Adobe tests an AI that edits photos by picking the right tool

Adobe Research and UC San Diego have built an AI system called EditAgent that edits photos the way a Photoshop pro might: by choosing the right tool for each job. In their test, a basic AI photo editor failed to give a usable result 42.7 percent of the time. Specialized tools fixed about two-thirds of those failures.

What changed

Most AI editing agents split a request into steps. When a step fails, they usually retry it with the same tool. A separate AI model then judges the result by looking at it. EditAgent works differently. For each step, it compares a general-purpose editor with specialized tools built for jobs like sharpening, rotating and zooming.

The team tested 600 image edits, rated blind by five professionals. The general-purpose editor, Flux-2-Klein-9B, often fell short, and extra retries barely helped. Switching tools did.

How it checks its own work

Instead of just looking at each result, the judging AI gets a record of hard numbers. It includes a close-up of the edited area, a map of what changed, and measured brightness and color statistics. In one example, a photo was meant to be gently brightened. The numbers showed the general editor overexposed it and made the colors warmer. The specialized tool stayed within the request.

• With this record, all seven language models tested beat every model trained on human ratings of edits.
• Qwen3.8-27B picked a usable result 81.3 percent of the time, versus 74.8 percent for EditScore-72B.
• Professionals preferred variants made in parallel from the original photo in 78.5 percent of the comparisons where they picked a side.

EditAgent also checks whether a later edit undid an earlier one. In one example, asking to "let in more daylight" made a frosted window clear again. The agent caught the conflict and fixed it.

The catch

The system needs about ten times as much computing power as a single editor. Its results also depend on which tools are plugged in. The researchers note that the agent keeps only verified results, which gave them about 26,000 training examples. Editors trained on that data did better across all three benchmarks, which could make the cost worth it over time.

Why it matters for Adobe users

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 steps. This research tackles a question that applies directly to those products: how well an agent picks tools and checks its own output. The material does not say whether EditAgent will ship in any Adobe product. Anyone editing photos for work should still review the results by eye.