Larry Gritz, author of BMRT, lead developer of the great and regrettably discontinued Entropy renderer, co-author of nVidia’s Gelato renderer and all time RMan guru, suggested a “Gritz Test” to distinguish between renderers on the Highend 3D ‘Rendering Theory’ mailing list a while ago:
“[…] I suggest a sort of Turing Test — treat the system as a black box and call it a ‘real’ NURB/subdiv[ision surface]/whatever [renderer] if you cannot distinguish the approximative nature of it. In other words, I consider PRMan (and Entropy) to render NURBS natively, because you specify NURBS and you get pixels that show no approximative errors. By the Turing Test (he didn’t suggest it, exactly, so maybe it‘s the ‘Gritz Test’ 
, it’s as ‘real’ as it gets, and the fact that there’s a discretization approximation under the covers doesn’t matter, since it’s not detectable from the outside.”
Larry in particular refers to the fact that if you set the number of subdivision steps manually in your modeling application, there is always a view imaginable (just think of zooming or dollying in on your geometry), where the chosen subdivision depth will be too small and reveal that the geometry rendered is not a subdivision surface [or NURB patch] but a heavy polygon mesh.
Moritz– nice little article.
You may remember me from such productions as ‘Animalia’ (I’m one of the joonya’s that sat next to you).
Question after reading the article– If Subdivision is recalculated frame by frame in order to optimize micropolygon numbers, is it ever likely to encounter artifacts caused by a timestep and resulting mesh change. That is to say if the topology of the surface is changing to incorporate more or less polygons will the change every lead to problems. I wondering both about possible silhouette issues and texture problems. I suppose what I’m really asking is if I was to crank the shading rate to 50 or 100 (assuming it lets me and the control surface is suitably low poly) would the mesh pop in a random fashion either side of the ‘limit surface’ or would the subdivision remain, by some miracle, relatively uniform .
Obviously I would assume such variations would be next to impossible to detect on a moving object with a suitable shading rate but all the same I wonder.
Myles
Yes, it is indeed so that in REYES, the dicing of a mesh will change every frame, if that mesh is deforming/transforming and/or the camera is transforming or changing parameters, relative to the gprim in question.
This can lead to crawling artifacts on very slowly deforming geometry (slowly, as in: little change over many frames and possibly motion blur not being used to worsen things). Particulalrly displacement/bump mapping is prone to expose the problem.
Note, however, that this is all happening in alignement with the primtives’s UV parameterization which (supposedly) is the same, direction wise, over a sequence of frames. This and the fact that motion blur is almost always used in images coming out of REYES renderers, helps hiding the issue, if there is any visible one.
Consider the alternative in a ray tracer: every pixel is sampled (many times). Were a ray hits the primitive (aka: the ray-tracer’s “shading rate”) has nothing to do with the UV parameterization of the surface at all. Even with correct filtering, you can get similar popping of features but it will likely be much more pronounced than in a REYES renderer.
la idea muy buena
RSS feed for comments on this post. / TrackBack URI