Adobe's EditAgent judge picks a usable edit 81.3% of the time, versus 74.8% for EditScore

Researchers from Adobe Research and the University of California, San Diego have introduced EditAgent, an image-editing agent that works more like a Photoshop professional than earlier systems. It splits a request into dependent steps, picks a suitable tool for each step, and compares the result with the original using measured data. Its main result is in the judge role. On a benchmark of 600 edits, where five professionals rated results blind, the off-the-shelf Qwen3.8-27B, with no fine-tuning, picked a candidate that the professionals also deemed usable in 81.3% of cases. The best of the purpose-trained models, EditScore-72B (fine-tuned on human ratings of edits), reached 74.8%.

Why the general-purpose editor falls short

Most image-editing agents are built simply: a complex request is cut into steps, a failed step is retried with the same editor, and a separate model judges the final result by eye. The authors show that this design hits a ceiling. In a test of 600 edits, the general-purpose editor Flux-2-Klein-9B failed to produce a usable result in 42.7% of cases. Narrow tools, for sharpening, rotation and zooming, fixed roughly two thirds of those failures. Retrying with the same editor did almost nothing.

The second finding concerns the order of work. Iterative refinement accumulates small errors, because each round builds on an already altered image. Generating variants in parallel from the original image worked better: in comparisons where professionals picked a side, they preferred the parallel variants in 78.5% of cases. The agent also needed half as many evaluation passes. EditAgent is built on these two findings.

A judge that gets to see the numbers

For each step, the agent pits the general-purpose editor against specialist tools and gets several candidates. A plain LLM then compares all the variants in a single pass. But it doesn't look only at the whole image. For each variant, the system assembles a verification record: a close-up of the edited region, a map of the changes, and measured brightness and color statistics. In the paper's example, a photo needed to be lightened slightly. The numbers showed that the general-purpose editor overexposed the shot and shifted the colors toward warm tones, while a specialist tool stayed within the requested limits. A model looking at the full frame misses errors like this.

For comparison, the researchers used models like EditScore, which are fine-tuned on human ratings of edits. Without the verification record, none of the seven LLMs tested matched the best fine-tuned model. With it, all seven beat every one of them. The authors conclude that what the judge sees can matter more than how it was trained.

After each round, the agent also checks whether later edits have undone earlier ones. In the paper's example, a request to "let in more daylight" made a window transparent again after it had previously been frosted. The agent notices the conflict and fixes it.

A tenfold cost and training data

EditAgent improved the open editors Qwen-Image-Edit-2511 and Flux-2-Klein-9B on three benchmarks and outperformed seven other agents. On the EdiVal benchmark, the Qwen-based variant slightly edged out the commercial Seedream 4.0 and GPT-Image-1.5 in the third round of edits. The authors state the limitations plainly: the agent needs roughly ten times more compute than a single editor, and its quality depends on the set of connected tools.

Part of that cost may pay for itself. The agent keeps only verified results, so its runs produced about 26,000 training examples. Editors fine-tuned on this data improved on all three benchmarks. Data generated by FLUX itself, by contrast, made FLUX worse. An expensive agent can thus serve as a data factory for fast models.

For Adobe, this is not an academic question. In June, the company added AI agents to Photoshop, Premiere and other Creative Cloud apps as a public beta: the user describes a task, and the agent carries out the steps. Choosing a tool and checking one's own output is exactly where such products succeed or fail. One caveat: Black Forest Labs has already replaced the model used in the tests. The new Flux 3 Image Klein has 5.6 billion parameters, fewer than Flux 2 Klein, but in the company's internal tests it outperforms all other open models and even the commercial Flux 2 Pro.