VOID

VOID is built for a harder version of video object removal. Instead of only hiding the removed object and patching the background, it also tries to correct the downstream interactions that object caused in the scene. That means the real question is not just "did the object disappear?" but "does the rest of the clip still make physical sense afterward?" The public demo and sample videos make that difference easy to see fast. If you want the full open workflow, the model card, GitHub repo, and notebook also give you a real local path for running the provided checkpoints.

Start with the Hugging Face Space and play through the built-in samples before you worry about your own footage. Look at cases where removing one object should change another object's motion, balance, or position, because that is where this workflow earns its slot. If the results look useful, move to the model repo and notebook for the local path. The fastest reproducible local test is to run one included sample first, then decide whether it is worth setting up custom quadmask generation for your own clips.

Treat the Space as the fast proof, not the whole setup. The practical local path is heavy, and generating your own quadmasks adds extra setup through SAM2 plus Gemini.

• Compare interaction-aware removal against plain video inpainting on short clips where simple background fill is not enough.
• Prototype cleanup workflows for ads, product demos, or creative edits where removing one object changes the rest of the scene.
• Pressure-test whether a public demo plus open weights is enough to justify a heavier local video editing workflow.
• Use the sample notebook and repo as a starting point for counterfactual video editing experiments on your own footage.

Try the demo

View model page