The early system was optimised for performance rather than visual accuracy. Every torch was sorted into 4×4 voxels, without consideration for elevation for practical reasons. The data structure was fixed, meaning dynamic lights were not supported.

When illuminating an object, it would only consider light from directly surrounding voxels, and only the n=5 closest light sources, leading to broken lights in many places.

A solution to many of these issues was, for the inspected papers, a dynamic spatial data structure that would automatically filter for level of detail.

A Barnes-Hut data structure subdivides the space into clusters that increase in density around areas of interest, while condensing areas more in the background or out of view overall.

Light sources were organised into a tree structure by spatial proximity to form these light clusters for the light cuts to use. For time reasons, no performant way was found to add or move light sources dynamically, so the data structure is rebuilt every frame, leading to much room for improvement.

In a landmark paper published at ACM SIGGRAPH in 2005, researchers published this method, which takes a tree structure and traverses the structure from the top-down to spare on performance.

If a node is judged accurate enough by an approximation metric, the sub-tree is ignored and condensed cluster values are used for rendering. If more accuracy is needed, the child clusters are explored, until reaching individual light sources in or very close to areas of interest.

This leads to an implicit level of detail filtering, that can further be tuned to fit the filtering method used in the baseline to define areas of interest more clearly.

This project was inspired by the paper “Stochastic Barnes-Hut Approximation for Fast Summation on the GPU“ (Madan et.al., 2025) published at ACM SIGGRAPH.

The light cuts system was applied from the paper “Lightcuts: a scalable approach to illumination“ (Walter et.al., 2005), also published at ACM SIGGRAPH.