Research map / Image generation / Text-to-image diffusion
Optimizing Prompts for Text-to-Image Generation
Yaru Hao, Zewen Chi, Li Dong, Furu Wei
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 RL (PPO) after supervised fine-tuning (ablation) | Promptist GPT-2 · COCO out-of-domain | absolute reward improvement: 0.28 -> 0.48 (+71%) | Improved |
| add RL after SFT (ablation) | Promptist GPT-2 · Lexica RTP | absolute reward improvement: 0.11 -> 0.25 (+127%) | Improved |
| add RL after SFT (ablation) | Promptist GPT-2 · Lexica MC | absolute reward improvement: 0.36 -> 0.47 (+31%) | Improved |
| add RL after SFT (ablation) | Promptist GPT-2 · Lexica MCM | absolute reward improvement: 0.16 -> 0.17 (+6%) | No clear effect |
| add RL after SFT (ablation) | Promptist GPT-2 · DiffusionDB | aesthetic score: 6.15 -> 6.26 | Improved |
| add RL after SFT (ablation) | Promptist GPT-2 · DiffusionDB | relevance (CLIP score): 0.25 -> 0.26 | Improved |
| use optimized prompts instead of human-engineered prompts | Stable Diffusion v1.4 · DiffusionDB | aesthetic score: 5.87 -> 6.26 | Improved |
| use optimized prompts instead of human-engineered prompts | Stable Diffusion v1.5 · Lexica | reward: -0.05 -> 0.05 | Improved |
| use learned prompt adaptation instead of appending common hand-selected tags | Stable Diffusion · DiffusionDB | reward: -0.16 -> 0.06 (Tag3 vs Ours) | Improved |
| add source prompt augmentation in SFT data (ablation) | SFT policy · Lexica and COCO | reward | Improved |
| add random modifiers to user input (MCM) instead of learned optimization (ablation) | Lexica prompts · Lexica MCM | reward | No clear effect |
| human preference: optimized prompts vs manually engineered prompts | Promptist · Human evaluation | preference | Mixed |
Closest papers by meaning
- Learning to Sample Effective and Diverse Prompts for Text-to-Image Generation (CVPR 2025 · similarity 0.92)
- Dynamic Prompt Optimizing for Text-to-Image Generation (CVPR 2024 · similarity 0.91)
- A User-Friendly Framework for Generating Model-Preferred Prompts in Text-to-Image Synthesis (AAAI 2024 · similarity 0.91)
- Preference Adaptive and Sequential Text-to-Image Generation (ICML 2025 · similarity 0.90)
- Subject-driven Text-to-Image Generation via Preference-based Reinforcement Learning (NeurIPS 2024 · similarity 0.90)
- Multimodal Large Language Models Make Text-to-Image Generative Models Align Better (NeurIPS 2024 · 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.
