CAMEO: A Conditional and Quality-Aware Multi-Agent Image Editing Orchestrator
| Authors | Yuhan Pu et al. |
| Year | 2026 |
| Field | Computer Vision |
| arXiv | 2604.03156 |
| Download | |
| Categories | cs.CV |
Abstract
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 \textbf{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.
Engineering Breakdown
Plain English
CAMEO addresses a critical limitation in modern image editing: single-step generation models produce structural artifacts, environment-inconsistent modifications, and require manual prompt tuning to achieve acceptable results. The paper proposes a multi-agent orchestration framework that reformulates conditional image editing (modifying images according to text prompts and reference guidance) as a structured, quality-aware process rather than a single monolithic generation step. This is particularly important for high-stakes scenarios like anomaly insertion in autonomous driving scenes or complex human pose transformation, where structural control is essential. The key insight is that treating editing as a multi-stage coordination problem—with explicit quality checking between stages—reduces artifacts and environment inconsistencies compared to existing approaches like Seedream and Nano Banana.
Core Technical Contribution
CAMEO's core novelty is reformulating conditional image editing from a single-stage generation paradigm into a multi-agent orchestration framework with explicit quality control loops. Rather than treating editing as one end-to-end transformation, the system decomposes the task into coordinated sub-problems where different agents handle different aspects (structural guidance, content modification, environmental consistency) and validate outputs between steps. This introduces a feedback mechanism that catches and corrects artifacts before they propagate through the pipeline, addressing a fundamental weakness in large-scale editing models. The framework is condition-aware (leveraging both text prompts and reference images) and quality-aware (implementing quality metrics to gate stage transitions), which represents a departure from prior work's implicit assumption that a single forward pass can produce acceptable results.
How It Works
CAMEO operates as a choreographed pipeline of specialized agents coordinated by an orchestrator. The input consists of a source image, a textual editing instruction, and optionally a reference image for guidance. The orchestrator receives these inputs and sequences multiple agent calls: an initial structure-preservation agent analyzes the source image to identify critical structural elements; a content modification agent applies the textual edits while respecting structural constraints; an environment consistency agent verifies that modifications align with the image's spatial and semantic context; and a quality assessment agent evaluates whether the output meets predefined quality metrics (artifact-free, structurally sound, environmentally coherent). If quality checks fail, the framework can route back to earlier stages with corrective prompts or adjusted parameters, creating a feedback loop. Only when all quality gates pass does the framework output the final edited image, effectively converting a single-pass process into a multi-step refinement process with explicit validation between each step.
Production Impact
For engineers building image editing systems, CAMEO offers a concrete solution to a persistent production pain point: reducing manual review and prompt engineering overhead. In autonomous driving pipelines requiring synthetic anomaly insertion (testing object detection under occlusion, for example), the quality gates eliminate the need for human reviewers to inspect and reject images with structural artifacts—the system catches these automatically. The multi-agent architecture also creates modularity: individual agents can be tuned or replaced independently, and quality criteria can be updated without retraining the entire pipeline. However, there are real trade-offs: orchestrating multiple agents increases end-to-end latency (multiple forward passes instead of one) and computational cost (proportional to the number of agents and feedback loops). Integration complexity increases because you need well-defined quality metrics and potentially a scheduler to manage agent coordination. For latency-critical applications (real-time editing in creative tools), the multi-stage approach may be prohibitive; for batch processing scenarios (automated dataset generation, synthetic data creation) where quality is paramount, the overhead becomes acceptable.
Limitations and When Not to Use This
The paper does not address several critical production scenarios: it assumes well-defined, measurable quality metrics can be formulated for the editing domain (this breaks down in subjective editing tasks where artistic intent matters more than artifact-free-ness). The framework's effectiveness likely depends heavily on the quality of the individual agents and the orchestration logic—if agents are weak or quality gates are poorly calibrated, the multi-stage approach becomes overhead without benefit. Computational cost and latency are mentioned but not quantified; it's unclear how many feedback loops are typical or what the actual speedup/slowdown is compared to single-stage baselines. The paper focuses on specific high-control scenarios (driving scenes, pose transformation) and doesn't demonstrate generalization to arbitrary conditional editing tasks. Failure modes under distribution shift (editing images from domains unseen during agent training) are not explored, and it's unclear how the framework behaves when quality gates are deadlocked (e.g., no agent output passes validation).
Research Context
CAMEO builds on the recent wave of large-scale conditional image generation models (Seedream, Nano Banana, and earlier work in diffusion-based editing) but identifies a fundamental limitation: single-stage generation lacks quality control mechanisms. It aligns with broader trends in AI systems research emphasizing agentic reasoning and multi-step problem decomposition (similar to chain-of-thought reasoning in language models, but applied to vision). The work implicitly acknowledges that scaling model size alone doesn't solve structural consistency—instead, architectural composition and explicit validation loops provide better guarantees. This opens a research direction around quality-aware generative systems where models are optimized not just for output realism but for navigating quality-condition spaces, and where orchestration logic becomes as important as the underlying models themselves.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::