Adobe Research and the University of California, San Diego have developed EditAgent, an AI system intended to handle image-editing tasks more like an experienced Photoshop user. Rather than repeatedly applying the same general-purpose editor, the system breaks a request into steps and selects specialized tools when they are better suited to a particular operation.
The researchers found that general-purpose editors often struggle with specific tasks. In a test of 600 image edits evaluated blindly by five professionals, Flux-2-Klein-9B failed to produce a usable result in 42.7 percent of cases. Specialized tools resolved about two-thirds of those failures, particularly in sharpening, rotation and zooming tasks.
The study also compared iterative editing with generating several alternatives from the original image. Professionals preferred parallel variants in 78.5 percent of the comparisons in which they selected one approach. According to the researchers, repeated refinement can accumulate errors because each new attempt builds on an already modified image. EditAgent therefore generates competing results and evaluates them rather than relying solely on repeated retries.
Its evaluation process gives a standard language model additional verification data for each image variant. This includes a close-up of the edited area, a map showing changed regions, and measurements related to brightness and color. The model uses that information to compare the candidates in a single pass.
On the 600-edit benchmark, the verification record significantly improved the language models' selections. Without the additional data, none of the seven tested models matched the strongest model fine-tuned on human ratings. With it, all seven performed better than the fine-tuned models. Qwen3.8-27B selected an image that professionals considered usable 81.3 percent of the time, compared with 74.8 percent for EditScore-72B.
EditAgent also checks whether later instructions have unintentionally reversed earlier edits. In one example described by the researchers, an instruction to allow more daylight caused a previously frosted window to become clear again. The system detected the conflict and corrected the result.
The system improved the open-source editors Qwen-Image-Edit-2511 and Flux-2-Klein-9B across three benchmarks, and it outperformed seven other AI agents. On the EdiVal benchmark, the Qwen-based version in its third editing round slightly exceeded the commercial Seedream 4.0 and GPT-Image-1.5 models.
The main limitation is computational cost. EditAgent requires about ten times the compute of a single editor, and its results depend on the tools connected to it. The researchers also used the system to generate roughly 26,000 verified training examples. Models fine-tuned on that data performed better across all three benchmarks, while training data generated directly by FLUX reduced FLUX's performance.
Adobe began adding AI agents to Photoshop, Premiere and other Creative Cloud applications in public beta in June. The researchers' work addresses a related challenge for those products: determining how an agent chooses the appropriate tool and verifies that its final output matches the user's request.
Sources: The Decoder.