PROMPTMINER: Black-box prompt stealing from text-to-image generation models with reinforcement learning and fuzzy optimization

The researchers propose PROMPTMINER, a black-box prompt stealing framework for text-to-image models such as Stable Diffusion and FLUX. The method splits the task into two stages: first reconstructing the subject with reinforcement learning, then recovering style modifiers with fuzzy search. Experiments are conducted on multiple datasets and diffusion backbones, with CLIP similarity up to 0.958 and SBERT text alignment up to 0.751, both exceeding the baselines. On real images with an unknown generator, its CLIP similarity is 7.5% higher than the strongest baseline, and it maintains strong performance under defensive perturbations. The code has been released on GitHub.