DrawingsDreamer: A Unified Multi-View Engineering Drawings Generation Model

• Engineering drawings demand structural validity, exact geometric parameterization, and rigorous cross-view alignment beyond the requirements of artistic creation.Source: §1 · p. 2
• Existing SVG models struggle with engineering coordinate dependencies, while CAD generation has largely focused on 3D models rather than directly generating 2D drawings.Optimization-based methods also face high computational overhead and can produce dense, unstructured anchor points that reduce editability.Source: §1 · p. 2
• DrawingsDreamer frames multi-view drawing synthesis as sequence modeling without raster image encoders, using hierarchical postfix tokenization to establish local coordinates before semantic boundaries.Source: §1 · p. 2
• The model is trained on a multi-view aligned corpus rendered from 3D CAD collections, with a progressive curriculum spanning topological repair, partial view completion, and unconstrained generation.Source: §1 · p. 2
• DrawingsDreamer achieves state-of-the-art geometric fidelity and syntactic accuracy across diverse conditional and unconditional multi-view generation tasks.Source: §1 · p. 2

2.1 SVG Generation

• Optimization-based SVG methods refine individual instances but can incur prohibitive overhead and produce unstructured outputs; learning-based work has expanded through datasets, RNNs, VAEs, and LLM sequence models.The cited optimization methods are [36], datasets [54] [7] [20] [63], RNNs [44] [13] [40] [16], VAEs [5] [34] [46] [47], and LLM approaches [48] [1] [62] [8] [49].Source: §2.1 · pp. 2–3
• Existing LLM-driven SVG models are constrained by context length and artistic-domain training, leaving aligned multi-view engineering drawings an unexplored generation task.Source: §2.1 · p. 3

2.2 Computer-Aided Design

• CAD generation methods predominantly reconstruct BRep models from point clouds, images, or text, while some work reconstructs CAD models from engineering drawings.Source: §2.2 · p. 3
• Directly generating multi-view aligned drawings from standard 3D CAD databases is presented as a missing link in automating manufacturing blueprints.Source: §2.2 · p. 3

3.1 Streamlined Representation for Multi-View Engineering Drawings

• The representation projects CAD objects into four principal views and uses postfix tokens to place geometric coordinates before semantic boundary tokens.Coordinates are globally scaled to preserve cross-view alignment, with the isometric view scaled independently, then quantized into discrete bins.Source: §3.1 · pp. 3–4, Table 1 · p. 4
• Coordinates, primitives, and paths are hierarchically grouped into view sequences, while task instructions vary the context and target views for unified multitask training.This formulation accommodates localized geometric restoration as well as full spatial projection.Source: §3.1 · pp. 4–5, Figure 2 · p. 4, Table 2 · p. 4

3.2 Instruction-Tuning with Task-Aware Curriculum Schedule

• A single Llama-family model is instruction-tuned across generation tasks, with loss applied only to target sequences and instruction/context tokens masked.LoRA provides the trainable parameters for optimization.Source: §3.2 · p. 5
• The curriculum keeps unconditional generation active while shifting conditional tasks from path-level syntax and topology, to partial-view completion, then to isometric–orthographic translation.The stages progress from local geometric priors to cross-view alignment and global spatial tasks.Source: §3.2 · pp. 5–6, Figure 3 · p. 5

4.1 Experimental Setup

• The evaluation uses LLaMA-3.2-3B with LoRA on 178,238 valid DeepCAD models, split into 160,414 training, 8,912 validation, and 8,912 test models.Training runs for 15 epochs with AdamW and batch size 16.Source: §4.1 · p. 6

4.2 Metrics

• Conditional tasks are assessed with command and parameter accuracy, raster-image PSNR, SSIM, and LPIPS, plus cross-view Chamfer Distance; unconditional generation uses JSD, coverage, and MMD.GPT-4o is evaluated with five randomly selected training exemplars, and the baselines also include OmniSVG and Zero123.Source: §4.2 · p. 7, §4.3 · p. 7

4.3 Performance Comparison with Existing Methods

• DrawingsDreamer is reported to outperform strong baselines on cross-view translation, while achieving competitive JSD and MMD in unconditional generation and strong partial-view completion results.The reported comparisons cover translation, completion, and unconditional generation; the supplied passages give no numerical cell values.Source: §4.3 · p. 7, Table 4 · p. 7, Table 3 · p. 7
• Qualitative examples show aligned orthographic reconstructions, consistent isometric views, and structurally aligned completed views despite complex internal contours.Figure 5 covers orthographic-to-isometric translation, while Figure 6 shows separate partial-completion cases.Source: §4.3 · p. 8, Figure 5 · p. 8, Figure 6 · p. 9

4.4 Ablation Studies

• Unified multitask training outperforms isolated task-specific optimization, supporting cross-task knowledge transfer for multi-view spatial reasoning.The comparison includes specialized training for tasks such as isometric-to-orthographic translation.Source: §4.4 · p. 9
• Random task sampling instead of the progressive three-stage curriculum leads to suboptimal convergence and degraded Chamfer Distance.The reported curriculum orders local syntax before global spatial translation.Source: §4.4 · p. 9
• Removing path-level semantic masking degrades geometric fidelity and command accuracy in generated multi-view sequences.The authors describe learning basic geometric structures as a stepping stone to coherent macroscopic generation.Source: §4.4 · p. 9
• Replacing hierarchical postfix tokenization with standard prefix command sequences lowers parameter accuracy.The postfix representation regresses coordinates before assigning semantic boundaries.Source: §4.4 · p. 9

5 Conclusion

• DrawingsDreamer frames multi-view engineering drawing synthesis as sequence modeling, using hierarchical postfix tokenization and a progressive task-aware curriculum to capture cross-view alignments and geometric dependencies.The conclusion reports state-of-the-art generation of geometrically precise parametric engineering drawings.Source: §5 · p. 9

Appendix

• The appendix lists additional material on data processing, prompting and task configurations, model capacity, evaluation metrics, qualitative results, and limitations and future work.The supplied appendix passage provides the topics, not their substantive details.Source: Appendix · p. 13

• Each CAD program is used to reconstruct a three-dimensional solid and render Front, Right, Top, and isometric views; samples with missing or empty views are discarded.The resulting views are fixed projections from the reconstructed solid.Source: A Data Processing Details · p. 13
• All views are normalized to a 1000 × 1000 canvas, with orthographic views sharing a scale and the isometric view normalized independently.The shared orthographic scale preserves relative sizes among those views.Source: A Data Processing Details · p. 13
• SVG paths are canonicalized by grouping recoverable closed contours into loops and sorting them spatially, while other paths use deterministic traversal order.The procedure preserves visible segments and reduces ambiguity from arbitrary export order.Source: A Data Processing Details · p. 13
• The pipeline retains move, line, and cubic Bézier commands; continuous coordinates x in [0, L] are converted to absolute quantized values using coordinate quantization resolution Q = 256.Source: A Data Processing Details · p. 13
• The four view token sequences are concatenated in Front, Right, Top, Iso order, with a view-specific ending token after each view.The resulting training instance stores the model identifier and quantization resolution alongside the sequence.Source: A Data Processing Details · p. 13

• GPT-4o baseline prompts prepend five randomly sampled input-output exemplars, including instructions and answers, to the target query.Figure A1 illustrates the system instruction and exemplar layout.Source: B Prompting Templates and Task… · p. 13, Figure A1 · p. 14
• Task-specific instruction prompts are prepended to context sequences for each multitask configuration.Figure A2 provides the textual templates used during unified multitask training.Source: B Prompting Templates and Task… · p. 13, Figure A2 · p. 15
• The sequence-construction algorithm normalizes orthographic views with a shared scale, normalizes the isometric view independently, canonicalizes paths, then quantizes and serializes them into a fixed view order.It returns the model identifier, quantization resolution, and concatenated token sequence.Source: B Prompting Templates and Task… · p. 16

C The Impact of Model Capacity

• The 7B variant uses the same LoRA configuration and progressive task-aware curriculum as the 3B model, with batch size adjusted for memory constraints.The primary experiments use a 3B model for performance and computational efficiency; the 7B experiment explores whether greater capacity can mitigate topological hallucinations and dimensional drift.Source: C The Impact of Model Capacity · p. 16
• 82.82% Parameter Accuracy versus 75.53% and 1.92 Chamfer Distance versus 2.38: the 7B model improves both metrics on Partial View Completion.The comparison is for {viso, vfront} → {vtop, vright}.Source: C The Impact of Model Capacity · p. 16, Table 7 · p. 17
• Qualitatively, the 7B model reduces counting errors in repetitive structures and coordinate drift in thin or elongated objects.The passage attributes these changes to stronger local topological memory and tighter spatial constraints.Source: C The Impact of Model Capacity · p. 16

D Evaluation Metric Details

• For conditional generation, evaluation covers only the views specified as targets: three orthographic views for ISO-to-three-view, ISO for three-view-to-ISO, and the two remaining orthographic views for partial completion.Metrics are averaged over target views for each sample and then across the evaluation set; missing target views count as failed predictions.Source: D Evaluation Metric Details · p. 17
• Symbolic accuracy measures operation-type matches at corresponding sequence positions and numerical-parameter reconstruction within a fixed coordinate tolerance τ.The metrics are Command Accuracy and Parameter Accuracy, respectively.Source: D Evaluation Metric Details · p. 17
• Raster-level metrics compare 256 × 256 white-background renderings per target view using PSNR, SSIM, LPIPS, and Chamfer Distance on extracted line pixels.Chamfer Distance directly measures geometric alignment between predicted and ground-truth line drawings.Source: D Evaluation Metric Details · p. 17
• Unconditional generation is compared with the test-set distribution using JSD, COV, and MMD on joint point clouds formed from all four views.MMD reflects sample quality, COV diversity, and JSD occupancy-distribution differences; lower MMD and JSD and higher COV indicate better distributional matching.Source: D Evaluation Metric Details · p. 18
• Zero123 and DrawingsDreamer use the same per-view raster evaluation protocol, while symbolic accuracy is not applicable to Zero123’s raster outputs.Both are evaluated on aligned 256 × 256 RGB target views using PSNR, SSIM, LPIPS, and CD; CD uses the same foreground-pixel thresholding and bidirectional Chamfer formulation.Source: D Evaluation Metric Details · p. 18

E Additional Qualitative Results

• Unconditional samples illustrate diverse, structurally valid, strictly aligned multi-view blueprints generated without contextual conditioning.The examples are provided in Figure A3.Source: E Additional Qualitative Results · p. 18

F Limitations and Future Work

• Autoregressive generation can miscount repetitive features, such as omitting a circular hole or adding a symmetric slot, when long repeated-token sequences challenge attention without explicit counting.Source: F Limitations and Future Work · p. 18
• Thin, flat, or elongated objects can exhibit dimensional inconsistency across views; small coordinate drift or quantization artifacts may amplify during translation.Reported errors include inaccurate lengths, irregular thicknesses, distorted hole radii, and misaligned coordinates.Source: F Limitations and Future Work · p. 19
• Future work could add explicit cross-view alignment rules or train with reinforcement-learning rewards that penalize geometric discrepancies.The proposed reward approach aims to internalize spatial alignment while retaining data-driven sequence modeling without hand-crafted heuristics.Source: F Limitations and Future Work · p. 19

• The authors state that the abstract and introduction accurately reflect the paper’s contributions and scope, and that the paper includes no theoretical results.Source: NeurIPS Paper Checklist · pp. 20–21
• The authors state that limitations are discussed in Appendix F.Source: NeurIPS Paper Checklist · p. 20
• The authors report that Section 4 provides the information needed to reproduce the main experimental results and specifies training and test details.Source: NeurIPS Paper Checklist · pp. 21–23
• The authors say they will provide data, code, and instructions for reproducing the results, and release the codebase with detailed README files.Source: NeurIPS Paper Checklist · pp. 22, 25
• The authors report consistent performance across multiple runs with fixed random seeds and state that Section 4.1 describes the compute resources used.Source: NeurIPS Paper Checklist · p. 23
• The checklist states that the work does not use crowdsourcing or human subjects and claims no societal impact beyond engineering drawings generation.Source: NeurIPS Paper Checklist · pp. 24–25