Researchers introduce ConceptEdit-12M, a 12-million-pair dataset covering more than 1,000 fine-grained edit concepts to improve image editing models, alongside a dense supervision training strategy that combines multiple non-interfering concepts in single image pairs to improve training efficiency. The team also built ConceptEdit-Bench, a new evaluation suite assessing model performance across diverse real-world editing scenarios, and reports strong results from models trained with the dense-supervision strategy.
