Adobe's EditAgent Makes AI Judge Its Own Photo Edits With Hard Numbers

Adobe Research and UC San Diego have built EditAgent, an AI system for image editing that picks the right tool for each sub-task and then verifies its own output with measured statistics rather than a visual guess. The work is laid out in a new paper, and its results speak in concrete terms: with a verification record in hand, every one of seven tested language models outperformed every fine-tuned evaluation model the researchers used for comparison.

The stakes for Adobe are direct. The company added AI agents to Photoshop, Premiere and other Creative Cloud apps in June, starting as a public beta. Users describe what they want, and the agent handles the individual steps. How well such agents select tools and check their own output applies immediately to Adobe's shipping products.

Why general-purpose editors fall short

Today's AI agents for image editing typically break a complex request into steps, retry failed steps with the same tools, and hand the result to a separate model that judges the output by eyeballing it. The researchers wanted a system that works more like a professional in Photoshop: choose the right tool for each task and compare the result against the original using hard numbers.

The failure rates of the general-purpose approach justify that ambition. In 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.

What did the researchers change first?

Before building EditAgent, the team tested a structural question: does an agent do better by refining results step by step, or by generating multiple variants in parallel? Iterative refinement lost. Each round builds on an already altered image, so small errors pile up. Professionals preferred the parallel variants generated from the original in 78.5 percent of the comparisons where they picked a side. The agent also needed only half as many evaluation passes. EditAgent is built on these two findings: specialized tools where generalists fail, and parallel variants instead of iterative refinement.

How does the number-based judging work?

EditAgent plans a request in steps and, for each step, pits the general-purpose editor against matching specialized tools. A standard language model then compares all variants in a single pass. For each variant, the model receives a verification record containing three things:

• a close-up of the edited area
• a change map of altered regions
• measured statistics on brightness and color

In one example from the paper, a photo was supposed to be gently brightened. The numbers showed the general-purpose editor overexposed it and shifted the colors warmer. The specialized tool stayed within spec.

The researchers tested the judge on their 600-edit benchmark. The model had to pick the best result from several candidates, and the team counted how often it chose an image the professionals also rated as usable. For comparison, they ran models like EditScore that were fine-tuned on human ratings of image edits.

The gap was decisive. Without the verification record, none of the seven language models tested could match the best fine-tuned model. With it, all seven beat every fine-tuned model. Qwen3.8-27B made a usable pick 81.3 percent of the time, compared to 74.8 percent for EditScore-72B. For the researchers, the conclusion is that what a judge model sees can matter more than how it was trained.

EditAgent also audits its own step sequence. After each round, it checks whether later edits accidentally undid earlier ones. In the paper's example, the instruction "let in more daylight" makes a previously frosted window clear again. The agent catches the conflict and fixes it.

How does EditAgent stack up against other systems?

On benchmarks, the system 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.

The cost is compute. The agent needs roughly ten times as much compute as a single editor, and its performance depends on which tools are plugged in. That dependency on the tool set is a structural limit: EditAgent is only as strong as the specialized instruments available to it.

Can the agent pay for itself as a data source?

The researchers propose one answer to the compute problem. Because EditAgent only keeps verified results, it produced about 26,000 training examples. Editors fine-tuned on that data performed better across all three benchmarks. The counterexample is instructive: training data generated directly by FLUX actually made FLUX worse.

That asymmetry is the paper's economic argument. Verified, tool-checked output can train faster editors, so the agent's high compute costs could pay off over time even if the agent itself never ships as a consumer feature. The tenfold compute penalty buys quality control that raw generation cannot match.

What has changed since the paper's tests?

The field has already moved under the paper's feet. Black Forest Labs has replaced the general-purpose model tested in the research: the new Flux 3 Image Klein has 5.6 billion parameters, smaller than Flux 2 Klein, yet it scores higher than all other open models and even the commercial Flux 2 Pro in Black Forest Labs' internal tests.

The paper's core finding is likely to outlast that specific model swap. A verification record built from close-up crops, change maps and measured brightness and color statistics turned ordinary language models into judges that beat purpose-trained evaluators. If the result holds outside Adobe's benchmark, the design lesson transfers: give a general model hard evidence about its own output, and the training regime matters less than the data on the screen.