) http://www.pasteall.org/pic/5915
This last image was produced by an infinite checkerboard renderer
using 16 jittered samples per pixel, with the Catmull-Rom filter
covering a 4x4 pixel neighborhood around each image pixel. It follows
an algorithm that is very similar to what Blender's source code seems
to be doing.
Of course, using more samples per pixel gradually eliminates the
noise, as in the image below, rendered with 121 samples per pixel.
Catmul-Rom (121spp) http://www.pasteall.org/pic/5916
There is really no trick here, except for my concern that I have
somehow misused Blender or I have been wrong to assume procedural
textures would be evaluated in sample-frequency, rather than in
More information about the Bf-committers