Adobe’s EditAgent improves image editing with professional workflows and AI tools
Adobe Research and UC San Diego have developed EditAgent, an AI system designed to refine image editing by replicating professional workflows in Photoshop. The system combines a general-purpose editor with specialized tools and quantitative verification to improve accuracy, reducing errors compared to traditional AI editors.
Editor, Lazyfounder
30 SEC SUMMARY
- Adobe Research and UC San Diego developed EditAgent, an AI system for image editing that mimics professional Photoshop workflows.
- EditAgent uses specialized tools and quantitative verification to improve accuracy over general-purpose AI editors.
- Tests show Flux-2-Klein-9B failed 42.7% of edits, but specialized tools resolved two-thirds of those failures.
- Professionals preferred EditAgent’s outputs in 78.5% of comparisons, and Qwen3.8-27B achieved 81.3% accuracy in selecting usable edits.
TABLE OF CONTENTS
KEY HIGHLIGHTS
- Adobe Research and UC San Diego created EditAgent, an AI agent designed to replicate professional image-editing workflows in Photoshop.
- EditAgent combines a general-purpose editor with specialized tools and quantitative verification to reduce errors.
- In tests, Flux-2-Klein-9B failed 42.7% of image edits, but specialized tools resolved two-thirds of these failures.
- Professionals preferred EditAgent’s parallel-generated variants in 78.5% of comparisons.
- Qwen3.8-27B achieved 81.3% accuracy in selecting usable edits, surpassing the fine-tuned EditScore-72B model.
- EditAgent generates high-quality training data, improving open-source editors across three benchmarks.
How EditAgent works
EditAgent mimics professional Photoshop workflows by breaking down complex editing requests into steps. For each step, it selects the most suitable tool—either a general-purpose editor or a specialized one—and verifies the result using quantitative metrics like brightness, color changes, and altered regions.
Unlike traditional AI editors that retry failed steps with the same tool, EditAgent generates multiple editing candidates in parallel. An unmodified language model then evaluates these candidates using verification records, which include close-ups of edited areas, change maps, and statistical measurements.
This method reduces errors significantly. In tests, Flux-2-Klein-9B, a general-purpose editor, failed to produce usable results 42.7% of the time. Specialized tools resolved two-thirds of these failures, particularly for tasks like sharpening, rotation, and zooming.
Performance and professional preference
EditAgent’s outputs were preferred by professionals in 78.5% of comparisons where a clear preference was expressed. The system’s parallel variant generation and quantitative verification contributed to this preference, as it minimized errors like overexposure or unintended color shifts.
The system also excels at selecting usable edits. Qwen3.8-27B, an unmodified language model, achieved 81.3% accuracy in identifying usable edits when provided with verification data. This surpassed the performance of EditScore-72B, a fine-tuned model designed for this task.
Without verification records, none of the seven language models tested could match the accuracy of the best fine-tuned model. However, with the records, all models outperformed even EditScore-72B, highlighting the importance of quantitative verification in AI-driven editing.
Impact on open-source editors and training data
EditAgent improved the performance of open-source editors like Qwen-Image-Edit-2511 and Flux-2-Klein-9B across three benchmarks, outperforming seven other AI agents. For example, the Qwen variant edged past commercial models like Seedream 4.0 and GPT-Image-1.5 in the EdiVal benchmark.
The system also generates high-quality training data. In testing, EditAgent produced approximately 26,000 training examples. Editors fine-tuned on this data showed improved performance across all benchmarks, suggesting a scalable method for enhancing AI models without manual annotation.
Adobe’s broader AI integration
Adobe has already integrated AI agents into its flagship products, including Photoshop, Premiere, and Creative Cloud, as part of a public beta launched in June. This move reflects Adobe’s push to embed agent-based workflows into its software, potentially reshaping how professionals interact with these tools.
Black Forest Labs, a collaborator in the research, has since replaced the general-purpose model tested in the paper with Flux 3 Image Klein. This updated model, with 5.6 billion parameters, outperforms its predecessors, Flux 2 Klein and Flux 2 Pro, in benchmark tests.
What this means
Lazyfounder analysis — our interpretation, not reported fact.
EditAgent represents a step forward in bridging the gap between AI-driven automation and professional-grade image editing. By combining general-purpose models with specialized tools and quantitative verification, Adobe Research has demonstrated a method to reduce errors and improve output quality—key challenges for AI in creative workflows.
For founders and operators in the AI or design space, this research underscores the potential of agent-based systems to generate high-quality training data at scale. The ability to produce 26,000 usable training examples without manual annotation could lower costs and accelerate development for startups building AI-powered design tools.
Adobe’s integration of AI agents into its Creative Cloud suite also signals a broader industry shift toward embedding these systems into professional software. Competitors in the design and photography space, such as Canva or Affinity Designer, may need to explore similar approaches to stay competitive. However, the higher compute requirements of EditAgent—roughly 10x more than a single editor—could pose challenges for startups with limited resources.
Key takeaways
- EditAgent’s approach of combining general-purpose AI with specialized tools and quantitative verification could set a new standard for AI-driven image editing, particularly for startups building photography or design tools.
- The system’s ability to generate high-quality training data at scale may reduce reliance on manual annotation, a potential cost advantage for AI-driven design platforms.
- Adobe’s integration of AI agents into Photoshop and Creative Cloud signals a shift toward agent-based workflows, which could influence how competitors like Canva or Figma approach automation.
Related on Lazyfounder
Sources
- The Decoder · 2026-11-01
Adobe is testing how well AI agents can use editing tools like a Photoshop professional would
This story is an original summary drafted with AI by Lazyfounder from the reporting listed above and checked by automated validation. Facts are attributed to their original publishers; sections marked as analysis are Lazyfounder's. Where a source is in another language, facts were machine-translated and quotations are reported, not reproduced. Read the original coverage via the links, and see our AI policy and corrections policy.
About the author
Editor, Lazyfounder
Tarun Mottlia edits LazyFounders, covering Indian startups, funding rounds, AI and product launches. Every story on the site is AI-assisted and checked against its cited sources before publication.
More stories by Tarun MottliaGet the LazyFounder Brief
Startup, funding and AI news in a five-minute read. Join the early-access list.
