The article introduces CFM for CNFs, addressing training challenges and improving results in various generative tasks. It discusses OT-CFM for dynamic OT approximation and SB-CFM for Schrödinger bridge inference. Experiments show improved training efficiency and performance in single-cell dynamics, image generation, and unsupervised translation.
다른 언어로
소스 콘텐츠 기반
arxiv.org
더 깊은 질문