B-repLer: Language-guided Editing of CAD Models

TL;DR

The paper tackles the challenge of language-guided CAD editing by proposing B-repLer, which edits directly in the B-rep latent space without construction history. It introduces BrepEDIT-240K, a large-scale dataset built with a multimodal generation pipeline and CAD tooling to produce paired pre-/post-edit B-rep models with rich text annotations. The core method combines a variational autoregressive transformer with a flow matching network to produce diverse, valid edits that respect B-rep topology and complex freeform surfaces. Experimental results show improvements over state-of-the-art baselines in edit realism, prompt alignment, and the ability to handle high-level semantic instructions, enabling a scalable, native text-guided editing workflow for CAD models.

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.

Figures (17)

• Figure 1: Text-driven B-rep editing. We present B-rep Latent Editor (B-repLer), the first framework to perform generative CAD editing directly in the Boundary Representation (B-rep), without the need for the construction history. Our method can interpret high-level user instructions (in orange) to perform fine-grained modifications on B-rep models with complex and freeform geometries. By operating in the B-rep latent space, B-repLer generates a diverse set of valid B-rep edits from a single fixed input (left). It also supports multi-edit for complex freeform surfaces and functional descriptions, while preserving identities in unchanged areas (right).
• Figure 2: B-rep Editing. Text-based B-rep editing is challenging. For example, (a) a training-free pipeline usually involves a multiple stages, where errors can occur and accumulate across different stages, limiting its controllability; (b) a trained mLLM yuan2025cadeditor used the expensive construction history as a text proxy to generate modified B-rep, while having limited shape and edit complexity; (c) we propose a novel native B-rep editing approach, where we link the text embedding to B-rep latent, thus enabling semantic-aware and fine-grained control over various primitives.
• Figure 3: Dataset comparison.BrepEDIT-240K has rich annotations across complex and practical CAD models, where CADEditor yuan2025cadeditor is intrinsically restricted by the sketch-and-extrude definition and limited annotation quality (zoom recommended).
• Figure 4: Dataset construction. We construct an annotated dataset in three main steps: (a) Given a B-rep model, we randomly select an operating face, and we then employ Fusion360 willis2021fusion to delete it. (b) For each successful deletion, we select the best viewpoint to render images before and after the operation, and (c) combine them with additional "Left" and "Right" texts to form an input image sent to an mLLM (i.e., Gemini 2.5 Pro gemini) to produce the corresponding editing instructions.
• Figure 5: Method Overview. We process the input B-rep model and user prompt using a pre-trained mLLM and a B-rep encoder to extract visual, textual, and B-rep features. These features are fused by a transformer encoder. Subsequently, a decoder autoregressively attends to the fused features to produce a target embedding, which conditions a flow matching network to generate the edited B-rep CAD model.