Hierarchical Neural Coding for Controllable CAD Model Generation

TL;DR

The paper introduces a three-level hierarchical neural coding framework for controllable CAD model generation, encoding Loop, Profile, and Solid information into discrete codebooks and utilizing two-stage cascaded transformers to generate or complete CAD models from partial input or a target code tree. By learning design concepts as a code tree and enabling editing, autocompletion, and design-preserving changes, the approach aims to preserve design intent while offering richer control than prior methods. Empirical results on a large sketch-and-extrude CAD dataset show superior generation quality and interactive capabilities, including code-tree editing and partial-input auto-completion, with favorable human judgments. This work advances intelligent, intent-aware CAD tooling by integrating hierarchical neural representations with autoregressive generation.

Abstract

This paper presents a novel generative model for Computer Aided Design (CAD) that 1) represents high-level design concepts of a CAD model as a three-level hierarchical tree of neural codes, from global part arrangement down to local curve geometry; and 2) controls the generation or completion of CAD models by specifying the target design using a code tree. Concretely, a novel variant of a vector quantized VAE with "masked skip connection" extracts design variations as neural codebooks at three levels. Two-stage cascaded auto-regressive transformers learn to generate code trees from incomplete CAD models and then complete CAD models following the intended design. Extensive experiments demonstrate superior performance on conventional tasks such as random generation while enabling novel interaction capabilities on conditional generation tasks. The code is available at https://github.com/samxuxiang/hnc-cad.

Figures (20)

• Figure 1: We propose three-level hierarchical neural coding for controllable CAD model generation. Our system learns high-level design concepts as discrete codes at different levels, enabling more diverse and higher-quality generation (top); novel user controls while specifying design intent (bottom-left); and autocompleting a partial CAD model under construction (bottom-right).
• Figure 2: Example failures of parametric CAD, editing a design (a) by shortening or extending (green) the table. Inconsistent areas are highlighted in red.
• Figure 3: Our hierarchical tree representation of a CAD model, with which a novel VQ-VAE learns codebooks at the levels of solid, profile, and loop.
• Figure 4: Left: VQ-VAE with masked skip connection for codebook learning. Given a CAD model as a construction sequence (e.g., x, y, S), an MLP and a Transformer encoder convert the input to latent codes ($T^E_t$), and a vector quantization extracts a code ($\mathbf{c}$) after average pooling. A Transformer decoder recovers the input sequence, conditioned on the vector-quantized code ($\mathbf{c}$) and the masked input sequence ($T^D_t$). Grey color represents input tokens that were masked out. Right: Controllable CAD generation module with two-stage auto-regressive generators. Given a partial CAD model, a model encoder converts it to latent embeddings ($T^E_t$). The first auto-regressive Transformer generates hierarchical neural codes ($T^C_t$) conditioned on the encoded embeddings. The second auto-regressive Transformer generates a new CAD model.
• Figure 5: Unconditional generation results by (a) DeepCAD, (b) SkexGen and (c) our method. The bottom three rows (red color) show complex samples with three or more sketch-extrude steps.