Tools · News
Adobe researchers found that better AI editing often means changing tools, not trying again
EditAgent uses specialized editing methods and visual evidence to improve complex image edits—and professional retouchers preferred its approach to repeated refinement.
Reporting updated Nov 2, 2026
THE TAKEAWAY
Researchers from Adobe Research and UC San Diego have developed EditAgent, a system that gives AI image-editing agents a larger toolbox and a better way to judge their own work. Instead of repeatedly asking the same model to fix a failed edit, it tries different editing methods and compares the results with visual evidence.
WHAT HAPPENED
Imagine asking an AI editor to move a person slightly to the left, widen the shot, remove someone in the background and change the lighting to sunset.
The first result looks good. But when the editor fixes the background, it changes the person's face. Another attempt repairs the face but loses the lighting.
Anyone who has worked through multiple rounds of AI image editing has probably experienced something similar.
EditAgent approaches that problem differently.
In a paper published October 29, researchers examined how AI editing systems choose their tools and evaluate the changes they produce.
Their findings were revealing.
On a benchmark containing 600 sets of image-editing candidates across 30 types of tasks, the general-purpose editor FLUX.2-klein-9B failed to produce a usable candidate in 42.7% of the sets.
Specialized editing methods recovered usable results in 175 of those 256 failures.
The lesson was straightforward: sometimes the problem isn't the prompt. It's the tool.
EditAgent addresses this by giving an editing agent access to several approaches, including masked edits, object removal, repositioning, color adjustments and geometric operations.
Instead of repeatedly editing the previous result, the system can generate alternatives from the same source image.
A visual-language model then compares those candidates, assisted by close-up views, maps showing what changed, and measurements of properties such as color and geometry.
WHY IT MATTERS
The researchers also investigated whether an AI editor should repeatedly refine one image or generate different alternatives.
In a blind comparison, five professional retouchers preferred the parallel approach in 78.5% of the cases where they chose a winner. It required no more than half as many judging calls.
For creators, that is a useful insight.
An intelligent editing assistant should recognize when a problem needs a better tool, not merely another prompt.
It should also understand what must remain unchanged. If you're moving an object, you don't want the entire background redesigned. If you're adjusting brightness, you don't want a character's features altered.
EditAgent brings those priorities closer to the way experienced artists work.
The research reports improvements across three image-editing benchmarks using Qwen-Image-Edit-2511 and FLUX.2-klein-9B.
It is a research system, however, not an announced new Photoshop feature. The additional candidate generation and verification also introduce computational costs that would matter in a production application.
Focused takeaway
Break down the request
Separate complex edits into manageable tasks.
Choose the right tools
Route operations to general or specialized editors.
Compare alternatives
Generate candidates in parallel rather than repeatedly changing one image.
Inspect the evidence
Evaluate close-ups, change maps and measured differences.
Preserve earlier work
Check that later changes don't undo previously accepted edits.
The researchers distinguish between what an editing agent can do and what its visual judge can inspect.
Source excerpt
VEYR ANGLE
AI editing doesn't just need models that create better images.
It needs tools that understand what the creator is trying to preserve.
Sometimes the smartest editing decision is knowing when to stop asking the same model to try again.
