Editorial illustration for Adobe Tests AI Agent EditAgent Using Editing Tools Like Photoshop Pros
Adobe's EditAgent AI Masters Photoshop Like Pro
Adobe Tests AI Agent EditAgent Using Editing Tools Like Photoshop Pros
Adobe Research and UC San Diego have built an AI system called EditAgent that treats image editing less like a chatbot task and more like a job for someone who actually knows Photoshop. Instead of running every request through one general-purpose model and hoping for the best, EditAgent breaks a complex edit into sub-steps and hands each one to a specialized tool, whether that's sharpening, zooming, or something else entirely. A standard language model then checks the results, but not by just looking at them. It compares measured statistics and close-up crops across image variants generated in parallel, a method that reportedly beats evaluation models built specifically for that job.
The tradeoff is compute. EditAgent produces noticeably more accurate results than earlier AI agents, but it needs roughly ten times the processing power to get there. Adobe's researchers see a workaround: use EditAgent's verified output as training data for smaller, faster models down the line. The bigger question the paper raises is why general-purpose editors keep stumbling on tasks a dedicated tool would handle easily, and how often that failure actually happens in practice.
EditAgent improved the open-source editors Qwen-Image-Edit-2511 and Flux-2-Klein-9B across three benchmarks and outperformed seven other AI agents. On the EdiVal benchmark, the Qwen variant in its third editing round edged past the commercial models Seedream 4.0 and GPT-Image-1.5. But the agent needs about ten times as much compute as a single editor, and its performance depends on which tools are plugged in.
Why this matters
EditAgent is a useful data point for anyone building on top of open-source image editors rather than just marveling at them. The fact that a general-purpose Qwen variant, wrapped in EditAgent's sub-step planning and parallel evaluation, caught up to Seedream 4.0 and GPT-Image-1.5 by round three on EdiVal tells us the bottleneck in image editing isn't always the base model, it's the orchestration around it. That's good news for teams without access to frontier commercial APIs: a smart scaffolding layer, specialized tools for sharpening or zooming, and a language model doing evaluation instead of a bespoke critic model can close gaps that used to require more parameters or more training data.
We'd push back on reading too much into "outperformed seven other AI agents" without knowing the compute cost of running parallel variants and crop-based evaluation at each step, that's the kind of detail that decides whether this is production-ready or a research curiosity. Still, the core idea, that an LLM can judge pixel-level edits using measured statistics rather than a trained scorer, is worth watching if you're designing agent pipelines for any visual task, not just Photoshop work.
Related Reading
Research & Benchmarks Microsoft AI Chief: "If It Isn't Safe We Shouldn't Build It
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
Related Topic Adobe to buy Semrush for USD 1.9 bn, merging AI tools with SEO platform
Business & Startups
Related Topic Adobe Firefly AI Assistant runs Photoshop, Premiere, Illustrator from one prompt
AI Tools & Apps
From Archives Study: Chinese AI models shift pro-government when prompted in Chinese
From Archives Google's Gemini video AI analyzes scenes to understand what people need at work
From Archives Open Source Suite Clones Adobe's Photoshop, Premiere, Lightroom
Open Source
From Archives Adobe’s Indigo app adds AI Playground with generative photo editing
LLMs & Generative AI
Common Questions Answered
How does EditAgent differ from traditional AI chatbot approaches to image editing?
EditAgent breaks down complex image editing tasks into specialized sub-steps rather than running requests through a single general-purpose model. Each sub-step is handed to a specialized tool like sharpening or zooming, with a standard language model checking the results afterward, mimicking how professional Photoshop users actually work.
What benchmarks did EditAgent use to measure its performance against other AI agents?
EditAgent was tested across three benchmarks and compared against seven other AI agents. On the EdiVal benchmark specifically, the Qwen variant of EditAgent outperformed commercial models Seedream 4.0 and GPT-Image-1.5 by the third editing round.
What are the computational limitations of EditAgent compared to single image editors?
EditAgent requires approximately ten times as much compute as a single editor to perform its tasks. This increased computational demand is a trade-off for the improved performance and accuracy achieved through its multi-step orchestration approach.
Why is EditAgent's success significant for teams using open-source image editors?
EditAgent demonstrates that the bottleneck in AI image editing isn't always the base model itself, but rather the orchestration around it. This finding is encouraging for teams without access to frontier commercial APIs, as they can improve results by optimizing how they coordinate available tools rather than needing more powerful base models.
Which open-source image editors did EditAgent improve according to the research?
EditAgent improved the open-source editors Qwen-Image-Edit-2511 and Flux-2-Klein-9B across the three benchmarks tested. Both variants showed significant performance gains when wrapped in EditAgent's sub-step planning and parallel evaluation system.
Further Reading
- EditAgent: Rethinking Agentic Image Editing - arXiv
- More time to spend on your craft: Adobe brings the power of its creative agent to Creative Cloud apps - Adobe Blog
- Adobe Delivers New AI Innovations, Assistants and Models for Creative Professionals - Adobe Newsroom
- Image editing just got smarter with AI in Photoshop and Firefly - Adobe Blog
