Adobe Research and UC San Diego researchers have built EditAgent, an AI system for image editing that picks specialized tools for each step and checks results against measured data. The work is described in a new paper, according to The Decoder.
The researchers tested 600 image edits, which five professionals rated blind. The general-purpose editor Flux-2-Klein-9B failed to produce a usable result 42.7 percent of the time. Specialized tools solved about two-thirds of those failures, especially for sharpening, rotation, and zooming. Retrying with the same editor barely helped.
The team also compared two ways of generating edits. Iterative refinement, which builds each round on the previous image, piles up small errors. Parallel variants generated from the original image were preferred by professionals in 78.5 percent of comparisons where they picked a side. The agent needed only half as many evaluation passes.
EditAgent plans each request in steps. For each step, it runs the general-purpose editor and matching specialized tools in parallel. A language model then compares the variants in a single pass. The judge receives a verification record for each variant, including a close-up of the edited area, a map of altered regions, and measured brightness and color statistics. In one example, the numbers showed the general-purpose edit