Natural language-conditioned reinforcement learning (RL) enables agents to follow human instructions.

Foundation models have emerged as powerful tools across various domains including language, vision, and multimodal tasks.

Image restoration, aiming to restore high-quality images from their degraded counterparts, is a fundamental computer vision problem and has been studied for many years. Due to its ill-posed nature, early research efforts [1,2,3]typically focus on developing single-task models, with each model handling only one specific degradation.