Adobe EditAgent: Why Measured Evidence Beats Trained AI Judges

In 600 image edits rated blind by five professionals, a general-purpose AI editor failed to produce a usable result 42.7% of the time, and retrying with the same editor barely helped. Adobe Research and UC San Diego’s EditAgent fixes much of that by working like a Photoshop professional: pick the right tool, then check the result with numbers. For anyone building agents that judge their own output, the bigger finding is about the judge.

How It Works

The Decoder reports on the research paper. The team started with two findings. Specialized tools solved roughly two-thirds of the general editor’s failures, especially for sharpening, rotation and zooming. And generating several variants from the original in parallel beat refining one image step by step, which piles up small errors. Professionals preferred the parallel variants in 78.5% of comparisons where they picked a side, and the agent needed half as many evaluation passes.

EditAgent plans a request in steps. For each step, it runs the general editor against matching specialist tools and has a standard language model compare all the variants in one pass. The judge gets a verification record for each: a close-up of the edited area, a change map of altered regions, and measured brightness and color statistics. In one example, a photo meant to be gently brightened was overexposed and shifted warmer by the general editor, while the specialist tool stayed within spec.

After each round, EditAgent checks whether later edits undid earlier ones. Asked to “let in more daylight”, one edit made a previously frosted window clear again, and the agent caught and fixed the conflict.

What This Means for You

The headline result is about evaluation. Without verification records, none of the seven language models tested could match the best model fine-tuned on human ratings of edits. With them, all seven beat every fine-tuned model. Qwen3.8-27B picked a usable result 81.3% of the time, against 74.8% for EditScore-72B. The researchers’ conclusion: what a judge model sees can matter more than how it was trained.

The costs are real. EditAgent needs about ten times the compute of a single editor, and its performance depends on which tools are plugged in. The researchers turn that into an asset: because it keeps only verified results, the agent produced about 26,000 training examples, and editors fine-tuned on them improved on all three benchmarks. Training data generated directly by FLUX made FLUX worse.

It matters beyond research. Adobe added agents to Photoshop, Premiere and other Creative Cloud apps in June as a public beta, and how well those agents choose tools and check their own output is the same question. Stack Archive’s look at AWS’s Agentic AI Lens as a Claude skill found the same pattern in another domain: structured evidence and enforced rules made a general model a more reliable reviewer.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.