Researchers from Tsinghua University and the Hong Kong University of Science and Technology have developed LiveEdit, a real-time streaming video editing framework. This system processes continuous video input in chunks, achieving 12.66 FPS with a 4-step inference process per chunk. LiveEdit maintains accuracy in edited regions and consistency in unchanged areas by employing a three-stage distillation process and an auto-regressive mask cache to avoid redundant computations on static parts of the video. AI
IMPACT Enables real-time applications like live effects and video conferencing by significantly speeding up diffusion-based video editing.
RANK_REASON The cluster describes a new research framework and its technical details, accepted to a computer vision conference. [lever_c_demoted from research: ic=1 ai=1.0]
- AR-oriented mask cache
- Diffusion Transformer
- Ditto-1M
- ECCV 2026
- Hong Kong University of Science and Technology
- LiveEdit
- Tsinghua University
- Wan2.1-T2V-1.3B
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →