Reproduce “Hierarchical Neural Coding for Controllable CAD Model Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiang Xu, Pradeep Kumar Jayaraman, Joseph G. Lambourne, Karl D.D. Willis, Yasutaka Furukawa
Researchers from Simon Fraser University and Autodesk Research developed a hierarchical neural coding framework that uses a three-level code tree (Solid, Profile, Loop) to capture design intent, enabling more realistic, diverse, and controllable CAD model generation with capabilities like design-preserving editing and auto-completion.
Key insights
- Encoding design intent into a multi-level hierarchical code tree effectively enables fine-grained and intuitive control over CAD model generation and editing.
- A VQ-VAE variant with a partially masked skip connection can successfully learn robust, instance-agnostic discrete representations that abstract high-level design characteristics.
- A cascaded auto-regressive architecture, by conditioning on both partial model input and a generated hierarchical code tree, facilitates intelligent design-preserving editing and auto-completion capabilities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.