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

A research team has introduced IMAGAgent, a novel framework designed to solve the persistent problem of semantic drift and structural distortion in multi-turn AI image editing. Unlike current tools that execute edits in isolated steps, IMAGAgent implements a closed-loop 'plan-execute-reflect' mechanism. This allows the system to maintain context and adapt throughout a complex editing session, preventing the error accumulation that typically ruins multi-step projects.

The core of IMAGAgent is a three-stage pipeline. First, a constraint-aware planning module uses a vision-language model (VLM) to intelligently parse a user's complex natural language request (e.g., 'make the dog smaller, change its collar to red, and add a ball') into a sequence of executable, atomic sub-tasks. Second, a tool-chain orchestration module dynamically calls upon a suite of specialized AI models—for tasks like segmentation, detection, and inpainting—to execute each step, using the current image state and historical context to guide its choices.

Finally, and crucially, a multi-expert reflection mechanism acts as a quality control checkpoint. A central LLM synthesizes critiques from VLMs about the edited image, generating holistic feedback. This triggers fine-grained self-corrections and logs outcomes to inform future decisions, creating a learning loop. Tested on the new MTEditBench and MagicBrush datasets, IMAGAgent demonstrated significantly superior performance in instruction consistency and editing precision compared to existing methods, marking a substantial step toward reliable, conversational image editing.