Teaching image restoration agents to learn from their own mistakes

Image restoration agents powered by large language models often fail on their first try because they lack experience with real degradation patterns. EvoIR-Agent solves this by building a hierarchy of experiences from trial-and-error attempts, then automatically refines this experience pool over time. The result: better restoration quality with fewer attempts, and the ability to handle new damage types without retraining—matching or beating specialized methods while staying flexible.