Hugging Face researchers have unveiled EditaLive, a novel framework designed for real-time character video editing in live streaming. Unlike conventional video editing that focuses on scene-level content, EditaLive prioritizes the human subject, addressing the unique challenges of live-streaming environments.
The framework builds on a pretrained image animation model, Wan-Animate, which naturally separates appearance from motion. By repurposing it as the base for instruction-based editing and training on the CharEdit-50K dataset, EditaLive enables reference-frame editing and video reconstruction.
To achieve real-time performance, the researchers adapted the model from offline bidirectional processing to causal streaming generation. They employed a distilled two-step sampling strategy with aligned self-rollout, reducing training-inference discrepancies. Fixed RoPE and align forcing, along with first-frame preserved sparse attention, help mitigate appearance drift and filter redundant historical information.
Extensive experiments demonstrate that EditaLive delivers state-of-the-art editing performance while faithfully preserving facial expressions and maintaining low-latency streaming inference, making it suitable for interactive live-stream applications.