💻 computer science

Autoregressive B-Rep Shape Generation with Parametric Surfaces

This paper introduces ParaCAD, an autoregressive framework that generates high-fidelity, watertight Boundary Representation (B-Rep) models directly from point clouds by tokenizing and synthesizing native parametric surfaces, thereby overcoming the geometric fidelity and usability limitations of existing point- or grid-based methods.

Original authors: Dafei Qin, Rui Xu, Zeyu Shen, Kaichun Qiao, Hongyang Lin, Qixuan Zhang, Huaijin Pi, Lan Xu, Jingyi Yu, Wenping Wang, Taku Komura

Published 2026-07-21

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Dafei Qin, Rui Xu, Zeyu Shen, Kaichun Qiao, Hongyang Lin, Qixuan Zhang, Huaijin Pi, Lan Xu, Jingyi Yu, Wenping Wang, Taku Komura

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to rebuild a complex Lego castle, but you only have a blurry, scattered pile of dust that used to be the castle. Your goal is to turn that messy dust back into a perfect, sturdy structure where every brick fits exactly. This is the challenge of "generative design" in computer science, a field where AI tries to create new 3D objects from scratch. Usually, when computers try to build these shapes, they use a method similar to pixel art: they approximate curves with tiny, flat squares or dots. It's like trying to draw a perfect circle using only square pixels; it looks okay from far away, but up close, the edges are jagged and the math is messy. This makes it hard for engineers to use the shapes for real-world tasks like manufacturing or precise engineering, because the computer's "circle" isn't a true circle—it's just a very good guess.

The paper you are about to read tackles this problem by teaching the AI to stop guessing and start knowing. Instead of building shapes out of tiny, approximate blocks, the researchers want the AI to speak the native language of 3D design: exact mathematical formulas. They are asking the computer to say, "This part is a perfect cylinder with a radius of 5," rather than "This part looks like a cylinder made of 1,000 tiny flat squares." This matters because when you need to build a real car part or a medical device, you need the math to be exact, not approximate. If the computer's shape is slightly off, the real-world object might not fit, or the factory machine might break.

The Magic Recipe: ParaCAD

Enter ParaCAD, a new AI framework that acts like a master architect who can look at a pile of dust (a point cloud) and instantly write down the exact blueprints for a perfect building. The researchers, a team from universities and tech companies in China and the US, developed this system to generate CAD (Computer-Aided Design) models that are not just visual approximations, but mathematically precise 3D solids.

Here is how they did it, broken down into three simple steps:

1. The "Native Language" Translator
Most AI models try to describe a 3D shape by breaking it into a grid of tiny dots or a cloud of points. It's like describing a song by listing the volume of every single air molecule. ParaCAD does something different. It uses a special "tokenizer" (a translator) that recognizes the shape's true identity. If the AI sees a curved surface, it doesn't just say "curvy blob"; it identifies it as a "Cylinder" or a "Sphere" and writes down the exact numbers (parameters) that define it. For the weird, free-form shapes that don't fit simple categories, it uses a clever compression trick to describe them precisely without losing detail. This ensures the computer understands the type of surface, not just its look.

2. The Storyteller AI
Once the AI can "speak" the language of exact shapes, it uses a technique called "autoregressive generation." Think of this like a very smart storyteller who writes a story one word at a time. The AI looks at the input dust (the point cloud), and then it predicts the next "word" (the next surface of the object) based on what it has already written. It keeps adding surfaces—planes, cylinders, cones—until the whole object is described. Because it is building the object piece by piece using exact math, it doesn't have to worry about the jagged edges that plague other methods.

3. The "Seamless Stitcher"
Here is the tricky part: just because the AI wrote down a list of perfect cylinders and planes doesn't mean they automatically fit together to make a solid, watertight object. Imagine you have a list of perfect pizza slices; you still need to arrange them so they form a whole pizza without gaps or overlaps. ParaCAD has a special "post-processing" step that acts like a master seamstress. It takes all the generated surfaces, cuts them where they need to meet, and sews them together using exact geometric rules. It checks the point cloud (the dust) to make sure the final shape matches the original input, discarding any pieces that don't fit. This ensures the final result is a solid, usable 3D model with no holes.

What They Found

The researchers tested ParaCAD against other top methods that try to turn point clouds into 3D models. The results were impressive. While other methods often produced shapes that looked okay but had "jagged" math or missing parts, ParaCAD produced models that were:

More Accurate: The shapes matched the original point clouds much better.
More "True": The AI correctly identified the types of surfaces (like knowing a part is a cone, not just a cone-shaped mess).
More Usable: The final models were "watertight" (no holes) and ready to be used in real engineering software.

In their tests, ParaCAD achieved a score of 0.784 on a metric called F1 (which measures how well the shape matches the ground truth), beating the next best method which scored 0.502. When it came to recognizing specific shapes like cylinders and spheres, ParaCAD was significantly better at getting the type right, with scores like 0.65 for cylinders compared to 0.39 for the competition.

Why This Matters

The paper suggests that this approach is a big step forward because it bridges the gap between "cool AI art" and "useful engineering." By keeping the shapes in their native mathematical form, the models generated by ParaCAD can be directly used for things like 3D printing, manufacturing, and simulation. The authors showed that you could even take a rough, noisy scan of an object (like a dusty toy), turn it into a perfect CAD model, and then use that model to perform real engineering tasks, like adding rounded edges or cutting holes, just as if a human designer had drawn it from scratch.

However, the authors are careful to note that this isn't a magic wand that solves everything. They admit that for extremely complex shapes, the system still sometimes struggles to get the pieces to fit perfectly, and they suggest that future work could make the system even better at handling these tricky cases. But for now, ParaCAD shows that teaching AI to speak the "exact math" of 3D design is a powerful way to create shapes that are not just pretty, but truly useful.

Technical Summary: Autoregressive B-Rep Shape Generation with Parametric Surfaces (ParaCAD)

Problem Statement

Computer-Aided Design (CAD) relies on Boundary Representation (B-Rep) to define solids through exact analytical formulations of vertices, curves, and surfaces. Preserving native surface types (e.g., planes, cylinders) and their parametric definitions is critical for downstream tasks such as precise editing, simulation, and manufacturing. However, existing deep learning methods for B-Rep generation largely depend on uniformly sampled point- or grid-based geometry representations (e.g., latent grids). These approaches suffer from compression errors that blur surface-type semantics and fail to faithfully preserve analytic primitives. Consequently, generated models often lack geometric fidelity, exhibit small holes due to amplified intersection errors, and are difficult to interpret or edit. Furthermore, many methods avoid primitive recovery, representing shapes as freeform geometry, which undermines the structural advantages of CAD.

Methodology

The authors propose ParaCAD, an autoregressive framework designed to generate B-Rep models directly from point clouds while operating natively on parametric surfaces. The pipeline consists of three core components:

1. Surface-Centric Tokenization

ParaCAD introduces a hybrid tokenizer that converts B-Rep faces into discrete token blocks, explicitly encoding surface types and parameters rather than approximating them via latent vectors.

Analytic Primitives: For common surfaces (plane, cylinder, cone, sphere, torus), the tokenizer uses a fixed-length block of 13 tokens.
- $s_0$ : Surface type token.
- $s_1$ – $s_4$ : Discretized UV domain bounds.
- $s_5$ : Type-specific intrinsic scalar (e.g., cone semi-angle, torus minor radius).
- $s_6$ – $s_{12}$ : Instance-level SIM(3) transform parameters (translation, rotation, isotropic scale) mapping the canonical surface to its 3D placement.
Freeform Surfaces: For general freeform patches, the method employs Finite Scalar Quantization (FSQ). A 32×32 grid is sampled, compressed via a convolutional autoencoder into a 4×4 latent grid, and quantized into 12 tokens ( $s_1$ – $s_{12}$ ) to fill the face block.
Ordering: Faces are sorted lexicographically by their placement coordinates to create a deterministic sequence for autoregressive modeling.

2. Autoregressive Generation

The model utilizes a decoder-only GPT architecture (15 Transformer blocks) conditioned on input point clouds.

Conditioning: Point clouds are encoded into 257 tokens via a pre-trained encoder and injected into the Transformer via cross-attention.
Training: The model is trained to minimize negative log-likelihood (cross-entropy loss) over the token sequence, predicting one surface token block at a time given the point cloud and previously generated tokens.
Inference: Given a point cloud, the model samples tokens sequentially until an end token is reached, which are then decoded back into parametric surfaces.

3. Geometric Post-Processing

Since the generative model outputs independent surface patches without explicit topological connectivity, a deterministic geometric solver assembles the final B-Rep.

Intersection and Splitting: UV domains are extended to ensure overlap. Surface-surface intersections are computed exactly (leveraging the parametric nature of the output) to split patches into smaller pieces.
Selection: A two-stage selection process identifies the optimal subset of pieces:
1. Point-Cloud Alignment: Pieces are scored based on distance and normal agreement with the input point cloud. Low-scoring pieces are discarded.
2. Topological Optimization: A binary integer programming problem maximizes surface coverage while enforcing 2-manifold constraints (ensuring edges are shared by exactly two faces) and penalizing open boundaries.
Merging: Selected pieces are merged based on their original affiliation to form a closed, watertight B-Rep solid.

Key Contributions

Surface-Centric Tokenizer: A novel discretization scheme that represents B-Rep faces as discrete tokens while preserving native surface types and continuous parameters. It explicitly encodes analytic primitives and supports freeform patches via compatible discretization.
Point-Cloud-Conditioned Autoregressive Generator: A model that predicts parametric surfaces with bounded UV domains. By removing explicit topological entities (edges/vertices) from the learning space, the task is simplified to predicting geometrically meaningful surfaces.
Exact Geometric Solver: A robust assembly pipeline that leverages exact surface-surface intersections and binary integer programming to construct valid, watertight B-Reps from generated parametric surfaces, maximizing geometric coverage while ensuring topological validity.

Results

The authors evaluated ParaCAD on a large-scale dataset of 583K B-Rep instances derived from the ABC dataset, conditioned on point clouds.

Geometric and Semantic Accuracy: ParaCAD outperformed baselines (including P2CAD, HoLa, and NVDNet) in Chamfer Distance (CD) and semantic fidelity metrics (Surface Precision, Recall, and F1). It achieved an F1 score of 0.784 compared to 0.502 for the next best baseline (NVDNet).
Surface-Type Consistency: The method demonstrated superior performance in recovering specific analytic types (e.g., F1 of 0.82 for planes, 0.65 for cylinders) compared to methods that approximate all surfaces as freeform or struggle with specific primitives.
Validity: The post-processing pipeline achieved a 55.81% validity rate for generated models, significantly higher than the 47.15% validity of the HoLa baseline. The pipeline was shown to reliably recover topologically valid solids from both raw ground truth and generated patches.
Applications: The generated B-Reps were demonstrated to support downstream CAD operations such as edge filleting, boolean intersections, and surface offsets, and were successfully used as drafts for metal casting simulations.

Significance

The paper claims that ParaCAD represents a shift from approximation-based generation to a "representation-first" approach that operates directly in the native parametric surface space. By preserving exact geometric primitives and parameters, the method bridges the gap between 3D generation and industrial design requirements. It enables the creation of CAD models that are not only geometrically accurate but also semantically meaningful and directly usable in downstream manufacturing and simulation workflows, addressing the limitations of grid-based and latent-vector approaches that sacrifice fidelity for network compatibility.