Adobe tests EditAgent to make AI image edits choose the right tool

Adobe Research and UC San Diego have tested EditAgent, an AI system designed to handle image edits more like a skilled Photoshop user would. Instead of repeatedly asking one general editor to fix a picture, it breaks a request into steps, tries specialized tools where appropriate, and uses measured evidence to decide which result is best.

That matters for creators and small teams because today’s chat-style image editors can be convenient but inconsistent. The research points to a future where AI editing is more precise, though the paper also says this approach requires far more compute than a single editor.

EditAgent treats image editing as a tool-selection problem

According to The Decoder’s summary of the paper, EditAgent combines a general-purpose image editor with task-specific tools for operations such as sharpening, rotating, and zooming. The agent plans an edit as a sequence of smaller jobs, then decides which editor or specialized tool should handle each one.

The researchers first looked at why a single general editor is not enough. In a 600-edit test reviewed blindly by five professionals, Flux-2-Klein-9B failed to deliver a usable result 42.7 percent of the time. Specialized tools fixed roughly two-thirds of those failed cases, with the strongest gains reported for sharpening, rotation, and zooming. Running more attempts through the same general editor did not help much.

For working designers, marketers, and freelancers, the practical takeaway is simple: prompt-based editing may improve when an agent can call the right instrument for the job, rather than forcing one model to do every operation. That is especially relevant for product images, social graphics, and campaign assets where a small unwanted change can ruin consistency.

Parallel variants beat step-by-step repair

The paper also compares two editing strategies: refining an image over multiple rounds, or generating multiple options from the original image in parallel. The researchers found that iterative editing can compound small mistakes because every later step starts from an already changed image.

Professionals preferred parallel variants made from the original in 78.5 percent of the comparisons where they chose a winner. The agent also needed only half as many evaluation passes in that setup, according to the source.

EditAgent builds on this by producing several candidate versions for each step and judging them together. For each option, the judging model receives a verification record that includes:

• a close-up crop of the edited area;
• a change map showing which regions were altered;
• measured brightness and color statistics.

In one example described in the source, an image was meant to be brightened gently. The measurements showed that the general editor overexposed the picture and made the colors warmer, while the specialized tool stayed within the requested bounds.

Standard language models became stronger judges with better evidence

A notable part of the research is that the judge does not have to be a purpose-built image-editing evaluator. The paper tested standard language models against fine-tuned evaluation models such as EditScore.

Without the verification record, none of the seven language models tested matched the best fine-tuned evaluator. With the record, all seven surpassed the fine-tuned models. The source reports that Qwen3.8-27B selected a usable result 81.3 percent of the time, compared with 74.8 percent for EditScore-72B.

That result is important beyond Adobe’s own research. It suggests that AI agents may improve not just by using larger models, but by giving existing models better structured evidence. This is a recurring theme in agentic software, where the system around the model can matter as much as the model itself. We have seen similar interest in agent infrastructure across other areas, from app automation to connected AI workflows, including recent moves such as Google Agent Compute Engine.

The catch is compute cost, but the data may pay off

EditAgent improved the open-source editors Qwen-Image-Edit-2511 and Flux-2-Klein-9B across three benchmarks, according to the source, and performed better than seven other AI agents. On EdiVal, the Qwen variant in its third editing round slightly passed the commercial models Seedream 4.0 and GPT-Image-1.5.

But the tradeoff is significant: the agent uses about ten times as much compute as running a single editor. Its quality also depends on which tools are available for it to use. That makes it less like a lightweight chat editor and more like a verification-heavy production pipeline.

The researchers argue that the compute may still be useful because EditAgent can generate training data. Since it keeps only verified outputs, it produced about 26,000 training examples. Editors fine-tuned on that data improved across all three benchmarks, while training data generated directly by FLUX made FLUX perform worse, according to the source.

Adobe has already added AI agents to Photoshop, Premiere, and other Creative Cloud apps in a public beta that began in June, where users describe a goal and the agent handles the steps. For Adobe users, including those already working with Adobe Firefly for image generation and editing workflows, the research raises a concrete question: how well can an assistant select tools, check its own work, and avoid undoing earlier edits?

The paper’s example of this problem is a request to “let in more daylight,” which accidentally makes a previously frosted window clear again. EditAgent detects the conflict and corrects it. That kind of self-checking may be as important as the edit itself, especially for brand assets and client work where accuracy matters more than speed.

For readers comparing creative AI tools, this research is also a reminder that image quality is not only about generation. Editing, verification, and repeatability are becoming core differentiators. See Emonarc’s Best AI Image Generators in 2026 guide for broader context, and our coverage of Adobe Lightroom mobile’s AI fix for closed eyes for another example of Adobe applying AI to practical photo-editing tasks.