Netflix, INSAIT unveil void AI model

INSAIT at Sofia University St. Kliment Ohridski and Netflix developed VOID, a new artificial intelligence model that removes objects from video while realistically recreating changes in the scene, the Ministry of Education and Science of Bulgaria said on Wednesday.

Unlike standard tools that simply fill in deleted parts, the model understands how objects interact. For example, if a person holding an object is removed, the model simulates how the object falls or moves naturally, as if the scene had been filmed without that person.

This could give video content producers much greater freedom and save significant financial resources on reshooting scenes, CE Report quotes BTA.

The technology is based on CogVideoX and uses a special quadmask approach that makes it possible to distinguish between objects, interaction zones and the background. This helps the system preserve the logic and dynamics of the shot without visible artefacts. Because almost no real training data are available, the Netflix and INSAIT teams used simulated scenes generated with Blender. This allows the model to learn what happens in the real world when an object disappears.

Compared with existing solutions, VOID delivers better visual consistency and more realistic object behaviour. The model is open-source, allowing developers and researchers around the world to work on and build on the technology.

This is the latest project pointing to INSAIT's role in creating technologies of global importance.