Title: Constant-Time Ray Queries on Gaussian Splatting via Directed Distance Fields

Abstract:

3D Gaussian Splatting (3DGS) enables the real-time rendering of novel viewpoints for a scene. However, similar to conventional rasterizers, it is limited to processing primary rays—those originating from the camera and passing through the image plane—and is incapable of tracing the secondary rays required for effects such as global illumination, ambient occlusion, and shadows. To overcome this limitation, we transform a trained 3DGS scene into a ray oracle by distilling a Directed Distance Function (DDF). This compact neural field accepts a ray, defined by its origin and direction, and outputs the distance to the nearest surface along with a binary indicator of whether a collision occurs. Each query is resolved in a single forward pass. The resulting field occupies only 52 MB of memory; because its size is independent of the number of Gaussians, both its computational cost and memory footprint remain constant as the scene complexity increases.

Our work highlights three key contributions. First, we investigate the optimal supervision signals for training a DDF. We find that depth maps rendered from Gaussians lack the sharpness needed to teach thin structures, whereas clean distance supervision successfully recovers these details. Second, we evaluate performance, demonstrating that the DDF operates 26 to 72 times faster than sphere tracing an equivalent signed distance field. Unlike a bounding volume hierarchy constructed over Gaussians, which scales with scene size, the DDF maintains consistent query times and memory usage regardless of complexity, even when leveraging dedicated RT-core hardware. Third, we present a mesh-free pipeline: images are used to generate a 3DGS scene, which yields a neural surface providing clean distances that the DDF then learns from.

We utilize the DDF as a secondary-ray oracle to achieve global illumination. Across 142 objects and real-world captured scenes, the method reproduces reference ray-traced shadows with a PSNR of 30.3 dB and ambient occlusion at 21.3 dB. The source code is publicly available at https://github.com/smlab-niser/ddf-gs.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC