Netflix researchers released VOID (Video Object and Interaction Deletion) on April 3, 2026, marking the company's first public AI model release. The open-source video editing model removes objects from videos while predicting and eliminating all physical interactions they induce, including shadows, reflections, and dynamic physics effects.
Physics-Aware Object Removal Sets VOID Apart
VOID goes beyond standard video inpainting by understanding and removing physical interactions. When a person holding a guitar is removed, the model predicts the guitar falling naturally. Removing someone carrying a mug causes objects to react appropriately, and deleting a person jumping into a pool eliminates splash effects on both the pool and ground.
The model is built on the CogVideoX architecture, fine-tuned for video inpainting with interaction-aware mask conditioning. It uses two transformer checkpoints trained sequentially and can run with Pass 1 alone or chain both passes for higher temporal consistency.
VOID Outperforms Competitors in User Preference Testing
In a survey of 25 people across multiple scenarios, VOID achieved a 64.8% preference rate compared to competing solutions. Runway, the second-place competitor, received just 18.4% preference, with other alternatives scoring significantly lower.
The model is available on Hugging Face under netflix/void-model with accompanying GitHub repository at Netflix/void-model. Netflix released both the preprint paper and model weights as open-weights, enabling researchers and developers to use and build upon the technology.
Applications Span Film Production to Automated Workflows
VOID targets several production use cases:
- Film editing and post-production without manual frame-by-frame work
- Automated video manipulation workflows
- VFX work requiring physics-accurate object removal
- Creative production with physics simulation capabilities without render farms
The Register noted that while some results may appear "janky" in certain scenarios, VOID represents a groundbreaking advance in understanding video physics. Community reaction on X has been strong, with technical discussions highlighting the physics-aware inpainting as a significant breakthrough versus standard erasure tools.
Key Takeaways
- Netflix released VOID on April 3, 2026 as its first public AI model, available as open-weights on Hugging Face
- The model removes objects from videos along with all physics interactions including shadows, reflections, and dynamic effects
- VOID achieved 64.8% user preference compared to 18.4% for second-place competitor Runway in testing
- Built on CogVideoX architecture with interaction-aware mask conditioning and two-pass processing
- Applications include film production, VFX work, and automated video editing without manual frame-by-frame intervention