Hugging Face researchers have proposed InfinityEdit, a new approach to video editing that overcomes the limitations of existing methods, which typically assume a fixed-length input clip. The new method is designed for open-ended streams, such as live game restyling or applying camera moves to ongoing shots, where edits must extend to future frames as they arrive.
The core of InfinityEdit is a lightweight edit adapter that equips a streaming video generator with unbounded editing ability. The adapter comprises three attention modules: history cross-attention guides denoising frames using input frames, temporal causal self-attention ensures temporal cues flow only from earlier to later frames, and edit cross-attention injects the edit request into generation.
During inference, the adapter is activated only in the chunk where an edit request arrives. Subsequent chunks are generated by the original model with a reset anchor frame, preserving the model's infinite generation ability while applying the edit. The researchers also designed a data-collection pipeline specifically for infinite video editing.
Extensive experiments demonstrate that InfinityEdit faithfully continues the stream under each edit and remains stable over unbounded edit sequences, addressing the two main challenges: faithful continuation rather than frame-wise rewrite, and stable generation quality as edits accumulate.