VOID removes objects, fixes scene interactions

This is a meaningful step beyond standard video inpainting: the hard problem is causality, and VOID is explicitly modeling that instead of hoping a fill-in model will guess right.

• –Counterfactual paired data from Kubric and HUMOTO gives the model supervision for downstream effects, which is the right signal for collisions, falls, and chain reactions.
• –VLM-guided masks are a practical way to expand the edit beyond the clicked object to everything the object influences.
• –The second-pass flow-warped refinement reads like a targeted fix for the most obvious failure mode in video diffusion: shape drift and morphing over time.
• –A 64.8% human preference win over Runway, ProPainter, and Gen-Omnimatte suggests this is not just a clever paper idea; it is already competitive in messy real-world edits.
• –The strongest near-term use cases are likely robotics simulation, ADAS scenario editing, and dataset generation where preserving physics matters more than perfect texture fill.