Adobe Research and UC San Diego build EditAgent, an image-editing agent that checks its own results

EditAgent breaks an image edit into dependent steps and generates several candidate results for each step in parallel, rather than retrying one editor again and again. Each candidate is produced by either the general-purpose editor Flux-2-Klein-9B or a specialized tool for tasks such as sharpening, rotation and zooming. The general-purpose editor failed to produce a usable result in 42.7 percent of the 600 edits in the benchmark, and specialized tools solved roughly two-thirds of those failures.

A standard language model then picks the best candidate using a verification record for each variant. That record includes a close-up of the edited area, a map of altered regions and measured brightness and color statistics. Without it, none of the seven language models tested matched the best fine-tuned evaluation model, while with it all seven beat that model. Qwen3.8-27B picked a usable result 81.3 percent of the time, compared with 74.8 percent for EditScore-72B.

The agent also checks whether later edits undo earlier ones, and it has produced about 26,000 verified training examples for faster editors. The paper says EditAgent needs about ten times the compute of a single editor.

Developers building image-editing agents can use measured statistics and close-up crops as judge inputs, which the paper suggests can matter more than how the judge model was trained. Running several candidates in parallel from the original image avoids the error buildup that step-by-step refinement caused in the tests.

The approach costs about ten times the compute of a single editor, so teams should weigh that against the accuracy gains. Adobe has added AI agents to Photoshop and Premiere as a public beta, so the tool-selection and self-checking questions in the paper apply to shipping products.