256 of 600 AI editing sets failed; 175 sets supplemented with specialized tools

Adobe Research and UC San Diego researchers proposed EditAgent, an agentic image-editing system that decomposes complex edit requests into multiple steps and selects appropriate tools for each, in a paper submitted to arXiv on October 29, 2026. The motivation came from a diagnostic experiment of 600 editing sets—20 each across 30 edit types—in which the general-purpose image editing model FLUX.2-klein-9B failed to produce usable results on 256 sets, while specialized methods recovered usable outputs on 175 of those failures. The system matters because it suggests routing edits to specialized tools via an agent layer can substantially close the gap left by general-purpose models. The paper is a research contribution, not a product release, so direct commercial impact on Adobe's offering is not yet demonstrated. What to watch next: whether Adobe productizes the approach, broader benchmark adoption, and follow-up work validating the method across additional edit types.