Despite voluminous research in global illumination and radiosity over the past decade, few practical applications have surfaced in the fields that stand the most to benefit: architecture and lighting design. Most designers who use rendering software employ it in a purely illustrative fashion to show geometry and style, not to predict lighting or true appearance. The designers cannot be blamed for this; rendering systems that promote flash over content have been the mainstay of the graphics industry for years, and the shortcuts employed are well-understood by the software community and well-supported by the hardware manufacturers.
Why has radiosity not yet taken off in the rendering market? Perhaps not enough time has passed since its introduction to the graphics community a decade ago . After all, it took ray-tracing nearly that long to become a mainstream, commercial rendering technique. Another possibility is that the method is too compute-intensive for most applications, or that it simply does not fulfill enough people's needs. For example, most radiosity systems are not well automated, and do not permit general reflectance models or curved surfaces. If we are unable to garner support even from the principal beneficiaries, designers, what does that say of our chances with the rest of the user community?
Acceptance of physically-based rendering is bound to improve, but researchers must first demonstrate the real-life applicability of their techniques. There have been few notable successes in applying radiosity to the needs of practicing designers . While much research has focused on improving efficiency of the basic radiosity method, problems associated with more realistic, complicated geometries have only recently gotten the attention they deserve . For whatever reason, it appears that radiosity has yet to fulfill its promise, and it is time to reexamine this technique in light of real-world applications and other alternatives for solving the rendering equation .
There are three closely related challenges to physically-based rendering for architecture and lighting design: accuracy, generality and practicality. The first challenge is that the calculation must be accurate; it must compute absolute values in physical units with reasonable certainty. Although recent research in global illumination has studied sources of calculation error , few researchers bother to compute in physical lighting units, and even fewer have compared their results to physical experiments . No matter how good the theory is, accuracy claims for simulation must ultimately be backed up with comparisons to what is being simulated. The second challenge is that a rendering program must be general. It is not necessary to simulate every physical lighting phenomenon, but it is important to do enough that the unsolvable rendering problems are either unimportant or truly exceptional. The third challenge for any rendering system is that it be practical. This includes a broad spectrum of requirements, from being reliable (i.e. debugged and tested) to being application-friendly, to producing good results in a reasonable time. All three of the above challenges must be met if a physically-based rendering package is to succeed, and all three must be treated with equal importance.
Radiance is the name of a rendering system developed by the author over the past nine years at the Lawrence Berkeley Laboratory (LBL) in California and the Ecole Polytechnique Federale de Lausanne (EPFL) in Switzerland. It began as a study in ray-tracing algorithms, and after demonstrating its potential for saving energy through better lighting design, acquired funding from the U.S. Department of Energy and later from the Swiss government. The first free software release was in 1989, and since then it has found an increasing number of users in the research and design community. Although it has never been a commercial product, Radiance has benefited enormously from the existence of an enthusiastic, active and growing user base, which has provided invaluable debugging help and stress-testing of the software. In fact, most of the enhancements made to the system were the outcome of real or perceived user requirements. This is in contrast to the much of the research community, which tends to respond to intriguing problems before it responds to critical ones. Nine years of user-stimulated software evolution gives us the confidence to claim we have a rendering system that goes a long way towards satisfying the needs of the design community. Further evidence has been provided by the two or three design companies who have abandoned their own in-house software (some of which cost over a million dollars to develop) in favor of Radiance.
In this paper, we describe the Radiance system design goals, followed with the principal techniques used to meet these goals. We follow with examples of how users have applied Radiance to specific problems, followed by conclusions and ideas for future directions.