Mark,
Awesome pictures. What are the limitations of your low -ad and high over sampling approach? When is appropriate and when is it not appropriate?
Mark,
One disadvantage is hard drive space: how many 3 GB images can I keep around? Another problem that occurs more with some pieces then others is speckles caused by particularly bright pixels. Since pfilt works on the full dynamic range, a single super-bright pixel in the full-res image can cause a significant bright spot in the final reduced-res image. Note the speckles in this image:
http://mark.technolope.org/image/p54.html
http://mark.technolope.org/image/img54_detail_900.png
That second image is at print resolution, and was filtered down from a 4x or 6x resolution original. That scene is very contrasty. Some of my more recent work mitigates the speckling by rendering in a closed room: allowing enough bounces to provide a more even background.
I find that "-aa 0" is appropriate when the octree for a scene with little overlapping geometry takes more than a few hundred MB to create. Or for scenes in which I want detailed interreflection among a great number of very tiny surfaces (probably the same thing). It seems to me to be an easy way to get rid of the splotches that plagued some of my earlier renderings.
Probably not something you are concerned about in your work, but when you use such a low -ad setting and such a high over sampling rate, I'd like to know what effect there is on overall light level accuracy in a typical interior space with daylighting and/or electric lighting?
I have not studied this at all. I could posit that the low -ad with high oversampling rate allows me to bias the interreflection toward the first (and most important) bounce. Shooting 36 rays into a final image pixel (what 6x oversampling does) but letting each of those only spawn 8 child rays (-ad 8) at the second and deeper levels effectively means that I'm doing a more accurate 1st bounce.
I guess there is no over sampling on an rtrace calculation, but for an irradiance calculation image, my -ad setting is usually in the thousands and my ambient bounces are at 4 or so (depending on the results of a parameter test) I'm guessing the -ab and -ad parameters can't be reduced if light level accuracy is important. Maybe I'm wrong? Sure would be nice to slash some of those crazy calculation times.
My impetus was to slash the computation times and fit the problem in memory. Of course, back then I had 2 GB of RAM *at most* and no access to anything better than a dual-proc machine. Now I am blessed with access to a 32 GB monster 8-core intel box and a number of medium-sized clusters.
One other advantage of "-aa 0" is that one pixel trace does not depend on another, so parallelization is trivial. I use rpiece almost exclusively for my big jobs now. OpenMP or pthreads anyone?
Mark
ยทยทยท
On Thu, 28 Dec 2006, Mark de la Fuente wrote: