Issue: Volume: 24 Issue: 10 (October 2001)

The Softening of Computer Graphics


When it comes to translucent objects, computer graphics techniques tend to take a hard line. Most rendering systems pay little heed to the fact that light scatters inside translucent materials and leaves the surface at a different location from where it entered-a physics phenomenon called subsurface scattering. Rather, the majority of rendering algorithms are built using bidirectional reflectance distribution functions (BRDF), which assume that light entering at a point on a surface reflects at that same point. In reality, this is true only for metal objects. In all other materials, particularly translucent ones, light scatters.

This light-scattering effect can be seen when a laser pointer is shone on a translucent material. The area around the point of illumination starts glowing because of the light scattering inside the material. By ignoring this physical reality, many computer-generated representations of translucent objects, such as snow, milk, or skin, have an unnatural, hard appearance. Even for materials that don't seem very translucent, the BDRF methods create a hard, CG-looking effect, because they don't locally blend surface features such as color and geometry.

It's not that software developers are unaware of the physics of light, but they often choose to overlook certain aspects of it given the incumbent computational challenges. Existing techniques for simulating subsurface scattering are computationally impractical, says Henrik Wann Jensen, a researcher in the Computer Graphics Lab at Stanford University. "Typically, such techniques have involved costly Monte Carlo ray-tracing or photon-mapping approaches that simulate the scattering of individual photons," he says. "These methods are particularly costly for materials such as milk and skin, where the photons can scatter hundreds of times before leaving the material."
Traditional shading techniques give "Diana the Huntress" a hard look (left) because they fail to capture light glowing through the nose and hair. The softer-looking BSSRDF model (right) renders translucency correctly.

In an effort to soften things up without breaking the computational bank, Jensen and Graphics Lab researchers Stephen Marschner, Marc Levoy, and Pat Hanrahan have developed a Bidirectional Scattering Surface Reflectance Distribution Function (BSSRDF) that samples the surface of a translucent material and approximates the diffusion of light inside it. Instead of following the paths of individual protons, the system relies on a rapid, image-based measurement technique for determining the optical properties of translucent materials. Consequently, it is many orders of magnitude faster than Monte Carlo-based raytracing. In fact, says Jensen, the researchers are able to render translucent materials almost as fast as the traditional "hard" rendering techniques. And, he says, "our model is accurate enough that it allows us to measure the scattering properties of materials simply by fitting our theory to the scattering profile observed when the material is illuminated with a laser pointer."

In developing the new technique, the researchers' main challenge was understanding how to efficiently compute the diffusion of light within materials. Toward this end, says Jensen, "we benefited from work in the medical physics community, where light-scattering in tissue is studied in order to improve the existing light-based techniques for treating skin disorders."

Since the BSSRDF samples only the surface of structures, it doesn't know anything about the internals of the materials. Thus, says Jensen, its efficacy is limited to materials where the internal structure is invisible. "In the future," he says, "we will most likely be looking into how the internal structure of a material can be rendered as well."

Although the technique is slightly more complicated to implement than traditional rendering, and while it takes a bit longer to compute, the visual payoff is significant enough that it should be the natural choice for translucent materials, says Jensen.

For example, he says, an immediate application is in movies. "Instead of trying to hack the appearance of translucent materials, which is very common but does not work very well, our model makes it possible to correctly render many of the subtle effects that give the soft, smooth appearance of natural materials."
Subtle lighting effects such as color bleeding below the nose and the soft appearance of the skin can be seen in the BSSRDF rendering of this face (top). The traditionally rendered model (bottom) appears less natural.

This can be particularly advantageous when it comes to skin. "Currently, there is a huge interest in the movie industry in rendering a convincing synthetic human. Skin is very translucent and looks too hard if rendered with traditional shading techniques."

Given the possibilities, several movie companies have already expressed interest in the technique, and, says Jensen, "I have already heard of films coming out where the BSSRDF model has been used to render such things as ice and skin." In addition, Jensen is consulting with Pixar to implement the model in its upcoming film Nemo, as well as into the company's RenderMan renderer.

In addition to the obvious commercial possibilities in the film industry, other applications stand to benefit from the new rendering approach as well. "Computer games could gain a lot from the BSSRDF model, although they may have to wait a few years because the computations are still beyond the capabilities of current [real-time] graphics hardware," says Jensen.

More immediate gains may be had by lighting designers and architects who deal with the classic hard-look problem in computer graphics on a daily basis. "I think they will appreciate this technology because it will allow them to experiment with lighting in new ways while still preserving the soft appearance of the material."

More information on the BSSRDF research can be found at