EditAgent widens its toolbox and beats sequential refinement in 78.5% of comparisons

Image-editing agents usually follow a "plan, edit, evaluate" loop: a complex request is split into steps, and when a step fails, the same editor retries it while the same judge scores a single result. Researchers from UCSD and Adobe Research (the work was done during an internship at Adobe) argue that such an agent hits two limits: what it can do and what its judge can see. Their system, EditAgent, expands both. In a blind comparison, professional retouchers preferred its parallel search to sequential refinement in 78.5% of cases, with no ties.

Where a single general-purpose editor breaks down

To measure the problem, the authors built a diagnostic benchmark: 600 sets covering 30 edit types, 20 sets per type, grouped into six categories. Each set contains a source image, an instruction and 2–4 candidates: samples from the general-purpose editor FLUX.2-klein-9B plus outputs from specialized methods (local mask-based editing, object removal, photometric operators). Five professional Photoshop retouchers jointly marked all acceptable candidates.

The results are unflattering for general-purpose models. Only the general editor's candidates were usable in 254 sets (42.3%), only the specialized ones in 175 (29.2%), both types in 90 (15.0%), and in 81 sets (13.5%) no candidate was acceptable. Since the general editor appears in every set, it failed in 256 cases (42.7%), and a specialized method rescued 175 of them. The general editor did worst on geometry, tone and image enhancement.

Simply drawing more samples doesn't help. When the same VLM judge picked the best result from a pool containing only the general editor's outputs, it landed a usable edit in 45.2% of sets. With specialized methods added to the pool, that rose to 68.2%. Of 296 sets where exactly one of the two selected results was usable, 217 times it came from the expanded pool.

Parallel search instead of sequential refinement

In a second experiment, with a budget of four generations, the authors compared two modes using one editor and one judge (Qwen3.8-27B). Sequential refinement edits its own previous output and spends three feedback calls plus one final selection. Parallel search generates candidates independently from the source image and needs no more than two VLM calls.

On 144 samples from Complex-Edit (difficulty levels 1, 4 and 8), retouchers chose the parallel variant 102 times and the sequential one 28 times; the remaining 14 were ties. The authors attribute this to error accumulation: in sequential refinement, small defects from the first iteration carry into later ones, while parallel candidates always start from the original frame.

A judge that sees the evidence

The second part of the work concerns evaluation. Judges that score one result at a time proved unstable: none of those tested, including the fine-tuned reward model EditScore-72B, leads across all edit groups. On the Winner-Set Hit@1 metric (the share of sets where the top-scored candidate is among the acceptable ones), Qwen3.8-27B is only 5.2 points above random choice in one group. The CLIP-I similarity-to-source metric wins where the result barely changes, and is no better than random where the edit should be visible.

EditAgent gives the judge a verification record for each candidate: a change map, close-ups of the edited region, measured brightness, color and geometry, and the results of programmatic checks. The judge knows what code has already confirmed and what is left to assess by eye. It compares the candidates in a single call.

The effect shows on EditReward-Bench. EditScore models reach 65.9–70.3%, and the 72B version with a four-pass ensemble reaches 76.3%. Off-the-shelf VLMs score 71.7–73.8% when rating candidates one by one, and 85.2–86.1% when comparing candidates in a single call. On the diagnostic benchmark, list-wise comparison lifted Qwen3.8-27B from 71.6% to 73.6%, but in two groups where the result differs little from the source, accuracy fell from 65.0% to 36.8%. The verification records fixed this.

What else the experiments showed, and the limitations

The architecture includes a planner, an analyzer, region grounding via a VLM and SAM 2.1, and a router. The router selects methods: mask-based editing, removal via OSOR, moving and scaling objects, photo and style filters, rotations and flips, zoom via cropping and super-resolution, or outpainting. An accepted result becomes the source for the next step.

With Qwen-Image-Edit-2511 and FLUX.2-klein-9B as base editors, EditAgent improved both models on three benchmarks, including EdiVal. In addition, 25.7 thousand accepted sessions proved useful as training data: editors fine-tuned on them outperformed models trained on data taken directly from FLUX.2-klein-9B.

The limitations are also clear. The diagnostic set is small, five retouchers annotated it, and the general-purpose editor is a single model. Parallel search saves judge calls but requires running several methods at once. For practitioners, the conclusion is this: when an edit fails, add a different tool and give the judge measurable evidence. A hard image-editing request is better not solved by retrying with the same editor.