acceptodds
Under review as a conference paper at ICLR 2027

BRepLift: B-Rep Reconstruction with Shared 3D Foundation Priors

12 views

Abstract

B-Rep reconstruction requires recovering precise parametric geometry and valid topology from unstructured 3D observations. Existing surface-based methods typically pass task-specific predictions between reconstruction stages, compressing rich learned representations into intermediate outputs and creating a representational bottleneck. Greater geometric variation and topological complexity further amplify ambiguity and error propagation. We propose BRepLift, a framework that adapts pretrained 3D foundation features into B-Rep-aware representations shared across primitive identification, geometry recovery, and topology reasoning. Stage-specific designs exploit these features through clustering-guided face identification and context-conditioned geometry and topology prediction, capturing both intra- and inter-primitive information. Experiments on three test sets with complementary geometric and topological complexity demonstrate consistent improvements over DualBrep in face, edge, and vertex F1 scores and B-Rep validity. In particular, reconstruction validity increases from 10.60% to 22.14% on ABC-ParseNet and from 17.34% to 29.46% on our curated test set.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Boudary representation, CAD reverse engineering, 3d foundational model

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_FbHeOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes BRepLift, a framework for reconstructing boundary representations from 3D meshes using shared 3D foundation features. The method employs a pretrained Utonia backbone with LoRA-based adaptation and face-level contrastive learning to obtain primitive-aware representations. These features are then used for clustering-guided face segmentation and subsequently shared with context-conditioned decoders for geometry and topology reconstruction.

Experiments on three test sets demonstrate improvements over existing methods, particularly DualBrep. The improvements are especially noticeable on geometrically complex shapes.

Critical strengths
  • The paper addresses the challenging problem of recovering structured CAD representations from unstructured 3D meshes. The proposed pipeline integrates primitive segmentation, geometry reconstruction, and topology prediction through a shared feature representation. The overall framework is reasonably motivated, and reusing geometric features across different reconstruction stages is an interesting design choice.
  • The method demonstrates improvements over strong existing baselines across multiple datasets with different geometric and topological complexities. The paper also provides component-level ablations and qualitative comparisons, which help demonstrate the effectiveness of the proposed pipeline.
Critical weaknesses

A central claim of the paper is that pretrained 3D foundation features provide useful geometric priors for B-Rep reconstruction. However, the current experiments do not sufficiently isolate the contribution of foundation pretraining itself. The ablation studies mainly compare pretrained Utonia features w/and w/o task-specific adaptation. While these results demonstrate the effectiveness of feature adaptation, they do not establish whether large-scale pretraining is necessary or beneficial compared with a backbone trained from scratch on the same B-Rep data. Personally, a more informative comparison would include:

  • a randomly initialized backbone trained from scratch
  • a frozen pretrained backbone
  • a pretrained backbone with LoRA adaptation
  • a pretrained backbone with full-parameter fine-tuning.

Such comparisons would help distinguish the contributions of pretraining and task-specific adaptation. Also, the benefit of sharing foundation features across reconstruction stages is not sufficiently isolated.

Questions for authors
  • Could the authors compare the current pretrained Utonia backbone against the same backbone trained from scratch on the B-Rep training dataset? This would help establish whether the improvements arise from large-scale pretraining or primarily from the task-specific architecture and supervision.
  • What motivates using LoRA adaptation only in the final backbone block rather than full-parameter fine-tuning? Can the authors provide a comparison between frozen pretrained features, LoRA adaptation, and full fine-tuning under the same supervision?
  • For personal curiosity: how much of the remaining B-Rep invalidity is caused by primitive segmentation errors, geometry prediction errors, or topology assembly and post-processing? A quantitative breakdown would help identify the main bottlenecks of the complete pipeline.
3 comments

W1 (Q1): Comparison of Pre-trained backbone

The comparison with training-from-scratch indicates that pretrained geometric priors substantially benefit both segmentation and reconstruction.

The from-scratch encoder uses the same backbone architecture and is trained from random initialization with the contrastive objective (Eq. 2) on the same B-Rep training data. The resulting encoder is then frozen, and downstream segmentation and reconstruction modules with the same architectures as ours are trained using the frozen features. We compare the from-scratch variant and ours on the ABC-ParseNet testset in three aspects: local cluster consistency, B-Rep face segmentation, and reconstruction.

Clustering. We measure local consistency using mean cluster purity, i.e., for each cluster, we compute the fraction of mesh vertices belonging to its dominant ground-truth B-Rep face mask, then average equally across clusters. Under the same clustering setting, the from-scratch encoder and ours achieve similar results (from-scratch vs. ours, 0.970 vs. 0.967), both exceeding unadapted Utonia (0.871). This suggests that task-specific adaptation is important for aligning local clusters with B-Rep faces.

Segmentation. Despite similar cluster purity, ours yields substantially better B-Rep face segmentation.

MethodMatched mIoU (%)P@ IoU 0.50R@ IoU 0.50F1@ IoU 0.50P@ IoU 0.75R@ IoU 0.75F1@ IoU 0.75
From-scratch65.390.7620.6970.7280.5490.4950.521
Ours81.880.9530.8490.8980.8480.7550.799

Reconstruction. With identical ground-truth contexts, ours outperforms the from-scratch across all reported geometry and topology metrics.

MethodFace CD ↓\downarrowEdge CD ↓\downarrowVertex CD ↓\downarrowFace F1 ↑\uparrowEdge F1 ↑\uparrowVertex F1 ↑\uparrowFace–Edge F1 ↑\uparrowEdge–Vertex F1 ↑\uparrow
From-scratch0.05400.07500.11580.7220.5910.4090.5630.405
Ours0.01820.02380.02990.9380.7330.7770.8040.803

Q2: Comparison of Adaptation Strategy

LoRA adaptation achieves performance comparable to full-parameter fine-tuning while reducing trainable parameters across the overall pipeline from 155.92M to 19.33M (an 87.6% reduction). We restrict adaptation to the final backbone block to preserve the pretrained representations in earlier blocks. Under the same primitive-aware contrastive supervision and clustering configuration, LoRA and full fine-tuning achieve similar cluster purity (0.9669 vs. 0.9686), supporting LoRA as a parameter-efficient choice for promoting local feature consistency.

We compare frozen pretrained features, LoRA-adapted features, and fully fine-tuned features on the ABC-ParSeNet test set. LoRA adaptation and full fine-tuning use the same primitive-aware contrastive supervision. After adaptation, each encoder is frozen, and the downstream segmentation and reconstruction modules are trained using the same architectures and training settings.

Segmentation. Both adaptation strategies substantially outperform frozen pretrained features. LoRA achieves nearly identical matched mIoU and F1 scores to full fine-tuning, with higher precision but lower recall.

MethodMatched mIoU (%)P@ IoU 0.50R@ IoU 0.50F1@ IoU 0.50P@ IoU 0.75R@ IoU 0.75F1@ IoU 0.75
Frozen pretrained71.450.8450.7520.7960.5730.5110.540
Ours81.880.9530.8490.8980.8480.7550.799
Fully finetuned82.020.9420.8610.8990.8340.7700.801

Reconstruction. With identical ground-truth contexts, LoRA and full fine-tuning achieve comparable reconstruction performance, with small differences across geometry and topology metrics.

MethodFace CD ↓\downarrowEdge CD ↓\downarrowVertex CD ↓\downarrowFace F1 ↑\uparrowEdge F1 ↑\uparrowVertex F1 ↑\uparrowFace–Edge F1 ↑\uparrowEdge–Vertex F1 ↑\uparrow
Frozen pretrained0.01920.02420.03160.9270.7470.7790.7990.812
Ours0.01820.02380.02990.9380.7330.7770.8040.803
Fully finetuned0.01860.02310.03000.9350.7540.7770.8070.805

The best results are shown in bold, and the second-best results in italics.

Q3: The benefit of sharing foundation features across reconstruction stages

Response: Please refer to the response of XqXd-W1.

Q4: Sources of Remaining B-Rep Invalidity

Response: We conduct a progressive oracle ablation on the ABC-ParSeNet test set, cumulatively replacing intermediate predictions with ground truth while keeping the remaining pipeline unchanged

  • + GT segmentation: Use ground-truth face segmentation, retaining predicted geometry and edge existence.
  • ++ GT topology: Replace edge-existence predictions with ground truth.
  • +++ GT geometry: Replace predicted face and edge grids with their ground-truth counterparts.
SettingB-Rep validity (%)
Full predicted pipeline22.14
+ GT segmentation28.81 (+6.67)
++ GT topology31.09 (+2.28)
+++ GT geometry53.54 (+22.45)

GT segmentation improves validity by 6.67 points, primarily by restoring missed small and detailed faces and providing more accurate contexts for geometry reconstruction. Adding GT edge existence yields a smaller gain of 2.28 points. Correct connectivity alone cannot resolve geometric inconsistencies that prevent successful trimming and sewing during post-processing. Adding the GT geometry produces the largest gain, indicating that geometric accuracy is a major bottleneck.

Even with the oracle setting, validity reaches only 53.54%, indicating limitations of the discretized point grid representation and post-processing. GT grids require B-spline fitting, which introduces approximation errors. The optimization may perturb otherwise accurate geometry. Moreover, shared-vertex identification and boundary-wire construction remain heuristic.

After reviewing the rebuttal, I find that the authors have substantially addressed my main concerns. In particular, the additional comparisons with training from scratch and full fine-tuning provide useful evidence for the benefits of foundation pretraining and the parameter efficiency of LoRA. The oracle ablation also helps clarify the main bottlenecks affecting B-Rep validity.

There are still few ablations remaining, but I appreciate the additional experiments and consider the response sufficiently convincing overall. I will maintain my rating of 3.

Reviewer_XqXdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

BRepLift reconstructs parametric B-Reps from input meshes using features adapted from Utonia. It applies a (B-Rep)-face-level contrastive objective, clusters (mesh) vertices, aggregates clusters into (B-Rep) face masks Ωnf\Omega_{n_f} through a query decoder, and reuses the same features inside (B-Rep) face and (B-Rep) edge-candidate contexts to predict structured surface and curve grids plus (B-Rep) edge existence; fitting and assembly follow DualBrep, and the three stages are trained separately.

The pipeline reports higher primitive F1 and validity than DualBrep on three test sets, with the largest absolute validity gains on ABC-ParseNet (10.60→22.14) and the curated Fusion360/CC3D-Ops set (17.34→29.46). This review concentrates on the two middle stages, where the submission does not yet make clear which parts are learned, how they are trained, what the clustering units represent, and how the reconstruction tokens are defined.

Critical strengths
  • S1 — A coherent interface for structured reconstruction. Using the predicted (B-Rep) face masks Ωnf\Omega_{n_f} only to select which shared (mesh) vertex features F′F' are gathered, rather than as a network input, keeps geometric context available for reconstruction (§3.2-3.3; Figures 1-2).

  • S2 — Useful system-level numbers, with disclosed limits. Validity improves on all three test sets, and the paper states the single-orientation evaluation (p6), the separately trained stages (p9), and shows trimming and smoothness failures (Figure 10).

Critical weaknesses
  • W1 — The central sharing claim is not yet separated from the adapted features themselves (p2, §1; §3.1-3.3, Tables 3 and 4). As far as I can tell, "sharing" means the reconstruction stage reuses the same adapted features F′F' that produced the (B-Rep) face masks Ωnf\Omega_{n_f}; I could not find a comparison in which the reconstruction stage encodes its own context under the same masks, capacity and budget, so part of the gain may come from the pretrained prior or from the adaptation alone.

    • Tables 3-4: Table 3 covers adaptation and clustering inside §3.2; Table 4 compares adapted with unadapted features and repeated context conditioning, all with ground-truth masks, and the "without Context-CA" variant keeps the first context layer (p9).
    • Needed: keep the masks, geometry inputs, capacity, budget and post-processing identical, let the reconstruction stage encode its own context, and report final validity for the predicted-mask pipeline.
  • W2 — The benefit of the cluster-to-(B-Rep)-face aggregation over the adapted features alone does not yet come through (§3.2, Eq.(2), Table 3). As far as I can tell, Eq.(2) already encourages vertices of the same (B-Rep) face to share features, so a good clustering of those features should already approximate faces; it would help to see what the query decoder adds beyond that.

    • Table 3: disabling clustering costs 0.25 mIoU and moves [email protected] from 0.898 to 0.860; the ablation removes clustering, not the learned aggregation.
    • Needed: compare against a threshold- or (mesh)-connectivity-based merge of the same clusters, and report the mask IoU.
  • W3 — How far the stage depends on over-segmentation, and the cluster count itself, are not yet made clear (§3.2, Eq.(3), Table 3). As far as I can tell, the decoder can only merge clusters, so the design relies on clusters being finer than (B-Rep) faces and on each cluster belonging to a single face; I could not find a sentence stating this premise, nor a reported cluster count.

    • Eq.(3): assignment is at cluster granularity (A~∈RK×Nmax⁡\tilde A\in\mathbb R^{K\times N_{\max}}, ℓk=arg⁡max⁡n\ell_k=\arg\max_n, Ωn=⋃k:ℓk=nSk\Omega_n=\bigcup_{k:\ell_k=n}S_k), so a cluster cannot be split across (B-Rep) faces.
    • Clustering settings: I could not find KK, the ratio K/NfK/N_f, or the linkage, distance and stopping rule.
    • Needed: state KK, K/NfK/N_f and the clustering settings, add one sentence that assignment is at cluster granularity, and report cluster purity and small-(B-Rep)-face recall.
  • W4 — What a cluster represents, and which features enter the clustering, are not yet clear (§3.2, p4, Eq.(1)). The text calls clusters "compact local regions", but the clustering input is described only as "adapted vertex features", so a reader is left to infer several basic properties of the units.

    • PE (p4): built from the vertex coordinates viv_i and pooled into pkp_k, so it locates a cluster in space rather than identifying it.
    • Clustering input (Eq.(1)): four concatenated backbone levels with only the last LoRA-adapted; I could not find whether clustering uses the concatenation or only the LoRA level, nor any (mesh) connectivity constraint, and coordinates and normals enter only afterwards, through the PE.
    • Query initialization (p5): half learnable, half sampled cluster tokens; I could not find a rationale or an ablation separating the two halves.
    • Needed: state the clustered tensor, any connectivity constraint, how zkz_k and pkp_k combine, and why half the queries are sampled from clusters.
  • W5 — The learned parts of §3.2 and §3.3 are not yet tied to a stated training objective (§3.2-3.3; Appendix B.2-B.3). Both stages contain clearly learnable components, and I could not find a statement of what trains them.

    • Learned parts (§3.2, §3.3): the cluster-token construction (Poolattn\mathrm{Pool_{attn}}, Transformer, ϕq\phi_q/ϕc\phi_c, half the queries) and decoder, including an undescribed (B-Rep) face-type head; the primitive queries, both attention branches, and the geometry and existence heads. Clustering, PE and arg⁡max⁡\arg\max appear fixed.
    • Losses: Eq.(2) in §3.1 is the only loss in the main text; I could not find a reference to Lprim\mathcal L_{prim} (Eq. 8, Appendix B.2) or Lrecon\mathcal L_{recon} (Appendix B.3) from §3.2 or §3.3, and λmask\lambda_{mask} is undefined there.
    • Frozen features: I could not find whether F′F' stays frozen during segmentation, or what accuracy the (B-Rep) face-type head reaches.
    • Needed: add one training-objectives paragraph per stage citing Eq.(8)/B.2 and B.3, with the frozen components, matching and no-object handling, and the λ\lambda values.
  • W6 — Several objects and terms in §3.2-3.3 are not yet defined, and the token layout of §3.3 is given without its rationale (§3.2-3.3, §4.1). I could not always tell which object a sentence refers to, and the representation of a (B-Rep) face is introduced without a reason for its shape.

    • face, edge (p5): apparently the (B-Rep) face mask Ωnf\Omega_{n_f} and edge candidate Ωne\Omega_{n_e}, not (mesh) faces; F/E/VF/E/V are reused for M=(V,F)M=(V,F) and B=(F,… )B=(\mathcal F,\dots), and I could not find whether the input mesh is triangulated.
    • mask: covers Ωnf\Omega_{n_f} (a lifted vertex set used only to gather features), the predicted A~\tilde A, and Ωne\Omega_{n_e}; the name suggests per-vertex prediction, while §3.2 says the module "predicts B-Rep face masks over these vertices".
    • cluster, grid, context, shared: each carries two or more meanings (§3.2-3.3, §4.1).
    • Token layout (p5): 2×22\times2 local tokens plus one global token, upsampled to 4×44\times4, 8×88\times8 and 16×1616\times16; edges use two local tokens plus one topology token. I could not find the rationale, the choice of starting resolution, or an ablation of either.
    • Needed: add a notation table and a conventions paragraph, rename the (B-Rep) face mask after what is computed, and explain the token layout, the global and topology tokens, and the benefit of progressive upsampling. This is a writing request, not a request for new experiments.

Minor Issues

  • Metric definitions: it would help to state the distance scale, the matched subset and the aggregation rule, since §4.1 requires CD below 0.01 for a successful match and averages CD over successful matches, while Table 4 reports 0.0182–0.1370, and the Ours row on the DualBrep test set at IoU 0.50 in Table 2 lists P=0.973, R=0.980 and F1=0.980, whose displayed values give a harmonic mean near 0.9765. The relative improvements quoted on p2 would then also be easier to check against Table 1.
  • Validity protocol and comparison coverage: as far as I can tell, this would also benefit from a short clarification. I could not find the validity checker, the CAD-kernel version and tolerances, the shell-versus-solid criterion, or the treatment of failed outputs; the exact DualBrep inference and assembly settings are unclear to me; and segmentation is compared with SED-Net only, while BGPSeg is cited but not compared.
Questions for authors

Please refer to previous sections.

4 comments
AuthorsOpenReview ↗

Response to Weakness 1-3

W1: Isolating the Benefit of Feature Sharing

Shared adapted features achieve performance comparable to reconstruction-specific adaptation, with slight advantages in geometric accuracy and topology metrics. To isolate the effect of reconstruction-specific adaptation, we train all reconstruction variants on the same fixed subset of 7,680 shapes, keeping the segmentation branch and its predicted masks fixed. We consider two alternatives: (a) a separate copy of our adapted encoder, with its LoRA parameters further finetuned during reconstruction, and (b) the pretrained Utonia backbone with newly initialized LoRA parameters trained for reconstruction. In both alternatives, the backbone weights remain frozen, while the LoRA parameters and reconstruction module are jointly optimized.

The results on the ABC-ParseNet test set show that shared adapted features achieve performance comparable to both alternatives. Neither reconstruction-specific adaptation yields consistent improvements across geometry, topology, and validity metrics. These support the sufficiency of the shared adapted representation for reconstruction under the evaluated setting.

MethodShape CD ↓\downarrowNC ↑\uparrowFace F1 ↑\uparrowEdge F1 ↑\uparrowVertex F1 ↑\uparrowFace-Edge F1 ↑\uparrowEdge-Vertex F1 ↑\uparrowValidity ratio ↑\uparrow
Ours0.03460.8530.6790.5470.5640.5860.58515.40
(a)0.03580.8500.6780.5440.5600.5860.58415.91
(b)0.03670.8510.6710.5410.5580.5690.57414.86

W2: Benefit of Learned Cluster-to-Face Aggregation

Our learned aggregation consistently outperforms heuristic aggregation across all reported segmentation metrics. We evaluated mesh-connectivity-constrained agglomerative clustering as a threshold-based merge method, starting from the same clusters and cluster tokens used by our learned aggregation. Two clusters are adjacent if a mesh edge connects their vertices. We apply Ward linkage [1] with the Euclidean metric in cluster-token feature space, allowing only adjacent groups to merge and stopping when no eligible pair has a linkage distance below the threshold. After sweeping multiple thresholds, we report the best-performing setting (threshold = 0.3) below. Results are evaluated on the ABC-ParSeNet test set.

SettingMatched mIoU (%)P @ IoU 0.50R @ IoU 0.50F1 @ IoU 0.50P @ IoU 0.75R @ IoU 0.75F1 @ IoU 0.75
Heuristic aggregation79.010.8020.4900.6080.6970.4260.529
Learned aggregation (Ours)81.880.9530.8490.8980.8480.7550.799

[1] Joe H. Ward Jr. Hierarchical Grouping to Optimize an Objective Function. Journal of the American Statistical Association, 58(301):236–244, 1963.


W3: Over-Segmentation Assumptions and Clustering Sensitivity

During clustering, we initialize each mesh vertex as a cluster seed and perform mesh-connectivity-constrained agglomerative clustering on adapted mesh vertex features f′\mathbf{f'}, using Ward linkage and Euclidean distance until K=512K=512 clusters remain. During the inference, the B-Rep face count is determined by queries with non-negative face-type predictions. Across the three test sets, the mean K/NfGTK/N_f^{\mathrm{GT}} is 33.98/56.11/39.65, and cluster purity is 0.984/0.967/0.965. We define small B-Rep faces when their face mask contains fewer than 1% of the overall mesh vertices. Our method achieves small-face recall of 0.802/0.601/0.609 at IoU 0.50 and 0.723/0.513/0.527 at IoU 0.75. We have clarified assignment operates at cluster granularity in the revised submission.

AuthorsOpenReview ↗

Response to Weakness 4-6


W4: Cluster Definition, Input Features, and Query Initialization

Each cluster represents a local group of mesh vertices. It is an intermediate segmentation unit, rather than a predicted B-Rep face. Specifically, clustering uses the adapted features F′\mathbf F', which concatenate features from all Utonia backbone blocks, including both frozen and LoRA-adapted blocks, as defined in Eq. 1. Mesh connectivity constrains the clustering process, as clarified in our response to W3.

The cluster token zkz_k is obtained by pooling the adapted features of mesh vertices within cluster kk. The segmentation decoder takes zk+pkz_k+p_k as input, where pkp_k is the pooled positional encoding of the vertices’ XYZ coordinates and does not include normal encoding.

Mixed query initialization better suits our reconstruction pipeline by reducing missed faces, which can compromise downstream B-Rep completeness. Compared with fully learnable queries on the ABC-ParSeNet test set, it achieves higher recall and F1 scores, at the cost of slightly lower matched mIoU and precision. These results support our choice to prioritize face coverage while maintaining a better precision–recall balance.

Query initializationMatched mIoU (%)P @ IoU 0.50R @ IoU 0.50F1 @ IoU 0.50P @ IoU 0.75R @ IoU 0.75F1 @ IoU 0.75
Half learnable + half sampled81.880.9530.8490.8980.8480.7550.799
All learnable81.910.9600.8280.8890.8570.7260.787

W5: Training Objectives and Frozen Components

We have added a training-objective paragraph for each stage in Secs. 3.2–3.3, explicitly identifying the trainable components and referencing the detailed objectives and loss weights in Appendices B.2–B.3. During both stages, the pretrained Utonia backbone and the learned LoRA parameters remain frozen, providing fixed adapted features F′\mathbf F', while the respective task-specific modules are trained.

No-object handling. During training, unmatched queries are assigned the no-object label −1-1 and included in the face-type classification loss. During inference, queries classified as no-object are discarded. Face-type classification accuracy is 90.03%/84.32%/84.79% across the three test sets. Segmentation P/R/F1 and matched mIoU are computed on the retained face queries without requiring their predicted face types to match the ground truth.

λ\lambda. We have added the loss weight related to segmentation loss in Appendix B.2.


W6: Notations

Notation and mask terminology. We have added a notation table in the revised manuscript and summarize it below. Basicaly, we denote the input triangle mesh by M=(V,F)\mathcal M=(V,F) and the B-Rep solid B\mathcal B with B-Rep face, edge and vertex sets F,E,V\mathcal F,\mathcal E,\mathcal V. In Sec. 3.2, each cluster is represented by a mesh vertex index set SkS_k. The cluster-to-query assignment score A~\widetilde{A} determines cluster assignments to queries. The assigments are then lifted to mesh vertices to get B-Rep face masks on the input mesh, each of which is represented by a mesh vertex index set Ωnf\Omega_{n_f}. Ωne\Omega_{n_e} denotes the mesh vertex index set associated with an B-Rep edge candidate. Ωnf,Ωne\Omega_{n_f}, \Omega_{n_e} guide feature sampling from the adapted features F′\mathbf{F}' during reconstruction, yielding contexts Cf,nf\mathbf C_{f,n_f} and Ce,ne\mathbf C_{e,n_e}.

SymbolDefinition
M=(V,F)\mathcal M =(V,F)Input triangle mesh, with vertex set VV and triangle set FF.
B, F,E,V\mathcal B,\ \mathcal F,\mathcal E,\mathcal VB-Rep solid and its face, edge, and vertex sets.
SkS_kIndex set of mesh vertices belonging to cluster kk.
A~\widetilde{A}Cluster-to-query assignment scores.
Ωnf,Ωne\Omega_{n_f},\Omega_{n_e}Mesh vertex index sets associated with the nfn_f-th B-Rep face and nen_e-th edge candidate.
Cf,nf,Ce,ne\mathbf C_{f,n_f},\mathbf C_{e,n_e}B-Rep face and edge reconstruction contexts.

Token layout and rationale. We initialize each B-Rep face with 2×22\times2 local tokens and each B-Rep edge candidate with 22 local tokens to limit the decoding cost. Local token grids are progressively upsampled and decoded to the output face and edge point grids. Progressive upsampling increases spatial resolution in later stages, allowing finer geometric predictions while keeping the earlier stages inexpensive. We haved clarified the global and topology tokens in the revised text. Basically, the global token is for reducing computational cost and the topology token is for B-Rep edge prediction.

AuthorsOpenReview ↗

Response to Minor issues


M1: Metric Definitions and Numerical Consistency

We appologize for these reporting errors. The correct CD threshold for a successful match is 0.1, consistent with the DualBrep evaluation protocol. In Table 2, the recall and F1 values for our method on the DualBrep test set at IoU 0.50 were wrongly copied. The correct precision, recall, and F1 are 0.9731, 0.9705, and 0.9718, respectively. We corrected both errors in the revised manuscript.


M2: Validity Evaluation and Reconstruction Protocol

Response: We use OpenCASCADE through pythonocc-core 7.9.0. Following DualBRep, a reconstruction is considered valid only if the exported STEP represents a TopAbs_SOLID and passes BRepCheck_Analyzer.IsValid(), with a geometric tolerance of 0.1. Shell and compound reconstructions are coutned as invalid since they are not watertight. These fails are included in metrics calculation.

Post-processing first identifies candidate B-Rep vertices from the face–edge adjacency and predicted edge grids. For each triple of mutually adjacent faces, endpoints of their shared edges are grouped if their mean pairwise distance is at most 0.2, with each group representing a candidate vertex.

We then optimize the predicted face and edge grids using the face–edge adjacency and endpoint groups to enforce face–edge alignment and shared-endpoint consistency. Each edge grid is adjusted through per-axis scaling and translation, parameterized by a six-dimensional vector, while each face grid is adjusted through a three-dimensional translation vector. Optimization runs for 200 steps using AdamW with a learning rate of 10−310^{-3}. B-spline surfaces and curves are then fitted to the optimized grids. The fitted curves are assembled into boundary wires to trim the surfaces, and the resulting faces are sewn together.


M3: BGPSeg results

Response: Since the publicly released BGPSeg training code is incomplete, we evaluate its released pretrained model on all three test sets and report the results below.

TestsetMatched mIoU (%)P @ IoU 0.50R @ IoU 0.50F1 @ IoU 0.50P @ IoU 0.75R @ IoU 0.75F1 @ IoU 0.75
DualBRep35.950.3180.3660.3110.3610.3070.260
ParseNet54.400.3190.5620.3840.2770.4920.335
Ours35.220.2650.3460.2760.2160.2840.226

I thank the authors for their efforts and clarifications during the rebuttal. Most of my concerns have been addressed, and I am willing to raise my rating to weak acceptance.

Reviewer_ac7YOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes BRepLift for converting 3D meshes to B-Reps by using shared 3D foundation priors. In this framework, primitive-aware 3D features are first extracted as geometric priors and used for face segmentation, geometry recovery, and topology recovery. Specifically, the PTv3-based model Utonia, adapted with LoRA, is used to extract vertex-wise features, and a decoder is trained to reconstruct surface and curve geometry and predict edge existence. Experimental results demonstrate that the proposed method surpasses previous methods on most evaluated metrics across multiple test sets.

Critical strengths

This paper is well-written and technically sound. Introducing 3D foundation priors makes sense, since these models are pretrained to represent 3D geometry well. In addition, the design of using shared 3D foundation priors across the segmentation and recovery stages can help reduce information loss and make the results more consistent across multiple stages.

Critical weaknesses

The ablation study is insufficient. The core motivation of the paper is to use shared 3D foundation priors, but the experimental section does not directly compare shared priors with priors extracted independently at each stage.

Questions for authors
  1. Could the authors further provide additional ablation experiments on shared priors?

  2. I noticed that the converted B-Rep results in the paper and supplementary material do not consist of standard straight lines or curves. Is there any way to make them more standard?

  3. There have been many recent works based on large generative models. Could the authors further discuss the advantages and disadvantages of directly using large generative models to predict B-Rep representations versus the current approach (large generative models may produce more standardized results)?

3 comments

W1 (Q1): Shared Versus Independently Extracted Priors

We have added ablations to support the sufficiency of the shared adapted representation for reconstruction. Our primitive-aware adaptation uses face-level contrastive supervision (Sec. 3.1) to encourage feature consistency within faces and separation across faces. We therefore focus on whether this representation can also support reconstruction without further task-specific adaptation. Detailed settings and results are provided in our response to Reviewer XqXd, W1.


Q2: Standardizing Reconstructed B-Rep edges

Yes. Our framework can produce more standardized geometry by incorporating analytic primitive fitting into post-processing, without changing the learned models. On the ABC-ParSeNet test set, this simple extension increased the proportion of analytic faces from 0% to 15.43% and analytic edges from 0.03% to 8.15%, with a net increase of 25 valid reconstructions.

Our original post-processing follows DualBRep by fitting B-spline surfaces and curves to face and edge point grids, without explicitly recovering analytic primitive types. We add an analytic-fitting stage after grid optimization that fits planes, cylinders, and cones to face grids, and straight lines and circular arcs to edge grids. An analytic fit is accepted only if its geometric deviation from the optimized grid is within 0.001τ0.001\tau and it remains consistent with adjacent primitives, where τ\tau is the bounding-box diagonal of the corresponding optimized grid. Otherwise, B-spline fitting is retained.


Q3: Comparison with Large Generative Models

Generative methods learn global CAD priors that support completion of incomplete or ambiguous inputs, but learned abstraction and stochastic sampling may compromise fidelity to local details. Our method explicitly localizes face instances and constructs per-face geometric contexts, providing local evidence for reconstructing small faces and fine details.

Geometric regularity depends on primitive representations and post-processing, rather than generative modeling alone. Like our method, BrepGPT [1] and HoLa [2] reconstruct geometry by B-spline fitting, whereas BrepForge [3] and ParaCAD [4] explicitly encode analytic primitive types and parameters. Preserving these properties during B-Rep assembly also requires compatible post-processing. Such analytic fitting can also be incorporated into our framework, as demonstrated in the Q2.

Among the methods examined, we could not identify a complete, publicly available implementation supporting point-cloud- or mesh-conditioned B-Rep reconstruction that would enable a fair experimental comparison. We would welcome a specific reference and discuss the comparison further.

[1] Li P, Zhang W, Quan W, et al. Brepgpt: Autoregressive b-rep generation with voronoi half-patch[J]. ACM Transactions on Graphics (TOG), 2025, 44(6): 1-18.

[2] Liu Y, Xu D, Yu X, et al. Hola: B-rep generation using a holistic latent representation[J]. ACM Transactions on Graphics (TOG), 2025, 44(4): 1-25.

[3] Li J, Fu Y, Chen F. BrepForge: Factorized B-rep Synthesis via Wireframe Composition and Boundary-Conditioned Surface Instantiation[C]//Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers. 2026: 1-11.

[4] Qin D, Xu R, Shen Z, et al. Autoregressive B-Rep Shape Generation with Parametric Surfaces[J]. arXiv preprint arXiv:2607.17093, 2026.

Thanks a lot for the responses, most of my concerns are addressed and I will keep my rating at weak accept. Regarding Q3, I understand that a direct comparison may not be fair, but the response only states that no suitable public implementation was found. Since the paper already cites program-based methods, e.g. CAD-Recode and cadrille, which provide public pretrained weights and point-cloud-conditioned inference code, could the authors clarify what makes program-based methods unsuitable for comparison?

Thank you for the follow-up and for acknowledging our responses. We apologize for the ambiguity. No suitable public implementation primarily referred to generative models that directly predicting B-Rep geometry and topology rather than program-based methods. We omitted CAD-Recode and cadrille because their sketch-and-extrude formulation has limited expressiveness for general fillets, chamfers, and free-form surfaces. We agree this comparison helps reveal the strengths and limitations of different approaches. We have therefore evaluated cadrille on ABC-ParSeNet.

MethodShape CD ↓\downarrowNC ↑\uparrowFace F1 ↑\uparrowEdge F1 ↑\uparrowVertex F1 ↑\uparrowFace-Edge F1 ↑\uparrowEdge-Vertex F1 ↑\uparrowValidity ratio ↑\uparrow
CADrille0.02780.8470.4350.3460.3070.3990.32968.27
Ours0.03330.8900.7450.5920.6220.6440.64922.14

cadrille achieves lower shape CD and substantially higher validity, demonstrating the strengths in overall shape approximation and producing valid CAD models. Our method achieves higher normal consistency and all primitive and topological F1 scores, indicating better recovery of local geometry and structure.

Comments on the paper

1 comment

We sincerely thank the reviewers for their follow-up responses, constructive suggestions, and thoughtful discussions, and we thank the Area Chair for overseeing the review process. Your suggestions have helped us strengthen the experiments and clarify the presentation. Thank you all for your time and careful consideration.

Best regards, The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.