Adobe and UC San Diego Develop AI Image Editing Agent

Adobe Research and UC San Diego have jointly developed 'EditAgent,' a system that enables AI to autonomously perform image editing tasks. The system is designed to allow AI agents to master specialized editing tools like Photoshop, enabling images to be completed in a workflow similar to how human professionals work.

Traditionally, AI-based image editing has relied on repeatedly calling a single general-purpose model. However, this approach requires using the same tool regardless of task type, limiting accuracy and efficiency. EditAgent transforms this approach by adopting a more human-like method: selecting and distinguishing the optimal specialized tool for each editing task based on its content.

The core mechanism involves a combination of tool selection and result evaluation. The agent selects appropriate specialized tools for each editing task and generates multiple processing results. Subsequently, a standard language model (LLM) evaluates and selects the most suitable result based on measured image data. In other words, the AI not only 'creates' but also takes on the function of 'selecting,' which is its distinguishing feature.

Meanwhile, challenges have emerged. Implementing this approach requires approximately 10 times more computational resources compared to conventional general-purpose methods. This creates a tradeoff where performance advantages come at the cost of significantly increased computational expenses, placing the technology at a stage where efficiency improvements are necessary for practical implementation.

The trend of AI agents combining multiple tools to autonomously manage complex tasks—rather than handling only single tasks—represents a broader direction in AI research overall. In the context of image editing, this represents an attempt for AI to genuinely enter creative domains such as photo retouching and asset processing. The fact that leading companies like Adobe in the image editing software space are developing such systems at the research stage can be viewed as an indication of industry direction.

The future focus will be on how much computational cost reduction can be achieved. A 10-fold increase in computational resources presents a significant barrier when considering cloud service integration and provision to general users. The next milestone to watch will be how EditAgent progresses from the research stage toward practical application, and how Adobe reflects this concept in its products.