Research map / 3D vision / 3D generation
ViewCraft3D: High-fidelity and View-Consistent 3D Vector Graphics Synthesis
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.
Generates view-consistent 3D Bézier-curve vector graphics from a single image by fitting curves to salient 3D mesh edges (from a TripoSG-reconstructed mesh) and refining with 3D SDS.
Problem
Existing 3D vector graphics generation relies on 2D priors and per-view SDS, causing cross-view inconsistency, messy strokes, and long runtimes.
Approach
Work directly in 3D: reconstruct a mesh from one image, extract salient edge points, cluster them by spatial proximity and orientation, fit 3D cubic Bézier curves with Chamfer distance, then refine missing details with SDS from a pretrained image-to-3D model so view consistency holds by construction.
Key results
- VC3D achieves CLIPScore 0.799 and Aesthetic Score 4.352, versus 0.729/4.122 for 3Doodle and 0.673/3.769 for Diff3DS.
- VC3D takes about 0.5 hours per object, compared with about 6 hours for 3Doodle and about 2 hours for Diff3DS (single RTX 4090).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use VC3D instead of 3Doodle | 3D vector graphics generation · 3D vector graphics evaluation (40 images) | CLIPScore: 0.729 -> 0.799 | Helped |
| use VC3D instead of Diff3DS | 3D vector graphics generation · 3D vector graphics evaluation (40 images) | CLIPScore: 0.673 -> 0.799 | Helped |
| use VC3D instead of 3Doodle | 3D vector graphics generation · 3D vector graphics evaluation (40 images) | Aesthetic Score: 4.122 -> 4.352 | Helped |
| use VC3D instead of Diff3DS | 3D vector graphics generation · 3D vector graphics evaluation (40 images) | Aesthetic Score: 3.769 -> 4.352 | Helped |
| use VC3D instead of 3Doodle | 3D vector graphics generation · 3D vector graphics evaluation | runtime: ~6 hours -> approximately 0.5 hours | Helped |
| use VC3D instead of Diff3DS | 3D vector graphics generation · 3D vector graphics evaluation | runtime: ~2 hours -> approximately 0.5 hours | Helped |
| add Chamfer Distance optimization to salient point cloud extraction and clustering (ablation) | VC3D · 3D vector graphics ablation (20 images) | CLIPScore: 0.779 -> 0.805 | Helped |
| add Chamfer Distance optimization to salient point cloud extraction and clustering (ablation) | VC3D · 3D vector graphics ablation (20 images) | Aesthetic Score: 4.096 -> 4.167 | Helped |
| add detail refinement with SDS loss to primary structure fitting (ablation) | VC3D · 3D vector graphics ablation (20 images) | CLIPScore: 0.805 -> 0.818 | Helped |
| add detail refinement with SDS loss to primary structure fitting (ablation) | VC3D · 3D vector graphics ablation (20 images) | Aesthetic Score: 4.167 -> 4.297 | Helped |
Limitations
- Curves lack occlusion relationships and share uniform transparency when rendered.
- Evaluation is small (40 inputs) and relies on CLIPScore and aesthetic score; no direct view-consistency metric reported.
- Ablation details (variant definitions) are ambiguous.
Experimental setup
Datasets: 40 images collected from prior works and online sources
Benchmarks: 40 input images rendered into 12 views
Metrics: CLIPScore Aesthetic Score generation time
Baselines: 3Doodle Diff3DS
Builds on: Dora sharp edge sampling EdgeGaussians DiffVG 3Doodle Score Distillation Sampling
Closest papers by meaning
- Empowering Vector Graphics with Consistently Arbitrary Viewing and View-dependent Visibility (CVPR 2025 · similarity 0.87)
- Diff3DS: Generating View-Consistent 3D Sketch via Differentiable Curve Rendering (ICLR 2025 · similarity 0.86)
- 3DCodeBench: Benchmarking Agentic Procedural 3D Modeling Via Code (NeurIPS 2026 · similarity 0.86)
- Rapid 3D Model Generation with Intuitive 3D Input (CVPR 2024 · similarity 0.86)
- GenVDM: Generating Vector Displacement Maps From a Single Image (CVPR 2025 · similarity 0.86)
- BG-Triangle: Bezier Gaussian Triangle for 3D Vectorization and Rendering (CVPR 2025 · similarity 0.85)
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.