Research map / Image generation / Text-to-image diffusion
Idea2Img: Iterative Self-Refinement with GPT-4V for Automatic Image Design and Generation
Zhengyuan Yang, Jianfeng Wang, Linjie Li, Kevin Lin, Chung-Ching Lin, Zicheng Liu, Lijuan Wang
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 iterative self-refinement (T=3) on top of initial-round Idea2Img prompt (ablation) | Idea2Img with SDXL v1.0 · Idea2Img eval set (104 ideas) | user preference score (%): 29.8 -> 56.7 | Improved |
| add iterative self-refinement on top of initial-round Idea2Img prompt (ablation) | Idea2Img with SD v1.5 · Idea2Img eval set (104 ideas) | user preference score (%): 43.3 -> 48.1 | Improved |
| add iterative self-refinement on top of initial-round Idea2Img prompt (ablation) | Idea2Img with SD v2.1 · Idea2Img eval set (104 ideas) | user preference score (%): 40.4 -> 46.2 | Improved |
| use GPT-4V initial prompt generation and draft selection instead of manual prompts (ablation) | SDXL v1.0 · Idea2Img eval set (104 ideas) | user preference score (%): 13.5 -> 29.8 | Improved |
| apply Idea2Img to stronger T2I models rather than weaker ones | Idea2Img · Idea2Img eval set (104 ideas) | iteration gain in user preference: +26.9 (SDXL) versus +4.8 (SD v1.5) | Improved |
| apply Idea2Img when users compare across T2I models | SDXL · preferred T2I model selection | user preference rate (%): 46.2% -> 65.4% | Improved |
| replace GPT-4V with open-source LLaVA-1.5-13B as the LMM (ablation) | Idea2Img · qualitative cases | refinement effectiveness | Hurt |
| add draft image selection among N drafts per round | Idea2Img · design rationale | noise in iterative refinement | Improved |
Closest papers by meaning
- Imagine Before You Draw: Visual Prompt Engineering for Image Generation (NeurIPS 2026 · similarity 0.87)
- IDEA-Bench: How Far are Generative Models from Professional Designing? (CVPR 2025 · similarity 0.86)
- DuoGen: Towards Autonomous Interleaved Multimodal Generation (CVPR 2026 · similarity 0.85)
- Breaking Dual Bottlenecks: Evolving Unified Multimodal Models into Self-Adaptive Interleaved Visual Reasoners (ICML 2026 · similarity 0.85)
- Benchmarking and Evolving Reason-Reflect-Rectify for Reflective Visual Generation (ICML 2026 · similarity 0.85)
- MMIG-Bench: Towards Comprehensive and Explainable Evaluation of Multi-Modal Image Generation Models (NeurIPS 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.
