Image-editing agents stop retrying the same tool: routed specialists plus verification records win 78.5% of expert comparisons

When an image-editing agent fails, the problem is often not how many times it retries but which tools it can call and what its judge can see. Earlier agents refined, retried or rolled back within a fixed tool set, and their judge scored one output by eye from the source image, the edited image and the instruction. On a 600-set expert-annotated diagnostic benchmark spanning 30 edit types, the general editor FLUX.2-klein-9B yields no usable result in 42.7% of sets and specialized methods recover 175 of them; winner sets contain only general-editor candidates in 42.3% of sets and only specialized candidates in 29.2%.

EditAgent routes each step to both a general editor and specialized methods, and hands the judge a record of what code has already verified. The specialized pool covers masked local editing and recoloring, OSOR object removal, object movement, subject scaling, photometric operators, style filters, flips and rotations, and zoom-in crops or canvas extension. A verification record holds the share of visibly changed pixels, a change map, a before-and-after close-up of the named region, measurements such as brightness, contrast, saturation, color shift and sharpness, and code-check outcomes; on the 515 diagnostic sets with a usable candidate, the Qwen3.8-27B judge rises from 71.6% pointwise to 81.3% listwise with records.

Generating candidates in parallel beats sequential refinement on judge calls and on expert preference. Sequential refinement keeps editing its own previous output, so early errors can accumulate into later edits, whereas parallel search edits every candidate directly from the step's source image. In a blind comparison on 144 Complex-Edit samples at complexity levels 1, 4 and 8, five professional Photoshop editors preferred parallel search in 102 cases and sequential refinement in 28, with 14 ties, or 78.5% of decided comparisons, while parallel search needs at most two VLM calls and sequential refinement always needs four.

Sessions the judge accepts can serve as training data for stronger standalone editors. The matched control edits the same sessions directly with the base editor FLUX.2-klein-9B, so the two datasets differ only in their final images. With FLUX.2-klein-9B as the general editor, EditAgent produces about 25.7K fully accepted sessions; on EdiVal-O at turn 3, Qwen-Image-Edit-2511 and FLUX.2-klein-base-9B finetuned on EditAgent data gain 1.95 and 2.00 points, while direct FLUX.2-klein-9B data degrades FLUX.2-klein-base-9B on all three benchmarks.

The report covers the diagnostic benchmark, three editing benchmarks and the data-engine experiments in the supplied material: the diagnostic set has 600 sets over 30 edit types labeled jointly by five professional Photoshop editors; benchmark experiments use the Qwen-Image-Edit-2511 and FLUX.2-klein-9B backbones, run EdiVal at 512 px input, and also run Qwen-Image-Edit-2511 at 1024 px before resizing back to 512 px. It is aimed at researchers working on multi-turn image-editing agents and editing-data construction.

The material states that EditAgent incurs roughly 10 times the inference time of standalone editing models and that its performance is inherently bounded by the capabilities of its underlying editing methods; the authors propose using it as a data engine to finetune stronger and more efficient standalone editors that could amortize this cost, but no measurement of that amortization is given.