AI image editing depends on more than understanding a request. An agent also has to choose suitable tools, judge whether an edit worked, and make sure one change does not undo another. EditAgent, developed by Adobe Research and UC San Diego, is designed to handle those steps by generating alternatives and checking them against evidence.
Why one editor may not be enough
In a test of 600 image edits, five professionals rated results without knowing which system made them. The general-purpose editor Flux-2-Klein-9B failed to produce a usable result 42.7 percent of the time. Specialized tools resolved roughly two-thirds of those failures, particularly for sharpening, rotation, and zooming.
Simply asking the same editor to try again offered little help. That points to a practical limitation of retry-based editing: when a tool is poorly suited to a task, repeating the attempt may not fix the underlying mismatch. EditAgent instead plans the request in steps and brings in tools suited to each one.
The researchers also compared sequential refinement with generating several alternatives from the original image. With iterative editing, each new pass starts from an image already changed by earlier passes, allowing small errors to accumulate. When professionals chose between results, they preferred parallel variants in 78.5 percent of comparisons. This approach also required half as many evaluation passes.
A judge that can inspect evidence
EditAgent asks a standard language model to compare candidate edits in one pass. For each option, the model receives a verification record: a close-up of the edited area, a map showing which regions changed, and measurements of brightness and color.
Those details can make a subtle failure easier to spot. In one example, the request was to brighten a photo gently. Measurements showed that the general-purpose editor had overexposed the image and made its colors warmer, while a specialized tool stayed within the requested bounds.
The team tested whether this evidence helped the judge select images that professionals had also deemed usable. Without a verification record, none of the seven language models they evaluated matched the best fine-tuned evaluation model. With the record, all seven outperformed every fine-tuned model in the comparison.
Qwen3.8-27B selected a usable result 81.3 percent of the time, compared with 74.8 percent for EditScore-72B. The finding suggests that the information provided to an evaluator can matter as much as how the evaluator was trained.
Checking that edits still fit together
A multi-step request can create conflicts: a later change may undo an earlier one. EditAgent checks for those regressions after each round. In the paper’s example, the instruction “let in more daylight” causes a window that was previously made frosted to become clear again. The agent identifies the conflict and corrects it.
That kind of check matters because a series of individually plausible edits may not produce a coherent final image. By reviewing the effect of each step in context, the system can look for whether earlier instructions remain satisfied as the task progresses.
Across three benchmarks, EditAgent improved the open-source editors Qwen-Image-Edit-2511 and Flux-2-Klein-9B, and outperformed seven other AI agents. On EdiVal, the Qwen version in its third editing round edged past Seedream 4.0 and GPT-Image-1.5.
More compute now, potential training value later
The approach has a cost: EditAgent uses about ten times as much compute as a single editor, and its results depend on which tools are available to it. That makes tool selection and evaluation part of the system’s performance, rather than details that can be ignored.
The researchers also used the agent to create training material. Because it keeps results that have been verified, EditAgent produced about 26,000 examples. Editors fine-tuned on those examples performed better across all three benchmarks, while training data generated directly by FLUX made FLUX worse.
That creates a possible path from a compute-intensive agent to faster editors trained on its checked outputs. The article also notes that Adobe added AI agents to Photoshop, Premiere, and other Creative Cloud apps in June, starting as a public beta. For those products, the same questions apply: can an agent choose the right editing tool and tell whether its output meets the request?