Research map / Multi-agent systems and games / LLM agents
Hybrid Agents for Image Restoration
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 HybridAgent, an interactive image restoration system with fast (Llama3.2-1B), slow (fine-tuned Co-Instruct MLLM) and feedback agents plus a three-stage LoRA-based tool suite including a mixed-distortion removal tool.
Problem
Existing image restoration agents rely on users selecting modes, process hybrid distortions step by step causing error propagation, and use heavyweight MLLMs even for simple, explicit requests.
Approach
Combine a lightweight LLM fast agent for explicit prompts with an MLLM slow agent for ambiguous prompts and an MLLM feedback agent that judges whether the image is clean, backed by restoration tools trained via foundation multi-task pretraining, task-specific LoRA with prompt reinitialization, and a mixed-distortion LoRA tool.
Key results
- On mixed distortions, HybridAgent reaches 26.49 PSNR / 0.766 SSIM on Rainstreak+Noise versus 23.17 / 0.591 for PromptIR.
- Mixed distortion removal averages 24.83 PSNR / 0.741 SSIM over the mixed test set versus 19.84 / 0.557 for step-by-step single-distortion tools.
- FastAgent route takes roughly 12% of SlowAgent's runtime for direct prompts (e.g., 0.08s vs 0.75s on denoising).
Surprising findings
- Step-by-step single-distortion removal for haze and low-light can be worse than mixed-distortion removal by large margins, e.g., low-light+JPEG 7.05 vs 21.65 PSNR averages.
- FastAgent success rate on VVC prompts (60.0%) and low-light (73.3%) is much lower than SlowAgent's, showing the lightweight router is less reliable on some tasks.
Limitations
- Mixed distortion removal tool has limited representation capacity for more complex or real-world distortions.
- FastAgent success rate is noticeably lower on some distortion types such as VVC and low-light.
Experimental setup
Datasets: DF2K Rain100H RainDrop RESIDE-6k LOL CBSD68 Urban100 Kodak24 McMaster EUVP
Benchmarks: CBSD68 Urban100 Kodak24 McMaster Rain100H RainDrop RESIDE-6k LOL EUVP
Metrics: PSNR SSIM LPIPS average inference time tool invocation success rate
Baselines: SwinIR Uformer Restormer AirNet PromptIR InstructIR
Builds on: PromptIR Co-Instruct LoRA Real-ESRGAN AgenticIR
Closest papers by meaning
- RestoreAgent: Autonomous Image Restoration Agent via Multimodal Large Language Models (NeurIPS 2024 · similarity 0.92)
- An Intelligent Agentic System for Complex Image Restoration Problems (ICLR 2025 · similarity 0.91)
- OPERA: An Agent for Image Restoration with End-to-End Joint Planning–Execution Optimization (NeurIPS 2026 · similarity 0.88)
- Parameter Efficient Adaptation for Image Restoration with Heterogeneous Mixture-of-Experts (NeurIPS 2024 · similarity 0.86)
- DreamClear: High-Capacity Real-World Image Restoration with Privacy-Safe Dataset Curation (NeurIPS 2024 · similarity 0.85)
- Complexity Experts are Task-Discriminative Learners for Any Image Restoration (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.