Tsinghua & HKUST Introduce LiveEdit: Real-Time Streaming Video Editing at 12.66 FPS

机器之心 · wechat · 2026-08-01

Tsinghua University and HKUST jointly proposed LiveEdit, a real-time streaming video editing framework based on diffusion models, accepted by ECCV 2026.

The framework compresses inference steps to 4 via a three-stage distillation process (foundation tuning, teacher forcing, DMD) and introduces an autoregressive-oriented mask cache to skip redundant computations in unedited regions. Ultimately, it achieves a streaming editing speed of 12.66 FPS while maintaining visual consistency, providing a technical foundation for low-latency scenarios like live streaming effects and video conferencing.

Original post →

More from Multimodal

Multimodal channel →