From Short Clips to Real-Time Streams: Identity-Aware Supervision and Context-Efficient Generation for Video Face Swapping

One-line takeaway

For real-time video face swapping over long streams, the paper proposes a Joint Rec-ID Loss that balances identity supervision and reconstruction by noise level, and Face-RoI generation that uses fewer spatial tokens and shorter temporal context.

论文面向长视频实时换脸：提出按噪声水平平衡身份监督与重建的 Joint Rec-ID Loss，并利用目标视频已有的内容，只用更少空间 token 和更短时间上下文生成人脸区域。