Look Before You Restore: Learning to Predict Before Acting for Agentic Image Restoration

Key points

• Agents for mixed-degradation image restoration typically learn what a tool does to an image only after running it: failed calls waste computation and harmful outputs go unnoticed until too late.1
• For each candidate tool, a tool-conditioned world model forecasts changes in residual degradation, image quality, and fidelity, together with the risk of harm and the uncertainty of the prediction.1
• PreActIR also audits whether performance is limited by the selector or by the candidate space, and studies targeted tool adaptation as a controlled way to expand that space.1
• In a controlled evaluation that fixes the candidate pool, verifier, and budget and employs reference-annotated degradation states, PreActIR uses 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.1

Sources (1)

• [1]Look Before You Restore: Learning to Predict Before Acting for Agentic Image Restoration
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 14, 03:50 PM

  We propose PreActIR, an agentic restoration framework that predicts tool consequences before execution.

  Agents for mixed-degradation image restoration typically learn what a tool does to an image only after running it: failed calls waste computation and harmful outputs go unnoticed until too late.

Extractive summary: sentences quoted from the sources.