Researchers Propose 5D Spatio-Directional Hash Encoding to Improve Neural Rendering

A new research paper from Philippe Weier, Lukas Bode, Philipp Slusallek, Adrián Jarabo, and Sébastien Speierer introduces a compact five-dimensional neural encoding designed to handle high-frequency signals across both spatial and directional domains. The approach extends the hash-grid encoding method pioneered by Müller et al. (2022) into the directional domain by representing directions on a hierarchical geodesic grid, where each vertex stores a learnable latent parameter. This sidesteps the distortions, singularities, and discontinuities that arise when conventional Cartesian-space encodings are applied to directional data.

The authors demonstrate the encoding's practical value in neural path guiding — a technique used in physically based rendering — where their method outperforms the current state of the art by up to a factor of 2 in variance reduction for the same sample count. For film and VFX rendering pipelines that rely on path-traced rendering, the improvement in noise reduction efficiency could translate to meaningful reductions in render time or sample budgets.