Netflix Open-Sources VOID AI Model for Object Removal in Videos

The Netflix AI research team has open-sourced VOID (Video Object and Interaction Deletion), an AI model capable of removing objects from videos while preserving realistic physical interactions, according to the model release on Hugging Face.

VOID is built on the CogVideoX 3D Transformer architecture and fine-tuned for video inpainting using an interaction-aware “quadmask” system. The quadmask encodes not only the primary object to remove but also regions affected by the object’s interactions, such as items that fall or move once the object is deleted. This structured approach allows the model to maintain physically plausible outcomes in edited videos.

The system operates in one or two inference passes. The first pass performs standard inpainting, while an optional second pass refines temporal consistency using optical flow-warped latent initialization to stabilize shapes across frames. This two-step process addresses common issues like object morphing in diffusion-based video models.

Training for VOID used synthetic paired video data generated through Blender-based physics simulations (HUMOTO) and Google’s Kubric framework. These datasets simulate real-world interactions by rendering scenes both with and without specific objects, ensuring physically accurate results. The model and code are available under an Apache 2.0 license on Hugging Face.