DrawingsDreamer: A Unified Multi-View Engineering Drawings Generation Model

Liu, Shurui; Chen, Weide; Yi, Changwang; Wu, Ancong

NeurIPS 2026 · Creative AI

DrawingsDreamer: A Unified Multi-ViewEngineering Drawings Generation Model

Shurui Liu¹, Weide Chen², Changwang Yi¹, Ancong Wu¹

¹School of Computer Science and Engineering, Sun Yat-sen University

²School of Intelligent Systems Engineering, Sun Yat-sen University

NeurIPS 2026

GitHub arXiv

DrawingsDreamer multi-view generation overview — **DrawingsDreamer** handles a full spectrum of multi-view modeling tasks. Orange-to-pink gradients denote the outputs, while blue gradients represent the inputs across top, front, right, and isometric projections.

Abstract

Scalable Vector Graphics (SVG) are essential for modern industrial Computer-Aided Design (CAD). However, existing autoregressive SVG generation models are predominantly tailored for artistic creation and struggle to maintain the rigorous geometric fidelity and cross-view spatial alignment required for engineering drawings.

To bridge this gap, we introduce DrawingsDreamer, a unified Large Language Model (LLM)-driven framework for multi-view vector-based engineering drawings generation. By formulating generation purely as sequence modeling, we eliminate the need for raster image encoders. A streamlined representation with hierarchical postfix tokenization guides the model to establish local geometric coordinates before assigning semantic boundaries. A progressive task-aware curriculum lets one model transition from localized structural repair to macroscopic generation.

4canonical views

1unified model

5task configurations

SVGstructured output

From geometry to aligned view sequences

Streamlined Representation

We project each CAD object into a canonical set of front, right, top, and isometric views. Hierarchical postfix tokenization makes the model regress coordinates first and close each primitive with an explicit semantic boundary.

Global coordinate normalization preserves alignment across orthographic views while keeping the vocabulary compact and LLM-friendly.

Progressive task-aware curriculum

The same autoregressive model is trained across microscopic repair and macroscopic generation. Sampling moves from local topology learning to partial view alignment and finally to unconstrained multi-view spatial translation.

DrawingsDreamer training pipeline — Training schedule and unified sequence modeling pipeline.

01 · Generate

Unconditional generation

Sample all four aligned views directly from the learned drawing prior.

02 · Translate

Cross-view translation

Translate between isometric and orthographic projections while retaining geometry.

03 · Complete

Partial view completion

Complete missing orthographic views from an isometric view and a partial context.

04 · Repair

Path-level restoration

Restore masked paths with local topological and syntactic validity.

05 · Structure

Postfix boundaries

Use explicit end tokens to organize primitives, paths, and views into one sequence.

Qualitative results

The model produces structured drawings across conditional translation, completion, and unconditional sampling tasks.

Isometric to orthographic translation — **Isometric → orthographic.** Translation from an isometric context to the three aligned orthographic views.

Orthographic to isometric translation — **Orthographic → isometric.** Reconstructing a consistent isometric view from orthographic inputs.

**Partial view completion.** Completing missing top and right views from an isometric and front-view context.

Unconditional multi-view generation — **Unconditional generation.** Diverse, structurally valid multi-view engineering drawings sampled without context.

Citation

If DrawingsDreamer helps your research, please cite the work:

@inproceedings{liu2026drawingsdreamer,
  title={DrawingsDreamer: A Unified Multi-View Engineering Drawings Generation Model},
  author={Liu, Shurui and Chen, Weide and Yi, Changwang and Wu, Ancong},
  booktitle={Advances in Neural Information Processing Systems},
  year={2026},
  url={https://drawingsdreamer.github.io/}
}