EditAgent: Adobe's agent edits images like a Photoshop pro

What's wrong with today's editing agents?

Agentic image editors split a request into steps, but when a step fails they mostly retry with the same tool. A judge model scores the result, usually just by looking at it. According to a new paper from Adobe Research and UC San Diego, that creates two limits: what the agent can do, and what its judge can see.

The researchers built a benchmark of 600 editing sets across 30 edit types, blind-rated by professional Photoshop editors. The general-purpose Flux-2-Klein-9B editor produced no usable result in 42.7% of sets. A specialized method recovered 175 of those 256 failures, mostly in geometry, tone and enhancement edits.

How EditAgent works

• Each step goes to several methods at once: the general editor plus matching specialized tools such as rotation, zoom or brightness adjustment.
• Variants are generated in parallel from the original image instead of building on each other. Experts preferred this over step-by-step refinement in 78.5% of decided comparisons, with half the judge calls.
• Every variant gets a verification record: a change map, a close-up of the edited region, measured brightness and color statistics, and a list of facts checked by code.
• An off-the-shelf VLM compares all candidates in a single call, with no fine-tuning.

The most interesting result concerns the judge. Without verification records, none of the seven general models tested reached the best fine-tuned scorer, EditScore-72B (74.8%). With records, all seven beat it. Qwen3.8-27B reached 81.3% and Qwen3.8-Flash-Next 82.8%.

Source: Adobe Research, UC San Diego (EditAgent, Figure 5). Hit rate of seven off-the-shelf VLM judges without and with verification records, next to fine-tuned reward models.

The trade-off

EditAgent improved the open Qwen-Image-Edit-2511 and Flux-2-Klein-9B editors on three benchmarks (EdiVal, Complex-Edit, ComplexBench-Edit). The cost is speed: the authors say it needs roughly ten times the inference time of a standalone editor, and its performance depends on the tools plugged in.

That is partly offset by its use as a data engine. Because it only accepts verified edits, editors fine-tuned on its output improved on all three benchmarks. According to The Decoder, training data generated directly by Flux made Flux worse (The Decoder). Adobe added agents to Photoshop and other Creative Cloud apps in June, so the research speaks directly to its own products.

The lesson for developers goes beyond image editing: an LLM-as-a-judge often needs better inputs, not more training. With measurements and targeted crops, a general model can be a more reliable judge than one trained for the job.