Netflix Open-Sources VOID: The AI That Erases Objects From Video and Rewrites the Physics They Left Behind

Netflix has open-sourced a video AI framework called VOID (Video Object and Interaction Decomposition) that solves one of the most technically demanding problems in visual effects: not just removing an object from a video, but seamlessly correcting for all the physical interactions that object was having with its environment.

What VOID Does

Standard video inpainting tools can fill in the pixel region where an object used to be. What they cannot do is account for the downstream physics — the shadow the object was casting, the light it was reflecting onto nearby surfaces, the motion blur it was creating in adjacent frames, or the depth-of-field effects it was contributing to. VOID addresses all of these simultaneously.

The system decomposes video into distinct physical interaction layers — direct illumination, shadow fields, specular reflections, and motion contributions — and uses a diffusion-based generative model to reconstruct each layer coherently after removal. The result is a scene that appears as though the object was never there, rather than a scene with an obviously patched hole in it.

Why Netflix Is Open-Sourcing It

Netflix's decision to open-source VOID follows a pattern the company has established with tools like Hollow (HDR tone mapping) and several of its audio processing frameworks. The strategic logic is straightforward: Netflix is primarily a content delivery and production company, not an AI tools company. Releasing non-core technology as open source builds ecosystem goodwill, attracts research talent, and invites external improvements that Netflix can then incorporate. The company's internal use cases — removing logos, brand marks, location-specific text from international content — don't require VOID to be proprietary.

Implications for the Industry

VOID's open-source release dramatically lowers the barrier for post-production workflows that previously required VFX artists to manually rotoscope and reconstruct removed elements frame by frame. The implications extend from streaming localization to advertising (swapping products in existing footage), sports broadcasting (removing trackside advertising for market-specific feeds), and archival restoration. For independent filmmakers, it provides VFX capabilities that were previously accessible only to studios with dedicated effects pipelines.

The research paper accompanying the release notes performance benchmarks that substantially outperform prior approaches on the established DAVIS and YouTube-VOS evaluation datasets. Netflix has also released a reference implementation and pre-trained model weights, making the system immediately usable rather than merely publishable.