Agentic image restoration systems typically have to run a tool to see what it does to a degraded image. Failed calls waste computation, and harmful outputs may go unnoticed until too late. A new preprint introduces PreActIR, which instead predicts the consequences of each candidate tool before executing it.
PreActIR's tool-conditioned world model forecasts changes in residual degradation, image quality, and fidelity, along with the risk of harm and the uncertainty of the prediction. The agent executes the top-ranked candidate, and an independent verifier decides whether to commit or roll back the output before the agent replans. The framework also audits whether performance is limited by the selector or by the candidate space, and explores targeted tool adaptation as a controlled way to expand that space.
In a controlled evaluation with a fixed candidate pool, verifier, and budget, PreActIR used 42.5% fewer real tool calls and 63.4% fewer rejected calls than reactive execute-then-evaluate trial and error, while improving Y-PSNR by 0.92 dB. This digest is based on a single arXiv preprint that is under review, so the findings have not yet been peer-reviewed.