InfinityEdit adapter targets edits on open-ended video streams

Video editors today take a fixed clip in and hand a fixed clip back. That assumption breaks the moment the source is an open-ended stream, like a live broadcast or an ongoing shot. A paper posted to arXiv on 21 August frames the alternative as its own task, which the authors call 'infinite video editing': given a preceding segment and an edit request, a model must generate the next segment while applying the edit, and keep doing that as more edits arrive.

The authors put the failure mode concretely: 'This pattern fails for open-ended streams, e.g., restyling a live game or applying a camera move to an ongoing shot.'

Their proposed fix is InfinityEdit, described as 'a lightweight edit adapter that equips a streaming video generator with unbounded editing ability.' Three attention modules do the work. History cross-attention conditions denoising on the input frames. Temporal causal self-attention keeps information flowing only from earlier frames to later ones. Edit cross-attention injects the instruction. The adapter is active only inside the chunk carrying a fresh edit request; subsequent chunks are handed back to the original model with a reset anchor frame, which is how the paper preserves 'the original model's infinite generation ability.'

The abstract reports no benchmark numbers, no named baselines, and does not identify the streaming generator the adapter attaches to. Its result claim is qualitative: InfinityEdit 'faithfully continues the stream under each edit, and stays stable over unbounded edit sequences.'