what is void-model fr?

Netflix's VOID project removes objects from videos and handles the physical consequences of their disappearance. If you erase a person holding a guitar, the tool doesn't just leave a floating guitar, it makes the guitar fall naturally to the ground. How it works: the pipeline first identifies what to remove and what parts of the scene were physically affected by that object. It uses a combination of segmentation models and a vision-language model to reason about interactions (like "if the person disappears, the cup they were holding will drop"). Then it generates the final video using a two-pass system: the first pass fills in the background where the object used to be, and an optional second pass smooths out flickering across frames for better temporal consistency. Who would use this: video editors, VFX artists, or content creators who need to remove unwanted elements from footage. A filmmaker might use it to erase a boom mic that drifted into frame, or a product demo team could remove a person holding an item so the item appears to sit on a surface on its own. The tool is especially suited for cases where removing an object should trigger realistic physics, things tipping over, falling, or settling into place. Notable choices: the project builds on an existing video-generation model (CogVideoX) and fine-tunes it specifically for this inpainting task. The training data is generated synthetically using Blender physics simulations and the Kubric framework, creating paired videos (with and without the object) plus masks. Rather than releasing the training datasets, the authors release the code to generate them, citing licensing constraints. Running inference requires a high-end GPU (40GB+ VRAM), and the easiest way to try it is through the provided Google Colab notebook.