EditAgent lets an image editing agent send each step to different kinds of editors and check candidates with verification records, instead of retrying with the same tools.
The problem: The paper says agentic editors retry failed steps with a narrow set of tools and a judge that scores outputs by eye. On its expert-annotated benchmark, a general-purpose editor gives no usable result in over 40% of cases.
The method: Each step goes to general editors or specialized methods, and each candidate gets a verification record. Candidates are generated in parallel and judged in one listwise call.
The results: Experts prefer this to sequential refinement in 78.5% of decided comparisons, with no more than half the judge calls. The paper also says it improves two open-weight editors on three benchmarks.
Sign in to comment.
TL;DR of the paper, generated automatically.
EditAgent lets an image editing agent send each step to different kinds of editors and check candidates with verification records, instead of retrying with the same tools.