Research map / 3D vision / Gaussian splatting
Empowering Vector Graphics with Consistently Arbitrary Viewing and View-dependent Visibility
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.
Dream3DVG generates text-conditioned, multi-view consistent 3D vector graphics (sketches and iconographies) by jointly optimizing an auxiliary 3DGS branch that supplies coarse-to-fine guidance and a visibility-aware 3DVG renderer that culls occluded strokes.
Problem
Text-to-vector graphics generation lacks 3D consistency across viewpoints, clear shape silhouettes, and view-dependent occlusion handling, since 2D VG methods have no 3D context and existing 3D curve methods render occluded strokes.
Approach
Use an auxiliary 3DGS optimized with ISM/SDS as a consistent intermediate guide for 3D Bezier-curve vector graphics, sample progressively more detailed guidance images by scheduling classifier-free guidance scale, and cull occluded curves via a learned importance MLP plus antipodal-depth visibility voting.
Key results
- On 3D sketch generation, Ours reaches CLIPtext 0.6705 and ALPIPS 0.1656, beating 3Doodle (0.6618, 0.1718), Diff3DS and CLIPasso on both metrics.
- For iconography, Ours achieves CLIPtext 0.6788 vs. SVGDreamer 0.6740 and VectorFusion 0.6683 (2D methods).
- Training takes about 1 hour per prompt on one NVIDIA A100 for 2000 iterations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Dream3DVG instead of 3Doodle | 3D sketch generation · Sketch | CLIPtext: 0.6618 -> 0.6705 | Helped |
| Use Dream3DVG instead of 3Doodle | 3D sketch generation · Sketch | ALPIPS: 0.1718 -> 0.1656 | Helped |
| Use Dream3DVG instead of SVGDreamer | vector graphics generation · Icon | CLIPtext: 0.6740 -> 0.6788 | Helped |
| Replace direct SDS optimization with 3DGS-rendered guidance (ablation) | Dream3DVG · 3D vector graphics generation | shape structure | Helped |
| Use guidance sampled from the 3DGS optimization trajectory instead of 3DGS renderings (ablation) | Dream3DVG · 3D vector graphics generation | shape consistency | Helped |
| Use coarse-to-fine scheduled CFG instead of fixed CFG (ablation) | Dream3DVG · 3D vector graphics generation | semantic shape details | Helped |
| Add importance filtering (ablation) | Dream3DVG visibility-aware rendering · 3D vector graphics generation | curve selection | Helped |
| Add antipodal-depth visibility voting (ablation) | Dream3DVG visibility-aware rendering · 3D vector graphics generation | occlusion-aware rendering | Helped |
| Use coarse-to-fine guidance instead of coarse guidance only (ablation) | Dream3DVG · 3D vector graphics generation | fine details | Helped |
| Use coarse-to-fine guidance instead of fine guidance only (ablation) | Dream3DVG · 3D vector graphics generation | geometry accuracy | Helped |
| Increase the number of paths from 16 to 128 (ablation) | Dream3DVG · 3D vector graphics generation | detail level: N=16; N=128 | Helped |
| Change 3DGS initialization shape or random seed (ablation) | Dream3DVG · 3D vector graphics generation | shape control | Mixed |
Surprising findings
- Ablation shows an SDS-only loss on control points yields intersecting, under-optimized paths, motivating direct 3DGS-derived image supervision.
Limitations
- Visibility is binary per curve, so curves spanning both front and back visibility are rendered incorrectly (e.g., plate under cake).
- Evaluation relies on CLIPtext and ALPIPS proxies on 40 prompts rather than 3D-aware metrics or user studies.
- CLIPasso produces messy results beyond 64 strokes, so stroke counts were fixed per method for comparisons.
Experimental setup
Datasets: 40 text prompts collected from prior works
Benchmarks: 40 text prompts from DreamFusion/SVGDreamer-style prompt sets
Metrics: CLIPtext ALPIPS
Baselines: CLIPasso 3Doodle Diff3DS DiffSketcher VectorFusion SVGDreamer
Builds on: DreamFusion ISM VectorFusion SVGDreamer PyTorch-SVGRender 3Doodle
Closest papers by meaning
- ViewCraft3D: High-fidelity and View-Consistent 3D Vector Graphics Synthesis (NeurIPS 2025 · similarity 0.87)
- EVPGS: Enhanced View Prior Guidance for Splatting-based Extrapolated View Synthesis (CVPR 2025 · similarity 0.84)
- BG-Triangle: Bezier Gaussian Triangle for 3D Vectorization and Rendering (CVPR 2025 · similarity 0.84)
- Chat2SVG: Vector Graphics Generation with Large Language Models and Image Diffusion Models (CVPR 2025 · similarity 0.84)
- DrawingsDreamer: A Unified Multi-View Engineering Drawings Generation Model (NeurIPS 2026 · similarity 0.84)
- StrokeNUWA—Tokenizing Strokes for Vector Graphic Synthesis (ICML 2024 · similarity 0.84)
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.