Adobe is testing how well AI agents can use editing tools like a Photoshop professional would
Reported by The Decoder1 source ↓
Adobe Research and UC San Diego developed EditAgent, which compares general-purpose and specialized editing tools using measured image data. In tests, it improved results from open-source editors but used about 10 times the compute of a single editor.
Key points
- EditAgent chooses tools for each editing step and checks whether later edits undo earlier ones
- On a 600-edit benchmark, Qwen3.8-27B picked a usable image 81.3% of the time, versus 74.8% for EditScore-72B
- Professionals preferred parallel variants made from the original over step-by-step refinements in 78.5% of comparisons where they chose a side
- The system improved two open-source editors across three benchmarks and outperformed seven other AI agents
- EditAgent used about 10 times the compute of a single editor, and results depended on the tools connected to it
Details
Adobe Research and UC San Diego describe EditAgent in a new paper as a system that plans an image-editing request in steps, then compares results from a general-purpose editor and specialized tools suited to each task. A language model judges the candidates using a verification record that includes a close-up of the edited area, a map of changed regions, and brightness and color measurements. The agent also checks whether later edits have undone earlier changes.
In a test of 600 edits, the general-purpose editor Flux-2-Klein-9B failed to produce a usable result 42.7% of the time. Specialized tools resolved roughly two-thirds of those failures, particularly for sharpening, rotation and zooming. In comparisons where professionals chose between approaches, they preferred parallel variants generated from the original image over step-by-step refinements in 78.5% of cases.
The researchers report that adding the verification record enabled all seven tested language models to outperform the fine-tuned image-editing judges in their test. Qwen3.8-27B selected a result professionals rated usable 81.3% of the time, compared with 74.8% for EditScore-72B. EditAgent improved Qwen-Image-Edit-2511 and Flux-2-Klein-9B across three benchmarks, but required about ten times the compute of a single editor; its results also depended on which tools were connected. The system generated about 26,000 verified training examples, and editors fine-tuned on that data performed better across the three benchmarks.
Written from coverage as of
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh