Netflix unveils VOID: the AI that erases and rewrites video physics

Netflix has recently made available to the public an artificial intelligence framework named VOID (Video Object and Interaction Deletion). This system stands out for its ability to remove objects from videos while automatically adjusting the physical effects those objects had on the surrounding scene. In other words, VOID does not simply erase an object; it also modifies the physical interactions, such as collisions, that the object initially caused.

The development of VOID is based on Alibaba's video diffusion model CogVideoX. This model has been refined using synthetic data from Google's Kubric and Adobe's HUMOTO, which are used for interaction detection. Gemini 3 Pro from Google plays a crucial role in analyzing the scene to identify the affected areas, while SAM2 from Meta is responsible for segmenting the objects to be removed. To perfect the result, an optional second pass uses optical flow to correct any potential shape distortions.

This project was conducted by researchers at Netflix in collaboration with INSAIT Sofia University. Resources related to VOID, including the source code, a detailed paper, and a demonstration, are available on platforms such as GitHub, arXiv, and Hugging Face. The system is distributed under the Apache 2.0 license, allowing for commercial use.