English: A comparison of classical radiosity and path tracing for a simple scene, with albedo of 0.85 for all surfaces.

The top image shows the scene rendered using radiosity, implemented with Python code (using the "numba" library to speed up some portions). Surfaces are divided into 32x32 meshes (32x64 for the sphere). The final rendering uses ray casting, and colors the pixels using bilinear interpolation of radiosity values (extrapolation at edges). Some very faint noise may be visible due to dithering added to avoid visible "mach bands" when the pixels are quantized to 256 levels. Sampling was used to estimate form factors (rather than analytical formulas). Matrix inversion (via the numpy library) was used to solve the radiosity equation. The image was rendered at 2048x2048 and downsampled to 512x512 to anti-alias the edges.

The bottom image shows the scene rendered with path tracing, using the PBRT renderer (the code accompanying the book Physically Based Rendering: From Theory to Implementation). 1024 paths were traced per pixel. A box filter was used (for consistency with the radiosity output, which used downsampling with a simple box filter).

Even though the rendering algorithms are very different, because both are physically-based methods the output is in close agreement, except for places where the mesh is still too coarse in the radiosity rendering. The path-traced image has visible noise and took much longer to render (7 minutes vs. 1 minute). Most of the time for the radiosity rendering was for estimating form factors, which used single-threaded code that was not optimized.