中文智能插入-V：基于闭环反馈双流框架的照片级真实感视频插入

arXiv:2605.23891v1 提出 Smart-Insertion-V，一种端到端双流框架，同时进行视频插入与图像风格迁移。通过图像流同步引导视频生成，并引入闭环机制，有效克服参考对象与源场景间的严重风格差异，实现和谐的无遮罩视频对象插入。

arXiv:2605.23891v1 Announce Type: new Abstract: Mask-free video object insertion has emerged as a challenging task, requiring harmonious integration of reference objects into source videos. However, existing methods struggle when references exhibit severe stylistic domain gaps with the source scene. To overcome this, we propose \textit{\textbf{Smart-Insertion-V}}, an end-to-end \textbf{Dual-Stream} framework that concurrently conducts video insertion and image style transfer. Within this framework, the image stream synchronously guides the video generation process, while a \textbf{Closed-loop