B-Rep Distance Functions (BR-DF) How to Represent a B-Rep Model by Volumetric Distance Functions?
Fuyang Zhang, Pradeep Kumar Jayaraman, Xiang Xu, Yasutaka Furukawa, [2pt] Simon Fraser University Autodesk Wayve[2pt] {fuyangz, furukawa}@sfu.ca {xiang.xu, pradeep.kumar.jayaraman}@autodesk.com · 2025
Read B-Rep Distance Functions (BR-DF) How to Represent a B-Rep Model by Volumetric Distance Functions? by Fuyang Zhang, Pradeep Kumar Jayaraman, Xiang Xu, Yasutaka Furukawa, [2pt] Simon Fraser University Autodesk Wayve[2pt] {fuyangz, furukawa}@sfu.ca {xiang.xu, pradeep.kumar.jayaraman}@autodesk.com (2025) 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-2511-14870.txt.
Introduction
Fuyang Zhang, Pradeep Kumar Jayaraman, Xiang Xu, and Yasutaka Furukawa propose a new way to store computer-aided design shapes. Their method, called B-Rep Distance Functions, turns the surface of a CAD model into a signed distance field. The corners, edges, and faces of the shape are each stored as their own distance fields.
An extended marching cubes method turns the stored fields back into a watertight CAD model, and the authors report that this step never fails. They also train a diffusion model that generates new shapes in this form. The approach matches leading CAD generation methods while adding a reliable path back to an editable model.
Catalog record
| CREATOR | Fuyang Zhang, Pradeep Kumar Jayaraman, Xiang Xu, Yasutaka Furukawa, [2pt] Simon Fraser University Autodesk Wayve[2pt] {fuyangz, furukawa}@sfu.ca {xiang.xu, pradeep.kumar.jayaraman}@autodesk.com |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 32 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-14870
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
B-Rep Distance Functions (BR-DF) How to Represent a B-Rep Model by Volumetric Distance Functions? Fuyang Zhang, Pradeep Kumar Jayaraman, Xiang Xu, Yasutaka Furukawa, [2pt] Simon Fraser University Autodesk Wayve[2pt] {fuyangz, furukawa}@sfu.ca {xiang.xu, pradeep.kumar.jayaraman}@autodesk.com Abstract This paper presents a novel geometric representation for CAD Boundary Representation (B-Rep) based on volumetric distance functions, dubbed B-Rep Distance Functions (BR-DF). BR-DF encodes the surface mesh geometry of a CAD model as signed distance function (SDF). B-Rep vertices, edges, faces and their topology information are encoded as per-face unsigned distance functions (UDFs). An extension of the Marching Cubes algorithm converts BR-DF directly into watertight CAD B-Rep model (strictly speaking a faceted B-Rep model). A surprising characteristic of BR-DF is that this conversion process never fails. Leveraging the volumetric nature of BR-DF, we propose a multi-branch latent diffusion with 3D U-Net backbone for jointly generating the SDF and per-face UDFs of a BR-DF model. Our approach achieves comparable CAD generation performance against SOTA methods while reaching the unprecedented 100% success rate in producing (faceted) B-Rep models. Figure 1 : BR-DF is a geometric representation for Boundary Representation (B-Rep) models. An SDF encodes surface geometry. UDFs encode vertices, edges, faces, and their connectivity. An extension of the Marching Cubes converts BR-DF to a faceted B-Rep model. 1 Introduction Volumetric distance functions are fundamental in geometry processing. The introduction of the Marching Cubes algorithm 16 and signed distance functions (SDFs) marked a significant breakthrough in surface reconstruction for the field of computer graphics 1 ; 12 . These concepts were later adopted by the computer vision community through level-set and geometric active contour methods 3 , allowing geometry optimization without explicitly managing the surface topology changes. In the era of deep learning and generative artificial intelligence (GenAI), signed distance functions remain central to geometry reconstruction and generation 2 ; 21 ; 33 , enabling high-quality 3D asserts creation through large-scale training 13 ; 37 ; 7 . Solid modeling in computer-aided design (CAD) focuses on designing and editing shapes intended for real-world manufacturing. Ensuring watertightness is therefore a critical requirement and a key challenge for GenAI models designed to synthesize CAD data structures, such as Boundary Representations (B-Reps) 30 . Signed distance functions, by their very nature, guarantee watertight representations, making them a compelling choice for representing CAD B-Rep model in solid modeling. This paper advances the use of such volumetric distance functions by introducing an innovative representation that is both equivalent to and mutually convertible with a faceted B-Rep model 24 , i.e., a segmented triangle mesh where each segment corresponds to a face in B-Rep, and the boundary curves between segments correspond to B-Rep edges. Our proposed representation, B-Rep Distance Functions (BR-DF), combines a signed distance function (SDF) with per-face unsigned distance functions (UDFs) to fully capture the geometry and topology of a watertight CAD model. Surprisingly, we find that a simple algorithm can reliably convert BR-DF into faceted B-Rep—a process that never fails . Specifically, we introduce a novel extension of the Marching Cubes algorithm that reconstructs the CAD surface boundary from SDF, and then conditioned on the surface mesh and UDFs, recovers the complete B-Rep vertices, edges, and faces together with their topology. Concretely, for a CAD model with F faces, any point on the zero iso-surface of SDF has an associated F -dimensional face UDF scalar fields. A B-Rep face is determined by the minimal UDF value in the face UDF fields. A B-Rep edge is defined when at least two faces share the same minimum UDF value, and similarly a B-Rep vertex occurs where at least three faces share the minimum UDF value. Since our representation is a set of scalar functions on volumetric grids, we present a simple latent diffusion with 3D U-Net backbone for BR-DF generation. Qualitative and quantitative evaluations on DeepCAD 32 and ABC 11 datasets demonstrate that BR-DF based generative model matches the performance of recent SOTA methods 32 ; 35 , while achieving the unprecedented 100% success rate of producing faceted B-Rep models. While faceted B-Reps still needs to be converted into B-Reps using standard mesh-to-BRep conversion tools available in CAD software, the topology generated by our method greatly aids in this process. We believe this representation will play a key role in developing generative models for CAD that guarantee watertightness. To summarize, our paper makes three key contributions: 1) a novel volumetric distance function representation of B-Rep models, named BR-DF; 2) a Marching Cubes and Triangles algorithm to convert BR-DF to faceted B-rep models with 100% success rate; 3) A simple B-Rep generative model based on BR-DF with performance competitive to the state-of-the-art approaches. This paper potentially marks a key innovation in CAD geometry representation—analogous to the impact that Marching Cubes had on raster geometry modeling 25 years ago. 2 Related Work This section reviews related work on distance functions and CAD model generation. 2.1 Signed Distance Functions Signed distance functions (SDFs) have been instrumental in geometry processing. The introduction of the Marching Cubes algorithm 15 enabled the creation of 3D mesh models from SDFs, setting the foundation for many surface reconstruction techniques. The Digital Michelangelo Project 1 is a notable example, which utilized SDFs to capture complex 3D geometry. In the Computer Vision domain, level-set and active contour methods utilized SDFs, allowing implicit surface representation that facilitated topology adaptation 3 . Over time, distance functions became less prominent in Computer Vision, as depth maps, point clouds, and mesh representations gained popularity; however, SDFs combined with Marching Cubes remain fundamental for surface extraction in geometry processing. The integration of deep learning has renewed interest in distance functions. Neural implicit representations, such as multilayer perceptron (MLP), predict SDF values for single-image reconstruction or scan completion 33 ; 21 ; 2 . Volumetric grids are a popular alternative for the SDF representation and can be scaled to high resolutions 20 ; 6 . NeuralRecon reconstructs 3D scenes in real-time from a monocular video by predicting local TSDF volumes 25 . Similarly, Atlas directly regresses a TSDF volume from a set of posed RGB images, utilizing a 2D convolutional neural network (CNN) to extract features from each image 19 . 2.2 Generative Models for B-Reps Recent advancements in boundary representation (B-Rep) generation have focused on learning the construction sequences of sketch-and-extrude operations. This approach allows the generated sequences to be parsed by solid modeling kernels, enabling the rebuilding of B-Reps. DeepCAD 32 pioneered this direction by releasing a dataset of sketch-and-extrude sequences, laying the groundwork for further research. Subsequent works introduced significant improvements: SkexGen 36 leveraged autoregressive Transformer models 28 with disentangled learning of geometry and topology, leading to a substantial enhancement in generation quality. Later, HNC-CAD 34 introduced hierarchical latent codes that capture natural hierarchies in the data structure, better preserving design intent and improving editability. A key advantage of the sketch-and-extrude paradigm is its inherent guarantee of watertightness, as extruded 2D sketches are constrained to form closed loops. However, these methods rely on the availability of explicit construction sequences alongside B-Reps, limiting their applicability to simpler, primitive-based shapes. Direct B-Rep generation creates 3D geometry while maintaining topological consistency among faces, edges, and vertices. SolidGen 8 employs Transformer models and pointer networks 29 to generate B-Rep entities in