Key points
- Adobe Research and UC San Diego built EditAgent, which splits edit requests into steps and tries both a general editor and specialized tools on each one.
- In a 600-edit test, the general-purpose editor Flux-2-Klein-9B gave unusable results 42.7 percent of the time; specialized tools fixed about two-thirds of those misses.
- With measured data supplied, all seven standard language models beat every fine-tuned judge; Qwen3.8-27B picked a usable result 81.3 percent of the time.
- The agent needs about ten times the compute of a single editor, but its verified outputs yielded roughly 26,000 examples for training faster editors.
Why it matters for marketers: Agentic editing quality may depend more on tool choice and measurable checks than on one bigger model, which affects how teams judge AI image tools and review workflows.
Adobe Research and UC San Diego have published a paper on EditAgent, an image editing agent that behaves less like a single prompt box and more like a Photoshop professional. It splits a request into steps, then for each step runs a general-purpose editor against specialized tools for jobs such as sharpening, rotation and zooming.
A standard language model then judges all the variants at once. Instead of just looking at the images, it receives a close-up of the edited area, a map of what changed, and measurements of brightness and color. The researchers report that with this record, all seven language models they tested outperformed judges fine-tuned on human ratings of edits. The agent also checks whether a later edit undid an earlier one.
For marketers and content teams, the finding is practical. In the team's test, the general-purpose editor failed often, and simply retrying with the same tool barely helped. Professionals also preferred generating several variants from the original image over refining one image repeatedly, since small errors pile up. Teams evaluating AI editing tools may want to ask how outputs are verified, not only which model sits underneath.
There are tradeoffs. EditAgent uses about ten times the compute of a single editor, and its results depend on which tools are connected. The researchers also used it to generate about 26,000 verified training examples, and editors tuned on them improved across all three benchmarks. The Decoder notes that Adobe added agents to Creative Cloud apps in a public beta in June, so how well agents pick tools and check their work bears directly on shipping products.
Sources
This brief is our own summary. Read the full reporting at the source.
