Adobe Research and UC San Diego have built EditAgent, an image-editing agent that pairs a general-purpose editor with specialized tools and selects results with measured data. In a 600-edit test, the base editor Flux-2-Klein-9B failed to produce a usable result in 42.7 percent of cases. For business users of creative automation, the point is practical: tool choice and verification may matter more than a larger model.
How EditAgent plans, compares and verifies edits
Typical editing agents split a request into steps and retry failed steps with the same editor, while a separate model judges output visually. EditAgent works differently: it plans a request in steps and, for each step, runs the general editor against matching specialized tools in parallel. A standard language model then compares all variants in one pass and keeps the best one for the next stage.
The judge does not rely on eyeballing alone. Each candidate comes with a verification record containing a close-up of the edited area, a change map of altered regions, and measured statistics on brightness and color. In one example, a photo needed gentle brightening: the numbers showed the general editor overexposed it and shifted colors warmer, while the specialized tool stayed within specification. After each round the agent also checks whether later edits undid earlier ones.
The design rests on two findings from blind ratings by five professionals. Specialized tools solved roughly two-thirds of the general editor failures, especially for sharpening, rotation and zooming, while more retries with the same editor barely helped. Parallel variants made from the original beat step-by-step refinement in 78.5 percent of comparisons where raters picked a side, because iterative work piles small errors onto an already altered image and needed twice as many evaluation passes.
What this means for automated creative work
For companies that automate image production, the approach changes the mix of cost, quality and speed. EditAgent improved the open-source editors Qwen-Image-Edit-2511 and Flux-2-Klein-9B across three benchmarks and outperformed seven other AI agents. On the EdiVal benchmark, the Qwen variant in its third editing round edged past commercial models Seedream 4.0 and GPT-Image-1.5. Small teams gain access to near-commercial quality from open models, while large content operations can standardize tool selection per task.
The limits are compute, tooling and oversight. The agent needs about ten times as much compute as a single editor, and performance depends on which tools are plugged in. Without the verification record, none of seven tested language models matched the best fine-tuned judge; with it, all seven beat every fine-tuned model, with Qwen3.8-27B making a usable pick 81.3 percent of the time against 74.8 percent for EditScore-72B. Buyers should ask which tools are included, how conflicts are detected, and how verification data is logged.
A marker to watch is whether verified agent output becomes training data for faster editors. EditAgent kept only verified results and produced about 26,000 training examples; editors fine-tuned on that data improved on all three benchmarks, while data generated directly by FLUX made FLUX worse. Adobe added agents to Photoshop, Premiere and other Creative Cloud apps in June as a public beta, so progress in tool choice will show up in shipping products.
