Research map / Vision-language and multimodal / Multimodal LLMs
RetouchAgent: Towards Interactive and Explainable Image Retouching with MLLM Agents
Shuo Zhang, Xinyu Yang
Read the paperCheck your idea against 137,217 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 RetouchAgent, an MLLM-agent framework for interactive, explainable image retouching driven by natural language instructions.
Problem
Deep learning image retouching is subjective and black-box, so existing methods lack interactivity and explainability; parameter-controlled methods have ambiguous semantics and no natural-language control.
Approach
Uses multimodal LLM agents to interpret natural-language retouching requests and explain the edits, aiming to combine interactivity with explainability beyond parameter-controlled and RL-based approaches.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add text-intent retrieval (lambda_text 0.4) to image-only retrieval (ablation) | RetouchAgent · MIT-Adobe FiveK Expert C (480p) | PSNR: 23.50 -> 26.02 | Improved |
| increase text weight beyond 0.4 to 0.8 (ablation) | RetouchAgent · MIT-Adobe FiveK Expert C (480p) | PSNR: 26.02 -> 25.48 | Hurt |
| use 3 ICL examples instead of fewer or more (ablation) | RetouchAgent · MIT-Adobe FiveK (5 experts avg, 480p) | PSNR | Improved |
| add system prompt constraints and examples (ablation) | RetouchAgent · MIT-Adobe FiveK (full dataset) | success rate: 26.8% -> 99.4% | Improved |
| use the designed operation library instead of GMIC (ablation) | RetouchAgent · MIT-Adobe FiveK Expert C (480p) | PSNR: 20.21 -> 26.02 | Improved |
| use the designed operation library instead of OpenCV (ablation) | RetouchAgent · MIT-Adobe FiveK Expert C (480p) | PSNR: 21.65 -> 26.02 | Improved |
| add curve operations to the operation library (ablation) | RetouchAgent · MIT-Adobe FiveK Expert C (480p) | PSNR: 23.12 -> 26.02 | Improved |
| add target image analysis to CoT (ablation) | RetouchAgent · MIT-Adobe FiveK Expert C (480p) | PSNR: 24.32 -> 26.02 | Improved |
| add expert example analysis to CoT (ablation) | RetouchAgent · MIT-Adobe FiveK Expert C (480p) | PSNR: 24.16 -> 26.02 | Improved |
| add difference analysis to CoT (ablation) | RetouchAgent · MIT-Adobe FiveK Expert C (480p) | PSNR: 23.82 -> 26.02 | Improved |
| add reflection agent (ablation) | RetouchAgent · MIT-Adobe FiveK Expert C (480p) | PSNR: 21.35 -> 26.02 | Improved |
| use full reflection history instead of only previous round (ablation) | RetouchAgent · MIT-Adobe FiveK Expert C (480p) | PSNR: 24.76 -> 26.02 | Improved |
Closest papers by meaning
- RetouchIQ: MLLM Agents for Instruction-Based Image Retouching with Generalist Reward (CVPR 2026 · similarity 0.92)
- JarvisArt: Liberating Human Artistic Creativity via an Intelligent Photo Retouching Agent (NeurIPS 2025 · similarity 0.87)
- Guiding Instruction-based Image Editing via Multimodal Large Language Models (ICLR 2024 · similarity 0.86)
- RetouchGPT: LLM-based Interactive High-Fidelity Face Retouching via Imperfection Prompting (AAAI 2025 · similarity 0.85)
- VisionThink: Smart and Efficient Vision Language Model via Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
- The Promise of RL for Autoregressive Image Editing (NeurIPS 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.
