Adobe Research and UC San Diego have developed EditAgent, an AI system designed to test how effectively artificial intelligence agents can utilize image editing tools like a professional Photoshop user. The researchers built the system to address the limitations of general-purpose editors, which failed to produce usable results in over forty percent of test cases during a 600-image evaluation. By breaking complex editing tasks into sub-steps and pairing a general editor with specialized tools for functions like sharpening, rotation, and zooming, EditAgent improves accuracy and mimics professional workflows.
The system relies on parallel generation and a novel verification method to judge image edits. Instead of using iterative refinement, which accumulates small errors across steps, EditAgent generates multiple variants in parallel from the original image. A standard language model then evaluates these variants using verification records that include close-up crops, change maps, and measured statistics on brightness and color. According to the research, this data-driven approach allowed standard language models to outperform specialized evaluation models that were fine-tuned on human ratings.
EditAgent successfully improved open-source editors like Qwen-Image-Edit-2511 and Flux-2-Klein-9B across multiple benchmarks, outperforming seven other AI agents. While the system requires about ten times the compute of standard methods, the verified output can be utilized to train faster editing models in the future. This research matters because it demonstrates a reliable method for enhancing the precision and practical utility of AI-driven image editing workflows.