DrawingsDreamer: A Unified Multi-View Engineering Drawings Generation Model

Shurui Liu; Weide Chen; Changwang Yi; Ancong Wu

cs.CVSep 28, 2026

DrawingsDreamer: A Unified Multi-View Engineering Drawings Generation Model

Authors: Shurui Liu, Weide Chen, Changwang Yi, Ancong Wu

Organizations: School of Computer Science and Engineering, Sun Yat-sen University, China · School of Intelligent Systems Engineering, Shenzhen Campus of Sun Yat-sen University, China

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 \textbf{DrawingsDreamer}, a unified Large Language Model (LLM)-driven framework for multi-view vector-based engineering drawings generation. By formulating the generation of multi-view engineering drawings purely as a sequence modeling task, we eliminate the need of raster image encoders. We propose a Streamlined Representation utilizing hierarchical postfix tokenization, which guides the model to establish local geometric coordinates before assigning semantic boundaries. Optimized via a progressive task-aware curriculum schedule, \textbf{DrawingsDreamer} effectively transitions from localized structural repair to macroscopic generation in a unified model. Extensive experiments demonstrate that our unified model achieves strong performance in both geometric fidelity and syntactic accuracy across diverse conditional and unconditional generation tasks.

Figures & tables

Figure 1: DrawingsDreamer handles a full spectrum of multi-view modeling tasks. The o r a n g e g r a d i e nts denote the outputs, while the b l u e g r a d i e n t s represent the inputs. In each example, the multi-view layout arranges the Top, Front, and Right orthographic views at the top-left, bottom-left, and bottom-right, respectively, with the Isometric view situated at the top-right.

Figure 2: Visualization of the Streamlined Representation, detailing its hierarchical data organization. Continuous coordinates are grouped into primitive boundaries (e.g., lines and curves), formed into closed loops, and ultimately structured into unified view-level sequences.

Hierarchy Level	Token Format	Geometric Semantics
Primitive Boundaries	$x,y$ <end_move>	Move ( M ): Initializes a new subpath at coordinate $(x,y)$ .
	$x,y$ <end_line>	Line ( L ): Draws a straight segment from the current position to $(x,y)$ .
	$x_{1},y_{1},x_{2},y_{2},x,y$ <end_curve>	Cubic Bézier ( C ): Draws a curve ending at $(x,y)$ using control points $(x_{1},y_{1})$ and $(x_{2},y_{2})$ .
Macro Boundaries	<end_path>	Path Boundary: Terminates a continuous geometric path $\mathcal{P}$ .
	<end_view>	View Boundary: Terminates a spatial projection view sequence $S_{v}$ .
	<mask_path>	Mask Placeholder: Represents a masked path $\mathcal{P}_{\text{mask}}\in S_{\text{context}}$ during training.

Table 1: Vocabulary of structural boundary tokens and their geometric semantics.

Task Configuration	Input Context ( $S_{\text{context}}$ )	Target Response ( $S_{\text{target}}$ )
Unconditional Generation	$\emptyset$	$\mathcal{V}$
Isometric to Orthographic	$\{v_{\text{iso}}\}$	$\mathcal{V}_{\text{ortho}}$
Orthographic to Isometric	$\mathcal{V}_{\text{ortho}}$	$\{v_{\text{iso}}\}$
Partial View Completion	$\{v_{\text{iso}},v_{i}\}\mid v_{i}\in\mathcal{V}_{\text{ortho}}$	$\mathcal{V}_{\text{ortho}}\setminus\{v_{i}\}$
Path Level Semantic Masking	$S_{\text{available}}\setminus\{\mathcal{P}_{\text{mask}}\}$	$\{\mathcal{P}_{\text{mask}}\}$

Table 2: Summary of unified multitask configurations.

\mathcal{V}_{\text{ortho}}

denotes the set of orthographic views, and

v_{i}

represents a single orthographic view.

Figure 3: Overview of the proposed training framework. We cast diverse multi-view generation sub-tasks into a unified sequence modeling paradigm, guided by a progressive microscopic-to-macroscopic optimization schedule. The task sampling distribution dynamically transitions from microscopic local topology learning (Stage I) to partial view alignment (Stage II), and ultimately to macroscopic global spatial translation (Stage III) across training epochs.

Figure 4: Qualitative results on Isometric to Orthographic translation ( $\{v_{\text{iso}}\}\rightarrow\mathcal{V}_{\text{ortho}}$ ).

Method	JSD $\downarrow$	COV $\uparrow$	MMD $\downarrow$
GPT-4o (5-shot)	1.17	33.12%	8.35
DrawingsDreamer	0.61	66.24%	5.64

Table 3: Quantitative comparison on Unconditional Generation (

\emptyset\rightarrow\mathcal{V}

). JSD and MMD are multipled by

10^{2}

. Red and green cells denote the best and second-best results, respectively.

Method	$\{v_{\text{iso}}\}\rightarrow\mathcal{V}_{\text{ortho}}$						$\mathcal{V}_{\text{ortho}}\rightarrow\{v_{\text{iso}}\}$
Method	$Acc_{cmd}\uparrow$	$Acc_{params}\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	CD $\downarrow$	$Acc_{cmd}\uparrow$	$Acc_{params}\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	CD $\downarrow$
GPT-4o(5shot)	46.44%	35.54%	33.53	0.80	0.19	0.34	11.33%	56.68%	11.04	0.72	0.23	0.23
Zero123	–	–	24.93	0.85	0.16	0.75	–	–	–	–	–	–
OmniSVG	–	–	20.01	0.88	0.10	0.47	–	–	–	–	–	–
DrawingsDreamer	89.77%	80.82%	50.16	0.94	0.03	0.02	73.53%	68.36%	32.03	0.88	0.06	0.02

Table 4: Quantitative comparison on Cross-View Translation tasks. GPT-4o is enhanced with few-shot in-context learning. Specifically, each prompt comprises five exemplars randomly chosen from the training set. These exemplars include instructions and answer. Red and green cells denote the best and second-best results, respectively.

Method	$\{v_{\text{iso}}\}\rightarrow\mathcal{V}_{\text{ortho}}$						$\{v_{\text{ortho}}\}\rightarrow\mathcal{V}_{\text{iso}}$
Method	$Acc_{cmd}\uparrow$	$Acc_{params}\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	CD $\downarrow$	$Acc_{cmd}{\uparrow}$	$Acc_{params}\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	CD $\downarrow$
w/o Path Mask	86.35%	75.64%	48.99	0.91	3.53	3.17	73.18%	62.89%	31.77	0.84	6.51	2.85
Prefix Format	88.11%	80.58%	47.12	0.87	4.01	2.83	67.11%	67.28%	30.60	0.86	6.67	2.74
Random Schedule	87.49%	73.42%	40.91	0.92	4.52	3.03	70.91%	68.23%	31.64	0.83	6.43	2.59
Isolated: $\{v_{\text{iso}}\}\rightarrow\mathcal{V}_{\text{ortho}}$	88.71%	79.22%	47.42	0.94	3.44	2.36	–	–	–	–	–	–
Isolated: $\{v_{\text{ortho}}\}\rightarrow\mathcal{V}_{\text{iso}}$	–	–	–	–	–	–	75.58%	66.85%	33.68	0.89	5.58	2.45

Table 5: Ablation Study and Quantitative Results. The evaluation is grouped by task: Cross-View Translation (top) and Partial View Completion alongside Unconditional Generation (bottom). LPIPS, CD, JSD, and MMD are multiplied by

10^{2}

. Red and green cells denote the best and second-best results, respectively.

Figure 5: Qualitative results on Orthographic to Isometric translation ( $\mathcal{V}_{\text{ortho}}\rightarrow\{v_{\text{iso}}\}$ ).

Figure 6: Qualitative results on Partial View Completion ( $\{v_{\text{iso}},v_{\text{front}}\}\rightarrow\{v_{\text{top}},v_{\text{right}}\}$ ). Both the left and right panels display separate generation cases.

Method	$Acc_{cmd}\uparrow$	$Acc_{params}\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	CD $\downarrow$
GPT-4o (5-shot)	65.48%	30.13%	16.52	0.83	0.16	0.30
DrawingsDreamer	92.82%	82.52%	58.61	0.97	0.02	0.01

Table 6: Quantitative evaluation on Partial View Completion (

\{v_{\text{iso}},v_{\text{front}}\}\rightarrow\{v_{\text{top}},v_{\text{right}}\}

). Red and green cells denote the best and second-best results

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Model	$\{v_{\text{iso}}\}\rightarrow\mathcal{V}_{\text{ortho}}$			$\{v_{\text{ortho}}\}\rightarrow\mathcal{V}_{\text{iso}}$
Model	$Acc_{cmd}\uparrow$	$Acc_{param}\uparrow$	CD $\downarrow$	$Acc_{cmd}\uparrow$	$Acc_{param}\uparrow$	CD $\downarrow$
DrawingsDreamer (3B)	89.77%	80.82%	2.41	75.53%	68.36%	2.38
DrawingsDreamer (7B)	91.16%	85.41%	2.04	82.82%	73.51%	1.92

Appendix

Table 7: Quantitative comparison between 3B and 7B parameter models on Cross-View Translation and Partial View Completion tasks. CD is multiplied by

10^{2}

. Red and green cells denote the best and second-best results, respectively.

Figure A3: Qualitative results of unconditional multi-view engineering drawing generation.

Figure A4: Failure cases during Partial View Completion.

Figure A5: Failure case during Isometric to Orthographic translation.

Figure A6: Failure case during Orthographic to Isometric reconstruction.

Explore similar work

Render-in-the-Loop: Vector Graphics Generation via Visual Self-Feedback
Apr 22, 2026Guotao Liang, Zhangcheng Wang, Juncheng Hu +5Scalable Vector Graphic Visual Generation
School of Software, Beihang University, China · 3 4Paradigm, China · Department of Computer Science, The University of Hong Kong, China
Single-Line Drawing Generation via Semantics-Driven Optimization
Jun 1, 2026Tanguy Magne, Alexandre Binninger, Ruben Wiersma +1Sketches Modern Text-To-Image Models
ETH Zurich, Zurich, Switzerland
VectorGym: A Multi-Task Benchmark for SVG Code Generation, Sketching and Editing
Feb 22, 2026Joan Rodriguez, Haotian Zhang, Abhay Puri +14Scalable Vector Graphic Sketches
Mila, Quebec AI Institute · ÉTS Montréal · QuiverAI +11

Apr 22, 2026, cs.CVJ/K move · Enter open · S save

Render-in-the-Loop: Vector Graphics Generation via Visual Self-Feedback

Guotao Liang, Zhangcheng Wang, Juncheng Hu, Haitao Zhou+4

School of Software, Beihang University, China · 3 4Paradigm, China · Department of Computer Science, The University of Hong Kong, China

Scalable Vector Graphic Visual Generation Latent Visual Reasoning Multimodal Large Language Models Sketches Renderer+1

Open paper AskPDF arXiv

Multimodal Large Language Models (MLLMs) have shown promising capabilities in generating Scalable Vector Graphics (SVG) via direct code synthesis. However, existing paradigms typically adopt an open-loop "blind drawing" approach, where models generate symbolic code sequences without perceiving intermediate visual outcomes. This methodology severely underutilizes the powerful visual priors embedded in MLLMs vision encoders, treating SVG generation as a disjointed textual sequence modeling task rather than an integrated visuo-spatial one. Consequently, models struggle to reason about partial canvas states and implicit occlusion relationships, which are visually explicit but textually ambiguous. To bridge this gap, we propose Render-in-the-Loop, a novel generation paradigm that reformulates SVG synthesis as a step-wise, visual-context-aware process. By rendering intermediate code states into a cumulative canvas, the model explicitly observes the evolving visual context at each step, leveraging on-the-fly feedback to guide subsequent generation. However, we demonstrate that applying this visual loop naively to off-the-shelf models is suboptimal due to their inability to leverage incremental visual-code mappings. To address this, we first utilize fine-grained path decomposition to construct dense multi-step visual trajectories, and then introduce a Visual Self-Feedback (VSF) training strategy to condition the next primitive generation on intermediate visual states. Furthermore, a Render-and-Verify (RaV) inference mechanism is proposed to effectively filter degenerate and redundant primitives. Our framework, instantiated on a multimodal foundation model, outperforms strong open-weight baselines on the standard MMSVGBench. This result highlights the remarkable data efficiency and generalization capability of our Render-in-the-Loop paradigm for both Text-to-SVG and Image-to-SVG tasks.

Abstract

Explore similar work

Render-in-the-Loop: Vector Graphics Generation via Visual Self-Feedback

Single-Line Drawing Generation via Semantics-Driven Optimization

VectorGym: A Multi-Task Benchmark for SVG Code Generation, Sketching and Editing