Research map / Image generation / Text-to-image diffusion
SVGDreamer: Text Guided SVG Generation with Diffusion Model
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 SVGDreamer, a text-to-SVG method combining attention-guided semantic image vectorization (SIVE) with a particle-based variational score distillation (VPSD) that uses LoRA and reward feedback.
Problem
Text-to-SVG generation methods built on score distillation produce entangled, over-smoothed, over-saturated, low-diversity vector graphics with limited editability and slow convergence.
Approach
Decompose SVG synthesis into foreground objects and background using cross-attention maps to initialize and mask control points (SIVE), and replace SDS with a particle-based variational score distillation that models control points and colors as distributions, fine-tuned via LoRA and reward-weighted sampling (VPSD).
Key results
- SVGDreamer achieves FID 59.13 vs 100.68 for VectorFusion and PSNR 14.54 vs 8.01, indicating higher diversity and less color over-saturation.
- CLIPScore 0.3001 and BLIPScore 0.4623 are the highest among compared methods, with aesthetic score 5.5432 vs 4.9845 for VectorFusion.
- Reward feedback improves HPS from 0.2685 to 0.2559 for the from-scratch variant and aesthetic score from 5.1822 to 5.2825 (from-scratch comparison).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SVGDreamer instead of VectorFusion | Text-to-SVG generation · Six styles | FID: 100.68 -> 59.13 | Helped |
| use SVGDreamer instead of VectorFusion | Text-to-SVG generation · Six styles | PSNR: 8.01 -> 14.54 | Helped |
| use SVGDreamer instead of VectorFusion | Text-to-SVG generation · Six styles | CLIPScore: 0.2720 -> 0.3001 | Helped |
| use SVGDreamer instead of VectorFusion | Text-to-SVG generation · Six styles | BLIPScore: 0.4291 -> 0.4623 | Helped |
| use SVGDreamer instead of VectorFusion | Text-to-SVG generation · Six styles | Aesthetic score: 4.9845 -> 5.5432 | Helped |
| use SVGDreamer instead of VectorFusion | Text-to-SVG generation · Six styles | HPS: 0.2450 -> 0.2685 | Helped |
| add reward feedback (ablation) | SVGDreamer from scratch · Six styles | FID: 84.04 -> 83.21 | Helped |
| add reward feedback (ablation) | SVGDreamer from scratch · Six styles | PSNR: 10.48 -> 10.51 | Helped |
| add reward feedback (ablation) | SVGDreamer from scratch · Six styles | CLIPScore: 0.2951 -> 0.2988 | Helped |
| add reward feedback (ablation) | SVGDreamer from scratch · Six styles | BLIPScore: 0.4311 -> 0.4335 | Helped |
| add reward feedback (ablation) | SVGDreamer from scratch · Six styles | Aesthetic score: 5.1822 -> 5.2825 | Helped |
| add reward feedback (ablation) | SVGDreamer from scratch · Six styles | HPS: 0.2484 -> 0.2559 | Helped |
Limitations
- Editability depends on the underlying text-to-image model's attention decomposition ability.
- Number of control points at the SIVE object level is set manually rather than automatically determined.
Experimental setup
Benchmarks: custom 6-style prompt set (10 prompts per style, 50 SVGs each)
Metrics: FID PSNR CLIPScore BLIPScore Aesthetic score HPS
Baselines: CLIPDraw VectorFusion DiffSketcher LIVE+VectorFusion VectorFusion (scratch)
Builds on: DiffVG VectorFusion DiffSketcher Variational Score Distillation LIVE DreamFusion
Closest papers by meaning
- Chat2SVG: Vector Graphics Generation with Large Language Models and Image Diffusion Models (CVPR 2025 · similarity 0.92)
- SVGBuilder: Component-Based Colored SVG Generation with Text-Guided Autoregressive Transformers (AAAI 2025 · similarity 0.91)
- StarVector: Generating Scalable Vector Graphics Code from Images and Text (AAAI 2025 · similarity 0.89)
- OmniSVG: A Unified Scalable Vector Graphics Generation Model (NeurIPS 2025 · similarity 0.89)
- VFIG: Vectorizing Complex Figures in SVG with Vision-Language Models (NeurIPS 2026 · similarity 0.89)
- InternSVG: Towards Unified SVG Tasks with Multimodal Large Language Models (ICLR 2026 · similarity 0.87)
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.