Much has been made of the importance of global illumination for producing realistic computer graphics renderings. The presence of shadows, specular reflections, caustics, and diffuse interreflection provide important cues to the human visual system and help determine relationships between objects. But how correct do these details need to be in order to be convincing? What does the human visual system really require in terms of accuracy to make its judgements? How true to life does an image need to be to appear real? We observe that distant global lighting interactions may be randomly perturbed and in some cases eliminated with no change in the visual perception of a scene's realism. The same cannot be said for local interactions, which are vital for low-level and high-level visual processing. We propose a simple method for determining when an interaction is local and therefore visually relevant for presenting a plausible (though not necessarily accurate) rendition of a scene. Specifically, we conjecture that the following criteria must be met for accurate lighting interactions between two objects to be relevant to perceived realism:

• The interacting objects must be "nearby" each other and both visible.
• The interactions must be consistent with a plausible context (environment).

We define objects to be nearby one another if the distance between them is less than the larger of the two objects, and visible if at least some part of each is unoccluded and distinguishable from the background. The second rule is more difficult and requires some explanation; two interactions are said to be contextually linked if they relate to a third object, which may be nearby, distant, or fictional. A ready example is a shadow generated by a particular light source. The shadows cast by two local objects onto a third object, such as the floor, must be consistent with one another; i.e., they must correspond to roughly the same light source position. Likewise, specular highlights in the two objects must be consistent with a single light source configuration or environment. However, these two separate phenomena, shadows and highlights, need not be consistent with one another, or even plausible, to appear realistic. Furthermore, objects not nearby to the two objects in question may be consistent with a different environment with little or no loss in realism.

Based on the above conjecture, we propose the following redefinition of local illumination in graphics rendering, and suggest methods whereby rendering costs may be reduced while satisfying the need for realism. (Indeed, many CG lighters and shader writers regularly employ these tricks, but we seek to formalize them for the purposes of automation.) Comprehensive Local Illumination is the interaction of light from some plausible but unidentified illuminating environment with nearby (local) objects, including caustics and interreflections between these local objects.

We specifically include higher order interactions in our definition of local illumination because color bleeding and caustics do matter for nearby objects. This suggests that global illumination methods based on linking distant interacting groups (e.g., Rushmeier, Patterson and Veerasamy, GI '93), may have a good approach. In fact, we believe that a simple modification to this approach will provide equally realistic results at a substantially reduced cost, by dropping all links whose length significantly greater than the largest dimension of the two groups involved.

Another simplification to rendering complexity is to group light sources in a hierarchy, replacing individual light sources with group sources and environment maps past a certain distance. To maintain consistency on nearby objects, we suggest using the same grouping methods used for global illumination to assign environment maps and surrogate light sources to locally interacting subsets of the environment. When two groups are linked by the above proximity criterion, heuristics may be employed to determine tolerable differences between the associated environment maps and source groupings. In this way, each group will be linked to a set of nearby groups for the purposes of local interreflection, and to an environment map and surrogate source list for the purposes of generating shadows and reflections. Between distant, unlinked groups, there need be no restriction on how different the source grouping and environment maps may be. Conventional ray tracing or scanline rendering techniques may then be employed with the following advantages:

1. Each group requires only local geometry for rendering -- memory requirements are greatly reduced by rendering groups one at a time and compositing them afterwards.
2. Global illumination is reduced to a small, sparse matrix for interacting groups, which may be solved quickly with little memory overhead using full-matrix radiosity.
3. Local interreflection may be computed using any combination of object-space and image-space algorithms, since complexity is held in check by grouping and image segmentation.
4. Expensive caustic rendering may be applied to obtain local specular-to-diffuse illumination, since distant interactions are ignored and combinatorial complexity is thus avoided.

Extending the methods described to animation requires a method to smoothly link and unlink groups as they move together and apart, and morph environment maps and source groupings in such a way that objectionable time-discontinuities are avoided. One very simple approach to this problem is to track positions of objects forward and backward in time, and at each frame interpolate between discontinuous links and maps from the past and future. Links may then be forged and broken at will, and discontinuous changes made to environment maps and source groupings, provided that each parameter maintains its value for a minimum span of time. Worst case, this doubles the amount of time spent rendering each frame, but this is a small price to pay for smooth animation.

None of this should be construed as an indictment of global illumination, a field that is very close to my heart. However, the goal of realism has never been very well-defined in graphics, and hence a good deal of effort has been spent to achieve physical accuracy, which is most assuredly overkill if all you want is to produce something that looks real to a human observer. Physical accuracy is indispensable if one is attempting to determine object visibility, glare, or evaluate aesthetics of a particular design or configuration. Fields such as architecture, lighting design, and flight and driving simulation therefore need to target all effects, local and global, in an illuminated environment. They may incorporate perceptual models to compute tone-mapping or even to cut corners during the calculation, but they must in the end reach the same percept as a real environment to achieve their goal. However, for graphics rendering for special effects and other "make-believe" applications, the goal is very different. The distinction between these two applications is paramount, and different techniques are appropriate to each.