Adobe 提出 EditAgent：让 AI 智能体像 Photoshop 专业用户那样选工具编辑图像

Adobe researchers built an AI agent for image editing that pairs a general-purpose editor with specialized tools. A standard language model compares results using measured data and picks the best one.

AI agents for image editing break complex requests into steps and typically retry failed steps with the same tools, while a separate model judges the output by eyeballing it.

Researchers at Adobe Research and UC San Diego want to change that with EditAgent, a system that works more like a pro in Photoshop, according to a new paper. It picks the right tool for each task and compares the result against the original using hard numbers.

Overview graphic comparing typical agentic image editing with EditAgent across three columns: Action Space, Observation Space, and Result. The top row shows repeated attempts with the same tools, while the bottom row shows parallel candidates from General, Local, Tone, and Style tools, judged by an off-the-shelf VLM using evidence in a single call. — Instead of retrying the same editor over and over, EditAgent generates candidates with different methods in parallel and has an unmodified language model compare them all at once. | Image: Adobe

General-purpose editors fail at too many tasks

In a test of 600 image edits, blind-rated by five professionals, the general-purpose editor Flux-2-Klein-9B failed to produce a usable result 42.7 percent of the time. Specialized tools solved roughly two-thirds of those failures, especially for sharpening, rotation, and zooming. More retries with the same editor barely helped.

The researchers also tested whether an agent does better by refining results step by step or by generating multiple variants in parallel. Iterative refinement piles up small errors because each round builds on an already altered image. Professionals preferred the parallel variants generated from the original in 78.5 percent of the comparisons where they picked a side. The agent needed only half as many evaluation passes. EditAgent is built on these two findings.

Hard numbers turn standard models into better judges

EditAgent plans a request in steps and, for each step, pits the general-purpose editor against matching specialized tools. A standard language model then compares all variants in a single pass.

Architecture diagram of EditAgent using the example of extinguishing a candle flame. Planner, Step Analyzer, and Router select methods from the Capability Pool. Three candidates are generated in parallel with close-ups and change maps as evidence, and a VLM accepts candidate 3. Below, the steps Remove Flame, Add Smoke, and Adjust Lighting follow in sequence. — EditAgent breaks a request into dependent steps and generates multiple candidates per step. A language model picks the best one based on verification data. | Image: Adobe

For each variant, the model gets a verification record with a close-up of the edited area, a change map of altered regions, and measured statistics on brightness and color. In one example, a photo was supposed to be gently brightened. The numbers showed the general-purpose editor overexposed it and shifted the colors warmer. The specialized tool stayed within spec.

Two examples of verification data. Top: a Removal Specialist and a General Editor remove a smiley patch from a checkered shirt, with the close-up revealing a blurred plaid pattern from the specialist. Bottom: a General Editor and a Photometric Operator brighten a dark still life, with measured brightness, clipping, and color shift values shown beneath. — Close-ups and measured image statistics reveal errors that a language model misses when just looking at the full image. | Image: Adobe

The researchers tested this on their 600-edit benchmark. The judge model had to pick the best result from several candidates, and they counted how often it chose an image the professionals also rated as usable. For comparison, they ran models like EditScore that were fine-tuned on human ratings of image edits.

Without the verification record, none of the seven language models tested could match the best fine-tuned model. With it, all seven beat every fine-tuned model. Qwen3.8-27B made a usable pick 81.3 percent of the time, compared to 74.8 percent for EditScore-72B. For the researchers, this suggests that what a judge model sees can matter more than how it was trained.

Dot plot of Winner-Set Hit@1 for seven Qwen and Gemma models, each with and without evidence, compared to fine-tuned evaluation models EditScore and EditReward. With evidence, scores rise by 6.3 to 9.3 percentage points, with Qwen3.8-Flash-Next reaching 82.8 percent. — With verification data, every standard language model tested outperforms the strongest fine-tuned evaluation model, EditScore-72B at 74.8 percent. | Image: Adobe

After each round, EditAgent also checks whether later edits accidentally undid earlier ones. In the paper's example, the instruction "let in more daylight" makes a previously frosted window clear again. The agent catches the conflict and fixes it.

Editing sequence of a bathroom in four images. In round 1, the window is frosted. In round 2, adding more daylight makes the glass clear again. The History Check catches the violation and restores the frosted glass while keeping the brighter lighting. — A post-check after each round ensures that later edits do not silently undo earlier changes. | Image: Adobe

EditAgent improved the open-source editors Qwen-Image-Edit-2511 and Flux-2-Klein-9B across three benchmarks and outperformed seven other AI agents. On the EdiVal benchmark, the Qwen variant in its third editing round edged past the commercial models Seedream 4.0 and GPT-Image-1.5. But the agent needs about ten times as much compute as a single editor, and its performance depends on which tools are plugged in.

The agent generates training data for faster editors

The researchers also use EditAgent as a data source. Because the agent only keeps verified results, it produced about 26,000 training examples. Editors fine-tuned on that data performed better across all three benchmarks. Training data generated directly by FLUX, on the other hand, actually made FLUX worse. That tradeoff could make the agent's high compute costs worth it over time.

Adobe added AI agents to Photoshop, Premiere, and other Creative Cloud apps in June, starting as a public beta. Users describe what they want, and the agent handles the individual steps. How well these agents pick the right tools and check their own output is a question that applies directly to Adobe's shipping products. Black Forest Labs has since replaced the general-purpose model tested in the paper: the new Flux 3 Image Klein has 5.6 billion parameters, smaller than Flux 2 Klein, yet it scores higher than all other open models and even the commercial Flux 2 Pro in Black Forest Labs' internal tests.

AI News Without the Hype – Curated by Humans

Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.