ImageEdit-R1: Boosting Multi-Agent Image Editing via Reinforcement Learning

Yiran Zhao; Yaoqi Ye; Xiang Liu; Michael Qizhe Shieh; Trung Bui

cs.CVMar 9, 2026

ImageEdit-R1: Boosting Multi-Agent Image Editing via Reinforcement Learning

Authors: Yiran Zhao, Yaoqi Ye, Xiang Liu, Michael Qizhe Shieh, Trung Bui

Organizations: National University of Singapore · Adobe Research

Abstract

With the rapid advancement of commercial multi-modal models, image editing has garnered significant attention due to its widespread applicability in daily life. Despite impressive progress, existing image editing systems, particularly closed-source or proprietary models, often struggle with complex, indirect, or multi-step user instructions. These limitations hinder their ability to perform nuanced, context-aware edits that align with human intent. In this work, we propose ImageEdit-R1, a multi-agent framework for intelligent image editing that leverages reinforcement learning to coordinate high-level decision-making across a set of specialized, pretrained vision-language and generative agents. Each agent is responsible for distinct capabilities--such as understanding user intent, identifying regions of interest, selecting appropriate editing actions, and synthesizing visual content--while reinforcement learning governs their collaboration to ensure coherent and goal-directed behavior. Unlike existing approaches that rely on monolithic models or hand-crafted pipelines, our method treats image editing as a sequential decision-making problem, enabling dynamic and context-aware editing strategies. Experimental results demonstrate that ImageEdit-R1 consistently outperforms both individual closed-source diffusion models and alternative multi-agent framework baselines across multiple image editing datasets.

Figures & tables

Figure 1: Overview of ImageEdit-R1 : \scriptsize{1}⃝ The decomposition agent analyzes the user instruction and input image to extract a structured representation of the desired edits, including editing actions, subjects, and goals. \scriptsize{2}⃝ The sequencing agent arranges these components into an ordered list of sub-requests, enabling interpretable and modular execution. \scriptsize{3}⃝ The editing agent , built on a diffusion model, performs the actual image edits by sequentially applying the sub-requests.

Figure 2

Edit Model	Method	PSR		RealEdit		UltraEdit		Average
Edit Model	Method	GPT-4o	Gemini-2.5	GPT-4o	Gemini-2.5	GPT-4o	Gemini-2.5	Average
Single Model
Single -DiT-model	Step1X-Edit	6.70	6.58	6.83	6.92	7.51	7.67	7.04
	ILLUME+	6.00	6.22	6.37	6.50	7.28	6.89	6.54
	ICEdit	5.64	5.73	5.75	6.26	7.44	7.13	6.33
Close -source-model	GPT-4o	9.12	7.96	8.35	8.19	8.71	8.48	8.47

Table 1: Main results of ImageEdit-R1 with evaluation range of 0 to 10. “ImageEdit-R1 (w/ RL)” refers to the variant without RL for the decomposition agent. The “Average” column reports the mean score across all three benchmarks and two evaluators. Relative improvements over the original model are shown in superscript. Bold indicates the best performance, while underline represent the second-best.

You are an expert image quality evaluator. You will be shown two images:

1. The ORIGINAL image (first image)

2. The EDITED image (second image)

Edit Request: "{edit_request}"

Please evaluate how well the edited image fulfills the edit request compared to the original image.

Evaluation Criteria:

1. Request Fulfillment (0–4 points): Did the edit successfully accomplish what was requested?

Table 2: Evaluation prompts when using LLM-as-a-Judge.

Panel of Figure 2 — Figure 2: Representative examples demonstrating the performance of ImageEdit-R1 compared to baselines on complex editing tasks.

Figure 3: Training rewards of decomposition agent on Qwen2.5-VL-7B-Instruction.

Method	Fulfillment (0–4)	Quality (0–3)	Preservation (0–3)
Human	2.84	2.38	2.73
VLM-based	3.43	2.41	2.26
Correlation	0.66	0.43	0.59

Figure 4: Human vs. VLM-based evaluation. Top rows show average scores; bottom row shows their correlation.

Table 8

Figure 5: Comparison of multi-turn and single-turn strategies in the image editing process.

Explore similar work

IMAGAgent: Orchestrating Multi-Turn Image Editing via Constraint-Aware Planning and Reflection
Feb 12, 2026Fei Shen, Chengyu Xie, Lihong Wang +5Image Editing Tool-Using Vision-Language Agents
National University of Singapore · Nanjing University of Science and Technology · Nanjing Forestry University
MSRAMIE: Multimodal Structured Reasoning Agent for Multi-instruction Image Editing
Mar 17, 2026Zhaoyuan Qiu, Ken Chen, Xiangwei Wang +3Image Editing Multimodal Agents
University Of Melbourne, Grattan Street, Parkville, Victoria 3010, Australia
Leveraging Verifier-Based Reinforcement Learning in Image Editing
Apr 30, 2026Hanzhong Guo, Jie Wu, Jie Liu +6Reward Modeling Reinforcement Learning
School of Computing and Data Science, The University of Hong Kong · ByteDance Seed · CUHK +1

Feb 12, 2026J/K move · Enter open · S save

IMAGAgent: Orchestrating Multi-Turn Image Editing via Constraint-Aware Planning and Reflection

Fei Shen, Chengyu Xie, Lihong Wang, Zhanyi Zhang+4

National University of Singapore · Nanjing University of Science and Technology · Nanjing Forestry University

Image Editing Tool-Using Vision-Language Agents Text-Guided Image Editing LLM Agent Orchestration

Open paper AskPDF arXiv

Existing multi-turn image editing paradigms are often confined to isolated single-step execution. Due to a lack of context-awareness and closed-loop feedback mechanisms, they are prone to error accumulation and semantic drift during multi-turn interactions, ultimately resulting in severe structural distortion of the generated images. To address this, we propose IMAGAgent, a multi-turn image editing agent framework based on a ``plan--execute--reflect'' closed-loop mechanism that achieves deep synergy among instruction parsing, tool scheduling, and adaptive correction within a unified pipeline. Specifically, we first present a constraint-aware planning module that leverages a vision-language model (VLM) to precisely decompose complex natural language instructions into a series of executable sub-tasks, governed by target singularity, semantic atomicity, and visual perceptibility. Then, the tool-chain orchestration module dynamically constructs execution paths based on the current image, the current sub-task, and the historical context, enabling adaptive scheduling and collaborative operation among heterogeneous operation models covering image retrieval, segmentation, detection, and editing. Finally, we devise a multi-expert collaborative reflection mechanism in which a central large language model (LLM) synthesizes evaluations from multiple vision-language models into holistic feedback, triggering fine-grained self-correction and recording feedback outcomes to optimize future decisions. Extensive experiments on our constructed MTEditBench and the MagicBrush dataset demonstrate that IMAGAgent achieves performance significantly superior to existing methods in terms of instruction consistency, editing precision, and overall quality. The code is available at https://github.com/hackermmzz/IMAGAgent.

Abstract

Explore similar work

IMAGAgent: Orchestrating Multi-Turn Image Editing via Constraint-Aware Planning and Reflection

MSRAMIE: Multimodal Structured Reasoning Agent for Multi-instruction Image Editing

Leveraging Verifier-Based Reinforcement Learning in Image Editing