Found first: a primary source the press has not covered yet.
Researchers from Zhejiang University and Alibaba Group introduce InfinityEdit, a method that attaches a lightweight adapter to a frozen streaming video generator so that open-ended editing works on live or continuously growing video streams. The adapter requires no retraining of the base model, and the paper reports stable edits sustained over more than 1,000 frames.
What the source says
InfinityEdit plugs three attention layers into a frozen 14B Helios-Distilled backbone: history cross-attention conditions denoising on prior frames, temporal causal self-attention enforces frame order, and edit cross-attention injects the edit instruction. The adapter fires only when an edit request arrives; subsequent segments resume from a reset anchor frame using the original weights. The authors evaluate on a custom benchmark of 200 source videos covering 15 edit types across four categories, run on 32 NVIDIA H20 GPUs. On VBench, InfinityEdit ranks first on Camera Motion (0.7654) and Temporal Flickering (0.9660) among six compared methods, and a VLM-as-judge evaluation on a 5-point scale puts it first on Edit Faithfulness (3.828), Visual Quality (3.765), Scene Identity Preservation (3.815), and Cross-Edit Coherence (3.840). Edit Faithfulness varies by a standard deviation of 0.023 across three sequential editing rounds.
Why it matters
Existing video editing methods assume a clip with fixed length. Live broadcasts, sports feeds, and the output of streaming generators have no editing layer that can carry an instruction forward as new frames arrive. InfinityEdit addresses this with a frozen-base-model adapter, meaning any compatible streaming generator can gain editing capability without retraining its weights. The low per-round variance in Edit Faithfulness is relevant here: the practical concern with stacking edits over a long sequence is whether errors accumulate, and the paper's stability results suggest they do not.