- Adobe Research and UC San Diego report that one general-purpose editor, Flux-2-Klein-9B, gave no usable result in 42.7% of the 600 edits in the authors' benchmark.
- Their agent gives its judge measurements, not just the image. The researchers say this made off-the-shelf language models better judges than models fine-tuned for the task.
- The agent needs about ten times the compute of one editor. If you evaluate agents that act for you, ask how they verify output, and what that costs.
A judge's evidence can outweigh its schooling
An AI that edits photos has a quiet weakness. When one attempt fails, it usually tries again with the same narrow set of tools. A second model then judges the result by looking at it.
A new paper from Adobe Research and UC San Diego argues that both habits cap what these agents can do. The authors call the limits the agent's action space, meaning what it can do, and its observation space, meaning what its judge can see. Their system, EditAgent, widens both.
The wider lesson is the publication's argument, not a finding of the paper. The paper covers image editing only. But if you are evaluating agents that act for you, it suggests asking what the agent's checker actually sees.
What the researchers measured
The team built a benchmark of 600 image edits. Five professionals rated the results without knowing which system made them.
One editor, Flux-2-Klein-9B, failed to give a usable result in 42.7 percent of those edits, according to The Decoder. The paper's abstract puts the failure rate for a general-purpose editor at over 40 percent. It also says no judge model was reliable across all types of edit.
The Decoder reports that adding specialised tools fixed about two out of every three of these failures. The gains were biggest for sharpening, rotation and zooming. Simply asking the same editor to try again added little.
How EditAgent works
EditAgent splits a request into steps. For each step it sends the job to the general editor and to matching specialised tools. This widens what the agent can do.
It then attaches a verification record to every candidate image. According to The Decoder, each record includes three things. One is a zoomed view of the edited area. Another is a map of which regions changed. The third is measured brightness and colour figures.
A standard language model then compares all candidates in one pass. In one example from the paper, a photo was meant to be brightened gently. The numbers showed the general editor had overexposed it and shifted the colours warmer. The specialised tool stayed within the request.
The agent also makes its candidates side by side from the original image. The authors say the alternative, improving one image round after round, lets small errors build up. Each round starts from a picture that has already been changed. Among comparisons where experts chose a side, they favoured the side-by-side method 78.5 percent of the time. It used no more than half the judge calls.
One more check runs after each round. It looks for later edits that undo earlier ones. In the paper's example, the instruction "let in more daylight" made a frosted window clear again, and the agent caught it.
Measurements versus training
The most useful finding concerns the judges. The Decoder describes how the test worked. Each model was shown several candidate edits and had to name the best one. Scoring then asked a simple question: was the chosen image one the professionals had also called usable?
They compared seven off-the-shelf language models with models fine-tuned on human ratings of edits, such as EditScore. When the language models lacked the verification record, none matched the best fine-tuned model. Once they had it, every one of the seven scored higher than every fine-tuned model.
Qwen3.8-27B picked a usable image 81.3 percent of the time. EditScore-72B reached 74.8 percent. The Decoder says the researchers read this as a sign that a judge's inputs can count for more than its training.
The point holds for this test only. It is an argument for building evidence into the process, not a claim about every kind of AI system.
The costs and limits
These are the authors' own results, in a paper submitted to arXiv on October 29. The Decoder reports that EditAgent needs roughly ten times the compute of a single editor. Its performance also depends on which tools are plugged in.
There may be a payoff. The agent keeps only results it has verified, and the researchers used those to build a training set of about 26,000 examples. Editors fine-tuned on that data improved across all three benchmarks. Training data generated directly by FLUX made FLUX worse.
What to ask your teams and vendors
First, ask how any AI agent you pay for checks its own work. If the answer is that another model looks at the output, ask what that model can see.
Second, ask what happens after a failure. An agent that retries the same step with the same tool is repeating its mistake.
Third, ask what verification costs. Ten times the compute is a real price. Compare it with the cost of a wrong result reaching a customer.
A checker that can only look is working from appearances. A checker with measurements has something to show.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Adobe Research and UC San Diego.
