An Intelligent Agentic System for Complex Image Restoration Problems

Kaiwen Zhu^*, Jinjin Gu^*, Zhiyuan You, Yu Qiao, Chao Dong

ICLR 2025

Paper | Project Page

Overview

Learning from exploration

Workflow

Examples

Restoration of real-world images

Restore a UDC image (from this work) by motion deblurring, defocus deblurring, and low light enhancement.

Restore an underwater image (from this work) by defocus deblurring, dehazing, and motion deblurring.

Effectiveness of planning with experience

Effectiveness of workflow designs

Installation

Please refer to INSTALL.md.

Usage

Setup

• Fill in the API key in config.yml.
• Run python src/app_eval.py and python src/app_comp.py in the directory DepictQA.

Data preparation

To generate complexly degraded images, run python -m dataset.synthesize. You should place clean images in dataset/HQ/ and corresponding depth maps in dataset/depth/. In the paper we use the MiO100 dataset. The degradation combinations are listed in dataset/degradations.txt. You can customize combinations in dataset/degradations.txt or degradation types in dataset/add_single_degradation.py.

The data used in the paper can be downloaded from this link.

Learning

To let the agent learn from exploration, run

• python -m exploration.exhaust_seq to generate images to explore;
• python -m exploration.explore to accumulate experience by evaluating images;
• python -m exploration.distill to summarize the experience and distill knowledge.

Inference

Run python -m pipeline.infer to restore an image (path specified in pipeline/infer.py).

BibTex

@inproceedings{agenticir,
      title={An Intelligent Agentic System for Complex Image Restoration Problems},
      author={Kaiwen Zhu and Jinjin Gu and Zhiyuan You and Yu Qiao and Chao Dong},
      booktitle={The Thirteenth International Conference on Learning Representations},
      year={2025},
      url={https://openreview.net/forum?id=3RLxccFPHz}
}