Netflix open-sources VOID to fix physics when erasing video objects

VOID is an open-source video editing model from Netflix that removes objects from footage alongside all the physical interactions they cause. Instead of acting like a crude digital eraser, VOID recalculates the scene so the laws of physics do not break when an element vanishes.

Most traditional inpainting tools fail as soon as an object is doing work in a scene. If you erase someone carrying a tower of boxes, the boxes hang in mid-air. Erase a chair, and a person hovers. VOID approaches this by reasoning about causality before generating pixels. Remove the person carrying a guitar, and the guitar drops naturally to the floor. Erase a bowler and the bowling ball, and the bowling pins remain standing undisturbed at the end of the lane.

How it reasons through causality

Built on top of the 5-billion-parameter CogVideoX-Fun-V1.5-5b-InP architecture, VOID splits video removal into causal reasoning and video diffusion.

Before any inpainting happens, a vision-language pipeline using SAM 2 and Gemini analyzes the clip to determine what the target object is physically affecting. It encodes the scene into a quadmask with four distinct pixel values: 0 for the object being removed, 63 for overlaps, 127 for causally affected regions like falling props, and 255 for untouched background.

Guided by this map and a text prompt describing the clean background, the diffusion model fills in the video. The inference runs in two passes. Pass 1 handles the primary inpainting and is usually enough for standard clips. Pass 2 is an optional refinement step that uses optical flow to warp noise from the first pass, preventing objects from morphing over longer clips.

Specs and hardware requirements

Netflix released VOID under the Apache 2.0 license, publishing both the code on GitHub and model weights on Hugging Face. The release supports Hugging Face Diffusers and includes an all-in-one Jupyter notebook for quick testing.

The main obstacle is hardware. Between the roughly 10GB CogVideoX base model and VOID's 22GB checkpoints, running the model locally demands serious VRAM. The default pipeline targets 384x672 resolution for up to 197 frames and strictly requires an enterprise GPU with at least 40GB of VRAM, such as an Nvidia A100. Consumer cards like the RTX 3090 or 4090 will not cut it locally.

If you have access to cloud compute like RunPod or Lambda Labs, you can clone the Netflix/void-model repository and run the notebook directly. Everyone else can check out the interactive comparisons on the project page.

Why it matters

For video editors and visual effects teams, removing unwanted objects or continuity errors usually means tedious frame-by-frame rotoscoping and manual physics patching. VOID is the first tool to automate the physics side of that workflow. It also holds surprising utility for robotics and autonomous driving researchers who need counterfactual video datasets to simulate world physics.

We think releasing this under an Apache 2.0 license is a refreshing move from a streaming giant with enormous VFX resources. The obvious limitation right now is compute: a 40GB VRAM barrier keeps it out of reach for everyday creators without cloud rentals. If open-source contributors can optimize or quantize the weights down to consumer GPUs, VOID could reshape everyday video post-production.