Adobe Research and UCSD Introduce EditAgent: Statistical Verification Data Enables General LLMs to Outperform Fine-Tuned Judges in Image Editing
Adobe Research and UC San Diego introduced EditAgent, an AI system that uses parallel candidate generation and hard statistical verification data to allow unmodified language models to outperform specialized fine-tuned evaluation models in image editing tasks, albeit at ten times the compute cost.
Researchers at Adobe Research (US) and UC San Diego (US) have introduced EditAgent, an AI image editing system designed to address the high failure rates of general-purpose editors when handling complex instructions. The core mechanism involves breaking editing tasks into sub-steps and generating multiple candidates in parallel by pairing a general-purpose editor with specialized tools for operations like sharpening, rotation, and zooming.
Unlike traditional methods that rely on retrying the same editor, EditAgent employs an unmodified standard language model as a judge. This model receives verification records containing close-ups of edited areas, change maps, and measured statistics on brightness and color. The study indicates that these hard numbers help the standard model identify errors such as overexposure or color temperature shifts caused by the general editor, allowing it to select superior results.
In a benchmark test of 600 image edits, blind-rated by five professionals, the results showed that when provided with verification data, all seven tested standard language models, including Qwen3.8-27B, outperformed the best fine-tuned evaluation model, EditScore-72B. Specifically, Qwen3.8-27B achieved an 81.3 percent accuracy rate in selecting usable results, compared to 74.8 percent for EditScore-72B. This suggests that the structured input data may matter more than the training method for evaluation models.