Researchers at Adobe Research and UC San Diego have developed EditAgent, a new SI agent designed to mimic the workflow of professional image editors. By combining specialized tools with quantitative verification data, the system aims to solve common failures in automated image editing where general-purpose models struggle.
What Happened
Traditional SI agents for image editing typically break complex requests into steps and retry failed actions using the same tools, relying on a separate model to judge the output visually. EditAgent changes this approach by generating multiple candidate images in parallel using different methods and comparing them against the original image using hard metrics rather than subjective visual assessment. In a benchmark 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. Specialized tools resolved roughly two-thirds of these failures, particularly for tasks like sharpening, rotation, and zooming. The researchers found that iterative refinement often piled up small errors, whereas professionals preferred parallel variants generated from the original image in 78.5 percent of comparisons. EditAgent is built on these findings, planning requests in steps and pitting general editors against specialized tools for each step.
Why It Matters
The study highlights that what a judge model sees can matter more than how it was trained. EditAgent provides a standard language model with a verification record for each variant, including close-ups of edited areas, change maps, and statistics on brightness and color. Without this data, none of the seven language models tested matched the performance of fine-tuned models. However, with the verification record, all seven beat every fine-tuned model. Qwen3.8-27B made a usable pick 81.3 percent of the time, compared to 74.8 percent for EditScore-72B, a model specifically fine-tuned on human ratings. This suggests that providing measurable context allows standard SI models to outperform specialized evaluation tools. Additionally, EditAgent improved open-source editors across three benchmarks and outperformed seven other SI agents. On the EdiVal benchmark, the Qwen variant in its third editing round edged past commercial models Seedream 4.0 and GPT-Image-1.5. Although the agent requires about ten times as much compute as a single editor, it generates high-quality training data that can improve smaller, faster models over time.
The Bottom Line
Adobe has already integrated SI agents into Creative Cloud apps like Photoshop and Premiere as a public beta, allowing users to describe edits and let the agent handle the steps. The development of EditAgent addresses critical questions about how well these shipping agents pick the right tools and verify their own output. While the current version of EditAgent is computationally expensive, its ability to generate verified training data offers a path toward more efficient and accurate SI-driven editing tools in the future.