Adobe Research's EditAgent uses verification data to boost AI edit judging

EditAgent's approach contrasts with typical AI editing agents that retry the same general-purpose editor repeatedly and judge output by eyeballing. In a test of 600 image edits blind-rated by five professionals, the general-purpose editor Flux-2-Klein-9B failed to produce a usable result 42.7 percent of the time, while specialized tools solved roughly two-thirds of those failures, especially for sharpening, rotation, and zooming.

The researchers also found that iterative refinement piles up small errors, whereas professionals preferred parallel variants generated from the original in 78.5 percent of comparisons where they picked a side. EditAgent builds on both findings to avoid retrying the same editor and to use hard numbers for judging. It also checks whether later edits undo earlier ones, catching conflicts such as a "let in more daylight" instruction making a previously frosted window clear again.

Beyond editing, EditAgent serves as a data source: because it only keeps verified results, it produced about 26,000 training examples, and editors fine-tuned on that data performed better across three benchmarks, whereas training data generated directly by FLUX made FLUX worse. That tradeoff could make the agent's high compute costs worth it over time. Adobe's June addition of AI agents to Photoshop and other Creative Cloud apps as a public beta makes how well these agents pick tools and check their own output a directly relevant question for its shipping products.