The Rendering Debate, Continued: The Trickle-down Effect of Modeling

As I tried to figure out what rendering software I would be using for the exterior shots, I considered the type of lighting I would be using for these. All of them are exterior sunny, clear-sky lighting scenarios, so I knew I wanted to use HDRI because of the way it can get you to 80% of a finished shot, with the remaining 20% consisting of sweetening with direct lighting and 2D compositing. I've been very happy with the global illumination (GI) of RmF from using it over the past couple of years, because of its efficient raytracing and very reasonable render times, and I've also been happy with the GI/Final Gather (FG) results I've gotten with Mental Ray, but the render times have been significantly greater, and I am hoping to get away with using RmF for my exterior shots.

I started off running some tests in RmF, using an Environment Light with all the geometry in my scene given the default Lambert shader. These tests never exported any images; my scene crashed Maya. My scene's file size was in the range of 15 Mb, which was considerably less than 40 Mb scene files I had successfully rendered in the fall of 2007 with RmF in Maya 8.5. I figured that the file size was probably not the main culprit of Maya crashing, but the geometry itself. I remembered that the scenes I was rendering a year and a half ago had consisted of approximately 80% NURBS, 19% polygons and 1% SubDivisions. I was primarily using NURBS because I knew that they would tesselate perfectly in RmF, and this was before I knew that the same was true for SubD surfaces. The tesselation of NURBS was automatic; I did not have to tag anything or tell the renderer to convert anything at render time. In the case of rendering my exterior city block this weekend, however, I was tagging specific objects in my scene to subdivide at render time.

I admit that I had a brute-force approach to modeling this scene; I created everything, from the brick facades down to individual window pane slats, with the intent of subdividing at render time. (See this post on subdividing.) I was doing this in order to get beveled edges to catch glints of light (and actually save myself the hassle of adding in lights to catch edges of objects) as well as give myself the highest resolution in my modeling so that I would get realistically rounded edges even at the closest of camera angles. The result was an exorbitant number of polygons in places where I would not need any subdividing at all. Take this window frame for example:
These pieces of geometry all have "shored up" edges (i.e. edge loop cuts created to help the geometry retain its basic shape while providing "beveled" or rounded edges.) It's a reasonable modeling strategy for a camera angle this close, but duplicate the geometry literally 50 times and throw it as far away from the camera as this...
...and you have an unnecessarily large number of polys that no one cares about. Because I was subdividing virtually every object in the scene (except for the glass, which is represented by the green transparent shader), I ended up crashing Maya over the most basic of renders. Again, this model had the default Lambert shader on it (except for the basic partially transparent green Blinn on the windows) - nothing close to the texture mapped, lit and raytraced scene I would be rendering for production.

I ended up having to completely remove much of the geometry that was to be subdivided, and I replaced it with basic geometry. Essentially, an object with edge loop cuts like this:
would get replaced by a primitive object like this:
After doing this to the majority of the scene (I decided to retain subdivided objects on the hero apartment, since I knew I would be rendering plenty of shots with closer camera angles), I was able to successfully render out some test frames in RmF.