Netflix Research publishes VOID, a video object removal model accounting for physical interactions.

Netflix and Sofia University launch VOID, a video object removal model based on CogVideoX that simulates physical consequences, outperforming Runway.

Researchers from Netflix and Sofia University are publishing VOID, a video object removal framework capable of reconstructing the physical consequences induced by this removal. Where existing tools are limited to erasing an object's appearance and correcting its immediate visual effects (shadows, reflections), VOID models what the scene would have produced if the object had never been present: an object held in the air falls, a collision does not occur, a domino effect is interrupted. The model relies on CogVideoX, a 5-billion-parameter video transformer, fine-tuned on pairs of counterfactual videos generated via the Kubric and HUMOTO simulation engines. Training utilizes a four-value mask system (quadmask) which distinguishes the object to be removed, the overlap zone, the areas affected by the removal, and the background to be preserved. During inference, a visual language model automatically generates these masks from a simple user click. A first pass produces the counterfactual video. An optional second pass, based on optical flow-aligned noise, corrects structural deformations that the model may introduce during the synthesis of new movements. In a human preference evaluation on 75 real-world videos, VOID was selected in 64.8% of cases, ahead of Runway (18.4%) and Generative Omnimatte (11.2%). The model and checkpoints are available on Hugging Face under an Apache 2.0 license. The full documentation can be found here