VOID (void-model.github.io) is an open-source generative video editing framework developed by Netflix Research that performs physically plausible video object removal. Unlike conventional video inpainting models that only reconstruct background pixels behind a target, VOID models counterfactual scene evolution—erasing an object while actively rewriting its downstream physical consequences (such as halting a falling domino chain, preventing collisions, or stopping moving liquids).

Standard video inpainting and object removal tools (such as ProPainter or Generative Omnimatte) treat removal purely as a spatial hole-filling task. When applied to dynamic scenes where the target object actively pushes, carries, or interacts with other elements, traditional models produce severe physical anomalies—leaving phantom forces, unexplained momentum, or floating shadows. VOID resolves this spatiotemporal entanglement by asking: “What would the video look like if the object had never been there?”

Technical Architecture & Capability Matrix

Dimension / Feature	VOID Technical Specification	Core Capability Advantage
Primary Task	Physical Interaction-Aware Object Deletion	Deletes target objects and re-simulates affected scene dynamics.
Causal Reasoning	VLM-Guided Spatial Masking (“Quadmasks”)	Vision-language models identify downstream dynamic impact zones.
Generation Pipeline	Two-Pass Video Diffusion Refinement	Pass 1 synthesizes new motion; Pass 2 refines appearance and temporal flow.
Training Datasets	Kubric Simulation & HUMOTO Mocap Pairs	Synthetic and real counterfactual video pairs with/without objects.
Generalization	Zero-Shot Causal Physics Extrapolation	Halts domino chains, releases held balloons, and stops machinery.
Open Source	Code, Paper & Hugging Face Interactive Spaces	Open research artifacts for VFX, robotics world models, and AI video.

Core Breakthroughs & Capabilities

• Counterfactual Physical Reasoning: When an object in motion is removed (e.g., intermediate dominoes or a car causing a collision), VOID does not leave subsequent objects reacting to an invisible force. It recalculates the causal trajectory so that untouched objects remain stationary or proceed along uninterrupted paths.

• VLM-Driven Impact Identification: Users simply click the primary object to delete. A vision-language model evaluates world knowledge to infer which surrounding regions (shadows, reflections, connected tools, affected secondary objects) will change, automatically generating expanded guidance masks.

• Two-Pass Motion & Appearance Synthesis:

  • Pass 1 (Trajectory Prediction): The base diffusion model synthesizes the new, physically consistent motion path.

  • Pass 2 (Appearance Refinement): A secondary refiner conditions on flow-warped noise to eliminate pixel morphing, flickering, and temporal blur.

• Emergent Generalization to Unseen Scenarios: Despite having no floating objects or specific appliances in its synthetic training sets, VOID demonstrates high-level common-sense reasoning—such as letting a balloon float upward when its holder is removed or preventing food in a blender from spinning when the operator is deleted.