THE ESSENTIALS IN 30 SECONDS

  • On the authors' 600-case benchmark, the general editor FLUX.2-klein-9B produced no usable result in 42.7% of cases; specialist tools recovered 175 of those 256 failures.
  • With a verification record attached to each candidate, off-the-shelf vision-language judges beat the fine-tuned EditScore-72B judge (74.8%), the best reaching 82.8%.
  • The authors say the agent needs roughly 10 times the inference time of a single editor, and propose using it to generate training data.

Researchers from Adobe Research and the University of California, San Diego have described EditAgent, an image-editing agent that does not rely on a single model. The paper, posted on arXiv on October 29, argues that current editing agents are limited both by the few tools they can call and by judges that only look at the final picture.

How it works

EditAgent breaks a complex request into a sequence of steps. For each step, it routes the work either to a general-purpose editor or to specialised methods, such as tools for sharpening, rotation, zooming or tone changes, as The Decoder explains. Rather than refining one image again and again, it produces several candidates in parallel from the same starting point.

A vision-language model then compares all candidates in one pass. The key addition is a verification record for each candidate; according to The Decoder, it includes a close-up of the edited area, a map of changed regions and measured brightness and colour statistics. The paper reports that with these records, unmodified judge models outperform a judge fine-tuned for the task: EditScore-72B reaches 74.8% on the authors' benchmark, while the best judge with records, Qwen3.8-Flash-Next, reaches 82.8%.

What the evaluation shows

The authors built a diagnostic benchmark of 600 editing sets across 30 edit types, labelled by five professional retouchers who did not know which method produced which image. On it, the general editor FLUX.2-klein-9B failed in 256 sets, or 42.7%, and a specialised method recovered 175 of those failures. In a separate comparison, the editors preferred parallel search over step-by-step refinement in 78.5% of the 130 cases where they picked a side. The paper says EditAgent improves two open-weight editors, Qwen-Image-Edit-2511 and FLUX.2-klein-9B, on three benchmarks.

Costs and limits

The approach is expensive. The authors acknowledge roughly 10 times the inference time of a standalone editor, and results are bounded by the quality of the tools plugged in. Their proposed answer is to use the agent as a data engine: about 25,700 verified editing sessions were turned into training samples, and editors fine-tuned on them gained around two to three points on the benchmarks reported.

EditAgent is an academic paper, not a feature Adobe has announced for its apps. The results come from the authors' own benchmarks and have not yet been independently reproduced.