Research map / Multi-agent systems and games / LLM agents

OPERA: An Agent for Image Restoration with End-to-End Joint Planning–Execution Optimization

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: GUI agent memory and trajectories · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes OPERA, an agent that uses GRPO-trained VLM planning to output full tool-invocation plans in one pass, plus joint end-to-end fine-tuning of restoration tools under those plans, for mixed-degradation image restoration.

Problem

Agent-based image restoration is limited by implicitly constrained planning spaces (one-to-one degradation-tool matching, greedy stepwise planning) and by independently pretrained tools that are not coordinated for sequential composition under mixed degradations.

Approach

Replace step-by-step agent planning with single-shot full-plan generation by a VLM trained with GRPO using final restoration quality (IQA-based) as reward, and jointly fine-tune the tool chain end-to-end under agent-generated plans so tools learn to cooperate.

Key results

• On AgenticIR Group C, Ours (Full) reaches 23.04 dB PSNR and 0.30 LPIPS, versus 19.77 dB and 0.43 for 4KAgent.
• Ours (Planning) alone is on par with or above prior agentic systems, e.g. 22.32 dB PSNR on Group A vs 21.48 for 4KAgent.
• On RTTS real haze, Ours (Full) achieves the best CLIP-IQA (0.463) but IPC-Dehaze has higher MUSIQ (59.60 vs 55.78); on LHP, Ours (Full) gets 30.93 dB PSNR, below FAD-Former (31.27).

Surprising findings

• In an exhaustive search study, 60.0% of high-performing plans contain at least one out-of-scope tool, and 77.6% apply some tool multiple times.
• Removing duplicate tools from high-performing plans raised average aggregated rank from 54.0 to 65.3.

Limitations

• Real-world benchmarks evaluated are dominated by single degradations, so the multi-tool design is not well tested on them; the paper itself notes this.
• Tool training memory scales roughly linearly with chain length, limiting feasible chain length (~12-14 estimated).
• Reward uses IQA metrics as a proxy for perceptual quality, which may not align with human judgment; no human evaluation reported.

Experimental setup

Datasets: MiOIR-Train MiOIR-Test AgenticIR Groups A/B/C RTTS LHP

Benchmarks: AgenticIR Groups A/B/C RTTS LHP MiOIR-Test

Metrics: PSNR SSIM LPIPS MANIQA CLIP-IQA MUSIQ

Baselines: AirNet PromptIR MiOIR DA-CLIP InstructIR AutoDIR AgenticIR MAIR 4KAgent GPT-5.4

Builds on: GRPO VisualQualityR1 AgenticIR MAIR 4KAgent Restormer

Closest papers by meaning

• An Intelligent Agentic System for Complex Image Restoration Problems (ICLR 2025 · similarity 0.88)
• Hybrid Agents for Image Restoration (CVPR 2026 · similarity 0.88)
• RestoreAgent: Autonomous Image Restoration Agent via Multimodal Large Language Models (NeurIPS 2024 · similarity 0.88)
• Complexity Experts are Task-Discriminative Learners for Any Image Restoration (CVPR 2025 · similarity 0.85)
• Continuous Expert Assembly: Instance-Conditioned Low-Rank Residuals for All-in-One Image Restoration (NeurIPS 2026 · similarity 0.85)
• IRPO: Boosting Image Restoration via Post-training GRPO (NeurIPS 2026 · similarity 0.84)