Streaming Video Editing with Easy Adaptation
SVEET edits video as a stream while training only on a pretrained bidirectional video diffusion model.
The paper says the method adapts video-to-video diffusion for autoregressive editing by separating backbone features and keeping conditional frames independent. Its auxiliary branch encodes source video frames with temporally independent self-attention, then injects those features into matching backbone blocks. The authors also use a decoupled training scheme to keep controllability and causality from interfering with each other. They report higher editing quality with real-time performance, reaching 15 FPS on a single H100 GPU without extra acceleration. HF Daily Papers' note
The paper says the method adapts video-to-video diffusion for autoregressive editing by separating backbone features and keeping conditional frames independent. Its auxiliary branch encodes source video frames with temporally independent self-attention, then injects those features into matching backbone blocks. The authors also use a decoupled training scheme to keep controllability and causality from interfering with each other. They report higher editing quality with real-time performance, reaching 15 FPS on a single H100 GPU without extra acceleration. HF Daily Papers' note
score 5