Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
157 tokens/sec
GPT-4o
43 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
4 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

SkexGen: Autoregressive Generation of CAD Construction Sequences with Disentangled Codebooks (2207.04632v1)

Published 11 Jul 2022 in cs.CV and cs.LG

Abstract: We present SkexGen, a novel autoregressive generative model for computer-aided design (CAD) construction sequences containing sketch-and-extrude modeling operations. Our model utilizes distinct Transformer architectures to encode topological, geometric, and extrusion variations of construction sequences into disentangled codebooks. Autoregressive Transformer decoders generate CAD construction sequences sharing certain properties specified by the codebook vectors. Extensive experiments demonstrate that our disentangled codebook representation generates diverse and high-quality CAD models, enhances user control, and enables efficient exploration of the design space. The code is available at https://samxuxiang.github.io/skexgen.

Citations (50)

Summary

  • The paper presents SkexGen, an autoregressive model that generates CAD construction sequences by disentangling design attributes via dedicated codebooks.
  • It employs dual Transformer architectures to separately control topology, geometry, and extrusion variations for nuanced design exploration.
  • Experimental results demonstrate that SkexGen outperforms previous models in quality and diversity, significantly enhancing key CAD metrics.

An Expert Overview of "SkexGen: Autoregressive Generation of CAD Construction Sequences with Disentangled Codebooks"

The research paper presents SkexGen, a sophisticated autoregressive generative model designed to generate computer-aided design (CAD) construction sequences. This model effectively combines sketch-and-extrude modeling operations through the innovative use of disentangled codebooks within distinct Transformer architectures.

Key Contributions and Architecture

SkexGen stands out by employing discrete codebooks to facilitate CAD model generation. It operates within a sketch-and-extrude modeling framework, where 2D sketches are transformed into 3D models through extrusion processes. The architecture utilizes a dual-branch system: one for topology and geometry, and another for extrusion variations. Each of these branches incorporates Transformer encoders and decoders to manage the generative tasks.

The model generates CAD sequences by leveraging the disentangled representations across three codebooks, which separately encode variations in topology, geometry, and extrusions. This separation allows for nuanced user control and efficient exploration of design spaces beyond what previous models could achieve. Unlike traditional parametric CAD models, which can be volatile under significant topological changes, SkexGen's approach maintains robustness while enabling intricate design variations.

Empirical Evaluations

The efficacy of SkexGen was validated through extensive experiments on a substantial dataset of CAD sequences. Comparisons with baseline models, such as DeepCAD, demonstrate that SkexGen consistently produces CAD models of superior quality and diversity. For instance, the paper reports strong performance metrics including notable improvements in Fréchet Inception Distance (FID) and other coverage metrics over state-of-the-art benchmarks.

The paper details how SkexGen excels in generating realistic and intricate designs that reflect human-like design choices, such as symmetry and proper use of arcs and lines. Unlike some competitive models that rely on sketch constraints, SkexGen allows exploration of design variability without needing extensive constraint labels, making it more practical for real-world applications.

Theoretical and Practical Implications

Practically, SkexGen's architecture enhances user interaction by allowing selective control over design properties via its disentangled codebooks. Designers can focus on specific aspects of a design, such as preserving geometric proportions while exploring different topologies, or maintaining extrusion features across different model geometries. This level of control is particularly useful for optimizing designs in fields like mechanical engineering where specific physical or functional characteristics need to be maintained.

Theoretically, SkexGen contributes to the advancement of generative models in the CAD domain by providing a framework that efficiently decouples and encodes different design dimensions. This approach paves the way for future developments where disentangled representations can be applied to even more complex generative tasks in 3D modeling and beyond.

Future Directions

The potential extensions of this work are manifold. Future research could explore integrating SkexGen with interactive design systems, thereby enhancing its utility in user-driven design applications. Additionally, the principles of disentangled codebooks may be extended to other domains where complex generative tasks are common, such as in procedural content generation in digital entertainment or architecture.

In conclusion, SkexGen represents a significant advancement in the field of CAD model generation, particularly in terms of effectively disentangling design attributes for enhanced generative control. Its architecture not only addresses some of the limitations of previous models but also opens new avenues for interactive and user-focused design exploration.

Youtube Logo Streamline Icon: https://streamlinehq.com