Most instruction-based video editing methods share a hidden assumption: you hand the model a fixed-length clip, and it performs an in-place rewrite — edited frames are aligned frame by frame with the source over a fixed time span. That works for pre-recorded clips, but it breaks for open-ended streams such as restyling a live game or applying a camera move to an ongoing shot: frames keep arriving, and the edit must extend into future frames rather than act on a static input. A paper submitted to arXiv on August 21, InfinityEdit (arXiv:2608.20910), tackles this setting head-on and names it infinite video editing (https://arxiv.org/abs/2608.20910).

The Task: From Editing a Clip to Editing a Stream

The paper defines the new task as follows: given a preceding segment and an edit request, the model must generate the next segment — one that continues the stream while applying the requested edit. As instructions arrive as an unbounded sequence, the process repeats indefinitely. Two core challenges follow. First, the edit must be a faithful continuation rather than a frame-wise rewrite. Second, generation quality must remain stable as edits accumulate — the output cannot degrade with every new instruction.

The Method: A Lightweight Adapter with Three Attention Modules

Rather than training yet another heavyweight editing model, InfinityEdit equips a streaming video generator with a lightweight edit adapter. The adapter contains three attention modules with clear division of labor:

  • History cross-attention guides the denoising frames using the input frames, anchoring new content to what has already appeared.
  • Temporal causal self-attention keeps temporal cues flowing only from earlier frames to later ones, preserving causality.
  • Edit cross-attention injects the edit request into generation.

The inference-time scheme is the clever part: 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. The edit gets applied while the original model's infinite generation ability is preserved — an edit-ignition design that fires once and then lets the stream run.

Data and Validation

Because this is a new task, the authors also designed a data-collection pipeline to construct training data for infinite video editing. In extensive experiments, the paper reports that InfinityEdit faithfully continues the stream under each edit and stays stable over unbounded edit sequences. The work was submitted on August 21 with 12 authors and an 18-page paper; it landed on Hugging Face Daily Papers the same week (30 upvotes in the August 24 snapshot, second-highest of the day).

So What

For streaming world models and interactive video, editability shifts from being a property of static assets to a runtime capability. When edits can fire at any moment and generation never stops, effects like restyling a live broadcast or re-composing a game view in real time — things that today require hardcoded rendering pipelines — become a matter of saying one sentence to the generative model. The lightweight-adapter route also hints at a broader lesson: layering capabilities on top of streaming generation infrastructure may be closer to the right answer for this class of problems than training a dedicated monolith every time.