Adobe and UC San Diego test EditAgent for AI image editing

Adobe Research and UC San Diego built EditAgent, an AI image-editing system that splits requests into steps, tests a general-purpose editor against specialized tools, and uses measured data to choose results. It produced more accurate edits than prior AI agents in tests, but requires about ten times the compute of a single editor.

How EditAgent edits images

EditAgent plans a request in steps. For each step, it pits a general-purpose editor against matching specialized tools for tasks such as sharpening, rotation, or zooming. Instead of retrying the same editor repeatedly, it generates candidates with different methods in parallel from the original image. A standard language model then compares all variants in one pass. In a test of 600 image edits blind-rated by five professionals, 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. Iterative refinement accumulated small errors because each round built on an already altered image, while professionals preferred parallel variants from the original in 78.5 percent of comparisons where they chose a side. The agent also needed half as many evaluation passes.

Verification data improves judging

For each variant, the model receives a verification record containing a close-up of the edited area, a change map of altered regions, and measured statistics on brightness and color. In one example, a photo was meant to be gently brightened; the numbers showed the general-purpose editor overexposed it and shifted colors warmer, while the specialized tool stayed within spec. Close-ups and measured statistics revealed errors a language model could miss when viewing only the full image. In the 600-edit benchmark, the judge had to pick the best result from several candidates, and researchers counted how often it chose an image professionals also rated usable. Without the verification record, none of seven language models matched 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 with 74.8 percent for EditScore-72B. After each round, EditAgent also checks whether later edits undo earlier ones; in one example, "let in more daylight" made a previously frosted window clear again, and the agent caught and fixed the conflict.

Tradeoffs and training data

EditAgent improved 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 Seedream 4.0 and GPT-Image-1.5. However, the agent needs about ten times as much compute as a single editor, and 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, while training data generated directly by FLUX made FLUX worse.

Adobe products and model updates

Adobe added AI agents to Photoshop, Premiere, and other Creative Cloud apps in June as a public beta. Users describe what they want, and the agent handles individual steps. The paper's findings about tool selection and self-checking apply to those products. Separately, Black Forest Labs replaced the general-purpose model tested with Flux 3 Image Klein, which has 5.6 billion parameters, is 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.

Sources: The Decoder