UniBRep: Learning Unified Geometry and Topology for Image-conditioned B-Rep Generation

Haiyang Ying; Allen Tu; Jiaye Wu; Tom Goldstein; Matthias Zwicker

cs.CVOct 2, 2026

UniBRep: Learning Unified Geometry and Topology for Image-conditioned B-Rep Generation

Authors: Haiyang Ying, Allen Tu, Jiaye Wu, Tom Goldstein, Matthias Zwicker

Organizations: University of Maryland, College Park, Maryland, USA · Amazon, Bellevue, Washington, USA

Abstract

Generating a boundary representation (B-rep) conditioned on a single image requires faithful reconstruction of geometry, valid topology, and support for complex shapes. We present UniBRep, a geometry-first framework that adapts a pretrained image-to-3D model to generate a feature mesh as a unified intermediate representation. Its surface provides a geometric scaffold, while spatially aligned learned features encode face-separation cues for topology recovery. Dual decoder branches generate the geometry and face-separation features; a geometry- and feature-guided construction pipeline then fits parametric surfaces, recovers boundary curves and connectivity, and assembles an explicit B-rep using a CAD kernel. Recovering topology from mesh regions avoids predefined architectural face-count limits, allowing face count to scale with shape complexity. On the standard DeepCAD benchmark, UniBRep produces valid B-reps for 80.49% of inputs and reduces face Chamfer distance from 0.1096 to 0.0345 relative to CADDreamer. In a matched comparison, UniBRep also outperforms the HoLa public demo across all reported metrics. Further evaluations demonstrate scalability to high-complexity shapes beyond the standard 30-face range, generalization to objects outside the CAD training distribution, and qualitative transfer to real photographs.

Figures & tables

Figure 1: Overview of UniBRep. Top: UniBRep adapts TRELLIS’s sparse-structure (SS) and structured latent (SLat) generators to produce a CAD-adapted SLat from a single input image. Dual geometry and feature SLat decoder branches map the shared latent to surface geometry and learned face-separation features, forming a feature mesh. Bottom: The construction pipeline partitions the feature mesh using geometry- and feature-guided segmentation, fits and parameterizes surfaces, extracts topology, and assembles an explicit B-rep using a CAD kernel.

Figure 2: Geometry- and feature-guided segmentation. We first partition the generated mesh using local geometric cues and fit analytic primitives to the resulting regions. Regions that fail primitive fitting are further segmented using learned face-separation features and then refitted.

$N$	Method	Chamfer Distance $\downarrow$			Geometry F-score $\uparrow$			Topology $\uparrow$		Valid Ratio $\uparrow$
$N$	Method	Vertex	Edge	Face	Vertex	Edge	Face	FE	EV	Valid Ratio $\uparrow$
$2{,}424$	CADDreamer [ LLL∗25 ]	0.3135	0.1824	0.1096	0.425	0.359	0.607	0.384	0.483	18.28% ∗
	NVDNet † [ LCP∗24 ]	0.2100	0.0895	0.0458	0.572	0.662	0.747	0.694	0.620	–
	Point2CAD † [ LOWS24 ]	0.2382	0.1127	0.0615	0.517	0.621	0.708	0.725	0.633	–
	UniBRep (Ours)	0.1328	0.0627	0.0345	0.742	0.790	0.844	0.812	0.783	80.49%
$818$	HoLa (demo) [ LXY∗25 ]	0.3227	0.1737	0.1028	0.213	0.420	0.560	0.651	0.340	70.66%

Table 1: Quantitative comparison of single-image-conditioned B-rep generation on DeepCAD [ WXZ21 ] . The

N{=}2{,}424

rows report results on the standard 7–30-face benchmark, while the

N{=}818

rows report the matched comparison on a randomly selected subset of test objects. The best results within each set are shown in bold . ∗ indicates that CADDreamer’s accepted outputs are OpenCascade compounds rather than necessarily watertight single solids. † denotes NVDNet evaluated with our generated geometry and Point2CAD with our generated geometry and segmentation; both output parametric surfaces rather than B-reps. See Sec. 4 for the full evaluation protocol.

Figure 3: Qualitative comparison on DeepCAD [ WXZ21 ] . Outputs and ground truth are rendered from a common evaluation view that differs from the input view. Method-specific inputs are described in Sec. 4.1 . UniBRep more faithfully recovers fine geometry and CAD structure, while the baselines [ LLL∗25 , LXY∗25 , LOWS24 , LCP∗24 ] often omit parts or produce disconnected and non-watertight surfaces.

Geo. FT	Feat. Ref.	Supervision	Chamfer Distance $\downarrow$			Geometry F-score $\uparrow$			Topology $\uparrow$		Valid Ratio $\uparrow$
Geo. FT	Feat. Ref.	Supervision	Vertex	Edge	Face	Vertex	Edge	Face	FE	EV	Valid Ratio $\uparrow$
		–	0.1827	0.0895	0.0506	0.627	0.682	0.757	0.731	0.701	47.65%
✓		–	0.1564	0.0695	0.0382	0.695	0.739	0.799	0.775	0.741	62.50%
✓	✓	RGB	0.1381	0.0640	0.0355	0.716	0.766	0.824	0.791	0.757	74.75%
✓	✓	Contrastive	0.1328	0.0627	0.0345	0.742	0.790	0.844	0.812	0.783	80.49%

Table 2: Component ablation on DeepCAD [ WXZ21 ] . Geo. FT denotes CAD geometry finetuning, Feat. Ref. denotes feature-guided segmentation refinement, and Supervision denotes RGB or contrastive feature supervision. Our full configuration achieves the best results.

Method	Chamfer Distance $\downarrow$		Primitive Accuracy $\uparrow$
Method	Mesh	B-rep	Overall	Mean
CADDreamer	0.0581	0.1096	0.8690	0.8288
UniBRep (Ours)	0.0319	0.0345	0.8820	0.8515

Table 3: B-rep reconstruction quality on DeepCAD [ WXZ21 ] . We compare the construction pipelines of CADDreamer [ LLL∗25 ] and UniBRep on cases whose final outputs satisfy each method’s acceptance criteria. For matched face pairs, Overall is the fraction with matching predicted and ground-truth surface types, while Mean is the unweighted mean of per-type accuracies.

Dataset	Method	Chamfer Distance $\downarrow$			Geometry F-score $\uparrow$			Topology $\uparrow$		Valid Ratio $\uparrow$
Dataset	Method	Vertex	Edge	Face	Vertex	Edge	Face	FE	EV	Valid Ratio $\uparrow$
DeepCAD [ WXZ21 ] 31+ Faces	CADDreamer [ LLL∗25 ]	0.3794	0.2605	0.1950	0.145	0.208	0.216	0.107	0.186	9.94%
	Point2CAD [ LOWS24 ]	0.2384	0.1271	0.0751	0.371	0.475	0.533	0.362	0.301	–
	NVDNet [ LCP∗24 ]	0.2073	0.1123	0.0603	0.397	0.466	0.501	0.314	0.293	–
	UniBRep (Ours)	0.1408	0.0887	0.0581	0.584	0.619	0.645	0.534	0.533	37.57%
BrepGen [ XLJ∗24 ] OOD Furniture	HoLa (demo) [ LXY∗25 ]	0.2365	0.1723	0.1018	0.354	0.518	0.544	0.602	0.350	22.5%

Table 4: Quantitative evaluation on high-complexity and out-of-distribution objects. DeepCAD 31+ Faces contains 181 high-complexity models with 31 or more faces, while BrepGen OOD Furniture contains 40 furniture objects outside the DeepCAD [ WXZ21 ] training distribution. HoLa is omitted from DeepCAD 31+ Faces because its released configuration supports at most 30 faces; CADDreamer is omitted from BrepGen OOD Furniture because it produced only one valid output (2.5%). The best results within each dataset are shown in bold .

Figure 4: Real-world image-conditioned B-rep generation. Given foreground objects segmented from real photographs, UniBRep generates plausible geometry and constructs structured B-reps.

Figure 5: Performance across B-rep face counts on DeepCAD and DeepCAD 31+ Faces [ WXZ21 ] . We compare UniBRep with CADDreamer, Point2CAD, NVDNet, and the HoLa demo [ LLL∗25 , LOWS24 , LCP∗24 , LXY∗25 ] . Metrics are plotted against ground-truth face count. Face counts are grouped into bins of five. For methods that generate B-reps, F-scores are averaged over valid outputs, and validity denotes the fraction of inputs yielding valid B-reps in each bin. Validity is omitted for Point2CAD and NVDNet because they output parametric surfaces rather than B-reps; the HoLa demo is evaluated on the matched subset of 818 test objects and supports at most 30 faces.

Figure 6: Qualitative results on DeepCAD 31+ Faces [ WXZ21 ] . Each example shows the input image, UniBRep’s predicted faces and edges, and the corresponding ground-truth faces and edges. The baselines [ LLL∗25 , LXY∗25 , LOWS24 , LCP∗24 ] are omitted because none produces a comparable B-rep for these examples: CADDreamer fails valid B-rep construction, the HoLa demo supports at most 30 faces, and Point2CAD and NVDNet output parametric surfaces rather than B-reps.

Figure 7: Scalability across stair counts. We vary the number of steps from two to nine. For the HoLa demo [ LXY∗25 ] , the invalid two-step output is shown as a mesh, while red Xs indicate that no B-rep is produced for the eight- and nine-step inputs. UniBRep preserves the stepped structure and produces valid B-reps for all stair counts.

Figure 8: Qualitative comparison on out-of-distribution furniture objects from BrepGen [ XLJ∗24 ] . The HoLa demo [ LXY∗25 ] often oversimplifies geometry or omits object parts. Red # symbols mark CADDreamer [ LLL∗25 ] outputs that fail valid B-rep construction. UniBRep better preserves overall shape, part layout, and CAD-like boundaries.

Figure 9: Representative failure cases on DeepCAD 31+ Faces [ WXZ21 ] . From top to bottom, the examples show inaccurate geometry generation, incorrect segmentation of thin structures, and invalid B-rep assembly. CADDreamer fails to produce valid B-Reps for these cases.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Component	Status
SS generator input layer	Frozen
SS generator transformer blocks	LoRA-adapted
SS generator output layer	Trainable
SS decoder	Frozen
SLat generator transformer blocks	LoRA-adapted
SLat generator output layer	Trainable

Appendix

Table 5: Module adaptation strategy. Frozen modules retain pretrained TRELLIS [ XLX∗25 ] weights, LoRA-adapted modules receive low-rank updates [ HSW∗22 ] , and Trainable modules are optimized directly.

Method	Protocol	$N$	Chamfer Distance $\downarrow$			Geometry F-score $\uparrow$			Topology $\uparrow$		Valid Ratio $\uparrow$
Method	Protocol	$N$	Vertex	Edge	Face	Vertex	Edge	Face	FE	EV	Valid Ratio $\uparrow$
HoLa [ LXY∗25 ]	Published	2,424	0.0684	0.0182	0.0101	0.790	0.820	0.860	0.810	0.800	86.37%
HoLa (demo) [ LXY∗25 ]	Matched	818	0.3227	0.1737	0.1028	0.213	0.420	0.560	0.651	0.340	70.66%
UniBRep (Ours)	Matched	818	0.1277	0.0593	0.0327	0.754	0.802	0.853	0.816	0.782	81.30%

Appendix

Table 6: Comparison with HoLa’s published and public-demo results on DeepCAD [ WXZ21 ] . The Published row lists HoLa’s reported values under the authors’ original protocol. The Matched rows evaluate the HoLa demo and UniBRep on 818 randomly selected test objects, with a shared camera viewpoint per object and the common protocol described in Sec. 4.3 . UniBRep receives an RGB rendering, while HoLa receives the same-view rendering with black edge overlays matching its paper. Because the protocols are not directly comparable, the published results are included only for context.

Setting	CD(V) $\downarrow$	CD(E) $\downarrow$	CD(F) $\downarrow$	F-score(V) $\uparrow$	F-score(E) $\uparrow$	F-score(F) $\uparrow$	FE $\uparrow$	EV $\uparrow$	Valid Ratio $\uparrow$
Default	0.1231	0.0531	0.0291	0.765	0.807	0.867	0.826	0.795	78.5%
Geometry seg. $\times 0.75$	0.1243	0.0543	0.0293	0.760	0.798	0.853	0.807	0.777	74.0%
Geometry seg. $\times 1.25$	0.1232	0.0553	0.0300	0.758	0.803	0.863	0.831	0.802	76.5%
Feature seg. $\times 0.75$	0.1234	0.0536	0.0299	0.759	0.802	0.860	0.828	0.792	79.0%
Feature seg. $\times 1.25$	0.1295	0.0592	0.0328	0.736	0.793	0.853	0.814	0.776	81.0%
Primitive support $+0.02$	0.1244	0.0557	0.0311	0.761	0.803	0.856	0.824	0.795	78.5%

Appendix

Table 7: Sensitivity to B-rep construction parameters on DeepCAD [ WXZ21 ] . Each parameter row varies only the indicated reconstruction setting from its default on 200 randomly selected test objects. Geometry seg. jointly scales the boundary-strength and normal-consistency thresholds (

g_{\mathrm{thresh}}{=}15.0

n_{\mathrm{thresh}}{=}0.25

). Feature seg. scales the mean-shift bandwidth used for feature-guided refinement (0.85). Primitive support changes the minimum RANSAC support-ratio threshold (0.9), while RANSAC

\epsilon

scales the maximum mean point-to-surface residual (0.002). Multi-seed reports the best of five post-processing runs.

Figure 10: Additional qualitative comparison on DeepCAD [ WXZ21 ] . We compare UniBRep with the baselines [ LLL∗25 , LXY∗25 , LOWS24 , LCP∗24 ] . Outputs and ground truth are rendered from a common evaluation view that differs from the input view.

Parametric CAD Modeling

Parametric CAD modeling represents 3D designs as editable, constraint-driven construction histories and develops methods to generate, reconstruct, and revise them.

See all 80 papers

HiDiGen: Hierarchical Diffusion for B-Rep Generation with Explicit Topological Constraints
Apr 3, 2026Shurui Liu, Weide Chen, Ancong Wu3D Shape Generation Constrained Generative Modeling
School of Computer Science and Engineering, Sun Yat-sen University, China · School of Intelligent Systems Engineering, Shenzhen Campus of Sun Yat-sen University, China
HiFi-BRep: High-Fidelity Latent Representation for Robust B-Rep Generation
Aug 17, 2026Junhao Hou, Chenqi Luo, Pufan Wang +53D Shape Generation Constrained Generative Modeling
State Key Lab of CAD&CG, Zhejiang University · Hangzhou Dianzi University · Guangzhou University
Towards Valid B-Rep Generation: Training-Free Wireframe Anomaly Detection and Repair
Aug 5, 2026Jingyu Wu, Youcheng Cai, Tengyu Luo +1Anomaly Detection
University of Science and Technology of China, Hefei, China

Apr 3, 2026J/K move · Enter open · S save

HiDiGen: Hierarchical Diffusion for B-Rep Generation with Explicit Topological Constraints

Shurui Liu, Weide Chen, Ancong Wu

School of Computer Science and Engineering, Sun Yat-sen University, China · School of Intelligent Systems Engineering, Shenzhen Campus of Sun Yat-sen University, China

3D Shape Generation Constrained Generative Modeling 3D Diffusion Models

Open paper AskPDF arXiv

Boundary representation (B-rep) is the standard 3D modeling format in CAD systems, encoding both geometric primitives and topological connectivity. Despite its prevalence, deep generative modeling of valid B-rep structures remains challenging due to the intricate interplay between discrete topology and continuous geometry. In this paper, we propose HiDiGen, a hierarchical generation framework that decouples geometry modeling into two stages, each guided by explicitly modeled topological constraints. Specifically, our approach first establishes face-edge incidence relations to define a coherent topological scaffold, upon which face proxies and initial edge curves are generated. Subsequently, multiple Transformer-based diffusion modules are employed to refine the geometry by generating precise face surfaces and vertex positions, with edge-vertex adjacencies dynamically established and enforced to preserve structural consistency. This progressive geometry hierarchy enables the generation of more novel and diverse shapes, while two-stage topological modeling ensures high validity. Experimental results show that HiDiGen achieves strong performance, generating novel, diverse, and topologically sound CAD models.

Abstract

Parametric CAD Modeling

Explore similar work

HiDiGen: Hierarchical Diffusion for B-Rep Generation with Explicit Topological Constraints

HiFi-BRep: High-Fidelity Latent Representation for Robust B-Rep Generation

Towards Valid B-Rep Generation: Training-Free Wireframe Anomaly Detection and Repair