A new AI framework edits video the same way it edits a single photo, then spreads that change across every frame.
VINCIE-NExT, described in a new research paper, tackles a basic data problem: training a model to edit video requires huge numbers of before-instruction-after video triplets that are expensive to collect. Image editing has no such shortage, with millions of paired examples already available. The researchers route video edits through the image domain instead, breaking the task into a chain: video to image, image to edited image, then back to video. A single edited image pair, either generated by the model or supplied by a user, acts as a visual template that a new position-encoding scheme propagates pixel-by-pixel to every frame. The system also supports "chain-of-editing," letting users trade more compute at inference time for sharper results without retraining. On the OpenVE-Bench benchmark, the approach reportedly beats prior methods across multiple editing categories.
The clever part is the workaround, not the editing itself. Video generation has gotten good fast, but video editing has lagged because nobody wants to pay for video-triplet labeling at scale. By routing edits through the image pipeline, where abundant data already exists, VINCIE-NExT sidesteps that bottleneck rather than solving it directly.
Worth remembering this is a benchmark result from a single paper, not a shipped tool, and the extra compute needed for chain-of-editing's quality boost isn't free. Still, routing problems through a better-resourced adjacent domain is a trick other video AI teams will likely borrow.