An Intelligent Agentic System for Complex Image Restoration Problems

26
citations
#676
in ICLR 2025
of 3827 papers
5
Top Authors
5
Data Points

Abstract

Real-world image restoration (IR) is inherently complex and often requires combining multiple specialized models to address diverse degradations. Inspired by human problem-solving, we propose AgenticIR, an agentic system that mimics the human approach to image processing by following five key stages: Perception, Scheduling, Execution, Reflection, and Rescheduling. AgenticIR leverages large language models (LLMs) and vision-language models (VLMs) that interact via text generation to dynamically operate a toolbox of IR models. We fine-tune VLMs for image quality analysis and employ LLMs for reasoning, guiding the system step by step. To compensate for LLMs' lack of specific IR knowledge and experience, we introduce a self-exploration method, allowing the LLM to observe and summarize restoration results into referenceable documents. Experiments demonstrate AgenticIR's potential in handling complex IR tasks, representing a promising path toward achieving general intelligence in visual processing.

Citation History

Jan 26, 2026
24
Feb 2, 2026
25+1
Feb 13, 2026
26+1
Feb 13, 2026
26
Feb 13, 2026
26