constructive trees

This paper is a modern answer to a problem older references mostly leave open: even if sphere tracing is mathematically safe, evaluating large constructive SDF trees can still be too expensive. The authors focus on the cost of field evaluation itself and introduce optimization nodes embedded directly in the construction tree, rather than relying only on external acceleration structures. That makes the paper valuable when you want to discuss performance as a property of the field representation and its composition, not just the number of march steps.

It is also distinctive because it explicitly preserves the conservative or Lipschitz-style guarantees that make sphere tracing trustworthy in the first place. The paper’s proxy and continuous level-of-detail nodes are designed to reduce evaluation cost without breaking the lower-bound behavior the marcher depends on, and the reported GPU results are measured in terms of sphere-tracing performance on rendered objects. So this is a strong citation for the modern frontier of SDF work: acceleration that respects the mathematical contract of distance-guided marching instead of treating it as disposable.