B-repLer: Semantic B-rep Latent Editor using Large Language Models

Yilin Liu, Niladri Shekhar Dutt, Changjian Li, Niloy J. Mitra

Abstract

Computer-Aided Design (CAD) models, given their compactness and precision, remain the industry standard for designing and fabricating engineering objects. However, language-guided CAD editing is still in its infancy, largely due to missing semantic connection between user commands and underlying shape geometry, a problem exacerbated by the shortage of paired text-and-edit CAD datasets. While recent Multimodal Large Language Models (mLLMs) have attempted to bridge this gap, their reliance on CAD construction history -- often an expensive and hard to obtain input -- severely limits their expressiveness and restricts their usage. We present B-repLer, a novel framework that directly connects natural language with editing CAD models by operating in a learned latent space. Importantly, our approach bypasses the need for construction history, enabling semantic edits on a wide range of geometries, from simple prismatic parts to complex freeform shapes defined by B-Spline surfaces. To facilitate this research, we introduce BrepEDIT-240K, the first large-scale dataset for this task. We demonstrate how this paired dataset can be automatically generated, (user) validated, and scaled by leveraging existing CAD tools, in conjunction with mLLMs, to create the required paired data without relying on any external annotations. Our results demonstrate that B-repLer can accurately perform complex edits on complex CAD shapes, even when the input edit specifications are high-level and ambiguous to interpret, consistently producing valid, high-quality CAD outputs enabling a class of text-guided edits not previously possible.

Introduction

In the domain of Computer-Aided Design (CAD), Boundary Representation (B-rep) models are the gold standard for precision and manufacturing readiness. A B-rep model defines an object through a collection of geometric surfaces—such as planes, cylinders, and complex B-splines—stitched together with precise topological connectivity. However, this precision comes with inherent fragility; even minor inconsistencies in vertex or curve definitions can render a model invalid for downstream engineering tasks.

Overview of B-repLer's semantic editing capabilities and diverse solution generation.

Current state-of-the-art tools for CAD editing often rely on construction history—the sequential "recipe" of sketches and extrusions used to build a part. In industrial practice, this history is frequently lost during file transfers or proprietary data stripping. Furthermore, existing models struggle to interpret "semantic" instructions that convey functional intent (e.g., "incorporate a seat for a bearing") rather than explicit geometric coordinates. B-repLer addresses these challenges by enabling direct, language-guided editing on final B-rep geometries without requiring their historical construction data.

Methodology: Latent Space Editing

B-repLer treats CAD editing as a sequence-to-sequence translation task within a learned latent space. This approach ensures that the output remains topologically valid and "watertight," a critical requirement for engineering applications.

Architectural Components

The framework utilizes a pre-trained B-rep autoencoder, HoLa-BRep, to map complex 3D shapes into a continuous latent representation. To process user intent, the model employs a multimodal fusion strategy, integrating four distinct inputs: the raw geometric data of the original part, user instructions encoded via the Qwen2.5 language model, rendered images of the part processed through DINOv2, and an optional 2D bounding box to focus the edit on a specific region.

Variational Autoregressive Transformer and Flow Matching

A core innovation of B-repLer is its hybrid generation strategy. Because a single natural language prompt can often be satisfied by multiple valid geometric solutions, the model uses a transformer to predict intermediate features. These features condition a Flow Matching network, which progressively refines the new B-rep tokens from noise. This method provides superior precision and diversity compared to purely deterministic or standard diffusion-based models.

Comparison of B-repLer against existing mLLM-based CAD editing baselines.

The BrepEDIT-240K Dataset

A significant barrier to progress in this field has been the lack of paired "text-and-edit" CAD datasets. To overcome this, the authors developed BrepEDIT-240K, comprising 240,000 addition and deletion operations across 52,000 unique shapes.

The dataset was constructed using an automated pipeline that leverages the Fusion360 CAD engine to perform "face deletions" on models from the ABC dataset. To generate high-quality natural language labels, the researchers utilized Gemini 2.5 Pro to analyze geometric differences and infer functional intent. This resulted in a multi-level annotation scheme that includes both concise functional descriptions (e.g., "improve structural integrity") and detailed geometric instructions.

Key Results and Evaluation

B-repLer was benchmarked against CADEditor, the previous state-of-the-art in generative CAD editing. The results demonstrate significant improvements across several metrics:

  • Semantic Alignment: B-repLer achieved a D-CLIP score of 0.34, significantly outperforming CADEditor's 0.25. This indicates that the generated edits are much more closely aligned with the user's textual intent.
  • Geometric Complexity: Unlike previous methods restricted to simple prismatic blocks, B-repLer successfully modifies complex B-spline (freeform) surfaces, which are essential for modern industrial design.
  • Model Validity: The system maintains a validity ratio of 97.13%, ensuring that the generated B-reps are compatible with professional CAD software.
  • Human Preference: In a blind user study, participants preferred B-repLer’s results 61.6% of the time, citing better adherence to functional requirements and higher geometric quality.

Conclusion and Future Directions

B-repLer represents a major step toward intuitive, language-driven CAD design. By bypassing the need for construction history and operating in a robust latent space, it enables a new class of semantic edits. However, limitations remain. The model currently struggles with precise counting (e.g., "add exactly five holes") and does not explicitly enforce engineering constraints like perfect symmetry. Furthermore, the model is "physics-blind," meaning it does not evaluate the structural or aerodynamic consequences of its edits. Future work will focus on integrating latent physical simulators to ensure that semantic edits are not only geometrically valid but also engineering-sound.

Summary updated

Abstract

Computer-Aided Design (CAD) models, given their compactness and precision, remain the industry standard for designing and fabricating engineering objects. However, language-guided CAD editing is still in its infancy, largely due to missing semantic connection between user commands and underlying shape geometry, a problem exacerbated by the shortage of paired text-and-edit CAD datasets. While recent Multimodal Large Language Models (mLLMs) have attempted to bridge this gap, their reliance on CAD construction history -- often an expensive and hard to obtain input -- severely limits their expressiveness and restricts their usage. We present B-repLer, a novel framework that directly connects natural language with editing CAD models by operating in a learned latent space. Importantly, our approach bypasses the need for construction history, enabling semantic edits on a wide range of geometries, from simple prismatic parts to complex freeform shapes defined by B-Spline surfaces. To facilitate this research, we introduce BrepEDIT-240K, the first large-scale dataset for this task. We demonstrate how this paired dataset can be automatically generated, (user) validated, and scaled by leveraging existing CAD tools, in conjunction with mLLMs, to create the required paired data without relying on any external annotations. Our results demonstrate that B-repLer can accurately perform complex edits on complex CAD shapes, even when the input edit specifications are high-level and ambiguous to interpret, consistently producing valid, high-quality CAD outputs enabling a class of text-guided edits not previously possible.
BibTeX
@inproceedings{liu2026b,
  author = {Yilin Liu and Niladri Shekhar Dutt and Changjian Li and Niloy J. Mitra},
  title = {{B-repLer: Semantic B-rep Latent Editor using Large Language Models}},
  booktitle = {SIGGRAPH (Conference Paper Track)},
  year = {2026},
  doi = {10.1145/3799902.3811166},
}
Figure