Adobe's EditAgent edits like a Photoshop pro: try several tools, then check the numbers

The trick to a better image-editing agent turns out to be the same one good engineers use: try more than one approach, and measure the result instead of eyeballing it.

EditAgent comes from a team at Adobe Research and UC San Diego, and its new paper is covered by The Decoder's Jonathan Kemper. Typical editing agents break a request into steps, rerun the same editor when a step fails, and let a second model rate the result visually.

What they found

Across a 600-edit benchmark scored blind by five pros, Flux-2-Klein-9B, the all-purpose editor in the test, delivered an unusable result in 42.7% of cases. Dedicated tools fixed roughly two in three of those misses, with sharpening, rotation and zoom the biggest wins, while extra retries with the same editor barely helped. Generating several variants in parallel from the original also beat refining one image round after round: pros preferred the parallel versions 78.5 percent of the time when they had a preference, and it took half as many evaluation passes.

Benchmarks vs. vibes

The interesting result is about judging. Every candidate ships with a zoomed crop of the changed region, a diff map, and measured brightness and color statistics. Without that record, none of the seven off-the-shelf language models tested matched the best judge trained on human ratings. With it, all seven beat it. Qwen3.8-27B picked a usable result 81.3 percent of the time against 74.8 percent for EditScore-72B. The agent also checks that later edits don't quietly undo earlier ones.

The costs are real: about ten times the compute of a single editor. The researchers offset that by using the agent's verified results, about 26,000 of them, to train faster editors, which improved on every benchmark.

The lesson travels well beyond Photoshop: give your evaluator evidence, not just the output.

Scoreboard update: AI +1. Agents learn to check their work like a Photoshop pro.

Stay undead.