InfinityEdit formalizes infinite video editing, in which a model receives a preceding video segment and an edit request and must produce the next segment that continues the stream under the edit; the process repeats as a stream of edit instructions arrives. The paper argues that prior instruction-based video editing relies on an in-place editing assumption: output frames are aligned to source frames over a fixed time span. That assumption breaks for open-ended streams such as live-game restyling or camera motion applied to an ongoing shot.1
Context
To address the task, the authors design a data-collection pipeline and propose InfinityEdit, a lightweight edit adapter added to a streaming video generator. The adapter contains three attention modules: history cross-attention for conditioning on input frames, temporal causal self-attention so information flows only from earlier to later frames, and edit cross-attention to inject the edit request. At inference, the adapter is activated only in the chunk where an edit arrives; subsequent chunks are generated by the original model with a reset anchor frame, preserving the base model's infinite generation ability.1
Assessment
The technical novelty lies in decoupling edit application from video continuation: the edit is applied locally in the arrival chunk, while continuity is delegated to the original streaming generator. This avoids accumulating adapter activations and keeps the adapter lightweight. The reset anchor frame is the key inference-side mechanism, but the abstract only names it and does not specify how the anchor is chosen or reset, so the exact continuity guarantee remains underspecified.1
Caveats
Because the available source is only the abstract, no quantitative evidence is accessible: there are no reported metrics, baselines, ablations, or dataset statistics. The claim that generation stays stable over unbounded edit sequences rests on unspecified extensive experiments, so an assessment of competitive performance or failure modes cannot be made. What would change the evaluation is evidence on long-horizon edit accumulation, comparison against frame-wise baselines under the same streaming protocol, and details of the data-collection pipeline.1
