CLR-Wire: Towards Continuous Latent Representations for 3D Curve Wireframe Generation

TL;DR

CLR-Wire addresses the challenge of generating 3D curve wireframes by encoding geometry and topology into a unified Continuous Latent Representation using Curve VAE and Wireframe VAE, followed by Latent Flow Matching to produce diverse wireframes from noise. A fixed-length latent Z_W combines curve latents Z_Curve with topology (Adj_V, V_Coords) and is decoded into variable-length wireframes, enabling both unconditional and conditional generation on point clouds or images via cross-attention and differential adjacency. The approach introduces a differential adjacency encoding and leverages an ODE-based flow to map Gaussian noise to the target latent distribution, achieving superior performance on metrics like COV, MMD, and 1-NN while enabling smooth latent-space interpolation. The method holds practical significance for CAD design, geometric reconstruction, and 3D content creation, providing a robust framework for complex, topology-rich wireframes and flexible conditioning scenarios. The key contributions include a fixed-length, continuous representation for geometry and topology, a three-stage learning and generation pipeline, and demonstrated improvements in diversity and accuracy with meaningful topological control.

Abstract

We introduce CLR-Wire, a novel framework for 3D curve-based wireframe generation that integrates geometry and topology into a unified Continuous Latent Representation. Unlike conventional methods that decouple vertices, edges, and faces, CLR-Wire encodes curves as Neural Parametric Curves along with their topological connectivity into a continuous and fixed-length latent space using an attention-driven variational autoencoder (VAE). This unified approach facilitates joint learning and generation of both geometry and topology. To generate wireframes, we employ a flow matching model to progressively map Gaussian noise to these latents, which are subsequently decoded into complete 3D wireframes. Our method provides fine-grained modeling of complex shapes and irregular topologies, and supports both unconditional generation and generation conditioned on point cloud or image inputs. Experimental results demonstrate that, compared with state-of-the-art generative approaches, our method achieves substantial improvements in accuracy, novelty, and diversity, offering an efficient and comprehensive solution for CAD design, geometric reconstruction, and 3D content creation.

Figures (28)

• Figure 1: Overview of wireframe generation. Our method employs Latent Flow Matching based on the proposed latent wireframe representation $Z_W$, which is decoded into a wireframe via the Wireframe Decoder. First, random noise is mapped to $Z_W$, and then decoded into adjacency $\text{Adj}_{V}$, endpoint coordinates $V_{\text{Coords}}$, and curve latents $Z_{\text{Curve}}$. The final wireframe is produced using $\text{Adj}_{V}$, $V_{\text{Coords}}$, and 3D curves decoded from $Z_{\text{Curve}}$. Furthermore, our method supports conditional inputs (e.g., sparse point clouds, partial point clouds, single-view images, or sketches), offering controllable and flexible 3D generation.
• Figure 2: 2D Curve normalization examples. Curves are normalized successively through translation, rotation, and scaling to align their start and end points, ensuring consistent spatial representation.
• Figure 3: Wireframe VAE pipeline. Given a 3D curve wireframe, we extract curve latents $Z_\text{Curve}$ from normalized curves via Curve Encoder, and the topology embeddings, which include adjacency $\text{Adj}_{V}$ embeddings and corresponding vertex coordinate $V_{\text{Coords}}$ embeddings. They are concatenated and transformed into a unified latent space $Z_W$ via Transformer blocks. In decoding phase, the $Z_W$ are decoded back to $\text{Adj}_{V}$, $V_{\text{Coords}}$, $Z_\text{Curve}$ and valid flags, which are used to compute losses against ground truths.
• Figure 4: Qualitative evaluation of our method with state-of-the-art methods, BrepGenbrepgen, DeepCADdeepcad, and 3DWire3dwire. Our method generates wireframes with richer curve details, maintaining geometric integrity and topological correctness, especially for complex shapes.
• Figure 5: Novelty analysis. We present the Chamfer Distance (CD) distribution for 4k randomly generated wireframes compared to their most similar samples in the training set. Visualizations at various CD percentile ranges highlight both close resemblance (low CD) and novelty (high CD), showing that our method produces realistic and diverse shapes.