What they did

The authors constructed IEA as a vision-language model that selects and applies image editing tools (e.g., brightness, crop, color balance) from a predefined set of 16. Training proceeded in three stages: (1) supervised fine-tuning on expert demonstrations distilled from a larger model, (2) GRPO (GrouP Relative Policy Optimization) with rewards for image-likeness improvement, tool usefulness, and intent summarization, and (3) large-scale synthetic fine-tuning to jointly master editing, refinement, and user intent summarization.

Evaluation used a dataset of edit instructions paired with original and target images. The authors compared IEA against both tool-calling baselines and generative models (including text-to-image and instruction-following diffusion models). Metrics included pixel-level distance (mean squared error), ROUGE-L for summary quality, and human ratings for instruction following and perceptual quality.