EditaLive! Unified Character Video Editing for Live Streaming
EditaLive turns human-focused video edits into a low-latency streaming process while preserving facial expressions.
The paper says existing video-editing methods struggle in live-streaming settings because they can disrupt expressions and rely on offline inference steps. EditaLive starts from Wan-Animate, uses reference-frame editing and reconstruction on the CharEdit-50K dataset, then adapts the model for causal streaming generation. Its distillation strategy compresses generation into a two-step sampler and adds attention controls meant to reduce appearance drift. The authors report state-of-the-art editing performance with real-time streaming inference. HF Daily Papers' note
The paper says existing video-editing methods struggle in live-streaming settings because they can disrupt expressions and rely on offline inference steps. EditaLive starts from Wan-Animate, uses reference-frame editing and reconstruction on the CharEdit-50K dataset, then adapts the model for causal streaming generation. Its distillation strategy compresses generation into a two-step sampler and adds attention controls meant to reduce appearance drift. The authors report state-of-the-art editing performance with real-time streaming inference. HF Daily Papers' note
score 4