Researchers unveil PROVE to extract AI image prompts without training
New black-box attack recovers prompts from images with 100% verifiable accuracy...
Researchers from Boston University, University of Massachusetts Amherst, and Adobe Research have developed PROVE (Prompt Recovery with Verified Evidence), a training-free black-box attack that reconstructs text-to-image prompts from generated outputs with verifiable accuracy. Unlike existing prompt inversion methods that rely on gradient optimization, autoregressive captioning, or reinforcement learning—which often produce hallucinated or unreadable outputs—PROVE composes prompts by extracting grounded scene descriptions directly from image evidence.
The method achieves superior performance across standard benchmarks (MS-COCO, Flickr30K, and Lexica) using state-of-the-art generators like Stable Diffusion and DALL-E, outperforming optimization, captioning, and RL-based baselines in image similarity (DINO, LPIPS) and text-image alignment (CLIP) metrics. Notably, PROVE operates without any training, fine-tuning, or generator access, making it a practical and scalable threat to copyrighted AI-generated content and prompt marketplaces.
- PROVE recovers prompts from AI-generated images using only black-box access, with no training or model fine-tuning required
- Achieves 10-20% higher accuracy than baselines on DINO/LPIPS metrics across MS-COCO and Flickr30K datasets
- Outputs are fully auditable, with each recovered prompt grounded in explicit image evidence
Why It Matters
PROVE exposes critical vulnerabilities in AI-generated content protection, threatening prompt marketplaces and copyright enforcement for artists and developers.