What did the EditAgent paper find?
A general-purpose AI image editor produced no usable result in 42.7% of 600 test cases, according to a new paper from Adobe Research and UC San Diego. The paper introduces EditAgent, a system that routes each editing step to several tools and gives its judge more evidence before picking a result.
The paper is titled “EditAgent: Rethinking Agentic Image Editing”. It was posted to arXiv on 29 October 2026, with authors from both institutions. The work describes a research prototype, not an announced Photoshop feature.
The benchmark targets requests that contain several changes at once. A user might ask to move a subject, widen the frame, remove people in the background and change the light. The authors say current agents often finish some parts and silently drop others, especially when a later instruction depends on an earlier one.
Five professional Photoshop editors labeled the outputs blind. Across 600 candidate sets and 30 edit types, the general editor failed in 256 sets. Specialized methods recovered 175 of those failures. In 81 sets, or 13.5%, no candidate from any method was usable.
The breakdown shows how often one tool type wins. General-editor candidates were the only usable options in 254 sets, or 42.3%. Specialized candidates were the only usable options in 175 sets, or 29.2%. Both kinds appeared in 90 sets, or 15%.
Those 81 unusable sets are part of the result. They suggest that some multi-part requests exceed what the current tool set can deliver, even when several methods are tried. The paper does not claim that EditAgent solves every request, only that it improves the share of usable outputs.
The comparison to a Photoshop professional is an analogy about process, not a claim that EditAgent drives Adobe's editor. The system works with open-weight image models and separate editing tools, and the paper does not describe a Creative Cloud integration.
Why do AI image editors fail at multi-part requests?
The paper frames the problem as two limits: the agent's action space and its judge's observation space. The action space is what the agent can do. The observation space is what the judge can see when it decides whether an edit worked.
Most agentic editors retry a failed step with the same model. That approach cannot fix an error caused by using the wrong tool. The authors write that “one editor is not enough” even when it is sampled more often.
The specialists matter most where a general model lacks precision. Masked local editing can change one region without repainting the rest. A removal specialist can delete an object and fill the hole. A move specialist can reposition a subject. A general editor asked to do all three may blur boundaries or leave artifacts.
The failures cluster by edit type. Geometry, tone and enhancement edits were the weakest groups for the general editor. Style and Effects sets often had only general-editor winners, at 73.8%. Enhancement and Geometry sets often had only specialized winners, at 72.5% and 66.7%.
More sampling did not close the gap. A pool that delivered several general-editor samples produced a usable edit in 45.2% of sets. A pool that also included specialists reached 68.2%. In the 296 sets with exactly one usable delivered edit, the specialist pool supplied 217.
The paper separates two kinds of error. Some come from choosing the wrong operation. Others come from judging the result badly. A better action space addresses the first. A better observation space addresses the second.
Sequential refinement adds another failure mode. Each round edits the previous output, so small errors accumulate. In one example from the paper, a standalone editor dropped a dog when it added an old man in a later turn, then added three baguettes when asked for two.
The authors tested the two strategies head to head. On 144 Complex-Edit samples, five professional editors preferred parallel candidates made from the original image in 102 cases. They preferred sequential refinement in 28, with 14 ties.
That is 78.5% of the 130 non-tied comparisons. Parallel search also needed at most two judge calls, while the sequential setup needed four.
How does EditAgent check its own work?
EditAgent plans a request as a series of dependent steps. For each step, it routes the work to a general editor and to matching specialists. The paper lists specialists for removal, recoloring, restyling, moving objects, scaling subjects, photometric changes and named filters.
A planner orders the steps, and an analyzer and router choose the methods for each one. If a step fails after its attempts, a re-planner can replace the remaining steps or tell the agent to keep the best attempt.
The selected methods generate candidates in parallel from the same source image. Each candidate gets a verification record. That record can include a change map, before-and-after close-ups, a target mask, luminance and color statistics, geometry measurements and the results of code checks.
By default, the system generates at most three candidates per attempt and allows three attempts per step. An off-the-shelf vision-language model then compares the candidates in one listwise call. It sees the instruction, the source image, every candidate and every record.
The paper says this tells the judge what code has verified and what still needs visual judgment. That distinction matters because some properties are measurable. Brightness, color shift, rotation and scale can be checked with numbers. Whether an edit looks natural often cannot.
The verification record is not a single score. It is a bundle of evidence, and the judge receives it alongside the image. The paper argues that this changes the judge's job from guessing to checking.
The records can expose errors that a full-image glance hides. In one example, a general editor overexposed a photo and shifted its colors warmer when asked for a gentle brightening. The measurements showed the drift. A specialized tool stayed within the requested range.
A second check covers earlier turns. Because each step is judged against its own source image, a later edit can undo an earlier one without notice. After every turn, EditAgent shows the judge the original image, earlier results and the current image.
If an earlier request no longer holds, the system runs one repair turn. The paper's example is a window that was frosted in the first turn and cleared by an instruction to let in more daylight in the second. The history check catches the conflict and repairs it.
Does better evidence beat a bigger judge model?
The authors tested whether a standard model judges better when it receives the verification record. Without records, none of the seven vision-language models they tested matched EditScore-72B. That model was fine-tuned on human ratings of image edits and scored 74.8%.
The authors compared four judges that score one output at a time: EditScore-72B, EditReward-7B, CLIP-I and Qwen3.8-27B with a pass or fail prompt. No single judge was the most accurate across all edit groups. That inconsistency motivated the evidence bundle.
With records, all seven beat every fine-tuned model in the comparison. Qwen3.8-27B’s top pick was usable 81.3% of the time. Qwen3.8-Flash-Next reached 82.8%.
The paper measures this with Winner-Set Hit@1. The score is the share of sets where the judge's top pick is one the editors marked usable. Ties share credit. The test covers 515 sets that had at least one usable candidate.
The gain ranged from 6.3 to 9.3 points across the seven judges. The authors conclude that what a judge can see may matter more than how the judge was trained.
Comparison alone is not a cure. Judging candidates side by side raised overall accuracy, but it hurt the Photometric and Enhancement groups when the judge had no measurements. The records restored both groups above their pointwise scores.
This is the paper's most portable idea for agent builders. A stronger model is expensive to train and slow to swap. A better observation format can improve several judges at once, including models that were not trained for the task.
Does EditAgent make editing better, or just slower?
EditAgent improved two open-weight editors, Qwen-Image-Edit-2511 and FLUX.2-klein-9B, on three benchmarks. The benchmarks were EdiVal, Complex-Edit and ComplexBench-Edit. They cover multi-turn sessions, instruction complexity and compositional requests.
The test sets are substantial. EdiVal has 572 three-turn sessions. Complex-Edit has 8,496 real and synthetic examples. ComplexBench-Edit has 763 examples with parallel constraints and multi-step chains.
On EdiVal's overall score at the third turn, EditAgent reached 59.45. The strongest agent baseline, RISE-Agent, reached 58.42. The paper reports the best score in all three turns for EditAgent in that comparison, using the same backbones and judge.
EdiVal's overall score, called EdiVal-O, runs from 0 to 100. A gain of about one point at turn three is modest in absolute terms. The paper reports it consistently across backbones and turns, which matters more than the single headline number.
The cost is inference time. The authors report roughly 10 times the inference time of a standalone editing model. Performance is also bounded by the tools that are plugged in, so a weak specialist limits what the agent can deliver.
An ablation shows where the gains come from. A standard agent scored 47.64 on EdiVal at turn three. Expanding the action space added 2.94 points, and expanding the observation space added 3.09. The full system reached 55.46.
The paper offers a way to amortize that cost. EditAgent keeps only edits its judge accepts, which produced about 25,700 accepted sessions, averaging four turns each. Those sessions became training data.
Editors fine-tuned on that data improved on all three benchmarks. At turn three on EdiVal, Qwen-Image-Edit-2511 gained 1.95 points and FLUX.2-klein-base-9B gained 2.00. Training directly on FLUX-generated edits instead degraded FLUX.2-klein-base-9B on all three benchmarks.
The authors argue that this makes EditAgent a data engine, not only a slower editor. A costly agent could produce cheaper standalone models that keep part of its quality gain.
How strong is the evidence?
The benchmark is the authors' own, not an independent leaderboard. It contains 600 candidate sets over 30 edit types, and 515 sets had at least one usable edit. Five professional Photoshop editors labeled the candidates.
The general editor in the diagnostic benchmark was FLUX.2-klein-9B. The labels came from the authors' panel of five retouchers, not from a crowd or a commercial grading service. The paper does not describe a public leaderboard or a code release.
That is a small panel by design. The parallel-versus-sequential test used 144 samples and five editors, with 14 ties. The results show a clear preference, but they do not measure performance in a real editing workflow.
The paper evaluates open-weight editors in a research setting. It does not report a shipped product, pricing or a comparison with Adobe's commercial editing tools. The 10-times inference cost means the full system is not a drop-in replacement for a one-shot model.
Some gains are modest. At turn three on EdiVal, the best full system scored 59.45 out of 100. The benchmark rewards usable edits, not professional-grade retouching. The paper's examples are selected from its test sets.
The paper is an arXiv preprint, and its benchmark has not been independently reproduced. Treat the percentages as results on the authors' test set, not universal failure rates for AI photo editing.
What does this mean for AI photo editing tools?
For users, the practical lesson is that complex, multi-part edits still need review. A single instruction can contain several dependent changes, and an editor may complete some while dropping others. Check edges, colors and anything the request said to preserve.
For tool builders, the paper suggests two design choices. Route steps to the right specialist instead of retrying a general model. Give the judge measurable evidence instead of a full image alone.
The compute trade-off is the open question. A 10-times slower agent is hard to offer at consumer prices. Its accepted edits, however, can train faster models that absorb part of the gain.
The paper cites several competing agentic editors, which makes reliability as important as raw image quality. Verification records offer a way to show why an edit was accepted, not just that a model produced it. That kind of audit trail could matter for professional and commercial workflows.
Watch whether Adobe or other vendors adopt verification records in shipping products. The idea is easy to describe and testable on existing editors. That makes it more likely to spread than a new model architecture.
The paper also shows where evaluation needs to go. A single pass or fail score hides which edit types fail. Reporting results by edit group, as this paper does, gives buyers and researchers a clearer picture.
The most useful claim is not that one system won a benchmark. It is that the information a judge receives can change the result more than the judge's size. That is a cheaper lever for anyone building an editing agent.