AI has learnt how to remove objects from videos without violating physics

It is about modelling cause and effect relationships.

Details

The development was presented by researchers from Netflix and their colleagues.

The system is called VOID (Video Object and Interaction Deletion).

Unlike conventional tools, it:

• doesn't just "delete" the object
• takes into account how it affected the surrounding elements
• rearranges the behaviour of the entire scene

For example:

• if you remove a support, the object should fall
• if you remove an element in a chain of dominoes, the fall will stop

The analysis showed:

• The AI first identifies the areas that will be affected by the change
• then creates a map of the change (including shadows and movements)
• then generates a new version of the scene
• at the final stage, it refines the movement and shape of objects

The system was trained on thousands of simulated scenes to understand cause-effect relationships.

Why it matters

The technology solves one of the key problems in video editing.

Researchers note:

• conventional tools often disrupt the physics of a scene
• which makes video look unnatural
• the new model can improve the quality of film and visual effects

It also opens up possibilities for:

• automatic video editing
• create realistic scenes without reshoots

Background

Previously, video editing systems worked mainly with pixels and not with the logic of what was happening.

This led to situations where the scene looked "broken" after an object was removed.

Source

The research is based on the development of a VOID model for object removal taking into account the physics of interactions. The work has been published on the arXiv preprint server (2026).