Adobe 正在测试 AI 智能体能否像 Photoshop 专业人员那样使用编辑工具

Adobe Research and UC San Diego built EditAgent, an AI system that breaks image editing tasks into sub-steps and pairs a general-purpose editor with specialized tools for things like sharpening or zooming. A standard language model evaluates image variants generated in parallel, using measured statistics and close-up crops. With this data, it outperforms evaluation models that were specifically trained for the job. EditAgent produces more accurate results than previous AI agents but needs about ten times the compute. The verified output can then be used to train faster editing models. 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. Ad 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. Ad 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. 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. Ad 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. 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. Ad 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. Ad