StreamingSwap brings real-time face swapping to 10-minute video streams
A paper submitted to arXiv on 19 October 2026 presents StreamingSwap, a reference-guided flow-matching framework for real-time streaming video face swapping. The authors introduce a Joint Rec-ID Loss that balances explicit identity supervision and latent reconstruction according to noise level, plus Face-RoI Generation and Short-Context Rollout Training that cut the spatial tokens and temporal context needed per frame. They also release HumanFaceVid, described as 1,000 hours of cinematic training data with evaluation sets of 200 short and 200 long videos, and report competitive identity transfer and visual quality at 20 FPS on a single NVIDIA RTX 4090 with stable generation over videos up to 10 minutes long; those results are self-reported, with no independent measurement supplied.
The authors present long-form, real-time identity swapping as the open problem their design targets, and report 20 FPS on a single consumer-class RTX 4090 GPU with stability over 10-minute videos, claims that have not been independently verified.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2610.26681 (cs)
[Submitted on 19 Oct 2026]
Title:From Short Clips to Real-Time Streams: Identity-Aware Supervision and Context-Efficient Generation for Video Face Swapping
Authors:Zekai Luo, Yuwei Liang, Ruiqi Ma, Haoyu Huang, Xuehao Wang, Huanyi Zheng, Mingyu Liu, Linhao Zhong, Chenchen Jing, Yuling Xi, Hao Chen
View a PDF of the paper titled From Short Clips to Real-Time Streams: Identity-Aware Supervision and Context-Efficient Generation for Video Face Swapping, by Zekai Luo and 10 other authors