IEA is a conversational image editing agent that maps natural language requests to sequences of tool calls on 16 parameterized editing tools. It is trained in three phases: supervised fine-tuning on distilled expert edit plans, GRPO reinforcement learning with rewards for likeness improvement, tool usefulness, and intent summarization, and large-scale synthetic fine-tuning that adds image refinement from follow-up feedback. This makes each edit explicit and inspectable rather than a black-box pixel generation.
The empirical results show that IEA reaches an average pixel distance of 0.1034 on the Image-Edit test set and a ROUGE-L of 0.2575 on Image-Summary, beating strong baseline models. In user studies it ranks highest among tool-calling methods for instruction following and surpasses generative methods in overall perceptual quality. The interest is that an interpretable tool-centric VLM can rival or beat generative approaches while giving users transparent edit traces.