B-repLer: Language-guided Editing of CAD Models

Yilin Liu, Nikil 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, validated, and scaled by leveraging existing CAD tools in conjunction with mLLMs. 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, consistently producing valid, high-quality CAD outputs.

Introduction

Computer-Aided Design (CAD) is the backbone of modern engineering and manufacturing. Despite the precision of these models, editing them through natural language instructions remains a significant challenge. There is a profound semantic gap between high-level human intent (e.g., "modify the top face for mounting purposes") and the low-level geometric operations required to execute such changes.

Previous attempts to bridge this gap using Large Language Models (LLMs) have largely relied on a model's construction history—the sequential list of operations like sketches and extrusions. This dependency is problematic because construction history is often unavailable in industry settings and typically limits AI models to simple prismatic shapes. Furthermore, tokenizing 3D coordinates into text often introduces quantization errors that compromise geometric precision.

Comparison of B-repLer with existing CAD editing baselines, highlighting its ability to handle complex geometry without construction history.

The B-repLer Framework

To overcome these limitations, we introduce B-repLer, a framework that operates directly on the Boundary Representation (B-rep) of a 3D shape. B-repLer utilizes a learned latent space to represent surfaces, curves, and vertices, bypassing the need for construction history.

Architecture and Methodology

  • Multimodal Encoding: The system integrates three distinct inputs: the source CAD model (encoded via a pre-trained B-rep autoencoder), a rendered image of the model for visual context, and the user's natural language prompt. These are processed through a Transformer encoder to create a unified representation.
  • Variational Autoregressive Decoding: Because a single text prompt can imply multiple valid geometric solutions, B-repLer uses an autoregressive transformer to model the one-to-many nature of CAD editing.
  • Flow Matching for Precision: To maintain the mathematical rigor required for CAD, the model employs a flow matching network. Instead of predicting discrete tokens, the transformer predicts intermediate features that condition the generation of continuous B-rep latent tokens, ensuring high-fidelity outputs.

BrepEDIT-240K: A Large-Scale Dataset

A major hurdle in CAD research is the lack of paired "before-and-after" datasets with corresponding text instructions. We address this by introducing BrepEDIT-240K, featuring 240,000 edit operations across 52,000 unique shapes.

The automated pipeline for generating the BrepEDIT-240K dataset, utilizing face deletion and mLLM-based instruction labeling.

Our dataset was created through an automated pipeline:

  1. Geometric Perturbation: We used the Fusion360 engine to randomly delete faces from existing CAD models, creating valid "before" and "after" states.
  2. mLLM Annotation: We leveraged a multimodal LLM (Gemini 2.5 Pro) to analyze rendered images of these states. The mLLM generated both low-level geometric descriptions and high-level functional intents, effectively acting as a synthetic expert annotator.

Results and Discussion

B-repLer demonstrates a significant leap in performance over existing baselines like CAD-Editor. Quantitatively, it achieved a D-CLIP score of 0.34 (compared to 0.25 for the baseline), indicating superior alignment between the generated edits and the text prompts. Crucially, the framework maintained a 97.13% validity rate for the generated CAD files.

In user studies, B-repLer's outputs were preferred 61.6% of the time. Furthermore, users overwhelmingly preferred the BrepEDIT-240K dataset over previous benchmarks, citing higher model complexity and practical utility.

Conclusion and Future Work

B-repLer establishes a new state-of-the-art for language-guided CAD editing by operating in a continuous latent space and supporting native B-spline surfaces. While the model is highly capable, future work will focus on enforcing strict geometric constraints (such as symmetry and parallelism) and incorporating physical simulation data to ensure that edits are not only geometrically valid but also structurally sound under real-world stress.

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 Nikil Dutt and Changjian Li and Niloy J. Mitra},
  title = {{B-repLer: Language-guided Editing of CAD Models}},
  booktitle = {SIGGRAPH (Conference Paper Track)},
  year = {2026},
  doi = {10.1145/3799902.3811166},
}
Figure