Research map / Image generation / Text-to-image diffusion
PROMPTMINER: Black-Box Prompt Stealing against Text-to-Image Generative Models via Reinforcement Learning and VLM-Guided Optimization
Mingzhe Li, Renhao Zhang, Zhiyang Wen, Siqi Pan, Bruno Castro da Silva, Juan Zhai, Shiqing Ma
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 |
|---|---|---|---|
| use PromptMiner instead of the best baseline per model on in-the-wild images | prompt stealing · DiffusionDB in-the-wild | CLIP similarity: 0.803 -> 0.863 | Improved |
| add potential-based reward shaping to the sparse image-similarity reward in RL phase (ablation) | PromptMiner Phase I · training dynamics | image similarity vs steps | Improved |
| increase Phase I query budget beyond about 100 on simple prompts (ablation) | PromptMiner Phase I · Flickr | CLIP similarity | No clear effect |
| increase Phase I query budget on modifier-rich prompts (ablation) | PromptMiner Phase I · Lexica | CLIP similarity | Improved |
| increase Phase II query budget (ablation) | PromptMiner Phase II · Flickr and Lexica | CLIP similarity | Mixed |
| apply random noise injection to target image | PromptMiner · Lexica, SD 3.5 Medium | CLIP similarity: 0.911 -> 0.903 | No clear effect |
| apply puzzle effect to target image | PromptMiner · Lexica, SD 3.5 Medium | CLIP similarity: 0.911 -> 0.902 | No clear effect |
| overlay textual watermark on target image | PromptMiner · Lexica, SD 3.5 Medium | CLIP similarity: 0.911 -> 0.887 | Hurt |
| two-phase RL subject plus VLM modifier search versus caption-only BLIP on SBERT | PromptMiner · MS COCO, SD 3.5 Medium | SBERT: 0.718 -> 0.751 | Improved |
| PromptMiner versus PromptStealer on Lexica text alignment with SD v1.5 | PromptMiner · Lexica, SD v1.5 | SBERT: 0.563 vs 0.563 | No clear effect |
Closest papers by meaning
- Learning to Sample Effective and Diverse Prompts for Text-to-Image Generation (CVPR 2025 · similarity 0.88)
- On Discrete Prompt Optimization for Diffusion Models (ICML 2024 · similarity 0.88)
- Prompting4Debugging: Red-Teaming Text-to-Image Diffusion Models by Finding Problematic Prompts (ICML 2024 · similarity 0.88)
- Prompting Hard or Hardly Prompting: Prompt Inversion for Text-to-Image Diffusion Models (CVPR 2024 · similarity 0.88)
- PromptCoT: Align Prompt Distribution via Adapted Chain-of-Thought (CVPR 2024 · similarity 0.87)
- Defending Multimodal Backdoored Models by Repulsive Visual Prompt Tuning (NeurIPS 2025 · 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.
