CAD Translator: An Effective Drive for Text to 3D Parametric Computer-Aided Design Generative Modeling
TLDR
A novel encoder-decoder framework for text-to-parametric CAD generation using contrastive learning and fusion embeddings.
Reasoning
The paper addresses an underexplored task with a clear methodology (Cascading Contrastive Strategy and CT-Mix) and a new dataset. However, the abstract lacks quantitative results and details on dataset size, limiting assessment of empirical strength.
Read-first score
Read-first score 46.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 72.
Field roles
Rank sensitivity
Stability: volatile; rank range: 62.
Keyword Scores
Deep Analysis
Innovations
- CAD Translator: an encoder-decoder framework for text-to-parametric CAD generative modeling with one-stage training
- Cascading Contrastive Strategy to align text and parametric CAD sequences in latent space
- CT-Mix: random mask operation on embeddings and linear interpolation to fuse text and CAD sequence embeddings
- Text2CAD dataset built using a Large Multimodal Model (LMM)
Methodology
The authors propose an encoder-decoder framework that first aligns text and parametric CAD command sequences via a cascading contrastive strategy in latent space, then applies CT-Mix, which randomly masks embeddings and fuses them via linear interpolation to strengthen cross-modal connections. The model is trained on a newly constructed Text2CAD dataset generated with a Large Multimodal Model.
Key Results
Thorough experiments on the Text2CAD dataset demonstrate the effectiveness of the proposed CAD Translator method.