New dataset and loss function improve object consistency in image editing models

Researchers have introduced ABO-Edit, a new dataset featuring over 12,000 image triplets designed to evaluate object consistency in text-guided image editing. They also proposed FlowMirror, a parameter-free auxiliary loss that leverages an overlooked property of image-editing rectified flow models. This property involves the conditioning embedding space, which predicts the final image even at high noise levels. By supervising this embedding space, FlowMirror enhances generation quality without requiring architectural changes. AI

IMPACT Enhances object consistency in image editing, potentially improving user control and realism in generative AI applications.

RANK_REASON The cluster describes a new academic paper detailing a novel dataset and methodology for improving image editing models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• 3D assets
• ABO-Edit
• alphaXiv
• arXiv
• CatalyzeX
• CORE Recommender
• DagsHub
• DiTailed
• FlowMirror
• Francesco Taioli
• Gotit.pub
• Hugging Face
• Influence Flower
• ScienceCast

• paper
• product

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →