The Pulse
EditAgent Gives AI Image Editors More Tools and Better Checks
EditAgent combines general-purpose and specialized image-editing methods, then gives a vision-language model evidence to help judge the results. In the paper’s benchmarks, verification records improved candidate selection, while the system’
AI.info Team · Written with AI · fact-checked ·
“We argue that such agents are bounded by two spaces: the action space of what the agent can do, and the observation space of what its judge can see.”
Haiyang Xu et al., authors of EditAgent: Rethinking Agentic Image Editing
That distinction frames EditAgent, a research system presented in a paper submitted to arXiv on October 29. Rather than repeatedly asking one image editor to try again, EditAgent sends each editing step to a selection of general-purpose and specialized methods. It then gives a vision-language model (VLM) judge several candidate results at once, alongside records showing what changed and what automated checks found.
When one editor cannot do the job
The authors begin with a diagnostic benchmark of 600 editing tasks across 30 types, from moving objects to changing color and correcting geometry. Five professional Photoshop retouchers judged which candidate edits were usable. A general-purpose editor failed to produce a usable result in 256 tasks, or 42.7%; specialized methods recovered 175 of those failures.
Those results varied by task. General-purpose editing candidates alone were usable in 73.8% of Style & Effects tasks, while specialized methods alone supplied the usable result in 72.5% of Enhancement tasks and 66.7% of Geometry tasks. The benchmark also found 81 tasks, or 13.5%, where none of the candidates worked. More samples from the same general editor did not close the gap: the authors report usable results in 45.2% of sets for that editor’s candidate pool, compared with 68.2% when specialized methods joined it.
Parallel candidates, checked side by side
EditAgent’s planner breaks a complex request into steps, and a router selects methods suited to each one. Its toolbox includes masked local edits, object removal and movement, color and lighting adjustments, style filters, rotations, and zooming through cropping or outpainting. The system generates candidate edits from the same source image rather than feeding one edited result into the next attempt.
That design matters because sequential editing can carry mistakes forward. In a blind comparison of 144 Complex-Edit examples, five professional editors preferred parallel search in 102 cases and sequential refinement in 28; 14 comparisons ended in a tie. Among the comparisons with a clear preference, parallel search won 78.5%, using no more than half as many VLM calls, the paper says.
The judge also receives a verification record for each candidate. Records can include a map of changed pixels, close-ups of the edited area, measurements such as color shifts, and results from programmatic checks. Such evidence can help distinguish a subtle but correct edit from a candidate that simply looks more like the original.
Evidence lifts the judge’s accuracy
On the authors’ diagnostic benchmark, the Qwen3.8-27B judge selected a usable edit in 71.6% of eligible sets when judging candidates individually. Comparing candidates together raised that figure to 73.6%; adding verification records brought it to 81.3%. The paper notes a trade-off: side-by-side judging alone hurt performance on some subtle photometric and enhancement tasks, while the records helped recover accuracy on those categories.
The researchers also tested EditAgent with two open-weight image editors, Qwen-Image-Edit-2511 and FLUX.2-klein-9B, across EdiVal, Complex-Edit, and ComplexBench-Edit. On EdiVal’s third-turn overall score, the paper reports Qwen-Image-Edit-2511 rising from 47.64 to 55.46 with EditAgent in the 512-pixel setting. The system’s accepted sessions also supplied about 25,700 examples for fine-tuning; models trained on that data improved across the three benchmarks, according to the paper.
The cost is roughly ten times the inference time
EditAgent’s added search and checking come with a substantial runtime cost. The authors estimate that it takes roughly 10 times as long as a standalone editing model, and say its results remain limited by the capabilities of the underlying editing methods. Their proposed use for the accepted sessions is to train standalone editors that could deliver some of the quality gains with less computation.
The paper presents EditAgent as a framework and reports benchmark results. Its central test is whether a system can choose a suitable editing method and give its judge enough evidence to assess the result. In the reported benchmark, the combination of specialized methods and verification records raised the usable-edit rate and improved candidate selection, but the added runtime remains a concrete constraint.