B-repLer: Text-Driven B-rep Editing

• B-repLer is a text-driven editing framework that converts natural language prompts into precise B-rep CAD model edits while preserving CAD validity.
• It employs a two-stage architecture: intent understanding with multimodal LLMs and latent-space modification via a bespoke HoLa-BRep representation.
• Empirical results on the BrepEDIT-10K dataset demonstrate improved validity, success rates, and F1 scores compared to baseline methods.

B-repLer is a text-driven editing framework for Boundary Representation (B-rep) CAD models that aims to transform high-level natural-language instructions into valid edited B-reps without leaving the native B-rep domain. It is presented as a finetuned multimodal LLM (mLLM) system for semantic editing, motivated by the fact that B-reps remain the industry-standard representation for engineering objects yet are fragile under direct manipulation, lack semantically annotated public corpora, and are difficult to edit from vague textual intent. The method addresses this by decomposing editing into intent understanding and localization on the one hand, and latent-space B-rep modification on the other, using a bespoke multimodal architecture built around a HoLa-BRep latent representation (Liu et al., 13 Aug 2025).

1. Problem setting and conceptual scope

B-repLer targets semantic editing rather than unconditional generation or reverse engineering. Its input is a pre-edit B-rep model, a rendered image of that model from a user-selected viewpoint, and a text prompt such as “strengthen the tip” or “remove the hole.” Its outputs are a localized 2D bounding box identifying the edit region, a refined low-level instruction prompt, and an edited B-rep model (Liu et al., 13 Aug 2025).

The framework is motivated by two structural difficulties. First, B-rep is fragile: invalidity can arise easily because geometry and topology must remain jointly consistent. Second, user prompts are often abstract rather than operational. B-repLer therefore does not ask for explicit CAD commands; instead, it uses an mLLM to infer an edit plan and then executes that plan in a learned B-rep latent space. The paper describes this as native semantic editing of B-rep CAD models, contrasting it with pipelines that pass through images or procedural CAD code (Liu et al., 13 Aug 2025).

A central design choice is to split the task into two subproblems: understanding and localizing user intent, then performing the actual B-rep modification. This suggests that the system treats language grounding and geometric editing as distinct failure modes rather than as a single end-to-end prediction problem. A related implication is that the method is structured to preserve CAD validity by editing a representation already optimized for B-rep coherence rather than manipulating raw geometry directly.

2. Two-stage architecture

The architecture is explicitly two-stage. In Stage I, a finetuned multimodal LLM, specifically Qwen2.5 VL 7B, receives the source rendering $I_b$ and the high-level prompt $T_{b\rightarrow a}$, and predicts a 2D bounding box $B \in \mathbb{R}^4$ together with a low-level instruct prompt $\hat{T}_{b\rightarrow a}$: $$(B, \hat{T}_{b\rightarrow a}) = \text{Qwen}(I_b, T_{b\rightarrow a}).$$ This stage converts ambiguous intent into a concrete edit target (Liu et al., 13 Aug 2025).

Stage II is a latent modifier Transformer that operates in HoLa-BRep latent space. The pre-edit B-rep $M_b$ is encoded into a latent token sequence

$\{H_b\}_{i=1}^{N} \in \mathbb{R}^{N \times 32},$

and the target edited model $M_a$ is likewise represented as

$\{H_a\}_{i=1}^{N'} \in \mathbb{R}^{N' \times 32}.$

The latent editor predicts the post-edit latent sequence autoregressively and decodes it back into a B-rep (Liu et al., 13 Aug 2025).

The use of HoLa-BRep is structurally important. HoLa compresses a full B-rep into a latent space defined on surfaces while reconstructing curves, vertices, and topology through neural surface-pair interactions rather than storing those entities as separate latent branches (Liu et al., 19 Apr 2025). In B-repLer, this means the edit model inherits a compact latent space that was designed to preserve CAD integrity, which is why the paper treats HoLa as the basis for validity-aware latent editing rather than as a generic geometry encoder (Liu et al., 13 Aug 2025).

3. Multimodal fusion and B-rep-specific alignment

B-repLer’s second stage fuses four modalities into a common 768D space: B-rep latent features through a trainable MLP, image features from a frozen DINOv2 encoder plus projection, bounding-box features from an MLP on the 4D box, and text features from a frozen Qwen text encoder plus projection. The resulting source sequence is

$$S_b := \{ \{\mathcal{F}_\text{brep}^i + \mathcal{F}_\text{roi}^i\}_{i=1}^N, \mathcal{F}_\text{img}, \mathcal{F}_\text{bbox}, \mathcal{F}_\text{text} \}.$$

The decoder then predicts a target sequence containing a start token, the post-edit image feature, the post-edit B-rep latent tokens, and an end token, with outputs projected back to 32D via $T_{b\rightarrow a}$0 (Liu et al., 13 Aug 2025).

A distinctive B-rep-aware module is RoI image-to-face alignment. For each B-rep face, the system extracts its 2D bounding box in the rendered view, applies RoIAlign to the DINO feature map, transforms that region into a 768D feature $T_{b\rightarrow a}$1, and adds it to the corresponding face latent. This mechanism addresses a specific mismatch: the 2D rendering and the 3D B-rep faces are not naturally aligned. The localized image feature therefore acts as visual grounding for face-level latent editing (Liu et al., 13 Aug 2025).

The decoder uses causal autoregressive decoding with cross-attention and masking, while Stage II training uses MSE on B-rep and DINO features. This design indicates that the target of learning is not only the edited latent geometry but also consistency with the post-edit visual representation. A plausible implication is that image-space supervision functions as an auxiliary regularizer on the latent edit trajectory, especially when multiple latent edits could produce similar topological outcomes.

4. BrepEDIT-10K and automatic supervision

Because no public dataset exists with text-annotated B-rep edits, B-repLer introduces BrepEDIT-10K, built on top of the ABC dataset. Each sample contains a pre-edit model $T_{b\rightarrow a}$2, a post-edit model $T_{b\rightarrow a}$3, paired renderings $T_{b\rightarrow a}$4 and $T_{b\rightarrow a}$5, a 2D bounding box $T_{b\rightarrow a}$6, detailed geometric instructions $T_{b\rightarrow a}$7, and abstract user-style prompts $T_{b\rightarrow a}$8. The dataset statistics reported are 10k shapes, 100k editing operations, 600k text annotations, about 10 valid edits per shape, 6 diverse text annotations per operation, and an 8:1:1 train/val/test split (Liu et al., 13 Aug 2025).

The data-generation pipeline combines CAD tools and mLLM annotation. In Fusion360, each face is tried as a deletion target; if the result remains valid, the before/after pair is retained. The system then chooses the best among 32 isometric views to show the edit clearly, projects the edited face to image space, and computes a 2D bounding box. Finally, Gemini Flash 2.0 is prompted with side-by-side before/after images and the boxed edit region to produce both detailed geometric instructions and higher-level user prompts for both edit directions (Liu et al., 13 Aug 2025).

This dataset construction method is directly tied to the two-stage model design. High-level prompts supervise intent understanding, while low-level instructions supervise geometric edit translation. At the same time, the operational basis of the dataset is restricted: the current implementation supports delete and add operations, since addition is obtained as the inverse of face deletion. The paper nonetheless reports semantically rich prompts, including more abstract instructions such as “increase aerodynamics,” which are translated into geometry-level edits rather than validated by physical analysis (Liu et al., 13 Aug 2025).

5. Training protocol and empirical performance

Stage I fine-tunes Qwen2.5 VL 7B with DoRA adapters using dropout 0.2, adapter rank 256, alpha 512, cross-entropy loss, batch size 32, learning rate $T_{b\rightarrow a}$9, and a cosine schedule, requiring about 23 hours on one H100. Stage II trains the latent modifier Transformer from scratch with batch size 1024, learning rate $B \in \mathbb{R}^4$0, AdamW, and MSE loss on B-rep and DINO features, taking about 48 hours on one H100 80GB (Liu et al., 13 Aug 2025).

Evaluation uses BrepEDIT-10K with a synthetic baseline denoted “mLLM + HoLa.” In that baseline, GPT4.1 first generates a post-edit image from the source rendering and text prompt, and HoLa then reconstructs a B-rep from that image. The paper reports the following results. For “mLLM + HoLa,” validity is 63.6%, success is 0%, and face/edge/vertex F1 are 0.142, 0.055, and 0.036. For B-repLer, validity is 69.8%, success is 53.4%, and face/edge/vertex F1 are 0.914, 0.892, and 0.886. In the ablation table, the full model reaches validity 70.4%, success 54.4%, and face/edge/vertex F1 of 0.909, 0.888, and 0.880 (Liu et al., 13 Aug 2025).

The success criterion is stringent. Matching is performed with Hungarian matching over faces, edges, and vertices based on Chamfer distance with a threshold of 0.1, and an edit is considered successful only if the predicted model exactly matches the ground truth, i.e., $B \in \mathbb{R}^4$1. Validity is defined using OpenCASCADE and requires the output B-rep to be manifold, closed, and free of self-intersections (Liu et al., 13 Aug 2025).

The localization and instruction-translation subsystem is evaluated separately. The reported bounding-box IoU is approximately 0.51 on unseen operations and approximately 0.31 on unseen objects; normalized box distance is approximately 0.08 on unseen operations and approximately 0.1 on unseen shapes; instruction translation cosine similarity is approximately 0.84, versus 0.67 for the raw user prompt compared directly to the target instruction. These figures support the paper’s claim that translating high-level prompts into low-level edit descriptions is an essential intermediate step rather than an auxiliary convenience (Liu et al., 13 Aug 2025).

6. Relation to the B-rep literature, misconceptions, and limitations

B-repLer belongs to the broader line of direct B-rep learning, but its task differs from recent work on unconditional generation, topology modeling, or continuous B-rep representations. BrepGen generates B-reps through a hierarchical tree with duplicated nodes and implicit topology recovery (Xu et al., 2024). HoLa unifies geometry and topology in a holistic latent space defined on surfaces and reconstructs curves and topology through a neural intersection network (Liu et al., 19 Apr 2025). GraphBrep explicitly models topology as a surface adjacency matrix and factorizes generation into surfaces, adjacency, and edges (Lai et al., 7 Jul 2025). DualBrep moves geometry and topology into a continuous dual-field representation based on an SDF and a UDF and reconstructs explicit B-reps through a learned rebuilder (Liu et al., 30 Jun 2026). Against this backdrop, B-repLer is best understood as a semantic editing system that reuses a learned B-rep latent space rather than as a new unconditional generator or reverse-engineering backbone (Liu et al., 13 Aug 2025).

Several misconceptions are addressed by the paper’s design and stated limitations. B-repLer does not execute symbolic CAD programs and does not require the user to specify exact operations. It also does not perform physical simulation: prompts such as “strengthen” or “improve aerodynamics” are translated into geometry-level edits, but the method is explicitly text-only and does not check functionality such as strength or aerodynamics (Liu et al., 13 Aug 2025). Likewise, the current implementation does not support arbitrary multi-step editing; it supports single-step delete and add operations, and it does not explicitly model relations such as symmetry, orthogonality, or parallelism between multiple faces (Liu et al., 13 Aug 2025).

Failure cases arise when the mLLM misunderstands the prompt, localizes the wrong region, or the latent modifier fails, in which case the output may remain unchanged or fail to reflect the intended edit. The paper also notes that validity remains challenging because editing noise lowers it further. These caveats are important because they locate the method’s contribution precisely: B-repLer demonstrates that text-based semantic editing of valid B-rep CAD models is possible with a native latent-space approach, but it does not yet provide complete semantic reasoning over engineering intent or guaranteed edit correctness (Liu et al., 13 Aug 2025).

From a broader research perspective, B-repLer suggests a shift in how CAD editing can be posed. Rather than viewing semantic editing as text-to-image generation followed by reconstruction, it formulates the problem as text-guided transformation of a validity-aware B-rep latent. This suggests that future systems may combine instruction grounding, region localization, and native B-rep latent manipulation as the core ingredients of practical language-driven CAD editing.