Unlocking the Potential of Image Editing via Concept Scaling and Dense Supervision
The paper’s bet is that image editors improve when they learn many fine-grained edit concepts from denser training examples.
The authors introduce a taxonomy with more than 1,000 edit concepts and a 12 million-pair dataset called ConceptEdit-12M. They say their synthesis approach avoids distribution collapse while keeping data quality high. A dense supervision strategy combines multiple non-interfering concepts in single image pairs, giving the model more learning signal per example. They also present ConceptEdit-Bench to test editing ability across granular real-world scenarios. HF Daily Papers' note
The authors introduce a taxonomy with more than 1,000 edit concepts and a 12 million-pair dataset called ConceptEdit-12M. They say their synthesis approach avoids distribution collapse while keeping data quality high. A dense supervision strategy combines multiple non-interfering concepts in single image pairs, giving the model more learning signal per example. They also present ConceptEdit-Bench to test editing ability across granular real-world scenarios. HF Daily Papers' note
score 5