Organizations: The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, Guangdong, China · Harbin Institute of Technology, Weihai, Shandong, China · Shenzhen University, Shenzhen, Guangdong, China · Claremont McKenna College, Claremont, California, USA · Research Institute of Petroleum Exploration and Development, CNPC, Beijing, China
Conditional image editing aims to modify a source image according to textual prompts and optional reference guidance. Such editing is crucial in scenarios requiring strict structural control (i.e., anomaly insertion in driving scenes and complex human pose transformation). Despite recent advances in large-scale editing models (i.e., Seedream, Nano Banana, etc), most approaches rely on single-step generation. This paradigm often lacks explicit quality control, may introduce excessive deviation from the original image, and frequently produces structural artifacts or environment-inconsistent modifications, typically requiring manual prompt tuning to achieve acceptable results. We propose CAMEO, a structured multi-agent framework that reformulates conditional editing as a quality-aware, feedback-driven process rather than a one-shot generation task. CAMEO decomposes editing into coordinated stages of planning, structured prompting, hypothesis generation, and adaptive reference grounding, where external guidance is invoked only when task complexity requires it. To overcome the lack of intrinsic quality control in existing methods, evaluation is embedded directly within the editing loop. Intermediate results are iteratively refined through structured feedback, forming a closed-loop process that progressively corrects structural and contextual inconsistencies. We evaluate CAMEO on anomaly insertion and human pose switching tasks. Across multiple strong editing backbones and independent evaluation models, CAMEO consistently achieves 20% more win rate on average compared to multiple state-of-the-art models, demonstrating improved robustness, controllability, and structural reliability in conditional image editing.
Figures & tables
Figure 1: Compare CAMEO with other State-of-the-Art image editing models
Figure 2: Representative failure cases illustrating common issues of conditional image editing on images from BDD100K under complex situations.
Figure 3: Overview of the CAMEO multi-agent workflow. The Strategic Director coordinates multiple agents to perform task interpretation, structured generation, quality evaluation, and iterative refinement.
Figure 4: Representative cases of how CAMEO improves semantic correctness and physical plausibility issues.
Figure 5: Representative cases of how CAMEO improves boundary blending and contextual coherence issues.
Table 1: Pairwise win/lose/tie statistics (%) on the road anomaly insertion task across multiple vision-language judges. Higher win rates indicate stronger performance of CAMEO over direct editing.
Table 2: Pairwise win/lose/tie statistics (%) on the human pose switching task across multiple vision-language judges. Higher win rates indicate stronger performance of CAMEO over direct editing.
Table 3: Average evaluation scores (1-10 scale) of CAMEO and direct editing baselines under four vision-language judges on the road anomaly insertion task. CAMEO consistently achieves higher scores across all judges.
Table 4: Average evaluation scores (1-10 scale) of CAMEO and direct editing baselines under four vision-language judges on the human pose switching task. CAMEO consistently achieves higher scores across most judges.
Figure 6: Qualitative comparison across methods on diverse human pose switching examples. Each input consists of an instruction and an original image. Instructions are: Row 1: a high front kick with the right leg extended forward, upper body leaning back for balance. Row 2: a running posture with the right leg stepping forward, left leg pushing back. Row 3: a side stretching pose bending the torso to the right with left arm reaching overhead. Row 4: a high front kick with the right leg extended forward, upper body leaning back for balance.
Baseline
Win
Lose
Tie
Net ↑
Qwen-Image-Edit-Plus
56.4
22.1
21.5
+34.3
Nano Banana Pro
39.8
28.6
31.6
+11.2
Flux-2-Pro
47.5
24.3
28.2
+23.2
Seedream-4.5
33.7
29.4
36.9
+4.3
Table 5: Human evaluation results (%) comparing CAMEO with direct editing baselines.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Qualitative comparison of the full CAMEO system and three ablation variants. Removing key components leads to degraded realism and quality of details.
Ablation Setting
Qwen3-VL-Plus
GPT-4o
Gemini-2.5
Claude-Opus-4.5
Full Model (CAMEO)
8.47
8.42
8.36
8.45
w/o Adaptive Reference Grounding
7.88
7.91
7.80
7.83
w/o Quality Control
7.05
7.18
7.10
6.96
w/o Iterative Refinement
7.96
7.99
7.92
7.94
Appendix
Table 6: Ablation study of CAMEO on the road anomaly insertion task. Average evaluation scores (1–10 scale) reported by four vision-language judges.
Ablation Setting
Qwen3-VL-Plus
GPT-4o
Gemini-2.5
Claude-Opus-4.5
Full Model (CAMEO)
8.39
8.33
8.27
8.35
w/o Adaptive Reference Grounding
7.83
7.86
7.74
7.79
w/o Quality Control
6.98
7.12
7.05
6.91
w/o Iterative Refinement
7.91
7.88
7.82
7.86
Appendix
Table 7: Ablation study of CAMEO on the human pose switching task. Average evaluation scores (1–10 scale) reported by four vision-language judges.
Figure 8: Screenshot of the human evaluation interface used in our study.
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.
Image editing models produce impressive results for individual edits, but their quality often deteriorates over multiple editing turns. Each editing turn introduces small distribution shifts that accumulate over successive edits, creating a gap between the input distributions encountered during the editor's original training and multi-turn inference. We introduce a training scheme that makes multi-turn image editing stable under this accumulating distribution shift. A key is to force the editor to operate on its own outputs from previous editing turns during training, bringing the principle of Self-Forcing to image editing. The main challenge is that self-generated editing trajectories deviate from ground-truth trajectories, making direct supervision unreliable. We address this challenge through a distillation objective that leverages a teacher image editor to provide a reliable training signal when the model operates on its own previous outputs. We introduce an asymmetric conditioning scheme: while the student performs edits sequentially on its own outputs, the teacher remains anchored to the original clean image and receives a compound editing instruction. Experiments show substantially reduced degradation over long multi-turn sequences while maintaining editing quality and instruction adherence.
Real-world e-commerce image editing often requires multiple, localized, and auditable operations rather than global restyling. This compositional nature poses a dual challenge: models must precisely apply all requested edits to the correct regions while preserving unmodified content, even under ambiguous instructions. Existing one-shot editors conflate intent resolution, spatial grounding, and synthesis into a single step, frequently resulting in partial execution failures, which is unacceptable for commercial scenarios. To address this, we introduce GMO-E2DIT, an agentic editing framework that couples a Vision-Language Model (VLM) with a mask-conditioned image editor to tackle structured multi-turn task completion. Given an underspecified instruction, the VLM agent constructs a region-grounded edit agenda, effectively decoupling cognitive reasoning from generative rendering. The framework then executes sub-programs via operation-aware masks and references, utilizing a reflection-driven loop to inspect intermediate results and determine the subsequent state. This iterative mechanism reliably preserves safe partial progress, retries unfinished operations, and recovers from errors. Furthermore, we develop a unified data pipeline providing aligned supervision for planning, execution, and reflection, alongside EComEditBench, a comprehensive benchmark for instruction-driven evaluation. Extensive experiments demonstrate that GMO-E2DIT achieves competitive performance compared to strong closed-source models, yielding superior instruction accuracy and edit fidelity over existing baselines.
Zipeng Guo, Xiaoan Liu, Lichen Ma +9
Wuhan University · Xi’an Jiaotong University · Sun Yat-sen University