InfinityEdit: Adapter for Infinite Streaming Video Editing Unveiled

On August 21, 2026, a paper titled InfinityEdit: Infinite Video Editing with a Lightweight Edit-Ignition Adapter was published on arXiv. The authors abandon the usual assumption that the video being edited already exists as a finished clip, and instead consider an open stream: edits are applied to future frames as they appear. To do this, they assembled a separate dataset and proposed a lightweight adapter made of three attention modules — history cross-attention to rely on previous frames, temporal causal self-attention to pass signals only from early frames to later ones, and edit cross-attention to embed the editing request. The authors did not release any code.

What it means

Our reference on video generators does not yet include tools that work in infinite streaming editing mode. The closest in terms of task — LTX-Video, CogVideoX, and Pyramid Flow — generate video from a text or image prompt, but all of them assume a finite clip rather than a continuously arriving stream of frames. LTX-Video solves the image-to-video task, weighs 26.6 GB in its largest file, and, as described by the developer, requires only 1 GB of VRAM. CogVideoX is a text-to-video generator with a 9.2 GB model, oriented toward CUDA and NVIDIA H100. Pyramid Flow also works as a text-to-video generator, with the largest weights file taking up 7.8 GB, and less than 8 GB of VRAM is enough to run it. None of the fact sheets mention a streaming mode or a mechanism for accumulating edits on top of already generated frames. InfinityEdit describes exactly the niche that is not yet covered in the reference: editing a live stream, where requests arrive sequentially and must be applied to frames that have not yet been generated. For now, this is a research work without code, and its applicability to existing generators cannot be assessed — the authors have published neither weights nor a repository.