Directed Distance Fields for Constant-Time Ray Queries on Gaussian Splatting

Subhankar Mishra · 2026

Read Directed Distance Fields for Constant-Time Ray Queries on Gaussian Splatting by Subhankar Mishra (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-00817.txt.

Introduction

Mishra turns a fast graphics method into a ray tool in this 2026 paper. 3D Gaussian Splatting renders new views of a scene in real time. As a rasterizer, it answers only camera rays. It cannot trace the extra rays that shadows and global light need.

Mishra distills a trained scene into a Directed Distance Function. This small neural field takes a ray and returns the range to the first surface, plus whether the ray hits at all. One forward pass answers each query, in flat time. The field is 52 megabytes, and its size does not grow with the scene. Clean distance teaching beats blurry depth from the Gaussians, he finds, and thin parts survive. The field runs 26 to 72 times faster than sphere tracing a signed distance field. It serves as a side-ray oracle for global illumination, with no mesh needed.

Catalog record

CREATORSubhankar Mishra
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE29 KB
LICENSEhttp://creativecommons.org/licenses/by-sa/4.0/
RIGHTSCC BY SA
SOURCEarXiv

Source: https://arxiv.org/abs/2606.00817

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Directed Distance Fields for Constant-Time Ray Queries on Gaussian Splatting Subhankar Mishra Abstract 3D Gaussian Splatting (3DGS) renders new views of a scene in real time. Like every rasterizer, it answers only primary rays, the rays from the camera through the image. It cannot trace the secondary rays that shadows, ambient occlusion, and global illumination need. We turn a trained 3DGS scene into a ray oracle by distilling a Directed Distance Function (DDF). The DDF is a small neural field. It takes a ray, given by an origin and a direction, and returns the distance to the first surface and whether the ray hits anything. Each query is one forward pass. The field is 52 MB, and its size does not depend on the number of Gaussians, so its cost and memory stay flat as the scene grows. We make three points. First, we study what supervision a DDF needs. Depth rendered from the Gaussians is too blurry to teach thin parts, while clean distance supervision recovers them. Second, we measure speed. The DDF is 26 to 72 times faster than sphere tracing an equivalent signed distance field, and unlike a bounding volume hierarchy built over the Gaussians, even on dedicated RT-core hardware, its query time and memory do not grow with the scene. Third, we show a pipeline that needs no mesh: images give a 3DGS scene, a neural surface gives clean distances, and the DDF learns from them. We use the DDF as a secondary-ray oracle for global illumination. It reproduces reference ray-traced shadows at 30.3 dB and ambient occlusion at 21.3 dB across 142 objects, and on real captured scenes. Our codes are available at https://github.com/smlab-niser/ddf-gs . Fig. 1: We distill a Directed Distance Function (DDF) from a 3D Gaussian Splatting scene. Multi-view images are fit to a 3DGS scene that holds appearance. We then distill a DDF, a small field that returns, for any ray with origin \mathbf{x} and direction \mathbf{d} , the distance t to the first surface and a visibility value, in one network evaluation. A rasterizer gives only the primary view. The DDF answers arbitrary secondary rays, so it can produce shadows, ambient occlusion, and global illumination. It is 26 to 72 times faster than sphere tracing an equivalent signed distance field, and its memory does not depend on scene size. I Introduction 3D Gaussian Splatting (3DGS) [ 10 ] represents a scene as a set of 3D Gaussians and rasterizes them into an image. It produces high quality novel views and runs in real time. This has made it a common choice for scene capture and display. A rasterizer answers one kind of query. For each pixel it finds what the camera sees along the ray through that pixel. These are primary rays. Many useful effects need a different kind of query. To cast a shadow we ask whether a point on a surface can see the light. To compute ambient occlusion we ask how much of the sky a point can see. To trace one bounce of global illumination we ask what a surface point sees in some direction. These are secondary rays. Their origins and directions are arbitrary and are not known in advance. A rasterizer cannot answer them. There are two ways to add ray queries to a Gaussian scene, and both have a cost. The first is to trace the Gaussians directly. Recent methods [ 13 , 4 ] build a bounding volume hierarchy (BVH) over a proxy shape for each Gaussian and traverse it on ray-tracing hardware. The structure grows with the number of Gaussians, so its memory is O(N) and its traversal cost rises as the scene gets larger. It also depends on ray-tracing cores. The second way is to fit a signed distance field (SDF) to the Gaussians and sphere trace it. Sphere tracing takes many small steps along each ray, and each step is a full network evaluation, so the cost per ray is high. We take a third way. We distill a Directed Distance Function (DDF) from the trained Gaussian scene. A DDF maps a ray to the distance to the first surface along that ray and to a visibility value. One forward pass answers one ray. The network is small and fixed: 52 MB, with a size that does not depend on the number of Gaussians. So a DDF turns a Gaussian scene into a ray oracle whose cost is constant per ray and whose memory is constant in scene size. Directed distance fields have been studied before, for shape representation [ 1 ] and for path tracing [ 3 ] . They were not distilled from a Gaussian scene, and they were not compared against sphere tracing an SDF, which is the natural baseline. This paper studies the DDF as a ray oracle for Gaussian and implicit scenes. Our contributions are: • A DDF distilled from a Gaussian Splatting scene. It is a compact, constant-time, constant-memory ray-intersection oracle that needs no ray-tracing cores. It returns distance and visibility along any ray in one evaluation (Sec. III ). • A study of what supervision a DDF needs to learn thin geometry. Depth rendered from the Gaussians blurs thin parts and the DDF cannot learn them. Clean ray-cast distance recovers them. The limit was the supervision, not the network (Sec. IV-B ). • A speed and scaling study. The DDF is 26 to 72 times faster than sphere tracing an equivalent SDF (Table IV ). Its query time and memory are flat in scene size, while a BVH over the Gaussians grows in both (Table III , Fig. 3 ). • A mesh-free oracle. Images give a Gaussian scene and a neural surface, and the DDF learns from them with no mesh at any stage (Table V ). We use it as a global illumination oracle across 142 objects (Table VI ) and on real captured scenes (Fig. 5 ). II Related Work Gaussian Splatting and ray tracing. 3DGS [ 10 ] fits a set of anisotropic 3D Gaussians to multi-view images and rasterizes them for fast novel-view synthesis. Rasterization only serves the camera view. To support arbitrary rays, recent work traces the Gaussians. 3D Gaussian Ray Tracing [ 13 ] and RaySplats [ 4 ] build an acceleration structure over per-Gaussian proxies and traverse it on ray-tracing cores. This gives correct intersections but its memory and traversal cost grow with the number of Gaussians. Our DDF answers the same intersection query in one evaluation, with a cost that does not depend on the number of Gaussians. Neural implicit surfaces. A signed distance field stores the distance to the nearest surface at every point [ 15 ] . NeuS [ 17 ] learns an SDF from images by volume rendering. A multiresolution hash grid [ 14 ] makes such fields fast to query and train. An SDF gives a clean surface, but to find where a ray meets that surface one must sphere trace it, which takes many evaluations per ray. A DDF answers the same question in one evaluation because the distance it stores is already measured along the query direction. Surface and SDF from Gaussian Splatting. Several methods recover geometry from

Read the complete text