Autoregressive B-Rep Shape Generation with Parametric Surfaces
ParaCAD employs a surface-centric autoregressive model to directly generate native parametric B-Rep surfaces conditioned on point clouds, improving geometric accuracy.
Key Findings
Methodology
ParaCAD introduces a surface tokenization scheme that explicitly encodes each B-Rep face by its surface type and continuous parameters, forming a discrete sequence compatible with GPT-based autoregressive modeling. The approach models common analytic primitives (planes, cylinders, cones, spheres, tori) with fixed intrinsic parameters via quantization, while freeform surfaces are encoded using a learned latent grid and finite scalar quantization. The model conditions on point clouds, generating sequences of surface tokens that represent the geometry. Post-processing involves geometric intersection and clipping to assemble a valid, watertight B-Rep. This method avoids the errors introduced by grid-based surface approximations, maintaining the original parametric semantics and enabling accurate, editable CAD models.
Key Results
- On a dataset of 583,000 point cloud-CAD pairs, ParaCAD reduces geometric error to 0.3mm, outperforming baseline latent grid methods at 0.8mm. It maintains high surface-type consistency and robustness across complex geometries, supporting downstream editing and manufacturing tasks. The generated B-Reps exhibit superior geometric fidelity, topological correctness, and watertightness, facilitating reliable CAD operations.
- Quantitative evaluations show that ParaCAD achieves a 65% reduction in geometric error compared to prior latent-based approaches. The model demonstrates strong generalization to complex freeform and multi-faceted surfaces, with stable performance under noisy point cloud conditions. Ablation studies confirm that explicit surface tokenization and geometric post-processing are critical for accuracy.
- In practical applications, ParaCAD supports high-quality parameter editing, topology repair, and manufacturability, significantly enhancing CAD usability. Its ability to generate precise, native-parameter surfaces from raw point clouds opens new avenues for automated reverse engineering, design, and digital twin creation.
Significance
This work addresses a longstanding challenge in CAD automation: how to generate high-fidelity, parameterized boundary representations directly from raw point cloud data. By operating in the native parametric surface space, ParaCAD preserves CAD semantics, enabling precise editing, simulation, and manufacturing. It overcomes the limitations of grid-based surface approximations, reducing errors and improving model reliability. The approach bridges the gap between raw sensor data and production-ready CAD models, promising to accelerate workflows in industrial design, reverse engineering, and digital manufacturing. Its ability to produce watertight, topologically valid B-Reps with explicit surface types marks a significant step forward for intelligent CAD systems.
Technical Contribution
The key technical innovations include: 1) a novel surface-centric tokenization scheme that explicitly encodes surface types and parameters, supporting both analytic primitives and freeform patches; 2) a GPT-based autoregressive model conditioned on point clouds, capable of sequentially generating surface tokens with high fidelity; 3) a robust geometric post-processing pipeline utilizing surface-surface intersections and clipping to assemble valid B-Reps. These contributions enable direct, high-precision generation of CAD models, maintaining semantic integrity and geometric correctness, which was not achievable with prior latent or grid-based methods.
Novelty
This is the first work to directly generate native parametric B-Rep surfaces conditioned on point clouds using a sequence modeling approach. Unlike existing methods that rely on latent space encoding or grid approximation, ParaCAD explicitly encodes surface types and parameters, preserving CAD semantics. Its surface tokenization and geometric assembly pipeline provide a new paradigm for high-fidelity CAD generation, bridging the gap between raw sensor data and production-quality models, representing a significant innovation in the field.
Limitations
- The model's performance degrades on extremely noisy or sparse point clouds, where geometric ambiguities hinder accurate surface prediction. Numerical stability issues in geometric intersection operations can lead to incomplete or topologically incorrect B-Reps. Additionally, training requires large annotated datasets, which are costly to produce, limiting scalability. Computational costs remain high for complex models, necessitating further optimization for industrial deployment.
- The current framework primarily handles common primitive and freeform surfaces but may struggle with highly intricate or degenerate geometries. Extending the approach to handle more diverse surface types and complex topologies remains an open challenge. Future work should focus on improving robustness, efficiency, and generalization to real-world noisy data.
Future Work
Future directions include integrating end-to-end learning of geometric intersections to reduce post-processing complexity, developing multi-resolution models for large-scale assemblies, and expanding surface type support. Additionally, exploring unsupervised or semi-supervised training paradigms could reduce data annotation costs. Applying the framework to real industrial datasets and integrating with CAD editing tools will facilitate practical adoption. Further research may also investigate real-time generation for interactive design workflows.
AI Executive Summary
In modern manufacturing and industrial design, creating accurate CAD models from raw sensor data remains a critical challenge. Traditional approaches often rely on approximations or manual modeling, which are time-consuming and prone to errors. Recent advances in deep learning have enabled some automation, but most methods either encode geometry in latent spaces or approximate surfaces via grids, leading to inaccuracies and loss of CAD semantics. This paper introduces ParaCAD, a groundbreaking framework that directly generates boundary representation (B-Rep) models with native parametric surfaces conditioned on point clouds.
The core innovation lies in a surface-centric tokenization scheme that explicitly encodes each face by its surface type and continuous parameters, supporting both analytic primitives and freeform patches. This tokenization enables a GPT-based autoregressive model to generate sequences of surface tokens conditioned on input point clouds, capturing complex geometries with high fidelity. The model's architecture leverages the strengths of transformer-based sequence prediction, ensuring stable and scalable generation.
Post-processing involves geometric intersection and clipping operations to assemble a valid, watertight B-Rep, preserving the original semantic and parametric information. Extensive experiments on a dataset of 583,000 point cloud-CAD pairs demonstrate that ParaCAD significantly outperforms existing methods, reducing geometric errors to 0.3mm and maintaining high topological correctness. The generated models support downstream CAD operations such as editing, topology repair, and manufacturing, indicating strong practical potential.
This work marks a major step forward in automated CAD modeling, bridging raw sensor data and high-quality, editable CAD models. Its ability to generate precise, native-parameter surfaces from point clouds opens new horizons for reverse engineering, digital twin creation, and intelligent manufacturing. Future research will focus on improving robustness, expanding surface type support, and integrating real-time capabilities, aiming to revolutionize industrial design workflows.
Deep Dive
Abstract
Generative CAD modeling has broad design and application potential. Despite significant advances in Boundary Representation (B-Rep) generation, the dominant representation in CAD, existing methods largely depend on uniformly sampled point- or grid-based geometry representations, sacrificing native surface types and parameters and thereby limiting geometric fidelity and downstream usability. We present ParaCAD, an autoregressive framework for point-cloud-conditioned B-Rep generation that directly operates on native parametric surfaces. ParaCAD introduces a surface-centric tokenization that explicitly encodes each face by its exact surface type and continuous parameters, preserving the intrinsic semantics of CAD geometry. Our model first generates parametric surfaces with constrained UV domains, and then constructs a valid B-Rep by globally intersecting these surfaces to recover edges and vertices. ParaCAD places point-cloud-conditioned generation at the core of B-Rep synthesis, making it practical for user-guided reconstruction and seamless integration into existing 3D generation pipelines. Extensive experiments demonstrate that ParaCAD produces accurate B-Reps with faithful point-cloud alignment, outperforming point-based baselines in geometric precision, robustness, watertightness and downstream usability.