Research map / 3D vision / 3D generation
Wired Perspectives: Multi-View Wire Art Embraces Generative AI
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 DreamWire, which optimises 3D Bezier wire curves via score distillation from Stable Diffusion/ControlNet with an MST connectivity loss to generate multi-view wire art from text or scribbles.
Problem
Creating multi-view wire art (static 3D sculptures that show different 2D images from different viewpoints) is hard for novices and prior methods are rule-based, fail to generalise, and need manual conflict resolution; the task also suffers from scarce training examples.
Approach
Represent wire art as cubic 3D Bezier curves rendered via a 2D differentiable Bezier renderer (DiffVG) by projecting control points, optimise them with multi-view score distillation sampling from a latent diffusion model (ControlNet for scribble conditions), and enforce connectivity with a Prim's minimum-spanning-tree loss.
Key results
- Ours-T reaches CLIP score 37.21 on text control, versus 34.81 for MVWA and 34.37 for ShadowArt on visual control inputs.
- On DINO-V2-Giant similarity, ShadowArt (83.82) and MVWA (78.86) outperform Ours-V (72.76), which the authors attribute to underfitting/overfitting of the optimisation.
- Setting lambda to 50 balances aesthetics and manufacturability; larger lambda reduces wire connectivity budget but lowers CLIP score and R-Precision.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DreamWire Our-V instead of ShadowArt for visual control | Visual control target sketches · Visual control target sketches | DINO-V2-Base similarity (%): 79.62 -> 69.08 | Hurt |
| use DreamWire Our-V instead of MVWA for visual control | Visual control target sketches · Visual control target sketches | DINO-V2-Giant similarity (%): 78.86 -> 72.76 | Hurt |
| use DreamWire Our-T for text control instead of ShadowArt | Text control · Text control | CLIP Score (%): 34.37 -> 37.21 | Helped |
| add MST regularization (ablation) | DreamWire · Text control | CLIP Score (%): 37.21 -> 36.52 | Hurt |
| increase λ in MST regularization (ablation) | DreamWire · Text control | CLIP Score and R-Precision | Hurt |
| increase λ in MST regularization (ablation) | DreamWire · Text control | wire connectivity budget | Helped |
| replace MST regularization with one Bézier curve containing 150 segments (ablation) | DreamWire · a simple drawing of a bicycle | visual succinctness | Hurt |
Surprising findings
- The proposed optimisation-based method yields lower DINO similarity than the traditional rule-based baselines under visual control, though it is competitive on text control.
Limitations
- Struggles to resolve conflicting visual inputs across views, so visual-control results trail rule-based baselines.
- Limited to three viewpoints due to conflicts in 3D wire space.
- Bezier curves may not fit targets with many non-smooth polylines.
Experimental setup
Datasets: QuickDraw
Benchmarks: Visual control input sets from ShadowArt and MVWA baselines QuickDraw-derived text prompts
Metrics: CLIP score R-Precision DINO-V2 similarity MST wire budget
Baselines: ShadowArt MVWA (Wire art rule-based method)
Builds on: DreamFusion VectorFusion DiffVG ControlNet SketchDreamer
Closest papers by meaning
- NEAT: Distilling 3D Wireframes from Neural Attraction Fields (CVPR 2024 · similarity 0.85)
- LineArt: A Knowledge-guided Training-free High-quality Appearance Transfer for Design Drawing with Diffusion Model (CVPR 2025 · similarity 0.82)
- Sculpt3D: Multi-View Consistent Text-to-3D Generation with Sparse 3D Prior (CVPR 2024 · similarity 0.82)
- ShapeCraft: LLM Agents for Structured, Textured and Interactive 3D Modeling (NeurIPS 2025 · similarity 0.81)
- Sharp-It: A Multi-view to Multi-view Diffusion Model for 3D Synthesis and Manipulation (CVPR 2025 · similarity 0.81)
- ArtiScene: Language-Driven Artistic 3D Scene Generation Through Image Intermediary (CVPR 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.