acceptodds
Withdrawn from ICLR 2027

BRepPreDiff: Diffusion-Based Self-Supervised Representation Learning for Boundary Representations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

10 views

Abstract

Diffusion models have recently demonstrated strong potential for generating structured boundary representation (B-Rep) models by modeling complex geometric distributions and dependencies. However, their use for transferable B-Rep representation learning remains largely unexplored. Meanwhile, large-scale Computer-Aided Design (CAD) datasets are predominantly unlabeled, and models stored in intermediate exchange formats typically lack their original modeling histories, making scalable semantic representation learning from CAD data particularly challenging. To address this problem, we introduce , a diffusion-based self-supervised pre-training framework for learning transferable B-Rep representations across downstream tasks, including face-level semantic segmentation and shape classification. BRepPreDiff consists of a and a strategy. The encoder employs edge-update attention to explicitly model geometric and topological interactions between adjacent B-Rep faces and their relations. During pre-training, we preserve the B-Rep topology while corrupting geometric features at sampled noise levels and train the encoder to predict the injected noise and reconstruct clean geometric features. This multi-level denoising objective enables the model to learn transferable geometric and structural features from unlabeled B-Rep data. For downstream adaptation, the pretrained encoder is jointly fine-tuned with lightweight MLP heads for face-level semantic segmentation and whole-model classification. BRepPreDiff achieves mIoU on Fusion360Seg and classification accuracy on TMCAD, exceeding the strongest compared baselines by 5.05 and 2.28 percentage points, respectively. We further introduce , a CAD dataset comprising 18K B-Rep models with face-level blend annotations, providing a large-scale benchmark for blend-face recognition and semantic supervision for design-intent recovery and feature-aware CAD analysis.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Boundary Representation, Self-Supervised Learning, Diffusion Models, Representation Learning, Computer-Aided Design

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_RvJ5OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes BRepPreDiff for learning transferable geometric and topological representations from large-scale unlabeled CAD/B-Rep data. The method employs a topology-aware B-Rep graph encoder with edge-update attention to model the relationships between adjacent faces and shared edges. During pre-training, Gaussian noise at different levels is applied to face and edge geometric features, and the model is jointly trained for noise prediction, clean-feature reconstruction, attribute prediction, and adjacency-relation prediction. The authors also introduce BLEND, a dataset containing approximately 18K B-Rep models, and evaluate the proposed method on multiple segmentation and classification tasks.

Critical strengths
  1. The problem is practically meaningful, as it aims to exploit large-scale unlabeled CAD data to alleviate the scarcity of face-level semantic annotations in downstream tasks.
  2. The edge-update attention mechanism explicitly models the interactions among faces, edges, and adjacent faces, making the architecture well aligned with the structural characteristics of B-Rep models.
Critical weaknesses
  1. The methodological contribution of diffusion-based representation learning is not sufficiently established. The pre-training procedure is more closely related to a multi-level denoising autoencoder with timestep conditioning, noise prediction, and auxiliary multi-task supervision. Although the method adopts a diffusion-style noise schedule, it does not perform reverse diffusion or use a learned score function for generation. The current experiments lack strict comparisons against fixed-level denoising, self-supervised Gaussian corruption, a variant without timestep conditioning, and a model trained only with clean-feature reconstruction. Therefore, it remains unclear whether the performance gains come from the diffusion objective itself, or simply from stronger geometric perturbations and additional auxiliary tasks. The authors should provide these ablations, or position the method more accurately as topology-preserving multi-level denoising pre-training.
  2. The paper lacks pre-training scaling experiments. The paper pre-trains on approximately 512K unlabeled B-Rep models, but reports results only for a fixed pre-training corpus, model capacity, and training budget. A key question for pre-training is whether scaling the data and computation leads to consistent improvements on downstream tasks. The different annotation ratios in Table 3 mainly measure downstream label efficiency and cannot replace scaling analyses with respect to pre-training data size or computational budget. The authors should report performance as a function of pre-training corpus size, model capacity, training steps, and noise schedule.
Questions for authors

See Weaknesses.

Reviewer_eqK7OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes BRepPreDiff, which learns transferable B-rep representations through diffusion-based denoising. Additionally, the paper constructs the BLEND dataset, which contains approximately 18,000 CAD models. Experiments span seven downstream datasets, and results on few-label learning and computational efficiency are reported.

Critical strengths
  1. The method is designed to align with the characteristics of B-rep data. By using adjacent edges to guide the aggregation of face features and then updating edges based on face context, it naturally integrates geometric and topological information. The results are competitive across multiple real-world CAD datasets.

  2. The BLEND dataset has practical value. By distinguishing between edge-rounded faces and vertex-rounded faces, it provides richer annotation resources for fine-grained rounding recognition and related CAD analysis.

Critical weaknesses
  1. The performance does not appear to stem directly from the core contribution. In Table 4, the same encoder trained from scratch achieved 85.93 mIoU on Fusion360Seg, while the complete method achieved 86.64; the results from training from scratch already outperformed the other methods in Table 1. Therefore, the current overall advantage likely stems primarily from the encoder or the input design. The paper lacks comparisons with masked reconstruction, standard denoising, and diffusion pre-training using the same encoder, inputs, and computational budget. Additionally, the low-label experiments lack a baseline of training the same encoder from scratch, making it difficult to assess the necessity of the diffusion objective.

  2. The effectiveness and cost-benefit analysis of the pre-training objective are insufficient. Section 3.3 explicitly state that discrete categories and adjacency attributes simultaneously serve as both uncorrupted inputs and prediction targets; thus, a shortcut exists to complete predictions by retaining input information. However, the paper lacks ablation studies involving the removal or masking of these attributes. Furthermore, Table 2 primarily measures the inference overhead during downstraem training with cached inputs and does not report the full pretraining cost, which does not yet sufficiently support the claim that diffusion-based self-supervised learning is more efficient.

  3. The figures in the paper are overly crude and very difficult to read, particularly Figure 1.

Questions for authors
  1. Given the same encoder, input features, and training budget, what evidence suggests that the gains stem from diffusion pretraining rather than encoder design or general geometric reconstruction? In particular, under low-label conditions, how significant are the gains relative to training the same encoder from scratch?

  2. When discrete attributes serve as both inputs and prediction targets, how can we prevent the shortcut of directly copying the inputs? How does downstream performance change after removing the category and relationship prediction losses?

  3. How many GPU-hours does full pretraining require? After accounting for the cost of pretraining, to which specific stages and use cases do the paper’s conclusions regarding efficiency advantages apply?

Reviewer_ZeAeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents BRepPreDiff, a diffusion-based self-supervised framework for learning transferable B-Rep representations. It combines topology-preserving, multi-level geometric denoising with a topology-aware encoder that jointly updates face and edge representations. The pretrained encoder is fine-tuned for face-level semantic segmentation and shape classification. The paper also introduces BLEND, a dataset of approximately 18,000 B-Rep models with annotations distinguishing edge-blend from vertex-blend faces, and evaluates the framework through downstream comparisons and ablation studies.

Critical strengths
  1. BLEND provides over one million face annotations across approximately 18,000 CAD models, distinguishing edge-blend from vertex-blend faces and supporting fine-grained blend recognition.

  2. The encoder explicitly couples adjacent face and edge representations. The fixed-edge and edge-only ablations in Table 4 support the benefit of updating edges using incident-face context.

  3. The evaluation covers seven benchmarks spanning face segmentation and shape classification, with additional experiments under limited supervision and on a dataset excluded from pre-training.

Critical weaknesses
  1. Eq. (2) adds Gaussian noise to geometric features at a sampled noise level: Xtq=αˉtX0q+1−αˉtϵq.X_t^q=\sqrt{\bar{\alpha}_t}X_0^q+\sqrt{1-\bar{\alpha}_t}\epsilon^q. The claimed benefit of multi-level denoising in Lines 61–65 is not isolated from the encoder design or the other pre-training objectives. Table 4 lacks same-encoder comparisons with fixed-noise denoising and masked reconstruction, while Eq. (15) combines noise prediction, clean-feature reconstruction, and attribute prediction without loss ablations. The categorical targets s,c,rs,c,r in Eqs. (13)–(14) are also supplied unchanged through the embeddings in Eqs. (3)–(4), allowing attribute prediction through retention of input information rather than geometric reasoning. The reported gains therefore do not establish the specific contribution of diffusion-based pre-training.

  2. The transferable benefit of pre-training remains unclear. Table 4 shows a full-supervision mIoU increase from 85.93 to 86.64, versus 36.25 for frozen-encoder linear probing, but the latter comparison changes both encoder adaptation and the prediction head. Table 3 lacks a same-encoder, from-scratch baseline under limited supervision. Run-to-run variability and total pre-training cost are also unreported; Table 2 measures downstream training efficiency.

  3. BLEND’s annotation reliability is insufficiently documented. Lines 312–314 define edge-blend and vertex-blend faces, and Lines 320–323 report manual annotation and cross-checking. The paper does not specify labeling rules for ambiguous faces, disagreement resolution, or annotator agreement. These omissions make it difficult to assess the reliability of the distinction that constitutes BLEND’s central contribution.

Questions for authors
  1. Under the corruption process in Eq. (2), Xtq=αˉtX0q+1−αˉtϵq,q∈f,e,X_t^q=\sqrt{\bar{\alpha}_t}X_0^q+\sqrt{1-\bar{\alpha}_t}\epsilon^q,\qquad q\in\\{f,e\\}, what evidence shows that sampling multiple noise levels improves transfer over fixed-noise denoising or masked reconstruction with the same encoder and training budget? Can the authors isolate the terms in Eq. (15),
=\lambda_{\mathrm{noise}}\mathcal{L}_{\mathrm{noise}} +\lambda_{\mathrm{recon}}\mathcal{L}_{\mathrm{recon}} +\lambda_{\mathrm{cat}}\mathcal{L}_{\mathrm{cat}} +\lambda_{\mathrm{rel}}\mathcal{L}_{\mathrm{rel}}?$$ Since $s,c,r$ enter the encoder unchanged in Eqs. (3)–(4), what prevents $\mathcal{L}_{\mathrm{cat}}$ and $\mathcal{L}_{\mathrm{rel}}$ from being solved by retaining these supplied attributes rather than learning geometric dependencies? 2. Table 4 shows a small gain from pre-training under full supervision and a large gap between frozen-encoder linear probing and full fine-tuning. What do these results reveal about the transferable features learned during pre-training, given that the prediction heads also differ? How can the limited-supervision gains in Table 3 be attributed to pre-training rather than encoder design? How stable are these gains across runs, and what computational cost does pre-training add? 3. What labeling rules were used when a face was difficult to classify as edge-blend, vertex-blend, or non-blend? Lines 320–323 mention cross-checking: how were conflicting labels resolved, and how was annotation consistency assessed?

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

Related papers

Discussion (0)

Sign in to comment.