EditAgent research involving Adobe evaluated 600 candidate sets across 30 editing categories with professional editors. Of 256 unusable results from Flux.2 Klein 9B, 175 could be recovered with professional editing tools. A parallel approach that tried both generative editing and tool calls was preferred over a sequential approach in 78.5% of non-tied comparisons. The team also trained smaller agents using about 25,700 accepted tool trajectories. These findings concern specific evaluation conditions; they are not an announced Photoshop feature or a success rate for every image-editing task.
Back to Tech
Adobe research uses editing tools to recover failures in generative image editing
The study evaluates layers, masks, and tool execution as alternatives to simply requesting another generated image.