EditAgent: Rethinking Agentic Image Editing

EditAgent. Rethinking Agentic Image Editing. By Haiyang Xu and seven others, from UC San Diego and Adobe Research. Published on the twenty-ninth of October.

Agentic image editors break a complex request into steps. When a step fails, though, they mostly retry with the same narrow set of tools, and with a judge that scores each output by eye, directly. The authors frame the limit as two spaces. The action space is what the agent can do. The observation space is what its judge can see. On their expert-annotated benchmark, a general-purpose editor yields no usable result in over forty percent of cases, and no judge is reliable across edit types. Both spaces, in other words, are too small.

EditAgent is built to widen both at once. On the action side, it routes each step not only to general editors but also to specialized methods. On the observation side, it attaches a verification record to every candidate it produces, so the judge has more to look at than the image alone.

Candidates are generated in parallel rather than one after another, and then judged together in a single listwise call. That replaces a loop of edit, inspect and retry with one round of many attempts and one comparison across them. The records matter for the judge as much as the routing matters for the editor. With the verification records in hand, off-the-shelf VLM judges outperform finetuned reward models, which means a stronger judge here comes from giving it more to observe, not from training it further.

Experts prefer the parallel, listwise approach to sequential refinement in seventy-eight point five percent of decided comparisons, and it uses no more than half the judge calls. EditAgent improves two open-weight editors on three benchmarks, and it also serves as a data engine for training stronger editors.

The reframing carries the paper. Retrying harder inside the same tools and the same view does not escape the bound. Giving the agent more ways to act and its judge more evidence to see is what moves the result, and it does so while spending fewer judge calls, not more.