NTH

CAD-WAM: Anticipating Geometry for Text-to-CAD Generation

AuthorsZhengyang Li, Weijie Zheng, Yuxin Cui, Sijie Wang, Wenjie Feng, Qiyu Kang

AffiliationsUniversity of Science and Technology of China · Nanyang Technological University

November 9, 2026 2 min read
Watch on YouTube
The one-line take

CAD-WAM turns text into editable CAD by predicting both the next modeling action and the geometry that action should create.

Key results

526,010
Step-wise training records

Records built from DeepCAD trajectories, pairing actions with geometric-state supervision.

1.142
Mean Chamfer distance

CAD-WAM test-set score; GPT-6 Astra scores 1.755.

13.2%
Chamfer-distance reduction

Relative reduction from adding next-state supervision versus action-only training.

0.666%
CAD-WAM invalidity rate

Share of attempted CAD-WAM outputs that are invalid.

0.025%
GPT-6 Astra invalidity rate

Lower invalidity rate than CAD-WAM in the evaluated comparison.

What the paper found

CAD-WAM treats text-to-CAD generation as a sequence of geometry-aware decisions, rather than simply predicting a list of commands. Its CAD-specific language packages a sketch, extrusion, and Boolean operation into one executable action. At each step, a pretrained SetVAE encodes the current solid into a 1,024-dimensional latent state; a shared Qwen2.5-1.5B-Instruct backbone uses that state and the text description to generate the next action, while an auxiliary head predicts the resulting geometry. During inference, a CAD kernel executes each action and the actual solid is re-encoded to guide the next decision. Training uses 526,010 step-wise records constructed from DeepCAD. On the DeepCAD test set, CAD-WAM achieves a mean Chamfer distance of 1.142, compared with 1.755 for OpenAI’s GPT-6 Astra, the strongest general-purpose model on that metric. Adding next-state supervision reduces mean Chamfer distance by 13.2% versus training on actions alone. CAD-WAM’s invalidity rate is 0.666%; GPT-6 Astra has a lower rate, at 0.025%. The results highlight the benefit of teaching a model to anticipate the geometric consequences of its actions, especially for designs requiring more operations.

Original abstract

Generating editable computer-aided design (CAD) models from text requires selecting operations that produce the intended changes to the geometry constructed so far. An incorrect operation can compromise subsequent steps, making reliable multi-step generation challenging. Existing methods based on large language models (LLMs) primarily learn to predict commands, without explicit supervision of the geometric states those commands should produce. Inspired by world-action models, we introduce CAD-WAM, which couples geometry-conditioned action generation with next-state prediction. We define complete modeling operations as actions in a CAD domain-specific language (DSL) and encode surface point clouds of intermediate solids into compact latent states using a pretrained variational autoencoder. An LLM backbone generates the next action from the target description and current geometric state, while an auxiliary head on the same backbone predicts the next state to provide geometric supervision during training. We construct step-wise training trajectories from parametric CAD models in the DeepCAD dataset, pairing each trajectory with an expert-level textual description. At inference, each generated action is executed by a CAD kernel, and the resulting solid is re-encoded to inform the next decision. On the DeepCAD test set, CAD-WAM outperforms the task-specific baselines in geometric accuracy and output validity, with larger gains on models requiring more operations, and produces more accurate geometry than general-purpose LLMs prompted with the same DSL. Ablations further show that next-state supervision reduces mean Chamfer distance by 13.2\% relative to an otherwise identical model trained only to predict actions. Code is available at https://github.com/Lzy-1037/CAD-WAM.

Read the original paper

More in Generative Models

Browse all 87 papers →
01Generative Model

dNaViT: Discrete Codes are General-Purpose Visual Representations

Xiaotong Li, Zhengrong Yue, Xu Huang, Jialv Zou, Minhao Jing, Wenhao Yan, Haodong Li, Hai Huang, Bo Zhao, Yan Feng, Peng Pei, Xunliang Cai

dNaViT turns images into reusable discrete codes designed to work across visual understanding, generation, and editing.

Read analysis