Adobe Research and researchers at UC San Diego have developed EditAgent to improve how AI handles complex image editing tasks. The agent uses specialized tools alongside a general-purpose editor to process difficult modifications that standard models often struggle with. A standard language model compares different variants using measured image data and verification records during this process.
In a test of 600 edits, a general-purpose editor failed 42.7 percent of the time. Specialized tools solved roughly two-thirds of these failures, particularly for tasks involving sharpening and rotation. Professionals preferred parallel variants generated from the original in 78.5 percent of comparisons. With verification records, the Qwen3.8-27B model achieved an 81.3 percent usable pick rate.
Verification records containing change maps and statistics helped all seven tested models beat fine-tuned models. The approach requires about ten times the compute compared to standard editing methods, which is a significant trade-off for the improved accuracy. This development shows how specialized tooling can help AI systems handle complex edits that general purpose models cannot reliably perform.
