SolidGen: An Autoregressive Model for Direct B-rep Synthesis

Abstract

The Boundary representation (B-rep) format is the de-facto shape representation in computer-aided design (CAD) to model watertight solid objects. Recent approaches to generating CAD models have focused on learning sketch-and-extrude modeling sequences that are executed by a solid modeling kernel in postprocess to recover a B-rep. In this paper we present a new approach that enables learning from and synthesizing B-reps without the need for supervision through CAD modeling sequence data. Our method SolidGen, is an autoregressive neural network that models the B-rep directly by predicting the vertices, edges and faces using Transformer-based and pointer neural networks. Key to achieving this is our Indexed Boundary Representation that references B-rep vertices, edges and faces in a well-defined hierarchy to capture the geometric and topological relations suitable for use with machine learning. SolidGen can be easily conditioned on contexts e.g., class labels thanks to its probabilistic modeling of the B-rep distribution. We demonstrate qualitatively, quantitatively and through perceptual evaluation by human subjects that SolidGen can produce high quality, realistic looking CAD models.

CLIP-Forge: Towards Zero-Shot Text-to-Shape Generation

“…3D Generative Models. Significant progress has been made in the field of generative models for the creation of 3D shapes in various formats such as voxels [70,9,66,30], CAD [71,29,36,72], implicit representations [42,7,54], meshes [48,19], and point clouds [49,3,75,37,74,34].…”

Learning to Generate 3D Shapes from a Single Example

“…Since the introduction of deep generative networks such as Generative Adversarial Network (GAN) [Goodfellow et al 2014] and Variational Autoencoder (VAE) [Kingma and Welling 2014], developing deep generative models for 3D shapes has attracted immense research interest. Existing works learn to produce 3D shapes in different representations, including voxels Wu et al 2016], point clouds [Achlioptas et al 2018;Cai et al 2020;Li et al 2021;Yang et al 2019], meshes [Gao et al 2021;Hertz et al 2020;Liu et al 2020;Nash et al 2020], implicit functions [Chen and Zhang 2019;Mescheder et al 2019;Park et al 2019], multi-charts [Ben-Hamu et al 2018Groueix et al 2018],structural primitives [Jones et al 2020;Li et al 2017;Mo et al 2019;Wu et al 2020], and parametric models Jayaraman et al 2022;]. Nearly all these methods are trained on a large dataset of category-specific 3D shapes, e.g., ShapeNet [Chang et al 2015].…”

“…It is such a laborious process that motivates the development of computer algorithms that create shapes-new, diverse, and highquality 3D shapes created in a fully automatic fashion. Leveraging recent advance in deep learning, research in this direction has been vibrant [Achlioptas et al 2018;Chen and Zhang 2019;Jayaraman et al 2022;Nash et al 2020;Wu et al 2016]: the general theme here Authors' addresses: Rundi Wu, Columbia University, New York City, NY, 10025, USA, rundi.wu@columbia.edu; Changxi Zheng, Columbia University, New York City, NY, 10025, USA, cxz@cs.columbia.edu. is to develop a generative model able to learn from a training dataset to generate 3D shapes.…”