Adobe Tests AI Agents That Can Edit Images Like Professional Photoshop Users

Adobe Research and the University of California, San Diego have developed EditAgent, an AI system designed to perform image-editing tasks more like a professional Photoshop user. Instead of repeatedly asking one image-generation model to fix an image, EditAgent breaks instructions into smaller tasks, selects specialised editing tools, generates multiple versions and uses measurable evidence to choose the best result.

The research, published on October 29, 2026, addresses a major weakness in AI image editing: models can produce attractive images but often struggle with precise adjustments, preserving existing details and following multiple instructions consistently. In tests involving 600 image-editing tasks, professional reviewers preferred parallel-generated edits over sequential refinements in 78.5% of comparisons where they expressed a preference. The findings could influence how Adobe develops AI assistants across Photoshop and its wider Creative Cloud ecosystem.

Key Takeaways

• Adobe Research and UC San Diego developed an AI image-editing system called EditAgent.
• The system breaks complicated editing instructions into smaller steps.
• It combines general-purpose image editors with specialised tools.
• It generates multiple editing options in parallel instead of repeatedly modifying one result.
• A language model evaluates the options using image measurements, change maps and close-up crops.
• Professional reviewers preferred parallel editing results in 78.5% of comparisons where they expressed a preference.
• The system improved two open-weight image editors across three benchmarks.
• EditAgent requires approximately 10 times more computing resources than a single editing model.
• Researchers generated approximately 26,000 verified editing examples to train other models.

What Is Adobe’s EditAgent?

EditAgent is an AI agent designed to make image editing more reliable and precise.

Traditional AI image editors typically work by receiving a natural-language instruction and generating an edited image. If the result is incorrect, the system may attempt another edit using the same model.

This approach can become inefficient when a task requires several precise adjustments.

For example, a user might ask an AI editor to brighten a photograph, sharpen a person’s face, remove an unwanted object and preserve the original background.

A general-purpose model may complete some of these tasks successfully but struggle with others. Repeatedly generating new versions can also introduce unwanted changes.

EditAgent takes a different approach.

It breaks the request into smaller operations and identifies the most appropriate tool for each operation.

The system can combine a general-purpose image editor with specialised methods for tasks such as sharpening, rotation and zooming.

It then evaluates the available results before deciding which version should be retained.

The objective is to make AI editing more controlled, measurable and consistent.

Why Existing AI Image Editors Struggle

The researchers tested a general-purpose image editor called Flux-2-Klein-9B across 600 image-editing tasks.

Five professional reviewers assessed the resulting images without being told which editing approach produced them.

The findings revealed important limitations.

The general-purpose editor failed to produce a usable result in 42.7% of the tested cases.

The researchers also found that specialised tools could solve approximately two-thirds of the failures, particularly in tasks involving sharpening, rotation and zooming.

Simply asking the original model to try again did not consistently solve the problem.

This suggests that image-editing agents need more than the ability to generate new versions. They also need to recognise when a different tool or editing method would be more appropriate.

For professional designers, the distinction is important.

An image that looks visually appealing may still be unsuitable if it changes the wrong object, damages important details or fails to follow the original instructions.

How EditAgent Works

EditAgent introduces two important changes to conventional AI image-editing workflows.

1. It Breaks Complex Instructions Into Smaller Tasks

The system first analyses the user’s request and divides it into individual editing operations.

For example, an instruction to improve a portrait might involve:

• Adjusting brightness and exposure.
• Correcting colour temperature.
• Sharpening important details.
• Removing unwanted objects.
• Preserving the original composition.

The agent then selects appropriate editing methods for each operation.

Instead of relying exclusively on one general-purpose model, it can use specialised tools when they are better suited to the task.

2. It Generates Multiple Versions in Parallel

Traditional AI editing agents often follow a sequential process.

They generate an image, inspect the result and make another change based on the already modified version.

The problem is that errors can accumulate.

EditAgent instead generates multiple candidate results from the original image in parallel.

A separate language model compares the alternatives and identifies which result best satisfies the instruction.

This allows the system to compare different editing methods without repeatedly building on potentially flawed outputs.

The researchers found that professional reviewers preferred parallel-generated results in 78.5% of comparisons where they expressed a preference over sequential refinement.

AI Models Use Measurements to Judge Image Quality

One of the most interesting parts of EditAgent is how it evaluates editing results.

Rather than relying only on a language model’s visual impression, the system provides additional information about each candidate.

Its verification records include:

• Close-up crops: Enlarged sections showing the edited area in greater detail.
• Change maps: Visual representations of which parts of the image were modified.
• Image measurements: Numerical information about brightness, colour and other visual properties.

The language model uses these records to compare different versions.

For example, if a user asks for a photograph to become slightly brighter, the system can examine brightness measurements and colour changes to determine whether an edit has overexposed the image or introduced an unwanted colour shift.

The researchers found that providing this additional verification information significantly improved the language models’ ability to judge image edits.

Without the verification records, the seven language models tested could not match the strongest specialised evaluation model.

With the additional information, all seven outperformed the specialised evaluation models.

The Qwen3.8-27B model achieved an 81.3% usable selection rate, compared with 74.8% for EditScore-72B, the strongest fine-tuned evaluation model in the comparison.

This suggests that improving the information available to an AI evaluator can be as important as training a larger or more specialised evaluation model.

EditAgent Produces Better Results Across Multiple Benchmarks

The researchers tested EditAgent using two open-weight image editors:

• Qwen-Image-Edit-2511
• Flux-2-Klein-9B

The system improved their performance across three image-editing benchmarks and outperformed seven other AI editing agents included in the evaluation.

In one comparison on the EdiVal benchmark, the Qwen-based EditAgent configuration in its third editing round slightly outperformed the commercial Seedream 4.0 and GPT-Image-1.5 models.

These results demonstrate the potential of combining different editing tools with structured evaluation.

However, the performance should be understood within the benchmarks and experimental conditions used by the researchers. It does not establish that EditAgent is better than every commercial image editor in all real-world editing tasks.

Researchers Also Use EditAgent to Train Better AI Editors

The system has another potential application beyond editing images.

Because EditAgent evaluates multiple outputs and retains verified results, it can generate useful training data for other AI models.

The researchers used it to produce approximately 26,000 verified image-editing examples.

They then used these examples to fine-tune image editors.

The resulting models performed better across all three benchmarks.

Interestingly, training with data generated directly by FLUX made FLUX’s performance worse in the researchers’ experiments.

This indicates that the quality of training examples matters. Simply generating more synthetic images does not necessarily produce a better editing model.

Verified examples selected through a structured evaluation process could provide more useful training material.

The Biggest Limitation: Computing Costs

Despite its improvements, EditAgent is computationally expensive.

The system requires approximately 10 times more computing resources than running a single image-editing model.

This is because it generates multiple candidates, evaluates them and performs additional verification.

That could make the approach difficult to deploy for every routine editing request, particularly in applications where users expect near-instant results.

The researchers see an opportunity to use EditAgent as a data-generation system.

It could perform expensive, high-quality editing and evaluation during development, producing verified examples that can subsequently train smaller and faster models.

Those models could then deliver improved editing performance with lower computing requirements.

Why This Research Matters for Adobe Photoshop

Adobe has been expanding its AI capabilities across Photoshop, Premiere, Illustrator and other Creative Cloud applications.

In June 2026, the company announced a wider rollout of its Creative Agent, allowing users to describe desired outcomes while AI assistants coordinate multi-step workflows across creative applications.

Photoshop’s AI Assistant can already help users perform operations such as replacing backgrounds, adjusting lighting and removing objects.

EditAgent’s research addresses a related challenge: ensuring that AI agents select the right tools and accurately verify their results.

If these techniques become part of commercial creative software, users could gain more reliable automated editing while retaining control over the final image.

However, EditAgent is a research system, and Adobe has not announced that this specific technology is being integrated into Photoshop as a commercial feature.

The Bigger Picture

The research reflects a shift in AI-powered creative software.

Early image-generation tools focused primarily on producing new images from text descriptions. The next stage involves giving AI systems the ability to understand existing images, operate professional editing tools and make precise changes without damaging other parts of the composition.

EditAgent demonstrates why tool selection and result verification are important components of this development.

For professional designers, photographers and creative agencies, the potential value is not simply generating images faster. It is reducing repetitive editing work while maintaining precision and control.

Looking Ahead

The next challenge will be reducing EditAgent’s computing requirements while maintaining the improvements demonstrated in the research. Wider testing across different image editors, editing tasks and professional workflows will also help establish how reliably the approach works outside controlled experiments.

For Adobe, the research could inform future generations of Creative Cloud AI assistants. If AI agents become better at choosing specialised tools, comparing alternatives and verifying their own work, creative professionals may be able to delegate more complicated editing operations without giving up control over the final result.