EditAgent: Adobe, UC San Diego agent beats trained edit judges with measured image data

Adobe Research and UC San Diego have built EditAgent, an image editing agent in which an unmodified language model judges edits using measured statistics and close-up crops. According to a new paper on arXiv, all seven standard language models tested outperformed every fine-tuned evaluation model once they received this data. The paper was covered by The Decoder.

How EditAgent works

EditAgent plans a request as a series of dependent steps. For each step, it pits a general-purpose editor against matching specialized tools, such as those for sharpening, rotation and zooming. It generates the candidates in parallel and has a language model compare all of them in a single pass, instead of retrying the same editor repeatedly.

Each candidate comes with a verification record: a close-up of the edited area, a change map of altered regions, and measured statistics on brightness and color. In one paper example, a photo was meant to be gently brightened. The numbers showed the general-purpose editor overexposed it and shifted the colors warmer, while the specialized tool stayed within spec.

After each round, the agent also checks whether later edits undid earlier ones. In the paper's example, the instruction "let in more daylight" made a previously frosted window clear again, and the agent caught and fixed the conflict.

Results

The researchers built a test of 600 image edits, blind-rated by five professionals.

• The general-purpose editor Flux-2-Klein-9B failed to produce a usable result 42.7 percent of the time.
• Specialized tools solved roughly two-thirds of those failures, especially for sharpening, rotation and zooming. More retries with the same editor barely helped.
• Professionals preferred parallel variants generated from the original image in 78.5 percent of the comparisons where they picked a side. Iterative refinement accumulates small errors because each round builds on an already altered image. The parallel approach needed half as many evaluation passes.
• As judges, none of the seven language models tested could match the best fine-tuned model without the verification record. With it, all seven beat every fine-tuned model. Qwen3.8-27B made a usable pick 81.3 percent of the time, against 74.8 percent for EditScore-72B.

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 the commercial models Seedream 4.0 and GPT-Image-1.5.

Limits and training data

The agent needs about ten times as much compute as a single editor, and its performance depends on which tools are plugged in. Because it keeps only verified results, it produced about 26,000 training examples. Editors fine-tuned on that data performed better across all three benchmarks. Training data generated directly by FLUX made FLUX worse.

The Decoder notes that Black Forest Labs has since replaced the general-purpose model used in the paper. Flux 3 Image Klein has 5.6 billion parameters, smaller than Flux 2 Klein, and scores higher than all other open models and the commercial Flux 2 Pro in Black Forest Labs' internal tests.

What this means

The results point to the judge's input as a lever. According to the researchers, what a judge model sees can matter more than how it was trained, since measured data lifted general-purpose models past a dedicated fine-tuned evaluator. The ten-times compute cost is substantial, but the 26,000 verified examples show a way to recover it: the agent's output can train faster editors. The Decoder points out that Adobe added AI agents to Photoshop, Premiere and other Creative Cloud apps in June, starting as a public beta, so how well agents choose tools and check their own output applies directly to Adobe's shipping products.