Research map / Image generation / Text-to-image diffusion
IntroSVG: Learning from Rendering Feedback for Text-to-SVG Generation via an Introspective Generator-Critic Framework
Feiyu Wang, Jiayuan Yang, Zhiyuan Zhao, Da Zhang, Bingyu Li, Peng Liu, Junyu Gao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DPO on generation preferences (ablation) | IntroSVG SFT model (no iteration) · Text-to-SVG benchmark | FID: 30.15 -> 29.76 | Improved |
| add SFT on mixed generation and critic data (ablation) | Qwen2.5-VL-7B · Text-to-SVG benchmark | FID: 71.10 -> 30.15 | Improved |
| add SFT on mixed generation and critic data (ablation) | Qwen2.5-VL-7B · Text-to-SVG benchmark | Aesthetic: 4.32 -> 4.80 | Improved |
| add the iterative generate-critique-refine loop (3 iterations) (ablation) | IntroSVG after SFT+DPO · Text-to-SVG benchmark | FID: 29.76 -> 26.18 | Improved |
| use absolute commands and integer coordinates instead of mixed commands and decimals (ablation) | Qwen2.5-VL-3B SVG training data · SVG data standardization ablation | FID: 121.50 -> 32.15 | Improved |
| use absolute instead of relative commands (integer precision) (ablation) | Qwen2.5-VL-3B SVG training data · SVG data standardization ablation | FID: 37.34 -> 32.15 | Improved |
| use integer instead of decimal coordinates (absolute commands) (ablation) | Qwen2.5-VL-3B SVG training data · SVG data standardization ablation | FID: 34.62 -> 32.15 | Improved |
| use integer instead of decimal coordinates (absolute commands) (ablation) | Qwen2.5-VL-3B SVG training data · SVG data standardization ablation | RSR%: 95.89 -> 98.62 | Improved |
| apply introspective loop as zero-shot prompting (Iter 0 to Iter 3) | Grok-4 · Text-to-SVG benchmark | FID: 41.39 -> 32.85 | Improved |
| apply introspective loop as zero-shot prompting (Iter 0 to Iter 3) | GPT-4o · Text-to-SVG benchmark | FID: 37.80 -> 36.34 | Improved |
| apply introspective loop as zero-shot prompting (Iter 0 to Iter 3) | GPT-5 · Text-to-SVG benchmark | CLIP-T2I: 0.2858 -> 0.2779 | Mixed |
| add more refinement iterations (Iter 0 to Iter 3) | IntroSVG · Text-to-SVG benchmark | FID: 29.76 -> 26.18 | Improved |
Closest papers by meaning
- Chat2SVG: Vector Graphics Generation with Large Language Models and Image Diffusion Models (CVPR 2025 · similarity 0.91)
- Rendering-Aware Reinforcement Learning for Vector Graphics Generation (NeurIPS 2025 · similarity 0.90)
- InternSVG: Towards Unified SVG Tasks with Multimodal Large Language Models (ICLR 2026 · similarity 0.89)
- Empowering LLMs to Understand and Generate Complex Vector Graphics (CVPR 2025 · similarity 0.89)
- SVGDreamer: Text Guided SVG Generation with Diffusion Model (CVPR 2024 · similarity 0.89)
- OmniSVG: A Unified Scalable Vector Graphics Generation Model (NeurIPS 2025 · similarity 0.89)
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.
