Contribution
Text2CAD is the first AI framework for generating parametric CAD designs from multi-level textual descriptions, supporting prompts from abstract shape descriptions to detailed parametric instructions.
Key Contributions
-
01
Novel Data Annotation Pipeline — Leverages open-source LLMs and VLMs to annotate the DeepCAD dataset with multi-level text prompts of varying complexity and parametric detail.
-
02
Text2CAD Transformer — An end-to-end Transformer-based autoregressive architecture for generating complete CAD design history from natural language prompts.
Data Annotation
Our pipeline generates multi-level text prompts describing CAD construction workflow with varying complexities via a two-stage approach:
- Stage 1: Shape description generation using VLM (LlaVA-NeXT).
- Stage 2: Multi-level textual annotation generation using LLM (Mixtral-50B).
Text2CAD Transformer
The Text2CAD Transformer converts natural language into parametric 3D CAD models by deducing intermediate design steps autoregressively. Given text prompt \(T\) and CAD subsequence \(\mathbf{C}_{1:t-1}\), a pretrained BERT Encoder with trainable Adaptive layer extracts \(T_{adapt}\), which passes through \(\mathbf{L}\) decoder blocks alongside CAD sequence embedding \(F^0_{t-1}\).
