Adobe and UCSD's EditAgent combines general and specialist image editing tools

Adobe Research and UC San Diego built EditAgent, an image-editing agent. It routes each step of a request to either a general-purpose editor or a specialized tool. It then uses measured image data, rather than visual impressions, to decide which result to keep. The arXiv paper was submitted on October 29, 2026. It targets a specific failure mode: existing agents mostly retry a failed step with the same narrow set of tools, and their judges score each output by eye.

The paper starts from a benchmark of 600 image edits, blind-rated by five professionals. In that test, the general-purpose editor Flux-2-Klein-9B failed to produce a usable result 42.7 percent of the time, according to The Decoder. Specialized tools solved roughly two-thirds of those failures, especially for sharpening, rotation and zooming. Retrying the same editor barely helped. The authors' abstract makes the same point more bluntly: such agents are limited by what their action space can do and by what their judge can see.

The second finding concerns how steps are chained. Iterative refinement, where each round edits the previous output, compounds small errors, because every round builds on an already altered image. EditAgent instead generates candidates in parallel from the original. A judge compares them in a single listwise call. Professionals preferred these parallel variants in 78.5 percent of decided comparisons, and the agent needed no more than half as many judge calls, according to the abstract.

The judging change is the most technically interesting part. For each candidate, the judge receives 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 example from the paper, a photo meant to be gently brightened was overexposed and shifted warmer by the general-purpose editor. The specialized tool avoided that fault. A language model looking at a full image can miss errors like this, while measured numbers expose them. The paper claims that with these records, off-the-shelf vision-language judges outperform reward models fine-tuned on human ratings. On the benchmark, Qwen3.8-27B picked a usable result 81.3 percent of the time, against 74.8 percent for the fine-tuned EditScore-72B. Without the records, none of the seven judges tested matched the best fine-tuned model. For the researchers, this suggests that what a judge is shown can matter more than how it was trained.

The paper also reports a post-check after each round. It looks for later edits that undo earlier ones. Its example is the instruction "let in more daylight" making a previously frosted window clear again, which the agent catches and corrects.

The results come with costs. EditAgent improved the open-weight editors Qwen-Image-Edit-2511 and Flux-2-Klein-9B across three benchmarks and beat seven other agents. On EdiVal, a third-round Qwen variant edged past the commercial Seedream 4.0 and GPT-Image-1.5. However, The Decoder notes that the agent needs about ten times the compute of a single editor. Its performance also depends on which tools are plugged in.

The paper also proposes using EditAgent as a data engine. Because it keeps only verified results, it produced about 26,000 training examples. Editors fine-tuned on that data did better across all three benchmarks. Training data generated directly by FLUX made FLUX worse. The Decoder suggests this tradeoff could make the agent's high compute costs worth it over time.

The paper does not say when or whether EditAgent will reach users. The Decoder links the research to Adobe's June addition of AI agents to Photoshop, Premiere and other Creative Cloud apps, which began as a public beta. It notes, however, that the paper's general-purpose model has since been replaced by Black Forest Labs' Flux 3 Image Klein. The company says that model scores higher on its internal tests than the commercial Flux 2 Pro.