Driving · 2026-10-15

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

Xu Zhang, Jun Wan, Lefei Zhang, Liangpei Zhang

arXiv:2610.20384PDF

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

PreActIR is an agentic image restoration framework that predicts candidate tool consequences before execution using a tool-conditioned world model (including predicted benefit/harm probabilities and uncertainty). The agent executes only the top-ranked candidate, and an independent verifier commits or rolls back its output before replanning. In controlled replays, it uses 42.5% fewer real tool calls and 63.4% fewer rejected calls than reactive trial-and-error, while improving Y-PSNR by 0.92 dB.

Why it matters

The authors address mixed-degradation restoration as a sequential decision problem where tool outcomes are unknown until after execution. PreActIR changes the interaction loop to a predict–act–verify cycle so the agent can rank candidates and estimate risk before spending real tool calls, with rollback based on verification.

Method

  • Tool-conditioned world model predicts per-candidate changes in residual degradation, image quality, and fidelity, plus content damage, harm/benefit probabilities, and predictive uncertainty.
  • Predict–act–verify loop: select top-ranked candidate from predicted utility, execute it once, and use an independent verifier to commit or roll back before replanning from the observed state.
  • Candidate-frontier audit and targeted tool adaptation: offline analysis separates selector headroom from tool-set limits; adapt an existing operator (HAT-S) and re-audit the expanded pool.

Limitation

The authors report that PreActIR improves fidelity efficiency and some metrics, but its PSNR on the paired real-world 100-image set falls behind other agents (e.g., SEAR), indicating a fidelity–perception trade-off on that set.

Abstract (from arXiv)

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. We propose PreActIR, an agentic restoration framework that predicts tool consequences before execution. 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. The agent executes the top-ranked candidate, and an independent verifier commits or rolls back its output before the agent replans from the observed state. 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. 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.

Related papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-14 · Indexed by Pub-AI: 2026-10-15 UTC

preprint

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

Xu Zhang, Jun Wan, Lefei Zhang, Liangpei Zhang

Preprint posted: 2026-10-14 · Original record ↗

arxiv
2610.20384
Explicit source versions (1)
  • v1 · 2026-10-14
  1. Preprint available

    Source date: 2026-10-14 · First detected: 2026-10-25 UTC · Historical baseline

    Source evidence (1)
Abstract · arxiv

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. We propose PreActIR, an agentic restoration framework that predicts tool consequences before execution. 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. The agent executes the top-ranked candidate, and an independent verifier commits or rolls back its output before the agent replans from the observed state. 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. 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.