Research map / 3D vision / 3D generation

DrawingsDreamer: A Unified Multi-View Engineering Drawings Generation Model

NeurIPS 2026 · 3D vision › 3D generation · approach: Feed-forward 3D asset generation and editing · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DrawingsDreamer, a LoRA-tuned Llama-3.2-3B that generates multi-view SVG engineering drawings as pure token sequences, using hierarchical postfix tokenization and a three-stage task-aware curriculum.

Problem

Generating multi-view engineering drawings as SVG with geometric fidelity and cross-view alignment, which existing artistic SVG generators and CAD BRep methods do not address.

Approach

Formulates multi-view engineering drawing generation as pure sequence modeling with an LLM (no raster encoder), using postfix tokenization that emits coordinates before their primitive boundary token, and a microscopic-to-macroscopic curriculum over unified multitask instruction formats.

Key results

• On iso-to-ortho translation, DrawingsDreamer reaches 89.77% Acc_cmd and 80.82% Acc_params, with PSNR 50.16 and CD 0.02, outperforming GPT-4o 5-shot (46.44% / 35.54%, PSNR 33.53).
• On unconditional generation, JSD 0.61 and COV 66.24% vs GPT-4o 1.17 and 33.12%.
• Scaling to 7B improves iso-to-ortho Acc_params from 80.82% to 85.41% and partial-view CD from 2.38 to 1.92.

Surprising findings

• The 3B model's partial-view completion results are reported inconsistently across tables (e.g., Acc_params 68.36% vs 82.52% for the same task), suggesting table mislabeling.
• Prefix-format ablation improves some metrics on the iso-to-ortho task (Acc_cmd 88.11% vs ours 89.77%, Acc_params 80.58% close to ours), so postfix gains are mainly on specific metrics.

Limitations

• Failures on repetitive structures (counting holes or slots) due to autoregressive limits.
• Dimensional and coordinate drift on thin or elongated objects due to quantization.
• Baselines are weak or not directly comparable; no direct baseline for the novel task, and Zero123 and OmniSVG are compared on different metrics.

Experimental setup

Datasets: DeepCAD

Benchmarks: DeepCAD-derived multi-view test split (8,912 models)

Metrics: Acc_cmd Acc_params PSNR SSIM LPIPS Chamfer Distance JSD COV MMD

Baselines: GPT-4o (5-shot) OmniSVG Zero123

Builds on: LoRA OmniSVG DeepCAD

Closest papers by meaning

• From 2D CAD Drawings to 3D Parametric Models: A Vision-Language Approach (AAAI 2025 · similarity 0.88)
• VCG-Bench: Towards A Unified Visual-Centric Benchmark for Structured Generation and Editing (ICML 2026 · similarity 0.87)
• Text-to-CAD Generation Through Infusing Visual Feedback in Large Language Models (ICML 2025 · similarity 0.86)
• Revisiting CAD Model Generation by Learning Raster Sketch (AAAI 2025 · similarity 0.86)
• CAD-Llama: Leveraging Large Language Models for Computer-Aided Design Parametric 3D Model Generation (CVPR 2025 · similarity 0.86)
• LineArt: A Knowledge-guided Training-free High-quality Appearance Transfer for Design Drawing with Diffusion Model (CVPR 2025 · similarity 0.85)