Netflix has open-sourced VOID, a new artificial intelligence framework capable of removing objects from videos while realistically reconstructing the physical interactions left behind. Released under the Apache 2.0 license, the tool allows developers and researchers to use it freely for commercial and creative purposes.
Unlike traditional video editing or inpainting applications that only fill missing areas with background pixels, VOID can simulate what the scene’s physics would look like if the removed object had never existed. For instance, deleting one car from a crash sequence produces footage showing the other vehicle continuing along its path naturally, without damaged debris or distortion.
Technically, VOID works through a multi-model pipeline that integrates several state-of-the-art AI components. It builds on Alibaba’s CogVideoX video diffusion model, customized with synthetic data from Google’s Kubric and Adobe’s HUMOTO datasets to enhance motion understanding and object interaction detection. Google’s Gemini 3 Pro identifies the affected scene areas, while Meta’s SAM2 provides precise object segmentation. The system then encodes these regions into a specialized “quadmask” to guide the diffusion model, with an optional refinement step using optical flow correction for realistic scene continuity.
The project was developed by a Netflix research team Saman Motamed, William Harvey, Benjamin Klein, Zhuoning Yuan, and Ta-Ying Cheng in collaboration with Luc Van Gool from INSAIT at Sofia University. In testing across multiple scenarios, VOID was preferred 64.8 percent of the time by participants compared with alternative approaches, including Runway.
Netflix has made the code, research paper, and demo available on GitHub, arXiv, and Hugging Face. Running VOID, however, requires a high-performance GPU with at least 40GB of VRAM. The company has not indicated whether it plans to integrate the technology into its production workflows.
Read Article: SEBI Set to Launch SETU Platform for Investment Advisors
