Overview
- Netflix, which released the research Monday, made VOID available on Hugging Face under an Apache 2.0 license.
- The system removes chosen objects and then regenerates each frame so the scene behaves plausibly, like turning a head-on crash into a clean drive or letting a guitar drop to the floor.
- In a paper-backed test with 25 participants, viewers preferred VOID’s results 64.8% of the time versus 18.4% for Runway, though the study size was small.
- Running the model demands roughly 40GB of GPU memory and a high-end card, and it builds on a 5‑billion‑parameter CogVideoX variant that uses a “quadmask” to focus edits on affected regions.
- Coverage flags open questions in crowded or complex footage and warns that convincing rewrites may weaken trust in video as evidence.