Organizations: College of Computer Science, Sichuan University · Engineering Research Center of Machine Learning and Industry Intelligence, Ministry of Education · Dongfang Electric (Chengdu) Innovation Research Co., Ltd. · Southwest China Research Institute of Electronic Equipment
Generative modeling for Computer-Aided Design (CAD) remains fundamentally challenged by the need to capture the intricate interplay between discrete topological connectivity and precise spatial geometry inherent in Boundary Representations (B-reps). Existing generative approaches largely formulate CAD construction as a sequence modeling problem, but consequently lack an explicit topological inductive bias. In this work, we present GraphCAD, a geometry-aware generative framework that centers modeling on the Edge-Adjacency Graph (EAG), a novel graph-based CAD representation tailored for generative tasks. GraphCAD adopts a hybrid graph diffusion formulation to jointly model the discrete and continuous components of EAG and introduces a geometry-aware denoising mechanism that renders node attributes into spatial point sets. These point sets are encoded by an SE(3)-equivariant backbone and fused into the node representations, thereby grounding the generative process in physical geometry. To support this study, we construct EAGBench, a large-scale dataset of engineering CAD models. Extensive experiments show that GraphCAD can generate complex CAD models with more than 50 faces and achieves state-of-the-art performance in geometric fidelity and distributional alignment compared with sequence-based baselines.
Figures & tables
Figure 1. Construction of the Edge-Adjacency Graph (EAG) representation. (a) A Boundary Representation (B-rep) of a CAD solid. (b) Construction of the Edge-Adjacency Graph, where each graph node corresponds to a topological edge in the B-rep and graph edges indicate shared B-rep vertices. (c) The EAG representation with node attributes, where each node encodes the curve type together with compact analytic geometric parameters.
Figure 2. Overview of the GraphCAD framework. Top: The generative pipeline. The model synthesizes CAD models by iteratively denoising the Edge-Adjacency Graph (EAG), our proposed topological representation of B-rep geometry. Bottom: The Geometry-Aware Denoising module. Unlike standard graph-based methods that treat node attributes as abstract vectors, our approach explicitly “ renders ” these attributes into spatial point sets through “ Curve Sampling ”. The sampled point sets are then processed by an SE(3)-equivariant Vector Neuron backbone to extract rotation-robust geometric embeddings.
Figure 3. Relationship between sketch-and-extrude CAD and the EAG representation. EAG encodes the resulting line, arc, and circle boundary primitives as typed nodes, while graph edges capture their topological adjacency.
Model
Deg ↓
Clus ↓
Orb ↓
Validity ↑
Uniqueness ↑
Novelty ↑
EAGBench-DC
GraphNVP
0.836
0.749
1.634
52.79
89.79
81.96
GDSS
0.659
0.721
1.310
80.08
85.21
79.50
ConGress
0.313
0.399
0.559
71.97
91.20
85.18
DiGress
0.242
0.383
0.183
95.31
93.42
89.34
GraphCAD (Ours)
0.176
0.248
0.120
98.05
95.76
92.21
Table 1. Quantitative evaluation of unconditional generation across two benchmarks. We measure distributional alignment between generated and reference Edge-Adjacency Graphs (EAGs) using maximum mean discrepancy over node degree ( Deg ), clustering coefficient ( Clus ), and orbit counts ( Orb ). Generative quality is further evaluated in terms of Validity , Uniqueness , and Novelty . Best results are shown in bold .
Figure 4. Visual categorization of the primary failure modes encountered during EAG-to-B-rep reconstruction.
Figure 5. Unconditional generation results on the EAGBench dataset. GraphCAD generates realistic CAD models with complex geometry and topology.
EAGBench-DC
EAGBench-FS
Model
IR ↓
CD ↓
HD ↓
JSD ↓
MMD ↓
IR ↓
CD ↓
HD ↓
JSD ↓
MMD ↓
Sequence-based
Transformer ( Vaswani et al., 2017 )
12.02
11.83
0.48
83.94
4.02
19.11
17.51
0.81
84.01
5.33
DeepCAD ( Wu et al., 2021 )
9.29
6.76
0.51
98.40
4.25
13.07
14.98
0.79
92.53
5.91
Text2CAD ( Khan et al., 2024b )
3.87
3.36
0.45
59.63
3.98
8.29
4.62
0.63
70.38
4.78
TGBFN ( Zheng et al., 2025 )
0.78
2.34
0.37
65.92
3.40
6.48
5.35
0.72
74.73
5.08
Table 2. Quantitative comparison of conditional generation. We evaluate reconstructed B-rep models in terms of topological validity (IR), geometric fidelity (CD, HD), and distributional quality (JSD, MMD). Note: For readability, CD, JSD, and MMD are each reported after scaling by 102 within the [−1,1]3 unit cube.
Graph Metrics
CAD Metrics
Model
Deg ↓
Clus ↓
Orb ↓
Validity ↑
Uniqueness ↑
Novelty ↑
CD ↓
HD ↓
JSD ↓
MMD ↓
Full Method (Ours)
0.164
0.236
0.119
98.33
96.71
90.38
1.73
0.28
33.22
1.97
w/o SE(3) Equivariance
0.234
0.377
0.166
97.27
93.98
89.16
2.17
0.44
42.37
2.67
w/o Loss Masking
0.316
0.395
0.582
73.17
92.11
88.46
2.48
0.51
47.56
2.93
w/o Geometric Rendering
0.352
0.411
0.612
66.45
87.93
80.38
5.70
0.66
67.23
3.62
Table 3. Ablation study on the EAGBench-DC dataset evaluating the contribution of individual components in GraphCAD. We compare the full model with variants that remove the SE(3)-equivariant backbone ( w/o SE(3) Equivariance ), semantic loss masking ( w/o Loss Masking ), and point-based geometric rendering ( w/o Geometric Rendering ).
Points ( K )
Computational Cost
Generation Quality
VRAM (GB)
Time (s)
IR ( ↓ )
CD ( ↓ )
0
4.40 ± 0.03
131 ± 1
27.59
5.47
10
4.60 ± 0.02
145 ± 2
2.52
2.81
20
4.64 ± 0.02
146 ± 1
2.67
2.23
50
4.93 ± 0.01
149 ± 2
1.84
2.29
100
5.06 ± 0.03
151 ± 2
1.78
1.94
Table 4. Impact of point sampling density ( K ). Trade-off between inference computational cost and generation quality on EAGBench-DC. Computational cost is measured on a single batch of 256 samples.
Figure 6. Qualitative comparison of CAD models from conditional generation. From top to bottom, we show results produced by DeepCAD, Text2CAD, TGBFN, BrepGiff, our proposed GraphCAD, and the Ground Truth (GT).
School of Computer Science and Engineering, Sun Yat-sen University, China · School of Intelligent Systems Engineering, Shenzhen Campus of Sun Yat-sen University, China
School of Artificial Intelligence and Robotics, Hunan University · College of Computer Science and Software Engineer, Shenzhen University · School of Computer Science and Technology, Harbin Institute of Technology, Shenzhen +1