Research map / Vision-language and multimodal / Multimodal LLMs

An Intelligent Agentic System for Complex Image Restoration Problems

ICLR 2025 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 1 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 AgenticIR, an LLM/VLM agent that perceives, schedules, executes, reflects and reschedules (with rollback) over a toolbox of single-degradation IR models, using a self-explored experience base to order tools.

Problem

Real-world image restoration involves complex mixed degradations that single models or all-in-one models handle poorly, and the order and choice of restoration tools strongly affects results while LLMs lack reliable knowledge of IR tool interactions.

Approach

An agentic loop in which a fine-tuned VLM (DepictQA) perceives and verifies degradations, GPT-4 schedules tool order using a knowledge base distilled from self-exploration success/fail statistics, and a DFS-style rollback reschedules failed steps.

Key results

• AgenticIR achieves PSNR 21.04 / MUSIQ 56.88 on Group A, outperforming all-in-one baselines such as AutoDIR (19.64 / 47.01) and MiOIR (20.84 / 47.82).
• Versus random tool invocation, AgenticIR improves Group A MUSIQ from 55.30 to 56.88 and Group C MUSIQ from 44.61 to 48.68.
• Experience-based scheduling makes GPT-4's order consistent; the presentation-order bias difference is reported as zero for Groups A and B with experience.

Measured effects reported in the paper

Surprising findings

• Without experience, GPT-4's scheduling output changed when the order of the same two tasks was merely presented differently, showing order sensitivity.
• On Group C, ablating reflection gave slightly higher PSNR (18.93 vs 18.85), and random invocation slightly beat AgenticIR on PSNR (18.87 vs 18.82).

Limitations

• Evaluated only on synthetic mixed degradations from eight types, with a small set of real-world examples shown qualitatively.
• Relies on closed-source GPT-4 (gpt-4-1106-preview) for scheduling, and DFS rescheduling can be costly.
• Blur types (defocus vs motion) are hard to distinguish, limiting DepictQA recall for them.

Experimental setup

Datasets: DIV2K Flickr2K MiO100

Benchmarks: MiO100 synthetic mixed-degradation test set (Groups A, B, C) real-world under-display camera, underwater, and internet rain images (qualitative)

Metrics: PSNR SSIM LPIPS MANIQA CLIP-IQA MUSIQ precision recall F1 scheduling entropy

Baselines: AirNet PromptIR MiOIR DA-CLIP InstructIR AutoDIR Random tool invocation

Builds on: DepictQA LoRA Tree of Thoughts HuggingGPT Reflexion

Closest papers by meaning

• Hybrid Agents for Image Restoration (CVPR 2026 · similarity 0.91)
• RestoreAgent: Autonomous Image Restoration Agent via Multimodal Large Language Models (NeurIPS 2024 · similarity 0.90)
• OPERA: An Agent for Image Restoration with End-to-End Joint Planning–Execution Optimization (NeurIPS 2026 · similarity 0.88)
• Complexity Experts are Task-Discriminative Learners for Any Image Restoration (CVPR 2025 · similarity 0.86)
• Simple Agents Outperform Experts in Biomedical Imaging Workflow Optimization (CVPR 2026 · similarity 0.85)
• EpiAgent: An Agent-Centric System for Ancient Inscription Restoration (CVPR 2026 · similarity 0.84)