Netflix Void Model: Video Object and Interaction Deletion (github.com) AI

Netflix has released VOID (Video Object and Interaction Deletion) on GitHub, an open-source pipeline built on CogVideoX that removes a target object from a video while also deleting the physical interactions the object causes (e.g., preventing a guitar from falling when the person is removed). The project includes a two-pass inpainting approach for temporal consistency, plus a mask-generation stage that uses SAM2 segmentation and a VLM (via Gemini) to produce “quadmasks” capturing both the object and interaction-affected regions. Instructions and sample data are provided, along with optional tooling to manually refine masks before running inference.