Adobe researchers build an AI photo editor that checks its own work with measurements

Researchers at Adobe, the San Jose software company, and UC San Diego have built an AI system that edits photos more like a professional, according to a research paper posted Oct. 29.  The paper has not yet been reviewed by outside experts, and the system, called EditAgent, is a research project, not a product.

Today's AI image editors, like those built into chatbots, take a typed request such as "brighten this photo" and change the picture.  Some act as agents, breaking a big request into steps.  But when a step fails, they usually retry with the same tool, and a second AI judges the result by simply looking at it, the researchers wrote.

EditAgent works more like a team.  A planner splits a request into ordered steps, and a locating tool finds the exact object or area to change.  It then runs several tools on the same photo at once:  a general AI editor plus specialists that remove or move objects, resize a subject, recolor an area, adjust brightness, contrast and color, apply styles and filters, rotate or flip the image, zoom in or extend the edges of the picture.

For each attempt, it builds what the paper calls a verification record:  a map of what changed, before-and-after close-ups and measured brightness, contrast, saturation and color shift.  An off-the-shelf AI judge compares all the attempts at once, picks one or sends feedback, with up to three tries per step.  After each round, it also checks that a new edit did not undo an earlier one.  In one example, a request to "let in more daylight" erased a frosted window added in an earlier step.

In 600 test edits of 30 kinds, rated blind by five professional Photoshop retouchers, a general-purpose editor alone produced nothing usable 42.7% of the time, the paper says.  Specialist tools rescued 175 of those 256 failures, mostly rotations and other geometry, lighting and touch-up jobs.  With the verification records, the judge picked a usable edit 81.3% of the time, versus 74.8% for a larger judge trained just for this task.  On a public test of multi-step editing, EditAgent's best version outscored commercial editors from ByteDance and OpenAI, the researchers reported.

The catch is cost.  EditAgent takes roughly 10 times as long to run as a single editing model, the researchers wrote.  On one test, it used 12.9 hours of time on Nvidia H100 chips, high-end chips widely used for AI, compared with 1.1 hours for the editor alone.  Chip time is a main cost of running AI services, so a tool that needs 10 times as much is harder to offer cheaply to millions of people.  But other research agents cost about the same, 9.4 to 14.7 hours on the same test, and the researchers argue the cost can pay for itself:  about 26,000 edits EditAgent approved were used to train single, faster editors, which then improved on all three tests.  The Decoder, a tech news site, said that tradeoff could make the extra computing worth it over time.

Ordinary users are unlikely to get EditAgent itself soon.  The project page says its code is coming "soon," for researchers.  What people can try now is Adobe's AI Assistant, a test version in Photoshop, Premiere, Illustrator and other apps since June.  Users describe the result they want, such as swapping a background or resizing images for every social media site, and the assistant carries out the steps, with results that stay adjustable, Adobe said.  The paper does not say whether EditAgent's methods will reach Adobe's products.

The paper does not address jobs.  The U.S. Bureau of Labor Statistics projects that jobs for graphic designers, a related field, will fall 2% from 2025 to 2035, to about 249,000 from 253,100, while jobs overall grow 3%.  "Automated design tools, such as artificial intelligence (AI), are projected to make graphic designers more productive and reduce the need for these workers," the agency says.  It still expects about 16,000 openings a year as workers retire or change careers.  In the study, human professionals remained the judges of which edits were usable.

Source:  arXiv preprint, Oct. 29, 2026, EditAgent project page, The Decoder, Nov. 1, 2026, Adobe, June 18, 2026, Adobe Help, June 23, 2026 and U.S. Bureau of Labor Statistics, Occupational Outlook Handbook.